Amend patch handling in ps3

[oweals/openwrt.git] / target / linux / ps3 / patches-2.6.27 / 001-perfmon-2.6.27.patch

diff --git a/Documentation/ABI/testing/sysfs-perfmon b/Documentation/ABI/testing/sysfs-perfmon
new file mode 100644
index 0000000..bde434c
--- /dev/null
+++ b/Documentation/ABI/testing/sysfs-perfmon
@@ -0,0 +1,87 @@
+What:          /sys/kernel/perfmon
+Date:          Nov 2007
+KernelVersion: 2.6.24
+Contact:       eranian@gmail.com
+
+Description:   provide the configuration interface for the perfmon2 subsystems.
+               The tree contains information about the detected hardware, current
+               state of the subsystem as well as some configuration parameters.
+       
+               The tree consists of the following entries:
+
+       /sys/kernel/perfmon/debug (read-write):
+
+               Enable perfmon2 debugging output via klogd. Debug messages produced during
+               PMU interrupt handling are not controlled by this entry. The traces a rate-limited
+               to avoid flooding of the console. It is possible to change the throttling
+               via /proc/sys/kernel/printk_ratelimit. The value is interpreted as a bitmask.
+               Each bit enables a particular type of debug messages. Refer to the file
+               include/linux/perfmon_kern.h for more information
+
+       /sys/kernel/perfmon/pmc_max_fast_arg (read-only):
+
+               Number of perfmon2 syscall arguments copied directly onto the
+               stack (copy_from_user) for pfm_write_pmcs(). Copying to the stack avoids
+               having to allocate a buffer. The unit is the number of pfarg_pmc_t
+               structures.
+
+       /sys/kernel/perfmon/pmd_max_fast_arg (read-only):
+
+               Number of perfmon2 syscall arguments copied directly onto the
+               stack (copy_from_user) for pfm_write_pmds()/pfm_read_pmds(). Copying
+               to the stack avoids having to allocate a buffer. The unit is the number
+               of pfarg_pmd_t structures.
+
+
+       /sys/kernel/perfmon/reset_stats (write-only):
+
+               Reset the statistics collected by perfmon2. Stats are available
+               per-cpu via debugfs.
+       
+       /sys/kernel/perfmon/smpl_buffer_mem_cur (read-only):
+
+               Reports the amount of memory currently dedicated to sampling
+               buffers by the kernel. The unit is byte.
+
+       /sys/kernel/perfmon/smpl_buffer_mem_max (read-write):
+
+               Maximum amount of kernel memory usable for sampling buffers. -1 means
+               everything that is available. Unit is byte.
+
+       /sys/kernel/perfmon/smpl_buffer_mem_cur (read-only):
+
+               Current utilization of kernel memory in bytes.
+
+       /sys/kernel/perfmon/sys_group (read-write):
+
+               Users group allowed to create a system-wide perfmon2 context (session).
+               -1 means any group. This control will be kept until we find a package
+               able to control capabilities via PAM.
+
+       /sys/kernel/perfmon/task_group (read-write):
+
+               Users group allowed to create a per-thread context (session).
+               -1 means any group. This control will be kept until we find a
+               package able to control capabilities via PAM.
+
+       /sys/kernel/perfmon/sys_sessions_count (read-only):
+
+               Number of system-wide contexts currently attached to CPUs.
+
+       /sys/kernel/perfmon/task_sessions_count (read-only):
+
+               Number of per-thread contexts currently attached to threads.
+
+       /sys/kernel/perfmon/version (read-only):
+
+               Perfmon2 interface revision number.
+
+       /sys/kernel/perfmon/arg_mem_max(read-write):
+
+               Maximum size of vector arguments expressed in bytes. Can be modified
+
+       /sys/kernel/perfmon/mode(read-write):
+
+               Bitmask to enable/disable certain perfmon2 features.
+               Currently defined:
+               - bit 0: if set, then reserved bitfield are ignored on PMC writes
diff --git a/Documentation/ABI/testing/sysfs-perfmon-fmt b/Documentation/ABI/testing/sysfs-perfmon-fmt
new file mode 100644
index 0000000..1b45270
--- /dev/null
+++ b/Documentation/ABI/testing/sysfs-perfmon-fmt
@@ -0,0 +1,18 @@
+What:          /sys/kernel/perfmon/formats
+Date:          2007
+KernelVersion: 2.6.24
+Contact:       eranian@gmail.com
+
+Description:   provide description of available perfmon2 custom sampling buffer formats
+               which are implemented as independent kernel modules. Each formats gets
+               a subdir which a few entries.
+
+               The name of the subdir is the name of the sampling format. The same name
+               must be passed to pfm_create_context() to use the format.
+
+               Each subdir XX contains the following entries:
+
+       /sys/kernel/perfmon/formats/XX/version (read-only):
+
+               Version number of the format in clear text and null terminated.
+
diff --git a/Documentation/ABI/testing/sysfs-perfmon-pmu b/Documentation/ABI/testing/sysfs-perfmon-pmu
new file mode 100644
index 0000000..a1afc7e
--- /dev/null
+++ b/Documentation/ABI/testing/sysfs-perfmon-pmu
@@ -0,0 +1,46 @@
+What:          /sys/kernel/perfmon/pmu
+Date:          Nov 2007
+KernelVersion: 2.6.24
+Contact:       eranian@gmail.com
+
+Description:   provide information about the currently loaded PMU description module.
+               The module contains the mapping of the actual performance counter registers
+               onto the logical PMU exposed by perfmon.  There is at most one PMU description
+               module loaded at any time.
+
+               The sysfs PMU tree provides a description of the mapping for each register.
+               There is one subdir per config and data registers along an entry for the
+               name of the PMU model.
+
+               The model entry is as follows:
+
+       /sys/kernel/perfmon/pmu_desc/model (read-only):
+
+               Name of the PMU model is clear text and zero terminated.
+       
+               Then for each logical PMU register, XX, gets a subtree with the following entries:
+
+       /sys/kernel/perfmon/pmu_desc/pm*XX/addr (read-only):
+
+               The physical address or index of the actual underlying hardware register.
+               On Itanium, it corresponds to the index. But on X86 processor, this is
+               the actual MSR address.
+
+       /sys/kernel/perfmon/pmu_desc/pm*XX/dfl_val (read-only):
+
+               The default value of the register in hexadecimal.
+
+       /sys/kernel/perfmon/pmu_desc/pm*XX/name (read-only):
+
+               The name of the hardware register.
+
+       /sys/kernel/perfmon/pmu_desc/pm*XX/rsvd_msk (read-only):
+
+               The bitmask of reserved bits, i.e., bits which cannot be changed by
+               applications. When a bit is set, it means the corresponding bit in the
+               actual register is reserved.
+
+       /sys/kernel/perfmon/pmu_desc/pm*XX/width (read-only):
+
+               the width in bits of the registers. This field is only relevant for counter
+               registers.
diff --git a/Documentation/kernel-parameters.txt b/Documentation/kernel-parameters.txt
index 1150444..2652b6c 100644
--- a/Documentation/kernel-parameters.txt
+++ b/Documentation/kernel-parameters.txt
@@ -1643,6 +1643,9 @@ and is between 256 and 4096 characters. It is defined in the file
                        Format: { 0 | 1 }
                        See arch/parisc/kernel/pdc_chassis.c
 
+       perfmon_debug   [PERFMON] Enables Perfmon debug messages. Needed
+                       to see traces of the early startup startup phase.
+
        pf.             [PARIDE]
                        See Documentation/paride.txt.
 
diff --git a/Documentation/perfmon2-debugfs.txt b/Documentation/perfmon2-debugfs.txt
new file mode 100644
index 0000000..b30cae8
--- /dev/null
+++ b/Documentation/perfmon2-debugfs.txt
@@ -0,0 +1,126 @@
+               The perfmon2 debug and statistics interface
+                ------------------------------------------
+                          Stephane Eranian
+                         <eranian@gmail.com>
+
+The perfmon2 interfaces exports a set of statistics which are used to tune and
+debug the implementation. The data is composed of a set of very simple metrics
+mostly aggregated counts and durations. They instruments key points in the
+perfmon2 code, such as context switch and interrupt handling.
+
+The data is accessible via the debug filesystem (debugfs). Thus you need to
+have the filesystem support enabled in your kernel. Furthermore since, 2.6.25,
+the perfmon2 statistics interface is an optional component. It needs to be
+explicitely enabled in the kernel config file (CONFIG_PERFMON_DEBUG_FS).
+
+To access the data, the debugs filesystem must be mounted. Supposing the mount
+point is /debugfs, you would need to do:
+       $ mount -t debugs none /debugfs
+
+The data is located under the perfmon subdirectory and is organized per CPU.
+For each CPU, the same set of metrics is available, one metric per file in
+clear ASCII text.
+
+The metrics are as follows:
+
+       ctxswin_count (read-only):
+       
+               Number of PMU context switch in.
+
+       ctxswin_ns (read-only):
+
+               Number of nanoseconds spent in the PMU context switch in
+               routine.  Dividing this number by the value of ctxswin_count,
+               yields average cost of the PMU context switch in.
+
+       ctxswout_count (read-only):
+
+               Number of PMU context switch out.
+
+       ctxswout_ns (read-only):
+
+               Number of nanoseconds spent in the PMU context switch in
+               routine. Dividing this number by the value of ctxswout_count,
+               yields average cost of the PMU context switch out.
+
+       fmt_handler_calls (read-only):
+
+               Number of calls to the sampling format routine that handles
+               PMU interrupts, i.e., typically the routine that records a
+               sample.
+
+       fmt_handler_ns (read-only):
+
+               Number of nanoseconds spent in the routine that handle PMU
+               interrupt in the sampling format. Dividing this number by
+               the number of calls provided by fmt_handler_calls, yields
+               average time spent in this routine.
+
+       ovfl_intr_all_count (read-only):
+
+               Number of PMU interrupts received by the kernel.
+
+
+       ovfl_intr_nmi_count (read-only):
+
+               Number of Non Maskeable Interrupts (NMI) received by the kernel
+               for perfmon. This is relevant only on X86 hardware.
+       
+       ovfl_intr_ns (read-only):
+
+               Number of nanoseconds spent in the perfmon2 PMU interrupt
+               handler routine. Dividing this number of ovfl_intr_all_count
+               yields the average time to handle one PMU interrupt.
+
+       ovfl_intr_regular_count (read-only):
+
+               Number of PMU interrupts which are actually processed by 
+               the perfmon interrupt handler. There may be spurious or replay
+               interrupts.
+
+       ovfl_intr_replay_count (read-only):
+
+               Number of PMU interrupts which were replayed on context switch  
+               in or on event set switching. Interrupts get replayed when they
+               were in flight at the time monitoring had to be stopped.
+
+       perfmon/ovfl_intr_spurious_count (read-only):
+
+               Number of PMU interrupts which were dropped because there was
+               no active context (session).
+
+       ovfl_notify_count (read-only):
+
+               Number of user level notifications sent. Notifications are
+               appended as messages to the context queue. Notifications may
+               be sent on PMU interrupts.
+
+       pfm_restart_count (read-only):
+
+               Number of times pfm_restart() is called.
+
+       reset_pmds_count (read-only):
+
+               Number of times pfm_reset_pmds() is called.
+
+       set_switch_count (read-only):
+
+               Number of event set switches.
+
+       set_switch_ns (read-only):
+
+               Number of nanoseconds spent in the set switching routine.
+               Dividing this number by set_switch_count yields the average
+               cost of switching sets.
+
+       handle_timeout_count (read-only):
+
+               Number of times the pfm_handle_timeout() routine is called.
+               It is used for timeout-based set switching.
+
+       handle_work_count (read-only):
+
+               Number of times pfm_handle_work() is called. The routine
+               handles asynchronous perfmon2 work for per-thread contexts
+               (sessions).
+
diff --git a/Documentation/perfmon2.txt b/Documentation/perfmon2.txt
new file mode 100644
index 0000000..4a8fada
--- /dev/null
+++ b/Documentation/perfmon2.txt
@@ -0,0 +1,213 @@
+              The perfmon2 hardware monitoring interface
+              ------------------------------------------
+                          Stephane Eranian
+                         <eranian@gmail.com>
+
+I/ Introduction
+
+   The perfmon2 interface provides access to the hardware performance counters of
+   major processors. Nowadays, all processors implement some flavors of performance
+   counters which capture micro-architectural level information such as the number
+   of elapsed cycles, number of cache misses, and so on.
+
+   The interface is implemented as a set of new system calls and a set of config files
+   in /sys.
+
+   It is possible to monitoring a single thread or a CPU. In either mode, applications
+   can count or collect samples. System-wide monitoring is supported by running a
+   monitoring session on each CPU. The interface support event-based sampling where the
+   sampling period is expressed as the number of occurrences of event, instead of just a
+   timeout.  This approach provides a much better granularity and flexibility.
+
+   For performance reason, it is possible to use a kernel-level sampling buffer to minimize
+   the overhead incurred by sampling. The format of the buffer, i.e., what is recorded, how
+   it is recorded, and how it is exported to user-land is controlled by a kernel module called
+   a custom sampling format. The current implementation comes with a default format but
+   it is possible to create additional formats. There is an in-kernel registration
+   interface for formats. Each format is identified by a simple string which a tool
+   can pass when a monitoring session is created.
+
+   The interface also provides support for event set and multiplexing to work around
+   hardware limitations in the number of available counters or in how events can be 
+   combined. Each set defines as many counters as the hardware can support. The kernel
+   then multiplexes the sets. The interface supports time-base switching but also
+   overflow based switching, i.e., after n overflows of designated counters.
+
+   Applications never manipulates the actual performance counter registers. Instead they see
+   a logical Performance Monitoring Unit (PMU) composed of a set of config register (PMC)
+   and a set of data registers (PMD). Note that PMD are not necessarily counters, they
+   can be buffers. The logical PMU is then mapped onto the actual PMU using a mapping
+   table which is implemented as a kernel module. The mapping is chosen once for each
+   new processor. It is visible in /sys/kernel/perfmon/pmu_desc. The kernel module
+   is automatically loaded on first use.
+
+   A monitoring session, or context, is uniquely identified by a file descriptor
+   obtained when the context is created. File sharing semantics apply to access
+   the context inside a process. A context is never inherited across fork. The file
+   descriptor can be used to received counter overflow notifications or when the
+   sampling buffer is full. It is possible to use poll/select on the descriptor 
+   to wait for notifications from multiplex contexts. Similarly, the descriptor
+   supports asynchronous notification via SIGIO.
+
+   Counters are always exported as being 64-bit wide regardless of what the underlying
+   hardware implements.
+
+II/ Kernel compilation
+
+    To enable perfmon2, you need to enable CONFIG_PERFMON
+
+III/ OProfile interactions
+
+    The set of features offered by perfmon2 is rich enough to support migrating
+    Oprofile on top of it. That means that PMU programming and low-level interrupt
+    handling could be done by perfmon2. The Oprofile sampling buffer management code
+    in the kernel as well as how samples are exported to users could remain through
+    the use of a custom sampling buffer format. This is how Oprofile work on Itanium.
+
+    The current interactions with Oprofile are:
+       - on X86: Both subsystems can be compiled into the same kernel. There is enforced
+                 mutual exclusion between the two subsystems. When there is an Oprofile
+                 session, no perfmon2 session can exist and vice-versa. Perfmon2 session
+                 encapsulates both per-thread and system-wide sessions here.
+
+       - On IA-64: Oprofile works on top of perfmon2. Oprofile being a system-wide monitoring
+                   tool, the regular per-thread vs. system-wide session restrictions apply.
+
+       - on PPC: no integration yet. You need to enable/disble one of the two subsystems
+       - on MIPS: no integration yet. You need to enable/disble one of the two subsystems
+
+IV/ User tools
+
+    We have released a simple monitoring tool to demonstrate the feature of the
+    interface. The tool is called pfmon and it comes with a simple helper library
+    called libpfm. The library comes with a set of examples to show how to use the
+    kernel perfmon2 interface. Visit http://perfmon2.sf.net for details.
+
+    There maybe other tools available for perfmon2.
+
+V/ How to program?
+
+   The best way to learn how to program perfmon2, is to take a look at the source
+   code for the examples in libpfm. The source code is available from:
+               http://perfmon2.sf.net
+
+VI/ System calls overview
+
+   The interface is implemented by the following system calls:
+
+   * int pfm_create_context(pfarg_ctx_t *ctx, char *fmt, void *arg, size_t arg_size)
+
+      This function create a perfmon2 context. The type of context is per-thread by
+      default unless PFM_FL_SYSTEM_WIDE is passed in ctx. The sampling format name
+      is passed in fmt. Arguments to the format are passed in arg which is of size
+      arg_size. Upon successful return, the file descriptor identifying the context
+      is returned.
+
+   * int pfm_write_pmds(int fd, pfarg_pmd_t *pmds, int n)
+   
+      This function is used to program the PMD registers. It is possible to pass
+      vectors of PMDs.
+
+   * int pfm_write_pmcs(int fd, pfarg_pmc_t *pmds, int n)
+   
+      This function is used to program the PMC registers. It is possible to pass
+      vectors of PMDs.
+
+   * int pfm_read_pmds(int fd, pfarg_pmd_t *pmds, int n)
+   
+      This function is used to read the PMD registers. It is possible to pass
+      vectors of PMDs.
+
+   * int pfm_load_context(int fd, pfarg_load_t *load)
+   
+      This function is used to attach the context to a thread or CPU.
+      Thread means kernel-visible thread (NPTL). The thread identification
+      as obtained by gettid must be passed to load->load_target.
+
+      To operate on another thread (not self), it is mandatory that the thread
+      be stopped via ptrace().
+
+      To attach to a CPU, the CPU number must be specified in load->load_target
+      AND the call must be issued on that CPU. To monitor a CPU, a thread MUST
+      be pinned on that CPU.
+
+      Until the context is attached, the actual counters are not accessed.
+
+   * int pfm_unload_context(int fd)
+
+     The context is detached for the thread or CPU is was attached to.
+     As a consequence monitoring is stopped.
+
+     When monitoring another thread, the thread MUST be stopped via ptrace()
+     for this function to succeed.
+
+   * int pfm_start(int fd, pfarg_start_t *st)
+
+     Start monitoring. The context must be attached for this function to succeed.
+     Optionally, it is possible to specify the event set on which to start using the
+     st argument, otherwise just pass NULL.
+
+     When monitoring another thread, the thread MUST be stopped via ptrace()
+     for this function to succeed.
+
+   * int pfm_stop(int fd)
+
+     Stop monitoring. The context must be attached for this function to succeed.
+
+     When monitoring another thread, the thread MUST be stopped via ptrace()
+     for this function to succeed.
+
+
+   * int pfm_create_evtsets(int fd, pfarg_setdesc_t *sets, int n)
+
+     This function is used to create or change event sets. By default set 0 exists.
+     It is possible to create/change multiple sets in one call.
+
+     The context must be detached for this call to succeed.
+
+     Sets are identified by a 16-bit integer. They are sorted based on this
+     set and switching occurs in a round-robin fashion.
+
+   * int pfm_delete_evtsets(int fd, pfarg_setdesc_t *sets, int n)
+   
+     Delete event sets. The context must be detached for this call to succeed.
+
+
+   * int pfm_getinfo_evtsets(int fd, pfarg_setinfo_t *sets, int n)
+
+     Retrieve information about event sets. In particular it is possible
+     to get the number of activation of a set. It is possible to retrieve
+     information about multiple sets in one call.
+
+
+   * int pfm_restart(int fd)
+
+     Indicate to the kernel that the application is done processing an overflow
+     notification. A consequence of this call could be that monitoring resumes.
+
+   * int read(fd, pfm_msg_t *msg, sizeof(pfm_msg_t))
+
+   the regular read() system call can be used with the context file descriptor to
+   receive overflow notification messages. Non-blocking read() is supported.
+
+   Each message carry information about the overflow such as which counter overflowed
+   and where the program was (interrupted instruction pointer).
+
+   * int close(int fd)
+
+   To destroy a context, the regular close() system call is used.
+
+
+VII/ /sys interface overview
+
+   Refer to Documentation/ABI/testing/sysfs-perfmon-* for a detailed description
+   of the sysfs interface of perfmon2.
+
+VIII/ debugfs interface overview
+
+  Refer to Documentation/perfmon2-debugfs.txt for a detailed description of the
+  debug and statistics interface of perfmon2.
+
+IX/ Documentation
+
+   Visit http://perfmon2.sf.net
diff --git a/MAINTAINERS b/MAINTAINERS
index 8dae455..fb38c2a 100644
--- a/MAINTAINERS
+++ b/MAINTAINERS
@@ -3239,6 +3239,14 @@ M:       balbir@linux.vnet.ibm.com
 L:     linux-kernel@vger.kernel.org
 S:     Maintained
 
+PERFMON SUBSYSTEM
+P:     Stephane Eranian
+M:     eranian@gmail.com
+L:     perfmon2-devel@lists.sf.net
+W:     http://perfmon2.sf.net
+T:     git kernel.org:/pub/scm/linux/kernel/git/eranian/linux-2.6
+S:     Maintained
+
 PERSONALITY HANDLING
 P:     Christoph Hellwig
 M:     hch@infradead.org
diff --git a/Makefile b/Makefile
index 16e3fbb..7bb1320 100644
--- a/Makefile
+++ b/Makefile
@@ -620,6 +620,7 @@ export mod_strip_cmd
 
 ifeq ($(KBUILD_EXTMOD),)
 core-y         += kernel/ mm/ fs/ ipc/ security/ crypto/ block/
+core-$(CONFIG_PERFMON) += perfmon/
 
 vmlinux-dirs   := $(patsubst %/,%,$(filter %/, $(init-y) $(init-m) \
                     $(core-y) $(core-m) $(drivers-y) $(drivers-m) \
diff --git a/arch/ia64/Kconfig b/arch/ia64/Kconfig
index 48e496f..1d79b01 100644
--- a/arch/ia64/Kconfig
+++ b/arch/ia64/Kconfig
@@ -470,14 +470,6 @@ config COMPAT_FOR_U64_ALIGNMENT
 config IA64_MCA_RECOVERY
        tristate "MCA recovery from errors other than TLB."
 
-config PERFMON
-       bool "Performance monitor support"
-       help
-         Selects whether support for the IA-64 performance monitor hardware
-         is included in the kernel.  This makes some kernel data-structures a
-         little bigger and slows down execution a bit, but it is generally
-         a good idea to turn this on.  If you're unsure, say Y.
-
 config IA64_PALINFO
        tristate "/proc/pal support"
        help
@@ -549,6 +541,8 @@ source "drivers/firmware/Kconfig"
 
 source "fs/Kconfig.binfmt"
 
+source "arch/ia64/perfmon/Kconfig"
+
 endmenu
 
 menu "Power management and ACPI"
diff --git a/arch/ia64/Makefile b/arch/ia64/Makefile
index 905d25b..9aa622d 100644
--- a/arch/ia64/Makefile
+++ b/arch/ia64/Makefile
@@ -57,6 +57,7 @@ core-$(CONFIG_IA64_GENERIC)   += arch/ia64/dig/
 core-$(CONFIG_IA64_HP_ZX1)     += arch/ia64/dig/
 core-$(CONFIG_IA64_HP_ZX1_SWIOTLB) += arch/ia64/dig/
 core-$(CONFIG_IA64_SGI_SN2)    += arch/ia64/sn/
+core-$(CONFIG_PERFMON)         += arch/ia64/perfmon/
 core-$(CONFIG_IA64_SGI_UV)     += arch/ia64/uv/
 core-$(CONFIG_KVM)             += arch/ia64/kvm/
 
diff --git a/arch/ia64/configs/generic_defconfig b/arch/ia64/configs/generic_defconfig
index 9f48397..ff9572a 100644
--- a/arch/ia64/configs/generic_defconfig
+++ b/arch/ia64/configs/generic_defconfig
@@ -209,7 +209,6 @@ CONFIG_IA32_SUPPORT=y
 CONFIG_COMPAT=y
 CONFIG_COMPAT_FOR_U64_ALIGNMENT=y
 CONFIG_IA64_MCA_RECOVERY=y
-CONFIG_PERFMON=y
 CONFIG_IA64_PALINFO=y
 # CONFIG_IA64_MC_ERR_INJECT is not set
 CONFIG_SGI_SN=y
@@ -234,6 +233,16 @@ CONFIG_BINFMT_ELF=y
 CONFIG_BINFMT_MISC=m
 
 #
+# Hardware Performance Monitoring support
+#
+CONFIG_PERFMON=y
+CONFIG_IA64_PERFMON_COMPAT=y
+CONFIG_IA64_PERFMON_GENERIC=m
+CONFIG_IA64_PERFMON_ITANIUM=y
+CONFIG_IA64_PERFMON_MCKINLEY=y
+CONFIG_IA64_PERFMON_MONTECITO=y
+
+#
 # Power management and ACPI
 #
 CONFIG_PM=y
diff --git a/arch/ia64/include/asm/Kbuild b/arch/ia64/include/asm/Kbuild
index ccbe8ae..cf64b3b 100644
--- a/arch/ia64/include/asm/Kbuild
+++ b/arch/ia64/include/asm/Kbuild
@@ -5,10 +5,12 @@ header-y += fpu.h
 header-y += fpswa.h
 header-y += ia64regs.h
 header-y += intel_intrin.h
-header-y += perfmon_default_smpl.h
 header-y += ptrace_offsets.h
 header-y += rse.h
 header-y += ucontext.h
+header-y += perfmon.h
+header-y += perfmon_compat.h
+header-y += perfmon_default_smpl.h
 
 unifdef-y += gcc_intrin.h
 unifdef-y += intrinsics.h
diff --git a/arch/ia64/include/asm/hw_irq.h b/arch/ia64/include/asm/hw_irq.h
index 5c99cbc..4a45cb0 100644
--- a/arch/ia64/include/asm/hw_irq.h
+++ b/arch/ia64/include/asm/hw_irq.h
@@ -67,9 +67,9 @@ extern int ia64_last_device_vector;
 #define IA64_NUM_DEVICE_VECTORS                (IA64_LAST_DEVICE_VECTOR - IA64_FIRST_DEVICE_VECTOR + 1)
 
 #define IA64_MCA_RENDEZ_VECTOR         0xe8    /* MCA rendez interrupt */
-#define IA64_PERFMON_VECTOR            0xee    /* performance monitor interrupt vector */
 #define IA64_TIMER_VECTOR              0xef    /* use highest-prio group 15 interrupt for timer */
 #define        IA64_MCA_WAKEUP_VECTOR          0xf0    /* MCA wakeup (must be >MCA_RENDEZ_VECTOR) */
+#define IA64_PERFMON_VECTOR            0xf1    /* performance monitor interrupt vector */
 #define IA64_IPI_LOCAL_TLB_FLUSH       0xfc    /* SMP flush local TLB */
 #define IA64_IPI_RESCHEDULE            0xfd    /* SMP reschedule */
 #define IA64_IPI_VECTOR                        0xfe    /* inter-processor interrupt vector */
diff --git a/arch/ia64/include/asm/perfmon.h b/arch/ia64/include/asm/perfmon.h
index 7f3333d..150c4b4 100644
--- a/arch/ia64/include/asm/perfmon.h
+++ b/arch/ia64/include/asm/perfmon.h
@@ -1,279 +1,59 @@
 /*
- * Copyright (C) 2001-2003 Hewlett-Packard Co
- *               Stephane Eranian <eranian@hpl.hp.com>
- */
-
-#ifndef _ASM_IA64_PERFMON_H
-#define _ASM_IA64_PERFMON_H
-
-/*
- * perfmon comamnds supported on all CPU models
- */
-#define PFM_WRITE_PMCS         0x01
-#define PFM_WRITE_PMDS         0x02
-#define PFM_READ_PMDS          0x03
-#define PFM_STOP               0x04
-#define PFM_START              0x05
-#define PFM_ENABLE             0x06 /* obsolete */
-#define PFM_DISABLE            0x07 /* obsolete */
-#define PFM_CREATE_CONTEXT     0x08
-#define PFM_DESTROY_CONTEXT    0x09 /* obsolete use close() */
-#define PFM_RESTART            0x0a
-#define PFM_PROTECT_CONTEXT    0x0b /* obsolete */
-#define PFM_GET_FEATURES       0x0c
-#define PFM_DEBUG              0x0d
-#define PFM_UNPROTECT_CONTEXT  0x0e /* obsolete */
-#define PFM_GET_PMC_RESET_VAL  0x0f
-#define PFM_LOAD_CONTEXT       0x10
-#define PFM_UNLOAD_CONTEXT     0x11
-
-/*
- * PMU model specific commands (may not be supported on all PMU models)
- */
-#define PFM_WRITE_IBRS         0x20
-#define PFM_WRITE_DBRS         0x21
-
-/*
- * context flags
- */
-#define PFM_FL_NOTIFY_BLOCK             0x01   /* block task on user level notifications */
-#define PFM_FL_SYSTEM_WIDE      0x02   /* create a system wide context */
-#define PFM_FL_OVFL_NO_MSG      0x80   /* do not post overflow/end messages for notification */
-
-/*
- * event set flags
- */
-#define PFM_SETFL_EXCL_IDLE      0x01   /* exclude idle task (syswide only) XXX: DO NOT USE YET */
-
-/*
- * PMC flags
- */
-#define PFM_REGFL_OVFL_NOTIFY  0x1     /* send notification on overflow */
-#define PFM_REGFL_RANDOM       0x2     /* randomize sampling interval   */
-
-/*
- * PMD/PMC/IBR/DBR return flags (ignored on input)
+ * Copyright (c) 2001-2007 Hewlett-Packard Development Company, L.P.
+ * Contributed by Stephane Eranian <eranian@hpl.hp.com>
  *
- * Those flags are used on output and must be checked in case EAGAIN is returned
- * by any of the calls using a pfarg_reg_t or pfarg_dbreg_t structure.
- */
-#define PFM_REG_RETFL_NOTAVAIL (1UL<<31) /* set if register is implemented but not available */
-#define PFM_REG_RETFL_EINVAL   (1UL<<30) /* set if register entry is invalid */
-#define PFM_REG_RETFL_MASK     (PFM_REG_RETFL_NOTAVAIL|PFM_REG_RETFL_EINVAL)
-
-#define PFM_REG_HAS_ERROR(flag)        (((flag) & PFM_REG_RETFL_MASK) != 0)
-
-typedef unsigned char pfm_uuid_t[16];  /* custom sampling buffer identifier type */
-
-/*
- * Request structure used to define a context
- */
-typedef struct {
-       pfm_uuid_t     ctx_smpl_buf_id;  /* which buffer format to use (if needed) */
-       unsigned long  ctx_flags;        /* noblock/block */
-       unsigned short ctx_nextra_sets;  /* number of extra event sets (you always get 1) */
-       unsigned short ctx_reserved1;    /* for future use */
-       int            ctx_fd;           /* return arg: unique identification for context */
-       void           *ctx_smpl_vaddr;  /* return arg: virtual address of sampling buffer, is used */
-       unsigned long  ctx_reserved2[11];/* for future use */
-} pfarg_context_t;
-
-/*
- * Request structure used to write/read a PMC or PMD
- */
-typedef struct {
-       unsigned int    reg_num;           /* which register */
-       unsigned short  reg_set;           /* event set for this register */
-       unsigned short  reg_reserved1;     /* for future use */
-
-       unsigned long   reg_value;         /* initial pmc/pmd value */
-       unsigned long   reg_flags;         /* input: pmc/pmd flags, return: reg error */
-
-       unsigned long   reg_long_reset;    /* reset after buffer overflow notification */
-       unsigned long   reg_short_reset;   /* reset after counter overflow */
-
-       unsigned long   reg_reset_pmds[4]; /* which other counters to reset on overflow */
-       unsigned long   reg_random_seed;   /* seed value when randomization is used */
-       unsigned long   reg_random_mask;   /* bitmask used to limit random value */
-       unsigned long   reg_last_reset_val;/* return: PMD last reset value */
-
-       unsigned long   reg_smpl_pmds[4];  /* which pmds are accessed when PMC overflows */
-       unsigned long   reg_smpl_eventid;  /* opaque sampling event identifier */
-
-       unsigned long   reg_reserved2[3];   /* for future use */
-} pfarg_reg_t;
-
-typedef struct {
-       unsigned int    dbreg_num;              /* which debug register */
-       unsigned short  dbreg_set;              /* event set for this register */
-       unsigned short  dbreg_reserved1;        /* for future use */
-       unsigned long   dbreg_value;            /* value for debug register */
-       unsigned long   dbreg_flags;            /* return: dbreg error */
-       unsigned long   dbreg_reserved2[1];     /* for future use */
-} pfarg_dbreg_t;
-
-typedef struct {
-       unsigned int    ft_version;     /* perfmon: major [16-31], minor [0-15] */
-       unsigned int    ft_reserved;    /* reserved for future use */
-       unsigned long   reserved[4];    /* for future use */
-} pfarg_features_t;
-
-typedef struct {
-       pid_t           load_pid;          /* process to load the context into */
-       unsigned short  load_set;          /* first event set to load */
-       unsigned short  load_reserved1;    /* for future use */
-       unsigned long   load_reserved2[3]; /* for future use */
-} pfarg_load_t;
-
-typedef struct {
-       int             msg_type;               /* generic message header */
-       int             msg_ctx_fd;             /* generic message header */
-       unsigned long   msg_ovfl_pmds[4];       /* which PMDs overflowed */
-       unsigned short  msg_active_set;         /* active set at the time of overflow */
-       unsigned short  msg_reserved1;          /* for future use */
-       unsigned int    msg_reserved2;          /* for future use */
-       unsigned long   msg_tstamp;             /* for perf tuning/debug */
-} pfm_ovfl_msg_t;
-
-typedef struct {
-       int             msg_type;               /* generic message header */
-       int             msg_ctx_fd;             /* generic message header */
-       unsigned long   msg_tstamp;             /* for perf tuning */
-} pfm_end_msg_t;
-
-typedef struct {
-       int             msg_type;               /* type of the message */
-       int             msg_ctx_fd;             /* unique identifier for the context */
-       unsigned long   msg_tstamp;             /* for perf tuning */
-} pfm_gen_msg_t;
-
-#define PFM_MSG_OVFL   1       /* an overflow happened */
-#define PFM_MSG_END    2       /* task to which context was attached ended */
-
-typedef union {
-       pfm_ovfl_msg_t  pfm_ovfl_msg;
-       pfm_end_msg_t   pfm_end_msg;
-       pfm_gen_msg_t   pfm_gen_msg;
-} pfm_msg_t;
-
-/*
- * Define the version numbers for both perfmon as a whole and the sampling buffer format.
+ * This file contains Itanium Processor Family specific definitions
+ * for the perfmon interface.
+ *
+ * This program is free software; you can redistribute it and/or
+ * modify it under the terms of version 2 of the GNU General Public
+ * License as published by the Free Software Foundation.
+ *
+ * This program is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+ * General Public License for more details.
+ *
+ * You should have received a copy of the GNU General Public License
+ * along with this program; if not, write to the Free Software
+ * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA
+ * 02111-1307 USA
  */
-#define PFM_VERSION_MAJ                 2U
-#define PFM_VERSION_MIN                 0U
-#define PFM_VERSION             (((PFM_VERSION_MAJ&0xffff)<<16)|(PFM_VERSION_MIN & 0xffff))
-#define PFM_VERSION_MAJOR(x)    (((x)>>16) & 0xffff)
-#define PFM_VERSION_MINOR(x)    ((x) & 0xffff)
-
+#ifndef _ASM_IA64_PERFMON_H_
+#define _ASM_IA64_PERFMON_H_
 
 /*
- * miscellaneous architected definitions
+ * arch-specific user visible interface definitions
  */
-#define PMU_FIRST_COUNTER      4       /* first counting monitor (PMC/PMD) */
-#define PMU_MAX_PMCS           256     /* maximum architected number of PMC registers */
-#define PMU_MAX_PMDS           256     /* maximum architected number of PMD registers */
-
-#ifdef __KERNEL__
-
-extern long perfmonctl(int fd, int cmd, void *arg, int narg);
-
-typedef struct {
-       void (*handler)(int irq, void *arg, struct pt_regs *regs);
-} pfm_intr_handler_desc_t;
-
-extern void pfm_save_regs (struct task_struct *);
-extern void pfm_load_regs (struct task_struct *);
 
-extern void pfm_exit_thread(struct task_struct *);
-extern int  pfm_use_debug_registers(struct task_struct *);
-extern int  pfm_release_debug_registers(struct task_struct *);
-extern void pfm_syst_wide_update_task(struct task_struct *, unsigned long info, int is_ctxswin);
-extern void pfm_inherit(struct task_struct *task, struct pt_regs *regs);
-extern void pfm_init_percpu(void);
-extern void pfm_handle_work(void);
-extern int  pfm_install_alt_pmu_interrupt(pfm_intr_handler_desc_t *h);
-extern int  pfm_remove_alt_pmu_interrupt(pfm_intr_handler_desc_t *h);
+#define PFM_ARCH_MAX_PMCS      (256+64)
+#define PFM_ARCH_MAX_PMDS      (256+64)
 
-
-
-/*
- * Reset PMD register flags
- */
-#define PFM_PMD_SHORT_RESET    0
-#define PFM_PMD_LONG_RESET     1
-
-typedef union {
-       unsigned int val;
-       struct {
-               unsigned int notify_user:1;     /* notify user program of overflow */
-               unsigned int reset_ovfl_pmds:1; /* reset overflowed PMDs */
-               unsigned int block_task:1;      /* block monitored task on kernel exit */
-               unsigned int mask_monitoring:1; /* mask monitors via PMCx.plm */
-               unsigned int reserved:28;       /* for future use */
-       } bits;
-} pfm_ovfl_ctrl_t;
-
-typedef struct {
-       unsigned char   ovfl_pmd;                       /* index of overflowed PMD  */
-       unsigned char   ovfl_notify;                    /* =1 if monitor requested overflow notification */
-       unsigned short  active_set;                     /* event set active at the time of the overflow */
-       pfm_ovfl_ctrl_t ovfl_ctrl;                      /* return: perfmon controls to set by handler */
-
-       unsigned long   pmd_last_reset;                 /* last reset value of of the PMD */
-       unsigned long   smpl_pmds[4];                   /* bitmask of other PMD of interest on overflow */
-       unsigned long   smpl_pmds_values[PMU_MAX_PMDS]; /* values for the other PMDs of interest */
-       unsigned long   pmd_value;                      /* current 64-bit value of the PMD */
-       unsigned long   pmd_eventid;                    /* eventid associated with PMD */
-} pfm_ovfl_arg_t;
-
-
-typedef struct {
-       char            *fmt_name;
-       pfm_uuid_t      fmt_uuid;
-       size_t          fmt_arg_size;
-       unsigned long   fmt_flags;
-
-       int             (*fmt_validate)(struct task_struct *task, unsigned int flags, int cpu, void *arg);
-       int             (*fmt_getsize)(struct task_struct *task, unsigned int flags, int cpu, void *arg, unsigned long *size);
-       int             (*fmt_init)(struct task_struct *task, void *buf, unsigned int flags, int cpu, void *arg);
-       int             (*fmt_handler)(struct task_struct *task, void *buf, pfm_ovfl_arg_t *arg, struct pt_regs *regs, unsigned long stamp);
-       int             (*fmt_restart)(struct task_struct *task, pfm_ovfl_ctrl_t *ctrl, void *buf, struct pt_regs *regs);
-       int             (*fmt_restart_active)(struct task_struct *task, pfm_ovfl_ctrl_t *ctrl, void *buf, struct pt_regs *regs);
-       int             (*fmt_exit)(struct task_struct *task, void *buf, struct pt_regs *regs);
-
-       struct list_head fmt_list;
-} pfm_buffer_fmt_t;
-
-extern int pfm_register_buffer_fmt(pfm_buffer_fmt_t *fmt);
-extern int pfm_unregister_buffer_fmt(pfm_uuid_t uuid);
+#define PFM_ARCH_PMD_STK_ARG   8
+#define PFM_ARCH_PMC_STK_ARG   8
 
 /*
- * perfmon interface exported to modules
+ * Itanium specific context flags
+ *
+ * bits[00-15]: generic flags (see asm/perfmon.h)
+ * bits[16-31]: arch-specific flags
  */
-extern int pfm_mod_read_pmds(struct task_struct *, void *req, unsigned int nreq, struct pt_regs *regs);
-extern int pfm_mod_write_pmcs(struct task_struct *, void *req, unsigned int nreq, struct pt_regs *regs);
-extern int pfm_mod_write_ibrs(struct task_struct *task, void *req, unsigned int nreq, struct pt_regs *regs);
-extern int pfm_mod_write_dbrs(struct task_struct *task, void *req, unsigned int nreq, struct pt_regs *regs);
+#define PFM_ITA_FL_INSECURE 0x10000 /* clear psr.sp on non system, non self */
 
 /*
- * describe the content of the local_cpu_date->pfm_syst_info field
+ * Itanium specific public event set flags (set_flags)
+ *
+ * event set flags layout:
+ * bits[00-15] : generic flags
+ * bits[16-31] : arch-specific flags
  */
-#define PFM_CPUINFO_SYST_WIDE  0x1     /* if set a system wide session exists */
-#define PFM_CPUINFO_DCR_PP     0x2     /* if set the system wide session has started */
-#define PFM_CPUINFO_EXCL_IDLE  0x4     /* the system wide session excludes the idle task */
+#define PFM_ITA_SETFL_EXCL_INTR        0x10000  /* exclude interrupt execution */
+#define PFM_ITA_SETFL_INTR_ONLY        0x20000  /* include only interrupt execution */
+#define PFM_ITA_SETFL_IDLE_EXCL 0x40000  /* stop monitoring in idle loop */
 
 /*
- * sysctl control structure. visible to sampling formats
+ * compatibility for version v2.0 of the interface
  */
-typedef struct {
-       int     debug;          /* turn on/off debugging via syslog */
-       int     debug_ovfl;     /* turn on/off debug printk in overflow handler */
-       int     fastctxsw;      /* turn on/off fast (unsecure) ctxsw */
-       int     expert_mode;    /* turn on/off value checking */
-} pfm_sysctl_t;
-extern pfm_sysctl_t pfm_sysctl;
-
-
-#endif /* __KERNEL__ */
+#include <asm/perfmon_compat.h>
 
-#endif /* _ASM_IA64_PERFMON_H */
+#endif /* _ASM_IA64_PERFMON_H_ */
diff --git a/arch/ia64/include/asm/perfmon_compat.h b/arch/ia64/include/asm/perfmon_compat.h
new file mode 100644
index 0000000..5c14514
--- /dev/null
+++ b/arch/ia64/include/asm/perfmon_compat.h
@@ -0,0 +1,167 @@
+/*
+ * Copyright (c) 2001-2006 Hewlett-Packard Development Company, L.P.
+ * Contributed by Stephane Eranian <eranian@hpl.hp.com>
+ *
+ * This header file contains perfmon interface definition
+ * that are now obsolete and should be dropped in favor
+ * of their equivalent functions as explained below.
+ *
+ * This program is free software; you can redistribute it and/or
+ * modify it under the terms of version 2 of the GNU General Public
+ * License as published by the Free Software Foundation.
+ *
+ * This program is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+ * General Public License for more details.
+ *
+ * You should have received a copy of the GNU General Public License
+ * along with this program; if not, write to the Free Software
+ * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA
+ * 02111-1307 USA
+  */
+
+#ifndef _ASM_IA64_PERFMON_COMPAT_H_
+#define _ASM_IA64_PERFMON_COMPAT_H_
+
+/*
+ * custom sampling buffer identifier type
+ */
+typedef __u8 pfm_uuid_t[16];
+
+/*
+ * obsolete perfmon commands. Supported only on IA-64 for
+ * backward compatiblity reasons with perfmon v2.0.
+ */
+#define PFM_WRITE_PMCS         0x01 /* use pfm_write_pmcs */
+#define PFM_WRITE_PMDS         0x02 /* use pfm_write_pmds */
+#define PFM_READ_PMDS          0x03 /* use pfm_read_pmds */
+#define PFM_STOP               0x04 /* use pfm_stop */
+#define PFM_START              0x05 /* use pfm_start */
+#define PFM_ENABLE             0x06 /* obsolete */
+#define PFM_DISABLE            0x07 /* obsolete */
+#define PFM_CREATE_CONTEXT     0x08 /* use pfm_create_context */
+#define PFM_DESTROY_CONTEXT    0x09 /* use close() */
+#define PFM_RESTART            0x0a /* use pfm_restart */
+#define PFM_PROTECT_CONTEXT    0x0b /* obsolete */
+#define PFM_GET_FEATURES       0x0c /* use /proc/sys/perfmon */
+#define PFM_DEBUG              0x0d /* /proc/sys/kernel/perfmon/debug */
+#define PFM_UNPROTECT_CONTEXT  0x0e /* obsolete */
+#define PFM_GET_PMC_RESET_VAL  0x0f /* use /proc/perfmon_map */
+#define PFM_LOAD_CONTEXT       0x10 /* use pfm_load_context */
+#define PFM_UNLOAD_CONTEXT     0x11 /* use pfm_unload_context */
+
+/*
+ * PMU model specific commands (may not be supported on all PMU models)
+ */
+#define PFM_WRITE_IBRS         0x20 /* obsolete: use PFM_WRITE_PMCS[256-263]*/
+#define PFM_WRITE_DBRS         0x21 /* obsolete: use PFM_WRITE_PMCS[264-271]*/
+
+/*
+ * argument to PFM_CREATE_CONTEXT
+ */
+struct pfarg_context {
+       pfm_uuid_t     ctx_smpl_buf_id;  /* buffer format to use */
+       unsigned long  ctx_flags;        /* noblock/block */
+       unsigned int   ctx_reserved1;    /* for future use */
+       int            ctx_fd;           /* return: fildesc */
+       void           *ctx_smpl_vaddr;  /* return: vaddr of buffer */
+       unsigned long  ctx_reserved3[11];/* for future use */
+};
+
+/*
+ * argument structure for PFM_WRITE_PMCS/PFM_WRITE_PMDS/PFM_WRITE_PMDS
+ */
+struct pfarg_reg {
+       unsigned int    reg_num;           /* which register */
+       unsigned short  reg_set;           /* event set for this register */
+       unsigned short  reg_reserved1;     /* for future use */
+
+       unsigned long   reg_value;         /* initial pmc/pmd value */
+       unsigned long   reg_flags;         /* input: flags, ret: error */
+
+       unsigned long   reg_long_reset;    /* reset value after notification */
+       unsigned long   reg_short_reset;   /* reset after counter overflow */
+
+       unsigned long   reg_reset_pmds[4]; /* registers to reset on overflow */
+       unsigned long   reg_random_seed;   /* seed for randomization */
+       unsigned long   reg_random_mask;   /* random range limit */
+       unsigned long   reg_last_reset_val;/* return: PMD last reset value */
+
+       unsigned long   reg_smpl_pmds[4];  /* pmds to be saved on overflow */
+       unsigned long   reg_smpl_eventid;  /* opaque sampling event id */
+       unsigned long   reg_ovfl_switch_cnt;/* #overflows to switch */
+
+       unsigned long   reg_reserved2[2];   /* for future use */
+};
+
+/*
+ * argument to PFM_WRITE_IBRS/PFM_WRITE_DBRS
+ */
+struct pfarg_dbreg {
+       unsigned int    dbreg_num;              /* which debug register */
+       unsigned short  dbreg_set;              /* event set */
+       unsigned short  dbreg_reserved1;        /* for future use */
+       unsigned long   dbreg_value;            /* value for debug register */
+       unsigned long   dbreg_flags;            /* return: dbreg error */
+       unsigned long   dbreg_reserved2[1];     /* for future use */
+};
+
+/*
+ * argument to PFM_GET_FEATURES
+ */
+struct pfarg_features {
+       unsigned int    ft_version;     /* major [16-31], minor [0-15] */
+       unsigned int    ft_reserved;    /* reserved for future use */
+       unsigned long   reserved[4];    /* for future use */
+};
+
+typedef struct {
+       int             msg_type;               /* generic message header */
+       int             msg_ctx_fd;             /* generic message header */
+       unsigned long   msg_ovfl_pmds[4];       /* which PMDs overflowed */
+       unsigned short  msg_active_set;         /* active set on overflow */
+       unsigned short  msg_reserved1;          /* for future use */
+       unsigned int    msg_reserved2;          /* for future use */
+       unsigned long   msg_tstamp;             /* for perf tuning/debug */
+} pfm_ovfl_msg_t;
+
+typedef struct {
+       int             msg_type;               /* generic message header */
+       int             msg_ctx_fd;             /* generic message header */
+       unsigned long   msg_tstamp;             /* for perf tuning */
+} pfm_end_msg_t;
+
+typedef struct {
+       int             msg_type;               /* type of the message */
+       int             msg_ctx_fd;             /* context file descriptor */
+       unsigned long   msg_tstamp;             /* for perf tuning */
+} pfm_gen_msg_t;
+
+typedef union {
+       int type;
+       pfm_ovfl_msg_t  pfm_ovfl_msg;
+       pfm_end_msg_t   pfm_end_msg;
+       pfm_gen_msg_t   pfm_gen_msg;
+} pfm_msg_t;
+
+/*
+ * PMD/PMC return flags in case of error (ignored on input)
+ *
+ * reg_flags layout:
+ * bit 00-15 : generic flags
+ * bits[16-23] : arch-specific flags (see asm/perfmon.h)
+ * bit 24-31 : error codes
+ *
+ * Those flags are used on output and must be checked in case EINVAL is
+ * returned by a command accepting a vector of values and each has a flag
+ * field, such as pfarg_reg or pfarg_reg
+ */
+#define PFM_REG_RETFL_NOTAVAIL (1<<31) /* not implemented or unaccessible */
+#define PFM_REG_RETFL_EINVAL   (1<<30) /* entry is invalid */
+#define PFM_REG_RETFL_MASK     (PFM_REG_RETFL_NOTAVAIL|\
+                                PFM_REG_RETFL_EINVAL)
+
+#define PFM_REG_HAS_ERROR(flag)        (((flag) & PFM_REG_RETFL_MASK) != 0)
+
+#endif /* _ASM_IA64_PERFMON_COMPAT_H_ */
diff --git a/arch/ia64/include/asm/perfmon_default_smpl.h b/arch/ia64/include/asm/perfmon_default_smpl.h
index 48822c0..8234f32 100644
--- a/arch/ia64/include/asm/perfmon_default_smpl.h
+++ b/arch/ia64/include/asm/perfmon_default_smpl.h
@@ -1,83 +1,106 @@
 /*
- * Copyright (C) 2002-2003 Hewlett-Packard Co
- *               Stephane Eranian <eranian@hpl.hp.com>
+ * Copyright (c) 2002-2006 Hewlett-Packard Development Company, L.P.
+ * Contributed by Stephane Eranian <eranian@hpl.hp.com>
  *
- * This file implements the default sampling buffer format
- * for Linux/ia64 perfmon subsystem.
+ * This file implements the old default sampling buffer format
+ * for the perfmon2 subsystem. For IA-64 only.
+ *
+ * It requires the use of the perfmon_compat.h header. It is recommended
+ * that applications be ported to the new format instead.
+ *
+ * This program is free software; you can redistribute it and/or
+ * modify it under the terms of version 2 of the GNU General Public
+ * License as published by the Free Software Foundation.
+ *
+ * This program is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+ * General Public License for more details.
+ *
+ * You should have received a copy of the GNU General Public License
+ * along with this program; if not, write to the Free Software
+ * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA
+ * 02111-1307 USA
  */
-#ifndef __PERFMON_DEFAULT_SMPL_H__
-#define __PERFMON_DEFAULT_SMPL_H__ 1
+#ifndef __ASM_IA64_PERFMON_DEFAULT_SMPL_H__
+#define __ASM_IA64_PERFMON_DEFAULT_SMPL_H__ 1
+
+#ifndef __ia64__
+#error "this file must be used for compatibility reasons only on IA-64"
+#endif
 
 #define PFM_DEFAULT_SMPL_UUID { \
-               0x4d, 0x72, 0xbe, 0xc0, 0x06, 0x64, 0x41, 0x43, 0x82, 0xb4, 0xd3, 0xfd, 0x27, 0x24, 0x3c, 0x97}
+               0x4d, 0x72, 0xbe, 0xc0, 0x06, 0x64, 0x41, 0x43, 0x82,\
+               0xb4, 0xd3, 0xfd, 0x27, 0x24, 0x3c, 0x97}
 
 /*
  * format specific parameters (passed at context creation)
  */
-typedef struct {
+struct pfm_default_smpl_arg {
        unsigned long buf_size;         /* size of the buffer in bytes */
        unsigned int  flags;            /* buffer specific flags */
        unsigned int  res1;             /* for future use */
        unsigned long reserved[2];      /* for future use */
-} pfm_default_smpl_arg_t;
+};
 
 /*
  * combined context+format specific structure. Can be passed
- * to PFM_CONTEXT_CREATE
+ * to PFM_CONTEXT_CREATE (not PFM_CONTEXT_CREATE2)
  */
-typedef struct {
-       pfarg_context_t         ctx_arg;
-       pfm_default_smpl_arg_t  buf_arg;
-} pfm_default_smpl_ctx_arg_t;
+struct pfm_default_smpl_ctx_arg {
+       struct pfarg_context            ctx_arg;
+       struct pfm_default_smpl_arg     buf_arg;
+};
 
 /*
  * This header is at the beginning of the sampling buffer returned to the user.
  * It is directly followed by the first record.
  */
-typedef struct {
-       unsigned long   hdr_count;              /* how many valid entries */
-       unsigned long   hdr_cur_offs;           /* current offset from top of buffer */
-       unsigned long   hdr_reserved2;          /* reserved for future use */
+struct pfm_default_smpl_hdr {
+       u64     hdr_count;      /* how many valid entries */
+       u64     hdr_cur_offs;   /* current offset from top of buffer */
+       u64     dr_reserved2;   /* reserved for future use */
 
-       unsigned long   hdr_overflows;          /* how many times the buffer overflowed */
-       unsigned long   hdr_buf_size;           /* how many bytes in the buffer */
+       u64     hdr_overflows;  /* how many times the buffer overflowed */
+       u64     hdr_buf_size;   /* how many bytes in the buffer */
 
-       unsigned int    hdr_version;            /* contains perfmon version (smpl format diffs) */
-       unsigned int    hdr_reserved1;          /* for future use */
-       unsigned long   hdr_reserved[10];       /* for future use */
-} pfm_default_smpl_hdr_t;
+       u32     hdr_version;    /* smpl format version*/
+       u32     hdr_reserved1;          /* for future use */
+       u64     hdr_reserved[10];       /* for future use */
+};
 
 /*
  * Entry header in the sampling buffer.  The header is directly followed
- * with the values of the PMD registers of interest saved in increasing 
- * index order: PMD4, PMD5, and so on. How many PMDs are present depends 
+ * with the values of the PMD registers of interest saved in increasing
+ * index order: PMD4, PMD5, and so on. How many PMDs are present depends
  * on how the session was programmed.
  *
  * In the case where multiple counters overflow at the same time, multiple
  * entries are written consecutively.
  *
- * last_reset_value member indicates the initial value of the overflowed PMD. 
+ * last_reset_value member indicates the initial value of the overflowed PMD.
  */
-typedef struct {
-        int             pid;                    /* thread id (for NPTL, this is gettid()) */
-        unsigned char   reserved1[3];           /* reserved for future use */
-        unsigned char   ovfl_pmd;               /* index of overflowed PMD */
-
-        unsigned long   last_reset_val;         /* initial value of overflowed PMD */
-        unsigned long   ip;                     /* where did the overflow interrupt happened  */
-        unsigned long   tstamp;                 /* ar.itc when entering perfmon intr. handler */
-
-        unsigned short  cpu;                    /* cpu on which the overfow occured */
-        unsigned short  set;                    /* event set active when overflow ocurred   */
-        int                    tgid;                   /* thread group id (for NPTL, this is getpid()) */
-} pfm_default_smpl_entry_t;
+struct pfm_default_smpl_entry {
+       pid_t   pid;            /* thread id (for NPTL, this is gettid()) */
+       uint8_t reserved1[3];   /* for future use */
+       uint8_t ovfl_pmd;       /* overflow pmd for this sample */
+       u64     last_reset_val; /* initial value of overflowed PMD */
+       unsigned long ip;       /* where did the overflow interrupt happened */
+       u64     tstamp;         /* overflow timetamp */
+       u16     cpu;            /* cpu on which the overfow occured */
+       u16     set;            /* event set active when overflow ocurred   */
+       pid_t   tgid;           /* thread group id (for NPTL, this is getpid()) */
+};
 
-#define PFM_DEFAULT_MAX_PMDS           64 /* how many pmds supported by data structures (sizeof(unsigned long) */
-#define PFM_DEFAULT_MAX_ENTRY_SIZE     (sizeof(pfm_default_smpl_entry_t)+(sizeof(unsigned long)*PFM_DEFAULT_MAX_PMDS))
-#define PFM_DEFAULT_SMPL_MIN_BUF_SIZE  (sizeof(pfm_default_smpl_hdr_t)+PFM_DEFAULT_MAX_ENTRY_SIZE)
+#define PFM_DEFAULT_MAX_PMDS           64 /* #pmds supported  */
+#define PFM_DEFAULT_MAX_ENTRY_SIZE     (sizeof(struct pfm_default_smpl_entry)+\
+                                        (sizeof(u64)*PFM_DEFAULT_MAX_PMDS))
+#define PFM_DEFAULT_SMPL_MIN_BUF_SIZE  (sizeof(struct pfm_default_smpl_hdr)+\
+                                        PFM_DEFAULT_MAX_ENTRY_SIZE)
 
 #define PFM_DEFAULT_SMPL_VERSION_MAJ   2U
-#define PFM_DEFAULT_SMPL_VERSION_MIN   0U
-#define PFM_DEFAULT_SMPL_VERSION       (((PFM_DEFAULT_SMPL_VERSION_MAJ&0xffff)<<16)|(PFM_DEFAULT_SMPL_VERSION_MIN & 0xffff))
+#define PFM_DEFAULT_SMPL_VERSION_MIN 1U
+#define PFM_DEFAULT_SMPL_VERSION (((PFM_DEFAULT_SMPL_VERSION_MAJ&0xffff)<<16)|\
+                                   (PFM_DEFAULT_SMPL_VERSION_MIN & 0xffff))
 
-#endif /* __PERFMON_DEFAULT_SMPL_H__ */
+#endif /* __ASM_IA64_PERFMON_DEFAULT_SMPL_H__ */
diff --git a/arch/ia64/include/asm/perfmon_kern.h b/arch/ia64/include/asm/perfmon_kern.h
new file mode 100644
index 0000000..fb40459
--- /dev/null
+++ b/arch/ia64/include/asm/perfmon_kern.h
@@ -0,0 +1,356 @@
+/*
+ * Copyright (c) 2001-2007 Hewlett-Packard Development Company, L.P.
+ * Contributed by Stephane Eranian <eranian@hpl.hp.com>
+ *
+ * This file contains Itanium Processor Family specific definitions
+ * for the perfmon interface.
+ *
+ * This program is free software; you can redistribute it and/or
+ * modify it under the terms of version 2 of the GNU General Public
+ * License as published by the Free Software Foundation.
+ *
+ * This program is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+ * General Public License for more details.
+ *
+ * You should have received a copy of the GNU General Public License
+ * along with this program; if not, write to the Free Software
+ * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA
+ * 02111-1307 USA
+  */
+#ifndef _ASM_IA64_PERFMON_KERN_H_
+#define _ASM_IA64_PERFMON_KERN_H_
+
+#ifdef __KERNEL__
+
+#ifdef CONFIG_PERFMON
+#include <asm/unistd.h>
+#include <asm/hw_irq.h>
+
+/*
+ * describe the content of the pfm_syst_info field
+ * layout:
+ * bits[00-15] : generic flags
+ * bits[16-31] : arch-specific flags
+ */
+#define PFM_ITA_CPUINFO_IDLE_EXCL 0x10000 /* stop monitoring in idle loop */
+
+/*
+ * For some CPUs, the upper bits of a counter must be set in order for the
+ * overflow interrupt to happen. On overflow, the counter has wrapped around,
+ * and the upper bits are cleared. This function may be used to set them back.
+ */
+static inline void pfm_arch_ovfl_reset_pmd(struct pfm_context *ctx,
+                                          unsigned int cnum)
+{}
+
+/*
+ * called from __pfm_interrupt_handler(). ctx is not NULL.
+ * ctx is locked. PMU interrupt is masked.
+ *
+ * must stop all monitoring to ensure handler has consistent view.
+ * must collect overflowed PMDs bitmask  into povfls_pmds and
+ * npend_ovfls. If no interrupt detected then npend_ovfls
+ * must be set to zero.
+ */
+static inline void pfm_arch_intr_freeze_pmu(struct pfm_context *ctx,
+                                           struct pfm_event_set *set)
+{
+       u64 tmp;
+
+       /*
+        * do not overwrite existing value, must
+        * process those first (coming from context switch replay)
+        */
+       if (set->npend_ovfls)
+               return;
+
+       ia64_srlz_d();
+
+       tmp =  ia64_get_pmc(0) & ~0xf;
+
+       set->povfl_pmds[0] = tmp;
+
+       set->npend_ovfls = ia64_popcnt(tmp);
+}
+
+static inline int pfm_arch_init_pmu_config(void)
+{
+       return 0;
+}
+
+static inline void pfm_arch_resend_irq(struct pfm_context *ctx)
+{
+       ia64_resend_irq(IA64_PERFMON_VECTOR);
+}
+
+static inline void pfm_arch_clear_pmd_ovfl_cond(struct pfm_context *ctx,
+                                               struct pfm_event_set *set)
+{}
+
+static inline void pfm_arch_serialize(void)
+{
+       ia64_srlz_d();
+}
+
+static inline void pfm_arch_intr_unfreeze_pmu(struct pfm_context *ctx)
+{
+       PFM_DBG_ovfl("state=%d", ctx->state);
+       ia64_set_pmc(0, 0);
+       /* no serialization */
+}
+
+static inline void pfm_arch_write_pmc(struct pfm_context *ctx,
+                                     unsigned int cnum, u64 value)
+{
+       if (cnum < 256) {
+               ia64_set_pmc(pfm_pmu_conf->pmc_desc[cnum].hw_addr, value);
+       } else if (cnum < 264) {
+               ia64_set_ibr(cnum-256, value);
+               ia64_dv_serialize_instruction();
+       } else {
+               ia64_set_dbr(cnum-264, value);
+               ia64_dv_serialize_instruction();
+       }
+}
+
+/*
+ * On IA-64, for per-thread context which have the ITA_FL_INSECURE
+ * flag, it is possible to start/stop monitoring directly from user evel
+ * without calling pfm_start()/pfm_stop. This allows very lightweight
+ * control yet the kernel sometimes needs to know if monitoring is actually
+ * on or off.
+ *
+ * Tracking of this information is normally done by pfm_start/pfm_stop
+ * in flags.started. Here we need to compensate by checking actual
+ * psr bit.
+ */
+static inline int pfm_arch_is_active(struct pfm_context *ctx)
+{
+       return ctx->flags.started
+              || ia64_getreg(_IA64_REG_PSR) & (IA64_PSR_UP|IA64_PSR_PP);
+}
+
+static inline void pfm_arch_write_pmd(struct pfm_context *ctx,
+                                     unsigned int cnum, u64 value)
+{
+       /*
+        * for a counting PMD, overflow bit must be cleared
+        */
+       if (pfm_pmu_conf->pmd_desc[cnum].type & PFM_REG_C64)
+               value &= pfm_pmu_conf->ovfl_mask;
+
+       /*
+        * for counters, write to upper bits are ignored, no need to mask
+        */
+       ia64_set_pmd(pfm_pmu_conf->pmd_desc[cnum].hw_addr, value);
+}
+
+static inline u64 pfm_arch_read_pmd(struct pfm_context *ctx, unsigned int cnum)
+{
+       return ia64_get_pmd(pfm_pmu_conf->pmd_desc[cnum].hw_addr);
+}
+
+static inline u64 pfm_arch_read_pmc(struct pfm_context *ctx, unsigned int cnum)
+{
+       return ia64_get_pmc(pfm_pmu_conf->pmc_desc[cnum].hw_addr);
+}
+
+static inline void pfm_arch_ctxswout_sys(struct task_struct *task,
+                                        struct pfm_context *ctx)
+{
+       struct pt_regs *regs;
+
+       regs = task_pt_regs(task);
+       ia64_psr(regs)->pp = 0;
+}
+
+static inline void pfm_arch_ctxswin_sys(struct task_struct *task,
+                                       struct pfm_context *ctx)
+{
+       struct pt_regs *regs;
+
+       if (!(ctx->active_set->flags & PFM_ITA_SETFL_INTR_ONLY)) {
+               regs = task_pt_regs(task);
+               ia64_psr(regs)->pp = 1;
+       }
+}
+
+/*
+ * On IA-64, the PMDs are NOT saved by pfm_arch_freeze_pmu()
+ * when entering the PMU interrupt handler, thus, we need
+ * to save them in pfm_switch_sets_from_intr()
+ */
+static inline void pfm_arch_save_pmds_from_intr(struct pfm_context *ctx,
+                                          struct pfm_event_set *set)
+{
+       pfm_save_pmds(ctx, set);
+}
+
+int pfm_arch_context_create(struct pfm_context *ctx, u32 ctx_flags);
+
+static inline void pfm_arch_context_free(struct pfm_context *ctx)
+{}
+
+int pfm_arch_ctxswout_thread(struct task_struct *task, struct pfm_context *ctx);
+void pfm_arch_ctxswin_thread(struct task_struct *task,
+                            struct pfm_context *ctx);
+
+void pfm_arch_unload_context(struct pfm_context *ctx);
+int pfm_arch_load_context(struct pfm_context *ctx);
+int pfm_arch_setfl_sane(struct pfm_context *ctx, u32 flags);
+
+void pfm_arch_mask_monitoring(struct pfm_context *ctx,
+                             struct pfm_event_set *set);
+void pfm_arch_unmask_monitoring(struct pfm_context *ctx,
+                               struct pfm_event_set *set);
+
+void pfm_arch_restore_pmds(struct pfm_context *ctx, struct pfm_event_set *set);
+void pfm_arch_restore_pmcs(struct pfm_context *ctx, struct pfm_event_set *set);
+
+void pfm_arch_stop(struct task_struct *task, struct pfm_context *ctx);
+void pfm_arch_start(struct task_struct *task, struct pfm_context *ctx);
+
+int  pfm_arch_init(void);
+void pfm_arch_init_percpu(void);
+char *pfm_arch_get_pmu_module_name(void);
+
+int __pfm_use_dbregs(struct task_struct *task);
+int  __pfm_release_dbregs(struct task_struct *task);
+int pfm_ia64_mark_dbregs_used(struct pfm_context *ctx,
+                             struct pfm_event_set *set);
+
+void pfm_arch_show_session(struct seq_file *m);
+
+static inline int pfm_arch_pmu_acquire(u64 *unavail_pmcs, u64 *unavail_pmds)
+{
+       return 0;
+}
+
+static inline void pfm_arch_pmu_release(void)
+{}
+
+/* not necessary on IA-64 */
+static inline void pfm_cacheflush(void *addr, unsigned int len)
+{}
+
+/*
+ * miscellaneous architected definitions
+ */
+#define PFM_ITA_FCNTR  4 /* first counting monitor (PMC/PMD) */
+
+/*
+ * private event set flags  (set_priv_flags)
+ */
+#define PFM_ITA_SETFL_USE_DBR  0x1000000 /* set uses debug registers */
+
+
+/*
+ * Itanium-specific data structures
+ */
+struct pfm_ia64_context_flags {
+       unsigned int use_dbr:1;  /* use range restrictions (debug registers) */
+       unsigned int insecure:1; /* insecure monitoring for non-self session */
+       unsigned int reserved:30;/* for future use */
+};
+
+struct pfm_arch_context {
+       struct pfm_ia64_context_flags flags;    /* arch specific ctx flags */
+       u64                      ctx_saved_psr_up;/* storage for psr_up */
+#ifdef CONFIG_IA64_PERFMON_COMPAT
+       void                    *ctx_smpl_vaddr; /* vaddr of user mapping */
+#endif
+};
+
+#ifdef CONFIG_IA64_PERFMON_COMPAT
+ssize_t pfm_arch_compat_read(struct pfm_context *ctx,
+                            char __user *buf,
+                            int non_block,
+                            size_t size);
+int pfm_ia64_compat_init(void);
+int pfm_smpl_buf_alloc_compat(struct pfm_context *ctx,
+                             size_t rsize, struct file *filp);
+#else
+static inline ssize_t pfm_arch_compat_read(struct pfm_context *ctx,
+                            char __user *buf,
+                            int non_block,
+                            size_t size)
+{
+       return -EINVAL;
+}
+
+static inline int pfm_smpl_buf_alloc_compat(struct pfm_context *ctx,
+                                           size_t rsize, struct file *filp)
+{
+       return -EINVAL;
+}
+#endif
+
+static inline void pfm_arch_arm_handle_work(struct task_struct *task)
+{
+       /*
+        * On IA-64, we ran out of bits in the bottom 7 bits of the
+        * threadinfo bitmask.Thus we used a 2-stage approach by piggybacking
+        * on NOTIFY_RESUME and then in do_notify_resume() we demultiplex and
+        * call pfm_handle_work() if needed
+        */
+       set_tsk_thread_flag(task, TIF_NOTIFY_RESUME);
+}
+
+static inline void pfm_arch_disarm_handle_work(struct task_struct *task)
+{
+       /*
+        * we cannot just clear TIF_NOTIFY_RESUME because other TIF flags are
+        * piggybackedonto it: TIF_PERFMON_WORK, TIF_RESTORE_RSE
+        *
+        * The tsk_clear_notify_resume() checks if any of those are set before
+        * clearing the * bit
+        */
+       tsk_clear_notify_resume(task);
+}
+
+static inline int pfm_arch_pmu_config_init(struct pfm_pmu_config *cfg)
+{
+       return 0;
+}
+
+extern struct pfm_ia64_pmu_info *pfm_ia64_pmu_info;
+
+#define PFM_ARCH_CTX_SIZE      (sizeof(struct pfm_arch_context))
+
+/*
+ * IA-64 does not need extra alignment requirements for the sampling buffer
+ */
+#define PFM_ARCH_SMPL_ALIGN_SIZE       0
+
+
+static inline void pfm_release_dbregs(struct task_struct *task)
+{
+       if (task->thread.flags & IA64_THREAD_DBG_VALID)
+               __pfm_release_dbregs(task);
+}
+
+#define pfm_use_dbregs(_t)     __pfm_use_dbregs(_t)
+
+static inline int pfm_arch_get_base_syscall(void)
+{
+       return __NR_pfm_create_context;
+}
+
+struct pfm_arch_pmu_info {
+       unsigned long mask_pmcs[PFM_PMC_BV]; /* modify on when masking */
+};
+
+DECLARE_PER_CPU(u32, pfm_syst_info);
+#else /* !CONFIG_PERFMON */
+/*
+ * perfmon ia64-specific hooks
+ */
+#define pfm_release_dbregs(_t)                 do { } while (0)
+#define pfm_use_dbregs(_t)                     (0)
+
+#endif /* CONFIG_PERFMON */
+
+#endif /* __KERNEL__ */
+#endif /* _ASM_IA64_PERFMON_KERN_H_ */
diff --git a/arch/ia64/include/asm/processor.h b/arch/ia64/include/asm/processor.h
index f88fa05..9d6af9c 100644
--- a/arch/ia64/include/asm/processor.h
+++ b/arch/ia64/include/asm/processor.h
@@ -42,7 +42,6 @@
 
 #define IA64_THREAD_FPH_VALID  (__IA64_UL(1) << 0)     /* floating-point high state valid? */
 #define IA64_THREAD_DBG_VALID  (__IA64_UL(1) << 1)     /* debug registers valid? */
-#define IA64_THREAD_PM_VALID   (__IA64_UL(1) << 2)     /* performance registers valid? */
 #define IA64_THREAD_UAC_NOPRINT        (__IA64_UL(1) << 3)     /* don't log unaligned accesses */
 #define IA64_THREAD_UAC_SIGBUS (__IA64_UL(1) << 4)     /* generate SIGBUS on unaligned acc. */
 #define IA64_THREAD_MIGRATION  (__IA64_UL(1) << 5)     /* require migration
@@ -321,14 +320,6 @@ struct thread_struct {
 #else
 # define INIT_THREAD_IA32
 #endif /* CONFIG_IA32_SUPPORT */
-#ifdef CONFIG_PERFMON
-       void *pfm_context;                   /* pointer to detailed PMU context */
-       unsigned long pfm_needs_checking;    /* when >0, pending perfmon work on kernel exit */
-# define INIT_THREAD_PM                .pfm_context =          NULL,     \
-                               .pfm_needs_checking =   0UL,
-#else
-# define INIT_THREAD_PM
-#endif
        __u64 dbr[IA64_NUM_DBG_REGS];
        __u64 ibr[IA64_NUM_DBG_REGS];
        struct ia64_fpreg fph[96];      /* saved/loaded on demand */
@@ -343,7 +334,6 @@ struct thread_struct {
        .task_size =    DEFAULT_TASK_SIZE,                      \
        .last_fph_cpu =  -1,                                    \
        INIT_THREAD_IA32                                        \
-       INIT_THREAD_PM                                          \
        .dbr =          {0, },                                  \
        .ibr =          {0, },                                  \
        .fph =          {{{{0}}}, }                             \
diff --git a/arch/ia64/include/asm/system.h b/arch/ia64/include/asm/system.h
index 927a381..ab5aeea 100644
--- a/arch/ia64/include/asm/system.h
+++ b/arch/ia64/include/asm/system.h
@@ -217,6 +217,7 @@ struct task_struct;
 extern void ia64_save_extra (struct task_struct *task);
 extern void ia64_load_extra (struct task_struct *task);
 
+
 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
 extern void ia64_account_on_switch (struct task_struct *prev, struct task_struct *next);
 # define IA64_ACCOUNT_ON_SWITCH(p,n) ia64_account_on_switch(p,n)
@@ -224,16 +225,9 @@ extern void ia64_account_on_switch (struct task_struct *prev, struct task_struct
 # define IA64_ACCOUNT_ON_SWITCH(p,n)
 #endif
 
-#ifdef CONFIG_PERFMON
-  DECLARE_PER_CPU(unsigned long, pfm_syst_info);
-# define PERFMON_IS_SYSWIDE() (__get_cpu_var(pfm_syst_info) & 0x1)
-#else
-# define PERFMON_IS_SYSWIDE() (0)
-#endif
-
-#define IA64_HAS_EXTRA_STATE(t)                                                        \
-       ((t)->thread.flags & (IA64_THREAD_DBG_VALID|IA64_THREAD_PM_VALID)       \
-        || IS_IA32_PROCESS(task_pt_regs(t)) || PERFMON_IS_SYSWIDE())
+#define IA64_HAS_EXTRA_STATE(t)                                                \
+       (((t)->thread.flags & IA64_THREAD_DBG_VALID)                    \
+        || IS_IA32_PROCESS(task_pt_regs(t)))
 
 #define __switch_to(prev,next,last) do {                                                        \
        IA64_ACCOUNT_ON_SWITCH(prev, next);                                                      \
@@ -241,6 +235,10 @@ extern void ia64_account_on_switch (struct task_struct *prev, struct task_struct
                ia64_save_extra(prev);                                                           \
        if (IA64_HAS_EXTRA_STATE(next))                                                          \
                ia64_load_extra(next);                                                           \
+       if (test_tsk_thread_flag(prev, TIF_PERFMON_CTXSW))                                       \
+               pfm_ctxsw_out(prev, next);                                                       \
+       if (test_tsk_thread_flag(next, TIF_PERFMON_CTXSW))                                       \
+               pfm_ctxsw_in(prev, next);                                                        \
        ia64_psr(task_pt_regs(next))->dfh = !ia64_is_local_fpu_owner(next);                      \
        (last) = ia64_switch_to((next));                                                         \
 } while (0)
diff --git a/arch/ia64/include/asm/thread_info.h b/arch/ia64/include/asm/thread_info.h
index 7c60fcd..3355332 100644
--- a/arch/ia64/include/asm/thread_info.h
+++ b/arch/ia64/include/asm/thread_info.h
@@ -110,6 +110,8 @@ extern void tsk_clear_notify_resume(struct task_struct *tsk);
 #define TIF_DB_DISABLED                19      /* debug trap disabled for fsyscall */
 #define TIF_FREEZE             20      /* is freezing for suspend */
 #define TIF_RESTORE_RSE                21      /* user RBS is newer than kernel RBS */
+#define TIF_PERFMON_CTXSW      22      /* perfmon needs ctxsw calls */
+#define TIF_PERFMON_WORK       23      /* work for pfm_handle_work() */
 
 #define _TIF_SYSCALL_TRACE     (1 << TIF_SYSCALL_TRACE)
 #define _TIF_SYSCALL_AUDIT     (1 << TIF_SYSCALL_AUDIT)
@@ -123,6 +125,8 @@ extern void tsk_clear_notify_resume(struct task_struct *tsk);
 #define _TIF_DB_DISABLED       (1 << TIF_DB_DISABLED)
 #define _TIF_FREEZE            (1 << TIF_FREEZE)
 #define _TIF_RESTORE_RSE       (1 << TIF_RESTORE_RSE)
+#define _TIF_PERFMON_CTXSW     (1 << TIF_PERFMON_CTXSW)
+#define _TIF_PERFMON_WORK      (1 << TIF_PERFMON_WORK)
 
 /* "work to do on user-return" bits */
 #define TIF_ALLWORK_MASK       (_TIF_SIGPENDING|_TIF_NOTIFY_RESUME|_TIF_SYSCALL_AUDIT|\
diff --git a/arch/ia64/include/asm/unistd.h b/arch/ia64/include/asm/unistd.h
index d535833..29a43bc 100644
--- a/arch/ia64/include/asm/unistd.h
+++ b/arch/ia64/include/asm/unistd.h
@@ -308,11 +308,23 @@
 #define __NR_dup3                      1316
 #define __NR_pipe2                     1317
 #define __NR_inotify_init1             1318
+#define __NR_pfm_create_context                1319
+#define __NR_pfm_write_pmcs            (__NR_pfm_create_context+1)
+#define __NR_pfm_write_pmds            (__NR_pfm_create_context+2)
+#define __NR_pfm_read_pmds             (__NR_pfm_create_context+3)
+#define __NR_pfm_load_context          (__NR_pfm_create_context+4)
+#define __NR_pfm_start                 (__NR_pfm_create_context+5)
+#define __NR_pfm_stop                  (__NR_pfm_create_context+6)
+#define __NR_pfm_restart               (__NR_pfm_create_context+7)
+#define __NR_pfm_create_evtsets                (__NR_pfm_create_context+8)
+#define __NR_pfm_getinfo_evtsets       (__NR_pfm_create_context+9)
+#define __NR_pfm_delete_evtsets        (__NR_pfm_create_context+10)
+#define __NR_pfm_unload_context                (__NR_pfm_create_context+11)
 
 #ifdef __KERNEL__
 
 
-#define NR_syscalls                    295 /* length of syscall table */
+#define NR_syscalls                    307 /* length of syscall table */
 
 /*
  * The following defines stop scripts/checksyscalls.sh from complaining about
diff --git a/arch/ia64/kernel/Makefile b/arch/ia64/kernel/Makefile
index 87fea11..b5ac54c 100644
--- a/arch/ia64/kernel/Makefile
+++ b/arch/ia64/kernel/Makefile
@@ -5,7 +5,7 @@
 extra-y        := head.o init_task.o vmlinux.lds
 
 obj-y := acpi.o entry.o efi.o efi_stub.o gate-data.o fsys.o ia64_ksyms.o irq.o irq_ia64.o      \
-        irq_lsapic.o ivt.o machvec.o pal.o patch.o process.o perfmon.o ptrace.o sal.o          \
+        irq_lsapic.o ivt.o machvec.o pal.o patch.o process.o ptrace.o sal.o            \
         salinfo.o setup.o signal.o sys_ia64.o time.o traps.o unaligned.o \
         unwind.o mca.o mca_asm.o topology.o
 
@@ -23,7 +23,6 @@ obj-$(CONFIG_IOSAPIC)         += iosapic.o
 obj-$(CONFIG_MODULES)          += module.o
 obj-$(CONFIG_SMP)              += smp.o smpboot.o
 obj-$(CONFIG_NUMA)             += numa.o
-obj-$(CONFIG_PERFMON)          += perfmon_default_smpl.o
 obj-$(CONFIG_IA64_CYCLONE)     += cyclone.o
 obj-$(CONFIG_CPU_FREQ)         += cpufreq/
 obj-$(CONFIG_IA64_MCA_RECOVERY)        += mca_recovery.o
diff --git a/arch/ia64/kernel/entry.S b/arch/ia64/kernel/entry.S
index 0dd6c14..f1c3e41 100644
--- a/arch/ia64/kernel/entry.S
+++ b/arch/ia64/kernel/entry.S
@@ -1697,6 +1697,18 @@ sys_call_table:
        data8 sys_dup3
        data8 sys_pipe2
        data8 sys_inotify_init1
+       data8 sys_pfm_create_context
+       data8 sys_pfm_write_pmcs                // 1320
+       data8 sys_pfm_write_pmds
+       data8 sys_pfm_read_pmds
+       data8 sys_pfm_load_context
+       data8 sys_pfm_start
+       data8 sys_pfm_stop                      // 1325
+       data8 sys_pfm_restart
+       data8 sys_pfm_create_evtsets
+       data8 sys_pfm_getinfo_evtsets
+       data8 sys_pfm_delete_evtsets
+       data8 sys_pfm_unload_context            // 1330
 
        .org sys_call_table + 8*NR_syscalls     // guard against failures to increase NR_syscalls
 #endif /* __IA64_ASM_PARAVIRTUALIZED_NATIVE */
diff --git a/arch/ia64/kernel/irq_ia64.c b/arch/ia64/kernel/irq_ia64.c
index 28d3d48..ede8024 100644
--- a/arch/ia64/kernel/irq_ia64.c
+++ b/arch/ia64/kernel/irq_ia64.c
@@ -40,10 +40,6 @@
 #include <asm/system.h>
 #include <asm/tlbflush.h>
 
-#ifdef CONFIG_PERFMON
-# include <asm/perfmon.h>
-#endif
-
 #define IRQ_DEBUG      0
 
 #define IRQ_VECTOR_UNASSIGNED  (0)
@@ -660,9 +656,6 @@ init_IRQ (void)
        }
 #endif
 #endif
-#ifdef CONFIG_PERFMON
-       pfm_init_percpu();
-#endif
        platform_irq_init();
 }
 
diff --git a/arch/ia64/kernel/perfmon_default_smpl.c b/arch/ia64/kernel/perfmon_default_smpl.c
deleted file mode 100644
index 5f637bb..0000000
--- a/arch/ia64/kernel/perfmon_default_smpl.c
+++ /dev/null
@@ -1,296 +0,0 @@
-/*
- * Copyright (C) 2002-2003 Hewlett-Packard Co
- *               Stephane Eranian <eranian@hpl.hp.com>
- *
- * This file implements the default sampling buffer format
- * for the Linux/ia64 perfmon-2 subsystem.
- */
-#include <linux/kernel.h>
-#include <linux/types.h>
-#include <linux/module.h>
-#include <linux/init.h>
-#include <asm/delay.h>
-#include <linux/smp.h>
-
-#include <asm/perfmon.h>
-#include <asm/perfmon_default_smpl.h>
-
-MODULE_AUTHOR("Stephane Eranian <eranian@hpl.hp.com>");
-MODULE_DESCRIPTION("perfmon default sampling format");
-MODULE_LICENSE("GPL");
-
-#define DEFAULT_DEBUG 1
-
-#ifdef DEFAULT_DEBUG
-#define DPRINT(a) \
-       do { \
-               if (unlikely(pfm_sysctl.debug >0)) { printk("%s.%d: CPU%d ", __func__, __LINE__, smp_processor_id()); printk a; } \
-       } while (0)
-
-#define DPRINT_ovfl(a) \
-       do { \
-               if (unlikely(pfm_sysctl.debug > 0 && pfm_sysctl.debug_ovfl >0)) { printk("%s.%d: CPU%d ", __func__, __LINE__, smp_processor_id()); printk a; } \
-       } while (0)
-
-#else
-#define DPRINT(a)
-#define DPRINT_ovfl(a)
-#endif
-
-static int
-default_validate(struct task_struct *task, unsigned int flags, int cpu, void *data)
-{
-       pfm_default_smpl_arg_t *arg = (pfm_default_smpl_arg_t*)data;
-       int ret = 0;
-
-       if (data == NULL) {
-               DPRINT(("[%d] no argument passed\n", task_pid_nr(task)));
-               return -EINVAL;
-       }
-
-       DPRINT(("[%d] validate flags=0x%x CPU%d\n", task_pid_nr(task), flags, cpu));
-
-       /*
-        * must hold at least the buffer header + one minimally sized entry
-        */
-       if (arg->buf_size < PFM_DEFAULT_SMPL_MIN_BUF_SIZE) return -EINVAL;
-
-       DPRINT(("buf_size=%lu\n", arg->buf_size));
-
-       return ret;
-}
-
-static int
-default_get_size(struct task_struct *task, unsigned int flags, int cpu, void *data, unsigned long *size)
-{
-       pfm_default_smpl_arg_t *arg = (pfm_default_smpl_arg_t *)data;
-
-       /*
-        * size has been validated in default_validate
-        */
-       *size = arg->buf_size;
-
-       return 0;
-}
-
-static int
-default_init(struct task_struct *task, void *buf, unsigned int flags, int cpu, void *data)
-{
-       pfm_default_smpl_hdr_t *hdr;
-       pfm_default_smpl_arg_t *arg = (pfm_default_smpl_arg_t *)data;
-
-       hdr = (pfm_default_smpl_hdr_t *)buf;
-
-       hdr->hdr_version      = PFM_DEFAULT_SMPL_VERSION;
-       hdr->hdr_buf_size     = arg->buf_size;
-       hdr->hdr_cur_offs     = sizeof(*hdr);
-       hdr->hdr_overflows    = 0UL;
-       hdr->hdr_count        = 0UL;
-
-       DPRINT(("[%d] buffer=%p buf_size=%lu hdr_size=%lu hdr_version=%u cur_offs=%lu\n",
-               task_pid_nr(task),
-               buf,
-               hdr->hdr_buf_size,
-               sizeof(*hdr),
-               hdr->hdr_version,
-               hdr->hdr_cur_offs));
-
-       return 0;
-}
-
-static int
-default_handler(struct task_struct *task, void *buf, pfm_ovfl_arg_t *arg, struct pt_regs *regs, unsigned long stamp)
-{
-       pfm_default_smpl_hdr_t *hdr;
-       pfm_default_smpl_entry_t *ent;
-       void *cur, *last;
-       unsigned long *e, entry_size;
-       unsigned int npmds, i;
-       unsigned char ovfl_pmd;
-       unsigned char ovfl_notify;
-
-       if (unlikely(buf == NULL || arg == NULL|| regs == NULL || task == NULL)) {
-               DPRINT(("[%d] invalid arguments buf=%p arg=%p\n", task->pid, buf, arg));
-               return -EINVAL;
-       }
-
-       hdr         = (pfm_default_smpl_hdr_t *)buf;
-       cur         = buf+hdr->hdr_cur_offs;
-       last        = buf+hdr->hdr_buf_size;
-       ovfl_pmd    = arg->ovfl_pmd;
-       ovfl_notify = arg->ovfl_notify;
-
-       /*
-        * precheck for sanity
-        */
-       if ((last - cur) < PFM_DEFAULT_MAX_ENTRY_SIZE) goto full;
-
-       npmds = hweight64(arg->smpl_pmds[0]);
-
-       ent = (pfm_default_smpl_entry_t *)cur;
-
-       prefetch(arg->smpl_pmds_values);
-
-       entry_size = sizeof(*ent) + (npmds << 3);
-
-       /* position for first pmd */
-       e = (unsigned long *)(ent+1);
-
-       hdr->hdr_count++;
-
-       DPRINT_ovfl(("[%d] count=%lu cur=%p last=%p free_bytes=%lu ovfl_pmd=%d ovfl_notify=%d npmds=%u\n",
-                       task->pid,
-                       hdr->hdr_count,
-                       cur, last,
-                       last-cur,
-                       ovfl_pmd,
-                       ovfl_notify, npmds));
-
-       /*
-        * current = task running at the time of the overflow.
-        *
-        * per-task mode:
-        *      - this is ususally the task being monitored.
-        *        Under certain conditions, it might be a different task
-        *
-        * system-wide:
-        *      - this is not necessarily the task controlling the session
-        */
-       ent->pid            = current->pid;
-       ent->ovfl_pmd       = ovfl_pmd;
-       ent->last_reset_val = arg->pmd_last_reset; //pmd[0].reg_last_reset_val;
-
-       /*
-        * where did the fault happen (includes slot number)
-        */
-       ent->ip = regs->cr_iip | ((regs->cr_ipsr >> 41) & 0x3);
-
-       ent->tstamp    = stamp;
-       ent->cpu       = smp_processor_id();
-       ent->set       = arg->active_set;
-       ent->tgid      = current->tgid;
-
-       /*
-        * selectively store PMDs in increasing index number
-        */
-       if (npmds) {
-               unsigned long *val = arg->smpl_pmds_values;
-               for(i=0; i < npmds; i++) {
-                       *e++ = *val++;
-               }
-       }
-
-       /*
-        * update position for next entry
-        */
-       hdr->hdr_cur_offs += entry_size;
-       cur               += entry_size;
-
-       /*
-        * post check to avoid losing the last sample
-        */
-       if ((last - cur) < PFM_DEFAULT_MAX_ENTRY_SIZE) goto full;
-
-       /*
-        * keep same ovfl_pmds, ovfl_notify
-        */
-       arg->ovfl_ctrl.bits.notify_user     = 0;
-       arg->ovfl_ctrl.bits.block_task      = 0;
-       arg->ovfl_ctrl.bits.mask_monitoring = 0;
-       arg->ovfl_ctrl.bits.reset_ovfl_pmds = 1; /* reset before returning from interrupt handler */
-
-       return 0;
-full:
-       DPRINT_ovfl(("sampling buffer full free=%lu, count=%lu, ovfl_notify=%d\n", last-cur, hdr->hdr_count, ovfl_notify));
-
-       /*
-        * increment number of buffer overflow.
-        * important to detect duplicate set of samples.
-        */
-       hdr->hdr_overflows++;
-
-       /*
-        * if no notification requested, then we saturate the buffer
-        */
-       if (ovfl_notify == 0) {
-               arg->ovfl_ctrl.bits.notify_user     = 0;
-               arg->ovfl_ctrl.bits.block_task      = 0;
-               arg->ovfl_ctrl.bits.mask_monitoring = 1;
-               arg->ovfl_ctrl.bits.reset_ovfl_pmds = 0;
-       } else {
-               arg->ovfl_ctrl.bits.notify_user     = 1;
-               arg->ovfl_ctrl.bits.block_task      = 1; /* ignored for non-blocking context */
-               arg->ovfl_ctrl.bits.mask_monitoring = 1;
-               arg->ovfl_ctrl.bits.reset_ovfl_pmds = 0; /* no reset now */
-       }
-       return -1; /* we are full, sorry */
-}
-
-static int
-default_restart(struct task_struct *task, pfm_ovfl_ctrl_t *ctrl, void *buf, struct pt_regs *regs)
-{
-       pfm_default_smpl_hdr_t *hdr;
-
-       hdr = (pfm_default_smpl_hdr_t *)buf;
-
-       hdr->hdr_count    = 0UL;
-       hdr->hdr_cur_offs = sizeof(*hdr);
-
-       ctrl->bits.mask_monitoring = 0;
-       ctrl->bits.reset_ovfl_pmds = 1; /* uses long-reset values */
-
-       return 0;
-}
-
-static int
-default_exit(struct task_struct *task, void *buf, struct pt_regs *regs)
-{
-       DPRINT(("[%d] exit(%p)\n", task_pid_nr(task), buf));
-       return 0;
-}
-
-static pfm_buffer_fmt_t default_fmt={
-       .fmt_name           = "default_format",
-       .fmt_uuid           = PFM_DEFAULT_SMPL_UUID,
-       .fmt_arg_size       = sizeof(pfm_default_smpl_arg_t),
-       .fmt_validate       = default_validate,
-       .fmt_getsize        = default_get_size,
-       .fmt_init           = default_init,
-       .fmt_handler        = default_handler,
-       .fmt_restart        = default_restart,
-       .fmt_restart_active = default_restart,
-       .fmt_exit           = default_exit,
-};
-
-static int __init
-pfm_default_smpl_init_module(void)
-{
-       int ret;
-
-       ret = pfm_register_buffer_fmt(&default_fmt);
-       if (ret == 0) {
-               printk("perfmon_default_smpl: %s v%u.%u registered\n",
-                       default_fmt.fmt_name,
-                       PFM_DEFAULT_SMPL_VERSION_MAJ,
-                       PFM_DEFAULT_SMPL_VERSION_MIN);
-       } else {
-               printk("perfmon_default_smpl: %s cannot register ret=%d\n",
-                       default_fmt.fmt_name,
-                       ret);
-       }
-
-       return ret;
-}
-
-static void __exit
-pfm_default_smpl_cleanup_module(void)
-{
-       int ret;
-       ret = pfm_unregister_buffer_fmt(default_fmt.fmt_uuid);
-
-       printk("perfmon_default_smpl: unregister %s=%d\n", default_fmt.fmt_name, ret);
-}
-
-module_init(pfm_default_smpl_init_module);
-module_exit(pfm_default_smpl_cleanup_module);
-
diff --git a/arch/ia64/kernel/perfmon_generic.h b/arch/ia64/kernel/perfmon_generic.h
deleted file mode 100644
index 6748947..0000000
--- a/arch/ia64/kernel/perfmon_generic.h
+++ /dev/null
@@ -1,45 +0,0 @@
-/*
- * This file contains the generic PMU register description tables
- * and pmc checker used by perfmon.c.
- *
- * Copyright (C) 2002-2003  Hewlett Packard Co
- *               Stephane Eranian <eranian@hpl.hp.com>
- */
-
-static pfm_reg_desc_t pfm_gen_pmc_desc[PMU_MAX_PMCS]={
-/* pmc0  */ { PFM_REG_CONTROL , 0, 0x1UL, -1UL, NULL, NULL, {0UL,0UL, 0UL, 0UL}, {0UL,0UL, 0UL, 0UL}},
-/* pmc1  */ { PFM_REG_CONTROL , 0, 0x0UL, -1UL, NULL, NULL, {0UL,0UL, 0UL, 0UL}, {0UL,0UL, 0UL, 0UL}},
-/* pmc2  */ { PFM_REG_CONTROL , 0, 0x0UL, -1UL, NULL, NULL, {0UL,0UL, 0UL, 0UL}, {0UL,0UL, 0UL, 0UL}},
-/* pmc3  */ { PFM_REG_CONTROL , 0, 0x0UL, -1UL, NULL, NULL, {0UL,0UL, 0UL, 0UL}, {0UL,0UL, 0UL, 0UL}},
-/* pmc4  */ { PFM_REG_COUNTING, 0, 0x0UL, -1UL, NULL, NULL, {RDEP(4),0UL, 0UL, 0UL}, {0UL,0UL, 0UL, 0UL}},
-/* pmc5  */ { PFM_REG_COUNTING, 0, 0x0UL, -1UL, NULL, NULL, {RDEP(5),0UL, 0UL, 0UL}, {0UL,0UL, 0UL, 0UL}},
-/* pmc6  */ { PFM_REG_COUNTING, 0, 0x0UL, -1UL, NULL, NULL, {RDEP(6),0UL, 0UL, 0UL}, {0UL,0UL, 0UL, 0UL}},
-/* pmc7  */ { PFM_REG_COUNTING, 0, 0x0UL, -1UL, NULL, NULL, {RDEP(7),0UL, 0UL, 0UL}, {0UL,0UL, 0UL, 0UL}},
-           { PFM_REG_END     , 0, 0x0UL, -1UL, NULL, NULL, {0,}, {0,}}, /* end marker */
-};
-
-static pfm_reg_desc_t pfm_gen_pmd_desc[PMU_MAX_PMDS]={
-/* pmd0  */ { PFM_REG_NOTIMPL , 0, 0x0UL, -1UL, NULL, NULL, {0,}, {0,}},
-/* pmd1  */ { PFM_REG_NOTIMPL , 0, 0x0UL, -1UL, NULL, NULL, {0,}, {0,}},
-/* pmd2  */ { PFM_REG_NOTIMPL , 0, 0x0UL, -1UL, NULL, NULL, {0,}, {0,}},
-/* pmd3  */ { PFM_REG_NOTIMPL , 0, 0x0UL, -1UL, NULL, NULL, {0,}, {0,}},
-/* pmd4  */ { PFM_REG_COUNTING, 0, 0x0UL, -1UL, NULL, NULL, {0UL,0UL, 0UL, 0UL}, {RDEP(4),0UL, 0UL, 0UL}},
-/* pmd5  */ { PFM_REG_COUNTING, 0, 0x0UL, -1UL, NULL, NULL, {0UL,0UL, 0UL, 0UL}, {RDEP(5),0UL, 0UL, 0UL}},
-/* pmd6  */ { PFM_REG_COUNTING, 0, 0x0UL, -1UL, NULL, NULL, {0UL,0UL, 0UL, 0UL}, {RDEP(6),0UL, 0UL, 0UL}},
-/* pmd7  */ { PFM_REG_COUNTING, 0, 0x0UL, -1UL, NULL, NULL, {0UL,0UL, 0UL, 0UL}, {RDEP(7),0UL, 0UL, 0UL}},
-           { PFM_REG_END     , 0, 0x0UL, -1UL, NULL, NULL, {0,}, {0,}}, /* end marker */
-};
-
-/*
- * impl_pmcs, impl_pmds are computed at runtime to minimize errors!
- */
-static pmu_config_t pmu_conf_gen={
-       .pmu_name   = "Generic",
-       .pmu_family = 0xff, /* any */
-       .ovfl_val   = (1UL << 32) - 1,
-       .num_ibrs   = 0, /* does not use */
-       .num_dbrs   = 0, /* does not use */
-       .pmd_desc   = pfm_gen_pmd_desc,
-       .pmc_desc   = pfm_gen_pmc_desc
-};
-
diff --git a/arch/ia64/kernel/perfmon_itanium.h b/arch/ia64/kernel/perfmon_itanium.h
deleted file mode 100644
index d1d508a..0000000
--- a/arch/ia64/kernel/perfmon_itanium.h
+++ /dev/null
@@ -1,115 +0,0 @@
-/*
- * This file contains the Itanium PMU register description tables
- * and pmc checker used by perfmon.c.
- *
- * Copyright (C) 2002-2003  Hewlett Packard Co
- *               Stephane Eranian <eranian@hpl.hp.com>
- */
-static int pfm_ita_pmc_check(struct task_struct *task, pfm_context_t *ctx, unsigned int cnum, unsigned long *val, struct pt_regs *regs);
-
-static pfm_reg_desc_t pfm_ita_pmc_desc[PMU_MAX_PMCS]={
-/* pmc0  */ { PFM_REG_CONTROL , 0, 0x1UL, -1UL, NULL, NULL, {0UL,0UL, 0UL, 0UL}, {0UL,0UL, 0UL, 0UL}},
-/* pmc1  */ { PFM_REG_CONTROL , 0, 0x0UL, -1UL, NULL, NULL, {0UL,0UL, 0UL, 0UL}, {0UL,0UL, 0UL, 0UL}},
-/* pmc2  */ { PFM_REG_CONTROL , 0, 0x0UL, -1UL, NULL, NULL, {0UL,0UL, 0UL, 0UL}, {0UL,0UL, 0UL, 0UL}},
-/* pmc3  */ { PFM_REG_CONTROL , 0, 0x0UL, -1UL, NULL, NULL, {0UL,0UL, 0UL, 0UL}, {0UL,0UL, 0UL, 0UL}},
-/* pmc4  */ { PFM_REG_COUNTING, 6, 0x0UL, -1UL, NULL, NULL, {RDEP(4),0UL, 0UL, 0UL}, {0UL,0UL, 0UL, 0UL}},
-/* pmc5  */ { PFM_REG_COUNTING, 6, 0x0UL, -1UL, NULL, NULL, {RDEP(5),0UL, 0UL, 0UL}, {0UL,0UL, 0UL, 0UL}},
-/* pmc6  */ { PFM_REG_COUNTING, 6, 0x0UL, -1UL, NULL, NULL, {RDEP(6),0UL, 0UL, 0UL}, {0UL,0UL, 0UL, 0UL}},
-/* pmc7  */ { PFM_REG_COUNTING, 6, 0x0UL, -1UL, NULL, NULL, {RDEP(7),0UL, 0UL, 0UL}, {0UL,0UL, 0UL, 0UL}},
-/* pmc8  */ { PFM_REG_CONFIG  , 0, 0xf00000003ffffff8UL, -1UL, NULL, NULL, {0UL,0UL, 0UL, 0UL}, {0UL,0UL, 0UL, 0UL}},
-/* pmc9  */ { PFM_REG_CONFIG  , 0, 0xf00000003ffffff8UL, -1UL, NULL, NULL, {0UL,0UL, 0UL, 0UL}, {0UL,0UL, 0UL, 0UL}},
-/* pmc10 */ { PFM_REG_MONITOR , 6, 0x0UL, -1UL, NULL, NULL, {RDEP(0)|RDEP(1),0UL, 0UL, 0UL}, {0UL,0UL, 0UL, 0UL}},
-/* pmc11 */ { PFM_REG_MONITOR , 6, 0x0000000010000000UL, -1UL, NULL, pfm_ita_pmc_check, {RDEP(2)|RDEP(3)|RDEP(17),0UL, 0UL, 0UL}, {0UL,0UL, 0UL, 0UL}},
-/* pmc12 */ { PFM_REG_MONITOR , 6, 0x0UL, -1UL, NULL, NULL, {RDEP(8)|RDEP(9)|RDEP(10)|RDEP(11)|RDEP(12)|RDEP(13)|RDEP(14)|RDEP(15)|RDEP(16),0UL, 0UL, 0UL}, {0UL,0UL, 0UL, 0UL}},
-/* pmc13 */ { PFM_REG_CONFIG  , 0, 0x0003ffff00000001UL, -1UL, NULL, pfm_ita_pmc_check, {0UL,0UL, 0UL, 0UL}, {0UL,0UL, 0UL, 0UL}},
-           { PFM_REG_END     , 0, 0x0UL, -1UL, NULL, NULL, {0,}, {0,}}, /* end marker */
-};
-
-static pfm_reg_desc_t pfm_ita_pmd_desc[PMU_MAX_PMDS]={
-/* pmd0  */ { PFM_REG_BUFFER  , 0, 0UL, -1UL, NULL, NULL, {RDEP(1),0UL, 0UL, 0UL}, {RDEP(10),0UL, 0UL, 0UL}},
-/* pmd1  */ { PFM_REG_BUFFER  , 0, 0UL, -1UL, NULL, NULL, {RDEP(0),0UL, 0UL, 0UL}, {RDEP(10),0UL, 0UL, 0UL}},
-/* pmd2  */ { PFM_REG_BUFFER  , 0, 0UL, -1UL, NULL, NULL, {RDEP(3)|RDEP(17),0UL, 0UL, 0UL}, {RDEP(11),0UL, 0UL, 0UL}},
-/* pmd3  */ { PFM_REG_BUFFER  , 0, 0UL, -1UL, NULL, NULL, {RDEP(2)|RDEP(17),0UL, 0UL, 0UL}, {RDEP(11),0UL, 0UL, 0UL}},
-/* pmd4  */ { PFM_REG_COUNTING, 0, 0UL, -1UL, NULL, NULL, {0UL,0UL, 0UL, 0UL}, {RDEP(4),0UL, 0UL, 0UL}},
-/* pmd5  */ { PFM_REG_COUNTING, 0, 0UL, -1UL, NULL, NULL, {0UL,0UL, 0UL, 0UL}, {RDEP(5),0UL, 0UL, 0UL}},
-/* pmd6  */ { PFM_REG_COUNTING, 0, 0UL, -1UL, NULL, NULL, {0UL,0UL, 0UL, 0UL}, {RDEP(6),0UL, 0UL, 0UL}},
-/* pmd7  */ { PFM_REG_COUNTING, 0, 0UL, -1UL, NULL, NULL, {0UL,0UL, 0UL, 0UL}, {RDEP(7),0UL, 0UL, 0UL}},
-/* pmd8  */ { PFM_REG_BUFFER  , 0, 0UL, -1UL, NULL, NULL, {RDEP(9)|RDEP(10)|RDEP(11)|RDEP(12)|RDEP(13)|RDEP(14)|RDEP(15)|RDEP(16),0UL, 0UL, 0UL}, {RDEP(12),0UL, 0UL, 0UL}},
-/* pmd9  */ { PFM_REG_BUFFER  , 0, 0UL, -1UL, NULL, NULL, {RDEP(8)|RDEP(10)|RDEP(11)|RDEP(12)|RDEP(13)|RDEP(14)|RDEP(15)|RDEP(16),0UL, 0UL, 0UL}, {RDEP(12),0UL, 0UL, 0UL}},
-/* pmd10 */ { PFM_REG_BUFFER  , 0, 0UL, -1UL, NULL, NULL, {RDEP(8)|RDEP(9)|RDEP(11)|RDEP(12)|RDEP(13)|RDEP(14)|RDEP(15)|RDEP(16),0UL, 0UL, 0UL}, {RDEP(12),0UL, 0UL, 0UL}},
-/* pmd11 */ { PFM_REG_BUFFER  , 0, 0UL, -1UL, NULL, NULL, {RDEP(8)|RDEP(9)|RDEP(10)|RDEP(12)|RDEP(13)|RDEP(14)|RDEP(15)|RDEP(16),0UL, 0UL, 0UL}, {RDEP(12),0UL, 0UL, 0UL}},
-/* pmd12 */ { PFM_REG_BUFFER  , 0, 0UL, -1UL, NULL, NULL, {RDEP(8)|RDEP(9)|RDEP(10)|RDEP(11)|RDEP(13)|RDEP(14)|RDEP(15)|RDEP(16),0UL, 0UL, 0UL}, {RDEP(12),0UL, 0UL, 0UL}},
-/* pmd13 */ { PFM_REG_BUFFER  , 0, 0UL, -1UL, NULL, NULL, {RDEP(8)|RDEP(9)|RDEP(10)|RDEP(11)|RDEP(12)|RDEP(14)|RDEP(15)|RDEP(16),0UL, 0UL, 0UL}, {RDEP(12),0UL, 0UL, 0UL}},
-/* pmd14 */ { PFM_REG_BUFFER  , 0, 0UL, -1UL, NULL, NULL, {RDEP(8)|RDEP(9)|RDEP(10)|RDEP(11)|RDEP(12)|RDEP(13)|RDEP(15)|RDEP(16),0UL, 0UL, 0UL}, {RDEP(12),0UL, 0UL, 0UL}},
-/* pmd15 */ { PFM_REG_BUFFER  , 0, 0UL, -1UL, NULL, NULL, {RDEP(8)|RDEP(9)|RDEP(10)|RDEP(11)|RDEP(12)|RDEP(13)|RDEP(14)|RDEP(16),0UL, 0UL, 0UL}, {RDEP(12),0UL, 0UL, 0UL}},
-/* pmd16 */ { PFM_REG_BUFFER  , 0, 0UL, -1UL, NULL, NULL, {RDEP(8)|RDEP(9)|RDEP(10)|RDEP(11)|RDEP(12)|RDEP(13)|RDEP(14)|RDEP(15),0UL, 0UL, 0UL}, {RDEP(12),0UL, 0UL, 0UL}},
-/* pmd17 */ { PFM_REG_BUFFER  , 0, 0UL, -1UL, NULL, NULL, {RDEP(2)|RDEP(3),0UL, 0UL, 0UL}, {RDEP(11),0UL, 0UL, 0UL}},
-           { PFM_REG_END     , 0, 0UL, -1UL, NULL, NULL, {0,}, {0,}}, /* end marker */
-};
-
-static int
-pfm_ita_pmc_check(struct task_struct *task, pfm_context_t *ctx, unsigned int cnum, unsigned long *val, struct pt_regs *regs)
-{
-       int ret;
-       int is_loaded;
-
-       /* sanitfy check */
-       if (ctx == NULL) return -EINVAL;
-
-       is_loaded = ctx->ctx_state == PFM_CTX_LOADED || ctx->ctx_state == PFM_CTX_MASKED;
-
-       /*
-        * we must clear the (instruction) debug registers if pmc13.ta bit is cleared
-        * before they are written (fl_using_dbreg==0) to avoid picking up stale information.
-        */
-       if (cnum == 13 && is_loaded && ((*val & 0x1) == 0UL) && ctx->ctx_fl_using_dbreg == 0) {
-
-               DPRINT(("pmc[%d]=0x%lx has active pmc13.ta cleared, clearing ibr\n", cnum, *val));
-
-               /* don't mix debug with perfmon */
-               if (task && (task->thread.flags & IA64_THREAD_DBG_VALID) != 0) return -EINVAL;
-
-               /*
-                * a count of 0 will mark the debug registers as in use and also
-                * ensure that they are properly cleared.
-                */
-               ret = pfm_write_ibr_dbr(1, ctx, NULL, 0, regs);
-               if (ret) return ret;
-       }
-
-       /*
-        * we must clear the (data) debug registers if pmc11.pt bit is cleared
-        * before they are written (fl_using_dbreg==0) to avoid picking up stale information.
-        */
-       if (cnum == 11 && is_loaded && ((*val >> 28)& 0x1) == 0 && ctx->ctx_fl_using_dbreg == 0) {
-
-               DPRINT(("pmc[%d]=0x%lx has active pmc11.pt cleared, clearing dbr\n", cnum, *val));
-
-               /* don't mix debug with perfmon */
-               if (task && (task->thread.flags & IA64_THREAD_DBG_VALID) != 0) return -EINVAL;
-
-               /*
-                * a count of 0 will mark the debug registers as in use and also
-                * ensure that they are properly cleared.
-                */
-               ret = pfm_write_ibr_dbr(0, ctx, NULL, 0, regs);
-               if (ret) return ret;
-       }
-       return 0;
-}
-
-/*
- * impl_pmcs, impl_pmds are computed at runtime to minimize errors!
- */
-static pmu_config_t pmu_conf_ita={
-       .pmu_name      = "Itanium",
-       .pmu_family    = 0x7,
-       .ovfl_val      = (1UL << 32) - 1,
-       .pmd_desc      = pfm_ita_pmd_desc,
-       .pmc_desc      = pfm_ita_pmc_desc,
-       .num_ibrs      = 8,
-       .num_dbrs      = 8,
-       .use_rr_dbregs = 1, /* debug register are use for range retrictions */
-};
-
-
diff --git a/arch/ia64/kernel/perfmon_mckinley.h b/arch/ia64/kernel/perfmon_mckinley.h
deleted file mode 100644
index c4bec7a..0000000
--- a/arch/ia64/kernel/perfmon_mckinley.h
+++ /dev/null
@@ -1,187 +0,0 @@
-/*
- * This file contains the McKinley PMU register description tables
- * and pmc checker used by perfmon.c.
- *
- * Copyright (C) 2002-2003  Hewlett Packard Co
- *               Stephane Eranian <eranian@hpl.hp.com>
- */
-static int pfm_mck_pmc_check(struct task_struct *task, pfm_context_t *ctx, unsigned int cnum, unsigned long *val, struct pt_regs *regs);
-
-static pfm_reg_desc_t pfm_mck_pmc_desc[PMU_MAX_PMCS]={
-/* pmc0  */ { PFM_REG_CONTROL , 0, 0x1UL, -1UL, NULL, NULL, {0UL,0UL, 0UL, 0UL}, {0UL,0UL, 0UL, 0UL}},
-/* pmc1  */ { PFM_REG_CONTROL , 0, 0x0UL, -1UL, NULL, NULL, {0UL,0UL, 0UL, 0UL}, {0UL,0UL, 0UL, 0UL}},
-/* pmc2  */ { PFM_REG_CONTROL , 0, 0x0UL, -1UL, NULL, NULL, {0UL,0UL, 0UL, 0UL}, {0UL,0UL, 0UL, 0UL}},
-/* pmc3  */ { PFM_REG_CONTROL , 0, 0x0UL, -1UL, NULL, NULL, {0UL,0UL, 0UL, 0UL}, {0UL,0UL, 0UL, 0UL}},
-/* pmc4  */ { PFM_REG_COUNTING, 6, 0x0000000000800000UL, 0xfffff7fUL, NULL, pfm_mck_pmc_check, {RDEP(4),0UL, 0UL, 0UL}, {0UL,0UL, 0UL, 0UL}},
-/* pmc5  */ { PFM_REG_COUNTING, 6, 0x0UL, 0xfffff7fUL, NULL,  pfm_mck_pmc_check, {RDEP(5),0UL, 0UL, 0UL}, {0UL,0UL, 0UL, 0UL}},
-/* pmc6  */ { PFM_REG_COUNTING, 6, 0x0UL, 0xfffff7fUL, NULL,  pfm_mck_pmc_check, {RDEP(6),0UL, 0UL, 0UL}, {0UL,0UL, 0UL, 0UL}},
-/* pmc7  */ { PFM_REG_COUNTING, 6, 0x0UL, 0xfffff7fUL, NULL,  pfm_mck_pmc_check, {RDEP(7),0UL, 0UL, 0UL}, {0UL,0UL, 0UL, 0UL}},
-/* pmc8  */ { PFM_REG_CONFIG  , 0, 0xffffffff3fffffffUL, 0xffffffff3ffffffbUL, NULL, pfm_mck_pmc_check, {0UL,0UL, 0UL, 0UL}, {0UL,0UL, 0UL, 0UL}},
-/* pmc9  */ { PFM_REG_CONFIG  , 0, 0xffffffff3ffffffcUL, 0xffffffff3ffffffbUL, NULL, pfm_mck_pmc_check, {0UL,0UL, 0UL, 0UL}, {0UL,0UL, 0UL, 0UL}},
-/* pmc10 */ { PFM_REG_MONITOR , 4, 0x0UL, 0xffffUL, NULL, pfm_mck_pmc_check, {RDEP(0)|RDEP(1),0UL, 0UL, 0UL}, {0UL,0UL, 0UL, 0UL}},
-/* pmc11 */ { PFM_REG_MONITOR , 6, 0x0UL, 0x30f01cf, NULL,  pfm_mck_pmc_check, {RDEP(2)|RDEP(3)|RDEP(17),0UL, 0UL, 0UL}, {0UL,0UL, 0UL, 0UL}},
-/* pmc12 */ { PFM_REG_MONITOR , 6, 0x0UL, 0xffffUL, NULL,  pfm_mck_pmc_check, {RDEP(8)|RDEP(9)|RDEP(10)|RDEP(11)|RDEP(12)|RDEP(13)|RDEP(14)|RDEP(15)|RDEP(16),0UL, 0UL, 0UL}, {0UL,0UL, 0UL, 0UL}},
-/* pmc13 */ { PFM_REG_CONFIG  , 0, 0x00002078fefefefeUL, 0x1e00018181818UL, NULL, pfm_mck_pmc_check, {0UL,0UL, 0UL, 0UL}, {0UL,0UL, 0UL, 0UL}},
-/* pmc14 */ { PFM_REG_CONFIG  , 0, 0x0db60db60db60db6UL, 0x2492UL, NULL, pfm_mck_pmc_check, {0UL,0UL, 0UL, 0UL}, {0UL,0UL, 0UL, 0UL}},
-/* pmc15 */ { PFM_REG_CONFIG  , 0, 0x00000000fffffff0UL, 0xfUL, NULL, pfm_mck_pmc_check, {0UL,0UL, 0UL, 0UL}, {0UL,0UL, 0UL, 0UL}},
-           { PFM_REG_END     , 0, 0x0UL, -1UL, NULL, NULL, {0,}, {0,}}, /* end marker */
-};
-
-static pfm_reg_desc_t pfm_mck_pmd_desc[PMU_MAX_PMDS]={
-/* pmd0  */ { PFM_REG_BUFFER  , 0, 0x0UL, -1UL, NULL, NULL, {RDEP(1),0UL, 0UL, 0UL}, {RDEP(10),0UL, 0UL, 0UL}},
-/* pmd1  */ { PFM_REG_BUFFER  , 0, 0x0UL, -1UL, NULL, NULL, {RDEP(0),0UL, 0UL, 0UL}, {RDEP(10),0UL, 0UL, 0UL}},
-/* pmd2  */ { PFM_REG_BUFFER  , 0, 0x0UL, -1UL, NULL, NULL, {RDEP(3)|RDEP(17),0UL, 0UL, 0UL}, {RDEP(11),0UL, 0UL, 0UL}},
-/* pmd3  */ { PFM_REG_BUFFER  , 0, 0x0UL, -1UL, NULL, NULL, {RDEP(2)|RDEP(17),0UL, 0UL, 0UL}, {RDEP(11),0UL, 0UL, 0UL}},
-/* pmd4  */ { PFM_REG_COUNTING, 0, 0x0UL, -1UL, NULL, NULL, {0UL,0UL, 0UL, 0UL}, {RDEP(4),0UL, 0UL, 0UL}},
-/* pmd5  */ { PFM_REG_COUNTING, 0, 0x0UL, -1UL, NULL, NULL, {0UL,0UL, 0UL, 0UL}, {RDEP(5),0UL, 0UL, 0UL}},
-/* pmd6  */ { PFM_REG_COUNTING, 0, 0x0UL, -1UL, NULL, NULL, {0UL,0UL, 0UL, 0UL}, {RDEP(6),0UL, 0UL, 0UL}},
-/* pmd7  */ { PFM_REG_COUNTING, 0, 0x0UL, -1UL, NULL, NULL, {0UL,0UL, 0UL, 0UL}, {RDEP(7),0UL, 0UL, 0UL}},
-/* pmd8  */ { PFM_REG_BUFFER  , 0, 0x0UL, -1UL, NULL, NULL, {RDEP(9)|RDEP(10)|RDEP(11)|RDEP(12)|RDEP(13)|RDEP(14)|RDEP(15)|RDEP(16),0UL, 0UL, 0UL}, {RDEP(12),0UL, 0UL, 0UL}},
-/* pmd9  */ { PFM_REG_BUFFER  , 0, 0x0UL, -1UL, NULL, NULL, {RDEP(8)|RDEP(10)|RDEP(11)|RDEP(12)|RDEP(13)|RDEP(14)|RDEP(15)|RDEP(16),0UL, 0UL, 0UL}, {RDEP(12),0UL, 0UL, 0UL}},
-/* pmd10 */ { PFM_REG_BUFFER  , 0, 0x0UL, -1UL, NULL, NULL, {RDEP(8)|RDEP(9)|RDEP(11)|RDEP(12)|RDEP(13)|RDEP(14)|RDEP(15)|RDEP(16),0UL, 0UL, 0UL}, {RDEP(12),0UL, 0UL, 0UL}},
-/* pmd11 */ { PFM_REG_BUFFER  , 0, 0x0UL, -1UL, NULL, NULL, {RDEP(8)|RDEP(9)|RDEP(10)|RDEP(12)|RDEP(13)|RDEP(14)|RDEP(15)|RDEP(16),0UL, 0UL, 0UL}, {RDEP(12),0UL, 0UL, 0UL}},
-/* pmd12 */ { PFM_REG_BUFFER  , 0, 0x0UL, -1UL, NULL, NULL, {RDEP(8)|RDEP(9)|RDEP(10)|RDEP(11)|RDEP(13)|RDEP(14)|RDEP(15)|RDEP(16),0UL, 0UL, 0UL}, {RDEP(12),0UL, 0UL, 0UL}},
-/* pmd13 */ { PFM_REG_BUFFER  , 0, 0x0UL, -1UL, NULL, NULL, {RDEP(8)|RDEP(9)|RDEP(10)|RDEP(11)|RDEP(12)|RDEP(14)|RDEP(15)|RDEP(16),0UL, 0UL, 0UL}, {RDEP(12),0UL, 0UL, 0UL}},
-/* pmd14 */ { PFM_REG_BUFFER  , 0, 0x0UL, -1UL, NULL, NULL, {RDEP(8)|RDEP(9)|RDEP(10)|RDEP(11)|RDEP(12)|RDEP(13)|RDEP(15)|RDEP(16),0UL, 0UL, 0UL}, {RDEP(12),0UL, 0UL, 0UL}},
-/* pmd15 */ { PFM_REG_BUFFER  , 0, 0x0UL, -1UL, NULL, NULL, {RDEP(8)|RDEP(9)|RDEP(10)|RDEP(11)|RDEP(12)|RDEP(13)|RDEP(14)|RDEP(16),0UL, 0UL, 0UL}, {RDEP(12),0UL, 0UL, 0UL}},
-/* pmd16 */ { PFM_REG_BUFFER  , 0, 0x0UL, -1UL, NULL, NULL, {RDEP(8)|RDEP(9)|RDEP(10)|RDEP(11)|RDEP(12)|RDEP(13)|RDEP(14)|RDEP(15),0UL, 0UL, 0UL}, {RDEP(12),0UL, 0UL, 0UL}},
-/* pmd17 */ { PFM_REG_BUFFER  , 0, 0x0UL, -1UL, NULL, NULL, {RDEP(2)|RDEP(3),0UL, 0UL, 0UL}, {RDEP(11),0UL, 0UL, 0UL}},
-           { PFM_REG_END     , 0, 0x0UL, -1UL, NULL, NULL, {0,}, {0,}}, /* end marker */
-};
-
-/*
- * PMC reserved fields must have their power-up values preserved
- */
-static int
-pfm_mck_reserved(unsigned int cnum, unsigned long *val, struct pt_regs *regs)
-{
-       unsigned long tmp1, tmp2, ival = *val;
-
-       /* remove reserved areas from user value */
-       tmp1 = ival & PMC_RSVD_MASK(cnum);
-
-       /* get reserved fields values */
-       tmp2 = PMC_DFL_VAL(cnum) & ~PMC_RSVD_MASK(cnum);
-
-       *val = tmp1 | tmp2;
-
-       DPRINT(("pmc[%d]=0x%lx, mask=0x%lx, reset=0x%lx, val=0x%lx\n",
-                 cnum, ival, PMC_RSVD_MASK(cnum), PMC_DFL_VAL(cnum), *val));
-       return 0;
-}
-
-/*
- * task can be NULL if the context is unloaded
- */
-static int
-pfm_mck_pmc_check(struct task_struct *task, pfm_context_t *ctx, unsigned int cnum, unsigned long *val, struct pt_regs *regs)
-{
-       int ret = 0, check_case1 = 0;
-       unsigned long val8 = 0, val14 = 0, val13 = 0;
-       int is_loaded;
-
-       /* first preserve the reserved fields */
-       pfm_mck_reserved(cnum, val, regs);
-
-       /* sanitfy check */
-       if (ctx == NULL) return -EINVAL;
-
-       is_loaded = ctx->ctx_state == PFM_CTX_LOADED || ctx->ctx_state == PFM_CTX_MASKED;
-
-       /*
-        * we must clear the debug registers if pmc13 has a value which enable
-        * memory pipeline event constraints. In this case we need to clear the
-        * the debug registers if they have not yet been accessed. This is required
-        * to avoid picking stale state.
-        * PMC13 is "active" if:
-        *      one of the pmc13.cfg_dbrpXX field is different from 0x3
-        * AND
-        *      at the corresponding pmc13.ena_dbrpXX is set.
-        */
-       DPRINT(("cnum=%u val=0x%lx, using_dbreg=%d loaded=%d\n", cnum, *val, ctx->ctx_fl_using_dbreg, is_loaded));
-
-       if (cnum == 13 && is_loaded
-           && (*val & 0x1e00000000000UL) && (*val & 0x18181818UL) != 0x18181818UL && ctx->ctx_fl_using_dbreg == 0) {
-
-               DPRINT(("pmc[%d]=0x%lx has active pmc13 settings, clearing dbr\n", cnum, *val));
-
-               /* don't mix debug with perfmon */
-               if (task && (task->thread.flags & IA64_THREAD_DBG_VALID) != 0) return -EINVAL;
-
-               /*
-                * a count of 0 will mark the debug registers as in use and also
-                * ensure that they are properly cleared.
-                */
-               ret = pfm_write_ibr_dbr(PFM_DATA_RR, ctx, NULL, 0, regs);
-               if (ret) return ret;
-       }
-       /*
-        * we must clear the (instruction) debug registers if any pmc14.ibrpX bit is enabled
-        * before they are (fl_using_dbreg==0) to avoid picking up stale information.
-        */
-       if (cnum == 14 && is_loaded && ((*val & 0x2222UL) != 0x2222UL) && ctx->ctx_fl_using_dbreg == 0) {
-
-               DPRINT(("pmc[%d]=0x%lx has active pmc14 settings, clearing ibr\n", cnum, *val));
-
-               /* don't mix debug with perfmon */
-               if (task && (task->thread.flags & IA64_THREAD_DBG_VALID) != 0) return -EINVAL;
-
-               /*
-                * a count of 0 will mark the debug registers as in use and also
-                * ensure that they are properly cleared.
-                */
-               ret = pfm_write_ibr_dbr(PFM_CODE_RR, ctx, NULL, 0, regs);
-               if (ret) return ret;
-
-       }
-
-       switch(cnum) {
-               case  4: *val |= 1UL << 23; /* force power enable bit */
-                        break;
-               case  8: val8 = *val;
-                        val13 = ctx->ctx_pmcs[13];
-                        val14 = ctx->ctx_pmcs[14];
-                        check_case1 = 1;
-                        break;
-               case 13: val8  = ctx->ctx_pmcs[8];
-                        val13 = *val;
-                        val14 = ctx->ctx_pmcs[14];
-                        check_case1 = 1;
-                        break;
-               case 14: val8  = ctx->ctx_pmcs[8];
-                        val13 = ctx->ctx_pmcs[13];
-                        val14 = *val;
-                        check_case1 = 1;
-                        break;
-       }
-       /* check illegal configuration which can produce inconsistencies in tagging
-        * i-side events in L1D and L2 caches
-        */
-       if (check_case1) {
-               ret =   ((val13 >> 45) & 0xf) == 0
-                  && ((val8 & 0x1) == 0)
-                  && ((((val14>>1) & 0x3) == 0x2 || ((val14>>1) & 0x3) == 0x0)
-                      ||(((val14>>4) & 0x3) == 0x2 || ((val14>>4) & 0x3) == 0x0));
-
-               if (ret) DPRINT((KERN_DEBUG "perfmon: failure check_case1\n"));
-       }
-
-       return ret ? -EINVAL : 0;
-}
-
-/*
- * impl_pmcs, impl_pmds are computed at runtime to minimize errors!
- */
-static pmu_config_t pmu_conf_mck={
-       .pmu_name      = "Itanium 2",
-       .pmu_family    = 0x1f,
-       .flags         = PFM_PMU_IRQ_RESEND,
-       .ovfl_val      = (1UL << 47) - 1,
-       .pmd_desc      = pfm_mck_pmd_desc,
-       .pmc_desc      = pfm_mck_pmc_desc,
-       .num_ibrs       = 8,
-       .num_dbrs       = 8,
-       .use_rr_dbregs = 1 /* debug register are use for range restrictions */
-};
-
-
diff --git a/arch/ia64/kernel/perfmon_montecito.h b/arch/ia64/kernel/perfmon_montecito.h
deleted file mode 100644
index 7f8da4c..0000000
--- a/arch/ia64/kernel/perfmon_montecito.h
+++ /dev/null
@@ -1,269 +0,0 @@
-/*
- * This file contains the Montecito PMU register description tables
- * and pmc checker used by perfmon.c.
- *
- * Copyright (c) 2005-2006 Hewlett-Packard Development Company, L.P.
- *               Contributed by Stephane Eranian <eranian@hpl.hp.com>
- */
-static int pfm_mont_pmc_check(struct task_struct *task, pfm_context_t *ctx, unsigned int cnum, unsigned long *val, struct pt_regs *regs);
-
-#define RDEP_MONT_ETB  (RDEP(38)|RDEP(39)|RDEP(48)|RDEP(49)|RDEP(50)|RDEP(51)|RDEP(52)|RDEP(53)|RDEP(54)|\
-                        RDEP(55)|RDEP(56)|RDEP(57)|RDEP(58)|RDEP(59)|RDEP(60)|RDEP(61)|RDEP(62)|RDEP(63))
-#define RDEP_MONT_DEAR  (RDEP(32)|RDEP(33)|RDEP(36))
-#define RDEP_MONT_IEAR  (RDEP(34)|RDEP(35))
-
-static pfm_reg_desc_t pfm_mont_pmc_desc[PMU_MAX_PMCS]={
-/* pmc0  */ { PFM_REG_CONTROL , 0, 0x0, -1, NULL, NULL, {0,0, 0, 0}, {0,0, 0, 0}},
-/* pmc1  */ { PFM_REG_CONTROL , 0, 0x0, -1, NULL, NULL, {0,0, 0, 0}, {0,0, 0, 0}},
-/* pmc2  */ { PFM_REG_CONTROL , 0, 0x0, -1, NULL, NULL, {0,0, 0, 0}, {0,0, 0, 0}},
-/* pmc3  */ { PFM_REG_CONTROL , 0, 0x0, -1, NULL, NULL, {0,0, 0, 0}, {0,0, 0, 0}},
-/* pmc4  */ { PFM_REG_COUNTING, 6, 0x2000000, 0x7c7fff7f, NULL, pfm_mont_pmc_check, {RDEP(4),0, 0, 0}, {0,0, 0, 0}},
-/* pmc5  */ { PFM_REG_COUNTING, 6, 0x2000000, 0x7c7fff7f, NULL, pfm_mont_pmc_check, {RDEP(5),0, 0, 0}, {0,0, 0, 0}},
-/* pmc6  */ { PFM_REG_COUNTING, 6, 0x2000000, 0x7c7fff7f, NULL, pfm_mont_pmc_check, {RDEP(6),0, 0, 0}, {0,0, 0, 0}},
-/* pmc7  */ { PFM_REG_COUNTING, 6, 0x2000000, 0x7c7fff7f, NULL, pfm_mont_pmc_check, {RDEP(7),0, 0, 0}, {0,0, 0, 0}},
-/* pmc8  */ { PFM_REG_COUNTING, 6, 0x2000000, 0x7c7fff7f, NULL, pfm_mont_pmc_check, {RDEP(8),0, 0, 0}, {0,0, 0, 0}},
-/* pmc9  */ { PFM_REG_COUNTING, 6, 0x2000000, 0x7c7fff7f, NULL, pfm_mont_pmc_check, {RDEP(9),0, 0, 0}, {0,0, 0, 0}},
-/* pmc10 */ { PFM_REG_COUNTING, 6, 0x2000000, 0x7c7fff7f, NULL, pfm_mont_pmc_check, {RDEP(10),0, 0, 0}, {0,0, 0, 0}},
-/* pmc11 */ { PFM_REG_COUNTING, 6, 0x2000000, 0x7c7fff7f, NULL, pfm_mont_pmc_check, {RDEP(11),0, 0, 0}, {0,0, 0, 0}},
-/* pmc12 */ { PFM_REG_COUNTING, 6, 0x2000000, 0x7c7fff7f, NULL, pfm_mont_pmc_check, {RDEP(12),0, 0, 0}, {0,0, 0, 0}},
-/* pmc13 */ { PFM_REG_COUNTING, 6, 0x2000000, 0x7c7fff7f, NULL, pfm_mont_pmc_check, {RDEP(13),0, 0, 0}, {0,0, 0, 0}},
-/* pmc14 */ { PFM_REG_COUNTING, 6, 0x2000000, 0x7c7fff7f, NULL, pfm_mont_pmc_check, {RDEP(14),0, 0, 0}, {0,0, 0, 0}},
-/* pmc15 */ { PFM_REG_COUNTING, 6, 0x2000000, 0x7c7fff7f, NULL, pfm_mont_pmc_check, {RDEP(15),0, 0, 0}, {0,0, 0, 0}},
-/* pmc16 */ { PFM_REG_NOTIMPL, },
-/* pmc17 */ { PFM_REG_NOTIMPL, },
-/* pmc18 */ { PFM_REG_NOTIMPL, },
-/* pmc19 */ { PFM_REG_NOTIMPL, },
-/* pmc20 */ { PFM_REG_NOTIMPL, },
-/* pmc21 */ { PFM_REG_NOTIMPL, },
-/* pmc22 */ { PFM_REG_NOTIMPL, },
-/* pmc23 */ { PFM_REG_NOTIMPL, },
-/* pmc24 */ { PFM_REG_NOTIMPL, },
-/* pmc25 */ { PFM_REG_NOTIMPL, },
-/* pmc26 */ { PFM_REG_NOTIMPL, },
-/* pmc27 */ { PFM_REG_NOTIMPL, },
-/* pmc28 */ { PFM_REG_NOTIMPL, },
-/* pmc29 */ { PFM_REG_NOTIMPL, },
-/* pmc30 */ { PFM_REG_NOTIMPL, },
-/* pmc31 */ { PFM_REG_NOTIMPL, },
-/* pmc32 */ { PFM_REG_CONFIG,  0, 0x30f01ffffffffffUL, 0x30f01ffffffffffUL, NULL, pfm_mont_pmc_check, {0,0, 0, 0}, {0,0, 0, 0}},
-/* pmc33 */ { PFM_REG_CONFIG,  0, 0x0,  0x1ffffffffffUL, NULL, pfm_mont_pmc_check, {0,0, 0, 0}, {0,0, 0, 0}},
-/* pmc34 */ { PFM_REG_CONFIG,  0, 0xf01ffffffffffUL, 0xf01ffffffffffUL, NULL, pfm_mont_pmc_check, {0,0, 0, 0}, {0,0, 0, 0}},
-/* pmc35 */ { PFM_REG_CONFIG,  0, 0x0,  0x1ffffffffffUL, NULL, pfm_mont_pmc_check, {0,0, 0, 0}, {0,0, 0, 0}},
-/* pmc36 */ { PFM_REG_CONFIG,  0, 0xfffffff0, 0xf, NULL, pfm_mont_pmc_check, {0,0, 0, 0}, {0,0, 0, 0}},
-/* pmc37 */ { PFM_REG_MONITOR, 4, 0x0, 0x3fff, NULL, pfm_mont_pmc_check, {RDEP_MONT_IEAR, 0, 0, 0}, {0, 0, 0, 0}},
-/* pmc38 */ { PFM_REG_CONFIG,  0, 0xdb6, 0x2492, NULL, pfm_mont_pmc_check, {0,0, 0, 0}, {0,0, 0, 0}},
-/* pmc39 */ { PFM_REG_MONITOR, 6, 0x0, 0xffcf, NULL, pfm_mont_pmc_check, {RDEP_MONT_ETB,0, 0, 0}, {0,0, 0, 0}},
-/* pmc40 */ { PFM_REG_MONITOR, 6, 0x2000000, 0xf01cf, NULL, pfm_mont_pmc_check, {RDEP_MONT_DEAR,0, 0, 0}, {0,0, 0, 0}},
-/* pmc41 */ { PFM_REG_CONFIG,  0, 0x00002078fefefefeUL, 0x1e00018181818UL, NULL, pfm_mont_pmc_check, {0,0, 0, 0}, {0,0, 0, 0}},
-/* pmc42 */ { PFM_REG_MONITOR, 6, 0x0, 0x7ff4f, NULL, pfm_mont_pmc_check, {RDEP_MONT_ETB,0, 0, 0}, {0,0, 0, 0}},
-           { PFM_REG_END    , 0, 0x0, -1, NULL, NULL, {0,}, {0,}}, /* end marker */
-};
-
-static pfm_reg_desc_t pfm_mont_pmd_desc[PMU_MAX_PMDS]={
-/* pmd0  */ { PFM_REG_NOTIMPL, }, 
-/* pmd1  */ { PFM_REG_NOTIMPL, },
-/* pmd2  */ { PFM_REG_NOTIMPL, },
-/* pmd3  */ { PFM_REG_NOTIMPL, },
-/* pmd4  */ { PFM_REG_COUNTING, 0, 0x0, -1, NULL, NULL, {0,0, 0, 0}, {RDEP(4),0, 0, 0}},
-/* pmd5  */ { PFM_REG_COUNTING, 0, 0x0, -1, NULL, NULL, {0,0, 0, 0}, {RDEP(5),0, 0, 0}},
-/* pmd6  */ { PFM_REG_COUNTING, 0, 0x0, -1, NULL, NULL, {0,0, 0, 0}, {RDEP(6),0, 0, 0}},
-/* pmd7  */ { PFM_REG_COUNTING, 0, 0x0, -1, NULL, NULL, {0,0, 0, 0}, {RDEP(7),0, 0, 0}},
-/* pmd8  */ { PFM_REG_COUNTING, 0, 0x0, -1, NULL, NULL, {0,0, 0, 0}, {RDEP(8),0, 0, 0}}, 
-/* pmd9  */ { PFM_REG_COUNTING, 0, 0x0, -1, NULL, NULL, {0,0, 0, 0}, {RDEP(9),0, 0, 0}},
-/* pmd10 */ { PFM_REG_COUNTING, 0, 0x0, -1, NULL, NULL, {0,0, 0, 0}, {RDEP(10),0, 0, 0}},
-/* pmd11 */ { PFM_REG_COUNTING, 0, 0x0, -1, NULL, NULL, {0,0, 0, 0}, {RDEP(11),0, 0, 0}},
-/* pmd12 */ { PFM_REG_COUNTING, 0, 0x0, -1, NULL, NULL, {0,0, 0, 0}, {RDEP(12),0, 0, 0}},
-/* pmd13 */ { PFM_REG_COUNTING, 0, 0x0, -1, NULL, NULL, {0,0, 0, 0}, {RDEP(13),0, 0, 0}},
-/* pmd14 */ { PFM_REG_COUNTING, 0, 0x0, -1, NULL, NULL, {0,0, 0, 0}, {RDEP(14),0, 0, 0}},
-/* pmd15 */ { PFM_REG_COUNTING, 0, 0x0, -1, NULL, NULL, {0,0, 0, 0}, {RDEP(15),0, 0, 0}},
-/* pmd16 */ { PFM_REG_NOTIMPL, },
-/* pmd17 */ { PFM_REG_NOTIMPL, },
-/* pmd18 */ { PFM_REG_NOTIMPL, },
-/* pmd19 */ { PFM_REG_NOTIMPL, },
-/* pmd20 */ { PFM_REG_NOTIMPL, },
-/* pmd21 */ { PFM_REG_NOTIMPL, },
-/* pmd22 */ { PFM_REG_NOTIMPL, },
-/* pmd23 */ { PFM_REG_NOTIMPL, },
-/* pmd24 */ { PFM_REG_NOTIMPL, },
-/* pmd25 */ { PFM_REG_NOTIMPL, },
-/* pmd26 */ { PFM_REG_NOTIMPL, },
-/* pmd27 */ { PFM_REG_NOTIMPL, },
-/* pmd28 */ { PFM_REG_NOTIMPL, },
-/* pmd29 */ { PFM_REG_NOTIMPL, },
-/* pmd30 */ { PFM_REG_NOTIMPL, },
-/* pmd31 */ { PFM_REG_NOTIMPL, },
-/* pmd32 */ { PFM_REG_BUFFER, 0, 0x0, -1, NULL, NULL, {RDEP(33)|RDEP(36),0, 0, 0}, {RDEP(40),0, 0, 0}},
-/* pmd33 */ { PFM_REG_BUFFER, 0, 0x0, -1, NULL, NULL, {RDEP(32)|RDEP(36),0, 0, 0}, {RDEP(40),0, 0, 0}},
-/* pmd34 */ { PFM_REG_BUFFER, 0, 0x0, -1, NULL, NULL, {RDEP(35),0, 0, 0}, {RDEP(37),0, 0, 0}},
-/* pmd35 */ { PFM_REG_BUFFER, 0, 0x0, -1, NULL, NULL, {RDEP(34),0, 0, 0}, {RDEP(37),0, 0, 0}},
-/* pmd36 */ { PFM_REG_BUFFER, 0, 0x0, -1, NULL, NULL, {RDEP(32)|RDEP(33),0, 0, 0}, {RDEP(40),0, 0, 0}},
-/* pmd37 */ { PFM_REG_NOTIMPL, },
-/* pmd38 */ { PFM_REG_BUFFER, 0, 0x0, -1, NULL, NULL, {RDEP_MONT_ETB,0, 0, 0}, {RDEP(39),0, 0, 0}},
-/* pmd39 */ { PFM_REG_BUFFER, 0, 0x0, -1, NULL, NULL, {RDEP_MONT_ETB,0, 0, 0}, {RDEP(39),0, 0, 0}},
-/* pmd40 */ { PFM_REG_NOTIMPL, },
-/* pmd41 */ { PFM_REG_NOTIMPL, },
-/* pmd42 */ { PFM_REG_NOTIMPL, },
-/* pmd43 */ { PFM_REG_NOTIMPL, },
-/* pmd44 */ { PFM_REG_NOTIMPL, },
-/* pmd45 */ { PFM_REG_NOTIMPL, },
-/* pmd46 */ { PFM_REG_NOTIMPL, },
-/* pmd47 */ { PFM_REG_NOTIMPL, },
-/* pmd48 */ { PFM_REG_BUFFER, 0, 0x0, -1, NULL, NULL, {RDEP_MONT_ETB,0, 0, 0}, {RDEP(39),0, 0, 0}},
-/* pmd49 */ { PFM_REG_BUFFER, 0, 0x0, -1, NULL, NULL, {RDEP_MONT_ETB,0, 0, 0}, {RDEP(39),0, 0, 0}},
-/* pmd50 */ { PFM_REG_BUFFER, 0, 0x0, -1, NULL, NULL, {RDEP_MONT_ETB,0, 0, 0}, {RDEP(39),0, 0, 0}},
-/* pmd51 */ { PFM_REG_BUFFER, 0, 0x0, -1, NULL, NULL, {RDEP_MONT_ETB,0, 0, 0}, {RDEP(39),0, 0, 0}},
-/* pmd52 */ { PFM_REG_BUFFER, 0, 0x0, -1, NULL, NULL, {RDEP_MONT_ETB,0, 0, 0}, {RDEP(39),0, 0, 0}},
-/* pmd53 */ { PFM_REG_BUFFER, 0, 0x0, -1, NULL, NULL, {RDEP_MONT_ETB,0, 0, 0}, {RDEP(39),0, 0, 0}},
-/* pmd54 */ { PFM_REG_BUFFER, 0, 0x0, -1, NULL, NULL, {RDEP_MONT_ETB,0, 0, 0}, {RDEP(39),0, 0, 0}},
-/* pmd55 */ { PFM_REG_BUFFER, 0, 0x0, -1, NULL, NULL, {RDEP_MONT_ETB,0, 0, 0}, {RDEP(39),0, 0, 0}},
-/* pmd56 */ { PFM_REG_BUFFER, 0, 0x0, -1, NULL, NULL, {RDEP_MONT_ETB,0, 0, 0}, {RDEP(39),0, 0, 0}},
-/* pmd57 */ { PFM_REG_BUFFER, 0, 0x0, -1, NULL, NULL, {RDEP_MONT_ETB,0, 0, 0}, {RDEP(39),0, 0, 0}},
-/* pmd58 */ { PFM_REG_BUFFER, 0, 0x0, -1, NULL, NULL, {RDEP_MONT_ETB,0, 0, 0}, {RDEP(39),0, 0, 0}},
-/* pmd59 */ { PFM_REG_BUFFER, 0, 0x0, -1, NULL, NULL, {RDEP_MONT_ETB,0, 0, 0}, {RDEP(39),0, 0, 0}},
-/* pmd60 */ { PFM_REG_BUFFER, 0, 0x0, -1, NULL, NULL, {RDEP_MONT_ETB,0, 0, 0}, {RDEP(39),0, 0, 0}},
-/* pmd61 */ { PFM_REG_BUFFER, 0, 0x0, -1, NULL, NULL, {RDEP_MONT_ETB,0, 0, 0}, {RDEP(39),0, 0, 0}},
-/* pmd62 */ { PFM_REG_BUFFER, 0, 0x0, -1, NULL, NULL, {RDEP_MONT_ETB,0, 0, 0}, {RDEP(39),0, 0, 0}},
-/* pmd63 */ { PFM_REG_BUFFER, 0, 0x0, -1, NULL, NULL, {RDEP_MONT_ETB,0, 0, 0}, {RDEP(39),0, 0, 0}},
-           { PFM_REG_END   , 0, 0x0, -1, NULL, NULL, {0,}, {0,}}, /* end marker */
-};
-
-/*
- * PMC reserved fields must have their power-up values preserved
- */
-static int
-pfm_mont_reserved(unsigned int cnum, unsigned long *val, struct pt_regs *regs)
-{
-       unsigned long tmp1, tmp2, ival = *val;
-
-       /* remove reserved areas from user value */
-       tmp1 = ival & PMC_RSVD_MASK(cnum);
-
-       /* get reserved fields values */
-       tmp2 = PMC_DFL_VAL(cnum) & ~PMC_RSVD_MASK(cnum);
-
-       *val = tmp1 | tmp2;
-
-       DPRINT(("pmc[%d]=0x%lx, mask=0x%lx, reset=0x%lx, val=0x%lx\n",
-                 cnum, ival, PMC_RSVD_MASK(cnum), PMC_DFL_VAL(cnum), *val));
-       return 0;
-}
-
-/*
- * task can be NULL if the context is unloaded
- */
-static int
-pfm_mont_pmc_check(struct task_struct *task, pfm_context_t *ctx, unsigned int cnum, unsigned long *val, struct pt_regs *regs)
-{
-       int ret = 0;
-       unsigned long val32 = 0, val38 = 0, val41 = 0;
-       unsigned long tmpval;
-       int check_case1 = 0;
-       int is_loaded;
-
-       /* first preserve the reserved fields */
-       pfm_mont_reserved(cnum, val, regs);
-
-       tmpval = *val;
-
-       /* sanity check */
-       if (ctx == NULL) return -EINVAL;
-
-       is_loaded = ctx->ctx_state == PFM_CTX_LOADED || ctx->ctx_state == PFM_CTX_MASKED;
-
-       /*
-        * we must clear the debug registers if pmc41 has a value which enable
-        * memory pipeline event constraints. In this case we need to clear the
-        * the debug registers if they have not yet been accessed. This is required
-        * to avoid picking stale state.
-        * PMC41 is "active" if:
-        *      one of the pmc41.cfg_dtagXX field is different from 0x3
-        * AND
-        *      at the corresponding pmc41.en_dbrpXX is set.
-        * AND
-        *      ctx_fl_using_dbreg == 0  (i.e., dbr not yet used)
-        */
-       DPRINT(("cnum=%u val=0x%lx, using_dbreg=%d loaded=%d\n", cnum, tmpval, ctx->ctx_fl_using_dbreg, is_loaded));
-
-       if (cnum == 41 && is_loaded 
-           && (tmpval & 0x1e00000000000UL) && (tmpval & 0x18181818UL) != 0x18181818UL && ctx->ctx_fl_using_dbreg == 0) {
-
-               DPRINT(("pmc[%d]=0x%lx has active pmc41 settings, clearing dbr\n", cnum, tmpval));
-
-               /* don't mix debug with perfmon */
-               if (task && (task->thread.flags & IA64_THREAD_DBG_VALID) != 0) return -EINVAL;
-
-               /*
-                * a count of 0 will mark the debug registers if:
-                * AND
-                */
-               ret = pfm_write_ibr_dbr(PFM_DATA_RR, ctx, NULL, 0, regs);
-               if (ret) return ret;
-       }
-       /*
-        * we must clear the (instruction) debug registers if:
-        *      pmc38.ig_ibrpX is 0 (enabled)
-        * AND
-        *      ctx_fl_using_dbreg == 0  (i.e., dbr not yet used)
-        */
-       if (cnum == 38 && is_loaded && ((tmpval & 0x492UL) != 0x492UL) && ctx->ctx_fl_using_dbreg == 0) {
-
-               DPRINT(("pmc38=0x%lx has active pmc38 settings, clearing ibr\n", tmpval));
-
-               /* don't mix debug with perfmon */
-               if (task && (task->thread.flags & IA64_THREAD_DBG_VALID) != 0) return -EINVAL;
-
-               /*
-                * a count of 0 will mark the debug registers as in use and also
-                * ensure that they are properly cleared.
-                */
-               ret = pfm_write_ibr_dbr(PFM_CODE_RR, ctx, NULL, 0, regs);
-               if (ret) return ret;
-
-       }
-       switch(cnum) {
-               case  32: val32 = *val;
-                         val38 = ctx->ctx_pmcs[38];
-                         val41 = ctx->ctx_pmcs[41];
-                         check_case1 = 1;
-                         break;
-               case  38: val38 = *val;
-                         val32 = ctx->ctx_pmcs[32];
-                         val41 = ctx->ctx_pmcs[41];
-                         check_case1 = 1;
-                         break;
-               case  41: val41 = *val;
-                         val32 = ctx->ctx_pmcs[32];
-                         val38 = ctx->ctx_pmcs[38];
-                         check_case1 = 1;
-                         break;
-       }
-       /* check illegal configuration which can produce inconsistencies in tagging
-        * i-side events in L1D and L2 caches
-        */
-       if (check_case1) {
-               ret =   (((val41 >> 45) & 0xf) == 0 && ((val32>>57) & 0x1) == 0)
-                    && ((((val38>>1) & 0x3) == 0x2 || ((val38>>1) & 0x3) == 0)
-                    ||  (((val38>>4) & 0x3) == 0x2 || ((val38>>4) & 0x3) == 0));
-               if (ret) {
-                       DPRINT(("invalid config pmc38=0x%lx pmc41=0x%lx pmc32=0x%lx\n", val38, val41, val32));
-                       return -EINVAL;
-               }
-       }
-       *val = tmpval;
-       return 0;
-}
-
-/*
- * impl_pmcs, impl_pmds are computed at runtime to minimize errors!
- */
-static pmu_config_t pmu_conf_mont={
-       .pmu_name        = "Montecito",
-       .pmu_family      = 0x20,
-       .flags           = PFM_PMU_IRQ_RESEND,
-       .ovfl_val        = (1UL << 47) - 1,
-       .pmd_desc        = pfm_mont_pmd_desc,
-       .pmc_desc        = pfm_mont_pmc_desc,
-       .num_ibrs        = 8,
-       .num_dbrs        = 8,
-       .use_rr_dbregs   = 1 /* debug register are use for range retrictions */
-};
diff --git a/arch/ia64/kernel/process.c b/arch/ia64/kernel/process.c
index 3ab8373..a7dfb39 100644
--- a/arch/ia64/kernel/process.c
+++ b/arch/ia64/kernel/process.c
@@ -28,6 +28,7 @@
 #include <linux/delay.h>
 #include <linux/kdebug.h>
 #include <linux/utsname.h>
+#include <linux/perfmon_kern.h>
 
 #include <asm/cpu.h>
 #include <asm/delay.h>
@@ -45,10 +46,6 @@
 
 #include "entry.h"
 
-#ifdef CONFIG_PERFMON
-# include <asm/perfmon.h>
-#endif
-
 #include "sigframe.h"
 
 void (*ia64_mark_idle)(int);
@@ -162,10 +159,8 @@ show_regs (struct pt_regs *regs)
 
 void tsk_clear_notify_resume(struct task_struct *tsk)
 {
-#ifdef CONFIG_PERFMON
-       if (tsk->thread.pfm_needs_checking)
+       if (test_ti_thread_flag(task_thread_info(tsk), TIF_PERFMON_WORK))
                return;
-#endif
        if (test_ti_thread_flag(task_thread_info(tsk), TIF_RESTORE_RSE))
                return;
        clear_ti_thread_flag(task_thread_info(tsk), TIF_NOTIFY_RESUME);
@@ -188,14 +183,9 @@ do_notify_resume_user(sigset_t *unused, struct sigscratch *scr, long in_syscall)
                return;
        }
 
-#ifdef CONFIG_PERFMON
-       if (current->thread.pfm_needs_checking)
-               /*
-                * Note: pfm_handle_work() allow us to call it with interrupts
-                * disabled, and may enable interrupts within the function.
-                */
-               pfm_handle_work();
-#endif
+       /* process perfmon asynchronous work (e.g. block thread or reset) */
+       if (test_thread_flag(TIF_PERFMON_WORK))
+               pfm_handle_work(task_pt_regs(current));
 
        /* deal with pending signal delivery */
        if (test_thread_flag(TIF_SIGPENDING)) {
@@ -212,22 +202,15 @@ do_notify_resume_user(sigset_t *unused, struct sigscratch *scr, long in_syscall)
        local_irq_disable();    /* force interrupt disable */
 }
 
-static int pal_halt        = 1;
 static int can_do_pal_halt = 1;
 
 static int __init nohalt_setup(char * str)
 {
-       pal_halt = can_do_pal_halt = 0;
+       can_do_pal_halt = 0;
        return 1;
 }
 __setup("nohalt", nohalt_setup);
 
-void
-update_pal_halt_status(int status)
-{
-       can_do_pal_halt = pal_halt && status;
-}
-
 /*
  * We use this if we don't have any better idle routine..
  */
@@ -236,6 +219,22 @@ default_idle (void)
 {
        local_irq_enable();
        while (!need_resched()) {
+#ifdef CONFIG_PERFMON
+               u64 psr = 0;
+               /*
+                * If requested, we stop the PMU to avoid
+                * measuring across the core idle loop.
+                *
+                * dcr.pp is not modified on purpose
+                * it is used when coming out of
+                * safe_halt() via interrupt
+                */
+               if ((__get_cpu_var(pfm_syst_info) & PFM_ITA_CPUINFO_IDLE_EXCL)) {
+                       psr = ia64_getreg(_IA64_REG_PSR);
+                       if (psr & IA64_PSR_PP)
+                               ia64_rsm(IA64_PSR_PP);
+               }
+#endif
                if (can_do_pal_halt) {
                        local_irq_disable();
                        if (!need_resched()) {
@@ -244,6 +243,12 @@ default_idle (void)
                        local_irq_enable();
                } else
                        cpu_relax();
+#ifdef CONFIG_PERFMON
+               if ((__get_cpu_var(pfm_syst_info) & PFM_ITA_CPUINFO_IDLE_EXCL)) {
+                       if (psr & IA64_PSR_PP)
+                               ia64_ssm(IA64_PSR_PP);
+               }
+#endif
        }
 }
 
@@ -344,22 +349,9 @@ cpu_idle (void)
 void
 ia64_save_extra (struct task_struct *task)
 {
-#ifdef CONFIG_PERFMON
-       unsigned long info;
-#endif
-
        if ((task->thread.flags & IA64_THREAD_DBG_VALID) != 0)
                ia64_save_debug_regs(&task->thread.dbr[0]);
 
-#ifdef CONFIG_PERFMON
-       if ((task->thread.flags & IA64_THREAD_PM_VALID) != 0)
-               pfm_save_regs(task);
-
-       info = __get_cpu_var(pfm_syst_info);
-       if (info & PFM_CPUINFO_SYST_WIDE)
-               pfm_syst_wide_update_task(task, info, 0);
-#endif
-
 #ifdef CONFIG_IA32_SUPPORT
        if (IS_IA32_PROCESS(task_pt_regs(task)))
                ia32_save_state(task);
@@ -369,22 +361,9 @@ ia64_save_extra (struct task_struct *task)
 void
 ia64_load_extra (struct task_struct *task)
 {
-#ifdef CONFIG_PERFMON
-       unsigned long info;
-#endif
-
        if ((task->thread.flags & IA64_THREAD_DBG_VALID) != 0)
                ia64_load_debug_regs(&task->thread.dbr[0]);
 
-#ifdef CONFIG_PERFMON
-       if ((task->thread.flags & IA64_THREAD_PM_VALID) != 0)
-               pfm_load_regs(task);
-
-       info = __get_cpu_var(pfm_syst_info);
-       if (info & PFM_CPUINFO_SYST_WIDE) 
-               pfm_syst_wide_update_task(task, info, 1);
-#endif
-
 #ifdef CONFIG_IA32_SUPPORT
        if (IS_IA32_PROCESS(task_pt_regs(task)))
                ia32_load_state(task);
@@ -510,8 +489,7 @@ copy_thread (int nr, unsigned long clone_flags,
         * call behavior where scratch registers are preserved across
         * system calls (unless used by the system call itself).
         */
-#      define THREAD_FLAGS_TO_CLEAR    (IA64_THREAD_FPH_VALID | IA64_THREAD_DBG_VALID \
-                                        | IA64_THREAD_PM_VALID)
+#      define THREAD_FLAGS_TO_CLEAR    (IA64_THREAD_FPH_VALID | IA64_THREAD_DBG_VALID)
 #      define THREAD_FLAGS_TO_SET      0
        p->thread.flags = ((current->thread.flags & ~THREAD_FLAGS_TO_CLEAR)
                           | THREAD_FLAGS_TO_SET);
@@ -533,10 +511,8 @@ copy_thread (int nr, unsigned long clone_flags,
        }
 #endif
 
-#ifdef CONFIG_PERFMON
-       if (current->thread.pfm_context)
-               pfm_inherit(p, child_ptregs);
-#endif
+       pfm_copy_thread(p);
+
        return retval;
 }
 
@@ -745,15 +721,13 @@ exit_thread (void)
 {
 
        ia64_drop_fpu(current);
-#ifdef CONFIG_PERFMON
-       /* if needed, stop monitoring and flush state to perfmon context */
-       if (current->thread.pfm_context)
-               pfm_exit_thread(current);
+
+       /* if needed, stop monitoring and flush state to perfmon context */
+       pfm_exit_thread();
 
        /* free debug register resources */
-       if (current->thread.flags & IA64_THREAD_DBG_VALID)
-               pfm_release_debug_registers(current);
-#endif
+       pfm_release_dbregs(current);
+
        if (IS_IA32_PROCESS(task_pt_regs(current)))
                ia32_drop_ia64_partial_page_list(current);
 }
diff --git a/arch/ia64/kernel/ptrace.c b/arch/ia64/kernel/ptrace.c
index 2a9943b..bb1ca1e 100644
--- a/arch/ia64/kernel/ptrace.c
+++ b/arch/ia64/kernel/ptrace.c
@@ -20,6 +20,7 @@
 #include <linux/security.h>
 #include <linux/audit.h>
 #include <linux/signal.h>
+#include <linux/perfmon_kern.h>
 #include <linux/regset.h>
 #include <linux/elf.h>
 
@@ -30,9 +31,6 @@
 #include <asm/system.h>
 #include <asm/uaccess.h>
 #include <asm/unwind.h>
-#ifdef CONFIG_PERFMON
-#include <asm/perfmon.h>
-#endif
 
 #include "entry.h"
 
@@ -2124,7 +2122,6 @@ access_uarea(struct task_struct *child, unsigned long addr,
                                "address 0x%lx\n", addr);
                return -1;
        }
-#ifdef CONFIG_PERFMON
        /*
         * Check if debug registers are used by perfmon. This
         * test must be done once we know that we can do the
@@ -2142,9 +2139,8 @@ access_uarea(struct task_struct *child, unsigned long addr,
         * IA64_THREAD_DBG_VALID. The registers are restored
         * by the PMU context switch code.
         */
-       if (pfm_use_debug_registers(child))
+       if (pfm_use_dbregs(child))
                return -1;
-#endif
 
        if (!(child->thread.flags & IA64_THREAD_DBG_VALID)) {
                child->thread.flags |= IA64_THREAD_DBG_VALID;
diff --git a/arch/ia64/kernel/setup.c b/arch/ia64/kernel/setup.c
index de636b2..677fa68 100644
--- a/arch/ia64/kernel/setup.c
+++ b/arch/ia64/kernel/setup.c
@@ -45,6 +45,7 @@
 #include <linux/cpufreq.h>
 #include <linux/kexec.h>
 #include <linux/crash_dump.h>
+#include <linux/perfmon_kern.h>
 
 #include <asm/ia32.h>
 #include <asm/machvec.h>
@@ -1051,6 +1052,8 @@ cpu_init (void)
        }
        platform_cpu_init();
        pm_idle = default_idle;
+
+       pfm_init_percpu();
 }
 
 void __init
diff --git a/arch/ia64/kernel/smpboot.c b/arch/ia64/kernel/smpboot.c
index d8f05e5..3d7a739 100644
--- a/arch/ia64/kernel/smpboot.c
+++ b/arch/ia64/kernel/smpboot.c
@@ -39,6 +39,7 @@
 #include <linux/efi.h>
 #include <linux/percpu.h>
 #include <linux/bitops.h>
+#include <linux/perfmon_kern.h>
 
 #include <asm/atomic.h>
 #include <asm/cache.h>
@@ -381,10 +382,6 @@ smp_callin (void)
        extern void ia64_init_itm(void);
        extern volatile int time_keeper_id;
 
-#ifdef CONFIG_PERFMON
-       extern void pfm_init_percpu(void);
-#endif
-
        cpuid = smp_processor_id();
        phys_id = hard_smp_processor_id();
        itc_master = time_keeper_id;
@@ -410,10 +407,6 @@ smp_callin (void)
 
        ia64_mca_cmc_vector_setup();    /* Setup vector on AP */
 
-#ifdef CONFIG_PERFMON
-       pfm_init_percpu();
-#endif
-
        local_irq_enable();
 
        if (!(sal_platform_features & IA64_SAL_PLATFORM_FEATURE_ITC_DRIFT)) {
@@ -751,6 +744,7 @@ int __cpu_disable(void)
        cpu_clear(cpu, cpu_online_map);
        local_flush_tlb_all();
        cpu_clear(cpu, cpu_callin_map);
+       pfm_cpu_disable();
        return 0;
 }
 
diff --git a/arch/ia64/kernel/sys_ia64.c b/arch/ia64/kernel/sys_ia64.c
index bcbb6d8..a0ed33a 100644
--- a/arch/ia64/kernel/sys_ia64.c
+++ b/arch/ia64/kernel/sys_ia64.c
@@ -284,3 +284,11 @@ sys_pciconfig_write (unsigned long bus, unsigned long dfn, unsigned long off, un
 }
 
 #endif /* CONFIG_PCI */
+
+#ifndef CONFIG_IA64_PERFMON_COMPAT
+asmlinkage long
+sys_perfmonctl (int fd, int cmd, void __user *arg, int count)
+{
+       return -ENOSYS;
+}
+#endif
diff --git a/arch/ia64/lib/Makefile b/arch/ia64/lib/Makefile
index 98771e2..077fd09 100644
--- a/arch/ia64/lib/Makefile
+++ b/arch/ia64/lib/Makefile
@@ -13,7 +13,6 @@ lib-y := __divsi3.o __udivsi3.o __modsi3.o __umodsi3.o                        \
 
 obj-$(CONFIG_ITANIUM)  += copy_page.o copy_user.o memcpy.o
 obj-$(CONFIG_MCKINLEY) += copy_page_mck.o memcpy_mck.o
-lib-$(CONFIG_PERFMON)  += carta_random.o
 
 AFLAGS___divdi3.o      =
 AFLAGS___udivdi3.o     = -DUNSIGNED
diff --git a/arch/ia64/oprofile/init.c b/arch/ia64/oprofile/init.c
index 125a602..892de6a 100644
--- a/arch/ia64/oprofile/init.c
+++ b/arch/ia64/oprofile/init.c
@@ -12,8 +12,8 @@
 #include <linux/init.h>
 #include <linux/errno.h>
  
-extern int perfmon_init(struct oprofile_operations * ops);
-extern void perfmon_exit(void);
+extern int op_perfmon_init(struct oprofile_operations * ops);
+extern void op_perfmon_exit(void);
 extern void ia64_backtrace(struct pt_regs * const regs, unsigned int depth);
 
 int __init oprofile_arch_init(struct oprofile_operations * ops)
@@ -22,7 +22,7 @@ int __init oprofile_arch_init(struct oprofile_operations * ops)
 
 #ifdef CONFIG_PERFMON
        /* perfmon_init() can fail, but we have no way to report it */
-       ret = perfmon_init(ops);
+       ret = op_perfmon_init(ops);
 #endif
        ops->backtrace = ia64_backtrace;
 
@@ -33,6 +33,6 @@ int __init oprofile_arch_init(struct oprofile_operations * ops)
 void oprofile_arch_exit(void)
 {
 #ifdef CONFIG_PERFMON
-       perfmon_exit();
+       op_perfmon_exit();
 #endif
 }
diff --git a/arch/ia64/oprofile/perfmon.c b/arch/ia64/oprofile/perfmon.c
index bc41dd3..6fa9d17 100644
--- a/arch/ia64/oprofile/perfmon.c
+++ b/arch/ia64/oprofile/perfmon.c
@@ -10,25 +10,30 @@
 #include <linux/kernel.h>
 #include <linux/oprofile.h>
 #include <linux/sched.h>
-#include <asm/perfmon.h>
+#include <linux/module.h>
+#include <linux/perfmon_kern.h>
 #include <asm/ptrace.h>
 #include <asm/errno.h>
 
 static int allow_ints;
 
 static int
-perfmon_handler(struct task_struct *task, void *buf, pfm_ovfl_arg_t *arg,
-                struct pt_regs *regs, unsigned long stamp)
+perfmon_handler(struct pfm_context *ctx,
+               unsigned long ip, u64 stamp, void *data)
 {
-       int event = arg->pmd_eventid;
+       struct pt_regs *regs;
+       struct pfm_ovfl_arg *arg;
+
+       regs = data;
+       arg = &ctx->ovfl_arg;
  
-       arg->ovfl_ctrl.bits.reset_ovfl_pmds = 1;
+       arg->ovfl_ctrl = PFM_OVFL_CTRL_RESET;
 
        /* the owner of the oprofile event buffer may have exited
         * without perfmon being shutdown (e.g. SIGSEGV)
         */
        if (allow_ints)
-               oprofile_add_sample(regs, event);
+               oprofile_add_sample(regs, arg->pmd_eventid);
        return 0;
 }
 
@@ -45,17 +50,13 @@ static void perfmon_stop(void)
        allow_ints = 0;
 }
 
-
-#define OPROFILE_FMT_UUID { \
-       0x77, 0x7a, 0x6e, 0x61, 0x20, 0x65, 0x73, 0x69, 0x74, 0x6e, 0x72, 0x20, 0x61, 0x65, 0x0a, 0x6c }
-
-static pfm_buffer_fmt_t oprofile_fmt = {
-       .fmt_name           = "oprofile_format",
-       .fmt_uuid           = OPROFILE_FMT_UUID,
-       .fmt_handler        = perfmon_handler,
+static struct pfm_smpl_fmt oprofile_fmt = {
+       .fmt_name = "OProfile",
+       .fmt_handler = perfmon_handler,
+       .fmt_flags = PFM_FMT_BUILTIN_FLAG,
+       .owner = THIS_MODULE
 };
 
-
 static char * get_cpu_type(void)
 {
        __u8 family = local_cpu_data->family;
@@ -75,9 +76,9 @@ static char * get_cpu_type(void)
 
 static int using_perfmon;
 
-int perfmon_init(struct oprofile_operations * ops)
+int __init op_perfmon_init(struct oprofile_operations * ops)
 {
-       int ret = pfm_register_buffer_fmt(&oprofile_fmt);
+       int ret = pfm_fmt_register(&oprofile_fmt);
        if (ret)
                return -ENODEV;
 
@@ -90,10 +91,10 @@ int perfmon_init(struct oprofile_operations * ops)
 }
 
 
-void perfmon_exit(void)
+void __exit op_perfmon_exit(void)
 {
        if (!using_perfmon)
                return;
 
-       pfm_unregister_buffer_fmt(oprofile_fmt.fmt_uuid);
+       pfm_fmt_unregister(&oprofile_fmt);
 }
diff --git a/arch/ia64/perfmon/Kconfig b/arch/ia64/perfmon/Kconfig
new file mode 100644
index 0000000..99c68bd
--- /dev/null
+++ b/arch/ia64/perfmon/Kconfig
@@ -0,0 +1,67 @@
+menu "Hardware Performance Monitoring support"
+config PERFMON
+       bool "Perfmon2 performance monitoring interface"
+       default n
+       help
+       Enables the perfmon2 interface to access the hardware
+       performance counters. See <http://perfmon2.sf.net/> for
+       more details.
+
+config PERFMON_DEBUG
+       bool "Perfmon debugging"
+       default n
+       depends on PERFMON
+       help
+       Enables perfmon debugging support
+
+config PERFMON_DEBUG_FS
+       bool "Enable perfmon statistics reporting via debugfs"
+       default y
+       depends on PERFMON && DEBUG_FS
+       help
+       Enable collection and reporting of perfmon timing statistics under
+       debugfs. This is used for debugging and performance analysis of the
+       subsystem. The debugfs filesystem must be mounted.
+
+config IA64_PERFMON_COMPAT
+       bool "Enable old perfmon-2 compatbility mode"
+       default n
+       depends on PERFMON
+       help
+       Enable this option to allow performance tools which used the old
+       perfmon-2 interface to continue to work. Old tools are those using
+       the obsolete commands and arguments. Check your programs and look
+       in include/asm-ia64/perfmon_compat.h for more information.
+
+config IA64_PERFMON_GENERIC
+       tristate "Generic IA-64 PMU support"
+       depends on PERFMON
+       default n
+       help
+       Enables generic IA-64 PMU support.
+       The generic PMU is defined by the IA-64 architecture document.
+       This option should only be necessary when running with a PMU that
+       is not yet explicitely supported. Even then, there is no guarantee
+       that this support will work.
+
+config IA64_PERFMON_ITANIUM
+       tristate "Itanium (Merced) Performance Monitoring support"
+       depends on PERFMON
+       default n
+       help
+       Enables Itanium (Merced) PMU support.
+
+config IA64_PERFMON_MCKINLEY
+       tristate "Itanium 2 (McKinley) Performance Monitoring  support"
+       depends on PERFMON
+       default n
+       help
+       Enables Itanium 2 (McKinley, Madison, Deerfield) PMU support.
+
+config IA64_PERFMON_MONTECITO
+       tristate "Itanium 2 9000 (Montecito) Performance Monitoring  support"
+       depends on PERFMON
+       default n
+       help
+       Enables support for Itanium 2 9000 (Montecito) PMU.
+endmenu
diff --git a/arch/ia64/perfmon/Makefile b/arch/ia64/perfmon/Makefile
new file mode 100644
index 0000000..c9cdf9f
--- /dev/null
+++ b/arch/ia64/perfmon/Makefile
@@ -0,0 +1,11 @@
+#
+# Copyright (c) 2005-2006 Hewlett-Packard Development Company, L.P.
+# Contributed by Stephane Eranian <eranian@hpl.hp.com>
+#
+obj-$(CONFIG_PERFMON)                  += perfmon.o
+obj-$(CONFIG_IA64_PERFMON_COMPAT)      += perfmon_default_smpl.o \
+                                          perfmon_compat.o
+obj-$(CONFIG_IA64_PERFMON_GENERIC)     += perfmon_generic.o
+obj-$(CONFIG_IA64_PERFMON_ITANIUM)     += perfmon_itanium.o
+obj-$(CONFIG_IA64_PERFMON_MCKINLEY)    += perfmon_mckinley.o
+obj-$(CONFIG_IA64_PERFMON_MONTECITO)   += perfmon_montecito.o
diff --git a/arch/ia64/perfmon/perfmon.c b/arch/ia64/perfmon/perfmon.c
new file mode 100644
index 0000000..3f59410
--- /dev/null
+++ b/arch/ia64/perfmon/perfmon.c
@@ -0,0 +1,946 @@
+/*
+ * This file implements the IA-64 specific
+ * support for the perfmon2 interface
+ *
+ * Copyright (c) 1999-2006 Hewlett-Packard Development Company, L.P.
+ * Contributed by Stephane Eranian <eranian@hpl.hp.com>
+ *
+ * This program is free software; you can redistribute it and/or
+ * modify it under the terms of version 2 of the GNU General Public
+ * License as published by the Free Software Foundation.
+ *
+ * This program is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+ * General Public License for more details.
+ *
+ * You should have received a copy of the GNU General Public License
+ * along with this program; if not, write to the Free Software
+ * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA
+ * 02111-1307 USA
+  */
+#include <linux/module.h>
+#include <linux/perfmon_kern.h>
+
+struct pfm_arch_session {
+       u32     pfs_sys_use_dbr;    /* syswide session uses dbr */
+       u32     pfs_ptrace_use_dbr; /* a thread uses dbr via ptrace()*/
+};
+
+DEFINE_PER_CPU(u32, pfm_syst_info);
+
+static struct pfm_arch_session pfm_arch_sessions;
+static __cacheline_aligned_in_smp DEFINE_SPINLOCK(pfm_arch_sessions_lock);
+
+static inline void pfm_clear_psr_pp(void)
+{
+       ia64_rsm(IA64_PSR_PP);
+}
+
+static inline void pfm_set_psr_pp(void)
+{
+       ia64_ssm(IA64_PSR_PP);
+}
+
+static inline void pfm_clear_psr_up(void)
+{
+       ia64_rsm(IA64_PSR_UP);
+}
+
+static inline void pfm_set_psr_up(void)
+{
+       ia64_ssm(IA64_PSR_UP);
+}
+
+static inline void pfm_set_psr_l(u64 val)
+{
+       ia64_setreg(_IA64_REG_PSR_L, val);
+}
+
+static inline void pfm_restore_ibrs(u64 *ibrs, unsigned int nibrs)
+{
+       unsigned int i;
+
+       for (i = 0; i < nibrs; i++) {
+               ia64_set_ibr(i, ibrs[i]);
+               ia64_dv_serialize_instruction();
+       }
+       ia64_srlz_i();
+}
+
+static inline void pfm_restore_dbrs(u64 *dbrs, unsigned int ndbrs)
+{
+       unsigned int i;
+
+       for (i = 0; i < ndbrs; i++) {
+               ia64_set_dbr(i, dbrs[i]);
+               ia64_dv_serialize_data();
+       }
+       ia64_srlz_d();
+}
+
+irqreturn_t pmu_interrupt_handler(int irq, void *arg)
+{
+       struct pt_regs *regs;
+       regs = get_irq_regs();
+       irq_enter();
+       pfm_interrupt_handler(instruction_pointer(regs), regs);
+       irq_exit();
+       return IRQ_HANDLED;
+}
+static struct irqaction perfmon_irqaction = {
+       .handler = pmu_interrupt_handler,
+       .flags = IRQF_DISABLED, /* means keep interrupts masked */
+       .name = "perfmon"
+};
+
+void pfm_arch_quiesce_pmu_percpu(void)
+{
+       u64 dcr;
+       /*
+        * make sure no measurement is active
+        * (may inherit programmed PMCs from EFI).
+        */
+       pfm_clear_psr_pp();
+       pfm_clear_psr_up();
+
+       /*
+        * ensure dcr.pp is cleared
+        */
+       dcr = ia64_getreg(_IA64_REG_CR_DCR);
+       ia64_setreg(_IA64_REG_CR_DCR, dcr & ~IA64_DCR_PP);
+
+       /*
+        * we run with the PMU not frozen at all times
+        */
+       ia64_set_pmc(0, 0);
+       ia64_srlz_d();
+}
+
+void pfm_arch_init_percpu(void)
+{
+       pfm_arch_quiesce_pmu_percpu();
+       /*
+        * program PMU interrupt vector
+        */
+       ia64_setreg(_IA64_REG_CR_PMV, IA64_PERFMON_VECTOR);
+       ia64_srlz_d();
+}
+
+int pfm_arch_context_create(struct pfm_context *ctx, u32 ctx_flags)
+{
+       struct pfm_arch_context *ctx_arch;
+
+       ctx_arch = pfm_ctx_arch(ctx);
+
+       ctx_arch->flags.use_dbr = 0;
+       ctx_arch->flags.insecure = (ctx_flags & PFM_ITA_FL_INSECURE) ? 1: 0;
+
+       PFM_DBG("insecure=%d", ctx_arch->flags.insecure);
+
+       return 0;
+}
+
+/*
+ * Called from pfm_ctxsw(). Task is guaranteed to be current.
+ * Context is locked. Interrupts are masked. Monitoring may be active.
+ * PMU access is guaranteed. PMC and PMD registers are live in PMU.
+ *
+ * Return:
+ *     non-zero : did not save PMDs (as part of stopping the PMU)
+ *            0 : saved PMDs (no need to save them in caller)
+ */
+int pfm_arch_ctxswout_thread(struct task_struct *task, struct pfm_context *ctx)
+{
+       struct pfm_arch_context *ctx_arch;
+       struct pfm_event_set *set;
+       u64 psr, tmp;
+
+       ctx_arch = pfm_ctx_arch(ctx);
+       set = ctx->active_set;
+
+       /*
+        * save current PSR: needed because we modify it
+        */
+       ia64_srlz_d();
+       psr = ia64_getreg(_IA64_REG_PSR);
+
+       /*
+        * stop monitoring:
+        * This is the last instruction which may generate an overflow
+        *
+        * we do not clear ipsr.up
+        */
+       pfm_clear_psr_up();
+       ia64_srlz_d();
+
+       /*
+        * extract overflow status bits
+        */
+       tmp =  ia64_get_pmc(0) & ~0xf;
+
+       /*
+        * keep a copy of psr.up (for reload)
+        */
+       ctx_arch->ctx_saved_psr_up = psr & IA64_PSR_UP;
+
+       /*
+        * save overflow status bits
+        */
+       set->povfl_pmds[0] = tmp;
+
+       /*
+        * record how many pending overflows
+        * XXX: assume identity mapping for counters
+        */
+       set->npend_ovfls = ia64_popcnt(tmp);
+
+       /*
+        * make sure the PMU is unfrozen for the next task
+        */
+       if (set->npend_ovfls) {
+               ia64_set_pmc(0, 0);
+               ia64_srlz_d();
+       }
+       return 1;
+}
+
+/*
+ * Called from pfm_ctxsw(). Task is guaranteed to be current.
+ * set cannot be NULL. Context is locked. Interrupts are masked.
+ * Caller has already restored all PMD and PMC registers.
+ *
+ * must reactivate monitoring
+ */
+void pfm_arch_ctxswin_thread(struct task_struct *task, struct pfm_context *ctx)
+{
+       struct pfm_arch_context *ctx_arch;
+
+       ctx_arch = pfm_ctx_arch(ctx);
+
+       /*
+        * when monitoring is not explicitly started
+        * then psr_up = 0, in which case we do not
+        * need to restore
+        */
+       if (likely(ctx_arch->ctx_saved_psr_up)) {
+               pfm_set_psr_up();
+               ia64_srlz_d();
+       }
+}
+
+int pfm_arch_reserve_session(struct pfm_context *ctx, u32 cpu)
+{
+       struct pfm_arch_context *ctx_arch;
+       int is_system;
+       int ret = 0;
+
+       ctx_arch = pfm_ctx_arch(ctx);
+       is_system = ctx->flags.system;
+
+       spin_lock(&pfm_arch_sessions_lock);
+
+       if (is_system && ctx_arch->flags.use_dbr) {
+               PFM_DBG("syswide context uses dbregs");
+
+               if (pfm_arch_sessions.pfs_ptrace_use_dbr) {
+                       PFM_DBG("cannot reserve syswide context: "
+                                 "dbregs in use by ptrace");
+                       ret = -EBUSY;
+               } else {
+                       pfm_arch_sessions.pfs_sys_use_dbr++;
+               }
+       }
+       spin_unlock(&pfm_arch_sessions_lock);
+
+       return ret;
+}
+
+void pfm_arch_release_session(struct pfm_context *ctx, u32 cpu)
+{
+       struct pfm_arch_context *ctx_arch;
+       int is_system;
+
+       ctx_arch = pfm_ctx_arch(ctx);
+       is_system = ctx->flags.system;
+
+       spin_lock(&pfm_arch_sessions_lock);
+
+       if (is_system && ctx_arch->flags.use_dbr)
+               pfm_arch_sessions.pfs_sys_use_dbr--;
+       spin_unlock(&pfm_arch_sessions_lock);
+}
+
+/*
+ * function called from pfm_load_context_*(). Task is not guaranteed to be
+ * current task. If not then other task is guaranteed stopped and off any CPU.
+ * context is locked and interrupts are masked.
+ *
+ * On PFM_LOAD_CONTEXT, the interface guarantees monitoring is stopped.
+ *
+ * For system-wide task is NULL
+ */
+int pfm_arch_load_context(struct pfm_context *ctx)
+{
+       struct pfm_arch_context *ctx_arch;
+       struct pt_regs *regs;
+       int ret = 0;
+
+       ctx_arch = pfm_ctx_arch(ctx);
+
+       /*
+        * cannot load a context which is using range restrictions,
+        * into a thread that is being debugged.
+        *
+        * if one set out of several is using the debug registers, then
+        * we assume the context as whole is using them.
+        */
+       if (ctx_arch->flags.use_dbr) {
+               if (ctx->flags.system) {
+                       spin_lock(&pfm_arch_sessions_lock);
+
+                       if (pfm_arch_sessions.pfs_ptrace_use_dbr) {
+                               PFM_DBG("cannot reserve syswide context: "
+                                       "dbregs in use by ptrace");
+                               ret = -EBUSY;
+                       } else {
+                               pfm_arch_sessions.pfs_sys_use_dbr++;
+                               PFM_DBG("pfs_sys_use_dbr=%u",
+                                       pfm_arch_sessions.pfs_sys_use_dbr);
+                       }
+                       spin_unlock(&pfm_arch_sessions_lock);
+
+               } else if (ctx->task->thread.flags & IA64_THREAD_DBG_VALID) {
+                       PFM_DBG("load_pid [%d] thread is debugged, cannot "
+                                 "use range restrictions", ctx->task->pid);
+                       ret = -EBUSY;
+               }
+               if (ret)
+                       return ret;
+       }
+
+       /*
+        * We need to intervene on context switch to toggle the
+        * psr.pp bit in system-wide. As such, we set the TIF
+        * flag so that pfm_arch_ctxswout_sys() and the
+        * pfm_arch_ctxswin_sys() functions get called
+        * from pfm_ctxsw_sys();
+        */
+       if (ctx->flags.system) {
+               set_thread_flag(TIF_PERFMON_CTXSW);
+               PFM_DBG("[%d] set TIF", current->pid);
+               return 0;
+       }
+
+       regs = task_pt_regs(ctx->task);
+
+       /*
+        * self-monitoring systematically allows user level control
+        */
+       if (ctx->task != current) {
+               /*
+                * when not current, task is stopped, so this is safe
+                */
+               ctx_arch->ctx_saved_psr_up = 0;
+               ia64_psr(regs)->up = ia64_psr(regs)->pp = 0;
+       } else
+               ctx_arch->flags.insecure = 1;
+
+       /*
+        * allow user level control (start/stop/read pmd) if:
+        *      - self-monitoring
+        *      - requested at context creation (PFM_IA64_FL_INSECURE)
+        *
+        * There is not security hole with PFM_IA64_FL_INSECURE because
+        * when not self-monitored, the caller must have permissions to
+        * attached to the task.
+        */
+       if (ctx_arch->flags.insecure) {
+               ia64_psr(regs)->sp = 0;
+               PFM_DBG("clearing psr.sp for [%d]", ctx->task->pid);
+       }
+       return 0;
+}
+
+int pfm_arch_setfl_sane(struct pfm_context *ctx, u32 flags)
+{
+#define PFM_SETFL_BOTH_SWITCH  (PFM_SETFL_OVFL_SWITCH|PFM_SETFL_TIME_SWITCH)
+#define PFM_ITA_SETFL_BOTH_INTR        (PFM_ITA_SETFL_INTR_ONLY|\
+                                PFM_ITA_SETFL_EXCL_INTR)
+
+/* exclude return value field */
+#define PFM_SETFL_ALL_MASK     (PFM_ITA_SETFL_BOTH_INTR \
+                                | PFM_SETFL_BOTH_SWITCH        \
+                                | PFM_ITA_SETFL_IDLE_EXCL)
+
+       if ((flags & ~PFM_SETFL_ALL_MASK)) {
+               PFM_DBG("invalid flags=0x%x", flags);
+               return -EINVAL;
+       }
+
+       if ((flags & PFM_ITA_SETFL_BOTH_INTR) == PFM_ITA_SETFL_BOTH_INTR) {
+               PFM_DBG("both excl intr and ontr only are set");
+               return -EINVAL;
+       }
+
+       if ((flags & PFM_ITA_SETFL_IDLE_EXCL) && !ctx->flags.system) {
+               PFM_DBG("idle exclude flag only for system-wide context");
+               return -EINVAL;
+       }
+       return 0;
+}
+
+/*
+ * function called from pfm_unload_context_*(). Context is locked.
+ * interrupts are masked. task is not guaranteed to be current task.
+ * Access to PMU is not guaranteed.
+ *
+ * function must do whatever arch-specific action is required on unload
+ * of a context.
+ *
+ * called for both system-wide and per-thread. task is NULL for ssytem-wide
+ */
+void pfm_arch_unload_context(struct pfm_context *ctx)
+{
+       struct pfm_arch_context *ctx_arch;
+       struct pt_regs *regs;
+
+       ctx_arch = pfm_ctx_arch(ctx);
+
+       if (ctx->flags.system) {
+               /*
+                * disable context switch hook
+                */
+               clear_thread_flag(TIF_PERFMON_CTXSW);
+
+               if (ctx_arch->flags.use_dbr) {
+                       spin_lock(&pfm_arch_sessions_lock);
+                       pfm_arch_sessions.pfs_sys_use_dbr--;
+                       PFM_DBG("sys_use_dbr=%u", pfm_arch_sessions.pfs_sys_use_dbr);
+                       spin_unlock(&pfm_arch_sessions_lock);
+               }
+       } else {
+               regs = task_pt_regs(ctx->task);
+
+               /*
+                * cancel user level control for per-task context
+                */
+               ia64_psr(regs)->sp = 1;
+               PFM_DBG("setting psr.sp for [%d]", ctx->task->pid);
+       }
+}
+
+/*
+ * mask monitoring by setting the privilege level to 0
+ * we cannot use psr.pp/psr.up for this, it is controlled by
+ * the user
+ */
+void pfm_arch_mask_monitoring(struct pfm_context *ctx, struct pfm_event_set *set)
+{
+       struct pfm_arch_pmu_info *arch_info;
+       unsigned long mask;
+       unsigned int i;
+
+       arch_info = pfm_pmu_info();
+       /*
+        * as an optimization we look at the first 64 PMC
+        * registers only starting at PMC4.
+        */
+       mask = arch_info->mask_pmcs[0] >> PFM_ITA_FCNTR;
+       for (i = PFM_ITA_FCNTR; mask; i++, mask >>= 1) {
+               if (likely(mask & 0x1))
+                       ia64_set_pmc(i, set->pmcs[i] & ~0xfUL);
+       }
+       /*
+        * make changes visisble
+        */
+       ia64_srlz_d();
+}
+
+/*
+ * function called from pfm_switch_sets(), pfm_context_load_thread(),
+ * pfm_context_load_sys(), pfm_ctxsw(), pfm_switch_sets()
+ * context is locked. Interrupts are masked. set cannot be NULL.
+ * Access to the PMU is guaranteed.
+ *
+ * function must restore all PMD registers from set.
+ */
+void pfm_arch_restore_pmds(struct pfm_context *ctx, struct pfm_event_set *set)
+{
+       struct pfm_arch_context *ctx_arch;
+       unsigned long *mask;
+       u16 i, num;
+
+       ctx_arch = pfm_ctx_arch(ctx);
+
+       if (ctx_arch->flags.insecure) {
+               num = ctx->regs.num_rw_pmd;
+               mask = ctx->regs.rw_pmds;
+       } else {
+               num = set->nused_pmds;
+               mask = set->used_pmds;
+       }
+       /*
+        * must restore all implemented read-write PMDS to avoid leaking
+        * information especially when PFM_IA64_FL_INSECURE is set.
+        *
+        * XXX: should check PFM_IA64_FL_INSECURE==0 and use used_pmd instead
+        */
+       for (i = 0; num; i++) {
+               if (likely(test_bit(i, mask))) {
+                       pfm_arch_write_pmd(ctx, i, set->pmds[i].value);
+                       num--;
+               }
+       }
+       ia64_srlz_d();
+}
+
+/*
+ * function called from pfm_switch_sets(), pfm_context_load_thread(),
+ * pfm_context_load_sys(), pfm_ctxsw(), pfm_switch_sets()
+ * context is locked. Interrupts are masked. set cannot be NULL.
+ * Access to the PMU is guaranteed.
+ *
+ * function must restore all PMC registers from set if needed
+ */
+void pfm_arch_restore_pmcs(struct pfm_context *ctx, struct pfm_event_set *set)
+{
+       struct pfm_arch_pmu_info *arch_info;
+       u64 mask2 = 0, val, plm;
+       unsigned long impl_mask, mask_pmcs;
+       unsigned int i;
+
+       arch_info = pfm_pmu_info();
+       /*
+        * as an optimization we only look at the first 64
+        * PMC registers. In fact, we should never scan the
+        * entire impl_pmcs because ibr/dbr are implemented
+        * separately.
+        *
+        * always skip PMC0-PMC3. PMC0 taken care of when saving
+        * state. PMC1-PMC3 not used until we get counters in
+        * the 60 and above index range.
+        */
+       impl_mask = ctx->regs.pmcs[0] >> PFM_ITA_FCNTR;
+       mask_pmcs = arch_info->mask_pmcs[0] >> PFM_ITA_FCNTR;
+       plm = ctx->state == PFM_CTX_MASKED ? ~0xf : ~0x0;
+
+       for (i = PFM_ITA_FCNTR;
+            impl_mask;
+            i++, impl_mask >>= 1, mask_pmcs >>= 1) {
+               if (likely(impl_mask & 0x1)) {
+                       mask2 = mask_pmcs & 0x1 ? plm : ~0;
+                       val = set->pmcs[i] & mask2;
+                       ia64_set_pmc(i, val);
+                       PFM_DBG_ovfl("pmc%u=0x%lx", i, val);
+               }
+       }
+       /*
+        * restore DBR/IBR
+        */
+       if (set->priv_flags & PFM_ITA_SETFL_USE_DBR) {
+               pfm_restore_ibrs(set->pmcs+256, 8);
+               pfm_restore_dbrs(set->pmcs+264, 8);
+       }
+       ia64_srlz_d();
+}
+
+void pfm_arch_unmask_monitoring(struct pfm_context *ctx, struct pfm_event_set *set)
+{
+       u64 psr;
+       int is_system;
+
+       is_system = ctx->flags.system;
+
+       psr = ia64_getreg(_IA64_REG_PSR);
+
+       /*
+        * monitoring is masked via the PMC.plm
+        *
+        * As we restore their value, we do not want each counter to
+        * restart right away. We stop monitoring using the PSR,
+        * restore the PMC (and PMD) and then re-establish the psr
+        * as it was. Note that there can be no pending overflow at
+        * this point, because monitoring is still MASKED.
+        *
+        * Because interrupts are masked we can avoid changing
+        * DCR.pp.
+        */
+       if (is_system)
+               pfm_clear_psr_pp();
+       else
+               pfm_clear_psr_up();
+
+       ia64_srlz_d();
+
+       pfm_arch_restore_pmcs(ctx, set);
+
+       /*
+        * restore psr
+        *
+        * monitoring may start right now but interrupts
+        * are still masked
+        */
+       pfm_set_psr_l(psr);
+       ia64_srlz_d();
+}
+
+/*
+ * Called from pfm_stop()
+ *
+ * For per-thread:
+ *   task is not necessarily current. If not current task, then
+ *   task is guaranteed stopped and off any cpu. Access to PMU
+ *   is not guaranteed. Interrupts are masked. Context is locked.
+ *   Set is the active set.
+ *
+ * must disable active monitoring. ctx cannot be NULL
+ */
+void pfm_arch_stop(struct task_struct *task, struct pfm_context *ctx)
+{
+       struct pfm_arch_context *ctx_arch;
+       struct pt_regs *regs;
+       u64 dcr, psr;
+
+       ctx_arch = pfm_ctx_arch(ctx);
+       regs = task_pt_regs(task);
+
+       if (!ctx->flags.system) {
+               /*
+                * in ZOMBIE state we always have task == current due to
+                * pfm_exit_thread()
+                */
+               ia64_psr(regs)->up = 0;
+               ctx_arch->ctx_saved_psr_up = 0;
+
+               /*
+                * in case of ZOMBIE state, there is no unload to clear
+                * insecure monitoring, so we do it in stop instead.
+                */
+               if (ctx->state == PFM_CTX_ZOMBIE)
+                       ia64_psr(regs)->sp = 1;
+
+               if (task == current) {
+                       pfm_clear_psr_up();
+                       ia64_srlz_d();
+               }
+       } else if (ctx->flags.started) { /* do not stop twice */
+               dcr = ia64_getreg(_IA64_REG_CR_DCR);
+               psr = ia64_getreg(_IA64_REG_PSR);
+
+               ia64_psr(regs)->pp = 0;
+               ia64_setreg(_IA64_REG_CR_DCR, dcr & ~IA64_DCR_PP);
+               pfm_clear_psr_pp();
+               ia64_srlz_d();
+
+               if (ctx->active_set->flags & PFM_ITA_SETFL_IDLE_EXCL) {
+                       PFM_DBG("disabling idle exclude");
+                       __get_cpu_var(pfm_syst_info) &= ~PFM_ITA_CPUINFO_IDLE_EXCL;
+               }
+       }
+}
+
+/*
+ * called from pfm_start()
+ *
+ * Interrupts are masked. Context is locked. Set is the active set.
+ *
+ * For per-thread:
+ *     Task is not necessarily current. If not current task, then task
+ *     is guaranteed stopped and off any cpu. No access to PMU is task
+ *     is not current.
+ *
+ * For system-wide:
+ *     task is always current
+ *
+ * must enable active monitoring.
+ */
+void pfm_arch_start(struct task_struct *task, struct pfm_context *ctx)
+{
+       struct pfm_arch_context *ctx_arch;
+       struct pt_regs *regs;
+       u64 dcr, dcr_pp, psr_pp;
+       u32 flags;
+
+       ctx_arch = pfm_ctx_arch(ctx);
+       regs = task_pt_regs(task);
+       flags = ctx->active_set->flags;
+
+       /*
+        * per-thread mode
+        */
+       if (!ctx->flags.system) {
+
+               ia64_psr(regs)->up = 1;
+
+               if (task == current) {
+                       pfm_set_psr_up();
+                       ia64_srlz_d();
+               } else {
+                       /*
+                        * activate monitoring at next ctxswin
+                        */
+                       ctx_arch->ctx_saved_psr_up = IA64_PSR_UP;
+               }
+               return;
+       }
+
+       /*
+        * system-wide mode
+        */
+       dcr = ia64_getreg(_IA64_REG_CR_DCR);
+       if (flags & PFM_ITA_SETFL_INTR_ONLY) {
+               dcr_pp = 1;
+               psr_pp = 0;
+       } else if (flags & PFM_ITA_SETFL_EXCL_INTR) {
+               dcr_pp = 0;
+               psr_pp = 1;
+       } else {
+               dcr_pp = psr_pp = 1;
+       }
+       PFM_DBG("dcr_pp=%lu psr_pp=%lu", dcr_pp, psr_pp);
+
+       /*
+        * update dcr_pp and psr_pp
+        */
+       if (dcr_pp)
+               ia64_setreg(_IA64_REG_CR_DCR, dcr | IA64_DCR_PP);
+       else
+               ia64_setreg(_IA64_REG_CR_DCR, dcr & ~IA64_DCR_PP);
+
+       if (psr_pp) {
+               pfm_set_psr_pp();
+               ia64_psr(regs)->pp = 1;
+       } else {
+               pfm_clear_psr_pp();
+               ia64_psr(regs)->pp = 0;
+       }
+       ia64_srlz_d();
+
+       if (ctx->active_set->flags & PFM_ITA_SETFL_IDLE_EXCL) {
+               PFM_DBG("enable idle exclude");
+               __get_cpu_var(pfm_syst_info) |= PFM_ITA_CPUINFO_IDLE_EXCL;
+       }
+}
+
+/*
+ * Only call this function when a process is trying to
+ * write the debug registers (reading is always allowed)
+ * called from arch/ia64/kernel/ptrace.c:access_uarea()
+ */
+int __pfm_use_dbregs(struct task_struct *task)
+{
+       struct pfm_arch_context *ctx_arch;
+       struct pfm_context *ctx;
+       unsigned long flags;
+       int ret = 0;
+
+       PFM_DBG("called for [%d]", task->pid);
+
+       ctx = task->pfm_context;
+
+       /*
+        * do it only once
+        */
+       if (task->thread.flags & IA64_THREAD_DBG_VALID) {
+               PFM_DBG("IA64_THREAD_DBG_VALID already set");
+               return 0;
+       }
+       if (ctx) {
+               spin_lock_irqsave(&ctx->lock, flags);
+               ctx_arch = pfm_ctx_arch(ctx);
+
+               if (ctx_arch->flags.use_dbr == 1) {
+                       PFM_DBG("PMU using dbregs already, no ptrace access");
+                       ret = -1;
+               }
+               spin_unlock_irqrestore(&ctx->lock, flags);
+               if (ret)
+                       return ret;
+       }
+
+       spin_lock(&pfm_arch_sessions_lock);
+
+       /*
+        * We cannot allow setting breakpoints when system wide monitoring
+        * sessions are using the debug registers.
+        */
+       if (!pfm_arch_sessions.pfs_sys_use_dbr)
+               pfm_arch_sessions.pfs_ptrace_use_dbr++;
+       else
+               ret = -1;
+
+       PFM_DBG("ptrace_use_dbr=%u  sys_use_dbr=%u by [%d] ret = %d",
+                 pfm_arch_sessions.pfs_ptrace_use_dbr,
+                 pfm_arch_sessions.pfs_sys_use_dbr,
+                 task->pid, ret);
+
+       spin_unlock(&pfm_arch_sessions_lock);
+       if (ret)
+               return ret;
+#ifndef CONFIG_SMP
+       /*
+        * in UP, we need to check whether the current
+        * owner of the PMU is not using the debug registers
+        * for monitoring. Because we are using a lazy
+        * save on ctxswout, we must force a save in this
+        * case because the debug registers are being
+        * modified by another task. We save the current
+        * PMD registers, and clear ownership. In ctxswin,
+        * full state will be reloaded.
+        *
+        * Note: we overwrite task.
+        */
+       task = __get_cpu_var(pmu_owner);
+       ctx = __get_cpu_var(pmu_ctx);
+
+       if (task == NULL)
+               return 0;
+
+       ctx_arch = pfm_ctx_arch(ctx);
+
+       if (ctx_arch->flags.use_dbr)
+               pfm_save_pmds_release(ctx);
+#endif
+       return 0;
+}
+
+/*
+ * This function is called for every task that exits with the
+ * IA64_THREAD_DBG_VALID set. This indicates a task which was
+ * able to use the debug registers for debugging purposes via
+ * ptrace(). Therefore we know it was not using them for
+ * perfmormance monitoring, so we only decrement the number
+ * of "ptraced" debug register users to keep the count up to date
+ */
+int __pfm_release_dbregs(struct task_struct *task)
+{
+       int ret;
+
+       spin_lock(&pfm_arch_sessions_lock);
+
+       if (pfm_arch_sessions.pfs_ptrace_use_dbr == 0) {
+               PFM_ERR("invalid release for [%d] ptrace_use_dbr=0", task->pid);
+               ret = -1;
+       }  else {
+               pfm_arch_sessions.pfs_ptrace_use_dbr--;
+               ret = 0;
+       }
+       spin_unlock(&pfm_arch_sessions_lock);
+
+       return ret;
+}
+
+int pfm_ia64_mark_dbregs_used(struct pfm_context *ctx,
+                             struct pfm_event_set *set)
+{
+       struct pfm_arch_context *ctx_arch;
+       struct task_struct *task;
+       struct thread_struct *thread;
+       int ret = 0, state;
+       int i, can_access_pmu = 0;
+       int is_loaded, is_system;
+
+       ctx_arch = pfm_ctx_arch(ctx);
+       state = ctx->state;
+       task = ctx->task;
+       is_loaded = state == PFM_CTX_LOADED || state == PFM_CTX_MASKED;
+       is_system = ctx->flags.system;
+       can_access_pmu = __get_cpu_var(pmu_owner) == task || is_system;
+
+       if (is_loaded == 0)
+               goto done;
+
+       if (is_system == 0) {
+               thread = &(task->thread);
+
+               /*
+                * cannot use debug registers for montioring if they are
+                * already used for debugging
+                */
+               if (thread->flags & IA64_THREAD_DBG_VALID) {
+                       PFM_DBG("debug registers already in use for [%d]",
+                                 task->pid);
+                       return -EBUSY;
+               }
+       }
+
+       /*
+        * check for debug registers in system wide mode
+        */
+       spin_lock(&pfm_arch_sessions_lock);
+
+       if (is_system) {
+               if (pfm_arch_sessions.pfs_ptrace_use_dbr)
+                       ret = -EBUSY;
+               else
+                       pfm_arch_sessions.pfs_sys_use_dbr++;
+       }
+
+       spin_unlock(&pfm_arch_sessions_lock);
+
+       if (ret != 0)
+               return ret;
+
+       /*
+        * clear hardware registers to make sure we don't
+        * pick up stale state.
+        */
+       if (can_access_pmu) {
+               PFM_DBG("clearing ibrs, dbrs");
+               for (i = 0; i < 8; i++) {
+                       ia64_set_ibr(i, 0);
+                       ia64_dv_serialize_instruction();
+               }
+               ia64_srlz_i();
+               for (i = 0; i < 8; i++) {
+                       ia64_set_dbr(i, 0);
+                       ia64_dv_serialize_data();
+               }
+               ia64_srlz_d();
+       }
+done:
+       /*
+        * debug registers are now in use
+        */
+       ctx_arch->flags.use_dbr = 1;
+       set->priv_flags |= PFM_ITA_SETFL_USE_DBR;
+       PFM_DBG("set%u use_dbr=1", set->id);
+       return 0;
+}
+EXPORT_SYMBOL(pfm_ia64_mark_dbregs_used);
+
+char *pfm_arch_get_pmu_module_name(void)
+{
+       switch (local_cpu_data->family) {
+       case 0x07:
+               return "perfmon_itanium";
+       case 0x1f:
+               return "perfmon_mckinley";
+       case 0x20:
+               return "perfmon_montecito";
+       default:
+               return "perfmon_generic";
+       }
+       return NULL;
+}
+
+/*
+ * global arch-specific intialization, called only once
+ */
+int __init pfm_arch_init(void)
+{
+       int ret;
+
+       spin_lock_init(&pfm_arch_sessions_lock);
+
+#ifdef CONFIG_IA64_PERFMON_COMPAT
+       ret = pfm_ia64_compat_init();
+       if (ret)
+               return ret;
+#endif
+       register_percpu_irq(IA64_PERFMON_VECTOR, &perfmon_irqaction);
+
+
+       return 0;
+}
diff --git a/arch/ia64/perfmon/perfmon_compat.c b/arch/ia64/perfmon/perfmon_compat.c
new file mode 100644
index 0000000..2fd3d3c
--- /dev/null
+++ b/arch/ia64/perfmon/perfmon_compat.c
@@ -0,0 +1,1210 @@
+/*
+ * This file implements the IA-64 specific
+ * support for the perfmon2 interface
+ *
+ * Copyright (c) 1999-2006 Hewlett-Packard Development Company, L.P.
+ * Contributed by Stephane Eranian <eranian@hpl.hp.com>
+ *
+ * This program is free software; you can redistribute it and/or
+ * modify it under the terms of version 2 of the GNU General Public
+ * License as published by the Free Software Foundation.
+ *
+ * This program is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+ * General Public License for more details.
+ *
+ * You should have received a copy of the GNU General Public License
+ * along with this program; if not, write to the Free Software
+ * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA
+ * 02111-1307 USA
+  */
+#include <linux/interrupt.h>
+#include <linux/module.h>
+#include <linux/file.h>
+#include <linux/fdtable.h>
+#include <linux/seq_file.h>
+#include <linux/vmalloc.h>
+#include <linux/proc_fs.h>
+#include <linux/perfmon_kern.h>
+#include <linux/uaccess.h>
+
+asmlinkage long sys_pfm_stop(int fd);
+asmlinkage long sys_pfm_start(int fd, struct pfarg_start __user *st);
+asmlinkage long sys_pfm_unload_context(int fd);
+asmlinkage long sys_pfm_restart(int fd);
+asmlinkage long sys_pfm_load_context(int fd, struct pfarg_load __user *ld);
+
+ssize_t pfm_sysfs_res_show(char *buf, size_t sz, int what);
+
+extern ssize_t __pfm_read(struct pfm_context *ctx,
+                         union pfarg_msg *msg_buf,
+                         int non_block);
+/*
+ * function providing some help for backward compatiblity with old IA-64
+ * applications. In the old model, certain attributes of a counter were
+ * passed via the PMC, now they are passed via the PMD.
+ */
+static int pfm_compat_update_pmd(struct pfm_context *ctx, u16 set_id, u16 cnum,
+                                u32 rflags,
+                                unsigned long *smpl_pmds,
+                                unsigned long *reset_pmds,
+                                u64 eventid)
+{
+       struct pfm_event_set *set;
+       int is_counting;
+       unsigned long *impl_pmds;
+       u32 flags = 0;
+       u16 max_pmd;
+
+       impl_pmds = ctx->regs.pmds;
+       max_pmd = ctx->regs.max_pmd;
+
+       /*
+        * given that we do not maintain PMC ->PMD dependencies
+        * we cannot figure out what to do in case PMCxx != PMDxx
+        */
+       if (cnum > max_pmd)
+               return 0;
+
+       /*
+        * assumes PMCxx controls PMDxx which is always true for counters
+        * on Itanium PMUs.
+        */
+       is_counting = pfm_pmu_conf->pmd_desc[cnum].type & PFM_REG_C64;
+       set = pfm_find_set(ctx, set_id, 0);
+
+       /*
+        * for v2.0, we only allowed counting PMD to generate
+        * user-level notifications. Same thing with randomization.
+        */
+       if (is_counting) {
+               if (rflags & PFM_REGFL_OVFL_NOTIFY)
+                       flags |= PFM_REGFL_OVFL_NOTIFY;
+               if (rflags & PFM_REGFL_RANDOM)
+                       flags |= PFM_REGFL_RANDOM;
+               /*
+                * verify validity of smpl_pmds
+                */
+               if (unlikely(bitmap_subset(smpl_pmds,
+                                          impl_pmds, max_pmd) == 0)) {
+                       PFM_DBG("invalid smpl_pmds=0x%llx for pmd%u",
+                                 (unsigned long long)smpl_pmds[0], cnum);
+                       return -EINVAL;
+               }
+               /*
+                * verify validity of reset_pmds
+                */
+               if (unlikely(bitmap_subset(reset_pmds,
+                                          impl_pmds, max_pmd) == 0)) {
+                       PFM_DBG("invalid reset_pmds=0x%lx for pmd%u",
+                                 reset_pmds[0], cnum);
+                       return -EINVAL;
+               }
+               /*
+                * ensures that a PFM_READ_PMDS succeeds with a
+                * corresponding PFM_WRITE_PMDS
+                */
+               __set_bit(cnum, set->used_pmds);
+
+       } else if (rflags & (PFM_REGFL_OVFL_NOTIFY|PFM_REGFL_RANDOM)) {
+               PFM_DBG("cannot set ovfl_notify or random on pmd%u", cnum);
+               return -EINVAL;
+       }
+
+       set->pmds[cnum].flags = flags;
+
+       if (is_counting) {
+               bitmap_copy(set->pmds[cnum].reset_pmds,
+                           reset_pmds,
+                           max_pmd);
+
+               bitmap_copy(set->pmds[cnum].smpl_pmds,
+                           smpl_pmds,
+                           max_pmd);
+
+               set->pmds[cnum].eventid = eventid;
+
+               /*
+                * update ovfl_notify
+                */
+               if (rflags & PFM_REGFL_OVFL_NOTIFY)
+                       __set_bit(cnum, set->ovfl_notify);
+               else
+                       __clear_bit(cnum, set->ovfl_notify);
+
+       }
+       PFM_DBG("pmd%u flags=0x%x eventid=0x%lx r_pmds=0x%lx s_pmds=0x%lx",
+                 cnum, flags,
+                 eventid,
+                 reset_pmds[0],
+                 smpl_pmds[0]);
+
+       return 0;
+}
+
+
+int __pfm_write_ibrs_old(struct pfm_context *ctx, void *arg, int count)
+{
+       struct pfarg_dbreg *req = arg;
+       struct pfarg_pmc pmc;
+       int i, ret = 0;
+
+       memset(&pmc, 0, sizeof(pmc));
+
+       for (i = 0; i < count; i++, req++) {
+               pmc.reg_num   = 256+req->dbreg_num;
+               pmc.reg_value = req->dbreg_value;
+               pmc.reg_flags = 0;
+               pmc.reg_set   = req->dbreg_set;
+
+               ret = __pfm_write_pmcs(ctx, &pmc, 1);
+
+               req->dbreg_flags &= ~PFM_REG_RETFL_MASK;
+               req->dbreg_flags |= pmc.reg_flags;
+
+               if (ret)
+                       return ret;
+       }
+       return 0;
+}
+
+static long pfm_write_ibrs_old(int fd, void __user *ureq, int count)
+{
+       struct pfm_context *ctx;
+       struct task_struct *task;
+       struct file *filp;
+       struct pfarg_dbreg *req = NULL;
+       void *fptr, *resume;
+       unsigned long flags;
+       size_t sz;
+       int ret, fput_needed;
+
+       if (count < 1 || count >= PFM_MAX_ARG_COUNT(req))
+               return -EINVAL;
+
+       sz = count*sizeof(*req);
+
+       filp = fget_light(fd, &fput_needed);
+       if (unlikely(filp == NULL)) {
+               PFM_DBG("invalid fd %d", fd);
+               return -EBADF;
+       }
+
+       ctx = filp->private_data;
+       ret = -EBADF;
+
+       if (unlikely(!ctx || filp->f_op != &pfm_file_ops)) {
+               PFM_DBG("fd %d not related to perfmon", fd);
+               goto error;
+       }
+
+       ret = pfm_get_args(ureq, sz, 0, NULL, (void **)&req, &fptr);
+       if (ret)
+               goto error;
+
+       spin_lock_irqsave(&ctx->lock, flags);
+
+       task = ctx->task;
+
+       ret = pfm_check_task_state(ctx, PFM_CMD_STOPPED, &flags, &resume);
+       if (ret == 0)
+               ret = __pfm_write_ibrs_old(ctx, req, count);
+
+       spin_unlock_irqrestore(&ctx->lock, flags);
+
+       if (resume)
+               pfm_resume_task(task, resume);
+
+       if (copy_to_user(ureq, req, sz))
+               ret = -EFAULT;
+
+       kfree(fptr);
+error:
+       fput_light(filp, fput_needed);
+       return ret;
+}
+
+int __pfm_write_dbrs_old(struct pfm_context *ctx, void *arg, int count)
+{
+       struct pfarg_dbreg *req = arg;
+       struct pfarg_pmc pmc;
+       int i, ret = 0;
+
+       memset(&pmc, 0, sizeof(pmc));
+
+       for (i = 0; i < count; i++, req++) {
+               pmc.reg_num   = 264+req->dbreg_num;
+               pmc.reg_value = req->dbreg_value;
+               pmc.reg_flags = 0;
+               pmc.reg_set   = req->dbreg_set;
+
+               ret = __pfm_write_pmcs(ctx, &pmc, 1);
+
+               req->dbreg_flags &= ~PFM_REG_RETFL_MASK;
+               req->dbreg_flags |= pmc.reg_flags;
+               if (ret)
+                       return ret;
+       }
+       return 0;
+}
+
+static long pfm_write_dbrs_old(int fd, void __user *ureq, int count)
+{
+       struct pfm_context *ctx;
+       struct task_struct *task;
+       struct file *filp;
+       struct pfarg_dbreg *req = NULL;
+       void *fptr, *resume;
+       unsigned long flags;
+       size_t sz;
+       int ret, fput_needed;
+
+       if (count < 1 || count >= PFM_MAX_ARG_COUNT(req))
+               return -EINVAL;
+
+       sz = count*sizeof(*req);
+
+       filp = fget_light(fd, &fput_needed);
+       if (unlikely(filp == NULL)) {
+               PFM_DBG("invalid fd %d", fd);
+               return -EBADF;
+       }
+
+       ctx = filp->private_data;
+       ret = -EBADF;
+
+       if (unlikely(!ctx || filp->f_op != &pfm_file_ops)) {
+               PFM_DBG("fd %d not related to perfmon", fd);
+               goto error;
+       }
+
+       ret = pfm_get_args(ureq, sz, 0, NULL, (void **)&req, &fptr);
+       if (ret)
+               goto error;
+
+       spin_lock_irqsave(&ctx->lock, flags);
+
+       task = ctx->task;
+
+       ret = pfm_check_task_state(ctx, PFM_CMD_STOPPED, &flags, &resume);
+       if (ret == 0)
+               ret = __pfm_write_dbrs_old(ctx, req, count);
+
+       spin_unlock_irqrestore(&ctx->lock, flags);
+
+       if (resume)
+               pfm_resume_task(task, resume);
+
+       if (copy_to_user(ureq, req, sz))
+               ret = -EFAULT;
+
+       kfree(fptr);
+error:
+       fput_light(filp, fput_needed);
+       return ret;
+}
+
+int __pfm_write_pmcs_old(struct pfm_context *ctx, struct pfarg_reg *req_old,
+                        int count)
+{
+       struct pfarg_pmc req;
+       unsigned int i;
+       int ret, error_code;
+
+       memset(&req, 0, sizeof(req));
+
+       for (i = 0; i < count; i++, req_old++) {
+               req.reg_num   = req_old->reg_num;
+               req.reg_set   = req_old->reg_set;
+               req.reg_flags = 0;
+               req.reg_value = req_old->reg_value;
+
+               ret = __pfm_write_pmcs(ctx, (void *)&req, 1);
+               req_old->reg_flags &= ~PFM_REG_RETFL_MASK;
+               req_old->reg_flags |= req.reg_flags;
+
+               if (ret)
+                       return ret;
+
+               ret = pfm_compat_update_pmd(ctx, req_old->reg_set,
+                                     req_old->reg_num,
+                                     (u32)req_old->reg_flags,
+                                     req_old->reg_smpl_pmds,
+                                     req_old->reg_reset_pmds,
+                                     req_old->reg_smpl_eventid);
+
+               error_code = ret ? PFM_REG_RETFL_EINVAL : 0;
+               req_old->reg_flags &= ~PFM_REG_RETFL_MASK;
+               req_old->reg_flags |= error_code;
+
+               if (ret)
+                       return ret;
+       }
+       return 0;
+}
+
+static long pfm_write_pmcs_old(int fd, void __user *ureq, int count)
+{
+       struct pfm_context *ctx;
+       struct task_struct *task;
+       struct file *filp;
+       struct pfarg_reg *req = NULL;
+       void *fptr, *resume;
+       unsigned long flags;
+       size_t sz;
+       int ret, fput_needed;
+
+       if (count < 1 || count >= PFM_MAX_ARG_COUNT(req))
+               return -EINVAL;
+
+       sz = count*sizeof(*req);
+
+       filp = fget_light(fd, &fput_needed);
+       if (unlikely(filp == NULL)) {
+               PFM_DBG("invalid fd %d", fd);
+               return -EBADF;
+       }
+
+       ctx = filp->private_data;
+       ret = -EBADF;
+
+       if (unlikely(!ctx || filp->f_op != &pfm_file_ops)) {
+               PFM_DBG("fd %d not related to perfmon", fd);
+               goto error;
+       }
+
+       ret = pfm_get_args(ureq, sz, 0, NULL, (void **)&req, &fptr);
+       if (ret)
+               goto error;
+
+       spin_lock_irqsave(&ctx->lock, flags);
+
+       task = ctx->task;
+
+       ret = pfm_check_task_state(ctx, PFM_CMD_STOPPED, &flags, &resume);
+       if (ret == 0)
+               ret = __pfm_write_pmcs_old(ctx, req, count);
+
+       spin_unlock_irqrestore(&ctx->lock, flags);
+
+       if (resume)
+               pfm_resume_task(task, resume);
+
+       if (copy_to_user(ureq, req, sz))
+               ret = -EFAULT;
+
+       kfree(fptr);
+
+error:
+       fput_light(filp, fput_needed);
+       return ret;
+}
+
+int __pfm_write_pmds_old(struct pfm_context *ctx, struct pfarg_reg *req_old,
+                        int count)
+{
+       struct pfarg_pmd req;
+       int i, ret;
+
+       memset(&req, 0, sizeof(req));
+
+       for (i = 0; i < count; i++, req_old++) {
+               req.reg_num   = req_old->reg_num;
+               req.reg_set   = req_old->reg_set;
+               req.reg_value = req_old->reg_value;
+               /* flags passed with pmcs in v2.0 */
+
+               req.reg_long_reset  = req_old->reg_long_reset;
+               req.reg_short_reset = req_old->reg_short_reset;
+               req.reg_random_mask = req_old->reg_random_mask;
+               /*
+                * reg_random_seed is ignored since v2.3
+                */
+
+               /*
+                * skip last_reset_val not used for writing
+                * skip smpl_pmds, reset_pmds, eventid, ovfl_swtch_cnt
+                * as set in pfm_write_pmcs_old.
+                *
+                * ovfl_switch_cnt ignored, not implemented in v2.0
+                */
+               ret = __pfm_write_pmds(ctx, (void *)&req, 1, 1);
+
+               req_old->reg_flags &= ~PFM_REG_RETFL_MASK;
+               req_old->reg_flags |= req.reg_flags;
+
+               if (ret)
+                       return ret;
+       }
+       return 0;
+}
+
+static long pfm_write_pmds_old(int fd, void __user *ureq, int count)
+{
+       struct pfm_context *ctx;
+       struct task_struct *task;
+       struct file *filp;
+       struct pfarg_reg *req = NULL;
+       void *fptr, *resume;
+       unsigned long flags;
+       size_t sz;
+       int ret, fput_needed;
+
+       if (count < 1 || count >= PFM_MAX_ARG_COUNT(req))
+               return -EINVAL;
+
+       sz = count*sizeof(*req);
+
+       filp = fget_light(fd, &fput_needed);
+       if (unlikely(filp == NULL)) {
+               PFM_DBG("invalid fd %d", fd);
+               return -EBADF;
+       }
+
+       ctx = filp->private_data;
+       ret = -EBADF;
+
+       if (unlikely(!ctx || filp->f_op != &pfm_file_ops)) {
+               PFM_DBG("fd %d not related to perfmon", fd);
+               goto error;
+       }
+
+       ret = pfm_get_args(ureq, sz, 0, NULL, (void **)&req, &fptr);
+       if (ret)
+               goto error;
+
+       spin_lock_irqsave(&ctx->lock, flags);
+
+       task = ctx->task;
+
+       ret = pfm_check_task_state(ctx, PFM_CMD_STOPPED, &flags, &resume);
+       if (ret == 0)
+               ret = __pfm_write_pmds_old(ctx, req, count);
+
+       spin_unlock_irqrestore(&ctx->lock, flags);
+
+       if (copy_to_user(ureq, req, sz))
+               ret = -EFAULT;
+
+       if (resume)
+               pfm_resume_task(task, resume);
+
+       kfree(fptr);
+error:
+       fput_light(filp, fput_needed);
+       return ret;
+}
+
+int __pfm_read_pmds_old(struct pfm_context *ctx, struct pfarg_reg *req_old,
+                       int count)
+{
+       struct pfarg_pmd req;
+       int i, ret;
+
+       memset(&req, 0, sizeof(req));
+
+       for (i = 0; i < count; i++, req_old++) {
+               req.reg_num   = req_old->reg_num;
+               req.reg_set   = req_old->reg_set;
+
+               /* skip value not used for reading */
+               req.reg_flags = req_old->reg_flags;
+
+               /* skip short/long_reset not used for reading */
+               /* skip last_reset_val not used for reading */
+               /* skip ovfl_switch_cnt not used for reading */
+
+               ret = __pfm_read_pmds(ctx, (void *)&req, 1);
+
+               req_old->reg_flags &= ~PFM_REG_RETFL_MASK;
+               req_old->reg_flags |= req.reg_flags;
+               if (ret)
+                       return ret;
+
+               /* update fields */
+               req_old->reg_value = req.reg_value;
+
+               req_old->reg_last_reset_val  = req.reg_last_reset_val;
+               req_old->reg_ovfl_switch_cnt = req.reg_ovfl_switch_cnt;
+       }
+       return 0;
+}
+
+static long pfm_read_pmds_old(int fd, void __user *ureq, int count)
+{
+       struct pfm_context *ctx;
+       struct task_struct *task;
+       struct file *filp;
+       struct pfarg_reg *req = NULL;
+       void *fptr, *resume;
+       unsigned long flags;
+       size_t sz;
+       int ret, fput_needed;
+
+       if (count < 1 || count >= PFM_MAX_ARG_COUNT(req))
+               return -EINVAL;
+
+       sz = count*sizeof(*req);
+
+       filp = fget_light(fd, &fput_needed);
+       if (unlikely(filp == NULL)) {
+               PFM_DBG("invalid fd %d", fd);
+               return -EBADF;
+       }
+
+       ctx = filp->private_data;
+       ret = -EBADF;
+
+       if (unlikely(!ctx || filp->f_op != &pfm_file_ops)) {
+               PFM_DBG("fd %d not related to perfmon", fd);
+               goto error;
+       }
+
+       ret = pfm_get_args(ureq, sz, 0, NULL, (void **)&req, &fptr);
+       if (ret)
+               goto error;
+
+       spin_lock_irqsave(&ctx->lock, flags);
+
+       task = ctx->task;
+
+       ret = pfm_check_task_state(ctx, PFM_CMD_STOPPED, &flags, &resume);
+       if (ret == 0)
+               ret = __pfm_read_pmds_old(ctx, req, count);
+
+       spin_unlock_irqrestore(&ctx->lock, flags);
+
+       if (resume)
+               pfm_resume_task(task, resume);
+
+       if (copy_to_user(ureq, req, sz))
+               ret = -EFAULT;
+
+       kfree(fptr);
+error:
+       fput_light(filp, fput_needed);
+       return ret;
+}
+
+/*
+ * OBSOLETE: use /proc/perfmon_map instead
+ */
+static long pfm_get_default_pmcs_old(int fd, void __user *ureq, int count)
+{
+       struct pfarg_reg *req = NULL;
+       void *fptr;
+       size_t sz;
+       int ret, i;
+       unsigned int cnum;
+
+       if (count < 1)
+               return -EINVAL;
+
+       /*
+        * ensure the pfm_pmu_conf does not disappear while
+        * we use it
+        */
+       ret = pfm_pmu_conf_get(1);
+       if (ret)
+               return ret;
+
+       sz = count*sizeof(*ureq);
+
+       ret = pfm_get_args(ureq, sz, 0, NULL, (void **)&req, &fptr);
+       if (ret)
+               goto error;
+
+
+       for (i = 0; i < count; i++, req++) {
+               cnum   = req->reg_num;
+
+               if (i >= PFM_MAX_PMCS ||
+                   (pfm_pmu_conf->pmc_desc[cnum].type & PFM_REG_I) == 0) {
+                       req->reg_flags = PFM_REG_RETFL_EINVAL;
+                       break;
+               }
+               req->reg_value = pfm_pmu_conf->pmc_desc[cnum].dfl_val;
+               req->reg_flags = 0;
+
+               PFM_DBG("pmc[%u]=0x%lx", cnum, req->reg_value);
+       }
+
+       if (copy_to_user(ureq, req, sz))
+               ret = -EFAULT;
+
+       kfree(fptr);
+error:
+       pfm_pmu_conf_put();
+
+       return ret;
+}
+
+/*
+ * allocate a sampling buffer and remaps it into the user address space of
+ * the task. This is only in compatibility mode
+ *
+ * function called ONLY on current task
+ */
+int pfm_smpl_buf_alloc_compat(struct pfm_context *ctx, size_t rsize,
+                             struct file *filp)
+{
+       struct mm_struct *mm = current->mm;
+       struct vm_area_struct *vma = NULL;
+       struct pfm_arch_context *ctx_arch;
+       size_t size;
+       int ret;
+       extern struct vm_operations_struct pfm_buf_map_vm_ops;
+
+       ctx_arch = pfm_ctx_arch(ctx);
+
+       /*
+        * allocate buffer + map desc
+        */
+       ret = pfm_smpl_buf_alloc(ctx, rsize);
+       if (ret)
+               return ret;
+
+       size = ctx->smpl_size;
+
+
+       /* allocate vma */
+       vma = kmem_cache_alloc(vm_area_cachep, GFP_KERNEL);
+       if (!vma) {
+               PFM_DBG("Cannot allocate vma");
+               goto error_kmem;
+       }
+       memset(vma, 0, sizeof(*vma));
+
+       /*
+        * partially initialize the vma for the sampling buffer
+        */
+       vma->vm_mm           = mm;
+       vma->vm_flags        = VM_READ | VM_MAYREAD | VM_RESERVED;
+       vma->vm_page_prot    = PAGE_READONLY;
+       vma->vm_ops          = &pfm_buf_map_vm_ops;
+       vma->vm_file         = filp;
+       vma->vm_private_data = ctx;
+       vma->vm_pgoff        = 0;
+
+       /*
+        * simulate effect of mmap()
+        */
+       get_file(filp);
+
+       /*
+        * Let's do the difficult operations next.
+        *
+        * now we atomically find some area in the address space and
+        * remap the buffer into it.
+        */
+       down_write(&current->mm->mmap_sem);
+
+       /* find some free area in address space, must have mmap sem held */
+       vma->vm_start = get_unmapped_area(NULL, 0, size, 0,
+                                         MAP_PRIVATE|MAP_ANONYMOUS);
+       if (vma->vm_start == 0) {
+               PFM_DBG("cannot find unmapped area of size %zu", size);
+               up_write(&current->mm->mmap_sem);
+               goto error;
+       }
+       vma->vm_end = vma->vm_start + size;
+
+       PFM_DBG("aligned_size=%zu mapped @0x%lx", size, vma->vm_start);
+       /*
+        * now insert the vma in the vm list for the process, must be
+        * done with mmap lock held
+        */
+       insert_vm_struct(mm, vma);
+
+       mm->total_vm  += size >> PAGE_SHIFT;
+
+       up_write(&current->mm->mmap_sem);
+
+       /*
+        * IMPORTANT: we do not issue the fput()
+        * because we want to increase the ref count
+        * on the descriptor to simulate what mmap()
+        * would do
+        */
+
+       /*
+        * used to propagate vaddr to syscall stub
+        */
+       ctx_arch->ctx_smpl_vaddr = (void *)vma->vm_start;
+
+       return 0;
+error:
+       kmem_cache_free(vm_area_cachep, vma);
+error_kmem:
+       pfm_smpl_buf_space_release(ctx, ctx->smpl_size);
+       vfree(ctx->smpl_addr);
+       return -ENOMEM;
+}
+
+#define PFM_DEFAULT_SMPL_UUID { \
+               0x4d, 0x72, 0xbe, 0xc0, 0x06, 0x64, 0x41, 0x43, 0x82,\
+               0xb4, 0xd3, 0xfd, 0x27, 0x24, 0x3c, 0x97}
+
+static pfm_uuid_t old_default_uuid = PFM_DEFAULT_SMPL_UUID;
+static pfm_uuid_t null_uuid;
+
+/*
+ * function invoked in case, pfm_context_create fails
+ * at the last operation, copy_to_user. It needs to
+ * undo memory allocations and free the file descriptor
+ */
+static void pfm_undo_create_context_fd(int fd, struct pfm_context *ctx)
+{
+       struct files_struct *files = current->files;
+       struct file *file;
+       int fput_needed;
+
+       file = fget_light(fd, &fput_needed);
+       /*
+        * there is no fd_uninstall(), so we do it
+        * here. put_unused_fd() does not remove the
+        * effect of fd_install().
+        */
+
+       spin_lock(&files->file_lock);
+       files->fd_array[fd] = NULL;
+       spin_unlock(&files->file_lock);
+
+       fput_light(file, fput_needed);
+
+       /*
+        * decrement ref count and kill file
+        */
+       put_filp(file);
+
+       put_unused_fd(fd);
+
+       pfm_free_context(ctx);
+}
+
+static int pfm_get_smpl_arg_old(pfm_uuid_t uuid, void __user *fmt_uarg,
+                               size_t usize, void **arg,
+                               struct pfm_smpl_fmt **fmt)
+{
+       struct pfm_smpl_fmt *f;
+       void *addr = NULL;
+       size_t sz;
+       int ret;
+
+       if (!memcmp(uuid, null_uuid, sizeof(pfm_uuid_t)))
+               return 0;
+
+       if (memcmp(uuid, old_default_uuid, sizeof(pfm_uuid_t))) {
+               PFM_DBG("compatibility mode supports only default sampling format");
+               return -EINVAL;
+       }
+       /*
+        * find fmt and increase refcount
+        */
+       f = pfm_smpl_fmt_get("default-old");
+       if (f == NULL) {
+               PFM_DBG("default-old buffer format not found");
+               return -EINVAL;
+       }
+
+       /*
+        * expected format argument size
+        */
+       sz = f->fmt_arg_size;
+
+       /*
+        * check user size matches expected size
+        * usize = -1 is for IA-64 backward compatibility
+        */
+       ret = -EINVAL;
+       if (sz != usize && usize != -1) {
+               PFM_DBG("invalid arg size %zu, format expects %zu",
+                       usize, sz);
+               goto error;
+       }
+
+       ret = -ENOMEM;
+       addr = kmalloc(sz, GFP_KERNEL);
+       if (addr == NULL)
+               goto error;
+
+       ret = -EFAULT;
+       if (copy_from_user(addr, fmt_uarg, sz))
+               goto error;
+
+       *arg = addr;
+       *fmt = f;
+       return 0;
+
+error:
+       kfree(addr);
+       pfm_smpl_fmt_put(f);
+       return ret;
+}
+
+static long pfm_create_context_old(int fd, void __user *ureq, int count)
+{
+       struct pfm_context *new_ctx;
+       struct pfm_arch_context *ctx_arch;
+       struct pfm_smpl_fmt *fmt = NULL;
+       struct pfarg_context req_old;
+       void __user *usmpl_arg;
+       void *smpl_arg = NULL;
+       struct pfarg_ctx req;
+       int ret;
+
+       if (count != 1)
+               return -EINVAL;
+
+       if (copy_from_user(&req_old, ureq, sizeof(req_old)))
+               return -EFAULT;
+
+       memset(&req, 0, sizeof(req));
+
+       /*
+        * sampling format args are following pfarg_context
+        */
+       usmpl_arg = ureq+sizeof(req_old);
+
+       ret = pfm_get_smpl_arg_old(req_old.ctx_smpl_buf_id, usmpl_arg, -1,
+                                  &smpl_arg, &fmt);
+       if (ret)
+               return ret;
+
+       req.ctx_flags = req_old.ctx_flags;
+
+       /*
+        * returns file descriptor if >=0, or error code */
+       ret = __pfm_create_context(&req, fmt, smpl_arg, PFM_COMPAT, &new_ctx);
+       if (ret >= 0) {
+               ctx_arch = pfm_ctx_arch(new_ctx);
+               req_old.ctx_fd = ret;
+               req_old.ctx_smpl_vaddr = ctx_arch->ctx_smpl_vaddr;
+       }
+
+       if (copy_to_user(ureq, &req_old, sizeof(req_old))) {
+               pfm_undo_create_context_fd(req_old.ctx_fd, new_ctx);
+               ret = -EFAULT;
+       }
+
+       kfree(smpl_arg);
+
+       return ret;
+}
+
+/*
+ * obsolete call: use /proc/perfmon
+ */
+static long pfm_get_features_old(int fd, void __user *arg, int count)
+{
+       struct pfarg_features req;
+       int ret = 0;
+
+       if (count != 1)
+               return -EINVAL;
+
+       memset(&req, 0, sizeof(req));
+
+       req.ft_version = PFM_VERSION;
+
+       if (copy_to_user(arg, &req, sizeof(req)))
+               ret = -EFAULT;
+
+       return ret;
+}
+
+static long pfm_debug_old(int fd, void __user *arg, int count)
+{
+       int m;
+
+       if (count != 1)
+               return -EINVAL;
+
+       if (get_user(m, (int __user *)arg))
+               return -EFAULT;
+
+
+       pfm_controls.debug = m == 0 ? 0 : 1;
+
+       PFM_INFO("debugging %s (timing reset)",
+                pfm_controls.debug ? "on" : "off");
+
+       if (m == 0)
+               for_each_online_cpu(m) {
+                       memset(&per_cpu(pfm_stats, m), 0,
+                              sizeof(struct pfm_stats));
+               }
+       return 0;
+}
+
+static long pfm_unload_context_old(int fd, void __user *arg, int count)
+{
+       if (count)
+               return -EINVAL;
+
+       return sys_pfm_unload_context(fd);
+}
+
+static long pfm_restart_old(int fd, void __user *arg, int count)
+{
+       if (count)
+               return -EINVAL;
+
+       return sys_pfm_restart(fd);
+}
+
+static long pfm_stop_old(int fd, void __user *arg, int count)
+{
+       if (count)
+               return -EINVAL;
+
+       return sys_pfm_stop(fd);
+}
+
+static long pfm_start_old(int fd, void __user *arg, int count)
+{
+       if (count > 1)
+               return -EINVAL;
+
+       return sys_pfm_start(fd, arg);
+}
+
+static long pfm_load_context_old(int fd, void __user *ureq, int count)
+{
+       if (count != 1)
+               return -EINVAL;
+
+       return sys_pfm_load_context(fd, ureq);
+}
+
+/*
+ * perfmon command descriptions
+ */
+struct pfm_cmd_desc {
+       long (*cmd_func)(int fd, void __user *arg, int count);
+};
+
+/*
+ * functions MUST be listed in the increasing order of
+ * their index (see permfon.h)
+ */
+#define PFM_CMD(name)  \
+       { .cmd_func = name,  \
+       }
+#define PFM_CMD_NONE           \
+       { .cmd_func = NULL   \
+       }
+
+static struct pfm_cmd_desc pfm_cmd_tab[] = {
+/* 0  */PFM_CMD_NONE,
+/* 1  */PFM_CMD(pfm_write_pmcs_old),
+/* 2  */PFM_CMD(pfm_write_pmds_old),
+/* 3  */PFM_CMD(pfm_read_pmds_old),
+/* 4  */PFM_CMD(pfm_stop_old),
+/* 5  */PFM_CMD(pfm_start_old),
+/* 6  */PFM_CMD_NONE,
+/* 7  */PFM_CMD_NONE,
+/* 8  */PFM_CMD(pfm_create_context_old),
+/* 9  */PFM_CMD_NONE,
+/* 10 */PFM_CMD(pfm_restart_old),
+/* 11 */PFM_CMD_NONE,
+/* 12 */PFM_CMD(pfm_get_features_old),
+/* 13 */PFM_CMD(pfm_debug_old),
+/* 14 */PFM_CMD_NONE,
+/* 15 */PFM_CMD(pfm_get_default_pmcs_old),
+/* 16 */PFM_CMD(pfm_load_context_old),
+/* 17 */PFM_CMD(pfm_unload_context_old),
+/* 18 */PFM_CMD_NONE,
+/* 19 */PFM_CMD_NONE,
+/* 20 */PFM_CMD_NONE,
+/* 21 */PFM_CMD_NONE,
+/* 22 */PFM_CMD_NONE,
+/* 23 */PFM_CMD_NONE,
+/* 24 */PFM_CMD_NONE,
+/* 25 */PFM_CMD_NONE,
+/* 26 */PFM_CMD_NONE,
+/* 27 */PFM_CMD_NONE,
+/* 28 */PFM_CMD_NONE,
+/* 29 */PFM_CMD_NONE,
+/* 30 */PFM_CMD_NONE,
+/* 31 */PFM_CMD_NONE,
+/* 32 */PFM_CMD(pfm_write_ibrs_old),
+/* 33 */PFM_CMD(pfm_write_dbrs_old),
+};
+#define PFM_CMD_COUNT ARRAY_SIZE(pfm_cmd_tab)
+
+/*
+ * system-call entry point (must return long)
+ */
+asmlinkage long sys_perfmonctl(int fd, int cmd, void __user *arg, int count)
+{
+       if (perfmon_disabled)
+               return -ENOSYS;
+
+       if (unlikely(cmd < 0 || cmd >= PFM_CMD_COUNT
+                    || pfm_cmd_tab[cmd].cmd_func == NULL)) {
+               PFM_DBG("invalid cmd=%d", cmd);
+               return -EINVAL;
+       }
+       return (long)pfm_cmd_tab[cmd].cmd_func(fd, arg, count);
+}
+
+/*
+ * Called from pfm_read() for a perfmon v2.0 context.
+ *
+ * compatibility mode pfm_read() routine. We need a separate
+ * routine because the definition of the message has changed.
+ * The pfm_msg and pfarg_msg structures are different.
+ *
+ * return: sizeof(pfm_msg_t) on success, -errno otherwise
+ */
+ssize_t pfm_arch_compat_read(struct pfm_context *ctx,
+                            char __user *buf,
+                            int non_block,
+                            size_t size)
+{
+       union pfarg_msg msg_buf;
+       pfm_msg_t old_msg_buf;
+       pfm_ovfl_msg_t *o_msg;
+       struct pfarg_ovfl_msg *n_msg;
+       int ret;
+
+       PFM_DBG("msg=%p size=%zu", buf, size);
+
+       /*
+        * cannot extract partial messages.
+        * check even when there is no message
+        *
+        * cannot extract more than one message per call. Bytes
+        * above sizeof(msg) are ignored.
+        */
+       if (size < sizeof(old_msg_buf)) {
+               PFM_DBG("message is too small size=%zu must be >=%zu)",
+                       size,
+                       sizeof(old_msg_buf));
+               return -EINVAL;
+       }
+
+       ret =  __pfm_read(ctx, &msg_buf, non_block);
+       if (ret < 1)
+               return ret;
+
+       /*
+        * force return value to old message size
+        */
+       ret = sizeof(old_msg_buf);
+
+       o_msg = &old_msg_buf.pfm_ovfl_msg;
+       n_msg = &msg_buf.pfm_ovfl_msg;
+
+       switch (msg_buf.type) {
+       case PFM_MSG_OVFL:
+               o_msg->msg_type   = PFM_MSG_OVFL;
+               o_msg->msg_ctx_fd = 0;
+               o_msg->msg_active_set = n_msg->msg_active_set;
+               o_msg->msg_tstamp = 0;
+
+               o_msg->msg_ovfl_pmds[0] = n_msg->msg_ovfl_pmds[0];
+               o_msg->msg_ovfl_pmds[1] = n_msg->msg_ovfl_pmds[1];
+               o_msg->msg_ovfl_pmds[2] = n_msg->msg_ovfl_pmds[2];
+               o_msg->msg_ovfl_pmds[3] = n_msg->msg_ovfl_pmds[3];
+               break;
+       case PFM_MSG_END:
+               o_msg->msg_type = PFM_MSG_END;
+               o_msg->msg_ctx_fd = 0;
+               o_msg->msg_tstamp = 0;
+               break;
+       default:
+               PFM_DBG("unknown msg type=%d", msg_buf.type);
+       }
+       if (copy_to_user(buf, &old_msg_buf, sizeof(old_msg_buf)))
+               ret = -EFAULT;
+       PFM_DBG_ovfl("ret=%d", ret);
+       return ret;
+}
+
+/*
+ * legacy /proc/perfmon simplified interface (we only maintain the
+ * global information (no more per-cpu stats, use
+ * /sys/devices/system/cpu/cpuXX/perfmon
+ */
+static struct proc_dir_entry   *perfmon_proc;
+
+static void *pfm_proc_start(struct seq_file *m, loff_t *pos)
+{
+       if (*pos == 0)
+               return (void *)1;
+
+       return NULL;
+}
+
+static void *pfm_proc_next(struct seq_file *m, void *v, loff_t *pos)
+{
+       ++*pos;
+       return pfm_proc_start(m, pos);
+}
+
+static void pfm_proc_stop(struct seq_file *m, void *v)
+{
+}
+
+/*
+ * this is a simplified version of the legacy /proc/perfmon.
+ * We have retained ONLY the key information that tools are actually
+ * using
+ */
+static void pfm_proc_show_header(struct seq_file *m)
+{
+       char buf[128];
+
+       pfm_sysfs_res_show(buf, sizeof(buf), 3);
+
+       seq_printf(m, "perfmon version            : %u.%u\n",
+               PFM_VERSION_MAJ, PFM_VERSION_MIN);
+
+       seq_printf(m, "model                      : %s", buf);
+}
+
+static int pfm_proc_show(struct seq_file *m, void *v)
+{
+       pfm_proc_show_header(m);
+       return 0;
+}
+
+struct seq_operations pfm_proc_seq_ops = {
+       .start = pfm_proc_start,
+       .next = pfm_proc_next,
+       .stop = pfm_proc_stop,
+       .show = pfm_proc_show
+};
+
+static int pfm_proc_open(struct inode *inode, struct file *file)
+{
+       return seq_open(file, &pfm_proc_seq_ops);
+}
+
+
+static struct file_operations pfm_proc_fops = {
+       .open = pfm_proc_open,
+       .read = seq_read,
+       .llseek = seq_lseek,
+       .release = seq_release,
+};
+
+/*
+ * called from pfm_arch_init(), global initialization, called once
+ */
+int __init pfm_ia64_compat_init(void)
+{
+       /*
+        * create /proc/perfmon
+        */
+       perfmon_proc = create_proc_entry("perfmon", S_IRUGO, NULL);
+       if (perfmon_proc == NULL) {
+               PFM_ERR("cannot create /proc entry, perfmon disabled");
+               return -1;
+       }
+       perfmon_proc->proc_fops = &pfm_proc_fops;
+       return 0;
+}
diff --git a/arch/ia64/perfmon/perfmon_default_smpl.c b/arch/ia64/perfmon/perfmon_default_smpl.c
new file mode 100644
index 0000000..b408a13
--- /dev/null
+++ b/arch/ia64/perfmon/perfmon_default_smpl.c
@@ -0,0 +1,273 @@
+/*
+ * Copyright (c) 2002-2006 Hewlett-Packard Development Company, L.P.
+ * Contributed by Stephane Eranian <eranian@hpl.hp.com>
+ *
+ * This file implements the old default sampling buffer format
+ * for the Linux/ia64 perfmon-2 subsystem. This is for backward
+ * compatibility only. use the new default format in perfmon/
+ *
+ * This program is free software; you can redistribute it and/or
+ * modify it under the terms of version 2 of the GNU General Public
+ * License as published by the Free Software Foundation.
+ *
+ * This program is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+ * General Public License for more details.
+ *
+ * You should have received a copy of the GNU General Public License
+ * along with this program; if not, write to the Free Software
+ * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA
+ * 02111-1307 USA
+  */
+#include <linux/kernel.h>
+#include <linux/types.h>
+#include <linux/module.h>
+#include <linux/init.h>
+#include <linux/delay.h>
+#include <linux/smp.h>
+#include <linux/sysctl.h>
+
+#ifdef MODULE
+#define FMT_FLAGS      0
+#else
+#define FMT_FLAGS      PFM_FMTFL_IS_BUILTIN
+#endif
+
+#include <linux/perfmon_kern.h>
+#include <asm/perfmon_default_smpl.h>
+
+MODULE_AUTHOR("Stephane Eranian <eranian@hpl.hp.com>");
+MODULE_DESCRIPTION("perfmon old default sampling format");
+MODULE_LICENSE("GPL");
+
+static int pfm_default_fmt_validate(u32 flags, u16 npmds, void *data)
+{
+       struct pfm_default_smpl_arg *arg = data;
+       size_t min_buf_size;
+
+       if (data == NULL) {
+               PFM_DBG("no argument passed");
+               return -EINVAL;
+       }
+
+       /*
+        * compute min buf size. All PMD are manipulated as 64bit entities
+        */
+       min_buf_size = sizeof(struct pfm_default_smpl_hdr)
+            + (sizeof(struct pfm_default_smpl_entry) + (npmds*sizeof(u64)));
+
+       PFM_DBG("validate flags=0x%x npmds=%u min_buf_size=%lu "
+                 "buf_size=%lu CPU%d", flags, npmds, min_buf_size,
+                 arg->buf_size, smp_processor_id());
+
+       /*
+        * must hold at least the buffer header + one minimally sized entry
+        */
+       if (arg->buf_size < min_buf_size)
+               return -EINVAL;
+
+       return 0;
+}
+
+static int pfm_default_fmt_get_size(unsigned int flags, void *data,
+                                   size_t *size)
+{
+       struct pfm_default_smpl_arg *arg = data;
+
+       /*
+        * size has been validated in default_validate
+        */
+       *size = arg->buf_size;
+
+       return 0;
+}
+
+static int pfm_default_fmt_init(struct pfm_context *ctx, void *buf,
+                               u32 flags, u16 npmds, void *data)
+{
+       struct pfm_default_smpl_hdr *hdr;
+       struct pfm_default_smpl_arg *arg = data;
+
+       hdr = buf;
+
+       hdr->hdr_version      = PFM_DEFAULT_SMPL_VERSION;
+       hdr->hdr_buf_size     = arg->buf_size;
+       hdr->hdr_cur_offs     = sizeof(*hdr);
+       hdr->hdr_overflows    = 0;
+       hdr->hdr_count        = 0;
+
+       PFM_DBG("buffer=%p buf_size=%lu hdr_size=%lu "
+                 "hdr_version=%u cur_offs=%lu",
+                 buf,
+                 hdr->hdr_buf_size,
+                 sizeof(*hdr),
+                 hdr->hdr_version,
+                 hdr->hdr_cur_offs);
+
+       return 0;
+}
+
+static int pfm_default_fmt_handler(struct pfm_context *ctx,
+                                  unsigned long ip, u64 tstamp, void *data)
+{
+       struct pfm_default_smpl_hdr *hdr;
+       struct pfm_default_smpl_entry *ent;
+       void *cur, *last, *buf;
+       u64 *e;
+       size_t entry_size;
+       u16 npmds, i, ovfl_pmd;
+       struct pfm_ovfl_arg *arg;
+
+       hdr = ctx->smpl_addr;
+       arg = &ctx->ovfl_arg;
+
+       buf = hdr;
+       cur = buf+hdr->hdr_cur_offs;
+       last = buf+hdr->hdr_buf_size;
+       ovfl_pmd = arg->ovfl_pmd;
+
+       /*
+        * precheck for sanity
+        */
+       if ((last - cur) < PFM_DEFAULT_MAX_ENTRY_SIZE)
+               goto full;
+
+       npmds = arg->num_smpl_pmds;
+
+       ent = cur;
+
+       prefetch(arg->smpl_pmds_values);
+
+       entry_size = sizeof(*ent) + (npmds << 3);
+
+       /* position for first pmd */
+       e = (unsigned long *)(ent+1);
+
+       hdr->hdr_count++;
+
+       PFM_DBG_ovfl("count=%lu cur=%p last=%p free_bytes=%lu "
+                      "ovfl_pmd=%d npmds=%u",
+                      hdr->hdr_count,
+                      cur, last,
+                      last-cur,
+                      ovfl_pmd,
+                      npmds);
+
+       /*
+        * current = task running at the time of the overflow.
+        *
+        * per-task mode:
+        *      - this is ususally the task being monitored.
+        *        Under certain conditions, it might be a different task
+        *
+        * system-wide:
+        *      - this is not necessarily the task controlling the session
+        */
+       ent->pid            = current->pid;
+       ent->ovfl_pmd       = ovfl_pmd;
+       ent->last_reset_val = arg->pmd_last_reset;
+
+       /*
+        * where did the fault happen (includes slot number)
+        */
+       ent->ip = ip;
+
+       ent->tstamp    = tstamp;
+       ent->cpu       = smp_processor_id();
+       ent->set       = arg->active_set;
+       ent->tgid      = current->tgid;
+
+       /*
+        * selectively store PMDs in increasing index number
+        */
+       if (npmds) {
+               u64 *val = arg->smpl_pmds_values;
+               for (i = 0; i < npmds; i++)
+                       *e++ = *val++;
+       }
+
+       /*
+        * update position for next entry
+        */
+       hdr->hdr_cur_offs += entry_size;
+       cur               += entry_size;
+
+       /*
+        * post check to avoid losing the last sample
+        */
+       if ((last - cur) < PFM_DEFAULT_MAX_ENTRY_SIZE)
+               goto full;
+
+       /*
+        * reset before returning from interrupt handler
+        */
+       arg->ovfl_ctrl = PFM_OVFL_CTRL_RESET;
+       return 0;
+full:
+       PFM_DBG_ovfl("smpl buffer full free=%lu, count=%lu",
+                      last-cur, hdr->hdr_count);
+
+       /*
+        * increment number of buffer overflow.
+        * important to detect duplicate set of samples.
+        */
+       hdr->hdr_overflows++;
+
+       /*
+        * request notification and masking of monitoring.
+        * Notification is still subject to the overflowed
+        */
+       arg->ovfl_ctrl = PFM_OVFL_CTRL_NOTIFY | PFM_OVFL_CTRL_MASK;
+
+       return -ENOBUFS; /* we are full, sorry */
+}
+
+static int pfm_default_fmt_restart(int is_active, u32 *ovfl_ctrl, void *buf)
+{
+       struct pfm_default_smpl_hdr *hdr;
+
+       hdr = buf;
+
+       hdr->hdr_count    = 0;
+       hdr->hdr_cur_offs = sizeof(*hdr);
+
+       *ovfl_ctrl = PFM_OVFL_CTRL_RESET;
+
+       return 0;
+}
+
+static int pfm_default_fmt_exit(void *buf)
+{
+       return 0;
+}
+
+static struct pfm_smpl_fmt default_fmt = {
+       .fmt_name = "default-old",
+       .fmt_version = 0x10000,
+       .fmt_arg_size = sizeof(struct pfm_default_smpl_arg),
+       .fmt_validate = pfm_default_fmt_validate,
+       .fmt_getsize = pfm_default_fmt_get_size,
+       .fmt_init = pfm_default_fmt_init,
+       .fmt_handler = pfm_default_fmt_handler,
+       .fmt_restart = pfm_default_fmt_restart,
+       .fmt_exit = pfm_default_fmt_exit,
+       .fmt_flags = FMT_FLAGS,
+       .owner = THIS_MODULE
+};
+
+static int pfm_default_fmt_init_module(void)
+{
+       int ret;
+
+       return pfm_fmt_register(&default_fmt);
+       return ret;
+}
+
+static void pfm_default_fmt_cleanup_module(void)
+{
+       pfm_fmt_unregister(&default_fmt);
+}
+
+module_init(pfm_default_fmt_init_module);
+module_exit(pfm_default_fmt_cleanup_module);
diff --git a/arch/ia64/perfmon/perfmon_generic.c b/arch/ia64/perfmon/perfmon_generic.c
new file mode 100644
index 0000000..47b1870
--- /dev/null
+++ b/arch/ia64/perfmon/perfmon_generic.c
@@ -0,0 +1,148 @@
+/*
+ * This file contains the generic PMU register description tables
+ * and pmc checker used by perfmon.c.
+ *
+ * Copyright (c) 2002-2006 Hewlett-Packard Development Company, L.P.
+ * contributed by Stephane Eranian <eranian@hpl.hp.com>
+ *
+ * This program is free software; you can redistribute it and/or
+ * modify it under the terms of version 2 of the GNU General Public
+ * License as published by the Free Software Foundation.
+ *
+ * This program is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+ * General Public License for more details.
+ *
+ * You should have received a copy of the GNU General Public License
+ * along with this program; if not, write to the Free Software
+ * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA
+ * 02111-1307 USA
+  */
+#include <linux/module.h>
+#include <linux/perfmon_kern.h>
+#include <asm/pal.h>
+
+MODULE_AUTHOR("Stephane Eranian <eranian@hpl.hp.com>");
+MODULE_DESCRIPTION("Generic IA-64 PMU description tables");
+MODULE_LICENSE("GPL");
+
+#define RDEP(x)        (1UL << (x))
+
+#define PFM_IA64GEN_MASK_PMCS  (RDEP(4)|RDEP(5)|RDEP(6)|RDEP(7))
+#define PFM_IA64GEN_RSVD       (0xffffffffffff0080UL)
+#define PFM_IA64GEN_NO64       (1UL<<5)
+
+/* forward declaration */
+static struct pfm_pmu_config pfm_ia64gen_pmu_conf;
+
+static struct pfm_arch_pmu_info pfm_ia64gen_pmu_info = {
+       .mask_pmcs = {PFM_IA64GEN_MASK_PMCS,},
+};
+
+static struct pfm_regmap_desc pfm_ia64gen_pmc_desc[] = {
+/* pmc0  */ PMX_NA,
+/* pmc1  */ PMX_NA,
+/* pmc2  */ PMX_NA,
+/* pmc3  */ PMX_NA,
+/* pmc4  */ PMC_D(PFM_REG_W64, "PMC4", 0x0, PFM_IA64GEN_RSVD, PFM_IA64GEN_NO64, 4),
+/* pmc5  */ PMC_D(PFM_REG_W64, "PMC5", 0x0, PFM_IA64GEN_RSVD, PFM_IA64GEN_NO64, 5),
+/* pmc6  */ PMC_D(PFM_REG_W64, "PMC6", 0x0, PFM_IA64GEN_RSVD, PFM_IA64GEN_NO64, 6),
+/* pmc7  */ PMC_D(PFM_REG_W64, "PMC7", 0x0, PFM_IA64GEN_RSVD, PFM_IA64GEN_NO64, 7)
+};
+#define PFM_IA64GEN_NUM_PMCS ARRAY_SIZE(pfm_ia64gen_pmc_desc)
+
+static struct pfm_regmap_desc pfm_ia64gen_pmd_desc[] = {
+/* pmd0  */ PMX_NA,
+/* pmd1  */ PMX_NA,
+/* pmd2  */ PMX_NA,
+/* pmd3  */ PMX_NA,
+/* pmd4  */ PMD_DP(PFM_REG_C, "PMD4", 4, 1ull << 4),
+/* pmd5  */ PMD_DP(PFM_REG_C, "PMD5", 5, 1ull << 5),
+/* pmd6  */ PMD_DP(PFM_REG_C, "PMD6", 6, 1ull << 6),
+/* pmd7  */ PMD_DP(PFM_REG_C, "PMD7", 7, 1ull << 7)
+};
+#define PFM_IA64GEN_NUM_PMDS ARRAY_SIZE(pfm_ia64gen_pmd_desc)
+
+static int pfm_ia64gen_pmc_check(struct pfm_context *ctx,
+                                struct pfm_event_set *set,
+                                struct pfarg_pmc *req)
+{
+#define PFM_IA64GEN_PMC_PM_POS6        (1UL<<6)
+       u64 tmpval;
+       int is_system;
+
+       is_system = ctx->flags.system;
+       tmpval = req->reg_value;
+
+       switch (req->reg_num) {
+       case  4:
+       case  5:
+       case  6:
+       case  7:
+               /* set pmc.oi for 64-bit emulation */
+               tmpval |= 1UL << 5;
+
+               if (is_system)
+                       tmpval |= PFM_IA64GEN_PMC_PM_POS6;
+               else
+                       tmpval &= ~PFM_IA64GEN_PMC_PM_POS6;
+               break;
+
+       }
+       req->reg_value = tmpval;
+
+       return 0;
+}
+
+/*
+ * matches anything
+ */
+static int pfm_ia64gen_probe_pmu(void)
+{
+       u64 pm_buffer[16];
+       pal_perf_mon_info_u_t pm_info;
+
+       /*
+        * call PAL_PERFMON_INFO to retrieve counter width which
+        * is implementation specific
+        */
+       if (ia64_pal_perf_mon_info(pm_buffer, &pm_info))
+               return -1;
+
+       pfm_ia64gen_pmu_conf.counter_width = pm_info.pal_perf_mon_info_s.width;
+
+       return 0;
+}
+
+/*
+ * impl_pmcs, impl_pmds are computed at runtime to minimize errors!
+ */
+static struct pfm_pmu_config pfm_ia64gen_pmu_conf = {
+       .pmu_name = "Generic IA-64",
+       .counter_width = 0, /* computed from PAL_PERFMON_INFO */
+       .pmd_desc = pfm_ia64gen_pmd_desc,
+       .pmc_desc = pfm_ia64gen_pmc_desc,
+       .probe_pmu = pfm_ia64gen_probe_pmu,
+       .num_pmc_entries = PFM_IA64GEN_NUM_PMCS,
+       .num_pmd_entries = PFM_IA64GEN_NUM_PMDS,
+       .pmc_write_check = pfm_ia64gen_pmc_check,
+       .version = "1.0",
+       .flags = PFM_PMU_BUILTIN_FLAG,
+       .owner = THIS_MODULE,
+       .pmu_info = &pfm_ia64gen_pmu_info
+       /* no read/write checkers */
+};
+
+static int __init pfm_gen_pmu_init_module(void)
+{
+       return pfm_pmu_register(&pfm_ia64gen_pmu_conf);
+}
+
+static void __exit pfm_gen_pmu_cleanup_module(void)
+{
+       pfm_pmu_unregister(&pfm_ia64gen_pmu_conf);
+}
+
+module_init(pfm_gen_pmu_init_module);
+module_exit(pfm_gen_pmu_cleanup_module);
diff --git a/arch/ia64/perfmon/perfmon_itanium.c b/arch/ia64/perfmon/perfmon_itanium.c
new file mode 100644
index 0000000..094b31b
--- /dev/null
+++ b/arch/ia64/perfmon/perfmon_itanium.c
@@ -0,0 +1,232 @@
+/*
+ * This file contains the Itanium PMU register description tables
+ * and pmc checker used by perfmon.c.
+ *
+ * Copyright (c) 2002-2006 Hewlett-Packard Development Company, L.P.
+ * Contributed by Stephane Eranian <eranian@hpl.hp.com>
+ *
+ * This program is free software; you can redistribute it and/or
+ * modify it under the terms of version 2 of the GNU General Public
+ * License as published by the Free Software Foundation.
+ *
+ * This program is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+ * General Public License for more details.
+ *
+ * You should have received a copy of the GNU General Public License
+ * along with this program; if not, write to the Free Software
+ * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA
+ * 02111-1307 USA
+  */
+#include <linux/module.h>
+#include <linux/perfmon_kern.h>
+
+MODULE_AUTHOR("Stephane Eranian <eranian@hpl.hp.com>");
+MODULE_DESCRIPTION("Itanium (Merced) PMU description tables");
+MODULE_LICENSE("GPL");
+
+#define RDEP(x)        (1ULL << (x))
+
+#define PFM_ITA_MASK_PMCS (RDEP(4)|RDEP(5)|RDEP(6)|RDEP(7)|RDEP(10)|RDEP(11)|\
+                          RDEP(12))
+
+#define PFM_ITA_NO64   (1ULL<<5)
+
+static struct pfm_arch_pmu_info pfm_ita_pmu_info = {
+       .mask_pmcs = {PFM_ITA_MASK_PMCS,},
+};
+/* reserved bits are 1 in the mask */
+#define PFM_ITA_RSVD 0xfffffffffc8000a0UL
+/*
+ * For debug registers, writing xBR(y) means we use also xBR(y+1). Hence using
+ * PMC256+y means we use PMC256+y+1.  Yet, we do not have dependency information
+ * but this is fine because they are handled separately in the IA-64 specific
+ * code.
+ */
+static struct pfm_regmap_desc pfm_ita_pmc_desc[] = {
+/* pmc0  */ PMX_NA,
+/* pmc1  */ PMX_NA,
+/* pmc2  */ PMX_NA,
+/* pmc3  */ PMX_NA,
+/* pmc4  */ PMC_D(PFM_REG_W64, "PMC4" , 0x20, PFM_ITA_RSVD, PFM_ITA_NO64, 4),
+/* pmc5  */ PMC_D(PFM_REG_W64, "PMC5" , 0x20, PFM_ITA_RSVD, PFM_ITA_NO64, 5),
+/* pmc6  */ PMC_D(PFM_REG_W64, "PMC6" , 0x20, PFM_ITA_RSVD, PFM_ITA_NO64, 6),
+/* pmc7  */ PMC_D(PFM_REG_W64, "PMC7" , 0x20, PFM_ITA_RSVD, PFM_ITA_NO64, 7),
+/* pmc8  */ PMC_D(PFM_REG_W  , "PMC8" , 0xfffffffe3ffffff8UL, 0xfff00000001c0000UL, 0, 8),
+/* pmc9  */ PMC_D(PFM_REG_W  , "PMC9" , 0xfffffffe3ffffff8UL, 0xfff00000001c0000UL, 0, 9),
+/* pmc10 */ PMC_D(PFM_REG_W  , "PMC10", 0x0, 0xfffffffff3f0ff30UL, 0, 10),
+/* pmc11 */ PMC_D(PFM_REG_W  , "PMC11", 0x10000000UL, 0xffffffffecf0ff30UL, 0, 11),
+/* pmc12 */ PMC_D(PFM_REG_W  , "PMC12", 0x0, 0xffffffffffff0030UL, 0, 12),
+/* pmc13 */ PMC_D(PFM_REG_W  , "PMC13", 0x3ffff00000001UL, 0xfffffffffffffffeUL, 0, 13),
+/* pmc14 */ PMX_NA,
+/* pmc15 */ PMX_NA,
+/* pmc16 */  PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA,
+/* pmc24 */  PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA,
+/* pmc32 */  PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA,
+/* pmc40 */  PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA,
+/* pmc48 */  PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA,
+/* pmc56 */  PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA,
+/* pmc64 */  PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA,
+/* pmc72 */  PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA,
+/* pmc80 */  PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA,
+/* pmc88 */  PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA,
+/* pmc96 */  PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA,
+/* pmc104 */ PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA,
+/* pmc112 */ PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA,
+/* pmc120 */ PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA,
+/* pmc128 */ PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA,
+/* pmc136 */ PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA,
+/* pmc144 */ PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA,
+/* pmc152 */ PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA,
+/* pmc160 */ PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA,
+/* pmc168 */ PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA,
+/* pmc176 */ PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA,
+/* pmc184 */ PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA,
+/* pmc192 */ PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA,
+/* pmc200 */ PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA,
+/* pmc208 */ PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA,
+/* pmc216 */ PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA,
+/* pmc224 */ PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA,
+/* pmc232 */ PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA,
+/* pmc240 */ PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA,
+/* pmc248 */ PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA,
+/* pmc256 */ PMC_D(PFM_REG_W  , "IBR0", 0x0, 0, 0, 0),
+/* pmc257 */ PMC_D(PFM_REG_W  , "IBR1", 0x0, 0x8000000000000000UL, 0, 1),
+/* pmc258 */ PMC_D(PFM_REG_W  , "IBR2", 0x0, 0, 0, 2),
+/* pmc259 */ PMC_D(PFM_REG_W  , "IBR3", 0x0, 0x8000000000000000UL, 0, 3),
+/* pmc260 */ PMC_D(PFM_REG_W  , "IBR4", 0x0, 0, 0, 4),
+/* pmc261 */ PMC_D(PFM_REG_W  , "IBR5", 0x0, 0x8000000000000000UL, 0, 5),
+/* pmc262 */ PMC_D(PFM_REG_W  , "IBR6", 0x0, 0, 0, 6),
+/* pmc263 */ PMC_D(PFM_REG_W  , "IBR7", 0x0, 0x8000000000000000UL, 0, 7),
+/* pmc264 */ PMC_D(PFM_REG_W  , "DBR0", 0x0, 0, 0, 0),
+/* pmc265 */ PMC_D(PFM_REG_W  , "DBR1", 0x0, 0xc000000000000000UL, 0, 1),
+/* pmc266 */ PMC_D(PFM_REG_W  , "DBR2", 0x0, 0, 0, 2),
+/* pmc267 */ PMC_D(PFM_REG_W  , "DBR3", 0x0, 0xc000000000000000UL, 0, 3),
+/* pmc268 */ PMC_D(PFM_REG_W  , "DBR4", 0x0, 0, 0, 4),
+/* pmc269 */ PMC_D(PFM_REG_W  , "DBR5", 0x0, 0xc000000000000000UL, 0, 5),
+/* pmc270 */ PMC_D(PFM_REG_W  , "DBR6", 0x0, 0, 0, 6),
+/* pmc271 */ PMC_D(PFM_REG_W  , "DBR7", 0x0, 0xc000000000000000UL, 0, 7)
+};
+#define PFM_ITA_NUM_PMCS ARRAY_SIZE(pfm_ita_pmc_desc)
+
+static struct pfm_regmap_desc pfm_ita_pmd_desc[] = {
+/* pmd0  */ PMD_DP(PFM_REG_I , "PMD0", 0, 1ull << 10),
+/* pmd1  */ PMD_DP(PFM_REG_I , "PMD1", 1, 1ull << 10),
+/* pmd2  */ PMD_DP(PFM_REG_I , "PMD2", 2, 1ull << 11),
+/* pmd3  */ PMD_DP(PFM_REG_I , "PMD3", 3, 1ull << 11),
+/* pmd4  */ PMD_DP(PFM_REG_C , "PMD4", 4, 1ull << 4),
+/* pmd5  */ PMD_DP(PFM_REG_C , "PMD5", 5, 1ull << 5),
+/* pmd6  */ PMD_DP(PFM_REG_C , "PMD6", 6, 1ull << 6),
+/* pmd7  */ PMD_DP(PFM_REG_C , "PMD7", 7, 1ull << 7),
+/* pmd8  */ PMD_DP(PFM_REG_I , "PMD8", 8, 1ull << 12),
+/* pmd9  */ PMD_DP(PFM_REG_I , "PMD9", 9, 1ull << 12),
+/* pmd10 */ PMD_DP(PFM_REG_I , "PMD10", 10, 1ull << 12),
+/* pmd11 */ PMD_DP(PFM_REG_I , "PMD11", 11, 1ull << 12),
+/* pmd12 */ PMD_DP(PFM_REG_I , "PMD12", 12, 1ull << 12),
+/* pmd13 */ PMD_DP(PFM_REG_I , "PMD13", 13, 1ull << 12),
+/* pmd14 */ PMD_DP(PFM_REG_I , "PMD14", 14, 1ull << 12),
+/* pmd15 */ PMD_DP(PFM_REG_I , "PMD15", 15, 1ull << 12),
+/* pmd16 */ PMD_DP(PFM_REG_I , "PMD16", 16, 1ull << 12),
+/* pmd17 */ PMD_DP(PFM_REG_I , "PMD17", 17, 1ull << 11)
+};
+#define PFM_ITA_NUM_PMDS ARRAY_SIZE(pfm_ita_pmd_desc)
+
+static int pfm_ita_pmc_check(struct pfm_context *ctx,
+                            struct pfm_event_set *set,
+                            struct pfarg_pmc *req)
+{
+#define PFM_ITA_PMC_PM_POS6    (1UL<<6)
+       struct pfm_arch_context *ctx_arch;
+       u64 tmpval;
+       u16 cnum;
+       int ret = 0, is_system;
+
+       tmpval = req->reg_value;
+       cnum   = req->reg_num;
+       ctx_arch = pfm_ctx_arch(ctx);
+       is_system = ctx->flags.system;
+
+       switch (cnum) {
+       case  4:
+       case  5:
+       case  6:
+       case  7:
+       case 10:
+       case 11:
+       case 12:
+               if (is_system)
+                       tmpval |= PFM_ITA_PMC_PM_POS6;
+               else
+                       tmpval &= ~PFM_ITA_PMC_PM_POS6;
+               break;
+       }
+
+       /*
+        * we must clear the (instruction) debug registers if pmc13.ta bit is
+        * cleared before they are written (fl_using_dbreg==0) to avoid
+        * picking up stale information.
+        */
+       if (cnum == 13 && ((tmpval & 0x1) == 0)
+               && ctx_arch->flags.use_dbr == 0) {
+               PFM_DBG("pmc13 has pmc13.ta cleared, clearing ibr");
+               ret = pfm_ia64_mark_dbregs_used(ctx, set);
+               if (ret)
+                       return ret;
+       }
+
+       /*
+        * we must clear the (data) debug registers if pmc11.pt bit is cleared
+        * before they are written (fl_using_dbreg==0) to avoid picking up
+        * stale information.
+        */
+       if (cnum == 11 && ((tmpval >> 28) & 0x1) == 0
+               && ctx_arch->flags.use_dbr == 0) {
+               PFM_DBG("pmc11 has pmc11.pt cleared, clearing dbr");
+               ret = pfm_ia64_mark_dbregs_used(ctx, set);
+               if (ret)
+                       return ret;
+       }
+
+       req->reg_value = tmpval;
+
+       return 0;
+}
+
+static int pfm_ita_probe_pmu(void)
+{
+       return local_cpu_data->family == 0x7 && !ia64_platform_is("hpsim")
+               ? 0 : -1;
+}
+
+/*
+ * impl_pmcs, impl_pmds are computed at runtime to minimize errors!
+ */
+static struct pfm_pmu_config pfm_ita_pmu_conf = {
+       .pmu_name = "Itanium",
+       .counter_width = 32,
+       .pmd_desc = pfm_ita_pmd_desc,
+       .pmc_desc = pfm_ita_pmc_desc,
+       .pmc_write_check = pfm_ita_pmc_check,
+       .num_pmc_entries = PFM_ITA_NUM_PMCS,
+       .num_pmd_entries = PFM_ITA_NUM_PMDS,
+       .probe_pmu = pfm_ita_probe_pmu,
+       .version = "1.0",
+       .flags = PFM_PMU_BUILTIN_FLAG,
+       .owner = THIS_MODULE,
+       .pmu_info = &pfm_ita_pmu_info
+};
+
+static int __init pfm_ita_pmu_init_module(void)
+{
+       return pfm_pmu_register(&pfm_ita_pmu_conf);
+}
+
+static void __exit pfm_ita_pmu_cleanup_module(void)
+{
+       pfm_pmu_unregister(&pfm_ita_pmu_conf);
+}
+
+module_init(pfm_ita_pmu_init_module);
+module_exit(pfm_ita_pmu_cleanup_module);
+
diff --git a/arch/ia64/perfmon/perfmon_mckinley.c b/arch/ia64/perfmon/perfmon_mckinley.c
new file mode 100644
index 0000000..dc59092
--- /dev/null
+++ b/arch/ia64/perfmon/perfmon_mckinley.c
@@ -0,0 +1,290 @@
+/*
+ * This file contains the McKinley PMU register description tables
+ * and pmc checker used by perfmon.c.
+ *
+ * Copyright (c) 2002-2006 Hewlett-Packard Development Company, L.P.
+ * Contributed by Stephane Eranian <eranian@hpl.hp.com>
+ *
+ * This program is free software; you can redistribute it and/or
+ * modify it under the terms of version 2 of the GNU General Public
+ * License as published by the Free Software Foundation.
+ *
+ * This program is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+ * General Public License for more details.
+ *
+ * You should have received a copy of the GNU General Public License
+ * along with this program; if not, write to the Free Software
+ * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA
+ * 02111-1307 USA
+  */
+#include <linux/module.h>
+#include <linux/perfmon_kern.h>
+
+MODULE_AUTHOR("Stephane Eranian <eranian@hpl.hp.com>");
+MODULE_DESCRIPTION("Itanium 2 (McKinley) PMU description tables");
+MODULE_LICENSE("GPL");
+
+#define RDEP(x)        (1UL << (x))
+
+#define PFM_MCK_MASK_PMCS (RDEP(4)|RDEP(5)|RDEP(6)|RDEP(7)|RDEP(10)|RDEP(11)|\
+                          RDEP(12))
+
+#define PFM_MCK_NO64   (1UL<<5)
+
+static struct pfm_arch_pmu_info pfm_mck_pmu_info = {
+       .mask_pmcs = {PFM_MCK_MASK_PMCS,},
+};
+
+/* reserved bits are 1 in the mask */
+#define PFM_ITA2_RSVD 0xfffffffffc8000a0UL
+
+/*
+ * For debug registers, writing xBR(y) means we use also xBR(y+1). Hence using
+ * PMC256+y means we use PMC256+y+1.  Yet, we do not have dependency information
+ * but this is fine because they are handled separately in the IA-64 specific
+ * code.
+ */
+static struct pfm_regmap_desc pfm_mck_pmc_desc[] = {
+/* pmc0  */ PMX_NA,
+/* pmc1  */ PMX_NA,
+/* pmc2  */ PMX_NA,
+/* pmc3  */ PMX_NA,
+/* pmc4  */ PMC_D(PFM_REG_W64, "PMC4" , 0x800020UL, 0xfffffffffc8000a0, PFM_MCK_NO64, 4),
+/* pmc5  */ PMC_D(PFM_REG_W64, "PMC5" , 0x20UL, PFM_ITA2_RSVD, PFM_MCK_NO64, 5),
+/* pmc6  */ PMC_D(PFM_REG_W64, "PMC6" , 0x20UL, PFM_ITA2_RSVD, PFM_MCK_NO64, 6),
+/* pmc7  */ PMC_D(PFM_REG_W64, "PMC7" , 0x20UL, PFM_ITA2_RSVD, PFM_MCK_NO64, 7),
+/* pmc8  */ PMC_D(PFM_REG_W  , "PMC8" , 0xffffffff3fffffffUL, 0xc0000004UL, 0, 8),
+/* pmc9  */ PMC_D(PFM_REG_W  , "PMC9" , 0xffffffff3ffffffcUL, 0xc0000004UL, 0, 9),
+/* pmc10 */ PMC_D(PFM_REG_W  , "PMC10", 0x0, 0xffffffffffff0000UL, 0, 10),
+/* pmc11 */ PMC_D(PFM_REG_W  , "PMC11", 0x0, 0xfffffffffcf0fe30UL, 0, 11),
+/* pmc12 */ PMC_D(PFM_REG_W  , "PMC12", 0x0, 0xffffffffffff0000UL, 0, 12),
+/* pmc13 */ PMC_D(PFM_REG_W  , "PMC13", 0x2078fefefefeUL, 0xfffe1fffe7e7e7e7UL, 0, 13),
+/* pmc14 */ PMC_D(PFM_REG_W  , "PMC14", 0x0db60db60db60db6UL, 0xffffffffffffdb6dUL, 0, 14),
+/* pmc15 */ PMC_D(PFM_REG_W  , "PMC15", 0xfffffff0UL, 0xfffffffffffffff0UL, 0, 15),
+/* pmc16 */  PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA,
+/* pmc24 */  PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA,
+/* pmc32 */  PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA,
+/* pmc40 */  PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA,
+/* pmc48 */  PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA,
+/* pmc56 */  PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA,
+/* pmc64 */  PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA,
+/* pmc72 */  PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA,
+/* pmc80 */  PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA,
+/* pmc88 */  PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA,
+/* pmc96 */  PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA,
+/* pmc104 */ PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA,
+/* pmc112 */ PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA,
+/* pmc120 */ PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA,
+/* pmc128 */ PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA,
+/* pmc136 */ PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA,
+/* pmc144 */ PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA,
+/* pmc152 */ PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA,
+/* pmc160 */ PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA,
+/* pmc168 */ PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA,
+/* pmc176 */ PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA,
+/* pmc184 */ PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA,
+/* pmc192 */ PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA,
+/* pmc200 */ PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA,
+/* pmc208 */ PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA,
+/* pmc216 */ PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA,
+/* pmc224 */ PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA,
+/* pmc232 */ PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA,
+/* pmc240 */ PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA,
+/* pmc248 */ PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA,
+/* pmc256 */ PMC_D(PFM_REG_W  , "IBR0", 0x0, 0, 0, 0),
+/* pmc257 */ PMC_D(PFM_REG_W  , "IBR1", 0x0, 0x8000000000000000UL, 0, 1),
+/* pmc258 */ PMC_D(PFM_REG_W  , "IBR2", 0x0, 0, 0, 2),
+/* pmc259 */ PMC_D(PFM_REG_W  , "IBR3", 0x0, 0x8000000000000000UL, 0, 3),
+/* pmc260 */ PMC_D(PFM_REG_W  , "IBR4", 0x0, 0, 0, 4),
+/* pmc261 */ PMC_D(PFM_REG_W  , "IBR5", 0x0, 0x8000000000000000UL, 0, 5),
+/* pmc262 */ PMC_D(PFM_REG_W  , "IBR6", 0x0, 0, 0, 6),
+/* pmc263 */ PMC_D(PFM_REG_W  , "IBR7", 0x0, 0x8000000000000000UL, 0, 7),
+/* pmc264 */ PMC_D(PFM_REG_W  , "DBR0", 0x0, 0, 0, 0),
+/* pmc265 */ PMC_D(PFM_REG_W  , "DBR1", 0x0, 0xc000000000000000UL, 0, 1),
+/* pmc266 */ PMC_D(PFM_REG_W  , "DBR2", 0x0, 0, 0, 2),
+/* pmc267 */ PMC_D(PFM_REG_W  , "DBR3", 0x0, 0xc000000000000000UL, 0, 3),
+/* pmc268 */ PMC_D(PFM_REG_W  , "DBR4", 0x0, 0, 0, 4),
+/* pmc269 */ PMC_D(PFM_REG_W  , "DBR5", 0x0, 0xc000000000000000UL, 0, 5),
+/* pmc270 */ PMC_D(PFM_REG_W  , "DBR6", 0x0, 0, 0, 6),
+/* pmc271 */ PMC_D(PFM_REG_W  , "DBR7", 0x0, 0xc000000000000000UL, 0, 7)
+};
+#define PFM_MCK_NUM_PMCS ARRAY_SIZE(pfm_mck_pmc_desc)
+
+static struct pfm_regmap_desc pfm_mck_pmd_desc[] = {
+/* pmd0  */ PMD_DP(PFM_REG_I, "PMD0", 0, 1ull << 10),
+/* pmd1  */ PMD_DP(PFM_REG_I, "PMD1", 1, 1ull << 10),
+/* pmd2  */ PMD_DP(PFM_REG_I, "PMD2", 2, 1ull << 11),
+/* pmd3  */ PMD_DP(PFM_REG_I, "PMD3", 3, 1ull << 11),
+/* pmd4  */ PMD_DP(PFM_REG_C, "PMD4", 4, 1ull << 4),
+/* pmd5  */ PMD_DP(PFM_REG_C, "PMD5", 5, 1ull << 5),
+/* pmd6  */ PMD_DP(PFM_REG_C, "PMD6", 6, 1ull << 6),
+/* pmd7  */ PMD_DP(PFM_REG_C, "PMD7", 7, 1ull << 7),
+/* pmd8  */ PMD_DP(PFM_REG_I, "PMD8", 8, 1ull << 12),
+/* pmd9  */ PMD_DP(PFM_REG_I, "PMD9", 9, 1ull << 12),
+/* pmd10 */ PMD_DP(PFM_REG_I, "PMD10", 10, 1ull << 12),
+/* pmd11 */ PMD_DP(PFM_REG_I, "PMD11", 11, 1ull << 12),
+/* pmd12 */ PMD_DP(PFM_REG_I, "PMD12", 12, 1ull << 12),
+/* pmd13 */ PMD_DP(PFM_REG_I, "PMD13", 13, 1ull << 12),
+/* pmd14 */ PMD_DP(PFM_REG_I, "PMD14", 14, 1ull << 12),
+/* pmd15 */ PMD_DP(PFM_REG_I, "PMD15", 15, 1ull << 12),
+/* pmd16 */ PMD_DP(PFM_REG_I, "PMD16", 16, 1ull << 12),
+/* pmd17 */ PMD_DP(PFM_REG_I, "PMD17", 17, 1ull << 11)
+};
+#define PFM_MCK_NUM_PMDS ARRAY_SIZE(pfm_mck_pmd_desc)
+
+static int pfm_mck_pmc_check(struct pfm_context *ctx,
+                            struct pfm_event_set *set,
+                            struct pfarg_pmc *req)
+{
+       struct pfm_arch_context *ctx_arch;
+       u64 val8 = 0, val14 = 0, val13 = 0;
+       u64 tmpval;
+       u16 cnum;
+       int ret = 0, check_case1 = 0;
+       int is_system;
+
+       tmpval = req->reg_value;
+       cnum = req->reg_num;
+       ctx_arch = pfm_ctx_arch(ctx);
+       is_system = ctx->flags.system;
+
+#define PFM_MCK_PMC_PM_POS6    (1UL<<6)
+#define PFM_MCK_PMC_PM_POS4    (1UL<<4)
+
+       switch (cnum) {
+       case  4:
+       case  5:
+       case  6:
+       case  7:
+       case 11:
+       case 12:
+               if (is_system)
+                       tmpval |= PFM_MCK_PMC_PM_POS6;
+               else
+                       tmpval &= ~PFM_MCK_PMC_PM_POS6;
+               break;
+
+       case  8:
+               val8 = tmpval;
+               val13 = set->pmcs[13];
+               val14 = set->pmcs[14];
+               check_case1 = 1;
+               break;
+
+       case 10:
+               if (is_system)
+                       tmpval |= PFM_MCK_PMC_PM_POS4;
+               else
+                       tmpval &= ~PFM_MCK_PMC_PM_POS4;
+               break;
+
+       case 13:
+               val8 = set->pmcs[8];
+               val13 = tmpval;
+               val14 = set->pmcs[14];
+               check_case1 = 1;
+               break;
+
+       case 14:
+               val8 = set->pmcs[8];
+               val13 = set->pmcs[13];
+               val14 = tmpval;
+               check_case1 = 1;
+               break;
+       }
+
+       /*
+        * check illegal configuration which can produce inconsistencies
+        * in tagging i-side events in L1D and L2 caches
+        */
+       if (check_case1) {
+               ret = (((val13 >> 45) & 0xf) == 0 && ((val8 & 0x1) == 0))
+                       && ((((val14>>1) & 0x3) == 0x2 || ((val14>>1) & 0x3) == 0x0)
+                       || (((val14>>4) & 0x3) == 0x2 || ((val14>>4) & 0x3) == 0x0));
+
+               if (ret) {
+                       PFM_DBG("perfmon: invalid config pmc8=0x%lx "
+                               "pmc13=0x%lx pmc14=0x%lx",
+                               val8, val13, val14);
+                       return -EINVAL;
+               }
+       }
+
+       /*
+        * check if configuration implicitely activates the use of
+        * the debug registers. If true, then we ensure that this is
+        * possible and that we do not pick up stale value in the HW
+        * registers.
+        *
+        * We postpone the checks of pmc13 and pmc14 to avoid side effects
+        * in case of errors
+        */
+
+       /*
+        * pmc13 is "active" if:
+        *      one of the pmc13.cfg_dbrpXX field is different from 0x3
+        * AND
+        *      at the corresponding pmc13.ena_dbrpXX is set.
+        */
+       if (cnum == 13 && (tmpval & 0x1e00000000000UL)
+           && (tmpval & 0x18181818UL) != 0x18181818UL
+               && ctx_arch->flags.use_dbr == 0) {
+               PFM_DBG("pmc13=0x%lx active", tmpval);
+               ret = pfm_ia64_mark_dbregs_used(ctx, set);
+               if (ret)
+                       return ret;
+       }
+
+       /*
+        *  if any pmc14.ibrpX bit is enabled we must clear the ibrs
+        */
+       if (cnum == 14 && ((tmpval & 0x2222UL) != 0x2222UL)
+               && ctx_arch->flags.use_dbr == 0) {
+               PFM_DBG("pmc14=0x%lx active", tmpval);
+               ret = pfm_ia64_mark_dbregs_used(ctx, set);
+               if (ret)
+                       return ret;
+       }
+
+       req->reg_value = tmpval;
+
+       return 0;
+}
+
+static int pfm_mck_probe_pmu(void)
+{
+       return local_cpu_data->family == 0x1f ? 0 : -1;
+}
+
+/*
+ * impl_pmcs, impl_pmds are computed at runtime to minimize errors!
+ */
+static struct pfm_pmu_config pfm_mck_pmu_conf = {
+       .pmu_name = "Itanium 2",
+       .counter_width = 47,
+       .pmd_desc = pfm_mck_pmd_desc,
+       .pmc_desc = pfm_mck_pmc_desc,
+       .pmc_write_check = pfm_mck_pmc_check,
+       .num_pmc_entries = PFM_MCK_NUM_PMCS,
+       .num_pmd_entries = PFM_MCK_NUM_PMDS,
+       .probe_pmu = pfm_mck_probe_pmu,
+       .version = "1.0",
+       .flags = PFM_PMU_BUILTIN_FLAG,
+       .owner = THIS_MODULE,
+       .pmu_info = &pfm_mck_pmu_info,
+};
+
+static int __init pfm_mck_pmu_init_module(void)
+{
+       return pfm_pmu_register(&pfm_mck_pmu_conf);
+}
+
+static void __exit pfm_mck_pmu_cleanup_module(void)
+{
+       pfm_pmu_unregister(&pfm_mck_pmu_conf);
+}
+
+module_init(pfm_mck_pmu_init_module);
+module_exit(pfm_mck_pmu_cleanup_module);
diff --git a/arch/ia64/perfmon/perfmon_montecito.c b/arch/ia64/perfmon/perfmon_montecito.c
new file mode 100644
index 0000000..3f76f73
--- /dev/null
+++ b/arch/ia64/perfmon/perfmon_montecito.c
@@ -0,0 +1,412 @@
+/*
+ * This file contains the McKinley PMU register description tables
+ * and pmc checker used by perfmon.c.
+ *
+ * Copyright (c) 2005-2006 Hewlett-Packard Development Company, L.P.
+ * Contributed Stephane Eranian <eranian@hpl.hp.com>
+ *
+ * This program is free software; you can redistribute it and/or
+ * modify it under the terms of version 2 of the GNU General Public
+ * License as published by the Free Software Foundation.
+ *
+ * This program is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+ * General Public License for more details.
+ *
+ * You should have received a copy of the GNU General Public License
+ * along with this program; if not, write to the Free Software
+ * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA
+ * 02111-1307 USA
+  */
+#include <linux/module.h>
+#include <linux/smp.h>
+#include <linux/perfmon_kern.h>
+
+MODULE_AUTHOR("Stephane Eranian <eranian@hpl.hp.com>");
+MODULE_DESCRIPTION("Dual-Core Itanium 2 (Montecito) PMU description table");
+MODULE_LICENSE("GPL");
+
+#define RDEP(x)        (1UL << (x))
+
+#define PFM_MONT_MASK_PMCS (RDEP(4)|RDEP(5)|RDEP(6)|RDEP(7)|\
+                           RDEP(8)|RDEP(9)|RDEP(10)|RDEP(11)|\
+                           RDEP(12)|RDEP(13)|RDEP(14)|RDEP(15)|\
+                           RDEP(37)|RDEP(39)|RDEP(40)|RDEP(42))
+
+#define PFM_MONT_NO64  (1UL<<5)
+
+static struct pfm_arch_pmu_info pfm_mont_pmu_info = {
+       .mask_pmcs = {PFM_MONT_MASK_PMCS,},
+};
+
+#define PFM_MONT_RSVD 0xffffffff838000a0UL
+/*
+ *
+ * For debug registers, writing xBR(y) means we use also xBR(y+1). Hence using
+ * PMC256+y means we use PMC256+y+1.  Yet, we do not have dependency information
+ * but this is fine because they are handled separately in the IA-64 specific
+ * code.
+ *
+ * For PMC4-PMC15, PMC40: we force pmc.ism=2 (IA-64 mode only)
+ */
+static struct pfm_regmap_desc pfm_mont_pmc_desc[] = {
+/* pmc0  */ PMX_NA,
+/* pmc1  */ PMX_NA,
+/* pmc2  */ PMX_NA,
+/* pmc3  */ PMX_NA,
+/* pmc4  */ PMC_D(PFM_REG_W64, "PMC4" , 0x2000020UL, PFM_MONT_RSVD, PFM_MONT_NO64, 4),
+/* pmc5  */ PMC_D(PFM_REG_W64, "PMC5" , 0x2000020UL, PFM_MONT_RSVD, PFM_MONT_NO64, 5),
+/* pmc6  */ PMC_D(PFM_REG_W64, "PMC6" , 0x2000020UL, PFM_MONT_RSVD, PFM_MONT_NO64, 6),
+/* pmc7  */ PMC_D(PFM_REG_W64, "PMC7" , 0x2000020UL, PFM_MONT_RSVD, PFM_MONT_NO64, 7),
+/* pmc8  */ PMC_D(PFM_REG_W64, "PMC8" , 0x2000020UL, PFM_MONT_RSVD, PFM_MONT_NO64, 8),
+/* pmc9  */ PMC_D(PFM_REG_W64, "PMC9" , 0x2000020UL, PFM_MONT_RSVD, PFM_MONT_NO64, 9),
+/* pmc10 */ PMC_D(PFM_REG_W64, "PMC10", 0x2000020UL, PFM_MONT_RSVD, PFM_MONT_NO64, 10),
+/* pmc11 */ PMC_D(PFM_REG_W64, "PMC11", 0x2000020UL, PFM_MONT_RSVD, PFM_MONT_NO64, 11),
+/* pmc12 */ PMC_D(PFM_REG_W64, "PMC12", 0x2000020UL, PFM_MONT_RSVD, PFM_MONT_NO64, 12),
+/* pmc13 */ PMC_D(PFM_REG_W64, "PMC13", 0x2000020UL, PFM_MONT_RSVD, PFM_MONT_NO64, 13),
+/* pmc14 */ PMC_D(PFM_REG_W64, "PMC14", 0x2000020UL, PFM_MONT_RSVD, PFM_MONT_NO64, 14),
+/* pmc15 */ PMC_D(PFM_REG_W64, "PMC15", 0x2000020UL, PFM_MONT_RSVD, PFM_MONT_NO64, 15),
+/* pmc16 */ PMX_NA,
+/* pmc17 */ PMX_NA,
+/* pmc18 */ PMX_NA,
+/* pmc19 */ PMX_NA,
+/* pmc20 */ PMX_NA,
+/* pmc21 */ PMX_NA,
+/* pmc22 */ PMX_NA,
+/* pmc23 */ PMX_NA,
+/* pmc24 */ PMX_NA,
+/* pmc25 */ PMX_NA,
+/* pmc26 */ PMX_NA,
+/* pmc27 */ PMX_NA,
+/* pmc28 */ PMX_NA,
+/* pmc29 */ PMX_NA,
+/* pmc30 */ PMX_NA,
+/* pmc31 */ PMX_NA,
+/* pmc32 */ PMC_D(PFM_REG_W , "PMC32", 0x30f01ffffffffffUL, 0xfcf0fe0000000000UL, 0, 32),
+/* pmc33 */ PMC_D(PFM_REG_W , "PMC33", 0x0, 0xfffffe0000000000UL, 0, 33),
+/* pmc34 */ PMC_D(PFM_REG_W , "PMC34", 0xf01ffffffffffUL, 0xfff0fe0000000000UL, 0, 34),
+/* pmc35 */ PMC_D(PFM_REG_W , "PMC35", 0x0,  0x1ffffffffffUL, 0, 35),
+/* pmc36 */ PMC_D(PFM_REG_W , "PMC36", 0xfffffff0UL, 0xfffffffffffffff0UL, 0, 36),
+/* pmc37 */ PMC_D(PFM_REG_W , "PMC37", 0x0, 0xffffffffffffc000UL, 0, 37),
+/* pmc38 */ PMC_D(PFM_REG_W , "PMC38", 0xdb6UL, 0xffffffffffffdb6dUL, 0, 38),
+/* pmc39 */ PMC_D(PFM_REG_W , "PMC39", 0x0, 0xffffffffffff0030UL, 0, 39),
+/* pmc40 */ PMC_D(PFM_REG_W , "PMC40", 0x2000000UL, 0xfffffffffff0fe30UL, 0, 40),
+/* pmc41 */ PMC_D(PFM_REG_W , "PMC41", 0x00002078fefefefeUL, 0xfffe1fffe7e7e7e7UL, 0, 41),
+/* pmc42 */ PMC_D(PFM_REG_W , "PMC42", 0x0, 0xfff800b0UL, 0, 42),
+/* pmc43 */  PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA,
+/* pmc48 */  PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA,
+/* pmc56 */  PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA,
+/* pmc64 */  PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA,
+/* pmc72 */  PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA,
+/* pmc80 */  PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA,
+/* pmc88 */  PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA,
+/* pmc96 */  PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA,
+/* pmc104 */ PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA,
+/* pmc112 */ PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA,
+/* pmc120 */ PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA,
+/* pmc128 */ PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA,
+/* pmc136 */ PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA,
+/* pmc144 */ PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA,
+/* pmc152 */ PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA,
+/* pmc160 */ PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA,
+/* pmc168 */ PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA,
+/* pmc176 */ PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA,
+/* pmc184 */ PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA,
+/* pmc192 */ PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA,
+/* pmc200 */ PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA,
+/* pmc208 */ PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA,
+/* pmc216 */ PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA,
+/* pmc224 */ PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA,
+/* pmc232 */ PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA,
+/* pmc240 */ PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA,
+/* pmc248 */ PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA, PMX_NA,
+/* pmc256 */ PMC_D(PFM_REG_W, "IBR0", 0x0, 0, 0, 0),
+/* pmc257 */ PMC_D(PFM_REG_W, "IBR1", 0x0, 0x8000000000000000UL, 0, 1),
+/* pmc258 */ PMC_D(PFM_REG_W, "IBR2", 0x0, 0, 0, 2),
+/* pmc259 */ PMC_D(PFM_REG_W, "IBR3", 0x0, 0x8000000000000000UL, 0, 3),
+/* pmc260 */ PMC_D(PFM_REG_W, "IBR4", 0x0, 0, 0, 4),
+/* pmc261 */ PMC_D(PFM_REG_W, "IBR5", 0x0, 0x8000000000000000UL, 0, 5),
+/* pmc262 */ PMC_D(PFM_REG_W, "IBR6", 0x0, 0, 0, 6),
+/* pmc263 */ PMC_D(PFM_REG_W, "IBR7", 0x0, 0x8000000000000000UL, 0, 7),
+/* pmc264 */ PMC_D(PFM_REG_W, "DBR0", 0x0, 0, 0, 0),
+/* pmc265 */ PMC_D(PFM_REG_W, "DBR1", 0x0, 0xc000000000000000UL, 0, 1),
+/* pmc266 */ PMC_D(PFM_REG_W, "DBR2", 0x0, 0, 0, 2),
+/* pmc267 */ PMC_D(PFM_REG_W, "DBR3", 0x0, 0xc000000000000000UL, 0, 3),
+/* pmc268 */ PMC_D(PFM_REG_W, "DBR4", 0x0, 0, 0, 4),
+/* pmc269 */ PMC_D(PFM_REG_W, "DBR5", 0x0, 0xc000000000000000UL, 0, 5),
+/* pmc270 */ PMC_D(PFM_REG_W, "DBR6", 0x0, 0, 0, 6),
+/* pmc271 */ PMC_D(PFM_REG_W, "DBR7", 0x0, 0xc000000000000000UL, 0, 7)
+};
+#define PFM_MONT_NUM_PMCS ARRAY_SIZE(pfm_mont_pmc_desc)
+
+static struct pfm_regmap_desc pfm_mont_pmd_desc[] = {
+/* pmd0  */ PMX_NA,
+/* pmd1  */ PMX_NA,
+/* pmd2  */ PMX_NA,
+/* pmd3  */ PMX_NA,
+/* pmd4  */ PMD_DP(PFM_REG_C, "PMD4", 4, 1ull << 4),
+/* pmd5  */ PMD_DP(PFM_REG_C, "PMD5", 5, 1ull << 5),
+/* pmd6  */ PMD_DP(PFM_REG_C, "PMD6", 6, 1ull << 6),
+/* pmd7  */ PMD_DP(PFM_REG_C, "PMD7", 7, 1ull << 7),
+/* pmd8  */ PMD_DP(PFM_REG_C, "PMD8", 8, 1ull << 8),
+/* pmd9  */ PMD_DP(PFM_REG_C, "PMD9", 9, 1ull << 9),
+/* pmd10 */ PMD_DP(PFM_REG_C, "PMD10", 10, 1ull << 10),
+/* pmd11 */ PMD_DP(PFM_REG_C, "PMD11", 11, 1ull << 11),
+/* pmd12 */ PMD_DP(PFM_REG_C, "PMD12", 12, 1ull << 12),
+/* pmd13 */ PMD_DP(PFM_REG_C, "PMD13", 13, 1ull << 13),
+/* pmd14 */ PMD_DP(PFM_REG_C, "PMD14", 14, 1ull << 14),
+/* pmd15 */ PMD_DP(PFM_REG_C, "PMD15", 15, 1ull << 15),
+/* pmd16 */ PMX_NA,
+/* pmd17 */ PMX_NA,
+/* pmd18 */ PMX_NA,
+/* pmd19 */ PMX_NA,
+/* pmd20 */ PMX_NA,
+/* pmd21 */ PMX_NA,
+/* pmd22 */ PMX_NA,
+/* pmd23 */ PMX_NA,
+/* pmd24 */ PMX_NA,
+/* pmd25 */ PMX_NA,
+/* pmd26 */ PMX_NA,
+/* pmd27 */ PMX_NA,
+/* pmd28 */ PMX_NA,
+/* pmd29 */ PMX_NA,
+/* pmd30 */ PMX_NA,
+/* pmd31 */ PMX_NA,
+/* pmd32 */ PMD_DP(PFM_REG_I, "PMD32", 32, 1ull << 40),
+/* pmd33 */ PMD_DP(PFM_REG_I, "PMD33", 33, 1ull << 40),
+/* pmd34 */ PMD_DP(PFM_REG_I, "PMD34", 34, 1ull << 37),
+/* pmd35 */ PMD_DP(PFM_REG_I, "PMD35", 35, 1ull << 37),
+/* pmd36 */ PMD_DP(PFM_REG_I, "PMD36", 36, 1ull << 40),
+/* pmd37 */ PMX_NA,
+/* pmd38 */ PMD_DP(PFM_REG_I, "PMD38", 38, (1ull<<39)|(1ull<<42)),
+/* pmd39 */ PMD_DP(PFM_REG_I, "PMD39", 39, (1ull<<39)|(1ull<<42)),
+/* pmd40 */ PMX_NA,
+/* pmd41 */ PMX_NA,
+/* pmd42 */ PMX_NA,
+/* pmd43 */ PMX_NA,
+/* pmd44 */ PMX_NA,
+/* pmd45 */ PMX_NA,
+/* pmd46 */ PMX_NA,
+/* pmd47 */ PMX_NA,
+/* pmd48 */ PMD_DP(PFM_REG_I, "PMD48", 48, (1ull<<39)|(1ull<<42)),
+/* pmd49 */ PMD_DP(PFM_REG_I, "PMD49", 49, (1ull<<39)|(1ull<<42)),
+/* pmd50 */ PMD_DP(PFM_REG_I, "PMD50", 50, (1ull<<39)|(1ull<<42)),
+/* pmd51 */ PMD_DP(PFM_REG_I, "PMD51", 51, (1ull<<39)|(1ull<<42)),
+/* pmd52 */ PMD_DP(PFM_REG_I, "PMD52", 52, (1ull<<39)|(1ull<<42)),
+/* pmd53 */ PMD_DP(PFM_REG_I, "PMD53", 53, (1ull<<39)|(1ull<<42)),
+/* pmd54 */ PMD_DP(PFM_REG_I, "PMD54", 54, (1ull<<39)|(1ull<<42)),
+/* pmd55 */ PMD_DP(PFM_REG_I, "PMD55", 55, (1ull<<39)|(1ull<<42)),
+/* pmd56 */ PMD_DP(PFM_REG_I, "PMD56", 56, (1ull<<39)|(1ull<<42)),
+/* pmd57 */ PMD_DP(PFM_REG_I, "PMD57", 57, (1ull<<39)|(1ull<<42)),
+/* pmd58 */ PMD_DP(PFM_REG_I, "PMD58", 58, (1ull<<39)|(1ull<<42)),
+/* pmd59 */ PMD_DP(PFM_REG_I, "PMD59", 59, (1ull<<39)|(1ull<<42)),
+/* pmd60 */ PMD_DP(PFM_REG_I, "PMD60", 60, (1ull<<39)|(1ull<<42)),
+/* pmd61 */ PMD_DP(PFM_REG_I, "PMD61", 61, (1ull<<39)|(1ull<<42)),
+/* pmd62 */ PMD_DP(PFM_REG_I, "PMD62", 62, (1ull<<39)|(1ull<<42)),
+/* pmd63 */ PMD_DP(PFM_REG_I, "PMD63", 63, (1ull<<39)|(1ull<<42))
+};
+#define PFM_MONT_NUM_PMDS ARRAY_SIZE(pfm_mont_pmd_desc)
+
+static int pfm_mont_has_ht;
+
+static int pfm_mont_pmc_check(struct pfm_context *ctx,
+                             struct pfm_event_set *set,
+                             struct pfarg_pmc *req)
+{
+       struct pfm_arch_context *ctx_arch;
+       u64 val32 = 0, val38 = 0, val41 = 0;
+       u64 tmpval;
+       u16 cnum;
+       int ret = 0, check_case1 = 0;
+       int is_system;
+
+       tmpval = req->reg_value;
+       cnum = req->reg_num;
+       ctx_arch = pfm_ctx_arch(ctx);
+       is_system = ctx->flags.system;
+
+#define PFM_MONT_PMC_PM_POS6   (1UL<<6)
+#define PFM_MONT_PMC_PM_POS4   (1UL<<4)
+
+       switch (cnum) {
+       case  4:
+       case  5:
+       case  6:
+       case  7:
+       case  8:
+       case  9:
+               if (is_system)
+                       tmpval |= PFM_MONT_PMC_PM_POS6;
+               else
+                       tmpval &= ~PFM_MONT_PMC_PM_POS6;
+               break;
+       case 10:
+       case 11:
+       case 12:
+       case 13:
+       case 14:
+       case 15:
+               if ((req->reg_flags & PFM_REGFL_NO_EMUL64) == 0) {
+                       if (pfm_mont_has_ht) {
+                               PFM_INFO("perfmon: Errata 121 PMD10/PMD15 cannot be used to overflow"
+                                        "when threads on on");
+                               return -EINVAL;
+                       }
+               }
+               if (is_system)
+                       tmpval |= PFM_MONT_PMC_PM_POS6;
+               else
+                       tmpval &= ~PFM_MONT_PMC_PM_POS6;
+               break;
+       case 39:
+       case 40:
+       case 42:
+               if (pfm_mont_has_ht && ((req->reg_value >> 8) & 0x7) == 4) {
+                       PFM_INFO("perfmon: Errata 120: IP-EAR not available when threads are on");
+                       return -EINVAL;
+               }
+               if (is_system)
+                       tmpval |= PFM_MONT_PMC_PM_POS6;
+               else
+                       tmpval &= ~PFM_MONT_PMC_PM_POS6;
+               break;
+
+       case  32:
+               val32 = tmpval;
+               val38 = set->pmcs[38];
+               val41 = set->pmcs[41];
+               check_case1 = 1;
+               break;
+
+       case  37:
+               if (is_system)
+                       tmpval |= PFM_MONT_PMC_PM_POS4;
+               else
+                       tmpval &= ~PFM_MONT_PMC_PM_POS4;
+               break;
+
+       case  38:
+               val38 = tmpval;
+               val32 = set->pmcs[32];
+               val41 = set->pmcs[41];
+               check_case1 = 1;
+               break;
+       case  41:
+               val41 = tmpval;
+               val32 = set->pmcs[32];
+               val38 = set->pmcs[38];
+               check_case1 = 1;
+               break;
+       }
+
+       if (check_case1) {
+               ret = (((val41 >> 45) & 0xf) == 0 && ((val32>>57) & 0x1) == 0)
+                    && ((((val38>>1) & 0x3) == 0x2 || ((val38>>1) & 0x3) == 0)
+                    || (((val38>>4) & 0x3) == 0x2 || ((val38>>4) & 0x3) == 0));
+               if (ret) {
+                       PFM_DBG("perfmon: invalid config pmc38=0x%lx "
+                               "pmc41=0x%lx pmc32=0x%lx",
+                               val38, val41, val32);
+                       return -EINVAL;
+               }
+       }
+
+       /*
+        * check if configuration implicitely activates the use of the
+        * debug registers. If true, then we ensure that this is possible
+        * and that we do not pick up stale value in the HW registers.
+        */
+
+       /*
+        *
+        * pmc41 is "active" if:
+        *      one of the pmc41.cfgdtagXX field is different from 0x3
+        * AND
+        *      the corsesponding pmc41.en_dbrpXX is set.
+        * AND
+        *      ctx_fl_use_dbr (dbr not yet used)
+        */
+       if (cnum == 41
+           && (tmpval & 0x1e00000000000)
+               && (tmpval & 0x18181818) != 0x18181818
+               && ctx_arch->flags.use_dbr == 0) {
+               PFM_DBG("pmc41=0x%lx active, clearing dbr", tmpval);
+               ret = pfm_ia64_mark_dbregs_used(ctx, set);
+               if (ret)
+                       return ret;
+       }
+       /*
+        * we must clear the (instruction) debug registers if:
+        *      pmc38.ig_ibrpX is 0 (enabled)
+        * and
+        *      fl_use_dbr == 0 (dbr not yet used)
+        */
+       if (cnum == 38 && ((tmpval & 0x492) != 0x492)
+               && ctx_arch->flags.use_dbr == 0) {
+               PFM_DBG("pmc38=0x%lx active pmc38, clearing ibr", tmpval);
+               ret = pfm_ia64_mark_dbregs_used(ctx, set);
+               if (ret)
+                       return ret;
+
+       }
+       req->reg_value = tmpval;
+       return 0;
+}
+
+static void pfm_handle_errata(void)
+{
+       pfm_mont_has_ht = 1;
+
+       PFM_INFO("activating workaround for errata 120 "
+                "(Disable IP-EAR when threads are on)");
+
+       PFM_INFO("activating workaround for Errata 121 "
+                "(PMC10-PMC15 cannot be used to overflow"
+                " when threads are on");
+}
+static int pfm_mont_probe_pmu(void)
+{
+       if (local_cpu_data->family != 0x20)
+               return -1;
+
+       /*
+        * the 2 errata must be activated when
+        * threads are/can be enabled
+        */
+       if (is_multithreading_enabled())
+               pfm_handle_errata();
+
+       return 0;
+}
+
+/*
+ * impl_pmcs, impl_pmds are computed at runtime to minimize errors!
+ */
+static struct pfm_pmu_config pfm_mont_pmu_conf = {
+       .pmu_name = "Montecito",
+       .counter_width = 47,
+       .pmd_desc = pfm_mont_pmd_desc,
+       .pmc_desc = pfm_mont_pmc_desc,
+       .num_pmc_entries = PFM_MONT_NUM_PMCS,
+       .num_pmd_entries = PFM_MONT_NUM_PMDS,
+       .pmc_write_check = pfm_mont_pmc_check,
+       .probe_pmu = pfm_mont_probe_pmu,
+       .version = "1.0",
+       .pmu_info = &pfm_mont_pmu_info,
+       .flags = PFM_PMU_BUILTIN_FLAG,
+       .owner = THIS_MODULE
+};
+
+static int __init pfm_mont_pmu_init_module(void)
+{
+       return pfm_pmu_register(&pfm_mont_pmu_conf);
+}
+
+static void __exit pfm_mont_pmu_cleanup_module(void)
+{
+       pfm_pmu_unregister(&pfm_mont_pmu_conf);
+}
+
+module_init(pfm_mont_pmu_init_module);
+module_exit(pfm_mont_pmu_cleanup_module);
diff --git a/arch/mips/Kconfig b/arch/mips/Kconfig
index 1e06d23..b87f445 100644
--- a/arch/mips/Kconfig
+++ b/arch/mips/Kconfig
@@ -1857,6 +1857,8 @@ config SECCOMP
 
          If unsure, say Y. Only embedded should say N here.
 
+source "arch/mips/perfmon/Kconfig"
+
 endmenu
 
 config RWSEM_GENERIC_SPINLOCK
diff --git a/arch/mips/Makefile b/arch/mips/Makefile
index 9aab51c..712acf7 100644
--- a/arch/mips/Makefile
+++ b/arch/mips/Makefile
@@ -154,6 +154,12 @@ endif
 endif
 
 #
+# Perfmon support
+#
+
+core-$(CONFIG_PERFMON)         += arch/mips/perfmon/
+
+#
 # Firmware support
 #
 libs-$(CONFIG_ARC)             += arch/mips/fw/arc/
diff --git a/arch/mips/kernel/process.c b/arch/mips/kernel/process.c
index 22fc19b..4467361 100644
--- a/arch/mips/kernel/process.c
+++ b/arch/mips/kernel/process.c
@@ -27,6 +27,7 @@
 #include <linux/completion.h>
 #include <linux/kallsyms.h>
 #include <linux/random.h>
+#include <linux/perfmon_kern.h>
 
 #include <asm/asm.h>
 #include <asm/bootinfo.h>
@@ -94,6 +95,7 @@ void start_thread(struct pt_regs * regs, unsigned long pc, unsigned long sp)
 
 void exit_thread(void)
 {
+  pfm_exit_thread();
 }
 
 void flush_thread(void)
@@ -162,6 +164,8 @@ int copy_thread(int nr, unsigned long clone_flags, unsigned long usp,
        if (clone_flags & CLONE_SETTLS)
                ti->tp_value = regs->regs[7];
 
+       pfm_copy_thread(p);
+
        return 0;
 }
 
diff --git a/arch/mips/kernel/scall32-o32.S b/arch/mips/kernel/scall32-o32.S
index 5e75a31..e96ddd6 100644
--- a/arch/mips/kernel/scall32-o32.S
+++ b/arch/mips/kernel/scall32-o32.S
@@ -653,6 +653,18 @@ einval:    li      v0, -EINVAL
        sys     sys_dup3                3
        sys     sys_pipe2               2
        sys     sys_inotify_init1       1
+       sys     sys_pfm_create_context  4       /* 4330 */
+       sys     sys_pfm_write_pmcs      3
+       sys     sys_pfm_write_pmds      4
+       sys     sys_pfm_read_pmds       3
+       sys     sys_pfm_load_context    2
+       sys     sys_pfm_start           2       /* 4335 */
+       sys     sys_pfm_stop            1
+       sys     sys_pfm_restart         1
+       sys     sys_pfm_create_evtsets  3
+       sys     sys_pfm_getinfo_evtsets 3
+       sys     sys_pfm_delete_evtsets  3       /* 4340 */
+       sys     sys_pfm_unload_context  1
        .endm
 
        /* We pre-compute the number of _instruction_ bytes needed to
diff --git a/arch/mips/kernel/scall64-64.S b/arch/mips/kernel/scall64-64.S
index 3d58204..adb2ba9 100644
--- a/arch/mips/kernel/scall64-64.S
+++ b/arch/mips/kernel/scall64-64.S
@@ -487,4 +487,16 @@ sys_call_table:
        PTR     sys_dup3
        PTR     sys_pipe2
        PTR     sys_inotify_init1
+       PTR     sys_pfm_create_context
+       PTR     sys_pfm_write_pmcs              /* 5290 */
+       PTR     sys_pfm_write_pmds
+       PTR     sys_pfm_read_pmds
+       PTR     sys_pfm_load_context
+       PTR     sys_pfm_start
+       PTR     sys_pfm_stop                    /* 5295 */
+       PTR     sys_pfm_restart
+       PTR     sys_pfm_create_evtsets
+       PTR     sys_pfm_getinfo_evtsets
+       PTR     sys_pfm_delete_evtsets
+       PTR     sys_pfm_unload_context          /* 5300 */
        .size   sys_call_table,.-sys_call_table
diff --git a/arch/mips/kernel/scall64-n32.S b/arch/mips/kernel/scall64-n32.S
index da7f1b6..6d12095 100644
--- a/arch/mips/kernel/scall64-n32.S
+++ b/arch/mips/kernel/scall64-n32.S
@@ -400,12 +400,12 @@ EXPORT(sysn32_call_table)
        PTR     sys_ioprio_set
        PTR     sys_ioprio_get
        PTR     compat_sys_utimensat
-       PTR     compat_sys_signalfd             /* 5280 */
+       PTR     compat_sys_signalfd             /* 6280 */
        PTR     sys_ni_syscall
        PTR     sys_eventfd
        PTR     sys_fallocate
        PTR     sys_timerfd_create
-       PTR     sys_timerfd_gettime             /* 5285 */
+       PTR     sys_timerfd_gettime             /* 6285 */
        PTR     sys_timerfd_settime
        PTR     sys_signalfd4
        PTR     sys_eventfd2
@@ -413,4 +413,16 @@ EXPORT(sysn32_call_table)
        PTR     sys_dup3                        /* 5290 */
        PTR     sys_pipe2
        PTR     sys_inotify_init1
+       PTR     sys_pfm_create_context
+       PTR     sys_pfm_write_pmcs
+       PTR     sys_pfm_write_pmds              /* 6295 */
+       PTR     sys_pfm_read_pmds
+       PTR     sys_pfm_load_context
+       PTR     sys_pfm_start
+       PTR     sys_pfm_stop
+       PTR     sys_pfm_restart                 /* 6300 */
+       PTR     sys_pfm_create_evtsets
+       PTR     sys_pfm_getinfo_evtsets
+       PTR     sys_pfm_delete_evtsets
+       PTR     sys_pfm_unload_context
        .size   sysn32_call_table,.-sysn32_call_table
diff --git a/arch/mips/kernel/scall64-o32.S b/arch/mips/kernel/scall64-o32.S
index d7cd1aa..e77f55a 100644
--- a/arch/mips/kernel/scall64-o32.S
+++ b/arch/mips/kernel/scall64-o32.S
@@ -535,4 +535,16 @@ sys_call_table:
        PTR     sys_dup3
        PTR     sys_pipe2
        PTR     sys_inotify_init1
+       PTR     sys_pfm_create_context          /* 4330 */
+       PTR     sys_pfm_write_pmcs
+       PTR     sys_pfm_write_pmds
+       PTR     sys_pfm_read_pmds
+       PTR     sys_pfm_load_context
+       PTR     sys_pfm_start                   /* 4335 */
+       PTR     sys_pfm_stop
+       PTR     sys_pfm_restart
+       PTR     sys_pfm_create_evtsets
+       PTR     sys_pfm_getinfo_evtsets
+       PTR     sys_pfm_delete_evtsets          /* 4340 */
+       PTR     sys_pfm_unload_context
        .size   sys_call_table,.-sys_call_table
diff --git a/arch/mips/kernel/signal.c b/arch/mips/kernel/signal.c
index a4e106c..6a7e60c 100644
--- a/arch/mips/kernel/signal.c
+++ b/arch/mips/kernel/signal.c
@@ -20,6 +20,7 @@
 #include <linux/unistd.h>
 #include <linux/compiler.h>
 #include <linux/uaccess.h>
+#include <linux/perfmon_kern.h>
 
 #include <asm/abi.h>
 #include <asm/asm.h>
@@ -694,8 +695,11 @@ static void do_signal(struct pt_regs *regs)
  * - triggered by the TIF_WORK_MASK flags
  */
 asmlinkage void do_notify_resume(struct pt_regs *regs, void *unused,
-       __u32 thread_info_flags)
+                                __u32 thread_info_flags)
 {
+       if (thread_info_flags & _TIF_PERFMON_WORK)
+               pfm_handle_work(regs);
+
        /* deal with pending signal delivery */
        if (thread_info_flags & (_TIF_SIGPENDING | _TIF_RESTORE_SIGMASK))
                do_signal(regs);
diff --git a/arch/mips/kernel/time.c b/arch/mips/kernel/time.c
index 1f467d5..163dfe4 100644
--- a/arch/mips/kernel/time.c
+++ b/arch/mips/kernel/time.c
@@ -49,10 +49,11 @@ int update_persistent_clock(struct timespec now)
        return rtc_mips_set_mmss(now.tv_sec);
 }
 
-static int null_perf_irq(void)
+int null_perf_irq(void)
 {
        return 0;
 }
+EXPORT_SYMBOL(null_perf_irq);
 
 int (*perf_irq)(void) = null_perf_irq;
 
diff --git a/arch/mips/kernel/traps.c b/arch/mips/kernel/traps.c
index b602ac6..9cbd75f 100644
--- a/arch/mips/kernel/traps.c
+++ b/arch/mips/kernel/traps.c
@@ -92,17 +92,15 @@ static void show_raw_backtrace(unsigned long reg29)
 #ifdef CONFIG_KALLSYMS
        printk("\n");
 #endif
-       while (!kstack_end(sp)) {
-               unsigned long __user *p =
-                       (unsigned long __user *)(unsigned long)sp++;
-               if (__get_user(addr, p)) {
-                       printk(" (Bad stack address)");
-                       break;
+#define IS_KVA01(a) ((((unsigned long)a) & 0xc0000000) == 0x80000000)
+       if (IS_KVA01(sp)) {
+               while (!kstack_end(sp)) {
+                       addr = *sp++;
+                       if (__kernel_text_address(addr))
+                               print_ip_sym(addr);
                }
-               if (__kernel_text_address(addr))
-                       print_ip_sym(addr);
+               printk("\n");
        }
-       printk("\n");
 }
 
 #ifdef CONFIG_KALLSYMS
diff --git a/arch/mips/mti-malta/malta-time.c b/arch/mips/mti-malta/malta-time.c
index 0b97d47..d8f36b5 100644
--- a/arch/mips/mti-malta/malta-time.c
+++ b/arch/mips/mti-malta/malta-time.c
@@ -27,6 +27,7 @@
 #include <linux/time.h>
 #include <linux/timex.h>
 #include <linux/mc146818rtc.h>
+#include <linux/perfmon_kern.h>
 
 #include <asm/mipsregs.h>
 #include <asm/mipsmtregs.h>
diff --git a/arch/mips/perfmon/Kconfig b/arch/mips/perfmon/Kconfig
new file mode 100644
index 0000000..b426eea
--- /dev/null
+++ b/arch/mips/perfmon/Kconfig
@@ -0,0 +1,61 @@
+menu "Hardware Performance Monitoring support"
+config PERFMON
+       bool "Perfmon2 performance monitoring interface"
+       default n
+       help
+       Enables the perfmon2 interface to access the hardware
+       performance counters. See <http://perfmon2.sf.net/> for
+       more details.
+
+config PERFMON_DEBUG
+       bool "Perfmon debugging"
+       default n
+       depends on PERFMON
+       help
+       Enables perfmon debugging support
+
+config PERFMON_DEBUG_FS
+       bool "Enable perfmon statistics reporting via debugfs"
+       default y
+       depends on PERFMON && DEBUG_FS
+       help
+       Enable collection and reporting of perfmon timing statistics under
+       debugfs. This is used for debugging and performance analysis of the
+       subsystem. The debugfs filesystem must be mounted.
+
+config PERFMON_FLUSH
+       bool "Flush sampling buffer when modified"
+       depends on PERFMON
+       default n
+       help
+       On some MIPS models, cache aliasing may cause invalid
+       data to be read from the perfmon sampling buffer. Use this option
+       to flush the buffer when it is modified to ensure valid data is
+       visible at the user level.
+
+config PERFMON_ALIGN
+       bool "Align sampling buffer to avoid cache aliasing"
+       depends on PERFMON
+       default n
+       help
+       On some MIPS models, cache aliasing may cause invalid
+       data to be read from the perfmon sampling buffer. By forcing a bigger
+       page alignment (4-page), one can guarantee the buffer virtual address
+       will conflict in the cache with the user level mapping of the buffer
+       thereby ensuring a consistent view by user programs.
+
+config PERFMON_DEBUG
+       bool "Perfmon debugging"
+       depends on PERFMON
+       default n
+       depends on PERFMON
+       help
+       Enables perfmon debugging support
+
+config PERFMON_MIPS64
+       tristate "Support for MIPS64 hardware performance counters"
+       depends on PERFMON
+       default n
+       help
+       Enables support for the MIPS64 hardware performance counters"
+endmenu
diff --git a/arch/mips/perfmon/Makefile b/arch/mips/perfmon/Makefile
new file mode 100644
index 0000000..153b83f
--- /dev/null
+++ b/arch/mips/perfmon/Makefile
@@ -0,0 +1,2 @@
+obj-$(CONFIG_PERFMON)          += perfmon.o
+obj-$(CONFIG_PERFMON_MIPS64)   += perfmon_mips64.o
diff --git a/arch/mips/perfmon/perfmon.c b/arch/mips/perfmon/perfmon.c
new file mode 100644
index 0000000..6615a77
--- /dev/null
+++ b/arch/mips/perfmon/perfmon.c
@@ -0,0 +1,313 @@
+/*
+ * This file implements the MIPS64 specific
+ * support for the perfmon2 interface
+ *
+ * Copyright (c) 2005 Philip J. Mucci
+ *
+ * based on versions for other architectures:
+ * Copyright (c) 2005 Hewlett-Packard Development Company, L.P.
+ * Contributed by Stephane Eranian <eranian@htrpl.hp.com>
+ *
+ * This program is free software; you can redistribute it and/or
+ * modify it under the terms of version 2 of the GNU General Public
+ * License as published by the Free Software Foundation.
+ *
+ * This program is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+ * General Public License for more details.
+ *
+ * You should have received a copy of the GNU General Public License
+ * along with this program; if not, write to the Free Software
+ * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA
+ * 02111-1307 USA
+  */
+#include <linux/interrupt.h>
+#include <linux/module.h>
+#include <linux/perfmon_kern.h>
+
+/*
+ * collect pending overflowed PMDs. Called from pfm_ctxsw()
+ * and from PMU interrupt handler. Must fill in set->povfl_pmds[]
+ * and set->npend_ovfls. Interrupts are masked
+ */
+static void __pfm_get_ovfl_pmds(struct pfm_context *ctx, struct pfm_event_set *set)
+{
+       u64 new_val, wmask;
+       u64 *used_mask, *intr_pmds;
+       u64 mask[PFM_PMD_BV];
+       unsigned int i, max;
+
+       max = ctx->regs.max_intr_pmd;
+       intr_pmds = ctx->regs.intr_pmds;
+       used_mask = set->used_pmds;
+
+       wmask = 1ULL << pfm_pmu_conf->counter_width;
+
+       bitmap_and(cast_ulp(mask),
+                  cast_ulp(intr_pmds),
+                  cast_ulp(used_mask),
+                  max);
+
+       /*
+        * check all PMD that can generate interrupts
+        * (that includes counters)
+        */
+       for (i = 0; i < max; i++) {
+               if (test_bit(i, mask)) {
+                       new_val = pfm_arch_read_pmd(ctx, i);
+
+                       PFM_DBG_ovfl("pmd%u new_val=0x%llx bit=%d\n",
+                                    i, (unsigned long long)new_val,
+                                    (new_val&wmask) ? 1 : 0);
+
+                       if (new_val & wmask) {
+                               __set_bit(i, set->povfl_pmds);
+                               set->npend_ovfls++;
+                       }
+               }
+       }
+}
+
+static void pfm_stop_active(struct task_struct *task, struct pfm_context *ctx,
+                           struct pfm_event_set *set)
+{
+       unsigned int i, max;
+
+       max = ctx->regs.max_pmc;
+
+       /*
+        * clear enable bits, assume all pmcs are enable pmcs
+        */
+       for (i = 0; i < max; i++) {
+               if (test_bit(i, set->used_pmcs))
+                       pfm_arch_write_pmc(ctx, i, 0);
+       }
+
+       if (set->npend_ovfls)
+               return;
+
+       __pfm_get_ovfl_pmds(ctx, set);
+}
+
+/*
+ * Called from pfm_ctxsw(). Task is guaranteed to be current.
+ * Context is locked. Interrupts are masked. Monitoring is active.
+ * PMU access is guaranteed. PMC and PMD registers are live in PMU.
+ *
+ * for per-thread:
+ *     must stop monitoring for the task
+ *
+ * Return:
+ *     non-zero : did not save PMDs (as part of stopping the PMU)
+ *            0 : saved PMDs (no need to save them in caller)
+ */
+int pfm_arch_ctxswout_thread(struct task_struct *task, struct pfm_context *ctx)
+{
+       /*
+        * disable lazy restore of PMC registers.
+        */
+       ctx->active_set->priv_flags |= PFM_SETFL_PRIV_MOD_PMCS;
+
+       /*
+        * if masked, monitoring is stopped, thus there is no
+        * need to stop the PMU again and there is no need to
+        * check for pending overflows. This is not just an
+        * optimization, this is also for correctness as you
+        * may end up detecting overflows twice.
+        */
+       if (ctx->state == PFM_CTX_MASKED)
+               return 1;
+
+       pfm_stop_active(task, ctx, ctx->active_set);
+
+       return 1;
+}
+
+/*
+ * Called from pfm_stop() and pfm_ctxsw()
+ * Interrupts are masked. Context is locked. Set is the active set.
+ *
+ * For per-thread:
+ *   task is not necessarily current. If not current task, then
+ *   task is guaranteed stopped and off any cpu. Access to PMU
+ *   is not guaranteed. Interrupts are masked. Context is locked.
+ *   Set is the active set.
+ *
+ * For system-wide:
+ *     task is current
+ *
+ * must disable active monitoring. ctx cannot be NULL
+ */
+void pfm_arch_stop(struct task_struct *task, struct pfm_context *ctx)
+{
+       /*
+        * no need to go through stop_save()
+        * if we are already stopped
+        */
+       if (!ctx->flags.started || ctx->state == PFM_CTX_MASKED)
+               return;
+
+       /*
+        * stop live registers and collect pending overflow
+        */
+       if (task == current)
+               pfm_stop_active(task, ctx, ctx->active_set);
+}
+
+/*
+ * called from pfm_start() or pfm_ctxsw() when idle task and
+ * EXCL_IDLE is on.
+ *
+ * Interrupts are masked. Context is locked. Set is the active set.
+ *
+ * For per-trhead:
+ *     Task is not necessarily current. If not current task, then task
+ *     is guaranteed stopped and off any cpu. Access to PMU is not guaranteed.
+ *
+ * For system-wide:
+ *     task is always current
+ *
+ * must enable active monitoring.
+ */
+void pfm_arch_start(struct task_struct *task, struct pfm_context *ctx)
+{
+       struct pfm_event_set *set;
+       unsigned int i, max_pmc;
+
+       if (task != current)
+               return;
+
+       set = ctx->active_set;
+       max_pmc = ctx->regs.max_pmc;
+
+       for (i = 0; i < max_pmc; i++) {
+               if (test_bit(i, set->used_pmcs))
+                       pfm_arch_write_pmc(ctx, i, set->pmcs[i]);
+       }
+}
+
+/*
+ * function called from pfm_switch_sets(), pfm_context_load_thread(),
+ * pfm_context_load_sys(), pfm_ctxsw(), pfm_switch_sets()
+ * context is locked. Interrupts are masked. set cannot be NULL.
+ * Access to the PMU is guaranteed.
+ *
+ * function must restore all PMD registers from set.
+ */
+void pfm_arch_restore_pmds(struct pfm_context *ctx, struct pfm_event_set *set)
+{
+       u64 ovfl_mask, val;
+       u64 *impl_pmds;
+       unsigned int i;
+       unsigned int max_pmd;
+
+       max_pmd = ctx->regs.max_pmd;
+       ovfl_mask = pfm_pmu_conf->ovfl_mask;
+       impl_pmds = ctx->regs.pmds;
+
+       /*
+        * must restore all pmds to avoid leaking
+        * information to user.
+        */
+       for (i = 0; i < max_pmd; i++) {
+
+               if (test_bit(i, impl_pmds) == 0)
+                       continue;
+
+               val = set->pmds[i].value;
+
+               /*
+                * set upper bits for counter to ensure
+                * overflow will trigger
+                */
+               val &= ovfl_mask;
+
+               pfm_arch_write_pmd(ctx, i, val);
+       }
+}
+
+/*
+ * function called from pfm_switch_sets(), pfm_context_load_thread(),
+ * pfm_context_load_sys(), pfm_ctxsw().
+ * Context is locked. Interrupts are masked. set cannot be NULL.
+ * Access to the PMU is guaranteed.
+ *
+ * function must restore all PMC registers from set, if needed.
+ */
+void pfm_arch_restore_pmcs(struct pfm_context *ctx, struct pfm_event_set *set)
+{
+       u64 *impl_pmcs;
+       unsigned int i, max_pmc;
+
+       max_pmc = ctx->regs.max_pmc;
+       impl_pmcs = ctx->regs.pmcs;
+
+       /*
+        * - by default no PMCS measures anything
+        * - on ctxswout, all used PMCs are disabled (cccr enable bit cleared)
+        * hence when masked we do not need to restore anything
+        */
+       if (ctx->state == PFM_CTX_MASKED || ctx->flags.started == 0)
+               return;
+
+       /*
+        * restore all pmcs
+        */
+       for (i = 0; i < max_pmc; i++)
+               if (test_bit(i, impl_pmcs))
+                       pfm_arch_write_pmc(ctx, i, set->pmcs[i]);
+}
+
+char *pfm_arch_get_pmu_module_name(void)
+{
+       switch (cpu_data->cputype) {
+#ifndef CONFIG_SMP
+       case CPU_34K:
+#if defined(CPU_74K)
+       case CPU_74K:
+#endif
+#endif
+       case CPU_SB1:
+       case CPU_SB1A:
+       case CPU_R12000:
+       case CPU_25KF:
+       case CPU_24K:
+       case CPU_20KC:
+       case CPU_5KC:
+               return "perfmon_mips64";
+       default:
+               return NULL;
+       }
+       return NULL;
+}
+
+int perfmon_perf_irq(void)
+{
+       /* BLATANTLY STOLEN FROM OPROFILE, then modified */
+       struct pt_regs *regs;
+       unsigned int counters = pfm_pmu_conf->regs_all.max_pmc;
+       unsigned int control;
+       unsigned int counter;
+
+       regs = get_irq_regs();
+       switch (counters) {
+#define HANDLE_COUNTER(n)                                              \
+       case n + 1:                                                     \
+               control = read_c0_perfctrl ## n();                      \
+               counter = read_c0_perfcntr ## n();                      \
+               if ((control & MIPS64_PMC_INT_ENABLE_MASK) &&           \
+                   (counter & MIPS64_PMD_INTERRUPT)) {                 \
+                       pfm_interrupt_handler(instruction_pointer(regs),\
+                                             regs);                    \
+                       return(1);                                      \
+               }
+               HANDLE_COUNTER(3)
+               HANDLE_COUNTER(2)
+               HANDLE_COUNTER(1)
+               HANDLE_COUNTER(0)
+       }
+
+       return 0;
+}
+EXPORT_SYMBOL(perfmon_perf_irq);
diff --git a/arch/mips/perfmon/perfmon_mips64.c b/arch/mips/perfmon/perfmon_mips64.c
new file mode 100644
index 0000000..78cb43d
--- /dev/null
+++ b/arch/mips/perfmon/perfmon_mips64.c
@@ -0,0 +1,218 @@
+/*
+ * This file contains the MIPS64 and decendent PMU register description tables
+ * and pmc checker used by perfmon.c.
+ *
+ * Copyright (c) 2005 Philip Mucci
+ *
+ * Based on perfmon_p6.c:
+ * Copyright (c) 2005-2006 Hewlett-Packard Development Company, L.P.
+ * Contributed by Stephane Eranian <eranian@hpl.hp.com>
+ *
+ * This program is free software; you can redistribute it and/or
+ * modify it under the terms of version 2 of the GNU General Public
+ * License as published by the Free Software Foundation.
+ *
+ * This program is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+ * General Public License for more details.
+ *
+ * You should have received a copy of the GNU General Public License
+ * along with this program; if not, write to the Free Software
+ * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA
+ * 02111-1307 USA
+ */
+#include <linux/module.h>
+#include <linux/perfmon_kern.h>
+
+MODULE_AUTHOR("Philip Mucci <mucci@cs.utk.edu>");
+MODULE_DESCRIPTION("MIPS64 PMU description tables");
+MODULE_LICENSE("GPL");
+
+/*
+ * reserved:
+ *     - bit 63-9
+ * RSVD: reserved bits must be 1
+ */
+#define PFM_MIPS64_PMC_RSVD 0xfffffffffffff810ULL
+#define PFM_MIPS64_PMC_VAL  (1ULL<<4)
+
+extern int null_perf_irq(struct pt_regs *regs);
+extern int (*perf_irq)(struct pt_regs *regs);
+extern int perfmon_perf_irq(struct pt_regs *regs);
+
+static struct pfm_arch_pmu_info pfm_mips64_pmu_info;
+
+static struct pfm_regmap_desc pfm_mips64_pmc_desc[] = {
+/* pmc0 */  PMC_D(PFM_REG_I64, "CP0_25_0", PFM_MIPS64_PMC_VAL, PFM_MIPS64_PMC_RSVD, 0, 0),
+/* pmc1 */  PMC_D(PFM_REG_I64, "CP0_25_1", PFM_MIPS64_PMC_VAL, PFM_MIPS64_PMC_RSVD, 0, 1),
+/* pmc2 */  PMC_D(PFM_REG_I64, "CP0_25_2", PFM_MIPS64_PMC_VAL, PFM_MIPS64_PMC_RSVD, 0, 2),
+/* pmc3 */  PMC_D(PFM_REG_I64, "CP0_25_3", PFM_MIPS64_PMC_VAL, PFM_MIPS64_PMC_RSVD, 0, 3)
+};
+#define PFM_MIPS64_NUM_PMCS ARRAY_SIZE(pfm_mips64_pmc_desc)
+
+static struct pfm_regmap_desc pfm_mips64_pmd_desc[] = {
+/* pmd0 */ PMD_D(PFM_REG_C, "CP0_25_0", 0),
+/* pmd1 */ PMD_D(PFM_REG_C, "CP0_25_1", 1),
+/* pmd2 */ PMD_D(PFM_REG_C, "CP0_25_2", 2),
+/* pmd3 */ PMD_D(PFM_REG_C, "CP0_25_3", 3)
+};
+#define PFM_MIPS64_NUM_PMDS ARRAY_SIZE(pfm_mips64_pmd_desc)
+
+static int pfm_mips64_probe_pmu(void)
+{
+       struct cpuinfo_mips *c = &current_cpu_data;
+
+       switch (c->cputype) {
+#ifndef CONFIG_SMP
+       case CPU_34K:
+#if defined(CPU_74K)
+       case CPU_74K:
+#endif
+#endif
+       case CPU_SB1:
+       case CPU_SB1A:
+       case CPU_R12000:
+       case CPU_25KF:
+       case CPU_24K:
+       case CPU_20KC:
+       case CPU_5KC:
+               return 0;
+               break;
+       default:
+               PFM_INFO("Unknown cputype 0x%x", c->cputype);
+       }
+       return -1;
+}
+
+/*
+ * impl_pmcs, impl_pmds are computed at runtime to minimize errors!
+ */
+static struct pfm_pmu_config pfm_mips64_pmu_conf = {
+       .pmu_name = "MIPS", /* placeholder */
+       .counter_width = 31,
+       .pmd_desc = pfm_mips64_pmd_desc,
+       .pmc_desc = pfm_mips64_pmc_desc,
+       .num_pmc_entries = PFM_MIPS64_NUM_PMCS,
+       .num_pmd_entries = PFM_MIPS64_NUM_PMDS,
+       .probe_pmu = pfm_mips64_probe_pmu,
+       .flags = PFM_PMU_BUILTIN_FLAG,
+       .owner = THIS_MODULE,
+       .pmu_info = &pfm_mips64_pmu_info
+};
+
+static inline int n_counters(void)
+{
+       if (!(read_c0_config1() & MIPS64_CONFIG_PMC_MASK))
+               return 0;
+       if (!(read_c0_perfctrl0() & MIPS64_PMC_CTR_MASK))
+               return 1;
+       if (!(read_c0_perfctrl1() & MIPS64_PMC_CTR_MASK))
+               return 2;
+       if (!(read_c0_perfctrl2() & MIPS64_PMC_CTR_MASK))
+               return 3;
+       return 4;
+}
+
+static int __init pfm_mips64_pmu_init_module(void)
+{
+       struct cpuinfo_mips *c = &current_cpu_data;
+       int i, ret, num;
+       u64 temp_mask;
+
+       switch (c->cputype) {
+       case CPU_5KC:
+               pfm_mips64_pmu_conf.pmu_name = "MIPS5KC";
+               break;
+       case CPU_R12000:
+               pfm_mips64_pmu_conf.pmu_name = "MIPSR12000";
+               break;
+       case CPU_20KC:
+               pfm_mips64_pmu_conf.pmu_name = "MIPS20KC";
+               break;
+       case CPU_24K:
+               pfm_mips64_pmu_conf.pmu_name = "MIPS24K";
+               break;
+       case CPU_25KF:
+               pfm_mips64_pmu_conf.pmu_name = "MIPS25KF";
+               break;
+       case CPU_SB1:
+               pfm_mips64_pmu_conf.pmu_name = "SB1";
+               break;
+       case CPU_SB1A:
+       pfm_mips64_pmu_conf.pmu_name = "SB1A";
+               break;
+#ifndef CONFIG_SMP
+       case CPU_34K:
+               pfm_mips64_pmu_conf.pmu_name = "MIPS34K";
+               break;
+#if defined(CPU_74K)
+       case CPU_74K:
+               pfm_mips64_pmu_conf.pmu_name = "MIPS74K";
+               break;
+#endif
+#endif
+       default:
+               PFM_INFO("Unknown cputype 0x%x", c->cputype);
+               return -1;
+       }
+
+       /* The R14k and older performance counters have to          */
+       /* be hard-coded, as there is no support for auto-detection */
+       if ((c->cputype == CPU_R12000) || (c->cputype == CPU_R14000))
+               num = 4;
+       else if (c->cputype == CPU_R10000)
+               num = 2;
+       else
+               num = n_counters();
+
+       if (num == 0) {
+               PFM_INFO("cputype 0x%x has no counters", c->cputype);
+               return -1;
+       }
+       /* mark remaining counters unavailable */
+       for (i = num; i < PFM_MIPS64_NUM_PMCS; i++)
+               pfm_mips64_pmc_desc[i].type = PFM_REG_NA;
+
+       for (i = num; i < PFM_MIPS64_NUM_PMDS; i++)
+               pfm_mips64_pmd_desc[i].type = PFM_REG_NA;
+
+       /* set the PMC_RSVD mask */
+       switch (c->cputype) {
+       case CPU_5KC:
+       case CPU_R10000:
+       case CPU_20KC:
+          /* 4-bits for event */
+          temp_mask = 0xfffffffffffffe10ULL;
+          break;
+       case CPU_R12000:
+       case CPU_R14000:
+          /* 5-bits for event */
+          temp_mask = 0xfffffffffffffc10ULL;
+          break;
+       default:
+          /* 6-bits for event */
+          temp_mask = 0xfffffffffffff810ULL;
+       }
+       for (i = 0; i < PFM_MIPS64_NUM_PMCS; i++)
+               pfm_mips64_pmc_desc[i].rsvd_msk = temp_mask;
+
+       pfm_mips64_pmu_conf.num_pmc_entries = num;
+       pfm_mips64_pmu_conf.num_pmd_entries = num;
+
+       pfm_mips64_pmu_info.pmu_style = c->cputype;
+
+       ret = pfm_pmu_register(&pfm_mips64_pmu_conf);
+       if (ret == 0)
+               perf_irq = perfmon_perf_irq;
+       return ret;
+}
+
+static void __exit pfm_mips64_pmu_cleanup_module(void)
+{
+       pfm_pmu_unregister(&pfm_mips64_pmu_conf);
+       perf_irq = null_perf_irq;
+}
+
+module_init(pfm_mips64_pmu_init_module);
+module_exit(pfm_mips64_pmu_cleanup_module);
diff --git a/arch/powerpc/Kconfig b/arch/powerpc/Kconfig
index 587da5e..a411389 100644
--- a/arch/powerpc/Kconfig
+++ b/arch/powerpc/Kconfig
@@ -230,6 +230,8 @@ source "init/Kconfig"
 source "arch/powerpc/sysdev/Kconfig"
 source "arch/powerpc/platforms/Kconfig"
 
+source "arch/powerpc/perfmon/Kconfig"
+
 menu "Kernel options"
 
 config HIGHMEM
diff --git a/arch/powerpc/Makefile b/arch/powerpc/Makefile
index c6be19e..7ea20cb 100644
--- a/arch/powerpc/Makefile
+++ b/arch/powerpc/Makefile
@@ -146,6 +146,7 @@ core-y                              += arch/powerpc/kernel/ \
                                   arch/powerpc/platforms/
 core-$(CONFIG_MATH_EMULATION)  += arch/powerpc/math-emu/
 core-$(CONFIG_XMON)            += arch/powerpc/xmon/
+core-$(CONFIG_PERFMON)         += arch/powerpc/perfmon/
 core-$(CONFIG_KVM)             += arch/powerpc/kvm/
 
 drivers-$(CONFIG_OPROFILE)     += arch/powerpc/oprofile/
diff --git a/arch/powerpc/include/asm/Kbuild b/arch/powerpc/include/asm/Kbuild
index 5ab7d7f..88cb533 100644
--- a/arch/powerpc/include/asm/Kbuild
+++ b/arch/powerpc/include/asm/Kbuild
@@ -21,6 +21,7 @@ header-y += resource.h
 header-y += sigcontext.h
 header-y += statfs.h
 header-y += ps3fb.h
+header-y += perfmon.h
 
 unifdef-y += bootx.h
 unifdef-y += byteorder.h
diff --git a/arch/powerpc/include/asm/cell-pmu.h b/arch/powerpc/include/asm/cell-pmu.h
index 8066eed..981db26 100644
--- a/arch/powerpc/include/asm/cell-pmu.h
+++ b/arch/powerpc/include/asm/cell-pmu.h
@@ -61,6 +61,11 @@
 
 /* Macros for the pm_status register. */
 #define CBE_PM_CTR_OVERFLOW_INTR(ctr)      (1 << (31 - ((ctr) & 7)))
+#define CBE_PM_OVERFLOW_CTRS(pm_status)    (((pm_status) >> 24) & 0xff)
+#define CBE_PM_ALL_OVERFLOW_INTR           0xff000000
+#define CBE_PM_INTERVAL_INTR               0x00800000
+#define CBE_PM_TRACE_BUFFER_FULL_INTR      0x00400000
+#define CBE_PM_TRACE_BUFFER_UNDERFLOW_INTR 0x00200000
 
 enum pm_reg_name {
        group_control,
diff --git a/arch/powerpc/include/asm/cell-regs.h b/arch/powerpc/include/asm/cell-regs.h
index fd6fd00..580786d 100644
--- a/arch/powerpc/include/asm/cell-regs.h
+++ b/arch/powerpc/include/asm/cell-regs.h
@@ -117,8 +117,9 @@ struct cbe_pmd_regs {
        u8      pad_0x0c1c_0x0c20 [4];                          /* 0x0c1c */
 #define CBE_PMD_FIR_MODE_M8            0x00800
        u64     fir_enable_mask;                                /* 0x0c20 */
-
-       u8      pad_0x0c28_0x0ca8 [0x0ca8 - 0x0c28];            /* 0x0c28 */
+       u8      pad_0x0c28_0x0c98 [0x0c98 - 0x0c28];            /* 0x0c28 */
+       u64     on_ramp_trace;                                  /* 0x0c98 */
+       u64     pad_0x0ca0;                                     /* 0x0ca0 */
        u64     ras_esc_0;                                      /* 0x0ca8 */
        u8      pad_0x0cb0_0x1000 [0x1000 - 0x0cb0];            /* 0x0cb0 */
 };
@@ -218,7 +219,11 @@ extern struct cbe_iic_regs __iomem *cbe_get_cpu_iic_regs(int cpu);
 
 
 struct cbe_mic_tm_regs {
-       u8      pad_0x0000_0x0040[0x0040 - 0x0000];             /* 0x0000 */
+       u8      pad_0x0000_0x0010[0x0010 - 0x0000];             /* 0x0000 */
+
+       u64     MBL_debug;                                      /* 0x0010 */
+
+       u8      pad_0x0018_0x0040[0x0040 - 0x0018];             /* 0x0018 */
 
        u64     mic_ctl_cnfg2;                                  /* 0x0040 */
 #define CBE_MIC_ENABLE_AUX_TRC         0x8000000000000000LL
@@ -303,6 +308,25 @@ struct cbe_mic_tm_regs {
 extern struct cbe_mic_tm_regs __iomem *cbe_get_mic_tm_regs(struct device_node *np);
 extern struct cbe_mic_tm_regs __iomem *cbe_get_cpu_mic_tm_regs(int cpu);
 
+/*
+ *
+ *  PPE Privileged MMIO Registers definition. (offset 0x500000 - 0x500fff)
+ *
+ */
+struct cbe_ppe_priv_regs {
+       u8      pad_0x0000_0x0858[0x0858 - 0x0000];             /* 0x0000 */
+
+       u64     L2_debug1;                                      /* 0x0858 */
+
+       u8      pad_0x0860_0x0958[0x0958 - 0x0860];             /* 0x0860 */
+
+       u64     ciu_dr1;                                        /* 0x0958 */
+
+       u8      pad_0x0960_0x1000[0x1000 - 0x0960];             /* 0x0960 */
+};
+
+extern struct cbe_ppe_priv_regs __iomem *cbe_get_cpu_ppe_priv_regs(int cpu);
+
 /* some utility functions to deal with SMT */
 extern u32 cbe_get_hw_thread_id(int cpu);
 extern u32 cbe_cpu_to_node(int cpu);
diff --git a/arch/powerpc/include/asm/paca.h b/arch/powerpc/include/asm/paca.h
index 6493a39..ba9ead4 100644
--- a/arch/powerpc/include/asm/paca.h
+++ b/arch/powerpc/include/asm/paca.h
@@ -97,6 +97,10 @@ struct paca_struct {
        u8 soft_enabled;                /* irq soft-enable flag */
        u8 hard_enabled;                /* set if irqs are enabled in MSR */
        u8 io_sync;                     /* writel() needs spin_unlock sync */
+#ifdef CONFIG_PERFMON
+       u8 pmu_except_pending;          /* PMU exception occurred while soft
+                                        * disabled */
+#endif
 
        /* Stuff for accurate time accounting */
        u64 user_time;                  /* accumulated usermode TB ticks */
diff --git a/arch/powerpc/include/asm/perfmon.h b/arch/powerpc/include/asm/perfmon.h
new file mode 100644
index 0000000..da0ae3b
--- /dev/null
+++ b/arch/powerpc/include/asm/perfmon.h
@@ -0,0 +1,33 @@
+/*
+ * Copyright (c) 2007 Hewlett-Packard Development Company, L.P.
+ * Contributed by Stephane Eranian <eranian@hpl.hp.com>
+ *
+ * This file contains powerpc specific definitions for the perfmon
+ * interface.
+ *
+ * This file MUST never be included directly. Use linux/perfmon.h.
+ *
+ * This program is free software; you can redistribute it and/or
+ * modify it under the terms of version 2 of the GNU General Public
+ * License as published by the Free Software Foundation.
+ *
+ * This program is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+ * General Public License for more details.
+ *
+ * You should have received a copy of the GNU General Public License
+ * along with this program; if not, write to the Free Software
+ * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA
+ * 02111-1307 USA
+ */
+#ifndef _ASM_POWERPC_PERFMON_H_
+#define _ASM_POWERPC_PERFMON_H_
+
+/*
+ * arch-specific user visible interface definitions
+ */
+#define PFM_ARCH_MAX_PMCS      (256+64) /* 256 HW 64 SW */
+#define PFM_ARCH_MAX_PMDS      (256+64) /* 256 HW 64 SW */
+
+#endif /* _ASM_POWERPC_PERFMON_H_ */
diff --git a/arch/powerpc/include/asm/perfmon_kern.h b/arch/powerpc/include/asm/perfmon_kern.h
new file mode 100644
index 0000000..65ec984
--- /dev/null
+++ b/arch/powerpc/include/asm/perfmon_kern.h
@@ -0,0 +1,390 @@
+/*
+ * Copyright (c) 2005 David Gibson, IBM Corporation.
+ *
+ * Based on other versions:
+ * Copyright (c) 2005 Hewlett-Packard Development Company, L.P.
+ * Contributed by Stephane Eranian <eranian@hpl.hp.com>
+ *
+ * This file contains powerpc specific definitions for the perfmon
+ * interface.
+ *
+ * This program is free software; you can redistribute it and/or
+ * modify it under the terms of version 2 of the GNU General Public
+ * License as published by the Free Software Foundation.
+ *
+ * This program is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+ * General Public License for more details.
+ *
+ * You should have received a copy of the GNU General Public License
+ * along with this program; if not, write to the Free Software
+ * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA
+ * 02111-1307 USA
+ */
+#ifndef _ASM_POWERPC_PERFMON_KERN_H_
+#define _ASM_POWERPC_PERFMON_KERN_H_
+
+#ifdef __KERNEL__
+
+#ifdef CONFIG_PERFMON
+
+#include <asm/pmc.h>
+#include <asm/unistd.h>
+
+#define HID0_PMC5_6_GR_MODE (1UL << (63 - 40))
+
+enum powerpc_pmu_type {
+       PFM_POWERPC_PMU_NONE,
+       PFM_POWERPC_PMU_604,
+       PFM_POWERPC_PMU_604e,
+       PFM_POWERPC_PMU_750,    /* XXX: Minor event set diffs between IBM and Moto. */
+       PFM_POWERPC_PMU_7400,
+       PFM_POWERPC_PMU_7450,
+       PFM_POWERPC_PMU_POWER4,
+       PFM_POWERPC_PMU_POWER5,
+       PFM_POWERPC_PMU_POWER5p,
+       PFM_POWERPC_PMU_POWER6,
+       PFM_POWERPC_PMU_CELL,
+};
+
+struct pfm_arch_pmu_info {
+       enum powerpc_pmu_type pmu_style;
+
+       void (*write_pmc)(unsigned int cnum, u64 value);
+       void (*write_pmd)(unsigned int cnum, u64 value);
+
+       u64 (*read_pmd)(unsigned int cnum);
+
+       void (*enable_counters)(struct pfm_context *ctx,
+                               struct pfm_event_set *set);
+       void (*disable_counters)(struct pfm_context *ctx,
+                                struct pfm_event_set *set);
+
+       void (*irq_handler)(struct pt_regs *regs, struct pfm_context *ctx);
+       void (*get_ovfl_pmds)(struct pfm_context *ctx,
+                             struct pfm_event_set *set);
+
+       /* The following routines are optional. */
+       void (*restore_pmcs)(struct pfm_context *ctx,
+                            struct pfm_event_set *set);
+       void (*restore_pmds)(struct pfm_context *ctx,
+                            struct pfm_event_set *set);
+
+       int  (*ctxswout_thread)(struct task_struct *task,
+                               struct pfm_context *ctx,
+                               struct pfm_event_set *set);
+       void (*ctxswin_thread)(struct task_struct *task,
+                              struct pfm_context *ctx,
+                              struct pfm_event_set *set);
+       int  (*load_context)(struct pfm_context *ctx);
+       void (*unload_context)(struct pfm_context *ctx);
+       int  (*acquire_pmu)(u64 *unavail_pmcs, u64 *unavail_pmds);
+       void (*release_pmu)(void);
+       void *platform_info;
+       void (*resend_irq)(struct pfm_context *ctx);
+};
+
+#ifdef CONFIG_PPC32
+#define PFM_ARCH_PMD_STK_ARG   6 /* conservative value */
+#define PFM_ARCH_PMC_STK_ARG   6 /* conservative value */
+#else
+#define PFM_ARCH_PMD_STK_ARG   8 /* conservative value */
+#define PFM_ARCH_PMC_STK_ARG   8 /* conservative value */
+#endif
+
+static inline void pfm_arch_resend_irq(struct pfm_context *ctx)
+{
+       struct pfm_arch_pmu_info *arch_info;
+
+       arch_info = pfm_pmu_info();
+       arch_info->resend_irq(ctx);
+}
+
+static inline void pfm_arch_serialize(void)
+{}
+
+static inline void pfm_arch_write_pmc(struct pfm_context *ctx,
+                                     unsigned int cnum,
+                                     u64 value)
+{
+       struct pfm_arch_pmu_info *arch_info;
+
+       arch_info =  pfm_pmu_info();
+
+       /*
+        * we only write to the actual register when monitoring is
+        * active (pfm_start was issued)
+        */
+       if (ctx && ctx->flags.started == 0)
+               return;
+
+       BUG_ON(!arch_info->write_pmc);
+
+       arch_info->write_pmc(cnum, value);
+}
+
+static inline void pfm_arch_write_pmd(struct pfm_context *ctx,
+                                     unsigned int cnum, u64 value)
+{
+       struct pfm_arch_pmu_info *arch_info;
+
+       arch_info = pfm_pmu_info();
+
+       value &= pfm_pmu_conf->ovfl_mask;
+
+       BUG_ON(!arch_info->write_pmd);
+
+       arch_info->write_pmd(cnum, value);
+}
+
+static inline u64 pfm_arch_read_pmd(struct pfm_context *ctx, unsigned int cnum)
+{
+       struct pfm_arch_pmu_info *arch_info;
+
+       arch_info = pfm_pmu_info();
+
+       BUG_ON(!arch_info->read_pmd);
+
+       return arch_info->read_pmd(cnum);
+}
+
+/*
+ * For some CPUs, the upper bits of a counter must be set in order for the
+ * overflow interrupt to happen. On overflow, the counter has wrapped around,
+ * and the upper bits are cleared. This function may be used to set them back.
+ */
+static inline void pfm_arch_ovfl_reset_pmd(struct pfm_context *ctx,
+                                          unsigned int cnum)
+{
+       u64 val = pfm_arch_read_pmd(ctx, cnum);
+
+       /* This masks out overflow bit 31 */
+       pfm_arch_write_pmd(ctx, cnum, val);
+}
+
+/*
+ * At certain points, perfmon needs to know if monitoring has been
+ * explicitely started/stopped by user via pfm_start/pfm_stop. The
+ * information is tracked in flags.started. However on certain
+ * architectures, it may be possible to start/stop directly from
+ * user level with a single assembly instruction bypassing
+ * the kernel. This function must be used to determine by
+ * an arch-specific mean if monitoring is actually started/stopped.
+ */
+static inline int pfm_arch_is_active(struct pfm_context *ctx)
+{
+       return ctx->flags.started;
+}
+
+static inline void pfm_arch_ctxswout_sys(struct task_struct *task,
+                                        struct pfm_context *ctx)
+{}
+
+static inline void pfm_arch_ctxswin_sys(struct task_struct *task,
+                                       struct pfm_context *ctx)
+{}
+
+void pfm_arch_init_percpu(void);
+int  pfm_arch_is_monitoring_active(struct pfm_context *ctx);
+int  pfm_arch_ctxswout_thread(struct task_struct *task, struct pfm_context *ctx);
+void pfm_arch_ctxswin_thread(struct task_struct *task, struct pfm_context *ctx);
+void pfm_arch_stop(struct task_struct *task, struct pfm_context *ctx);
+void pfm_arch_start(struct task_struct *task, struct pfm_context *ctx);
+void pfm_arch_restore_pmds(struct pfm_context *ctx, struct pfm_event_set *set);
+void pfm_arch_restore_pmcs(struct pfm_context *ctx, struct pfm_event_set *set);
+void pfm_arch_clear_pmd_ovfl_cond(struct pfm_context *ctx, struct pfm_event_set *set);
+int  pfm_arch_get_ovfl_pmds(struct pfm_context *ctx,
+                           struct pfm_event_set *set);
+char *pfm_arch_get_pmu_module_name(void);
+/*
+ * called from __pfm_interrupt_handler(). ctx is not NULL.
+ * ctx is locked. PMU interrupt is masked.
+ *
+ * must stop all monitoring to ensure handler has consistent view.
+ * must collect overflowed PMDs bitmask  into povfls_pmds and
+ * npend_ovfls. If no interrupt detected then npend_ovfls
+ * must be set to zero.
+ */
+static inline void pfm_arch_intr_freeze_pmu(struct pfm_context *ctx, struct pfm_event_set *set)
+{
+       pfm_arch_stop(current, ctx);
+}
+
+void powerpc_irq_handler(struct pt_regs *regs);
+
+/*
+ * unfreeze PMU from pfm_do_interrupt_handler()
+ * ctx may be NULL for spurious
+ */
+static inline void pfm_arch_intr_unfreeze_pmu(struct pfm_context *ctx)
+{
+       struct pfm_arch_pmu_info *arch_info;
+
+       if (!ctx)
+               return;
+
+       PFM_DBG_ovfl("state=%d", ctx->state);
+
+       ctx->flags.started = 1;
+
+       if (ctx->state == PFM_CTX_MASKED)
+               return;
+
+       arch_info = pfm_pmu_info();
+       BUG_ON(!arch_info->enable_counters);
+       arch_info->enable_counters(ctx, ctx->active_set);
+}
+
+/*
+ * PowerPC does not save the PMDs during pfm_arch_intr_freeze_pmu(), thus
+ * this routine needs to do it when switching sets on overflow
+ */
+static inline void pfm_arch_save_pmds_from_intr(struct pfm_context *ctx,
+                                          struct pfm_event_set *set)
+{
+       pfm_save_pmds(ctx, set);
+}
+
+/*
+ * this function is called from the PMU interrupt handler ONLY.
+ * On PPC, the PMU is frozen via arch_stop, masking would be implemented
+ * via arch-stop as well. Given that the PMU is already stopped when
+ * entering the interrupt handler, we do not need to stop it again, so
+ * this function is a nop.
+ */
+static inline void pfm_arch_mask_monitoring(struct pfm_context *ctx,
+                                           struct pfm_event_set *set)
+{}
+
+/*
+ * Simply need to start the context in order to unmask.
+ */
+static inline void pfm_arch_unmask_monitoring(struct pfm_context *ctx,
+                                             struct pfm_event_set *set)
+{
+       pfm_arch_start(current, ctx);
+}
+
+
+static inline int pfm_arch_pmu_config_init(struct pfm_pmu_config *cfg)
+{
+       return 0;
+}
+
+static inline int pfm_arch_context_create(struct pfm_context *ctx,
+                                         u32 ctx_flags)
+{
+       return 0;
+}
+
+static inline void pfm_arch_context_free(struct pfm_context *ctx)
+{}
+
+/* not necessary on PowerPC */
+static inline void pfm_cacheflush(void *addr, unsigned int len)
+{}
+
+/*
+ * function called from pfm_setfl_sane(). Context is locked
+ * and interrupts are masked.
+ * The value of flags is the value of ctx_flags as passed by
+ * user.
+ *
+ * function must check arch-specific set flags.
+ * Return:
+ *     1 when flags are valid
+ *      0 on error
+ */
+static inline int pfm_arch_setfl_sane(struct pfm_context *ctx, u32 flags)
+{
+       return 0;
+}
+
+static inline int pfm_arch_init(void)
+{
+       return 0;
+}
+
+static inline int pfm_arch_load_context(struct pfm_context *ctx)
+{
+       struct pfm_arch_pmu_info *arch_info;
+       int rc = 0;
+
+       arch_info = pfm_pmu_info();
+       if (arch_info->load_context)
+               rc = arch_info->load_context(ctx);
+
+       return rc;
+}
+
+static inline void pfm_arch_unload_context(struct pfm_context *ctx)
+{
+       struct pfm_arch_pmu_info *arch_info;
+
+       arch_info = pfm_pmu_info();
+       if (arch_info->unload_context)
+               arch_info->unload_context(ctx);
+}
+
+static inline int pfm_arch_pmu_acquire(u64 *unavail_pmcs, u64 *unavail_pmds)
+{
+       struct pfm_arch_pmu_info *arch_info;
+       int rc = 0;
+
+       arch_info = pfm_pmu_info();
+       if (arch_info->acquire_pmu) {
+               rc = arch_info->acquire_pmu(unavail_pmcs, unavail_pmds);
+               if (rc)
+                       return rc;
+       }
+
+       return reserve_pmc_hardware(powerpc_irq_handler);
+}
+
+static inline void pfm_arch_pmu_release(void)
+{
+       struct pfm_arch_pmu_info *arch_info;
+
+       arch_info = pfm_pmu_info();
+       if (arch_info->release_pmu)
+               arch_info->release_pmu();
+
+       release_pmc_hardware();
+}
+
+static inline void pfm_arch_arm_handle_work(struct task_struct *task)
+{}
+
+static inline void pfm_arch_disarm_handle_work(struct task_struct *task)
+{}
+
+static inline int pfm_arch_get_base_syscall(void)
+{
+       return __NR_pfm_create_context;
+}
+
+struct pfm_arch_context {
+       /* Cell: Most recent value of the pm_status
+        * register read by the interrupt handler.
+        *
+        * Interrupt handler sets last_read_updated if it
+        * just read and updated last_read_pm_status
+        */
+       u32 last_read_pm_status;
+       u32 last_read_updated;
+       u64 powergs_pmc5, powergs_pmc6;
+       u64 delta_tb, delta_tb_start;
+       u64 delta_purr, delta_purr_start;
+};
+
+#define PFM_ARCH_CTX_SIZE sizeof(struct pfm_arch_context)
+/*
+ * PowerPC does not need extra alignment requirements for the sampling buffer
+ */
+#define PFM_ARCH_SMPL_ALIGN_SIZE       0
+
+#endif /* CONFIG_PERFMON */
+
+#endif /* __KERNEL__ */
+#endif /* _ASM_POWERPC_PERFMON_KERN_H_ */
diff --git a/arch/powerpc/include/asm/reg.h b/arch/powerpc/include/asm/reg.h
index c6d1ab6..a9f3ad0 100644
--- a/arch/powerpc/include/asm/reg.h
+++ b/arch/powerpc/include/asm/reg.h
@@ -698,6 +698,7 @@
 #define PV_POWER5      0x003A
 #define PV_POWER5p     0x003B
 #define PV_970FX       0x003C
+#define PV_POWER6      0x003E
 #define PV_630         0x0040
 #define PV_630p        0x0041
 #define PV_970MP       0x0044
diff --git a/arch/powerpc/include/asm/systbl.h b/arch/powerpc/include/asm/systbl.h
index f6cc7a4..0164841 100644
--- a/arch/powerpc/include/asm/systbl.h
+++ b/arch/powerpc/include/asm/systbl.h
@@ -322,3 +322,15 @@ SYSCALL_SPU(epoll_create1)
 SYSCALL_SPU(dup3)
 SYSCALL_SPU(pipe2)
 SYSCALL(inotify_init1)
+SYSCALL(pfm_create_context)
+SYSCALL(pfm_write_pmcs)
+SYSCALL(pfm_write_pmds)
+SYSCALL(pfm_read_pmds)
+SYSCALL(pfm_load_context)
+SYSCALL(pfm_start)
+SYSCALL(pfm_stop)
+SYSCALL(pfm_restart)
+SYSCALL(pfm_create_evtsets)
+SYSCALL(pfm_getinfo_evtsets)
+SYSCALL(pfm_delete_evtsets)
+SYSCALL(pfm_unload_context)
diff --git a/arch/powerpc/include/asm/thread_info.h b/arch/powerpc/include/asm/thread_info.h
index 9665a26..6cda9f9 100644
--- a/arch/powerpc/include/asm/thread_info.h
+++ b/arch/powerpc/include/asm/thread_info.h
@@ -130,10 +130,12 @@ static inline struct thread_info *current_thread_info(void)
 #define _TIF_FREEZE            (1<<TIF_FREEZE)
 #define _TIF_RUNLATCH          (1<<TIF_RUNLATCH)
 #define _TIF_ABI_PENDING       (1<<TIF_ABI_PENDING)
+#define _TIF_PERFMON_WORK      (1<<TIF_PERFMON_WORK)
+#define _TIF_PERFMON_CTXSW     (1<<TIF_PERFMON_CTXSW)
 #define _TIF_SYSCALL_T_OR_A    (_TIF_SYSCALL_TRACE|_TIF_SYSCALL_AUDIT|_TIF_SECCOMP)
 
 #define _TIF_USER_WORK_MASK    (_TIF_SIGPENDING | _TIF_NEED_RESCHED | \
-                                _TIF_NOTIFY_RESUME)
+                                _TIF_NOTIFY_RESUME | _TIF_PERFMON_WORK)
 #define _TIF_PERSYSCALL_MASK   (_TIF_RESTOREALL|_TIF_NOERROR)
 
 /* Bits in local_flags */
diff --git a/arch/powerpc/include/asm/unistd.h b/arch/powerpc/include/asm/unistd.h
index e07d0c7..6226cba 100644
--- a/arch/powerpc/include/asm/unistd.h
+++ b/arch/powerpc/include/asm/unistd.h
@@ -341,10 +341,22 @@
 #define __NR_dup3              316
 #define __NR_pipe2             317
 #define __NR_inotify_init1     318
+#define __NR_pfm_create_context        319
+#define __NR_pfm_write_pmcs    320
+#define __NR_pfm_write_pmds    321
+#define __NR_pfm_read_pmds     322
+#define __NR_pfm_load_context  323
+#define __NR_pfm_start         324
+#define __NR_pfm_stop          325
+#define __NR_pfm_restart       326
+#define __NR_pfm_create_evtsets        327
+#define __NR_pfm_getinfo_evtsets 328
+#define __NR_pfm_delete_evtsets 329
+#define __NR_pfm_unload_context        330
 
 #ifdef __KERNEL__
 
-#define __NR_syscalls          319
+#define __NR_syscalls          331
 
 #define __NR__exit __NR_exit
 #define NR_syscalls    __NR_syscalls
diff --git a/arch/powerpc/kernel/entry_32.S b/arch/powerpc/kernel/entry_32.S
index 1cbbf70..198645f 100644
--- a/arch/powerpc/kernel/entry_32.S
+++ b/arch/powerpc/kernel/entry_32.S
@@ -39,7 +39,7 @@
  * MSR_KERNEL is > 0x10000 on 4xx/Book-E since it include MSR_CE.
  */
 #if MSR_KERNEL >= 0x10000
-#define LOAD_MSR_KERNEL(r, x)  lis r,(x)@h; ori r,r,(x)@l
+#define LOAD_MSR_KERNEL(r, x)  lis r,(x)@ha; ori r,r,(x)@l
 #else
 #define LOAD_MSR_KERNEL(r, x)  li r,(x)
 #endif
diff --git a/arch/powerpc/kernel/entry_64.S b/arch/powerpc/kernel/entry_64.S
index 2d802e9..77a090d 100644
--- a/arch/powerpc/kernel/entry_64.S
+++ b/arch/powerpc/kernel/entry_64.S
@@ -643,6 +643,10 @@ user_work:
        b       .ret_from_except_lite
 
 1:     bl      .save_nvgprs
+#ifdef CONFIG_PERFMON
+       addi    r3,r1,STACK_FRAME_OVERHEAD
+       bl      .pfm_handle_work
+#endif /* CONFIG_PERFMON */
        addi    r3,r1,STACK_FRAME_OVERHEAD
        bl      .do_signal
        b       .ret_from_except
diff --git a/arch/powerpc/kernel/irq.c b/arch/powerpc/kernel/irq.c
index d972dec..b255fba 100644
--- a/arch/powerpc/kernel/irq.c
+++ b/arch/powerpc/kernel/irq.c
@@ -104,6 +104,24 @@ static inline notrace void set_soft_enabled(unsigned long enable)
        : : "r" (enable), "i" (offsetof(struct paca_struct, soft_enabled)));
 }
 
+#ifdef CONFIG_PERFMON
+static inline unsigned long get_pmu_except_pending(void)
+{
+       unsigned long pending;
+
+       __asm__ __volatile__("lbz %0,%1(13)"
+       : "=r" (pending) : "i" (offsetof(struct paca_struct, pmu_except_pending)));
+
+       return pending;
+}
+
+static inline void set_pmu_except_pending(unsigned long pending)
+{
+       __asm__ __volatile__("stb %0,%1(13)"
+       : : "r" (pending), "i" (offsetof(struct paca_struct, pmu_except_pending)));
+}
+#endif /* CONFIG_PERFMON */
+
 notrace void raw_local_irq_restore(unsigned long en)
 {
        /*
@@ -162,6 +180,19 @@ notrace void raw_local_irq_restore(unsigned long en)
                lv1_get_version_info(&tmp);
        }
 
+#ifdef CONFIG_PERFMON
+       /*
+        * If a PMU exception occurred while interrupts were soft disabled,
+        * force a PMU exception.
+        */
+       if (get_pmu_except_pending()) {
+               set_pmu_except_pending(0);
+               /* Make sure we trigger the edge detection circuitry */
+               mtspr(SPRN_MMCR0, mfspr(SPRN_MMCR0) & ~MMCR0_PMAO);
+               mtspr(SPRN_MMCR0, mfspr(SPRN_MMCR0) | MMCR0_PMAO);
+       }
+#endif /* CONFIG_PERFMON */
+
        __hard_irq_enable();
 }
 EXPORT_SYMBOL(raw_local_irq_restore);
diff --git a/arch/powerpc/kernel/process.c b/arch/powerpc/kernel/process.c
index 957bded..32dbc8e 100644
--- a/arch/powerpc/kernel/process.c
+++ b/arch/powerpc/kernel/process.c
@@ -33,6 +33,7 @@
 #include <linux/mqueue.h>
 #include <linux/hardirq.h>
 #include <linux/utsname.h>
+#include <linux/perfmon_kern.h>
 
 #include <asm/pgtable.h>
 #include <asm/uaccess.h>
@@ -393,9 +394,14 @@ struct task_struct *__switch_to(struct task_struct *prev,
                new_thread->start_tb = current_tb;
        }
 #endif
-
        local_irq_save(flags);
 
+       if (test_tsk_thread_flag(prev, TIF_PERFMON_CTXSW))
+               pfm_ctxsw_out(prev, new);
+
+       if (test_tsk_thread_flag(new, TIF_PERFMON_CTXSW))
+               pfm_ctxsw_in(prev, new);
+
        account_system_vtime(current);
        account_process_vtime(current);
        calculate_steal_time();
@@ -544,6 +550,7 @@ void show_regs(struct pt_regs * regs)
 void exit_thread(void)
 {
        discard_lazy_cpu_state();
+       pfm_exit_thread();
 }
 
 void flush_thread(void)
@@ -669,6 +676,7 @@ int copy_thread(int nr, unsigned long clone_flags, unsigned long usp,
 #else
        kregs->nip = (unsigned long)ret_from_fork;
 #endif
+       pfm_copy_thread(p);
 
        return 0;
 }
diff --git a/arch/powerpc/perfmon/Kconfig b/arch/powerpc/perfmon/Kconfig
new file mode 100644
index 0000000..3f4bbf2
--- /dev/null
+++ b/arch/powerpc/perfmon/Kconfig
@@ -0,0 +1,67 @@
+menu "Hardware Performance Monitoring support"
+config PERFMON
+       bool "Perfmon2 performance monitoring interface"
+       default n
+       help
+       Enables the perfmon2 interface to access the hardware
+       performance counters. See <http://perfmon2.sf.net/> for
+       more details.
+
+config PERFMON_DEBUG
+       bool "Perfmon debugging"
+       default n
+       depends on PERFMON
+       help
+       Enables perfmon debugging support
+
+config PERFMON_DEBUG_FS
+       bool "Enable perfmon statistics reporting via debugfs"
+       default y
+       depends on PERFMON && DEBUG_FS
+       help
+       Enable collection and reporting of perfmon timing statistics under
+       debugfs. This is used for debugging and performance analysis of the
+       subsystem. The debugfs filesystem must be mounted.
+
+config PERFMON_POWER4
+       tristate "Support for Power4 hardware performance counters"
+       depends on PERFMON && PPC64
+       default n
+       help
+       Enables support for the Power 4 hardware performance counters
+       If unsure, say M.
+
+config PERFMON_POWER5
+       tristate "Support for Power5 hardware performance counters"
+       depends on PERFMON && PPC64
+       default n
+       help
+       Enables support for the Power 5 hardware performance counters
+       If unsure, say M.
+
+config PERFMON_POWER6
+       tristate "Support for Power6 hardware performance counters"
+       depends on PERFMON && PPC64
+       default n
+       help
+       Enables support for the Power 6 hardware performance counters
+       If unsure, say M.
+
+config PERFMON_PPC32
+       tristate "Support for PPC32 hardware performance counters"
+       depends on PERFMON && PPC32
+       default n
+       help
+       Enables support for the PPC32 hardware performance counters
+       If unsure, say M.
+
+config PERFMON_CELL
+       tristate "Support for Cell hardware performance counters"
+       depends on PERFMON && PPC_CELL
+       select PS3_LPM if PPC_PS3
+       default n
+       help
+       Enables support for the Cell hardware performance counters.
+       If unsure, say M.
+
+endmenu
diff --git a/arch/powerpc/perfmon/Makefile b/arch/powerpc/perfmon/Makefile
new file mode 100644
index 0000000..300661f
--- /dev/null
+++ b/arch/powerpc/perfmon/Makefile
@@ -0,0 +1,6 @@
+obj-$(CONFIG_PERFMON)          += perfmon.o
+obj-$(CONFIG_PERFMON_POWER4)   += perfmon_power4.o
+obj-$(CONFIG_PERFMON_POWER5)   += perfmon_power5.o
+obj-$(CONFIG_PERFMON_POWER6)   += perfmon_power6.o
+obj-$(CONFIG_PERFMON_PPC32)    += perfmon_ppc32.o
+obj-$(CONFIG_PERFMON_CELL)     += perfmon_cell.o
diff --git a/arch/powerpc/perfmon/perfmon.c b/arch/powerpc/perfmon/perfmon.c
new file mode 100644
index 0000000..51a8b6a
--- /dev/null
+++ b/arch/powerpc/perfmon/perfmon.c
@@ -0,0 +1,334 @@
+/*
+ * This file implements the powerpc specific
+ * support for the perfmon2 interface
+ *
+ * Copyright (c) 2005 David Gibson, IBM Corporation.
+ *
+ * based on versions for other architectures:
+ * Copyright (c) 2005-2006 Hewlett-Packard Development Company, L.P.
+ * Contributed by Stephane Eranian <eranian@hpl.hp.com>
+ *
+ * This program is free software; you can redistribute it and/or
+ * modify it under the terms of version 2 of the GNU General Public
+ * License as published by the Free Software Foundation.
+ *
+ * This program is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+ * General Public License for more details.
+ *
+ * You should have received a copy of the GNU General Public License
+ * along with this program; if not, write to the Free Software
+ * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA
+ * 02111-1307 USA
+ */
+#include <linux/interrupt.h>
+#include <linux/perfmon_kern.h>
+
+static void pfm_stop_active(struct task_struct *task,
+                           struct pfm_context *ctx, struct pfm_event_set *set)
+{
+       struct pfm_arch_pmu_info *arch_info;
+
+       arch_info = pfm_pmu_info();
+       BUG_ON(!arch_info->disable_counters || !arch_info->get_ovfl_pmds);
+
+       arch_info->disable_counters(ctx, set);
+
+       if (set->npend_ovfls)
+               return;
+
+       arch_info->get_ovfl_pmds(ctx, set);
+}
+
+/*
+ * Called from pfm_save_pmds(). Interrupts are masked.  Registers are
+ * already saved away.
+ */
+void pfm_arch_clear_pmd_ovfl_cond(struct pfm_context *ctx,
+                                 struct pfm_event_set *set)
+{
+       int i, num;
+       u64 *used_pmds, *intr_pmds;
+
+       num = set->nused_pmds;
+       used_pmds = set->used_pmds;
+       intr_pmds = ctx->regs.intr_pmds;
+
+       for (i = 0; num; i++)
+               if (likely(test_bit(i, used_pmds))) {
+                       if (likely(test_bit(i, intr_pmds)))
+                               pfm_write_pmd(ctx, i, 0);
+                       num--;
+               }
+}
+
+/*
+ * Called from pfm_ctxsw(). Task is guaranteed to be current.
+ * Context is locked. Interrupts are masked. Monitoring is active.
+ * PMU access is guaranteed. PMC and PMD registers are live in PMU.
+ *
+ * for per-thread:
+ *     must stop monitoring for the task
+ * Return:
+ *     non-zero : did not save PMDs (as part of stopping the PMU)
+ *            0 : saved PMDs (no need to save them in caller)
+ */
+int pfm_arch_ctxswout_thread(struct task_struct *task, struct pfm_context *ctx)
+{
+       struct pfm_arch_pmu_info *arch_info;
+
+       arch_info = pfm_pmu_info();
+       /*
+        * disable lazy restore of the PMC/PMD registers.
+        */
+       ctx->active_set->priv_flags |= PFM_SETFL_PRIV_MOD_BOTH;
+
+       if (ctx->state == PFM_CTX_MASKED)
+               return 1;
+
+       pfm_stop_active(task, ctx, ctx->active_set);
+
+       if (arch_info->ctxswout_thread)
+               arch_info->ctxswout_thread(task, ctx, ctx->active_set);
+
+       return pfm_arch_is_active(ctx);
+}
+
+/*
+ * Called from pfm_ctxsw
+ */
+void pfm_arch_ctxswin_thread(struct task_struct *task, struct pfm_context *ctx)
+{
+       struct pfm_arch_pmu_info *arch_info;
+
+       arch_info = pfm_pmu_info();
+       if (ctx->state != PFM_CTX_MASKED && ctx->flags.started == 1) {
+               BUG_ON(!arch_info->enable_counters);
+               arch_info->enable_counters(ctx, ctx->active_set);
+       }
+
+       if (arch_info->ctxswin_thread)
+               arch_info->ctxswin_thread(task, ctx, ctx->active_set);
+}
+
+/*
+ * Called from pfm_stop() and idle notifier
+ *
+ * Interrupts are masked. Context is locked. Set is the active set.
+ *
+ * For per-thread:
+ *   task is not necessarily current. If not current task, then
+ *   task is guaranteed stopped and off any cpu. Access to PMU
+ *   is not guaranteed. Interrupts are masked. Context is locked.
+ *   Set is the active set.
+ *
+ * For system-wide:
+ *     task is current
+ *
+ * must disable active monitoring. ctx cannot be NULL
+ */
+void pfm_arch_stop(struct task_struct *task, struct pfm_context *ctx)
+{
+       /*
+        * no need to go through stop_save()
+        * if we are already stopped
+        */
+       if (!ctx->flags.started || ctx->state == PFM_CTX_MASKED)
+               return;
+
+       /*
+        * stop live registers and collect pending overflow
+        */
+       if (task == current)
+               pfm_stop_active(task, ctx, ctx->active_set);
+}
+
+/*
+ * Enable active monitoring. Called from pfm_start() and
+ * pfm_arch_unmask_monitoring().
+ *
+ * Interrupts are masked. Context is locked. Set is the active set.
+ *
+ * For per-thread:
+ *     Task is not necessarily current. If not current task, then task
+ *     is guaranteed stopped and off any cpu. No access to PMU if task
+ *     is not current.
+ *
+ * For system-wide:
+ *     Task is always current
+ */
+void pfm_arch_start(struct task_struct *task, struct pfm_context *ctx)
+{
+       struct pfm_arch_pmu_info *arch_info;
+
+       arch_info = pfm_pmu_info();
+       if (task != current)
+               return;
+
+       BUG_ON(!arch_info->enable_counters);
+
+       arch_info->enable_counters(ctx, ctx->active_set);
+}
+
+/*
+ * function called from pfm_switch_sets(), pfm_context_load_thread(),
+ * pfm_context_load_sys(), pfm_ctxsw(), pfm_switch_sets()
+ * context is locked. Interrupts are masked. set cannot be NULL.
+ * Access to the PMU is guaranteed.
+ *
+ * function must restore all PMD registers from set.
+ */
+void pfm_arch_restore_pmds(struct pfm_context *ctx, struct pfm_event_set *set)
+{
+       struct pfm_arch_pmu_info *arch_info;
+       u64 *used_pmds;
+       u16 i, num;
+
+       arch_info = pfm_pmu_info();
+
+       /* The model-specific module can override the default
+        * restore-PMD method.
+        */
+       if (arch_info->restore_pmds)
+               return arch_info->restore_pmds(ctx, set);
+
+       num = set->nused_pmds;
+       used_pmds = set->used_pmds;
+
+       for (i = 0; num; i++) {
+               if (likely(test_bit(i, used_pmds))) {
+                       pfm_write_pmd(ctx, i, set->pmds[i].value);
+                       num--;
+               }
+       }
+}
+
+/*
+ * function called from pfm_switch_sets(), pfm_context_load_thread(),
+ * pfm_context_load_sys(), pfm_ctxsw(), pfm_switch_sets()
+ * context is locked. Interrupts are masked. set cannot be NULL.
+ * Access to the PMU is guaranteed.
+ *
+ * function must restore all PMC registers from set, if needed.
+ */
+void pfm_arch_restore_pmcs(struct pfm_context *ctx, struct pfm_event_set *set)
+{
+       struct pfm_arch_pmu_info *arch_info;
+       u64 *impl_pmcs;
+       unsigned int i, max_pmc, reg;
+
+       arch_info = pfm_pmu_info();
+       /* The model-specific module can override the default
+        * restore-PMC method.
+        */
+       if (arch_info->restore_pmcs)
+               return arch_info->restore_pmcs(ctx, set);
+
+       /* The "common" powerpc model's enable the counters simply by writing
+        * all the control registers. Therefore, if we're masked or stopped we
+        * don't need to bother restoring the PMCs now.
+        */
+       if (ctx->state == PFM_CTX_MASKED || ctx->flags.started == 0)
+               return;
+
+       max_pmc = ctx->regs.max_pmc;
+       impl_pmcs = ctx->regs.pmcs;
+
+       /*
+        * Restore all pmcs in reverse order to ensure the counters aren't
+        * enabled before their event selectors are set correctly.
+        */
+       reg = max_pmc - 1;
+       for (i = 0; i < max_pmc; i++) {
+               if (test_bit(reg, impl_pmcs))
+                       pfm_arch_write_pmc(ctx, reg, set->pmcs[reg]);
+               reg--;
+       }
+}
+
+char *pfm_arch_get_pmu_module_name(void)
+{
+       unsigned int pvr = mfspr(SPRN_PVR);
+
+       switch (PVR_VER(pvr)) {
+       case 0x0004: /* 604 */
+       case 0x0009: /* 604e;  */
+       case 0x000A: /* 604ev */
+       case 0x0008: /* 750/740 */
+       case 0x7000: /* 750FX */
+       case 0x7001:
+       case 0x7002: /* 750GX */
+       case 0x000C: /* 7400 */
+       case 0x800C: /* 7410 */
+       case 0x8000: /* 7451/7441 */
+       case 0x8001: /* 7455/7445 */
+       case 0x8002: /* 7457/7447 */
+       case 0x8003: /* 7447A */
+       case 0x8004: /* 7448 */
+               return("perfmon_ppc32");
+       case PV_POWER4:
+       case PV_POWER4p:
+               return "perfmon_power4";
+       case PV_POWER5:
+               return "perfmon_power5";
+       case PV_POWER5p:
+               if (PVR_REV(pvr) < 0x300)
+                       /* PMU behaves like POWER5 */
+                       return "perfmon_power5";
+               else
+                       /* PMU behaves like POWER6 */
+                       return "perfmon_power6";
+       case PV_POWER6:
+               return "perfmon_power6";
+       case PV_970:
+       case PV_970FX:
+       case PV_970MP:
+               return "perfmon_ppc970";
+       case PV_BE:
+               return "perfmon_cell";
+       }
+       return NULL;
+}
+
+void pfm_arch_init_percpu(void)
+{
+#ifdef CONFIG_PPC64
+       extern void ppc64_enable_pmcs(void);
+       ppc64_enable_pmcs();
+#endif
+}
+
+/**
+ * powerpc_irq_handler
+ *
+ * Get the perfmon context that belongs to the current CPU, and call the
+ * model-specific interrupt handler.
+ **/
+void powerpc_irq_handler(struct pt_regs *regs)
+{
+       struct pfm_arch_pmu_info *arch_info;
+       struct pfm_context *ctx;
+
+       if (! regs->softe) {
+               /*
+                * We got a PMU interrupt while interrupts were soft
+                * disabled.  Disable hardware interrupts by clearing
+                * MSR_EE and also clear PMAO because we will need to set
+                * that again later when interrupts are re-enabled and
+                * raw_local_irq_restore() sees that the pmu_except_pending
+                * flag is set.
+                */
+               regs->msr &= ~MSR_EE;
+               get_paca()->pmu_except_pending = 1;
+               mtspr(SPRN_MMCR0, mfspr(SPRN_MMCR0) & ~MMCR0_PMAO);
+               return;
+       }
+
+       arch_info = pfm_pmu_info();
+       if (arch_info->irq_handler) {
+               ctx = __get_cpu_var(pmu_ctx);
+               if (likely(ctx))
+                       arch_info->irq_handler(regs, ctx);
+       }
+}
diff --git a/arch/powerpc/perfmon/perfmon_cell.c b/arch/powerpc/perfmon/perfmon_cell.c
new file mode 100644
index 0000000..e1ae12c
--- /dev/null
+++ b/arch/powerpc/perfmon/perfmon_cell.c
@@ -0,0 +1,1449 @@
+/*
+ * This file contains the Cell PMU register description tables
+ * and pmc checker used by perfmon.c.
+ *
+ * Copyright IBM Corporation 2007
+ * (C) Copyright 2007 TOSHIBA CORPORATION
+ *
+ * Based on other Perfmon2 PMU modules.
+ * Copyright (c) 2005-2006 Hewlett-Packard Development Company, L.P.
+ * Contributed by Stephane Eranian <eranian@hpl.hp.com>
+ *
+ * This program is free software; you can redistribute it and/or
+ * modify it under the terms of version 2 of the GNU General Public
+ * License as published by the Free Software Foundation.
+ *
+ * This program is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+ * General Public License for more details.
+ *
+ * You should have received a copy of the GNU General Public License
+ * along with this program; if not, write to the Free Software
+ * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA
+ * 02111-1307 USA
+ */
+
+#include <linux/module.h>
+#include <linux/perfmon_kern.h>
+#include <linux/io.h>
+#include <asm/cell-pmu.h>
+#include <asm/cell-regs.h>
+#include <asm/machdep.h>
+#include <asm/rtas.h>
+#include <asm/ps3.h>
+#include <asm/spu.h>
+
+MODULE_AUTHOR("Kevin Corry <kevcorry@us.ibm.com>, "
+             "Carl Love <carll@us.ibm.com>");
+MODULE_DESCRIPTION("Cell PMU description table");
+MODULE_LICENSE("GPL");
+
+struct pfm_cell_platform_pmu_info {
+       u32  (*read_ctr)(u32 cpu, u32 ctr);
+       void (*write_ctr)(u32 cpu, u32 ctr, u32 val);
+       void (*write_pm07_control)(u32 cpu, u32 ctr, u32 val);
+       void (*write_pm)(u32 cpu, enum pm_reg_name reg, u32 val);
+       void (*enable_pm)(u32 cpu);
+       void (*disable_pm)(u32 cpu);
+       void (*enable_pm_interrupts)(u32 cpu, u32 thread, u32 mask);
+       u32  (*get_and_clear_pm_interrupts)(u32 cpu);
+       u32  (*get_hw_thread_id)(int cpu);
+       struct cbe_ppe_priv_regs __iomem *(*get_cpu_ppe_priv_regs)(int cpu);
+       struct cbe_pmd_regs __iomem *(*get_cpu_pmd_regs)(int cpu);
+       struct cbe_mic_tm_regs __iomem *(*get_cpu_mic_tm_regs)(int cpu);
+       int (*rtas_token)(const char *service);
+       int (*rtas_call)(int token, int param1, int param2, int *param3, ...);
+};
+
+/*
+ * Mapping from Perfmon logical control registers to Cell hardware registers.
+ */
+static struct pfm_regmap_desc pfm_cell_pmc_desc[] = {
+       /* Per-counter control registers. */
+       PMC_D(PFM_REG_I, "pm0_control",       0, 0, 0, 0),
+       PMC_D(PFM_REG_I, "pm1_control",       0, 0, 0, 0),
+       PMC_D(PFM_REG_I, "pm2_control",       0, 0, 0, 0),
+       PMC_D(PFM_REG_I, "pm3_control",       0, 0, 0, 0),
+       PMC_D(PFM_REG_I, "pm4_control",       0, 0, 0, 0),
+       PMC_D(PFM_REG_I, "pm5_control",       0, 0, 0, 0),
+       PMC_D(PFM_REG_I, "pm6_control",       0, 0, 0, 0),
+       PMC_D(PFM_REG_I, "pm7_control",       0, 0, 0, 0),
+
+       /* Per-counter RTAS arguments. Each of these registers has three fields.
+        *   bits 63-48: debug-bus word
+        *   bits 47-32: sub-unit
+        *   bits 31-0 : full signal number
+        *   (MSB = 63, LSB = 0)
+        */
+       PMC_D(PFM_REG_I, "pm0_event",         0, 0, 0, 0),
+       PMC_D(PFM_REG_I, "pm1_event",         0, 0, 0, 0),
+       PMC_D(PFM_REG_I, "pm2_event",         0, 0, 0, 0),
+       PMC_D(PFM_REG_I, "pm3_event",         0, 0, 0, 0),
+       PMC_D(PFM_REG_I, "pm4_event",         0, 0, 0, 0),
+       PMC_D(PFM_REG_I, "pm5_event",         0, 0, 0, 0),
+       PMC_D(PFM_REG_I, "pm6_event",         0, 0, 0, 0),
+       PMC_D(PFM_REG_I, "pm7_event",         0, 0, 0, 0),
+
+       /* Global control registers. Same order as enum pm_reg_name. */
+       PMC_D(PFM_REG_I, "group_control",     0, 0, 0, 0),
+       PMC_D(PFM_REG_I, "debug_bus_control", 0, 0, 0, 0),
+       PMC_D(PFM_REG_I, "trace_address",     0, 0, 0, 0),
+       PMC_D(PFM_REG_I, "ext_trace_timer",   0, 0, 0, 0),
+       PMC_D(PFM_REG_I, "pm_status",         0, 0, 0, 0),
+       /* set the interrupt overflow bit for the four 32 bit counters
+        * that is currently supported.  Will need to fix when 32 and 16
+        * bit counters are supported.
+        */
+       PMC_D(PFM_REG_I, "pm_control",        0xF0000000, 0xF0000000, 0, 0),
+       PMC_D(PFM_REG_I, "pm_interval",       0, 0, 0, 0), /* FIX: Does user-space also need read access to this one? */
+       PMC_D(PFM_REG_I, "pm_start_stop",     0, 0, 0, 0),
+};
+#define PFM_PM_NUM_PMCS        ARRAY_SIZE(pfm_cell_pmc_desc)
+
+#define CELL_PMC_GROUP_CONTROL    16
+#define CELL_PMC_PM_STATUS        20
+#define CELL_PMC_PM_CONTROL       21
+#define CELL_PMC_PM_CONTROL_CNTR_MASK   0x01E00000UL
+#define CELL_PMC_PM_CONTROL_CNTR_16     0x01E00000UL
+
+/*
+ * Mapping from Perfmon logical data counters to Cell hardware counters.
+ */
+static struct pfm_regmap_desc pfm_cell_pmd_desc[] = {
+       PMD_D(PFM_REG_C, "pm0", 0),
+       PMD_D(PFM_REG_C, "pm1", 0),
+       PMD_D(PFM_REG_C, "pm2", 0),
+       PMD_D(PFM_REG_C, "pm3", 0),
+       PMD_D(PFM_REG_C, "pm4", 0),
+       PMD_D(PFM_REG_C, "pm5", 0),
+       PMD_D(PFM_REG_C, "pm6", 0),
+       PMD_D(PFM_REG_C, "pm7", 0),
+};
+#define PFM_PM_NUM_PMDS        ARRAY_SIZE(pfm_cell_pmd_desc)
+
+#define PFM_EVENT_PMC_BUS_WORD(x)      (((x) >> 48) & 0x00ff)
+#define PFM_EVENT_PMC_FULL_SIGNAL_NUMBER(x) ((x) & 0xffffffff)
+#define PFM_EVENT_PMC_SIGNAL_GROUP(x) (((x) & 0xffffffff) / 100)
+#define PFM_PM_CTR_INPUT_MUX_BIT(pm07_control) (((pm07_control) >> 26) & 0x1f)
+#define PFM_PM_CTR_INPUT_MUX_GROUP_INDEX(pm07_control) ((pm07_control) >> 31)
+#define PFM_GROUP_CONTROL_GROUP0_WORD(grp_ctrl) ((grp_ctrl) >> 30)
+#define PFM_GROUP_CONTROL_GROUP1_WORD(grp_ctrl) (((grp_ctrl) >> 28) & 0x3)
+#define PFM_NUM_OF_GROUPS 2
+#define PFM_PPU_IU1_THREAD1_BASE_BIT 19
+#define PFM_PPU_XU_THREAD1_BASE_BIT  16
+#define PFM_COUNTER_CTRL_PMC_PPU_TH0 0x100000000ULL
+#define PFM_COUNTER_CTRL_PMC_PPU_TH1 0x200000000ULL
+
+/*
+ * Debug-bus signal handling.
+ *
+ * Some Cell systems have firmware that can handle the debug-bus signal
+ * routing. For systems without this firmware, we have a minimal in-kernel
+ * implementation as well.
+ */
+
+/* The firmware only sees physical CPUs, so divide by 2 if SMT is on. */
+#ifdef CONFIG_SCHED_SMT
+#define RTAS_CPU(cpu) ((cpu) / 2)
+#else
+#define RTAS_CPU(cpu) (cpu)
+#endif
+#define RTAS_BUS_WORD(x)      (u16)(((x) >> 48) & 0x0000ffff)
+#define RTAS_SUB_UNIT(x)      (u16)(((x) >> 32) & 0x0000ffff)
+#define RTAS_SIGNAL_NUMBER(x) (s32)( (x)        & 0xffffffff)
+#define RTAS_SIGNAL_GROUP(x)  (RTAS_SIGNAL_NUMBER(x) / 100)
+
+#define subfunc_RESET          1
+#define subfunc_ACTIVATE       2
+
+#define passthru_ENABLE                1
+#define passthru_DISABLE       2
+
+/**
+ * struct cell_rtas_arg
+ *
+ * @cpu: Processor to modify. Linux numbers CPUs based on SMT IDs, but the
+ *       firmware only sees the physical CPUs. So this value should be the
+ *       SMT ID (from smp_processor_id() or get_cpu()) divided by 2.
+ * @sub_unit: Hardware subunit this applies to (if applicable).
+ * @signal_group: Signal group to enable/disable on the trace bus.
+ * @bus_word: For signal groups that propagate via the trace bus, this trace
+ *            bus word will be used. This is a mask of (1 << TraceBusWord).
+ *            For other signal groups, this specifies the trigger or event bus.
+ * @bit: Trigger/Event bit, if applicable for the signal group.
+ *
+ * An array of these structures are passed to rtas_call() to set up the
+ * signals on the debug bus.
+ **/
+struct cell_rtas_arg {
+       u16 cpu;
+       u16 sub_unit;
+       s16 signal_group;
+       u8 bus_word;
+       u8 bit;
+};
+
+/**
+ * rtas_reset_signals
+ *
+ * Use the firmware RTAS call to disable signal pass-thru and to reset the
+ * debug-bus signals.
+ **/
+static int rtas_reset_signals(u32 cpu)
+{
+       struct cell_rtas_arg signal;
+       u64 real_addr = virt_to_phys(&signal);
+       int rc;
+       struct pfm_cell_platform_pmu_info *info =
+               ((struct pfm_arch_pmu_info *)
+                (pfm_pmu_conf->pmu_info))->platform_info;
+
+       memset(&signal, 0, sizeof(signal));
+       signal.cpu = RTAS_CPU(cpu);
+       rc = info->rtas_call(info->rtas_token("ibm,cbe-perftools"),
+                      5, 1, NULL,
+                      subfunc_RESET,
+                      passthru_DISABLE,
+                      real_addr >> 32,
+                      real_addr & 0xffffffff,
+                      sizeof(signal));
+
+       return rc;
+}
+
+/**
+ * rtas_activate_signals
+ *
+ * Use the firmware RTAS call to enable signal pass-thru and to activate the
+ * desired signal groups on the debug-bus.
+ **/
+static int rtas_activate_signals(struct cell_rtas_arg *signals,
+                                int num_signals)
+{
+       u64 real_addr = virt_to_phys(signals);
+       int rc;
+       struct pfm_cell_platform_pmu_info *info =
+               ((struct pfm_arch_pmu_info *)
+                (pfm_pmu_conf->pmu_info))->platform_info;
+
+       rc = info->rtas_call(info->rtas_token("ibm,cbe-perftools"),
+                      5, 1, NULL,
+                      subfunc_ACTIVATE,
+                      passthru_ENABLE,
+                      real_addr >> 32,
+                      real_addr & 0xffffffff,
+                      num_signals * sizeof(*signals));
+
+       return rc;
+}
+
+#define HID1_RESET_MASK                        (~0x00000001ffffffffUL)
+#define PPU_IU1_WORD0_HID1_EN_MASK     (~0x00000001f0c0802cUL)
+#define PPU_IU1_WORD0_HID1_EN_WORD     ( 0x00000001f0400000UL)
+#define PPU_IU1_WORD1_HID1_EN_MASK     (~0x000000010fc08023UL)
+#define PPU_IU1_WORD1_HID1_EN_WORD     ( 0x000000010f400001UL)
+#define PPU_XU_WORD0_HID1_EN_MASK      (~0x00000001f038402cUL)
+#define PPU_XU_WORD0_HID1_EN_WORD      ( 0x00000001f0080008UL)
+#define PPU_XU_WORD1_HID1_EN_MASK      (~0x000000010f074023UL)
+#define PPU_XU_WORD1_HID1_EN_WORD      ( 0x000000010f030002UL)
+
+/* The bus_word field in the cell_rtas_arg structure is a bit-mask
+ * indicating which debug-bus word(s) to use.
+ */
+enum {
+       BUS_WORD_0 = 1,
+       BUS_WORD_1 = 2,
+       BUS_WORD_2 = 4,
+       BUS_WORD_3 = 8,
+};
+
+/* Definitions of the signal-groups that the built-in signal-activation
+ * code can handle.
+ */
+enum {
+       SIG_GROUP_NONE = 0,
+
+       /* 2.x PowerPC Processor Unit (PPU) Signal Groups */
+       SIG_GROUP_PPU_BASE = 20,
+       SIG_GROUP_PPU_IU1 = 21,
+       SIG_GROUP_PPU_XU = 22,
+
+       /* 3.x PowerPC Storage Subsystem (PPSS) Signal Groups */
+       SIG_GROUP_PPSS_BASE = 30,
+
+       /* 4.x Synergistic Processor Unit (SPU) Signal Groups */
+       SIG_GROUP_SPU_BASE = 40,
+
+       /* 5.x Memory Flow Controller (MFC) Signal Groups */
+       SIG_GROUP_MFC_BASE = 50,
+
+       /* 6.x Element )nterconnect Bus (EIB) Signal Groups */
+       SIG_GROUP_EIB_BASE = 60,
+
+       /* 7.x Memory Interface Controller (MIC) Signal Groups */
+       SIG_GROUP_MIC_BASE = 70,
+
+       /* 8.x Cell Broadband Engine Interface (BEI) Signal Groups */
+       SIG_GROUP_BEI_BASE = 80,
+};
+
+/**
+ * rmw_spr
+ *
+ * Read-modify-write for a special-purpose-register.
+ **/
+#define rmw_spr(spr_id, a_mask, o_mask) \
+       do { \
+               u64 value = mfspr(spr_id); \
+               value &= (u64)(a_mask); \
+               value |= (u64)(o_mask); \
+               mtspr((spr_id), value); \
+       } while (0)
+
+/**
+ * rmw_mmio_reg64
+ *
+ * Read-modify-write for a 64-bit MMIO register.
+ **/
+#define rmw_mmio_reg64(mem, a_mask, o_mask) \
+       do { \
+               u64 value = in_be64(&(mem)); \
+               value &= (u64)(a_mask); \
+               value |= (u64)(o_mask); \
+               out_be64(&(mem), value); \
+       } while (0)
+
+/**
+ * rmwb_mmio_reg64
+ *
+ * Set or unset a specified bit within a 64-bit MMIO register.
+ **/
+#define rmwb_mmio_reg64(mem, bit_num, set_bit) \
+       rmw_mmio_reg64((mem), ~(1UL << (63 - (bit_num))), \
+                      ((set_bit) << (63 - (bit_num))))
+
+/**
+ * passthru
+ *
+ * Enable or disable passthru mode in all the Cell signal islands.
+ **/
+static int passthru(u32 cpu, u64 enable)
+{
+       struct cbe_ppe_priv_regs __iomem *ppe_priv_regs;
+       struct cbe_pmd_regs __iomem *pmd_regs;
+       struct cbe_mic_tm_regs __iomem *mic_tm_regs;
+       struct pfm_cell_platform_pmu_info *info =
+               ((struct pfm_arch_pmu_info *)
+                (pfm_pmu_conf->pmu_info))->platform_info;
+
+       ppe_priv_regs = info->get_cpu_ppe_priv_regs(cpu);
+       pmd_regs = info->get_cpu_pmd_regs(cpu);
+       mic_tm_regs = info->get_cpu_mic_tm_regs(cpu);
+
+       if (!ppe_priv_regs || !pmd_regs || !mic_tm_regs) {
+               PFM_ERR("Error getting Cell PPE, PMD, and MIC "
+                       "register maps: 0x%p, 0x%p, 0x%p",
+                       ppe_priv_regs, pmd_regs, mic_tm_regs);
+               return -EINVAL;
+       }
+
+       rmwb_mmio_reg64(ppe_priv_regs->L2_debug1, 61, enable);
+       rmwb_mmio_reg64(ppe_priv_regs->ciu_dr1, 5, enable);
+       rmwb_mmio_reg64(pmd_regs->on_ramp_trace, 39, enable);
+       rmwb_mmio_reg64(mic_tm_regs->MBL_debug, 20, enable);
+
+       return 0;
+}
+
+#define passthru_enable(cpu)  passthru(cpu, 1)
+#define passthru_disable(cpu) passthru(cpu, 0)
+
+static inline void reset_signal_registers(u32 cpu)
+{
+       rmw_spr(SPRN_HID1, HID1_RESET_MASK, 0);
+}
+
+/**
+ * celleb_reset_signals
+ *
+ * Non-rtas version of resetting the debug-bus signals.
+ **/
+static int celleb_reset_signals(u32 cpu)
+{
+       int rc;
+       rc = passthru_disable(cpu);
+       if (!rc)
+               reset_signal_registers(cpu);
+       return rc;
+}
+
+/**
+ * ppu_selection
+ *
+ * Write the HID1 register to connect the specified PPU signal-group to the
+ * debug-bus.
+ **/
+static int ppu_selection(struct cell_rtas_arg *signal)
+{
+       u64 hid1_enable_word = 0;
+       u64 hid1_enable_mask = 0;
+
+       switch (signal->signal_group) {
+
+       case SIG_GROUP_PPU_IU1: /* 2.1 PPU Instruction Unit - Group 1 */
+               switch (signal->bus_word) {
+               case BUS_WORD_0:
+                       hid1_enable_mask = PPU_IU1_WORD0_HID1_EN_MASK;
+                       hid1_enable_word = PPU_IU1_WORD0_HID1_EN_WORD;
+                       break;
+               case BUS_WORD_1:
+                       hid1_enable_mask = PPU_IU1_WORD1_HID1_EN_MASK;
+                       hid1_enable_word = PPU_IU1_WORD1_HID1_EN_WORD;
+                       break;
+               default:
+                       PFM_ERR("Invalid bus-word (0x%x) for signal-group %d.",
+                               signal->bus_word, signal->signal_group);
+                       return -EINVAL;
+               }
+               break;
+
+       case SIG_GROUP_PPU_XU:  /* 2.2 PPU Execution Unit */
+               switch (signal->bus_word) {
+               case BUS_WORD_0:
+                       hid1_enable_mask = PPU_XU_WORD0_HID1_EN_MASK;
+                       hid1_enable_word = PPU_XU_WORD0_HID1_EN_WORD;
+                       break;
+               case BUS_WORD_1:
+                       hid1_enable_mask = PPU_XU_WORD1_HID1_EN_MASK;
+                       hid1_enable_word = PPU_XU_WORD1_HID1_EN_WORD;
+                       break;
+               default:
+                       PFM_ERR("Invalid bus-word (0x%x) for signal-group %d.",
+                               signal->bus_word, signal->signal_group);
+                       return -EINVAL;
+               }
+               break;
+
+       default:
+               PFM_ERR("Signal-group %d not implemented.",
+                       signal->signal_group);
+               return -EINVAL;
+       }
+
+       rmw_spr(SPRN_HID1, hid1_enable_mask, hid1_enable_word);
+
+       return 0;
+}
+
+/**
+ * celleb_activate_signals
+ *
+ * Non-rtas version of activating the debug-bus signals.
+ **/
+static int celleb_activate_signals(struct cell_rtas_arg *signals,
+                                  int num_signals)
+{
+       int i, rc = -EINVAL;
+
+       for (i = 0; i < num_signals; i++) {
+               switch (signals[i].signal_group) {
+
+               /* 2.x PowerPC Processor Unit (PPU) Signal Selection */
+               case SIG_GROUP_PPU_IU1:
+               case SIG_GROUP_PPU_XU:
+                       rc = ppu_selection(signals + i);
+                       if (rc)
+                               return rc;
+                       break;
+
+               default:
+                       PFM_ERR("Signal-group %d not implemented.",
+                               signals[i].signal_group);
+                       return -EINVAL;
+               }
+       }
+
+       if (0 < i)
+               rc = passthru_enable(signals[0].cpu);
+
+       return rc;
+}
+
+/**
+ * ps3_reset_signals
+ *
+ * ps3 version of resetting the debug-bus signals.
+ **/
+static int ps3_reset_signals(u32 cpu)
+{
+#ifdef CONFIG_PPC_PS3
+       return ps3_set_signal(0, 0, 0, 0);
+#else
+       return 0;
+#endif
+}
+
+/**
+ * ps3_activate_signals
+ *
+ * ps3 version of activating the debug-bus signals.
+ **/
+static int ps3_activate_signals(struct cell_rtas_arg *signals,
+                               int num_signals)
+{
+#ifdef CONFIG_PPC_PS3
+       int i;
+
+       for (i = 0; i < num_signals; i++)
+               ps3_set_signal(signals[i].signal_group, signals[i].bit,
+                              signals[i].sub_unit, signals[i].bus_word);
+#endif
+       return 0;
+}
+
+
+/**
+ * reset_signals
+ *
+ * Call to the firmware (if available) to reset the debug-bus signals.
+ * Otherwise call the built-in version.
+ **/
+int reset_signals(u32 cpu)
+{
+       int rc;
+
+       if (machine_is(celleb))
+               rc = celleb_reset_signals(cpu);
+       else if (machine_is(ps3))
+               rc = ps3_reset_signals(cpu);
+       else
+               rc = rtas_reset_signals(cpu);
+
+       return rc;
+}
+
+/**
+ * activate_signals
+ *
+ * Call to the firmware (if available) to activate the debug-bus signals.
+ * Otherwise call the built-in version.
+ **/
+int activate_signals(struct cell_rtas_arg *signals, int num_signals)
+{
+       int rc;
+
+       if (machine_is(celleb))
+               rc = celleb_activate_signals(signals, num_signals);
+       else if (machine_is(ps3))
+               rc = ps3_activate_signals(signals, num_signals);
+       else
+               rc = rtas_activate_signals(signals, num_signals);
+
+       return rc;
+}
+
+/**
+ *  pfm_cell_pmc_check
+ *
+ * Verify that we are going to write a valid value to the specified PMC.
+ **/
+int pfm_cell_pmc_check(struct pfm_context *ctx,
+                      struct pfm_event_set *set,
+                      struct pfarg_pmc *req)
+{
+       u16 cnum, reg_num = req->reg_num;
+       s16 signal_group = RTAS_SIGNAL_GROUP(req->reg_value);
+       u8 bus_word = RTAS_BUS_WORD(req->reg_value);
+
+       if (reg_num < NR_CTRS || reg_num >= (NR_CTRS * 2))
+               return -EINVAL;
+
+       switch (signal_group) {
+       case SIG_GROUP_PPU_IU1:
+       case SIG_GROUP_PPU_XU:
+               if ((bus_word != 0) && (bus_word != 1)) {
+                       PFM_ERR("Invalid bus word (%d) for signal-group %d",
+                               bus_word, signal_group);
+                       return -EINVAL;
+               }
+               break;
+       default:
+               PFM_ERR("Signal-group %d not implemented.", signal_group);
+               return -EINVAL;
+       }
+
+       for (cnum = NR_CTRS; cnum < (NR_CTRS * 2); cnum++) {
+               if (test_bit(cnum, cast_ulp(set->used_pmcs)) &&
+                   bus_word == RTAS_BUS_WORD(set->pmcs[cnum]) &&
+                   signal_group != RTAS_SIGNAL_GROUP(set->pmcs[cnum])) {
+                       PFM_ERR("Impossible signal-group combination: "
+                               "(%u,%u,%d) (%u,%u,%d)",
+                               reg_num, bus_word, signal_group, cnum,
+                               RTAS_BUS_WORD(set->pmcs[cnum]),
+                               RTAS_SIGNAL_GROUP(set->pmcs[cnum]));
+                       return  -EBUSY;
+               }
+       }
+
+       return 0;
+}
+
+/**
+ * write_pm07_event
+ *
+ * Pull out the RTAS arguments from the 64-bit register value and make the
+ * RTAS activate-signals call.
+ **/
+static void write_pm07_event(int cpu, unsigned int ctr, u64 value)
+{
+       struct cell_rtas_arg signal;
+       s32 signal_number;
+       int rc;
+
+       signal_number = RTAS_SIGNAL_NUMBER(value);
+       if (!signal_number) {
+               /* Don't include counters that are counting cycles. */
+               return;
+       }
+
+       signal.cpu = RTAS_CPU(cpu);
+       signal.bus_word = 1 << RTAS_BUS_WORD(value);
+       signal.sub_unit = RTAS_SUB_UNIT(value);
+       signal.signal_group = signal_number / 100;
+       signal.bit = abs(signal_number) % 100;
+
+       rc = activate_signals(&signal, 1);
+       if (rc) {
+               PFM_WARN("%s(%d, %u, %lu): Error calling "
+                        "activate_signals(): %d\n", __func__,
+                        cpu, ctr, (unsigned long)value, rc);
+               /* FIX: Could we change this routine to return an error? */
+       }
+}
+
+/**
+ * pfm_cell_probe_pmu
+ *
+ * Simply check the processor version register to see if we're currently
+ * on a Cell system.
+ **/
+static int pfm_cell_probe_pmu(void)
+{
+       unsigned long pvr = mfspr(SPRN_PVR);
+
+       if (PVR_VER(pvr) != PV_BE)
+               return -1;
+
+       return 0;
+}
+
+/**
+ * pfm_cell_write_pmc
+ **/
+static void pfm_cell_write_pmc(unsigned int cnum, u64 value)
+{
+       int cpu = smp_processor_id();
+       struct pfm_cell_platform_pmu_info *info =
+               ((struct pfm_arch_pmu_info *)
+                (pfm_pmu_conf->pmu_info))->platform_info;
+
+       if (cnum < NR_CTRS) {
+               info->write_pm07_control(cpu, cnum, value);
+
+       } else if (cnum < NR_CTRS * 2) {
+               write_pm07_event(cpu, cnum - NR_CTRS, value);
+
+       } else if (cnum == CELL_PMC_PM_STATUS) {
+               /* The pm_status register must be treated separately from
+                * the other "global" PMCs. This call will ensure that
+                * the interrupts are routed to the correct CPU, as well
+                * as writing the desired value to the pm_status register.
+                */
+               info->enable_pm_interrupts(cpu, info->get_hw_thread_id(cpu),
+                                          value);
+
+       } else if (cnum < PFM_PM_NUM_PMCS) {
+               info->write_pm(cpu, cnum - (NR_CTRS * 2), value);
+       }
+}
+
+/**
+ * pfm_cell_write_pmd
+ **/
+static void pfm_cell_write_pmd(unsigned int cnum, u64 value)
+{
+       int cpu = smp_processor_id();
+       struct pfm_cell_platform_pmu_info *info =
+               ((struct pfm_arch_pmu_info *)
+                (pfm_pmu_conf->pmu_info))->platform_info;
+
+       if (cnum < NR_CTRS)
+               info->write_ctr(cpu, cnum, value);
+}
+
+/**
+ * pfm_cell_read_pmd
+ **/
+static u64 pfm_cell_read_pmd(unsigned int cnum)
+{
+       int cpu = smp_processor_id();
+       struct pfm_cell_platform_pmu_info *info =
+               ((struct pfm_arch_pmu_info *)
+                (pfm_pmu_conf->pmu_info))->platform_info;
+
+       if (cnum < NR_CTRS)
+               return info->read_ctr(cpu, cnum);
+
+       return -EINVAL;
+}
+
+/**
+ * pfm_cell_enable_counters
+ *
+ * Just need to turn on the global disable bit in pm_control.
+ **/
+static void pfm_cell_enable_counters(struct pfm_context *ctx,
+                                    struct pfm_event_set *set)
+{
+       struct pfm_cell_platform_pmu_info *info =
+               ((struct pfm_arch_pmu_info *)
+                (pfm_pmu_conf->pmu_info))->platform_info;
+
+       info->enable_pm(smp_processor_id());
+}
+
+/**
+ * pfm_cell_disable_counters
+ *
+ * Just need to turn off the global disable bit in pm_control.
+ **/
+static void pfm_cell_disable_counters(struct pfm_context *ctx,
+                                     struct pfm_event_set *set)
+{
+       struct pfm_cell_platform_pmu_info *info =
+               ((struct pfm_arch_pmu_info *)
+                (pfm_pmu_conf->pmu_info))->platform_info;
+
+       info->disable_pm(smp_processor_id());
+       if (machine_is(ps3))
+               reset_signals(smp_processor_id());
+}
+
+/*
+ * Return the thread id of the specified ppu signal.
+ */
+static inline u32 get_target_ppu_thread_id(u32 group, u32 bit)
+{
+       if ((group == SIG_GROUP_PPU_IU1 &&
+            bit < PFM_PPU_IU1_THREAD1_BASE_BIT) ||
+           (group == SIG_GROUP_PPU_XU &&
+            bit < PFM_PPU_XU_THREAD1_BASE_BIT))
+               return 0;
+       else
+               return 1;
+}
+
+/*
+ * Return whether the specified counter is for PPU signal group.
+ */
+static inline int is_counter_for_ppu_sig_grp(u32 counter_control, u32 sig_grp)
+{
+       if (!(counter_control & CBE_PM_CTR_INPUT_CONTROL) &&
+           (counter_control & CBE_PM_CTR_ENABLE) &&
+           ((sig_grp == SIG_GROUP_PPU_IU1) || (sig_grp == SIG_GROUP_PPU_XU)))
+               return 1;
+       else
+               return 0;
+}
+
+/*
+ * Search ppu signal groups.
+ */
+static int get_ppu_signal_groups(struct pfm_event_set *set,
+                                u32 *ppu_sig_grp0, u32 *ppu_sig_grp1)
+{
+       u64 pm_event, *used_pmcs = set->used_pmcs;
+       int i, j;
+       u32 grp0_wd, grp1_wd, wd, sig_grp;
+
+       *ppu_sig_grp0 = 0;
+       *ppu_sig_grp1 = 0;
+       grp0_wd = PFM_GROUP_CONTROL_GROUP0_WORD(
+               set->pmcs[CELL_PMC_GROUP_CONTROL]);
+       grp1_wd = PFM_GROUP_CONTROL_GROUP1_WORD(
+               set->pmcs[CELL_PMC_GROUP_CONTROL]);
+
+       for (i = 0, j = 0; (i < NR_CTRS) && (j < PFM_NUM_OF_GROUPS); i++) {
+               if (test_bit(i + NR_CTRS, used_pmcs)) {
+                       pm_event = set->pmcs[i + NR_CTRS];
+                       wd = PFM_EVENT_PMC_BUS_WORD(pm_event);
+                       sig_grp = PFM_EVENT_PMC_SIGNAL_GROUP(pm_event);
+                       if ((sig_grp == SIG_GROUP_PPU_IU1) ||
+                           (sig_grp == SIG_GROUP_PPU_XU)) {
+
+                               if (wd == grp0_wd && *ppu_sig_grp0 == 0) {
+                                       *ppu_sig_grp0 = sig_grp;
+                                       j++;
+                               } else if (wd == grp1_wd &&
+                                          *ppu_sig_grp1 == 0) {
+                                       *ppu_sig_grp1 = sig_grp;
+                                       j++;
+                               }
+                       }
+               }
+       }
+       return j;
+}
+
+/**
+ * pfm_cell_restore_pmcs
+ *
+ * Write all control register values that are saved in the specified event
+ * set. We could use the pfm_arch_write_pmc() function to restore each PMC
+ * individually (as is done in other architectures), but that results in
+ * multiple RTAS calls. As an optimization, we will setup the RTAS argument
+ * array so we can do all event-control registers in one RTAS call.
+ *
+ * In per-thread mode,
+ * The counter enable bit of the pmX_control PMC is enabled while the target
+ * task runs on the target HW thread.
+ **/
+void pfm_cell_restore_pmcs(struct pfm_context *ctx, struct pfm_event_set *set)
+{
+       u64 ctr_ctrl;
+       u64 *used_pmcs = set->used_pmcs;
+       int i;
+       int cpu = smp_processor_id();
+       u32 current_th_id;
+       struct pfm_cell_platform_pmu_info *info =
+               ((struct pfm_arch_pmu_info *)
+                (pfm_pmu_conf->pmu_info))->platform_info;
+
+       for (i = 0; i < NR_CTRS; i++) {
+               ctr_ctrl = set->pmcs[i];
+
+               if (ctr_ctrl & PFM_COUNTER_CTRL_PMC_PPU_TH0) {
+                       current_th_id = info->get_hw_thread_id(cpu);
+
+                       /*
+                        * Set the counter enable bit down if the current
+                        * HW thread is NOT 0
+                        **/
+                       if (current_th_id)
+                               ctr_ctrl = ctr_ctrl & ~CBE_PM_CTR_ENABLE;
+
+               } else if (ctr_ctrl & PFM_COUNTER_CTRL_PMC_PPU_TH1) {
+                       current_th_id = info->get_hw_thread_id(cpu);
+
+                       /*
+                        * Set the counter enable bit down if the current
+                        * HW thread is 0
+                        **/
+                       if (!current_th_id)
+                               ctr_ctrl = ctr_ctrl & ~CBE_PM_CTR_ENABLE;
+               }
+
+               /* Write the per-counter control register. If the PMC is not
+                * in use, then it will simply clear the register, which will
+                * disable the associated counter.
+                */
+               info->write_pm07_control(cpu, i, ctr_ctrl);
+
+               if (test_bit(i + NR_CTRS, used_pmcs))
+                       write_pm07_event(cpu, 0, set->pmcs[i + NR_CTRS]);
+       }
+
+       /* Write all the global PMCs. Need to call pfm_cell_write_pmc()
+        * instead of cbe_write_pm() due to special handling for the
+        * pm_status register.
+        */
+       for (i *= 2; i < PFM_PM_NUM_PMCS; i++)
+               pfm_cell_write_pmc(i, set->pmcs[i]);
+}
+
+/**
+ * pfm_cell_restore_pmds
+ *
+ * Write to pm_control register before writing to counter registers
+ * so that we can decide the counter width berfore writing to the couters.
+ **/
+void pfm_cell_restore_pmds(struct pfm_context *ctx, struct pfm_event_set *set)
+{
+       u64 *used_pmds;
+       unsigned int i, max_pmd;
+       int cpu = smp_processor_id();
+       struct pfm_cell_platform_pmu_info *info =
+               ((struct pfm_arch_pmu_info *)
+                (pfm_pmu_conf->pmu_info))->platform_info;
+
+       /*
+        * Write pm_control register value
+        */
+       info->write_pm(cpu, pm_control,
+                      set->pmcs[CELL_PMC_PM_CONTROL] &
+                      ~CBE_PM_ENABLE_PERF_MON);
+       PFM_DBG("restore pm_control(0x%lx) before restoring pmds",
+               set->pmcs[CELL_PMC_PM_CONTROL]);
+
+       max_pmd = ctx->regs.max_pmd;
+       used_pmds = set->used_pmds;
+
+       for (i = 0; i < max_pmd; i++)
+               if (test_bit(i, used_pmds) &&
+                   !(pfm_pmu_conf->pmd_desc[i].type & PFM_REG_RO))
+                       pfm_cell_write_pmd(i, set->pmds[i].value);
+}
+
+/**
+ * pfm_cell_get_cntr_width
+ *
+ * This function check the 16bit counter field in pm_control pmc.
+ *
+ * Return value
+ *  16 : all counters are 16bit width.
+ *  32 : all counters are 32bit width.
+ *  0  : several counter width exists.
+ **/
+static int pfm_cell_get_cntr_width(struct pfm_context *ctx,
+                                  struct pfm_event_set *s)
+{
+       int width = 0;
+       int tmp = 0;
+       u64 cntr_field;
+
+       if (ctx->flags.switch_ovfl || ctx->flags.switch_time) {
+               list_for_each_entry(s, &ctx->set_list, list) {
+                       cntr_field = s->pmcs[CELL_PMC_PM_CONTROL] &
+                               CELL_PMC_PM_CONTROL_CNTR_MASK;
+
+                       if (cntr_field == CELL_PMC_PM_CONTROL_CNTR_16)
+                               tmp = 16;
+                       else if (cntr_field == 0x0)
+                               tmp = 32;
+                       else
+                               return 0;
+
+                       if (tmp != width && width != 0)
+                               return 0;
+
+                       width = tmp;
+               }
+       } else {
+               cntr_field = s->pmcs[CELL_PMC_PM_CONTROL] &
+                       CELL_PMC_PM_CONTROL_CNTR_MASK;
+
+               if (cntr_field == CELL_PMC_PM_CONTROL_CNTR_16)
+                       width = 16;
+               else if (cntr_field == 0x0)
+                       width = 32;
+               else
+                       width = 0;
+       }
+       return width;
+}
+
+/**
+ * pfm_cell_check_cntr_ovfl_mask
+ *
+ * Return value
+ *  1  : cntr_ovfl interrupt is used.
+ *  0  : cntr_ovfl interrupt is not used.
+ **/
+static int pfm_cell_check_cntr_ovfl(struct pfm_context *ctx,
+                                   struct pfm_event_set *s)
+{
+       if (ctx->flags.switch_ovfl || ctx->flags.switch_time) {
+               list_for_each_entry(s, &ctx->set_list, list) {
+                       if (CBE_PM_OVERFLOW_CTRS(s->pmcs[CELL_PMC_PM_STATUS]))
+                               return 1;
+               }
+       } else {
+               if (CBE_PM_OVERFLOW_CTRS(s->pmcs[CELL_PMC_PM_STATUS]))
+                       return 1;
+       }
+       return 0;
+}
+
+#ifdef CONFIG_PPC_PS3
+/**
+ * update_sub_unit_field
+ *
+ **/
+static inline u64 update_sub_unit_field(u64 pm_event, u64 spe_id)
+{
+       return ((pm_event & 0xFFFF0000FFFFFFFF) | (spe_id << 32));
+}
+
+/**
+ * pfm_get_spe_id
+ *
+ **/
+static u64 pfm_get_spe_id(void *arg)
+{
+       struct spu *spu = arg;
+       u64 spe_id;
+
+       if (machine_is(ps3))
+               spe_id = ps3_get_spe_id(arg);
+       else
+               spe_id = spu->spe_id;
+
+       return spe_id;
+}
+
+/**
+ * pfm_spu_number_to_id
+ *
+ **/
+static int pfm_spu_number_to_id(int number, u64 *spe_id)
+{
+       struct spu *spu;
+       int i;
+
+       for (i = 0; i < MAX_NUMNODES; i++) {
+               if (cbe_spu_info[i].n_spus == 0)
+                       continue;
+
+               list_for_each_entry(spu, &cbe_spu_info[i].spus, cbe_list)
+                       if (spu->number == number) {
+                               *spe_id = pfm_get_spe_id(spu);
+                               return 0;
+                       }
+       }
+       return -ENODEV;
+}
+
+/**
+ * pfm_update_pmX_event_subunit_field
+ *
+ * In system wide mode,
+ * This function updates the subunit field of SPE pmX_event.
+ **/
+static int pfm_update_pmX_event_subunit_field(struct pfm_context *ctx)
+{
+       struct pfm_event_set *set;
+       int i, last_pmc, ret;
+       u64 signal_group, spe_id;
+       int sub_unit;
+       u64 *used_pmcs;
+
+       last_pmc = NR_CTRS + 8;
+       ret = 0;
+       list_for_each_entry(set, &ctx->set_list, list) {
+
+               used_pmcs = set->used_pmcs;
+               for (i = NR_CTRS; i < last_pmc; i++) {
+                       if (!test_bit(i, used_pmcs))
+                               continue;
+
+                       signal_group = PFM_EVENT_PMC_SIGNAL_GROUP(set->pmcs[i]);
+
+                       /*
+                        * If the target event is a SPE signal group event,
+                        * The sub_unit field in pmX_event pmc is changed to the
+                        * specified spe_id.
+                        */
+                       if (SIG_GROUP_SPU_BASE < signal_group &&
+                           signal_group < SIG_GROUP_EIB_BASE) {
+                               sub_unit = RTAS_SUB_UNIT(set->pmcs[i]);
+
+                               ret = pfm_spu_number_to_id(sub_unit, &spe_id);
+                               if (ret)
+                                       return ret;
+
+                               set->pmcs[i] = update_sub_unit_field(
+                                       set->pmcs[i], spe_id);
+                       }
+               }
+       }
+       return 0;
+}
+#endif
+
+/**
+ * pfm_cell_load_context
+ *
+ * In per-thread mode,
+ *  The pmX_control PMCs which are used for PPU IU/XU event are marked with
+ *  the thread id(PFM_COUNTER_CTRL_PMC_PPU_TH0/TH1).
+ **/
+static int pfm_cell_load_context(struct pfm_context *ctx)
+{
+       int i;
+       u32 ppu_sig_grp[PFM_NUM_OF_GROUPS] = {SIG_GROUP_NONE, SIG_GROUP_NONE};
+       u32 bit;
+       int index;
+       u32 target_th_id;
+       int ppu_sig_num = 0;
+       struct pfm_event_set *s;
+       int cntr_width = 32;
+       int ret = 0;
+
+       if (pfm_cell_check_cntr_ovfl(ctx, ctx->active_set)) {
+               cntr_width = pfm_cell_get_cntr_width(ctx, ctx->active_set);
+
+               /*
+                * Counter overflow interrupt works with only 32bit counter,
+                * because perfmon core uses pfm_cell_pmu_conf.counter_width
+                * to deal with the counter overflow. we can't change the
+                * counter width here.
+                */
+               if (cntr_width != 32)
+                       return -EINVAL;
+       }
+
+       if (ctx->flags.system) {
+#ifdef CONFIG_PPC_PS3
+               if (machine_is(ps3))
+                       ret = pfm_update_pmX_event_subunit_field(ctx);
+#endif
+               return ret;
+       }
+
+       list_for_each_entry(s, &ctx->set_list, list) {
+               ppu_sig_num = get_ppu_signal_groups(s, &ppu_sig_grp[0],
+                                                   &ppu_sig_grp[1]);
+
+               for (i = 0; i < NR_CTRS; i++) {
+                       index = PFM_PM_CTR_INPUT_MUX_GROUP_INDEX(s->pmcs[i]);
+                       if (ppu_sig_num &&
+                           (ppu_sig_grp[index] != SIG_GROUP_NONE) &&
+                           is_counter_for_ppu_sig_grp(s->pmcs[i],
+                                                      ppu_sig_grp[index])) {
+
+                               bit = PFM_PM_CTR_INPUT_MUX_BIT(s->pmcs[i]);
+                               target_th_id = get_target_ppu_thread_id(
+                                       ppu_sig_grp[index], bit);
+                               if (!target_th_id)
+                                       s->pmcs[i] |=
+                                               PFM_COUNTER_CTRL_PMC_PPU_TH0;
+                               else
+                                       s->pmcs[i] |=
+                                               PFM_COUNTER_CTRL_PMC_PPU_TH1;
+                               PFM_DBG("set:%d mark ctr:%d target_thread:%d",
+                                       s->id, i, target_th_id);
+                       }
+               }
+       }
+
+       return ret;
+}
+
+/**
+ * pfm_cell_unload_context
+ *
+ * For system-wide contexts and self-monitored contexts, make the RTAS call
+ * to reset the debug-bus signals.
+ *
+ * For non-self-monitored contexts, the monitored thread will already have
+ * been taken off the CPU and we don't need to do anything additional.
+ **/
+static void pfm_cell_unload_context(struct pfm_context *ctx)
+{
+       if (ctx->task == current || ctx->flags.system)
+               reset_signals(smp_processor_id());
+}
+
+/**
+ * pfm_cell_ctxswout_thread
+ *
+ * When a monitored thread is switched out (self-monitored or externally
+ * monitored) we need to reset the debug-bus signals so the next context that
+ * gets switched in can start from a clean set of signals.
+ **/
+int pfm_cell_ctxswout_thread(struct task_struct *task,
+                            struct pfm_context *ctx, struct pfm_event_set *set)
+{
+       reset_signals(smp_processor_id());
+       return 0;
+}
+
+/**
+ * pfm_cell_get_ovfl_pmds
+ *
+ * Determine which counters in this set have overflowed and fill in the
+ * set->povfl_pmds mask and set->npend_ovfls count. On Cell, the pm_status
+ * register contains a bit for each counter to indicate overflow. However,
+ * those 8 bits are in the reverse order than what Perfmon2 is expecting,
+ * so we need to reverse the order of the overflow bits.
+ **/
+static void pfm_cell_get_ovfl_pmds(struct pfm_context *ctx,
+                                  struct pfm_event_set *set)
+{
+       struct pfm_arch_context *ctx_arch = pfm_ctx_arch(ctx);
+       u32 pm_status, ovfl_ctrs;
+       u64 povfl_pmds = 0;
+       int i;
+       struct pfm_cell_platform_pmu_info *info =
+               ((struct pfm_arch_pmu_info *)
+                (pfm_pmu_conf->pmu_info))->platform_info;
+
+       if (!ctx_arch->last_read_updated)
+               /* This routine was not called via the interrupt handler.
+                * Need to start by getting interrupts and updating
+                * last_read_pm_status.
+                */
+               ctx_arch->last_read_pm_status =
+ info->get_and_clear_pm_interrupts(smp_processor_id());
+
+       /* Reset the flag that the interrupt handler last read pm_status. */
+       ctx_arch->last_read_updated = 0;
+
+       pm_status = ctx_arch->last_read_pm_status &
+                   set->pmcs[CELL_PMC_PM_STATUS];
+       ovfl_ctrs = CBE_PM_OVERFLOW_CTRS(pm_status);
+
+       /* Reverse the order of the bits in ovfl_ctrs
+        * and store the result in povfl_pmds.
+        */
+       for (i = 0; i < PFM_PM_NUM_PMDS; i++) {
+               povfl_pmds = (povfl_pmds << 1) | (ovfl_ctrs & 1);
+               ovfl_ctrs >>= 1;
+       }
+
+       /* Mask povfl_pmds with set->used_pmds to get set->povfl_pmds.
+        * Count the bits set in set->povfl_pmds to get set->npend_ovfls.
+        */
+       bitmap_and(set->povfl_pmds, &povfl_pmds,
+                  set->used_pmds, PFM_PM_NUM_PMDS);
+       set->npend_ovfls = bitmap_weight(set->povfl_pmds, PFM_PM_NUM_PMDS);
+}
+
+/**
+ * pfm_cell_acquire_pmu
+ *
+ * acquire PMU resource.
+ * This acquisition is done when the first context is created.
+ **/
+int pfm_cell_acquire_pmu(u64 *unavail_pmcs, u64 *unavail_pmds)
+{
+#ifdef CONFIG_PPC_PS3
+       int ret;
+
+       if (machine_is(ps3)) {
+               PFM_DBG("");
+               ret = ps3_lpm_open(PS3_LPM_TB_TYPE_INTERNAL, NULL, 0);
+               if (ret) {
+                       PFM_ERR("Can't create PS3 lpm. error:%d", ret);
+                       return -EFAULT;
+               }
+       }
+#endif
+       return 0;
+}
+
+/**
+ * pfm_cell_release_pmu
+ *
+ * release PMU resource.
+ * actual release happens when last context is destroyed
+ **/
+void pfm_cell_release_pmu(void)
+{
+#ifdef CONFIG_PPC_PS3
+       if (machine_is(ps3)) {
+               if (ps3_lpm_close())
+                       PFM_ERR("Can't delete PS3 lpm.");
+       }
+#endif
+}
+
+/**
+ * handle_trace_buffer_interrupts
+ *
+ * This routine is for processing just the interval timer and trace buffer
+ * overflow interrupts. Performance counter interrupts are handled by the
+ * perf_irq_handler() routine, which reads and saves the pm_status register.
+ * This routine should not read the actual pm_status register, but rather
+ * the value passed in.
+ **/
+static void handle_trace_buffer_interrupts(unsigned long iip,
+                                          struct pt_regs *regs,
+                                          struct pfm_context *ctx,
+                                          u32 pm_status)
+{
+       /* FIX: Currently ignoring trace-buffer interrupts. */
+       return;
+}
+
+/**
+ * pfm_cell_irq_handler
+ *
+ * Handler for all Cell performance-monitor interrupts.
+ **/
+static void pfm_cell_irq_handler(struct pt_regs *regs, struct pfm_context *ctx)
+{
+       struct pfm_arch_context *ctx_arch = pfm_ctx_arch(ctx);
+       u32 last_read_pm_status;
+       int cpu = smp_processor_id();
+       struct pfm_cell_platform_pmu_info *info =
+               ((struct pfm_arch_pmu_info *)
+                (pfm_pmu_conf->pmu_info))->platform_info;
+
+       /* Need to disable and reenable the performance counters to get the
+        * desired behavior from the hardware. This is specific to the Cell
+        * PMU hardware.
+        */
+       info->disable_pm(cpu);
+
+       /* Read the pm_status register to get the interrupt bits. If a
+        * perfmormance counter overflow interrupt occurred, call the core
+        * perfmon interrupt handler to service the counter overflow. If the
+        * interrupt was for the interval timer or the trace_buffer,
+        * call the interval timer and trace buffer interrupt handler.
+        *
+        * The value read from the pm_status register is stored in the
+        * pmf_arch_context structure for use by other routines. Note that
+        * reading the pm_status register resets the interrupt flags to zero.
+        * Hence, it is important that the register is only read in one place.
+        *
+        * The pm_status reg interrupt reg format is:
+        * [pmd0:pmd1:pmd2:pmd3:pmd4:pmd5:pmd6:pmd7:intt:tbf:tbu:]
+        * - pmd0 to pm7 are the perf counter overflow interrupts.
+        * - intt is the interval timer overflowed interrupt.
+        * - tbf is the trace buffer full interrupt.
+        * - tbu is the trace buffer underflow interrupt.
+        * - The pmd0 bit is the MSB of the 32 bit register.
+        */
+       ctx_arch->last_read_pm_status = last_read_pm_status =
+                       info->get_and_clear_pm_interrupts(cpu);
+
+       /* Set flag for pfm_cell_get_ovfl_pmds() routine so it knows
+        * last_read_pm_status was updated by the interrupt handler.
+        */
+       ctx_arch->last_read_updated = 1;
+
+       if (last_read_pm_status & CBE_PM_ALL_OVERFLOW_INTR)
+               /* At least one counter overflowed. */
+               pfm_interrupt_handler(instruction_pointer(regs), regs);
+
+       if (last_read_pm_status & (CBE_PM_INTERVAL_INTR |
+                                  CBE_PM_TRACE_BUFFER_FULL_INTR |
+                                  CBE_PM_TRACE_BUFFER_UNDERFLOW_INTR))
+               /* Trace buffer or interval timer overflow. */
+               handle_trace_buffer_interrupts(instruction_pointer(regs),
+                                              regs, ctx, last_read_pm_status);
+
+       /* The interrupt settings is the value written to the pm_status
+        * register. It is saved in the context when the register is
+        * written.
+        */
+       info->enable_pm_interrupts(cpu, info->get_hw_thread_id(cpu),
+       ctx->active_set->pmcs[CELL_PMC_PM_STATUS]);
+
+       /* The writes to the various performance counters only writes to a
+        * latch. The new values (interrupt setting bits, reset counter value
+        * etc.) are not copied to the actual registers until the performance
+        * monitor is enabled. In order to get this to work as desired, the
+        * permormance monitor needs to be disabled while writting to the
+        * latches. This is a HW design issue.
+        */
+       info->enable_pm(cpu);
+}
+
+
+static struct pfm_cell_platform_pmu_info ps3_platform_pmu_info = {
+#ifdef CONFIG_PPC_PS3
+       .read_ctr                    = ps3_read_ctr,
+       .write_ctr                   = ps3_write_ctr,
+       .write_pm07_control          = ps3_write_pm07_control,
+       .write_pm                    = ps3_write_pm,
+       .enable_pm                   = ps3_enable_pm,
+       .disable_pm                  = ps3_disable_pm,
+       .enable_pm_interrupts        = ps3_enable_pm_interrupts,
+       .get_and_clear_pm_interrupts = ps3_get_and_clear_pm_interrupts,
+       .get_hw_thread_id            = ps3_get_hw_thread_id,
+       .get_cpu_ppe_priv_regs       = NULL,
+       .get_cpu_pmd_regs            = NULL,
+       .get_cpu_mic_tm_regs         = NULL,
+       .rtas_token                  = NULL,
+       .rtas_call                   = NULL,
+#endif
+};
+
+static struct pfm_cell_platform_pmu_info native_platform_pmu_info = {
+#ifdef CONFIG_PPC_CELL_NATIVE
+       .read_ctr                    = cbe_read_ctr,
+       .write_ctr                   = cbe_write_ctr,
+       .write_pm07_control          = cbe_write_pm07_control,
+       .write_pm                    = cbe_write_pm,
+       .enable_pm                   = cbe_enable_pm,
+       .disable_pm                  = cbe_disable_pm,
+       .enable_pm_interrupts        = cbe_enable_pm_interrupts,
+       .get_and_clear_pm_interrupts = cbe_get_and_clear_pm_interrupts,
+       .get_hw_thread_id            = cbe_get_hw_thread_id,
+       .get_cpu_ppe_priv_regs       = cbe_get_cpu_ppe_priv_regs,
+       .get_cpu_pmd_regs            = cbe_get_cpu_pmd_regs,
+       .get_cpu_mic_tm_regs         = cbe_get_cpu_mic_tm_regs,
+       .rtas_token                  = rtas_token,
+       .rtas_call                   = rtas_call,
+#endif
+};
+
+static struct pfm_arch_pmu_info pfm_cell_pmu_info = {
+       .pmu_style        = PFM_POWERPC_PMU_CELL,
+       .acquire_pmu      = pfm_cell_acquire_pmu,
+       .release_pmu      = pfm_cell_release_pmu,
+       .write_pmc        = pfm_cell_write_pmc,
+       .write_pmd        = pfm_cell_write_pmd,
+       .read_pmd         = pfm_cell_read_pmd,
+       .enable_counters  = pfm_cell_enable_counters,
+       .disable_counters = pfm_cell_disable_counters,
+       .irq_handler      = pfm_cell_irq_handler,
+       .get_ovfl_pmds    = pfm_cell_get_ovfl_pmds,
+       .restore_pmcs     = pfm_cell_restore_pmcs,
+       .restore_pmds     = pfm_cell_restore_pmds,
+       .ctxswout_thread  = pfm_cell_ctxswout_thread,
+       .load_context     = pfm_cell_load_context,
+       .unload_context   = pfm_cell_unload_context,
+};
+
+static struct pfm_pmu_config pfm_cell_pmu_conf = {
+       .pmu_name = "Cell",
+       .version = "0.1",
+       .counter_width = 32,
+       .pmd_desc = pfm_cell_pmd_desc,
+       .pmc_desc = pfm_cell_pmc_desc,
+       .num_pmc_entries = PFM_PM_NUM_PMCS,
+       .num_pmd_entries = PFM_PM_NUM_PMDS,
+       .probe_pmu  = pfm_cell_probe_pmu,
+       .pmu_info = &pfm_cell_pmu_info,
+       .flags = PFM_PMU_BUILTIN_FLAG,
+       .owner = THIS_MODULE,
+};
+
+/**
+ * pfm_cell_platform_probe
+ *
+ * If we're on a system without the firmware rtas call available, set up the
+ * PMC write-checker for all the pmX_event control registers.
+ **/
+static void pfm_cell_platform_probe(void)
+{
+       if (machine_is(celleb)) {
+               int cnum;
+               pfm_cell_pmu_conf.pmc_write_check = pfm_cell_pmc_check;
+               for (cnum = NR_CTRS; cnum < (NR_CTRS * 2); cnum++)
+                       pfm_cell_pmc_desc[cnum].type |= PFM_REG_WC;
+       }
+
+       if (machine_is(ps3))
+               pfm_cell_pmu_info.platform_info = &ps3_platform_pmu_info;
+       else
+               pfm_cell_pmu_info.platform_info = &native_platform_pmu_info;
+}
+
+static int __init pfm_cell_pmu_init_module(void)
+{
+       pfm_cell_platform_probe();
+       return pfm_pmu_register(&pfm_cell_pmu_conf);
+}
+
+static void __exit pfm_cell_pmu_cleanup_module(void)
+{
+       pfm_pmu_unregister(&pfm_cell_pmu_conf);
+}
+
+module_init(pfm_cell_pmu_init_module);
+module_exit(pfm_cell_pmu_cleanup_module);
diff --git a/arch/powerpc/perfmon/perfmon_power4.c b/arch/powerpc/perfmon/perfmon_power4.c
new file mode 100644
index 0000000..eba9e8c
--- /dev/null
+++ b/arch/powerpc/perfmon/perfmon_power4.c
@@ -0,0 +1,309 @@
+/*
+ * This file contains the POWER4 PMU register description tables
+ * and pmc checker used by perfmon.c.
+ *
+ * Copyright (c) 2007, IBM Corporation.
+ *
+ * Based on a simple modification of perfmon_power5.c for POWER4 by
+ * Corey Ashford <cjashfor@us.ibm.com>.
+ *
+ * This program is free software; you can redistribute it and/or
+ * modify it under the terms of version 2 of the GNU General Public
+ * License as published by the Free Software Foundation.
+ *
+ * This program is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+ * General Public License for more details.
+ *
+ * You should have received a copy of the GNU General Public License
+ * along with this program; if not, write to the Free Software
+ * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA
+ * 02111-1307 USA
+ */
+#include <linux/module.h>
+#include <linux/perfmon_kern.h>
+
+MODULE_AUTHOR("Corey Ashford <cjashfor@us.ibm.com>");
+MODULE_DESCRIPTION("POWER4 PMU description table");
+MODULE_LICENSE("GPL");
+
+static struct pfm_regmap_desc pfm_power4_pmc_desc[] = {
+/* mmcr0 */ PMC_D(PFM_REG_I, "MMCR0", MMCR0_FC, 0, 0, SPRN_MMCR0),
+/* mmcr1 */ PMC_D(PFM_REG_I, "MMCR1", 0, 0, 0, SPRN_MMCR1),
+/* mmcra */ PMC_D(PFM_REG_I, "MMCRA", 0, 0, 0, SPRN_MMCRA)
+};
+#define PFM_PM_NUM_PMCS        ARRAY_SIZE(pfm_power4_pmc_desc)
+
+/* The TB and PURR registers are read-only. Also, note that the TB register
+ * actually consists of both the 32-bit SPRN_TBRU and SPRN_TBRL registers.
+ * For Perfmon2's purposes, we'll treat it as a single 64-bit register.
+ */
+static struct pfm_regmap_desc pfm_power4_pmd_desc[] = {
+/* tb    */ PMD_D((PFM_REG_I|PFM_REG_RO), "TB", SPRN_TBRL),
+/* pmd1  */ PMD_D(PFM_REG_C, "PMC1", SPRN_PMC1),
+/* pmd2  */ PMD_D(PFM_REG_C, "PMC2", SPRN_PMC2),
+/* pmd3  */ PMD_D(PFM_REG_C, "PMC3", SPRN_PMC3),
+/* pmd4  */ PMD_D(PFM_REG_C, "PMC4", SPRN_PMC4),
+/* pmd5  */ PMD_D(PFM_REG_C, "PMC5", SPRN_PMC5),
+/* pmd6  */ PMD_D(PFM_REG_C, "PMC6", SPRN_PMC6),
+/* pmd7  */ PMD_D(PFM_REG_C, "PMC7", SPRN_PMC7),
+/* pmd8  */ PMD_D(PFM_REG_C, "PMC8", SPRN_PMC8)
+};
+#define PFM_PM_NUM_PMDS        ARRAY_SIZE(pfm_power4_pmd_desc)
+
+static int pfm_power4_probe_pmu(void)
+{
+       unsigned long pvr = mfspr(SPRN_PVR);
+       int ver = PVR_VER(pvr);
+
+       if ((ver == PV_POWER4) || (ver == PV_POWER4p))
+               return 0;
+
+       return -1;
+}
+
+static void pfm_power4_write_pmc(unsigned int cnum, u64 value)
+{
+       switch (pfm_pmu_conf->pmc_desc[cnum].hw_addr) {
+       case SPRN_MMCR0:
+               mtspr(SPRN_MMCR0, value);
+               break;
+       case SPRN_MMCR1:
+               mtspr(SPRN_MMCR1, value);
+               break;
+       case SPRN_MMCRA:
+               mtspr(SPRN_MMCRA, value);
+               break;
+       default:
+               BUG();
+       }
+}
+
+static void pfm_power4_write_pmd(unsigned int cnum, u64 value)
+{
+       u64 ovfl_mask = pfm_pmu_conf->ovfl_mask;
+
+       switch (pfm_pmu_conf->pmd_desc[cnum].hw_addr) {
+       case SPRN_PMC1:
+               mtspr(SPRN_PMC1, value & ovfl_mask);
+               break;
+       case SPRN_PMC2:
+               mtspr(SPRN_PMC2, value & ovfl_mask);
+               break;
+       case SPRN_PMC3:
+               mtspr(SPRN_PMC3, value & ovfl_mask);
+               break;
+       case SPRN_PMC4:
+               mtspr(SPRN_PMC4, value & ovfl_mask);
+               break;
+       case SPRN_PMC5:
+               mtspr(SPRN_PMC5, value & ovfl_mask);
+               break;
+       case SPRN_PMC6:
+               mtspr(SPRN_PMC6, value & ovfl_mask);
+               break;
+       case SPRN_PMC7:
+               mtspr(SPRN_PMC7, value & ovfl_mask);
+               break;
+       case SPRN_PMC8:
+               mtspr(SPRN_PMC8, value & ovfl_mask);
+               break;
+       case SPRN_TBRL:
+       case SPRN_PURR:
+               /* Ignore writes to read-only registers. */
+               break;
+       default:
+               BUG();
+       }
+}
+
+static u64 pfm_power4_read_pmd(unsigned int cnum)
+{
+       switch (pfm_pmu_conf->pmd_desc[cnum].hw_addr) {
+       case SPRN_PMC1:
+               return mfspr(SPRN_PMC1);
+       case SPRN_PMC2:
+               return mfspr(SPRN_PMC2);
+       case SPRN_PMC3:
+               return mfspr(SPRN_PMC3);
+       case SPRN_PMC4:
+               return mfspr(SPRN_PMC4);
+       case SPRN_PMC5:
+               return mfspr(SPRN_PMC5);
+       case SPRN_PMC6:
+               return mfspr(SPRN_PMC6);
+       case SPRN_PMC7:
+               return mfspr(SPRN_PMC7);
+       case SPRN_PMC8:
+               return mfspr(SPRN_PMC8);
+       case SPRN_TBRL:
+               return ((u64)mfspr(SPRN_TBRU) << 32) | mfspr(SPRN_TBRL);
+       case SPRN_PURR:
+               if (cpu_has_feature(CPU_FTR_PURR))
+                       return mfspr(SPRN_PURR);
+               else
+                       return 0;
+       default:
+               BUG();
+       }
+}
+
+/* forward decl */
+static void pfm_power4_disable_counters(struct pfm_context *ctx,
+                                       struct pfm_event_set *set);
+
+/**
+ * pfm_power4_enable_counters
+ *
+ **/
+static void pfm_power4_enable_counters(struct pfm_context *ctx,
+                                      struct pfm_event_set *set)
+{
+       unsigned int i, max_pmc;
+
+       /* Make sure the counters are disabled before touching the other
+          control registers */
+       pfm_power4_disable_counters(ctx, set);
+
+       max_pmc = ctx->regs.max_pmc;
+
+       /* Write MMCR0 last, and a fairly easy way to do this is to write
+          the registers in the reverse order */
+       for (i = max_pmc; i != 0; i--)
+               if (test_bit(i - 1, set->used_pmcs))
+                       pfm_power4_write_pmc(i - 1, set->pmcs[i - 1]);
+}
+
+/**
+ * pfm_power4_disable_counters
+ *
+ **/
+static void pfm_power4_disable_counters(struct pfm_context *ctx,
+                                       struct pfm_event_set *set)
+{
+       /* Set the Freeze Counters bit */
+       mtspr(SPRN_MMCR0, mfspr(SPRN_MMCR0) | MMCR0_FC);
+       asm volatile ("sync");
+}
+
+/**
+ * pfm_power4_get_ovfl_pmds
+ *
+ * Determine which counters in this set have overflowed and fill in the
+ * set->povfl_pmds mask and set->npend_ovfls count.
+ **/
+static void pfm_power4_get_ovfl_pmds(struct pfm_context *ctx,
+                                    struct pfm_event_set *set)
+{
+       unsigned int i;
+       unsigned int max_pmd = ctx->regs.max_intr_pmd;
+       u64 *used_pmds = set->used_pmds;
+       u64 *cntr_pmds = ctx->regs.cnt_pmds;
+       u64 width_mask = 1 << pfm_pmu_conf->counter_width;
+       u64 new_val, mask[PFM_PMD_BV];
+
+       bitmap_and(cast_ulp(mask), cast_ulp(cntr_pmds),
+                  cast_ulp(used_pmds), max_pmd);
+
+       for (i = 0; i < max_pmd; i++) {
+               if (test_bit(i, mask)) {
+                       new_val = pfm_power4_read_pmd(i);
+                       if (new_val & width_mask) {
+                               set_bit(i, set->povfl_pmds);
+                               set->npend_ovfls++;
+                       }
+               }
+       }
+}
+
+static void pfm_power4_irq_handler(struct pt_regs *regs,
+                                  struct pfm_context *ctx)
+{
+       u32 mmcr0;
+
+       /* Disable the counters (set the freeze bit) to not polute
+        * the counts.
+        */
+       mmcr0 = mfspr(SPRN_MMCR0);
+       mtspr(SPRN_MMCR0, (mmcr0 | MMCR0_FC));
+
+       /* Set the PMM bit (see comment below). */
+       mtmsrd(mfmsr() | MSR_PMM);
+
+       pfm_interrupt_handler(instruction_pointer(regs), regs);
+
+       mmcr0 = mfspr(SPRN_MMCR0);
+
+       /*
+        * Reset the perfmon trigger if
+        * not in masking mode.
+        */
+       if (ctx->state != PFM_CTX_MASKED)
+               mmcr0 |= MMCR0_PMXE;
+
+       /*
+        * We must clear the PMAO bit on some (GQ) chips. Just do it
+        * all the time.
+        */
+       mmcr0 &= ~MMCR0_PMAO;
+
+       /*
+        * Now clear the freeze bit, counting will not start until we
+        * rfid from this exception, because only at that point will
+        * the PMM bit be cleared.
+        */
+       mmcr0 &= ~MMCR0_FC;
+       mtspr(SPRN_MMCR0, mmcr0);
+}
+
+static void pfm_power4_resend_irq(struct pfm_context *ctx)
+{
+       /*
+        * Assert the PMAO bit to cause a PMU interrupt.  Make sure we
+        * trigger the edge detection circuitry for PMAO
+        */
+       mtspr(SPRN_MMCR0, mfspr(SPRN_MMCR0) & ~MMCR0_PMAO);
+       mtspr(SPRN_MMCR0, mfspr(SPRN_MMCR0) | MMCR0_PMAO);
+}
+
+struct pfm_arch_pmu_info pfm_power4_pmu_info = {
+       .pmu_style        = PFM_POWERPC_PMU_POWER4,
+       .write_pmc        = pfm_power4_write_pmc,
+       .write_pmd        = pfm_power4_write_pmd,
+       .read_pmd         = pfm_power4_read_pmd,
+       .irq_handler      = pfm_power4_irq_handler,
+       .get_ovfl_pmds    = pfm_power4_get_ovfl_pmds,
+       .enable_counters  = pfm_power4_enable_counters,
+       .disable_counters = pfm_power4_disable_counters,
+       .resend_irq       = pfm_power4_resend_irq
+};
+
+/*
+ * impl_pmcs, impl_pmds are computed at runtime to minimize errors!
+ */
+static struct pfm_pmu_config pfm_power4_pmu_conf = {
+       .pmu_name = "POWER4",
+       .counter_width = 31,
+       .pmd_desc = pfm_power4_pmd_desc,
+       .pmc_desc = pfm_power4_pmc_desc,
+       .num_pmc_entries = PFM_PM_NUM_PMCS,
+       .num_pmd_entries = PFM_PM_NUM_PMDS,
+       .probe_pmu  = pfm_power4_probe_pmu,
+       .pmu_info = &pfm_power4_pmu_info,
+       .flags = PFM_PMU_BUILTIN_FLAG,
+       .owner = THIS_MODULE
+};
+
+static int __init pfm_power4_pmu_init_module(void)
+{
+       return pfm_pmu_register(&pfm_power4_pmu_conf);
+}
+
+static void __exit pfm_power4_pmu_cleanup_module(void)
+{
+       pfm_pmu_unregister(&pfm_power4_pmu_conf);
+}
+
+module_init(pfm_power4_pmu_init_module);
+module_exit(pfm_power4_pmu_cleanup_module);
diff --git a/arch/powerpc/perfmon/perfmon_power5.c b/arch/powerpc/perfmon/perfmon_power5.c
new file mode 100644
index 0000000..f4bb1ac
--- /dev/null
+++ b/arch/powerpc/perfmon/perfmon_power5.c
@@ -0,0 +1,326 @@
+/*
+ * This file contains the POWER5 PMU register description tables
+ * and pmc checker used by perfmon.c.
+ *
+ * Copyright (c) 2005 David Gibson, IBM Corporation.
+ *
+ * Based on perfmon_p6.c:
+ * Copyright (c) 2005-2006 Hewlett-Packard Development Company, L.P.
+ * Contributed by Stephane Eranian <eranian@hpl.hp.com>
+ *
+ * This program is free software; you can redistribute it and/or
+ * modify it under the terms of version 2 of the GNU General Public
+ * License as published by the Free Software Foundation.
+ *
+ * This program is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+ * General Public License for more details.
+ *
+ * You should have received a copy of the GNU General Public License
+ * along with this program; if not, write to the Free Software
+ * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA
+ * 02111-1307 USA
+ */
+#include <linux/module.h>
+#include <linux/perfmon_kern.h>
+
+MODULE_AUTHOR("David Gibson <dwg@au1.ibm.com>");
+MODULE_DESCRIPTION("POWER5 PMU description table");
+MODULE_LICENSE("GPL");
+
+static struct pfm_regmap_desc pfm_power5_pmc_desc[] = {
+/* mmcr0 */ PMC_D(PFM_REG_I, "MMCR0", MMCR0_FC, 0, 0, SPRN_MMCR0),
+/* mmcr1 */ PMC_D(PFM_REG_I, "MMCR1", 0, 0, 0, SPRN_MMCR1),
+/* mmcra */ PMC_D(PFM_REG_I, "MMCRA", 0, 0, 0, SPRN_MMCRA)
+};
+#define PFM_PM_NUM_PMCS        ARRAY_SIZE(pfm_power5_pmc_desc)
+
+/* The TB and PURR registers are read-only. Also, note that the TB register
+ * actually consists of both the 32-bit SPRN_TBRU and SPRN_TBRL registers.
+ * For Perfmon2's purposes, we'll treat it as a single 64-bit register.
+ */
+static struct pfm_regmap_desc pfm_power5_pmd_desc[] = {
+/* tb    */ PMD_D((PFM_REG_I|PFM_REG_RO), "TB", SPRN_TBRL),
+/* pmd1  */ PMD_D(PFM_REG_C, "PMC1", SPRN_PMC1),
+/* pmd2  */ PMD_D(PFM_REG_C, "PMC2", SPRN_PMC2),
+/* pmd3  */ PMD_D(PFM_REG_C, "PMC3", SPRN_PMC3),
+/* pmd4  */ PMD_D(PFM_REG_C, "PMC4", SPRN_PMC4),
+/* pmd5  */ PMD_D(PFM_REG_C, "PMC5", SPRN_PMC5),
+/* pmd6  */ PMD_D(PFM_REG_C, "PMC6", SPRN_PMC6),
+/* purr  */ PMD_D((PFM_REG_I|PFM_REG_RO), "PURR", SPRN_PURR),
+};
+#define PFM_PM_NUM_PMDS        ARRAY_SIZE(pfm_power5_pmd_desc)
+
+/* forward decl */
+static void pfm_power5_disable_counters(struct pfm_context *ctx,
+                                       struct pfm_event_set *set);
+
+static int pfm_power5_probe_pmu(void)
+{
+       unsigned long pvr = mfspr(SPRN_PVR);
+
+       switch (PVR_VER(pvr)) {
+       case PV_POWER5:
+               return 0;
+       case PV_POWER5p:
+               return (PVR_REV(pvr) < 0x300) ? 0 : -1;
+       default:
+               return -1;
+       }
+}
+
+static void pfm_power5_write_pmc(unsigned int cnum, u64 value)
+{
+       switch (pfm_pmu_conf->pmc_desc[cnum].hw_addr) {
+       case SPRN_MMCR0:
+               mtspr(SPRN_MMCR0, value);
+               break;
+       case SPRN_MMCR1:
+               mtspr(SPRN_MMCR1, value);
+               break;
+       case SPRN_MMCRA:
+               mtspr(SPRN_MMCRA, value);
+               break;
+       default:
+               BUG();
+       }
+}
+
+static void pfm_power5_write_pmd(unsigned int cnum, u64 value)
+{
+       u64 ovfl_mask = pfm_pmu_conf->ovfl_mask;
+
+       switch (pfm_pmu_conf->pmd_desc[cnum].hw_addr) {
+       case SPRN_PMC1:
+               mtspr(SPRN_PMC1, value & ovfl_mask);
+               break;
+       case SPRN_PMC2:
+               mtspr(SPRN_PMC2, value & ovfl_mask);
+               break;
+       case SPRN_PMC3:
+               mtspr(SPRN_PMC3, value & ovfl_mask);
+               break;
+       case SPRN_PMC4:
+               mtspr(SPRN_PMC4, value & ovfl_mask);
+               break;
+       case SPRN_PMC5:
+               mtspr(SPRN_PMC5, value & ovfl_mask);
+               break;
+       case SPRN_PMC6:
+               mtspr(SPRN_PMC6, value & ovfl_mask);
+               break;
+       case SPRN_TBRL:
+       case SPRN_PURR:
+               /* Ignore writes to read-only registers. */
+               break;
+       default:
+               BUG();
+       }
+}
+
+static u64 pfm_power5_read_pmd(unsigned int cnum)
+{
+       switch (pfm_pmu_conf->pmd_desc[cnum].hw_addr) {
+       case SPRN_PMC1:
+               return mfspr(SPRN_PMC1);
+       case SPRN_PMC2:
+               return mfspr(SPRN_PMC2);
+       case SPRN_PMC3:
+               return mfspr(SPRN_PMC3);
+       case SPRN_PMC4:
+               return mfspr(SPRN_PMC4);
+       case SPRN_PMC5:
+               return mfspr(SPRN_PMC5);
+       case SPRN_PMC6:
+               return mfspr(SPRN_PMC6);
+       case SPRN_TBRL:
+               return ((u64)mfspr(SPRN_TBRU) << 32) | mfspr(SPRN_TBRL);
+       case SPRN_PURR:
+               if (cpu_has_feature(CPU_FTR_PURR))
+                       return mfspr(SPRN_PURR);
+               else
+                       return 0;
+       default:
+               BUG();
+       }
+}
+
+/**
+ * pfm_power5_enable_counters
+ *
+ **/
+static void pfm_power5_enable_counters(struct pfm_context *ctx,
+                                      struct pfm_event_set *set)
+{
+       unsigned int i, max_pmc;
+
+       /*
+        * Make sure the counters are disabled before touching the
+        * other control registers
+        */
+       pfm_power5_disable_counters(ctx, set);
+
+       max_pmc = ctx->regs.max_pmc;
+
+       /*
+        * Write MMCR0 last, and a fairly easy way to do
+        * this is to write the registers in the reverse
+        * order
+        */
+       for (i = max_pmc; i != 0; i--)
+               if (test_bit(i - 1, set->used_pmcs))
+                       pfm_power5_write_pmc(i - 1, set->pmcs[i - 1]);
+}
+
+/**
+ * pfm_power5_disable_counters
+ *
+ * Just need to zero all the control registers.
+ **/
+static void pfm_power5_disable_counters(struct pfm_context *ctx,
+                                       struct pfm_event_set *set)
+{
+       /* Set the Freeze Counters bit */
+       mtspr(SPRN_MMCR0, mfspr(SPRN_MMCR0) | MMCR0_FC);
+       asm volatile ("sync");
+}
+
+/**
+ * pfm_power5_get_ovfl_pmds
+ *
+ * Determine which counters in this set have overflowed and fill in the
+ * set->povfl_pmds mask and set->npend_ovfls count.
+ **/
+static void pfm_power5_get_ovfl_pmds(struct pfm_context *ctx,
+                                    struct pfm_event_set *set)
+{
+       unsigned int i;
+       unsigned int max = ctx->regs.max_intr_pmd;
+       u64 *used_pmds = set->used_pmds;
+       u64 *intr_pmds = ctx->regs.intr_pmds;
+       u64 width_mask = 1 << pfm_pmu_conf->counter_width;
+       u64 new_val, mask[PFM_PMD_BV];
+
+       bitmap_and(cast_ulp(mask), cast_ulp(intr_pmds),
+                  cast_ulp(used_pmds), max);
+       /*
+        * If either PMC5 or PMC6 are not being used, just zero out the unused
+        * ones so that they won't interrupt again for another 2^31 counts.
+        * Note that if no other counters overflowed, set->npend_ovfls will
+        * be zero upon returning from this call (i.e. a spurious
+        * interrupt), but that should be ok.
+        *
+        * If neither PMC5 nor PMC6 are used, the counters should be frozen
+        * via MMCR0_FC5_6 and zeroed out.
+        *
+        * If both PMC5 and PMC6 are used, they can be handled correctly by
+        * the loop that follows.
+        */
+
+       if  (!test_bit(5, cast_ulp(used_pmds)))
+               mtspr(SPRN_PMC5, 0);
+       if  (!test_bit(6, cast_ulp(used_pmds)))
+               mtspr(SPRN_PMC6, 0);
+
+       for (i = 0; i < max; i++) {
+               if (test_bit(i, mask)) {
+                       new_val = pfm_power5_read_pmd(i);
+                       if (new_val & width_mask) {
+                               set_bit(i, set->povfl_pmds);
+                               set->npend_ovfls++;
+                       }
+               }
+       }
+}
+
+static void pfm_power5_irq_handler(struct pt_regs *regs,
+                                  struct pfm_context *ctx)
+{
+       u32 mmcr0;
+
+       /* Disable the counters (set the freeze bit) to not polute
+        * the counts.
+        */
+       mmcr0 = mfspr(SPRN_MMCR0);
+       mtspr(SPRN_MMCR0, (mmcr0 | MMCR0_FC));
+
+       /* Set the PMM bit (see comment below). */
+       mtmsrd(mfmsr() | MSR_PMM);
+
+       pfm_interrupt_handler(instruction_pointer(regs), regs);
+
+       mmcr0 = mfspr(SPRN_MMCR0);
+
+       /*
+        * Reset the perfmon trigger if
+        * not in masking mode.
+        */
+       if (ctx->state != PFM_CTX_MASKED)
+               mmcr0 |= MMCR0_PMXE;
+
+       /*
+        * We must clear the PMAO bit on some (GQ) chips. Just do it
+        * all the time.
+        */
+       mmcr0 &= ~MMCR0_PMAO;
+
+       /*
+        * Now clear the freeze bit, counting will not start until we
+        * rfid from this exception, because only at that point will
+        * the PMM bit be cleared.
+        */
+       mmcr0 &= ~MMCR0_FC;
+       mtspr(SPRN_MMCR0, mmcr0);
+}
+
+static void pfm_power5_resend_irq(struct pfm_context *ctx)
+{
+       /*
+        * Assert the PMAO bit to cause a PMU interrupt.  Make sure we
+        * trigger the edge detection circuitry for PMAO
+        */
+       mtspr(SPRN_MMCR0, mfspr(SPRN_MMCR0) & ~MMCR0_PMAO);
+       mtspr(SPRN_MMCR0, mfspr(SPRN_MMCR0) | MMCR0_PMAO);
+}
+
+struct pfm_arch_pmu_info pfm_power5_pmu_info = {
+       .pmu_style        = PFM_POWERPC_PMU_POWER5,
+       .write_pmc        = pfm_power5_write_pmc,
+       .write_pmd        = pfm_power5_write_pmd,
+       .read_pmd         = pfm_power5_read_pmd,
+       .irq_handler      = pfm_power5_irq_handler,
+       .get_ovfl_pmds    = pfm_power5_get_ovfl_pmds,
+       .enable_counters  = pfm_power5_enable_counters,
+       .disable_counters = pfm_power5_disable_counters,
+       .resend_irq       = pfm_power5_resend_irq
+};
+
+/*
+ * impl_pmcs, impl_pmds are computed at runtime to minimize errors!
+ */
+static struct pfm_pmu_config pfm_power5_pmu_conf = {
+       .pmu_name = "POWER5",
+       .counter_width = 31,
+       .pmd_desc = pfm_power5_pmd_desc,
+       .pmc_desc = pfm_power5_pmc_desc,
+       .num_pmc_entries = PFM_PM_NUM_PMCS,
+       .num_pmd_entries = PFM_PM_NUM_PMDS,
+       .probe_pmu  = pfm_power5_probe_pmu,
+       .pmu_info = &pfm_power5_pmu_info,
+       .flags = PFM_PMU_BUILTIN_FLAG,
+       .owner = THIS_MODULE
+};
+
+static int __init pfm_power5_pmu_init_module(void)
+{
+       return pfm_pmu_register(&pfm_power5_pmu_conf);
+}
+
+static void __exit pfm_power5_pmu_cleanup_module(void)
+{
+       pfm_pmu_unregister(&pfm_power5_pmu_conf);
+}
+
+module_init(pfm_power5_pmu_init_module);
+module_exit(pfm_power5_pmu_cleanup_module);
diff --git a/arch/powerpc/perfmon/perfmon_power6.c b/arch/powerpc/perfmon/perfmon_power6.c
new file mode 100644
index 0000000..7882feb
--- /dev/null
+++ b/arch/powerpc/perfmon/perfmon_power6.c
@@ -0,0 +1,520 @@
+/*
+ * This file contains the POWER6 PMU register description tables
+ * and pmc checker used by perfmon.c.
+ *
+ * Copyright (c) 2007, IBM Corporation
+ *
+ * Based on perfmon_power5.c, and written by Carl Love <carll@us.ibm.com>
+ * and Kevin Corry <kevcorry@us.ibm.com>.  Some fixes and refinement by
+ * Corey Ashford <cjashfor@us.ibm.com>
+ *
+ * This program is free software; you can redistribute it and/or
+ * modify it under the terms of version 2 of the GNU General Public
+ * License as published by the Free Software Foundation.
+ *
+ * This program is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+ * General Public License for more details.
+ *
+ * You should have received a copy of the GNU General Public License
+ * along with this program; if not, write to the Free Software
+ * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA
+ * 02111-1307 USA
+ */
+#include <linux/module.h>
+#include <linux/perfmon_kern.h>
+
+MODULE_AUTHOR("Corey Ashford <cjashfor@us.ibm.com>");
+MODULE_DESCRIPTION("POWER6 PMU description table");
+MODULE_LICENSE("GPL");
+
+static struct pfm_regmap_desc pfm_power6_pmc_desc[] = {
+/* mmcr0 */ PMC_D(PFM_REG_I, "MMCR0", MMCR0_FC, 0, 0, SPRN_MMCR0),
+/* mmcr1 */ PMC_D(PFM_REG_I, "MMCR1", 0, 0, 0, SPRN_MMCR1),
+/* mmcra */ PMC_D(PFM_REG_I, "MMCRA", 0, 0, 0, SPRN_MMCRA)
+};
+#define PFM_PM_NUM_PMCS        ARRAY_SIZE(pfm_power6_pmc_desc)
+#define PFM_DELTA_TB    10000   /* Not a real registers */
+#define PFM_DELTA_PURR  10001
+
+/*
+ * counters wrap to zero at transition from 2^32-1 to 2^32.  Note:
+ * interrupt generated at transition from 2^31-1 to 2^31
+ */
+#define OVERFLOW_VALUE    0x100000000UL
+
+/* The TB and PURR registers are read-only. Also, note that the TB register
+ * actually consists of both the 32-bit SPRN_TBRU and SPRN_TBRL registers.
+ * For Perfmon2's purposes, we'll treat it as a single 64-bit register.
+ */
+static struct pfm_regmap_desc pfm_power6_pmd_desc[] = {
+       /* On POWER 6 PMC5 and PMC6 are not writable, they do not
+        * generate interrupts, and do not qualify their counts
+        * based on problem mode, supervisor mode or hypervisor mode.
+        * These two counters are implemented as virtual counters
+        * to make the appear to work like the other counters.  A
+        * kernel timer is used sample the real PMC5 and PMC6 and
+        * update the virtual counters.
+        */
+/* tb    */ PMD_D((PFM_REG_I|PFM_REG_RO), "TB", SPRN_TBRL),
+/* pmd1  */ PMD_D(PFM_REG_C, "PMC1", SPRN_PMC1),
+/* pmd2  */ PMD_D(PFM_REG_C, "PMC2", SPRN_PMC2),
+/* pmd3  */ PMD_D(PFM_REG_C, "PMC3", SPRN_PMC3),
+/* pmd4  */ PMD_D(PFM_REG_C, "PMC4", SPRN_PMC4),
+/* pmd5  */ PMD_D((PFM_REG_I|PFM_REG_V), "PMC5", SPRN_PMC5),
+/* pmd6  */ PMD_D((PFM_REG_I|PFM_REG_V), "PMC6", SPRN_PMC6),
+/* purr  */ PMD_D((PFM_REG_I|PFM_REG_RO), "PURR", SPRN_PURR),
+/* delta purr */ PMD_D((PFM_REG_I|PFM_REG_V), "DELTA_TB", PFM_DELTA_TB),
+/* delta tb   */ PMD_D((PFM_REG_I|PFM_REG_V), "DELTA_PURR", PFM_DELTA_PURR),
+};
+
+#define PFM_PM_NUM_PMDS        ARRAY_SIZE(pfm_power6_pmd_desc)
+
+u32 pmc5_start_save[NR_CPUS];
+u32 pmc6_start_save[NR_CPUS];
+
+static struct timer_list pmc5_6_update[NR_CPUS];
+u64 enable_cntrs_cnt;
+u64 disable_cntrs_cnt;
+u64 call_delta;
+u64 pm5_6_interrupt;
+u64 pm1_4_interrupt;
+/* need ctx_arch for kernel timer.  Can't get it in context of the kernel
+ * timer.
+ */
+struct pfm_arch_context *pmc5_6_ctx_arch[NR_CPUS];
+long int update_time;
+
+static void delta(int cpu_num, struct pfm_arch_context *ctx_arch)
+{
+       u32 tmp5, tmp6;
+
+       call_delta++;
+
+       tmp5 = (u32) mfspr(SPRN_PMC5);
+       tmp6 = (u32) mfspr(SPRN_PMC6);
+
+       /*
+        * The following difference calculation relies on 32-bit modular
+        * arithmetic for the deltas to come out correct (especially in the
+        * presence of a 32-bit counter wrap).
+        */
+       ctx_arch->powergs_pmc5 += (u64)(tmp5 - pmc5_start_save[cpu_num]);
+       ctx_arch->powergs_pmc6 += (u64)(tmp6 - pmc6_start_save[cpu_num]);
+
+       pmc5_start_save[cpu_num] = tmp5;
+       pmc6_start_save[cpu_num] = tmp6;
+
+       return;
+}
+
+
+static void pmc5_6_updater(unsigned long cpu_num)
+{
+       /* update the virtual pmd 5 and pmd 6 counters */
+
+       delta(cpu_num, pmc5_6_ctx_arch[cpu_num]);
+       mod_timer(&pmc5_6_update[cpu_num], jiffies + update_time);
+}
+
+
+static int pfm_power6_probe_pmu(void)
+{
+       unsigned long pvr = mfspr(SPRN_PVR);
+
+       switch (PVR_VER(pvr)) {
+       case PV_POWER6:
+               return 0;
+       case PV_POWER5p:
+               /* If this is a POWER5+ and the revision is less than 0x300,
+                  don't treat it as a POWER6. */
+               return (PVR_REV(pvr) < 0x300) ? -1 : 0;
+       default:
+               return -1;
+       }
+}
+
+static void pfm_power6_write_pmc(unsigned int cnum, u64 value)
+{
+       switch (pfm_pmu_conf->pmc_desc[cnum].hw_addr) {
+       case SPRN_MMCR0:
+               mtspr(SPRN_MMCR0, value);
+               break;
+       case SPRN_MMCR1:
+               mtspr(SPRN_MMCR1, value);
+               break;
+       case SPRN_MMCRA:
+               mtspr(SPRN_MMCRA, value);
+               break;
+       default:
+               BUG();
+       }
+}
+
+static void pfm_power6_write_pmd(unsigned int cnum, u64 value)
+{
+       /* On POWER 6 PMC5 and PMC6 are implemented as
+        * virtual counters.  See comment in pfm_power6_pmd_desc
+        * definition.
+        */
+       u64 ovfl_mask = pfm_pmu_conf->ovfl_mask;
+
+       switch (pfm_pmu_conf->pmd_desc[cnum].hw_addr) {
+       case SPRN_PMC1:
+               mtspr(SPRN_PMC1, value & ovfl_mask);
+               break;
+       case SPRN_PMC2:
+               mtspr(SPRN_PMC2, value & ovfl_mask);
+               break;
+       case SPRN_PMC3:
+               mtspr(SPRN_PMC3, value & ovfl_mask);
+               break;
+       case SPRN_PMC4:
+               mtspr(SPRN_PMC4, value & ovfl_mask);
+               break;
+       case SPRN_TBRL:
+       case SPRN_PURR:
+               /* Ignore writes to read-only registers. */
+               break;
+       default:
+               BUG();
+       }
+}
+
+static u64 pfm_power6_sread(struct pfm_context *ctx, unsigned int cnum)
+{
+       struct pfm_arch_context *ctx_arch = pfm_ctx_arch(ctx);
+       int cpu_num = smp_processor_id();
+
+       /* On POWER 6 PMC5 and PMC6 are implemented as
+        * virtual counters.  See comment in pfm_power6_pmd_desc
+        * definition.
+        */
+
+       switch (pfm_pmu_conf->pmd_desc[cnum].hw_addr) {
+       case SPRN_PMC5:
+               return ctx_arch->powergs_pmc5 + (u64)((u32)mfspr(SPRN_PMC5) - pmc5_start_save[cpu_num]);
+               break;
+
+       case SPRN_PMC6:
+               return ctx_arch->powergs_pmc6 + (u64)((u32)mfspr(SPRN_PMC6) - pmc6_start_save[cpu_num]);
+               break;
+
+       case PFM_DELTA_TB:
+               return ctx_arch->delta_tb
+                       + (((u64)mfspr(SPRN_TBRU) << 32) | mfspr(SPRN_TBRL))
+                       - ctx_arch->delta_tb_start;
+               break;
+
+       case PFM_DELTA_PURR:
+               return ctx_arch->delta_purr
+                       + mfspr(SPRN_PURR)
+                       - ctx_arch->delta_purr_start;
+               break;
+
+       default:
+               BUG();
+       }
+}
+
+void pfm_power6_swrite(struct pfm_context *ctx, unsigned int cnum,
+       u64 val)
+{
+       struct pfm_arch_context *ctx_arch = pfm_ctx_arch(ctx);
+       int cpu_num = smp_processor_id();
+
+       switch (pfm_pmu_conf->pmd_desc[cnum].hw_addr) {
+       case SPRN_PMC5:
+               pmc5_start_save[cpu_num] = mfspr(SPRN_PMC5);
+               ctx_arch->powergs_pmc5 = val;
+               break;
+
+       case SPRN_PMC6:
+               pmc6_start_save[cpu_num] = mfspr(SPRN_PMC6);
+               ctx_arch->powergs_pmc6 = val;
+               break;
+
+       case PFM_DELTA_TB:
+               ctx_arch->delta_tb_start =
+                       (((u64)mfspr(SPRN_TBRU) << 32) | mfspr(SPRN_TBRL));
+               ctx_arch->delta_tb = val;
+               break;
+
+       case PFM_DELTA_PURR:
+               ctx_arch->delta_purr_start = mfspr(SPRN_PURR);
+               ctx_arch->delta_purr = val;
+               break;
+
+       default:
+               BUG();
+       }
+}
+
+static u64 pfm_power6_read_pmd(unsigned int cnum)
+{
+       switch (pfm_pmu_conf->pmd_desc[cnum].hw_addr) {
+       case SPRN_PMC1:
+               return mfspr(SPRN_PMC1);
+       case SPRN_PMC2:
+               return mfspr(SPRN_PMC2);
+       case SPRN_PMC3:
+               return mfspr(SPRN_PMC3);
+       case SPRN_PMC4:
+               return mfspr(SPRN_PMC4);
+       case SPRN_TBRL:
+               return ((u64)mfspr(SPRN_TBRU) << 32) | mfspr(SPRN_TBRL);
+       case SPRN_PURR:
+               if (cpu_has_feature(CPU_FTR_PURR))
+                       return mfspr(SPRN_PURR);
+               else
+                       return 0;
+       default:
+               BUG();
+       }
+}
+
+
+/**
+ * pfm_power6_enable_counters
+ *
+ **/
+static void pfm_power6_enable_counters(struct pfm_context *ctx,
+                                      struct pfm_event_set *set)
+{
+
+       unsigned int i, max_pmc;
+       int cpu_num = smp_processor_id();
+       struct pfm_arch_context *ctx_arch;
+
+       enable_cntrs_cnt++;
+
+       /* need the ctx passed down to the routine */
+       ctx_arch = pfm_ctx_arch(ctx);
+       max_pmc = ctx->regs.max_pmc;
+
+       /* Write MMCR0 last, and a fairly easy way to do this is to write
+          the registers in the reverse order */
+       for (i = max_pmc; i != 0; i--)
+               if (test_bit(i - 1, set->used_pmcs))
+                       pfm_power6_write_pmc(i - 1, set->pmcs[i - 1]);
+
+       /* save current free running HW event count */
+       pmc5_start_save[cpu_num] = mfspr(SPRN_PMC5);
+       pmc6_start_save[cpu_num] = mfspr(SPRN_PMC6);
+
+       ctx_arch->delta_purr_start = mfspr(SPRN_PURR);
+
+       if (cpu_has_feature(CPU_FTR_PURR))
+               ctx_arch->delta_tb_start =
+                       ((u64)mfspr(SPRN_TBRU) << 32) | mfspr(SPRN_TBRL);
+       else
+               ctx_arch->delta_tb_start = 0;
+
+       /* Start kernel timer for this cpu to periodically update
+        * the virtual counters.
+        */
+       init_timer(&pmc5_6_update[cpu_num]);
+       pmc5_6_update[cpu_num].function = pmc5_6_updater;
+       pmc5_6_update[cpu_num].data = (unsigned long) cpu_num;
+       pmc5_6_update[cpu_num].expires = jiffies + update_time;
+       /* context for this timer, timer will be removed if context
+        * is switched because the counters will be stopped first.
+        * NEEDS WORK, I think this is all ok, a little concerned about a
+        * race between the kernel timer going off right as the counters
+        * are being stopped and the context switching.  Need to think
+        * about this.
+        */
+       pmc5_6_ctx_arch[cpu_num] = ctx_arch;
+       add_timer(&pmc5_6_update[cpu_num]);
+}
+
+/**
+ * pfm_power6_disable_counters
+ *
+ **/
+static void pfm_power6_disable_counters(struct pfm_context *ctx,
+                                       struct pfm_event_set *set)
+{
+       struct pfm_arch_context *ctx_arch;
+       int cpu_num = smp_processor_id();
+
+       disable_cntrs_cnt++;
+
+       /* Set the Freeze Counters bit */
+       mtspr(SPRN_MMCR0, mfspr(SPRN_MMCR0) | MMCR0_FC);
+       asm volatile ("sync");
+
+       /* delete kernel update timer */
+       del_timer_sync(&pmc5_6_update[cpu_num]);
+
+       /* Update the virtual pmd 5 and 6 counters from the free running
+        * HW counters
+        */
+       ctx_arch = pfm_ctx_arch(ctx);
+       delta(cpu_num, ctx_arch);
+
+       ctx_arch->delta_tb +=
+               (((u64)mfspr(SPRN_TBRU) << 32) | mfspr(SPRN_TBRL))
+               - ctx_arch->delta_tb_start;
+
+       ctx_arch->delta_purr += mfspr(SPRN_PURR)
+               - ctx_arch->delta_purr_start;
+}
+
+/**
+ * pfm_power6_get_ovfl_pmds
+ *
+ * Determine which counters in this set have overflowed and fill in the
+ * set->povfl_pmds mask and set->npend_ovfls count.
+ **/
+static void pfm_power6_get_ovfl_pmds(struct pfm_context *ctx,
+                                    struct pfm_event_set *set)
+{
+       unsigned int i;
+       unsigned int first_intr_pmd = ctx->regs.first_intr_pmd;
+       unsigned int max_intr_pmd = ctx->regs.max_intr_pmd;
+       u64 *used_pmds = set->used_pmds;
+       u64 *cntr_pmds = ctx->regs.cnt_pmds;
+       u64 width_mask = 1 << pfm_pmu_conf->counter_width;
+       u64 new_val, mask[PFM_PMD_BV];
+
+       bitmap_and(cast_ulp(mask), cast_ulp(cntr_pmds), cast_ulp(used_pmds), max_intr_pmd);
+
+       /* max_intr_pmd is actually the last interrupting pmd register + 1 */
+       for (i = first_intr_pmd; i < max_intr_pmd; i++) {
+               if (test_bit(i, mask)) {
+                       new_val = pfm_power6_read_pmd(i);
+                       if (new_val & width_mask) {
+                               set_bit(i, set->povfl_pmds);
+                               set->npend_ovfls++;
+                       }
+               }
+       }
+}
+
+static void pfm_power6_irq_handler(struct pt_regs *regs,
+                                  struct pfm_context *ctx)
+{
+       u32 mmcr0;
+       u64 mmcra;
+
+       /* Disable the counters (set the freeze bit) to not polute
+        * the counts.
+        */
+       mmcr0 = mfspr(SPRN_MMCR0);
+       mtspr(SPRN_MMCR0, (mmcr0 | MMCR0_FC));
+       mmcra = mfspr(SPRN_MMCRA);
+
+       /* Set the PMM bit (see comment below). */
+       mtmsrd(mfmsr() | MSR_PMM);
+
+       pm1_4_interrupt++;
+
+       pfm_interrupt_handler(instruction_pointer(regs), regs);
+
+       mmcr0 = mfspr(SPRN_MMCR0);
+
+       /*
+        * Reset the perfmon trigger if
+        * not in masking mode.
+        */
+       if (ctx->state != PFM_CTX_MASKED)
+               mmcr0 |= MMCR0_PMXE;
+
+       /*
+        * Clear the PMU Alert Occurred bit
+        */
+       mmcr0 &= ~MMCR0_PMAO;
+
+       /* Clear the appropriate bits in the MMCRA. */
+       mmcra &= ~(POWER6_MMCRA_THRM | POWER6_MMCRA_OTHER);
+       mtspr(SPRN_MMCRA, mmcra);
+
+       /*
+        * Now clear the freeze bit, counting will not start until we
+        * rfid from this exception, because only at that point will
+        * the PMM bit be cleared.
+        */
+       mmcr0 &= ~MMCR0_FC;
+       mtspr(SPRN_MMCR0, mmcr0);
+}
+
+static void pfm_power6_resend_irq(struct pfm_context *ctx)
+{
+       /*
+        * Assert the PMAO bit to cause a PMU interrupt.  Make sure we
+        * trigger the edge detection circuitry for PMAO
+        */
+       mtspr(SPRN_MMCR0, mfspr(SPRN_MMCR0) & ~MMCR0_PMAO);
+       mtspr(SPRN_MMCR0, mfspr(SPRN_MMCR0) | MMCR0_PMAO);
+}
+
+struct pfm_arch_pmu_info pfm_power6_pmu_info = {
+       .pmu_style        = PFM_POWERPC_PMU_POWER6,
+       .write_pmc        = pfm_power6_write_pmc,
+       .write_pmd        = pfm_power6_write_pmd,
+       .read_pmd         = pfm_power6_read_pmd,
+       .irq_handler      = pfm_power6_irq_handler,
+       .get_ovfl_pmds    = pfm_power6_get_ovfl_pmds,
+       .enable_counters  = pfm_power6_enable_counters,
+       .disable_counters = pfm_power6_disable_counters,
+       .resend_irq       = pfm_power6_resend_irq
+};
+
+/*
+ * impl_pmcs, impl_pmds are computed at runtime to minimize errors!
+ */
+static struct pfm_pmu_config pfm_power6_pmu_conf = {
+       .pmu_name = "POWER6",
+       .counter_width = 31,
+       .pmd_desc = pfm_power6_pmd_desc,
+       .pmc_desc = pfm_power6_pmc_desc,
+       .num_pmc_entries = PFM_PM_NUM_PMCS,
+       .num_pmd_entries = PFM_PM_NUM_PMDS,
+       .probe_pmu  = pfm_power6_probe_pmu,
+       .pmu_info = &pfm_power6_pmu_info,
+       .pmd_sread = pfm_power6_sread,
+       .pmd_swrite = pfm_power6_swrite,
+       .flags = PFM_PMU_BUILTIN_FLAG,
+       .owner = THIS_MODULE
+};
+
+static int __init pfm_power6_pmu_init_module(void)
+{
+       int ret;
+       disable_cntrs_cnt = 0;
+       enable_cntrs_cnt = 0;
+       call_delta = 0;
+       pm5_6_interrupt = 0;
+       pm1_4_interrupt = 0;
+
+       /* calculate the time for updating counters 5 and 6 */
+
+       /*
+        * MAX_EVENT_RATE assumes a max instruction issue rate of 2
+        * instructions per clock cycle.  Experience shows that this factor
+        * of 2 is more than adequate.
+        */
+
+# define MAX_EVENT_RATE (ppc_proc_freq * 2)
+
+       /*
+        * Calculate the time, in jiffies, it takes for event counter 5 or
+        * 6 to completely wrap when counting at the max event rate, and
+        * then figure on sampling at twice that rate.
+        */
+       update_time = (((unsigned long)HZ * OVERFLOW_VALUE)
+                      / ((unsigned long)MAX_EVENT_RATE)) / 2;
+
+       ret =  pfm_pmu_register(&pfm_power6_pmu_conf);
+       return ret;
+}
+
+static void __exit pfm_power6_pmu_cleanup_module(void)
+{
+       pfm_pmu_unregister(&pfm_power6_pmu_conf);
+}
+
+module_init(pfm_power6_pmu_init_module);
+module_exit(pfm_power6_pmu_cleanup_module);
diff --git a/arch/powerpc/perfmon/perfmon_ppc32.c b/arch/powerpc/perfmon/perfmon_ppc32.c
new file mode 100644
index 0000000..76f0b84
--- /dev/null
+++ b/arch/powerpc/perfmon/perfmon_ppc32.c
@@ -0,0 +1,340 @@
+/*
+ * This file contains the PPC32 PMU register description tables
+ * and pmc checker used by perfmon.c.
+ *
+ * Philip Mucci, mucci@cs.utk.edu
+ *
+ * Based on code from:
+ * Copyright (c) 2005 David Gibson, IBM Corporation.
+ *
+ * Based on perfmon_p6.c:
+ * Copyright (c) 2005-2006 Hewlett-Packard Development Company, L.P.
+ * Contributed by Stephane Eranian <eranian@hpl.hp.com>
+ *
+ * This program is free software; you can redistribute it and/or
+ * modify it under the terms of version 2 of the GNU General Public
+ * License as published by the Free Software Foundation.
+ *
+ * This program is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+ * General Public License for more details.
+ *
+ * You should have received a copy of the GNU General Public License
+ * along with this program; if not, write to the Free Software
+ * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA
+ * 02111-1307 USA
+ */
+#include <linux/module.h>
+#include <linux/perfmon_kern.h>
+#include <asm/reg.h>
+
+MODULE_AUTHOR("Philip Mucci <mucci@cs.utk.edu>");
+MODULE_DESCRIPTION("PPC32 PMU description table");
+MODULE_LICENSE("GPL");
+
+static struct pfm_pmu_config pfm_ppc32_pmu_conf;
+
+static struct pfm_regmap_desc pfm_ppc32_pmc_desc[] = {
+/* mmcr0 */ PMC_D(PFM_REG_I, "MMCR0", 0x0, 0, 0, SPRN_MMCR0),
+/* mmcr1 */ PMC_D(PFM_REG_I, "MMCR1", 0x0, 0, 0, SPRN_MMCR1),
+/* mmcr2 */ PMC_D(PFM_REG_I, "MMCR2", 0x0, 0, 0, SPRN_MMCR2),
+};
+#define PFM_PM_NUM_PMCS        ARRAY_SIZE(pfm_ppc32_pmc_desc)
+
+static struct pfm_regmap_desc pfm_ppc32_pmd_desc[] = {
+/* pmd0  */ PMD_D(PFM_REG_C, "PMC1", SPRN_PMC1),
+/* pmd1  */ PMD_D(PFM_REG_C, "PMC2", SPRN_PMC2),
+/* pmd2  */ PMD_D(PFM_REG_C, "PMC3", SPRN_PMC3),
+/* pmd3  */ PMD_D(PFM_REG_C, "PMC4", SPRN_PMC4),
+/* pmd4  */ PMD_D(PFM_REG_C, "PMC5", SPRN_PMC5),
+/* pmd5  */ PMD_D(PFM_REG_C, "PMC6", SPRN_PMC6),
+};
+#define PFM_PM_NUM_PMDS        ARRAY_SIZE(pfm_ppc32_pmd_desc)
+
+static void perfmon_perf_irq(struct pt_regs *regs)
+{
+       u32 mmcr0;
+
+       /* BLATANTLY STOLEN FROM OPROFILE, then modified */
+
+       /* set the PMM bit (see comment below) */
+       mtmsr(mfmsr() | MSR_PMM);
+
+       pfm_interrupt_handler(instruction_pointer(regs), regs);
+
+       /* The freeze bit was set by the interrupt.
+        * Clear the freeze bit, and reenable the interrupt.
+        * The counters won't actually start until the rfi clears
+        * the PMM bit.
+        */
+
+       /* Unfreezes the counters on this CPU, enables the interrupt,
+        * enables the counters to trigger the interrupt, and sets the
+        * counters to only count when the mark bit is not set.
+        */
+       mmcr0 = mfspr(SPRN_MMCR0);
+
+       mmcr0 &= ~(MMCR0_FC | MMCR0_FCM0);
+       mmcr0 |= (MMCR0_FCECE | MMCR0_PMC1CE | MMCR0_PMCnCE | MMCR0_PMXE);
+
+       mtspr(SPRN_MMCR0, mmcr0);
+}
+
+static int pfm_ppc32_probe_pmu(void)
+{
+       enum ppc32_pmu_type pm_type;
+       int nmmcr = 0, npmds = 0, intsok = 0, i;
+       unsigned int pvr;
+       char *str;
+
+       pvr = mfspr(SPRN_PVR);
+
+       switch (PVR_VER(pvr)) {
+       case 0x0004: /* 604 */
+               str = "PPC604";
+               pm_type = PFM_POWERPC_PMU_604;
+               nmmcr = 1;
+               npmds = 2;
+               break;
+       case 0x0009: /* 604e;  */
+       case 0x000A: /* 604ev */
+               str = "PPC604e";
+               pm_type = PFM_POWERPC_PMU_604e;
+               nmmcr = 2;
+               npmds = 4;
+               break;
+       case 0x0008: /* 750/740 */
+               str = "PPC750";
+               pm_type = PFM_POWERPC_PMU_750;
+               nmmcr = 2;
+               npmds = 4;
+               break;
+       case 0x7000: /* 750FX */
+       case 0x7001:
+               str = "PPC750";
+               pm_type = PFM_POWERPC_PMU_750;
+               nmmcr = 2;
+               npmds = 4;
+               if ((pvr & 0xFF0F) >= 0x0203)
+                       intsok = 1;
+               break;
+       case 0x7002: /* 750GX */
+               str = "PPC750";
+               pm_type = PFM_POWERPC_PMU_750;
+               nmmcr = 2;
+               npmds = 4;
+               intsok = 1;
+       case 0x000C: /* 7400 */
+               str = "PPC7400";
+               pm_type = PFM_POWERPC_PMU_7400;
+               nmmcr = 3;
+               npmds = 4;
+               break;
+       case 0x800C: /* 7410 */
+               str = "PPC7410";
+               pm_type = PFM_POWERPC_PMU_7400;
+               nmmcr = 3;
+               npmds = 4;
+               if ((pvr & 0xFFFF) >= 0x01103)
+                       intsok = 1;
+               break;
+       case 0x8000: /* 7451/7441 */
+       case 0x8001: /* 7455/7445 */
+       case 0x8002: /* 7457/7447 */
+       case 0x8003: /* 7447A */
+       case 0x8004: /* 7448 */
+               str = "PPC7450";
+               pm_type = PFM_POWERPC_PMU_7450;
+               nmmcr = 3; npmds = 6;
+               intsok = 1;
+               break;
+       default:
+               PFM_INFO("Unknown PVR_VER(0x%x)\n", PVR_VER(pvr));
+               return -1;
+       }
+
+       /*
+        * deconfigure unimplemented registers
+        */
+       for (i = npmds; i < PFM_PM_NUM_PMDS; i++)
+               pfm_ppc32_pmd_desc[i].type = PFM_REG_NA;
+
+       for (i = nmmcr; i < PFM_PM_NUM_PMCS; i++)
+               pfm_ppc32_pmc_desc[i].type = PFM_REG_NA;
+
+       /*
+        * update PMU description structure
+        */
+       pfm_ppc32_pmu_conf.pmu_name = str;
+       pfm_ppc32_pmu_info.pmu_style = pm_type;
+       pfm_ppc32_pmu_conf.num_pmc_entries = nmmcr;
+       pfm_ppc32_pmu_conf.num_pmd_entries = npmds;
+
+       if (intsok == 0)
+               PFM_INFO("Interrupts unlikely to work\n");
+
+       return reserve_pmc_hardware(perfmon_perf_irq);
+}
+
+static void pfm_ppc32_write_pmc(unsigned int cnum, u64 value)
+{
+       switch (pfm_pmu_conf->pmc_desc[cnum].hw_addr) {
+       case SPRN_MMCR0:
+               mtspr(SPRN_MMCR0, value);
+               break;
+       case SPRN_MMCR1:
+               mtspr(SPRN_MMCR1, value);
+               break;
+       case SPRN_MMCR2:
+               mtspr(SPRN_MMCR2, value);
+               break;
+       default:
+               BUG();
+       }
+}
+
+static void pfm_ppc32_write_pmd(unsigned int cnum, u64 value)
+{
+       switch (pfm_pmu_conf->pmd_desc[cnum].hw_addr) {
+       case SPRN_PMC1:
+               mtspr(SPRN_PMC1, value);
+               break;
+       case SPRN_PMC2:
+               mtspr(SPRN_PMC2, value);
+               break;
+       case SPRN_PMC3:
+               mtspr(SPRN_PMC3, value);
+               break;
+       case SPRN_PMC4:
+               mtspr(SPRN_PMC4, value);
+               break;
+       case SPRN_PMC5:
+               mtspr(SPRN_PMC5, value);
+               break;
+       case SPRN_PMC6:
+               mtspr(SPRN_PMC6, value);
+               break;
+       default:
+               BUG();
+       }
+}
+
+static u64 pfm_ppc32_read_pmd(unsigned int cnum)
+{
+       switch (pfm_pmu_conf->pmd_desc[cnum].hw_addr) {
+       case SPRN_PMC1:
+               return mfspr(SPRN_PMC1);
+       case SPRN_PMC2:
+               return mfspr(SPRN_PMC2);
+       case SPRN_PMC3:
+               return mfspr(SPRN_PMC3);
+       case SPRN_PMC4:
+               return mfspr(SPRN_PMC4);
+       case SPRN_PMC5:
+               return mfspr(SPRN_PMC5);
+       case SPRN_PMC6:
+               return mfspr(SPRN_PMC6);
+       default:
+               BUG();
+       }
+}
+
+/**
+ * pfm_ppc32_enable_counters
+ *
+ * Just need to load the current values into the control registers.
+ **/
+static void pfm_ppc32_enable_counters(struct pfm_context *ctx,
+                                     struct pfm_event_set *set)
+{
+       unsigned int i, max_pmc;
+
+       max_pmc = pfm_pmu_conf->regs.max_pmc;
+
+       for (i = 0; i < max_pmc; i++)
+               if (test_bit(i, set->used_pmcs))
+                       pfm_ppc32_write_pmc(i, set->pmcs[i]);
+}
+
+/**
+ * pfm_ppc32_disable_counters
+ *
+ * Just need to zero all the control registers.
+ **/
+static void pfm_ppc32_disable_counters(struct pfm_context *ctx,
+                                      struct pfm_event_set *set)
+{
+       unsigned int i, max;
+
+       max = pfm_pmu_conf->regs.max_pmc;
+
+       for (i = 0; i < max; i++)
+               if (test_bit(i, set->used_pmcs))
+                       pfm_ppc32_write_pmc(ctx, 0);
+}
+
+/**
+ * pfm_ppc32_get_ovfl_pmds
+ *
+ * Determine which counters in this set have overflowed and fill in the
+ * set->povfl_pmds mask and set->npend_ovfls count.
+ **/
+static void pfm_ppc32_get_ovfl_pmds(struct pfm_context *ctx,
+                                   struct pfm_event_set *set)
+{
+       unsigned int i;
+       unsigned int max_pmd = pfm_pmu_conf->regs.max_cnt_pmd;
+       u64 *used_pmds = set->used_pmds;
+       u64 *cntr_pmds = pfm_pmu_conf->regs.cnt_pmds;
+       u64 width_mask = 1 << pfm_pmu_conf->counter_width;
+       u64 new_val, mask[PFM_PMD_BV];
+
+       bitmap_and(cast_ulp(mask), cast_ulp(cntr_pmds),
+                  cast_ulp(used_pmds), max_pmd);
+
+       for (i = 0; i < max_pmd; i++) {
+               if (test_bit(i, mask)) {
+                       new_val = pfm_ppc32_read_pmd(i);
+                       if (new_val & width_mask) {
+                               set_bit(i, set->povfl_pmds);
+                               set->npend_ovfls++;
+                       }
+               }
+       }
+}
+
+struct pfm_arch_pmu_info pfm_ppc32_pmu_info = {
+       .pmu_style        = PFM_POWERPC_PMU_NONE,
+       .write_pmc        = pfm_ppc32_write_pmc,
+       .write_pmd        = pfm_ppc32_write_pmd,
+       .read_pmd         = pfm_ppc32_read_pmd,
+       .get_ovfl_pmds    = pfm_ppc32_get_ovfl_pmds,
+       .enable_counters  = pfm_ppc32_enable_counters,
+       .disable_counters = pfm_ppc32_disable_counters,
+};
+
+static struct pfm_pmu_config pfm_ppc32_pmu_conf = {
+       .counter_width = 31,
+       .pmd_desc = pfm_ppc32_pmd_desc,
+       .pmc_desc = pfm_ppc32_pmc_desc,
+       .probe_pmu  = pfm_ppc32_probe_pmu,
+       .flags = PFM_PMU_BUILTIN_FLAG,
+       .owner = THIS_MODULE,
+       .version = "0.1",
+       .arch_info = &pfm_ppc32_pmu_info,
+};
+
+static int __init pfm_ppc32_pmu_init_module(void)
+{
+       return pfm_pmu_register(&pfm_ppc32_pmu_conf);
+}
+
+static void __exit pfm_ppc32_pmu_cleanup_module(void)
+{
+       release_pmc_hardware();
+       pfm_pmu_unregister(&pfm_ppc32_pmu_conf);
+}
+
+module_init(pfm_ppc32_pmu_init_module);
+module_exit(pfm_ppc32_pmu_cleanup_module);
diff --git a/arch/powerpc/platforms/cell/cbe_regs.c b/arch/powerpc/platforms/cell/cbe_regs.c
index dbc338f..e24320e 100644
--- a/arch/powerpc/platforms/cell/cbe_regs.c
+++ b/arch/powerpc/platforms/cell/cbe_regs.c
@@ -33,6 +33,7 @@ static struct cbe_regs_map
        struct cbe_iic_regs __iomem *iic_regs;
        struct cbe_mic_tm_regs __iomem *mic_tm_regs;
        struct cbe_pmd_shadow_regs pmd_shadow_regs;
+       struct cbe_ppe_priv_regs __iomem *ppe_priv_regs;
 } cbe_regs_maps[MAX_CBE];
 static int cbe_regs_map_count;
 
@@ -145,6 +146,23 @@ struct cbe_mic_tm_regs __iomem *cbe_get_cpu_mic_tm_regs(int cpu)
 }
 EXPORT_SYMBOL_GPL(cbe_get_cpu_mic_tm_regs);
 
+struct cbe_ppe_priv_regs __iomem *cbe_get_ppe_priv_regs(struct device_node *np)
+{
+       struct cbe_regs_map *map = cbe_find_map(np);
+       if (map == NULL)
+               return NULL;
+       return map->ppe_priv_regs;
+}
+
+struct cbe_ppe_priv_regs __iomem *cbe_get_cpu_ppe_priv_regs(int cpu)
+{
+       struct cbe_regs_map *map = cbe_thread_map[cpu].regs;
+       if (map == NULL)
+               return NULL;
+       return map->ppe_priv_regs;
+}
+EXPORT_SYMBOL_GPL(cbe_get_cpu_ppe_priv_regs);
+
 u32 cbe_get_hw_thread_id(int cpu)
 {
        return cbe_thread_map[cpu].thread_id;
@@ -206,6 +224,11 @@ void __init cbe_fill_regs_map(struct cbe_regs_map *map)
                for_each_node_by_type(np, "mic-tm")
                        if (of_get_parent(np) == be)
                                map->mic_tm_regs = of_iomap(np, 0);
+
+               for_each_node_by_type(np, "ppe-mmio")
+                       if (of_get_parent(np) == be)
+                               map->ppe_priv_regs = of_iomap(np, 0);
+
        } else {
                struct device_node *cpu;
                /* That hack must die die die ! */
@@ -227,6 +250,10 @@ void __init cbe_fill_regs_map(struct cbe_regs_map *map)
                prop = of_get_property(cpu, "mic-tm", NULL);
                if (prop != NULL)
                        map->mic_tm_regs = ioremap(prop->address, prop->len);
+
+               prop = of_get_property(cpu, "ppe-mmio", NULL);
+               if (prop != NULL)
+                       map->ppe_priv_regs = ioremap(prop->address, prop->len);
        }
 }
 
diff --git a/arch/sparc/include/asm/hypervisor.h b/arch/sparc/include/asm/hypervisor.h
index 109ae24..bafe5a6 100644
--- a/arch/sparc/include/asm/hypervisor.h
+++ b/arch/sparc/include/asm/hypervisor.h
@@ -2713,6 +2713,30 @@ extern unsigned long sun4v_ldc_revoke(unsigned long channel,
  */
 #define HV_FAST_SET_PERFREG            0x101
 
+#define HV_N2_PERF_SPARC_CTL           0x0
+#define HV_N2_PERF_DRAM_CTL0           0x1
+#define HV_N2_PERF_DRAM_CNT0           0x2
+#define HV_N2_PERF_DRAM_CTL1           0x3
+#define HV_N2_PERF_DRAM_CNT1           0x4
+#define HV_N2_PERF_DRAM_CTL2           0x5
+#define HV_N2_PERF_DRAM_CNT2           0x6
+#define HV_N2_PERF_DRAM_CTL3           0x7
+#define HV_N2_PERF_DRAM_CNT3           0x8
+
+#define HV_FAST_N2_GET_PERFREG         0x104
+#define HV_FAST_N2_SET_PERFREG         0x105
+
+#ifndef __ASSEMBLY__
+extern unsigned long sun4v_niagara_getperf(unsigned long reg,
+                                          unsigned long *val);
+extern unsigned long sun4v_niagara_setperf(unsigned long reg,
+                                          unsigned long val);
+extern unsigned long sun4v_niagara2_getperf(unsigned long reg,
+                                           unsigned long *val);
+extern unsigned long sun4v_niagara2_setperf(unsigned long reg,
+                                           unsigned long val);
+#endif
+
 /* MMU statistics services.
  *
  * The hypervisor maintains MMU statistics and privileged code provides
diff --git a/arch/sparc/include/asm/irq_64.h b/arch/sparc/include/asm/irq_64.h
index e3dd930..6cf3aec 100644
--- a/arch/sparc/include/asm/irq_64.h
+++ b/arch/sparc/include/asm/irq_64.h
@@ -67,6 +67,9 @@ extern void virt_irq_free(unsigned int virt_irq);
 extern void __init init_IRQ(void);
 extern void fixup_irqs(void);
 
+extern int register_perfctr_intr(void (*handler)(struct pt_regs *));
+extern void release_perfctr_intr(void (*handler)(struct pt_regs *));
+
 static inline void set_softint(unsigned long bits)
 {
        __asm__ __volatile__("wr        %0, 0x0, %%set_softint"
diff --git a/arch/sparc/include/asm/perfmon.h b/arch/sparc/include/asm/perfmon.h
new file mode 100644
index 0000000..f20cbfa
--- /dev/null
+++ b/arch/sparc/include/asm/perfmon.h
@@ -0,0 +1,11 @@
+#ifndef _SPARC64_PERFMON_H_
+#define _SPARC64_PERFMON_H_
+
+/*
+ * arch-specific user visible interface definitions
+ */
+
+#define PFM_ARCH_MAX_PMCS      2
+#define PFM_ARCH_MAX_PMDS      3
+
+#endif /* _SPARC64_PERFMON_H_ */
diff --git a/arch/sparc/include/asm/perfmon_kern.h b/arch/sparc/include/asm/perfmon_kern.h
new file mode 100644
index 0000000..033eff5
--- /dev/null
+++ b/arch/sparc/include/asm/perfmon_kern.h
@@ -0,0 +1,286 @@
+#ifndef _SPARC64_PERFMON_KERN_H_
+#define _SPARC64_PERFMON_KERN_H_
+
+#ifdef __KERNEL__
+
+#ifdef CONFIG_PERFMON
+
+#include <linux/irq.h>
+#include <asm/system.h>
+
+#define PFM_ARCH_PMD_STK_ARG   2
+#define PFM_ARCH_PMC_STK_ARG   1
+
+struct pfm_arch_pmu_info {
+       u32 pmu_style;
+};
+
+static inline void pfm_arch_resend_irq(struct pfm_context *ctx)
+{
+}
+
+static inline void pfm_arch_clear_pmd_ovfl_cond(struct pfm_context *ctx,
+                                               struct pfm_event_set *set)
+{}
+
+static inline void pfm_arch_serialize(void)
+{
+}
+
+/*
+ * SPARC does not save the PMDs during pfm_arch_intr_freeze_pmu(), thus
+ * this routine needs to do it when switching sets on overflow
+ */
+static inline void pfm_arch_save_pmds_from_intr(struct pfm_context *ctx,
+                                               struct pfm_event_set *set)
+{
+       pfm_save_pmds(ctx, set);
+}
+
+extern void pfm_arch_write_pmc(struct pfm_context *ctx,
+                              unsigned int cnum, u64 value);
+extern u64 pfm_arch_read_pmc(struct pfm_context *ctx, unsigned int cnum);
+
+static inline void pfm_arch_write_pmd(struct pfm_context *ctx,
+                                     unsigned int cnum, u64 value)
+{
+       u64 pic;
+
+       value &= pfm_pmu_conf->ovfl_mask;
+
+       read_pic(pic);
+
+       switch (cnum) {
+       case 0:
+               pic = (pic & 0xffffffff00000000UL) |
+                       (value & 0xffffffffUL);
+               break;
+       case 1:
+               pic = (pic & 0xffffffffUL) |
+                       (value << 32UL);
+               break;
+       default:
+               BUG();
+       }
+
+       write_pic(pic);
+}
+
+static inline u64 pfm_arch_read_pmd(struct pfm_context *ctx,
+                                   unsigned int cnum)
+{
+       u64 pic;
+
+       read_pic(pic);
+
+       switch (cnum) {
+       case 0:
+               return pic & 0xffffffffUL;
+       case 1:
+               return pic >> 32UL;
+       default:
+               BUG();
+               return 0;
+       }
+}
+
+/*
+ * For some CPUs, the upper bits of a counter must be set in order for the
+ * overflow interrupt to happen. On overflow, the counter has wrapped around,
+ * and the upper bits are cleared. This function may be used to set them back.
+ */
+static inline void pfm_arch_ovfl_reset_pmd(struct pfm_context *ctx,
+                                          unsigned int cnum)
+{
+       u64 val = pfm_arch_read_pmd(ctx, cnum);
+
+       /* This masks out overflow bit 31 */
+       pfm_arch_write_pmd(ctx, cnum, val);
+}
+
+/*
+ * At certain points, perfmon needs to know if monitoring has been
+ * explicitely started/stopped by user via pfm_start/pfm_stop. The
+ * information is tracked in ctx.flags.started. However on certain
+ * architectures, it may be possible to start/stop directly from
+ * user level with a single assembly instruction bypassing
+ * the kernel. This function must be used to determine by
+ * an arch-specific mean if monitoring is actually started/stopped.
+ */
+static inline int pfm_arch_is_active(struct pfm_context *ctx)
+{
+       return ctx->flags.started;
+}
+
+static inline void pfm_arch_ctxswout_sys(struct task_struct *task,
+                                        struct pfm_context *ctx)
+{
+}
+
+static inline void pfm_arch_ctxswin_sys(struct task_struct *task,
+                                       struct pfm_context *ctx)
+{
+}
+
+static inline void pfm_arch_ctxswin_thread(struct task_struct *task,
+                                          struct pfm_context *ctx)
+{
+}
+
+int  pfm_arch_is_monitoring_active(struct pfm_context *ctx);
+int  pfm_arch_ctxswout_thread(struct task_struct *task,
+                             struct pfm_context *ctx);
+void pfm_arch_stop(struct task_struct *task, struct pfm_context *ctx);
+void pfm_arch_start(struct task_struct *task, struct pfm_context *ctx);
+void pfm_arch_restore_pmds(struct pfm_context *ctx, struct pfm_event_set *set);
+void pfm_arch_restore_pmcs(struct pfm_context *ctx, struct pfm_event_set *set);
+char *pfm_arch_get_pmu_module_name(void);
+
+static inline void pfm_arch_intr_freeze_pmu(struct pfm_context *ctx,
+                                           struct pfm_event_set *set)
+{
+       pfm_arch_stop(current, ctx);
+       /*
+        * we mark monitoring as stopped to avoid
+        * certain side effects especially in
+        * pfm_switch_sets_from_intr() on
+        * pfm_arch_restore_pmcs()
+        */
+       ctx->flags.started = 0;
+}
+
+/*
+ * unfreeze PMU from pfm_do_interrupt_handler()
+ * ctx may be NULL for spurious
+ */
+static inline void pfm_arch_intr_unfreeze_pmu(struct pfm_context *ctx)
+{
+       if (!ctx)
+               return;
+
+       PFM_DBG_ovfl("state=%d", ctx->state);
+
+       ctx->flags.started = 1;
+
+       if (ctx->state == PFM_CTX_MASKED)
+               return;
+
+       pfm_arch_restore_pmcs(ctx, ctx->active_set);
+}
+
+/*
+ * this function is called from the PMU interrupt handler ONLY.
+ * On SPARC, the PMU is frozen via arch_stop, masking would be implemented
+ * via arch-stop as well. Given that the PMU is already stopped when
+ * entering the interrupt handler, we do not need to stop it again, so
+ * this function is a nop.
+ */
+static inline void pfm_arch_mask_monitoring(struct pfm_context *ctx,
+                                           struct pfm_event_set *set)
+{
+}
+
+/*
+ * on MIPS masking/unmasking uses the start/stop mechanism, so we simply
+ * need to start here.
+ */
+static inline void pfm_arch_unmask_monitoring(struct pfm_context *ctx,
+                                             struct pfm_event_set *set)
+{
+       pfm_arch_start(current, ctx);
+}
+
+static inline void pfm_arch_pmu_config_remove(void)
+{
+}
+
+static inline int pfm_arch_context_create(struct pfm_context *ctx,
+                                         u32 ctx_flags)
+{
+       return 0;
+}
+
+static inline void pfm_arch_context_free(struct pfm_context *ctx)
+{
+}
+
+/*
+ * function called from pfm_setfl_sane(). Context is locked
+ * and interrupts are masked.
+ * The value of flags is the value of ctx_flags as passed by
+ * user.
+ *
+ * function must check arch-specific set flags.
+ * Return:
+ *     1 when flags are valid
+ *      0 on error
+ */
+static inline int pfm_arch_setfl_sane(struct pfm_context *ctx, u32 flags)
+{
+       return 0;
+}
+
+static inline int pfm_arch_init(void)
+{
+       return 0;
+}
+
+static inline void pfm_arch_init_percpu(void)
+{
+}
+
+static inline int pfm_arch_load_context(struct pfm_context *ctx)
+{
+       return 0;
+}
+
+static inline void pfm_arch_unload_context(struct pfm_context *ctx)
+{}
+
+extern void perfmon_interrupt(struct pt_regs *);
+
+static inline int pfm_arch_pmu_acquire(u64 *unavail_pmcs, u64 *unavail_pmds)
+{
+       return register_perfctr_intr(perfmon_interrupt);
+}
+
+static inline void pfm_arch_pmu_release(void)
+{
+       release_perfctr_intr(perfmon_interrupt);
+}
+
+static inline void pfm_arch_arm_handle_work(struct task_struct *task)
+{}
+
+static inline void pfm_arch_disarm_handle_work(struct task_struct *task)
+{}
+
+static inline int pfm_arch_pmu_config_init(struct pfm_pmu_config *cfg)
+{
+       return 0;
+}
+
+static inline int pfm_arch_get_base_syscall(void)
+{
+       return __NR_pfm_create_context;
+}
+
+struct pfm_arch_context {
+       /* empty */
+};
+
+#define PFM_ARCH_CTX_SIZE      sizeof(struct pfm_arch_context)
+/*
+ * SPARC needs extra alignment for the sampling buffer
+ */
+#define PFM_ARCH_SMPL_ALIGN_SIZE       (16 * 1024)
+
+static inline void pfm_cacheflush(void *addr, unsigned int len)
+{
+}
+
+#endif /* CONFIG_PERFMON */
+
+#endif /* __KERNEL__ */
+
+#endif /* _SPARC64_PERFMON_KERN_H_ */
diff --git a/arch/sparc/include/asm/system_64.h b/arch/sparc/include/asm/system_64.h
index db9e742..2a9ddb9 100644
--- a/arch/sparc/include/asm/system_64.h
+++ b/arch/sparc/include/asm/system_64.h
@@ -30,6 +30,9 @@ enum sparc_cpu {
 #define ARCH_SUN4C_SUN4 0
 #define ARCH_SUN4 0
 
+extern char *sparc_cpu_type;
+extern char *sparc_fpu_type;
+extern char *sparc_pmu_type;
 extern char reboot_command[];
 
 /* These are here in an effort to more fully work around Spitfire Errata
@@ -104,15 +107,13 @@ do {      __asm__ __volatile__("ba,pt     %%xcc, 1f\n\t" \
 #define write_pcr(__p) __asm__ __volatile__("wr        %0, 0x0, %%pcr" : : "r" (__p))
 #define read_pic(__p)  __asm__ __volatile__("rd %%pic, %0" : "=r" (__p))
 
-/* Blackbird errata workaround.  See commentary in
- * arch/sparc64/kernel/smp.c:smp_percpu_timer_interrupt()
- * for more information.
- */
-#define reset_pic()                                                    \
-       __asm__ __volatile__("ba,pt     %xcc, 99f\n\t"          \
+/* Blackbird errata workaround.  */
+#define write_pic(val)                                         \
+       __asm__ __volatile__("ba,pt     %%xcc, 99f\n\t"         \
                             ".align    64\n"                   \
-                         "99:wr        %g0, 0x0, %pic\n\t"     \
-                            "rd        %pic, %g0")
+                         "99:wr        %0, 0x0, %%pic\n\t"     \
+                            "rd        %%pic, %%g0" : : "r" (val))
+#define reset_pic()    write_pic(0)
 
 #ifndef __ASSEMBLY__
 
@@ -145,14 +146,10 @@ do {                                              \
         * and 2 stores in this critical code path.  -DaveM
         */
 #define switch_to(prev, next, last)                                    \
-do {   if (test_thread_flag(TIF_PERFCTR)) {                            \
-               unsigned long __tmp;                                    \
-               read_pcr(__tmp);                                        \
-               current_thread_info()->pcr_reg = __tmp;                 \
-               read_pic(__tmp);                                        \
-               current_thread_info()->kernel_cntd0 += (unsigned int)(__tmp);\
-               current_thread_info()->kernel_cntd1 += ((__tmp) >> 32); \
-       }                                                               \
+do {   if (test_tsk_thread_flag(prev, TIF_PERFMON_CTXSW))              \
+               pfm_ctxsw_out(prev, next);                              \
+       if (test_tsk_thread_flag(next, TIF_PERFMON_CTXSW))              \
+               pfm_ctxsw_in(prev, next);                               \
        flush_tlb_pending();                                            \
        save_and_clear_fpu();                                           \
        /* If you are tempted to conditionalize the following */        \
@@ -197,11 +194,6 @@ do {       if (test_thread_flag(TIF_PERFCTR)) {                            \
                "l1", "l2", "l3", "l4", "l5", "l6", "l7",               \
          "i0", "i1", "i2", "i3", "i4", "i5",                           \
          "o0", "o1", "o2", "o3", "o4", "o5",       "o7");              \
-       /* If you fuck with this, update ret_from_syscall code too. */  \
-       if (test_thread_flag(TIF_PERFCTR)) {                            \
-               write_pcr(current_thread_info()->pcr_reg);              \
-               reset_pic();                                            \
-       }                                                               \
 } while(0)
 
 static inline unsigned long xchg32(__volatile__ unsigned int *m, unsigned int val)
diff --git a/arch/sparc/include/asm/thread_info_64.h b/arch/sparc/include/asm/thread_info_64.h
index c0a737d..53857f7 100644
--- a/arch/sparc/include/asm/thread_info_64.h
+++ b/arch/sparc/include/asm/thread_info_64.h
@@ -58,11 +58,6 @@ struct thread_info {
        unsigned long           gsr[7];
        unsigned long           xfsr[7];
 
-       __u64                   __user *user_cntd0;
-       __u64                   __user *user_cntd1;
-       __u64                   kernel_cntd0, kernel_cntd1;
-       __u64                   pcr_reg;
-
        struct restart_block    restart_block;
 
        struct pt_regs          *kern_una_regs;
@@ -96,15 +91,10 @@ struct thread_info {
 #define TI_RWIN_SPTRS  0x000003c8
 #define TI_GSR         0x00000400
 #define TI_XFSR                0x00000438
-#define TI_USER_CNTD0  0x00000470
-#define TI_USER_CNTD1  0x00000478
-#define TI_KERN_CNTD0  0x00000480
-#define TI_KERN_CNTD1  0x00000488
-#define TI_PCR         0x00000490
-#define TI_RESTART_BLOCK 0x00000498
-#define TI_KUNA_REGS   0x000004c0
-#define TI_KUNA_INSN   0x000004c8
-#define TI_FPREGS      0x00000500
+#define TI_RESTART_BLOCK 0x00000470
+#define TI_KUNA_REGS   0x00000498
+#define TI_KUNA_INSN   0x000004a0
+#define TI_FPREGS      0x000004c0
 
 /* We embed this in the uppermost byte of thread_info->flags */
 #define FAULT_CODE_WRITE       0x01    /* Write access, implies D-TLB     */
@@ -222,11 +212,11 @@ register struct thread_info *current_thread_info_reg asm("g6");
 #define TIF_NOTIFY_RESUME      1       /* callback before returning to user */
 #define TIF_SIGPENDING         2       /* signal pending */
 #define TIF_NEED_RESCHED       3       /* rescheduling necessary */
-#define TIF_PERFCTR            4       /* performance counters active */
+/* Bit 4 is available */
 #define TIF_UNALIGNED          5       /* allowed to do unaligned accesses */
 /* flag bit 6 is available */
 #define TIF_32BIT              7       /* 32-bit binary */
-/* flag bit 8 is available */
+#define TIF_PERFMON_WORK       8       /* work for pfm_handle_work() */
 #define TIF_SECCOMP            9       /* secure computing */
 #define TIF_SYSCALL_AUDIT      10      /* syscall auditing active */
 /* flag bit 11 is available */
@@ -237,22 +227,24 @@ register struct thread_info *current_thread_info_reg asm("g6");
 #define TIF_ABI_PENDING                12
 #define TIF_MEMDIE             13
 #define TIF_POLLING_NRFLAG     14
+#define TIF_PERFMON_CTXSW      15      /* perfmon needs ctxsw calls */
 
 #define _TIF_SYSCALL_TRACE     (1<<TIF_SYSCALL_TRACE)
 #define _TIF_NOTIFY_RESUME     (1<<TIF_NOTIFY_RESUME)
 #define _TIF_SIGPENDING                (1<<TIF_SIGPENDING)
 #define _TIF_NEED_RESCHED      (1<<TIF_NEED_RESCHED)
-#define _TIF_PERFCTR           (1<<TIF_PERFCTR)
 #define _TIF_UNALIGNED         (1<<TIF_UNALIGNED)
 #define _TIF_32BIT             (1<<TIF_32BIT)
+#define _TIF_PERFMON_WORK      (1<<TIF_PERFMON_WORK)
 #define _TIF_SECCOMP           (1<<TIF_SECCOMP)
 #define _TIF_SYSCALL_AUDIT     (1<<TIF_SYSCALL_AUDIT)
 #define _TIF_ABI_PENDING       (1<<TIF_ABI_PENDING)
 #define _TIF_POLLING_NRFLAG    (1<<TIF_POLLING_NRFLAG)
+#define _TIF_PERFMON_CTXSW     (1<<TIF_PERFMON_CTXSW)
 
 #define _TIF_USER_WORK_MASK    ((0xff << TI_FLAG_WSAVED_SHIFT) | \
                                 _TIF_DO_NOTIFY_RESUME_MASK | \
-                                _TIF_NEED_RESCHED | _TIF_PERFCTR)
+                                _TIF_NEED_RESCHED)
 #define _TIF_DO_NOTIFY_RESUME_MASK     (_TIF_NOTIFY_RESUME | _TIF_SIGPENDING)
 
 /*
diff --git a/arch/sparc/include/asm/unistd_32.h b/arch/sparc/include/asm/unistd_32.h
index 648643a..efe4d86 100644
--- a/arch/sparc/include/asm/unistd_32.h
+++ b/arch/sparc/include/asm/unistd_32.h
@@ -338,8 +338,20 @@
 #define __NR_dup3              320
 #define __NR_pipe2             321
 #define __NR_inotify_init1     322
+#define __NR_pfm_create_context 323
+#define __NR_pfm_write_pmcs    324
+#define __NR_pfm_write_pmds    325
+#define __NR_pfm_read_pmds     326
+#define __NR_pfm_load_context  327
+#define __NR_pfm_start         328
+#define __NR_pfm_stop          329
+#define __NR_pfm_restart       330
+#define __NR_pfm_create_evtsets        331
+#define __NR_pfm_getinfo_evtsets 332
+#define __NR_pfm_delete_evtsets        333
+#define __NR_pfm_unload_context        334
 
-#define NR_SYSCALLS            323
+#define NR_SYSCALLS            325
 
 /* Sparc 32-bit only has the "setresuid32", "getresuid32" variants,
  * it never had the plain ones and there is no value to adding those
diff --git a/arch/sparc/include/asm/unistd_64.h b/arch/sparc/include/asm/unistd_64.h
index c5cc0e0..cbbb0b5 100644
--- a/arch/sparc/include/asm/unistd_64.h
+++ b/arch/sparc/include/asm/unistd_64.h
@@ -340,8 +340,20 @@
 #define __NR_dup3              320
 #define __NR_pipe2             321
 #define __NR_inotify_init1     322
+#define __NR_pfm_create_context 323
+#define __NR_pfm_write_pmcs    324
+#define __NR_pfm_write_pmds    325
+#define __NR_pfm_read_pmds     326
+#define __NR_pfm_load_context  327
+#define __NR_pfm_start         328
+#define __NR_pfm_stop          329
+#define __NR_pfm_restart       330
+#define __NR_pfm_create_evtsets        331
+#define __NR_pfm_getinfo_evtsets 332
+#define __NR_pfm_delete_evtsets        333
+#define __NR_pfm_unload_context        334
 
-#define NR_SYSCALLS            323
+#define NR_SYSCALLS            335
 
 #ifdef __KERNEL__
 #define __ARCH_WANT_IPC_PARSE_VERSION
diff --git a/arch/sparc/kernel/systbls.S b/arch/sparc/kernel/systbls.S
index e1b9233..727e4e7 100644
--- a/arch/sparc/kernel/systbls.S
+++ b/arch/sparc/kernel/systbls.S
@@ -81,4 +81,6 @@ sys_call_table:
 /*305*/        .long sys_set_mempolicy, sys_kexec_load, sys_move_pages, sys_getcpu, sys_epoll_pwait
 /*310*/        .long sys_utimensat, sys_signalfd, sys_timerfd_create, sys_eventfd, sys_fallocate
 /*315*/        .long sys_timerfd_settime, sys_timerfd_gettime, sys_signalfd4, sys_eventfd2, sys_epoll_create1
-/*320*/        .long sys_dup3, sys_pipe2, sys_inotify_init1
+/*320*/        .long sys_dup3, sys_pipe2, sys_inotify_init1, sys_pfm_create_context, sys_pfm_write_pmcs, sys_pfm_write_pmds
+/*325*/ .long sys_pfm_write_pmds, sys_pfm_read_pmds, sys_pfm_load_context, sys_pfm_start, sys_pfm_stop
+/*330*/        .long sys_pfm_restart, sys_pfm_create_evtsets, sys_pfm_getinfo_evtsets, sys_pfm_delete_evtsets, sys_pfm_unload_context
diff --git a/arch/sparc64/Kconfig b/arch/sparc64/Kconfig
index 36b4b7a..5555d1e 100644
--- a/arch/sparc64/Kconfig
+++ b/arch/sparc64/Kconfig
@@ -401,6 +401,8 @@ source "drivers/sbus/char/Kconfig"
 
 source "fs/Kconfig"
 
+source "arch/sparc64/perfmon/Kconfig"
+
 source "arch/sparc64/Kconfig.debug"
 
 source "security/Kconfig"
diff --git a/arch/sparc64/Makefile b/arch/sparc64/Makefile
index b785a39..646731c 100644
--- a/arch/sparc64/Makefile
+++ b/arch/sparc64/Makefile
@@ -32,6 +32,8 @@ core-y                                += arch/sparc64/math-emu/
 libs-y                         += arch/sparc64/prom/ arch/sparc64/lib/
 drivers-$(CONFIG_OPROFILE)     += arch/sparc64/oprofile/
 
+core-$(CONFIG_PERFMON)         += arch/sparc64/perfmon/
+
 boot := arch/sparc64/boot
 
 image tftpboot.img vmlinux.aout: vmlinux
diff --git a/arch/sparc64/kernel/cpu.c b/arch/sparc64/kernel/cpu.c
index 0097c08..f839f84 100644
--- a/arch/sparc64/kernel/cpu.c
+++ b/arch/sparc64/kernel/cpu.c
@@ -20,16 +20,17 @@
 DEFINE_PER_CPU(cpuinfo_sparc, __cpu_data) = { 0 };
 
 struct cpu_iu_info {
-  short manuf;
-  short impl;
-  char* cpu_name;   /* should be enough I hope... */
+       short manuf;
+       short impl;
+       char *cpu_name;
+       char *pmu_name;
 };
 
 struct cpu_fp_info {
-  short manuf;
-  short impl;
-  char fpu_vers;
-  char* fp_name;
+       short manuf;
+       short impl;
+       char fpu_vers;
+       char* fp_name;
 };
 
 static struct cpu_fp_info linux_sparc_fpu[] = {
@@ -49,23 +50,24 @@ static struct cpu_fp_info linux_sparc_fpu[] = {
 #define NSPARCFPU  ARRAY_SIZE(linux_sparc_fpu)
 
 static struct cpu_iu_info linux_sparc_chips[] = {
-  { 0x17, 0x10, "TI UltraSparc I   (SpitFire)"},
-  { 0x22, 0x10, "TI UltraSparc I   (SpitFire)"},
-  { 0x17, 0x11, "TI UltraSparc II  (BlackBird)"},
-  { 0x17, 0x12, "TI UltraSparc IIi (Sabre)"},
-  { 0x17, 0x13, "TI UltraSparc IIe (Hummingbird)"},
-  { 0x3e, 0x14, "TI UltraSparc III (Cheetah)"},
-  { 0x3e, 0x15, "TI UltraSparc III+ (Cheetah+)"},
-  { 0x3e, 0x16, "TI UltraSparc IIIi (Jalapeno)"},
-  { 0x3e, 0x18, "TI UltraSparc IV (Jaguar)"},
-  { 0x3e, 0x19, "TI UltraSparc IV+ (Panther)"},
-  { 0x3e, 0x22, "TI UltraSparc IIIi+ (Serrano)"},
-};
+   { 0x17, 0x10, "TI UltraSparc I   (SpitFire)",    "ultra12"},
+   { 0x22, 0x10, "TI UltraSparc I   (SpitFire)",    "ultra12"},
+   { 0x17, 0x11, "TI UltraSparc II  (BlackBird)",   "ultra12"},
+   { 0x17, 0x12, "TI UltraSparc IIi (Sabre)",       "ultra12"},
+   { 0x17, 0x13, "TI UltraSparc IIe (Hummingbird)", "ultra12"},
+   { 0x3e, 0x14, "TI UltraSparc III (Cheetah)",     "ultra3"},
+   { 0x3e, 0x15, "TI UltraSparc III+ (Cheetah+)",   "ultra3+"},
+   { 0x3e, 0x16, "TI UltraSparc IIIi (Jalapeno)",   "ultra3i"},
+   { 0x3e, 0x18, "TI UltraSparc IV (Jaguar)",       "ultra4"},
+   { 0x3e, 0x19, "TI UltraSparc IV+ (Panther)",     "ultra4+"},
+   { 0x3e, 0x22, "TI UltraSparc IIIi+ (Serrano)",   "ultra3+"},
+ };
 
 #define NSPARCCHIPS  ARRAY_SIZE(linux_sparc_chips)
 
 char *sparc_cpu_type;
 char *sparc_fpu_type;
+char *sparc_pmu_type;
 
 static void __init sun4v_cpu_probe(void)
 {
@@ -73,11 +75,13 @@ static void __init sun4v_cpu_probe(void)
        case SUN4V_CHIP_NIAGARA1:
                sparc_cpu_type = "UltraSparc T1 (Niagara)";
                sparc_fpu_type = "UltraSparc T1 integrated FPU";
+               sparc_pmu_type = "niagara";
                break;
 
        case SUN4V_CHIP_NIAGARA2:
                sparc_cpu_type = "UltraSparc T2 (Niagara2)";
                sparc_fpu_type = "UltraSparc T2 integrated FPU";
+               sparc_pmu_type = "niagara2";
                break;
 
        default:
@@ -85,6 +89,7 @@ static void __init sun4v_cpu_probe(void)
                       prom_cpu_compatible);
                sparc_cpu_type = "Unknown SUN4V CPU";
                sparc_fpu_type = "Unknown SUN4V FPU";
+               sparc_pmu_type = "Unknown SUN4V PMU";
                break;
        }
 }
@@ -117,6 +122,8 @@ retry:
                        if (linux_sparc_chips[i].impl == impl) {
                                sparc_cpu_type =
                                        linux_sparc_chips[i].cpu_name;
+                               sparc_pmu_type =
+                                       linux_sparc_chips[i].pmu_name;
                                break;
                        }
                }
@@ -134,7 +141,7 @@ retry:
                        printk("DEBUG: manuf[%lx] impl[%lx]\n",
                               manuf, impl);
                }
-               sparc_cpu_type = "Unknown CPU";
+               sparc_pmu_type = "Unknown PMU";
        }
 
        for (i = 0; i < NSPARCFPU; i++) {
diff --git a/arch/sparc64/kernel/hvcalls.S b/arch/sparc64/kernel/hvcalls.S
index a2810f3..b9f508c 100644
--- a/arch/sparc64/kernel/hvcalls.S
+++ b/arch/sparc64/kernel/hvcalls.S
@@ -884,3 +884,44 @@ sun4v_mmu_demap_all:
        retl
         nop
        .size   sun4v_mmu_demap_all, .-sun4v_mmu_demap_all
+
+       .globl  sun4v_niagara_getperf
+       .type   sun4v_niagara_getperf,#function
+sun4v_niagara_getperf:
+       mov     %o0, %o4
+       mov     HV_FAST_GET_PERFREG, %o5
+       ta      HV_FAST_TRAP
+       stx     %o1, [%o4]
+       retl
+        nop
+       .size   sun4v_niagara_getperf, .-sun4v_niagara_getperf
+
+       .globl  sun4v_niagara_setperf
+       .type   sun4v_niagara_setperf,#function
+sun4v_niagara_setperf:
+       mov     HV_FAST_SET_PERFREG, %o5
+       ta      HV_FAST_TRAP
+       retl
+        nop
+       .size   sun4v_niagara_setperf, .-sun4v_niagara_setperf
+
+       .globl  sun4v_niagara2_getperf
+       .type   sun4v_niagara2_getperf,#function
+sun4v_niagara2_getperf:
+       mov     %o0, %o4
+       mov     HV_FAST_N2_GET_PERFREG, %o5
+       ta      HV_FAST_TRAP
+       stx     %o1, [%o4]
+       retl
+        nop
+       .size   sun4v_niagara2_getperf, .-sun4v_niagara2_getperf
+
+       .globl  sun4v_niagara2_setperf
+       .type   sun4v_niagara2_setperf,#function
+sun4v_niagara2_setperf:
+       mov     HV_FAST_N2_SET_PERFREG, %o5
+       ta      HV_FAST_TRAP
+       retl
+        nop
+       .size   sun4v_niagara2_setperf, .-sun4v_niagara2_setperf
+
diff --git a/arch/sparc64/kernel/irq.c b/arch/sparc64/kernel/irq.c
index 7495bc7..e2bcca5 100644
--- a/arch/sparc64/kernel/irq.c
+++ b/arch/sparc64/kernel/irq.c
@@ -749,6 +749,20 @@ void handler_irq(int irq, struct pt_regs *regs)
        irq_exit();
        set_irq_regs(old_regs);
 }
+static void unhandled_perf_irq(struct pt_regs *regs)
+{
+       unsigned long pcr, pic;
+
+       read_pcr(pcr);
+       read_pic(pic);
+
+       write_pcr(0);
+
+       printk(KERN_EMERG "CPU %d: Got unexpected perf counter IRQ.\n",
+              smp_processor_id());
+       printk(KERN_EMERG "CPU %d: PCR[%016lx] PIC[%016lx]\n",
+              smp_processor_id(), pcr, pic);
+}
 
 void do_softirq(void)
 {
@@ -776,6 +790,55 @@ void do_softirq(void)
        local_irq_restore(flags);
 }
 
+/* Almost a direct copy of the powerpc PMC code.  */
+static DEFINE_SPINLOCK(perf_irq_lock);
+static void *perf_irq_owner_caller; /* mostly for debugging */
+static void (*perf_irq)(struct pt_regs *regs) = unhandled_perf_irq;
+
+/* Invoked from level 15 PIL handler in trap table.  */
+void perfctr_irq(int irq, struct pt_regs *regs)
+{
+       clear_softint(1 << irq);
+       perf_irq(regs);
+}
+
+int register_perfctr_intr(void (*handler)(struct pt_regs *))
+{
+       int ret;
+
+       if (!handler)
+               return -EINVAL;
+
+       spin_lock(&perf_irq_lock);
+       if (perf_irq != unhandled_perf_irq) {
+               printk(KERN_WARNING "register_perfctr_intr: "
+                      "perf IRQ busy (reserved by caller %p)\n",
+                      perf_irq_owner_caller);
+               ret = -EBUSY;
+               goto out;
+       }
+
+       perf_irq_owner_caller = __builtin_return_address(0);
+       perf_irq = handler;
+
+       ret = 0;
+out:
+       spin_unlock(&perf_irq_lock);
+
+       return ret;
+}
+EXPORT_SYMBOL_GPL(register_perfctr_intr);
+
+void release_perfctr_intr(void (*handler)(struct pt_regs *))
+{
+       spin_lock(&perf_irq_lock);
+       perf_irq_owner_caller = NULL;
+       perf_irq = unhandled_perf_irq;
+       spin_unlock(&perf_irq_lock);
+}
+EXPORT_SYMBOL_GPL(release_perfctr_intr);
+
+
 #ifdef CONFIG_HOTPLUG_CPU
 void fixup_irqs(void)
 {
diff --git a/arch/sparc64/kernel/process.c b/arch/sparc64/kernel/process.c
index 15f4178..7282d21 100644
--- a/arch/sparc64/kernel/process.c
+++ b/arch/sparc64/kernel/process.c
@@ -30,6 +30,7 @@
 #include <linux/cpu.h>
 #include <linux/elfcore.h>
 #include <linux/sysrq.h>
+#include <linux/perfmon_kern.h>
 
 #include <asm/oplib.h>
 #include <asm/uaccess.h>
@@ -385,11 +386,7 @@ void exit_thread(void)
                        t->utraps[0]--;
        }
 
-       if (test_and_clear_thread_flag(TIF_PERFCTR)) {
-               t->user_cntd0 = t->user_cntd1 = NULL;
-               t->pcr_reg = 0;
-               write_pcr(0);
-       }
+       pfm_exit_thread();
 }
 
 void flush_thread(void)
@@ -411,13 +408,6 @@ void flush_thread(void)
 
        set_thread_wsaved(0);
 
-       /* Turn off performance counters if on. */
-       if (test_and_clear_thread_flag(TIF_PERFCTR)) {
-               t->user_cntd0 = t->user_cntd1 = NULL;
-               t->pcr_reg = 0;
-               write_pcr(0);
-       }
-
        /* Clear FPU register state. */
        t->fpsaved[0] = 0;
        
@@ -631,16 +621,6 @@ int copy_thread(int nr, unsigned long clone_flags, unsigned long sp,
                t->kregs->u_regs[UREG_FP] =
                  ((unsigned long) child_sf) - STACK_BIAS;
 
-               /* Special case, if we are spawning a kernel thread from
-                * a userspace task (usermode helper, NFS or similar), we
-                * must disable performance counters in the child because
-                * the address space and protection realm are changing.
-                */
-               if (t->flags & _TIF_PERFCTR) {
-                       t->user_cntd0 = t->user_cntd1 = NULL;
-                       t->pcr_reg = 0;
-                       t->flags &= ~_TIF_PERFCTR;
-               }
                t->flags |= ((long)ASI_P << TI_FLAG_CURRENT_DS_SHIFT);
                t->kregs->u_regs[UREG_G6] = (unsigned long) t;
                t->kregs->u_regs[UREG_G4] = (unsigned long) t->task;
@@ -673,6 +653,8 @@ int copy_thread(int nr, unsigned long clone_flags, unsigned long sp,
        if (clone_flags & CLONE_SETTLS)
                t->kregs->u_regs[UREG_G7] = regs->u_regs[UREG_I3];
 
+       pfm_copy_thread(p);
+
        return 0;
 }
 
diff --git a/arch/sparc64/kernel/rtrap.S b/arch/sparc64/kernel/rtrap.S
index 97a993c..c2af29d 100644
--- a/arch/sparc64/kernel/rtrap.S
+++ b/arch/sparc64/kernel/rtrap.S
@@ -65,55 +65,14 @@ __handle_user_windows:
                ba,pt                   %xcc, __handle_user_windows_continue
 
                 andn                   %l1, %l4, %l1
-__handle_perfctrs:
-               call                    update_perfctrs
-                wrpr                   %g0, RTRAP_PSTATE, %pstate
-               wrpr                    %g0, RTRAP_PSTATE_IRQOFF, %pstate
-               ldub                    [%g6 + TI_WSAVED], %o2
-               brz,pt                  %o2, 1f
-                nop
-               /* Redo userwin+sched+sig checks */
-               call                    fault_in_user_windows
-
-                wrpr                   %g0, RTRAP_PSTATE, %pstate
-               wrpr                    %g0, RTRAP_PSTATE_IRQOFF, %pstate
-               ldx                     [%g6 + TI_FLAGS], %l0
-               andcc                   %l0, _TIF_NEED_RESCHED, %g0
-               be,pt                   %xcc, 1f
-
-                nop
-               call                    schedule
-                wrpr                   %g0, RTRAP_PSTATE, %pstate
-               wrpr                    %g0, RTRAP_PSTATE_IRQOFF, %pstate
-               ldx                     [%g6 + TI_FLAGS], %l0
-1:             andcc                   %l0, _TIF_DO_NOTIFY_RESUME_MASK, %g0
-
-               be,pt                   %xcc, __handle_perfctrs_continue
-                sethi                  %hi(TSTATE_PEF), %o0
-               mov                     %l5, %o1
-               add                     %sp, PTREGS_OFF, %o0
-               mov                     %l0, %o2
-               call                    do_notify_resume
-
-                wrpr                   %g0, RTRAP_PSTATE, %pstate
-               wrpr                    %g0, RTRAP_PSTATE_IRQOFF, %pstate
-               /* Signal delivery can modify pt_regs tstate, so we must
-                * reload it.
-                */
-               ldx                     [%sp + PTREGS_OFF + PT_V9_TSTATE], %l1
-               sethi                   %hi(0xf << 20), %l4
-               and                     %l1, %l4, %l4
-               andn                    %l1, %l4, %l1
-               ba,pt                   %xcc, __handle_perfctrs_continue
-
-                sethi                  %hi(TSTATE_PEF), %o0
 __handle_userfpu:
                rd                      %fprs, %l5
                andcc                   %l5, FPRS_FEF, %g0
                sethi                   %hi(TSTATE_PEF), %o0
                be,a,pn                 %icc, __handle_userfpu_continue
                 andn                   %l1, %o0, %l1
-               ba,a,pt                 %xcc, __handle_userfpu_continue
+               ba,pt                   %xcc, __handle_userfpu_continue
+                nop
 
 __handle_signal:
                mov                     %l5, %o1
@@ -202,12 +161,8 @@ __handle_signal_continue:
                brnz,pn                 %o2, __handle_user_windows
                 nop
 __handle_user_windows_continue:
-               ldx                     [%g6 + TI_FLAGS], %l5
-               andcc                   %l5, _TIF_PERFCTR, %g0
                sethi                   %hi(TSTATE_PEF), %o0
-               bne,pn                  %xcc, __handle_perfctrs
-__handle_perfctrs_continue:
-                andcc                  %l1, %o0, %g0
+               andcc                   %l1, %o0, %g0
 
                /* This fpdepth clear is necessary for non-syscall rtraps only */
 user_nowork:
diff --git a/arch/sparc64/kernel/setup.c b/arch/sparc64/kernel/setup.c
index c8b03a4..248aa1f 100644
--- a/arch/sparc64/kernel/setup.c
+++ b/arch/sparc64/kernel/setup.c
@@ -352,6 +352,7 @@ static int show_cpuinfo(struct seq_file *m, void *__unused)
        seq_printf(m, 
                   "cpu\t\t: %s\n"
                   "fpu\t\t: %s\n"
+                  "pmu\t\t: %s\n"
                   "prom\t\t: %s\n"
                   "type\t\t: %s\n"
                   "ncpus probed\t: %d\n"
@@ -364,6 +365,7 @@ static int show_cpuinfo(struct seq_file *m, void *__unused)
                   ,
                   sparc_cpu_type,
                   sparc_fpu_type,
+                  sparc_pmu_type,
                   prom_version,
                   ((tlb_type == hypervisor) ?
                    "sun4v" :
diff --git a/arch/sparc64/kernel/signal.c b/arch/sparc64/kernel/signal.c
index ec82d76..cea1082 100644
--- a/arch/sparc64/kernel/signal.c
+++ b/arch/sparc64/kernel/signal.c
@@ -23,6 +23,7 @@
 #include <linux/tty.h>
 #include <linux/binfmts.h>
 #include <linux/bitops.h>
+#include <linux/perfmon_kern.h>
 
 #include <asm/uaccess.h>
 #include <asm/ptrace.h>
@@ -608,6 +609,9 @@ static void do_signal(struct pt_regs *regs, unsigned long orig_i0)
 
 void do_notify_resume(struct pt_regs *regs, unsigned long orig_i0, unsigned long thread_info_flags)
 {
+       if (thread_info_flags & _TIF_PERFMON_WORK)
+               pfm_handle_work(regs);
+
        if (thread_info_flags & _TIF_SIGPENDING)
                do_signal(regs, orig_i0);
        if (thread_info_flags & _TIF_NOTIFY_RESUME) {
diff --git a/arch/sparc64/kernel/sys_sparc.c b/arch/sparc64/kernel/sys_sparc.c
index 39749e3..384004b 100644
--- a/arch/sparc64/kernel/sys_sparc.c
+++ b/arch/sparc64/kernel/sys_sparc.c
@@ -26,7 +26,6 @@
 
 #include <asm/uaccess.h>
 #include <asm/utrap.h>
-#include <asm/perfctr.h>
 #include <asm/unistd.h>
 
 #include "entry.h"
@@ -791,106 +790,10 @@ asmlinkage long sys_rt_sigaction(int sig,
        return ret;
 }
 
-/* Invoked by rtrap code to update performance counters in
- * user space.
- */
-asmlinkage void update_perfctrs(void)
-{
-       unsigned long pic, tmp;
-
-       read_pic(pic);
-       tmp = (current_thread_info()->kernel_cntd0 += (unsigned int)pic);
-       __put_user(tmp, current_thread_info()->user_cntd0);
-       tmp = (current_thread_info()->kernel_cntd1 += (pic >> 32));
-       __put_user(tmp, current_thread_info()->user_cntd1);
-       reset_pic();
-}
-
 asmlinkage long sys_perfctr(int opcode, unsigned long arg0, unsigned long arg1, unsigned long arg2)
 {
-       int err = 0;
-
-       switch(opcode) {
-       case PERFCTR_ON:
-               current_thread_info()->pcr_reg = arg2;
-               current_thread_info()->user_cntd0 = (u64 __user *) arg0;
-               current_thread_info()->user_cntd1 = (u64 __user *) arg1;
-               current_thread_info()->kernel_cntd0 =
-                       current_thread_info()->kernel_cntd1 = 0;
-               write_pcr(arg2);
-               reset_pic();
-               set_thread_flag(TIF_PERFCTR);
-               break;
-
-       case PERFCTR_OFF:
-               err = -EINVAL;
-               if (test_thread_flag(TIF_PERFCTR)) {
-                       current_thread_info()->user_cntd0 =
-                               current_thread_info()->user_cntd1 = NULL;
-                       current_thread_info()->pcr_reg = 0;
-                       write_pcr(0);
-                       clear_thread_flag(TIF_PERFCTR);
-                       err = 0;
-               }
-               break;
-
-       case PERFCTR_READ: {
-               unsigned long pic, tmp;
-
-               if (!test_thread_flag(TIF_PERFCTR)) {
-                       err = -EINVAL;
-                       break;
-               }
-               read_pic(pic);
-               tmp = (current_thread_info()->kernel_cntd0 += (unsigned int)pic);
-               err |= __put_user(tmp, current_thread_info()->user_cntd0);
-               tmp = (current_thread_info()->kernel_cntd1 += (pic >> 32));
-               err |= __put_user(tmp, current_thread_info()->user_cntd1);
-               reset_pic();
-               break;
-       }
-
-       case PERFCTR_CLRPIC:
-               if (!test_thread_flag(TIF_PERFCTR)) {
-                       err = -EINVAL;
-                       break;
-               }
-               current_thread_info()->kernel_cntd0 =
-                       current_thread_info()->kernel_cntd1 = 0;
-               reset_pic();
-               break;
-
-       case PERFCTR_SETPCR: {
-               u64 __user *user_pcr = (u64 __user *)arg0;
-
-               if (!test_thread_flag(TIF_PERFCTR)) {
-                       err = -EINVAL;
-                       break;
-               }
-               err |= __get_user(current_thread_info()->pcr_reg, user_pcr);
-               write_pcr(current_thread_info()->pcr_reg);
-               current_thread_info()->kernel_cntd0 =
-                       current_thread_info()->kernel_cntd1 = 0;
-               reset_pic();
-               break;
-       }
-
-       case PERFCTR_GETPCR: {
-               u64 __user *user_pcr = (u64 __user *)arg0;
-
-               if (!test_thread_flag(TIF_PERFCTR)) {
-                       err = -EINVAL;
-                       break;
-               }
-               err |= __put_user(current_thread_info()->pcr_reg, user_pcr);
-               break;
-       }
-
-       default:
-               err = -EINVAL;
-               break;
-       };
-       return err;
+       /* Superceded by perfmon2 */
+       return -ENOSYS;
 }
 
 /*
diff --git a/arch/sparc64/kernel/syscalls.S b/arch/sparc64/kernel/syscalls.S
index a2f2427..b20bf1e 100644
--- a/arch/sparc64/kernel/syscalls.S
+++ b/arch/sparc64/kernel/syscalls.S
@@ -117,26 +117,9 @@ ret_from_syscall:
        stb     %g0, [%g6 + TI_NEW_CHILD]
        ldx     [%g6 + TI_FLAGS], %l0
        call    schedule_tail
-        mov    %g7, %o0
-       andcc   %l0, _TIF_PERFCTR, %g0
-       be,pt   %icc, 1f
-        nop
-       ldx     [%g6 + TI_PCR], %o7
-       wr      %g0, %o7, %pcr
-
-       /* Blackbird errata workaround.  See commentary in
-        * smp.c:smp_percpu_timer_interrupt() for more
-        * information.
-        */
-       ba,pt   %xcc, 99f
-        nop
-
-       .align  64
-99:    wr      %g0, %g0, %pic
-       rd      %pic, %g0
-
-1:     ba,pt   %xcc, ret_sys_call
-        ldx    [%sp + PTREGS_OFF + PT_V9_I0], %o0
+       mov             %g7, %o0
+       ba,pt           %xcc, ret_sys_call
+       ldx     [%sp + PTREGS_OFF + PT_V9_I0], %o0
 
        .globl  sparc_exit
        .type   sparc_exit,#function
diff --git a/arch/sparc64/kernel/systbls.S b/arch/sparc64/kernel/systbls.S
index 0fdbf3b..1a1a296 100644
--- a/arch/sparc64/kernel/systbls.S
+++ b/arch/sparc64/kernel/systbls.S
@@ -82,7 +82,9 @@ sys_call_table32:
        .word compat_sys_set_mempolicy, compat_sys_kexec_load, compat_sys_move_pages, sys_getcpu, compat_sys_epoll_pwait
 /*310*/        .word compat_sys_utimensat, compat_sys_signalfd, sys_timerfd_create, sys_eventfd, compat_sys_fallocate
        .word compat_sys_timerfd_settime, compat_sys_timerfd_gettime, compat_sys_signalfd4, sys_eventfd2, sys_epoll_create1
-/*320*/        .word sys_dup3, sys_pipe2, sys_inotify_init1
+/*320*/        .word sys_dup3, sys_pipe2, sys_inotify_init1, sys_pfm_create_context, sys_pfm_write_pmcs
+       .word sys_pfm_write_pmds, sys_pfm_read_pmds, sys_pfm_load_context, sys_pfm_start, sys_pfm_stop
+/*330*/        .word sys_pfm_restart, sys_pfm_create_evtsets, sys_pfm_getinfo_evtsets, sys_pfm_delete_evtsets, sys_pfm_unload_context
 
 #endif /* CONFIG_COMPAT */
 
@@ -156,4 +158,6 @@ sys_call_table:
        .word sys_set_mempolicy, sys_kexec_load, sys_move_pages, sys_getcpu, sys_epoll_pwait
 /*310*/        .word sys_utimensat, sys_signalfd, sys_timerfd_create, sys_eventfd, sys_fallocate
        .word sys_timerfd_settime, sys_timerfd_gettime, sys_signalfd4, sys_eventfd2, sys_epoll_create1
-/*320*/        .word sys_dup3, sys_pipe2, sys_inotify_init1
+/*320*/        .word sys_dup3, sys_pipe2, sys_inotify_init1, sys_pfm_create_context, sys_pfm_write_pmcs
+       .word sys_pfm_write_pmds, sys_pfm_read_pmds, sys_pfm_load_context, sys_pfm_start, sys_pfm_stop
+/*330*/        .word sys_pfm_restart, sys_pfm_create_evtsets, sys_pfm_getinfo_evtsets, sys_pfm_delete_evtsets, sys_pfm_unload_context
diff --git a/arch/sparc64/kernel/traps.c b/arch/sparc64/kernel/traps.c
index c824df1..be45d09 100644
--- a/arch/sparc64/kernel/traps.c
+++ b/arch/sparc64/kernel/traps.c
@@ -2470,86 +2470,90 @@ extern void tsb_config_offsets_are_bolixed_dave(void);
 /* Only invoked on boot processor. */
 void __init trap_init(void)
 {
-       /* Compile time sanity check. */
-       if (TI_TASK != offsetof(struct thread_info, task) ||
-           TI_FLAGS != offsetof(struct thread_info, flags) ||
-           TI_CPU != offsetof(struct thread_info, cpu) ||
-           TI_FPSAVED != offsetof(struct thread_info, fpsaved) ||
-           TI_KSP != offsetof(struct thread_info, ksp) ||
-           TI_FAULT_ADDR != offsetof(struct thread_info, fault_address) ||
-           TI_KREGS != offsetof(struct thread_info, kregs) ||
-           TI_UTRAPS != offsetof(struct thread_info, utraps) ||
-           TI_EXEC_DOMAIN != offsetof(struct thread_info, exec_domain) ||
-           TI_REG_WINDOW != offsetof(struct thread_info, reg_window) ||
-           TI_RWIN_SPTRS != offsetof(struct thread_info, rwbuf_stkptrs) ||
-           TI_GSR != offsetof(struct thread_info, gsr) ||
-           TI_XFSR != offsetof(struct thread_info, xfsr) ||
-           TI_USER_CNTD0 != offsetof(struct thread_info, user_cntd0) ||
-           TI_USER_CNTD1 != offsetof(struct thread_info, user_cntd1) ||
-           TI_KERN_CNTD0 != offsetof(struct thread_info, kernel_cntd0) ||
-           TI_KERN_CNTD1 != offsetof(struct thread_info, kernel_cntd1) ||
-           TI_PCR != offsetof(struct thread_info, pcr_reg) ||
-           TI_PRE_COUNT != offsetof(struct thread_info, preempt_count) ||
-           TI_NEW_CHILD != offsetof(struct thread_info, new_child) ||
-           TI_SYS_NOERROR != offsetof(struct thread_info, syscall_noerror) ||
-           TI_RESTART_BLOCK != offsetof(struct thread_info, restart_block) ||
-           TI_KUNA_REGS != offsetof(struct thread_info, kern_una_regs) ||
-           TI_KUNA_INSN != offsetof(struct thread_info, kern_una_insn) ||
-           TI_FPREGS != offsetof(struct thread_info, fpregs) ||
-           (TI_FPREGS & (64 - 1)))
-               thread_info_offsets_are_bolixed_dave();
-
-       if (TRAP_PER_CPU_THREAD != offsetof(struct trap_per_cpu, thread) ||
-           (TRAP_PER_CPU_PGD_PADDR !=
-            offsetof(struct trap_per_cpu, pgd_paddr)) ||
-           (TRAP_PER_CPU_CPU_MONDO_PA !=
-            offsetof(struct trap_per_cpu, cpu_mondo_pa)) ||
-           (TRAP_PER_CPU_DEV_MONDO_PA !=
-            offsetof(struct trap_per_cpu, dev_mondo_pa)) ||
-           (TRAP_PER_CPU_RESUM_MONDO_PA !=
-            offsetof(struct trap_per_cpu, resum_mondo_pa)) ||
-           (TRAP_PER_CPU_RESUM_KBUF_PA !=
-            offsetof(struct trap_per_cpu, resum_kernel_buf_pa)) ||
-           (TRAP_PER_CPU_NONRESUM_MONDO_PA !=
-            offsetof(struct trap_per_cpu, nonresum_mondo_pa)) ||
-           (TRAP_PER_CPU_NONRESUM_KBUF_PA !=
-            offsetof(struct trap_per_cpu, nonresum_kernel_buf_pa)) ||
-           (TRAP_PER_CPU_FAULT_INFO !=
-            offsetof(struct trap_per_cpu, fault_info)) ||
-           (TRAP_PER_CPU_CPU_MONDO_BLOCK_PA !=
-            offsetof(struct trap_per_cpu, cpu_mondo_block_pa)) ||
-           (TRAP_PER_CPU_CPU_LIST_PA !=
-            offsetof(struct trap_per_cpu, cpu_list_pa)) ||
-           (TRAP_PER_CPU_TSB_HUGE !=
-            offsetof(struct trap_per_cpu, tsb_huge)) ||
-           (TRAP_PER_CPU_TSB_HUGE_TEMP !=
-            offsetof(struct trap_per_cpu, tsb_huge_temp)) ||
-           (TRAP_PER_CPU_IRQ_WORKLIST_PA !=
-            offsetof(struct trap_per_cpu, irq_worklist_pa)) ||
-           (TRAP_PER_CPU_CPU_MONDO_QMASK !=
-            offsetof(struct trap_per_cpu, cpu_mondo_qmask)) ||
-           (TRAP_PER_CPU_DEV_MONDO_QMASK !=
-            offsetof(struct trap_per_cpu, dev_mondo_qmask)) ||
-           (TRAP_PER_CPU_RESUM_QMASK !=
-            offsetof(struct trap_per_cpu, resum_qmask)) ||
-           (TRAP_PER_CPU_NONRESUM_QMASK !=
-            offsetof(struct trap_per_cpu, nonresum_qmask)))
-               trap_per_cpu_offsets_are_bolixed_dave();
-
-       if ((TSB_CONFIG_TSB !=
-            offsetof(struct tsb_config, tsb)) ||
-           (TSB_CONFIG_RSS_LIMIT !=
-            offsetof(struct tsb_config, tsb_rss_limit)) ||
-           (TSB_CONFIG_NENTRIES !=
-            offsetof(struct tsb_config, tsb_nentries)) ||
-           (TSB_CONFIG_REG_VAL !=
-            offsetof(struct tsb_config, tsb_reg_val)) ||
-           (TSB_CONFIG_MAP_VADDR !=
-            offsetof(struct tsb_config, tsb_map_vaddr)) ||
-           (TSB_CONFIG_MAP_PTE !=
-            offsetof(struct tsb_config, tsb_map_pte)))
-               tsb_config_offsets_are_bolixed_dave();
-
+       BUILD_BUG_ON(TI_TASK != offsetof(struct thread_info, task));
+       BUILD_BUG_ON(TI_FLAGS != offsetof(struct thread_info, flags));
+       BUILD_BUG_ON(TI_CPU != offsetof(struct thread_info, cpu));
+       BUILD_BUG_ON(TI_FPSAVED != offsetof(struct thread_info, fpsaved));
+       BUILD_BUG_ON(TI_KSP != offsetof(struct thread_info, ksp));
+       BUILD_BUG_ON(TI_FAULT_ADDR !=
+                    offsetof(struct thread_info, fault_address));
+       BUILD_BUG_ON(TI_KREGS != offsetof(struct thread_info, kregs));
+       BUILD_BUG_ON(TI_UTRAPS != offsetof(struct thread_info, utraps));
+       BUILD_BUG_ON(TI_EXEC_DOMAIN !=
+                    offsetof(struct thread_info, exec_domain));
+       BUILD_BUG_ON(TI_REG_WINDOW !=
+                    offsetof(struct thread_info, reg_window));
+       BUILD_BUG_ON(TI_RWIN_SPTRS !=
+                    offsetof(struct thread_info, rwbuf_stkptrs));
+       BUILD_BUG_ON(TI_GSR != offsetof(struct thread_info, gsr));
+       BUILD_BUG_ON(TI_XFSR != offsetof(struct thread_info, xfsr));
+       BUILD_BUG_ON(TI_PRE_COUNT !=
+                    offsetof(struct thread_info, preempt_count));
+       BUILD_BUG_ON(TI_NEW_CHILD !=
+                    offsetof(struct thread_info, new_child));
+       BUILD_BUG_ON(TI_SYS_NOERROR !=
+                    offsetof(struct thread_info, syscall_noerror));
+       BUILD_BUG_ON(TI_RESTART_BLOCK !=
+                    offsetof(struct thread_info, restart_block));
+       BUILD_BUG_ON(TI_KUNA_REGS !=
+                    offsetof(struct thread_info, kern_una_regs));
+       BUILD_BUG_ON(TI_KUNA_INSN !=
+                    offsetof(struct thread_info, kern_una_insn));
+       BUILD_BUG_ON(TI_FPREGS != offsetof(struct thread_info, fpregs));
+       BUILD_BUG_ON((TI_FPREGS & (64 - 1)));
+ 
+       BUILD_BUG_ON(TRAP_PER_CPU_THREAD !=
+                    offsetof(struct trap_per_cpu, thread));
+       BUILD_BUG_ON(TRAP_PER_CPU_PGD_PADDR !=
+                    offsetof(struct trap_per_cpu, pgd_paddr));
+       BUILD_BUG_ON(TRAP_PER_CPU_CPU_MONDO_PA !=
+                    offsetof(struct trap_per_cpu, cpu_mondo_pa));
+       BUILD_BUG_ON(TRAP_PER_CPU_DEV_MONDO_PA !=
+                    offsetof(struct trap_per_cpu, dev_mondo_pa));
+       BUILD_BUG_ON(TRAP_PER_CPU_RESUM_MONDO_PA !=
+                    offsetof(struct trap_per_cpu, resum_mondo_pa));
+       BUILD_BUG_ON(TRAP_PER_CPU_RESUM_KBUF_PA !=
+                    offsetof(struct trap_per_cpu, resum_kernel_buf_pa));
+       BUILD_BUG_ON(TRAP_PER_CPU_NONRESUM_MONDO_PA !=
+                    offsetof(struct trap_per_cpu, nonresum_mondo_pa));
+       BUILD_BUG_ON(TRAP_PER_CPU_NONRESUM_KBUF_PA !=
+                    offsetof(struct trap_per_cpu, nonresum_kernel_buf_pa));
+       BUILD_BUG_ON(TRAP_PER_CPU_FAULT_INFO !=
+                    offsetof(struct trap_per_cpu, fault_info));
+       BUILD_BUG_ON(TRAP_PER_CPU_CPU_MONDO_BLOCK_PA !=
+                    offsetof(struct trap_per_cpu, cpu_mondo_block_pa));
+       BUILD_BUG_ON(TRAP_PER_CPU_CPU_LIST_PA !=
+                    offsetof(struct trap_per_cpu, cpu_list_pa));
+       BUILD_BUG_ON(TRAP_PER_CPU_TSB_HUGE !=
+                    offsetof(struct trap_per_cpu, tsb_huge));
+       BUILD_BUG_ON(TRAP_PER_CPU_TSB_HUGE_TEMP !=
+                    offsetof(struct trap_per_cpu, tsb_huge_temp));
+#if 0
+       BUILD_BUG_ON(TRAP_PER_CPU_IRQ_WORKLIST !=
+                    offsetof(struct trap_per_cpu, irq_worklist));
+#endif
+       BUILD_BUG_ON(TRAP_PER_CPU_CPU_MONDO_QMASK !=
+                    offsetof(struct trap_per_cpu, cpu_mondo_qmask));
+       BUILD_BUG_ON(TRAP_PER_CPU_DEV_MONDO_QMASK !=
+                    offsetof(struct trap_per_cpu, dev_mondo_qmask));
+       BUILD_BUG_ON(TRAP_PER_CPU_RESUM_QMASK !=
+                    offsetof(struct trap_per_cpu, resum_qmask));
+       BUILD_BUG_ON(TRAP_PER_CPU_NONRESUM_QMASK !=
+                    offsetof(struct trap_per_cpu, nonresum_qmask));
+ 
+       BUILD_BUG_ON(TSB_CONFIG_TSB !=
+                     offsetof(struct tsb_config, tsb));
+       BUILD_BUG_ON(TSB_CONFIG_RSS_LIMIT !=
+                    offsetof(struct tsb_config, tsb_rss_limit));
+       BUILD_BUG_ON(TSB_CONFIG_NENTRIES !=
+                    offsetof(struct tsb_config, tsb_nentries));
+       BUILD_BUG_ON(TSB_CONFIG_REG_VAL !=
+                    offsetof(struct tsb_config, tsb_reg_val));
+       BUILD_BUG_ON(TSB_CONFIG_MAP_VADDR !=
+                    offsetof(struct tsb_config, tsb_map_vaddr));
+       BUILD_BUG_ON(TSB_CONFIG_MAP_PTE !=
+                    offsetof(struct tsb_config, tsb_map_pte));
+ 
        /* Attach to the address space of init_task.  On SMP we
         * do this in smp.c:smp_callin for other cpus.
         */
diff --git a/arch/sparc64/kernel/ttable.S b/arch/sparc64/kernel/ttable.S
index 1ade3d6..2a31ffa 100644
--- a/arch/sparc64/kernel/ttable.S
+++ b/arch/sparc64/kernel/ttable.S
@@ -66,7 +66,7 @@ tl0_irq6:     BTRAP(0x46)
 tl0_irq7:      BTRAP(0x47) BTRAP(0x48) BTRAP(0x49)
 tl0_irq10:     BTRAP(0x4a) BTRAP(0x4b) BTRAP(0x4c) BTRAP(0x4d)
 tl0_irq14:     TRAP_IRQ(timer_interrupt, 14)
-tl0_irq15:     TRAP_IRQ(handler_irq, 15)
+tl0_irq15:     TRAP_IRQ(perfctr_irq, 15)
 tl0_resv050:   BTRAP(0x50) BTRAP(0x51) BTRAP(0x52) BTRAP(0x53) BTRAP(0x54) BTRAP(0x55)
 tl0_resv056:   BTRAP(0x56) BTRAP(0x57) BTRAP(0x58) BTRAP(0x59) BTRAP(0x5a) BTRAP(0x5b)
 tl0_resv05c:   BTRAP(0x5c) BTRAP(0x5d) BTRAP(0x5e) BTRAP(0x5f)
diff --git a/arch/sparc64/perfmon/Kconfig b/arch/sparc64/perfmon/Kconfig
new file mode 100644
index 0000000..4672024
--- /dev/null
+++ b/arch/sparc64/perfmon/Kconfig
@@ -0,0 +1,26 @@
+menu "Hardware Performance Monitoring support"
+config PERFMON
+       bool "Perfmon2 performance monitoring interface"
+       default n
+       help
+       Enables the perfmon2 interface to access the hardware
+       performance counters. See <http://perfmon2.sf.net/> for
+       more details.
+
+config PERFMON_DEBUG
+       bool "Perfmon debugging"
+       depends on PERFMON
+       default n
+       help
+       Enables perfmon debugging support
+
+config PERFMON_DEBUG_FS
+       bool "Enable perfmon statistics reporting via debugfs"
+       default y
+       depends on PERFMON && DEBUG_FS
+       help
+       Enable collection and reporting of perfmon timing statistics under
+       debugfs. This is used for debugging and performance analysis of the
+       subsystem. The debugfs filesystem must be mounted.
+
+endmenu
diff --git a/arch/sparc64/perfmon/Makefile b/arch/sparc64/perfmon/Makefile
new file mode 100644
index 0000000..ad2d907
--- /dev/null
+++ b/arch/sparc64/perfmon/Makefile
@@ -0,0 +1 @@
+obj-$(CONFIG_PERFMON)  += perfmon.o
diff --git a/arch/sparc64/perfmon/perfmon.c b/arch/sparc64/perfmon/perfmon.c
new file mode 100644
index 0000000..9e29833
--- /dev/null
+++ b/arch/sparc64/perfmon/perfmon.c
@@ -0,0 +1,422 @@
+/* perfmon.c: sparc64 perfmon support
+ *
+ * Copyright (C) 2007 David S. Miller (davem@davemloft.net)
+ */
+
+#include <linux/kernel.h>
+#include <linux/module.h>
+#include <linux/irq.h>
+#include <linux/perfmon_kern.h>
+
+#include <asm/system.h>
+#include <asm/spitfire.h>
+#include <asm/hypervisor.h>
+
+struct pcr_ops {
+       void (*write)(u64);
+       u64 (*read)(void);
+};
+
+static void direct_write_pcr(u64 val)
+{
+       write_pcr(val);
+}
+
+static u64 direct_read_pcr(void)
+{
+       u64 pcr;
+
+       read_pcr(pcr);
+
+       return pcr;
+}
+
+static struct pcr_ops direct_pcr_ops = {
+       .write  = direct_write_pcr,
+       .read   = direct_read_pcr,
+};
+
+/* Using the hypervisor call is needed so that we can set the
+ * hypervisor trace bit correctly, which is hyperprivileged.
+ */
+static void n2_write_pcr(u64 val)
+{
+       unsigned long ret;
+
+       ret = sun4v_niagara2_setperf(HV_N2_PERF_SPARC_CTL, val);
+       if (val != HV_EOK)
+               write_pcr(val);
+}
+
+static u64 n2_read_pcr(void)
+{
+       u64 pcr;
+
+       read_pcr(pcr);
+
+       return pcr;
+}
+
+static struct pcr_ops n2_pcr_ops = {
+       .write  = n2_write_pcr,
+       .read   = n2_read_pcr,
+};
+
+static struct pcr_ops *pcr_ops;
+
+void pfm_arch_write_pmc(struct pfm_context *ctx,
+                       unsigned int cnum, u64 value)
+{
+       /*
+        * we only write to the actual register when monitoring is
+        * active (pfm_start was issued)
+        */
+       if (ctx && ctx->flags.started == 0)
+               return;
+
+       pcr_ops->write(value);
+}
+
+u64 pfm_arch_read_pmc(struct pfm_context *ctx, unsigned int cnum)
+{
+       return pcr_ops->read();
+}
+
+/*
+ * collect pending overflowed PMDs. Called from pfm_ctxsw()
+ * and from PMU interrupt handler. Must fill in set->povfl_pmds[]
+ * and set->npend_ovfls. Interrupts are masked
+ */
+static void __pfm_get_ovfl_pmds(struct pfm_context *ctx, struct pfm_event_set *set)
+{
+       unsigned int max = ctx->regs.max_intr_pmd;
+       u64 wmask = 1ULL << pfm_pmu_conf->counter_width;
+       u64 *intr_pmds = ctx->regs.intr_pmds;
+       u64 *used_mask = set->used_pmds;
+       u64 mask[PFM_PMD_BV];
+       unsigned int i;
+
+       bitmap_and(cast_ulp(mask),
+                  cast_ulp(intr_pmds),
+                  cast_ulp(used_mask),
+                  max);
+
+       /*
+        * check all PMD that can generate interrupts
+        * (that includes counters)
+        */
+       for (i = 0; i < max; i++) {
+               if (test_bit(i, mask)) {
+                       u64 new_val = pfm_arch_read_pmd(ctx, i);
+
+                       PFM_DBG_ovfl("pmd%u new_val=0x%llx bit=%d\n",
+                                    i, (unsigned long long)new_val,
+                                    (new_val&wmask) ? 1 : 0);
+
+                       if (new_val & wmask) {
+                               __set_bit(i, set->povfl_pmds);
+                               set->npend_ovfls++;
+                       }
+               }
+       }
+}
+
+static void pfm_stop_active(struct task_struct *task, struct pfm_context *ctx,
+                           struct pfm_event_set *set)
+{
+       unsigned int i, max = ctx->regs.max_pmc;
+
+       /*
+        * clear enable bits, assume all pmcs are enable pmcs
+        */
+       for (i = 0; i < max; i++) {
+               if (test_bit(i, set->used_pmcs))
+                       pfm_arch_write_pmc(ctx, i, 0);
+       }
+
+       if (set->npend_ovfls)
+               return;
+
+       __pfm_get_ovfl_pmds(ctx, set);
+}
+
+/*
+ * Called from pfm_ctxsw(). Task is guaranteed to be current.
+ * Context is locked. Interrupts are masked. Monitoring is active.
+ * PMU access is guaranteed. PMC and PMD registers are live in PMU.
+ *
+ * for per-thread:
+ *     must stop monitoring for the task
+ *
+ * Return:
+ *     non-zero : did not save PMDs (as part of stopping the PMU)
+ *            0 : saved PMDs (no need to save them in caller)
+ */
+int pfm_arch_ctxswout_thread(struct task_struct *task, struct pfm_context *ctx)
+{
+       /*
+        * disable lazy restore of PMC registers.
+        */
+       ctx->active_set->priv_flags |= PFM_SETFL_PRIV_MOD_PMCS;
+
+       pfm_stop_active(task, ctx, ctx->active_set);
+
+       return 1;
+}
+
+/*
+ * Called from pfm_stop() and idle notifier
+ *
+ * Interrupts are masked. Context is locked. Set is the active set.
+ *
+ * For per-thread:
+ *   task is not necessarily current. If not current task, then
+ *   task is guaranteed stopped and off any cpu. Access to PMU
+ *   is not guaranteed. Interrupts are masked. Context is locked.
+ *   Set is the active set.
+ *
+ * For system-wide:
+ *     task is current
+ *
+ * must disable active monitoring. ctx cannot be NULL
+ */
+void pfm_arch_stop(struct task_struct *task, struct pfm_context *ctx)
+{
+       /*
+        * no need to go through stop_save()
+        * if we are already stopped
+        */
+       if (!ctx->flags.started || ctx->state == PFM_CTX_MASKED)
+               return;
+
+       /*
+        * stop live registers and collect pending overflow
+        */
+       if (task == current)
+               pfm_stop_active(task, ctx, ctx->active_set);
+}
+
+/*
+ * Enable active monitoring. Called from pfm_start() and
+ * pfm_arch_unmask_monitoring().
+ *
+ * Interrupts are masked. Context is locked. Set is the active set.
+ *
+ * For per-trhead:
+ *     Task is not necessarily current. If not current task, then task
+ *     is guaranteed stopped and off any cpu. Access to PMU is not guaranteed.
+ *
+ * For system-wide:
+ *     task is always current
+ *
+ * must enable active monitoring.
+ */
+void pfm_arch_start(struct task_struct *task, struct pfm_context *ctx)
+{
+       struct pfm_event_set *set;
+       unsigned int max_pmc = ctx->regs.max_pmc;
+       unsigned int i;
+
+       if (task != current)
+               return;
+
+       set = ctx->active_set;
+       for (i = 0; i < max_pmc; i++) {
+               if (test_bit(i, set->used_pmcs))
+                       pfm_arch_write_pmc(ctx, i, set->pmcs[i]);
+       }
+}
+
+/*
+ * function called from pfm_switch_sets(), pfm_context_load_thread(),
+ * pfm_context_load_sys(), pfm_ctxsw(), pfm_switch_sets()
+ * context is locked. Interrupts are masked. set cannot be NULL.
+ * Access to the PMU is guaranteed.
+ *
+ * function must restore all PMD registers from set.
+ */
+void pfm_arch_restore_pmds(struct pfm_context *ctx, struct pfm_event_set *set)
+{
+       unsigned int max_pmd = ctx->regs.max_pmd;
+       u64 ovfl_mask = pfm_pmu_conf->ovfl_mask;
+       u64 *impl_pmds = ctx->regs.pmds;
+       unsigned int i;
+
+       /*
+        * must restore all pmds to avoid leaking
+        * information to user.
+        */
+       for (i = 0; i < max_pmd; i++) {
+               u64 val;
+
+               if (test_bit(i, impl_pmds) == 0)
+                       continue;
+
+               val = set->pmds[i].value;
+
+               /*
+                * set upper bits for counter to ensure
+                * overflow will trigger
+                */
+               val &= ovfl_mask;
+
+               pfm_arch_write_pmd(ctx, i, val);
+       }
+}
+
+/*
+ * function called from pfm_switch_sets(), pfm_context_load_thread(),
+ * pfm_context_load_sys(), pfm_ctxsw().
+ * Context is locked. Interrupts are masked. set cannot be NULL.
+ * Access to the PMU is guaranteed.
+ *
+ * function must restore all PMC registers from set, if needed.
+ */
+void pfm_arch_restore_pmcs(struct pfm_context *ctx, struct pfm_event_set *set)
+{
+       unsigned int max_pmc = ctx->regs.max_pmc;
+       u64 *impl_pmcs = ctx->regs.pmcs;
+       unsigned int i;
+
+       /* If we're masked or stopped we don't need to bother restoring
+        * the PMCs now.
+        */
+       if (ctx->state == PFM_CTX_MASKED || ctx->flags.started == 0)
+               return;
+
+       /*
+        * restore all pmcs
+        */
+       for (i = 0; i < max_pmc; i++)
+               if (test_bit(i, impl_pmcs))
+                       pfm_arch_write_pmc(ctx, i, set->pmcs[i]);
+}
+
+char *pfm_arch_get_pmu_module_name(void)
+{
+       return NULL;
+}
+
+void perfmon_interrupt(struct pt_regs *regs)
+{
+       pfm_interrupt_handler(instruction_pointer(regs), regs);
+}
+
+static struct pfm_regmap_desc pfm_sparc64_pmc_desc[] = {
+       PMC_D(PFM_REG_I, "PCR", 0, 0, 0, 0),
+};
+
+static struct pfm_regmap_desc pfm_sparc64_pmd_desc[] = {
+       PMD_D(PFM_REG_C, "PIC0", 0),
+       PMD_D(PFM_REG_C, "PIC1", 0),
+};
+
+static int pfm_sparc64_probe(void)
+{
+       return 0;
+}
+
+static struct pfm_pmu_config pmu_sparc64_pmu_conf = {
+       .counter_width  = 31,
+       .pmd_desc       = pfm_sparc64_pmd_desc,
+       .num_pmd_entries = 2,
+       .pmc_desc       = pfm_sparc64_pmc_desc,
+       .num_pmc_entries = 1,
+       .probe_pmu      = pfm_sparc64_probe,
+       .flags          = PFM_PMU_BUILTIN_FLAG,
+       .owner          = THIS_MODULE,
+};
+
+static unsigned long perf_hsvc_group;
+static unsigned long perf_hsvc_major;
+static unsigned long perf_hsvc_minor;
+
+static int __init register_perf_hsvc(void)
+{
+       if (tlb_type == hypervisor) {
+               switch (sun4v_chip_type) {
+               case SUN4V_CHIP_NIAGARA1:
+                       perf_hsvc_group = HV_GRP_N2_CPU;
+                       break;
+
+               case SUN4V_CHIP_NIAGARA2:
+                       perf_hsvc_group = HV_GRP_N2_CPU;
+                       break;
+
+               default:
+                       return -ENODEV;
+               }
+
+
+               perf_hsvc_major = 1;
+               perf_hsvc_minor = 0;
+               if (sun4v_hvapi_register(perf_hsvc_group,
+                                        perf_hsvc_major,
+                                        &perf_hsvc_minor)) {
+                       printk("perfmon: Could not register N2 hvapi.\n");
+                       return -ENODEV;
+               }
+       }
+       return 0;
+}
+
+static void unregister_perf_hsvc(void)
+{
+       if (tlb_type != hypervisor)
+               return;
+       sun4v_hvapi_unregister(perf_hsvc_group);
+}
+
+static int __init pfm_sparc64_pmu_init(void)
+{
+       u64 mask;
+       int err;
+
+       err = register_perf_hsvc();
+       if (err)
+               return err;
+
+       if (tlb_type == hypervisor &&
+           sun4v_chip_type == SUN4V_CHIP_NIAGARA2)
+               pcr_ops = &n2_pcr_ops;
+       else
+               pcr_ops = &direct_pcr_ops;
+
+       if (!strcmp(sparc_pmu_type, "ultra12"))
+               mask = (0xf << 11) | (0xf << 4) | 0x7;
+       else if (!strcmp(sparc_pmu_type, "ultra3") ||
+                !strcmp(sparc_pmu_type, "ultra3i") ||
+                !strcmp(sparc_pmu_type, "ultra3+") ||
+                !strcmp(sparc_pmu_type, "ultra4+"))
+               mask = (0x3f << 11) | (0x3f << 4) | 0x7;
+       else if (!strcmp(sparc_pmu_type, "niagara2"))
+               mask = ((1UL << 63) | (1UL << 62) |
+                       (1UL << 31) | (0xfUL << 27) | (0xffUL << 19) |
+                       (1UL << 18) | (0xfUL << 14) | (0xff << 6) |
+                       (0x3UL << 4) | 0x7UL);
+       else if (!strcmp(sparc_pmu_type, "niagara"))
+               mask = ((1UL << 9) | (1UL << 8) |
+                       (0x7UL << 4) | 0x7UL);
+       else {
+               err = -ENODEV;
+               goto out_err;
+       }
+
+       pmu_sparc64_pmu_conf.pmu_name = sparc_pmu_type;
+       pfm_sparc64_pmc_desc[0].rsvd_msk = ~mask;
+
+       return pfm_pmu_register(&pmu_sparc64_pmu_conf);
+
+out_err:
+       unregister_perf_hsvc();
+       return err;
+}
+
+static void __exit pfm_sparc64_pmu_exit(void)
+{
+       unregister_perf_hsvc();
+       return pfm_pmu_unregister(&pmu_sparc64_pmu_conf);
+}
+
+module_init(pfm_sparc64_pmu_init);
+module_exit(pfm_sparc64_pmu_exit);
diff --git a/arch/x86/Kconfig b/arch/x86/Kconfig
index ed92864..3a2b544 100644
--- a/arch/x86/Kconfig
+++ b/arch/x86/Kconfig
@@ -1378,6 +1378,8 @@ config COMPAT_VDSO
 
          If unsure, say Y.
 
+source "arch/x86/perfmon/Kconfig"
+
 endmenu
 
 config ARCH_ENABLE_MEMORY_HOTPLUG
diff --git a/arch/x86/Makefile b/arch/x86/Makefile
index f5631da..c868ad6 100644
--- a/arch/x86/Makefile
+++ b/arch/x86/Makefile
@@ -150,6 +150,8 @@ core-$(CONFIG_LGUEST_GUEST) += arch/x86/lguest/
 core-y += arch/x86/kernel/
 core-y += arch/x86/mm/
 
+core-$(CONFIG_PERFMON) += arch/x86/perfmon/
+
 # Remaining sub architecture files
 core-y += $(mcore-y)
 
diff --git a/arch/x86/ia32/ia32entry.S b/arch/x86/ia32/ia32entry.S
index ffc1bb4..58e00cb 100644
--- a/arch/x86/ia32/ia32entry.S
+++ b/arch/x86/ia32/ia32entry.S
@@ -832,4 +832,16 @@ ia32_sys_call_table:
        .quad sys_dup3                  /* 330 */
        .quad sys_pipe2
        .quad sys_inotify_init1
+       .quad sys_pfm_create_context
+       .quad sys_pfm_write_pmcs
+       .quad sys_pfm_write_pmds        /* 335 */
+       .quad sys_pfm_read_pmds
+       .quad sys_pfm_load_context
+       .quad sys_pfm_start
+       .quad sys_pfm_stop
+       .quad sys_pfm_restart           /* 340 */
+       .quad sys_pfm_create_evtsets
+       .quad sys_pfm_getinfo_evtsets
+       .quad sys_pfm_delete_evtsets
+       .quad sys_pfm_unload_context
 ia32_syscall_end:
diff --git a/arch/x86/kernel/apic_32.c b/arch/x86/kernel/apic_32.c
index f88bd0d..53fe335 100644
--- a/arch/x86/kernel/apic_32.c
+++ b/arch/x86/kernel/apic_32.c
@@ -28,6 +28,7 @@
 #include <linux/acpi_pmtmr.h>
 #include <linux/module.h>
 #include <linux/dmi.h>
+#include <linux/perfmon_kern.h>
 
 #include <asm/atomic.h>
 #include <asm/smp.h>
@@ -669,6 +670,7 @@ u8 setup_APIC_eilvt_ibs(u8 vector, u8 msg_type, u8 mask)
        setup_APIC_eilvt(APIC_EILVT_LVTOFF_IBS, vector, msg_type, mask);
        return APIC_EILVT_LVTOFF_IBS;
 }
+EXPORT_SYMBOL(setup_APIC_eilvt_ibs);
 
 /*
  * Local APIC start and shutdown
@@ -1367,6 +1369,9 @@ void __init apic_intr_init(void)
 #ifdef CONFIG_X86_MCE_P4THERMAL
        alloc_intr_gate(THERMAL_APIC_VECTOR, thermal_interrupt);
 #endif
+#ifdef CONFIG_PERFMON
+       set_intr_gate(LOCAL_PERFMON_VECTOR, pmu_interrupt);
+#endif
 }
 
 /**
diff --git a/arch/x86/kernel/apic_64.c b/arch/x86/kernel/apic_64.c
index 446c062..574cd3b 100644
--- a/arch/x86/kernel/apic_64.c
+++ b/arch/x86/kernel/apic_64.c
@@ -228,6 +228,7 @@ u8 setup_APIC_eilvt_ibs(u8 vector, u8 msg_type, u8 mask)
        setup_APIC_eilvt(APIC_EILVT_LVTOFF_IBS, vector, msg_type, mask);
        return APIC_EILVT_LVTOFF_IBS;
 }
+EXPORT_SYMBOL(setup_APIC_eilvt_ibs);
 
 /*
  * Program the next event, relative to now
diff --git a/arch/x86/kernel/cpu/common.c b/arch/x86/kernel/cpu/common.c
index 4e456bd..5b6d6ca 100644
--- a/arch/x86/kernel/cpu/common.c
+++ b/arch/x86/kernel/cpu/common.c
@@ -5,6 +5,7 @@
 #include <linux/module.h>
 #include <linux/percpu.h>
 #include <linux/bootmem.h>
+#include <linux/perfmon_kern.h>
 #include <asm/processor.h>
 #include <asm/i387.h>
 #include <asm/msr.h>
@@ -726,6 +727,8 @@ void __cpuinit cpu_init(void)
        current_thread_info()->status = 0;
        clear_used_math();
        mxcsr_feature_mask_init();
+
+       pfm_init_percpu();
 }
 
 #ifdef CONFIG_HOTPLUG_CPU
diff --git a/arch/x86/kernel/entry_32.S b/arch/x86/kernel/entry_32.S
index 109792b..0b6e34c 100644
--- a/arch/x86/kernel/entry_32.S
+++ b/arch/x86/kernel/entry_32.S
@@ -513,7 +513,7 @@ ENDPROC(system_call)
        ALIGN
        RING0_PTREGS_FRAME              # can't unwind into user space anyway
 work_pending:
-       testb $_TIF_NEED_RESCHED, %cl
+       testw $(_TIF_NEED_RESCHED|_TIF_PERFMON_WORK), %cx
        jz work_notifysig
 work_resched:
        call schedule
diff --git a/arch/x86/kernel/entry_64.S b/arch/x86/kernel/entry_64.S
index 89434d4..34e44f5 100644
--- a/arch/x86/kernel/entry_64.S
+++ b/arch/x86/kernel/entry_64.S
@@ -888,7 +888,13 @@ END(error_interrupt)
 ENTRY(spurious_interrupt)
        apicinterrupt SPURIOUS_APIC_VECTOR,smp_spurious_interrupt
 END(spurious_interrupt)
-                               
+
+#ifdef CONFIG_PERFMON
+ENTRY(pmu_interrupt)
+       apicinterrupt LOCAL_PERFMON_VECTOR,smp_pmu_interrupt
+END(pmu_interrupt)
+#endif
+
 /*
  * Exception entry points.
  */            
diff --git a/arch/x86/kernel/irqinit_64.c b/arch/x86/kernel/irqinit_64.c
index 1f26fd9..83f6bc1 100644
--- a/arch/x86/kernel/irqinit_64.c
+++ b/arch/x86/kernel/irqinit_64.c
@@ -11,6 +11,7 @@
 #include <linux/kernel_stat.h>
 #include <linux/sysdev.h>
 #include <linux/bitops.h>
+#include <linux/perfmon_kern.h>
 
 #include <asm/acpi.h>
 #include <asm/atomic.h>
@@ -217,6 +218,10 @@ void __init native_init_IRQ(void)
        alloc_intr_gate(SPURIOUS_APIC_VECTOR, spurious_interrupt);
        alloc_intr_gate(ERROR_APIC_VECTOR, error_interrupt);
 
+#ifdef CONFIG_PERFMON
+       alloc_intr_gate(LOCAL_PERFMON_VECTOR, pmu_interrupt);
+#endif
+
        if (!acpi_ioapic)
                setup_irq(2, &irq2);
 }
diff --git a/arch/x86/kernel/process_32.c b/arch/x86/kernel/process_32.c
index 31f40b2..ed27150 100644
--- a/arch/x86/kernel/process_32.c
+++ b/arch/x86/kernel/process_32.c
@@ -36,6 +36,7 @@
 #include <linux/personality.h>
 #include <linux/tick.h>
 #include <linux/percpu.h>
+#include <linux/perfmon_kern.h>
 #include <linux/prctl.h>
 
 #include <asm/uaccess.h>
@@ -277,6 +278,7 @@ void exit_thread(void)
                tss->x86_tss.io_bitmap_base = INVALID_IO_BITMAP_OFFSET;
                put_cpu();
        }
+       pfm_exit_thread();
 }
 
 void flush_thread(void)
@@ -334,6 +336,8 @@ int copy_thread(int nr, unsigned long clone_flags, unsigned long sp,
 
        savesegment(gs, p->thread.gs);
 
+       pfm_copy_thread(p);
+
        tsk = current;
        if (unlikely(test_tsk_thread_flag(tsk, TIF_IO_BITMAP))) {
                p->thread.io_bitmap_ptr = kmemdup(tsk->thread.io_bitmap_ptr,
@@ -448,6 +452,9 @@ __switch_to_xtra(struct task_struct *prev_p, struct task_struct *next_p,
        prev = &prev_p->thread;
        next = &next_p->thread;
 
+       if (test_tsk_thread_flag(prev_p, TIF_PERFMON_CTXSW))
+               pfm_ctxsw_out(prev_p, next_p);
+
        debugctl = prev->debugctlmsr;
        if (next->ds_area_msr != prev->ds_area_msr) {
                /* we clear debugctl to make sure DS
@@ -460,6 +467,9 @@ __switch_to_xtra(struct task_struct *prev_p, struct task_struct *next_p,
        if (next->debugctlmsr != debugctl)
                update_debugctlmsr(next->debugctlmsr);
 
+       if (test_tsk_thread_flag(next_p, TIF_PERFMON_CTXSW))
+               pfm_ctxsw_in(prev_p, next_p);
+
        if (test_tsk_thread_flag(next_p, TIF_DEBUG)) {
                set_debugreg(next->debugreg0, 0);
                set_debugreg(next->debugreg1, 1);
diff --git a/arch/x86/kernel/process_64.c b/arch/x86/kernel/process_64.c
index e12e0e4..97d49ce 100644
--- a/arch/x86/kernel/process_64.c
+++ b/arch/x86/kernel/process_64.c
@@ -36,6 +36,7 @@
 #include <linux/kprobes.h>
 #include <linux/kdebug.h>
 #include <linux/tick.h>
+#include <linux/perfmon_kern.h>
 #include <linux/prctl.h>
 
 #include <asm/uaccess.h>
@@ -240,6 +241,7 @@ void exit_thread(void)
                t->io_bitmap_max = 0;
                put_cpu();
        }
+       pfm_exit_thread();
 }
 
 void flush_thread(void)
@@ -344,6 +346,8 @@ int copy_thread(int nr, unsigned long clone_flags, unsigned long sp,
        savesegment(es, p->thread.es);
        savesegment(ds, p->thread.ds);
 
+       pfm_copy_thread(p);
+
        if (unlikely(test_tsk_thread_flag(me, TIF_IO_BITMAP))) {
                p->thread.io_bitmap_ptr = kmalloc(IO_BITMAP_BYTES, GFP_KERNEL);
                if (!p->thread.io_bitmap_ptr) {
@@ -472,6 +476,9 @@ static inline void __switch_to_xtra(struct task_struct *prev_p,
        prev = &prev_p->thread,
        next = &next_p->thread;
 
+       if (test_tsk_thread_flag(prev_p, TIF_PERFMON_CTXSW))
+               pfm_ctxsw_out(prev_p, next_p);
+
        debugctl = prev->debugctlmsr;
        if (next->ds_area_msr != prev->ds_area_msr) {
                /* we clear debugctl to make sure DS
@@ -484,6 +491,9 @@ static inline void __switch_to_xtra(struct task_struct *prev_p,
        if (next->debugctlmsr != debugctl)
                update_debugctlmsr(next->debugctlmsr);
 
+       if (test_tsk_thread_flag(next_p, TIF_PERFMON_CTXSW))
+               pfm_ctxsw_in(prev_p, next_p);
+
        if (test_tsk_thread_flag(next_p, TIF_DEBUG)) {
                loaddebug(next, 0);
                loaddebug(next, 1);
diff --git a/arch/x86/kernel/signal_32.c b/arch/x86/kernel/signal_32.c
index 6fb5bcd..53e6665 100644
--- a/arch/x86/kernel/signal_32.c
+++ b/arch/x86/kernel/signal_32.c
@@ -18,6 +18,7 @@
 #include <linux/sched.h>
 #include <linux/wait.h>
 #include <linux/elf.h>
+#include <linux/perfmon_kern.h>
 #include <linux/smp.h>
 #include <linux/mm.h>
 
@@ -657,6 +658,10 @@ static void do_signal(struct pt_regs *regs)
 void
 do_notify_resume(struct pt_regs *regs, void *unused, __u32 thread_info_flags)
 {
+       /* process perfmon asynchronous work (e.g. block thread or reset) */
+       if (thread_info_flags & _TIF_PERFMON_WORK)
+               pfm_handle_work(regs);
+
        /* deal with pending signal delivery */
        if (thread_info_flags & _TIF_SIGPENDING)
                do_signal(regs);
diff --git a/arch/x86/kernel/signal_64.c b/arch/x86/kernel/signal_64.c
index ca316b5..6e9fa74 100644
--- a/arch/x86/kernel/signal_64.c
+++ b/arch/x86/kernel/signal_64.c
@@ -19,6 +19,7 @@
 #include <linux/stddef.h>
 #include <linux/personality.h>
 #include <linux/compiler.h>
+#include <linux/perfmon_kern.h>
 #include <asm/processor.h>
 #include <asm/ucontext.h>
 #include <asm/uaccess.h>
@@ -549,12 +550,17 @@ static void do_signal(struct pt_regs *regs)
 void do_notify_resume(struct pt_regs *regs, void *unused,
                      __u32 thread_info_flags)
 {
+
 #ifdef CONFIG_X86_MCE
        /* notify userspace of pending MCEs */
        if (thread_info_flags & _TIF_MCE_NOTIFY)
                mce_notify_user();
 #endif /* CONFIG_X86_MCE */
 
+       /* process perfmon asynchronous work (e.g. block thread or reset) */
+       if (thread_info_flags & _TIF_PERFMON_WORK)
+               pfm_handle_work(regs);
+
        /* deal with pending signal delivery */
        if (thread_info_flags & _TIF_SIGPENDING)
                do_signal(regs);
diff --git a/arch/x86/kernel/smpboot.c b/arch/x86/kernel/smpboot.c
index 7985c5b..9ddf6db 100644
--- a/arch/x86/kernel/smpboot.c
+++ b/arch/x86/kernel/smpboot.c
@@ -42,6 +42,7 @@
 #include <linux/init.h>
 #include <linux/smp.h>
 #include <linux/module.h>
+#include <linux/perfmon_kern.h>
 #include <linux/sched.h>
 #include <linux/percpu.h>
 #include <linux/bootmem.h>
@@ -1382,6 +1383,7 @@ int __cpu_disable(void)
        remove_cpu_from_maps(cpu);
        unlock_vector_lock();
        fixup_irqs(cpu_online_map);
+       pfm_cpu_disable();
        return 0;
 }
 
diff --git a/arch/x86/kernel/syscall_table_32.S b/arch/x86/kernel/syscall_table_32.S
index d44395f..e1384a9 100644
--- a/arch/x86/kernel/syscall_table_32.S
+++ b/arch/x86/kernel/syscall_table_32.S
@@ -332,3 +332,15 @@ ENTRY(sys_call_table)
        .long sys_dup3                  /* 330 */
        .long sys_pipe2
        .long sys_inotify_init1
+       .long sys_pfm_create_context
+       .long sys_pfm_write_pmcs
+       .long sys_pfm_write_pmds        /* 335 */
+       .long sys_pfm_read_pmds
+       .long sys_pfm_load_context
+       .long sys_pfm_start
+       .long sys_pfm_stop
+       .long sys_pfm_restart           /* 340 */
+       .long sys_pfm_create_evtsets
+       .long sys_pfm_getinfo_evtsets
+       .long sys_pfm_delete_evtsets
+       .long sys_pfm_unload_context
diff --git a/arch/x86/oprofile/nmi_int.c b/arch/x86/oprofile/nmi_int.c
index 8a5f161..10faef5 100644
--- a/arch/x86/oprofile/nmi_int.c
+++ b/arch/x86/oprofile/nmi_int.c
@@ -16,6 +16,7 @@
 #include <linux/moduleparam.h>
 #include <linux/kdebug.h>
 #include <linux/cpu.h>
+#include <linux/perfmon_kern.h>
 #include <asm/nmi.h>
 #include <asm/msr.h>
 #include <asm/apic.h>
@@ -217,12 +218,18 @@ static int nmi_setup(void)
        int err = 0;
        int cpu;
 
-       if (!allocate_msrs())
+       if (pfm_session_allcpus_acquire())
+               return -EBUSY;
+
+       if (!allocate_msrs()) {
+               pfm_session_allcpus_release();
                return -ENOMEM;
+       }
 
        err = register_die_notifier(&profile_exceptions_nb);
        if (err) {
                free_msrs();
+               pfm_session_allcpus_release();
                return err;
        }
 
@@ -304,6 +311,7 @@ static void nmi_shutdown(void)
        model->shutdown(msrs);
        free_msrs();
        put_cpu_var(cpu_msrs);
+       pfm_session_allcpus_release();
 }
 
 static void nmi_cpu_start(void *dummy)
diff --git a/arch/x86/perfmon/Kconfig b/arch/x86/perfmon/Kconfig
new file mode 100644
index 0000000..08842e6
--- /dev/null
+++ b/arch/x86/perfmon/Kconfig
@@ -0,0 +1,89 @@
+menu "Hardware Performance Monitoring support"
+config PERFMON
+       bool "Perfmon2 performance monitoring interface"
+       select X86_LOCAL_APIC
+       default n
+       help
+       Enables the perfmon2 interface to access the hardware
+       performance counters. See <http://perfmon2.sf.net/> for
+       more details.
+
+config PERFMON_DEBUG
+       bool "Perfmon debugging"
+       default n
+       depends on PERFMON
+       help
+       Enables perfmon debugging support
+
+config PERFMON_DEBUG_FS
+       bool "Enable perfmon statistics reporting via debugfs"
+       default y
+       depends on PERFMON && DEBUG_FS
+       help
+       Enable collection and reporting of perfmon timing statistics under
+       debugfs. This is used for debugging and performance analysis of the
+       subsystem.The debugfs filesystem must be mounted.
+
+config X86_PERFMON_P6
+       tristate "Support for Intel P6/Pentium M processor hardware performance counters"
+       depends on PERFMON && X86_32
+       default n
+       help
+       Enables support for Intel P6-style hardware performance counters.
+       To be used for with Intel Pentium III, PentiumPro, Pentium M processors.
+
+config X86_PERFMON_P4
+       tristate "Support for Intel Pentium 4/Xeon hardware performance counters"
+       depends on PERFMON
+       default n
+       help
+       Enables support for Intel Pentium 4/Xeon (Netburst) hardware performance
+       counters.
+
+config X86_PERFMON_PEBS_P4
+       tristate "Support for Intel Netburst Precise Event-Based Sampling (PEBS)"
+       depends on PERFMON && X86_PERFMON_P4
+       default n
+       help
+       Enables support for Precise Event-Based Sampling (PEBS) on the Intel
+       Netburst processors such as Pentium 4, Xeon which support it.
+
+config  X86_PERFMON_CORE
+       tristate "Support for Intel Core-based performance counters"
+       depends on PERFMON
+       default n
+       help
+       Enables support for Intel Core-based performance counters. Enable
+       this option to support Intel Core 2 processors.
+
+config X86_PERFMON_PEBS_CORE
+       tristate "Support for Intel Core Precise Event-Based Sampling (PEBS)"
+       depends on PERFMON && X86_PERFMON_CORE
+       default n
+       help
+       Enables support for Precise Event-Based Sampling (PEBS) on the Intel
+       Core processors.
+
+config  X86_PERFMON_INTEL_ATOM
+       tristate "Support for Intel Atom processor"
+       depends on PERFMON
+       default n
+       help
+       Enables support for Intel Atom processors.
+
+config  X86_PERFMON_INTEL_ARCH
+       tristate "Support for Intel architectural perfmon v1/v2"
+       depends on PERFMON
+       default n
+       help
+       Enables support for Intel architectural performance counters.
+       This feature was introduced with Intel Core Solo/Core Duo processors.
+
+config X86_PERFMON_AMD64
+       tristate "Support AMD Athlon64/Opteron64 hardware performance counters"
+       depends on PERFMON
+       default n
+       help
+       Enables support for Athlon64/Opterton64 hardware performance counters.
+       Support for family 6, 15 and 16(10H) processors.
+endmenu
diff --git a/arch/x86/perfmon/Makefile b/arch/x86/perfmon/Makefile
new file mode 100644
index 0000000..1cbed3e
--- /dev/null
+++ b/arch/x86/perfmon/Makefile
@@ -0,0 +1,13 @@
+#
+# Copyright (c) 2005-2007 Hewlett-Packard Development Company, L.P.
+# Contributed by Stephane Eranian <eranian@hpl.hp.com>
+#
+obj-$(CONFIG_PERFMON)                  += perfmon.o
+obj-$(CONFIG_X86_PERFMON_P6)           += perfmon_p6.o
+obj-$(CONFIG_X86_PERFMON_P4)           += perfmon_p4.o
+obj-$(CONFIG_X86_PERFMON_CORE)         += perfmon_intel_core.o
+obj-$(CONFIG_X86_PERFMON_INTEL_ARCH)   += perfmon_intel_arch.o
+obj-$(CONFIG_X86_PERFMON_PEBS_P4)      += perfmon_pebs_p4_smpl.o
+obj-$(CONFIG_X86_PERFMON_PEBS_CORE)    += perfmon_pebs_core_smpl.o
+obj-$(CONFIG_X86_PERFMON_AMD64)        += perfmon_amd64.o
+obj-$(CONFIG_X86_PERFMON_INTEL_ATOM)   += perfmon_intel_atom.o
diff --git a/arch/x86/perfmon/perfmon.c b/arch/x86/perfmon/perfmon.c
new file mode 100644
index 0000000..e727fed
--- /dev/null
+++ b/arch/x86/perfmon/perfmon.c
@@ -0,0 +1,761 @@
+/*
+ * This file implements the X86 specific support for the perfmon2 interface
+ *
+ * Copyright (c) 2005-2007 Hewlett-Packard Development Company, L.P.
+ * Contributed by Stephane Eranian <eranian@hpl.hp.com>
+ *
+ * Copyright (c) 2007 Advanced Micro Devices, Inc.
+ * Contributed by Robert Richter <robert.richter@amd.com>
+ *
+ * This program is free software; you can redistribute it and/or
+ * modify it under the terms of version 2 of the GNU General Public
+ * License as published by the Free Software Foundation.
+ *
+ * This program is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+ * General Public License for more details.
+ *
+ * You should have received a copy of the GNU General Public License
+ * along with this program; if not, write to the Free Software
+ * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA
+ * 02111-1307 USA
+ */
+#include <linux/interrupt.h>
+#include <linux/perfmon_kern.h>
+#include <linux/kprobes.h>
+#include <linux/kdebug.h>
+#include <linux/nmi.h>
+
+#include <asm/apic.h>
+
+DEFINE_PER_CPU(unsigned long, real_iip);
+DEFINE_PER_CPU(int, pfm_using_nmi);
+DEFINE_PER_CPU(unsigned long, saved_lvtpc);
+
+/**
+ * pfm_arch_ctxswin_thread - thread context switch in
+ * @task: task switched in
+ * @ctx: context for the task
+ *
+ * Called from pfm_ctxsw(). Task is guaranteed to be current.
+ * set cannot be NULL. Context is locked. Interrupts are masked.
+ *
+ * Caller has already restored all PMD and PMC registers, if
+ * necessary (i.e., lazy restore scheme).
+ *
+ * On x86, the only common code just needs to unsecure RDPMC if necessary
+ *
+ * On model-specific features, e.g., PEBS, IBS, are taken care of in the
+ * corresponding PMU description module
+ */
+void pfm_arch_ctxswin_thread(struct task_struct *task, struct pfm_context *ctx)
+{
+       struct pfm_arch_context *ctx_arch;
+
+       ctx_arch = pfm_ctx_arch(ctx);
+
+       /*
+        *  restore saved real iip
+        */
+       if (ctx->active_set->npend_ovfls)
+               __get_cpu_var(real_iip) = ctx_arch->saved_real_iip;
+
+       /*
+        * enable RDPMC on this CPU
+        */
+       if (ctx_arch->flags.insecure)
+               set_in_cr4(X86_CR4_PCE);
+}
+
+/**
+ * pfm_arch_ctxswout_thread - context switch out thread
+ * @task: task switched out
+ * @ctx : context switched out
+ *
+ * Called from pfm_ctxsw(). Task is guaranteed to be current.
+ * Context is locked. Interrupts are masked. Monitoring may be active.
+ * PMU access is guaranteed. PMC and PMD registers are live in PMU.
+ *
+ * Return:
+ *     non-zero : did not save PMDs (as part of stopping the PMU)
+ *            0 : saved PMDs (no need to save them in caller)
+ */
+int pfm_arch_ctxswout_thread(struct task_struct *task, struct pfm_context *ctx)
+{
+       struct pfm_arch_context *ctx_arch;
+       struct pfm_arch_pmu_info *pmu_info;
+
+       ctx_arch = pfm_ctx_arch(ctx);
+       pmu_info = pfm_pmu_info();
+
+       /*
+        * disable lazy restore of PMCS on ctxswin because
+        * we modify some of them.
+        */
+       ctx->active_set->priv_flags |= PFM_SETFL_PRIV_MOD_PMCS;
+
+       if (ctx->active_set->npend_ovfls)
+               ctx_arch->saved_real_iip = __get_cpu_var(real_iip);
+
+       /*
+        * disable RDPMC on this CPU
+        */
+       if (ctx_arch->flags.insecure)
+               clear_in_cr4(X86_CR4_PCE);
+
+       if (ctx->state == PFM_CTX_MASKED)
+               return 1;
+
+       return pmu_info->stop_save(ctx, ctx->active_set);
+}
+
+/**
+ * pfm_arch_stop - deactivate monitoring
+ * @task: task to stop
+ * @ctx: context to stop
+ *
+ * Called from pfm_stop()
+ * Interrupts are masked. Context is locked. Set is the active set.
+ *
+ * For per-thread:
+ *   task is not necessarily current. If not current task, then
+ *   task is guaranteed stopped and off any cpu. Access to PMU
+ *   is not guaranteed.
+ *
+ * For system-wide:
+ *     task is current
+ *
+ * must disable active monitoring. ctx cannot be NULL
+ */
+void pfm_arch_stop(struct task_struct *task, struct pfm_context *ctx)
+{
+       struct pfm_arch_pmu_info *pmu_info;
+
+       pmu_info = pfm_pmu_info();
+
+       /*
+        * no need to go through stop_save()
+        * if we are already stopped
+        */
+       if (!ctx->flags.started || ctx->state == PFM_CTX_MASKED)
+               return;
+
+       if (task != current)
+               return;
+
+       pmu_info->stop_save(ctx, ctx->active_set);
+}
+
+
+/**
+ * pfm_arch_start - activate monitoring
+ * @task: task to start
+ * @ctx: context to stop
+ *
+ * Interrupts are masked. Context is locked.
+ *
+ * For per-thread:
+ *     Task is not necessarily current. If not current task, then task
+ *     is guaranteed stopped and off any cpu. No access to PMU is task
+ *     is not current.
+ *
+ * For system-wide:
+ *     task is always current
+ */
+void pfm_arch_start(struct task_struct *task, struct pfm_context *ctx)
+{
+       struct pfm_event_set *set;
+
+       set = ctx->active_set;
+
+       if (task != current)
+               return;
+
+       /*
+        * cannot restore PMC if no access to PMU. Will be done
+        * when the thread is switched back in
+        */
+
+       pfm_arch_restore_pmcs(ctx, set);
+}
+
+/**
+ * pfm_arch_restore_pmds - reload PMD registers
+ * @ctx: context to restore from
+ * @set: current event set
+ *
+ * function called from pfm_switch_sets(), pfm_context_load_thread(),
+ * pfm_context_load_sys(), pfm_ctxsw()
+ *
+ * Context is locked. Interrupts are masked. Set cannot be NULL.
+ * Access to the PMU is guaranteed.
+ */
+void pfm_arch_restore_pmds(struct pfm_context *ctx, struct pfm_event_set *set)
+{
+       struct pfm_arch_pmu_info *pmu_info;
+       u16 i, num;
+
+       pmu_info = pfm_pmu_info();
+
+       num = set->nused_pmds;
+
+       /*
+        * model-specific override
+        */
+       if (pmu_info->restore_pmds) {
+               pmu_info->restore_pmds(ctx, set);
+               return;
+       }
+
+       /*
+        * we can restore only the PMD we use because:
+        *
+        *      - can only read with pfm_read_pmds() the registers
+        *        declared used via pfm_write_pmds(), smpl_pmds, reset_pmds
+        *
+        *      - if cr4.pce=1, only counters are exposed to user. RDPMC
+        *        does not work with other types of PMU registers.Thus, no
+        *        address is ever exposed by counters
+        *
+        *      - there is never a dependency between one pmd register and
+        *        another
+        */
+       for (i = 0; num; i++) {
+               if (likely(test_bit(i, cast_ulp(set->used_pmds)))) {
+                       pfm_write_pmd(ctx, i, set->pmds[i].value);
+                       num--;
+               }
+       }
+}
+
+/**
+ * pfm_arch_restore_pmcs - reload PMC registers
+ * @ctx: context to restore from
+ * @set: current event set
+ *
+ * function called from pfm_switch_sets(), pfm_context_load_thread(),
+ * pfm_context_load_sys(), pfm_ctxsw().
+ *
+ * Context is locked. Interrupts are masked. set cannot be NULL.
+ * Access to the PMU is guaranteed.
+ *
+ * function must restore all PMC registers from set
+ */
+void pfm_arch_restore_pmcs(struct pfm_context *ctx, struct pfm_event_set *set)
+{
+       struct pfm_arch_pmu_info *pmu_info;
+       u64 *mask;
+       u16 i, num;
+
+       pmu_info = pfm_pmu_info();
+
+       /*
+        * we need to restore PMCs only when:
+        *      - context is not masked
+        *      - monitoring activated
+        *
+        * Masking monitoring after an overflow does not change the
+        * value of flags.started
+        */
+       if (ctx->state == PFM_CTX_MASKED || !ctx->flags.started)
+               return;
+
+       /*
+        * model-specific override
+        */
+       if (pmu_info->restore_pmcs) {
+               pmu_info->restore_pmcs(ctx, set);
+               return;
+       }
+       /*
+        * restore all pmcs
+        *
+        * It is not possible to restore only the pmcs we used because
+        * certain PMU models (e.g. Pentium 4) have dependencies. Thus
+        * we do not want one application using stale PMC coming from
+        * another one.
+        *
+        * On PMU models where there is no dependencies between pmc, then
+        * it is possible to optimize by only restoring the registers that
+        * are used, and this can be done with the models-specific override
+        * for this function.
+        *
+        * The default code takes the safest approach, i.e., assume the worse
+        */
+       mask = ctx->regs.pmcs;
+       num = ctx->regs.num_pmcs;
+       for (i = 0; num; i++) {
+               if (test_bit(i, cast_ulp(mask))) {
+                       pfm_arch_write_pmc(ctx, i, set->pmcs[i]);
+                       num--;
+               }
+       }
+}
+
+/**
+ * smp_pmu_interrupt - lowest level PMU interrupt handler for X86
+ * @regs: machine state
+ *
+ * The PMU interrupt is handled through an interrupt gate, therefore
+ * the CPU automatically clears the EFLAGS.IF, i.e., masking interrupts.
+ *
+ * The perfmon interrupt handler MUST run with interrupts disabled due
+ * to possible race with other, higher priority interrupts, such as timer
+ * or IPI function calls.
+ *
+ * See description in IA-32 architecture manual, Vol 3 section 5.8.1
+ */
+void smp_pmu_interrupt(struct pt_regs *regs)
+{
+       struct pfm_arch_pmu_info *pmu_info;
+       struct pfm_context *ctx;
+       unsigned long iip;
+       int using_nmi;
+
+       using_nmi = __get_cpu_var(pfm_using_nmi);
+
+       ack_APIC_irq();
+
+       irq_enter();
+
+       /*
+        * when using NMI, pfm_handle_nmi() gets called
+        * first. It stops monitoring and record the
+        * iip into real_iip, then it repost the interrupt
+        * using the lower priority vector LOCAL_PERFMON_VECTOR
+        *
+        * On some processors, e.g., P4, it may be that some
+        * state is already recorded from pfm_handle_nmi()
+        * and it only needs to be copied back into the normal
+        * fields so it can be used transparently by higher level
+        * code.
+        */
+       if (using_nmi) {
+               ctx = __get_cpu_var(pmu_ctx);
+               pmu_info = pfm_pmu_info();
+               iip = __get_cpu_var(real_iip);
+               if (ctx && pmu_info->nmi_copy_state)
+                       pmu_info->nmi_copy_state(ctx);
+       } else
+               iip = instruction_pointer(regs);
+
+       pfm_interrupt_handler(iip, regs);
+
+       /*
+        * On Intel P6, Pentium M, P4, Intel Core:
+        *      - it is necessary to clear the MASK field for the LVTPC
+        *        vector. Otherwise interrupts remain masked. See
+        *        section 8.5.1
+        * AMD X86-64:
+        *      - the documentation does not stipulate the behavior.
+        *        To be safe, we also rewrite the vector to clear the
+        *        mask field
+        */
+       if (!using_nmi && current_cpu_data.x86_vendor == X86_VENDOR_INTEL)
+               apic_write(APIC_LVTPC, LOCAL_PERFMON_VECTOR);
+
+       irq_exit();
+}
+
+/**
+ * pfm_handle_nmi - PMU NMI handler notifier callback
+ * @nb ; notifier block
+ * @val: type of die notifier
+ * @data: die notifier-specific data
+ *
+ * called from notify_die() notifier from an trap handler path. We only
+ * care about NMI related callbacks, and ignore everything else.
+ *
+ * Cannot grab any locks, include the perfmon context lock
+ *
+ * Must detect if NMI interrupt comes from perfmon, and if so it must
+ * stop the PMU and repost a lower-priority interrupt. The perfmon interrupt
+ * handler needs to grab the context lock, thus is cannot be run directly
+ * from the NMI interrupt call path.
+ */
+static int __kprobes pfm_handle_nmi(struct notifier_block *nb,
+                                   unsigned long val,
+                                   void *data)
+{
+       struct die_args *args = data;
+       struct pfm_context *ctx;
+       struct pfm_arch_pmu_info *pmu_info;
+
+       /*
+        * only NMI related calls
+        */
+       if (val != DIE_NMI_IPI)
+               return NOTIFY_DONE;
+
+       /*
+        * perfmon not using NMI
+        */
+       if (!__get_cpu_var(pfm_using_nmi))
+               return NOTIFY_DONE;
+
+       /*
+        * No context
+        */
+       ctx = __get_cpu_var(pmu_ctx);
+       if (!ctx) {
+               PFM_DBG_ovfl("no ctx");
+               return NOTIFY_DONE;
+       }
+
+       /*
+        * Detect if we have overflows, i.e., NMI interrupt
+        * caused by PMU
+        */
+       pmu_info = pfm_pmu_conf->pmu_info;
+       if (!pmu_info->has_ovfls(ctx)) {
+               PFM_DBG_ovfl("no ovfl");
+               return NOTIFY_DONE;
+       }
+
+       /*
+        * we stop the PMU to avoid further overflow before this
+        * one is treated by lower priority interrupt handler
+        */
+       pmu_info->quiesce();
+
+       /*
+        * record actual instruction pointer
+        */
+       __get_cpu_var(real_iip) = instruction_pointer(args->regs);
+
+       /*
+        * post lower priority interrupt (LOCAL_PERFMON_VECTOR)
+        */
+       pfm_arch_resend_irq(ctx);
+
+       pfm_stats_inc(ovfl_intr_nmi_count);
+
+       /*
+        * we need to rewrite the APIC vector on Intel
+        */
+       if (current_cpu_data.x86_vendor == X86_VENDOR_INTEL)
+               apic_write(APIC_LVTPC, APIC_DM_NMI);
+
+       /*
+        * the notification was for us
+        */
+       return NOTIFY_STOP;
+}
+
+static struct notifier_block pfm_nmi_nb = {
+       .notifier_call = pfm_handle_nmi
+};
+
+/**
+ * pfm_arch_get_pmu_module_name - get PMU description module name for autoload
+ *
+ * called from pfm_pmu_request_module
+ */
+char *pfm_arch_get_pmu_module_name(void)
+{
+       switch (current_cpu_data.x86) {
+       case 6:
+               switch (current_cpu_data.x86_model) {
+               case 3: /* Pentium II */
+               case 7 ... 11:
+               case 13:
+                       return "perfmon_p6";
+               case 15: /* Merom */
+               case 23: /* Penryn */
+                       return "perfmon_intel_core";
+               case 28: /* Atom/Silverthorne */
+                       return "perfmon_intel_atom";
+               case 29: /* Dunnington */
+                       return "perfmon_intel_core";
+               default:
+                       goto try_arch;
+               }
+       case 15:
+       case 16:
+               /* All Opteron processors */
+               if (current_cpu_data.x86_vendor == X86_VENDOR_AMD)
+                       return "perfmon_amd64";
+
+               switch (current_cpu_data.x86_model) {
+               case 0 ... 6:
+                       return "perfmon_p4";
+               }
+               /* FALL THROUGH */
+       default:
+try_arch:
+               if (boot_cpu_has(X86_FEATURE_ARCH_PERFMON))
+                       return "perfmon_intel_arch";
+               return NULL;
+       }
+       return NULL;
+}
+
+/**
+ * pfm_arch_resend_irq - post perfmon interrupt on regular vector
+ *
+ * called from pfm_ctxswin_thread() and pfm_handle_nmi()
+ */
+void pfm_arch_resend_irq(struct pfm_context *ctx)
+{
+       unsigned long val, dest;
+       /*
+        * we cannot use hw_resend_irq() because it goes to
+        * the I/O APIC. We need to go to the Local APIC.
+        *
+        * The "int vec" is not the right solution either
+        * because it triggers a software intr. We need
+        * to regenerate the interrupt and have it pended
+        * until we unmask interrupts.
+        *
+        * Instead we send ourself an IPI on the perfmon
+        * vector.
+        */
+       val  = APIC_DEST_SELF|APIC_INT_ASSERT|
+              APIC_DM_FIXED|LOCAL_PERFMON_VECTOR;
+
+       dest = apic_read(APIC_ID);
+       apic_write(APIC_ICR2, dest);
+       apic_write(APIC_ICR, val);
+}
+
+/**
+ * pfm_arch_pmu_acquire_percpu - setup APIC per CPU
+ * @data: contains pmu flags
+ */
+static void pfm_arch_pmu_acquire_percpu(void *data)
+{
+
+       struct pfm_arch_pmu_info *pmu_info;
+       unsigned int tmp, vec;
+       unsigned long flags = (unsigned long)data;
+       unsigned long lvtpc;
+
+       pmu_info = pfm_pmu_conf->pmu_info;
+
+       /*
+        * we only reprogram the LVTPC vector if we have detected
+        * no sharing, otherwise it means the APIC is already programmed
+        * and we use whatever vector (likely NMI) is there
+        */
+       if (!(flags & PFM_X86_FL_SHARING)) {
+               if (flags & PFM_X86_FL_USE_NMI)
+                       vec = APIC_DM_NMI;
+               else
+                       vec = LOCAL_PERFMON_VECTOR;
+
+               tmp = apic_read(APIC_LVTERR);
+               apic_write(APIC_LVTERR, tmp | APIC_LVT_MASKED);
+               apic_write(APIC_LVTPC, vec);
+               apic_write(APIC_LVTERR, tmp);
+       }
+       lvtpc = (unsigned long)apic_read(APIC_LVTPC);
+
+       __get_cpu_var(pfm_using_nmi) = lvtpc == APIC_DM_NMI;
+
+       PFM_DBG("LTVPC=0x%lx using_nmi=%d", lvtpc, __get_cpu_var(pfm_using_nmi));
+
+       /*
+        * invoke model specific acquire routine. May be used for
+        * model-specific initializations
+        */
+       if (pmu_info->acquire_pmu_percpu)
+               pmu_info->acquire_pmu_percpu();
+}
+
+/**
+ * pfm_arch_pmu_acquire - acquire PMU resource from system
+ * @unavail_pmcs : bitmask to use to set unavailable pmcs
+ * @unavail_pmds : bitmask to use to set unavailable pmds
+ *
+ * interrupts are not masked
+ *
+ * Grab PMU registers from lower level MSR allocator
+ *
+ * Program the APIC according the possible interrupt vector
+ * either LOCAL_PERFMON_VECTOR or NMI
+ */
+int pfm_arch_pmu_acquire(u64 *unavail_pmcs, u64 *unavail_pmds)
+{
+       struct pfm_arch_pmu_info *pmu_info;
+       struct pfm_regmap_desc *d;
+       u16 i, nlost;
+
+       pmu_info = pfm_pmu_conf->pmu_info;
+       pmu_info->flags &= ~PFM_X86_FL_SHARING;
+
+       nlost = 0;
+
+       d = pfm_pmu_conf->pmc_desc;
+       for (i = 0; i < pfm_pmu_conf->num_pmc_entries;  i++, d++) {
+               if (!(d->type & PFM_REG_I))
+                       continue;
+
+               if (d->type & PFM_REG_V)
+                       continue;
+               /*
+                * reserve register with lower-level allocator
+                */
+               if (!reserve_evntsel_nmi(d->hw_addr)) {
+                       PFM_DBG("pmc%d(%s) already used", i, d->desc);
+                       __set_bit(i, cast_ulp(unavail_pmcs));
+                       nlost++;
+                       continue;
+               }
+       }
+       PFM_DBG("nlost=%d info_flags=0x%x\n", nlost, pmu_info->flags);
+       /*
+        * some PMU models (e.g., P6) do not support sharing
+        * so check if we found less than the expected number of PMC registers
+        */
+       if (nlost) {
+               if (pmu_info->flags & PFM_X86_FL_NO_SHARING) {
+                       PFM_INFO("PMU already used by another subsystem, "
+                                "PMU does not support sharing, "
+                                "try disabling Oprofile or "
+                                "reboot with nmi_watchdog=0");
+                       goto undo;
+               }
+               pmu_info->flags |= PFM_X86_FL_SHARING;
+       }
+
+       d = pfm_pmu_conf->pmd_desc;
+       for (i = 0; i < pfm_pmu_conf->num_pmd_entries;  i++, d++) {
+               if (!(d->type & PFM_REG_I))
+                       continue;
+
+               if (d->type & PFM_REG_V)
+                       continue;
+
+               if (!reserve_perfctr_nmi(d->hw_addr)) {
+                       PFM_DBG("pmd%d(%s) already used", i, d->desc);
+                       __set_bit(i, cast_ulp(unavail_pmds));
+               }
+       }
+       /*
+        * program APIC on each CPU
+        */
+       on_each_cpu(pfm_arch_pmu_acquire_percpu,
+                   (void *)(unsigned long)pmu_info->flags , 1);
+
+       return 0;
+undo:
+       /*
+        * must undo reservation of pmcs in case of error
+        */
+       d = pfm_pmu_conf->pmc_desc;
+       for (i = 0; i < pfm_pmu_conf->num_pmc_entries;  i++, d++) {
+               if (!(d->type & (PFM_REG_I|PFM_REG_V)))
+                       continue;
+               if (!test_bit(i, cast_ulp(unavail_pmcs)))
+                       release_evntsel_nmi(d->hw_addr);
+       }
+       return -EBUSY;
+}
+/**
+ * pfm-arch_pmu_release_percpu - clear NMI state for one CPU
+ *
+ */
+static void pfm_arch_pmu_release_percpu(void *data)
+{
+       struct pfm_arch_pmu_info *pmu_info;
+
+       pmu_info = pfm_pmu_conf->pmu_info;
+
+       __get_cpu_var(pfm_using_nmi) = 0;
+
+       /*
+        * invoke model specific release routine.
+        * May be used to undo certain initializations
+        * or free some model-specific ressources.
+        */
+       if (pmu_info->release_pmu_percpu)
+               pmu_info->release_pmu_percpu();
+}
+
+/**
+ * pfm_arch_pmu_release - release PMU resource to system
+ *
+ * called from pfm_pmu_release()
+ * interrupts are not masked
+ *
+ * On x86, we return the PMU registers to the MSR allocator
+ */
+void pfm_arch_pmu_release(void)
+{
+       struct pfm_regmap_desc *d;
+       u16 i, n;
+
+       d = pfm_pmu_conf->pmc_desc;
+       n = pfm_pmu_conf->regs_all.num_pmcs;
+       for (i = 0; n; i++, d++) {
+               if (!test_bit(i, cast_ulp(pfm_pmu_conf->regs_all.pmcs)))
+                       continue;
+               release_evntsel_nmi(d->hw_addr);
+               n--;
+               PFM_DBG("pmc%u released", i);
+       }
+       d = pfm_pmu_conf->pmd_desc;
+       n = pfm_pmu_conf->regs_all.num_pmds;
+       for (i = 0; n; i++, d++) {
+               if (!test_bit(i, cast_ulp(pfm_pmu_conf->regs_all.pmds)))
+                       continue;
+               release_perfctr_nmi(d->hw_addr);
+               n--;
+               PFM_DBG("pmd%u released", i);
+       }
+
+       /* clear NMI variable if used */
+       if (__get_cpu_var(pfm_using_nmi))
+               on_each_cpu(pfm_arch_pmu_release_percpu, NULL , 1);
+}
+
+/**
+ * pfm_arch_pmu_config_init - validate PMU description structure
+ * @cfg: PMU description structure
+ *
+ * return:
+ *     0 if valid
+ *     errno otherwise
+ *
+ * called from pfm_pmu_register()
+ */
+int pfm_arch_pmu_config_init(struct pfm_pmu_config *cfg)
+{
+       struct pfm_arch_pmu_info *pmu_info;
+
+       pmu_info = pfm_pmu_info();
+       if (!pmu_info) {
+               PFM_DBG("%s missing pmu_info", cfg->pmu_name);
+               return -EINVAL;
+       }
+       if (!pmu_info->has_ovfls) {
+               PFM_DBG("%s missing has_ovfls callback", cfg->pmu_name);
+               return -EINVAL;
+       }
+       if (!pmu_info->quiesce) {
+               PFM_DBG("%s missing quiesce callback", cfg->pmu_name);
+               return -EINVAL;
+       }
+       if (!pmu_info->stop_save) {
+               PFM_DBG("%s missing stop_save callback", cfg->pmu_name);
+               return -EINVAL;
+       }
+       return 0;
+}
+
+/**
+ * pfm_arch_init - one time global arch-specific initialization
+ *
+ * called from pfm_init()
+ */
+int __init pfm_arch_init(void)
+{
+       /*
+        * we need to register our NMI handler when the kernels boots
+        * to avoid a deadlock condition with the NMI watchdog or Oprofile
+        * if we were to try and register/unregister on-demand.
+        */
+       register_die_notifier(&pfm_nmi_nb);
+       return 0;
+}
diff --git a/arch/x86/perfmon/perfmon_amd64.c b/arch/x86/perfmon/perfmon_amd64.c
new file mode 100644
index 0000000..f9b5f9c
--- /dev/null
+++ b/arch/x86/perfmon/perfmon_amd64.c
@@ -0,0 +1,754 @@
+/*
+ * This file contains the PMU description for the Athlon64 and Opteron64
+ * processors. It supports 32 and 64-bit modes.
+ *
+ * Copyright (c) 2005-2007 Hewlett-Packard Development Company, L.P.
+ * Contributed by Stephane Eranian <eranian@hpl.hp.com>
+ *
+ * Copyright (c) 2007 Advanced Micro Devices, Inc.
+ * Contributed by Robert Richter <robert.richter@amd.com>
+ *
+ * This program is free software; you can redistribute it and/or
+ * modify it under the terms of version 2 of the GNU General Public
+ * License as published by the Free Software Foundation.
+ *
+ * This program is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+ * General Public License for more details.
+ *
+ * You should have received a copy of the GNU General Public License
+ * along with this program; if not, write to the Free Software
+ * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA
+ * 02111-1307 USA
+  */
+#include <linux/module.h>
+#include <linux/vmalloc.h>
+#include <linux/topology.h>
+#include <linux/kprobes.h>
+#include <linux/pci.h>
+#include <linux/perfmon_kern.h>
+#include <asm/hw_irq.h>
+#include <asm/apic.h>
+
+MODULE_AUTHOR("Stephane Eranian <eranian@hpl.hp.com>");
+MODULE_AUTHOR("Robert Richter <robert.richter@amd.com>");
+MODULE_DESCRIPTION("AMD64 PMU description table");
+MODULE_LICENSE("GPL");
+
+#define PCI_DEVICE_ID_AMD_10H_NB_MISC  0x1203
+
+static int force_nmi;
+MODULE_PARM_DESC(force_nmi, "bool: force use of NMI for PMU interrupt");
+module_param(force_nmi, bool, 0600);
+
+#define HAS_IBS                0x01    /* has IBS  support */
+
+static u8  ibs_eilvt_off, ibs_status;  /* AMD: extended interrupt LVT offset */
+
+static void pfm_amd64_restore_pmcs(struct pfm_context *ctx,
+                                  struct pfm_event_set *set);
+static void __kprobes pfm_amd64_quiesce(void);
+static int pfm_amd64_has_ovfls(struct pfm_context *ctx);
+static int pfm_amd64_stop_save(struct pfm_context *ctx,
+                              struct pfm_event_set *set);
+
+#define IBSFETCHCTL_PMC        4 /* pmc4 */
+#define IBSFETCHCTL_PMD        4 /* pmd4 */
+#define IBSOPSCTL_PMC  5 /* pmc5 */
+#define IBSOPSCTL_PMD  7 /* pmd7 */
+
+static u64 enable_mask[PFM_MAX_PMCS];
+static u16 max_enable;
+
+static struct pfm_arch_pmu_info pfm_amd64_pmu_info = {
+       .stop_save = pfm_amd64_stop_save,
+       .has_ovfls = pfm_amd64_has_ovfls,
+       .quiesce = pfm_amd64_quiesce,
+       .restore_pmcs = pfm_amd64_restore_pmcs
+};
+
+#define PFM_AMD64_IBSFETCHVAL  (1ULL<<49) /* valid fetch sample */
+#define PFM_AMD64_IBSFETCHEN   (1ULL<<48) /* fetch sampling enabled */
+#define PFM_AMD64_IBSOPVAL     (1ULL<<18) /* valid execution sample */
+#define PFM_AMD64_IBSOPEN      (1ULL<<17) /* execution sampling enabled */
+
+/*
+ * force Local APIC interrupt on overflow
+ */
+#define PFM_K8_VAL     (1ULL<<20)
+#define PFM_K8_NO64    (1ULL<<20)
+
+/*
+ * reserved bits must be 1
+ *
+ * for family 15:
+ * - upper 32 bits are reserved
+ * - bit 20, bit 21
+ *
+ * for family 16:
+ * - bits 36-39 are reserved
+ * - bits 42-63 are reserved
+ * - bit 20, bit 21
+ *
+ * for IBS registers:
+ *     IBSFETCHCTL: all bits are reserved except bits 57, 48, 15:0
+ *     IBSOPSCTL  : all bits are reserved except bits 17, 15:0
+ */
+#define PFM_K8_RSVD    ((~((1ULL<<32)-1)) | (1ULL<<20) | (1ULL<<21))
+#define PFM_16_RSVD ((0x3fffffULL<<42) | (0xfULL<<36) | (1ULL<<20) | (1ULL<<21))
+#define PFM_AMD64_IBSFETCHCTL_RSVD     (~((1ULL<<48)|(1ULL<<57)|0xffffULL))
+#define PFM_AMD64_IBSOPCTL_RSVD                (~((1ULL<<17)|0xffffULL))
+
+static struct pfm_regmap_desc pfm_amd64_pmc_desc[] = {
+/* pmc0  */ PMC_D(PFM_REG_I64, "PERFSEL0", PFM_K8_VAL, PFM_K8_RSVD, PFM_K8_NO64, MSR_K7_EVNTSEL0),
+/* pmc1  */ PMC_D(PFM_REG_I64, "PERFSEL1", PFM_K8_VAL, PFM_K8_RSVD, PFM_K8_NO64, MSR_K7_EVNTSEL1),
+/* pmc2  */ PMC_D(PFM_REG_I64, "PERFSEL2", PFM_K8_VAL, PFM_K8_RSVD, PFM_K8_NO64, MSR_K7_EVNTSEL2),
+/* pmc3  */ PMC_D(PFM_REG_I64, "PERFSEL3", PFM_K8_VAL, PFM_K8_RSVD, PFM_K8_NO64, MSR_K7_EVNTSEL3),
+/* pmc4  */ PMC_D(PFM_REG_I,   "IBSFETCHCTL", 0, PFM_AMD64_IBSFETCHCTL_RSVD, 0, MSR_AMD64_IBSFETCHCTL),
+/* pmc5  */ PMC_D(PFM_REG_I,   "IBSOPCTL",    0, PFM_AMD64_IBSOPCTL_RSVD,    0, MSR_AMD64_IBSOPCTL),
+};
+#define PFM_AMD_NUM_PMCS ARRAY_SIZE(pfm_amd64_pmc_desc)
+
+#define PFM_REG_IBS (PFM_REG_I|PFM_REG_INTR)
+
+/*
+ * AMD64 counters are 48 bits, upper bits are reserved
+ */
+#define PFM_AMD64_CTR_RSVD     (~((1ULL<<48)-1))
+
+#define PFM_AMD_D(n) \
+       { .type = PFM_REG_C,                    \
+         .desc = "PERFCTR"#n,                  \
+         .hw_addr = MSR_K7_PERFCTR0+n,         \
+         .rsvd_msk = PFM_AMD64_CTR_RSVD,       \
+         .dep_pmcs[0] = 1ULL << n              \
+       }
+
+#define PFM_AMD_IBSO(t, s, a) \
+       { .type = t,                    \
+         .desc = s,                    \
+         .hw_addr = a,                 \
+         .rsvd_msk = 0,                \
+         .dep_pmcs[0] = 1ULL << 5      \
+       }
+
+#define PFM_AMD_IBSF(t, s, a) \
+       { .type = t,                    \
+         .desc = s,                    \
+         .hw_addr = a,                 \
+         .rsvd_msk = 0,                \
+         .dep_pmcs[0] = 1ULL << 6      \
+       }
+
+static struct pfm_regmap_desc pfm_amd64_pmd_desc[] = {
+/* pmd0  */ PFM_AMD_D(0),
+/* pmd1  */ PFM_AMD_D(1),
+/* pmd2  */ PFM_AMD_D(2),
+/* pmd3  */ PFM_AMD_D(3),
+/* pmd4  */ PFM_AMD_IBSF(PFM_REG_IBS, "IBSFETCHCTL",   MSR_AMD64_IBSFETCHCTL),
+/* pmd5  */ PFM_AMD_IBSF(PFM_REG_IRO, "IBSFETCHLINAD", MSR_AMD64_IBSFETCHLINAD),
+/* pmd6  */ PFM_AMD_IBSF(PFM_REG_IRO, "IBSFETCHPHYSAD", MSR_AMD64_IBSFETCHPHYSAD),
+/* pmd7  */ PFM_AMD_IBSO(PFM_REG_IBS, "IBSOPCTL",      MSR_AMD64_IBSOPCTL),
+/* pmd8  */ PFM_AMD_IBSO(PFM_REG_IRO, "IBSOPRIP",      MSR_AMD64_IBSOPRIP),
+/* pmd9  */ PFM_AMD_IBSO(PFM_REG_IRO, "IBSOPDATA",     MSR_AMD64_IBSOPDATA),
+/* pmd10 */ PFM_AMD_IBSO(PFM_REG_IRO, "IBSOPDATA2",    MSR_AMD64_IBSOPDATA2),
+/* pmd11 */ PFM_AMD_IBSO(PFM_REG_IRO, "IBSOPDATA3",    MSR_AMD64_IBSOPDATA3),
+/* pmd12 */ PFM_AMD_IBSO(PFM_REG_IRO, "IBSDCLINAD",    MSR_AMD64_IBSDCLINAD),
+/* pmd13 */ PFM_AMD_IBSO(PFM_REG_IRO, "IBSDCPHYSAD",   MSR_AMD64_IBSDCPHYSAD),
+};
+#define PFM_AMD_NUM_PMDS ARRAY_SIZE(pfm_amd64_pmd_desc)
+
+static struct pfm_context **pfm_nb_sys_owners;
+static struct pfm_context *pfm_nb_task_owner;
+
+static struct pfm_pmu_config pfm_amd64_pmu_conf;
+
+#define is_ibs_pmc(x) (x == 4 || x == 5)
+
+static void pfm_amd64_setup_eilvt_per_cpu(void *info)
+{
+       u8 lvt_off;
+
+       /* program the IBS vector to the perfmon vector */
+       lvt_off =  setup_APIC_eilvt_ibs(LOCAL_PERFMON_VECTOR,
+                                       APIC_EILVT_MSG_FIX, 0);
+       PFM_DBG("APIC_EILVT%d set to 0x%x", lvt_off, LOCAL_PERFMON_VECTOR);
+       ibs_eilvt_off = lvt_off;
+}
+
+static int pfm_amd64_setup_eilvt(void)
+{
+#define IBSCTL_LVTOFFSETVAL            (1 << 8)
+#define IBSCTL                         0x1cc
+       struct pci_dev *cpu_cfg;
+       int nodes;
+       u32 value = 0;
+
+       /* per CPU setup */
+       on_each_cpu(pfm_amd64_setup_eilvt_per_cpu, NULL, 1);
+
+       nodes = 0;
+       cpu_cfg = NULL;
+       do {
+               cpu_cfg = pci_get_device(PCI_VENDOR_ID_AMD,
+                                        PCI_DEVICE_ID_AMD_10H_NB_MISC,
+                                        cpu_cfg);
+               if (!cpu_cfg)
+                       break;
+               ++nodes;
+               pci_write_config_dword(cpu_cfg, IBSCTL, ibs_eilvt_off
+                                      | IBSCTL_LVTOFFSETVAL);
+               pci_read_config_dword(cpu_cfg, IBSCTL, &value);
+               if (value != (ibs_eilvt_off | IBSCTL_LVTOFFSETVAL)) {
+                       PFM_DBG("Failed to setup IBS LVT offset, "
+                               "IBSCTL = 0x%08x", value);
+                       return 1;
+               }
+       } while (1);
+
+       if (!nodes) {
+               PFM_DBG("No CPU node configured for IBS");
+               return 1;
+       }
+
+#ifdef CONFIG_NUMA
+       /* Sanity check */
+       /* Works only for 64bit with proper numa implementation. */
+       if (nodes != num_possible_nodes()) {
+               PFM_DBG("Failed to setup CPU node(s) for IBS, "
+                       "found: %d, expected %d",
+                       nodes, num_possible_nodes());
+               return 1;
+       }
+#endif
+       return 0;
+}
+
+/*
+ * There can only be one user per socket for the Northbridge (NB) events,
+ * so we enforce mutual exclusion as follows:
+ *     - per-thread : only one context machine-wide can use NB events
+ *     - system-wide: only one context per processor socket
+ *
+ * Exclusion is enforced at:
+ *     - pfm_load_context()
+ *     - pfm_write_pmcs() for attached contexts
+ *
+ * Exclusion is released at:
+ *     - pfm_unload_context() or any calls that implicitely uses it
+ *
+ * return:
+ *     0  : successfully acquire NB access
+ *     < 0:  errno, failed to acquire NB access
+ */
+static int pfm_amd64_acquire_nb(struct pfm_context *ctx)
+{
+       struct pfm_context **entry, *old;
+       int proc_id;
+
+#ifdef CONFIG_SMP
+       proc_id = cpu_data(smp_processor_id()).phys_proc_id;
+#else
+       proc_id = 0;
+#endif
+
+       if (ctx->flags.system)
+               entry = &pfm_nb_sys_owners[proc_id];
+       else
+               entry = &pfm_nb_task_owner;
+
+       old = cmpxchg(entry, NULL, ctx);
+       if (!old) {
+               if (ctx->flags.system)
+                       PFM_DBG("acquired Northbridge event access on socket %u", proc_id);
+               else
+                       PFM_DBG("acquired Northbridge event access globally");
+       } else if (old != ctx) {
+               if (ctx->flags.system)
+                       PFM_DBG("NorthBridge event conflict on socket %u", proc_id);
+               else
+                       PFM_DBG("global NorthBridge event conflict");
+               return -EBUSY;
+       }
+       return 0;
+}
+
+/*
+ * invoked from pfm_write_pmcs() when pfm_nb_sys_owners is not NULL,i.e.,
+ * when we have detected a multi-core processor.
+ *
+ * context is locked, interrupts are masked
+ */
+static int pfm_amd64_pmc_write_check(struct pfm_context *ctx,
+                            struct pfm_event_set *set,
+                            struct pfarg_pmc *req)
+{
+       unsigned int event;
+
+       /*
+        * delay checking NB event until we load the context
+        */
+       if (ctx->state == PFM_CTX_UNLOADED)
+               return 0;
+
+       /*
+        * check event is NB event
+        */
+       event = (unsigned int)(req->reg_value & 0xff);
+       if (event < 0xee)
+               return 0;
+
+       return pfm_amd64_acquire_nb(ctx);
+}
+
+/*
+ * invoked on pfm_load_context().
+ * context is locked, interrupts are masked
+ */
+static int pfm_amd64_load_context(struct pfm_context *ctx)
+{
+       struct pfm_event_set *set;
+       unsigned int i, n;
+
+       /*
+        * scan all sets for NB events
+        */
+       list_for_each_entry(set, &ctx->set_list, list) {
+               n = set->nused_pmcs;
+               for (i = 0; n; i++) {
+                       if (!test_bit(i, cast_ulp(set->used_pmcs)))
+                               continue;
+
+                       if (!is_ibs_pmc(i) && (set->pmcs[i] & 0xff) >= 0xee)
+                               goto found;
+                       n--;
+               }
+       }
+       return 0;
+found:
+       return pfm_amd64_acquire_nb(ctx);
+}
+
+/*
+ * invoked on pfm_unload_context()
+ */
+static void pfm_amd64_unload_context(struct pfm_context *ctx)
+{
+       struct pfm_context **entry, *old;
+       int proc_id;
+
+#ifdef CONFIG_SMP
+       proc_id = cpu_data(smp_processor_id()).phys_proc_id;
+#else
+       proc_id = 0;
+#endif
+
+       /*
+        * unload always happens on the monitored CPU in system-wide
+        */
+       if (ctx->flags.system)
+               entry = &pfm_nb_sys_owners[proc_id];
+       else
+               entry = &pfm_nb_task_owner;
+
+       old = cmpxchg(entry, ctx, NULL);
+       if (old == ctx) {
+               if (ctx->flags.system)
+                       PFM_DBG("released NorthBridge on socket %u", proc_id);
+               else
+                       PFM_DBG("released NorthBridge events globally");
+       }
+}
+
+/*
+ * detect if we need to activate NorthBridge event access control
+ */
+static int pfm_amd64_setup_nb_event_control(void)
+{
+       unsigned int c, n = 0;
+       unsigned int max_phys = 0;
+
+#ifdef CONFIG_SMP
+       for_each_possible_cpu(c) {
+               if (cpu_data(c).phys_proc_id > max_phys)
+                       max_phys = cpu_data(c).phys_proc_id;
+       }
+#else
+       max_phys = 0;
+#endif
+       if (max_phys > 255) {
+               PFM_INFO("socket id %d is too big to handle", max_phys);
+               return -ENOMEM;
+       }
+
+       n = max_phys + 1;
+       if (n < 2)
+               return 0;
+
+       pfm_nb_sys_owners = vmalloc(n * sizeof(*pfm_nb_sys_owners));
+       if (!pfm_nb_sys_owners)
+               return -ENOMEM;
+
+       memset(pfm_nb_sys_owners, 0, n * sizeof(*pfm_nb_sys_owners));
+       pfm_nb_task_owner = NULL;
+
+       /*
+        * activate write-checker for PMC registers
+        */
+       for (c = 0; c < PFM_AMD_NUM_PMCS; c++) {
+               if (!is_ibs_pmc(c))
+                       pfm_amd64_pmc_desc[c].type |= PFM_REG_WC;
+       }
+
+       pfm_amd64_pmu_info.load_context = pfm_amd64_load_context;
+       pfm_amd64_pmu_info.unload_context = pfm_amd64_unload_context;
+
+       pfm_amd64_pmu_conf.pmc_write_check = pfm_amd64_pmc_write_check;
+
+       PFM_INFO("NorthBridge event access control enabled");
+
+       return 0;
+}
+
+/*
+ * disable registers which are not available on
+ * the host (applies to IBS registers)
+ */
+static void pfm_amd64_check_registers(void)
+{
+       u16 i;
+
+       PFM_DBG("has_ibs=%d", !!(ibs_status & HAS_IBS));
+
+       __set_bit(0, cast_ulp(enable_mask));
+       __set_bit(1, cast_ulp(enable_mask));
+       __set_bit(2, cast_ulp(enable_mask));
+       __set_bit(3, cast_ulp(enable_mask));
+       max_enable = 3+1;
+
+
+       /*
+        * remove IBS registers if feature not present
+        */
+       if (!(ibs_status & HAS_IBS)) {
+               pfm_amd64_pmc_desc[4].type = PFM_REG_NA;
+               pfm_amd64_pmc_desc[5].type = PFM_REG_NA;
+               for (i = 4; i < 14; i++)
+                       pfm_amd64_pmd_desc[i].type = PFM_REG_NA;
+       } else {
+               __set_bit(16, cast_ulp(enable_mask));
+               __set_bit(17, cast_ulp(enable_mask));
+               max_enable = 17 + 1;
+       }
+
+       /*
+        * adjust reserved bit fields for family 16
+        */
+       if (current_cpu_data.x86 == 16) {
+               for (i = 0; i < PFM_AMD_NUM_PMCS; i++)
+                       if (pfm_amd64_pmc_desc[i].rsvd_msk == PFM_K8_RSVD)
+                               pfm_amd64_pmc_desc[i].rsvd_msk = PFM_16_RSVD;
+       }
+}
+
+static int pfm_amd64_probe_pmu(void)
+{
+       u64 val = 0;
+       if (current_cpu_data.x86_vendor != X86_VENDOR_AMD) {
+               PFM_INFO("not an AMD processor");
+               return -1;
+       }
+
+       switch (current_cpu_data.x86) {
+       case 16:
+       case 15:
+       case  6:
+               break;
+       default:
+               PFM_INFO("unsupported family=%d", current_cpu_data.x86);
+               return -1;
+       }
+
+       /* check for IBS */
+       if (cpu_has(&current_cpu_data, X86_FEATURE_IBS)) {
+               ibs_status |= HAS_IBS;
+               rdmsrl(MSR_AMD64_IBSCTL, val);
+       }
+
+       PFM_INFO("found family=%d IBSCTL=0x%llx", current_cpu_data.x86, (unsigned long long)val);
+
+       /*
+        * check for local APIC (required)
+        */
+       if (!cpu_has_apic) {
+               PFM_INFO("no local APIC, unsupported");
+               return -1;
+       }
+
+       if (current_cpu_data.x86_max_cores > 1
+           && pfm_amd64_setup_nb_event_control())
+               return -1;
+
+       if (force_nmi)
+               pfm_amd64_pmu_info.flags |= PFM_X86_FL_USE_NMI;
+
+       if (ibs_status & HAS_IBS) {
+               /* Setup extended interrupt */
+               if (pfm_amd64_setup_eilvt()) {
+                       PFM_INFO("Failed to initialize extended interrupts "
+                                "for IBS");
+                       ibs_status  &= ~HAS_IBS;
+                       PFM_INFO("Unable to use IBS");
+               } else {
+                       PFM_INFO("IBS supported");
+               }
+       }
+
+       pfm_amd64_check_registers();
+
+       return 0;
+}
+
+/*
+ * detect is counters have overflowed.
+ * return:
+ *     0 : no overflow
+ *     1 : at least one overflow
+ */
+static int __kprobes pfm_amd64_has_ovfls(struct pfm_context *ctx)
+{
+       struct pfm_regmap_desc *xrd;
+       u64 *cnt_mask;
+       u64 wmask, val;
+       u16 i, num;
+
+       /*
+        * Check for IBS events
+        */
+       if (ibs_status & HAS_IBS) {
+               rdmsrl(MSR_AMD64_IBSFETCHCTL, val);
+               if (val & PFM_AMD64_IBSFETCHVAL)
+                       return 1;
+               rdmsrl(MSR_AMD64_IBSOPCTL, val);
+               if (val & PFM_AMD64_IBSOPVAL)
+                       return 1;
+       }
+       /*
+        * Check regular counters
+        */
+       cnt_mask = ctx->regs.cnt_pmds;
+       num = ctx->regs.num_counters;
+       wmask = 1ULL << pfm_pmu_conf->counter_width;
+       xrd = pfm_amd64_pmd_desc;
+
+       for (i = 0; num; i++) {
+               if (test_bit(i, cast_ulp(cnt_mask))) {
+                       rdmsrl(xrd[i].hw_addr, val);
+                       if (!(val & wmask))
+                               return 1;
+                       num--;
+               }
+       }
+       return 0;
+}
+
+/*
+ * Must check for IBS event BEFORE stop_save_p6 because
+ * stopping monitoring does destroy IBS state information
+ * in IBSFETCHCTL/IBSOPCTL because they are tagged as enable
+ * registers.
+ */
+static int pfm_amd64_stop_save(struct pfm_context *ctx, struct pfm_event_set *set)
+{
+       struct pfm_arch_pmu_info *pmu_info;
+       u64 used_mask[PFM_PMC_BV];
+       u64 *cnt_pmds;
+       u64 val, wmask, ovfl_mask;
+       u32 i, count, use_ibs;
+
+       pmu_info = pfm_pmu_info();
+
+       /*
+        * IBS used if:
+        *   - on family 10h processor with IBS
+        *   - at least one of the IBS PMD registers is used
+        */
+       use_ibs = (ibs_status & HAS_IBS)
+               && (test_bit(IBSFETCHCTL_PMD, cast_ulp(set->used_pmds))
+                   || test_bit(IBSOPSCTL_PMD, cast_ulp(set->used_pmds)));
+
+       wmask = 1ULL << pfm_pmu_conf->counter_width;
+
+       bitmap_and(cast_ulp(used_mask),
+                  cast_ulp(set->used_pmcs),
+                  cast_ulp(enable_mask),
+                  max_enable);
+
+       count = bitmap_weight(cast_ulp(used_mask), max_enable);
+
+       /*
+        * stop monitoring
+        * Unfortunately, this is very expensive!
+        * wrmsrl() is serializing.
+        *
+        * With IBS, we need to do read-modify-write to preserve the content
+        * for OpsCTL and FetchCTL because they are also used as PMDs and saved
+        * below
+        */
+       if (use_ibs) {
+               for (i = 0; count; i++) {
+                       if (test_bit(i, cast_ulp(used_mask))) {
+                               if (i == IBSFETCHCTL_PMC) {
+                                       rdmsrl(pfm_pmu_conf->pmc_desc[i].hw_addr, val);
+                                       val &= ~PFM_AMD64_IBSFETCHEN;
+                               } else if (i == IBSOPSCTL_PMC) {
+                                       rdmsrl(pfm_pmu_conf->pmc_desc[i].hw_addr, val);
+                                       val &= ~PFM_AMD64_IBSOPEN;
+                               } else
+                                       val = 0;
+                               wrmsrl(pfm_pmu_conf->pmc_desc[i].hw_addr, val);
+                               count--;
+                       }
+               }
+       } else {
+               for (i = 0; count; i++) {
+                       if (test_bit(i, cast_ulp(used_mask))) {
+                               wrmsrl(pfm_pmu_conf->pmc_desc[i].hw_addr, 0);
+                               count--;
+                       }
+               }
+       }
+
+       /*
+        * if we already having a pending overflow condition, we simply
+        * return to take care of this first.
+        */
+       if (set->npend_ovfls)
+               return 1;
+
+       ovfl_mask = pfm_pmu_conf->ovfl_mask;
+       cnt_pmds = ctx->regs.cnt_pmds;
+
+       /*
+        * check for pending overflows and save PMDs (combo)
+        * we employ used_pmds because we also need to save
+        * and not just check for pending interrupts.
+        *
+        * Must check for counting PMDs because of virtual PMDs and IBS
+        */
+       count = set->nused_pmds;
+       for (i = 0; count; i++) {
+               if (test_bit(i, cast_ulp(set->used_pmds))) {
+                       val = pfm_arch_read_pmd(ctx, i);
+                       if (likely(test_bit(i, cast_ulp(cnt_pmds)))) {
+                               if (!(val & wmask)) {
+                                       __set_bit(i, cast_ulp(set->povfl_pmds));
+                                       set->npend_ovfls++;
+                               }
+                               val = (set->pmds[i].value & ~ovfl_mask) | (val & ovfl_mask);
+                       }
+                       set->pmds[i].value = val;
+                       count--;
+               }
+       }
+
+       /*
+        * check if IBS contains valid data, and mark the corresponding
+        * PMD has overflowed
+        */
+       if (use_ibs) {
+               if (set->pmds[IBSFETCHCTL_PMD].value & PFM_AMD64_IBSFETCHVAL) {
+                       __set_bit(IBSFETCHCTL_PMD, cast_ulp(set->povfl_pmds));
+                       set->npend_ovfls++;
+               }
+               if (set->pmds[IBSOPSCTL_PMD].value & PFM_AMD64_IBSOPVAL) {
+                       __set_bit(IBSOPSCTL_PMD, cast_ulp(set->povfl_pmds));
+                       set->npend_ovfls++;
+               }
+       }
+       /* 0 means: no need to save PMDs at upper level */
+       return 0;
+}
+
+/**
+ * pfm_amd64_quiesce_pmu -- stop monitoring without grabbing any lock
+ *
+ * called from NMI interrupt handler to immediately stop monitoring
+ * cannot grab any lock, including perfmon related locks
+ */
+static void __kprobes pfm_amd64_quiesce(void)
+{
+       /*
+        * quiesce PMU by clearing available registers that have
+        * the start/stop capability
+        */
+       if (test_bit(0, cast_ulp(pfm_pmu_conf->regs_all.pmcs)))
+               wrmsrl(MSR_K7_EVNTSEL0, 0);
+       if (test_bit(1, cast_ulp(pfm_pmu_conf->regs_all.pmcs)))
+               wrmsrl(MSR_K7_EVNTSEL0+1, 0);
+       if (test_bit(2, cast_ulp(pfm_pmu_conf->regs_all.pmcs)))
+               wrmsrl(MSR_K7_EVNTSEL0+2, 0);
+       if (test_bit(3, cast_ulp(pfm_pmu_conf->regs_all.pmcs)))
+               wrmsrl(MSR_K7_EVNTSEL0+3, 0);
+
+       if (test_bit(4, cast_ulp(pfm_pmu_conf->regs_all.pmcs)))
+               wrmsrl(MSR_AMD64_IBSFETCHCTL, 0);
+       if (test_bit(5, cast_ulp(pfm_pmu_conf->regs_all.pmcs)))
+               wrmsrl(MSR_AMD64_IBSOPCTL, 0);
+}
+
+/**
+ * pfm_amd64_restore_pmcs - reload PMC registers
+ * @ctx: context to restore from
+ * @set: current event set
+ *
+ * optimized version of pfm_arch_restore_pmcs(). On AMD64, we can
+ * afford to only restore the pmcs registers we use, because they are
+ * all independent from each other.
+ */
+static void pfm_amd64_restore_pmcs(struct pfm_context *ctx,
+                                  struct pfm_event_set *set)
+{
+       u64 *mask;
+       u16 i, num;
+
+       mask = set->used_pmcs;
+       num = set->nused_pmcs;
+       for (i = 0; num; i++) {
+               if (test_bit(i, cast_ulp(mask))) {
+                       wrmsrl(pfm_amd64_pmc_desc[i].hw_addr, set->pmcs[i]);
+                       num--;
+               }
+       }
+}
+
+static struct pfm_pmu_config pfm_amd64_pmu_conf = {
+       .pmu_name = "AMD64",
+       .counter_width = 47,
+       .pmd_desc = pfm_amd64_pmd_desc,
+       .pmc_desc = pfm_amd64_pmc_desc,
+       .num_pmc_entries = PFM_AMD_NUM_PMCS,
+       .num_pmd_entries = PFM_AMD_NUM_PMDS,
+       .probe_pmu = pfm_amd64_probe_pmu,
+       .version = "1.2",
+       .pmu_info = &pfm_amd64_pmu_info,
+       .flags = PFM_PMU_BUILTIN_FLAG,
+       .owner = THIS_MODULE,
+};
+
+static int __init pfm_amd64_pmu_init_module(void)
+{
+       return pfm_pmu_register(&pfm_amd64_pmu_conf);
+}
+
+static void __exit pfm_amd64_pmu_cleanup_module(void)
+{
+       if (pfm_nb_sys_owners)
+               vfree(pfm_nb_sys_owners);
+
+       pfm_pmu_unregister(&pfm_amd64_pmu_conf);
+}
+
+module_init(pfm_amd64_pmu_init_module);
+module_exit(pfm_amd64_pmu_cleanup_module);
diff --git a/arch/x86/perfmon/perfmon_intel_arch.c b/arch/x86/perfmon/perfmon_intel_arch.c
new file mode 100644
index 0000000..e27a732
--- /dev/null
+++ b/arch/x86/perfmon/perfmon_intel_arch.c
@@ -0,0 +1,610 @@
+/*
+ * This file contains the Intel architectural perfmon v1, v2, v3
+ * description tables.
+ *
+ * Architectural perfmon was introduced with Intel Core Solo/Duo
+ * processors.
+ *
+ * Copyright (c) 2006-2007 Hewlett-Packard Development Company, L.P.
+ * Contributed by Stephane Eranian <eranian@hpl.hp.com>
+ *
+ * This program is free software; you can redistribute it and/or
+ * modify it under the terms of version 2 of the GNU General Public
+ * License as published by the Free Software Foundation.
+ *
+ * This program is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+ * General Public License for more details.
+ *
+ * You should have received a copy of the GNU General Public License
+ * along with this program; if not, write to the Free Software
+ * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA
+ * 02111-1307 USA
+ */
+#include <linux/module.h>
+#include <linux/kprobes.h>
+#include <linux/perfmon_kern.h>
+#include <linux/nmi.h>
+#include <asm/msr.h>
+#include <asm/apic.h>
+
+MODULE_AUTHOR("Stephane Eranian <eranian@hpl.hp.com>");
+MODULE_DESCRIPTION("Intel architectural perfmon v1");
+MODULE_LICENSE("GPL");
+
+static int force, force_nmi;
+MODULE_PARM_DESC(force, "bool: force module to load succesfully");
+MODULE_PARM_DESC(force_nmi, "bool: force use of NMI for PMU interrupt");
+module_param(force, bool, 0600);
+module_param(force_nmi, bool, 0600);
+
+static u64 enable_mask[PFM_MAX_PMCS];
+static u16 max_enable;
+
+/*
+ * - upper 32 bits are reserved
+ * - INT: APIC enable bit is reserved (forced to 1)
+ * - bit 21 is reserved
+ *
+ * RSVD: reserved bits are 1
+ */
+#define PFM_IA_PMC_RSVD        ((~((1ULL<<32)-1)) \
+                       | (1ULL<<20) \
+                       | (1ULL<<21))
+
+/*
+ * force Local APIC interrupt on overflow
+ * disable with NO_EMUL64
+ */
+#define PFM_IA_PMC_VAL (1ULL<<20)
+#define PFM_IA_NO64    (1ULL<<20)
+
+/*
+ * architectuture specifies that:
+ * IA32_PMCx MSR        : starts at 0x0c1 & occupy a contiguous block of MSR
+ * IA32_PERFEVTSELx MSR : starts at 0x186 & occupy a contiguous block of MSR
+ * MSR_GEN_FIXED_CTR0   : starts at 0x309 & occupy a contiguous block of MSR
+ */
+#define MSR_GEN_SEL_BASE       MSR_P6_EVNTSEL0
+#define MSR_GEN_PMC_BASE       MSR_P6_PERFCTR0
+#define MSR_GEN_FIXED_PMC_BASE MSR_CORE_PERF_FIXED_CTR0
+
+/*
+ * layout of EAX for CPUID.0xa leaf function
+ */
+struct pmu_eax {
+       unsigned int version:8;         /* architectural perfmon version */
+       unsigned int num_cnt:8;         /* number of generic counters */
+       unsigned int cnt_width:8;       /* width of generic counters */
+       unsigned int ebx_length:8;      /* number of architected events */
+};
+
+/*
+ * layout of EDX for CPUID.0xa leaf function when perfmon v2 is detected
+ */
+struct pmu_edx {
+       unsigned int num_cnt:5;         /* number of fixed counters */
+       unsigned int cnt_width:8;       /* width of fixed counters */
+       unsigned int reserved:19;
+};
+
+static void pfm_intel_arch_restore_pmcs(struct pfm_context *ctx,
+                                       struct pfm_event_set *set);
+static int pfm_intel_arch_stop_save(struct pfm_context *ctx,
+                                   struct pfm_event_set *set);
+static int pfm_intel_arch_has_ovfls(struct pfm_context *ctx);
+static void __kprobes pfm_intel_arch_quiesce(void);
+
+/*
+ * physical addresses of MSR controlling the perfevtsel and counter registers
+ */
+struct pfm_arch_pmu_info pfm_intel_arch_pmu_info = {
+       .stop_save = pfm_intel_arch_stop_save,
+       .has_ovfls = pfm_intel_arch_has_ovfls,
+       .quiesce = pfm_intel_arch_quiesce,
+       .restore_pmcs = pfm_intel_arch_restore_pmcs
+};
+
+#define PFM_IA_C(n) {                   \
+       .type = PFM_REG_I64,            \
+       .desc = "PERFEVTSEL"#n,         \
+       .dfl_val = PFM_IA_PMC_VAL,      \
+       .rsvd_msk = PFM_IA_PMC_RSVD,    \
+       .no_emul64_msk = PFM_IA_NO64,   \
+       .hw_addr = MSR_GEN_SEL_BASE+(n) \
+       }
+
+#define PFM_IA_D(n) \
+       { .type = PFM_REG_C,                    \
+         .desc = "PMC"#n,                      \
+         .hw_addr = MSR_P6_PERFCTR0+n,         \
+         .dep_pmcs[0] = 1ULL << n              \
+       }
+
+#define PFM_IA_FD(n) \
+       { .type = PFM_REG_C,                    \
+         .desc = "FIXED_CTR"#n,                \
+         .hw_addr = MSR_CORE_PERF_FIXED_CTR0+n,\
+         .dep_pmcs[0] = 1ULL << 16             \
+       }
+
+static struct pfm_regmap_desc pfm_intel_arch_pmc_desc[] = {
+/* pmc0  */ PFM_IA_C(0),  PFM_IA_C(1),   PFM_IA_C(2),  PFM_IA_C(3),
+/* pmc4  */ PFM_IA_C(4),  PFM_IA_C(5),   PFM_IA_C(6),  PFM_IA_C(7),
+/* pmc8  */ PFM_IA_C(8),  PFM_IA_C(9),  PFM_IA_C(10), PFM_IA_C(11),
+/* pmc12 */ PFM_IA_C(12), PFM_IA_C(13), PFM_IA_C(14), PFM_IA_C(15),
+
+/* pmc16 */ { .type = PFM_REG_I,
+             .desc = "FIXED_CTRL",
+             .dfl_val = 0x8888888888888888ULL, /* force PMI */
+             .rsvd_msk = 0, /* set dynamically */
+             .no_emul64_msk = 0,
+             .hw_addr = MSR_CORE_PERF_FIXED_CTR_CTRL
+           },
+};
+#define PFM_IA_MAX_PMCS        ARRAY_SIZE(pfm_intel_arch_pmc_desc)
+
+static struct pfm_regmap_desc pfm_intel_arch_pmd_desc[] = {
+/* pmd0  */  PFM_IA_D(0),  PFM_IA_D(1),  PFM_IA_D(2),  PFM_IA_D(3),
+/* pmd4  */  PFM_IA_D(4),  PFM_IA_D(5),  PFM_IA_D(6),  PFM_IA_D(7),
+/* pmd8  */  PFM_IA_D(8),  PFM_IA_D(9), PFM_IA_D(10), PFM_IA_D(11),
+/* pmd12 */ PFM_IA_D(12), PFM_IA_D(13), PFM_IA_D(14), PFM_IA_D(15),
+
+/* pmd16 */ PFM_IA_FD(0), PFM_IA_FD(1), PFM_IA_FD(2), PFM_IA_FD(3),
+/* pmd20 */ PFM_IA_FD(4), PFM_IA_FD(5), PFM_IA_FD(6), PFM_IA_FD(7),
+/* pmd24 */ PFM_IA_FD(8), PFM_IA_FD(9), PFM_IA_FD(10), PFM_IA_FD(11),
+/* pmd28 */ PFM_IA_FD(16), PFM_IA_FD(17), PFM_IA_FD(18), PFM_IA_FD(19)
+};
+#define PFM_IA_MAX_PMDS        ARRAY_SIZE(pfm_intel_arch_pmd_desc)
+
+#define PFM_IA_MAX_CNT         16 /* # generic counters in mapping table */
+#define PFM_IA_MAX_FCNT                16 /* # of fixed counters in mapping table */
+#define PFM_IA_FCNT_BASE       16 /* base index of fixed counters PMD */
+
+static struct pfm_pmu_config pfm_intel_arch_pmu_conf;
+
+static void pfm_intel_arch_check_errata(void)
+{
+       /*
+        * Core Duo errata AE49 (no fix). Both counters share a single
+        * enable bit in PERFEVTSEL0
+        */
+       if (current_cpu_data.x86 == 6 && current_cpu_data.x86_model == 14)
+               pfm_intel_arch_pmu_info.flags |= PFM_X86_FL_NO_SHARING;
+}
+
+static inline void set_enable_mask(unsigned int i)
+{
+       __set_bit(i, cast_ulp(enable_mask));
+
+       /* max_enable = highest + 1 */
+       if ((i+1) > max_enable)
+               max_enable = i+ 1;
+}
+
+static void pfm_intel_arch_setup_generic(unsigned int version,
+                                        unsigned int width,
+                                        unsigned int count)
+{
+       u64 rsvd;
+       unsigned int i;
+
+       /*
+        * first we handle the generic counters:
+        *
+        * - ensure HW does not have more registers than hardcoded in the tables
+        * - adjust rsvd_msk to actual counter width
+        * - initialize enable_mask (list of PMC with start/stop capability)
+        * - mark unused hardcoded generic counters as unimplemented
+        */
+
+       /*
+        * min of number of Hw counters and hardcoded in the tables
+        */
+       if (count >= PFM_IA_MAX_CNT) {
+               printk(KERN_INFO "perfmon: Limiting number of generic counters"
+                                " to %u, HW supports %u",
+                                PFM_IA_MAX_CNT, count);
+               count = PFM_IA_MAX_CNT;
+       }
+
+       /*
+        * adjust rsvd_msk for generic counters based on actual width
+        * initialize enable_mask (1 per pmd)
+        */
+       rsvd = ~((1ULL << width)-1);
+       for (i = 0; i < count; i++) {
+               pfm_intel_arch_pmd_desc[i].rsvd_msk = rsvd;
+               set_enable_mask(i);
+       }
+
+       /*
+        * handle version 3 new anythread bit (21)
+        */
+       if (version == 3) {
+               for (i = 0; i < count; i++)
+                       pfm_intel_arch_pmc_desc[i].rsvd_msk &= ~(1ULL << 21);
+       }
+
+
+       /*
+        * mark unused generic counters as not available
+        */
+       for (i = count ; i < PFM_IA_MAX_CNT; i++) {
+               pfm_intel_arch_pmd_desc[i].type = PFM_REG_NA;
+               pfm_intel_arch_pmc_desc[i].type = PFM_REG_NA;
+       }
+}
+
+static void pfm_intel_arch_setup_fixed(unsigned int version,
+                                      unsigned int width,
+                                      unsigned int count)
+{
+       u64 rsvd, dfl;
+       unsigned int i;
+
+       /*
+        * handle the fixed counters (if any):
+        *
+        * - ensure HW does not have more registers than hardcoded in the tables
+        * - adjust rsvd_msk to actual counter width
+        * - initialize enable_mask (list of PMC with start/stop capability)
+        * - mark unused hardcoded generic counters as unimplemented
+        */
+       if (count >= PFM_IA_MAX_FCNT) {
+               printk(KERN_INFO "perfmon: Limiting number of fixed counters"
+                                " to %u, HW supports %u",
+                                PFM_IA_MAX_FCNT, count);
+               count = PFM_IA_MAX_FCNT;
+       }
+       /*
+        * adjust rsvd_msk for fixed counters based on actual width
+        */
+       rsvd = ~((1ULL << width)-1);
+       for (i = 0; i < count; i++)
+               pfm_intel_arch_pmd_desc[PFM_IA_FCNT_BASE+i].rsvd_msk = rsvd;
+
+       /*
+        * handle version new anythread bit (bit 2)
+        */
+       if (version == 3)
+               rsvd = 1ULL << 3;
+       else
+               rsvd = 3ULL << 2;
+
+       pfm_intel_arch_pmc_desc[16].rsvd_msk = 0;
+       for (i = 0; i < count; i++)
+               pfm_intel_arch_pmc_desc[16].rsvd_msk |= rsvd << (i<<2);
+
+       /*
+        * mark unused fixed counters as unimplemented
+        *
+        * update the rsvd_msk, dfl_val in FIXED_CTRL:
+        *      - rsvd_msk: set all 4 bits
+        *      - dfl_val : clear all 4 bits
+        */
+       dfl = pfm_intel_arch_pmc_desc[16].dfl_val;
+       rsvd = pfm_intel_arch_pmc_desc[16].rsvd_msk;
+
+       for (i = count ; i < PFM_IA_MAX_FCNT; i++) {
+               pfm_intel_arch_pmd_desc[PFM_IA_FCNT_BASE+i].type = PFM_REG_NA;
+               rsvd |= 0xfULL << (i<<2);
+               dfl &= ~(0xfULL << (i<<2));
+       }
+
+       /*
+        * FIXED_CTR_CTRL unavailable when no fixed counters are defined
+        */
+       if (!count) {
+               pfm_intel_arch_pmc_desc[16].type = PFM_REG_NA;
+       } else {
+               /* update rsvd_mask and dfl_val */
+               pfm_intel_arch_pmc_desc[16].rsvd_msk = rsvd;
+               pfm_intel_arch_pmc_desc[16].dfl_val = dfl;
+               set_enable_mask(16);
+       }
+}
+
+static int pfm_intel_arch_probe_pmu(void)
+{
+       union {
+               unsigned int val;
+               struct pmu_eax eax;
+               struct pmu_edx edx;
+       } eax, edx;
+       unsigned int ebx, ecx;
+       unsigned int width = 0;
+
+       edx.val = 0;
+
+       if (!(cpu_has_arch_perfmon || force)) {
+               PFM_INFO("no support for Intel architectural PMU");
+               return -1;
+       }
+
+       if (!cpu_has_apic) {
+               PFM_INFO("no Local APIC, try rebooting with lapic option");
+               return -1;
+       }
+
+       /* cpuid() call protected by cpu_has_arch_perfmon */
+       cpuid(0xa, &eax.val, &ebx, &ecx, &edx.val);
+
+       /*
+        * reject processors supported by perfmon_intel_core
+        *
+        * We need to do this explicitely to avoid depending
+        * on the link order in case, the modules are compiled as
+        * builtin.
+        *
+        * non Intel processors are rejected by cpu_has_arch_perfmon
+        */
+       if (current_cpu_data.x86 == 6 && !force) {
+               switch (current_cpu_data.x86_model) {
+               case 15: /* Merom: use perfmon_intel_core  */
+               case 23: /* Penryn: use perfmon_intel_core */
+                       return -1;
+               default:
+                       break;
+               }
+       }
+
+       /*
+        * some 6/15 models have buggy BIOS
+        */
+       if (eax.eax.version == 0
+           && current_cpu_data.x86 == 6 && current_cpu_data.x86_model == 15) {
+               PFM_INFO("buggy v2 BIOS, adjusting for 2 generic counters");
+               eax.eax.version = 2;
+               eax.eax.num_cnt = 2;
+               eax.eax.cnt_width = 40;
+       }
+
+       /*
+        * Intel Atom processors have a buggy firmware which does not report
+        * the correct number of fixed counters
+        */
+       if (eax.eax.version == 3 && edx.edx.num_cnt < 3
+           && current_cpu_data.x86 == 6 && current_cpu_data.x86_model == 28) {
+               PFM_INFO("buggy v3 BIOS, adjusting for 3 fixed counters");
+               edx.edx.num_cnt = 3;
+       }
+
+       /*
+        * some v2 BIOSes are incomplete
+        */
+       if (eax.eax.version == 2 && !edx.edx.num_cnt) {
+               PFM_INFO("buggy v2 BIOS, adjusting for 3 fixed counters");
+               edx.edx.num_cnt = 3;
+               edx.edx.cnt_width = 40;
+       }
+
+       /*
+        * no fixed counters on earlier versions
+        */
+       if (eax.eax.version < 2) {
+               edx.val = 0;
+       } else {
+               /*
+                * use the min value of both widths until we support
+                * variable width counters
+                */
+               width = eax.eax.cnt_width < edx.edx.cnt_width ?
+                       eax.eax.cnt_width : edx.edx.cnt_width;
+       }
+
+       PFM_INFO("detected architecural perfmon v%d", eax.eax.version);
+       PFM_INFO("num_gen=%d width=%d num_fixed=%d width=%d",
+                 eax.eax.num_cnt,
+                 eax.eax.cnt_width,
+                 edx.edx.num_cnt,
+                 edx.edx.cnt_width);
+
+
+       pfm_intel_arch_setup_generic(eax.eax.version,
+                                    width,
+                                    eax.eax.num_cnt);
+
+       pfm_intel_arch_setup_fixed(eax.eax.version,
+                                  width,
+                                  edx.edx.num_cnt);
+
+       if (force_nmi)
+               pfm_intel_arch_pmu_info.flags |= PFM_X86_FL_USE_NMI;
+
+       pfm_intel_arch_check_errata();
+
+       return 0;
+}
+
+/**
+ * pfm_intel_arch_has_ovfls - check for pending overflow condition
+ * @ctx: context to work on
+ *
+ * detect if counters have overflowed.
+ * return:
+ *     0 : no overflow
+ *     1 : at least one overflow
+ */
+static int __kprobes pfm_intel_arch_has_ovfls(struct pfm_context *ctx)
+{
+       u64 *cnt_mask;
+       u64 wmask, val;
+       u16 i, num;
+
+       cnt_mask = ctx->regs.cnt_pmds;
+       num = ctx->regs.num_counters;
+       wmask = 1ULL << pfm_pmu_conf->counter_width;
+
+       /*
+        * we can leverage the fact that we know the mapping
+        * to hardcode the MSR address and avoid accessing
+        * more cachelines
+        *
+        * We need to check cnt_mask because not all registers
+        * may be available.
+        */
+       for (i = 0; num; i++) {
+               if (test_bit(i, cast_ulp(cnt_mask))) {
+                       rdmsrl(pfm_intel_arch_pmd_desc[i].hw_addr, val);
+                       if (!(val & wmask))
+                               return 1;
+                       num--;
+               }
+       }
+       return 0;
+}
+
+static int pfm_intel_arch_stop_save(struct pfm_context *ctx,
+                                   struct pfm_event_set *set)
+{
+       u64 used_mask[PFM_PMC_BV];
+       u64 *cnt_pmds;
+       u64 val, wmask, ovfl_mask;
+       u32 i, count;
+
+       wmask = 1ULL << pfm_pmu_conf->counter_width;
+
+       bitmap_and(cast_ulp(used_mask),
+                  cast_ulp(set->used_pmcs),
+                  cast_ulp(enable_mask),
+                  max_enable);
+
+       count = bitmap_weight(cast_ulp(used_mask), max_enable);
+
+       /*
+        * stop monitoring
+        * Unfortunately, this is very expensive!
+        * wrmsrl() is serializing.
+        */
+       for (i = 0; count; i++) {
+               if (test_bit(i, cast_ulp(used_mask))) {
+                       wrmsrl(pfm_pmu_conf->pmc_desc[i].hw_addr, 0);
+                       count--;
+               }
+       }
+
+       /*
+        * if we already having a pending overflow condition, we simply
+        * return to take care of this first.
+        */
+       if (set->npend_ovfls)
+               return 1;
+
+       ovfl_mask = pfm_pmu_conf->ovfl_mask;
+       cnt_pmds = ctx->regs.cnt_pmds;
+
+       /*
+        * check for pending overflows and save PMDs (combo)
+        * we employ used_pmds because we also need to save
+        * and not just check for pending interrupts.
+        *
+        * Must check for counting PMDs because of virtual PMDs
+        */
+       count = set->nused_pmds;
+       for (i = 0; count; i++) {
+               if (test_bit(i, cast_ulp(set->used_pmds))) {
+                       val = pfm_arch_read_pmd(ctx, i);
+                       if (likely(test_bit(i, cast_ulp(cnt_pmds)))) {
+                               if (!(val & wmask)) {
+                                       __set_bit(i, cast_ulp(set->povfl_pmds));
+                                       set->npend_ovfls++;
+                               }
+                               val = (set->pmds[i].value & ~ovfl_mask)
+                                       | (val & ovfl_mask);
+                       }
+                       set->pmds[i].value = val;
+                       count--;
+               }
+       }
+       /* 0 means: no need to save PMDs at upper level */
+       return 0;
+}
+
+/**
+ * pfm_intel_arch_quiesce - stop monitoring without grabbing any lock
+ *
+ * called from NMI interrupt handler to immediately stop monitoring
+ * cannot grab any lock, including perfmon related locks
+ */
+static void __kprobes pfm_intel_arch_quiesce(void)
+{
+       u16 i;
+
+       /*
+        * PMC16 is the fixed control control register so it has a
+        * distinct MSR address
+        *
+        * We do not use the hw_addr field in the table to avoid touching
+        * too many cachelines
+        */
+       for (i = 0; i < pfm_pmu_conf->regs_all.max_pmc; i++) {
+               if (test_bit(i, cast_ulp(pfm_pmu_conf->regs_all.pmcs))) {
+                       if (i == 16)
+                               wrmsrl(MSR_CORE_PERF_FIXED_CTR_CTRL, 0);
+                       else
+                               wrmsrl(MSR_P6_EVNTSEL0+i, 0);
+               }
+       }
+}
+
+/**
+ * pfm_intel_arch_restore_pmcs - reload PMC registers
+ * @ctx: context to restore from
+ * @set: current event set
+ *
+ * optimized version of pfm_arch_restore_pmcs(). On architectural perfmon,
+ * we can afford to only restore the pmcs registers we use, because they
+ * are all independent from each other.
+ */
+static void pfm_intel_arch_restore_pmcs(struct pfm_context *ctx,
+                                       struct pfm_event_set *set)
+{
+       u64 *mask;
+       u16 i, num;
+
+       mask = set->used_pmcs;
+       num = set->nused_pmcs;
+       for (i = 0; num; i++) {
+               if (test_bit(i, cast_ulp(mask))) {
+                       wrmsrl(pfm_pmu_conf->pmc_desc[i].hw_addr, set->pmcs[i]);
+                       num--;
+               }
+       }
+}
+/*
+ * Counters may have model-specific width. Yet the documentation says
+ * that only the lower 32 bits can be written to due to the specification
+ * of wrmsr. bits [32-(w-1)] are sign extensions of bit 31. Bits [w-63] must
+ * not be set (see rsvd_msk for PMDs). As such the effective width of a
+ * counter is 31 bits only regardless of what CPUID.0xa returns.
+ *
+ * See IA-32 Intel Architecture Software developer manual Vol 3B chapter 18
+ */
+static struct pfm_pmu_config pfm_intel_arch_pmu_conf = {
+       .pmu_name = "Intel architectural",
+       .pmd_desc = pfm_intel_arch_pmd_desc,
+       .counter_width   = 31,
+       .num_pmc_entries = PFM_IA_MAX_PMCS,
+       .num_pmd_entries = PFM_IA_MAX_PMDS,
+       .pmc_desc = pfm_intel_arch_pmc_desc,
+       .probe_pmu = pfm_intel_arch_probe_pmu,
+       .version = "1.0",
+       .flags = PFM_PMU_BUILTIN_FLAG,
+       .owner = THIS_MODULE,
+       .pmu_info = &pfm_intel_arch_pmu_info
+};
+
+static int __init pfm_intel_arch_pmu_init_module(void)
+{
+       return pfm_pmu_register(&pfm_intel_arch_pmu_conf);
+}
+
+static void __exit pfm_intel_arch_pmu_cleanup_module(void)
+{
+       pfm_pmu_unregister(&pfm_intel_arch_pmu_conf);
+}
+
+module_init(pfm_intel_arch_pmu_init_module);
+module_exit(pfm_intel_arch_pmu_cleanup_module);
diff --git a/arch/x86/perfmon/perfmon_intel_atom.c b/arch/x86/perfmon/perfmon_intel_atom.c
new file mode 100644
index 0000000..9b94863
--- /dev/null
+++ b/arch/x86/perfmon/perfmon_intel_atom.c
@@ -0,0 +1,541 @@
+/*
+ * perfmon support for Intel Atom (architectural perfmon v3 + PEBS)
+ *
+ * Copyright (c) 2008 Google,Inc
+ * Contributed by Stephane Eranian <eranian@gmail.com>
+ *
+ * This program is free software; you can redistribute it and/or
+ * modify it under the terms of version 2 of the GNU General Public
+ * License as published by the Free Software Foundation.
+ *
+ * This program is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+ * General Public License for more details.
+ *
+ * You should have received a copy of the GNU General Public License
+ * along with this program; if not, write to the Free Software
+ * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA
+ * 02111-1307 USA
+ */
+#include <linux/module.h>
+#include <linux/kprobes.h>
+#include <linux/perfmon_kern.h>
+#include <asm/msr.h>
+
+MODULE_AUTHOR("Stephane Eranian <eranian@gmail.com>");
+MODULE_DESCRIPTION("Intel Atom");
+MODULE_LICENSE("GPL");
+
+static int force, force_nmi;
+MODULE_PARM_DESC(force, "bool: force module to load succesfully");
+MODULE_PARM_DESC(force_nmi, "bool: force use of NMI for PMU interrupt");
+module_param(force, bool, 0600);
+module_param(force_nmi, bool, 0600);
+
+/*
+ * - upper 32 bits are reserved
+ * - INT: APIC enable bit is reserved (forced to 1)
+ *
+ * RSVD: reserved bits are 1
+ */
+#define PFM_ATOM_PMC_RSVD      ((~((1ULL<<32)-1)) | (1ULL<<20))
+
+/*
+ * force Local APIC interrupt on overflow
+ * disable with NO_EMUL64
+ */
+#define PFM_ATOM_PMC_VAL       (1ULL<<20)
+#define PFM_ATOM_NO64  (1ULL<<20)
+
+/*
+ * Atom counters are 40-bits. 40-bits can be read but ony 31 can be written
+ * to due to a limitation of wrmsr. Bits [[63-32] are sign extensions of bit 31.
+ * Bits [63-40] must not be set
+ *
+ * See IA-32 Intel Architecture Software developer manual Vol 3B chapter 18
+ */
+#define PFM_ATOM_PMD_WIDTH     31
+#define PFM_ATOM_PMD_RSVD      ~((1ULL << 40)-1)
+
+static void pfm_intel_atom_acquire_pmu_percpu(void);
+static void pfm_intel_atom_release_pmu_percpu(void);
+static void pfm_intel_atom_restore_pmcs(struct pfm_context *ctx,
+                                       struct pfm_event_set *set);
+static int pfm_intel_atom_stop_save(struct pfm_context *ctx,
+                                   struct pfm_event_set *set);
+static int pfm_intel_atom_has_ovfls(struct pfm_context *ctx);
+static void __kprobes pfm_intel_atom_quiesce(void);
+
+struct pfm_arch_pmu_info pfm_intel_atom_pmu_info = {
+       .stop_save = pfm_intel_atom_stop_save,
+       .has_ovfls = pfm_intel_atom_has_ovfls,
+       .quiesce = pfm_intel_atom_quiesce,
+       .restore_pmcs = pfm_intel_atom_restore_pmcs,
+       .acquire_pmu_percpu = pfm_intel_atom_acquire_pmu_percpu,
+       .release_pmu_percpu = pfm_intel_atom_release_pmu_percpu
+
+};
+
+#define PFM_ATOM_C(n) {                  \
+       .type = PFM_REG_I64,             \
+       .desc = "PERFEVTSEL"#n,          \
+       .dfl_val = PFM_ATOM_PMC_VAL,       \
+       .rsvd_msk = PFM_ATOM_PMC_RSVD,     \
+       .no_emul64_msk = PFM_ATOM_NO64,    \
+       .hw_addr = MSR_P6_EVNTSEL0 + (n) \
+       }
+
+
+static struct pfm_regmap_desc pfm_intel_atom_pmc_desc[] = {
+/* pmc0  */ PFM_ATOM_C(0),
+/* pmc1  */ PFM_ATOM_C(1),
+/* pmc2  */ PMX_NA, PMX_NA,
+/* pmc4  */ PMX_NA, PMX_NA, PMX_NA, PMX_NA,
+/* pmc8  */ PMX_NA, PMX_NA, PMX_NA, PMX_NA,
+/* pmc12 */ PMX_NA, PMX_NA, PMX_NA, PMX_NA,
+/* pmc16 */ { .type = PFM_REG_I,
+             .desc = "FIXED_CTRL",
+             .dfl_val = 0x0000000000000888ULL,  /* force PMI */
+             .rsvd_msk = 0xfffffffffffffcccULL, /* 3 fixed counters defined */
+             .no_emul64_msk = 0,
+             .hw_addr = MSR_CORE_PERF_FIXED_CTR_CTRL
+           },
+/* pmc17  */{ .type = PFM_REG_W,
+             .desc = "PEBS_ENABLE",
+             .dfl_val = 0,
+             .rsvd_msk = 0xfffffffffffffffeULL,
+             .no_emul64_msk = 0,
+             .hw_addr = MSR_IA32_PEBS_ENABLE
+           }
+};
+#define PFM_ATOM_MAX_PMCS ARRAY_SIZE(pfm_intel_atom_pmc_desc)
+
+#define PFM_ATOM_D(n) \
+       { .type = PFM_REG_C,                    \
+         .desc = "PMC"#n,                      \
+         .rsvd_msk = PFM_ATOM_PMD_RSVD,        \
+         .hw_addr = MSR_P6_PERFCTR0+n,         \
+         .dep_pmcs[0] = 1ULL << n              \
+       }
+
+#define PFM_ATOM_FD(n) \
+       { .type = PFM_REG_C,                    \
+         .desc = "FIXED_CTR"#n,                \
+         .rsvd_msk = PFM_ATOM_PMD_RSVD,        \
+         .hw_addr = MSR_CORE_PERF_FIXED_CTR0+n,\
+         .dep_pmcs[0] = 1ULL << 16             \
+       }
+
+static struct pfm_regmap_desc pfm_intel_atom_pmd_desc[] = {
+/* pmd0  */ PFM_ATOM_D(0),
+/* pmd1  */ PFM_ATOM_D(1),
+/* pmd2  */ PMX_NA,
+/* pmd3  */ PMX_NA,
+/* pmd4  */ PMX_NA, PMX_NA, PMX_NA, PMX_NA,
+/* pmd8  */ PMX_NA, PMX_NA, PMX_NA, PMX_NA,
+/* pmd12 */ PMX_NA, PMX_NA, PMX_NA, PMX_NA,
+/* pmd16 */ PFM_ATOM_FD(0),
+/* pmd17 */ PFM_ATOM_FD(1),
+/* pmd18 */ PFM_ATOM_FD(2)
+};
+#define PFM_ATOM_MAX_PMDS ARRAY_SIZE(pfm_intel_atom_pmd_desc)
+
+static struct pfm_pmu_config pfm_intel_atom_pmu_conf;
+
+static int pfm_intel_atom_probe_pmu(void)
+{
+       if (force)
+               goto doit;
+
+       if (current_cpu_data.x86_vendor != X86_VENDOR_INTEL)
+               return -1;
+
+       if (current_cpu_data.x86 != 6)
+               return -1;
+
+       if (current_cpu_data.x86_model != 28)
+               return -1;
+doit:
+       /*
+        * having APIC is mandatory, so disregard force option
+        */
+       if (!cpu_has_apic) {
+               PFM_INFO("no Local APIC, try rebooting with lapic option");
+               return -1;
+       }
+
+       PFM_INFO("detected Intel Atom PMU");
+
+       if (force_nmi)
+               pfm_intel_atom_pmu_info.flags |= PFM_X86_FL_USE_NMI;
+
+       return 0;
+}
+
+/**
+ * pfm_intel_atom_has_ovfls - check for pending overflow condition
+ * @ctx: context to work on
+ *
+ * detect if counters have overflowed.
+ * return:
+ *     0 : no overflow
+ *     1 : at least one overflow
+ */
+static int __kprobes pfm_intel_atom_has_ovfls(struct pfm_context *ctx)
+{
+       struct pfm_regmap_desc *d;
+       u64 ovf;
+
+       d = pfm_pmu_conf->pmd_desc;
+       /*
+        * read global overflow status register
+        * if sharing PMU, then not all bit are ours so must
+        * check only the ones we actually use
+        */
+       rdmsrl(MSR_CORE_PERF_GLOBAL_STATUS, ovf);
+
+       /*
+        * for pmd0, we also check PEBS overflow on bit 62
+        */
+       if ((d[0].type & PFM_REG_I) && (ovf & ((1ull << 62) | 1ull)))
+               return 1;
+
+       if ((d[1].type & PFM_REG_I) && (ovf & 2ull))
+               return 1;
+
+       if ((d[16].type & PFM_REG_I) && (ovf & (1ull << 32)))
+               return 1;
+
+       if ((d[17].type & PFM_REG_I) && (ovf & (2ull << 32)))
+               return 1;
+
+       if ((d[18].type & PFM_REG_I) && (ovf & (4ull << 32)))
+               return 1;
+
+       return 0;
+}
+
+/**
+ * pfm_intel_atom_stop_save - stop monitoring, collect pending overflow, save pmds
+ * @ctx: context to work on
+ * @set: active set
+ *
+ * return:
+ *     1: caller needs to save pmds
+ *     0: caller does not need to save pmds, they have been saved by this call
+ */
+static int pfm_intel_atom_stop_save(struct pfm_context *ctx,
+                                   struct pfm_event_set *set)
+{
+#define PFM_ATOM_WMASK (1ULL << 31)
+#define PFM_ATOM_OMASK ((1ULL << 31)-1)
+       u64 clear_ovf = 0;
+       u64 ovf, ovf2, val;
+
+       /*
+        * read global overflow status register
+        * if sharing PMU, then not all bit are ours so must
+        * check only the ones we actually use.
+        *
+        * XXX: Atom seems to have a bug with the stickyness of
+        *      GLOBAL_STATUS. If we read GLOBAL_STATUS after we
+        *      clear the generic counters, then their bits in
+        *      GLOBAL_STATUS are cleared. This should not be the
+        *      case accoding to architected PMU. To workaround
+        *      the problem, we read GLOBAL_STATUS BEFORE we stop
+        *      all monitoring.
+        */
+       rdmsrl(MSR_CORE_PERF_GLOBAL_STATUS, ovf);
+
+       /*
+        * stop monitoring
+        */
+       if (test_bit(0, cast_ulp(set->used_pmcs)))
+               wrmsrl(MSR_P6_EVNTSEL0, 0);
+
+       if (test_bit(1, cast_ulp(set->used_pmcs)))
+               wrmsrl(MSR_P6_EVNTSEL1, 0);
+
+       if (test_bit(16, cast_ulp(set->used_pmcs)))
+               wrmsrl(MSR_CORE_PERF_FIXED_CTR_CTRL, 0);
+
+       if (test_bit(17, cast_ulp(set->used_pmcs)))
+               wrmsrl(MSR_IA32_PEBS_ENABLE, 0);
+
+       /*
+        * XXX: related to bug mentioned above
+        *
+        * read GLOBAL_STATUS again to avoid race condition
+        * with overflows happening after first read and
+        * before stop. That avoids missing overflows on
+        * the fixed counters and PEBS
+        */
+       rdmsrl(MSR_CORE_PERF_GLOBAL_STATUS, ovf2);
+       ovf |= ovf2;
+
+       /*
+        * if we already have a pending overflow condition, we simply
+        * return to take care of it first.
+        */
+       if (set->npend_ovfls)
+               return 1;
+
+       /*
+        * check PMD 0,1,16,17,18 for overflow and save their value
+        */
+       if (test_bit(0, cast_ulp(set->used_pmds))) {
+               rdmsrl(MSR_P6_PERFCTR0, val);
+               if (ovf & ((1ull<<62)|1ull)) {
+                       __set_bit(0, cast_ulp(set->povfl_pmds));
+                       set->npend_ovfls++;
+                       clear_ovf = (1ull << 62) | 1ull;
+               }
+               val = (set->pmds[0].value & ~PFM_ATOM_OMASK)
+                   | (val & PFM_ATOM_OMASK);
+               set->pmds[0].value = val;
+       }
+
+       if (test_bit(1, cast_ulp(set->used_pmds))) {
+               rdmsrl(MSR_P6_PERFCTR1, val);
+                if (ovf & 2ull) {
+                       __set_bit(1, cast_ulp(set->povfl_pmds));
+                       set->npend_ovfls++;
+                       clear_ovf |= 2ull;
+               }
+               val = (set->pmds[1].value & ~PFM_ATOM_OMASK)
+                   | (val & PFM_ATOM_OMASK);
+               set->pmds[1].value = val;
+       }
+
+       if (test_bit(16, cast_ulp(set->used_pmds))) {
+               rdmsrl(MSR_CORE_PERF_FIXED_CTR0, val);
+               if (ovf & (1ull << 32)) {
+                       __set_bit(16, cast_ulp(set->povfl_pmds));
+                       set->npend_ovfls++;
+                       clear_ovf |= 1ull << 32;
+               }
+               val = (set->pmds[16].value & ~PFM_ATOM_OMASK)
+                   | (val & PFM_ATOM_OMASK);
+               set->pmds[16].value = val;
+       }
+
+       if (test_bit(17, cast_ulp(set->used_pmds))) {
+               rdmsrl(MSR_CORE_PERF_FIXED_CTR0+1, val);
+               if (ovf & (2ull << 32)) {
+                       __set_bit(17, cast_ulp(set->povfl_pmds));
+                       set->npend_ovfls++;
+                       clear_ovf |= 2ull << 32;
+               }
+               val = (set->pmds[17].value & ~PFM_ATOM_OMASK)
+                   | (val & PFM_ATOM_OMASK);
+               set->pmds[17].value = val;
+       }
+
+       if (test_bit(18, cast_ulp(set->used_pmds))) {
+               rdmsrl(MSR_CORE_PERF_FIXED_CTR0+2, val);
+               if (ovf & (4ull << 32)) {
+                       __set_bit(18, cast_ulp(set->povfl_pmds));
+                       set->npend_ovfls++;
+                       clear_ovf |= 4ull << 32;
+               }
+               val = (set->pmds[18].value & ~PFM_ATOM_OMASK)
+                   | (val & PFM_ATOM_OMASK);
+               set->pmds[18].value = val;
+       }
+
+       if (clear_ovf)
+               wrmsrl(MSR_CORE_PERF_GLOBAL_OVF_CTRL, clear_ovf);
+
+       /* 0 means: no need to save PMDs at upper level */
+       return 0;
+}
+
+/**
+ * pfm_intel_atom_quiesce - stop monitoring without grabbing any lock
+ *
+ * called from NMI interrupt handler to immediately stop monitoring
+ * cannot grab any lock, including perfmon related locks
+ */
+static void __kprobes pfm_intel_atom_quiesce(void)
+{
+       /*
+        * quiesce PMU by clearing available registers that have
+        * the start/stop capability
+        */
+       if (test_bit(0, cast_ulp(pfm_pmu_conf->regs_all.pmcs)))
+               wrmsrl(MSR_P6_EVNTSEL0, 0);
+
+       if (test_bit(1, cast_ulp(pfm_pmu_conf->regs_all.pmcs)))
+               wrmsrl(MSR_P6_EVNTSEL1, 0);
+
+       if (test_bit(16, cast_ulp(pfm_pmu_conf->regs_all.pmcs)))
+               wrmsrl(MSR_CORE_PERF_FIXED_CTR_CTRL, 0);
+
+       if (test_bit(17, cast_ulp(pfm_pmu_conf->regs_all.pmcs)))
+               wrmsrl(MSR_IA32_PEBS_ENABLE, 0);
+}
+
+/**
+ * pfm_intel_atom_restore_pmcs - reload PMC registers
+ * @ctx: context to restore from
+ * @set: current event set
+ *
+ * restores pmcs and also PEBS Data Save area pointer
+ */
+static void pfm_intel_atom_restore_pmcs(struct pfm_context *ctx,
+                                       struct pfm_event_set *set)
+{
+       struct pfm_arch_context *ctx_arch;
+       u64 clear_ovf = 0;
+
+       ctx_arch = pfm_ctx_arch(ctx);
+       /*
+        * must restore DS pointer before restoring PMCs
+        * as this can potentially reactivate monitoring
+        */
+       if (ctx_arch->flags.use_ds)
+               wrmsrl(MSR_IA32_DS_AREA, (unsigned long)ctx_arch->ds_area);
+
+       if (test_bit(0, cast_ulp(set->used_pmcs))) {
+               wrmsrl(MSR_P6_EVNTSEL0, set->pmcs[0]);
+               clear_ovf = 1ull;
+       }
+
+       if (test_bit(1, cast_ulp(set->used_pmcs))) {
+               wrmsrl(MSR_P6_EVNTSEL1, set->pmcs[1]);
+               clear_ovf |= 2ull;
+       }
+
+       if (test_bit(16, cast_ulp(set->used_pmcs))) {
+               wrmsrl(MSR_CORE_PERF_FIXED_CTR_CTRL, set->pmcs[16]);
+               clear_ovf |= 7ull << 32;
+       }
+
+       if (test_bit(17, cast_ulp(set->used_pmcs))) {
+               wrmsrl(MSR_IA32_PEBS_ENABLE, set->pmcs[17]);
+               clear_ovf |= 1ull << 62;
+       }
+
+       if (clear_ovf)
+               wrmsrl(MSR_CORE_PERF_GLOBAL_OVF_CTRL, clear_ovf);
+}
+
+static int pfm_intel_atom_pmc17_check(struct pfm_context *ctx,
+                                     struct pfm_event_set *set,
+                                     struct pfarg_pmc *req)
+{
+       struct pfm_arch_context *ctx_arch;
+       ctx_arch = pfm_ctx_arch(ctx);
+
+       /*
+        * if user activates PEBS_ENABLE, then we need to have a valid
+        * DS Area setup. This only happens when the PEBS sampling format is
+        * used in which case PFM_X86_USE_PEBS is set. We must reject all other
+        * requests.
+        *
+        * Otherwise we may pickup stale MSR_IA32_DS_AREA values. It appears
+        * that a value of 0 for this MSR does crash the system with
+        * PEBS_ENABLE=1.
+        */
+       if (!ctx_arch->flags.use_pebs && req->reg_value) {
+               PFM_DBG("pmc17 useable only with a PEBS sampling format");
+               return -EINVAL;
+       }
+       return 0;
+}
+
+DEFINE_PER_CPU(u64, saved_global_ctrl);
+
+/**
+ * pfm_intel_atom_acquire_pmu_percpu - acquire PMU resource per CPU
+ *
+ * For Atom, it is necessary to enable all available
+ * registers. The firmware rightfully has the fixed counters
+ * disabled for backward compatibility with architectural perfmon
+ * v1
+ *
+ * This function is invoked on each online CPU
+ */
+static void pfm_intel_atom_acquire_pmu_percpu(void)
+{
+       struct pfm_regmap_desc *d;
+       u64 mask = 0;
+       unsigned int i;
+
+       /*
+        * build bitmask of registers that are available to
+        * us. In some cases, there may be fewer registers than
+        * what Atom supports due to sharing with other kernel
+        * subsystems, such as NMI
+        */
+       d = pfm_pmu_conf->pmd_desc;
+       for (i=0; i < 16; i++) {
+               if ((d[i].type & PFM_REG_I) == 0)
+                       continue;
+               mask |= 1ull << i;
+       }
+       for (i=16; i < PFM_ATOM_MAX_PMDS; i++) {
+               if ((d[i].type & PFM_REG_I) == 0)
+                       continue;
+               mask |= 1ull << (32+i-16);
+       }
+
+       /*
+        * keep a local copy of the current MSR_CORE_PERF_GLOBAL_CTRL
+        */
+       rdmsrl(MSR_CORE_PERF_GLOBAL_CTRL, __get_cpu_var(saved_global_ctrl));
+
+       PFM_DBG("global=0x%llx set to 0x%llx",
+               __get_cpu_var(saved_global_ctrl),
+               mask);
+
+       /*
+        * enable all registers
+        *
+        * No need to quiesce PMU. If there is a overflow, it will be
+        * treated as spurious by the handler
+        */
+       wrmsrl(MSR_CORE_PERF_GLOBAL_CTRL, mask);
+}
+
+/**
+ * pfm_intel_atom_release_pmu_percpu - release PMU resource per CPU
+ *
+ * For Atom, we restore MSR_CORE_PERF_GLOBAL_CTRL to its orginal value
+ */
+static void pfm_intel_atom_release_pmu_percpu(void)
+{
+       PFM_DBG("global_ctrl restored to 0x%llx\n",
+                       __get_cpu_var(saved_global_ctrl));
+
+       wrmsrl(MSR_CORE_PERF_GLOBAL_CTRL, __get_cpu_var(saved_global_ctrl));
+}
+
+static struct pfm_pmu_config pfm_intel_atom_pmu_conf = {
+       .pmu_name = "Intel Atom",
+       .pmd_desc = pfm_intel_atom_pmd_desc,
+       .counter_width   = PFM_ATOM_PMD_WIDTH,
+       .num_pmc_entries = PFM_ATOM_MAX_PMCS,
+       .num_pmd_entries = PFM_ATOM_MAX_PMDS,
+       .pmc_desc = pfm_intel_atom_pmc_desc,
+       .probe_pmu = pfm_intel_atom_probe_pmu,
+       .version = "1.0",
+       .flags = PFM_PMU_BUILTIN_FLAG,
+       .owner = THIS_MODULE,
+       .pmc_write_check = pfm_intel_atom_pmc17_check,
+       .pmu_info = &pfm_intel_atom_pmu_info
+};
+
+static int __init pfm_intel_atom_pmu_init_module(void)
+{
+       return pfm_pmu_register(&pfm_intel_atom_pmu_conf);
+}
+
+static void __exit pfm_intel_atom_pmu_cleanup_module(void)
+{
+       pfm_pmu_unregister(&pfm_intel_atom_pmu_conf);
+}
+
+module_init(pfm_intel_atom_pmu_init_module);
+module_exit(pfm_intel_atom_pmu_cleanup_module);
diff --git a/arch/x86/perfmon/perfmon_intel_core.c b/arch/x86/perfmon/perfmon_intel_core.c
new file mode 100644
index 0000000..fddc436
--- /dev/null
+++ b/arch/x86/perfmon/perfmon_intel_core.c
@@ -0,0 +1,449 @@
+/*
+ * This file contains the Intel Core PMU registers description tables.
+ * Intel Core-based processors support architectural perfmon v2 + PEBS
+ *
+ * Copyright (c) 2006-2007 Hewlett-Packard Development Company, L.P.
+ * Contributed by Stephane Eranian <eranian@hpl.hp.com>
+ */
+#include <linux/module.h>
+#include <linux/kprobes.h>
+#include <linux/perfmon_kern.h>
+#include <linux/nmi.h>
+
+MODULE_AUTHOR("Stephane Eranian <eranian@hpl.hp.com>");
+MODULE_DESCRIPTION("Intel Core");
+MODULE_LICENSE("GPL");
+
+static int force_nmi;
+MODULE_PARM_DESC(force_nmi, "bool: force use of NMI for PMU interrupt");
+module_param(force_nmi, bool, 0600);
+
+/*
+ * - upper 32 bits are reserved
+ * - INT: APIC enable bit is reserved (forced to 1)
+ * - bit 21 is reserved
+ *
+ *   RSVD: reserved bits must be 1
+ */
+#define PFM_CORE_PMC_RSVD ((~((1ULL<<32)-1)) \
+                       | (1ULL<<20)   \
+                       | (1ULL<<21))
+
+/*
+ * Core counters are 40-bits
+ */
+#define PFM_CORE_CTR_RSVD      (~((1ULL<<40)-1))
+
+/*
+ * force Local APIC interrupt on overflow
+ * disable with NO_EMUL64
+ */
+#define PFM_CORE_PMC_VAL       (1ULL<<20)
+#define PFM_CORE_NO64          (1ULL<<20)
+
+#define PFM_CORE_NA { .reg_type = PFM_REGT_NA}
+
+#define PFM_CORE_CA(m, c, t) \
+       { \
+         .addrs[0] = m, \
+         .ctr = c, \
+         .reg_type = t \
+       }
+
+struct pfm_ds_area_intel_core {
+       u64     bts_buf_base;
+       u64     bts_index;
+       u64     bts_abs_max;
+       u64     bts_intr_thres;
+       u64     pebs_buf_base;
+       u64     pebs_index;
+       u64     pebs_abs_max;
+       u64     pebs_intr_thres;
+       u64     pebs_cnt_reset;
+};
+
+static void pfm_core_restore_pmcs(struct pfm_context *ctx,
+                                 struct pfm_event_set *set);
+static int pfm_core_has_ovfls(struct pfm_context *ctx);
+static int pfm_core_stop_save(struct pfm_context *ctx,
+                             struct pfm_event_set *set);
+static void __kprobes pfm_core_quiesce(void);
+
+static u64 enable_mask[PFM_MAX_PMCS];
+static u16 max_enable;
+
+struct pfm_arch_pmu_info pfm_core_pmu_info = {
+       .stop_save = pfm_core_stop_save,
+       .has_ovfls = pfm_core_has_ovfls,
+       .quiesce = pfm_core_quiesce,
+       .restore_pmcs = pfm_core_restore_pmcs
+};
+
+static struct pfm_regmap_desc pfm_core_pmc_desc[] = {
+/* pmc0  */ {
+             .type = PFM_REG_I64,
+             .desc = "PERFEVTSEL0",
+             .dfl_val = PFM_CORE_PMC_VAL,
+             .rsvd_msk = PFM_CORE_PMC_RSVD,
+             .no_emul64_msk = PFM_CORE_NO64,
+             .hw_addr = MSR_P6_EVNTSEL0
+           },
+/* pmc1  */ {
+             .type = PFM_REG_I64,
+             .desc = "PERFEVTSEL1",
+             .dfl_val = PFM_CORE_PMC_VAL,
+             .rsvd_msk = PFM_CORE_PMC_RSVD,
+             .no_emul64_msk = PFM_CORE_NO64,
+             .hw_addr = MSR_P6_EVNTSEL1
+           },
+/* pmc2  */ PMX_NA, PMX_NA,
+/* pmc4  */ PMX_NA, PMX_NA, PMX_NA, PMX_NA,
+/* pmc8  */ PMX_NA, PMX_NA, PMX_NA, PMX_NA,
+/* pmc12 */ PMX_NA, PMX_NA, PMX_NA, PMX_NA,
+/* pmc16 */ { .type = PFM_REG_I,
+             .desc = "FIXED_CTRL",
+             .dfl_val = 0x888ULL,
+             .rsvd_msk = 0xfffffffffffffcccULL,
+             .no_emul64_msk = 0,
+             .hw_addr = MSR_CORE_PERF_FIXED_CTR_CTRL
+           },
+/* pmc17  */ { .type = PFM_REG_W,
+             .desc = "PEBS_ENABLE",
+             .dfl_val = 0,
+             .rsvd_msk = 0xfffffffffffffffeULL,
+             .no_emul64_msk = 0,
+             .hw_addr = MSR_IA32_PEBS_ENABLE
+           }
+};
+
+#define PFM_CORE_D(n) \
+       { .type = PFM_REG_C,                    \
+         .desc = "PMC"#n,                      \
+         .rsvd_msk =  PFM_CORE_CTR_RSVD,       \
+         .hw_addr = MSR_P6_PERFCTR0+n,         \
+         .dep_pmcs[0] = 1ULL << n              \
+       }
+
+#define PFM_CORE_FD(n) \
+       { .type = PFM_REG_C,                    \
+         .desc = "FIXED_CTR"#n,                \
+         .rsvd_msk =  PFM_CORE_CTR_RSVD,       \
+         .hw_addr = MSR_CORE_PERF_FIXED_CTR0+n,\
+         .dep_pmcs[0] = 1ULL << 16             \
+       }
+
+static struct pfm_regmap_desc pfm_core_pmd_desc[] = {
+/* pmd0  */ PFM_CORE_D(0),
+/* pmd1  */ PFM_CORE_D(1),
+/* pmd2  */ PMX_NA, PMX_NA,
+/* pmd4  */ PMX_NA, PMX_NA, PMX_NA, PMX_NA,
+/* pmd8  */ PMX_NA, PMX_NA, PMX_NA, PMX_NA,
+/* pmd12 */ PMX_NA, PMX_NA, PMX_NA, PMX_NA,
+/* pmd16 */ PFM_CORE_FD(0),
+/* pmd17 */ PFM_CORE_FD(1),
+/* pmd18 */ PFM_CORE_FD(2)
+};
+#define PFM_CORE_NUM_PMCS      ARRAY_SIZE(pfm_core_pmc_desc)
+#define PFM_CORE_NUM_PMDS      ARRAY_SIZE(pfm_core_pmd_desc)
+
+static struct pfm_pmu_config pfm_core_pmu_conf;
+
+static int pfm_core_probe_pmu(void)
+{
+       /*
+        * Check for Intel Core processor explicitely
+        * Checking for cpu_has_perfmon is not enough as this
+        * matches intel Core Duo/Core Solo but none supports
+        * PEBS.
+        *
+        * Intel Core = arch perfmon v2 + PEBS
+        */
+       if (current_cpu_data.x86_vendor != X86_VENDOR_INTEL) {
+               PFM_INFO("not an AMD processor");
+               return -1;
+       }
+
+       if (current_cpu_data.x86 != 6)
+               return -1;
+
+       switch (current_cpu_data.x86_model) {
+       case 15: /* Merom */
+               break;
+       case 23: /* Penryn */
+               break;
+       case 29: /* Dunnington */
+               break;
+       default:
+               return -1;
+       }
+
+       if (!cpu_has_apic) {
+               PFM_INFO("no Local APIC, unsupported");
+               return -1;
+       }
+
+       PFM_INFO("nmi_watchdog=%d nmi_active=%d force_nmi=%d",
+               nmi_watchdog, atomic_read(&nmi_active), force_nmi);
+
+       /*
+        * Intel Core processors implement DS and PEBS, no need to check
+        */
+       if (cpu_has_pebs)
+               PFM_INFO("PEBS supported, enabled");
+
+       /*
+        * initialize bitmask of register with enable capability, i.e.,
+        * startstop. This is used to restrict the number of registers to
+        * touch on start/stop
+        * max_enable: number of bits to scan in enable_mask = highest + 1
+        *
+        * may be adjusted in pfm_arch_pmu_acquire()
+        */
+       __set_bit(0, cast_ulp(enable_mask));
+       __set_bit(1, cast_ulp(enable_mask));
+       __set_bit(16, cast_ulp(enable_mask));
+       __set_bit(17, cast_ulp(enable_mask));
+       max_enable = 17+1;
+
+       if (force_nmi)
+               pfm_core_pmu_info.flags |= PFM_X86_FL_USE_NMI;
+
+       return 0;
+}
+
+static int pfm_core_pmc17_check(struct pfm_context *ctx,
+                            struct pfm_event_set *set,
+                            struct pfarg_pmc *req)
+{
+       struct pfm_arch_context *ctx_arch;
+       ctx_arch = pfm_ctx_arch(ctx);
+
+       /*
+        * if user activates PEBS_ENABLE, then we need to have a valid
+        * DS Area setup. This only happens when the PEBS sampling format is
+        * used in which case PFM_X86_USE_PEBS is set. We must reject all other
+        * requests.
+        *
+        * Otherwise we may pickup stale MSR_IA32_DS_AREA values. It appears
+        * that a value of 0 for this MSR does crash the system with
+        * PEBS_ENABLE=1.
+        */
+       if (!ctx_arch->flags.use_pebs && req->reg_value) {
+               PFM_DBG("pmc17 useable only with a PEBS sampling format");
+               return -EINVAL;
+       }
+       return 0;
+}
+
+/*
+ * detect is counters have overflowed.
+ * return:
+ *     0 : no overflow
+ *     1 : at least one overflow
+ *
+ * used by Intel Core-based processors
+ */
+static int __kprobes pfm_core_has_ovfls(struct pfm_context *ctx)
+{
+       struct pfm_arch_pmu_info *pmu_info;
+       u64 *cnt_mask;
+       u64 wmask, val;
+       u16 i, num;
+
+       pmu_info = &pfm_core_pmu_info;
+       cnt_mask = ctx->regs.cnt_pmds;
+       num = ctx->regs.num_counters;
+       wmask = 1ULL << pfm_pmu_conf->counter_width;
+
+       for (i = 0; num; i++) {
+               if (test_bit(i, cast_ulp(cnt_mask))) {
+                       rdmsrl(pfm_core_pmd_desc[i].hw_addr, val);
+                       if (!(val & wmask))
+                               return 1;
+                       num--;
+               }
+       }
+       return 0;
+}
+
+static int pfm_core_stop_save(struct pfm_context *ctx,
+                             struct pfm_event_set *set)
+{
+       struct pfm_arch_context *ctx_arch;
+       struct pfm_ds_area_intel_core *ds = NULL;
+       u64 used_mask[PFM_PMC_BV];
+       u64 *cnt_mask;
+       u64 val, wmask, ovfl_mask;
+       u16 count, has_ovfl;
+       u16 i, pebs_idx = ~0;
+
+       ctx_arch = pfm_ctx_arch(ctx);
+
+       wmask = 1ULL << pfm_pmu_conf->counter_width;
+
+       /*
+        * used enable pmc bitmask
+        */
+       bitmap_and(cast_ulp(used_mask),
+                       cast_ulp(set->used_pmcs),
+                       cast_ulp(enable_mask),
+                       max_enable);
+
+       count = bitmap_weight(cast_ulp(used_mask), max_enable);
+       /*
+        * stop monitoring
+        * Unfortunately, this is very expensive!
+        * wrmsrl() is serializing.
+        */
+       for (i = 0; count; i++) {
+               if (test_bit(i, cast_ulp(used_mask))) {
+                       wrmsrl(pfm_pmu_conf->pmc_desc[i].hw_addr, 0);
+                       count--;
+               }
+       }
+       /*
+        * if we already having a pending overflow condition, we simply
+        * return to take care of this first.
+        */
+       if (set->npend_ovfls)
+               return 1;
+
+       ovfl_mask = pfm_pmu_conf->ovfl_mask;
+       cnt_mask = ctx->regs.cnt_pmds;
+
+       if (ctx_arch->flags.use_pebs) {
+               ds = ctx_arch->ds_area;
+               pebs_idx = 0; /* PMC0/PMD0 */
+               PFM_DBG("ds=%p pebs_idx=0x%llx thres=0x%llx",
+                       ds,
+                       (unsigned long long)ds->pebs_index,
+                       (unsigned long long)ds->pebs_intr_thres);
+       }
+
+       /*
+        * Check for pending overflows and save PMDs (combo)
+        * We employ used_pmds and not intr_pmds because we must
+        * also saved on PMD registers.
+        * Must check for counting PMDs because of virtual PMDs
+        *
+        * XXX: should use the ovf_status register instead, yet
+        *      we would have to check if NMI is used and fallback
+        *      to individual pmd inspection.
+        */
+       count = set->nused_pmds;
+
+       for (i = 0; count; i++) {
+               if (test_bit(i, cast_ulp(set->used_pmds))) {
+                       val = pfm_arch_read_pmd(ctx, i);
+                       if (likely(test_bit(i, cast_ulp(cnt_mask)))) {
+                               if (i == pebs_idx)
+                                       has_ovfl = (ds->pebs_index >=
+                                                   ds->pebs_intr_thres);
+                               else
+                                       has_ovfl = !(val & wmask);
+                               if (has_ovfl) {
+                                       __set_bit(i, cast_ulp(set->povfl_pmds));
+                                       set->npend_ovfls++;
+                               }
+                               val = (set->pmds[i].value & ~ovfl_mask)
+                                   | (val & ovfl_mask);
+                       }
+                       set->pmds[i].value = val;
+                       count--;
+               }
+       }
+       /* 0 means: no need to save PMDs at upper level */
+       return 0;
+}
+
+/**
+ * pfm_core_quiesce - stop monitoring without grabbing any lock
+ *
+ * called from NMI interrupt handler to immediately stop monitoring
+ * cannot grab any lock, including perfmon related locks
+ */
+static void __kprobes pfm_core_quiesce(void)
+{
+       /*
+        * quiesce PMU by clearing available registers that have
+        * the start/stop capability
+        */
+       if (test_bit(0, cast_ulp(pfm_pmu_conf->regs_all.pmcs)))
+               wrmsrl(MSR_P6_EVNTSEL0, 0);
+       if (test_bit(1, cast_ulp(pfm_pmu_conf->regs_all.pmcs)))
+               wrmsrl(MSR_P6_EVNTSEL1, 0);
+       if (test_bit(16, cast_ulp(pfm_pmu_conf->regs_all.pmcs)))
+               wrmsrl(MSR_CORE_PERF_FIXED_CTR_CTRL, 0);
+       if (test_bit(17, cast_ulp(pfm_pmu_conf->regs_all.pmcs)))
+               wrmsrl(MSR_IA32_PEBS_ENABLE, 0);
+}
+/**
+ * pfm_core_restore_pmcs - reload PMC registers
+ * @ctx: context to restore from
+ * @set: current event set
+ *
+ * optimized version of pfm_arch_restore_pmcs(). On Core, we can
+ * afford to only restore the pmcs registers we use, because they are
+ * all independent from each other.
+ */
+static void pfm_core_restore_pmcs(struct pfm_context *ctx,
+                                 struct pfm_event_set *set)
+{
+       struct pfm_arch_context *ctx_arch;
+       u64 *mask;
+       u16 i, num;
+
+       ctx_arch = pfm_ctx_arch(ctx);
+
+       /*
+        * must restore DS pointer before restoring PMCs
+        * as this can potentially reactivate monitoring
+        */
+       if (ctx_arch->flags.use_ds)
+               wrmsrl(MSR_IA32_DS_AREA, (unsigned long)ctx_arch->ds_area);
+
+       mask = set->used_pmcs;
+       num = set->nused_pmcs;
+       for (i = 0; num; i++) {
+               if (test_bit(i, cast_ulp(mask))) {
+                       wrmsrl(pfm_pmu_conf->pmc_desc[i].hw_addr, set->pmcs[i]);
+                       num--;
+               }
+       }
+}
+
+/*
+ * Counters may have model-specific width which can be probed using
+ * the CPUID.0xa leaf. Yet, the documentation says: "
+ * In the initial implementation, only the read bit width is reported
+ * by CPUID, write operations are limited to the low 32 bits.
+ * Bits [w-32] are sign extensions of bit 31. As such the effective width
+ * of a counter is 31 bits only.
+ */
+static struct pfm_pmu_config pfm_core_pmu_conf = {
+       .pmu_name = "Intel Core",
+       .pmd_desc = pfm_core_pmd_desc,
+       .counter_width = 31,
+       .num_pmc_entries = PFM_CORE_NUM_PMCS,
+       .num_pmd_entries = PFM_CORE_NUM_PMDS,
+       .pmc_desc = pfm_core_pmc_desc,
+       .probe_pmu = pfm_core_probe_pmu,
+       .version = "1.2",
+       .flags = PFM_PMU_BUILTIN_FLAG,
+       .owner = THIS_MODULE,
+       .pmu_info = &pfm_core_pmu_info,
+       .pmc_write_check = pfm_core_pmc17_check
+};
+
+static int __init pfm_core_pmu_init_module(void)
+{
+       return pfm_pmu_register(&pfm_core_pmu_conf);
+}
+
+static void __exit pfm_core_pmu_cleanup_module(void)
+{
+       pfm_pmu_unregister(&pfm_core_pmu_conf);
+}
+
+module_init(pfm_core_pmu_init_module);
+module_exit(pfm_core_pmu_cleanup_module);
diff --git a/arch/x86/perfmon/perfmon_p4.c b/arch/x86/perfmon/perfmon_p4.c
new file mode 100644
index 0000000..1ffcf3c
--- /dev/null
+++ b/arch/x86/perfmon/perfmon_p4.c
@@ -0,0 +1,913 @@
+/*
+ * This file contains the P4/Xeon PMU register description tables
+ * for both 32 and 64 bit modes.
+ *
+ * Copyright (c) 2005 Intel Corporation
+ * Contributed by Bryan Wilkerson <bryan.p.wilkerson@intel.com>
+ *
+ * This program is free software; you can redistribute it and/or
+ * modify it under the terms of version 2 of the GNU General Public
+ * License as published by the Free Software Foundation.
+ *
+ * This program is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+ * General Public License for more details.
+ *
+ * You should have received a copy of the GNU General Public License
+ * along with this program; if not, write to the Free Software
+ * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA
+ * 02111-1307 USA
+  */
+#include <linux/module.h>
+#include <linux/perfmon_kern.h>
+#include <linux/kprobes.h>
+#include <linux/nmi.h>
+#include <asm/msr.h>
+#include <asm/apic.h>
+
+MODULE_AUTHOR("Bryan Wilkerson <bryan.p.wilkerson@intel.com>");
+MODULE_DESCRIPTION("P4/Xeon/EM64T PMU description table");
+MODULE_LICENSE("GPL");
+
+static int force;
+MODULE_PARM_DESC(force, "bool: force module to load succesfully");
+module_param(force, bool, 0600);
+
+static int force_nmi;
+MODULE_PARM_DESC(force_nmi, "bool: force use of NMI for PMU interrupt");
+module_param(force_nmi, bool, 0600);
+
+/*
+ * For extended register information in addition to address that is used
+ * at runtime to figure out the mapping of reg addresses to logical procs
+ * and association of registers to hardware specific features
+ */
+struct pfm_p4_regmap {
+       /*
+        * one each for the logical CPUs.  Index 0 corresponds to T0 and
+        * index 1 corresponds to T1.  Index 1 can be zero if no T1
+        * complement reg exists.
+        */
+       unsigned long addrs[2]; /* 2 = number of threads */
+       unsigned int ctr;       /* for CCCR/PERFEVTSEL, associated counter */
+       unsigned int reg_type;
+};
+
+/*
+ * bitmask for pfm_p4_regmap.reg_type
+ */
+#define PFM_REGT_NA            0x0000  /* not available */
+#define PFM_REGT_EN            0x0001  /* has enable bit (cleared on ctxsw) */
+#define PFM_REGT_ESCR          0x0002  /* P4: ESCR */
+#define PFM_REGT_CCCR          0x0004  /* P4: CCCR */
+#define PFM_REGT_PEBS          0x0010  /* PEBS related */
+#define PFM_REGT_NOHT          0x0020  /* unavailable with HT */
+#define PFM_REGT_CTR           0x0040  /* counter */
+
+/*
+ * architecture specific context extension.
+ * located at: (struct pfm_arch_context *)(ctx+1)
+ */
+struct pfm_arch_p4_context {
+       u32     npend_ovfls;    /* P4 NMI #pending ovfls */
+       u32     reserved;
+       u64     povfl_pmds[PFM_PMD_BV]; /* P4 NMI overflowed counters */
+       u64     saved_cccrs[PFM_MAX_PMCS];
+};
+
+/*
+ * ESCR reserved bitmask:
+ * - bits 31 - 63 reserved
+ * - T1_OS and T1_USR bits are reserved - set depending on logical proc
+ *      user mode application should use T0_OS and T0_USR to indicate
+ * RSVD: reserved bits must be 1
+ */
+#define PFM_ESCR_RSVD  ~0x000000007ffffffcULL
+
+/*
+ * CCCR default value:
+ *     - OVF_PMI_T0=1 (bit 26)
+ *     - OVF_PMI_T1=0 (bit 27) (set if necessary in pfm_write_reg())
+ *     - all other bits are zero
+ *
+ * OVF_PMI is forced to zero if PFM_REGFL_NO_EMUL64 is set on CCCR
+ */
+#define PFM_CCCR_DFL   (1ULL<<26) | (3ULL<<16)
+
+/*
+ * CCCR reserved fields:
+ *     - bits 0-11, 25-29, 31-63
+ *     - OVF_PMI (26-27), override with REGFL_NO_EMUL64
+ *
+ * RSVD: reserved bits must be 1
+ */
+#define PFM_CCCR_RSVD     ~((0xfull<<12)  \
+                       | (0x7full<<18) \
+                       | (0x1ull<<30))
+
+#define PFM_P4_NO64    (3ULL<<26) /* use 3 even in non HT mode */
+
+#define PEBS_PMD       8  /* thread0: IQ_CTR4, thread1: IQ_CTR5 */
+
+/*
+ * With HyperThreading enabled:
+ *
+ *  The ESCRs and CCCRs are divided in half with the top half
+ *  belonging to logical processor 0 and the bottom half going to
+ *  logical processor 1. Thus only half of the PMU resources are
+ *  accessible to applications.
+ *
+ *  PEBS is not available due to the fact that:
+ *     - MSR_PEBS_MATRIX_VERT is shared between the threads
+ *      - IA32_PEBS_ENABLE is shared between the threads
+ *
+ * With HyperThreading disabled:
+ *
+ * The full set of PMU resources is exposed to applications.
+ *
+ * The mapping is chosen such that PMCxx -> MSR is the same
+ * in HT and non HT mode, if register is present in HT mode.
+ *
+ */
+#define PFM_REGT_NHTESCR (PFM_REGT_ESCR|PFM_REGT_NOHT)
+#define PFM_REGT_NHTCCCR (PFM_REGT_CCCR|PFM_REGT_NOHT|PFM_REGT_EN)
+#define PFM_REGT_NHTPEBS (PFM_REGT_PEBS|PFM_REGT_NOHT|PFM_REGT_EN)
+#define PFM_REGT_NHTCTR  (PFM_REGT_CTR|PFM_REGT_NOHT)
+#define PFM_REGT_ENAC    (PFM_REGT_CCCR|PFM_REGT_EN)
+
+static void pfm_p4_write_pmc(struct pfm_context *ctx, unsigned int cnum, u64 value);
+static void pfm_p4_write_pmd(struct pfm_context *ctx, unsigned int cnum, u64 value);
+static u64 pfm_p4_read_pmd(struct pfm_context *ctx, unsigned int cnum);
+static u64 pfm_p4_read_pmc(struct pfm_context *ctx, unsigned int cnum);
+static int pfm_p4_create_context(struct pfm_context *ctx, u32 ctx_flags);
+static void pfm_p4_free_context(struct pfm_context *ctx);
+static int pfm_p4_has_ovfls(struct pfm_context *ctx);
+static int pfm_p4_stop_save(struct pfm_context *ctx, struct pfm_event_set *set);
+static void pfm_p4_restore_pmcs(struct pfm_context *ctx, struct pfm_event_set *set);
+static void pfm_p4_nmi_copy_state(struct pfm_context *ctx);
+static void __kprobes pfm_p4_quiesce(void);
+
+static u64 enable_mask[PFM_MAX_PMCS];
+static u16 max_enable;
+
+static struct pfm_p4_regmap pmc_addrs[PFM_MAX_PMCS] = {
+       /*pmc 0 */    {{MSR_P4_BPU_ESCR0, MSR_P4_BPU_ESCR1}, 0, PFM_REGT_ESCR}, /*   BPU_ESCR0,1 */
+       /*pmc 1 */    {{MSR_P4_IS_ESCR0, MSR_P4_IS_ESCR1}, 0, PFM_REGT_ESCR}, /*    IS_ESCR0,1 */
+       /*pmc 2 */    {{MSR_P4_MOB_ESCR0, MSR_P4_MOB_ESCR1}, 0, PFM_REGT_ESCR}, /*   MOB_ESCR0,1 */
+       /*pmc 3 */    {{MSR_P4_ITLB_ESCR0, MSR_P4_ITLB_ESCR1}, 0, PFM_REGT_ESCR}, /*  ITLB_ESCR0,1 */
+       /*pmc 4 */    {{MSR_P4_PMH_ESCR0, MSR_P4_PMH_ESCR1}, 0, PFM_REGT_ESCR}, /*   PMH_ESCR0,1 */
+       /*pmc 5 */    {{MSR_P4_IX_ESCR0, MSR_P4_IX_ESCR1}, 0, PFM_REGT_ESCR}, /*    IX_ESCR0,1 */
+       /*pmc 6 */    {{MSR_P4_FSB_ESCR0, MSR_P4_FSB_ESCR1}, 0, PFM_REGT_ESCR}, /*   FSB_ESCR0,1 */
+       /*pmc 7 */    {{MSR_P4_BSU_ESCR0, MSR_P4_BSU_ESCR1}, 0, PFM_REGT_ESCR}, /*   BSU_ESCR0,1 */
+       /*pmc 8 */    {{MSR_P4_MS_ESCR0, MSR_P4_MS_ESCR1}, 0, PFM_REGT_ESCR}, /*    MS_ESCR0,1 */
+       /*pmc 9 */    {{MSR_P4_TC_ESCR0, MSR_P4_TC_ESCR1}, 0, PFM_REGT_ESCR}, /*    TC_ESCR0,1 */
+       /*pmc 10*/    {{MSR_P4_TBPU_ESCR0, MSR_P4_TBPU_ESCR1}, 0, PFM_REGT_ESCR}, /*  TBPU_ESCR0,1 */
+       /*pmc 11*/    {{MSR_P4_FLAME_ESCR0, MSR_P4_FLAME_ESCR1}, 0, PFM_REGT_ESCR}, /* FLAME_ESCR0,1 */
+       /*pmc 12*/    {{MSR_P4_FIRM_ESCR0, MSR_P4_FIRM_ESCR1}, 0, PFM_REGT_ESCR}, /*  FIRM_ESCR0,1 */
+       /*pmc 13*/    {{MSR_P4_SAAT_ESCR0, MSR_P4_SAAT_ESCR1}, 0, PFM_REGT_ESCR}, /*  SAAT_ESCR0,1 */
+       /*pmc 14*/    {{MSR_P4_U2L_ESCR0, MSR_P4_U2L_ESCR1}, 0, PFM_REGT_ESCR}, /*   U2L_ESCR0,1 */
+       /*pmc 15*/    {{MSR_P4_DAC_ESCR0, MSR_P4_DAC_ESCR1}, 0, PFM_REGT_ESCR}, /*   DAC_ESCR0,1 */
+       /*pmc 16*/    {{MSR_P4_IQ_ESCR0, MSR_P4_IQ_ESCR1}, 0, PFM_REGT_ESCR}, /*    IQ_ESCR0,1 (only model 1 and 2) */
+       /*pmc 17*/    {{MSR_P4_ALF_ESCR0, MSR_P4_ALF_ESCR1}, 0, PFM_REGT_ESCR}, /*   ALF_ESCR0,1 */
+       /*pmc 18*/    {{MSR_P4_RAT_ESCR0, MSR_P4_RAT_ESCR1}, 0, PFM_REGT_ESCR}, /*   RAT_ESCR0,1 */
+       /*pmc 19*/    {{MSR_P4_SSU_ESCR0, 0}, 0, PFM_REGT_ESCR}, /*   SSU_ESCR0   */
+       /*pmc 20*/    {{MSR_P4_CRU_ESCR0, MSR_P4_CRU_ESCR1}, 0, PFM_REGT_ESCR}, /*   CRU_ESCR0,1 */
+       /*pmc 21*/    {{MSR_P4_CRU_ESCR2, MSR_P4_CRU_ESCR3}, 0, PFM_REGT_ESCR}, /*   CRU_ESCR2,3 */
+       /*pmc 22*/    {{MSR_P4_CRU_ESCR4, MSR_P4_CRU_ESCR5}, 0, PFM_REGT_ESCR}, /*   CRU_ESCR4,5 */
+
+       /*pmc 23*/    {{MSR_P4_BPU_CCCR0, MSR_P4_BPU_CCCR2}, 0, PFM_REGT_ENAC}, /*   BPU_CCCR0,2 */
+       /*pmc 24*/    {{MSR_P4_BPU_CCCR1, MSR_P4_BPU_CCCR3}, 1, PFM_REGT_ENAC}, /*   BPU_CCCR1,3 */
+       /*pmc 25*/    {{MSR_P4_MS_CCCR0, MSR_P4_MS_CCCR2}, 2, PFM_REGT_ENAC}, /*    MS_CCCR0,2 */
+       /*pmc 26*/    {{MSR_P4_MS_CCCR1, MSR_P4_MS_CCCR3}, 3, PFM_REGT_ENAC}, /*    MS_CCCR1,3 */
+       /*pmc 27*/    {{MSR_P4_FLAME_CCCR0, MSR_P4_FLAME_CCCR2}, 4, PFM_REGT_ENAC}, /* FLAME_CCCR0,2 */
+       /*pmc 28*/    {{MSR_P4_FLAME_CCCR1, MSR_P4_FLAME_CCCR3}, 5, PFM_REGT_ENAC}, /* FLAME_CCCR1,3 */
+       /*pmc 29*/    {{MSR_P4_IQ_CCCR0, MSR_P4_IQ_CCCR2}, 6, PFM_REGT_ENAC}, /*    IQ_CCCR0,2 */
+       /*pmc 30*/    {{MSR_P4_IQ_CCCR1, MSR_P4_IQ_CCCR3}, 7, PFM_REGT_ENAC}, /*    IQ_CCCR1,3 */
+       /*pmc 31*/    {{MSR_P4_IQ_CCCR4, MSR_P4_IQ_CCCR5}, 8, PFM_REGT_ENAC}, /*    IQ_CCCR4,5 */
+       /* non HT extensions */
+       /*pmc 32*/    {{MSR_P4_BPU_ESCR1,    0},  0, PFM_REGT_NHTESCR}, /*   BPU_ESCR1   */
+       /*pmc 33*/    {{MSR_P4_IS_ESCR1,     0},  0, PFM_REGT_NHTESCR}, /*    IS_ESCR1   */
+       /*pmc 34*/    {{MSR_P4_MOB_ESCR1,    0},  0, PFM_REGT_NHTESCR}, /*   MOB_ESCR1   */
+       /*pmc 35*/    {{MSR_P4_ITLB_ESCR1,   0},  0, PFM_REGT_NHTESCR}, /*  ITLB_ESCR1   */
+       /*pmc 36*/    {{MSR_P4_PMH_ESCR1,    0},  0, PFM_REGT_NHTESCR}, /*   PMH_ESCR1   */
+       /*pmc 37*/    {{MSR_P4_IX_ESCR1,     0},  0, PFM_REGT_NHTESCR}, /*    IX_ESCR1   */
+       /*pmc 38*/    {{MSR_P4_FSB_ESCR1,    0},  0, PFM_REGT_NHTESCR}, /*   FSB_ESCR1   */
+       /*pmc 39*/    {{MSR_P4_BSU_ESCR1,    0},  0, PFM_REGT_NHTESCR}, /*   BSU_ESCR1   */
+       /*pmc 40*/    {{MSR_P4_MS_ESCR1,     0},  0, PFM_REGT_NHTESCR}, /*    MS_ESCR1   */
+       /*pmc 41*/    {{MSR_P4_TC_ESCR1,     0},  0, PFM_REGT_NHTESCR}, /*    TC_ESCR1   */
+       /*pmc 42*/    {{MSR_P4_TBPU_ESCR1,   0},  0, PFM_REGT_NHTESCR}, /*  TBPU_ESCR1   */
+       /*pmc 43*/    {{MSR_P4_FLAME_ESCR1,  0},  0, PFM_REGT_NHTESCR}, /* FLAME_ESCR1   */
+       /*pmc 44*/    {{MSR_P4_FIRM_ESCR1,   0},  0, PFM_REGT_NHTESCR}, /*  FIRM_ESCR1   */
+       /*pmc 45*/    {{MSR_P4_SAAT_ESCR1,   0},  0, PFM_REGT_NHTESCR}, /*  SAAT_ESCR1   */
+       /*pmc 46*/    {{MSR_P4_U2L_ESCR1,    0},  0, PFM_REGT_NHTESCR}, /*   U2L_ESCR1   */
+       /*pmc 47*/    {{MSR_P4_DAC_ESCR1,    0},  0, PFM_REGT_NHTESCR}, /*   DAC_ESCR1   */
+       /*pmc 48*/    {{MSR_P4_IQ_ESCR1,     0},  0, PFM_REGT_NHTESCR}, /*    IQ_ESCR1   (only model 1 and 2) */
+       /*pmc 49*/    {{MSR_P4_ALF_ESCR1,    0},  0, PFM_REGT_NHTESCR}, /*   ALF_ESCR1   */
+       /*pmc 50*/    {{MSR_P4_RAT_ESCR1,    0},  0, PFM_REGT_NHTESCR}, /*   RAT_ESCR1   */
+       /*pmc 51*/    {{MSR_P4_CRU_ESCR1,    0},  0, PFM_REGT_NHTESCR}, /*   CRU_ESCR1   */
+       /*pmc 52*/    {{MSR_P4_CRU_ESCR3,    0},  0, PFM_REGT_NHTESCR}, /*   CRU_ESCR3   */
+       /*pmc 53*/    {{MSR_P4_CRU_ESCR5,    0},  0, PFM_REGT_NHTESCR}, /*   CRU_ESCR5   */
+       /*pmc 54*/    {{MSR_P4_BPU_CCCR1,    0},  9, PFM_REGT_NHTCCCR}, /*   BPU_CCCR1   */
+       /*pmc 55*/    {{MSR_P4_BPU_CCCR3,    0}, 10, PFM_REGT_NHTCCCR}, /*   BPU_CCCR3   */
+       /*pmc 56*/    {{MSR_P4_MS_CCCR1,     0}, 11, PFM_REGT_NHTCCCR}, /*    MS_CCCR1   */
+       /*pmc 57*/    {{MSR_P4_MS_CCCR3,     0}, 12, PFM_REGT_NHTCCCR}, /*    MS_CCCR3   */
+       /*pmc 58*/    {{MSR_P4_FLAME_CCCR1,  0}, 13, PFM_REGT_NHTCCCR}, /* FLAME_CCCR1   */
+       /*pmc 59*/    {{MSR_P4_FLAME_CCCR3,  0}, 14, PFM_REGT_NHTCCCR}, /* FLAME_CCCR3   */
+       /*pmc 60*/    {{MSR_P4_IQ_CCCR2,     0}, 15, PFM_REGT_NHTCCCR}, /*    IQ_CCCR2   */
+       /*pmc 61*/    {{MSR_P4_IQ_CCCR3,     0}, 16, PFM_REGT_NHTCCCR}, /*    IQ_CCCR3   */
+       /*pmc 62*/    {{MSR_P4_IQ_CCCR5,     0}, 17, PFM_REGT_NHTCCCR}, /*    IQ_CCCR5   */
+       /*pmc 63*/    {{0x3f2,     0}, 0, PFM_REGT_NHTPEBS},/* PEBS_MATRIX_VERT */
+       /*pmc 64*/    {{0x3f1,     0}, 0, PFM_REGT_NHTPEBS} /* PEBS_ENABLE   */
+};
+
+static struct pfm_p4_regmap pmd_addrs[PFM_MAX_PMDS] = {
+       /*pmd 0 */    {{MSR_P4_BPU_PERFCTR0, MSR_P4_BPU_PERFCTR2}, 0, PFM_REGT_CTR},  /*   BPU_CTR0,2  */
+       /*pmd 1 */    {{MSR_P4_BPU_PERFCTR1, MSR_P4_BPU_PERFCTR3}, 0, PFM_REGT_CTR},  /*   BPU_CTR1,3  */
+       /*pmd 2 */    {{MSR_P4_MS_PERFCTR0, MSR_P4_MS_PERFCTR2}, 0, PFM_REGT_CTR},  /*    MS_CTR0,2  */
+       /*pmd 3 */    {{MSR_P4_MS_PERFCTR1, MSR_P4_MS_PERFCTR3}, 0, PFM_REGT_CTR},  /*    MS_CTR1,3  */
+       /*pmd 4 */    {{MSR_P4_FLAME_PERFCTR0, MSR_P4_FLAME_PERFCTR2}, 0, PFM_REGT_CTR},  /* FLAME_CTR0,2  */
+       /*pmd 5 */    {{MSR_P4_FLAME_PERFCTR1, MSR_P4_FLAME_PERFCTR3}, 0, PFM_REGT_CTR},  /* FLAME_CTR1,3  */
+       /*pmd 6 */    {{MSR_P4_IQ_PERFCTR0, MSR_P4_IQ_PERFCTR2}, 0, PFM_REGT_CTR},  /*    IQ_CTR0,2  */
+       /*pmd 7 */    {{MSR_P4_IQ_PERFCTR1, MSR_P4_IQ_PERFCTR3}, 0, PFM_REGT_CTR},  /*    IQ_CTR1,3  */
+       /*pmd 8 */    {{MSR_P4_IQ_PERFCTR4, MSR_P4_IQ_PERFCTR5}, 0, PFM_REGT_CTR},  /*    IQ_CTR4,5  */
+       /*
+        * non HT extensions
+        */
+       /*pmd 9 */    {{MSR_P4_BPU_PERFCTR2,     0}, 0, PFM_REGT_NHTCTR},  /*   BPU_CTR2    */
+       /*pmd 10*/    {{MSR_P4_BPU_PERFCTR3,     0}, 0, PFM_REGT_NHTCTR},  /*   BPU_CTR3    */
+       /*pmd 11*/    {{MSR_P4_MS_PERFCTR2,     0}, 0, PFM_REGT_NHTCTR},  /*    MS_CTR2    */
+       /*pmd 12*/    {{MSR_P4_MS_PERFCTR3,     0}, 0, PFM_REGT_NHTCTR},  /*    MS_CTR3    */
+       /*pmd 13*/    {{MSR_P4_FLAME_PERFCTR2,     0}, 0, PFM_REGT_NHTCTR},  /* FLAME_CTR2    */
+       /*pmd 14*/    {{MSR_P4_FLAME_PERFCTR3,     0}, 0, PFM_REGT_NHTCTR},  /* FLAME_CTR3    */
+       /*pmd 15*/    {{MSR_P4_IQ_PERFCTR2,     0}, 0, PFM_REGT_NHTCTR},  /*    IQ_CTR2    */
+       /*pmd 16*/    {{MSR_P4_IQ_PERFCTR3,     0}, 0, PFM_REGT_NHTCTR},  /*    IQ_CTR3    */
+       /*pmd 17*/    {{MSR_P4_IQ_PERFCTR5,     0}, 0, PFM_REGT_NHTCTR},  /*    IQ_CTR5    */
+};
+
+static struct pfm_arch_pmu_info pfm_p4_pmu_info = {
+       .write_pmc = pfm_p4_write_pmc,
+       .write_pmd = pfm_p4_write_pmd,
+       .read_pmc = pfm_p4_read_pmc,
+       .read_pmd = pfm_p4_read_pmd,
+       .create_context = pfm_p4_create_context,
+       .free_context = pfm_p4_free_context,
+       .has_ovfls = pfm_p4_has_ovfls,
+       .stop_save = pfm_p4_stop_save,
+       .restore_pmcs = pfm_p4_restore_pmcs,
+       .nmi_copy_state = pfm_p4_nmi_copy_state,
+       .quiesce = pfm_p4_quiesce
+};
+
+static struct pfm_regmap_desc pfm_p4_pmc_desc[] = {
+/* pmc0  */ PMC_D(PFM_REG_I, "BPU_ESCR0"  , 0x0, PFM_ESCR_RSVD, 0, MSR_P4_BPU_ESCR0),
+/* pmc1  */ PMC_D(PFM_REG_I, "IS_ESCR0"   , 0x0, PFM_ESCR_RSVD, 0, MSR_P4_IQ_ESCR0),
+/* pmc2  */ PMC_D(PFM_REG_I, "MOB_ESCR0"  , 0x0, PFM_ESCR_RSVD, 0, MSR_P4_MOB_ESCR0),
+/* pmc3  */ PMC_D(PFM_REG_I, "ITLB_ESCR0" , 0x0, PFM_ESCR_RSVD, 0, MSR_P4_ITLB_ESCR0),
+/* pmc4  */ PMC_D(PFM_REG_I, "PMH_ESCR0"  , 0x0, PFM_ESCR_RSVD, 0, MSR_P4_PMH_ESCR0),
+/* pmc5  */ PMC_D(PFM_REG_I, "IX_ESCR0"   , 0x0, PFM_ESCR_RSVD, 0, MSR_P4_IX_ESCR0),
+/* pmc6  */ PMC_D(PFM_REG_I, "FSB_ESCR0"  , 0x0, PFM_ESCR_RSVD, 0, MSR_P4_FSB_ESCR0),
+/* pmc7  */ PMC_D(PFM_REG_I, "BSU_ESCR0"  , 0x0, PFM_ESCR_RSVD, 0, MSR_P4_BSU_ESCR0),
+/* pmc8  */ PMC_D(PFM_REG_I, "MS_ESCR0"   , 0x0, PFM_ESCR_RSVD, 0, MSR_P4_MS_ESCR0),
+/* pmc9  */ PMC_D(PFM_REG_I, "TC_ESCR0"   , 0x0, PFM_ESCR_RSVD, 0, MSR_P4_TC_ESCR0),
+/* pmc10 */ PMC_D(PFM_REG_I, "TBPU_ESCR0" , 0x0, PFM_ESCR_RSVD, 0, MSR_P4_TBPU_ESCR0),
+/* pmc11 */ PMC_D(PFM_REG_I, "FLAME_ESCR0", 0x0, PFM_ESCR_RSVD, 0, MSR_P4_FLAME_ESCR0),
+/* pmc12 */ PMC_D(PFM_REG_I, "FIRM_ESCR0" , 0x0, PFM_ESCR_RSVD, 0, MSR_P4_FIRM_ESCR0),
+/* pmc13 */ PMC_D(PFM_REG_I, "SAAT_ESCR0" , 0x0, PFM_ESCR_RSVD, 0, MSR_P4_SAAT_ESCR0),
+/* pmc14 */ PMC_D(PFM_REG_I, "U2L_ESCR0"  , 0x0, PFM_ESCR_RSVD, 0, MSR_P4_U2L_ESCR0),
+/* pmc15 */ PMC_D(PFM_REG_I, "DAC_ESCR0"  , 0x0, PFM_ESCR_RSVD, 0, MSR_P4_DAC_ESCR0),
+/* pmc16 */ PMC_D(PFM_REG_I, "IQ_ESCR0"   , 0x0, PFM_ESCR_RSVD, 0, MSR_P4_IQ_ESCR0), /* only model 1 and 2*/
+/* pmc17 */ PMC_D(PFM_REG_I, "ALF_ESCR0"  , 0x0, PFM_ESCR_RSVD, 0, MSR_P4_ALF_ESCR0),
+/* pmc18 */ PMC_D(PFM_REG_I, "RAT_ESCR0"  , 0x0, PFM_ESCR_RSVD, 0, MSR_P4_RAT_ESCR0),
+/* pmc19 */ PMC_D(PFM_REG_I, "SSU_ESCR0"  , 0x0, PFM_ESCR_RSVD, 0, MSR_P4_SSU_ESCR0),
+/* pmc20 */ PMC_D(PFM_REG_I, "CRU_ESCR0"  , 0x0, PFM_ESCR_RSVD, 0, MSR_P4_CRU_ESCR0),
+/* pmc21 */ PMC_D(PFM_REG_I, "CRU_ESCR2"  , 0x0, PFM_ESCR_RSVD, 0, MSR_P4_CRU_ESCR2),
+/* pmc22 */ PMC_D(PFM_REG_I, "CRU_ESCR4"  , 0x0, PFM_ESCR_RSVD, 0, MSR_P4_CRU_ESCR4),
+/* pmc23 */ PMC_D(PFM_REG_I64, "BPU_CCCR0"  , PFM_CCCR_DFL, PFM_CCCR_RSVD, PFM_P4_NO64, MSR_P4_BPU_CCCR0),
+/* pmc24 */ PMC_D(PFM_REG_I64, "BPU_CCCR1"  , PFM_CCCR_DFL, PFM_CCCR_RSVD, PFM_P4_NO64, MSR_P4_BPU_CCCR1),
+/* pmc25 */ PMC_D(PFM_REG_I64, "MS_CCCR0"   , PFM_CCCR_DFL, PFM_CCCR_RSVD, PFM_P4_NO64, MSR_P4_MS_CCCR0),
+/* pmc26 */ PMC_D(PFM_REG_I64, "MS_CCCR1"   , PFM_CCCR_DFL, PFM_CCCR_RSVD, PFM_P4_NO64, MSR_P4_MS_CCCR1),
+/* pmc27 */ PMC_D(PFM_REG_I64, "FLAME_CCCR0", PFM_CCCR_DFL, PFM_CCCR_RSVD, PFM_P4_NO64, MSR_P4_FLAME_CCCR0),
+/* pmc28 */ PMC_D(PFM_REG_I64, "FLAME_CCCR1", PFM_CCCR_DFL, PFM_CCCR_RSVD, PFM_P4_NO64, MSR_P4_FLAME_CCCR1),
+/* pmc29 */ PMC_D(PFM_REG_I64, "IQ_CCCR0"   , PFM_CCCR_DFL, PFM_CCCR_RSVD, PFM_P4_NO64, MSR_P4_IQ_CCCR0),
+/* pmc30 */ PMC_D(PFM_REG_I64, "IQ_CCCR1"   , PFM_CCCR_DFL, PFM_CCCR_RSVD, PFM_P4_NO64, MSR_P4_IQ_CCCR1),
+/* pmc31 */ PMC_D(PFM_REG_I64, "IQ_CCCR4"   , PFM_CCCR_DFL, PFM_CCCR_RSVD, PFM_P4_NO64, MSR_P4_IQ_CCCR4),
+               /* No HT extension */
+/* pmc32 */ PMC_D(PFM_REG_I, "BPU_ESCR1"  , 0x0, PFM_ESCR_RSVD, 0, MSR_P4_BPU_ESCR1),
+/* pmc33 */ PMC_D(PFM_REG_I, "IS_ESCR1"   , 0x0, PFM_ESCR_RSVD, 0, MSR_P4_IS_ESCR1),
+/* pmc34 */ PMC_D(PFM_REG_I, "MOB_ESCR1"  , 0x0, PFM_ESCR_RSVD, 0, MSR_P4_MOB_ESCR1),
+/* pmc35 */ PMC_D(PFM_REG_I, "ITLB_ESCR1" , 0x0, PFM_ESCR_RSVD, 0, MSR_P4_ITLB_ESCR1),
+/* pmc36 */ PMC_D(PFM_REG_I, "PMH_ESCR1"  , 0x0, PFM_ESCR_RSVD, 0, MSR_P4_PMH_ESCR1),
+/* pmc37 */ PMC_D(PFM_REG_I, "IX_ESCR1"   , 0x0, PFM_ESCR_RSVD, 0, MSR_P4_IX_ESCR1),
+/* pmc38 */ PMC_D(PFM_REG_I, "FSB_ESCR1"  , 0x0, PFM_ESCR_RSVD, 0, MSR_P4_FSB_ESCR1),
+/* pmc39 */ PMC_D(PFM_REG_I, "BSU_ESCR1"  , 0x0, PFM_ESCR_RSVD, 0, MSR_P4_BSU_ESCR1),
+/* pmc40 */ PMC_D(PFM_REG_I, "MS_ESCR1"   , 0x0, PFM_ESCR_RSVD, 0, MSR_P4_MS_ESCR1),
+/* pmc41 */ PMC_D(PFM_REG_I, "TC_ESCR1"   , 0x0, PFM_ESCR_RSVD, 0, MSR_P4_TC_ESCR1),
+/* pmc42 */ PMC_D(PFM_REG_I, "TBPU_ESCR1" , 0x0, PFM_ESCR_RSVD, 0, MSR_P4_TBPU_ESCR1),
+/* pmc43 */ PMC_D(PFM_REG_I, "FLAME_ESCR1", 0x0, PFM_ESCR_RSVD, 0, MSR_P4_FLAME_ESCR1),
+/* pmc44 */ PMC_D(PFM_REG_I, "FIRM_ESCR1" , 0x0, PFM_ESCR_RSVD, 0, MSR_P4_FIRM_ESCR1),
+/* pmc45 */ PMC_D(PFM_REG_I, "SAAT_ESCR1" , 0x0, PFM_ESCR_RSVD, 0, MSR_P4_SAAT_ESCR1),
+/* pmc46 */ PMC_D(PFM_REG_I, "U2L_ESCR1"  , 0x0, PFM_ESCR_RSVD, 0, MSR_P4_U2L_ESCR1),
+/* pmc47 */ PMC_D(PFM_REG_I, "DAC_ESCR1"  , 0x0, PFM_ESCR_RSVD, 0, MSR_P4_DAC_ESCR1),
+/* pmc48 */ PMC_D(PFM_REG_I, "IQ_ESCR1"   , 0x0, PFM_ESCR_RSVD, 0, MSR_P4_IQ_ESCR1), /* only model 1 and 2 */
+/* pmc49 */ PMC_D(PFM_REG_I, "ALF_ESCR1"  , 0x0, PFM_ESCR_RSVD, 0, MSR_P4_ALF_ESCR1),
+/* pmc50 */ PMC_D(PFM_REG_I, "RAT_ESCR1"  , 0x0, PFM_ESCR_RSVD, 0, MSR_P4_RAT_ESCR1),
+/* pmc51 */ PMC_D(PFM_REG_I, "CRU_ESCR1"  , 0x0, PFM_ESCR_RSVD, 0, MSR_P4_CRU_ESCR1),
+/* pmc52 */ PMC_D(PFM_REG_I, "CRU_ESCR3"  , 0x0, PFM_ESCR_RSVD, 0, MSR_P4_CRU_ESCR3),
+/* pmc53 */ PMC_D(PFM_REG_I, "CRU_ESCR5"  , 0x0, PFM_ESCR_RSVD, 0, MSR_P4_CRU_ESCR5),
+/* pmc54 */ PMC_D(PFM_REG_I64, "BPU_CCCR2"  , PFM_CCCR_DFL, PFM_CCCR_RSVD, PFM_P4_NO64, MSR_P4_BPU_CCCR2),
+/* pmc55 */ PMC_D(PFM_REG_I64, "BPU_CCCR3"  , PFM_CCCR_DFL, PFM_CCCR_RSVD, PFM_P4_NO64, MSR_P4_BPU_CCCR3),
+/* pmc56 */ PMC_D(PFM_REG_I64, "MS_CCCR2"   , PFM_CCCR_DFL, PFM_CCCR_RSVD, PFM_P4_NO64, MSR_P4_MS_CCCR2),
+/* pmc57 */ PMC_D(PFM_REG_I64, "MS_CCCR3"   , PFM_CCCR_DFL, PFM_CCCR_RSVD, PFM_P4_NO64, MSR_P4_MS_CCCR3),
+/* pmc58 */ PMC_D(PFM_REG_I64, "FLAME_CCCR2", PFM_CCCR_DFL, PFM_CCCR_RSVD, PFM_P4_NO64, MSR_P4_FLAME_CCCR2),
+/* pmc59 */ PMC_D(PFM_REG_I64, "FLAME_CCCR3", PFM_CCCR_DFL, PFM_CCCR_RSVD, PFM_P4_NO64, MSR_P4_FLAME_CCCR3),
+/* pmc60 */ PMC_D(PFM_REG_I64, "IQ_CCCR2"   , PFM_CCCR_DFL, PFM_CCCR_RSVD, PFM_P4_NO64, MSR_P4_IQ_CCCR2),
+/* pmc61 */ PMC_D(PFM_REG_I64, "IQ_CCCR3"   , PFM_CCCR_DFL, PFM_CCCR_RSVD, PFM_P4_NO64, MSR_P4_IQ_CCCR3),
+/* pmc62 */ PMC_D(PFM_REG_I64, "IQ_CCCR5"   , PFM_CCCR_DFL, PFM_CCCR_RSVD, PFM_P4_NO64, MSR_P4_IQ_CCCR5),
+/* pmc63 */ PMC_D(PFM_REG_I, "PEBS_MATRIX_VERT", 0, 0xffffffffffffffecULL, 0, 0x3f2),
+/* pmc64 */ PMC_D(PFM_REG_I, "PEBS_ENABLE", 0, 0xfffffffff8ffe000ULL, 0, 0x3f1)
+};
+#define PFM_P4_NUM_PMCS ARRAY_SIZE(pfm_p4_pmc_desc)
+
+/*
+ * See section 15.10.6.6 for details about the IQ block
+ */
+static struct pfm_regmap_desc pfm_p4_pmd_desc[] = {
+/* pmd0  */ PMD_D(PFM_REG_C, "BPU_CTR0", MSR_P4_BPU_PERFCTR0),
+/* pmd1  */ PMD_D(PFM_REG_C, "BPU_CTR1", MSR_P4_BPU_PERFCTR1),
+/* pmd2  */ PMD_D(PFM_REG_C, "MS_CTR0", MSR_P4_MS_PERFCTR0),
+/* pmd3  */ PMD_D(PFM_REG_C, "MS_CTR1", MSR_P4_MS_PERFCTR1),
+/* pmd4  */ PMD_D(PFM_REG_C, "FLAME_CTR0", MSR_P4_FLAME_PERFCTR0),
+/* pmd5  */ PMD_D(PFM_REG_C, "FLAME_CTR1", MSR_P4_FLAME_PERFCTR1),
+/* pmd6  */ PMD_D(PFM_REG_C, "IQ_CTR0", MSR_P4_IQ_PERFCTR0),
+/* pmd7  */ PMD_D(PFM_REG_C, "IQ_CTR1", MSR_P4_IQ_PERFCTR1),
+/* pmd8  */ PMD_D(PFM_REG_C, "IQ_CTR4", MSR_P4_IQ_PERFCTR4),
+               /* no HT extension */
+/* pmd9  */ PMD_D(PFM_REG_C, "BPU_CTR2", MSR_P4_BPU_PERFCTR2),
+/* pmd10 */ PMD_D(PFM_REG_C, "BPU_CTR3", MSR_P4_BPU_PERFCTR3),
+/* pmd11 */ PMD_D(PFM_REG_C, "MS_CTR2", MSR_P4_MS_PERFCTR2),
+/* pmd12 */ PMD_D(PFM_REG_C, "MS_CTR3", MSR_P4_MS_PERFCTR3),
+/* pmd13 */ PMD_D(PFM_REG_C, "FLAME_CTR2", MSR_P4_FLAME_PERFCTR2),
+/* pmd14 */ PMD_D(PFM_REG_C, "FLAME_CTR3", MSR_P4_FLAME_PERFCTR3),
+/* pmd15 */ PMD_D(PFM_REG_C, "IQ_CTR2", MSR_P4_IQ_PERFCTR2),
+/* pmd16 */ PMD_D(PFM_REG_C, "IQ_CTR3", MSR_P4_IQ_PERFCTR3),
+/* pmd17 */ PMD_D(PFM_REG_C, "IQ_CTR5", MSR_P4_IQ_PERFCTR5)
+};
+#define PFM_P4_NUM_PMDS ARRAY_SIZE(pfm_p4_pmd_desc)
+
+/*
+ * Due to hotplug CPU support, threads may not necessarily
+ * be activated at the time the module is inserted. We need
+ * to check whether  they could be activated by looking at
+ * the present CPU (present != online).
+ */
+static int pfm_p4_probe_pmu(void)
+{
+       unsigned int i;
+       int ht_enabled;
+
+       /*
+        * only works on Intel processors
+        */
+       if (current_cpu_data.x86_vendor != X86_VENDOR_INTEL) {
+               PFM_INFO("not running on Intel processor");
+               return -1;
+       }
+
+       if (current_cpu_data.x86 != 15) {
+               PFM_INFO("unsupported family=%d", current_cpu_data.x86);
+               return -1;
+       }
+
+       switch (current_cpu_data.x86_model) {
+       case 0 ... 2:
+               break;
+       case 3 ... 6:
+               /*
+                * IQ_ESCR0, IQ_ESCR1 only present on model 1, 2
+                */
+               pfm_p4_pmc_desc[16].type = PFM_REG_NA;
+               pfm_p4_pmc_desc[48].type = PFM_REG_NA;
+               break;
+       default:
+               /*
+                * do not know if they all work the same, so reject
+                * for now
+                */
+               if (!force) {
+                       PFM_INFO("unsupported model %d",
+                                current_cpu_data.x86_model);
+                       return -1;
+               }
+       }
+
+       /*
+        * check for local APIC (required)
+        */
+       if (!cpu_has_apic) {
+               PFM_INFO("no local APIC, unsupported");
+               return -1;
+       }
+#ifdef CONFIG_SMP
+       ht_enabled = (cpus_weight(__get_cpu_var(cpu_core_map))
+                  / current_cpu_data.x86_max_cores) > 1;
+#else
+       ht_enabled = 0;
+#endif
+       if (cpu_has_ht) {
+
+               PFM_INFO("HyperThreading supported, status %s",
+                        ht_enabled ? "on": "off");
+               /*
+                * disable registers not supporting HT
+                */
+               if (ht_enabled) {
+                       PFM_INFO("disabling half the registers for HT");
+                       for (i = 0; i < PFM_P4_NUM_PMCS; i++) {
+                               if (pmc_addrs[(i)].reg_type & PFM_REGT_NOHT)
+                                       pfm_p4_pmc_desc[i].type = PFM_REG_NA;
+                       }
+                       for (i = 0; i < PFM_P4_NUM_PMDS; i++) {
+                               if (pmd_addrs[(i)].reg_type & PFM_REGT_NOHT)
+                                       pfm_p4_pmd_desc[i].type = PFM_REG_NA;
+                       }
+               }
+       }
+
+       if (cpu_has_ds) {
+               PFM_INFO("Data Save Area (DS) supported");
+
+               if (cpu_has_pebs) {
+                       /*
+                        * PEBS does not work with HyperThreading enabled
+                        */
+                       if (ht_enabled)
+                               PFM_INFO("PEBS supported, status off (because of HT)");
+                       else
+                               PFM_INFO("PEBS supported, status on");
+               }
+       }
+
+       /*
+        * build enable mask
+        */
+       for (i = 0; i < PFM_P4_NUM_PMCS; i++) {
+               if (pmc_addrs[(i)].reg_type & PFM_REGT_EN) {
+                       __set_bit(i, cast_ulp(enable_mask));
+                       max_enable = i + 1;
+               }
+       }
+
+       if (force_nmi)
+               pfm_p4_pmu_info.flags |= PFM_X86_FL_USE_NMI;
+       return 0;
+}
+static inline int get_smt_id(void)
+{
+#ifdef CONFIG_SMP
+       int cpu = smp_processor_id();
+       return (cpu != first_cpu(__get_cpu_var(cpu_sibling_map)));
+#else
+       return 0;
+#endif
+}
+
+static void __pfm_write_reg_p4(const struct pfm_p4_regmap *xreg, u64 val)
+{
+       u64 pmi;
+       int smt_id;
+
+       smt_id = get_smt_id();
+       /*
+        * HT is only supported by P4-style PMU
+        *
+        * Adjust for T1 if necessary:
+        *
+        * - move the T0_OS/T0_USR bits into T1 slots
+        * - move the OVF_PMI_T0 bits into T1 slot
+        *
+        * The P4/EM64T T1 is cleared by description table.
+        * User only works with T0.
+        */
+       if (smt_id) {
+               if (xreg->reg_type & PFM_REGT_ESCR) {
+
+                       /* copy T0_USR & T0_OS to T1 */
+                       val |= ((val & 0xc) >> 2);
+
+                       /* clear bits T0_USR & T0_OS */
+                       val &= ~0xc;
+
+               } else if (xreg->reg_type & PFM_REGT_CCCR) {
+                       pmi = (val >> 26) & 0x1;
+                       if (pmi) {
+                               val &= ~(1UL<<26);
+                               val |= 1UL<<27;
+                       }
+               }
+       }
+       if (xreg->addrs[smt_id])
+               wrmsrl(xreg->addrs[smt_id], val);
+}
+
+void __pfm_read_reg_p4(const struct pfm_p4_regmap *xreg, u64 *val)
+{
+       int smt_id;
+
+       smt_id = get_smt_id();
+
+       if (likely(xreg->addrs[smt_id])) {
+               rdmsrl(xreg->addrs[smt_id], *val);
+               /*
+                * HT is only supported by P4-style PMU
+                *
+                * move the Tx_OS and Tx_USR bits into
+                * T0 slots setting the T1 slots to zero
+                */
+               if (xreg->reg_type & PFM_REGT_ESCR) {
+                       if (smt_id)
+                               *val |= (((*val) & 0x3) << 2);
+
+                       /*
+                        * zero out bits that are reserved
+                        * (including T1_OS and T1_USR)
+                        */
+                       *val &= PFM_ESCR_RSVD;
+               }
+       } else {
+               *val = 0;
+       }
+}
+static void pfm_p4_write_pmc(struct pfm_context *ctx, unsigned int cnum, u64 value)
+{
+       __pfm_write_reg_p4(&pmc_addrs[cnum], value);
+}
+
+static void pfm_p4_write_pmd(struct pfm_context *ctx, unsigned int cnum, u64 value)
+{
+       __pfm_write_reg_p4(&pmd_addrs[cnum], value);
+}
+
+static u64 pfm_p4_read_pmd(struct pfm_context *ctx, unsigned int cnum)
+{
+       u64 tmp;
+       __pfm_read_reg_p4(&pmd_addrs[cnum], &tmp);
+       return tmp;
+}
+
+static u64 pfm_p4_read_pmc(struct pfm_context *ctx, unsigned int cnum)
+{
+       u64 tmp;
+       __pfm_read_reg_p4(&pmc_addrs[cnum], &tmp);
+       return tmp;
+}
+
+struct pfm_ds_area_p4 {
+       unsigned long   bts_buf_base;
+       unsigned long   bts_index;
+       unsigned long   bts_abs_max;
+       unsigned long   bts_intr_thres;
+       unsigned long   pebs_buf_base;
+       unsigned long   pebs_index;
+       unsigned long   pebs_abs_max;
+       unsigned long   pebs_intr_thres;
+       u64             pebs_cnt_reset;
+};
+
+
+static int pfm_p4_stop_save(struct pfm_context *ctx, struct pfm_event_set *set)
+{
+       struct pfm_arch_pmu_info *pmu_info;
+       struct pfm_arch_context *ctx_arch;
+       struct pfm_ds_area_p4 *ds = NULL;
+       u64 used_mask[PFM_PMC_BV];
+       u16 i, j, count, pebs_idx = ~0;
+       u16 max_pmc;
+       u64 cccr, ctr1, ctr2, ovfl_mask;
+
+       pmu_info = &pfm_p4_pmu_info;
+       ctx_arch = pfm_ctx_arch(ctx);
+       max_pmc = ctx->regs.max_pmc;
+       ovfl_mask = pfm_pmu_conf->ovfl_mask;
+
+       /*
+        * build used enable PMC bitmask
+        * if user did not set any CCCR, then mask is
+        * empty and there is nothing to do because nothing
+        * was started
+        */
+       bitmap_and(cast_ulp(used_mask),
+                  cast_ulp(set->used_pmcs),
+                  cast_ulp(enable_mask),
+                  max_enable);
+
+       count = bitmap_weight(cast_ulp(used_mask), max_enable);
+
+       PFM_DBG_ovfl("npend=%u ena_mask=0x%llx u_pmcs=0x%llx count=%u num=%u",
+               set->npend_ovfls,
+               (unsigned long long)enable_mask[0],
+               (unsigned long long)set->used_pmcs[0],
+               count, max_enable);
+
+       /*
+        * ensures we do not destroy pending overflow
+        * information. If pended interrupts are already
+        * known, then we just stop monitoring.
+        */
+       if (set->npend_ovfls) {
+               /*
+                * clear enable bit
+                * unfortunately, this is very expensive!
+                */
+               for (i = 0; count; i++) {
+                       if (test_bit(i, cast_ulp(used_mask))) {
+                               __pfm_write_reg_p4(pmc_addrs+i, 0);
+                               count--;
+                       }
+               }
+               /* need save PMDs at upper level */
+               return 1;
+       }
+
+       if (ctx_arch->flags.use_pebs) {
+               ds = ctx_arch->ds_area;
+               pebs_idx = PEBS_PMD;
+               PFM_DBG("ds=%p pebs_idx=0x%llx thres=0x%llx",
+                       ds,
+                       (unsigned long long)ds->pebs_index,
+                       (unsigned long long)ds->pebs_intr_thres);
+       }
+
+       /*
+        * stop monitoring AND collect pending overflow information AND
+        * save pmds.
+        *
+        * We need to access the CCCR twice, once to get overflow info
+        * and a second to stop monitoring (which destroys the OVF flag)
+        * Similarly, we need to read the counter twice to check whether
+        * it did overflow between the CCR read and the CCCR write.
+        */
+       for (i = 0; count; i++) {
+               if (i != pebs_idx && test_bit(i, cast_ulp(used_mask))) {
+                       /*
+                        * controlled counter
+                        */
+                       j = pmc_addrs[i].ctr;
+
+                       /* read CCCR (PMC) value */
+                       __pfm_read_reg_p4(pmc_addrs+i, &cccr);
+
+                       /* read counter (PMD) controlled by PMC */
+                       __pfm_read_reg_p4(pmd_addrs+j, &ctr1);
+
+                       /* clear CCCR value: stop counter but destroy OVF */
+                       __pfm_write_reg_p4(pmc_addrs+i, 0);
+
+                       /* read counter controlled by CCCR again */
+                       __pfm_read_reg_p4(pmd_addrs+j, &ctr2);
+
+                       /*
+                        * there is an overflow if either:
+                        *      - CCCR.ovf is set (and we just cleared it)
+                        *      - ctr2 < ctr1
+                        * in that case we set the bit corresponding to the
+                        * overflowed PMD  in povfl_pmds.
+                        */
+                       if ((cccr & (1ULL<<31)) || (ctr2 < ctr1)) {
+                               __set_bit(j, cast_ulp(set->povfl_pmds));
+                               set->npend_ovfls++;
+                       }
+                       ctr2 = (set->pmds[j].value & ~ovfl_mask) | (ctr2 & ovfl_mask);
+                       set->pmds[j].value = ctr2;
+                       count--;
+               }
+       }
+       /*
+        * check for PEBS buffer full and set the corresponding PMD overflow
+        */
+       if (ctx_arch->flags.use_pebs) {
+               PFM_DBG("ds=%p pebs_idx=0x%lx thres=0x%lx", ds, ds->pebs_index, ds->pebs_intr_thres);
+               if (ds->pebs_index >= ds->pebs_intr_thres
+                   && test_bit(PEBS_PMD, cast_ulp(set->used_pmds))) {
+                       __set_bit(PEBS_PMD, cast_ulp(set->povfl_pmds));
+                       set->npend_ovfls++;
+               }
+       }
+       /* 0 means: no need to save the PMD at higher level */
+       return 0;
+}
+
+static int pfm_p4_create_context(struct pfm_context *ctx, u32 ctx_flags)
+{
+       struct pfm_arch_context *ctx_arch;
+
+       ctx_arch = pfm_ctx_arch(ctx);
+
+       ctx_arch->data = kzalloc(sizeof(struct pfm_arch_p4_context), GFP_KERNEL);
+       if (!ctx_arch->data)
+               return -ENOMEM;
+
+       return 0;
+}
+
+static void pfm_p4_free_context(struct pfm_context *ctx)
+{
+       struct pfm_arch_context *ctx_arch;
+
+       ctx_arch = pfm_ctx_arch(ctx);
+       /*
+        * we do not check if P4, because it would be NULL and
+        * kfree can deal with NULL
+        */
+       kfree(ctx_arch->data);
+}
+
+/*
+ * detect is counters have overflowed.
+ * return:
+ *     0 : no overflow
+ *     1 : at least one overflow
+ *
+ * used by Intel P4
+ */
+static int __kprobes pfm_p4_has_ovfls(struct pfm_context *ctx)
+{
+       struct pfm_arch_pmu_info *pmu_info;
+       struct pfm_p4_regmap *xrc, *xrd;
+       struct pfm_arch_context *ctx_arch;
+       struct pfm_arch_p4_context *p4;
+       u64 ena_mask[PFM_PMC_BV];
+       u64 cccr, ctr1, ctr2;
+       int n, i, j;
+
+       pmu_info = &pfm_p4_pmu_info;
+
+       ctx_arch = pfm_ctx_arch(ctx);
+       xrc = pmc_addrs;
+       xrd = pmd_addrs;
+       p4 = ctx_arch->data;
+
+       bitmap_and(cast_ulp(ena_mask),
+                  cast_ulp(ctx->regs.pmcs),
+                  cast_ulp(enable_mask),
+                  max_enable);
+
+       n = bitmap_weight(cast_ulp(ena_mask), max_enable);
+
+       for (i = 0; n; i++) {
+               if (!test_bit(i, cast_ulp(ena_mask)))
+                       continue;
+               /*
+                * controlled counter
+                */
+               j = xrc[i].ctr;
+
+               /* read CCCR (PMC) value */
+               __pfm_read_reg_p4(xrc+i, &cccr);
+
+               /* read counter (PMD) controlled by PMC */
+               __pfm_read_reg_p4(xrd+j, &ctr1);
+
+               /* clear CCCR value: stop counter but destroy OVF */
+               __pfm_write_reg_p4(xrc+i, 0);
+
+               /* read counter controlled by CCCR again */
+               __pfm_read_reg_p4(xrd+j, &ctr2);
+
+               /*
+                * there is an overflow if either:
+                *      - CCCR.ovf is set (and we just cleared it)
+                *      - ctr2 < ctr1
+                * in that case we set the bit corresponding to the
+                * overflowed PMD in povfl_pmds.
+                */
+               if ((cccr & (1ULL<<31)) || (ctr2 < ctr1)) {
+                       __set_bit(j, cast_ulp(p4->povfl_pmds));
+                       p4->npend_ovfls++;
+               }
+               p4->saved_cccrs[i] = cccr;
+               n--;
+       }
+       /*
+        * if there was no overflow, then it means the NMI was not really
+        * for us, so we have to resume monitoring
+        */
+       if (unlikely(!p4->npend_ovfls)) {
+               for (i = 0; n; i++) {
+                       if (!test_bit(i, cast_ulp(ena_mask)))
+                               continue;
+                       __pfm_write_reg_p4(xrc+i, p4->saved_cccrs[i]);
+               }
+       }
+       return 0;
+}
+
+void pfm_p4_restore_pmcs(struct pfm_context *ctx, struct pfm_event_set *set)
+{
+       struct pfm_arch_pmu_info *pmu_info;
+       struct pfm_arch_context *ctx_arch;
+       u64 *mask;
+       u16 i, num;
+
+       ctx_arch = pfm_ctx_arch(ctx);
+       pmu_info = pfm_pmu_info();
+
+       /*
+        * must restore DS pointer before restoring PMCs
+        * as this can potentially reactivate monitoring
+        */
+       if (ctx_arch->flags.use_ds)
+               wrmsrl(MSR_IA32_DS_AREA, (unsigned long)ctx_arch->ds_area);
+
+       /*
+        * must restore everything because there are some dependencies
+        * (e.g., ESCR and CCCR)
+        */
+       num = ctx->regs.num_pmcs;
+       mask = ctx->regs.pmcs;
+       for (i = 0; num; i++) {
+               if (test_bit(i, cast_ulp(mask))) {
+                       pfm_arch_write_pmc(ctx, i, set->pmcs[i]);
+                       num--;
+               }
+       }
+}
+
+/*
+ * invoked only when NMI is used. Called from the LOCAL_PERFMON_VECTOR
+ * handler to copy P4 overflow state captured when the NMI triggered.
+ * Given that on P4, stopping monitoring destroy the overflow information
+ * we save it in pfm_has_ovfl_p4() where monitoring is also stopped.
+ *
+ * Here we propagate the overflow state to current active set. The
+ * freeze_pmu() call we not overwrite this state because npend_ovfls
+ * is non-zero.
+ */
+static void pfm_p4_nmi_copy_state(struct pfm_context *ctx)
+{
+       struct pfm_arch_context *ctx_arch;
+       struct pfm_event_set *set;
+       struct pfm_arch_p4_context *p4;
+
+       ctx_arch = pfm_ctx_arch(ctx);
+       p4 = ctx_arch->data;
+       set = ctx->active_set;
+
+       if (p4->npend_ovfls) {
+               set->npend_ovfls = p4->npend_ovfls;
+
+               bitmap_copy(cast_ulp(set->povfl_pmds),
+                           cast_ulp(p4->povfl_pmds),
+                           ctx->regs.max_pmd);
+
+               p4->npend_ovfls = 0;
+       }
+}
+
+/**
+ * pfm_p4_quiesce - stop monitoring without grabbing any lock
+ *
+ * called from NMI interrupt handler to immediately stop monitoring
+ * cannot grab any lock, including perfmon related locks
+ */
+static void __kprobes pfm_p4_quiesce(void)
+{
+       u16 i;
+       /*
+        * quiesce PMU by clearing available registers that have
+        * the start/stop capability
+        */
+       for (i = 0; i < pfm_pmu_conf->regs_all.max_pmc; i++) {
+               if (test_bit(i, cast_ulp(pfm_pmu_conf->regs_all.pmcs))
+                   && test_bit(i, cast_ulp(enable_mask)))
+                       __pfm_write_reg_p4(pmc_addrs+i, 0);
+       }
+}
+
+
+static struct pfm_pmu_config pfm_p4_pmu_conf = {
+       .pmu_name = "Intel P4",
+       .counter_width = 40,
+       .pmd_desc = pfm_p4_pmd_desc,
+       .pmc_desc = pfm_p4_pmc_desc,
+       .num_pmc_entries = PFM_P4_NUM_PMCS,
+       .num_pmd_entries = PFM_P4_NUM_PMDS,
+       .probe_pmu = pfm_p4_probe_pmu,
+       .version = "1.0",
+       .flags = PFM_PMU_BUILTIN_FLAG,
+       .owner = THIS_MODULE,
+       .pmu_info = &pfm_p4_pmu_info
+};
+
+static int __init pfm_p4_pmu_init_module(void)
+{
+       return pfm_pmu_register(&pfm_p4_pmu_conf);
+}
+
+static void __exit pfm_p4_pmu_cleanup_module(void)
+{
+       pfm_pmu_unregister(&pfm_p4_pmu_conf);
+}
+
+module_init(pfm_p4_pmu_init_module);
+module_exit(pfm_p4_pmu_cleanup_module);
diff --git a/arch/x86/perfmon/perfmon_p6.c b/arch/x86/perfmon/perfmon_p6.c
new file mode 100644
index 0000000..47c0a46
--- /dev/null
+++ b/arch/x86/perfmon/perfmon_p6.c
@@ -0,0 +1,310 @@
+/*
+ * This file contains the P6 family processor PMU register description tables
+ *
+ * This module supports original P6 processors
+ * (Pentium II, Pentium Pro, Pentium III) and Pentium M.
+ *
+ * Copyright (c) 2005-2007 Hewlett-Packard Development Company, L.P.
+ * Contributed by Stephane Eranian <eranian@hpl.hp.com>
+ *
+ * This program is free software; you can redistribute it and/or
+ * modify it under the terms of version 2 of the GNU General Public
+ * License as published by the Free Software Foundation.
+ *
+ * This program is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+ * General Public License for more details.
+ *
+ * You should have received a copy of the GNU General Public License
+ * along with this program; if not, write to the Free Software
+ * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA
+ * 02111-1307 USA
+  */
+#include <linux/module.h>
+#include <linux/kprobes.h>
+#include <linux/perfmon_kern.h>
+#include <linux/nmi.h>
+#include <asm/msr.h>
+
+MODULE_AUTHOR("Stephane Eranian <eranian@hpl.hp.com>");
+MODULE_DESCRIPTION("P6 PMU description table");
+MODULE_LICENSE("GPL");
+
+static int force_nmi;
+MODULE_PARM_DESC(force_nmi, "bool: force use of NMI for PMU interrupt");
+module_param(force_nmi, bool, 0600);
+
+/*
+ * - upper 32 bits are reserved
+ * - INT: APIC enable bit is reserved (forced to 1)
+ * - bit 21 is reserved
+ * - bit 22 is reserved on PEREVNTSEL1
+ *
+ * RSVD: reserved bits are 1
+ */
+#define PFM_P6_PMC0_RSVD ((~((1ULL<<32)-1)) | (1ULL<<20) | (1ULL<<21))
+#define PFM_P6_PMC1_RSVD ((~((1ULL<<32)-1)) | (1ULL<<20) | (3ULL<<21))
+
+/*
+ * force Local APIC interrupt on overflow
+ * disable with NO_EMUL64
+ */
+#define PFM_P6_PMC_VAL  (1ULL<<20)
+#define PFM_P6_NO64    (1ULL<<20)
+
+
+static void __kprobes pfm_p6_quiesce(void);
+static int pfm_p6_has_ovfls(struct pfm_context *ctx);
+static int pfm_p6_stop_save(struct pfm_context *ctx,
+                           struct pfm_event_set *set);
+
+static u64 enable_mask[PFM_MAX_PMCS];
+static u16 max_enable;
+
+/*
+ * PFM_X86_FL_NO_SHARING: because of the single enable bit on MSR_P6_EVNTSEL0
+ * the PMU cannot be shared with NMI watchdog or Oprofile
+ */
+struct pfm_arch_pmu_info pfm_p6_pmu_info = {
+       .stop_save = pfm_p6_stop_save,
+       .has_ovfls = pfm_p6_has_ovfls,
+       .quiesce = pfm_p6_quiesce,
+       .flags = PFM_X86_FL_NO_SHARING,
+};
+
+static struct pfm_regmap_desc pfm_p6_pmc_desc[] = {
+/* pmc0  */ PMC_D(PFM_REG_I64, "PERFEVTSEL0", PFM_P6_PMC_VAL, PFM_P6_PMC0_RSVD, PFM_P6_NO64, MSR_P6_EVNTSEL0),
+/* pmc1  */ PMC_D(PFM_REG_I64, "PERFEVTSEL1", PFM_P6_PMC_VAL, PFM_P6_PMC1_RSVD, PFM_P6_NO64, MSR_P6_EVNTSEL1)
+};
+#define PFM_P6_NUM_PMCS        ARRAY_SIZE(pfm_p6_pmc_desc)
+
+#define PFM_P6_D(n) \
+       { .type = PFM_REG_C,                    \
+         .desc = "PERFCTR"#n,                  \
+         .hw_addr = MSR_P6_PERFCTR0+n,         \
+         .rsvd_msk = 0,                        \
+         .dep_pmcs[0] = 1ULL << n              \
+       }
+
+static struct pfm_regmap_desc pfm_p6_pmd_desc[] = {
+/* pmd0  */ PFM_P6_D(0),
+/* pmd1  */ PFM_P6_D(1)
+};
+#define PFM_P6_NUM_PMDS ARRAY_SIZE(pfm_p6_pmd_desc)
+
+static int pfm_p6_probe_pmu(void)
+{
+       int high, low;
+
+       if (current_cpu_data.x86_vendor != X86_VENDOR_INTEL) {
+               PFM_INFO("not an Intel processor");
+               return -1;
+       }
+
+       /*
+        * check for P6 processor family
+        */
+       if (current_cpu_data.x86 != 6) {
+               PFM_INFO("unsupported family=%d", current_cpu_data.x86);
+               return -1;
+       }
+
+       switch (current_cpu_data.x86_model) {
+       case 1: /* Pentium Pro */
+       case 3:
+       case 5: /* Pentium II Deschutes */
+       case 7 ... 11:
+               break;
+       case 13:
+               /* for Pentium M, we need to check if PMU exist */
+               rdmsr(MSR_IA32_MISC_ENABLE, low, high);
+               if (low & (1U << 7))
+                       break;
+       default:
+               PFM_INFO("unsupported CPU model %d",
+                        current_cpu_data.x86_model);
+               return -1;
+
+       }
+
+       if (!cpu_has_apic) {
+               PFM_INFO("no Local APIC, try rebooting with lapic");
+               return -1;
+       }
+       __set_bit(0, cast_ulp(enable_mask));
+       __set_bit(1, cast_ulp(enable_mask));
+       max_enable = 1 + 1;
+       /*
+        * force NMI interrupt?
+        */
+       if (force_nmi)
+               pfm_p6_pmu_info.flags |= PFM_X86_FL_USE_NMI;
+
+       return 0;
+}
+
+/**
+ * pfm_p6_has_ovfls - check for pending overflow condition
+ * @ctx: context to work on
+ *
+ * detect if counters have overflowed.
+ * return:
+ *     0 : no overflow
+ *     1 : at least one overflow
+ */
+static int __kprobes pfm_p6_has_ovfls(struct pfm_context *ctx)
+{
+       u64 *cnt_mask;
+       u64 wmask, val;
+       u16 i, num;
+
+       cnt_mask = ctx->regs.cnt_pmds;
+       num = ctx->regs.num_counters;
+       wmask = 1ULL << pfm_pmu_conf->counter_width;
+
+       /*
+        * we can leverage the fact that we know the mapping
+        * to hardcode the MSR address and avoid accessing
+        * more cachelines
+        *
+        * We need to check cnt_mask because not all registers
+        * may be available.
+        */
+       for (i = 0; num; i++) {
+               if (test_bit(i, cast_ulp(cnt_mask))) {
+                       rdmsrl(MSR_P6_PERFCTR0+i, val);
+                       if (!(val & wmask))
+                               return 1;
+                       num--;
+               }
+       }
+       return 0;
+}
+
+/**
+ * pfm_p6_stop_save -- stop monitoring and save PMD values
+ * @ctx: context to work on
+ * @set: current event set
+ *
+ * return value:
+ *     0 - no need to save PMDs in caller
+ *     1 - need to save PMDs in caller
+ */
+static int pfm_p6_stop_save(struct pfm_context *ctx, struct pfm_event_set *set)
+{
+       struct pfm_arch_pmu_info *pmu_info;
+       u64 used_mask[PFM_PMC_BV];
+       u64 *cnt_pmds;
+       u64 val, wmask, ovfl_mask;
+       u32 i, count;
+
+       pmu_info = pfm_pmu_info();
+
+       wmask = 1ULL << pfm_pmu_conf->counter_width;
+       bitmap_and(cast_ulp(used_mask),
+                  cast_ulp(set->used_pmcs),
+                  cast_ulp(enable_mask),
+                  max_enable);
+
+       count = bitmap_weight(cast_ulp(used_mask), ctx->regs.max_pmc);
+
+       /*
+        * stop monitoring
+        * Unfortunately, this is very expensive!
+        * wrmsrl() is serializing.
+        */
+       for (i = 0; count; i++) {
+               if (test_bit(i, cast_ulp(used_mask))) {
+                       wrmsrl(MSR_P6_EVNTSEL0+i, 0);
+                       count--;
+               }
+       }
+
+       /*
+        * if we already having a pending overflow condition, we simply
+        * return to take care of this first.
+        */
+       if (set->npend_ovfls)
+               return 1;
+
+       ovfl_mask = pfm_pmu_conf->ovfl_mask;
+       cnt_pmds = ctx->regs.cnt_pmds;
+
+       /*
+        * check for pending overflows and save PMDs (combo)
+        * we employ used_pmds because we also need to save
+        * and not just check for pending interrupts.
+        *
+        * Must check for counting PMDs because of virtual PMDs
+        */
+       count = set->nused_pmds;
+       for (i = 0; count; i++) {
+               if (test_bit(i, cast_ulp(set->used_pmds))) {
+                       val = pfm_arch_read_pmd(ctx, i);
+                       if (likely(test_bit(i, cast_ulp(cnt_pmds)))) {
+                               if (!(val & wmask)) {
+                                       __set_bit(i, cast_ulp(set->povfl_pmds));
+                                       set->npend_ovfls++;
+                               }
+                               val = (set->pmds[i].value & ~ovfl_mask) | (val & ovfl_mask);
+                       }
+                       set->pmds[i].value = val;
+                       count--;
+               }
+       }
+       /* 0 means: no need to save PMDs at upper level */
+       return 0;
+}
+
+/**
+ * pfm_p6_quiesce_pmu -- stop monitoring without grabbing any lock
+ *
+ * called from NMI interrupt handler to immediately stop monitoring
+ * cannot grab any lock, including perfmon related locks
+ */
+static void __kprobes pfm_p6_quiesce(void)
+{
+       /*
+        * quiesce PMU by clearing available registers that have
+        * the start/stop capability
+        *
+        * P6 processors only have enable bit on PERFEVTSEL0
+        */
+       if (test_bit(0, cast_ulp(pfm_pmu_conf->regs_all.pmcs)))
+               wrmsrl(MSR_P6_EVNTSEL0, 0);
+}
+
+/*
+ * Counters have 40 bits implemented. However they are designed such
+ * that bits [32-39] are sign extensions of bit 31. As such the
+ * effective width of a counter for P6-like PMU is 31 bits only.
+ *
+ * See IA-32 Intel Architecture Software developer manual Vol 3B
+ */
+static struct pfm_pmu_config pfm_p6_pmu_conf = {
+       .pmu_name = "Intel P6 processor Family",
+       .counter_width = 31,
+       .pmd_desc = pfm_p6_pmd_desc,
+       .pmc_desc = pfm_p6_pmc_desc,
+       .num_pmc_entries = PFM_P6_NUM_PMCS,
+       .num_pmd_entries = PFM_P6_NUM_PMDS,
+       .probe_pmu = pfm_p6_probe_pmu,
+       .version = "1.0",
+       .flags = PFM_PMU_BUILTIN_FLAG,
+       .owner = THIS_MODULE,
+       .pmu_info = &pfm_p6_pmu_info
+};
+
+static int __init pfm_p6_pmu_init_module(void)
+{
+       return pfm_pmu_register(&pfm_p6_pmu_conf);
+}
+
+static void __exit pfm_p6_pmu_cleanup_module(void)
+{
+       pfm_pmu_unregister(&pfm_p6_pmu_conf);
+}
+
+module_init(pfm_p6_pmu_init_module);
+module_exit(pfm_p6_pmu_cleanup_module);
diff --git a/arch/x86/perfmon/perfmon_pebs_core_smpl.c b/arch/x86/perfmon/perfmon_pebs_core_smpl.c
new file mode 100644
index 0000000..eeb9174
--- /dev/null
+++ b/arch/x86/perfmon/perfmon_pebs_core_smpl.c
@@ -0,0 +1,256 @@
+/*
+ * Copyright (c) 2005-2007 Hewlett-Packard Development Company, L.P.
+ * Contributed by Stephane Eranian <eranian@hpl.hp.com>
+ *
+ * This file implements the Precise Event Based Sampling (PEBS)
+ * sampling format for Intel Core and Atom  processors.
+ *
+ * This program is free software; you can redistribute it and/or
+ * modify it under the terms of version 2 of the GNU General Public
+ * License as published by the Free Software Foundation.
+ *
+ * This program is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+ * General Public License for more details.
+ *
+ * You should have received a copy of the GNU General Public License
+ * along with this program; if not, write to the Free Software
+ * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA
+ * 02111-1307 USA
+  */
+#include <linux/kernel.h>
+#include <linux/types.h>
+#include <linux/module.h>
+#include <linux/init.h>
+#include <linux/smp.h>
+#include <linux/perfmon_kern.h>
+
+#include <asm/msr.h>
+#include <asm/perfmon_pebs_core_smpl.h>
+
+MODULE_AUTHOR("Stephane Eranian <eranian@hpl.hp.com>");
+MODULE_DESCRIPTION("Intel Core Precise Event-Based Sampling (PEBS)");
+MODULE_LICENSE("GPL");
+
+#define ALIGN_PEBS(a, order) \
+               ((a)+(1UL<<(order))-1) & ~((1UL<<(order))-1)
+
+#define PEBS_PADDING_ORDER 8 /* log2(256) padding for PEBS alignment constraint */
+
+static int pfm_pebs_core_fmt_validate(u32 flags, u16 npmds, void *data)
+{
+       struct pfm_pebs_core_smpl_arg *arg = data;
+       size_t min_buf_size;
+
+       /*
+        * need to define at least the size of the buffer
+        */
+       if (data == NULL) {
+               PFM_DBG("no argument passed");
+               return -EINVAL;
+       }
+
+       /*
+        * compute min buf size. npmds is the maximum number
+        * of implemented PMD registers.
+        */
+       min_buf_size = sizeof(struct pfm_pebs_core_smpl_hdr)
+                    + sizeof(struct pfm_pebs_core_smpl_entry)
+                    + (1UL<<PEBS_PADDING_ORDER); /* padding for alignment */
+
+       PFM_DBG("validate flags=0x%x min_buf_size=%zu buf_size=%zu",
+                 flags,
+                 min_buf_size,
+                 arg->buf_size);
+
+       /*
+        * must hold at least the buffer header + one minimally sized entry
+        */
+       if (arg->buf_size < min_buf_size)
+               return -EINVAL;
+
+       return 0;
+}
+
+static int pfm_pebs_core_fmt_get_size(unsigned int flags, void *data, size_t *size)
+{
+       struct pfm_pebs_core_smpl_arg *arg = data;
+
+       /*
+        * size has been validated in pfm_pebs_core_fmt_validate()
+        */
+       *size = arg->buf_size + (1UL<<PEBS_PADDING_ORDER);
+
+       return 0;
+}
+
+static int pfm_pebs_core_fmt_init(struct pfm_context *ctx, void *buf,
+                            u32 flags, u16 npmds, void *data)
+{
+       struct pfm_arch_context *ctx_arch;
+       struct pfm_pebs_core_smpl_hdr *hdr;
+       struct pfm_pebs_core_smpl_arg *arg = data;
+       u64 pebs_start, pebs_end;
+       struct pfm_ds_area_core *ds;
+
+       ctx_arch = pfm_ctx_arch(ctx);
+
+       hdr = buf;
+       ds = &hdr->ds;
+
+       /*
+        * align PEBS buffer base
+        */
+       pebs_start = ALIGN_PEBS((unsigned long)(hdr+1), PEBS_PADDING_ORDER);
+       pebs_end = pebs_start + arg->buf_size + 1;
+
+       hdr->version = PFM_PEBS_CORE_SMPL_VERSION;
+       hdr->buf_size = arg->buf_size;
+       hdr->overflows = 0;
+
+       /*
+        * express PEBS buffer base as offset from the end of the header
+        */
+       hdr->start_offs = pebs_start - (unsigned long)(hdr+1);
+
+       /*
+        * PEBS buffer boundaries
+        */
+       ds->pebs_buf_base = pebs_start;
+       ds->pebs_abs_max = pebs_end;
+
+       /*
+        * PEBS starting position
+        */
+       ds->pebs_index = pebs_start;
+
+       /*
+        * PEBS interrupt threshold
+        */
+       ds->pebs_intr_thres = pebs_start
+                           + arg->intr_thres
+                           * sizeof(struct pfm_pebs_core_smpl_entry);
+
+       /*
+        * save counter reset value for PEBS counter
+        */
+       ds->pebs_cnt_reset = arg->cnt_reset;
+
+       /*
+        * keep track of DS AREA
+        */
+       ctx_arch->ds_area = ds;
+       ctx_arch->flags.use_ds = 1;
+       ctx_arch->flags.use_pebs = 1;
+
+       PFM_DBG("buffer=%p buf_size=%llu offs=%llu pebs_start=0x%llx "
+                 "pebs_end=0x%llx ds=%p pebs_thres=0x%llx cnt_reset=0x%llx",
+                 buf,
+                 (unsigned long long)hdr->buf_size,
+                 (unsigned long long)hdr->start_offs,
+                 (unsigned long long)pebs_start,
+                 (unsigned long long)pebs_end,
+                 ds,
+                 (unsigned long long)ds->pebs_intr_thres,
+                 (unsigned long long)ds->pebs_cnt_reset);
+
+       return 0;
+}
+
+static int pfm_pebs_core_fmt_handler(struct pfm_context *ctx,
+                              unsigned long ip, u64 tstamp, void *data)
+{
+       struct pfm_pebs_core_smpl_hdr *hdr;
+       struct pfm_ovfl_arg *arg;
+
+       hdr = ctx->smpl_addr;
+       arg = &ctx->ovfl_arg;
+
+       PFM_DBG_ovfl("buffer full");
+       /*
+        * increment number of buffer overflows.
+        * important to detect duplicate set of samples.
+        */
+       hdr->overflows++;
+
+       /*
+        * request notification and masking of monitoring.
+        * Notification is still subject to the overflowed
+        * register having the FL_NOTIFY flag set.
+        */
+       arg->ovfl_ctrl = PFM_OVFL_CTRL_NOTIFY | PFM_OVFL_CTRL_MASK;
+
+       return -ENOBUFS; /* we are full, sorry */
+}
+
+static int pfm_pebs_core_fmt_restart(int is_active, u32 *ovfl_ctrl,
+                               void *buf)
+{
+       struct pfm_pebs_core_smpl_hdr *hdr = buf;
+
+       /*
+        * reset index to base of buffer
+        */
+       hdr->ds.pebs_index = hdr->ds.pebs_buf_base;
+
+       *ovfl_ctrl = PFM_OVFL_CTRL_RESET;
+
+       return 0;
+}
+
+static int pfm_pebs_core_fmt_exit(void *buf)
+{
+       return 0;
+}
+
+static struct pfm_smpl_fmt pebs_core_fmt = {
+       .fmt_name = PFM_PEBS_CORE_SMPL_NAME,
+       .fmt_version = 0x1,
+       .fmt_arg_size = sizeof(struct pfm_pebs_core_smpl_arg),
+       .fmt_validate = pfm_pebs_core_fmt_validate,
+       .fmt_getsize = pfm_pebs_core_fmt_get_size,
+       .fmt_init = pfm_pebs_core_fmt_init,
+       .fmt_handler = pfm_pebs_core_fmt_handler,
+       .fmt_restart = pfm_pebs_core_fmt_restart,
+       .fmt_exit = pfm_pebs_core_fmt_exit,
+       .fmt_flags = PFM_FMT_BUILTIN_FLAG,
+       .owner = THIS_MODULE,
+};
+
+static int __init pfm_pebs_core_fmt_init_module(void)
+{
+       if (!cpu_has_pebs) {
+               PFM_INFO("processor does not have PEBS support");
+               return -1;
+       }
+       /*
+        * cpu_has_pebs is not enough to identify Intel Core PEBS
+        * which is different fro Pentium 4 PEBS. Therefore we do
+        * a more detailed check here
+        */
+       if (current_cpu_data.x86 != 6) {
+               PFM_INFO("not a supported Intel processor");
+               return -1;
+       }
+
+       switch (current_cpu_data.x86_model) {
+       case 15: /* Merom */
+       case 23: /* Penryn */
+       case 28: /* Atom (Silverthorne) */
+       case 29: /* Dunnington */
+               break;
+       default:
+               PFM_INFO("not a supported Intel processor");
+               return -1;
+       }
+       return pfm_fmt_register(&pebs_core_fmt);
+}
+
+static void __exit pfm_pebs_core_fmt_cleanup_module(void)
+{
+       pfm_fmt_unregister(&pebs_core_fmt);
+}
+
+module_init(pfm_pebs_core_fmt_init_module);
+module_exit(pfm_pebs_core_fmt_cleanup_module);
diff --git a/arch/x86/perfmon/perfmon_pebs_p4_smpl.c b/arch/x86/perfmon/perfmon_pebs_p4_smpl.c
new file mode 100644
index 0000000..f4e9fd2
--- /dev/null
+++ b/arch/x86/perfmon/perfmon_pebs_p4_smpl.c
@@ -0,0 +1,253 @@
+/*
+ * Copyright (c) 2005-2007 Hewlett-Packard Development Company, L.P.
+ * Contributed by Stephane Eranian <eranian@hpl.hp.com>
+ *
+ * This file implements the Precise Event Based Sampling (PEBS)
+ * sampling format. It supports the following processors:
+ *     - 32-bit Pentium 4 or other Netburst-based processors
+ *     - 64-bit Pentium 4 or other Netburst-based processors
+ *
+ * This program is free software; you can redistribute it and/or
+ * modify it under the terms of version 2 of the GNU General Public
+ * License as published by the Free Software Foundation.
+ *
+ * This program is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+ * General Public License for more details.
+ *
+ * You should have received a copy of the GNU General Public License
+ * along with this program; if not, write to the Free Software
+ * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA
+ * 02111-1307 USA
+  */
+#include <linux/kernel.h>
+#include <linux/types.h>
+#include <linux/module.h>
+#include <linux/init.h>
+#include <linux/smp.h>
+#include <linux/perfmon_kern.h>
+
+#include <asm/msr.h>
+#include <asm/perfmon_pebs_p4_smpl.h>
+
+MODULE_AUTHOR("Stephane Eranian <eranian@hpl.hp.com>");
+MODULE_DESCRIPTION("Intel P4 Precise Event-Based Sampling (PEBS)");
+MODULE_LICENSE("GPL");
+
+#define ALIGN_PEBS(a, order) \
+               ((a)+(1UL<<(order))-1) & ~((1UL<<(order))-1)
+
+#define PEBS_PADDING_ORDER 8 /* log2(256) padding for PEBS alignment constraint */
+
+static int pfm_pebs_p4_fmt_validate(u32 flags, u16 npmds, void *data)
+{
+       struct pfm_pebs_p4_smpl_arg *arg = data;
+       size_t min_buf_size;
+
+       /*
+        * need to define at least the size of the buffer
+        */
+       if (data == NULL) {
+               PFM_DBG("no argument passed");
+               return -EINVAL;
+       }
+
+       /*
+        * compute min buf size. npmds is the maximum number
+        * of implemented PMD registers.
+        */
+       min_buf_size = sizeof(struct pfm_pebs_p4_smpl_hdr)
+               + sizeof(struct pfm_pebs_p4_smpl_entry)
+               + (1UL<<PEBS_PADDING_ORDER); /* padding for alignment */
+
+       PFM_DBG("validate flags=0x%x min_buf_size=%zu buf_size=%zu",
+                 flags,
+                 min_buf_size,
+                 arg->buf_size);
+
+       /*
+        * must hold at least the buffer header + one minimally sized entry
+        */
+       if (arg->buf_size < min_buf_size)
+               return -EINVAL;
+
+       return 0;
+}
+
+static int pfm_pebs_p4_fmt_get_size(unsigned int flags, void *data, size_t *size)
+{
+       struct pfm_pebs_p4_smpl_arg *arg = data;
+
+       /*
+        * size has been validated in pfm_pebs_p4_fmt_validate()
+        */
+       *size = arg->buf_size + (1UL<<PEBS_PADDING_ORDER);
+
+       return 0;
+}
+
+static int pfm_pebs_p4_fmt_init(struct pfm_context *ctx, void *buf,
+                               u32 flags, u16 npmds, void *data)
+{
+       struct pfm_arch_context *ctx_arch;
+       struct pfm_pebs_p4_smpl_hdr *hdr;
+       struct pfm_pebs_p4_smpl_arg *arg = data;
+       unsigned long pebs_start, pebs_end;
+       struct pfm_ds_area_p4 *ds;
+
+       ctx_arch = pfm_ctx_arch(ctx);
+
+       hdr = buf;
+       ds = &hdr->ds;
+
+       /*
+        * align PEBS buffer base
+        */
+       pebs_start = ALIGN_PEBS((unsigned long)(hdr+1), PEBS_PADDING_ORDER);
+       pebs_end = pebs_start + arg->buf_size + 1;
+
+       hdr->version = PFM_PEBS_P4_SMPL_VERSION;
+       hdr->buf_size = arg->buf_size;
+       hdr->overflows = 0;
+
+       /*
+        * express PEBS buffer base as offset from the end of the header
+        */
+       hdr->start_offs = pebs_start - (unsigned long)(hdr+1);
+
+       /*
+        * PEBS buffer boundaries
+        */
+       ds->pebs_buf_base = pebs_start;
+       ds->pebs_abs_max = pebs_end;
+
+       /*
+        * PEBS starting position
+        */
+       ds->pebs_index = pebs_start;
+
+       /*
+        * PEBS interrupt threshold
+        */
+       ds->pebs_intr_thres = pebs_start
+               + arg->intr_thres * sizeof(struct pfm_pebs_p4_smpl_entry);
+
+       /*
+        * save counter reset value for PEBS counter
+        */
+       ds->pebs_cnt_reset = arg->cnt_reset;
+
+       /*
+        * keep track of DS AREA
+        */
+       ctx_arch->ds_area = ds;
+       ctx_arch->flags.use_pebs = 1;
+       ctx_arch->flags.use_ds = 1;
+
+       PFM_DBG("buffer=%p buf_size=%llu offs=%llu pebs_start=0x%lx "
+                 "pebs_end=0x%lx ds=%p pebs_thres=0x%lx cnt_reset=0x%llx",
+                 buf,
+                 (unsigned long long)hdr->buf_size,
+                 (unsigned long long)hdr->start_offs,
+                 pebs_start,
+                 pebs_end,
+                 ds,
+                 ds->pebs_intr_thres,
+                 (unsigned long long)ds->pebs_cnt_reset);
+
+       return 0;
+}
+
+static int pfm_pebs_p4_fmt_handler(struct pfm_context *ctx,
+                                  unsigned long ip, u64 tstamp, void *data)
+{
+       struct pfm_pebs_p4_smpl_hdr *hdr;
+       struct pfm_ovfl_arg *arg;
+
+       hdr = ctx->smpl_addr;
+       arg = &ctx->ovfl_arg;
+
+       PFM_DBG_ovfl("buffer full");
+       /*
+        * increment number of buffer overflows.
+        * important to detect duplicate set of samples.
+        */
+       hdr->overflows++;
+
+       /*
+        * request notification and masking of monitoring.
+        * Notification is still subject to the overflowed
+        * register having the FL_NOTIFY flag set.
+        */
+       arg->ovfl_ctrl = PFM_OVFL_CTRL_NOTIFY | PFM_OVFL_CTRL_MASK;
+
+       return -ENOBUFS; /* we are full, sorry */
+}
+
+static int pfm_pebs_p4_fmt_restart(int is_active, u32 *ovfl_ctrl,
+                               void *buf)
+{
+       struct pfm_pebs_p4_smpl_hdr *hdr = buf;
+
+       /*
+        * reset index to base of buffer
+        */
+       hdr->ds.pebs_index = hdr->ds.pebs_buf_base;
+
+       *ovfl_ctrl = PFM_OVFL_CTRL_RESET;
+
+       return 0;
+}
+
+static int pfm_pebs_p4_fmt_exit(void *buf)
+{
+       return 0;
+}
+
+static struct pfm_smpl_fmt pebs_p4_fmt = {
+       .fmt_name = PFM_PEBS_P4_SMPL_NAME,
+       .fmt_version = 0x1,
+       .fmt_arg_size = sizeof(struct pfm_pebs_p4_smpl_arg),
+       .fmt_validate = pfm_pebs_p4_fmt_validate,
+       .fmt_getsize = pfm_pebs_p4_fmt_get_size,
+       .fmt_init = pfm_pebs_p4_fmt_init,
+       .fmt_handler = pfm_pebs_p4_fmt_handler,
+       .fmt_restart = pfm_pebs_p4_fmt_restart,
+       .fmt_exit = pfm_pebs_p4_fmt_exit,
+       .fmt_flags = PFM_FMT_BUILTIN_FLAG,
+       .owner = THIS_MODULE,
+};
+
+static int __init pfm_pebs_p4_fmt_init_module(void)
+{
+       int ht_enabled;
+
+       if (!cpu_has_pebs) {
+               PFM_INFO("processor does not have PEBS support");
+               return -1;
+       }
+       if (current_cpu_data.x86 != 15) {
+               PFM_INFO("not an Intel Pentium 4");
+               return -1;
+       }
+#ifdef CONFIG_SMP
+       ht_enabled = (cpus_weight(__get_cpu_var(cpu_core_map))
+                  / current_cpu_data.x86_max_cores) > 1;
+#else
+       ht_enabled = 0;
+#endif
+       if (ht_enabled) {
+               PFM_INFO("PEBS not available because HyperThreading is on");
+               return -1;
+       }
+       return pfm_fmt_register(&pebs_p4_fmt);
+}
+
+static void __exit pfm_pebs_p4_fmt_cleanup_module(void)
+{
+       pfm_fmt_unregister(&pebs_p4_fmt);
+}
+
+module_init(pfm_pebs_p4_fmt_init_module);
+module_exit(pfm_pebs_p4_fmt_cleanup_module);
diff --git a/include/asm-mips/Kbuild b/include/asm-mips/Kbuild
index 7897f05..7ed16fc 100644
--- a/include/asm-mips/Kbuild
+++ b/include/asm-mips/Kbuild
@@ -1,3 +1,4 @@
 include include/asm-generic/Kbuild.asm
 
 header-y += cachectl.h sgidefs.h sysmips.h
+header-y += perfmon.h
diff --git a/include/asm-mips/perfmon.h b/include/asm-mips/perfmon.h
new file mode 100644
index 0000000..7915c17
--- /dev/null
+++ b/include/asm-mips/perfmon.h
@@ -0,0 +1,34 @@
+/*
+ * Copyright (c) 2007 Hewlett-Packard Development Company, L.P.
+ * Contributed by Stephane Eranian <eranian@hpl.hp.com>
+ *
+ * This file contains mips64 specific definitions for the perfmon
+ * interface.
+ *
+ * This file MUST never be included directly. Use linux/perfmon.h.
+ *
+ * This program is free software; you can redistribute it and/or
+ * modify it under the terms of version 2 of the GNU General Public
+ * License as published by the Free Software Foundation.
+ *
+ * This program is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+ * General Public License for more details.
+ *
+ * You should have received a copy of the GNU General Public License
+ * along with this program; if not, write to the Free Software
+ * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA
+ * 02111-1307 USA
+  */
+#ifndef _ASM_MIPS64_PERFMON_H_
+#define _ASM_MIPS64_PERFMON_H_
+
+/*
+ * arch-specific user visible interface definitions
+ */
+
+#define PFM_ARCH_MAX_PMCS      (256+64) /* 256 HW 64 SW */
+#define PFM_ARCH_MAX_PMDS      (256+64) /* 256 HW 64 SW */
+
+#endif /* _ASM_MIPS64_PERFMON_H_ */
diff --git a/include/asm-mips/perfmon_kern.h b/include/asm-mips/perfmon_kern.h
new file mode 100644
index 0000000..7d213df
--- /dev/null
+++ b/include/asm-mips/perfmon_kern.h
@@ -0,0 +1,412 @@
+/*
+ * Copyright (c) 2005 Philip Mucci.
+ *
+ * Based on other versions:
+ * Copyright (c) 2005-2006 Hewlett-Packard Development Company, L.P.
+ * Contributed by Stephane Eranian <eranian@hpl.hp.com>
+ *
+ * This file contains mips64 specific definitions for the perfmon
+ * interface.
+ *
+ * This program is free software; you can redistribute it and/or
+ * modify it under the terms of version 2 of the GNU General Public
+ * License as published by the Free Software Foundation.
+ *
+ * This program is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+ * General Public License for more details.
+ *
+ * You should have received a copy of the GNU General Public License
+ * along with this program; if not, write to the Free Software
+ * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA
+ * 02111-1307 USA
+  */
+#ifndef _ASM_MIPS64_PERFMON_KERN_H_
+#define _ASM_MIPS64_PERFMON_KERN_H_
+
+#ifdef __KERNEL__
+
+#ifdef CONFIG_PERFMON
+#include <linux/unistd.h>
+#include <asm/cacheflush.h>
+
+#define PFM_ARCH_PMD_STK_ARG   2
+#define PFM_ARCH_PMC_STK_ARG   2
+
+struct pfm_arch_pmu_info {
+       u32 pmu_style;
+};
+
+#define MIPS64_CONFIG_PMC_MASK (1 << 4)
+#define MIPS64_PMC_INT_ENABLE_MASK (1 << 4)
+#define MIPS64_PMC_CNT_ENABLE_MASK (0xf)
+#define MIPS64_PMC_EVT_MASK (0x7 << 6)
+#define MIPS64_PMC_CTR_MASK (1 << 31)
+#define MIPS64_PMD_INTERRUPT (1 << 31)
+
+/* Coprocessor register 25 contains the PMU interface. */
+/* Sel 0 is control for counter 0 */
+/* Sel 1 is count for counter 0. */
+/* Sel 2 is control for counter 1. */
+/* Sel 3 is count for counter 1. */
+
+/*
+
+31 30 29 28 27 26 25 24 23 22 21 20 19 18 17 16 15 14 13 12 11 10 9 8 7 6 5 4  3 2 1 0
+M  0--------------------------------------------------------------0 Event-- IE U S K EXL
+
+M 31 If this bit is one, another pair of Performance Control
+and Counter registers is implemented at a MTC0
+
+Event 8:5 Counter event enabled for this counter. Possible events
+are listed in Table 6-30. R/W Undefined
+
+IE 4 Counter Interrupt Enable. This bit masks bit 31 of the
+associated count register from the interrupt exception
+request output. R/W 0
+
+U 3 Count in User Mode. When this bit is set, the specified
+event is counted in User Mode. R/W Undefined
+
+S 2 Count in Supervisor Mode. When this bit is set, the
+specified event is counted in Supervisor Mode. R/W Undefined
+
+K 1 Count in Kernel Mode. When this bit is set, count the
+event in Kernel Mode when EXL and ERL both are 0. R/W Undefined
+
+EXL 0 Count when EXL. When this bit is set, count the event
+when EXL = 1 and ERL = 0. R/W Undefined
+*/
+
+static inline void pfm_arch_resend_irq(struct pfm_context *ctx)
+{}
+
+static inline void pfm_arch_clear_pmd_ovfl_cond(struct pfm_context *ctx,
+                                               struct pfm_event_set *set)
+{}
+
+static inline void pfm_arch_serialize(void)
+{}
+
+
+/*
+ * MIPS does not save the PMDs during pfm_arch_intr_freeze_pmu(), thus
+ * this routine needs to do it when switching sets on overflow
+ */
+static inline void pfm_arch_save_pmds_from_intr(struct pfm_context *ctx,
+                                               struct pfm_event_set *set)
+{
+       pfm_save_pmds(ctx, set);
+}
+
+static inline void pfm_arch_write_pmc(struct pfm_context *ctx,
+                                     unsigned int cnum, u64 value)
+{
+       /*
+        * we only write to the actual register when monitoring is
+        * active (pfm_start was issued)
+        */
+       if (ctx && (ctx->flags.started == 0))
+               return;
+
+       switch (pfm_pmu_conf->pmc_desc[cnum].hw_addr) {
+       case 0:
+               write_c0_perfctrl0(value);
+               break;
+       case 1:
+               write_c0_perfctrl1(value);
+               break;
+       case 2:
+               write_c0_perfctrl2(value);
+               break;
+       case 3:
+               write_c0_perfctrl3(value);
+               break;
+       default:
+               BUG();
+       }
+}
+
+static inline void pfm_arch_write_pmd(struct pfm_context *ctx,
+                                     unsigned int cnum, u64 value)
+{
+       value &= pfm_pmu_conf->ovfl_mask;
+
+       switch (pfm_pmu_conf->pmd_desc[cnum].hw_addr) {
+       case 0:
+               write_c0_perfcntr0(value);
+               break;
+       case 1:
+               write_c0_perfcntr1(value);
+               break;
+       case 2:
+               write_c0_perfcntr2(value);
+               break;
+       case 3:
+               write_c0_perfcntr3(value);
+               break;
+       default:
+               BUG();
+       }
+}
+
+static inline u64 pfm_arch_read_pmd(struct pfm_context *ctx, unsigned int cnum)
+{
+       switch (pfm_pmu_conf->pmd_desc[cnum].hw_addr) {
+       case 0:
+               return read_c0_perfcntr0();
+               break;
+       case 1:
+               return read_c0_perfcntr1();
+               break;
+       case 2:
+               return read_c0_perfcntr2();
+               break;
+       case 3:
+               return read_c0_perfcntr3();
+               break;
+       default:
+               BUG();
+               return 0;
+       }
+}
+
+static inline u64 pfm_arch_read_pmc(struct pfm_context *ctx, unsigned int cnum)
+{
+       switch (pfm_pmu_conf->pmc_desc[cnum].hw_addr) {
+       case 0:
+               return read_c0_perfctrl0();
+               break;
+       case 1:
+               return read_c0_perfctrl1();
+               break;
+       case 2:
+               return read_c0_perfctrl2();
+               break;
+       case 3:
+               return read_c0_perfctrl3();
+               break;
+       default:
+               BUG();
+               return 0;
+       }
+}
+
+/*
+ * For some CPUs, the upper bits of a counter must be set in order for the
+ * overflow interrupt to happen. On overflow, the counter has wrapped around,
+ * and the upper bits are cleared. This function may be used to set them back.
+ */
+static inline void pfm_arch_ovfl_reset_pmd(struct pfm_context *ctx,
+                                          unsigned int cnum)
+{
+  u64 val;
+  val = pfm_arch_read_pmd(ctx, cnum);
+  /* This masks out overflow bit 31 */
+  pfm_arch_write_pmd(ctx, cnum, val);
+}
+
+/*
+ * At certain points, perfmon needs to know if monitoring has been
+ * explicitely started/stopped by user via pfm_start/pfm_stop. The
+ * information is tracked in ctx.flags.started. However on certain
+ * architectures, it may be possible to start/stop directly from
+ * user level with a single assembly instruction bypassing
+ * the kernel. This function must be used to determine by
+ * an arch-specific mean if monitoring is actually started/stopped.
+ */
+static inline int pfm_arch_is_active(struct pfm_context *ctx)
+{
+       return ctx->flags.started;
+}
+
+static inline void pfm_arch_ctxswout_sys(struct task_struct *task,
+                                        struct pfm_context *ctx)
+{}
+
+static inline void pfm_arch_ctxswin_sys(struct task_struct *task,
+                                       struct pfm_context *ctx)
+{}
+
+static inline void pfm_arch_ctxswin_thread(struct task_struct *task,
+                                          struct pfm_context *ctx)
+{}
+int  pfm_arch_ctxswout_thread(struct task_struct *task,
+                             struct pfm_context *ctx);
+
+int  pfm_arch_is_monitoring_active(struct pfm_context *ctx);
+void pfm_arch_stop(struct task_struct *task, struct pfm_context *ctx);
+void pfm_arch_start(struct task_struct *task, struct pfm_context *ctx);
+void pfm_arch_restore_pmds(struct pfm_context *ctx, struct pfm_event_set *set);
+void pfm_arch_restore_pmcs(struct pfm_context *ctx, struct pfm_event_set *set);
+char *pfm_arch_get_pmu_module_name(void);
+
+static inline void pfm_arch_intr_freeze_pmu(struct pfm_context *ctx,
+                                           struct pfm_event_set *set)
+{
+       pfm_arch_stop(current, ctx);
+       /*
+        * we mark monitoring as stopped to avoid
+        * certain side effects especially in
+        * pfm_switch_sets_from_intr() on
+        * pfm_arch_restore_pmcs()
+        */
+       ctx->flags.started = 0;
+}
+
+/*
+ * unfreeze PMU from pfm_do_interrupt_handler()
+ * ctx may be NULL for spurious
+ */
+static inline void pfm_arch_intr_unfreeze_pmu(struct pfm_context *ctx)
+{
+       if (!ctx)
+               return;
+
+       PFM_DBG_ovfl("state=%d", ctx->state);
+
+       ctx->flags.started = 1;
+
+       if (ctx->state == PFM_CTX_MASKED)
+               return;
+
+       pfm_arch_restore_pmcs(ctx, ctx->active_set);
+}
+
+/*
+ * this function is called from the PMU interrupt handler ONLY.
+ * On MIPS, the PMU is frozen via arch_stop, masking would be implemented
+ * via arch-stop as well. Given that the PMU is already stopped when
+ * entering the interrupt handler, we do not need to stop it again, so
+ * this function is a nop.
+ */
+static inline void pfm_arch_mask_monitoring(struct pfm_context *ctx,
+                                           struct pfm_event_set *set)
+{}
+
+/*
+ * on MIPS masking/unmasking uses the start/stop mechanism, so we simply
+ * need to start here.
+ */
+static inline void pfm_arch_unmask_monitoring(struct pfm_context *ctx,
+                                             struct pfm_event_set *set)
+{
+       pfm_arch_start(current, ctx);
+}
+
+static inline int pfm_arch_context_create(struct pfm_context *ctx,
+                                         u32 ctx_flags)
+{
+       return 0;
+}
+
+static inline void pfm_arch_context_free(struct pfm_context *ctx)
+{}
+
+
+
+
+
+/*
+ * function called from pfm_setfl_sane(). Context is locked
+ * and interrupts are masked.
+ * The value of flags is the value of ctx_flags as passed by
+ * user.
+ *
+ * function must check arch-specific set flags.
+ * Return:
+ *     1 when flags are valid
+ *      0 on error
+ */
+static inline int
+pfm_arch_setfl_sane(struct pfm_context *ctx, u32 flags)
+{
+       return 0;
+}
+
+static inline int pfm_arch_init(void)
+{
+       return 0;
+}
+
+static inline void pfm_arch_init_percpu(void)
+{}
+
+static inline int pfm_arch_load_context(struct pfm_context *ctx)
+{
+       return 0;
+}
+
+static inline void pfm_arch_unload_context(struct pfm_context *ctx)
+{}
+
+static inline int pfm_arch_pmu_acquire(u64 *unavail_pmcs, u64 *unavail_pmds)
+{
+       return 0;
+}
+
+static inline void pfm_arch_pmu_release(void)
+{}
+
+#ifdef CONFIG_PERFMON_FLUSH
+/*
+ * due to cache aliasing problem on MIPS, it is necessary to flush
+ * pages out of the cache when they are modified.
+ */
+static inline void pfm_cacheflush(void *addr, unsigned int len)
+{
+       unsigned long start, end;
+
+       start = (unsigned long)addr & PAGE_MASK;
+       end = ((unsigned long)addr + len + PAGE_SIZE - 1) & PAGE_MASK;
+
+       while (start < end) {
+               flush_data_cache_page(start);
+               start += PAGE_SIZE;
+       }
+}
+#else
+static inline void pfm_cacheflush(void *addr, unsigned int len)
+{}
+#endif
+
+static inline void pfm_arch_arm_handle_work(struct task_struct *task)
+{}
+
+static inline void pfm_arch_disarm_handle_work(struct task_struct *task)
+{}
+
+static inline int pfm_arch_pmu_config_init(struct pfm_pmu_config *cfg)
+{
+       return 0;
+}
+
+static inline int pfm_arch_get_base_syscall(void)
+{
+       if (test_thread_flag(TIF_32BIT_ADDR)) {
+               if (test_thread_flag(TIF_32BIT_REGS))
+                       return __NR_O32_Linux+330;
+               return __NR_N32_Linux+293;
+       }
+       return __NR_64_Linux+289;
+}
+
+struct pfm_arch_context {
+       /* empty */
+};
+
+#define PFM_ARCH_CTX_SIZE      sizeof(struct pfm_arch_context)
+/*
+ * MIPS may need extra alignment requirements for the sampling buffer
+ */
+#ifdef CONFIG_PERFMON_SMPL_ALIGN
+#define PFM_ARCH_SMPL_ALIGN_SIZE       0x4000
+#else
+#define PFM_ARCH_SMPL_ALIGN_SIZE       0
+#endif
+
+#endif /* CONFIG_PERFMON */
+
+#endif /* __KERNEL__ */
+#endif /* _ASM_MIPS64_PERFMON_KERN_H_ */
diff --git a/include/asm-mips/system.h b/include/asm-mips/system.h
index a944eda..470cdfc 100644
--- a/include/asm-mips/system.h
+++ b/include/asm-mips/system.h
@@ -67,6 +67,10 @@ do {                                                                 \
        __mips_mt_fpaff_switch_to(prev);                                \
        if (cpu_has_dsp)                                                \
                __save_dsp(prev);                                       \
+       if (test_tsk_thread_flag(prev, TIF_PERFMON_CTXSW))              \
+               pfm_ctxsw_out(prev, next);                              \
+       if (test_tsk_thread_flag(next, TIF_PERFMON_CTXSW))              \
+               pfm_ctxsw_in(prev, next);                               \
        (last) = resume(prev, next, task_thread_info(next));            \
 } while (0)
 
diff --git a/include/asm-mips/thread_info.h b/include/asm-mips/thread_info.h
index bb30606..34fd6aa 100644
--- a/include/asm-mips/thread_info.h
+++ b/include/asm-mips/thread_info.h
@@ -114,6 +114,7 @@ register struct thread_info *__current_thread_info __asm__("$28");
 #define TIF_NEED_RESCHED       2       /* rescheduling necessary */
 #define TIF_SYSCALL_AUDIT      3       /* syscall auditing active */
 #define TIF_SECCOMP            4       /* secure computing */
+#define TIF_PERFMON_WORK       5       /* work for pfm_handle_work() */
 #define TIF_RESTORE_SIGMASK    9       /* restore signal mask in do_signal() */
 #define TIF_USEDFPU            16      /* FPU was used by this task this quantum (SMP) */
 #define TIF_POLLING_NRFLAG     17      /* true if poll_idle() is polling TIF_NEED_RESCHED */
@@ -124,6 +125,7 @@ register struct thread_info *__current_thread_info __asm__("$28");
 #define TIF_32BIT_REGS         22      /* also implies 16/32 fprs */
 #define TIF_32BIT_ADDR         23      /* 32-bit address space (o32/n32) */
 #define TIF_FPUBOUND           24      /* thread bound to FPU-full CPU set */
+#define TIF_PERFMON_CTXSW      25      /* perfmon needs ctxsw calls */
 #define TIF_SYSCALL_TRACE      31      /* syscall trace active */
 
 #define _TIF_SYSCALL_TRACE     (1<<TIF_SYSCALL_TRACE)
@@ -140,6 +142,8 @@ register struct thread_info *__current_thread_info __asm__("$28");
 #define _TIF_32BIT_REGS                (1<<TIF_32BIT_REGS)
 #define _TIF_32BIT_ADDR                (1<<TIF_32BIT_ADDR)
 #define _TIF_FPUBOUND          (1<<TIF_FPUBOUND)
+#define _TIF_PERFMON_WORK      (1<<TIF_PERFMON_WORK)
+#define _TIF_PERFMON_CTXSW     (1<<TIF_PERFMON_CTXSW)
 
 /* work to do on interrupt/exception return */
 #define _TIF_WORK_MASK         (0x0000ffef & ~_TIF_SECCOMP)
diff --git a/include/asm-mips/unistd.h b/include/asm-mips/unistd.h
index a73e153..200f654 100644
--- a/include/asm-mips/unistd.h
+++ b/include/asm-mips/unistd.h
@@ -350,11 +350,23 @@
 #define __NR_dup3                      (__NR_Linux + 327)
 #define __NR_pipe2                     (__NR_Linux + 328)
 #define __NR_inotify_init1             (__NR_Linux + 329)
+#define __NR_pfm_create_context         (__NR_Linux + 330)
+#define __NR_pfm_write_pmcs            (__NR_pfm_create_context+1)
+#define __NR_pfm_write_pmds            (__NR_pfm_create_context+2)
+#define __NR_pfm_read_pmds             (__NR_pfm_create_context+3)
+#define __NR_pfm_load_context          (__NR_pfm_create_context+4)
+#define __NR_pfm_start                 (__NR_pfm_create_context+5)
+#define __NR_pfm_stop                  (__NR_pfm_create_context+6)
+#define __NR_pfm_restart               (__NR_pfm_create_context+7)
+#define __NR_pfm_create_evtsets                (__NR_pfm_create_context+8)
+#define __NR_pfm_getinfo_evtsets       (__NR_pfm_create_context+9)
+#define __NR_pfm_delete_evtsets        (__NR_pfm_create_context+10)
+#define __NR_pfm_unload_context                (__NR_pfm_create_context+11)
 
 /*
  * Offset of the last Linux o32 flavoured syscall
  */
-#define __NR_Linux_syscalls            329
+#define __NR_Linux_syscalls            341
 
 #endif /* _MIPS_SIM == _MIPS_SIM_ABI32 */
 
@@ -656,16 +668,28 @@
 #define __NR_dup3                      (__NR_Linux + 286)
 #define __NR_pipe2                     (__NR_Linux + 287)
 #define __NR_inotify_init1             (__NR_Linux + 288)
+#define __NR_pfm_create_context         (__NR_Linux + 289)
+#define __NR_pfm_write_pmcs            (__NR_pfm_create_context+1)
+#define __NR_pfm_write_pmds            (__NR_pfm_create_context+2)
+#define __NR_pfm_read_pmds             (__NR_pfm_create_context+3)
+#define __NR_pfm_load_context          (__NR_pfm_create_context+4)
+#define __NR_pfm_start                 (__NR_pfm_create_context+5)
+#define __NR_pfm_stop                  (__NR_pfm_create_context+6)
+#define __NR_pfm_restart               (__NR_pfm_create_context+7)
+#define __NR_pfm_create_evtsets                (__NR_pfm_create_context+8)
+#define __NR_pfm_getinfo_evtsets       (__NR_pfm_create_context+9)
+#define __NR_pfm_delete_evtsets        (__NR_pfm_create_context+10)
+#define __NR_pfm_unload_context                (__NR_pfm_create_context+11)
 
 /*
  * Offset of the last Linux 64-bit flavoured syscall
  */
-#define __NR_Linux_syscalls            288
+#define __NR_Linux_syscalls            300
 
 #endif /* _MIPS_SIM == _MIPS_SIM_ABI64 */
 
 #define __NR_64_Linux                  5000
-#define __NR_64_Linux_syscalls         288
+#define __NR_64_Linux_syscalls         300
 
 #if _MIPS_SIM == _MIPS_SIM_NABI32
 
@@ -966,16 +990,28 @@
 #define __NR_dup3                      (__NR_Linux + 290)
 #define __NR_pipe2                     (__NR_Linux + 291)
 #define __NR_inotify_init1             (__NR_Linux + 292)
+#define __NR_pfm_create_context         (__NR_Linux + 293)
+#define __NR_pfm_write_pmcs            (__NR_pfm_create_context+1)
+#define __NR_pfm_write_pmds            (__NR_pfm_create_context+2)
+#define __NR_pfm_read_pmds             (__NR_pfm_create_context+3)
+#define __NR_pfm_load_context          (__NR_pfm_create_context+4)
+#define __NR_pfm_start                 (__NR_pfm_create_context+5)
+#define __NR_pfm_stop                  (__NR_pfm_create_context+6)
+#define __NR_pfm_restart               (__NR_pfm_create_context+7)
+#define __NR_pfm_create_evtsets                (__NR_pfm_create_context+8)
+#define __NR_pfm_getinfo_evtsets       (__NR_pfm_create_context+9)
+#define __NR_pfm_delete_evtsets        (__NR_pfm_create_context+10)
+#define __NR_pfm_unload_context                (__NR_pfm_create_context+11)
 
 /*
  * Offset of the last N32 flavoured syscall
  */
-#define __NR_Linux_syscalls            292
+#define __NR_Linux_syscalls            304
 
 #endif /* _MIPS_SIM == _MIPS_SIM_NABI32 */
 
 #define __NR_N32_Linux                 6000
-#define __NR_N32_Linux_syscalls                292
+#define __NR_N32_Linux_syscalls                304
 
 #ifdef __KERNEL__
 
diff --git a/include/asm-x86/Kbuild b/include/asm-x86/Kbuild
index 4a8e80c..d7d819e 100644
--- a/include/asm-x86/Kbuild
+++ b/include/asm-x86/Kbuild
@@ -9,6 +9,7 @@ header-y += prctl.h
 header-y += ptrace-abi.h
 header-y += sigcontext32.h
 header-y += ucontext.h
+header-y += perfmon.h
 header-y += processor-flags.h
 
 unifdef-y += e820.h
diff --git a/include/asm-x86/ia32_unistd.h b/include/asm-x86/ia32_unistd.h
index 61cea9e..275e015 100644
--- a/include/asm-x86/ia32_unistd.h
+++ b/include/asm-x86/ia32_unistd.h
@@ -8,11 +8,12 @@
  * the number. This should be otherwise in sync with asm-x86/unistd_32.h. -AK
  */
 
-#define __NR_ia32_restart_syscall 0
-#define __NR_ia32_exit           1
-#define __NR_ia32_read           3
-#define __NR_ia32_write                  4
-#define __NR_ia32_sigreturn    119
-#define __NR_ia32_rt_sigreturn 173
+#define __NR_ia32_restart_syscall    0
+#define __NR_ia32_exit              1
+#define __NR_ia32_read              3
+#define __NR_ia32_write                     4
+#define __NR_ia32_sigreturn         119
+#define __NR_ia32_rt_sigreturn      173
+#define __NR_ia32_pfm_create_context 333
 
 #endif /* _ASM_X86_64_IA32_UNISTD_H_ */
diff --git a/include/asm-x86/irq_vectors.h b/include/asm-x86/irq_vectors.h
index a48c7f2..892fe8f 100644
--- a/include/asm-x86/irq_vectors.h
+++ b/include/asm-x86/irq_vectors.h
@@ -92,6 +92,11 @@
 #define LOCAL_TIMER_VECTOR     0xef
 
 /*
+ * Perfmon PMU interrupt vector
+ */
+#define LOCAL_PERFMON_VECTOR   0xee
+
+/*
  * First APIC vector available to drivers: (vectors 0x30-0xee) we
  * start at 0x31(0x41) to spread out vectors evenly between priority
  * levels. (0x80 is the syscall vector)
diff --git a/include/asm-x86/mach-default/entry_arch.h b/include/asm-x86/mach-default/entry_arch.h
index 9283b60..ac31c2d 100644
--- a/include/asm-x86/mach-default/entry_arch.h
+++ b/include/asm-x86/mach-default/entry_arch.h
@@ -32,4 +32,8 @@ BUILD_INTERRUPT(spurious_interrupt,SPURIOUS_APIC_VECTOR)
 BUILD_INTERRUPT(thermal_interrupt,THERMAL_APIC_VECTOR)
 #endif
 
+#ifdef CONFIG_PERFMON
+BUILD_INTERRUPT(pmu_interrupt,LOCAL_PERFMON_VECTOR)
+#endif
+
 #endif
diff --git a/include/asm-x86/perfmon.h b/include/asm-x86/perfmon.h
new file mode 100644
index 0000000..906f4b2
--- /dev/null
+++ b/include/asm-x86/perfmon.h
@@ -0,0 +1,34 @@
+/*
+ * Copyright (c) 2007 Hewlett-Packard Development Company, L.P.
+ * Contributed by Stephane Eranian <eranian@hpl.hp.com>
+ *
+ * This file contains i386/x86_64 specific definitions for the perfmon
+ * interface.
+ *
+ * This file MUST never be included directly. Use linux/perfmon.h.
+ *
+ * This program is free software; you can redistribute it and/or
+ * modify it under the terms of version 2 of the GNU General Public
+ * License as published by the Free Software Foundation.
+ *
+ * This program is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+ * General Public License for more details.
+ *
+ * You should have received a copy of the GNU General Public License
+ * along with this program; if not, write to the Free Software
+ * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA
+ * 02111-1307 USA
+  */
+#ifndef _ASM_X86_PERFMON__H_
+#define _ASM_X86_PERFMON__H_
+
+/*
+ * arch-specific user visible interface definitions
+ */
+
+#define PFM_ARCH_MAX_PMCS      (256+64) /* 256 HW 64 SW */
+#define PFM_ARCH_MAX_PMDS      (256+64) /* 256 HW 64 SW */
+
+#endif /* _ASM_X86_PERFMON_H_ */
diff --git a/include/asm-x86/perfmon_kern.h b/include/asm-x86/perfmon_kern.h
new file mode 100644
index 0000000..0e5d3a5
--- /dev/null
+++ b/include/asm-x86/perfmon_kern.h
@@ -0,0 +1,548 @@
+/*
+ * Copyright (c) 2005-2006 Hewlett-Packard Development Company, L.P.
+ * Contributed by Stephane Eranian <eranian@hpl.hp.com>
+ *
+ * Copyright (c) 2007 Advanced Micro Devices, Inc.
+ * Contributed by Robert Richter <robert.richter@amd.com>
+ *
+ * This file contains X86 Processor Family specific definitions
+ * for the perfmon interface. This covers P6, Pentium M, P4/Xeon
+ * (32-bit and 64-bit, i.e., EM64T) and AMD X86-64.
+ *
+ * This program is free software; you can redistribute it and/or
+ * modify it under the terms of version 2 of the GNU General Public
+ * License as published by the Free Software Foundation.
+ *
+ * This program is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+ * General Public License for more details.
+ *
+ * You should have received a copy of the GNU General Public License
+ * along with this program; if not, write to the Free Software
+ * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA
+ * 02111-1307 USA
+  */
+#ifndef _ASM_X86_PERFMON_KERN_H_
+#define _ASM_X86_PERFMON_KERN_H_
+
+#ifdef CONFIG_PERFMON
+#include <linux/unistd.h>
+#ifdef CONFIG_4KSTACKS
+#define PFM_ARCH_PMD_STK_ARG   2
+#define PFM_ARCH_PMC_STK_ARG   2
+#else
+#define PFM_ARCH_PMD_STK_ARG   4 /* about 700 bytes of stack space */
+#define PFM_ARCH_PMC_STK_ARG   4 /* about 200 bytes of stack space */
+#endif
+
+struct pfm_arch_pmu_info {
+       u32 flags;              /* PMU feature flags */
+       /*
+        * mandatory model-specific callbacks
+        */
+       int  (*stop_save)(struct pfm_context *ctx, struct pfm_event_set *set);
+       int  (*has_ovfls)(struct pfm_context *ctx);
+       void (*quiesce)(void);
+
+       /*
+        * optional model-specific callbacks
+        */
+       void (*acquire_pmu_percpu)(void);
+       void (*release_pmu_percpu)(void);
+       int (*create_context)(struct pfm_context *ctx, u32 ctx_flags);
+       void (*free_context)(struct pfm_context *ctx);
+       int (*load_context)(struct pfm_context *ctx);
+       void (*unload_context)(struct pfm_context *ctx);
+       void (*write_pmc)(struct pfm_context *ctx, unsigned int cnum, u64 value);
+       void (*write_pmd)(struct pfm_context *ctx, unsigned int cnum, u64 value);
+       u64  (*read_pmd)(struct pfm_context *ctx, unsigned int cnum);
+       u64  (*read_pmc)(struct pfm_context *ctx, unsigned int cnum);
+       void (*nmi_copy_state)(struct pfm_context *ctx);
+       void (*restore_pmcs)(struct pfm_context *ctx,
+                            struct pfm_event_set *set);
+       void (*restore_pmds)(struct pfm_context *ctx,
+                            struct pfm_event_set *set);
+};
+
+/*
+ * PMU feature flags
+ */
+#define PFM_X86_FL_USE_NMI     0x01    /* user asking for NMI */
+#define PFM_X86_FL_NO_SHARING  0x02    /* no sharing with other subsystems */
+#define PFM_X86_FL_SHARING     0x04    /* PMU is being shared */
+
+struct pfm_x86_ctx_flags {
+       unsigned int insecure:1;  /* rdpmc per-thread self-monitoring */
+       unsigned int use_pebs:1;  /* PEBS used */
+       unsigned int use_ds:1;    /* DS used */
+       unsigned int reserved:29; /* for future use */
+};
+
+struct pfm_arch_context {
+       u64 saved_real_iip;             /* instr pointer of last NMI intr */
+       struct pfm_x86_ctx_flags flags; /* flags */
+       void *ds_area;                  /* address of DS area (to go away) */
+       void *data;                     /* model-specific data */
+};
+
+/*
+ * functions implemented as inline on x86
+ */
+
+/**
+ * pfm_arch_write_pmc - write a single PMC register
+ * @ctx: context to work on
+ * @cnum: PMC index
+ * @value: PMC 64-bit value
+ *
+ * in certain situations, ctx may be NULL
+ */
+static inline void pfm_arch_write_pmc(struct pfm_context *ctx,
+                                     unsigned int cnum, u64 value)
+{
+       struct pfm_arch_pmu_info *pmu_info;
+
+       pmu_info = pfm_pmu_info();
+
+       /*
+        * we only write to the actual register when monitoring is
+        * active (pfm_start was issued)
+        */
+       if (ctx && ctx->flags.started == 0)
+               return;
+
+       /*
+        * model-specific override, if any
+        */
+       if (pmu_info->write_pmc) {
+               pmu_info->write_pmc(ctx, cnum, value);
+               return;
+       }
+
+       PFM_DBG_ovfl("pfm_arch_write_pmc(0x%lx, 0x%Lx)",
+                    pfm_pmu_conf->pmc_desc[cnum].hw_addr,
+                    (unsigned long long) value);
+
+       wrmsrl(pfm_pmu_conf->pmc_desc[cnum].hw_addr, value);
+}
+
+/**
+ * pfm_arch_write_pmd - write a single PMD register
+ * @ctx: context to work on
+ * @cnum: PMD index
+ * @value: PMD 64-bit value
+ */
+static inline void pfm_arch_write_pmd(struct pfm_context *ctx,
+                                     unsigned int cnum, u64 value)
+{
+       struct pfm_arch_pmu_info *pmu_info;
+
+       pmu_info = pfm_pmu_info();
+
+       /*
+        * to make sure the counter overflows, we set the
+        * upper bits. we also clear any other unimplemented
+        * bits as this may cause crash on some processors.
+        */
+       if (pfm_pmu_conf->pmd_desc[cnum].type & PFM_REG_C64)
+               value = (value | ~pfm_pmu_conf->ovfl_mask)
+                     & ~pfm_pmu_conf->pmd_desc[cnum].rsvd_msk;
+
+       PFM_DBG_ovfl("pfm_arch_write_pmd(0x%lx, 0x%Lx)",
+                    pfm_pmu_conf->pmd_desc[cnum].hw_addr,
+                    (unsigned long long) value);
+
+       /*
+        * model-specific override, if any
+        */
+       if (pmu_info->write_pmd) {
+               pmu_info->write_pmd(ctx, cnum, value);
+               return;
+       }
+
+       wrmsrl(pfm_pmu_conf->pmd_desc[cnum].hw_addr, value);
+}
+
+/**
+ * pfm_arch_read_pmd - read a single PMD register
+ * @ctx: context to work on
+ * @cnum: PMD index
+ *
+ * return value is register 64-bit value
+ */
+static inline u64 pfm_arch_read_pmd(struct pfm_context *ctx, unsigned int cnum)
+{
+       struct pfm_arch_pmu_info *pmu_info;
+       u64 tmp;
+
+       pmu_info = pfm_pmu_info();
+
+       /*
+        * model-specific override, if any
+        */
+       if (pmu_info->read_pmd)
+               tmp = pmu_info->read_pmd(ctx, cnum);
+       else
+               rdmsrl(pfm_pmu_conf->pmd_desc[cnum].hw_addr, tmp);
+
+       PFM_DBG_ovfl("pfm_arch_read_pmd(0x%lx) = 0x%Lx",
+                    pfm_pmu_conf->pmd_desc[cnum].hw_addr,
+                    (unsigned long long) tmp);
+       return tmp;
+}
+
+/**
+ * pfm_arch_read_pmc - read a single PMC register
+ * @ctx: context to work on
+ * @cnum: PMC index
+ *
+ * return value is register 64-bit value
+ */
+static inline u64 pfm_arch_read_pmc(struct pfm_context *ctx, unsigned int cnum)
+{
+       struct pfm_arch_pmu_info *pmu_info;
+       u64 tmp;
+
+       pmu_info = pfm_pmu_info();
+
+       /*
+        * model-specific override, if any
+        */
+       if (pmu_info->read_pmc)
+               tmp = pmu_info->read_pmc(ctx, cnum);
+       else
+               rdmsrl(pfm_pmu_conf->pmc_desc[cnum].hw_addr, tmp);
+
+       PFM_DBG_ovfl("pfm_arch_read_pmc(0x%lx) = 0x%016Lx",
+                    pfm_pmu_conf->pmc_desc[cnum].hw_addr,
+                    (unsigned long long) tmp);
+       return tmp;
+}
+
+/**
+ * pfm_arch_is_active - return non-zero is monitoring has been started
+ * @ctx: context to check
+ *
+ * At certain points, perfmon needs to know if monitoring has been
+ * explicitly started.
+ *
+ * On x86, there is not other way but to use pfm_start/pfm_stop
+ * to activate monitoring, thus we can simply check flags.started
+ */
+static inline int pfm_arch_is_active(struct pfm_context *ctx)
+{
+       return ctx->flags.started;
+}
+
+
+/**
+ * pfm_arch_unload_context - detach context from thread or CPU
+ * @ctx: context to detach
+ *
+ * in system-wide ctx->task is NULL, otherwise it points to the
+ * attached thread
+ */
+static inline void pfm_arch_unload_context(struct pfm_context *ctx)
+{
+       struct pfm_arch_pmu_info *pmu_info;
+       struct pfm_arch_context *ctx_arch;
+
+       ctx_arch = pfm_ctx_arch(ctx);
+       pmu_info = pfm_pmu_info();
+
+       if (ctx_arch->flags.insecure) {
+               PFM_DBG("clear cr4.pce");
+               clear_in_cr4(X86_CR4_PCE);
+       }
+
+       if (pmu_info->unload_context)
+               pmu_info->unload_context(ctx);
+}
+
+/**
+ * pfm_arch_load_context - attach context to thread or CPU
+ * @ctx: context to attach
+ */
+static inline int pfm_arch_load_context(struct pfm_context *ctx)
+{
+       struct pfm_arch_pmu_info *pmu_info;
+       struct pfm_arch_context *ctx_arch;
+       int ret = 0;
+
+       ctx_arch = pfm_ctx_arch(ctx);
+       pmu_info = pfm_pmu_info();
+
+       /*
+        * RDPMC authorized in system-wide and
+        * per-thread self-monitoring.
+        *
+        * RDPMC only gives access to counts.
+        *
+        * The context-switch routine code does not restore
+        * all the PMD registers (optimization), thus there
+        * is a possible leak of counts there in per-thread
+        * mode.
+        */
+       if (ctx->task == current || ctx->flags.system) {
+               PFM_DBG("set cr4.pce");
+               set_in_cr4(X86_CR4_PCE);
+               ctx_arch->flags.insecure = 1;
+       }
+
+       if (pmu_info->load_context)
+               ret = pmu_info->load_context(ctx);
+
+       return ret;
+}
+
+void pfm_arch_restore_pmcs(struct pfm_context *ctx, struct pfm_event_set *set);
+void pfm_arch_start(struct task_struct *task, struct pfm_context *ctx);
+void pfm_arch_stop(struct task_struct *task, struct pfm_context *ctx);
+
+/**
+ * pfm_arch_unmask_monitoring - unmask monitoring
+ * @ctx: context to mask
+ * @set: current event set
+ *
+ * masking is slightly different from stopping in that, it does not undo
+ * the pfm_start() issued by user. This is used in conjunction with
+ * sampling. Masking means stop monitoring, but do not authorize user
+ * to issue pfm_start/stop during that time. Unmasking is achieved via
+ * pfm_restart() and also may also depend on the sampling format used.
+ *
+ * on x86 masking/unmasking use the start/stop mechanism, except
+ * that flags.started is not modified.
+ */
+static inline void pfm_arch_unmask_monitoring(struct pfm_context *ctx,
+                                             struct pfm_event_set *set)
+{
+       pfm_arch_start(current, ctx);
+}
+
+/**
+ * pfm_arch_intr_freeze_pmu - stop monitoring when handling PMU interrupt
+ * @ctx: current context
+ * @set: current event set
+ *
+ * called from __pfm_interrupt_handler().
+ * ctx is not NULL. ctx is locked. interrupts are masked
+ *
+ * The following actions must take place:
+ *  - stop all monitoring to ensure handler has consistent view.
+ *  - collect overflowed PMDs bitmask into povfls_pmds and
+ *    npend_ovfls. If no interrupt detected then npend_ovfls
+ *    must be set to zero.
+ */
+static inline void pfm_arch_intr_freeze_pmu(struct pfm_context *ctx,
+                                           struct pfm_event_set *set)
+{
+       /*
+        * on X86, freezing is equivalent to stopping
+        */
+       pfm_arch_stop(current, ctx);
+
+       /*
+        * we mark monitoring as stopped to avoid
+        * certain side effects especially in
+        * pfm_switch_sets_from_intr() and
+        * pfm_arch_restore_pmcs()
+        */
+       ctx->flags.started = 0;
+}
+
+/**
+ * pfm_arch_intr_unfreeze_pmu - conditionally reactive monitoring
+ * @ctx: current context
+ *
+ * current context may be not when dealing when spurious interrupts
+ *
+ * Must re-activate monitoring if context is not MASKED.
+ * interrupts are masked.
+ */
+static inline void pfm_arch_intr_unfreeze_pmu(struct pfm_context *ctx)
+{
+       if (ctx == NULL)
+               return;
+
+       PFM_DBG_ovfl("state=%d", ctx->state);
+
+       /*
+        * restore flags.started which is cleared in
+        * pfm_arch_intr_freeze_pmu()
+        */
+       ctx->flags.started = 1;
+
+       if (ctx->state == PFM_CTX_MASKED)
+               return;
+
+       pfm_arch_restore_pmcs(ctx, ctx->active_set);
+}
+
+/**
+ * pfm_arch_setfl_sane - check arch/model specific event set flags
+ * @ctx: context to work on
+ * @flags: event set flags as passed by user
+ *
+ * called from pfm_setfl_sane(). Context is locked. Interrupts are masked.
+ *
+ * Return:
+ *      0 when flags are valid
+ *      1 on error
+ */
+static inline int pfm_arch_setfl_sane(struct pfm_context *ctx, u32 flags)
+{
+       return 0;
+}
+
+/**
+ * pfm_arch_ovfl_reset_pmd - reset pmd on overflow
+ * @ctx: current context
+ * @cnum: PMD index
+ *
+ * On some CPUs, the upper bits of a counter must be set in order for the
+ * overflow interrupt to happen. On overflow, the counter has wrapped around,
+ * and the upper bits are cleared. This function may be used to set them back.
+ *
+ * For x86, the current version loses whatever is remaining in the counter,
+ * which is usually has a small count. In order not to loose this count,
+ * we do a read-modify-write to set the upper bits while preserving the
+ * low-order bits. This is slow but works.
+ */
+static inline void pfm_arch_ovfl_reset_pmd(struct pfm_context *ctx, unsigned int cnum)
+{
+       u64 val;
+       val = pfm_arch_read_pmd(ctx, cnum);
+       pfm_arch_write_pmd(ctx, cnum, val);
+}
+
+/**
+ * pfm_arch_context_create - create context
+ * @ctx: newly created context
+ * @flags: context flags as passed by user
+ *
+ * called from __pfm_create_context()
+ */
+static inline int pfm_arch_context_create(struct pfm_context *ctx, u32 ctx_flags)
+{
+       struct pfm_arch_pmu_info *pmu_info;
+
+       pmu_info = pfm_pmu_info();
+
+       if (pmu_info->create_context)
+               return pmu_info->create_context(ctx, ctx_flags);
+
+       return 0;
+}
+
+/**
+ * pfm_arch_context_free - free context
+ * @ctx: context to free
+ */
+static inline void pfm_arch_context_free(struct pfm_context *ctx)
+{
+       struct pfm_arch_pmu_info *pmu_info;
+
+       pmu_info = pfm_pmu_info();
+
+       if (pmu_info->free_context)
+               pmu_info->free_context(ctx);
+}
+
+/*
+ * pfm_arch_clear_pmd_ovfl_cond - alter the pmds in such a way that they
+ * will not cause cause interrupts when unused.
+ *
+ * This is a nop on x86
+ */
+static inline void pfm_arch_clear_pmd_ovfl_cond(struct pfm_context *ctx,
+                                               struct pfm_event_set *set)
+{}
+
+/*
+ * functions implemented in arch/x86/perfmon/perfmon.c
+ */
+int  pfm_arch_init(void);
+void pfm_arch_resend_irq(struct pfm_context *ctx);
+
+int  pfm_arch_ctxswout_thread(struct task_struct *task, struct pfm_context *ctx);
+void pfm_arch_ctxswin_thread(struct task_struct *task, struct pfm_context *ctx);
+
+void pfm_arch_restore_pmds(struct pfm_context *ctx, struct pfm_event_set *set);
+int  pfm_arch_pmu_config_init(struct pfm_pmu_config *cfg);
+void pfm_arch_pmu_config_remove(void);
+char *pfm_arch_get_pmu_module_name(void);
+int pfm_arch_pmu_acquire(u64 *unavail_pmcs, u64 *unavail_pmds);
+void pfm_arch_pmu_release(void);
+
+/*
+ * pfm_arch_serialize - make PMU modifications visible to subsequent instructions
+ *
+ * This is a nop on x86
+ */
+static inline void pfm_arch_serialize(void)
+{}
+
+/*
+ * on x86, the PMDs are already saved by pfm_arch_freeze_pmu()
+ * when entering the PMU interrupt handler, thus, we do not need
+ * to save them again in pfm_switch_sets_from_intr()
+ */
+static inline void pfm_arch_save_pmds_from_intr(struct pfm_context *ctx,
+                                               struct pfm_event_set *set)
+{}
+
+
+static inline void pfm_arch_ctxswout_sys(struct task_struct *task,
+                                        struct pfm_context *ctx)
+{}
+
+static inline void pfm_arch_ctxswin_sys(struct task_struct *task,
+                                       struct pfm_context *ctx)
+{}
+
+static inline void pfm_arch_init_percpu(void)
+{}
+
+static inline void pfm_cacheflush(void *addr, unsigned int len)
+{}
+
+/*
+ * this function is called from the PMU interrupt handler ONLY.
+ * On x86, the PMU is frozen via arch_stop, masking would be implemented
+ * via arch-stop as well. Given that the PMU is already stopped when
+ * entering the interrupt handler, we do not need to stop it again, so
+ * this function is a nop.
+ */
+static inline void pfm_arch_mask_monitoring(struct pfm_context *ctx,
+                                           struct pfm_event_set *set)
+{}
+
+
+static inline void pfm_arch_arm_handle_work(struct task_struct *task)
+{}
+
+static inline void pfm_arch_disarm_handle_work(struct task_struct *task)
+{}
+
+static inline int pfm_arch_get_base_syscall(void)
+{
+#ifdef __x86_64__
+       /* 32-bit syscall definition coming from ia32_unistd.h */
+       if (test_thread_flag(TIF_IA32))
+               return __NR_ia32_pfm_create_context;
+#endif
+       return __NR_pfm_create_context;
+}
+
+#define PFM_ARCH_CTX_SIZE      (sizeof(struct pfm_arch_context))
+/*
+ * x86 does not need extra alignment requirements for the sampling buffer
+ */
+#define PFM_ARCH_SMPL_ALIGN_SIZE       0
+
+asmlinkage void  pmu_interrupt(void);
+
+#endif /* CONFIG_PEFMON */
+
+#endif /* _ASM_X86_PERFMON_KERN_H_ */
diff --git a/include/asm-x86/perfmon_pebs_core_smpl.h b/include/asm-x86/perfmon_pebs_core_smpl.h
new file mode 100644
index 0000000..4a12e0d
--- /dev/null
+++ b/include/asm-x86/perfmon_pebs_core_smpl.h
@@ -0,0 +1,164 @@
+/*
+ * Copyright (c) 2005-2007 Hewlett-Packard Development Company, L.P.
+ * Contributed by Stephane Eranian <eranian@hpl.hp.com>
+ *
+ * This program is free software; you can redistribute it and/or
+ * modify it under the terms of version 2 of the GNU General Public
+ * License as published by the Free Software Foundation.
+ *
+ * This program is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+ * General Public License for more details.
+ *
+ * You should have received a copy of the GNU General Public License
+ * along with this program; if not, write to the Free Software
+ * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA
+ * 02111-1307 USA
+ *
+ * This file implements the sampling format to support Intel
+ * Precise Event Based Sampling (PEBS) feature of Intel Core
+ * processors, such as Intel Core 2.
+ *
+ * What is PEBS?
+ * ------------
+ *  This is a hardware feature to enhance sampling by providing
+ *  better precision as to where a sample is taken. This avoids the
+ *  typical skew in the instruction one can observe with any
+ *  interrupt-based sampling technique.
+ *
+ *  PEBS also lowers sampling overhead significantly by having the
+ *  processor store samples instead of the OS. PMU interrupt are only
+ *  generated after multiple samples are written.
+ *
+ *  Another benefit of PEBS is that samples can be captured inside
+ *  critical sections where interrupts are masked.
+ *
+ * How does it work?
+ *  PEBS effectively implements a Hw buffer. The Os must pass a region
+ *  of memory where samples are to be stored. The region can have any
+ *  size. The OS must also specify the sampling period to reload. The PMU
+ *  will interrupt when it reaches the end of the buffer or a specified
+ *  threshold location inside the memory region.
+ *
+ *  The description of the buffer is stored in the Data Save Area (DS).
+ *  The samples are stored sequentially in the buffer. The format of the
+ *  buffer is fixed and specified in the PEBS documentation.  The sample
+ *  format does not change between 32-bit and 64-bit modes unlike on the
+ *  Pentium 4 version of PEBS.
+ *
+ *  PEBS does not work when HyperThreading is enabled due to certain MSR
+ *  being shared being to two threads.
+ *
+ *  What does the format do?
+ *   It provides access to the PEBS feature for both 32-bit and 64-bit
+ *   processors that support it.
+ *
+ *   The same code and data structures are used for both 32-bit and 64-bi
+ *   modes. A single format name is used for both modes. In 32-bit mode,
+ *   some of the extended registers are written to zero in each sample.
+ *
+ *   It is important to realize that the format provides a zero-copy
+ *   environment for the samples, i.e,, the OS never touches the
+ *   samples. Whatever the processor write is directly accessible to
+ *   the user.
+ *
+ *   Parameters to the buffer can be passed via pfm_create_context() in
+ *   the pfm_pebs_smpl_arg structure.
+ */
+#ifndef __PERFMON_PEBS_CORE_SMPL_H__
+#define __PERFMON_PEBS_CORE_SMPL_H__ 1
+
+/*
+ * The 32-bit and 64-bit formats are identical, thus we use only
+ * one name for the format.
+ */
+#define PFM_PEBS_CORE_SMPL_NAME        "pebs_core"
+
+/*
+ * format specific parameters (passed at context creation)
+ *
+ * intr_thres: index from start of buffer of entry where the
+ * PMU interrupt must be triggered. It must be several samples
+ * short of the end of the buffer.
+ */
+struct pfm_pebs_core_smpl_arg {
+       u64 cnt_reset;    /* counter reset value */
+       size_t buf_size;  /* size of the PEBS buffer in bytes */
+       size_t intr_thres;/* index of PEBS interrupt threshold entry */
+       u64 reserved[6];  /* for future use */
+};
+
+/*
+ * Data Save Area (32 and 64-bit mode)
+ *
+ * The DS area is exposed to the user. To determine the number
+ * of samples available in PEBS, it is necessary to substract
+ * pebs_index from pebs_base.
+ *
+ * Layout of the structure is mandated by hardware and specified
+ * in the Intel documentation.
+ */
+struct pfm_ds_area_core {
+       u64 bts_buf_base;
+       u64 bts_index;
+       u64 bts_abs_max;
+       u64 bts_intr_thres;
+       u64 pebs_buf_base;
+       u64 pebs_index;
+       u64 pebs_abs_max;
+       u64 pebs_intr_thres;
+       u64 pebs_cnt_reset;
+};
+
+/*
+ * This header is at the beginning of the sampling buffer returned to the user.
+ *
+ * Because of PEBS alignement constraints, the actual PEBS buffer area does
+ * not necessarily begin right after the header. The hdr_start_offs must be
+ * used to compute the first byte of the buffer. The offset is defined as
+ * the number of bytes between the end of the header and the beginning of
+ * the buffer. As such the formula is:
+ *     actual_buffer = (unsigned long)(hdr+1)+hdr->hdr_start_offs
+ */
+struct pfm_pebs_core_smpl_hdr {
+       u64 overflows;                  /* #overflows for buffer */
+       size_t buf_size;                /* bytes in the buffer */
+       size_t start_offs;              /* actual buffer start offset */
+       u32 version;                    /* smpl format version */
+       u32 reserved1;                  /* for future use */
+       u64 reserved2[5];               /* for future use */
+       struct pfm_ds_area_core ds;     /* data save area */
+};
+
+/*
+ * Sample format as mandated by Intel documentation.
+ * The same format is used in both 32 and 64 bit modes.
+ */
+struct pfm_pebs_core_smpl_entry {
+       u64     eflags;
+       u64     ip;
+       u64     eax;
+       u64     ebx;
+       u64     ecx;
+       u64     edx;
+       u64     esi;
+       u64     edi;
+       u64     ebp;
+       u64     esp;
+       u64     r8;     /* 0 in 32-bit mode */
+       u64     r9;     /* 0 in 32-bit mode */
+       u64     r10;    /* 0 in 32-bit mode */
+       u64     r11;    /* 0 in 32-bit mode */
+       u64     r12;    /* 0 in 32-bit mode */
+       u64     r13;    /* 0 in 32-bit mode */
+       u64     r14;    /* 0 in 32-bit mode */
+       u64     r15;    /* 0 in 32-bit mode */
+};
+
+#define PFM_PEBS_CORE_SMPL_VERSION_MAJ 1U
+#define PFM_PEBS_CORE_SMPL_VERSION_MIN 0U
+#define PFM_PEBS_CORE_SMPL_VERSION (((PFM_PEBS_CORE_SMPL_VERSION_MAJ&0xffff)<<16)|\
+                                  (PFM_PEBS_CORE_SMPL_VERSION_MIN & 0xffff))
+
+#endif /* __PERFMON_PEBS_CORE_SMPL_H__ */
diff --git a/include/asm-x86/perfmon_pebs_p4_smpl.h b/include/asm-x86/perfmon_pebs_p4_smpl.h
new file mode 100644
index 0000000..26b51b4
--- /dev/null
+++ b/include/asm-x86/perfmon_pebs_p4_smpl.h
@@ -0,0 +1,193 @@
+/*
+ * Copyright (c) 2005-2006 Hewlett-Packard Development Company, L.P.
+ * Contributed by Stephane Eranian <eranian@hpl.hp.com>
+ *
+ * This program is free software; you can redistribute it and/or
+ * modify it under the terms of version 2 of the GNU General Public
+ * License as published by the Free Software Foundation.
+ *
+ * This program is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+ * General Public License for more details.
+ *
+ * You should have received a copy of the GNU General Public License
+ * along with this program; if not, write to the Free Software
+ * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA
+ * 02111-1307 USA
+ *
+ * This file implements the sampling format to support Intel
+ * Precise Event Based Sampling (PEBS) feature of Pentium 4
+ * and other Netburst-based processors. Not to be used for
+ * Intel Core-based processors.
+ *
+ * What is PEBS?
+ * ------------
+ *  This is a hardware feature to enhance sampling by providing
+ *  better precision as to where a sample is taken. This avoids the
+ *  typical skew in the instruction one can observe with any
+ *  interrupt-based sampling technique.
+ *
+ *  PEBS also lowers sampling overhead significantly by having the
+ *  processor store samples instead of the OS. PMU interrupt are only
+ *  generated after multiple samples are written.
+ *
+ *  Another benefit of PEBS is that samples can be captured inside
+ *  critical sections where interrupts are masked.
+ *
+ * How does it work?
+ *  PEBS effectively implements a Hw buffer. The Os must pass a region
+ *  of memory where samples are to be stored. The region can have any
+ *  size. The OS must also specify the sampling period to reload. The PMU
+ *  will interrupt when it reaches the end of the buffer or a specified
+ *  threshold location inside the memory region.
+ *
+ *  The description of the buffer is stored in the Data Save Area (DS).
+ *  The samples are stored sequentially in the buffer. The format of the
+ *  buffer is fixed and specified in the PEBS documentation.  The sample
+ *  format changes between 32-bit and 64-bit modes due to extended register
+ *  file.
+ *
+ *  PEBS does not work when HyperThreading is enabled due to certain MSR
+ *  being shared being to two threads.
+ *
+ *  What does the format do?
+ *   It provides access to the PEBS feature for both 32-bit and 64-bit
+ *   processors that support it.
+ *
+ *   The same code is used for both 32-bit and 64-bit modes, but different
+ *   format names are used because the two modes are not compatible due to
+ *   data model and register file differences. Similarly the public data
+ *   structures describing the samples are different.
+ *
+ *   It is important to realize that the format provides a zero-copy environment
+ *   for the samples, i.e,, the OS never touches the samples. Whatever the
+ *   processor write is directly accessible to the user.
+ *
+ *   Parameters to the buffer can be passed via pfm_create_context() in
+ *   the pfm_pebs_smpl_arg structure.
+ *
+ *   It is not possible to mix a 32-bit PEBS application on top of a 64-bit
+ *   host kernel.
+ */
+#ifndef __PERFMON_PEBS_P4_SMPL_H__
+#define __PERFMON_PEBS_P4_SMPL_H__ 1
+
+#ifdef __i386__
+/*
+ * The 32-bit and 64-bit formats are not compatible, thus we have
+ * two different identifications so that 32-bit programs running on
+ * 64-bit OS will fail to use the 64-bit PEBS support.
+ */
+#define PFM_PEBS_P4_SMPL_NAME  "pebs32_p4"
+#else
+#define PFM_PEBS_P4_SMPL_NAME  "pebs64_p4"
+#endif
+
+/*
+ * format specific parameters (passed at context creation)
+ *
+ * intr_thres: index from start of buffer of entry where the
+ * PMU interrupt must be triggered. It must be several samples
+ * short of the end of the buffer.
+ */
+struct pfm_pebs_p4_smpl_arg {
+       u64 cnt_reset;    /* counter reset value */
+       size_t buf_size;  /* size of the PEBS buffer in bytes */
+       size_t intr_thres;/* index of PEBS interrupt threshold entry */
+       u64 reserved[6];  /* for future use */
+};
+
+/*
+ * Data Save Area (32 and 64-bit mode)
+ *
+ * The DS area must be exposed to the user because this is the only
+ * way to report on the number of valid entries recorded by the CPU.
+ * This is required when the buffer is not full, i..e, there was not
+ * PMU interrupt.
+ *
+ * Layout of the structure is mandated by hardware and specified in
+ * the Intel documentation.
+ */
+struct pfm_ds_area_p4 {
+       unsigned long   bts_buf_base;
+       unsigned long   bts_index;
+       unsigned long   bts_abs_max;
+       unsigned long   bts_intr_thres;
+       unsigned long   pebs_buf_base;
+       unsigned long   pebs_index;
+       unsigned long   pebs_abs_max;
+       unsigned long   pebs_intr_thres;
+       u64             pebs_cnt_reset;
+};
+
+/*
+ * This header is at the beginning of the sampling buffer returned to the user.
+ *
+ * Because of PEBS alignement constraints, the actual PEBS buffer area does
+ * not necessarily begin right after the header. The hdr_start_offs must be
+ * used to compute the first byte of the buffer. The offset is defined as
+ * the number of bytes between the end of the header and the beginning of
+ * the buffer. As such the formula is:
+ *     actual_buffer = (unsigned long)(hdr+1)+hdr->hdr_start_offs
+ */
+struct pfm_pebs_p4_smpl_hdr {
+       u64 overflows;                  /* #overflows for buffer */
+       size_t buf_size;                /* bytes in the buffer */
+       size_t start_offs;              /* actual buffer start offset */
+       u32 version;                    /* smpl format version */
+       u32 reserved1;                  /* for future use */
+       u64 reserved2[5];               /* for future use */
+       struct pfm_ds_area_p4 ds;       /* data save area */
+};
+
+/*
+ * 64-bit PEBS record format is described in
+ * http://www.intel.com/technology/64bitextensions/30083502.pdf
+ *
+ * The format does not peek at samples. The sample structure is only
+ * used to ensure that the buffer is large enough to accomodate one
+ * sample.
+ */
+#ifdef __i386__
+struct pfm_pebs_p4_smpl_entry {
+       u32     eflags;
+       u32     ip;
+       u32     eax;
+       u32     ebx;
+       u32     ecx;
+       u32     edx;
+       u32     esi;
+       u32     edi;
+       u32     ebp;
+       u32     esp;
+};
+#else
+struct pfm_pebs_p4_smpl_entry {
+       u64     eflags;
+       u64     ip;
+       u64     eax;
+       u64     ebx;
+       u64     ecx;
+       u64     edx;
+       u64     esi;
+       u64     edi;
+       u64     ebp;
+       u64     esp;
+       u64     r8;
+       u64     r9;
+       u64     r10;
+       u64     r11;
+       u64     r12;
+       u64     r13;
+       u64     r14;
+       u64     r15;
+};
+#endif
+
+#define PFM_PEBS_P4_SMPL_VERSION_MAJ 1U
+#define PFM_PEBS_P4_SMPL_VERSION_MIN 0U
+#define PFM_PEBS_P4_SMPL_VERSION (((PFM_PEBS_P4_SMPL_VERSION_MAJ&0xffff)<<16)|\
+                                  (PFM_PEBS_P4_SMPL_VERSION_MIN & 0xffff))
+
+#endif /* __PERFMON_PEBS_P4_SMPL_H__ */
diff --git a/include/asm-x86/thread_info.h b/include/asm-x86/thread_info.h
index da0a675..b3a6ae9 100644
--- a/include/asm-x86/thread_info.h
+++ b/include/asm-x86/thread_info.h
@@ -71,6 +71,7 @@ struct thread_info {
  * Warning: layout of LSW is hardcoded in entry.S
  */
 #define TIF_SYSCALL_TRACE      0       /* syscall trace active */
+#define TIF_PERFMON_WORK       1       /* work for pfm_handle_work() */
 #define TIF_SIGPENDING         2       /* signal pending */
 #define TIF_NEED_RESCHED       3       /* rescheduling necessary */
 #define TIF_SINGLESTEP         4       /* reenable singlestep on user return*/
@@ -91,6 +92,7 @@ struct thread_info {
 #define TIF_DEBUGCTLMSR                25      /* uses thread_struct.debugctlmsr */
 #define TIF_DS_AREA_MSR                26      /* uses thread_struct.ds_area_msr */
 #define TIF_BTS_TRACE_TS       27      /* record scheduling event timestamps */
+#define TIF_PERFMON_CTXSW      28      /* perfmon needs ctxsw calls */
 
 #define _TIF_SYSCALL_TRACE     (1 << TIF_SYSCALL_TRACE)
 #define _TIF_SIGPENDING                (1 << TIF_SIGPENDING)
@@ -112,6 +114,8 @@ struct thread_info {
 #define _TIF_DEBUGCTLMSR       (1 << TIF_DEBUGCTLMSR)
 #define _TIF_DS_AREA_MSR       (1 << TIF_DS_AREA_MSR)
 #define _TIF_BTS_TRACE_TS      (1 << TIF_BTS_TRACE_TS)
+#define _TIF_PERFMON_WORK      (1<<TIF_PERFMON_WORK)
+#define _TIF_PERFMON_CTXSW     (1<<TIF_PERFMON_CTXSW)
 
 /* work to do in syscall_trace_enter() */
 #define _TIF_WORK_SYSCALL_ENTRY        \
@@ -133,12 +137,12 @@ struct thread_info {
 
 /* Only used for 64 bit */
 #define _TIF_DO_NOTIFY_MASK                                            \
-       (_TIF_SIGPENDING|_TIF_MCE_NOTIFY)
+       (_TIF_SIGPENDING|_TIF_MCE_NOTIFY|_TIF_PERFMON_WORK)
 
 /* flags to check in __switch_to() */
 #define _TIF_WORK_CTXSW                                                        \
        (_TIF_IO_BITMAP|_TIF_DEBUGCTLMSR|_TIF_DS_AREA_MSR|_TIF_BTS_TRACE_TS| \
-                                                               _TIF_NOTSC)
+        _TIF_NOTSC|_TIF_PERFMON_CTXSW)
 
 #define _TIF_WORK_CTXSW_PREV _TIF_WORK_CTXSW
 #define _TIF_WORK_CTXSW_NEXT (_TIF_WORK_CTXSW|_TIF_DEBUG)
diff --git a/include/asm-x86/unistd_32.h b/include/asm-x86/unistd_32.h
index d739467..5d8cca1 100644
--- a/include/asm-x86/unistd_32.h
+++ b/include/asm-x86/unistd_32.h
@@ -338,9 +338,23 @@
 #define __NR_dup3              330
 #define __NR_pipe2             331
 #define __NR_inotify_init1     332
+#define __NR_pfm_create_context        333
+#define __NR_pfm_write_pmcs    (__NR_pfm_create_context+1)
+#define __NR_pfm_write_pmds    (__NR_pfm_create_context+2)
+#define __NR_pfm_read_pmds     (__NR_pfm_create_context+3)
+#define __NR_pfm_load_context  (__NR_pfm_create_context+4)
+#define __NR_pfm_start         (__NR_pfm_create_context+5)
+#define __NR_pfm_stop          (__NR_pfm_create_context+6)
+#define __NR_pfm_restart       (__NR_pfm_create_context+7)
+#define __NR_pfm_create_evtsets        (__NR_pfm_create_context+8)
+#define __NR_pfm_getinfo_evtsets (__NR_pfm_create_context+9)
+#define __NR_pfm_delete_evtsets (__NR_pfm_create_context+10)
+#define __NR_pfm_unload_context        (__NR_pfm_create_context+11)
 
 #ifdef __KERNEL__
 
+#define NR_syscalls 345
+
 #define __ARCH_WANT_IPC_PARSE_VERSION
 #define __ARCH_WANT_OLD_READDIR
 #define __ARCH_WANT_OLD_STAT
diff --git a/include/asm-x86/unistd_64.h b/include/asm-x86/unistd_64.h
index 3a341d7..75dac98 100644
--- a/include/asm-x86/unistd_64.h
+++ b/include/asm-x86/unistd_64.h
@@ -653,7 +653,30 @@ __SYSCALL(__NR_dup3, sys_dup3)
 __SYSCALL(__NR_pipe2, sys_pipe2)
 #define __NR_inotify_init1                     294
 __SYSCALL(__NR_inotify_init1, sys_inotify_init1)
-
+#define __NR_pfm_create_context                        295
+__SYSCALL(__NR_pfm_create_context, sys_pfm_create_context)
+#define __NR_pfm_write_pmcs                    (__NR_pfm_create_context+1)
+__SYSCALL(__NR_pfm_write_pmcs, sys_pfm_write_pmcs)
+#define __NR_pfm_write_pmds                    (__NR_pfm_create_context+2)
+__SYSCALL(__NR_pfm_write_pmds, sys_pfm_write_pmds)
+#define __NR_pfm_read_pmds                     (__NR_pfm_create_context+3)
+ __SYSCALL(__NR_pfm_read_pmds, sys_pfm_read_pmds)
+#define __NR_pfm_load_context                  (__NR_pfm_create_context+4)
+__SYSCALL(__NR_pfm_load_context, sys_pfm_load_context)
+#define __NR_pfm_start                         (__NR_pfm_create_context+5)
+__SYSCALL(__NR_pfm_start, sys_pfm_start)
+#define __NR_pfm_stop                          (__NR_pfm_create_context+6)
+__SYSCALL(__NR_pfm_stop, sys_pfm_stop)
+#define __NR_pfm_restart                       (__NR_pfm_create_context+7)
+__SYSCALL(__NR_pfm_restart, sys_pfm_restart)
+#define __NR_pfm_create_evtsets                        (__NR_pfm_create_context+8)
+__SYSCALL(__NR_pfm_create_evtsets, sys_pfm_create_evtsets)
+#define __NR_pfm_getinfo_evtsets               (__NR_pfm_create_context+9)
+__SYSCALL(__NR_pfm_getinfo_evtsets, sys_pfm_getinfo_evtsets)
+#define __NR_pfm_delete_evtsets                (__NR_pfm_create_context+10)
+__SYSCALL(__NR_pfm_delete_evtsets, sys_pfm_delete_evtsets)
+#define __NR_pfm_unload_context                        (__NR_pfm_create_context+11)
+__SYSCALL(__NR_pfm_unload_context, sys_pfm_unload_context)
 
 #ifndef __NO_STUBS
 #define __ARCH_WANT_OLD_READDIR
diff --git a/include/linux/Kbuild b/include/linux/Kbuild
index b68ec09..d37036a 100644
--- a/include/linux/Kbuild
+++ b/include/linux/Kbuild
@@ -162,6 +162,8 @@ header-y += video_decoder.h
 header-y += video_encoder.h
 header-y += videotext.h
 header-y += x25.h
+header-y += perfmon.h
+header-y += perfmon_dfl_smpl.h
 
 unifdef-y += acct.h
 unifdef-y += adb.h
diff --git a/include/linux/perfmon.h b/include/linux/perfmon.h
new file mode 100644
index 0000000..5d9b977
--- /dev/null
+++ b/include/linux/perfmon.h
@@ -0,0 +1,213 @@
+/*
+ * Copyright (c) 2001-2006 Hewlett-Packard Development Company, L.P.
+ * Contributed by Stephane Eranian <eranian@hpl.hp.com>
+ *
+ * This program is free software; you can redistribute it and/or
+ * modify it under the terms of version 2 of the GNU General Public
+ * License as published by the Free Software Foundation.
+ *
+ * This program is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+ * General Public License for more details.
+ *
+ * You should have received a copy of the GNU General Public License
+ * along with this program; if not, write to the Free Software
+ * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA
+ * 02111-1307 USA
+ */
+
+#ifndef __LINUX_PERFMON_H__
+#define __LINUX_PERFMON_H__
+
+/*
+ * This file contains all the user visible generic definitions for the
+ * interface. Model-specific user-visible definitions are located in
+ * the asm/perfmon.h file.
+ */
+
+/*
+ * include arch-specific user interface definitions
+ */
+#include <asm/perfmon.h>
+
+/*
+ * defined by each arch
+ */
+#define PFM_MAX_PMCS   PFM_ARCH_MAX_PMCS
+#define PFM_MAX_PMDS   PFM_ARCH_MAX_PMDS
+
+/*
+ * number of elements for each type of bitvector
+ * all bitvectors use u64 fixed size type on all architectures.
+ */
+#define PFM_BVSIZE(x)  (((x)+(sizeof(__u64)<<3)-1) / (sizeof(__u64)<<3))
+#define PFM_PMD_BV     PFM_BVSIZE(PFM_MAX_PMDS)
+#define PFM_PMC_BV     PFM_BVSIZE(PFM_MAX_PMCS)
+
+/*
+ * register flags layout:
+ * bit[00-15] : generic flags
+ * bit[16-31] : arch-specific flags
+ *
+ * PFM_REGFL_NO_EMUL64: must be set on the PMC controlling the PMD
+ */
+#define PFM_REGFL_OVFL_NOTIFY  0x1     /* PMD: send notification on event */
+#define PFM_REGFL_RANDOM       0x2     /* PMD: randomize value after event */
+#define PFM_REGFL_NO_EMUL64    0x4     /* PMC: no 64-bit emulation */
+
+/*
+ * event set flags layout:
+ * bits[00-15] : generic flags
+ * bits[16-31] : arch-specific flags (see asm/perfmon.h)
+ */
+#define PFM_SETFL_OVFL_SWITCH  0x01 /* enable switch on overflow */
+#define PFM_SETFL_TIME_SWITCH  0x02 /* enable switch on timeout */
+
+/*
+ * argument to pfm_create_context() system call
+ * structure shared with user level
+ */
+struct pfarg_ctx {
+       __u32           ctx_flags;        /* noblock/block/syswide */
+       __u32           ctx_reserved1;    /* for future use */
+       __u64           ctx_reserved2[7]; /* for future use */
+};
+
+/*
+ * context flags layout:
+ * bits[00-15]: generic flags
+ * bits[16-31]: arch-specific flags (see perfmon_const.h)
+ */
+#define PFM_FL_NOTIFY_BLOCK             0x01   /* block task on user notifications */
+#define PFM_FL_SYSTEM_WIDE      0x02   /* create a system wide context */
+#define PFM_FL_OVFL_NO_MSG      0x80   /* no overflow msgs */
+
+/*
+ * argument to pfm_write_pmcs() system call.
+ * structure shared with user level
+ */
+struct pfarg_pmc {
+       __u16 reg_num;          /* which register */
+       __u16 reg_set;          /* event set for this register */
+       __u32 reg_flags;        /* REGFL flags */
+       __u64 reg_value;        /* pmc value */
+       __u64 reg_reserved2[4]; /* for future use */
+};
+
+/*
+ * argument to pfm_write_pmds() and pfm_read_pmds() system calls.
+ * structure shared with user level
+ */
+struct pfarg_pmd {
+       __u16 reg_num;          /* which register */
+       __u16 reg_set;          /* event set for this register */
+       __u32 reg_flags;        /* REGFL flags */
+       __u64 reg_value;        /* initial pmc/pmd value */
+       __u64 reg_long_reset;   /* value to reload after notification */
+       __u64 reg_short_reset;  /* reset after counter overflow */
+       __u64 reg_last_reset_val;       /* return: PMD last reset value */
+       __u64 reg_ovfl_switch_cnt;      /* #overflows before switch */
+       __u64 reg_reset_pmds[PFM_PMD_BV]; /* reset on overflow */
+       __u64 reg_smpl_pmds[PFM_PMD_BV];  /* record in sample */
+       __u64 reg_smpl_eventid; /* opaque event identifier */
+       __u64 reg_random_mask;  /* bitmask used to limit random value */
+       __u32 reg_random_seed;  /* seed for randomization (OBSOLETE) */
+       __u32 reg_reserved2[7]; /* for future use */
+};
+
+/*
+ * optional argument to pfm_start() system call. Pass NULL if not needed.
+ * structure shared with user level
+ */
+struct pfarg_start {
+       __u16 start_set;        /* event set to start with */
+       __u16 start_reserved1;  /* for future use */
+       __u32 start_reserved2;  /* for future use */
+       __u64 reserved3[3];     /* for future use */
+};
+
+/*
+ * argument to pfm_load_context() system call.
+ * structure shared with user level
+ */
+struct pfarg_load {
+       __u32   load_pid;          /* thread or CPU to attach to */
+       __u16   load_set;          /* set to load first */
+       __u16   load_reserved1;    /* for future use */
+       __u64   load_reserved2[3]; /* for future use */
+};
+
+/*
+ * argument to pfm_create_evtsets() and pfm_delete_evtsets() system calls.
+ * structure shared with user level.
+ */
+struct pfarg_setdesc {
+       __u16   set_id;           /* which set */
+       __u16   set_reserved1;    /* for future use */
+       __u32   set_flags;        /* SETFL flags  */
+       __u64   set_timeout;      /* switch timeout in nsecs */
+       __u64   reserved[6];      /* for future use */
+};
+
+/*
+ * argument to pfm_getinfo_evtsets() system call.
+ * structure shared with user level
+ */
+struct pfarg_setinfo {
+       __u16   set_id;                 /* which set */
+       __u16   set_reserved1;          /* for future use */
+       __u32   set_flags;              /* out: SETFL flags */
+       __u64   set_ovfl_pmds[PFM_PMD_BV]; /* out: last ovfl PMDs */
+       __u64   set_runs;               /* out: #times the set was active */
+       __u64   set_timeout;            /* out: eff/leftover timeout (nsecs) */
+       __u64   set_act_duration;       /* out: time set was active in nsecs */
+       __u64   set_avail_pmcs[PFM_PMC_BV];/* out: available PMCs */
+       __u64   set_avail_pmds[PFM_PMD_BV];/* out: available PMDs */
+       __u64   set_reserved3[6];       /* for future use */
+};
+
+/*
+ * default value for the user and group security parameters in
+ * /proc/sys/kernel/perfmon/sys_group
+ * /proc/sys/kernel/perfmon/task_group
+ */
+#define PFM_GROUP_PERM_ANY     -1      /* any user/group */
+
+/*
+ * overflow notification message.
+ * structure shared with user level
+ */
+struct pfarg_ovfl_msg {
+       __u32           msg_type;       /* message type: PFM_MSG_OVFL */
+       __u32           msg_ovfl_pid;   /* process id */
+       __u16           msg_active_set; /* active set at overflow */
+       __u16           msg_ovfl_cpu;   /* cpu of PMU interrupt */
+       __u32           msg_ovfl_tid;   /* thread id */
+       __u64           msg_ovfl_ip;    /* IP on PMU intr */
+       __u64           msg_ovfl_pmds[PFM_PMD_BV];/* overflowed PMDs */
+};
+
+#define PFM_MSG_OVFL   1       /* an overflow happened */
+#define PFM_MSG_END    2       /* task to which context was attached ended */
+
+/*
+ * generic notification message (union).
+ * union shared with user level
+ */
+union pfarg_msg {
+       __u32   type;
+       struct pfarg_ovfl_msg pfm_ovfl_msg;
+};
+
+/*
+ * perfmon version number
+ */
+#define PFM_VERSION_MAJ                 2U
+#define PFM_VERSION_MIN                 82U
+#define PFM_VERSION             (((PFM_VERSION_MAJ&0xffff)<<16)|\
+                                 (PFM_VERSION_MIN & 0xffff))
+#define PFM_VERSION_MAJOR(x)    (((x)>>16) & 0xffff)
+#define PFM_VERSION_MINOR(x)    ((x) & 0xffff)
+
+#endif /* __LINUX_PERFMON_H__ */
diff --git a/include/linux/perfmon_dfl_smpl.h b/include/linux/perfmon_dfl_smpl.h
new file mode 100644
index 0000000..e0817a8
--- /dev/null
+++ b/include/linux/perfmon_dfl_smpl.h
@@ -0,0 +1,78 @@
+/*
+ * Copyright (c) 2005-2006 Hewlett-Packard Development Company, L.P.
+ *               Contributed by Stephane Eranian <eranian@hpl.hp.com>
+ *
+ * This file implements the new dfl sampling buffer format
+ * for perfmon2 subsystem.
+ *
+ * This program is free software; you can redistribute it and/or
+ * modify it under the terms of version 2 of the GNU General Public
+ * License as published by the Free Software Foundation.
+ *
+ * This program is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+ * General Public License for more details.
+ *
+ * You should have received a copy of the GNU General Public License
+ * along with this program; if not, write to the Free Software
+ * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA
+ * 02111-1307 USA
+ */
+#ifndef __PERFMON_DFL_SMPL_H__
+#define __PERFMON_DFL_SMPL_H__ 1
+
+/*
+ * format specific parameters (passed at context creation)
+ */
+struct pfm_dfl_smpl_arg {
+       __u64 buf_size;         /* size of the buffer in bytes */
+       __u32 buf_flags;        /* buffer specific flags */
+       __u32 reserved1;        /* for future use */
+       __u64 reserved[6];      /* for future use */
+};
+
+/*
+ * This header is at the beginning of the sampling buffer returned to the user.
+ * It is directly followed by the first record.
+ */
+struct pfm_dfl_smpl_hdr {
+       __u64 hdr_count;        /* how many valid entries */
+       __u64 hdr_cur_offs;     /* current offset from top of buffer */
+       __u64 hdr_overflows;    /* #overflows for buffer */
+       __u64 hdr_buf_size;     /* bytes in the buffer */
+       __u64 hdr_min_buf_space;/* minimal buffer size (internal use) */
+       __u32 hdr_version;      /* smpl format version */
+       __u32 hdr_buf_flags;    /* copy of buf_flags */
+       __u64 hdr_reserved[10]; /* for future use */
+};
+
+/*
+ * Entry header in the sampling buffer.  The header is directly followed
+ * with the values of the PMD registers of interest saved in increasing
+ * index order: PMD4, PMD5, and so on. How many PMDs are present depends
+ * on how the session was programmed.
+ *
+ * In the case where multiple counters overflow at the same time, multiple
+ * entries are written consecutively.
+ *
+ * last_reset_value member indicates the initial value of the overflowed PMD.
+ */
+struct pfm_dfl_smpl_entry {
+       __u32   pid;            /* thread id (for NPTL, this is gettid()) */
+       __u16   ovfl_pmd;       /* index of overflowed PMD for this sample */
+       __u16   reserved;       /* for future use */
+       __u64   last_reset_val; /* initial value of overflowed PMD */
+       __u64   ip;             /* where did the overflow intr happened */
+       __u64   tstamp;         /* overflow timetamp */
+       __u16   cpu;            /* cpu on which the overfow occurred */
+       __u16   set;            /* event set active when overflow ocurred */
+       __u32   tgid;           /* thread group id (getpid() for NPTL) */
+};
+
+#define PFM_DFL_SMPL_VERSION_MAJ 1U
+#define PFM_DFL_SMPL_VERSION_MIN 0U
+#define PFM_DFL_SMPL_VERSION (((PFM_DFL_SMPL_VERSION_MAJ&0xffff)<<16)|\
+                               (PFM_DFL_SMPL_VERSION_MIN & 0xffff))
+
+#endif /* __PERFMON_DFL_SMPL_H__ */
diff --git a/include/linux/perfmon_fmt.h b/include/linux/perfmon_fmt.h
new file mode 100644
index 0000000..82a6a90
--- /dev/null
+++ b/include/linux/perfmon_fmt.h
@@ -0,0 +1,74 @@
+/*
+ * Copyright (c) 2001-2006 Hewlett-Packard Development Company, L.P.
+ * Contributed by Stephane Eranian <eranian@hpl.hp.com>
+ *
+ * Interface for custom sampling buffer format modules
+ *
+ * This program is free software; you can redistribute it and/or
+ * modify it under the terms of version 2 of the GNU General Public
+ * License as published by the Free Software Foundation.
+ *
+ * This program is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+ * General Public License for more details.
+ *
+ * You should have received a copy of the GNU General Public License
+ * along with this program; if not, write to the Free Software
+ * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA
+ * 02111-1307 USA
+ */
+#ifndef __PERFMON_FMT_H__
+#define __PERFMON_FMT_H__ 1
+
+#include <linux/kobject.h>
+
+typedef int (*fmt_validate_t)(u32 flags, u16 npmds, void *arg);
+typedef        int (*fmt_getsize_t)(u32 flags, void *arg, size_t *size);
+typedef int (*fmt_init_t)(struct pfm_context *ctx, void *buf, u32 flags,
+                         u16 nmpds, void *arg);
+typedef int (*fmt_restart_t)(int is_active, u32 *ovfl_ctrl, void *buf);
+typedef int (*fmt_exit_t)(void *buf);
+typedef int (*fmt_handler_t)(struct pfm_context *ctx,
+                            unsigned long ip, u64 stamp, void *data);
+
+struct pfm_smpl_fmt {
+       char            *fmt_name;      /* name of the format (required) */
+       size_t          fmt_arg_size;   /* size of fmt args for ctx create */
+       u32             fmt_flags;      /* format specific flags */
+       u32             fmt_version;    /* format version number */
+
+       fmt_validate_t  fmt_validate;   /* validate context flags */
+       fmt_getsize_t   fmt_getsize;    /* get size for sampling buffer */
+       fmt_init_t      fmt_init;       /* initialize buffer area */
+       fmt_handler_t   fmt_handler;    /* overflow handler (required) */
+       fmt_restart_t   fmt_restart;    /* restart after notification  */
+       fmt_exit_t      fmt_exit;       /* context termination */
+
+       struct list_head fmt_list;      /* internal use only */
+
+       struct kobject  kobj;           /* sysfs internal use only */
+       struct module   *owner;         /* pointer to module owner */
+       u32             fmt_qdepth;     /* Max notify queue depth (required) */
+};
+#define to_smpl_fmt(n) container_of(n, struct pfm_smpl_fmt, kobj)
+
+#define PFM_FMTFL_IS_BUILTIN   0x1     /* fmt is compiled in */
+/*
+ * we need to know whether the format is builtin or compiled
+ * as a module
+ */
+#ifdef MODULE
+#define PFM_FMT_BUILTIN_FLAG   0       /* not built as a module */
+#else
+#define PFM_FMT_BUILTIN_FLAG   PFM_PMUFL_IS_BUILTIN /* built as a module */
+#endif
+
+int pfm_fmt_register(struct pfm_smpl_fmt *fmt);
+int pfm_fmt_unregister(struct pfm_smpl_fmt *fmt);
+void pfm_sysfs_builtin_fmt_add(void);
+
+int  pfm_sysfs_add_fmt(struct pfm_smpl_fmt *fmt);
+void pfm_sysfs_remove_fmt(struct pfm_smpl_fmt *fmt);
+
+#endif /* __PERFMON_FMT_H__ */
diff --git a/include/linux/perfmon_kern.h b/include/linux/perfmon_kern.h
new file mode 100644
index 0000000..6c3b527
--- /dev/null
+++ b/include/linux/perfmon_kern.h
@@ -0,0 +1,551 @@
+/*
+ * Copyright (c) 2001-2006 Hewlett-Packard Development Company, L.P.
+ * Contributed by Stephane Eranian <eranian@hpl.hp.com>
+ *
+ * This program is free software; you can redistribute it and/or
+ * modify it under the terms of version 2 of the GNU General Public
+ * License as published by the Free Software Foundation.
+ *
+ * This program is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+ * General Public License for more details.
+ *
+ * You should have received a copy of the GNU General Public License
+ * along with this program; if not, write to the Free Software
+ * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA
+ * 02111-1307 USA
+ */
+
+#ifndef __LINUX_PERFMON_KERN_H__
+#define __LINUX_PERFMON_KERN_H__
+/*
+ * This file contains all the definitions of data structures, variables, macros
+ * that are to be shared between generic code and arch-specific code
+ *
+ * For generic only definitions, use perfmon/perfmon_priv.h
+ */
+#ifdef CONFIG_PERFMON
+
+#include <linux/file.h>
+#include <linux/sched.h>
+#include <linux/perfmon.h>
+
+/*
+ * system adminstrator configuration controls available via
+ * the /sys/kerne/perfmon interface
+ */
+struct pfm_controls {
+       u32     debug;          /* debugging control bitmask */
+       gid_t   sys_group;      /* gid to create a syswide context */
+       gid_t   task_group;     /* gid to create a per-task context */
+       u32     flags;          /* control flags (see below) */
+       size_t  arg_mem_max;    /* maximum vector argument size */
+       size_t  smpl_buffer_mem_max; /* max buf mem, -1 for infinity */
+};
+extern struct pfm_controls pfm_controls;
+
+/*
+ * control flags
+ */
+#define PFM_CTRL_FL_RW_EXPERT  0x1 /* bypass reserved fields on read/write */
+
+/*
+ * software PMD
+ */
+struct pfm_pmd {
+       u64 value;                      /* 64-bit value */
+       u64 lval;                       /* last reset value */
+       u64 ovflsw_thres;               /* #ovfls left before switch */
+       u64 long_reset;                 /* long reset value on overflow */
+       u64 short_reset;                /* short reset value on overflow */
+       u64 reset_pmds[PFM_PMD_BV];     /* pmds to reset on overflow */
+       u64 smpl_pmds[PFM_PMD_BV];      /* pmds to record on overflow */
+       u64 mask;                       /* range mask for random value */
+       u64 ovflsw_ref_thres;           /* #ovfls before next set */
+       u64 eventid;                    /* opaque event identifier */
+       u32 flags;                      /* notify/do not notify */
+};
+
+/*
+ * event_set: encapsulates the full PMU state
+ */
+struct pfm_event_set {
+       struct list_head list;          /* ordered chain of sets */
+       u16 id;                         /* set identification */
+       u16 nused_pmds;                 /* max number of used PMDs */
+       u16 nused_pmcs;                 /* max number of used PMCs */
+       u16 pad1;                       /* paddding */
+       u32 flags;                      /* public flags */
+       u32 priv_flags;                 /* private flags (see below) */
+       u64 runs;                       /* # of activations */
+       u32 npend_ovfls;                /* number of pending PMD overflow */
+       u32 pad2;                       /* padding */
+       u64 used_pmds[PFM_PMD_BV];      /* used PMDs */
+       u64 povfl_pmds[PFM_PMD_BV];     /* pending overflowed PMDs */
+       u64 ovfl_pmds[PFM_PMD_BV];      /* last overflowed PMDs */
+       u64 reset_pmds[PFM_PMD_BV];     /* PMDs to reset after overflow */
+       u64 ovfl_notify[PFM_PMD_BV];    /* notify on overflow */
+       u64 used_pmcs[PFM_PMC_BV];      /* used PMCs */
+       u64 pmcs[PFM_MAX_PMCS];         /* PMC values */
+
+       struct pfm_pmd pmds[PFM_MAX_PMDS];
+
+       ktime_t hrtimer_exp;            /* switch timeout reference */
+       ktime_t hrtimer_rem;            /* per-thread remainder timeout */
+
+       u64 duration_start;             /* start time in ns */
+       u64 duration;                   /* total active ns */
+};
+
+/*
+ * common private event set flags (priv_flags)
+ *
+ * upper 16 bits: for arch-specific use
+ * lower 16 bits: for common use
+ */
+#define PFM_SETFL_PRIV_MOD_PMDS 0x1 /* PMD register(s) modified */
+#define PFM_SETFL_PRIV_MOD_PMCS 0x2 /* PMC register(s) modified */
+#define PFM_SETFL_PRIV_SWITCH  0x4 /* must switch set on restart */
+#define PFM_SETFL_PRIV_MOD_BOTH        (PFM_SETFL_PRIV_MOD_PMDS \
+                               | PFM_SETFL_PRIV_MOD_PMCS)
+
+/*
+ * context flags
+ */
+struct pfm_context_flags {
+       unsigned int block:1;           /* task blocks on user notifications */
+       unsigned int system:1;          /* do system wide monitoring */
+       unsigned int no_msg:1;          /* no message sent on overflow */
+       unsigned int switch_ovfl:1;     /* switch set on counter ovfl */
+       unsigned int switch_time:1;     /* switch set on timeout */
+       unsigned int started:1;         /* pfm_start() issued */
+       unsigned int work_type:2;       /* type of work for pfm_handle_work */
+       unsigned int mmap_nlock:1;      /* no lock in pfm_release_buf_space */
+       unsigned int ia64_v20_compat:1; /* context is IA-64 v2.0 mode */
+       unsigned int can_restart:8;     /* allowed to issue a PFM_RESTART */
+       unsigned int reset_count:8;     /* number of pending resets */
+       unsigned int is_self:1;         /* per-thread and self-montoring */
+       unsigned int reserved:5;        /* for future use */
+};
+
+/*
+ * values for work_type (TIF_PERFMON_WORK must be set)
+ */
+#define PFM_WORK_NONE  0       /* nothing to do */
+#define PFM_WORK_RESET 1       /* reset overflowed counters */
+#define PFM_WORK_BLOCK 2       /* block current thread */
+#define PFM_WORK_ZOMBIE        3       /* cleanup zombie context */
+
+/*
+ * overflow description argument passed to sampling format
+ */
+struct pfm_ovfl_arg {
+       u16 ovfl_pmd;           /* index of overflowed PMD  */
+       u16 active_set;         /* set active at the time of the overflow */
+       u32 ovfl_ctrl;          /* control flags */
+       u64 pmd_last_reset;     /* last reset value of overflowed PMD */
+       u64 smpl_pmds_values[PFM_MAX_PMDS]; /* values of other PMDs */
+       u64 pmd_eventid;        /* eventid associated with PMD */
+       u16 num_smpl_pmds;      /* number of PMDS in smpl_pmd_values */
+};
+/*
+ * depth of message queue
+ *
+ * Depth cannot be bigger than 255 (see reset_count)
+ */
+#define PFM_MSGS_ORDER         3 /* log2(number of messages) */
+#define PFM_MSGS_COUNT         (1<<PFM_MSGS_ORDER) /* number of messages */
+#define PFM_MSGQ_MASK          (PFM_MSGS_COUNT-1)
+
+/*
+ * perfmon context state
+ */
+#define PFM_CTX_UNLOADED       1 /* context is not loaded onto any task */
+#define PFM_CTX_LOADED         2 /* context is loaded onto a task */
+#define PFM_CTX_MASKED         3 /* context is loaded, monitoring is masked */
+#define PFM_CTX_ZOMBIE         4 /* context lost owner but still attached */
+
+/*
+ * registers description
+ */
+struct pfm_regdesc {
+       u64 pmcs[PFM_PMC_BV];           /* available PMC */
+       u64 pmds[PFM_PMD_BV];           /* available PMD */
+       u64 rw_pmds[PFM_PMD_BV];        /* available RW PMD */
+       u64 intr_pmds[PFM_PMD_BV];      /* PMD generating intr */
+       u64 cnt_pmds[PFM_PMD_BV];       /* PMD counters */
+       u16 max_pmc;                    /* highest+1 avail PMC */
+       u16 max_pmd;                    /* highest+1 avail PMD */
+       u16 max_rw_pmd;                 /* highest+1 avail RW PMD */
+       u16 first_intr_pmd;             /* first intr PMD */
+       u16 max_intr_pmd;               /* highest+1 intr PMD */
+       u16 num_rw_pmd;                 /* number of avail RW PMD */
+       u16 num_pmcs;                   /* number of logical PMCS */
+       u16 num_pmds;                   /* number of logical PMDS */
+       u16 num_counters;               /* number of counting PMD */
+};
+
+/*
+ * context: contains all the state of a session
+ */
+struct pfm_context {
+       spinlock_t              lock;           /* context protection */
+
+       struct pfm_context_flags flags;
+       u32                     state;          /* current state */
+       struct task_struct      *task;          /* attached task */
+
+       struct completion       restart_complete;/* block on notification */
+       u64                     last_act;       /* last activation */
+       u32                     last_cpu;       /* last CPU used (SMP only) */
+       u32                     cpu;            /* cpu bound to context */
+
+       struct pfm_smpl_fmt     *smpl_fmt;      /* sampling format callbacks */
+       void                    *smpl_addr;     /* user smpl buffer base */
+       size_t                  smpl_size;      /* user smpl buffer size */
+       void                    *smpl_real_addr;/* actual smpl buffer base */
+       size_t                  smpl_real_size; /* actual smpl buffer size */
+
+       wait_queue_head_t       msgq_wait;      /* pfm_read() wait queue */
+
+       union pfarg_msg         msgq[PFM_MSGS_COUNT];
+       int                     msgq_head;
+       int                     msgq_tail;
+
+       struct fasync_struct    *async_queue;   /* async notification */
+
+       struct pfm_event_set    *active_set;    /* active set */
+       struct list_head        set_list;       /* ordered list of sets */
+
+       struct pfm_regdesc      regs;           /* registers available to context */
+
+       /*
+        * save stack space by allocating temporary variables for
+        * pfm_overflow_handler() in pfm_context
+        */
+       struct pfm_ovfl_arg     ovfl_arg;
+       u64                     tmp_ovfl_notify[PFM_PMD_BV];
+};
+
+/*
+ * ovfl_ctrl bitmask (used by interrupt handler)
+ */
+#define PFM_OVFL_CTRL_NOTIFY   0x1     /* notify user */
+#define PFM_OVFL_CTRL_RESET    0x2     /* reset overflowed pmds */
+#define PFM_OVFL_CTRL_MASK     0x4     /* mask monitoring */
+#define PFM_OVFL_CTRL_SWITCH   0x8     /* switch sets */
+
+/*
+ * logging
+ */
+#define PFM_ERR(f, x...)  printk(KERN_ERR     "perfmon: " f "\n", ## x)
+#define PFM_WARN(f, x...) printk(KERN_WARNING "perfmon: " f "\n", ## x)
+#define PFM_LOG(f, x...)  printk(KERN_NOTICE  "perfmon: " f "\n", ## x)
+#define PFM_INFO(f, x...) printk(KERN_INFO    "perfmon: " f "\n", ## x)
+
+/*
+ * debugging
+ *
+ * Printk rate limiting is enforced to avoid getting flooded with too many
+ * error messages on the console (which could render the machine unresponsive).
+ * To get full debug output (turn off ratelimit):
+ *     $ echo 0 >/proc/sys/kernel/printk_ratelimit
+ *
+ * debug is a bitmask where bits are defined as follows:
+ * bit  0: enable non-interrupt code degbug messages
+ * bit  1: enable interrupt code debug messages
+ */
+#ifdef CONFIG_PERFMON_DEBUG
+#define _PFM_DBG(lm, f, x...) \
+       do { \
+               if (unlikely((pfm_controls.debug & lm) && printk_ratelimit())) { \
+                       preempt_disable(); \
+                       printk("perfmon: %s.%d: CPU%d [%d]: " f "\n", \
+                              __func__, __LINE__, \
+                              smp_processor_id(), current->pid , ## x); \
+                       preempt_enable(); \
+               } \
+       } while (0)
+
+#define PFM_DBG(f, x...) _PFM_DBG(0x1, f, ##x)
+#define PFM_DBG_ovfl(f, x...) _PFM_DBG(0x2, f, ## x)
+#else
+#define PFM_DBG(f, x...)       do {} while (0)
+#define PFM_DBG_ovfl(f, x...)  do {} while (0)
+#endif
+
+extern struct pfm_pmu_config  *pfm_pmu_conf;
+extern int perfmon_disabled;
+
+static inline struct pfm_arch_context *pfm_ctx_arch(struct pfm_context *c)
+{
+       return (struct pfm_arch_context *)(c+1);
+}
+
+int  pfm_get_args(void __user *ureq, size_t sz, size_t lsz, void *laddr,
+                 void **req, void **to_free);
+
+int pfm_get_smpl_arg(char __user *fmt_uname, void __user *uaddr, size_t usize,
+                    void **arg, struct pfm_smpl_fmt **fmt);
+
+int __pfm_write_pmcs(struct pfm_context *ctx, struct pfarg_pmc *req,
+                    int count);
+int __pfm_write_pmds(struct pfm_context *ctx, struct pfarg_pmd *req, int count,
+                    int compat);
+int __pfm_read_pmds(struct pfm_context *ctx, struct pfarg_pmd *req, int count);
+
+int __pfm_load_context(struct pfm_context *ctx, struct pfarg_load *req,
+                      struct task_struct *task);
+int __pfm_unload_context(struct pfm_context *ctx, int *can_release);
+
+int __pfm_stop(struct pfm_context *ctx, int *release_info);
+int  __pfm_restart(struct pfm_context *ctx, int *unblock);
+int __pfm_start(struct pfm_context *ctx, struct pfarg_start *start);
+
+void pfm_free_context(struct pfm_context *ctx);
+
+void pfm_smpl_buf_space_release(struct pfm_context *ctx, size_t size);
+
+int pfm_check_task_state(struct pfm_context *ctx, int check_mask,
+                        unsigned long *flags, void **resume);
+/*
+ * check_mask bitmask values for pfm_check_task_state()
+ */
+#define PFM_CMD_STOPPED                0x01    /* command needs thread stopped */
+#define PFM_CMD_UNLOADED       0x02    /* command needs ctx unloaded */
+#define PFM_CMD_UNLOAD         0x04    /* command is unload */
+
+int __pfm_create_context(struct pfarg_ctx *req,
+                        struct pfm_smpl_fmt *fmt,
+                        void *fmt_arg,
+                        int mode,
+                        struct pfm_context **new_ctx);
+
+struct pfm_event_set *pfm_find_set(struct pfm_context *ctx, u16 set_id,
+                                  int alloc);
+
+int pfm_pmu_conf_get(int autoload);
+void pfm_pmu_conf_put(void);
+
+int pfm_session_allcpus_acquire(void);
+void pfm_session_allcpus_release(void);
+
+int pfm_smpl_buf_alloc(struct pfm_context *ctx, size_t rsize);
+void pfm_smpl_buf_free(struct pfm_context *ctx);
+
+struct pfm_smpl_fmt *pfm_smpl_fmt_get(char *name);
+void pfm_smpl_fmt_put(struct pfm_smpl_fmt *fmt);
+
+void pfm_interrupt_handler(unsigned long iip, struct pt_regs *regs);
+
+void pfm_resume_task(struct task_struct *t, void *data);
+
+#include <linux/perfmon_pmu.h>
+#include <linux/perfmon_fmt.h>
+
+extern const struct file_operations pfm_file_ops;
+/*
+ * upper limit for count in calls that take vector arguments. This is used
+ * to prevent for multiplication overflow when we compute actual storage size
+ */
+#define PFM_MAX_ARG_COUNT(m) (INT_MAX/sizeof(*(m)))
+
+#define cast_ulp(_x) ((unsigned long *)_x)
+
+#define PFM_NORMAL      0
+#define PFM_COMPAT      1
+
+void __pfm_exit_thread(void);
+void pfm_ctxsw_in(struct task_struct *prev, struct task_struct *next);
+void pfm_ctxsw_out(struct task_struct *prev, struct task_struct *next);
+void pfm_handle_work(struct pt_regs *regs);
+void __pfm_init_percpu(void *dummy);
+void pfm_save_pmds(struct pfm_context *ctx, struct pfm_event_set *set);
+
+static inline void pfm_exit_thread(void)
+{
+       if (current->pfm_context)
+               __pfm_exit_thread();
+}
+
+/*
+ * include arch-specific kernel level definitions
+ */
+#include <asm/perfmon_kern.h>
+
+static inline void pfm_copy_thread(struct task_struct *task)
+{
+       /*
+        * context or perfmon TIF state  is NEVER inherited
+        * in child task. Holds for per-thread and system-wide
+        */
+       task->pfm_context = NULL;
+       clear_tsk_thread_flag(task, TIF_PERFMON_CTXSW);
+       clear_tsk_thread_flag(task, TIF_PERFMON_WORK);
+       pfm_arch_disarm_handle_work(task);
+}
+
+
+/*
+ * read a single PMD register.
+ *
+ * virtual PMD registers have special handler.
+ * Depends on definitions in asm/perfmon_kern.h
+ */
+static inline u64 pfm_read_pmd(struct pfm_context *ctx, unsigned int cnum)
+{
+       if (unlikely(pfm_pmu_conf->pmd_desc[cnum].type & PFM_REG_V))
+               return pfm_pmu_conf->pmd_sread(ctx, cnum);
+
+       return pfm_arch_read_pmd(ctx, cnum);
+}
+/*
+ * write a single PMD register.
+ *
+ * virtual PMD registers have special handler.
+ * Depends on definitions in asm/perfmon_kern.h
+ */
+static inline void pfm_write_pmd(struct pfm_context *ctx, unsigned int cnum,
+                                u64 value)
+{
+       /*
+        * PMD writes are ignored for read-only registers
+        */
+       if (pfm_pmu_conf->pmd_desc[cnum].type & PFM_REG_RO)
+               return;
+
+       if (pfm_pmu_conf->pmd_desc[cnum].type & PFM_REG_V) {
+               pfm_pmu_conf->pmd_swrite(ctx, cnum, value);
+               return;
+       }
+       /*
+        * clear unimplemented bits
+        */
+       value &= ~pfm_pmu_conf->pmd_desc[cnum].rsvd_msk;
+
+       pfm_arch_write_pmd(ctx, cnum, value);
+}
+
+void __pfm_init_percpu(void *dummy);
+
+static inline void pfm_init_percpu(void)
+{
+       __pfm_init_percpu(NULL);
+}
+
+/*
+ * pfm statistics are available via debugfs
+ * and perfmon subdir.
+ *
+ * When adding/removing new stats, make sure you also
+ * update the name table in perfmon_debugfs.c
+ */
+enum pfm_stats_names {
+       PFM_ST_ovfl_intr_all_count = 0,
+       PFM_ST_ovfl_intr_ns,
+       PFM_ST_ovfl_intr_spurious_count,
+       PFM_ST_ovfl_intr_replay_count,
+       PFM_ST_ovfl_intr_regular_count,
+       PFM_ST_handle_work_count,
+       PFM_ST_ovfl_notify_count,
+       PFM_ST_reset_pmds_count,
+       PFM_ST_pfm_restart_count,
+       PFM_ST_fmt_handler_calls,
+       PFM_ST_fmt_handler_ns,
+       PFM_ST_set_switch_count,
+       PFM_ST_set_switch_ns,
+       PFM_ST_set_switch_exp,
+       PFM_ST_ctxswin_count,
+       PFM_ST_ctxswin_ns,
+       PFM_ST_handle_timeout_count,
+       PFM_ST_ovfl_intr_nmi_count,
+       PFM_ST_ctxswout_count,
+       PFM_ST_ctxswout_ns,
+       PFM_ST_LAST     /* last entry marked */
+};
+#define PFM_NUM_STATS PFM_ST_LAST
+
+struct pfm_stats {
+       u64 v[PFM_NUM_STATS];
+       struct dentry *dirs[PFM_NUM_STATS];
+       struct dentry *cpu_dir;
+       char cpu_name[8];
+};
+
+#ifdef CONFIG_PERFMON_DEBUG_FS
+#define pfm_stats_get(x)  __get_cpu_var(pfm_stats).v[PFM_ST_##x]
+#define pfm_stats_inc(x)  __get_cpu_var(pfm_stats).v[PFM_ST_##x]++
+#define pfm_stats_add(x, y)  __get_cpu_var(pfm_stats).v[PFM_ST_##x] += (y)
+void pfm_reset_stats(int cpu);
+#else
+#define pfm_stats_get(x)
+#define pfm_stats_inc(x)
+#define pfm_stats_add(x, y)
+static inline void pfm_reset_stats(int cpu)
+{}
+#endif
+
+
+
+DECLARE_PER_CPU(struct pfm_context *, pmu_ctx);
+DECLARE_PER_CPU(struct pfm_stats, pfm_stats);
+DECLARE_PER_CPU(struct task_struct *, pmu_owner);
+
+void pfm_cpu_disable(void);
+
+
+/*
+ * max vector argument elements for local storage (no kmalloc/kfree)
+ * The PFM_ARCH_PM*_ARG should be defined in perfmon_kern.h.
+ * If not, default (conservative) values are used
+ */
+#ifndef PFM_ARCH_PMC_STK_ARG
+#define PFM_ARCH_PMC_STK_ARG   1
+#endif
+
+#ifndef PFM_ARCH_PMD_STK_ARG
+#define PFM_ARCH_PMD_STK_ARG   1
+#endif
+
+#define PFM_PMC_STK_ARG        PFM_ARCH_PMC_STK_ARG
+#define PFM_PMD_STK_ARG        PFM_ARCH_PMD_STK_ARG
+
+#else /* !CONFIG_PERFMON */
+
+
+/*
+ * perfmon hooks are nops when CONFIG_PERFMON is undefined
+ */
+static inline void pfm_cpu_disable(void)
+{}
+
+static inline void pfm_exit_thread(void)
+{}
+
+static inline void pfm_handle_work(struct pt_regs *regs)
+{}
+
+static inline void pfm_copy_thread(struct task_struct *t)
+{}
+
+static inline void pfm_ctxsw_in(struct task_struct *p, struct task_struct *n)
+{}
+
+static inline void pfm_ctxsw_out(struct task_struct *p, struct task_struct *n)
+{}
+
+static inline void pfm_session_allcpus_release(void)
+{}
+
+static inline int pfm_session_allcpus_acquire(void)
+{
+       return 0;
+}
+
+static inline void pfm_init_percpu(void)
+{}
+
+#endif /* CONFIG_PERFMON */
+
+#endif /* __LINUX_PERFMON_KERN_H__ */
diff --git a/include/linux/perfmon_pmu.h b/include/linux/perfmon_pmu.h
new file mode 100644
index 0000000..3f5f9e8
--- /dev/null
+++ b/include/linux/perfmon_pmu.h
@@ -0,0 +1,192 @@
+/*
+ * Copyright (c) 2006 Hewlett-Packard Development Company, L.P.
+ * Contributed by Stephane Eranian <eranian@hpl.hp.com>
+ *
+ * Interface for PMU description modules
+ *
+ * This program is free software; you can redistribute it and/or
+ * modify it under the terms of version 2 of the GNU General Public
+ * License as published by the Free Software Foundation.
+ *
+ * This program is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+ * General Public License for more details.
+ *
+ * You should have received a copy of the GNU General Public License
+ * along with this program; if not, write to the Free Software
+ * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA
+ * 02111-1307 USA
+ */
+#ifndef __PERFMON_PMU_H__
+#define __PERFMON_PMU_H__ 1
+
+/*
+ * generic information about a PMC or PMD register
+ *
+ * Dependency bitmasks:
+ *  They are used to allow lazy save/restore in the context switch
+ *  code. To avoid picking up stale configuration from a previous
+ *  thread. Usng the bitmask, the generic read/write routines can
+ *  ensure that all registers needed to support the measurement are
+ *  restored properly on context switch in.
+ */
+struct pfm_regmap_desc {
+       u16  type;              /* role of the register */
+       u16  reserved1;         /* for future use */
+       u32  reserved2;         /* for future use */
+       u64  dfl_val;           /* power-on default value (quiescent) */
+       u64  rsvd_msk;          /* reserved bits: 1 means reserved */
+       u64  no_emul64_msk;     /* bits to clear for PFM_REGFL_NO_EMUL64 */
+       unsigned long hw_addr;  /* HW register address or index */
+       struct kobject  kobj;   /* for internal use only */
+       char *desc;             /* HW register description string */
+       u64 dep_pmcs[PFM_PMC_BV];/* depending PMC registers */
+};
+#define to_reg(n) container_of(n, struct pfm_regmap_desc, kobj)
+
+/*
+ * pfm_reg_desc helper macros
+ */
+#define PMC_D(t, d, v, r, n, h) \
+       { .type = t,          \
+         .desc = d,          \
+         .dfl_val = v,       \
+         .rsvd_msk = r,      \
+         .no_emul64_msk = n, \
+         .hw_addr = h        \
+       }
+
+#define PMD_D(t, d, h)        \
+       { .type = t,          \
+         .desc = d,          \
+         .rsvd_msk = 0,      \
+         .no_emul64_msk = 0, \
+         .hw_addr = h        \
+       }
+
+#define PMD_DR(t, d, h, r)    \
+       { .type = t,          \
+         .desc = d,          \
+         .rsvd_msk = r,      \
+         .no_emul64_msk = 0, \
+         .hw_addr = h        \
+       }
+
+#define PMX_NA \
+       { .type = PFM_REG_NA }
+
+#define PMD_DP(t, d, h, p)    \
+       { .type = t,          \
+         .desc = d,          \
+         .rsvd_msk = 0,      \
+         .no_emul64_msk = 0, \
+         .dep_pmcs[0] = p,   \
+         .hw_addr = h        \
+       }
+
+/*
+ * type of a PMU register (16-bit bitmask) for use with pfm_reg_desc.type
+ */
+#define PFM_REG_NA     0x00  /* not avail. (not impl.,no access) must be 0 */
+#define PFM_REG_I      0x01  /* PMC/PMD: implemented */
+#define PFM_REG_WC     0x02  /* PMC: has write_checker */
+#define PFM_REG_C64    0x04  /* PMD: 64-bit virtualization */
+#define PFM_REG_RO     0x08  /* PMD: read-only (writes ignored) */
+#define PFM_REG_V      0x10  /* PMD: virtual reg */
+#define PFM_REG_INTR   0x20  /* PMD: register can generate interrupt */
+#define PFM_REG_SYS    0x40  /* PMC/PMD: register is for system-wide only */
+#define PFM_REG_THR    0x80  /* PMC/PMD: register is for per-thread only */
+#define PFM_REG_NO64   0x100 /* PMC: supports PFM_REGFL_NO_EMUL64 */
+
+/*
+ * define some shortcuts for common types
+ */
+#define PFM_REG_W      (PFM_REG_WC|PFM_REG_I)
+#define PFM_REG_W64    (PFM_REG_WC|PFM_REG_NO64|PFM_REG_I)
+#define PFM_REG_C      (PFM_REG_C64|PFM_REG_INTR|PFM_REG_I)
+#define PFM_REG_I64    (PFM_REG_NO64|PFM_REG_I)
+#define PFM_REG_IRO    (PFM_REG_I|PFM_REG_RO)
+
+typedef int (*pfm_pmc_check_t)(struct pfm_context *ctx,
+                              struct pfm_event_set *set,
+                              struct pfarg_pmc *req);
+
+typedef int (*pfm_pmd_check_t)(struct pfm_context *ctx,
+                              struct pfm_event_set *set,
+                              struct pfarg_pmd *req);
+
+
+typedef u64 (*pfm_sread_t)(struct pfm_context *ctx, unsigned int cnum);
+typedef void (*pfm_swrite_t)(struct pfm_context *ctx, unsigned int cnum, u64 val);
+
+/*
+ * structure used by pmu description modules
+ *
+ * probe_pmu() routine return value:
+ *     - 1 means recognized PMU
+ *     - 0 means not recognized PMU
+ */
+struct pfm_pmu_config {
+       char *pmu_name;                         /* PMU family name */
+       char *version;                          /* config module version */
+
+       int counter_width;                      /* width of hardware counter */
+
+       struct pfm_regmap_desc  *pmc_desc;      /* PMC register descriptions */
+       struct pfm_regmap_desc  *pmd_desc;      /* PMD register descriptions */
+
+       pfm_pmc_check_t         pmc_write_check;/* write checker (optional) */
+       pfm_pmd_check_t         pmd_write_check;/* write checker (optional) */
+       pfm_pmd_check_t         pmd_read_check; /* read checker (optional) */
+
+       pfm_sread_t             pmd_sread;      /* virtual pmd read */
+       pfm_swrite_t            pmd_swrite;     /* virtual pmd write */
+
+       int                     (*probe_pmu)(void);/* probe PMU routine */
+
+       u16                     num_pmc_entries;/* #entries in pmc_desc */
+       u16                     num_pmd_entries;/* #entries in pmd_desc */
+
+       void                    *pmu_info;      /* model-specific infos */
+       u32                     flags;          /* set of flags */
+
+       struct module           *owner;         /* pointer to module struct */
+
+       /*
+        * fields computed internally, do not set in module
+        */
+       struct pfm_regdesc      regs_all;       /* regs available to all */
+       struct pfm_regdesc      regs_thr;       /* regs avail per-thread */
+       struct pfm_regdesc      regs_sys;       /* regs avail system-wide */
+
+       u64                     ovfl_mask;      /* overflow mask */
+};
+
+static inline void *pfm_pmu_info(void)
+{
+       return pfm_pmu_conf->pmu_info;
+}
+
+/*
+ * pfm_pmu_config flags
+ */
+#define PFM_PMUFL_IS_BUILTIN   0x1     /* pmu config is compiled in */
+
+/*
+ * we need to know whether the PMU description is builtin or compiled
+ * as a module
+ */
+#ifdef MODULE
+#define PFM_PMU_BUILTIN_FLAG   0       /* not built as a module */
+#else
+#define PFM_PMU_BUILTIN_FLAG   PFM_PMUFL_IS_BUILTIN /* built as a module */
+#endif
+
+int pfm_pmu_register(struct pfm_pmu_config *cfg);
+void pfm_pmu_unregister(struct pfm_pmu_config *cfg);
+
+int pfm_sysfs_remove_pmu(struct pfm_pmu_config *pmu);
+int pfm_sysfs_add_pmu(struct pfm_pmu_config *pmu);
+
+#endif /* __PERFMON_PMU_H__ */
diff --git a/include/linux/sched.h b/include/linux/sched.h
index 3d9120c..8fb3b55 100644
--- a/include/linux/sched.h
+++ b/include/linux/sched.h
@@ -96,6 +96,7 @@ struct exec_domain;
 struct futex_pi_state;
 struct robust_list_head;
 struct bio;
+struct pfm_context;
 
 /*
  * List of flags we want to share for kernel threads,
@@ -1301,6 +1302,9 @@ struct task_struct {
        int latency_record_count;
        struct latency_record latency_record[LT_SAVECOUNT];
 #endif
+#ifdef CONFIG_PERFMON
+       struct pfm_context *pfm_context;
+#endif
 };
 
 /*
diff --git a/include/linux/syscalls.h b/include/linux/syscalls.h
index d6ff145..e308523 100644
--- a/include/linux/syscalls.h
+++ b/include/linux/syscalls.h
@@ -29,6 +29,13 @@ struct msqid_ds;
 struct new_utsname;
 struct nfsctl_arg;
 struct __old_kernel_stat;
+struct pfarg_ctx;
+struct pfarg_pmc;
+struct pfarg_pmd;
+struct pfarg_start;
+struct pfarg_load;
+struct pfarg_setinfo;
+struct pfarg_setdesc;
 struct pollfd;
 struct rlimit;
 struct rusage;
@@ -625,4 +632,27 @@ asmlinkage long sys_fallocate(int fd, int mode, loff_t offset, loff_t len);
 
 int kernel_execve(const char *filename, char *const argv[], char *const envp[]);
 
+asmlinkage long sys_pfm_create_context(struct pfarg_ctx __user *ureq,
+                                      void __user *uarg, size_t smpl_size);
+asmlinkage long sys_pfm_write_pmcs(int fd, struct pfarg_pmc __user *ureq,
+                                  int count);
+asmlinkage long sys_pfm_write_pmds(int fd, struct pfarg_pmd __user *ureq,
+                                  int count);
+asmlinkage long sys_pfm_read_pmds(int fd, struct pfarg_pmd __user *ureq,
+                                 int count);
+asmlinkage long sys_pfm_restart(int fd);
+asmlinkage long sys_pfm_stop(int fd);
+asmlinkage long sys_pfm_start(int fd, struct pfarg_start __user *ureq);
+asmlinkage long sys_pfm_load_context(int fd, struct pfarg_load __user *ureq);
+asmlinkage long sys_pfm_unload_context(int fd);
+asmlinkage long sys_pfm_delete_evtsets(int fd,
+                                      struct pfarg_setinfo __user *ureq,
+                                      int count);
+asmlinkage long sys_pfm_create_evtsets(int fd,
+                                      struct pfarg_setdesc __user *ureq,
+                                      int count);
+asmlinkage long sys_pfm_getinfo_evtsets(int fd,
+                                       struct pfarg_setinfo __user *ureq,
+                                       int count);
+
 #endif
diff --git a/kernel/sched.c b/kernel/sched.c
index ad1962d..1bc8fcf 100644
--- a/kernel/sched.c
+++ b/kernel/sched.c
@@ -71,6 +71,7 @@
 #include <linux/debugfs.h>
 #include <linux/ctype.h>
 #include <linux/ftrace.h>
+#include <linux/perfmon_kern.h>
 
 #include <asm/tlb.h>
 #include <asm/irq_regs.h>
diff --git a/kernel/sys_ni.c b/kernel/sys_ni.c
index 08d6e1b..61f4155 100644
--- a/kernel/sys_ni.c
+++ b/kernel/sys_ni.c
@@ -126,6 +126,19 @@ cond_syscall(sys_vm86);
 cond_syscall(compat_sys_ipc);
 cond_syscall(compat_sys_sysctl);
 
+cond_syscall(sys_pfm_create_context);
+cond_syscall(sys_pfm_write_pmcs);
+cond_syscall(sys_pfm_write_pmds);
+cond_syscall(sys_pfm_read_pmds);
+cond_syscall(sys_pfm_restart);
+cond_syscall(sys_pfm_start);
+cond_syscall(sys_pfm_stop);
+cond_syscall(sys_pfm_load_context);
+cond_syscall(sys_pfm_unload_context);
+cond_syscall(sys_pfm_create_evtsets);
+cond_syscall(sys_pfm_delete_evtsets);
+cond_syscall(sys_pfm_getinfo_evtsets);
+
 /* arch-specific weak syscall entries */
 cond_syscall(sys_pciconfig_read);
 cond_syscall(sys_pciconfig_write);
diff --git a/perfmon/Makefile b/perfmon/Makefile
new file mode 100644
index 0000000..32ff037
--- /dev/null
+++ b/perfmon/Makefile
@@ -0,0 +1,12 @@
+#
+# Copyright (c) 2005-2006 Hewlett-Packard Development Company, L.P.
+# Contributed by Stephane Eranian <eranian@hpl.hp.com>
+#
+obj-y = perfmon_init.o perfmon_rw.o perfmon_res.o           \
+       perfmon_pmu.o perfmon_sysfs.o perfmon_syscalls.o    \
+       perfmon_file.o perfmon_ctxsw.o perfmon_intr.o       \
+       perfmon_dfl_smpl.o perfmon_sets.o perfmon_hotplug.o \
+       perfmon_msg.o perfmon_smpl.o perfmon_attach.o       \
+       perfmon_activate.o perfmon_ctx.o perfmon_fmt.o
+
+obj-$(CONFIG_PERFMON_DEBUG_FS) +=  perfmon_debugfs.o
diff --git a/perfmon/perfmon_activate.c b/perfmon/perfmon_activate.c
new file mode 100644
index 0000000..d9f501d
--- /dev/null
+++ b/perfmon/perfmon_activate.c
@@ -0,0 +1,265 @@
+/*
+ * perfmon_activate.c: perfmon2 start/stop functions
+ *
+ * This file implements the perfmon2 interface which
+ * provides access to the hardware performance counters
+ * of the host processor.
+ *
+ *
+ * The initial version of perfmon.c was written by
+ * Ganesh Venkitachalam, IBM Corp.
+ *
+ * Then it was modified for perfmon-1.x by Stephane Eranian and
+ * David Mosberger, Hewlett Packard Co.
+ *
+ * Version Perfmon-2.x is a complete rewrite of perfmon-1.x
+ * by Stephane Eranian, Hewlett Packard Co.
+ *
+ * Copyright (c) 1999-2006 Hewlett-Packard Development Company, L.P.
+ * Contributed by Stephane Eranian <eranian@hpl.hp.com>
+ *                David Mosberger-Tang <davidm@hpl.hp.com>
+ *
+ * More information about perfmon available at:
+ *     http://perfmon2.sf.net
+ *
+ * This program is free software; you can redistribute it and/or
+ * modify it under the terms of version 2 of the GNU General Public
+ * License as published by the Free Software Foundation.
+ *
+ * This program is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+ * General Public License for more details.
+ *
+ * You should have received a copy of the GNU General Public License
+ * along with this program; if not, write to the Free Software
+ * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA
+ * 02111-1307 USA
+ */
+#include <linux/kernel.h>
+#include <linux/perfmon_kern.h>
+#include "perfmon_priv.h"
+
+/**
+ * __pfm_start - activate monitoring
+ * @ctx: context to operate on
+ * @start: pfarg_start as passed by user
+ *
+ * When operating in per-thread mode and not self-monitoring, the monitored
+ * thread must be stopped. Activation will be effective next time the thread
+ * is context switched in.
+ *
+ * The pfarg_start argument is optional and may be used to designate
+ * the initial event set to activate. When not provided, the last active
+ * set is used. For the first activation, set0 is used when start is NULL.
+ *
+ * On some architectures, e.g., IA-64, it may be possible to start monitoring
+ * without calling this function under certain conditions (per-thread and self
+ * monitoring). In this case, either set0 or the last active set is used.
+ *
+ * the context is locked and interrupts are disabled.
+ */
+int __pfm_start(struct pfm_context *ctx, struct pfarg_start *start)
+{
+       struct task_struct *task, *owner_task;
+       struct pfm_event_set *new_set, *old_set;
+       int is_self;
+
+       task = ctx->task;
+
+       /*
+        * UNLOADED: error
+        * LOADED  : normal start, nop if started unless set is different
+        * MASKED  : nop or change set when unmasking
+        * ZOMBIE  : cannot happen
+        */
+       if (ctx->state == PFM_CTX_UNLOADED)
+               return -EINVAL;
+
+       old_set = new_set = ctx->active_set;
+
+       /*
+        * always the case for system-wide
+        */
+       if (task == NULL)
+               task = current;
+
+       is_self = task == current;
+
+       /*
+        * argument is provided?
+        */
+       if (start) {
+               /*
+                * find the set to load first
+                */
+               new_set = pfm_find_set(ctx, start->start_set, 0);
+               if (new_set == NULL) {
+                       PFM_DBG("event set%u does not exist",
+                               start->start_set);
+                       return -EINVAL;
+               }
+       }
+
+       PFM_DBG("cur_set=%u req_set=%u", old_set->id, new_set->id);
+
+       /*
+        * if we need to change the active set we need
+        * to check if we can access the PMU
+        */
+       if (new_set != old_set) {
+
+               owner_task = __get_cpu_var(pmu_owner);
+               /*
+                * system-wide: must run on the right CPU
+                * per-thread : must be the owner of the PMU context
+                *
+                * pfm_switch_sets() returns with monitoring stopped
+                */
+               if (is_self) {
+                       pfm_switch_sets(ctx, new_set, PFM_PMD_RESET_LONG, 1);
+               } else {
+                       /*
+                        * In a UP kernel, the PMU may contain the state
+                        * of the task we want to operate on, yet the task
+                        * may be switched out (lazy save). We need to save
+                        * current state (old_set), switch active_set and
+                        * mark it for reload.
+                        */
+                       if (owner_task == task)
+                               pfm_save_pmds(ctx, old_set);
+                       ctx->active_set = new_set;
+                       new_set->priv_flags |= PFM_SETFL_PRIV_MOD_BOTH;
+               }
+       }
+
+       /*
+        * mark as started
+        * must be done before calling pfm_arch_start()
+        */
+       ctx->flags.started = 1;
+
+       pfm_arch_start(task, ctx);
+
+       /*
+        * we check whether we had a pending ovfl before restarting.
+        * If so we need to regenerate the interrupt to make sure we
+        * keep recorded samples. For non-self monitoring this check
+        * is done in the pfm_ctxswin_thread() routine.
+        *
+        * we check new_set/old_set because pfm_switch_sets() already
+        * takes care of replaying the pending interrupts
+        */
+       if (is_self && new_set != old_set && new_set->npend_ovfls) {
+               pfm_arch_resend_irq(ctx);
+               pfm_stats_inc(ovfl_intr_replay_count);
+       }
+
+       /*
+        * always start with full timeout
+        */
+       new_set->hrtimer_rem = new_set->hrtimer_exp;
+
+       /*
+        * activate timeout for system-wide, self-montoring
+        * Always start with full timeout
+        * Timeout is at least one tick away, so no risk of
+        * having hrtimer_start() trying to wakeup softirqd
+        * and thus causing troubles. This cannot happen anmyway
+        * because cb_mode = HRTIMER_CB_IRQSAFE_NO_SOFTIRQ
+        */
+       if (is_self && new_set->flags & PFM_SETFL_TIME_SWITCH) {
+               hrtimer_start(&__get_cpu_var(pfm_hrtimer),
+                             new_set->hrtimer_rem,
+                             HRTIMER_MODE_REL);
+
+               PFM_DBG("set%u started timeout=%lld",
+                       new_set->id,
+                       (unsigned long long)new_set->hrtimer_rem.tv64);
+       }
+
+       /*
+        * we restart total duration even if context was
+        * already started. In that case, counts are simply
+        * reset.
+        *
+        * For per-thread, if not self-monitoring, the statement
+        * below will have no effect because thread is stopped.
+        * The field is reset of ctxsw in.
+        */
+       new_set->duration_start = sched_clock();
+
+       return 0;
+}
+
+/**
+ * __pfm_stop - stop monitoring
+ * @ctx: context to operate on
+ * @release_info: infos for caller (see below)
+ *
+ * When operating in per-thread* mode and when not self-monitoring,
+ * the monitored thread must be stopped.
+ *
+ * the context is locked and interrupts are disabled.
+ *
+ * release_info value upon return:
+ *     - bit 0 : unused
+ *     - bit 1 : when set, must cancel hrtimer
+ */
+int __pfm_stop(struct pfm_context *ctx, int *release_info)
+{
+       struct pfm_event_set *set;
+       struct task_struct *task;
+       u64 now;
+       int state;
+
+       *release_info = 0;
+
+       now = sched_clock();
+       state = ctx->state;
+       set = ctx->active_set;
+
+       /*
+        * context must be attached (zombie cannot happen)
+        */
+       if (state == PFM_CTX_UNLOADED)
+               return -EINVAL;
+
+       task = ctx->task;
+
+       PFM_DBG("ctx_task=[%d] ctx_state=%d is_system=%d",
+               task ? task->pid : -1,
+               state,
+               !task);
+
+       /*
+        * this happens for system-wide context
+        */
+       if (task == NULL)
+               task = current;
+
+       /*
+        * compute elapsed time
+        *
+        * unless masked, compute elapsed duration, stop timeout
+        */
+       if (task == current && state == PFM_CTX_LOADED) {
+               /*
+                * timeout cancel must be deferred until context is
+                * unlocked to avoid race with pfm_handle_switch_timeout()
+                */
+               if (set->flags & PFM_SETFL_TIME_SWITCH)
+                       *release_info |= 0x2;
+
+               set->duration += now - set->duration_start;
+       }
+
+       pfm_arch_stop(task, ctx);
+
+       ctx->flags.started = 0;
+       /*
+        * starting now, in-flight PMU interrupt for this context
+        * are treated as spurious
+        */
+       return 0;
+}
diff --git a/perfmon/perfmon_attach.c b/perfmon/perfmon_attach.c
new file mode 100644
index 0000000..bbd1d1e
--- /dev/null
+++ b/perfmon/perfmon_attach.c
@@ -0,0 +1,474 @@
+/*
+ * perfmon_attach.c: perfmon2 load/unload functions
+ *
+ * This file implements the perfmon2 interface which
+ * provides access to the hardware performance counters
+ * of the host processor.
+ *
+ *
+ * The initial version of perfmon.c was written by
+ * Ganesh Venkitachalam, IBM Corp.
+ *
+ * Then it was modified for perfmon-1.x by Stephane Eranian and
+ * David Mosberger, Hewlett Packard Co.
+ *
+ * Version Perfmon-2.x is a complete rewrite of perfmon-1.x
+ * by Stephane Eranian, Hewlett Packard Co.
+ *
+ * Copyright (c) 1999-2006 Hewlett-Packard Development Company, L.P.
+ * Contributed by Stephane Eranian <eranian@hpl.hp.com>
+ *                David Mosberger-Tang <davidm@hpl.hp.com>
+ *
+ * More information about perfmon available at:
+ *     http://perfmon2.sf.net
+ *
+ * This program is free software; you can redistribute it and/or
+ * modify it under the terms of version 2 of the GNU General Public
+ * License as published by the Free Software Foundation.
+ *
+ * This program is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+ * General Public License for more details.
+ *
+ * You should have received a copy of the GNU General Public License
+ * along with this program; if not, write to the Free Software
+ * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA
+ * 02111-1307 USA
+ */
+#include <linux/kernel.h>
+#include <linux/fs.h>
+#include <linux/perfmon_kern.h>
+#include "perfmon_priv.h"
+
+/**
+ * __pfm_load_context_sys - attach context to a CPU in system-wide mode
+ * @ctx: context to operate on
+ * @set_id: set to activate first
+ * @cpu: CPU to monitor
+ *
+ * The cpu specified in the pfarg_load.load_pid  argument must be the current
+ * CPU.
+ *
+ * The function must be called with the context locked and interrupts disabled.
+ */
+static int pfm_load_ctx_sys(struct pfm_context *ctx, u16 set_id, u32 cpu)
+{
+       struct pfm_event_set *set;
+       int mycpu;
+       int ret;
+
+       mycpu = smp_processor_id();
+
+       /*
+        * system-wide: check we are running on the desired CPU
+        */
+       if (cpu != mycpu) {
+               PFM_DBG("wrong CPU: asking %u but on %u", cpu, mycpu);
+               return -EINVAL;
+       }
+
+       /*
+        * initialize sets
+        */
+       set = pfm_prepare_sets(ctx, set_id);
+       if (!set) {
+               PFM_DBG("event set%u does not exist", set_id);
+               return -EINVAL;
+       }
+
+       PFM_DBG("set=%u set_flags=0x%x", set->id, set->flags);
+
+       ctx->cpu = mycpu;
+       ctx->task = NULL;
+       ctx->active_set = set;
+
+       /*
+        * perform any architecture specific actions
+        */
+       ret = pfm_arch_load_context(ctx);
+       if (ret)
+               goto error_noload;
+
+       /*
+        * now reserve the session, before we can proceed with
+        * actually accessing the PMU hardware
+        */
+       ret = pfm_session_acquire(1, mycpu);
+       if (ret)
+               goto error;
+
+
+       /*
+        * caller must be on monitored CPU to access PMU, thus this is
+        * a form of self-monitoring
+        */
+       ctx->flags.is_self = 1;
+
+       set->runs++;
+
+       /*
+        * load PMD from set
+        * load PMC from set
+        */
+       pfm_arch_restore_pmds(ctx, set);
+       pfm_arch_restore_pmcs(ctx, set);
+
+       /*
+        * set new ownership
+        */
+       pfm_set_pmu_owner(NULL, ctx);
+
+       /*
+        * reset pending work
+        */
+       ctx->flags.work_type = PFM_WORK_NONE;
+       ctx->flags.reset_count = 0;
+
+       /*
+        * reset message queue
+        */
+       ctx->msgq_head = ctx->msgq_tail = 0;
+
+       ctx->state = PFM_CTX_LOADED;
+
+       return 0;
+error:
+       pfm_arch_unload_context(ctx);
+error_noload:
+       return ret;
+}
+
+/**
+ * __pfm_load_context_thread - attach context to a thread
+ * @ctx: context to operate on
+ * @set_id: first set
+ * @task: threadf to attach to
+ *
+ * The function must be called with the context locked and interrupts disabled.
+ */
+static int pfm_load_ctx_thread(struct pfm_context *ctx, u16 set_id,
+                              struct task_struct *task)
+{
+       struct pfm_event_set *set;
+       struct pfm_context *old;
+       int ret;
+
+       PFM_DBG("load_pid=%d set=%u", task->pid, set_id);
+       /*
+        * per-thread:
+        *   - task to attach to is checked in sys_pfm_load_context() to avoid
+        *     locking issues. if found, and not self,  task refcount was
+        *     incremented.
+        */
+       old = cmpxchg(&task->pfm_context, NULL, ctx);
+       if (old) {
+               PFM_DBG("load_pid=%d has a context "
+                       "old=%p new=%p cur=%p",
+                       task->pid,
+                       old,
+                       ctx,
+                       task->pfm_context);
+               return -EEXIST;
+       }
+
+       /*
+        * initialize sets
+        */
+       set = pfm_prepare_sets(ctx, set_id);
+       if (!set) {
+               PFM_DBG("event set%u does not exist", set_id);
+               return -EINVAL;
+       }
+
+
+       ctx->task = task;
+       ctx->cpu = -1;
+       ctx->active_set = set;
+
+       /*
+        * perform any architecture specific actions
+        */
+       ret = pfm_arch_load_context(ctx);
+       if (ret)
+               goto error_noload;
+
+       /*
+        * now reserve the session, before we can proceed with
+        * actually accessing the PMU hardware
+        */
+       ret = pfm_session_acquire(0, -1);
+       if (ret)
+               goto error;
+
+
+       set->runs++;
+       if (ctx->task != current) {
+
+               ctx->flags.is_self = 0;
+
+               /* force a full reload */
+               ctx->last_act = PFM_INVALID_ACTIVATION;
+               ctx->last_cpu = -1;
+               set->priv_flags |= PFM_SETFL_PRIV_MOD_BOTH;
+
+       } else {
+               pfm_check_save_prev_ctx();
+
+               ctx->last_cpu = smp_processor_id();
+               __get_cpu_var(pmu_activation_number)++;
+               ctx->last_act = __get_cpu_var(pmu_activation_number);
+
+               ctx->flags.is_self = 1;
+
+               /*
+                * load PMD from set
+                * load PMC from set
+                */
+               pfm_arch_restore_pmds(ctx, set);
+               pfm_arch_restore_pmcs(ctx, set);
+
+               /*
+                * set new ownership
+                */
+               pfm_set_pmu_owner(ctx->task, ctx);
+       }
+       set_tsk_thread_flag(task, TIF_PERFMON_CTXSW);
+
+       /*
+        * reset pending work
+        */
+       ctx->flags.work_type = PFM_WORK_NONE;
+       ctx->flags.reset_count = 0;
+
+       /*
+        * reset message queue
+        */
+       ctx->msgq_head = ctx->msgq_tail = 0;
+
+       ctx->state = PFM_CTX_LOADED;
+
+       return 0;
+
+error:
+       pfm_arch_unload_context(ctx);
+       ctx->task = NULL;
+error_noload:
+       /*
+        * detach context
+        */
+       task->pfm_context = NULL;
+       return ret;
+}
+
+/**
+ * __pfm_load_context - attach context to a CPU or thread
+ * @ctx: context to operate on
+ * @load: pfarg_load as passed by user
+ * @task: thread to attach to, NULL for system-wide
+ */
+int __pfm_load_context(struct pfm_context *ctx, struct pfarg_load *load,
+                      struct task_struct *task)
+{
+       if (ctx->flags.system)
+               return pfm_load_ctx_sys(ctx, load->load_set, load->load_pid);
+       return pfm_load_ctx_thread(ctx, load->load_set, task);
+}
+
+/**
+ * pfm_update_ovfl_pmds - account for pending ovfls on PMDs
+ * @ctx: context to operate on
+ *
+ * This function is always called after pfm_stop has been issued
+ */
+static void pfm_update_ovfl_pmds(struct pfm_context *ctx)
+{
+       struct pfm_event_set *set;
+       u64 *cnt_pmds;
+       u64 ovfl_mask;
+       u16 num_ovfls, i, first;
+
+       ovfl_mask = pfm_pmu_conf->ovfl_mask;
+       first = ctx->regs.first_intr_pmd;
+       cnt_pmds = ctx->regs.cnt_pmds;
+
+       /*
+        * look for pending interrupts and adjust PMD values accordingly
+        */
+       list_for_each_entry(set, &ctx->set_list, list) {
+
+               if (!set->npend_ovfls)
+                       continue;
+
+               num_ovfls = set->npend_ovfls;
+               PFM_DBG("set%u nintrs=%u", set->id, num_ovfls);
+
+               for (i = first; num_ovfls; i++) {
+                       if (test_bit(i, cast_ulp(set->povfl_pmds))) {
+                               /* only correct value for counters */
+                               if (test_bit(i, cast_ulp(cnt_pmds)))
+                                       set->pmds[i].value += 1 + ovfl_mask;
+                               num_ovfls--;
+                       }
+                       PFM_DBG("pmd%u set=%u val=0x%llx",
+                               i,
+                               set->id,
+                               (unsigned long long)set->pmds[i].value);
+               }
+               /*
+                * we need to clear to prevent a pfm_getinfo_evtsets() from
+                * returning stale data even after the context is unloaded
+                */
+               set->npend_ovfls = 0;
+               bitmap_zero(cast_ulp(set->povfl_pmds), ctx->regs.max_intr_pmd);
+       }
+}
+
+
+/**
+ * __pfm_unload_context - detach context from CPU or thread
+ * @ctx: context to operate on
+ * @release_info: pointer to return info (see below)
+ *
+ * The function must be called with the context locked and interrupts disabled.
+ *
+ * release_info value upon return:
+ *     - bit 0: when set, must free context
+ *     - bit 1: when set, must cancel hrtimer
+ */
+int __pfm_unload_context(struct pfm_context *ctx, int *release_info)
+{
+       struct task_struct *task;
+       int ret;
+
+       PFM_DBG("ctx_state=%d task [%d]",
+               ctx->state,
+               ctx->task ? ctx->task->pid : -1);
+
+       *release_info = 0;
+
+       /*
+        * unload only when necessary
+        */
+       if (ctx->state == PFM_CTX_UNLOADED)
+               return 0;
+
+       task = ctx->task;
+
+       /*
+        * stop monitoring
+        */
+       ret = __pfm_stop(ctx, release_info);
+       if (ret)
+               return ret;
+
+       ctx->state = PFM_CTX_UNLOADED;
+       ctx->flags.can_restart = 0;
+
+       /*
+        * save active set
+        * UP:
+        *      if not current task and due to lazy, state may
+        *      still be live
+        * for system-wide, guaranteed to run on correct CPU
+        */
+       if (__get_cpu_var(pmu_ctx) == ctx) {
+               /*
+                * pending overflows have been saved by pfm_stop()
+                */
+               pfm_save_pmds(ctx, ctx->active_set);
+               pfm_set_pmu_owner(NULL, NULL);
+               PFM_DBG("released ownership");
+       }
+
+       /*
+        * account for pending overflows
+        */
+       pfm_update_ovfl_pmds(ctx);
+
+       /*
+        * arch-specific unload operations
+        */
+       pfm_arch_unload_context(ctx);
+
+       /*
+        * per-thread: disconnect from monitored task
+        */
+       if (task) {
+               task->pfm_context = NULL;
+               ctx->task = NULL;
+               clear_tsk_thread_flag(task, TIF_PERFMON_CTXSW);
+               clear_tsk_thread_flag(task, TIF_PERFMON_WORK);
+               pfm_arch_disarm_handle_work(task);
+       }
+       /*
+        * session can be freed, must have interrupts enabled
+        * thus we release in the caller. Bit 0 signals to the
+        * caller that the session can be released.
+        */
+       *release_info |= 0x1;
+
+       return 0;
+}
+
+/**
+ * __pfm_exit_thread - detach and free context on thread exit
+ */
+void __pfm_exit_thread(void)
+{
+       struct pfm_context *ctx;
+       unsigned long flags;
+       int free_ok = 0, release_info = 0;
+       int ret;
+
+       ctx  = current->pfm_context;
+
+       BUG_ON(ctx->flags.system);
+
+       spin_lock_irqsave(&ctx->lock, flags);
+
+       PFM_DBG("state=%d is_self=%d", ctx->state, ctx->flags.is_self);
+
+       /*
+        * __pfm_unload_context() cannot fail
+        * in the context states we are interested in
+        */
+       switch (ctx->state) {
+       case PFM_CTX_LOADED:
+       case PFM_CTX_MASKED:
+               __pfm_unload_context(ctx, &release_info);
+               /*
+                * end notification only sent for non
+                * self-monitoring context
+                */
+               if (!ctx->flags.is_self)
+                       pfm_end_notify(ctx);
+               break;
+       case PFM_CTX_ZOMBIE:
+               __pfm_unload_context(ctx, &release_info);
+               free_ok = 1;
+               break;
+       default:
+               BUG_ON(ctx->state != PFM_CTX_LOADED);
+               break;
+       }
+       spin_unlock_irqrestore(&ctx->lock, flags);
+
+       /*
+        * cancel timer now that context is unlocked
+        */
+       if (release_info & 0x2) {
+               ret = hrtimer_cancel(&__get_cpu_var(pfm_hrtimer));
+               PFM_DBG("timeout cancel=%d", ret);
+       }
+
+       if (release_info & 0x1)
+               pfm_session_release(0, 0);
+
+       /*
+        * All memory free operations (especially for vmalloc'ed memory)
+        * MUST be done with interrupts ENABLED.
+        */
+       if (free_ok)
+               pfm_free_context(ctx);
+}
diff --git a/perfmon/perfmon_ctx.c b/perfmon/perfmon_ctx.c
new file mode 100644
index 0000000..afe6078
--- /dev/null
+++ b/perfmon/perfmon_ctx.c
@@ -0,0 +1,314 @@
+/*
+ * perfmon_ctx.c: perfmon2 context functions
+ *
+ * This file implements the perfmon2 interface which
+ * provides access to the hardware performance counters
+ * of the host processor.
+ *
+ *
+ * The initial version of perfmon.c was written by
+ * Ganesh Venkitachalam, IBM Corp.
+ *
+ * Then it was modified for perfmon-1.x by Stephane Eranian and
+ * David Mosberger, Hewlett Packard Co.
+ *
+ * Version Perfmon-2.x is a complete rewrite of perfmon-1.x
+ * by Stephane Eranian, Hewlett Packard Co.
+ *
+ * Copyright (c) 1999-2006 Hewlett-Packard Development Company, L.P.
+ * Contributed by Stephane Eranian <eranian@hpl.hp.com>
+ *                David Mosberger-Tang <davidm@hpl.hp.com>
+ *
+ * More information about perfmon available at:
+ *     http://perfmon2.sf.net
+ *
+ * This program is free software; you can redistribute it and/or
+ * modify it under the terms of version 2 of the GNU General Public
+ * License as published by the Free Software Foundation.
+ *
+ * This program is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+ * General Public License for more details.
+ *
+ * You should have received a copy of the GNU General Public License
+ * along with this program; if not, write to the Free Software
+ * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA
+ * 02111-1307 USA
+ */
+#include <linux/kernel.h>
+#include <linux/fs.h>
+#include <linux/perfmon_kern.h>
+#include "perfmon_priv.h"
+
+/*
+ * context memory pool pointer
+ */
+static struct kmem_cache *pfm_ctx_cachep;
+
+/**
+ * pfm_free_context - de-allocate context and associated resources
+ * @ctx: context to free
+ */
+void pfm_free_context(struct pfm_context *ctx)
+{
+       pfm_arch_context_free(ctx);
+
+       pfm_free_sets(ctx);
+
+       pfm_smpl_buf_free(ctx);
+
+       PFM_DBG("free ctx @0x%p", ctx);
+       kmem_cache_free(pfm_ctx_cachep, ctx);
+       /*
+        * decrease refcount on:
+        *      - PMU description table
+        *      - sampling format
+        */
+       pfm_pmu_conf_put();
+       pfm_pmu_release();
+}
+
+/**
+ * pfm_ctx_flags_sane - check if context flags passed by user are okay
+ * @ctx_flags: flags passed user on pfm_create_context
+ *
+ * return:
+ *      0 if successful
+ *     <0 and error code otherwise
+ */
+static inline int pfm_ctx_flags_sane(u32 ctx_flags)
+{
+       if (ctx_flags & PFM_FL_SYSTEM_WIDE) {
+               if (ctx_flags & PFM_FL_NOTIFY_BLOCK) {
+                       PFM_DBG("cannot use blocking mode in syswide mode");
+                       return -EINVAL;
+               }
+       }
+       return 0;
+}
+
+/**
+ * pfm_ctx_permissions - check authorization to create new context
+ * @ctx_flags: context flags passed by user
+ *
+ * check for permissions to create a context.
+ *
+ * A sysadmin may decide to restrict creation of per-thread
+ * and/or system-wide context to a group of users using the
+ * group id via /sys/kernel/perfmon/task_group  and
+ * /sys/kernel/perfmon/sys_group.
+ *
+ * Once we identify a user level package which can be used
+ * to grant/revoke Linux capabilites at login via PAM, we will
+ * be able to use capabilities. We would also need to increase
+ * the size of cap_t to support more than 32 capabilities (it
+ * is currently defined as u32 and 32 capabilities are alrady
+ * defined).
+ */
+static inline int pfm_ctx_permissions(u32 ctx_flags)
+{
+       if ((ctx_flags & PFM_FL_SYSTEM_WIDE)
+           && pfm_controls.sys_group != PFM_GROUP_PERM_ANY
+           && !in_group_p(pfm_controls.sys_group)) {
+               PFM_DBG("user group not allowed to create a syswide ctx");
+               return -EPERM;
+       } else if (pfm_controls.task_group != PFM_GROUP_PERM_ANY
+                  && !in_group_p(pfm_controls.task_group)) {
+               PFM_DBG("user group not allowed to create a task context");
+               return -EPERM;
+       }
+       return 0;
+}
+
+/**
+ * __pfm_create_context - allocate and initialize a perfmon context
+ * @req : pfarg_ctx from user
+ * @fmt : pointer sampling format, NULL if not used
+ * @fmt_arg: pointer to argument to sampling format, NULL if not used
+ * @mode: PFM_NORMAL or PFM_COMPAT(IA-64 v2.0 compatibility)
+ * @ctx : address of new context upon succesful return, undefined otherwise
+ *
+ * function used to allocate a new context. A context is allocated along
+ * with the default event set. If a sampling format is used, the buffer
+ * may be allocated and initialized.
+ *
+ * The file descriptor identifying the context is allocated and returned
+ * to caller.
+ *
+ * This function operates with no locks and interrupts are enabled.
+ * return:
+ *     >=0: the file descriptor to identify the context
+ *     <0 : the error code
+ */
+int __pfm_create_context(struct pfarg_ctx *req,
+                        struct pfm_smpl_fmt *fmt,
+                        void *fmt_arg,
+                        int mode,
+                        struct pfm_context **new_ctx)
+{
+       struct pfm_context *ctx;
+       struct file *filp = NULL;
+       u32 ctx_flags;
+       int fd = 0, ret;
+
+       ctx_flags = req->ctx_flags;
+
+       /* Increase refcount on PMU description */
+       ret = pfm_pmu_conf_get(1);
+       if (ret < 0)
+               goto error_conf;
+
+       ret = pfm_ctx_flags_sane(ctx_flags);
+       if (ret < 0)
+               goto error_alloc;
+
+       ret = pfm_ctx_permissions(ctx_flags);
+       if (ret < 0)
+               goto error_alloc;
+
+       /*
+        * we can use GFP_KERNEL and potentially sleep because we do
+        * not hold any lock at this point.
+        */
+       might_sleep();
+       ret = -ENOMEM;
+       ctx = kmem_cache_zalloc(pfm_ctx_cachep, GFP_KERNEL);
+       if (!ctx)
+               goto error_alloc;
+
+       PFM_DBG("alloc ctx @0x%p", ctx);
+
+       INIT_LIST_HEAD(&ctx->set_list);
+       spin_lock_init(&ctx->lock);
+       init_completion(&ctx->restart_complete);
+       init_waitqueue_head(&ctx->msgq_wait);
+
+       /*
+        * context is unloaded
+        */
+       ctx->state = PFM_CTX_UNLOADED;
+
+       /*
+        * initialization of context's flags
+        * must be done before pfm_find_set()
+        */
+       ctx->flags.block = (ctx_flags & PFM_FL_NOTIFY_BLOCK) ? 1 : 0;
+       ctx->flags.system = (ctx_flags & PFM_FL_SYSTEM_WIDE) ? 1: 0;
+       ctx->flags.no_msg = (ctx_flags & PFM_FL_OVFL_NO_MSG) ? 1: 0;
+       ctx->flags.ia64_v20_compat = mode == PFM_COMPAT ? 1 : 0;
+
+       ret = pfm_pmu_acquire(ctx);
+       if (ret)
+               goto error_file;
+       /*
+        * check if PMU is usable
+        */
+       if (!(ctx->regs.num_pmcs && ctx->regs.num_pmcs)) {
+               PFM_DBG("no usable PMU registers");
+               ret = -EBUSY;
+               goto error_file;
+       }
+
+       /*
+        * link to format, must be done first for correct
+        * error handling in pfm_context_free()
+        */
+       ctx->smpl_fmt = fmt;
+
+       ret = -ENFILE;
+       fd = pfm_alloc_fd(&filp);
+       if (fd < 0)
+               goto error_file;
+
+       /*
+        * initialize arch-specific section
+        * must be done before fmt_init()
+        */
+       ret = pfm_arch_context_create(ctx, ctx_flags);
+       if (ret)
+               goto error_set;
+
+       ret = -ENOMEM;
+
+       /*
+        * add initial set
+        */
+       if (pfm_create_initial_set(ctx))
+               goto error_set;
+
+       /*
+        * does the user want to sample?
+        * must be done after pfm_pmu_acquire() because
+        * needs ctx->regs
+        */
+       if (fmt) {
+               ret = pfm_setup_smpl_fmt(ctx, ctx_flags, fmt_arg, filp);
+               if (ret)
+                       goto error_set;
+       }
+
+       filp->private_data = ctx;
+
+       ctx->last_act = PFM_INVALID_ACTIVATION;
+       ctx->last_cpu = -1;
+
+       /*
+        * initialize notification message queue
+        */
+       ctx->msgq_head = ctx->msgq_tail = 0;
+
+       PFM_DBG("flags=0x%x system=%d notify_block=%d no_msg=%d"
+               " use_fmt=%d ctx_fd=%d mode=%d",
+               ctx_flags,
+               ctx->flags.system,
+               ctx->flags.block,
+               ctx->flags.no_msg,
+               !!fmt,
+               fd, mode);
+
+       if (new_ctx)
+               *new_ctx = ctx;
+
+       /*
+        * we defer the fd_install until we are certain the call succeeded
+        * to ensure we do not have to undo its effect. Neither put_filp()
+        * nor put_unused_fd() undoes the effect of fd_install().
+        */
+       fd_install(fd, filp);
+
+       return fd;
+
+error_set:
+       put_filp(filp);
+       put_unused_fd(fd);
+error_file:
+       /*
+        * calls the right *_put() functions
+        * calls pfm_release_pmu()
+        */
+       pfm_free_context(ctx);
+       return ret;
+error_alloc:
+       pfm_pmu_conf_put();
+error_conf:
+       pfm_smpl_fmt_put(fmt);
+       return ret;
+}
+
+/**
+ * pfm_init_ctx -- initialize context SLAB
+ *
+ * called from pfm_init
+ */
+int __init pfm_init_ctx(void)
+{
+       pfm_ctx_cachep = kmem_cache_create("pfm_context",
+                                  sizeof(struct pfm_context)+PFM_ARCH_CTX_SIZE,
+                                  SLAB_HWCACHE_ALIGN, 0, NULL);
+       if (!pfm_ctx_cachep) {
+               PFM_ERR("cannot initialize context slab");
+               return -ENOMEM;
+       }
+       return 0;
+}
diff --git a/perfmon/perfmon_ctxsw.c b/perfmon/perfmon_ctxsw.c
new file mode 100644
index 0000000..9a28d13
--- /dev/null
+++ b/perfmon/perfmon_ctxsw.c
@@ -0,0 +1,342 @@
+/*
+ * perfmon_cxtsw.c: perfmon2 context switch code
+ *
+ * This file implements the perfmon2 interface which
+ * provides access to the hardware performance counters
+ * of the host processor.
+ *
+ * The initial version of perfmon.c was written by
+ * Ganesh Venkitachalam, IBM Corp.
+ *
+ * Then it was modified for perfmon-1.x by Stephane Eranian and
+ * David Mosberger, Hewlett Packard Co.
+ *
+ * Version Perfmon-2.x is a complete rewrite of perfmon-1.x
+ * by Stephane Eranian, Hewlett Packard Co.
+ *
+ * Copyright (c) 1999-2006 Hewlett-Packard Development Company, L.P.
+ * Contributed by Stephane Eranian <eranian@hpl.hp.com>
+ *                David Mosberger-Tang <davidm@hpl.hp.com>
+ *
+ * More information about perfmon available at:
+ *     http://perfmon2.sf.net
+ *
+ * This program is free software; you can redistribute it and/or
+ * modify it under the terms of version 2 of the GNU General Public
+ * License as published by the Free Software Foundation.
+ *
+ * This program is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+ * General Public License for more details.
+ *
+ * You should have received a copy of the GNU General Public License
+ * along with this program; if not, write to the Free Software
+ * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA
+ * 02111-1307 USA
+ */
+#include <linux/kernel.h>
+#include <linux/perfmon_kern.h>
+#include "perfmon_priv.h"
+
+void pfm_save_pmds(struct pfm_context *ctx, struct pfm_event_set *set)
+{
+       u64 val, ovfl_mask;
+       u64 *used_pmds, *cnt_pmds;
+       u16 i, num;
+
+       ovfl_mask = pfm_pmu_conf->ovfl_mask;
+       num = set->nused_pmds;
+       cnt_pmds = ctx->regs.cnt_pmds;
+       used_pmds = set->used_pmds;
+
+       /*
+        * save HW PMD, for counters, reconstruct 64-bit value
+        */
+       for (i = 0; num; i++) {
+               if (test_bit(i, cast_ulp(used_pmds))) {
+                       val = pfm_read_pmd(ctx, i);
+                       if (likely(test_bit(i, cast_ulp(cnt_pmds))))
+                               val = (set->pmds[i].value & ~ovfl_mask) |
+                                       (val & ovfl_mask);
+                       set->pmds[i].value = val;
+                       num--;
+               }
+       }
+       pfm_arch_clear_pmd_ovfl_cond(ctx, set);
+}
+
+/*
+ * interrupts are  disabled (no preemption)
+ */
+void __pfm_ctxswin_thread(struct task_struct *task,
+                                struct pfm_context *ctx, u64 now)
+{
+       u64 cur_act;
+       struct pfm_event_set *set;
+       int reload_pmcs, reload_pmds;
+       int mycpu, is_active;
+
+       mycpu = smp_processor_id();
+
+       cur_act = __get_cpu_var(pmu_activation_number);
+       /*
+        * we need to lock context because it could be accessed
+        * from another CPU. Normally the schedule() functions
+        * has masked interrupts which should be enough to
+        * protect against PMU interrupts.
+        */
+       spin_lock(&ctx->lock);
+
+       is_active = pfm_arch_is_active(ctx);
+
+       set = ctx->active_set;
+
+       /*
+        * in case fo zombie, we do not complete ctswin of the
+        * PMU, and we force a call to pfm_handle_work() to finish
+        * cleanup, i.e., free context + smpl_buff. The reason for
+        * deferring to pfm_handle_work() is that it is not possible
+        * to vfree() with interrupts disabled.
+        */
+       if (unlikely(ctx->state == PFM_CTX_ZOMBIE)) {
+               pfm_post_work(task, ctx, PFM_WORK_ZOMBIE);
+               goto done;
+       }
+
+       /*
+        * if we were the last user of the PMU on that CPU,
+        * then nothing to do except restore psr
+        */
+       if (ctx->last_cpu == mycpu && ctx->last_act == cur_act) {
+               /*
+                * check for forced reload conditions
+                */
+               reload_pmcs = set->priv_flags & PFM_SETFL_PRIV_MOD_PMCS;
+               reload_pmds = set->priv_flags & PFM_SETFL_PRIV_MOD_PMDS;
+       } else {
+#ifndef CONFIG_SMP
+               pfm_check_save_prev_ctx();
+#endif
+               reload_pmcs = 1;
+               reload_pmds = 1;
+       }
+       /* consumed */
+       set->priv_flags &= ~PFM_SETFL_PRIV_MOD_BOTH;
+
+       if (reload_pmds)
+               pfm_arch_restore_pmds(ctx, set);
+
+       /*
+        * need to check if had in-flight interrupt in
+        * pfm_ctxswout_thread(). If at least one bit set, then we must replay
+        * the interrupt to avoid losing some important performance data.
+        *
+        * npend_ovfls is cleared in interrupt handler
+        */
+       if (set->npend_ovfls) {
+               pfm_arch_resend_irq(ctx);
+               pfm_stats_inc(ovfl_intr_replay_count);
+       }
+
+       if (reload_pmcs)
+               pfm_arch_restore_pmcs(ctx, set);
+
+       /*
+        * record current activation for this context
+        */
+       __get_cpu_var(pmu_activation_number)++;
+       ctx->last_cpu = mycpu;
+       ctx->last_act = __get_cpu_var(pmu_activation_number);
+
+       /*
+        * establish new ownership.
+        */
+       pfm_set_pmu_owner(task, ctx);
+
+       pfm_arch_ctxswin_thread(task, ctx);
+       /*
+        * set->duration does not count when context in MASKED state.
+        * set->duration_start is reset in unmask_monitoring()
+        */
+       set->duration_start = now;
+
+       /*
+        * re-arm switch timeout, if necessary
+        * Timeout is active only if monitoring is active,
+        * i.e., LOADED + started
+        *
+        * We reload the remainder timeout or the full timeout.
+        * Remainder is recorded on context switch out or in
+        * pfm_load_context()
+        */
+       if (ctx->state == PFM_CTX_LOADED
+           && (set->flags & PFM_SETFL_TIME_SWITCH) && is_active) {
+               pfm_restart_timer(ctx, set);
+               /* careful here as pfm_restart_timer may switch sets */
+       }
+done:
+       spin_unlock(&ctx->lock);
+}
+
+/*
+ * interrupts are masked, runqueue lock is held.
+ *
+ * In UP. we simply stop monitoring and leave the state
+ * in place, i.e., lazy save
+ */
+void __pfm_ctxswout_thread(struct task_struct *task,
+                                 struct pfm_context *ctx, u64 now)
+{
+       struct pfm_event_set *set;
+       int need_save_pmds, is_active;
+
+       /*
+        * we need to lock context because it could be accessed
+        * from another CPU. Normally the schedule() functions
+        * has masked interrupts which should be enough to
+        * protect against PMU interrupts.
+        */
+
+       spin_lock(&ctx->lock);
+
+       is_active = pfm_arch_is_active(ctx);
+       set = ctx->active_set;
+
+       /*
+        * stop monitoring and
+        * collect pending overflow information
+        * needed on ctxswin. We cannot afford to lose
+        * a PMU interrupt.
+        */
+       need_save_pmds = pfm_arch_ctxswout_thread(task, ctx);
+
+       if (ctx->state == PFM_CTX_LOADED) {
+               /*
+                * accumulate only when set is actively monitoring,
+                */
+               set->duration += now - set->duration_start;
+
+               /*
+                * record remaining timeout
+                * reload in pfm_ctxsw_in()
+                */
+               if (is_active && (set->flags & PFM_SETFL_TIME_SWITCH)) {
+                       struct hrtimer *h = NULL;
+                       h = &__get_cpu_var(pfm_hrtimer);
+                       hrtimer_cancel(h);
+                       set->hrtimer_rem = hrtimer_get_remaining(h);
+                       PFM_DBG_ovfl("hrtimer=%lld",
+                                    (long long)set->hrtimer_rem.tv64);
+               }
+       }
+
+#ifdef CONFIG_SMP
+       /*
+        * in SMP, release ownership of this PMU.
+        * PMU interrupts are masked, so nothing
+        * can happen.
+        */
+       pfm_set_pmu_owner(NULL, NULL);
+
+       /*
+        * On some architectures, it is necessary to read the
+        * PMD registers to check for pending overflow in
+        * pfm_arch_ctxswout_thread(). In that case, saving of
+        * the PMDs  may be  done there and not here.
+        */
+       if (need_save_pmds)
+               pfm_save_pmds(ctx, set);
+#endif
+       spin_unlock(&ctx->lock);
+}
+
+/*
+ *
+ */
+static void __pfm_ctxswout_sys(struct task_struct *prev,
+                              struct task_struct *next)
+{
+       struct pfm_context *ctx;
+
+       ctx = __get_cpu_var(pmu_ctx);
+       BUG_ON(!ctx);
+
+       /*
+        * propagate TIF_PERFMON_CTXSW to ensure that:
+        * - previous task has TIF_PERFMON_CTXSW cleared, in case it is
+        *   scheduled onto another CPU where there is syswide monitoring
+        * - next task has TIF_PERFMON_CTXSW set to ensure it will come back
+        *   here when context switched out
+        */
+       clear_tsk_thread_flag(prev, TIF_PERFMON_CTXSW);
+       set_tsk_thread_flag(next, TIF_PERFMON_CTXSW);
+
+       /*
+        * nothing to do until actually started
+        * XXX: assumes no mean to start from user level
+        */
+       if (!ctx->flags.started)
+               return;
+
+       pfm_arch_ctxswout_sys(prev, ctx);
+}
+
+/*
+ *
+ */
+static void __pfm_ctxswin_sys(struct task_struct *prev,
+                             struct task_struct *next)
+{
+       struct pfm_context *ctx;
+
+       ctx = __get_cpu_var(pmu_ctx);
+       BUG_ON(!ctx);
+
+       /*
+        * nothing to do until actually started
+        * XXX: assumes no mean to start from user level
+        */
+       if (!ctx->flags.started)
+               return;
+
+       pfm_arch_ctxswin_sys(next, ctx);
+}
+
+void pfm_ctxsw_out(struct task_struct *prev,
+                  struct task_struct *next)
+{
+       struct pfm_context *ctxp;
+       u64 now;
+
+       now = sched_clock();
+
+       ctxp = prev->pfm_context;
+
+       if (ctxp)
+               __pfm_ctxswout_thread(prev, ctxp, now);
+       else
+               __pfm_ctxswout_sys(prev, next);
+
+       pfm_stats_inc(ctxswout_count);
+       pfm_stats_add(ctxswout_ns, sched_clock() - now);
+}
+
+void pfm_ctxsw_in(struct task_struct *prev,
+                 struct task_struct *next)
+{
+       struct pfm_context *ctxn;
+       u64 now;
+
+       now = sched_clock();
+
+       ctxn = next->pfm_context;
+
+       if (ctxn)
+               __pfm_ctxswin_thread(next, ctxn, now);
+       else
+               __pfm_ctxswin_sys(prev, next);
+
+       pfm_stats_inc(ctxswin_count);
+       pfm_stats_add(ctxswin_ns, sched_clock() - now);
+}
diff --git a/perfmon/perfmon_debugfs.c b/perfmon/perfmon_debugfs.c
new file mode 100644
index 0000000..e4d2fad
--- /dev/null
+++ b/perfmon/perfmon_debugfs.c
@@ -0,0 +1,168 @@
+/*
+ * perfmon_debugfs.c: perfmon2 statistics interface to debugfs
+ *
+ * This file implements the perfmon2 interface which
+ * provides access to the hardware performance counters
+ * of the host processor.
+ *
+ * The initial version of perfmon.c was written by
+ * Ganesh Venkitachalam, IBM Corp.
+ *
+ * Then it was modified for perfmon-1.x by Stephane Eranian and
+ * David Mosberger, Hewlett Packard Co.
+ *
+ * Version Perfmon-2.x is a complete rewrite of perfmon-1.x
+ * by Stephane Eranian, Hewlett Packard Co.
+ *
+ * Copyright (c) 2007 Hewlett-Packard Development Company, L.P.
+ * Contributed by Stephane Eranian <eranian@hpl.hp.com>
+ *
+ * More information about perfmon available at:
+ *     http://perfmon2.sf.net
+ *
+ * This program is free software; you can redistribute it and/or
+ * modify it under the terms of version 2 of the GNU General Public
+ * License as published by the Free Software Foundation.
+ *
+ * This program is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+ * General Public License for more details.
+ *
+ * You should have received a copy of the GNU General Public License
+ * along with this program; if not, write to the Free Software
+ * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA
+ * 02111-1307 USA
+ */
+#include <linux/kernel.h>
+#include <linux/debugfs.h>
+#include <linux/perfmon_kern.h>
+
+/*
+ * to make the statistics visible to user space:
+ * $ mount -t debugfs none /mnt
+ * $ cd /mnt/perfmon
+ * then choose a CPU subdir
+ */
+DECLARE_PER_CPU(struct pfm_stats, pfm_stats);
+
+static struct dentry *pfm_debugfs_dir;
+
+void pfm_reset_stats(int cpu)
+{
+       struct pfm_stats *st;
+       unsigned long flags;
+
+       st = &per_cpu(pfm_stats, cpu);
+
+       local_irq_save(flags);
+       memset(st->v, 0, sizeof(st->v));
+       local_irq_restore(flags);
+}
+
+static const char *pfm_stats_strs[] = {
+       "ovfl_intr_all_count",
+       "ovfl_intr_ns",
+       "ovfl_intr_spurious_count",
+       "ovfl_intr_replay_count",
+       "ovfl_intr_regular_count",
+       "handle_work_count",
+       "ovfl_notify_count",
+       "reset_pmds_count",
+       "pfm_restart_count",
+       "fmt_handler_calls",
+       "fmt_handler_ns",
+       "set_switch_count",
+       "set_switch_ns",
+       "set_switch_exp",
+       "ctxswin_count",
+       "ctxswin_ns",
+       "handle_timeout_count",
+       "ovfl_intr_nmi_count",
+       "ctxswout_count",
+       "ctxswout_ns",
+};
+#define PFM_NUM_STRS ARRAY_SIZE(pfm_stats_strs)
+
+void pfm_debugfs_del_cpu(int cpu)
+{
+       struct pfm_stats *st;
+       int i;
+
+       st = &per_cpu(pfm_stats, cpu);
+
+       for (i = 0; i < PFM_NUM_STATS; i++) {
+               if (st->dirs[i])
+                       debugfs_remove(st->dirs[i]);
+               st->dirs[i] = NULL;
+       }
+       if (st->cpu_dir)
+               debugfs_remove(st->cpu_dir);
+       st->cpu_dir = NULL;
+}
+
+int pfm_debugfs_add_cpu(int cpu)
+{
+       struct pfm_stats *st;
+       int i;
+
+       /*
+        * sanity check between stats names and the number
+        * of entries in the pfm_stats value array.
+        */
+       if (PFM_NUM_STRS != PFM_NUM_STATS) {
+               PFM_ERR("PFM_NUM_STRS != PFM_NUM_STATS error");
+               return -1;
+       }
+
+       st = &per_cpu(pfm_stats, cpu);
+       sprintf(st->cpu_name, "cpu%d", cpu);
+
+       st->cpu_dir = debugfs_create_dir(st->cpu_name, pfm_debugfs_dir);
+       if (!st->cpu_dir)
+               return -1;
+
+       for (i = 0; i < PFM_NUM_STATS; i++) {
+               st->dirs[i] = debugfs_create_u64(pfm_stats_strs[i],
+                                                S_IRUGO,
+                                                st->cpu_dir,
+                                                &st->v[i]);
+               if (!st->dirs[i])
+                       goto error;
+       }
+       pfm_reset_stats(cpu);
+       return 0;
+error:
+       while (i >= 0) {
+               debugfs_remove(st->dirs[i]);
+               i--;
+       }
+       debugfs_remove(st->cpu_dir);
+       return -1;
+}
+
+/*
+ * called once from pfm_init()
+ */
+int __init pfm_init_debugfs(void)
+{
+       int cpu1, cpu2, ret;
+
+       pfm_debugfs_dir = debugfs_create_dir("perfmon", NULL);
+       if (!pfm_debugfs_dir)
+               return -1;
+
+       for_each_online_cpu(cpu1) {
+               ret = pfm_debugfs_add_cpu(cpu1);
+               if (ret)
+                       goto error;
+       }
+       return 0;
+error:
+       for_each_online_cpu(cpu2) {
+               if (cpu2 == cpu1)
+                       break;
+               pfm_debugfs_del_cpu(cpu2);
+       }
+       return -1;
+}
diff --git a/perfmon/perfmon_dfl_smpl.c b/perfmon/perfmon_dfl_smpl.c
new file mode 100644
index 0000000..8c83489
--- /dev/null
+++ b/perfmon/perfmon_dfl_smpl.c
@@ -0,0 +1,298 @@
+/*
+ * Copyright (c) 1999-2006 Hewlett-Packard Development Company, L.P.
+ *               Contributed by Stephane Eranian <eranian@hpl.hp.com>
+ *
+ * This file implements the new default sampling buffer format
+ * for the perfmon2 subsystem.
+ *
+ * This program is free software; you can redistribute it and/or
+ * modify it under the terms of version 2 of the GNU General Public
+ * License as published by the Free Software Foundation.
+ *
+ * This program is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+ * General Public License for more details.
+ *
+ * You should have received a copy of the GNU General Public License
+ * along with this program; if not, write to the Free Software
+ * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA
+ * 02111-1307 USA
+ */
+#include <linux/kernel.h>
+#include <linux/types.h>
+#include <linux/module.h>
+#include <linux/init.h>
+#include <linux/smp.h>
+
+#include <linux/perfmon_kern.h>
+#include <linux/perfmon_dfl_smpl.h>
+
+MODULE_AUTHOR("Stephane Eranian <eranian@hpl.hp.com>");
+MODULE_DESCRIPTION("new perfmon default sampling format");
+MODULE_LICENSE("GPL");
+
+static int pfm_dfl_fmt_validate(u32 ctx_flags, u16 npmds, void *data)
+{
+       struct pfm_dfl_smpl_arg *arg = data;
+       u64 min_buf_size;
+
+       if (data == NULL) {
+               PFM_DBG("no argument passed");
+               return -EINVAL;
+       }
+
+       /*
+        * sanity check in case size_t is smaller then u64
+        */
+#if BITS_PER_LONG == 4
+#define MAX_SIZE_T     (1ULL<<(sizeof(size_t)<<3))
+       if (sizeof(size_t) < sizeof(arg->buf_size)) {
+               if (arg->buf_size >= MAX_SIZE_T)
+                       return -ETOOBIG;
+       }
+#endif
+
+       /*
+        * compute min buf size. npmds is the maximum number
+        * of implemented PMD registers.
+        */
+       min_buf_size = sizeof(struct pfm_dfl_smpl_hdr)
+               + (sizeof(struct pfm_dfl_smpl_entry) + (npmds*sizeof(u64)));
+
+       PFM_DBG("validate ctx_flags=0x%x flags=0x%x npmds=%u "
+               "min_buf_size=%llu buf_size=%llu\n",
+               ctx_flags,
+               arg->buf_flags,
+               npmds,
+               (unsigned long long)min_buf_size,
+               (unsigned long long)arg->buf_size);
+
+       /*
+        * must hold at least the buffer header + one minimally sized entry
+        */
+       if (arg->buf_size < min_buf_size)
+               return -EINVAL;
+
+       return 0;
+}
+
+static int pfm_dfl_fmt_get_size(u32 flags, void *data, size_t *size)
+{
+       struct pfm_dfl_smpl_arg *arg = data;
+
+       /*
+        * size has been validated in default_validate
+        * we can never loose bits from buf_size.
+        */
+       *size = (size_t)arg->buf_size;
+
+       return 0;
+}
+
+static int pfm_dfl_fmt_init(struct pfm_context *ctx, void *buf, u32 ctx_flags,
+                           u16 npmds, void *data)
+{
+       struct pfm_dfl_smpl_hdr *hdr;
+       struct pfm_dfl_smpl_arg *arg = data;
+
+       hdr = buf;
+
+       hdr->hdr_version = PFM_DFL_SMPL_VERSION;
+       hdr->hdr_buf_size = arg->buf_size;
+       hdr->hdr_buf_flags = arg->buf_flags;
+       hdr->hdr_cur_offs = sizeof(*hdr);
+       hdr->hdr_overflows = 0;
+       hdr->hdr_count = 0;
+       hdr->hdr_min_buf_space = sizeof(struct pfm_dfl_smpl_entry) + (npmds*sizeof(u64));
+       /*
+        * due to cache aliasing, it may be necessary to flush the cache
+        * on certain architectures (e.g., MIPS)
+        */
+       pfm_cacheflush(hdr, sizeof(*hdr));
+
+       PFM_DBG("buffer=%p buf_size=%llu hdr_size=%zu hdr_version=%u.%u "
+                 "min_space=%llu npmds=%u",
+                 buf,
+                 (unsigned long long)hdr->hdr_buf_size,
+                 sizeof(*hdr),
+                 PFM_VERSION_MAJOR(hdr->hdr_version),
+                 PFM_VERSION_MINOR(hdr->hdr_version),
+                 (unsigned long long)hdr->hdr_min_buf_space,
+                 npmds);
+
+       return 0;
+}
+
+/*
+ * called from pfm_overflow_handler() to record a new sample
+ *
+ * context is locked, interrupts are disabled (no preemption)
+ */
+static int pfm_dfl_fmt_handler(struct pfm_context *ctx,
+                              unsigned long ip, u64 tstamp, void *data)
+{
+       struct pfm_dfl_smpl_hdr *hdr;
+       struct pfm_dfl_smpl_entry *ent;
+       struct pfm_ovfl_arg *arg;
+       void *cur, *last;
+       u64 *e;
+       size_t entry_size, min_size;
+       u16 npmds, i;
+       u16 ovfl_pmd;
+       void *buf;
+
+       hdr = ctx->smpl_addr;
+       arg = &ctx->ovfl_arg;
+
+        buf = hdr;
+       cur = buf+hdr->hdr_cur_offs;
+       last = buf+hdr->hdr_buf_size;
+       ovfl_pmd = arg->ovfl_pmd;
+       min_size = hdr->hdr_min_buf_space;
+
+       /*
+        * precheck for sanity
+        */
+       if ((last - cur) < min_size)
+               goto full;
+
+       npmds = arg->num_smpl_pmds;
+
+       ent = (struct pfm_dfl_smpl_entry *)cur;
+
+       entry_size = sizeof(*ent) + (npmds << 3);
+
+       /* position for first pmd */
+       e = (u64 *)(ent+1);
+
+       hdr->hdr_count++;
+
+       PFM_DBG_ovfl("count=%llu cur=%p last=%p free_bytes=%zu ovfl_pmd=%d "
+                    "npmds=%u",
+                    (unsigned long long)hdr->hdr_count,
+                    cur, last,
+                    (last-cur),
+                    ovfl_pmd,
+                    npmds);
+
+       /*
+        * current = task running at the time of the overflow.
+        *
+        * per-task mode:
+        *      - this is usually the task being monitored.
+        *        Under certain conditions, it might be a different task
+        *
+        * system-wide:
+        *      - this is not necessarily the task controlling the session
+        */
+       ent->pid = current->pid;
+       ent->ovfl_pmd = ovfl_pmd;
+       ent->last_reset_val = arg->pmd_last_reset;
+
+       /*
+        * where did the fault happen (includes slot number)
+        */
+       ent->ip = ip;
+
+       ent->tstamp = tstamp;
+       ent->cpu = smp_processor_id();
+       ent->set = arg->active_set;
+       ent->tgid = current->tgid;
+
+       /*
+        * selectively store PMDs in increasing index number
+        */
+       if (npmds) {
+               u64 *val = arg->smpl_pmds_values;
+               for (i = 0; i < npmds; i++)
+                       *e++ = *val++;
+       }
+
+       /*
+        * update position for next entry
+        */
+       hdr->hdr_cur_offs += entry_size;
+       cur += entry_size;
+
+       pfm_cacheflush(hdr, sizeof(*hdr));
+       pfm_cacheflush(ent, entry_size);
+
+       /*
+        * post check to avoid losing the last sample
+        */
+       if ((last - cur) < min_size)
+               goto full;
+
+       /* reset before returning from interrupt handler */
+       arg->ovfl_ctrl = PFM_OVFL_CTRL_RESET;
+
+       return 0;
+full:
+       PFM_DBG_ovfl("sampling buffer full free=%zu, count=%llu",
+                    last-cur,
+                    (unsigned long long)hdr->hdr_count);
+
+       /*
+        * increment number of buffer overflows.
+        * important to detect duplicate set of samples.
+        */
+       hdr->hdr_overflows++;
+
+       /*
+        * request notification and masking of monitoring.
+        * Notification is still subject to the overflowed
+        * register having the FL_NOTIFY flag set.
+        */
+       arg->ovfl_ctrl = PFM_OVFL_CTRL_NOTIFY | PFM_OVFL_CTRL_MASK;
+
+       return -ENOBUFS; /* we are full, sorry */
+}
+
+static int pfm_dfl_fmt_restart(int is_active, u32 *ovfl_ctrl, void *buf)
+{
+       struct pfm_dfl_smpl_hdr *hdr;
+
+       hdr = buf;
+
+       hdr->hdr_count = 0;
+       hdr->hdr_cur_offs = sizeof(*hdr);
+
+       pfm_cacheflush(hdr, sizeof(*hdr));
+
+       *ovfl_ctrl = PFM_OVFL_CTRL_RESET;
+
+       return 0;
+}
+
+static int pfm_dfl_fmt_exit(void *buf)
+{
+       return 0;
+}
+
+static struct pfm_smpl_fmt dfl_fmt = {
+       .fmt_name = "default",
+       .fmt_version = 0x10000,
+       .fmt_arg_size = sizeof(struct pfm_dfl_smpl_arg),
+       .fmt_validate = pfm_dfl_fmt_validate,
+       .fmt_getsize = pfm_dfl_fmt_get_size,
+       .fmt_init = pfm_dfl_fmt_init,
+       .fmt_handler = pfm_dfl_fmt_handler,
+       .fmt_restart = pfm_dfl_fmt_restart,
+       .fmt_exit = pfm_dfl_fmt_exit,
+       .fmt_flags = PFM_FMT_BUILTIN_FLAG,
+       .owner = THIS_MODULE
+};
+
+static int pfm_dfl_fmt_init_module(void)
+{
+       return pfm_fmt_register(&dfl_fmt);
+}
+
+static void pfm_dfl_fmt_cleanup_module(void)
+{
+       pfm_fmt_unregister(&dfl_fmt);
+}
+
+module_init(pfm_dfl_fmt_init_module);
+module_exit(pfm_dfl_fmt_cleanup_module);
diff --git a/perfmon/perfmon_file.c b/perfmon/perfmon_file.c
new file mode 100644
index 0000000..1cde81b
--- /dev/null
+++ b/perfmon/perfmon_file.c
@@ -0,0 +1,751 @@
+/*
+ * perfmon_file.c: perfmon2 file input/output functions
+ *
+ * This file implements the perfmon2 interface which
+ * provides access to the hardware performance counters
+ * of the host processor.
+ *
+ * The initial version of perfmon.c was written by
+ * Ganesh Venkitachalam, IBM Corp.
+ *
+ * Then it was modified for perfmon-1.x by Stephane Eranian and
+ * David Mosberger, Hewlett Packard Co.
+ *
+ * Version Perfmon-2.x is a complete rewrite of perfmon-1.x
+ * by Stephane Eranian, Hewlett Packard Co.
+ *
+ * Copyright (c) 1999-2006 Hewlett-Packard Development Company, L.P.
+ * Contributed by Stephane Eranian <eranian@hpl.hp.com>
+ *                David Mosberger-Tang <davidm@hpl.hp.com>
+ *
+ * More information about perfmon available at:
+ *     http://perfmon2.sf.net
+ *
+ * This program is free software; you can redistribute it and/or
+ * modify it under the terms of version 2 of the GNU General Public
+ * License as published by the Free Software Foundation.
+ *
+ * This program is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+ * General Public License for more details.
+ *
+ * You should have received a copy of the GNU General Public License
+ * along with this program; if not, write to the Free Software
+ * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA
+ * 02111-1307 USA
+ */
+#include <linux/kernel.h>
+#include <linux/module.h>
+#include <linux/file.h>
+#include <linux/poll.h>
+#include <linux/vfs.h>
+#include <linux/pagemap.h>
+#include <linux/mount.h>
+#include <linux/perfmon_kern.h>
+#include "perfmon_priv.h"
+
+#define PFMFS_MAGIC 0xa0b4d889 /* perfmon filesystem magic number */
+
+struct pfm_controls pfm_controls = {
+       .sys_group = PFM_GROUP_PERM_ANY,
+       .task_group = PFM_GROUP_PERM_ANY,
+       .arg_mem_max = PAGE_SIZE,
+       .smpl_buffer_mem_max = ~0,
+};
+EXPORT_SYMBOL(pfm_controls);
+
+static int __init enable_debug(char *str)
+{
+       pfm_controls.debug = 1;
+       PFM_INFO("debug output enabled\n");
+       return 1;
+}
+__setup("perfmon_debug", enable_debug);
+
+static int pfmfs_delete_dentry(struct dentry *dentry)
+{
+       return 1;
+}
+
+static struct dentry_operations pfmfs_dentry_operations = {
+       .d_delete = pfmfs_delete_dentry,
+};
+
+int pfm_buf_map_pagefault(struct vm_area_struct *vma, struct vm_fault *vmf)
+{
+       void *kaddr;
+       unsigned long address;
+       struct pfm_context *ctx;
+       size_t size;
+
+       address = (unsigned long)vmf->virtual_address;
+
+       ctx = vma->vm_private_data;
+       if (ctx == NULL) {
+               PFM_DBG("no ctx");
+               return VM_FAULT_SIGBUS;
+       }
+       /*
+        * size available to user (maybe different from real_smpl_size
+        */
+       size = ctx->smpl_size;
+
+       if ((address < vma->vm_start) ||
+           (address >= (vma->vm_start + size)))
+               return VM_FAULT_SIGBUS;
+
+       kaddr = ctx->smpl_addr + (address - vma->vm_start);
+
+       vmf->page = vmalloc_to_page(kaddr);
+       get_page(vmf->page);
+
+       PFM_DBG("[%d] start=%p ref_count=%d",
+               current->pid,
+               kaddr, page_count(vmf->page));
+
+       return 0;
+}
+
+/*
+ * we need to determine whther or not we are closing the last reference
+ * to the file and thus are going to end up in pfm_close() which eventually
+ * calls pfm_release_buf_space(). In that function, we update the accouting
+ * for locked_vm given that we are actually freeing the sampling buffer. The
+ * issue is that there are multiple paths leading to pfm_release_buf_space(),
+ * from exit(), munmap(), close(). The path coming from munmap() is problematic
+ * becuse do_munmap() grabs mmap_sem in write-mode which is also what
+ * pfm_release_buf_space does. To avoid deadlock, we need to determine where
+ * we are calling from and skip the locking. The vm_ops->close() callback
+ * is invoked for each remove_vma() independently of the number of references
+ * left on the file descriptor, therefore simple reference counter does not
+ * work. We need to determine if this is the last call, and then set a flag
+ * to skip the locking.
+ */
+static void pfm_buf_map_close(struct vm_area_struct *vma)
+{
+       struct file *file;
+       struct pfm_context *ctx;
+
+       file = vma->vm_file;
+       ctx = vma->vm_private_data;
+
+       /*
+        * if file is going to close, then pfm_close() will
+        * be called, do not lock in pfm_release_buf
+        */
+       if (atomic_read(&file->f_count) == 1)
+               ctx->flags.mmap_nlock = 1;
+}
+
+/*
+ * we do not have a close callback because, the locked
+ * memory accounting must be done when the actual buffer
+ * is freed. Munmap does not free the page backing the vma
+ * because they may still be in use by the PMU interrupt handler.
+ */
+struct vm_operations_struct pfm_buf_map_vm_ops = {
+       .fault = pfm_buf_map_pagefault,
+       .close = pfm_buf_map_close
+};
+
+static int pfm_mmap_buffer(struct pfm_context *ctx, struct vm_area_struct *vma,
+                          size_t size)
+{
+       if (ctx->smpl_addr == NULL) {
+               PFM_DBG("no sampling buffer to map");
+               return -EINVAL;
+       }
+
+       if (size > ctx->smpl_size) {
+               PFM_DBG("mmap size=%zu >= actual buf size=%zu",
+                       size,
+                       ctx->smpl_size);
+               return -EINVAL;
+       }
+
+       vma->vm_ops = &pfm_buf_map_vm_ops;
+       vma->vm_private_data = ctx;
+
+       return 0;
+}
+
+static int pfm_mmap(struct file *file, struct vm_area_struct *vma)
+{
+       size_t size;
+       struct pfm_context *ctx;
+       unsigned long flags;
+       int ret;
+
+       PFM_DBG("pfm_file_ops");
+
+       ctx  = file->private_data;
+       size = (vma->vm_end - vma->vm_start);
+
+       if (ctx == NULL)
+               return -EINVAL;
+
+       ret = -EINVAL;
+
+       spin_lock_irqsave(&ctx->lock, flags);
+
+       if (vma->vm_flags & VM_WRITE) {
+               PFM_DBG("cannot map buffer for writing");
+               goto done;
+       }
+
+       PFM_DBG("vm_pgoff=%lu size=%zu vm_start=0x%lx",
+               vma->vm_pgoff,
+               size,
+               vma->vm_start);
+
+       ret = pfm_mmap_buffer(ctx, vma, size);
+       if (ret == 0)
+               vma->vm_flags |= VM_RESERVED;
+
+       PFM_DBG("ret=%d vma_flags=0x%lx vma_start=0x%lx vma_size=%lu",
+               ret,
+               vma->vm_flags,
+               vma->vm_start,
+               vma->vm_end-vma->vm_start);
+done:
+       spin_unlock_irqrestore(&ctx->lock, flags);
+
+       return ret;
+}
+
+/*
+ * Extract one message from queue.
+ *
+ * return:
+ *     -EAGAIN:  when non-blocking and nothing is* in the queue.
+ *     -ERESTARTSYS: when blocking and signal is pending
+ *     Otherwise returns size of message (sizeof(pfarg_msg))
+ */
+ssize_t __pfm_read(struct pfm_context *ctx, union pfarg_msg *msg_buf, int non_block)
+{
+       ssize_t ret = 0;
+       unsigned long flags;
+       DECLARE_WAITQUEUE(wait, current);
+
+       /*
+        * we must masks interrupts to avoid a race condition
+        * with the PMU interrupt handler.
+        */
+       spin_lock_irqsave(&ctx->lock, flags);
+
+       while (pfm_msgq_is_empty(ctx)) {
+
+               /*
+                * handle non-blocking reads
+                * return -EAGAIN
+                */
+               ret = -EAGAIN;
+               if (non_block)
+                       break;
+
+               add_wait_queue(&ctx->msgq_wait, &wait);
+               set_current_state(TASK_INTERRUPTIBLE);
+
+               spin_unlock_irqrestore(&ctx->lock, flags);
+
+               schedule();
+
+               /*
+                * during this window, another thread may call
+                * pfm_read() and steal our message
+                */
+
+               spin_lock_irqsave(&ctx->lock, flags);
+
+               remove_wait_queue(&ctx->msgq_wait, &wait);
+               set_current_state(TASK_RUNNING);
+
+               /*
+                * check for pending signals
+                * return -ERESTARTSYS
+                */
+               ret = -ERESTARTSYS;
+               if (signal_pending(current))
+                       break;
+
+               /*
+                * we may have a message
+                */
+               ret = 0;
+       }
+
+       /*
+        * extract message
+        */
+       if (ret == 0) {
+               /*
+                * copy the oldest message into msg_buf.
+                * We cannot directly call copy_to_user()
+                * because interrupts masked. This is done
+                * in the caller
+                */
+               pfm_get_next_msg(ctx, msg_buf);
+
+               ret = sizeof(*msg_buf);
+
+               PFM_DBG("extracted type=%d", msg_buf->type);
+       }
+
+       spin_unlock_irqrestore(&ctx->lock, flags);
+
+       PFM_DBG("blocking=%d ret=%zd", non_block, ret);
+
+       return ret;
+}
+
+static ssize_t pfm_read(struct file *filp, char __user *buf, size_t size,
+                       loff_t *ppos)
+{
+       struct pfm_context *ctx;
+       union pfarg_msg msg_buf;
+       int non_block, ret;
+
+       PFM_DBG_ovfl("buf=%p size=%zu", buf, size);
+
+       ctx = filp->private_data;
+       if (ctx == NULL) {
+               PFM_ERR("no ctx for pfm_read");
+               return -EINVAL;
+       }
+
+       non_block = filp->f_flags & O_NONBLOCK;
+
+#ifdef CONFIG_IA64_PERFMON_COMPAT
+       /*
+        * detect IA-64 v2.0 context read (message size is different)
+        * nops on all other architectures
+        */
+       if (unlikely(ctx->flags.ia64_v20_compat))
+               return pfm_arch_compat_read(ctx,  buf, non_block, size);
+#endif
+       /*
+        * cannot extract partial messages.
+        * check even when there is no message
+        *
+        * cannot extract more than one message per call. Bytes
+        * above sizeof(msg) are ignored.
+        */
+       if (size < sizeof(msg_buf)) {
+               PFM_DBG("message is too small size=%zu must be >=%zu)",
+                       size,
+                       sizeof(msg_buf));
+               return -EINVAL;
+       }
+
+       ret =  __pfm_read(ctx, &msg_buf, non_block);
+       if (ret > 0) {
+               if (copy_to_user(buf, &msg_buf, sizeof(msg_buf)))
+                       ret = -EFAULT;
+       }
+       PFM_DBG_ovfl("ret=%d", ret);
+       return ret;
+}
+
+static ssize_t pfm_write(struct file *file, const char __user *ubuf,
+                         size_t size, loff_t *ppos)
+{
+       PFM_DBG("pfm_write called");
+       return -EINVAL;
+}
+
+static unsigned int pfm_poll(struct file *filp, poll_table *wait)
+{
+       struct pfm_context *ctx;
+       unsigned long flags;
+       unsigned int mask = 0;
+
+       PFM_DBG("pfm_file_ops");
+
+       if (filp->f_op != &pfm_file_ops) {
+               PFM_ERR("pfm_poll bad magic");
+               return 0;
+       }
+
+       ctx = filp->private_data;
+       if (ctx == NULL) {
+               PFM_ERR("pfm_poll no ctx");
+               return 0;
+       }
+
+       PFM_DBG("before poll_wait");
+
+       poll_wait(filp, &ctx->msgq_wait, wait);
+
+       /*
+        * pfm_msgq_is_empty() is non-atomic
+        *
+        * filp is protected by fget() at upper level
+        * context cannot be closed by another thread.
+        *
+        * There may be a race with a PMU interrupt adding
+        * messages to the queue. But we are interested in
+        * queue not empty, so adding more messages should
+        * not really be a problem.
+        *
+        * There may be a race with another thread issuing
+        * a read() and stealing messages from the queue thus
+        * may return the wrong answer. This could potentially
+        * lead to a blocking read, because nothing is
+        * available in the queue
+        */
+       spin_lock_irqsave(&ctx->lock, flags);
+
+       if (!pfm_msgq_is_empty(ctx))
+               mask =  POLLIN | POLLRDNORM;
+
+       spin_unlock_irqrestore(&ctx->lock, flags);
+
+       PFM_DBG("after poll_wait mask=0x%x", mask);
+
+       return mask;
+}
+
+static int pfm_ioctl(struct inode *inode, struct file *file, unsigned int cmd,
+                    unsigned long arg)
+{
+       PFM_DBG("pfm_ioctl called");
+       return -EINVAL;
+}
+
+/*
+ * interrupt cannot be masked when entering this function
+ */
+static inline int __pfm_fasync(int fd, struct file *filp,
+                              struct pfm_context *ctx, int on)
+{
+       int ret;
+
+       PFM_DBG("in  fd=%d on=%d async_q=%p",
+               fd,
+               on,
+               ctx->async_queue);
+
+       ret = fasync_helper(fd, filp, on, &ctx->async_queue);
+
+       PFM_DBG("out fd=%d on=%d async_q=%p ret=%d",
+               fd,
+               on,
+               ctx->async_queue, ret);
+
+       return ret;
+}
+
+static int pfm_fasync(int fd, struct file *filp, int on)
+{
+       struct pfm_context *ctx;
+       int ret;
+
+       PFM_DBG("pfm_file_ops");
+
+       ctx = filp->private_data;
+       if (ctx == NULL) {
+               PFM_ERR("pfm_fasync no ctx");
+               return -EBADF;
+       }
+
+       /*
+        * we cannot mask interrupts during this call because this may
+        * may go to sleep if memory is not readily avalaible.
+        *
+        * We are protected from the context disappearing by the
+        * get_fd()/put_fd() done in caller. Serialization of this function
+        * is ensured by caller.
+        */
+       ret = __pfm_fasync(fd, filp, ctx, on);
+
+       PFM_DBG("pfm_fasync called on fd=%d on=%d async_queue=%p ret=%d",
+               fd,
+               on,
+               ctx->async_queue, ret);
+
+       return ret;
+}
+
+#ifdef CONFIG_SMP
+static void __pfm_close_remote_cpu(void *info)
+{
+       struct pfm_context *ctx = info;
+       int can_release;
+
+       BUG_ON(ctx != __get_cpu_var(pmu_ctx));
+
+       /*
+        * we are in IPI interrupt handler which has always higher
+        * priority than PMU interrupt, therefore we do not need to
+        * mask interrupts. context locking is not needed because we
+        * are in close(), no more user references.
+        *
+        * can_release is ignored, release done on calling CPU
+        */
+       __pfm_unload_context(ctx, &can_release);
+
+       /*
+        * we cannot free context here because we are in_interrupt().
+        * we free on the calling CPU
+        */
+}
+
+static int pfm_close_remote_cpu(u32 cpu, struct pfm_context *ctx)
+{
+       BUG_ON(irqs_disabled());
+       return smp_call_function_single(cpu, __pfm_close_remote_cpu, ctx, 1);
+}
+#endif /* CONFIG_SMP */
+
+/*
+ * called either on explicit close() or from exit_files().
+ * Only the LAST user of the file gets to this point, i.e., it is
+ * called only ONCE.
+ *
+ * IMPORTANT: we get called ONLY when the refcnt on the file gets to zero
+ * (fput()),i.e, last task to access the file. Nobody else can access the
+ * file at this point.
+ *
+ * When called from exit_files(), the VMA has been freed because exit_mm()
+ * is executed before exit_files().
+ *
+ * When called from exit_files(), the current task is not yet ZOMBIE but we
+ * flush the PMU state to the context.
+ */
+int __pfm_close(struct pfm_context *ctx, struct file *filp)
+{
+       unsigned long flags;
+       int state;
+       int can_free = 1, can_unload = 1;
+       int is_system, can_release = 0;
+       u32 cpu;
+
+       /*
+        * no risk of ctx of filp disappearing so we can operate outside
+        * of spin_lock(). fasync_helper() runs with interrupts masked,
+        * thus there is no risk with the PMU interrupt handler
+        *
+        * In case of zombie, we will not have the async struct anymore
+        * thus kill_fasync() will not do anything
+        *
+        * fd is not used when removing the entry so we pass -1
+        */
+       if (filp->f_flags & FASYNC)
+               __pfm_fasync (-1, filp, ctx, 0);
+
+       spin_lock_irqsave(&ctx->lock, flags);
+
+       state = ctx->state;
+       is_system = ctx->flags.system;
+       cpu = ctx->cpu;
+
+       PFM_DBG("state=%d", state);
+
+       /*
+        * check if unload is needed
+        */
+       if (state == PFM_CTX_UNLOADED)
+               goto doit;
+
+#ifdef CONFIG_SMP
+       /*
+        * we need to release the resource on the ORIGINAL cpu.
+        * we need to release the context lock to avoid deadlocks
+        * on the original CPU, especially in the context switch
+        * routines. It is safe to unlock because we are in close(),
+        * in other words, there is no more access from user level.
+        * we can also unmask interrupts on this CPU because the
+        * context is running on the original CPU. Context will be
+        * unloaded and the session will be released on the original
+        * CPU. Upon return, the caller is guaranteed that the context
+        * is gone from original CPU.
+        */
+       if (is_system && cpu != smp_processor_id()) {
+               spin_unlock_irqrestore(&ctx->lock, flags);
+               pfm_close_remote_cpu(cpu, ctx);
+               can_release = 1;
+               goto free_it;
+       }
+
+       if (!is_system && ctx->task != current) {
+               /*
+                * switch context to zombie state
+                */
+               ctx->state = PFM_CTX_ZOMBIE;
+
+               PFM_DBG("zombie ctx for [%d]", ctx->task->pid);
+               /*
+                * must check if other thread is using block overflow
+                * notification mode. If so make sure it will not block
+                * because there will not be any pfm_restart() issued.
+                * When the thread notices the ZOMBIE state, it will clean
+                * up what is left of the context
+                */
+               if (state == PFM_CTX_MASKED && ctx->flags.block) {
+                       /*
+                        * force task to wake up from MASKED state
+                        */
+                       PFM_DBG("waking up [%d]", ctx->task->pid);
+
+                       complete(&ctx->restart_complete);
+               }
+               /*
+                * PMU session will be release by monitored task when it notices
+                * ZOMBIE state as part of pfm_unload_context()
+                */
+               can_unload = can_free = 0;
+       }
+#endif
+       if (can_unload)
+               __pfm_unload_context(ctx, &can_release);
+doit:
+       spin_unlock_irqrestore(&ctx->lock, flags);
+
+#ifdef CONFIG_SMP
+free_it:
+#endif
+       if (can_release)
+               pfm_session_release(is_system, cpu);
+
+       if (can_free)
+               pfm_free_context(ctx);
+
+       return 0;
+}
+
+static int pfm_close(struct inode *inode, struct file *filp)
+{
+       struct pfm_context *ctx;
+
+       PFM_DBG("called filp=%p", filp);
+
+       ctx = filp->private_data;
+       if (ctx == NULL) {
+               PFM_ERR("no ctx");
+               return -EBADF;
+       }
+       return __pfm_close(ctx, filp);
+}
+
+static int pfm_no_open(struct inode *irrelevant, struct file *dontcare)
+{
+       PFM_DBG("pfm_file_ops");
+
+       return -ENXIO;
+}
+
+
+const struct file_operations pfm_file_ops = {
+       .llseek = no_llseek,
+       .read = pfm_read,
+       .write = pfm_write,
+       .poll = pfm_poll,
+       .ioctl = pfm_ioctl,
+       .open = pfm_no_open, /* special open to disallow open via /proc */
+       .fasync = pfm_fasync,
+       .release = pfm_close,
+       .mmap = pfm_mmap
+};
+
+static int pfmfs_get_sb(struct file_system_type *fs_type,
+                       int flags, const char *dev_name,
+                       void *data, struct vfsmount *mnt)
+{
+       return get_sb_pseudo(fs_type, "pfm:", NULL, PFMFS_MAGIC, mnt);
+}
+
+static struct file_system_type pfm_fs_type = {
+       .name     = "pfmfs",
+       .get_sb   = pfmfs_get_sb,
+       .kill_sb  = kill_anon_super,
+};
+
+/*
+ * pfmfs should _never_ be mounted by userland - too much of security hassle,
+ * no real gain from having the whole whorehouse mounted. So we don't need
+ * any operations on the root directory. However, we need a non-trivial
+ * d_name - pfm: will go nicely and kill the special-casing in procfs.
+ */
+static struct vfsmount *pfmfs_mnt;
+
+int __init pfm_init_fs(void)
+{
+       int err = register_filesystem(&pfm_fs_type);
+       if (!err) {
+               pfmfs_mnt = kern_mount(&pfm_fs_type);
+               err = PTR_ERR(pfmfs_mnt);
+               if (IS_ERR(pfmfs_mnt))
+                       unregister_filesystem(&pfm_fs_type);
+               else
+                       err = 0;
+       }
+       return err;
+}
+
+int pfm_alloc_fd(struct file **cfile)
+{
+       int fd, ret = 0;
+       struct file *file = NULL;
+       struct inode * inode;
+       char name[32];
+       struct qstr this;
+
+       fd = get_unused_fd();
+       if (fd < 0)
+               return -ENFILE;
+
+       ret = -ENFILE;
+
+       file = get_empty_filp();
+       if (!file)
+               goto out;
+
+       /*
+        * allocate a new inode
+        */
+       inode = new_inode(pfmfs_mnt->mnt_sb);
+       if (!inode)
+               goto out;
+
+       PFM_DBG("new inode ino=%ld @%p", inode->i_ino, inode);
+
+       inode->i_sb = pfmfs_mnt->mnt_sb;
+       inode->i_mode = S_IFCHR|S_IRUGO;
+       inode->i_uid = current->fsuid;
+       inode->i_gid = current->fsgid;
+
+       sprintf(name, "[%lu]", inode->i_ino);
+       this.name = name;
+       this.hash = inode->i_ino;
+       this.len = strlen(name);
+
+       ret = -ENOMEM;
+
+       /*
+        * allocate a new dcache entry
+        */
+       file->f_dentry = d_alloc(pfmfs_mnt->mnt_sb->s_root, &this);
+       if (!file->f_dentry)
+               goto out;
+
+       file->f_dentry->d_op = &pfmfs_dentry_operations;
+
+       d_add(file->f_dentry, inode);
+       file->f_vfsmnt = mntget(pfmfs_mnt);
+       file->f_mapping = inode->i_mapping;
+
+       file->f_op = &pfm_file_ops;
+       file->f_mode = FMODE_READ;
+       file->f_flags = O_RDONLY;
+       file->f_pos  = 0;
+
+       *cfile = file;
+
+       return fd;
+out:
+       if (file)
+               put_filp(file);
+       put_unused_fd(fd);
+       return ret;
+}
diff --git a/perfmon/perfmon_fmt.c b/perfmon/perfmon_fmt.c
new file mode 100644
index 0000000..27c4340
--- /dev/null
+++ b/perfmon/perfmon_fmt.c
@@ -0,0 +1,219 @@
+/*
+ * perfmon_fmt.c: perfmon2 sampling buffer format management
+ *
+ * This file implements the perfmon2 interface which
+ * provides access to the hardware performance counters
+ * of the host processor.
+ *
+ * The initial version of perfmon.c was written by
+ * Ganesh Venkitachalam, IBM Corp.
+ *
+ * Then it was modified for perfmon-1.x by Stephane Eranian and
+ * David Mosberger, Hewlett Packard Co.
+ *
+ * Version Perfmon-2.x is a complete rewrite of perfmon-1.x
+ * by Stephane Eranian, Hewlett Packard Co.
+ *
+ * Copyright (c) 1999-2006 Hewlett-Packard Development Company, L.P.
+ * Contributed by Stephane Eranian <eranian@hpl.hp.com>
+ *                David Mosberger-Tang <davidm@hpl.hp.com>
+ *
+ * More information about perfmon available at:
+ *     http://perfmon2.sf.net
+ *
+ * This program is free software; you can redistribute it and/or
+ * modify it under the terms of version 2 of the GNU General Public
+ * License as published by the Free Software Foundation.
+ *
+ * This program is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+ * General Public License for more details.
+ *
+ * You should have received a copy of the GNU General Public License
+ * along with this program; if not, write to the Free Software
+ * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA
+ * 02111-1307 USA
+ */
+#include <linux/module.h>
+#include <linux/perfmon_kern.h>
+#include "perfmon_priv.h"
+
+static __cacheline_aligned_in_smp DEFINE_SPINLOCK(pfm_smpl_fmt_lock);
+static LIST_HEAD(pfm_smpl_fmt_list);
+
+static inline int fmt_is_mod(struct pfm_smpl_fmt *f)
+{
+       return !(f->fmt_flags & PFM_FMTFL_IS_BUILTIN);
+}
+
+static struct pfm_smpl_fmt *pfm_find_fmt(char *name)
+{
+       struct pfm_smpl_fmt *entry;
+
+       list_for_each_entry(entry, &pfm_smpl_fmt_list, fmt_list) {
+               if (!strcmp(entry->fmt_name, name))
+                       return entry;
+       }
+       return NULL;
+}
+/*
+ * find a buffer format based on its name
+ */
+struct pfm_smpl_fmt *pfm_smpl_fmt_get(char *name)
+{
+       struct pfm_smpl_fmt *fmt;
+
+       spin_lock(&pfm_smpl_fmt_lock);
+
+       fmt = pfm_find_fmt(name);
+
+       /*
+        * increase module refcount
+        */
+       if (fmt && fmt_is_mod(fmt) && !try_module_get(fmt->owner))
+               fmt = NULL;
+
+       spin_unlock(&pfm_smpl_fmt_lock);
+
+       return fmt;
+}
+
+void pfm_smpl_fmt_put(struct pfm_smpl_fmt *fmt)
+{
+       if (fmt == NULL || !fmt_is_mod(fmt))
+               return;
+       BUG_ON(fmt->owner == NULL);
+
+       spin_lock(&pfm_smpl_fmt_lock);
+       module_put(fmt->owner);
+       spin_unlock(&pfm_smpl_fmt_lock);
+}
+
+int pfm_fmt_register(struct pfm_smpl_fmt *fmt)
+{
+       int ret = 0;
+
+       if (perfmon_disabled) {
+               PFM_INFO("perfmon disabled, cannot add sampling format");
+               return -ENOSYS;
+       }
+
+       /* some sanity checks */
+       if (fmt == NULL) {
+               PFM_INFO("perfmon: NULL format for register");
+               return -EINVAL;
+       }
+
+       if (fmt->fmt_name == NULL) {
+               PFM_INFO("perfmon: format has no name");
+               return -EINVAL;
+       }
+
+       if (fmt->fmt_qdepth > PFM_MSGS_COUNT) {
+               PFM_INFO("perfmon: format %s requires %u msg queue depth (max %d)",
+                      fmt->fmt_name,
+                      fmt->fmt_qdepth,
+                      PFM_MSGS_COUNT);
+               return -EINVAL;
+       }
+
+       /*
+        * fmt is missing the initialization of .owner = THIS_MODULE
+        * this is only valid when format is compiled as a module
+        */
+       if (fmt->owner == NULL && fmt_is_mod(fmt)) {
+               PFM_INFO("format %s has no module owner", fmt->fmt_name);
+               return -EINVAL;
+       }
+       /*
+        * we need at least a handler
+        */
+       if (fmt->fmt_handler == NULL) {
+               PFM_INFO("format %s has no handler", fmt->fmt_name);
+               return -EINVAL;
+       }
+
+       /*
+        * format argument size cannot be bigger than PAGE_SIZE
+        */
+       if (fmt->fmt_arg_size > PAGE_SIZE) {
+               PFM_INFO("format %s arguments too big", fmt->fmt_name);
+               return -EINVAL;
+       }
+
+       spin_lock(&pfm_smpl_fmt_lock);
+
+       /*
+        * because of sysfs, we cannot have two formats with the same name
+        */
+       if (pfm_find_fmt(fmt->fmt_name)) {
+               PFM_INFO("format %s already registered", fmt->fmt_name);
+               ret = -EBUSY;
+               goto out;
+       }
+
+       ret = pfm_sysfs_add_fmt(fmt);
+       if (ret) {
+               PFM_INFO("sysfs cannot add format entry for %s", fmt->fmt_name);
+               goto out;
+       }
+
+       list_add(&fmt->fmt_list, &pfm_smpl_fmt_list);
+
+       PFM_INFO("added sampling format %s", fmt->fmt_name);
+out:
+       spin_unlock(&pfm_smpl_fmt_lock);
+
+       return ret;
+}
+EXPORT_SYMBOL(pfm_fmt_register);
+
+int pfm_fmt_unregister(struct pfm_smpl_fmt *fmt)
+{
+       struct pfm_smpl_fmt *fmt2;
+       int ret = 0;
+
+       if (!fmt || !fmt->fmt_name) {
+               PFM_DBG("invalid fmt");
+               return -EINVAL;
+       }
+
+       spin_lock(&pfm_smpl_fmt_lock);
+
+       fmt2 = pfm_find_fmt(fmt->fmt_name);
+       if (!fmt) {
+               PFM_INFO("unregister failed, format not registered");
+               ret = -EINVAL;
+               goto out;
+       }
+       list_del_init(&fmt->fmt_list);
+
+       pfm_sysfs_remove_fmt(fmt);
+
+       PFM_INFO("removed sampling format: %s", fmt->fmt_name);
+
+out:
+       spin_unlock(&pfm_smpl_fmt_lock);
+       return ret;
+
+}
+EXPORT_SYMBOL(pfm_fmt_unregister);
+
+/*
+ * we defer adding the builtin formats to /sys/kernel/perfmon/formats
+ * until after the pfm sysfs subsystem is initialized. This function
+ * is called from pfm_init_sysfs()
+ */
+void __init pfm_sysfs_builtin_fmt_add(void)
+{
+       struct pfm_smpl_fmt *entry;
+
+       /*
+        * locking not needed, kernel not fully booted
+        * when called
+        */
+       list_for_each_entry(entry, &pfm_smpl_fmt_list, fmt_list) {
+               pfm_sysfs_add_fmt(entry);
+       }
+}
diff --git a/perfmon/perfmon_hotplug.c b/perfmon/perfmon_hotplug.c
new file mode 100644
index 0000000..eaaba81
--- /dev/null
+++ b/perfmon/perfmon_hotplug.c
@@ -0,0 +1,151 @@
+/*
+ * perfmon_hotplug.c: handling of CPU hotplug
+ *
+ * The initial version of perfmon.c was written by
+ * Ganesh Venkitachalam, IBM Corp.
+ *
+ * Then it was modified for perfmon-1.x by Stephane Eranian and
+ * David Mosberger, Hewlett Packard Co.
+ *
+ * Version Perfmon-2.x is a complete rewrite of perfmon-1.x
+ * by Stephane Eranian, Hewlett Packard Co.
+ *
+ * Copyright (c) 1999-2006 Hewlett-Packard Development Company, L.P.
+ * Contributed by Stephane Eranian <eranian@hpl.hp.com>
+ *                David Mosberger-Tang <davidm@hpl.hp.com>
+ *
+ * More information about perfmon available at:
+ *     http://perfmon2.sf.net
+ *
+ * This program is free software; you can redistribute it and/or
+ * modify it under the terms of version 2 of the GNU General Public
+ * License as published by the Free Software Foundation.
+ *
+ * This program is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+ * General Public License for more details.
+ *
+ * You should have received a copy of the GNU General Public License
+ * along with this program; if not, write to the Free Software
+ * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA
+ * 02111-1307 USA
+ */
+#include <linux/kernel.h>
+#include <linux/perfmon_kern.h>
+#include <linux/cpu.h>
+#include "perfmon_priv.h"
+
+#ifndef CONFIG_HOTPLUG_CPU
+void pfm_cpu_disable(void)
+{}
+
+int __init pfm_init_hotplug(void)
+{
+       return 0;
+}
+#else /* CONFIG_HOTPLUG_CPU */
+/*
+ * CPU hotplug event nofication callback
+ *
+ * We use the callback to do manage the sysfs interface.
+ * Note that the actual shutdown of monitoring on the CPU
+ * is done in pfm_cpu_disable(), see comments there for more
+ * information.
+ */
+static int pfm_cpu_notify(struct notifier_block *nfb,
+                         unsigned long action, void *hcpu)
+{
+       unsigned int cpu = (unsigned long)hcpu;
+       int ret = NOTIFY_OK;
+
+       pfm_pmu_conf_get(0);
+
+       switch (action) {
+       case CPU_ONLINE:
+               pfm_debugfs_add_cpu(cpu);
+               PFM_INFO("CPU%d is online", cpu);
+               break;
+       case CPU_UP_PREPARE:
+               PFM_INFO("CPU%d prepare online", cpu);
+               break;
+       case CPU_UP_CANCELED:
+               pfm_debugfs_del_cpu(cpu);
+               PFM_INFO("CPU%d is up canceled", cpu);
+               break;
+       case CPU_DOWN_PREPARE:
+               PFM_INFO("CPU%d prepare offline", cpu);
+               break;
+       case CPU_DOWN_FAILED:
+               PFM_INFO("CPU%d is down failed", cpu);
+               break;
+       case CPU_DEAD:
+               pfm_debugfs_del_cpu(cpu);
+               PFM_INFO("CPU%d is offline", cpu);
+               break;
+       }
+       pfm_pmu_conf_put();
+       return ret;
+}
+
+/*
+ * called from cpu_disable() to detach the perfmon context
+ * from the CPU going down.
+ *
+ * We cannot use the cpu hotplug notifier because we MUST run
+ * on the CPU that is going down to save the PMU state
+ */
+void pfm_cpu_disable(void)
+{
+       struct pfm_context *ctx;
+       unsigned long flags;
+       int is_system, release_info = 0;
+       u32 cpu;
+       int r;
+
+       ctx = __get_cpu_var(pmu_ctx);
+       if (ctx == NULL)
+               return;
+
+       is_system = ctx->flags.system;
+       cpu = ctx->cpu;
+
+       /*
+        * context is LOADED or MASKED
+        *
+        * we unload from CPU. That stops monitoring and does
+        * all the bookeeping of saving values and updating duration
+        */
+       spin_lock_irqsave(&ctx->lock, flags);
+       if (is_system)
+               __pfm_unload_context(ctx, &release_info);
+       spin_unlock_irqrestore(&ctx->lock, flags);
+
+       /*
+        * cancel timer
+        */
+       if (release_info & 0x2) {
+               r = hrtimer_cancel(&__get_cpu_var(pfm_hrtimer));
+               PFM_DBG("timeout cancel=%d", r);
+       }
+
+       if (release_info & 0x1)
+               pfm_session_release(is_system, cpu);
+}
+
+static struct notifier_block pfm_cpu_notifier = {
+       .notifier_call = pfm_cpu_notify
+};
+
+int __init pfm_init_hotplug(void)
+{
+       int ret = 0;
+       /*
+        * register CPU hotplug event notifier
+        */
+       ret = register_cpu_notifier(&pfm_cpu_notifier);
+       if (!ret)
+               PFM_LOG("CPU hotplug support enabled");
+       return ret;
+}
+#endif /* CONFIG_HOTPLUG_CPU */
diff --git a/perfmon/perfmon_init.c b/perfmon/perfmon_init.c
new file mode 100644
index 0000000..bbb6e4d
--- /dev/null
+++ b/perfmon/perfmon_init.c
@@ -0,0 +1,131 @@
+/*
+ * perfmon.c: perfmon2 global initialization functions
+ *
+ * This file implements the perfmon2 interface which
+ * provides access to the hardware performance counters
+ * of the host processor.
+ *
+ *
+ * The initial version of perfmon.c was written by
+ * Ganesh Venkitachalam, IBM Corp.
+ *
+ * Then it was modified for perfmon-1.x by Stephane Eranian and
+ * David Mosberger, Hewlett Packard Co.
+ *
+ * Version Perfmon-2.x is a complete rewrite of perfmon-1.x
+ * by Stephane Eranian, Hewlett Packard Co.
+ *
+ * Copyright (c) 1999-2006 Hewlett-Packard Development Company, L.P.
+ * Contributed by Stephane Eranian <eranian@hpl.hp.com>
+ *                David Mosberger-Tang <davidm@hpl.hp.com>
+ *
+ * More information about perfmon available at:
+ *     http://perfmon2.sf.net
+ *
+ * This program is free software; you can redistribute it and/or
+ * modify it under the terms of version 2 of the GNU General Public
+ * License as published by the Free Software Foundation.
+ *
+ * This program is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+ * General Public License for more details.
+ *
+ * You should have received a copy of the GNU General Public License
+ * along with this program; if not, write to the Free Software
+ * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA
+ * 02111-1307 USA
+ */
+#include <linux/kernel.h>
+#include <linux/perfmon_kern.h>
+#include "perfmon_priv.h"
+
+/*
+ * external variables
+ */
+DEFINE_PER_CPU(struct task_struct *, pmu_owner);
+DEFINE_PER_CPU(struct pfm_context  *, pmu_ctx);
+DEFINE_PER_CPU(u64, pmu_activation_number);
+DEFINE_PER_CPU(struct pfm_stats, pfm_stats);
+DEFINE_PER_CPU(struct hrtimer, pfm_hrtimer);
+
+
+int perfmon_disabled;  /* >0 if perfmon is disabled */
+
+/*
+ * called from cpu_init() and pfm_pmu_register()
+ */
+void __pfm_init_percpu(void *dummy)
+{
+       struct hrtimer *h;
+
+       h = &__get_cpu_var(pfm_hrtimer);
+
+       pfm_arch_init_percpu();
+
+       /*
+        * initialize per-cpu high res timer
+        */
+       hrtimer_init(h, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
+#ifdef CONFIG_HIGH_RES_TIMERS
+       /*
+        * avoid potential deadlock on the runqueue lock
+        * during context switch when multiplexing. Situation
+        * arises on architectures which run switch_to() with
+        * the runqueue lock held, e.g., x86. On others, e.g.,
+        * IA-64, the problem does not exist.
+        * Setting the callback mode to HRTIMER_CB_IRQSAFE_UNOCKED
+        * such that the callback routine is only called on hardirq
+        * context not on softirq, thus the context switch will not
+        * end up trying to wakeup the softirqd
+        */
+       h->cb_mode = HRTIMER_CB_IRQSAFE_UNLOCKED;
+#endif
+       h->function = pfm_handle_switch_timeout;
+}
+
+/*
+ * global initialization routine, executed only once
+ */
+int __init pfm_init(void)
+{
+       PFM_LOG("version %u.%u", PFM_VERSION_MAJ, PFM_VERSION_MIN);
+
+       if (pfm_init_ctx())
+               goto error_disable;
+
+
+       if (pfm_init_sets())
+               goto error_disable;
+
+       if (pfm_init_fs())
+               goto error_disable;
+
+       if (pfm_init_sysfs())
+               goto error_disable;
+
+       /* not critical, so no error checking */
+       pfm_init_debugfs();
+
+       /*
+        * one time, arch-specific global initialization
+        */
+       if (pfm_arch_init())
+               goto error_disable;
+
+       if (pfm_init_hotplug())
+               goto error_disable;
+       return 0;
+
+error_disable:
+       PFM_ERR("perfmon is disabled due to initialization error");
+       perfmon_disabled = 1;
+       return -1;
+}
+
+/*
+ * must use subsys_initcall() to ensure that the perfmon2 core
+ * is initialized before any PMU description module when they are
+ * compiled in.
+ */
+subsys_initcall(pfm_init);
diff --git a/perfmon/perfmon_intr.c b/perfmon/perfmon_intr.c
new file mode 100644
index 0000000..c5e3cda
--- /dev/null
+++ b/perfmon/perfmon_intr.c
@@ -0,0 +1,648 @@
+/*
+ * perfmon_intr.c: perfmon2 interrupt handling
+ *
+ * This file implements the perfmon2 interface which
+ * provides access to the hardware performance counters
+ * of the host processor.
+ *
+ * The initial version of perfmon.c was written by
+ * Ganesh Venkitachalam, IBM Corp.
+ *
+ * Then it was modified for perfmon-1.x by Stephane Eranian and
+ * David Mosberger, Hewlett Packard Co.
+ *
+ * Version Perfmon-2.x is a complete rewrite of perfmon-1.x
+ * by Stephane Eranian, Hewlett Packard Co.
+ *
+ * Copyright (c) 1999-2006 Hewlett-Packard Development Company, L.P.
+ * Contributed by Stephane Eranian <eranian@hpl.hp.com>
+ *                David Mosberger-Tang <davidm@hpl.hp.com>
+ *
+ * More information about perfmon available at:
+ *     http://perfmon2.sf.net
+ *
+ * This program is free software; you can redistribute it and/or
+ * modify it under the terms of version 2 of the GNU General Public
+ * License as published by the Free Software Foundation.
+ *
+ * This program is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+ * General Public License for more details.
+ *
+ * You should have received a copy of the GNU General Public License
+ * along with this program; if not, write to the Free Software
+ * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA
+ * 02111-1307 USA
+ */
+#include <linux/kernel.h>
+#include <linux/module.h>
+#include <linux/perfmon_kern.h>
+#include "perfmon_priv.h"
+
+/**
+ * pfm_intr_process_64bit_ovfls - handle 64-bit counter emulation
+ * @ctx: context to operate on
+ * @set: set to operate on
+ *
+ * The function returns the number of 64-bit overflows detected.
+ *
+ * 64-bit software pmds are updated for overflowed pmd registers
+ * the set->reset_pmds is updated to the list of pmds to reset
+ *
+ * In any case, set->npend_ovfls is cleared
+ */
+static u16 pfm_intr_process_64bit_ovfls(struct pfm_context *ctx,
+                                       struct pfm_event_set *set,
+                                       u32 *ovfl_ctrl)
+{
+       u16 i, num_ovfls, max_pmd, max_intr;
+       u16 num_64b_ovfls, has_ovfl_sw, must_switch;
+       u64 ovfl_thres, old_val, new_val, ovfl_mask;
+
+       num_64b_ovfls = must_switch = 0;
+
+       ovfl_mask = pfm_pmu_conf->ovfl_mask;
+       max_pmd = ctx->regs.max_pmd;
+       max_intr = ctx->regs.max_intr_pmd;
+
+       num_ovfls = set->npend_ovfls;
+       has_ovfl_sw = set->flags & PFM_SETFL_OVFL_SWITCH;
+
+       bitmap_zero(cast_ulp(set->reset_pmds), max_pmd);
+
+       for (i = ctx->regs.first_intr_pmd; num_ovfls; i++) {
+               /*
+                * skip pmd which did not overflow
+                */
+               if (!test_bit(i, cast_ulp(set->povfl_pmds)))
+                       continue;
+
+               num_ovfls--;
+
+               /*
+                * Update software value for counters ONLY
+                *
+                * Note that the pmd is not necessarily 0 at this point as
+                * qualified events may have happened before the PMU was
+                * frozen. The residual count is not taken into consideration
+                * here but will be with any read of the pmd
+                */
+               ovfl_thres = set->pmds[i].ovflsw_thres;
+
+               if (likely(test_bit(i, cast_ulp(ctx->regs.cnt_pmds)))) {
+                       old_val = new_val = set->pmds[i].value;
+                       new_val += 1 + ovfl_mask;
+                       set->pmds[i].value = new_val;
+               }  else {
+                       /*
+                        * for non counters which interrupt, e.g., AMD IBS,
+                        * we consider this equivalent to a 64-bit counter
+                        * overflow.
+                        */
+                       old_val = 1; new_val = 0;
+               }
+
+               /*
+                * check for 64-bit overflow condition
+                */
+               if (likely(old_val > new_val)) {
+                       num_64b_ovfls++;
+                       if (has_ovfl_sw && ovfl_thres > 0) {
+                               if (ovfl_thres == 1)
+                                       must_switch = 1;
+                               set->pmds[i].ovflsw_thres = ovfl_thres - 1;
+                       }
+
+                       /*
+                        * what to reset because of this overflow
+                        * - the overflowed register
+                        * - its reset_smpls
+                        */
+                       __set_bit(i, cast_ulp(set->reset_pmds));
+
+                       bitmap_or(cast_ulp(set->reset_pmds),
+                                 cast_ulp(set->reset_pmds),
+                                 cast_ulp(set->pmds[i].reset_pmds),
+                                 max_pmd);
+               } else {
+                       /*
+                        * only keep track of 64-bit overflows or
+                        * assimilated
+                        */
+                       __clear_bit(i, cast_ulp(set->povfl_pmds));
+
+                       /*
+                        * on some PMU, it may be necessary to re-arm the PMD
+                        */
+                       pfm_arch_ovfl_reset_pmd(ctx, i);
+               }
+
+               PFM_DBG_ovfl("ovfl=%s pmd%u new=0x%llx old=0x%llx "
+                            "hw_pmd=0x%llx o_pmds=0x%llx must_switch=%u "
+                            "o_thres=%llu o_thres_ref=%llu",
+                            old_val > new_val ? "64-bit" : "HW",
+                            i,
+                            (unsigned long long)new_val,
+                            (unsigned long long)old_val,
+                            (unsigned long long)pfm_read_pmd(ctx, i),
+                            (unsigned long long)set->povfl_pmds[0],
+                            must_switch,
+                            (unsigned long long)set->pmds[i].ovflsw_thres,
+                            (unsigned long long)set->pmds[i].ovflsw_ref_thres);
+       }
+       /*
+        * update public bitmask of 64-bit overflowed pmds
+        */
+       if (num_64b_ovfls)
+               bitmap_copy(cast_ulp(set->ovfl_pmds), cast_ulp(set->povfl_pmds),
+                           max_intr);
+
+       if (must_switch)
+               *ovfl_ctrl |= PFM_OVFL_CTRL_SWITCH;
+
+       /*
+        * mark the overflows as consumed
+        */
+       set->npend_ovfls = 0;
+       bitmap_zero(cast_ulp(set->povfl_pmds), max_intr);
+
+       return num_64b_ovfls;
+}
+
+/**
+ * pfm_intr_get_smpl_pmds_values - copy 64-bit pmd values for sampling format
+ * @ctx: context to work on
+ * @set: current event set
+ * @arg: overflow arg to be passed to format
+ * @smpl_pmds: list of PMDs of interest for the overflowed register
+ *
+ * build an array of 46-bit PMD values based on smpl_pmds. Values are
+ * stored in increasing order of the PMD indexes
+ */
+static void pfm_intr_get_smpl_pmds_values(struct pfm_context *ctx,
+                                         struct pfm_event_set *set,
+                                         struct pfm_ovfl_arg *arg,
+                                         u64 *smpl_pmds)
+{
+       u16 j, k, max_pmd;
+       u64 new_val, ovfl_mask;
+       u64 *cnt_pmds;
+
+       cnt_pmds = ctx->regs.cnt_pmds;
+       max_pmd = ctx->regs.max_pmd;
+       ovfl_mask = pfm_pmu_conf->ovfl_mask;
+
+       for (j = k = 0; j < max_pmd; j++) {
+
+               if (!test_bit(j, cast_ulp(smpl_pmds)))
+                       continue;
+
+               new_val = pfm_read_pmd(ctx, j);
+
+               /* for counters, build 64-bit value */
+               if (test_bit(j, cast_ulp(cnt_pmds)))
+                       new_val = (set->pmds[j].value & ~ovfl_mask)
+                               | (new_val & ovfl_mask);
+
+               arg->smpl_pmds_values[k++] = new_val;
+
+               PFM_DBG_ovfl("s_pmd_val[%u]=pmd%u=0x%llx", k, j,
+                            (unsigned long long)new_val);
+       }
+       arg->num_smpl_pmds = k;
+}
+
+/**
+ * pfm_intr_process_smpl_fmt -- handle sampling format callback
+ * @ctx: context to work on
+ * @set: current event set
+ * @ip: interrupted instruction pointer
+ * @now: timestamp
+ * @num_ovfls: number of 64-bit overflows
+ * @ovfl_ctrl: set of controls for interrupt handler tail processing
+ * @regs: register state
+ *
+ * Prepare argument (ovfl_arg) to be passed to sampling format callback, then
+ * invoke the callback (fmt_handler)
+ */
+static int pfm_intr_process_smpl_fmt(struct pfm_context *ctx,
+                                   struct pfm_event_set *set,
+                                   unsigned long ip,
+                                   u64 now,
+                                   u64 num_ovfls,
+                                   u32 *ovfl_ctrl,
+                                   struct pt_regs *regs)
+{
+       struct pfm_ovfl_arg *ovfl_arg;
+       u64 start_cycles, end_cycles;
+       u16 i, max_pmd;
+       int ret = 0;
+
+       ovfl_arg = &ctx->ovfl_arg;
+
+       ovfl_arg->active_set = set->id;
+       max_pmd = ctx->regs.max_pmd;
+
+       /*
+        * first_intr_pmd: first PMD which can generate PMU interrupts
+        */
+       for (i = ctx->regs.first_intr_pmd; num_ovfls; i++) {
+               /*
+                * skip pmd which did not have 64-bit overflows
+                */
+               if (!test_bit(i, cast_ulp(set->ovfl_pmds)))
+                       continue;
+
+               num_ovfls--;
+
+               /*
+                * prepare argument to fmt_handler
+                */
+               ovfl_arg->ovfl_pmd = i;
+               ovfl_arg->ovfl_ctrl = 0;
+
+               ovfl_arg->pmd_last_reset = set->pmds[i].lval;
+               ovfl_arg->pmd_eventid = set->pmds[i].eventid;
+               ovfl_arg->num_smpl_pmds = 0;
+
+               /*
+                * copy values of pmds of interest, if any
+                * Sampling format may use them
+                * We do not initialize the unused smpl_pmds_values
+                */
+               if (!bitmap_empty(cast_ulp(set->pmds[i].smpl_pmds), max_pmd))
+                       pfm_intr_get_smpl_pmds_values(ctx, set, ovfl_arg,
+                                                     set->pmds[i].smpl_pmds);
+
+               pfm_stats_inc(fmt_handler_calls);
+
+               /*
+                * call format record (handler) routine
+                */
+               start_cycles = sched_clock();
+               ret = (*ctx->smpl_fmt->fmt_handler)(ctx, ip, now, regs);
+               end_cycles = sched_clock();
+
+               /*
+                * The reset_pmds mask is constructed automatically
+                * on overflow. When the actual reset takes place
+                * depends on the masking, switch and notification
+                * status. It may be deferred until pfm_restart().
+                */
+               *ovfl_ctrl |= ovfl_arg->ovfl_ctrl;
+
+               pfm_stats_add(fmt_handler_ns, end_cycles - start_cycles);
+       }
+       /*
+        * when the format cannot handle the rest of the overflow, we abort
+        */
+       if (ret)
+               PFM_DBG_ovfl("handler aborted at PMD%u ret=%d", i, ret);
+       return ret;
+}
+/**
+ * pfm_overflow_handler - main overflow processing routine.
+ * @ctx: context to work on (always current context)
+ * @set: current event set
+ * @ip: interrupt instruction pointer
+ * @regs: machine state
+ *
+ * set->num_ovfl_pmds is 0 when returning from this function even though
+ * set->ovfl_pmds[] may have bits set. When leaving set->num_ovfl_pmds
+ * must never be used to determine if there was a pending overflow.
+ */
+static void pfm_overflow_handler(struct pfm_context *ctx,
+                                struct pfm_event_set *set,
+                                unsigned long ip,
+                                struct pt_regs *regs)
+{
+       struct pfm_event_set *set_orig;
+       u64 now;
+       u32 ovfl_ctrl;
+       u16 max_intr, max_pmd;
+       u16 num_ovfls;
+       int ret, has_notify;
+
+       /*
+        * take timestamp
+        */
+       now = sched_clock();
+
+       max_pmd = ctx->regs.max_pmd;
+       max_intr = ctx->regs.max_intr_pmd;
+
+       set_orig = set;
+       ovfl_ctrl = 0;
+
+       /*
+        * skip ZOMBIE case
+        */
+       if (unlikely(ctx->state == PFM_CTX_ZOMBIE))
+               goto stop_monitoring;
+
+       PFM_DBG_ovfl("intr_pmds=0x%llx npend=%u ip=%p, blocking=%d "
+                    "u_pmds=0x%llx use_fmt=%u",
+                    (unsigned long long)set->povfl_pmds[0],
+                    set->npend_ovfls,
+                    (void *)ip,
+                    ctx->flags.block,
+                    (unsigned long long)set->used_pmds[0],
+                    !!ctx->smpl_fmt);
+
+       /*
+        * return number of 64-bit overflows
+        */
+       num_ovfls = pfm_intr_process_64bit_ovfls(ctx, set, &ovfl_ctrl);
+
+       /*
+        * there were no 64-bit overflows
+        * nothing else to do
+        */
+       if (!num_ovfls)
+               return;
+
+       /*
+        * tmp_ovfl_notify = ovfl_pmds & ovfl_notify
+        * with:
+        *   - ovfl_pmds: last 64-bit overflowed pmds
+        *   - ovfl_notify: notify on overflow registers
+        */
+       bitmap_and(cast_ulp(ctx->tmp_ovfl_notify),
+                  cast_ulp(set->ovfl_pmds),
+                  cast_ulp(set->ovfl_notify),
+                  max_intr);
+
+       has_notify = !bitmap_empty(cast_ulp(ctx->tmp_ovfl_notify), max_intr);
+
+       /*
+        * check for sampling format and invoke fmt_handler
+        */
+       if (likely(ctx->smpl_fmt)) {
+               pfm_intr_process_smpl_fmt(ctx, set, ip, now, num_ovfls,
+                                         &ovfl_ctrl, regs);
+       } else {
+               /*
+                * When no sampling format is used, the default
+                * is:
+                *      - mask monitoring if not switching
+                *      - notify user if requested
+                *
+                * If notification is not requested, monitoring is masked
+                * and overflowed registers are not reset (saturation).
+                * This mimics the behavior of the default sampling format.
+                */
+               ovfl_ctrl |= PFM_OVFL_CTRL_NOTIFY;
+               if (has_notify || !(ovfl_ctrl & PFM_OVFL_CTRL_SWITCH))
+                       ovfl_ctrl |= PFM_OVFL_CTRL_MASK;
+       }
+
+       PFM_DBG_ovfl("set%u o_notify=0x%llx o_pmds=0x%llx "
+                    "r_pmds=0x%llx ovfl_ctrl=0x%x",
+                    set->id,
+                    (unsigned long long)ctx->tmp_ovfl_notify[0],
+                    (unsigned long long)set->ovfl_pmds[0],
+                    (unsigned long long)set->reset_pmds[0],
+                    ovfl_ctrl);
+
+       /*
+        * execute the various controls
+        *        ORDER MATTERS
+        */
+
+
+       /*
+        * mask monitoring
+        */
+       if (ovfl_ctrl & PFM_OVFL_CTRL_MASK) {
+               pfm_mask_monitoring(ctx, set);
+               /*
+                * when masking, reset is deferred until
+                * pfm_restart()
+                */
+               ovfl_ctrl &= ~PFM_OVFL_CTRL_RESET;
+
+               /*
+                * when masking, switching is deferred until
+                * pfm_restart and we need to remember it
+                */
+               if (ovfl_ctrl & PFM_OVFL_CTRL_SWITCH) {
+                       set->priv_flags |= PFM_SETFL_PRIV_SWITCH;
+                       ovfl_ctrl &= ~PFM_OVFL_CTRL_SWITCH;
+               }
+       }
+
+       /*
+        * switch event set
+        */
+       if (ovfl_ctrl & PFM_OVFL_CTRL_SWITCH) {
+               pfm_switch_sets_from_intr(ctx);
+               /* update view of active set */
+               set = ctx->active_set;
+       }
+       /*
+        * send overflow notification
+        *
+        * only necessary if at least one overflowed
+        * register had the notify flag set
+        */
+       if (has_notify && (ovfl_ctrl & PFM_OVFL_CTRL_NOTIFY)) {
+               /*
+                * block on notify, not on masking
+                */
+               if (ctx->flags.block)
+                       pfm_post_work(current, ctx, PFM_WORK_BLOCK);
+
+               /*
+                * send notification and passed original set id
+                * if error, queue full, for instance, then default
+                * to masking monitoring, i.e., saturate
+                */
+               ret = pfm_ovfl_notify(ctx, set_orig, ip);
+               if (unlikely(ret)) {
+                       if (ctx->state == PFM_CTX_LOADED) {
+                               pfm_mask_monitoring(ctx, set);
+                               ovfl_ctrl &= ~PFM_OVFL_CTRL_RESET;
+                       }
+               } else {
+                       ctx->flags.can_restart++;
+                       PFM_DBG_ovfl("can_restart=%u", ctx->flags.can_restart);
+               }
+       }
+
+       /*
+        * reset overflowed registers
+        */
+       if (ovfl_ctrl & PFM_OVFL_CTRL_RESET) {
+               u16 nn;
+               nn = bitmap_weight(cast_ulp(set->reset_pmds), max_pmd);
+               if (nn)
+                       pfm_reset_pmds(ctx, set, nn, PFM_PMD_RESET_SHORT);
+       }
+       return;
+
+stop_monitoring:
+       /*
+        * Does not happen for a system-wide context nor for a
+        * self-monitored context. We cannot attach to kernel-only
+        * thread, thus it is safe to set TIF bits, i.e., the thread
+        * will eventually leave the kernel or die and either we will
+        * catch the context and clean it up in pfm_handler_work() or
+        * pfm_exit_thread().
+        *
+        * Mask until we get to pfm_handle_work()
+        */
+       pfm_mask_monitoring(ctx, set);
+
+       PFM_DBG_ovfl("ctx is zombie, converted to spurious");
+       pfm_post_work(current, ctx, PFM_WORK_ZOMBIE);
+}
+
+/**
+ * __pfm_interrupt_handler - 1st level interrupt handler
+ * @ip: interrupted instruction pointer
+ * @regs: machine state
+ *
+ * Function is static because we use a wrapper to easily capture timing infos.
+ *
+ *
+ * Context locking necessary to avoid concurrent accesses from other CPUs
+ *     - For per-thread, we must prevent pfm_restart() which works when
+ *       context is LOADED or MASKED
+ */
+static void __pfm_interrupt_handler(unsigned long ip, struct pt_regs *regs)
+{
+       struct task_struct *task;
+       struct pfm_context *ctx;
+       struct pfm_event_set *set;
+
+
+       task = __get_cpu_var(pmu_owner);
+       ctx = __get_cpu_var(pmu_ctx);
+
+       /*
+        * verify if there is a context on this CPU
+        */
+       if (unlikely(ctx == NULL)) {
+               PFM_DBG_ovfl("no ctx");
+               goto spurious;
+       }
+
+       /*
+        * we need to lock context because it could be accessed
+        * from another CPU. Depending on the priority level of
+        * the PMU interrupt or the arch, it may be necessary to
+        * mask interrupts alltogether to avoid race condition with
+        * the timer interrupt in case of time-based set switching,
+        * for instance.
+        */
+       spin_lock(&ctx->lock);
+
+       set = ctx->active_set;
+
+       /*
+        * For SMP per-thread, it is not possible to have
+        * owner != NULL && task != current.
+        *
+        * For UP per-thread, because of lazy save, it
+        * is possible to receive an interrupt in another task
+        * which is not using the PMU. This means
+        * that the interrupt was in-flight at the
+        * time of pfm_ctxswout_thread(). In that
+        * case, it will be replayed when the task
+        * is scheduled again. Hence we convert to spurious.
+        *
+        * The basic rule is that an overflow is always
+        * processed in the context of the task that
+        * generated it for all per-thread contexts.
+        *
+        * for system-wide, task is always NULL
+        */
+#ifndef CONFIG_SMP
+       if (unlikely((task && current->pfm_context != ctx))) {
+               PFM_DBG_ovfl("spurious: not owned by current task");
+               goto spurious;
+       }
+#endif
+       if (unlikely(ctx->state == PFM_CTX_MASKED)) {
+               PFM_DBG_ovfl("spurious: monitoring masked");
+               goto spurious;
+       }
+
+       /*
+        * check that monitoring is active, otherwise convert
+        * to spurious
+        */
+       if (unlikely(!pfm_arch_is_active(ctx))) {
+               PFM_DBG_ovfl("spurious: monitoring non active");
+               goto spurious;
+       }
+
+       /*
+        * freeze PMU and collect overflowed PMD registers
+        * into set->povfl_pmds. Number of overflowed PMDs
+        * reported in set->npend_ovfls
+        */
+       pfm_arch_intr_freeze_pmu(ctx, set);
+
+       /*
+        * no overflow detected, interrupt may have come
+        * from the previous thread running on this CPU
+        */
+       if (unlikely(!set->npend_ovfls)) {
+               PFM_DBG_ovfl("no npend_ovfls");
+               goto spurious;
+       }
+
+       pfm_stats_inc(ovfl_intr_regular_count);
+
+       /*
+        * invoke actual handler
+        */
+       pfm_overflow_handler(ctx, set, ip, regs);
+
+       /*
+        * unfreeze PMU, monitoring may not actual be restarted
+        * if context is MASKED
+        */
+       pfm_arch_intr_unfreeze_pmu(ctx);
+
+       spin_unlock(&ctx->lock);
+
+       return;
+
+spurious:
+       /* ctx may be NULL */
+       pfm_arch_intr_unfreeze_pmu(ctx);
+       if (ctx)
+               spin_unlock(&ctx->lock);
+
+       pfm_stats_inc(ovfl_intr_spurious_count);
+}
+
+
+/**
+ * pfm_interrupt_handler - 1st level interrupt handler
+ * @ip: interrupt instruction pointer
+ * @regs: machine state
+ *
+ * Function called from the low-level assembly code or arch-specific perfmon
+ * code. Simple wrapper used for timing purpose. Actual work done in
+ * __pfm_overflow_handler()
+ */
+void pfm_interrupt_handler(unsigned long ip, struct pt_regs *regs)
+{
+       u64 start;
+
+       pfm_stats_inc(ovfl_intr_all_count);
+
+       BUG_ON(!irqs_disabled());
+
+       start = sched_clock();
+
+       __pfm_interrupt_handler(ip, regs);
+
+       pfm_stats_add(ovfl_intr_ns, sched_clock() - start);
+}
+EXPORT_SYMBOL(pfm_interrupt_handler);
+
diff --git a/perfmon/perfmon_msg.c b/perfmon/perfmon_msg.c
new file mode 100644
index 0000000..b8a1e4c
--- /dev/null
+++ b/perfmon/perfmon_msg.c
@@ -0,0 +1,229 @@
+/*
+ * perfmon_msg.c: perfmon2 notification message queue management
+ *
+ * This file implements the perfmon2 interface which
+ * provides access to the hardware performance counters
+ * of the host processor.
+ *
+ * The initial version of perfmon.c was written by
+ * Ganesh Venkitachalam, IBM Corp.
+ *
+ * Then it was modified for perfmon-1.x by Stephane Eranian and
+ * David Mosberger, Hewlett Packard Co.
+ *
+ * Version Perfmon-2.x is a complete rewrite of perfmon-1.x
+ * by Stephane Eranian, Hewlett Packard Co.
+ *
+ * Copyright (c) 1999-2006 Hewlett-Packard Development Company, L.P.
+ * Contributed by Stephane Eranian <eranian@hpl.hp.com>
+ *                David Mosberger-Tang <davidm@hpl.hp.com>
+ *
+ * More information about perfmon available at:
+ *     http://perfmon2.sf.net
+ *
+ * This program is free software; you can redistribute it and/or
+ * modify it under the terms of version 2 of the GNU General Public
+ * License as published by the Free Software Foundation.
+ *
+ * This program is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+ * General Public License for more details.
+ *
+ * You should have received a copy of the GNU General Public License
+ * along with this program; if not, write to the Free Software
+ * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA
+ * 02111-1307 USA
+ */
+#include <linux/kernel.h>
+#include <linux/poll.h>
+#include <linux/perfmon_kern.h>
+
+/**
+ * pfm_get_new_msg - get a new message slot from the queue
+ * @ctx: context to operate on
+ *
+ * if queue if full NULL is returned
+ */
+static union pfarg_msg *pfm_get_new_msg(struct pfm_context *ctx)
+{
+       int next;
+
+       next = ctx->msgq_head & PFM_MSGQ_MASK;
+
+       if ((ctx->msgq_head - ctx->msgq_tail) == PFM_MSGS_COUNT)
+               return NULL;
+
+       /*
+        * move to next possible slot
+        */
+       ctx->msgq_head++;
+
+       PFM_DBG_ovfl("head=%d tail=%d msg=%d",
+               ctx->msgq_head & PFM_MSGQ_MASK,
+               ctx->msgq_tail & PFM_MSGQ_MASK,
+               next);
+
+       return ctx->msgq+next;
+}
+
+/**
+ * pfm_notify_user - wakeup any thread wiating on msg queue,  post SIGIO
+ * @ctx: context to operate on
+ *
+ * message is already enqueued
+ */
+static void pfm_notify_user(struct pfm_context *ctx)
+{
+       if (ctx->state == PFM_CTX_ZOMBIE) {
+               PFM_DBG("no notification, context is zombie");
+               return;
+       }
+
+       PFM_DBG_ovfl("waking up");
+
+       wake_up_interruptible(&ctx->msgq_wait);
+
+       /*
+        * it is safe to call kill_fasync() from an interrupt
+        * handler. kill_fasync()  grabs two RW locks (fasync_lock,
+        * tasklist_lock) in read mode. There is conflict only in
+        * case the PMU interrupt occurs during a write mode critical
+        * section. This cannot happen because for both locks, the
+        * write mode is always using interrupt masking (write_lock_irq).
+        */
+       kill_fasync(&ctx->async_queue, SIGIO, POLL_IN);
+}
+
+/**
+ * pfm_ovfl_notify - send overflow notification
+ * @ctx: context to operate on
+ * @set: which set the overflow comes from
+ * @ip: overflow interrupt instruction address (IIP)
+ *
+ * Appends an overflow notification message to context queue.
+ * call pfm_notify() to wakeup any threads and/or send a signal
+ *
+ * Context is locked and interrupts are disabled (no preemption).
+ */
+int pfm_ovfl_notify(struct pfm_context *ctx,
+                       struct pfm_event_set *set,
+                       unsigned long ip)
+{
+       union pfarg_msg *msg = NULL;
+       u64 *ovfl_pmds;
+
+       if (!ctx->flags.no_msg) {
+               msg = pfm_get_new_msg(ctx);
+               if (msg == NULL) {
+                       /*
+                        * when message queue fills up it is because the user
+                        * did not extract the message, yet issued
+                        * pfm_restart(). At this point, we stop sending
+                        * notification, thus the user will not be able to get
+                        * new samples when using the default format.
+                        */
+                       PFM_DBG_ovfl("no more notification msgs");
+                       return -1;
+               }
+
+               msg->pfm_ovfl_msg.msg_type = PFM_MSG_OVFL;
+               msg->pfm_ovfl_msg.msg_ovfl_pid = current->pid;
+               msg->pfm_ovfl_msg.msg_active_set = set->id;
+
+               ovfl_pmds = msg->pfm_ovfl_msg.msg_ovfl_pmds;
+
+               /*
+                * copy bitmask of all pmd that interrupted last
+                */
+               bitmap_copy(cast_ulp(ovfl_pmds), cast_ulp(set->ovfl_pmds),
+                           ctx->regs.max_intr_pmd);
+
+               msg->pfm_ovfl_msg.msg_ovfl_cpu = smp_processor_id();
+               msg->pfm_ovfl_msg.msg_ovfl_tid = current->tgid;
+               msg->pfm_ovfl_msg.msg_ovfl_ip = ip;
+
+               pfm_stats_inc(ovfl_notify_count);
+       }
+
+       PFM_DBG_ovfl("ip=0x%lx o_pmds=0x%llx",
+                    ip,
+                    (unsigned long long)set->ovfl_pmds[0]);
+
+       pfm_notify_user(ctx);
+       return 0;
+}
+
+/**
+ * pfm_end_notify_user - notify of thread termination
+ * @ctx: context to operate on
+ *
+ * In per-thread mode, when not self-monitoring, perfmon
+ * sends a 'end' notification message when the monitored
+ * thread where the context is attached is exiting.
+ *
+ * This helper message alleviates the need to track the activity
+ * of the thread/process when it is not directly related, i.e.,
+ * was attached. In other words, no needto keep the thread
+ * ptraced.
+ *
+ * The context must be locked and interrupts disabled.
+ */
+int pfm_end_notify(struct pfm_context *ctx)
+{
+       union pfarg_msg *msg;
+
+       msg = pfm_get_new_msg(ctx);
+       if (msg == NULL) {
+               PFM_ERR("%s no more msgs", __func__);
+               return -1;
+       }
+       /* no leak */
+       memset(msg, 0, sizeof(*msg));
+
+       msg->type = PFM_MSG_END;
+
+       PFM_DBG("end msg: msg=%p no_msg=%d",
+               msg,
+               ctx->flags.no_msg);
+
+       pfm_notify_user(ctx);
+       return 0;
+}
+
+/**
+ * pfm_get_next_msg - copy the oldest message from the queue and move tail
+ * @ctx: context to use
+ * @m: where to copy the message into
+ *
+ * The tail of the queue is moved as a consequence of this call
+ */
+void pfm_get_next_msg(struct pfm_context *ctx, union pfarg_msg *m)
+{
+       union pfarg_msg *next;
+
+       PFM_DBG_ovfl("in head=%d tail=%d",
+               ctx->msgq_head & PFM_MSGQ_MASK,
+               ctx->msgq_tail & PFM_MSGQ_MASK);
+
+       /*
+        * get oldest message
+        */
+       next = ctx->msgq + (ctx->msgq_tail & PFM_MSGQ_MASK);
+
+       /*
+        * move tail forward
+        */
+       ctx->msgq_tail++;
+
+       /*
+        * copy message, we cannot simply point to it
+        * as it may be re-used before we copy it out
+        */
+       *m = *next;
+
+       PFM_DBG_ovfl("out head=%d tail=%d type=%d",
+               ctx->msgq_head & PFM_MSGQ_MASK,
+               ctx->msgq_tail & PFM_MSGQ_MASK,
+               m->type);
+}
diff --git a/perfmon/perfmon_pmu.c b/perfmon/perfmon_pmu.c
new file mode 100644
index 0000000..df7a9c9
--- /dev/null
+++ b/perfmon/perfmon_pmu.c
@@ -0,0 +1,590 @@
+/*
+ * perfmon_pmu.c: perfmon2 PMU configuration management
+ *
+ * This file implements the perfmon2 interface which
+ * provides access to the hardware performance counters
+ * of the host processor.
+ *
+ * The initial version of perfmon.c was written by
+ * Ganesh Venkitachalam, IBM Corp.
+ *
+ * Then it was modified for perfmon-1.x by Stephane Eranian and
+ * David Mosberger, Hewlett Packard Co.
+ *
+ * Version Perfmon-2.x is a complete rewrite of perfmon-1.x
+ * by Stephane Eranian, Hewlett Packard Co.
+ *
+ * Copyright (c) 1999-2006 Hewlett-Packard Development Company, L.P.
+ * Contributed by Stephane Eranian <eranian@hpl.hp.com>
+ *                David Mosberger-Tang <davidm@hpl.hp.com>
+ *
+ * More information about perfmon available at:
+ *     http://perfmon2.sf.net
+ *
+ * This program is free software; you can redistribute it and/or
+ * modify it under the terms of version 2 of the GNU General Public
+ * License as published by the Free Software Foundation.
+ *
+ * This program is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+ * General Public License for more details.
+ *
+ * You should have received a copy of the GNU General Public License
+ * along with this program; if not, write to the Free Software
+ * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA
+ * 02111-1307 USA
+ */
+#include <linux/module.h>
+#include <linux/perfmon_kern.h>
+#include "perfmon_priv.h"
+
+#ifndef CONFIG_MODULE_UNLOAD
+#define module_refcount(n)     1
+#endif
+
+static __cacheline_aligned_in_smp int request_mod_in_progress;
+static __cacheline_aligned_in_smp DEFINE_SPINLOCK(pfm_pmu_conf_lock);
+
+static __cacheline_aligned_in_smp DEFINE_SPINLOCK(pfm_pmu_acq_lock);
+static u32 pfm_pmu_acquired;
+
+/*
+ * perfmon core must acces PMU information ONLY through pfm_pmu_conf
+ * if pfm_pmu_conf is NULL, then no description is registered
+ */
+struct pfm_pmu_config  *pfm_pmu_conf;
+EXPORT_SYMBOL(pfm_pmu_conf);
+
+static inline int pmu_is_module(struct pfm_pmu_config *c)
+{
+       return !(c->flags & PFM_PMUFL_IS_BUILTIN);
+}
+/**
+ * pfm_pmu_regdesc_init -- initialize regdesc structure from PMU table
+ * @regs: the regdesc structure to initialize
+ * @excl_type: the register type(s) to exclude from this regdesc
+ * @unvail_pmcs: unavailable PMC registers
+ * @unavail_pmds: unavailable PMD registers
+ *
+ * Return:
+ *     0 success
+ *     errno in case of error
+ */
+static int pfm_pmu_regdesc_init(struct pfm_regdesc *regs, int excl_type,
+                               u64 *unavail_pmcs, u64 *unavail_pmds)
+{
+       struct pfm_regmap_desc *d;
+       u16 n, n2, n_counters, i;
+       int first_intr_pmd = -1, max1, max2, max3;
+
+       /*
+        * compute the number of implemented PMC from the
+        * description table
+        */
+       n = 0;
+       max1 = max2 = -1;
+       d = pfm_pmu_conf->pmc_desc;
+       for (i = 0; i < pfm_pmu_conf->num_pmc_entries;  i++, d++) {
+               if (!(d->type & PFM_REG_I))
+                       continue;
+
+               if (test_bit(i, cast_ulp(unavail_pmcs)))
+                       continue;
+
+               if (d->type & excl_type)
+                       continue;
+
+               __set_bit(i, cast_ulp(regs->pmcs));
+
+               max1 = i;
+               n++;
+       }
+
+       if (!n) {
+               PFM_INFO("%s PMU description has no PMC registers",
+                        pfm_pmu_conf->pmu_name);
+               return -EINVAL;
+       }
+
+       regs->max_pmc = max1 + 1;
+       regs->num_pmcs = n;
+
+       n = n_counters = n2 = 0;
+       max1 = max2 = max3 = -1;
+       d = pfm_pmu_conf->pmd_desc;
+       for (i = 0; i < pfm_pmu_conf->num_pmd_entries;  i++, d++) {
+               if (!(d->type & PFM_REG_I))
+                       continue;
+
+               if (test_bit(i, cast_ulp(unavail_pmds)))
+                       continue;
+
+               if (d->type & excl_type)
+                       continue;
+
+               __set_bit(i, cast_ulp(regs->pmds));
+               max1 = i;
+               n++;
+
+               /*
+                * read-write registers
+                */
+               if (!(d->type & PFM_REG_RO)) {
+                       __set_bit(i, cast_ulp(regs->rw_pmds));
+                       max3 = i;
+                       n2++;
+               }
+
+               /*
+                * counter registers
+                */
+               if (d->type & PFM_REG_C64) {
+                       __set_bit(i, cast_ulp(regs->cnt_pmds));
+                       n_counters++;
+               }
+
+               /*
+                * PMD with intr capabilities
+                */
+               if (d->type & PFM_REG_INTR) {
+                       __set_bit(i, cast_ulp(regs->intr_pmds));
+                       if (first_intr_pmd == -1)
+                               first_intr_pmd = i;
+                       max2 = i;
+               }
+       }
+
+       if (!n) {
+               PFM_INFO("%s PMU description has no PMD registers",
+                        pfm_pmu_conf->pmu_name);
+               return -EINVAL;
+       }
+
+       regs->max_pmd = max1 + 1;
+       regs->first_intr_pmd = first_intr_pmd;
+       regs->max_intr_pmd  = max2 + 1;
+
+       regs->num_counters = n_counters;
+       regs->num_pmds = n;
+       regs->max_rw_pmd = max3 + 1;
+       regs->num_rw_pmd = n2;
+
+       return 0;
+}
+
+/**
+ * pfm_pmu_regdesc_init_all -- initialize all regdesc structures
+ * @una_pmcs : unavailable PMC registers
+ * @una_pmds : unavailable PMD registers
+ *
+ * Return:
+ *     0 sucess
+ *     errno if error
+ *
+ * We maintain 3 regdesc:
+ *     regs_all: all available registers
+ *     regs_sys: registers available to system-wide contexts only
+ *     regs_thr: registers available to per-thread contexts only
+ */
+static int pfm_pmu_regdesc_init_all(u64 *una_pmcs, u64 *una_pmds)
+{
+       int ret;
+
+       memset(&pfm_pmu_conf->regs_all, 0, sizeof(struct pfm_regdesc));
+       memset(&pfm_pmu_conf->regs_thr, 0, sizeof(struct pfm_regdesc));
+       memset(&pfm_pmu_conf->regs_sys, 0, sizeof(struct pfm_regdesc));
+
+       ret = pfm_pmu_regdesc_init(&pfm_pmu_conf->regs_all,
+                                  0,
+                                  una_pmcs, una_pmds);
+       if (ret)
+               return ret;
+
+       PFM_DBG("regs_all.pmcs=0x%llx",
+               (unsigned long long)pfm_pmu_conf->regs_all.pmcs[0]);
+
+       ret = pfm_pmu_regdesc_init(&pfm_pmu_conf->regs_thr,
+                                  PFM_REG_SYS,
+                                  una_pmcs, una_pmds);
+       if (ret)
+               return ret;
+       PFM_DBG("regs.thr.pmcs=0x%llx",
+               (unsigned long long)pfm_pmu_conf->regs_thr.pmcs[0]);
+
+       ret = pfm_pmu_regdesc_init(&pfm_pmu_conf->regs_sys,
+                                   PFM_REG_THR,
+                                   una_pmcs, una_pmds);
+
+       PFM_DBG("regs_sys.pmcs=0x%llx",
+               (unsigned long long)pfm_pmu_conf->regs_sys.pmcs[0]);
+
+       return ret;
+}
+
+int pfm_pmu_register(struct pfm_pmu_config *cfg)
+{
+       u16 i, nspec, nspec_ro, num_pmcs, num_pmds, num_wc = 0;
+       int type, ret = -EBUSY;
+
+       if (perfmon_disabled) {
+               PFM_INFO("perfmon disabled, cannot add PMU description");
+               return -ENOSYS;
+       }
+
+       nspec = nspec_ro = num_pmds = num_pmcs = 0;
+
+       /* some sanity checks */
+       if (cfg == NULL || cfg->pmu_name == NULL) {
+               PFM_INFO("PMU config descriptor is invalid");
+               return -EINVAL;
+       }
+
+       /* must have a probe */
+       if (cfg->probe_pmu == NULL) {
+               PFM_INFO("PMU config has no probe routine");
+               return -EINVAL;
+       }
+
+       /*
+        * execute probe routine before anything else as it
+        * may update configuration tables
+        */
+       if ((*cfg->probe_pmu)() == -1) {
+               PFM_INFO("%s PMU detection failed", cfg->pmu_name);
+               return -EINVAL;
+       }
+
+       if (!(cfg->flags & PFM_PMUFL_IS_BUILTIN) && cfg->owner == NULL) {
+               PFM_INFO("PMU config %s is missing owner", cfg->pmu_name);
+               return -EINVAL;
+       }
+
+       if (!cfg->num_pmd_entries) {
+               PFM_INFO("%s needs to define num_pmd_entries", cfg->pmu_name);
+               return -EINVAL;
+       }
+
+       if (!cfg->num_pmc_entries) {
+               PFM_INFO("%s needs to define num_pmc_entries", cfg->pmu_name);
+               return -EINVAL;
+       }
+
+       if (!cfg->counter_width) {
+               PFM_INFO("PMU config %s, zero width counters", cfg->pmu_name);
+               return -EINVAL;
+       }
+
+       /*
+        * REG_RO, REG_V not supported on PMC registers
+        */
+       for (i = 0; i < cfg->num_pmc_entries;  i++) {
+
+               type = cfg->pmc_desc[i].type;
+
+               if (type & PFM_REG_I)
+                       num_pmcs++;
+
+               if (type & PFM_REG_WC)
+                       num_wc++;
+
+               if (type & PFM_REG_V) {
+                       PFM_INFO("PFM_REG_V is not supported on "
+                                "PMCs (PMC%d)", i);
+                       return -EINVAL;
+               }
+               if (type & PFM_REG_RO) {
+                       PFM_INFO("PFM_REG_RO meaningless on "
+                                "PMCs (PMC%u)", i);
+                       return -EINVAL;
+               }
+       }
+
+       if (num_wc && cfg->pmc_write_check == NULL) {
+               PFM_INFO("some PMCs have write-checker but no callback provided\n");
+               return -EINVAL;
+       }
+
+       /*
+        * check virtual PMD registers
+        */
+       num_wc = 0;
+       for (i = 0; i < cfg->num_pmd_entries;  i++) {
+
+               type = cfg->pmd_desc[i].type;
+
+               if (type & PFM_REG_I)
+                       num_pmds++;
+
+               if (type & PFM_REG_V) {
+                       nspec++;
+                       if (type & PFM_REG_RO)
+                               nspec_ro++;
+               }
+
+               if (type & PFM_REG_WC)
+                       num_wc++;
+       }
+
+       if (num_wc && cfg->pmd_write_check == NULL) {
+               PFM_INFO("PMD have write-checker but no callback provided\n");
+               return -EINVAL;
+       }
+
+       if (nspec && cfg->pmd_sread == NULL) {
+               PFM_INFO("PMU config is missing pmd_sread()");
+               return -EINVAL;
+       }
+
+       nspec = nspec - nspec_ro;
+       if (nspec && cfg->pmd_swrite == NULL) {
+               PFM_INFO("PMU config is missing pmd_swrite()");
+               return -EINVAL;
+       }
+
+       if (num_pmcs >= PFM_MAX_PMCS) {
+               PFM_INFO("%s PMCS registers exceed name space [0-%u]",
+                        cfg->pmu_name,
+                        PFM_MAX_PMCS);
+               return -EINVAL;
+       }
+       if (num_pmds >= PFM_MAX_PMDS) {
+               PFM_INFO("%s PMDS registers exceed name space [0-%u]",
+                        cfg->pmu_name,
+                        PFM_MAX_PMDS);
+               return -EINVAL;
+       }
+       spin_lock(&pfm_pmu_conf_lock);
+
+       if (pfm_pmu_conf)
+               goto unlock;
+
+       if (!cfg->version)
+               cfg->version = "0.0";
+
+       pfm_pmu_conf = cfg;
+       pfm_pmu_conf->ovfl_mask = (1ULL << cfg->counter_width) - 1;
+
+       ret = pfm_arch_pmu_config_init(cfg);
+       if (ret)
+               goto unlock;
+
+       ret = pfm_sysfs_add_pmu(pfm_pmu_conf);
+       if (ret)
+               pfm_pmu_conf = NULL;
+
+unlock:
+       spin_unlock(&pfm_pmu_conf_lock);
+
+       if (ret) {
+               PFM_INFO("register %s PMU error %d", cfg->pmu_name, ret);
+       } else {
+               PFM_INFO("%s PMU installed", cfg->pmu_name);
+               /*
+                * (re)initialize PMU on each PMU now that we have a description
+                */
+               on_each_cpu(__pfm_init_percpu, cfg, 0);
+       }
+       return ret;
+}
+EXPORT_SYMBOL(pfm_pmu_register);
+
+/*
+ * remove PMU description. Caller must pass address of current
+ * configuration. This is mostly for sanity checking as only
+ * one config can exist at any time.
+ *
+ * We are using the module refcount mechanism to protect against
+ * removal while the configuration is being used. As long as there is
+ * one context, a PMU configuration cannot be removed. The protection is
+ * managed in module logic.
+ */
+void pfm_pmu_unregister(struct pfm_pmu_config *cfg)
+{
+       if (!(cfg || pfm_pmu_conf))
+               return;
+
+       spin_lock(&pfm_pmu_conf_lock);
+
+       BUG_ON(module_refcount(pfm_pmu_conf->owner));
+
+       if (cfg->owner == pfm_pmu_conf->owner) {
+               pfm_sysfs_remove_pmu(pfm_pmu_conf);
+               pfm_pmu_conf = NULL;
+       }
+
+       spin_unlock(&pfm_pmu_conf_lock);
+}
+EXPORT_SYMBOL(pfm_pmu_unregister);
+
+static int pfm_pmu_request_module(void)
+{
+       char *mod_name;
+       int ret;
+
+       mod_name = pfm_arch_get_pmu_module_name();
+       if (mod_name == NULL)
+               return -ENOSYS;
+
+       ret = request_module(mod_name);
+
+       PFM_DBG("mod=%s ret=%d\n", mod_name, ret);
+       return ret;
+}
+
+/*
+ * autoload:
+ *     0     : do not try to autoload the PMU description module
+ *     not 0 : try to autoload the PMU description module
+ */
+int pfm_pmu_conf_get(int autoload)
+{
+       int ret;
+
+       spin_lock(&pfm_pmu_conf_lock);
+
+       if (request_mod_in_progress) {
+               ret = -ENOSYS;
+               goto skip;
+       }
+
+       if (autoload && pfm_pmu_conf == NULL) {
+
+               request_mod_in_progress = 1;
+
+               spin_unlock(&pfm_pmu_conf_lock);
+
+               pfm_pmu_request_module();
+
+               spin_lock(&pfm_pmu_conf_lock);
+
+               request_mod_in_progress = 0;
+
+               /*
+                * request_module() may succeed but the module
+                * may not have registered properly so we need
+                * to check
+                */
+       }
+
+       ret = pfm_pmu_conf == NULL ? -ENOSYS : 0;
+       if (!ret && pmu_is_module(pfm_pmu_conf)
+           && !try_module_get(pfm_pmu_conf->owner))
+               ret = -ENOSYS;
+
+skip:
+       spin_unlock(&pfm_pmu_conf_lock);
+
+       return ret;
+}
+
+void pfm_pmu_conf_put(void)
+{
+       if (pfm_pmu_conf == NULL || !pmu_is_module(pfm_pmu_conf))
+               return;
+
+       spin_lock(&pfm_pmu_conf_lock);
+       module_put(pfm_pmu_conf->owner);
+       spin_unlock(&pfm_pmu_conf_lock);
+}
+
+
+/*
+ * acquire PMU resource from lower-level PMU register allocator
+ * (currently perfctr-watchdog.c)
+ *
+ * acquisition is done when the first context is created (and not
+ * when it is loaded). We grab all that is defined in the description
+ * module and then we make adjustments at the arch-specific level.
+ *
+ * The PMU resource is released when the last perfmon context is
+ * destroyed.
+ *
+ * interrupts are not masked
+ */
+int pfm_pmu_acquire(struct pfm_context *ctx)
+{
+       u64 unavail_pmcs[PFM_PMC_BV];
+       u64 unavail_pmds[PFM_PMD_BV];
+       int ret = 0;
+
+       spin_lock(&pfm_pmu_acq_lock);
+
+       PFM_DBG("pmu_acquired=%u", pfm_pmu_acquired);
+
+       pfm_pmu_acquired++;
+
+       /*
+        * we need to initialize regdesc each  time we re-acquire
+        * the PMU for the first time as there may have been changes
+        * in the list of available registers, e.g., NMI may have
+        * been disabled. Checking on PMU module insert is not
+        * enough
+        */
+       if (pfm_pmu_acquired == 1) {
+               memset(unavail_pmcs, 0, sizeof(unavail_pmcs));
+               memset(unavail_pmds, 0, sizeof(unavail_pmds));
+
+               ret = pfm_arch_pmu_acquire(unavail_pmcs, unavail_pmds);
+               if (ret) {
+                       pfm_pmu_acquired--;
+               } else {
+                       pfm_pmu_regdesc_init_all(unavail_pmcs, unavail_pmds);
+
+                       /* available PMU ressources */
+                       PFM_DBG("PMU acquired: %u PMCs, %u PMDs, %u counters",
+                               pfm_pmu_conf->regs_all.num_pmcs,
+                               pfm_pmu_conf->regs_all.num_pmds,
+                               pfm_pmu_conf->regs_all.num_counters);
+               }
+       }
+       spin_unlock(&pfm_pmu_acq_lock);
+
+       /*
+        * copy the regdesc that corresponds to the context
+        * we copy and not just point because it helps with
+        * memory locality. the regdesc structure is accessed
+        * very frequently in performance critical code such
+        * as context switch and interrupt handling. By using
+        * a local copy, we increase memory footprint, but
+        * increase chance to have local memory access,
+        * especially for system-wide contexts.
+        */
+       if (ctx->flags.system)
+               ctx->regs = pfm_pmu_conf->regs_sys;
+       else
+               ctx->regs = pfm_pmu_conf->regs_thr;
+
+       return ret;
+}
+
+/*
+ * release the PMU resource
+ *
+ * actual release happens when last context is destroyed
+ *
+ * interrupts are not masked
+ */
+void pfm_pmu_release(void)
+{
+       BUG_ON(irqs_disabled());
+
+       /*
+        * we need to use a spinlock because release takes some time
+        * and we may have a race with pfm_pmu_acquire()
+        */
+       spin_lock(&pfm_pmu_acq_lock);
+
+       PFM_DBG("pmu_acquired=%d", pfm_pmu_acquired);
+
+       /*
+        * we decouple test and decrement because if we had errors
+        * in pfm_pmu_acquire(), we still come here on pfm_context_free()
+        * but with pfm_pmu_acquire=0
+        */
+       if (pfm_pmu_acquired > 0 && --pfm_pmu_acquired == 0) {
+               pfm_arch_pmu_release();
+               PFM_DBG("PMU released");
+       }
+       spin_unlock(&pfm_pmu_acq_lock);
+}
diff --git a/perfmon/perfmon_priv.h b/perfmon/perfmon_priv.h
new file mode 100644
index 0000000..5b485de
--- /dev/null
+++ b/perfmon/perfmon_priv.h
@@ -0,0 +1,182 @@
+/*
+ * Copyright (c) 2001-2006 Hewlett-Packard Development Company, L.P.
+ * Contributed by Stephane Eranian <eranian@hpl.hp.com>
+ *
+ * This program is free software; you can redistribute it and/or
+ * modify it under the terms of version 2 of the GNU General Public
+ * License as published by the Free Software Foundation.
+ *
+ * This program is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+ * General Public License for more details.
+ *
+ * You should have received a copy of the GNU General Public License
+ * along with this program; if not, write to the Free Software
+ * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA
+ * 02111-1307 USA
+ */
+
+#ifndef __PERFMON_PRIV_H__
+#define __PERFMON_PRIV_H__
+/*
+ * This file contains all the definitions of data structures, variables, macros
+ * that are to private to the generic code, i.e., not shared with any code that
+ * lives under arch/ or include/asm-XX
+ *
+ * For shared definitions, use include/linux/perfmon_kern.h
+ */
+
+#ifdef CONFIG_PERFMON
+
+/*
+ * type of PMD reset for pfm_reset_pmds() or pfm_switch_sets*()
+ */
+#define PFM_PMD_RESET_SHORT    1       /* use short reset value */
+#define PFM_PMD_RESET_LONG     2       /* use long reset value  */
+
+/*
+ * context lazy save/restore activation count
+ */
+#define PFM_INVALID_ACTIVATION ((u64)~0)
+
+DECLARE_PER_CPU(u64, pmu_activation_number);
+DECLARE_PER_CPU(struct hrtimer, pfm_hrtimer);
+
+static inline void pfm_set_pmu_owner(struct task_struct *task,
+                                    struct pfm_context *ctx)
+{
+       __get_cpu_var(pmu_owner) = task;
+       __get_cpu_var(pmu_ctx) = ctx;
+}
+
+static inline int pfm_msgq_is_empty(struct pfm_context *ctx)
+{
+       return ctx->msgq_head == ctx->msgq_tail;
+}
+
+void pfm_get_next_msg(struct pfm_context *ctx, union pfarg_msg *m);
+int pfm_end_notify(struct pfm_context *ctx);
+int pfm_ovfl_notify(struct pfm_context *ctx, struct pfm_event_set *set,
+                   unsigned long ip);
+
+int pfm_alloc_fd(struct file **cfile);
+
+int __pfm_delete_evtsets(struct pfm_context *ctx, void *arg, int count);
+int __pfm_getinfo_evtsets(struct pfm_context *ctx, struct pfarg_setinfo *req,
+                         int count);
+int __pfm_create_evtsets(struct pfm_context *ctx, struct pfarg_setdesc *req,
+                       int count);
+
+
+int pfm_init_ctx(void);
+
+int pfm_pmu_acquire(struct pfm_context *ctx);
+void pfm_pmu_release(void);
+
+int pfm_session_acquire(int is_system, u32 cpu);
+void pfm_session_release(int is_system, u32 cpu);
+
+int pfm_smpl_buf_space_acquire(struct pfm_context *ctx, size_t size);
+int pfm_smpl_buf_load_context(struct pfm_context *ctx);
+void pfm_smpl_buf_unload_context(struct pfm_context *ctx);
+
+int  pfm_init_sysfs(void);
+
+#ifdef CONFIG_PERFMON_DEBUG_FS
+int  pfm_init_debugfs(void);
+int pfm_debugfs_add_cpu(int mycpu);
+void pfm_debugfs_del_cpu(int mycpu);
+#else
+static inline int pfm_init_debugfs(void)
+{
+       return 0;
+}
+static inline int pfm_debugfs_add_cpu(int mycpu)
+{
+       return 0;
+}
+
+static inline void pfm_debugfs_del_cpu(int mycpu)
+{}
+#endif
+
+
+void pfm_reset_pmds(struct pfm_context *ctx, struct pfm_event_set *set,
+                   int num_pmds,
+                   int reset_mode);
+
+struct pfm_event_set *pfm_prepare_sets(struct pfm_context *ctx, u16 load_set);
+int pfm_init_sets(void);
+
+ssize_t pfm_sysfs_res_show(char *buf, size_t sz, int what);
+
+void pfm_free_sets(struct pfm_context *ctx);
+int pfm_create_initial_set(struct pfm_context *ctx);
+void pfm_switch_sets_from_intr(struct pfm_context *ctx);
+void pfm_restart_timer(struct pfm_context *ctx, struct pfm_event_set *set);
+enum hrtimer_restart pfm_handle_switch_timeout(struct hrtimer *t);
+
+enum hrtimer_restart pfm_switch_sets(struct pfm_context *ctx,
+                   struct pfm_event_set *new_set,
+                   int reset_mode,
+                   int no_restart);
+
+/**
+ * pfm_save_prev_ctx - check if previous context exists and save state
+ *
+ * called from pfm_load_ctx_thread() and __pfm_ctxsin_thread() to
+ * check if previous context exists. If so saved its PMU state. This is used
+ * only for UP kernels.
+ *
+ * PMU ownership is not cleared because the function is always called while
+ * trying to install a new owner.
+ */
+static inline void pfm_check_save_prev_ctx(void)
+{
+#ifdef CONFIG_SMP
+       struct pfm_event_set *set;
+       struct pfm_context *ctxp;
+
+       ctxp = __get_cpu_var(pmu_ctx);
+       if (!ctxp)
+               return;
+       /*
+        * in UP per-thread, due to lazy save
+        * there could be a context from another
+        * task. We need to push it first before
+        * installing our new state
+        */
+       set = ctxp->active_set;
+       pfm_save_pmds(ctxp, set);
+       /*
+        * do not clear ownership because we rewrite
+        * right away
+        */
+#endif
+}
+
+
+int pfm_init_fs(void);
+
+int pfm_init_hotplug(void);
+
+void pfm_mask_monitoring(struct pfm_context *ctx, struct pfm_event_set *set);
+void pfm_resume_after_ovfl(struct pfm_context *ctx);
+int pfm_setup_smpl_fmt(struct pfm_context *ctx, u32 ctx_flags, void *fmt_arg,
+                      struct file *filp);
+
+static inline void pfm_post_work(struct task_struct *task,
+                                struct pfm_context *ctx, int type)
+{
+       ctx->flags.work_type = type;
+       set_tsk_thread_flag(task, TIF_PERFMON_WORK);
+       pfm_arch_arm_handle_work(task);
+}
+
+#define PFM_PMC_STK_ARG        PFM_ARCH_PMC_STK_ARG
+#define PFM_PMD_STK_ARG        PFM_ARCH_PMD_STK_ARG
+
+#endif /* CONFIG_PERFMON */
+
+#endif /* __PERFMON_PRIV_H__ */
diff --git a/perfmon/perfmon_res.c b/perfmon/perfmon_res.c
new file mode 100644
index 0000000..7b0382b
--- /dev/null
+++ b/perfmon/perfmon_res.c
@@ -0,0 +1,450 @@
+/*
+ * perfmon_res.c:  perfmon2 resource allocations
+ *
+ * This file implements the perfmon2 interface which
+ * provides access to the hardware performance counters
+ * of the host processor.
+ *
+ * The initial version of perfmon.c was written by
+ * Ganesh Venkitachalam, IBM Corp.
+ *
+ * Then it was modified for perfmon-1.x by Stephane Eranian and
+ * David Mosberger, Hewlett Packard Co.
+ *
+ * Version Perfmon-2.x is a complete rewrite of perfmon-1.x
+ * by Stephane Eranian, Hewlett Packard Co.
+ *
+ * Copyright (c) 1999-2006 Hewlett-Packard Development Company, L.P.
+ * Contributed by Stephane Eranian <eranian@hpl.hp.com>
+ *                David Mosberger-Tang <davidm@hpl.hp.com>
+ *
+ * More information about perfmon available at:
+ *     http://perfmon2.sf.net
+ *
+ * This program is free software; you can redistribute it and/or
+ * modify it under the terms of version 2 of the GNU General Public
+ * License as published by the Free Software Foundation.
+ *
+ * This program is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+ * General Public License for more details.
+ *
+ * You should have received a copy of the GNU General Public License
+ * along with this program; if not, write to the Free Software
+ * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA
+ * 02111-1307 USA
+ */
+#include <linux/kernel.h>
+#include <linux/module.h>
+#include <linux/perfmon_kern.h>
+#include "perfmon_priv.h"
+
+/*
+ * global information about all sessions
+ * mostly used to synchronize between system wide and per-process
+ */
+struct pfm_resources {
+       size_t          smpl_buf_mem_cur;/* current smpl buf mem usage */
+       cpumask_t       sys_cpumask;     /* bitmask of used cpus */
+       u32             thread_sessions; /* #num loaded per-thread sessions */
+};
+
+static struct pfm_resources pfm_res;
+
+static __cacheline_aligned_in_smp DEFINE_SPINLOCK(pfm_res_lock);
+
+/**
+ * pfm_smpl_buf_space_acquire - check memory resource usage for sampling buffer
+ * @ctx: context of interest
+ * @size: size fo requested buffer
+ *
+ * sampling buffer allocated by perfmon must be
+ * checked against max locked memory usage thresholds
+ * for security reasons.
+ *
+ * The first level check is against the system wide limit
+ * as indicated by the system administrator in /sys/kernel/perfmon
+ *
+ * The second level check is on a per-process basis using
+ * RLIMIT_MEMLOCK limit.
+ *
+ * Operating on the current task only.
+ */
+int pfm_smpl_buf_space_acquire(struct pfm_context *ctx, size_t size)
+{
+       struct mm_struct *mm;
+       unsigned long locked;
+       unsigned long buf_mem, buf_mem_max;
+       unsigned long flags;
+
+       spin_lock_irqsave(&pfm_res_lock, flags);
+
+       /*
+        * check against global buffer limit
+        */
+       buf_mem_max = pfm_controls.smpl_buffer_mem_max;
+       buf_mem = pfm_res.smpl_buf_mem_cur + size;
+
+       if (buf_mem <= buf_mem_max) {
+               pfm_res.smpl_buf_mem_cur = buf_mem;
+
+               PFM_DBG("buf_mem_max=%lu current_buf_mem=%lu",
+                       buf_mem_max,
+                       buf_mem);
+       }
+
+       spin_unlock_irqrestore(&pfm_res_lock, flags);
+
+       if (buf_mem > buf_mem_max) {
+               PFM_DBG("smpl buffer memory threshold reached");
+               return -ENOMEM;
+       }
+
+       /*
+        * check against per-process RLIMIT_MEMLOCK
+        */
+       mm = get_task_mm(current);
+
+       down_write(&mm->mmap_sem);
+
+       locked  = mm->locked_vm << PAGE_SHIFT;
+       locked += size;
+
+       if (locked > current->signal->rlim[RLIMIT_MEMLOCK].rlim_cur) {
+
+               PFM_DBG("RLIMIT_MEMLOCK reached ask_locked=%lu rlim_cur=%lu",
+                       locked,
+                       current->signal->rlim[RLIMIT_MEMLOCK].rlim_cur);
+
+               up_write(&mm->mmap_sem);
+               mmput(mm);
+               goto unres;
+       }
+
+       mm->locked_vm = locked >> PAGE_SHIFT;
+
+       up_write(&mm->mmap_sem);
+
+       mmput(mm);
+
+       return 0;
+
+unres:
+       /*
+        * remove global buffer memory allocation
+        */
+       spin_lock_irqsave(&pfm_res_lock, flags);
+
+       pfm_res.smpl_buf_mem_cur -= size;
+
+       spin_unlock_irqrestore(&pfm_res_lock, flags);
+
+       return -ENOMEM;
+}
+/**
+ * pfm_smpl_buf_space_release - release resource usage for sampling buffer
+ * @ctx: perfmon context of interest
+ *
+ * There exist multiple paths leading to this function. We need to
+ * be very careful withlokcing on the mmap_sem as it may already be
+ * held by the time we come here.
+ * The following paths exist:
+ *
+ * exit path:
+ * sys_exit_group
+ *    do_group_exit
+ *     do_exit
+ *      exit_mm
+ *       mmput
+ *        exit_mmap
+ *         remove_vma
+ *          fput
+ *           __fput
+ *            pfm_close
+ *             __pfm_close
+ *              pfm_context_free
+ *              pfm_release_buf_space
+ * munmap path:
+ * sys_munmap
+ *  do_munmap
+ *   remove_vma
+ *    fput
+ *     __fput
+ *      pfm_close
+ *       __pfm_close
+ *        pfm_context_free
+ *         pfm_release_buf_space
+ *
+ * close path:
+ * sys_close
+ *  filp_close
+ *   fput
+ *    __fput
+ *     pfm_close
+ *      __pfm_close
+ *       pfm_context_free
+ *        pfm_release_buf_space
+ *
+ * The issue is that on the munmap() path, the mmap_sem is already held
+ * in write-mode by the time we come here. To avoid the deadlock, we need
+ * to know where we are coming from and skip down_write(). If is fairly
+ * difficult to know this because of the lack of good hooks and
+ * the fact that, there may not have been any mmap() of the sampling buffer
+ * (i.e. create_context() followed by close() or exit()).
+ *
+ * We use a set flag ctx->flags.mmap_nlock which is toggled in the vm_ops
+ * callback in remove_vma() which is called systematically for the call, so
+ * on all but the pure close() path. The exit path does not already hold
+ * the lock but this is exit so there is no task->mm by the time we come here.
+ *
+ * The mmap_nlock is set only when unmapping and this is the LAST reference
+ * to the file (i.e., close() followed by munmap()).
+ */
+void pfm_smpl_buf_space_release(struct pfm_context *ctx, size_t size)
+{
+       unsigned long flags;
+       struct mm_struct *mm;
+
+       mm = get_task_mm(current);
+       if (mm) {
+               if (ctx->flags.mmap_nlock == 0) {
+                       PFM_DBG("doing down_write");
+                       down_write(&mm->mmap_sem);
+               }
+
+               mm->locked_vm -= size >> PAGE_SHIFT;
+
+               PFM_DBG("size=%zu locked_vm=%lu", size, mm->locked_vm);
+
+               if (ctx->flags.mmap_nlock == 0)
+                       up_write(&mm->mmap_sem);
+
+               mmput(mm);
+       }
+
+       spin_lock_irqsave(&pfm_res_lock, flags);
+
+       pfm_res.smpl_buf_mem_cur -= size;
+
+       spin_unlock_irqrestore(&pfm_res_lock, flags);
+}
+
+/**
+ * pfm_session_acquire - reserve a per-thread or per-cpu session
+ * @is_system: true if per-cpu session
+ * @cpu: cpu number for per-cpu session
+ *
+ * return:
+ *      0    : success
+ *     -EBUSY: if conflicting session exist
+ */
+int pfm_session_acquire(int is_system, u32 cpu)
+{
+       unsigned long flags;
+       u32 nsys_cpus;
+       int ret = 0;
+
+       /*
+        * validy checks on cpu_mask have been done upstream
+        */
+       spin_lock_irqsave(&pfm_res_lock, flags);
+
+       nsys_cpus = cpus_weight(pfm_res.sys_cpumask);
+
+       PFM_DBG("in  sys=%u task=%u is_sys=%d cpu=%u",
+               nsys_cpus,
+               pfm_res.thread_sessions,
+               is_system,
+               cpu);
+
+       if (is_system) {
+               /*
+                * cannot mix system wide and per-task sessions
+                */
+               if (pfm_res.thread_sessions > 0) {
+                       PFM_DBG("%u conflicting thread_sessions",
+                               pfm_res.thread_sessions);
+                       ret = -EBUSY;
+                       goto abort;
+               }
+
+               if (cpu_isset(cpu, pfm_res.sys_cpumask)) {
+                       PFM_DBG("conflicting session on CPU%u", cpu);
+                       ret = -EBUSY;
+                       goto abort;
+               }
+
+               PFM_DBG("reserved session on CPU%u", cpu);
+
+               cpu_set(cpu, pfm_res.sys_cpumask);
+               nsys_cpus++;
+       } else {
+               if (nsys_cpus) {
+                       ret = -EBUSY;
+                       goto abort;
+               }
+               pfm_res.thread_sessions++;
+       }
+
+       PFM_DBG("out sys=%u task=%u is_sys=%d cpu=%u",
+               nsys_cpus,
+               pfm_res.thread_sessions,
+               is_system,
+               cpu);
+
+abort:
+       spin_unlock_irqrestore(&pfm_res_lock, flags);
+
+       return ret;
+}
+
+/**
+ * pfm_session_release - release a per-cpu or per-thread session
+ * @is_system: true if per-cpu session
+ * @cpu: cpu number for per-cpu session
+ *
+ * called from __pfm_unload_context()
+ */
+void pfm_session_release(int is_system, u32 cpu)
+{
+       unsigned long flags;
+
+       spin_lock_irqsave(&pfm_res_lock, flags);
+
+       PFM_DBG("in sys_sessions=%u thread_sessions=%u syswide=%d cpu=%u",
+               cpus_weight(pfm_res.sys_cpumask),
+               pfm_res.thread_sessions,
+               is_system, cpu);
+
+       if (is_system)
+               cpu_clear(cpu, pfm_res.sys_cpumask);
+       else
+               pfm_res.thread_sessions--;
+
+       PFM_DBG("out sys_sessions=%u thread_sessions=%u syswide=%d cpu=%u",
+               cpus_weight(pfm_res.sys_cpumask),
+               pfm_res.thread_sessions,
+               is_system, cpu);
+
+       spin_unlock_irqrestore(&pfm_res_lock, flags);
+}
+
+/**
+ * pfm_session_allcpus_acquire - acquire per-cpu sessions on all available cpus
+ *
+ * currently used by Oprofile on X86
+ */
+int pfm_session_allcpus_acquire(void)
+{
+       unsigned long flags;
+       u32 nsys_cpus, cpu;
+       int ret = -EBUSY;
+
+       spin_lock_irqsave(&pfm_res_lock, flags);
+
+       nsys_cpus = cpus_weight(pfm_res.sys_cpumask);
+
+       PFM_DBG("in  sys=%u task=%u",
+               nsys_cpus,
+               pfm_res.thread_sessions);
+
+       if (nsys_cpus) {
+               PFM_DBG("already some system-wide sessions");
+               goto abort;
+       }
+
+       /*
+        * cannot mix system wide and per-task sessions
+        */
+       if (pfm_res.thread_sessions) {
+               PFM_DBG("%u conflicting thread_sessions",
+                       pfm_res.thread_sessions);
+               goto abort;
+       }
+
+       for_each_online_cpu(cpu) {
+               cpu_set(cpu, pfm_res.sys_cpumask);
+               nsys_cpus++;
+       }
+
+       PFM_DBG("out sys=%u task=%u",
+               nsys_cpus,
+               pfm_res.thread_sessions);
+
+       ret = 0;
+abort:
+       spin_unlock_irqrestore(&pfm_res_lock, flags);
+
+       return ret;
+}
+EXPORT_SYMBOL(pfm_session_allcpus_acquire);
+
+/**
+ * pfm_session_allcpus_release - relase per-cpu sessions on all cpus
+ *
+ * currently used by Oprofile code
+ */
+void pfm_session_allcpus_release(void)
+{
+       unsigned long flags;
+       u32 nsys_cpus, cpu;
+
+       spin_lock_irqsave(&pfm_res_lock, flags);
+
+       nsys_cpus = cpus_weight(pfm_res.sys_cpumask);
+
+       PFM_DBG("in  sys=%u task=%u",
+               nsys_cpus,
+               pfm_res.thread_sessions);
+
+       /*
+        * XXX: could use __cpus_clear() with nbits
+        */
+       for_each_online_cpu(cpu) {
+               cpu_clear(cpu, pfm_res.sys_cpumask);
+               nsys_cpus--;
+       }
+
+       PFM_DBG("out sys=%u task=%u",
+               nsys_cpus,
+               pfm_res.thread_sessions);
+
+       spin_unlock_irqrestore(&pfm_res_lock, flags);
+}
+EXPORT_SYMBOL(pfm_session_allcpus_release);
+
+/**
+ * pfm_sysfs_res_show - return currnt resourcde usage for sysfs
+ * @buf: buffer to hold string in return
+ * @sz: size of buf
+ * @what: what to produce
+ *        what=0 : thread_sessions
+ *        what=1 : cpus_weight(sys_cpumask)
+ *        what=2 : smpl_buf_mem_cur
+ *        what=3 : pmu model name
+ *
+ * called from perfmon_sysfs.c
+ * return number of bytes written into buf (up to sz)
+ */
+ssize_t pfm_sysfs_res_show(char *buf, size_t sz, int what)
+{
+       unsigned long flags;
+
+       spin_lock_irqsave(&pfm_res_lock, flags);
+
+       switch (what) {
+       case 0: snprintf(buf, sz, "%u\n", pfm_res.thread_sessions);
+               break;
+       case 1: snprintf(buf, sz, "%d\n", cpus_weight(pfm_res.sys_cpumask));
+               break;
+       case 2: snprintf(buf, sz, "%zu\n", pfm_res.smpl_buf_mem_cur);
+               break;
+       case 3:
+               snprintf(buf, sz, "%s\n",
+                       pfm_pmu_conf ?  pfm_pmu_conf->pmu_name
+                                    :  "unknown\n");
+       }
+       spin_unlock_irqrestore(&pfm_res_lock, flags);
+       return strlen(buf);
+}
diff --git a/perfmon/perfmon_rw.c b/perfmon/perfmon_rw.c
new file mode 100644
index 0000000..3168eb7
--- /dev/null
+++ b/perfmon/perfmon_rw.c
@@ -0,0 +1,733 @@
+/*
+ * perfmon.c: perfmon2 PMC/PMD read/write system calls
+ *
+ * This file implements the perfmon2 interface which
+ * provides access to the hardware performance counters
+ * of the host processor.
+ *
+ * The initial version of perfmon.c was written by
+ * Ganesh Venkitachalam, IBM Corp.
+ *
+ * Then it was modified for perfmon-1.x by Stephane Eranian and
+ * David Mosberger, Hewlett Packard Co.
+ *
+ * Version Perfmon-2.x is a complete rewrite of perfmon-1.x
+ * by Stephane Eranian, Hewlett Packard Co.
+ *
+ * Copyright (c) 1999-2006 Hewlett-Packard Development Company, L.P.
+ * Contributed by Stephane Eranian <eranian@hpl.hp.com>
+ *                David Mosberger-Tang <davidm@hpl.hp.com>
+ *
+ * More information about perfmon available at:
+ *     http://perfmon2.sf.net/
+ *
+ * This program is free software; you can redistribute it and/or
+ * modify it under the terms of version 2 of the GNU General Public
+ * License as published by the Free Software Foundation.
+ *
+ * This program is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+ * General Public License for more details.
+ *
+ * You should have received a copy of the GNU General Public License
+ * along with this program; if not, write to the Free Software
+ * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA
+ * 02111-1307 USA
+ */
+#include <linux/module.h>
+#include <linux/kernel.h>
+#include <linux/perfmon_kern.h>
+#include "perfmon_priv.h"
+
+#define PFM_REGFL_PMC_ALL      (PFM_REGFL_NO_EMUL64)
+#define PFM_REGFL_PMD_ALL      (PFM_REGFL_RANDOM|PFM_REGFL_OVFL_NOTIFY)
+
+/**
+ * update_used_reg -- updated used_pmcs for a single PMD
+ * @set: set to update
+ * @cnum: new PMD to add
+ *
+ * This function adds the pmds and pmcs depending on PMD cnum
+ */
+static inline void update_used_reg(struct pfm_context *ctx,
+                                  struct pfm_event_set *set, u16 cnum)
+{
+       bitmap_or(cast_ulp(set->used_pmcs),
+                 cast_ulp(set->used_pmcs),
+                 cast_ulp(pfm_pmu_conf->pmd_desc[cnum].dep_pmcs),
+                 ctx->regs.max_pmc);
+}
+
+/**
+ * update_used -- update used_pmcs bitmask
+ * @set: event set to update
+ * @bv: bitmask to inspect for new PMD registers
+ *
+ * This function updates the used_pmcs bitmask for
+ * the set using bv, a bitmask of pmds. For each pmd in bv,
+ * its depending pmcs are added to used_pmcs.
+ */
+static void update_used_pmcs(struct pfm_context *ctx,
+                            struct pfm_event_set *set, unsigned long *bv)
+{
+       u16 max_pmd;
+       int n, p, q;
+
+       max_pmd = ctx->regs.max_pmd;
+
+       n = bitmap_weight(bv, max_pmd);
+       for(p = 0; n; n--, p = q+1) {
+               q = find_next_bit(bv, max_pmd, p);
+               update_used_reg(ctx, set, q);
+       }
+}
+
+/**
+ * update_changes -- update nused_pmcs, nused_pmds, write newly touched pmcs
+ * @ctx: context to use
+ * @set: event set to use
+ * @old_used_pmcs: former used_pmc bitmask
+ * @can_access: non-zero if PMU is accessible, i.e., can be written to
+ *
+ * This function updates nused_pmcs and nused_pmds after the last modificiation
+ * to an event set. When new pmcs are used, then they must be initialized such
+ * that we do not pick up stale values from another session.
+ */
+static inline int update_changes(struct pfm_context *ctx, struct pfm_event_set *set,
+                                unsigned long *old_used_pmcs)
+{
+       struct pfarg_pmc req;
+       u16 max_pmc, max_pmd;
+       int n, p, q, ret = 0;
+
+       max_pmd = ctx->regs.max_pmd;
+       max_pmc = ctx->regs.max_pmc;
+
+       /*
+        * update used counts
+        */
+       set->nused_pmds = bitmap_weight(cast_ulp(set->used_pmds), max_pmd);
+       set->nused_pmcs = bitmap_weight(cast_ulp(set->used_pmcs), max_pmc);
+
+       PFM_DBG("set%u u_pmds=0x%llx nu_pmds=%u u_pmcs=0x%llx nu_pmcs=%u",
+                       set->id,
+                       (unsigned long long)set->used_pmds[0],
+                       set->nused_pmds,
+                       (unsigned long long)set->used_pmcs[0],
+                       set->nused_pmcs);
+
+       memset(&req, 0, sizeof(req));
+
+       n = bitmap_weight(cast_ulp(set->used_pmcs), max_pmc);
+       for(p = 0; n; n--, p = q+1) {
+               q = find_next_bit(cast_ulp(set->used_pmcs), max_pmc, p);
+
+               if (test_bit(q, cast_ulp(old_used_pmcs)))
+                       continue;
+
+               req.reg_num = q;
+               req.reg_value = set->pmcs[q];
+
+               ret = __pfm_write_pmcs(ctx, &req, 1);
+               if (ret)
+                       break;
+       }
+       return ret;
+}
+
+/**
+ * handle_smpl_bv - checks sampling bitmasks for new PMDs
+ * @ctx: context to use
+ * @set: set to use
+ * @bv: sampling bitmask
+ *
+ * scans the smpl bitmask looking for new PMDs (not yet used), if found
+ * invoke pfm_write_pmds() on them to get them initialized and marked used
+ */
+static int handle_smpl_bv(struct pfm_context *ctx, struct pfm_event_set *set,
+                         unsigned long *bv)
+{
+       struct pfarg_pmd req;
+       int p, q, n, ret = 0;
+       u16 max_pmd;
+
+       memset(&req, 0, sizeof(req));
+
+       max_pmd = ctx->regs.max_pmd;
+
+       n = bitmap_weight(cast_ulp(bv), max_pmd);
+
+       for(p = 0; n; n--, p = q+1) {
+               q = find_next_bit(cast_ulp(bv), max_pmd, p);
+
+               if (test_bit(q, cast_ulp(set->used_pmds)))
+                       continue;
+
+               req.reg_num = q;
+               req.reg_value = 0;
+
+               ret = __pfm_write_pmds(ctx, &req, 1, 0);
+               if (ret)
+                       break;
+       }
+       return ret;
+}
+
+/**
+ * is_invalid -- check if register index is within limits
+ * @cnum: register index
+ * @impl: bitmask of implemented registers
+ * @max: highest implemented registers + 1
+ *
+ * return:
+ *    0 is register index is valid
+ *    1 if invalid
+ */
+static inline int is_invalid(u16 cnum, unsigned long *impl, u16 max)
+{
+       return cnum >= max || !test_bit(cnum, impl);
+}
+
+/**
+ * __pfm_write_pmds - modified data registers
+ * @ctx: context to operate on
+ * @req: pfarg_pmd_t request from user
+ * @count: number of element in the pfarg_pmd_t vector
+ * @compat: used only on IA-64 to maintain backward compatibility with v2.0
+ *
+ * The function succeeds whether the context is attached or not.
+ * When attached to another thread, that thread must be stopped.
+ *
+ * The context is locked and interrupts are disabled.
+ */
+int __pfm_write_pmds(struct pfm_context *ctx, struct pfarg_pmd *req, int count,
+                    int compat)
+{
+       struct pfm_event_set *set, *active_set;
+       u64 old_used_pmcs[PFM_PMC_BV];
+       unsigned long *smpl_pmds, *reset_pmds, *impl_pmds, *impl_rw_pmds;
+       u32 req_flags, flags;
+       u16 cnum, pmd_type, max_pmd;
+       u16 set_id;
+       int i, can_access_pmu;
+       int ret;
+       pfm_pmd_check_t wr_func;
+
+       active_set = ctx->active_set;
+       max_pmd = ctx->regs.max_pmd;
+       impl_pmds = cast_ulp(ctx->regs.pmds);
+       impl_rw_pmds = cast_ulp(ctx->regs.rw_pmds);
+       wr_func = pfm_pmu_conf->pmd_write_check;
+       set = list_first_entry(&ctx->set_list, struct pfm_event_set, list);
+
+       can_access_pmu = 0;
+
+       /*
+        * we cannot access the actual PMD registers when monitoring is masked
+        */
+       if (unlikely(ctx->state == PFM_CTX_LOADED))
+               can_access_pmu = __get_cpu_var(pmu_owner) == ctx->task
+                       || ctx->flags.system;
+
+       bitmap_copy(cast_ulp(old_used_pmcs),
+                   cast_ulp(set->used_pmcs),
+                   ctx->regs.max_pmc);
+
+       ret = -EINVAL;
+       for (i = 0; i < count; i++, req++) {
+
+               cnum = req->reg_num;
+               set_id = req->reg_set;
+               req_flags = req->reg_flags;
+               smpl_pmds = cast_ulp(req->reg_smpl_pmds);
+               reset_pmds = cast_ulp(req->reg_reset_pmds);
+               flags = 0;
+
+               /*
+                * cannot write to unexisting
+                * writes to read-only register are ignored
+                */
+               if (unlikely(is_invalid(cnum, impl_pmds, max_pmd))) {
+                       PFM_DBG("pmd%u is not available", cnum);
+                       goto error;
+               }
+
+               pmd_type = pfm_pmu_conf->pmd_desc[cnum].type;
+
+               /*
+                * ensure only valid flags are set
+                */
+               if (req_flags & ~(PFM_REGFL_PMD_ALL)) {
+                       PFM_DBG("pmd%u: invalid flags=0x%x",
+                               cnum, req_flags);
+                       goto error;
+               }
+
+               /*
+                * OVFL_NOTIFY is valid for all types of PMD.
+                * non counting PMD may trigger PMU interrupt
+                * and thus may trigger recording of a sample.
+                * This is true with IBS on AMD family 16.
+                */
+               if (req_flags & PFM_REGFL_OVFL_NOTIFY)
+                       flags |= PFM_REGFL_OVFL_NOTIFY;
+
+               /*
+                * We allow randomization to non counting PMD
+                */
+               if (req_flags & PFM_REGFL_RANDOM)
+                       flags |= PFM_REGFL_RANDOM;
+
+               /*
+                * verify validity of smpl_pmds
+                */
+               if (unlikely(!bitmap_subset(smpl_pmds, impl_pmds, PFM_MAX_PMDS))) {
+                       PFM_DBG("invalid smpl_pmds=0x%llx for pmd%u",
+                               (unsigned long long)req->reg_smpl_pmds[0],
+                               cnum);
+                       goto error;
+               }
+
+               /*
+                * verify validity of reset_pmds
+                * check against impl_rw_pmds because it is not
+                * possible to reset read-only PMDs
+                */
+               if (unlikely(!bitmap_subset(reset_pmds, impl_rw_pmds, PFM_MAX_PMDS))) {
+                       PFM_DBG("invalid reset_pmds=0x%llx for pmd%u",
+                               (unsigned long long)req->reg_reset_pmds[0],
+                               cnum);
+                       goto error;
+               }
+
+               /*
+                * locate event set
+                */
+               if (set_id != set->id) {
+                       /* update number of used register for previous set */
+                       if (i) {
+                               ret = update_changes(ctx, set, cast_ulp(old_used_pmcs));
+                               if (ret)
+                                       goto error;
+                       }
+
+                       set = pfm_find_set(ctx, set_id, 0);
+                       if (set == NULL) {
+                               PFM_DBG("event set%u does not exist",
+                                               set_id);
+                               goto error;
+                       }
+                       bitmap_copy(cast_ulp(old_used_pmcs),
+                                   cast_ulp(set->used_pmcs),
+                                   ctx->regs.max_pmc);
+               }
+
+               /*
+                * execute write checker, if any
+                */
+               if (unlikely(wr_func && (pmd_type & PFM_REG_WC))) {
+                       ret = (*wr_func)(ctx, set, req);
+                       if (ret)
+                               goto error;
+
+               }
+
+
+               /*
+                * now commit changes to software state
+                */
+
+               if (unlikely(compat))
+                       goto skip_set;
+
+               if (bitmap_weight(smpl_pmds, max_pmd)) {
+                       ret = handle_smpl_bv(ctx, set, smpl_pmds);
+                       if (ret)
+                               goto error;
+                       update_used_pmcs(ctx, set, cast_ulp(smpl_pmds));
+               }
+
+               bitmap_copy(cast_ulp(set->pmds[cnum].smpl_pmds),
+                           smpl_pmds,
+                           max_pmd);
+
+
+               if (bitmap_weight(reset_pmds, max_pmd)) {
+                       ret = handle_smpl_bv(ctx, set, reset_pmds);
+                       if (ret)
+                               goto error;
+                       update_used_pmcs(ctx, set, cast_ulp(reset_pmds));
+               }
+
+               bitmap_copy(cast_ulp(set->pmds[cnum].reset_pmds),
+                          reset_pmds,
+                          max_pmd);
+
+               set->pmds[cnum].flags = flags;
+
+               __set_bit(cnum, cast_ulp(set->used_pmds));
+               update_used_reg(ctx, set, cnum);
+
+               /*
+                * we reprogram the PMD hence, we clear any pending
+                * ovfl. Does affect ovfl switch on restart but new
+                * value has already been established here
+                */
+               if (test_bit(cnum, cast_ulp(set->povfl_pmds))) {
+                       set->npend_ovfls--;
+                       __clear_bit(cnum, cast_ulp(set->povfl_pmds));
+               }
+               __clear_bit(cnum, cast_ulp(set->ovfl_pmds));
+
+               /*
+                * update ovfl_notify
+                */
+               if (flags & PFM_REGFL_OVFL_NOTIFY)
+                       __set_bit(cnum, cast_ulp(set->ovfl_notify));
+               else
+                       __clear_bit(cnum, cast_ulp(set->ovfl_notify));
+
+               /*
+                * establish new switch count
+                */
+               set->pmds[cnum].ovflsw_thres = req->reg_ovfl_switch_cnt;
+               set->pmds[cnum].ovflsw_ref_thres = req->reg_ovfl_switch_cnt;
+skip_set:
+
+               /*
+                * set last value to new value for all types of PMD
+                */
+               set->pmds[cnum].lval = req->reg_value;
+               set->pmds[cnum].value = req->reg_value;
+
+               /*
+                * update reset values (not just for counters)
+                */
+               set->pmds[cnum].long_reset = req->reg_long_reset;
+               set->pmds[cnum].short_reset = req->reg_short_reset;
+
+               /*
+                * update randomization mask
+                */
+               set->pmds[cnum].mask = req->reg_random_mask;
+
+               set->pmds[cnum].eventid = req->reg_smpl_eventid;
+
+               if (set == active_set) {
+                       set->priv_flags |= PFM_SETFL_PRIV_MOD_PMDS;
+                       if (can_access_pmu)
+                               pfm_write_pmd(ctx, cnum, req->reg_value);
+               }
+
+
+               PFM_DBG("set%u pmd%u=0x%llx flags=0x%x a_pmu=%d "
+                       "ctx_pmd=0x%llx s_reset=0x%llx "
+                       "l_reset=0x%llx s_pmds=0x%llx "
+                       "r_pmds=0x%llx o_pmds=0x%llx "
+                       "o_thres=%llu compat=%d eventid=%llx",
+                       set->id,
+                       cnum,
+                       (unsigned long long)req->reg_value,
+                       set->pmds[cnum].flags,
+                       can_access_pmu,
+                       (unsigned long long)set->pmds[cnum].value,
+                       (unsigned long long)set->pmds[cnum].short_reset,
+                       (unsigned long long)set->pmds[cnum].long_reset,
+                       (unsigned long long)set->pmds[cnum].smpl_pmds[0],
+                       (unsigned long long)set->pmds[cnum].reset_pmds[0],
+                       (unsigned long long)set->ovfl_pmds[0],
+                       (unsigned long long)set->pmds[cnum].ovflsw_thres,
+                       compat,
+                       (unsigned long long)set->pmds[cnum].eventid);
+       }
+       ret = 0;
+
+error:
+       update_changes(ctx, set, cast_ulp(old_used_pmcs));
+
+       /*
+        * make changes visible
+        */
+       if (can_access_pmu)
+               pfm_arch_serialize();
+
+       return ret;
+}
+
+/**
+ * __pfm_write_pmcs - modified config registers
+ * @ctx: context to operate on
+ * @req: pfarg_pmc_t request from user
+ * @count: number of element in the pfarg_pmc_t vector
+ *
+ *
+ * The function succeeds whether the context is * attached or not.
+ * When attached to another thread, that thread must be stopped.
+ *
+ * The context is locked and interrupts are disabled.
+ */
+int __pfm_write_pmcs(struct pfm_context *ctx, struct pfarg_pmc *req, int count)
+{
+       struct pfm_event_set *set, *active_set;
+       u64 value, dfl_val, rsvd_msk;
+       unsigned long *impl_pmcs;
+       int i, can_access_pmu;
+       int ret;
+       u16 set_id;
+       u16 cnum, pmc_type, max_pmc;
+       u32 flags, expert;
+       pfm_pmc_check_t wr_func;
+
+       active_set = ctx->active_set;
+
+       wr_func = pfm_pmu_conf->pmc_write_check;
+       max_pmc = ctx->regs.max_pmc;
+       impl_pmcs = cast_ulp(ctx->regs.pmcs);
+       set = list_first_entry(&ctx->set_list, struct pfm_event_set, list);
+
+       expert = pfm_controls.flags & PFM_CTRL_FL_RW_EXPERT;
+
+       can_access_pmu = 0;
+
+       /*
+        * we cannot access the actual PMC registers when monitoring is masked
+        */
+       if (unlikely(ctx->state == PFM_CTX_LOADED))
+               can_access_pmu = __get_cpu_var(pmu_owner) == ctx->task
+                       || ctx->flags.system;
+
+       ret = -EINVAL;
+
+       for (i = 0; i < count; i++, req++) {
+
+               cnum = req->reg_num;
+               set_id = req->reg_set;
+               value = req->reg_value;
+               flags = req->reg_flags;
+
+               /*
+                * no access to unavailable PMC register
+                */
+               if (unlikely(is_invalid(cnum, impl_pmcs, max_pmc))) {
+                       PFM_DBG("pmc%u is not available", cnum);
+                       goto error;
+               }
+
+               pmc_type = pfm_pmu_conf->pmc_desc[cnum].type;
+               dfl_val = pfm_pmu_conf->pmc_desc[cnum].dfl_val;
+               rsvd_msk = pfm_pmu_conf->pmc_desc[cnum].rsvd_msk;
+
+               /*
+                * ensure only valid flags are set
+                */
+               if (flags & ~PFM_REGFL_PMC_ALL) {
+                       PFM_DBG("pmc%u: invalid flags=0x%x", cnum, flags);
+                       goto error;
+               }
+
+               /*
+                * locate event set
+                */
+               if (set_id != set->id) {
+                       set = pfm_find_set(ctx, set_id, 0);
+                       if (set == NULL) {
+                               PFM_DBG("event set%u does not exist",
+                                       set_id);
+                               goto error;
+                       }
+               }
+
+               /*
+                * set reserved bits to default values
+                * (reserved bits must be 1 in rsvd_msk)
+                *
+                * bypass via /sys/kernel/perfmon/mode = 1
+                */
+               if (likely(!expert))
+                       value = (value & ~rsvd_msk) | (dfl_val & rsvd_msk);
+
+               if (flags & PFM_REGFL_NO_EMUL64) {
+                       if (!(pmc_type & PFM_REG_NO64)) {
+                               PFM_DBG("pmc%u no support for "
+                                       "PFM_REGFL_NO_EMUL64", cnum);
+                               goto error;
+                       }
+                       value &= ~pfm_pmu_conf->pmc_desc[cnum].no_emul64_msk;
+               }
+
+               /*
+                * execute write checker, if any
+                */
+               if (likely(wr_func && (pmc_type & PFM_REG_WC))) {
+                       req->reg_value = value;
+                       ret = (*wr_func)(ctx, set, req);
+                       if (ret)
+                               goto error;
+                       value = req->reg_value;
+               }
+
+               /*
+                * Now we commit the changes
+                */
+
+               /*
+                * mark PMC register as used
+                * We do not track associated PMC register based on
+                * the fact that they will likely need to be written
+                * in order to become useful at which point the statement
+                * below will catch that.
+                *
+                * The used_pmcs bitmask is only useful on architectures where
+                * the PMC needs to be modified for particular bits, especially
+                * on overflow or to stop/start.
+                */
+               if (!test_bit(cnum, cast_ulp(set->used_pmcs))) {
+                       __set_bit(cnum, cast_ulp(set->used_pmcs));
+                       set->nused_pmcs++;
+               }
+
+               set->pmcs[cnum] = value;
+
+               if (set == active_set) {
+                       set->priv_flags |= PFM_SETFL_PRIV_MOD_PMCS;
+                       if (can_access_pmu)
+                               pfm_arch_write_pmc(ctx, cnum, value);
+               }
+
+               PFM_DBG("set%u pmc%u=0x%llx a_pmu=%d "
+                       "u_pmcs=0x%llx nu_pmcs=%u",
+                       set->id,
+                       cnum,
+                       (unsigned long long)value,
+                       can_access_pmu,
+                       (unsigned long long)set->used_pmcs[0],
+                       set->nused_pmcs);
+       }
+       ret = 0;
+error:
+       /*
+        * make sure the changes are visible
+        */
+       if (can_access_pmu)
+               pfm_arch_serialize();
+
+       return ret;
+}
+
+/**
+ * __pfm_read_pmds - read data registers
+ * @ctx: context to operate on
+ * @req: pfarg_pmd_t request from user
+ * @count: number of element in the pfarg_pmd_t vector
+ *
+ *
+ * The function succeeds whether the context is attached or not.
+ * When attached to another thread, that thread must be stopped.
+ *
+ * The context is locked and interrupts are disabled.
+ */
+int __pfm_read_pmds(struct pfm_context *ctx, struct pfarg_pmd *req, int count)
+{
+       u64 val = 0, lval, ovfl_mask, hw_val;
+       u64 sw_cnt;
+       unsigned long *impl_pmds;
+       struct pfm_event_set *set, *active_set;
+       int i, ret, can_access_pmu = 0;
+       u16 cnum, pmd_type, set_id, max_pmd;
+
+       ovfl_mask = pfm_pmu_conf->ovfl_mask;
+       impl_pmds = cast_ulp(ctx->regs.pmds);
+       max_pmd   = ctx->regs.max_pmd;
+       active_set = ctx->active_set;
+       set = list_first_entry(&ctx->set_list, struct pfm_event_set, list);
+
+       if (likely(ctx->state == PFM_CTX_LOADED)) {
+               can_access_pmu = __get_cpu_var(pmu_owner) == ctx->task
+                       || ctx->flags.system;
+
+               if (can_access_pmu)
+                       pfm_arch_serialize();
+       }
+
+       /*
+        * on both UP and SMP, we can only read the PMD from the hardware
+        * register when the task is the owner of the local PMU.
+        */
+       ret = -EINVAL;
+       for (i = 0; i < count; i++, req++) {
+
+               cnum = req->reg_num;
+               set_id = req->reg_set;
+
+               if (unlikely(is_invalid(cnum, impl_pmds, max_pmd))) {
+                       PFM_DBG("pmd%u is not implemented/unaccessible", cnum);
+                       goto error;
+               }
+
+               pmd_type = pfm_pmu_conf->pmd_desc[cnum].type;
+
+               /*
+                * locate event set
+                */
+               if (set_id != set->id) {
+                       set = pfm_find_set(ctx, set_id, 0);
+                       if (set == NULL) {
+                               PFM_DBG("event set%u does not exist",
+                                       set_id);
+                               goto error;
+                       }
+               }
+               /*
+                * it is not possible to read a PMD which was not requested:
+                *      - explicitly written via pfm_write_pmds()
+                *      - provided as a reg_smpl_pmds[] to another PMD during
+                *        pfm_write_pmds()
+                *
+                * This is motivated by security and for optimization purposes:
+                *      - on context switch restore, we can restore only what
+                *        we use (except when regs directly readable at user
+                *        level, e.g., IA-64 self-monitoring, I386 RDPMC).
+                *      - do not need to maintain PMC -> PMD dependencies
+                */
+               if (unlikely(!test_bit(cnum, cast_ulp(set->used_pmds)))) {
+                       PFM_DBG("pmd%u cannot read, because not used", cnum);
+                       goto error;
+               }
+
+               val = set->pmds[cnum].value;
+               lval = set->pmds[cnum].lval;
+
+               /*
+                * extract remaining ovfl to switch
+                */
+               sw_cnt = set->pmds[cnum].ovflsw_thres;
+
+               /*
+                * If the task is not the current one, then we check if the
+                * PMU state is still in the local live register due to lazy
+                * ctxsw. If true, then we read directly from the registers.
+                */
+               if (set == active_set && can_access_pmu) {
+                       hw_val = pfm_read_pmd(ctx, cnum);
+                       if (pmd_type & PFM_REG_C64)
+                               val = (val & ~ovfl_mask) | (hw_val & ovfl_mask);
+                       else
+                               val = hw_val;
+               }
+
+               PFM_DBG("set%u pmd%u=0x%llx sw_thr=%llu lval=0x%llx",
+                       set->id,
+                       cnum,
+                       (unsigned long long)val,
+                       (unsigned long long)sw_cnt,
+                       (unsigned long long)lval);
+
+               req->reg_value = val;
+               req->reg_last_reset_val = lval;
+               req->reg_ovfl_switch_cnt = sw_cnt;
+       }
+       ret = 0;
+error:
+       return ret;
+}
diff --git a/perfmon/perfmon_sets.c b/perfmon/perfmon_sets.c
new file mode 100644
index 0000000..24534cb
--- /dev/null
+++ b/perfmon/perfmon_sets.c
@@ -0,0 +1,873 @@
+/*
+ * perfmon_sets.c: perfmon2 event sets and multiplexing functions
+ *
+ * This file implements the perfmon2 interface which
+ * provides access to the hardware performance counters
+ * of the host processor.
+ *
+ * The initial version of perfmon.c was written by
+ * Ganesh Venkitachalam, IBM Corp.
+ *
+ * Then it was modified for perfmon-1.x by Stephane Eranian and
+ * David Mosberger, Hewlett Packard Co.
+ *
+ * Version Perfmon-2.x is a complete rewrite of perfmon-1.x
+ * by Stephane Eranian, Hewlett Packard Co.
+ *
+ * Copyright (c) 1999-2006 Hewlett-Packard Development Company, L.P.
+ * Contributed by Stephane Eranian <eranian@hpl.hp.com>
+ *                David Mosberger-Tang <davidm@hpl.hp.com>
+ *
+ * More information about perfmon available at:
+ *     http://perfmon2.sf.net
+ *
+ * This program is free software; you can redistribute it and/or
+ * modify it under the terms of version 2 of the GNU General Public
+ * License as published by the Free Software Foundation.
+ *
+ * This program is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+ * General Public License for more details.
+ *
+ * You should have received a copy of the GNU General Public License
+ * along with this program; if not, write to the Free Software
+ * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA
+ * 02111-1307 USA
+ */
+#include <linux/kernel.h>
+#include <linux/perfmon_kern.h>
+#include "perfmon_priv.h"
+
+static struct kmem_cache       *pfm_set_cachep;
+
+/**
+ * pfm_reload_switch_thresholds - reload overflow-based switch thresholds per set
+ * @set: the set for which to reload thresholds
+ *
+ */
+static void pfm_reload_switch_thresholds(struct pfm_context *ctx,
+                                        struct pfm_event_set *set)
+{
+       u64 *used_pmds;
+       u16 i, max, first;
+
+       used_pmds = set->used_pmds;
+       first = ctx->regs.first_intr_pmd;
+       max = ctx->regs.max_intr_pmd;
+
+       for (i = first; i < max; i++) {
+               if (test_bit(i, cast_ulp(used_pmds))) {
+                       set->pmds[i].ovflsw_thres = set->pmds[i].ovflsw_ref_thres;
+
+                       PFM_DBG("set%u pmd%u ovflsw_thres=%llu",
+                               set->id,
+                               i,
+                               (unsigned long long)set->pmds[i].ovflsw_thres);
+               }
+       }
+}
+
+/**
+ * pfm_prepare_sets - initialize sets on pfm_load_context
+ * @ctx : context to operate on
+ * @load_set: set to activate first
+ *
+ * connect all sets, reset internal fields
+ */
+struct pfm_event_set *pfm_prepare_sets(struct pfm_context *ctx, u16 load_set)
+{
+       struct pfm_event_set *set, *p;
+       u16 max;
+
+       /*
+        * locate first set to activate
+        */
+       set = pfm_find_set(ctx, load_set, 0);
+       if (!set)
+               return NULL;
+
+       if (set->flags & PFM_SETFL_OVFL_SWITCH)
+               pfm_reload_switch_thresholds(ctx, set);
+
+       max = ctx->regs.max_intr_pmd;
+
+       list_for_each_entry(p, &ctx->set_list, list) {
+               /*
+                * cleanup bitvectors
+                */
+               bitmap_zero(cast_ulp(p->ovfl_pmds), max);
+               bitmap_zero(cast_ulp(p->povfl_pmds), max);
+
+               p->npend_ovfls = 0;
+
+               /*
+                * we cannot just use plain clear because of arch-specific flags
+                */
+               p->priv_flags &= ~(PFM_SETFL_PRIV_MOD_BOTH|PFM_SETFL_PRIV_SWITCH);
+               /*
+                * neither duration nor runs are reset because typically loading/unloading
+                * does not mean counts are reset. To reset, the set must be modified
+                */
+       }
+       return set;
+}
+
+/*
+ * called by hrtimer_interrupt()
+ *
+ * This is the only function where we come with
+ * cpu_base->lock held before ctx->lock
+ *
+ * interrupts are disabled
+ */
+enum hrtimer_restart pfm_handle_switch_timeout(struct hrtimer *t)
+{
+       struct pfm_event_set *set;
+       struct pfm_context *ctx;
+       unsigned long flags;
+       enum hrtimer_restart ret = HRTIMER_NORESTART;
+
+       /*
+        * prevent against race with unload
+        */
+       ctx  = __get_cpu_var(pmu_ctx);
+       if (!ctx)
+               return HRTIMER_NORESTART;
+
+       spin_lock_irqsave(&ctx->lock, flags);
+
+       set = ctx->active_set;
+
+       /*
+        * switching occurs only when context is attached
+        */
+       if (ctx->state != PFM_CTX_LOADED)
+               goto done;
+       /*
+        * timer does not run while monitoring is inactive (not started)
+        */
+       if (!pfm_arch_is_active(ctx))
+               goto done;
+
+       pfm_stats_inc(handle_timeout_count);
+
+       ret  = pfm_switch_sets(ctx, NULL, PFM_PMD_RESET_SHORT, 0);
+done:
+       spin_unlock_irqrestore(&ctx->lock, flags);
+       return ret;
+}
+
+/*
+ *
+ * always operating on the current task
+ * interrupts are masked
+ *
+ * input:
+ *     - new_set: new set to switch to, if NULL follow normal chain
+ */
+enum hrtimer_restart pfm_switch_sets(struct pfm_context *ctx,
+                                    struct pfm_event_set *new_set,
+                                    int reset_mode,
+                                    int no_restart)
+{
+       struct pfm_event_set *set;
+       u64 now, end;
+       u32 new_flags;
+       int is_system, is_active, nn;
+       enum hrtimer_restart ret = HRTIMER_NORESTART;
+
+       now = sched_clock();
+       set = ctx->active_set;
+       is_active = pfm_arch_is_active(ctx);
+
+       /*
+        * if no set is explicitly requested,
+        * use the set_switch_next field
+        */
+       if (!new_set) {
+               /*
+                * we use round-robin unless the user specified
+                * a particular set to go to.
+                */
+               new_set = list_first_entry(&set->list, struct pfm_event_set, list);
+               if (&new_set->list == &ctx->set_list)
+                       new_set = list_first_entry(&ctx->set_list, struct pfm_event_set, list);
+       }
+
+       PFM_DBG_ovfl("state=%d act=%d cur_set=%u cur_runs=%llu cur_npend=%d next_set=%u "
+                 "next_runs=%llu new_npend=%d reset_mode=%d reset_pmds=%llx",
+                 ctx->state,
+                 is_active,
+                 set->id,
+                 (unsigned long long)set->runs,
+                 set->npend_ovfls,
+                 new_set->id,
+                 (unsigned long long)new_set->runs,
+                 new_set->npend_ovfls,
+                 reset_mode,
+                 (unsigned long long)new_set->reset_pmds[0]);
+
+       is_system = ctx->flags.system;
+       new_flags = new_set->flags;
+
+       /*
+        * nothing more to do
+        */
+       if (new_set == set)
+               goto skip_same_set;
+
+       if (is_active) {
+               pfm_arch_stop(current, ctx);
+               pfm_save_pmds(ctx, set);
+               /*
+                * compute elapsed ns for active set
+                */
+               set->duration += now - set->duration_start;
+       }
+
+       pfm_arch_restore_pmds(ctx, new_set);
+       /*
+        * if masked, we must restore the pmcs such that they
+        * do not capture anything.
+        */
+       pfm_arch_restore_pmcs(ctx, new_set);
+
+       if (new_set->npend_ovfls) {
+               pfm_arch_resend_irq(ctx);
+               pfm_stats_inc(ovfl_intr_replay_count);
+       }
+
+       new_set->priv_flags &= ~PFM_SETFL_PRIV_MOD_BOTH;
+
+skip_same_set:
+       new_set->runs++;
+       /*
+        * reset switch threshold
+        */
+       if (new_flags & PFM_SETFL_OVFL_SWITCH)
+               pfm_reload_switch_thresholds(ctx, new_set);
+
+       /*
+        * reset overflowed PMD registers in new set
+        */
+       nn = bitmap_weight(cast_ulp(new_set->reset_pmds), ctx->regs.max_pmd);
+       if (nn)
+               pfm_reset_pmds(ctx, new_set, nn, reset_mode);
+
+
+       /*
+        * This is needed when coming from pfm_start()
+        *
+        * When switching to the same set, there is no
+        * need to restart
+        */
+       if (no_restart)
+               goto skip_restart;
+
+       if (is_active) {
+               /*
+                * do not need to restart when same set
+                */
+               if (new_set != set) {
+                       ctx->active_set = new_set;
+                       new_set->duration_start = now;
+                       pfm_arch_start(current, ctx);
+               }
+               /*
+                * install new timeout if necessary
+                */
+               if (new_flags & PFM_SETFL_TIME_SWITCH) {
+                       struct hrtimer *h;
+                       h = &__get_cpu_var(pfm_hrtimer);
+                       hrtimer_forward(h, h->base->get_time(), new_set->hrtimer_exp);
+                       new_set->hrtimer_rem = new_set->hrtimer_exp;
+                       ret = HRTIMER_RESTART;
+               }
+       }
+
+skip_restart:
+       ctx->active_set = new_set;
+
+       end = sched_clock();
+
+       pfm_stats_inc(set_switch_count);
+       pfm_stats_add(set_switch_ns, end - now);
+
+       return ret;
+}
+
+/*
+ * called from __pfm_overflow_handler() to switch event sets.
+ * monitoring is stopped, task is current, interrupts are masked.
+ * compared to pfm_switch_sets(), this version is simplified because
+ * it knows about the call path. There is no need to stop monitoring
+ * because it is already frozen by PMU handler.
+ */
+void pfm_switch_sets_from_intr(struct pfm_context *ctx)
+{
+       struct pfm_event_set *set, *new_set;
+       u64 now, end;
+       u32 new_flags;
+       int is_system, n;
+
+       now = sched_clock();
+       set = ctx->active_set;
+       new_set = list_first_entry(&set->list, struct pfm_event_set, list);
+       if (&new_set->list == &ctx->set_list)
+               new_set = list_first_entry(&ctx->set_list, struct pfm_event_set, list);
+
+       PFM_DBG_ovfl("state=%d cur_set=%u cur_runs=%llu cur_npend=%d next_set=%u "
+                 "next_runs=%llu new_npend=%d new_r_pmds=%llx",
+                 ctx->state,
+                 set->id,
+                 (unsigned long long)set->runs,
+                 set->npend_ovfls,
+                 new_set->id,
+                 (unsigned long long)new_set->runs,
+                 new_set->npend_ovfls,
+                 (unsigned long long)new_set->reset_pmds[0]);
+
+       is_system = ctx->flags.system;
+       new_flags = new_set->flags;
+
+       /*
+        * nothing more to do
+        */
+       if (new_set == set)
+               goto skip_same_set;
+
+       /*
+        * switch on intr only when set has OVFL_SWITCH
+        */
+       BUG_ON(set->flags & PFM_SETFL_TIME_SWITCH);
+
+       /*
+        * when called from PMU intr handler, monitoring
+        * is already stopped
+        *
+        * save current PMD registers, we use a special
+        * form for performance reason. On some architectures,
+        * such as x86, the pmds are already saved when entering
+        * the PMU interrupt handler via pfm-arch_intr_freeze()
+        * so we don't need to save them again. On the contrary,
+        * on IA-64, they are not saved by freeze, thus we have to
+        * to it here.
+        */
+       pfm_arch_save_pmds_from_intr(ctx, set);
+
+       /*
+        * compute elapsed ns for active set
+        */
+       set->duration += now - set->duration_start;
+
+       pfm_arch_restore_pmds(ctx, new_set);
+
+       /*
+        * must not be restored active as we are still executing in the
+        * PMU interrupt handler. activation is deferred to unfreeze PMU
+        */
+       pfm_arch_restore_pmcs(ctx, new_set);
+
+       /*
+        * check for pending interrupt on incoming set.
+        * interrupts are masked so handler call deferred
+        */
+       if (new_set->npend_ovfls) {
+               pfm_arch_resend_irq(ctx);
+               pfm_stats_inc(ovfl_intr_replay_count);
+       }
+       /*
+        * no need to restore anything, that is already done
+        */
+       new_set->priv_flags &= ~PFM_SETFL_PRIV_MOD_BOTH;
+       /*
+        * reset duration counter
+        */
+       new_set->duration_start = now;
+
+skip_same_set:
+       new_set->runs++;
+
+       /*
+        * reset switch threshold
+        */
+       if (new_flags & PFM_SETFL_OVFL_SWITCH)
+               pfm_reload_switch_thresholds(ctx, new_set);
+
+       /*
+        * reset overflowed PMD registers
+        */
+       n = bitmap_weight(cast_ulp(new_set->reset_pmds), ctx->regs.max_pmd);
+       if (n)
+               pfm_reset_pmds(ctx, new_set, n, PFM_PMD_RESET_SHORT);
+
+       /*
+        * XXX: isactive?
+        *
+        * Came here following a interrupt which triggered a switch, i.e.,
+        * previous set was using OVFL_SWITCH, thus we just need to arm
+        * check if the next set is using timeout, and if so arm the timer.
+        *
+        * Timeout is always at least one tick away. No risk of having to
+        * invoke the timeout handler right now. In any case, cb_mode is
+        * set to HRTIMER_CB_IRQSAFE_NO_SOFTIRQ such that hrtimer_start
+        * will not try to wakeup the softirqd which could cause a locking
+        * problem.
+        */
+       if (new_flags & PFM_SETFL_TIME_SWITCH) {
+               hrtimer_start(&__get_cpu_var(pfm_hrtimer), set->hrtimer_exp, HRTIMER_MODE_REL);
+               PFM_DBG("armed new timeout for set%u", new_set->id);
+       }
+
+       ctx->active_set = new_set;
+
+       end = sched_clock();
+
+       pfm_stats_inc(set_switch_count);
+       pfm_stats_add(set_switch_ns, end - now);
+}
+
+
+static int pfm_setfl_sane(struct pfm_context *ctx, u32 flags)
+{
+#define PFM_SETFL_BOTH_SWITCH  (PFM_SETFL_OVFL_SWITCH|PFM_SETFL_TIME_SWITCH)
+       int ret;
+
+       ret = pfm_arch_setfl_sane(ctx, flags);
+       if (ret)
+               return ret;
+
+       if ((flags & PFM_SETFL_BOTH_SWITCH) == PFM_SETFL_BOTH_SWITCH) {
+               PFM_DBG("both switch ovfl and switch time are set");
+               return -EINVAL;
+       }
+       return 0;
+}
+
+/*
+ * it is never possible to change the identification of an existing set
+ */
+static int pfm_change_evtset(struct pfm_context *ctx,
+                              struct pfm_event_set *set,
+                              struct pfarg_setdesc *req)
+{
+       struct timeval tv;
+       struct timespec ts;
+       ktime_t kt;
+       long d, res_ns;
+       s32 rem;
+       u32 flags;
+       int ret;
+       u16 set_id;
+
+       BUG_ON(ctx->state == PFM_CTX_LOADED);
+
+       set_id = req->set_id;
+       flags = req->set_flags;
+
+       ret = pfm_setfl_sane(ctx, flags);
+       if (ret) {
+               PFM_DBG("invalid flags 0x%x set %u", flags, set_id);
+               return -EINVAL;
+       }
+
+       /*
+        * compute timeout value
+        */
+       if (flags & PFM_SETFL_TIME_SWITCH) {
+               /*
+                * timeout value of zero is illegal
+                */
+               if (req->set_timeout == 0) {
+                       PFM_DBG("invalid timeout 0");
+                       return -EINVAL;
+               }
+
+               hrtimer_get_res(CLOCK_MONOTONIC, &ts);
+               res_ns = (long)ktime_to_ns(timespec_to_ktime(ts));
+
+               /*
+                * round-up to multiple of clock resolution
+                * timeout = ((req->set_timeout+res_ns-1)/res_ns)*res_ns;
+                *
+                * u64 division missing on 32-bit arch, so use div_s64_rem
+                */
+               d = div_s64_rem(req->set_timeout, res_ns, &rem);
+
+               PFM_DBG("set%u flags=0x%x req_timeout=%lluns "
+                               "HZ=%u TICK_NSEC=%lu clock_res=%ldns rem=%dns",
+                               set_id,
+                               flags,
+                               (unsigned long long)req->set_timeout,
+                               HZ, TICK_NSEC,
+                               res_ns,
+                               rem);
+
+               /*
+                * Only accept timeout, we can actually achieve.
+                * users can invoke clock_getres(CLOCK_MONOTONIC)
+                * to figure out resolution and adjust timeout
+                */
+               if (rem) {
+                       PFM_DBG("set%u invalid timeout=%llu",
+                               set_id,
+                               (unsigned long long)req->set_timeout);
+                       return -EINVAL;
+               }
+
+               tv = ns_to_timeval(req->set_timeout);
+               kt = timeval_to_ktime(tv);
+               set->hrtimer_exp = kt;
+       } else {
+               set->hrtimer_exp = ktime_set(0, 0);
+       }
+
+       /*
+        * commit changes
+        */
+       set->id = set_id;
+       set->flags = flags;
+       set->priv_flags = 0;
+
+       /*
+        * activation and duration counters are reset as
+        * most likely major things will change in the set
+        */
+       set->runs = 0;
+       set->duration = 0;
+
+       return 0;
+}
+
+/*
+ * this function does not modify the next field
+ */
+static void pfm_initialize_set(struct pfm_context *ctx,
+                              struct pfm_event_set *set)
+{
+       u64 *impl_pmcs;
+       u16 i, max_pmc;
+
+       max_pmc = ctx->regs.max_pmc;
+       impl_pmcs =  ctx->regs.pmcs;
+
+       /*
+        * install default values for all PMC  registers
+        */
+       for (i = 0; i < max_pmc; i++) {
+               if (test_bit(i, cast_ulp(impl_pmcs))) {
+                       set->pmcs[i] = pfm_pmu_conf->pmc_desc[i].dfl_val;
+                       PFM_DBG("set%u pmc%u=0x%llx",
+                               set->id,
+                               i,
+                               (unsigned long long)set->pmcs[i]);
+               }
+       }
+
+       /*
+        * PMD registers are set to 0 when the event set is allocated,
+        * hence we do not need to explicitly initialize them.
+        *
+        * For virtual PMD registers (i.e., those tied to a SW resource)
+        * their value becomes meaningful once the context is attached.
+        */
+}
+
+/*
+ * look for an event set using its identification. If the set does not
+ * exist:
+ *     - if alloc == 0 then return error
+ *     - if alloc == 1  then allocate set
+ *
+ * alloc is one ONLY when coming from pfm_create_evtsets() which can only
+ * be called when the context is detached, i.e. monitoring is stopped.
+ */
+struct pfm_event_set *pfm_find_set(struct pfm_context *ctx, u16 set_id, int alloc)
+{
+       struct pfm_event_set *set = NULL, *prev, *new_set;
+
+       PFM_DBG("looking for set=%u", set_id);
+
+       prev = NULL;
+       list_for_each_entry(set, &ctx->set_list, list) {
+               if (set->id == set_id)
+                       return set;
+               if (set->id > set_id)
+                       break;
+               prev = set;
+       }
+
+       if (!alloc)
+               return NULL;
+
+       /*
+        * we are holding the context spinlock and interrupts
+        * are unmasked. We must use GFP_ATOMIC as we cannot
+        * sleep while holding a spin lock.
+        */
+       new_set = kmem_cache_zalloc(pfm_set_cachep, GFP_ATOMIC);
+       if (!new_set)
+               return NULL;
+
+       new_set->id = set_id;
+
+       INIT_LIST_HEAD(&new_set->list);
+
+       if (prev == NULL) {
+               list_add(&(new_set->list), &ctx->set_list);
+       } else {
+               PFM_DBG("add after set=%u", prev->id);
+               list_add(&(new_set->list), &prev->list);
+       }
+       return new_set;
+}
+
+/**
+ * pfm_create_initial_set - create initial set from __pfm_c reate_context
+ * @ctx: context to atatched the set to
+ */
+int pfm_create_initial_set(struct pfm_context *ctx)
+{
+       struct pfm_event_set *set;
+
+       /*
+        * create initial set0
+        */
+       if (!pfm_find_set(ctx, 0, 1))
+               return -ENOMEM;
+
+       set = list_first_entry(&ctx->set_list, struct pfm_event_set, list);
+
+       pfm_initialize_set(ctx, set);
+
+       return 0;
+}
+
+/*
+ * context is unloaded for this command. Interrupts are enabled
+ */
+int __pfm_create_evtsets(struct pfm_context *ctx, struct pfarg_setdesc *req,
+                       int count)
+{
+       struct pfm_event_set *set;
+       u16 set_id;
+       int i, ret;
+
+       for (i = 0; i < count; i++, req++) {
+               set_id = req->set_id;
+
+               PFM_DBG("set_id=%u", set_id);
+
+               set = pfm_find_set(ctx, set_id, 1);
+               if (set == NULL)
+                       goto error_mem;
+
+               ret = pfm_change_evtset(ctx, set, req);
+               if (ret)
+                       goto error_params;
+
+               pfm_initialize_set(ctx, set);
+       }
+       return 0;
+error_mem:
+       PFM_DBG("cannot allocate set %u", set_id);
+       return -ENOMEM;
+error_params:
+       return ret;
+}
+
+int __pfm_getinfo_evtsets(struct pfm_context *ctx, struct pfarg_setinfo *req,
+                                int count)
+{
+       struct pfm_event_set *set;
+       int i, is_system, is_loaded, is_self, ret;
+       u16 set_id;
+       u64 end;
+
+       end = sched_clock();
+
+       is_system = ctx->flags.system;
+       is_loaded = ctx->state == PFM_CTX_LOADED;
+       is_self   = ctx->task == current || is_system;
+
+       ret = -EINVAL;
+       for (i = 0; i < count; i++, req++) {
+
+               set_id = req->set_id;
+
+               list_for_each_entry(set, &ctx->set_list, list) {
+                       if (set->id == set_id)
+                               goto found;
+                       if (set->id > set_id)
+                               goto error;
+               }
+found:
+               req->set_flags = set->flags;
+
+               /*
+                * compute leftover timeout
+                *
+                * lockdep may complain about lock inversion
+                * because of get_remaining() however, this
+                * applies to self-montoring only, thus the
+                * thread cannot be in the timeout handler
+                * and here at the same time given that we
+                * run with interrupts disabled
+                */
+               if (is_loaded && is_self) {
+                       struct hrtimer *h;
+                       h = &__get_cpu_var(pfm_hrtimer);
+                       req->set_timeout = ktime_to_ns(hrtimer_get_remaining(h));
+               } else {
+                       /*
+                        * hrtimer_rem zero when not using
+                        * timeout-based switching
+                        */
+                       req->set_timeout = ktime_to_ns(set->hrtimer_rem);
+               }
+
+               req->set_runs = set->runs;
+               req->set_act_duration = set->duration;
+
+               /*
+                * adjust for active set if needed
+                */
+               if (is_system && is_loaded && ctx->flags.started
+                   && set == ctx->active_set)
+                       req->set_act_duration  += end - set->duration_start;
+
+               /*
+                * copy the list of pmds which last overflowed
+                */
+               bitmap_copy(cast_ulp(req->set_ovfl_pmds),
+                           cast_ulp(set->ovfl_pmds),
+                           PFM_MAX_PMDS);
+
+               /*
+                * copy bitmask of available PMU registers
+                *
+                * must copy over the entire vector to avoid
+                * returning bogus upper bits pass by user
+                */
+               bitmap_copy(cast_ulp(req->set_avail_pmcs),
+                           cast_ulp(ctx->regs.pmcs),
+                           PFM_MAX_PMCS);
+
+               bitmap_copy(cast_ulp(req->set_avail_pmds),
+                           cast_ulp(ctx->regs.pmds),
+                           PFM_MAX_PMDS);
+
+               PFM_DBG("set%u flags=0x%x eff_usec=%llu runs=%llu "
+                       "a_pmcs=0x%llx a_pmds=0x%llx",
+                       set_id,
+                       set->flags,
+                       (unsigned long long)req->set_timeout,
+                       (unsigned long long)set->runs,
+                       (unsigned long long)ctx->regs.pmcs[0],
+                       (unsigned long long)ctx->regs.pmds[0]);
+       }
+       ret = 0;
+error:
+       return ret;
+}
+
+/*
+ * context is unloaded for this command. Interrupts are enabled
+ */
+int __pfm_delete_evtsets(struct pfm_context *ctx, void *arg, int count)
+{
+       struct pfarg_setdesc *req = arg;
+       struct pfm_event_set *set;
+       u16 set_id;
+       int i, ret;
+
+       ret = -EINVAL;
+       for (i = 0; i < count; i++, req++) {
+               set_id = req->set_id;
+
+               list_for_each_entry(set, &ctx->set_list, list) {
+                       if (set->id == set_id)
+                               goto found;
+                       if (set->id > set_id)
+                               goto error;
+               }
+               goto error;
+found:
+               /*
+                * clear active set if necessary.
+                * will be updated when context is loaded
+                */
+               if (set == ctx->active_set)
+                       ctx->active_set = NULL;
+
+               list_del(&set->list);
+
+               kmem_cache_free(pfm_set_cachep, set);
+
+               PFM_DBG("set%u deleted", set_id);
+       }
+       ret = 0;
+error:
+       return ret;
+}
+
+/*
+ * called from pfm_context_free() to free all sets
+ */
+void pfm_free_sets(struct pfm_context *ctx)
+{
+       struct pfm_event_set *set, *tmp;
+
+       list_for_each_entry_safe(set, tmp, &ctx->set_list, list) {
+               list_del(&set->list);
+               kmem_cache_free(pfm_set_cachep, set);
+       }
+}
+
+/**
+ * pfm_restart_timer - restart hrtimer taking care of expired timeout
+ * @ctx : context to work with
+ * @set : current active set
+ *
+ * Must be called on the processor on which the timer is to be armed.
+ * Assumes context is locked and interrupts are masked
+ *
+ * Upon return the active set for the context may have changed
+ */
+void pfm_restart_timer(struct pfm_context *ctx, struct pfm_event_set *set)
+{
+       struct hrtimer *h;
+       enum hrtimer_restart ret;
+
+       h = &__get_cpu_var(pfm_hrtimer);
+
+       PFM_DBG_ovfl("hrtimer=%lld", (long long)ktime_to_ns(set->hrtimer_rem));
+
+       if (ktime_to_ns(set->hrtimer_rem) > 0) {
+               hrtimer_start(h, set->hrtimer_rem, HRTIMER_MODE_REL);
+       } else {
+               /*
+                * timer was not re-armed because it has already expired
+                * timer was not enqueued, we need to switch set now
+                */
+               pfm_stats_inc(set_switch_exp);
+
+               ret = pfm_switch_sets(ctx, NULL, 1, 0);
+               set = ctx->active_set;
+               if (ret == HRTIMER_RESTART)
+                       hrtimer_start(h, set->hrtimer_rem, HRTIMER_MODE_REL);
+       }
+}
+
+int __init pfm_init_sets(void)
+{
+       pfm_set_cachep = kmem_cache_create("pfm_event_set",
+                                          sizeof(struct pfm_event_set),
+                                          SLAB_HWCACHE_ALIGN, 0, NULL);
+       if (!pfm_set_cachep) {
+               PFM_ERR("cannot initialize event set slab");
+               return -ENOMEM;
+       }
+       return 0;
+}
diff --git a/perfmon/perfmon_smpl.c b/perfmon/perfmon_smpl.c
new file mode 100644
index 0000000..e31fb15
--- /dev/null
+++ b/perfmon/perfmon_smpl.c
@@ -0,0 +1,865 @@
+/*
+ * perfmon_smpl.c: perfmon2 sampling management
+ *
+ * This file implements the perfmon2 interface which
+ * provides access to the hardware performance counters
+ * of the host processor.
+ *
+ *
+ * The initial version of perfmon.c was written by
+ * Ganesh Venkitachalam, IBM Corp.
+ *
+ * Then it was modified for perfmon-1.x by Stephane Eranian and
+ * David Mosberger, Hewlett Packard Co.
+ *
+ * Version Perfmon-2.x is a complete rewrite of perfmon-1.x
+ * by Stephane Eranian, Hewlett Packard Co.
+ *
+ * Copyright (c) 1999-2006 Hewlett-Packard Development Company, L.P.
+ * Contributed by Stephane Eranian <eranian@hpl.hp.com>
+ *                David Mosberger-Tang <davidm@hpl.hp.com>
+ *
+ * More information about perfmon available at:
+ *     http://perfmon2.sf.net
+ *
+ * This program is free software; you can redistribute it and/or
+ * modify it under the terms of version 2 of the GNU General Public
+ * License as published by the Free Software Foundation.
+ *
+ * This program is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+ * General Public License for more details.
+ *
+ * You should have received a copy of the GNU General Public License
+ * along with this program; if not, write to the Free Software
+ * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA
+ * 02111-1307 USA
+ */
+#include <linux/module.h>
+#include <linux/kernel.h>
+#include <linux/vmalloc.h>
+#include <linux/fs.h>
+#include <linux/mm.h>
+#include <linux/random.h>
+#include <linux/uaccess.h>
+#include <linux/perfmon_kern.h>
+
+#include "perfmon_priv.h"
+
+/**
+ * pfm_smpl_buf_alloc - allocate memory for sampling buffer
+ * @ctx: context to operate on
+ * @rsize: requested size
+ *
+ * called from pfm_smpl_buffer_alloc_old() (IA64-COMPAT)
+ * and pfm_setup_smpl_fmt()
+ *
+ * interrupts are enabled, context is not locked.
+ *
+ * function is not static because it is called from the IA-64
+ * compatibility module (perfmon_compat.c)
+ */
+int pfm_smpl_buf_alloc(struct pfm_context *ctx, size_t rsize)
+{
+#if PFM_ARCH_SMPL_ALIGN_SIZE > 0
+#define PFM_ALIGN_SMPL(a, f) (void *)((((unsigned long)(a))+(f-1)) & ~(f-1))
+#else
+#define PFM_ALIGN_SMPL(a, f) (a)
+#endif
+       void *addr, *real_addr;
+       size_t size, real_size;
+       int ret;
+
+       might_sleep();
+
+       /*
+        * align page boundary
+        */
+       size = PAGE_ALIGN(rsize);
+
+       /*
+        * On some arch, it may be necessary to get an alignment greater
+        * than page size to avoid certain cache effects (e.g., MIPS).
+        * This is the reason for PFM_ARCH_SMPL_ALIGN_SIZE.
+        */
+       real_size = size + PFM_ARCH_SMPL_ALIGN_SIZE;
+
+       PFM_DBG("req_size=%zu size=%zu real_size=%zu",
+               rsize,
+               size,
+               real_size);
+
+       ret = pfm_smpl_buf_space_acquire(ctx, real_size);
+       if (ret)
+               return ret;
+
+       /*
+        * vmalloc can sleep. we do not hold
+        * any spinlock and interrupts are enabled
+        */
+       real_addr = addr = vmalloc(real_size);
+       if (!real_addr) {
+               PFM_DBG("cannot allocate sampling buffer");
+               goto unres;
+       }
+
+       /*
+        * align the useable sampling buffer address to the arch requirement
+        * This is a nop on most architectures
+        */
+       addr = PFM_ALIGN_SMPL(real_addr, PFM_ARCH_SMPL_ALIGN_SIZE);
+
+       memset(addr, 0, real_size);
+
+       /*
+        * due to cache aliasing, it may be necessary to flush the pages
+        * on certain architectures (e.g., MIPS)
+        */
+       pfm_cacheflush(addr, real_size);
+
+       /*
+        * what needs to be freed
+        */
+       ctx->smpl_real_addr = real_addr;
+       ctx->smpl_real_size = real_size;
+
+       /*
+        * what is actually available to user
+        */
+       ctx->smpl_addr = addr;
+       ctx->smpl_size = size;
+
+       PFM_DBG("addr=%p real_addr=%p", addr, real_addr);
+
+       return 0;
+unres:
+       /*
+        * smpl_addr is NULL, no double freeing possible in pfm_context_free()
+        */
+       pfm_smpl_buf_space_release(ctx, real_size);
+
+       return -ENOMEM;
+}
+
+/**
+ * pfm_smpl_buf_free - free resources associated with sampling
+ * @ctx: context to operate on
+ */
+void pfm_smpl_buf_free(struct pfm_context *ctx)
+{
+       struct pfm_smpl_fmt *fmt;
+
+       fmt = ctx->smpl_fmt;
+
+       /*
+        * some formats may not use a buffer, yet they may
+        * need to be called on exit
+        */
+       if (fmt) {
+               if (fmt->fmt_exit)
+                       (*fmt->fmt_exit)(ctx->smpl_addr);
+               /*
+                * decrease refcount of sampling format
+                */
+               pfm_smpl_fmt_put(fmt);
+       }
+
+       if (ctx->smpl_addr) {
+               pfm_smpl_buf_space_release(ctx, ctx->smpl_real_size);
+
+               PFM_DBG("free buffer real_addr=0x%p real_size=%zu",
+                       ctx->smpl_real_addr,
+                       ctx->smpl_real_size);
+
+               vfree(ctx->smpl_real_addr);
+       }
+}
+
+/**
+ * pfm_setup_smpl_fmt - initialization of sampling format and buffer
+ * @ctx: context to operate on
+ * @fmt_arg: smapling format arguments
+ * @ctx_flags: context flags as passed by user
+ * @filp: file descriptor associated with context
+ *
+ * called from __pfm_create_context()
+ */
+int pfm_setup_smpl_fmt(struct pfm_context *ctx, u32 ctx_flags, void *fmt_arg,
+                      struct file *filp)
+{
+       struct pfm_smpl_fmt *fmt;
+       size_t size = 0;
+       int ret = 0;
+
+       fmt = ctx->smpl_fmt;
+
+       /*
+        * validate parameters
+        */
+       if (fmt->fmt_validate) {
+               ret = (*fmt->fmt_validate)(ctx_flags,
+                                          ctx->regs.num_pmds,
+                                          fmt_arg);
+               PFM_DBG("validate(0x%x,%p)=%d", ctx_flags, fmt_arg, ret);
+               if (ret)
+                       goto error;
+       }
+
+       /*
+        * check if buffer format needs buffer allocation
+        */
+       size = 0;
+       if (fmt->fmt_getsize) {
+               ret = (*fmt->fmt_getsize)(ctx_flags, fmt_arg, &size);
+               if (ret) {
+                       PFM_DBG("cannot get size ret=%d", ret);
+                       goto error;
+               }
+       }
+
+       /*
+        * allocate buffer
+        * v20_compat is for IA-64 backward compatibility with perfmon v2.0
+        */
+       if (size) {
+#ifdef CONFIG_IA64_PERFMON_COMPAT
+               /*
+                * backward compatibility with perfmon v2.0 on Ia-64
+                */
+               if (ctx->flags.ia64_v20_compat)
+                       ret = pfm_smpl_buf_alloc_compat(ctx, size, filp);
+               else
+#endif
+                       ret = pfm_smpl_buf_alloc(ctx, size);
+
+               if (ret)
+                       goto error;
+
+       }
+
+       if (fmt->fmt_init) {
+               ret = (*fmt->fmt_init)(ctx, ctx->smpl_addr, ctx_flags,
+                                      ctx->regs.num_pmds,
+                                      fmt_arg);
+       }
+       /*
+        * if there was an error, the buffer/resource will be freed by
+        * via pfm_context_free()
+        */
+error:
+       return ret;
+}
+
+void pfm_mask_monitoring(struct pfm_context *ctx, struct pfm_event_set *set)
+{
+       u64 now;
+
+       now = sched_clock();
+
+       /*
+        * we save the PMD values such that we can read them while
+        * MASKED without having the thread stopped
+        * because monitoring is stopped
+        *
+        * pfm_save_pmds() could be avoided if we knew
+        * that pfm_arch_intr_freeze() had saved them already
+        */
+       pfm_save_pmds(ctx, set);
+       pfm_arch_mask_monitoring(ctx, set);
+       /*
+        * accumulate the set duration up to this point
+        */
+       set->duration += now - set->duration_start;
+
+       ctx->state = PFM_CTX_MASKED;
+
+       /*
+        * need to stop timer and remember remaining time
+        * will be reloaded in pfm_unmask_monitoring
+        * hrtimer is cancelled in the tail of the interrupt
+        * handler once the context is unlocked
+        */
+       if (set->flags & PFM_SETFL_TIME_SWITCH) {
+               struct hrtimer *h = &__get_cpu_var(pfm_hrtimer);
+               hrtimer_cancel(h);
+               set->hrtimer_rem = hrtimer_get_remaining(h);
+       }
+       PFM_DBG_ovfl("can_restart=%u", ctx->flags.can_restart);
+}
+
+/**
+ * pfm_unmask_monitoring - unmask monitoring
+ * @ctx: context to work with
+ * @set: current active set
+ *
+ * interrupts are masked when entering this function.
+ * context must be in MASKED state when calling.
+ *
+ * Upon return, the active set may have changed when using timeout
+ * based switching.
+ */
+static void pfm_unmask_monitoring(struct pfm_context *ctx, struct pfm_event_set *set)
+{
+       if (ctx->state != PFM_CTX_MASKED)
+               return;
+
+       PFM_DBG_ovfl("unmasking monitoring");
+
+       /*
+        * must be done before calling
+        * pfm_arch_unmask_monitoring()
+        */
+       ctx->state = PFM_CTX_LOADED;
+
+       /*
+        * we need to restore the PMDs because they
+        * may have been modified by user while MASKED in
+        * which case the actual registers have no yet
+        * been updated
+        */
+       pfm_arch_restore_pmds(ctx, set);
+
+       /*
+        * call arch specific handler
+        */
+       pfm_arch_unmask_monitoring(ctx, set);
+
+       /*
+        * clear force reload flag. May have been set
+        * in pfm_write_pmcs or pfm_write_pmds
+        */
+       set->priv_flags &= ~PFM_SETFL_PRIV_MOD_BOTH;
+
+       /*
+        * reset set duration timer
+        */
+       set->duration_start = sched_clock();
+
+       /*
+        * restart hrtimer if needed
+        */
+       if (set->flags & PFM_SETFL_TIME_SWITCH) {
+               pfm_restart_timer(ctx, set);
+               /* careful here as pfm_restart_timer may switch sets */
+       }
+}
+
+void pfm_reset_pmds(struct pfm_context *ctx,
+                   struct pfm_event_set *set,
+                   int num_pmds,
+                   int reset_mode)
+{
+       u64 val, mask, new_seed;
+       struct pfm_pmd *reg;
+       unsigned int i, not_masked;
+
+       not_masked = ctx->state != PFM_CTX_MASKED;
+
+       PFM_DBG_ovfl("%s r_pmds=0x%llx not_masked=%d",
+               reset_mode == PFM_PMD_RESET_LONG ? "long" : "short",
+               (unsigned long long)set->reset_pmds[0],
+               not_masked);
+
+       pfm_stats_inc(reset_pmds_count);
+
+       for (i = 0; num_pmds; i++) {
+               if (test_bit(i, cast_ulp(set->reset_pmds))) {
+                       num_pmds--;
+
+                       reg = set->pmds + i;
+
+                       val = reset_mode == PFM_PMD_RESET_LONG ?
+                              reg->long_reset : reg->short_reset;
+
+                       if (reg->flags & PFM_REGFL_RANDOM) {
+                               mask = reg->mask;
+                               new_seed = random32();
+
+                               /* construct a full 64-bit random value: */
+                               if ((unlikely(mask >> 32) != 0))
+                                       new_seed |= (u64)random32() << 32;
+
+                               /* counter values are negative numbers! */
+                               val -= (new_seed & mask);
+                       }
+
+                       set->pmds[i].value = val;
+                       reg->lval = val;
+
+                       /*
+                        * not all PMD to reset are necessarily
+                        * counters
+                        */
+                       if (not_masked)
+                               pfm_write_pmd(ctx, i, val);
+
+                       PFM_DBG_ovfl("set%u pmd%u sval=0x%llx",
+                                       set->id,
+                                       i,
+                                       (unsigned long long)val);
+               }
+       }
+
+       /*
+        * done with reset
+        */
+       bitmap_zero(cast_ulp(set->reset_pmds), i);
+
+       /*
+        * make changes visible
+        */
+       if (not_masked)
+               pfm_arch_serialize();
+}
+
+/*
+ * called from pfm_handle_work() and __pfm_restart()
+ * for system-wide and per-thread context to resume
+ * monitoring after a user level notification.
+ *
+ * In both cases, the context is locked and interrupts
+ * are disabled.
+ */
+void pfm_resume_after_ovfl(struct pfm_context *ctx)
+{
+       struct pfm_smpl_fmt *fmt;
+       u32 rst_ctrl;
+       struct pfm_event_set *set;
+       u64 *reset_pmds;
+       void *hdr;
+       int state, ret;
+
+       hdr = ctx->smpl_addr;
+       fmt = ctx->smpl_fmt;
+       state = ctx->state;
+       set = ctx->active_set;
+       ret = 0;
+
+       if (hdr) {
+               rst_ctrl = 0;
+               prefetch(hdr);
+       } else {
+               rst_ctrl = PFM_OVFL_CTRL_RESET;
+       }
+
+       /*
+        * if using a sampling buffer format and it has a restart callback,
+        * then invoke it. hdr may be NULL, if the format does not use a
+        * perfmon buffer
+        */
+       if (fmt && fmt->fmt_restart)
+               ret = (*fmt->fmt_restart)(state == PFM_CTX_LOADED, &rst_ctrl,
+                                         hdr);
+
+       reset_pmds = set->reset_pmds;
+
+       PFM_DBG("fmt_restart=%d reset_count=%d set=%u r_pmds=0x%llx switch=%d "
+               "ctx_state=%d",
+               ret,
+               ctx->flags.reset_count,
+               set->id,
+               (unsigned long long)reset_pmds[0],
+               (set->priv_flags & PFM_SETFL_PRIV_SWITCH),
+               state);
+
+       if (!ret) {
+               /*
+                * switch set if needed
+                */
+               if (set->priv_flags & PFM_SETFL_PRIV_SWITCH) {
+                       set->priv_flags &= ~PFM_SETFL_PRIV_SWITCH;
+                       pfm_switch_sets(ctx, NULL, PFM_PMD_RESET_LONG, 0);
+                       set = ctx->active_set;
+               } else if (rst_ctrl & PFM_OVFL_CTRL_RESET) {
+                       int nn;
+                       nn = bitmap_weight(cast_ulp(set->reset_pmds),
+                                          ctx->regs.max_pmd);
+                       if (nn)
+                               pfm_reset_pmds(ctx, set, nn, PFM_PMD_RESET_LONG);
+               }
+
+               if (!(rst_ctrl & PFM_OVFL_CTRL_MASK))
+                       pfm_unmask_monitoring(ctx, set);
+               else
+                       PFM_DBG("stopping monitoring?");
+               ctx->state = PFM_CTX_LOADED;
+       }
+}
+
+/*
+ * This function is called when we need to perform asynchronous
+ * work on a context. This function is called ONLY when about to
+ * return to user mode (very much like with signal handling).
+ *
+ * There are several reasons why we come here:
+ *
+ *  - per-thread mode, not self-monitoring, to reset the counters
+ *    after a pfm_restart()
+ *
+ *  - we are zombie and we need to cleanup our state
+ *
+ *  - we need to block after an overflow notification
+ *    on a context with the PFM_OVFL_NOTIFY_BLOCK flag
+ *
+ * This function is never called for a system-wide context.
+ *
+ * pfm_handle_work() can be called with interrupts enabled
+ * (TIF_NEED_RESCHED) or disabled. The down_interruptible
+ * call may sleep, therefore we must re-enable interrupts
+ * to avoid deadlocks. It is safe to do so because this function
+ * is called ONLY when returning to user level, in which case
+ * there is no risk of kernel stack overflow due to deep
+ * interrupt nesting.
+ */
+void pfm_handle_work(struct pt_regs *regs)
+{
+       struct pfm_context *ctx;
+       unsigned long flags, dummy_flags;
+       int type, ret, info;
+
+#ifdef CONFIG_PPC
+       /*
+        * This is just a temporary fix. Obviously we'd like to fix the powerpc
+        * code to make that check before calling __pfm_handle_work() to
+        * prevent the function call overhead, but the call is made from
+        * assembly code, so it will take a little while to figure out how to
+        * perform the check correctly.
+        */
+       if (!test_thread_flag(TIF_PERFMON_WORK))
+               return;
+#endif
+
+       if (!user_mode(regs))
+               return;
+
+       clear_thread_flag(TIF_PERFMON_WORK);
+
+       pfm_stats_inc(handle_work_count);
+
+       ctx = current->pfm_context;
+       if (ctx == NULL) {
+               PFM_DBG("[%d] has no ctx", current->pid);
+               return;
+       }
+
+       BUG_ON(ctx->flags.system);
+
+       spin_lock_irqsave(&ctx->lock, flags);
+
+       type = ctx->flags.work_type;
+       ctx->flags.work_type = PFM_WORK_NONE;
+
+       PFM_DBG("work_type=%d reset_count=%d",
+               type,
+               ctx->flags.reset_count);
+
+       switch (type) {
+       case PFM_WORK_ZOMBIE:
+               goto do_zombie;
+       case PFM_WORK_RESET:
+               /* simply reset, no blocking */
+               goto skip_blocking;
+       case PFM_WORK_NONE:
+               PFM_DBG("unexpected PFM_WORK_NONE");
+               goto nothing_todo;
+       case PFM_WORK_BLOCK:
+               break;
+       default:
+               PFM_DBG("unkown type=%d", type);
+               goto nothing_todo;
+       }
+
+       /*
+        * restore interrupt mask to what it was on entry.
+        * Could be enabled/disabled.
+        */
+       spin_unlock_irqrestore(&ctx->lock, flags);
+
+       /*
+        * force interrupt enable because of down_interruptible()
+        */
+       local_irq_enable();
+
+       PFM_DBG("before block sleeping");
+
+       /*
+        * may go through without blocking on SMP systems
+        * if restart has been received already by the time we call down()
+        */
+       ret = wait_for_completion_interruptible(&ctx->restart_complete);
+
+       PFM_DBG("after block sleeping ret=%d", ret);
+
+       /*
+        * lock context and mask interrupts again
+        * We save flags into a dummy because we may have
+        * altered interrupts mask compared to entry in this
+        * function.
+        */
+       spin_lock_irqsave(&ctx->lock, dummy_flags);
+
+       if (ctx->state == PFM_CTX_ZOMBIE)
+               goto do_zombie;
+
+       /*
+        * in case of interruption of down() we don't restart anything
+        */
+       if (ret < 0)
+               goto nothing_todo;
+
+skip_blocking:
+       /*
+        * iterate over the number of pending resets
+        * There are certain situations where there may be
+        * multiple notifications sent before a pfm_restart().
+        * As such, it may be that multiple pfm_restart() are
+        * issued before the monitored thread gets to
+        * pfm_handle_work(). To avoid losing restarts, pfm_restart()
+        * increments a counter (reset_counts). Here, we take this
+        * into account by potentially calling pfm_resume_after_ovfl()
+        * multiple times. It is up to the sampling format to take the
+        * appropriate actions.
+        */
+       while (ctx->flags.reset_count) {
+               pfm_resume_after_ovfl(ctx);
+               /* careful as active set may have changed */
+               ctx->flags.reset_count--;
+       }
+
+nothing_todo:
+       /*
+        * restore flags as they were upon entry
+        */
+       spin_unlock_irqrestore(&ctx->lock, flags);
+       return;
+
+do_zombie:
+       PFM_DBG("context is zombie, bailing out");
+
+       __pfm_unload_context(ctx, &info);
+
+       /*
+        * keep the spinlock check happy
+        */
+       spin_unlock(&ctx->lock);
+
+       /*
+        * enable interrupt for vfree()
+        */
+       local_irq_enable();
+
+       /*
+        * cancel timer now that context is unlocked
+        */
+       if (info & 0x2) {
+               ret = hrtimer_cancel(&__get_cpu_var(pfm_hrtimer));
+               PFM_DBG("timeout cancel=%d", ret);
+       }
+
+       /*
+        * actual context free
+        */
+       pfm_free_context(ctx);
+
+       /*
+        * restore interrupts as they were upon entry
+        */
+       local_irq_restore(flags);
+
+       /* always true */
+       if (info & 0x1)
+               pfm_session_release(0, 0);
+}
+
+/**
+ * __pfm_restart - resume monitoring after user-level notification
+ * @ctx: context to operate on
+ * @info: return information used to free resource once unlocked
+ *
+ * function called from sys_pfm_restart(). It is used when overflow
+ * notification is requested. For each notification received, the user
+ * must call pfm_restart() to indicate to the kernel that it is done
+ * processing the notification.
+ *
+ * When the caller is doing user level sampling, this function resets
+ * the overflowed counters and resumes monitoring which is normally stopped
+ * during notification (always the consequence of a counter overflow).
+ *
+ * When using a sampling format, the format restart() callback is invoked,
+ * overflowed PMDS may be reset based upon decision from sampling format.
+ *
+ * When operating in per-thread mode, and when not self-monitoring, the
+ * monitored thread DOES NOT need to be stopped, unlike for many other calls.
+ *
+ * This means that the effect of the restart may not necessarily be observed
+ * right when returning from the call. For instance, counters may not already
+ * be reset in the other thread.
+ *
+ * When operating in system-wide, the caller must be running on the monitored
+ * CPU.
+ *
+ * The context is locked and interrupts are disabled.
+ *
+ * info value upon return:
+ *     - bit 0: when set, mudt issue complete() on restart semaphore
+ */
+int __pfm_restart(struct pfm_context *ctx, int *info)
+{
+       int state;
+
+       state = ctx->state;
+
+       PFM_DBG("state=%d can_restart=%d reset_count=%d",
+               state,
+               ctx->flags.can_restart,
+               ctx->flags.reset_count);
+
+       *info = 0;
+
+       switch (state) {
+       case PFM_CTX_MASKED:
+               break;
+       case PFM_CTX_LOADED:
+               if (ctx->smpl_addr && ctx->smpl_fmt->fmt_restart)
+                       break;
+       default:
+               PFM_DBG("invalid state=%d", state);
+               return -EBUSY;
+       }
+
+       /*
+        * first check if allowed to restart, i.e., notifications received
+        */
+       if (!ctx->flags.can_restart) {
+               PFM_DBG("no restart can_restart=0");
+               return -EBUSY;
+       }
+
+       pfm_stats_inc(pfm_restart_count);
+
+       /*
+        * at this point, the context is either LOADED or MASKED
+        */
+       ctx->flags.can_restart--;
+
+       /*
+        * handle self-monitoring case and system-wide
+        */
+       if (ctx->task == current || ctx->flags.system) {
+               pfm_resume_after_ovfl(ctx);
+               return 0;
+       }
+
+       /*
+        * restart another task
+        */
+
+       /*
+        * if blocking, then post the semaphore if PFM_CTX_MASKED, i.e.
+        * the task is blocked or on its way to block. That's the normal
+        * restart path. If the monitoring is not masked, then the task
+        * can be actively monitoring and we cannot directly intervene.
+        * Therefore we use the trap mechanism to catch the task and
+        * force it to reset the buffer/reset PMDs.
+        *
+        * if non-blocking, then we ensure that the task will go into
+        * pfm_handle_work() before returning to user mode.
+        *
+        * We cannot explicitly reset another task, it MUST always
+        * be done by the task itself. This works for system wide because
+        * the tool that is controlling the session is logically doing
+        * "self-monitoring".
+        */
+       if (ctx->flags.block && state == PFM_CTX_MASKED) {
+               PFM_DBG("unblocking [%d]", ctx->task->pid);
+               /*
+                * It is not possible to call complete() with the context locked
+                * otherwise we have a potential deadlock with the PMU context
+                * switch code due to a lock inversion between task_rq_lock()
+                * and the context lock.
+                * Instead we mark whether or not we need to issue the complete
+                * and we invoke the function once the context lock is released
+                * in sys_pfm_restart()
+                */
+               *info = 1;
+       } else {
+               PFM_DBG("[%d] armed exit trap", ctx->task->pid);
+               pfm_post_work(ctx->task, ctx, PFM_WORK_RESET);
+       }
+       ctx->flags.reset_count++;
+       return 0;
+}
+
+/**
+ * pfm_get_smpl_arg -- copy user arguments to pfm_create_context() related to sampling format
+ * @name: format name as passed by user
+ * @fmt_arg: format optional argument as passed by user
+ * @uszie: size of structure pass in fmt_arg
+ * @arg: kernel copy of fmt_arg
+ * @fmt: pointer to sampling format upon success
+ *
+ * arg is kmalloc'ed, thus it needs a kfree by caller
+ */
+int pfm_get_smpl_arg(char __user *fmt_uname, void __user *fmt_uarg, size_t usize, void **arg,
+                    struct pfm_smpl_fmt **fmt)
+{
+       struct pfm_smpl_fmt *f;
+       char *fmt_name;
+       void *addr = NULL;
+       size_t sz;
+       int ret;
+
+       fmt_name = getname(fmt_uname);
+       if (!fmt_name) {
+               PFM_DBG("getname failed");
+               return -ENOMEM;
+       }
+
+       /*
+        * find fmt and increase refcount
+        */
+       f = pfm_smpl_fmt_get(fmt_name);
+
+       putname(fmt_name);
+
+       if (f == NULL) {
+               PFM_DBG("buffer format not found");
+               return -EINVAL;
+       }
+
+       /*
+        * expected format argument size
+        */
+       sz = f->fmt_arg_size;
+
+       /*
+        * check user size matches expected size
+        * usize = -1 is for IA-64 backward compatibility
+        */
+       ret = -EINVAL;
+       if (sz != usize && usize != -1) {
+               PFM_DBG("invalid arg size %zu, format expects %zu",
+                       usize, sz);
+               goto error;
+       }
+
+       if (sz) {
+               ret = -ENOMEM;
+               addr = kmalloc(sz, GFP_KERNEL);
+               if (addr == NULL)
+                       goto error;
+
+               ret = -EFAULT;
+               if (copy_from_user(addr, fmt_uarg, sz))
+                       goto error;
+       }
+       *arg = addr;
+       *fmt = f;
+       return 0;
+
+error:
+       kfree(addr);
+       pfm_smpl_fmt_put(f);
+       return ret;
+}
diff --git a/perfmon/perfmon_syscalls.c b/perfmon/perfmon_syscalls.c
new file mode 100644
index 0000000..8777b58
--- /dev/null
+++ b/perfmon/perfmon_syscalls.c
@@ -0,0 +1,1060 @@
+/*
+ * perfmon_syscalls.c: perfmon2 system call interface
+ *
+ * This file implements the perfmon2 interface which
+ * provides access to the hardware performance counters
+ * of the host processor.
+ *
+ * The initial version of perfmon.c was written by
+ * Ganesh Venkitachalam, IBM Corp.
+ *
+ * Then it was modified for perfmon-1.x by Stephane Eranian and
+ * David Mosberger, Hewlett Packard Co.
+ *
+ * Version Perfmon-2.x is a complete rewrite of perfmon-1.x
+ * by Stephane Eranian, Hewlett Packard Co.
+ *
+ * Copyright (c) 1999-2006 Hewlett-Packard Development Company, L.P.
+ * Contributed by Stephane Eranian <eranian@hpl.hp.com>
+ *                David Mosberger-Tang <davidm@hpl.hp.com>
+ *
+ * More information about perfmon available at:
+ *     http://perfmon2.sf.net
+ *
+ * This program is free software; you can redistribute it and/or
+ * modify it under the terms of version 2 of the GNU General Public
+ * License as published by the Free Software Foundation.
+ *
+ * This program is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+ * General Public License for more details.
+ *
+ * You should have received a copy of the GNU General Public License
+ * along with this program; if not, write to the Free Software
+ * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA
+ * 02111-1307 USA
+ */
+#include <linux/kernel.h>
+#include <linux/fs.h>
+#include <linux/ptrace.h>
+#include <linux/perfmon_kern.h>
+#include <linux/uaccess.h>
+#include "perfmon_priv.h"
+
+/*
+ * Context locking rules:
+ * ---------------------
+ *     - any thread with access to the file descriptor of a context can
+ *       potentially issue perfmon calls
+ *
+ *     - calls must be serialized to guarantee correctness
+ *
+ *     - as soon as a context is attached to a thread or CPU, it may be
+ *       actively monitoring. On some architectures, such as IA-64, this
+ *       is true even though the pfm_start() call has not been made. This
+ *       comes from the fact that on some architectures, it is possible to
+ *       start/stop monitoring from userland.
+ *
+ *     - If monitoring is active, then there can PMU interrupts. Because
+ *       context accesses must be serialized, the perfmon system calls
+ *       must mask interrupts as soon as the context is attached.
+ *
+ *     - perfmon system calls that operate with the context unloaded cannot
+ *       assume it is actually unloaded when they are called. They first need
+ *       to check and for that they need interrupts masked. Then, if the
+ *       context is actually unloaded, they can unmask interrupts.
+ *
+ *     - interrupt masking holds true for other internal perfmon functions as
+ *       well. Except for PMU interrupt handler because those interrupts
+ *       cannot be nested.
+ *
+ *     - we mask ALL interrupts instead of just the PMU interrupt because we
+ *       also need to protect against timer interrupts which could trigger
+ *       a set switch.
+ */
+#ifdef CONFIG_UTRACE
+#include <linux/utrace.h>
+
+static u32
+stopper_quiesce(struct utrace_attached_engine *engine, struct task_struct *tsk)
+{
+       PFM_DBG("quiesced [%d]", tsk->pid);
+       complete(engine->data);
+       return UTRACE_ACTION_RESUME;
+}
+
+void
+pfm_resume_task(struct task_struct *t, void *data)
+{
+       PFM_DBG("utrace detach [%d]", t->pid);
+       (void) utrace_detach(t, data);
+}
+
+static const struct utrace_engine_ops utrace_ops =
+{
+       .report_quiesce = stopper_quiesce,
+};
+
+static int pfm_wait_task_stopped(struct task_struct *task, void **data)
+{
+       DECLARE_COMPLETION_ONSTACK(done);
+       struct utrace_attached_engine *eng;
+       int ret;
+
+       eng = utrace_attach(task, UTRACE_ATTACH_CREATE, &utrace_ops, &done);
+       if (IS_ERR(eng))
+               return PTR_ERR(eng);
+
+       ret = utrace_set_flags(task, eng,
+                              UTRACE_ACTION_QUIESCE | UTRACE_EVENT(QUIESCE));
+       PFM_DBG("wait quiesce [%d]", task->pid);
+       if (!ret)
+               ret = wait_for_completion_interruptible(&done);
+
+       if (ret)
+               (void) utrace_detach(task, eng);
+       else
+               *data = eng;
+       return 0;
+}
+#else /* !CONFIG_UTRACE */
+static int pfm_wait_task_stopped(struct task_struct *task, void **data)
+{
+       int ret;
+
+       *data = NULL;
+
+       /*
+        * returns 0 if cannot attach
+        */
+       ret = ptrace_may_access(task, PTRACE_MODE_ATTACH);
+       PFM_DBG("may_attach=%d", ret);
+       if (!ret)
+               return -EPERM;
+
+       ret = ptrace_check_attach(task, 0);
+       PFM_DBG("check_attach=%d", ret);
+       return ret;
+}
+void pfm_resume_task(struct task_struct *t, void *data)
+{}
+#endif
+
+struct pfm_syscall_cookie {
+       struct file *filp;
+       int fput_needed;
+};
+
+/*
+ * cannot attach if :
+ *     - kernel task
+ *     - task not owned by caller (checked by ptrace_may_attach())
+ *     - task is dead or zombie
+ *     - cannot use blocking notification when self-monitoring
+ */
+static int pfm_task_incompatible(struct pfm_context *ctx,
+                                struct task_struct *task)
+{
+       /*
+        * cannot attach to a kernel thread
+        */
+       if (!task->mm) {
+               PFM_DBG("cannot attach to kernel thread [%d]", task->pid);
+               return -EPERM;
+       }
+
+       /*
+        * cannot use block on notification when
+        * self-monitoring.
+        */
+       if (ctx->flags.block && task == current) {
+               PFM_DBG("cannot use block on notification when self-monitoring"
+                       "[%d]", task->pid);
+               return -EINVAL;
+       }
+       /*
+        * cannot attach to a zombie task
+        */
+       if (task->exit_state == EXIT_ZOMBIE || task->exit_state == EXIT_DEAD) {
+               PFM_DBG("cannot attach to zombie/dead task [%d]", task->pid);
+               return -EBUSY;
+       }
+       return 0;
+}
+
+/**
+ * pfm_get_task -- check permission and acquire task to monitor
+ * @ctx: perfmon context
+ * @pid: identification of the task to check
+ * @task: upon return, a pointer to the task to monitor
+ *
+ * This function  is used in per-thread mode only AND when not
+ * self-monitoring. It finds the task to monitor and checks
+ * that the caller has permissions to attach. It also checks
+ * that the task is stopped via ptrace so that we can safely
+ * modify its state.
+ *
+ * task refcount is incremented when succesful.
+ */
+static int pfm_get_task(struct pfm_context *ctx, pid_t pid,
+                       struct task_struct **task, void **data)
+{
+       struct task_struct *p;
+       int ret = 0, ret1 = 0;
+
+       *data = NULL;
+
+       /*
+        * When attaching to another thread we must ensure
+        * that the thread is actually stopped. 
+        *
+        * As a consequence, only the ptracing parent can actually
+        * attach a context to a thread. Obviously, this constraint
+        * does not exist for self-monitoring threads.
+        *
+        * We use ptrace_may_attach() to check for permission.
+        */
+       read_lock(&tasklist_lock);
+
+       p = find_task_by_vpid(pid);
+       if (p)
+               get_task_struct(p);
+
+       read_unlock(&tasklist_lock);
+
+       if (!p) {
+               PFM_DBG("task not found %d", pid);
+               return -ESRCH;
+       }
+
+       ret = pfm_task_incompatible(ctx, p);
+       if (ret)
+               goto error;
+
+       ret = pfm_wait_task_stopped(p, data);
+       if (ret)
+               goto error;
+
+       *task = p;
+
+       return 0;
+error:
+       if (!(ret1 || ret))
+               ret = -EPERM;
+
+       put_task_struct(p);
+
+       return ret;
+}
+
+/*
+ * context must be locked when calling this function
+ */
+int pfm_check_task_state(struct pfm_context *ctx, int check_mask,
+                        unsigned long *flags, void **resume)
+{
+       struct task_struct *task;
+       unsigned long local_flags, new_flags;
+       int state, ret;
+
+       *resume = NULL;
+
+recheck:
+       /*
+        * task is NULL for system-wide context
+        */
+       task = ctx->task;
+       state = ctx->state;
+       local_flags = *flags;
+
+       PFM_DBG("state=%d check_mask=0x%x", state, check_mask);
+       /*
+        * if the context is detached, then we do not touch
+        * hardware, therefore there is not restriction on when we can
+        * access it.
+        */
+       if (state == PFM_CTX_UNLOADED)
+               return 0;
+       /*
+        * no command can operate on a zombie context.
+        * A context becomes zombie when the file that identifies
+        * it is closed while the context is still attached to the
+        * thread it monitors.
+        */
+       if (state == PFM_CTX_ZOMBIE)
+               return -EINVAL;
+
+       /*
+        * at this point, state is PFM_CTX_LOADED or PFM_CTX_MASKED
+        */
+
+       /*
+        * some commands require the context to be unloaded to operate
+        */
+       if (check_mask & PFM_CMD_UNLOADED)  {
+               PFM_DBG("state=%d, cmd needs context unloaded", state);
+               return -EBUSY;
+       }
+
+       /*
+        * self-monitoring always ok.
+        */
+       if (task == current)
+               return 0;
+
+       /*
+        * for syswide, the calling thread must be running on the cpu
+        * the context is bound to.
+        */
+       if (ctx->flags.system) {
+               if (ctx->cpu != smp_processor_id())
+                       return -EBUSY;
+               return 0;
+       }
+
+       /*
+        * at this point, monitoring another thread
+        */
+
+       /*
+        * the pfm_unload_context() command is allowed on masked context
+        */
+       if (state == PFM_CTX_MASKED && !(check_mask & PFM_CMD_UNLOAD))
+               return 0;
+
+       /*
+        * When we operate on another thread, we must wait for it to be
+        * stopped and completely off any CPU as we need to access the
+        * PMU state (or machine state).
+        *
+        * A thread can be put in the STOPPED state in various ways
+        * including PTRACE_ATTACH, or when it receives a SIGSTOP signal.
+        * We enforce that the thread must be ptraced, so it is stopped
+        * AND it CANNOT wake up while we operate on it because this
+        * would require an action from the ptracing parent which is the
+        * thread that is calling this function.
+        *
+        * The dependency on ptrace, imposes that only the ptracing
+        * parent can issue command on a thread. This is unfortunate
+        * but we do not know of a better way of doing this.
+        */
+       if (check_mask & PFM_CMD_STOPPED) {
+
+               spin_unlock_irqrestore(&ctx->lock, local_flags);
+
+               /*
+                * check that the thread is ptraced AND STOPPED
+                */
+               ret = pfm_wait_task_stopped(task, resume);
+
+               spin_lock_irqsave(&ctx->lock, new_flags);
+
+               /*
+                * flags may be different than when we released the lock
+                */
+               *flags = new_flags;
+
+               if (ret)
+                       return ret;
+               /*
+                * we must recheck to verify if state has changed
+                */
+               if (unlikely(ctx->state != state)) {
+                       PFM_DBG("old_state=%d new_state=%d",
+                               state,
+                               ctx->state);
+                       goto recheck;
+               }
+       }
+       return 0;
+}
+
+/*
+ * pfm_get_args - Function used to copy the syscall argument into kernel memory.
+ * @ureq: user argument
+ * @sz: user argument size
+ * @lsz: size of stack buffer
+ * @laddr: stack buffer address
+ * @req: point to start of kernel copy of the argument
+ * @ptr_free: address of kernel copy to free
+ *
+ * There are two options:
+ *     - use a stack buffer described by laddr (addresses) and lsz (size)
+ *     - allocate memory
+ *
+ * return:
+ *     < 0 : in case of error (ptr_free may not be updated)
+ *       0 : success
+ *      - req: points to base of kernel copy of arguments
+ *     - ptr_free: address of buffer to free by caller on exit.
+ *                 NULL if using the stack buffer
+ *
+ * when ptr_free is not NULL upon return, the caller must kfree()
+ */
+int pfm_get_args(void __user *ureq, size_t sz, size_t lsz, void *laddr,
+                void **req, void **ptr_free)
+{
+       void *addr;
+
+       /*
+        * check syadmin argument limit
+        */
+       if (unlikely(sz > pfm_controls.arg_mem_max)) {
+               PFM_DBG("argument too big %zu max=%zu",
+                       sz,
+                       pfm_controls.arg_mem_max);
+               return -E2BIG;
+       }
+
+       /*
+        * check if vector fits on stack buffer
+        */
+       if (sz > lsz) {
+               addr = kmalloc(sz, GFP_KERNEL);
+               if (unlikely(addr == NULL))
+                       return -ENOMEM;
+               *ptr_free = addr;
+       } else {
+               addr = laddr;
+               *req = laddr;
+               *ptr_free = NULL;
+       }
+
+       /*
+        * bring the data in
+        */
+       if (unlikely(copy_from_user(addr, ureq, sz))) {
+               if (addr != laddr)
+                       kfree(addr);
+               return -EFAULT;
+       }
+
+       /*
+        * base address of kernel buffer
+        */
+       *req = addr;
+
+       return 0;
+}
+
+/**
+ * pfm_acquire_ctx_from_fd -- get ctx from file descriptor
+ * @fd: file descriptor
+ * @ctx: pointer to pointer of context updated on return
+ * @cookie: opaque structure to use for release
+ *
+ * This helper function extracts the ctx from the file descriptor.
+ * It also increments the refcount of the file structure. Thus
+ * it updates the cookie so the refcount can be decreased when
+ * leaving the perfmon syscall via pfm_release_ctx_from_fd
+ */
+static int pfm_acquire_ctx_from_fd(int fd, struct pfm_context **ctx,
+                                  struct pfm_syscall_cookie *cookie)
+{
+       struct file *filp;
+       int fput_needed;
+
+       filp = fget_light(fd, &fput_needed);
+       if (unlikely(filp == NULL)) {
+               PFM_DBG("invalid fd %d", fd);
+               return -EBADF;
+       }
+
+       *ctx = filp->private_data;
+
+       if (unlikely(!*ctx || filp->f_op != &pfm_file_ops)) {
+               PFM_DBG("fd %d not related to perfmon", fd);
+               return -EBADF;
+       }
+       cookie->filp = filp;
+       cookie->fput_needed = fput_needed;
+
+       return 0;
+}
+
+/**
+ * pfm_release_ctx_from_fd -- decrease refcount of file associated with context
+ * @cookie: the cookie structure initialized by pfm_acquire_ctx_from_fd
+ */
+static inline void pfm_release_ctx_from_fd(struct pfm_syscall_cookie *cookie)
+{
+       fput_light(cookie->filp, cookie->fput_needed);
+}
+
+/*
+ * unlike the other perfmon system calls, this one returns a file descriptor
+ * or a value < 0 in case of error, very much like open() or socket()
+ */
+asmlinkage long sys_pfm_create_context(struct pfarg_ctx __user *ureq,
+                                      char __user *fmt_name,
+                                      void __user *fmt_uarg, size_t fmt_size)
+{
+       struct pfarg_ctx req;
+       struct pfm_smpl_fmt *fmt = NULL;
+       void *fmt_arg = NULL;
+       int ret;
+
+       PFM_DBG("req=%p fmt=%p fmt_arg=%p size=%zu",
+               ureq, fmt_name, fmt_uarg, fmt_size);
+
+       if (perfmon_disabled)
+               return -ENOSYS;
+
+       if (copy_from_user(&req, ureq, sizeof(req)))
+               return -EFAULT;
+
+       if (fmt_name) {
+               ret = pfm_get_smpl_arg(fmt_name, fmt_uarg, fmt_size, &fmt_arg, &fmt);
+               if (ret)
+                       goto abort;
+       }
+
+       ret = __pfm_create_context(&req, fmt, fmt_arg, PFM_NORMAL, NULL);
+
+       kfree(fmt_arg);
+abort:
+       return ret;
+}
+
+asmlinkage long sys_pfm_write_pmcs(int fd, struct pfarg_pmc __user *ureq, int count)
+{
+       struct pfm_context *ctx;
+       struct task_struct *task;
+       struct pfm_syscall_cookie cookie;
+       struct pfarg_pmc pmcs[PFM_PMC_STK_ARG];
+       struct pfarg_pmc *req;
+       void *fptr, *resume;
+       unsigned long flags;
+       size_t sz;
+       int ret;
+
+       PFM_DBG("fd=%d req=%p count=%d", fd, ureq, count);
+
+       if (count < 0 || count >= PFM_MAX_ARG_COUNT(ureq)) {
+               PFM_DBG("invalid arg count %d", count);
+               return -EINVAL;
+       }
+
+       sz = count*sizeof(*ureq);
+
+       ret = pfm_acquire_ctx_from_fd(fd, &ctx, &cookie);
+       if (ret)
+               return ret;
+
+       ret = pfm_get_args(ureq, sz, sizeof(pmcs), pmcs, (void **)&req, &fptr);
+       if (ret)
+               goto error;
+
+       spin_lock_irqsave(&ctx->lock, flags);
+
+       task = ctx->task;
+
+       ret = pfm_check_task_state(ctx, PFM_CMD_STOPPED, &flags, &resume);
+       if (!ret)
+               ret = __pfm_write_pmcs(ctx, req, count);
+
+       spin_unlock_irqrestore(&ctx->lock, flags);
+
+       if (resume)
+               pfm_resume_task(task, resume);
+
+       /*
+        * This function may be on the critical path.
+        * We want to avoid the branch if unecessary.
+        */
+       if (fptr)
+               kfree(fptr);
+error:
+       pfm_release_ctx_from_fd(&cookie);
+       return ret;
+}
+
+asmlinkage long sys_pfm_write_pmds(int fd, struct pfarg_pmd __user *ureq, int count)
+{
+       struct pfm_context *ctx;
+       struct task_struct *task;
+       struct pfm_syscall_cookie cookie;
+       struct pfarg_pmd pmds[PFM_PMD_STK_ARG];
+       struct pfarg_pmd *req;
+       void *fptr, *resume;
+       unsigned long flags;
+       size_t sz;
+       int ret;
+
+       PFM_DBG("fd=%d req=%p count=%d", fd, ureq, count);
+
+       if (count < 0 || count >= PFM_MAX_ARG_COUNT(ureq)) {
+               PFM_DBG("invalid arg count %d", count);
+               return -EINVAL;
+       }
+
+       sz = count*sizeof(*ureq);
+
+       ret = pfm_acquire_ctx_from_fd(fd, &ctx, &cookie);
+       if (ret)
+               return ret;
+
+       ret = pfm_get_args(ureq, sz, sizeof(pmds), pmds, (void **)&req, &fptr);
+       if (ret)
+               goto error;
+
+       spin_lock_irqsave(&ctx->lock, flags);
+
+       task = ctx->task;
+
+       ret = pfm_check_task_state(ctx, PFM_CMD_STOPPED, &flags, &resume);
+       if (!ret)
+               ret = __pfm_write_pmds(ctx, req, count, 0);
+
+       spin_unlock_irqrestore(&ctx->lock, flags);
+
+       if (resume)
+               pfm_resume_task(task, resume);
+
+       if (fptr)
+               kfree(fptr);
+error:
+       pfm_release_ctx_from_fd(&cookie);
+       return ret;
+}
+
+asmlinkage long sys_pfm_read_pmds(int fd, struct pfarg_pmd __user *ureq, int count)
+{
+       struct pfm_context *ctx;
+       struct task_struct *task;
+       struct pfm_syscall_cookie cookie;
+       struct pfarg_pmd pmds[PFM_PMD_STK_ARG];
+       struct pfarg_pmd *req;
+       void *fptr, *resume;
+       unsigned long flags;
+       size_t sz;
+       int ret;
+
+       PFM_DBG("fd=%d req=%p count=%d", fd, ureq, count);
+
+       if (count < 0 || count >= PFM_MAX_ARG_COUNT(ureq))
+               return -EINVAL;
+
+       sz = count*sizeof(*ureq);
+
+       ret = pfm_acquire_ctx_from_fd(fd, &ctx, &cookie);
+       if (ret)
+               return ret;
+
+       ret = pfm_get_args(ureq, sz, sizeof(pmds), pmds, (void **)&req, &fptr);
+       if (ret)
+               goto error;
+
+       spin_lock_irqsave(&ctx->lock, flags);
+
+       task = ctx->task;
+
+       ret = pfm_check_task_state(ctx, PFM_CMD_STOPPED, &flags, &resume);
+       if (!ret)
+               ret = __pfm_read_pmds(ctx, req, count);
+
+       spin_unlock_irqrestore(&ctx->lock, flags);
+
+       if (copy_to_user(ureq, req, sz))
+               ret = -EFAULT;
+
+       if (resume)
+               pfm_resume_task(task, resume);
+
+       if (fptr)
+               kfree(fptr);
+error:
+       pfm_release_ctx_from_fd(&cookie);
+       return ret;
+}
+
+asmlinkage long sys_pfm_restart(int fd)
+{
+       struct pfm_context *ctx;
+       struct task_struct *task;
+       struct pfm_syscall_cookie cookie;
+       void *resume;
+       unsigned long flags;
+       int ret, info;
+
+       PFM_DBG("fd=%d", fd);
+
+       ret = pfm_acquire_ctx_from_fd(fd, &ctx, &cookie);
+       if (ret)
+               return ret;
+
+       spin_lock_irqsave(&ctx->lock, flags);
+
+       task = ctx->task;
+
+       ret = pfm_check_task_state(ctx, 0, &flags, &resume);
+       if (!ret)
+               ret = __pfm_restart(ctx, &info);
+
+       spin_unlock_irqrestore(&ctx->lock, flags);
+
+       if (resume)
+               pfm_resume_task(task, resume);
+       /*
+        * In per-thread mode with blocking notification, i.e.
+        * ctx->flags.blocking=1, we need to defer issuing the
+        * complete to unblock the blocked monitored thread.
+        * Otherwise we have a potential deadlock due to a lock
+        * inversion between the context lock and the task_rq_lock()
+        * which can happen if one thread is in this call and the other
+        * (the monitored thread) is in the context switch code.
+        *
+        * It is safe to access the context outside the critical section
+        * because:
+        *      - we are protected by the fget_light(), thus the context
+        *        cannot disappear
+        */
+       if (ret == 0 && info == 1)
+               complete(&ctx->restart_complete);
+
+       pfm_release_ctx_from_fd(&cookie);
+       return ret;
+}
+
+asmlinkage long sys_pfm_stop(int fd)
+{
+       struct pfm_context *ctx;
+       struct task_struct *task;
+       struct pfm_syscall_cookie cookie;
+       void *resume;
+       unsigned long flags;
+       int ret;
+       int release_info;
+
+       PFM_DBG("fd=%d", fd);
+
+       ret = pfm_acquire_ctx_from_fd(fd, &ctx, &cookie);
+       if (ret)
+               return ret;
+
+       spin_lock_irqsave(&ctx->lock, flags);
+
+       task = ctx->task;
+
+       ret = pfm_check_task_state(ctx, PFM_CMD_STOPPED, &flags, &resume);
+       if (!ret)
+               ret = __pfm_stop(ctx, &release_info);
+
+       spin_unlock_irqrestore(&ctx->lock, flags);
+
+       if (resume)
+               pfm_resume_task(task, resume);
+
+       /*
+        * defer cancellation of timer to avoid race
+        * with pfm_handle_switch_timeout()
+        *
+        * applies only when self-monitoring
+        */
+       if (release_info & 0x2)
+               hrtimer_cancel(&__get_cpu_var(pfm_hrtimer));
+
+       pfm_release_ctx_from_fd(&cookie);
+       return ret;
+}
+
+asmlinkage long sys_pfm_start(int fd, struct pfarg_start __user *ureq)
+{
+       struct pfm_context *ctx;
+       struct task_struct *task;
+       struct pfm_syscall_cookie cookie;
+       void *resume;
+       struct pfarg_start req;
+       unsigned long flags;
+       int ret;
+
+       PFM_DBG("fd=%d req=%p", fd, ureq);
+
+       ret = pfm_acquire_ctx_from_fd(fd, &ctx, &cookie);
+       if (ret)
+               return ret;
+
+       /*
+        * the one argument is actually optional
+        */
+       if (ureq && copy_from_user(&req, ureq, sizeof(req)))
+               return -EFAULT;
+
+       spin_lock_irqsave(&ctx->lock, flags);
+
+       task = ctx->task;
+
+       ret = pfm_check_task_state(ctx, PFM_CMD_STOPPED, &flags, &resume);
+       if (!ret)
+               ret = __pfm_start(ctx, ureq ? &req : NULL);
+
+       spin_unlock_irqrestore(&ctx->lock, flags);
+
+       if (resume)
+               pfm_resume_task(task, resume);
+
+       pfm_release_ctx_from_fd(&cookie);
+       return ret;
+}
+
+asmlinkage long sys_pfm_load_context(int fd, struct pfarg_load __user *ureq)
+{
+       struct pfm_context *ctx;
+       struct task_struct *task;
+       struct pfm_syscall_cookie cookie;
+       void *resume, *dummy_resume;
+       unsigned long flags;
+       struct pfarg_load req;
+       int ret;
+
+       PFM_DBG("fd=%d req=%p", fd, ureq);
+
+       if (copy_from_user(&req, ureq, sizeof(req)))
+               return -EFAULT;
+
+       ret = pfm_acquire_ctx_from_fd(fd, &ctx, &cookie);
+       if (ret)
+               return ret;
+
+       task = current;
+
+       /*
+        * in per-thread mode (not self-monitoring), get a reference
+        * on task to monitor. This must be done with interrupts enabled
+        * Upon succesful return, refcount on task is increased.
+        *
+        * fget_light() is protecting the context.
+        */
+       if (!ctx->flags.system && req.load_pid != current->pid) {
+               ret = pfm_get_task(ctx, req.load_pid, &task, &resume);
+               if (ret)
+                       goto error;
+       }
+
+       /*
+        * irqsave is required to avoid race in case context is already
+        * loaded or with switch timeout in the case of self-monitoring
+        */
+       spin_lock_irqsave(&ctx->lock, flags);
+
+       ret = pfm_check_task_state(ctx, PFM_CMD_UNLOADED, &flags, &dummy_resume);
+       if (!ret)
+               ret = __pfm_load_context(ctx, &req, task);
+
+       spin_unlock_irqrestore(&ctx->lock, flags);
+
+       if (resume)
+               pfm_resume_task(task, resume);
+
+       /*
+        * in per-thread mode (not self-monitoring), we need
+        * to decrease refcount on task to monitor:
+        *   - load successful: we have a reference to the task in ctx->task
+        *   - load failed    : undo the effect of pfm_get_task()
+        */
+       if (task != current)
+               put_task_struct(task);
+error:
+       pfm_release_ctx_from_fd(&cookie);
+       return ret;
+}
+
+asmlinkage long sys_pfm_unload_context(int fd)
+{
+       struct pfm_context *ctx;
+       struct task_struct *task;
+       struct pfm_syscall_cookie cookie;
+       void *resume;
+       unsigned long flags;
+       int ret;
+       int is_system, release_info = 0;
+       u32 cpu;
+
+       PFM_DBG("fd=%d", fd);
+
+       ret = pfm_acquire_ctx_from_fd(fd, &ctx, &cookie);
+       if (ret)
+               return ret;
+
+       is_system = ctx->flags.system;
+
+       spin_lock_irqsave(&ctx->lock, flags);
+
+       cpu = ctx->cpu;
+       task = ctx->task;
+
+       ret = pfm_check_task_state(ctx, PFM_CMD_STOPPED|PFM_CMD_UNLOAD,
+                                  &flags, &resume);
+       if (!ret)
+               ret = __pfm_unload_context(ctx, &release_info);
+
+       spin_unlock_irqrestore(&ctx->lock, flags);
+
+       if (resume)
+               pfm_resume_task(task, resume);
+
+       /*
+        * cancel time now that context is unlocked
+        * avoid race with pfm_handle_switch_timeout()
+        */
+       if (release_info & 0x2) {
+               int r;
+               r = hrtimer_cancel(&__get_cpu_var(pfm_hrtimer));
+               PFM_DBG("timeout cancel=%d", r);
+       }
+
+       if (release_info & 0x1)
+               pfm_session_release(is_system, cpu);
+
+       pfm_release_ctx_from_fd(&cookie);
+       return ret;
+}
+
+asmlinkage long sys_pfm_create_evtsets(int fd, struct pfarg_setdesc __user *ureq, int count)
+{
+       struct pfm_context *ctx;
+       struct pfm_syscall_cookie cookie;
+       struct pfarg_setdesc *req;
+       void *fptr, *resume;
+       unsigned long flags;
+       size_t sz;
+       int ret;
+
+       PFM_DBG("fd=%d req=%p count=%d", fd, ureq, count);
+
+       if (count < 0 || count >= PFM_MAX_ARG_COUNT(ureq))
+               return -EINVAL;
+
+       sz = count*sizeof(*ureq);
+
+       ret = pfm_acquire_ctx_from_fd(fd, &ctx, &cookie);
+       if (ret)
+               return ret;
+
+       ret = pfm_get_args(ureq, sz, 0, NULL, (void **)&req, &fptr);
+       if (ret)
+               goto error;
+
+       /*
+        * must mask interrupts because we do not know the state of context,
+        * could be attached and we could be getting PMU interrupts. So
+        * we mask and lock context and we check and possibly relax masking
+        */
+       spin_lock_irqsave(&ctx->lock, flags);
+
+       ret = pfm_check_task_state(ctx, PFM_CMD_UNLOADED, &flags, &resume);
+       if (!ret)
+               ret = __pfm_create_evtsets(ctx, req, count);
+
+       spin_unlock_irqrestore(&ctx->lock, flags);
+       /*
+        * context must be unloaded for this command. The resume pointer
+        * is necessarily NULL, thus no need to call pfm_resume_task()
+        */
+       kfree(fptr);
+
+error:
+       pfm_release_ctx_from_fd(&cookie);
+       return ret;
+}
+
+asmlinkage long  sys_pfm_getinfo_evtsets(int fd, struct pfarg_setinfo __user *ureq, int count)
+{
+       struct pfm_context *ctx;
+       struct task_struct *task;
+       struct pfm_syscall_cookie cookie;
+       struct pfarg_setinfo *req;
+       void *fptr, *resume;
+       unsigned long flags;
+       size_t sz;
+       int ret;
+
+       PFM_DBG("fd=%d req=%p count=%d", fd, ureq, count);
+
+       if (count < 0 || count >= PFM_MAX_ARG_COUNT(ureq))
+               return -EINVAL;
+
+       sz = count*sizeof(*ureq);
+
+       ret = pfm_acquire_ctx_from_fd(fd, &ctx, &cookie);
+       if (ret)
+               return ret;
+
+       ret = pfm_get_args(ureq, sz, 0, NULL, (void **)&req, &fptr);
+       if (ret)
+               goto error;
+
+       /*
+        * this command operates even when context is loaded, so we need
+        * to keep interrupts masked to avoid a race with PMU interrupt
+        * which may switch the active set
+        */
+       spin_lock_irqsave(&ctx->lock, flags);
+
+       task = ctx->task;
+
+       ret = pfm_check_task_state(ctx, 0, &flags, &resume);
+       if (!ret)
+               ret = __pfm_getinfo_evtsets(ctx, req, count);
+
+       spin_unlock_irqrestore(&ctx->lock, flags);
+
+       if (resume)
+               pfm_resume_task(task, resume);
+
+       if (copy_to_user(ureq, req, sz))
+               ret = -EFAULT;
+
+       kfree(fptr);
+error:
+       pfm_release_ctx_from_fd(&cookie);
+       return ret;
+}
+
+asmlinkage long sys_pfm_delete_evtsets(int fd, struct pfarg_setinfo __user *ureq, int count)
+{
+       struct pfm_context *ctx;
+       struct pfm_syscall_cookie cookie;
+       struct pfarg_setinfo *req;
+       void *fptr, *resume;
+       unsigned long flags;
+       size_t sz;
+       int ret;
+
+       PFM_DBG("fd=%d req=%p count=%d", fd, ureq, count);
+
+       if (count < 0 || count >= PFM_MAX_ARG_COUNT(ureq))
+               return -EINVAL;
+
+       sz = count*sizeof(*ureq);
+
+       ret = pfm_acquire_ctx_from_fd(fd, &ctx, &cookie);
+       if (ret)
+               return ret;
+
+       ret = pfm_get_args(ureq, sz, 0, NULL, (void **)&req, &fptr);
+       if (ret)
+               goto error;
+
+       /*
+        * must mask interrupts because we do not know the state of context,
+        * could be attached and we could be getting PMU interrupts
+        */
+       spin_lock_irqsave(&ctx->lock, flags);
+
+       ret = pfm_check_task_state(ctx, PFM_CMD_UNLOADED, &flags, &resume);
+       if (!ret)
+               ret = __pfm_delete_evtsets(ctx, req, count);
+
+       spin_unlock_irqrestore(&ctx->lock, flags);
+       /*
+        * context must be unloaded for this command. The resume pointer
+        * is necessarily NULL, thus no need to call pfm_resume_task()
+        */
+       kfree(fptr);
+
+error:
+       pfm_release_ctx_from_fd(&cookie);
+       return ret;
+}
diff --git a/perfmon/perfmon_sysfs.c b/perfmon/perfmon_sysfs.c
new file mode 100644
index 0000000..7353c3b
--- /dev/null
+++ b/perfmon/perfmon_sysfs.c
@@ -0,0 +1,525 @@
+/*
+ * perfmon_sysfs.c: perfmon2 sysfs interface
+ *
+ * This file implements the perfmon2 interface which
+ * provides access to the hardware performance counters
+ * of the host processor.
+ *
+ * The initial version of perfmon.c was written by
+ * Ganesh Venkitachalam, IBM Corp.
+ *
+ * Then it was modified for perfmon-1.x by Stephane Eranian and
+ * David Mosberger, Hewlett Packard Co.
+ *
+ * Version Perfmon-2.x is a complete rewrite of perfmon-1.x
+ * by Stephane Eranian, Hewlett Packard Co.
+ *
+ * Copyright (c) 1999-2006 Hewlett-Packard Development Company, L.P.
+ * Contributed by Stephane Eranian <eranian@hpl.hp.com>
+ *                David Mosberger-Tang <davidm@hpl.hp.com>
+ *
+ * More information about perfmon available at:
+ *     http://perfmon2.sf.net
+ *
+ * This program is free software; you can redistribute it and/or
+ * modify it under the terms of version 2 of the GNU General Public
+ * License as published by the Free Software Foundation.
+ *
+ * This program is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+ * General Public License for more details.
+ *
+ * You should have received a copy of the GNU General Public License
+ * along with this program; if not, write to the Free Software
+ * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA
+ * 02111-1307 USA
+ */
+#include <linux/kernel.h>
+#include <linux/module.h> /* for EXPORT_SYMBOL */
+#include <linux/perfmon_kern.h>
+#include "perfmon_priv.h"
+
+struct pfm_attribute {
+       struct attribute attr;
+       ssize_t (*show)(void *, struct pfm_attribute *attr, char *);
+       ssize_t (*store)(void *, const char *, size_t);
+};
+#define to_attr(n) container_of(n, struct pfm_attribute, attr);
+
+#define PFM_RO_ATTR(_name, _show) \
+       struct kobj_attribute attr_##_name = __ATTR(_name, 0444, _show, NULL)
+
+#define PFM_RW_ATTR(_name, _show, _store)                      \
+       struct kobj_attribute attr_##_name = __ATTR(_name, 0644, _show, _store)
+
+#define PFM_ROS_ATTR(_name, _show) \
+       struct pfm_attribute attr_##_name = __ATTR(_name, 0444, _show, NULL)
+
+#define is_attr_name(a, n) (!strcmp((a)->attr.name, n))
+int pfm_sysfs_add_pmu(struct pfm_pmu_config *pmu);
+
+static struct kobject *pfm_kernel_kobj, *pfm_fmt_kobj;
+static struct kobject *pfm_pmu_kobj;
+
+static ssize_t pfm_regs_attr_show(struct kobject *kobj,
+               struct attribute *attr, char *buf)
+{
+       struct pfm_regmap_desc *reg = to_reg(kobj);
+       struct pfm_attribute *attribute = to_attr(attr);
+       return attribute->show ? attribute->show(reg, attribute, buf) : -EIO;
+}
+
+static ssize_t pfm_fmt_attr_show(struct kobject *kobj,
+               struct attribute *attr, char *buf)
+{
+       struct pfm_smpl_fmt *fmt = to_smpl_fmt(kobj);
+       struct pfm_attribute *attribute = to_attr(attr);
+       return attribute->show ? attribute->show(fmt, attribute, buf) : -EIO;
+}
+
+static struct sysfs_ops pfm_regs_sysfs_ops = {
+       .show  = pfm_regs_attr_show
+};
+
+static struct sysfs_ops pfm_fmt_sysfs_ops = {
+       .show = pfm_fmt_attr_show
+};
+
+static struct kobj_type pfm_regs_ktype = {
+       .sysfs_ops = &pfm_regs_sysfs_ops,
+};
+
+static struct kobj_type pfm_fmt_ktype = {
+       .sysfs_ops = &pfm_fmt_sysfs_ops,
+};
+
+static ssize_t pfm_controls_show(struct kobject *kobj, struct kobj_attribute *attr, char *buf)
+{
+       int base;
+
+       if (is_attr_name(attr, "version"))
+               return snprintf(buf, PAGE_SIZE, "%u.%u\n",  PFM_VERSION_MAJ, PFM_VERSION_MIN);
+
+       if (is_attr_name(attr, "task_sessions_count"))
+               return pfm_sysfs_res_show(buf, PAGE_SIZE, 0);
+
+       if (is_attr_name(attr, "debug"))
+               return snprintf(buf, PAGE_SIZE, "%d\n", pfm_controls.debug);
+
+       if (is_attr_name(attr, "task_group"))
+               return snprintf(buf, PAGE_SIZE, "%d\n", pfm_controls.task_group);
+
+       if (is_attr_name(attr, "mode"))
+               return snprintf(buf, PAGE_SIZE, "%d\n", pfm_controls.flags);
+
+       if (is_attr_name(attr, "arg_mem_max"))
+               return snprintf(buf, PAGE_SIZE, "%zu\n", pfm_controls.arg_mem_max);
+
+       if (is_attr_name(attr, "syscall")) {
+               base = pfm_arch_get_base_syscall();
+               return snprintf(buf, PAGE_SIZE, "%d\n",  base);
+       }
+
+       if (is_attr_name(attr, "sys_sessions_count"))
+               return pfm_sysfs_res_show(buf, PAGE_SIZE, 1);
+
+       if (is_attr_name(attr, "smpl_buffer_mem_max"))
+               return snprintf(buf, PAGE_SIZE, "%zu\n", pfm_controls.smpl_buffer_mem_max);
+
+       if (is_attr_name(attr, "smpl_buffer_mem_cur"))
+               return pfm_sysfs_res_show(buf, PAGE_SIZE, 2);
+
+       if (is_attr_name(attr, "sys_group"))
+               return snprintf(buf, PAGE_SIZE, "%d\n", pfm_controls.sys_group);
+
+       /* XXX: could be set to write-only */
+       if (is_attr_name(attr, "reset_stats")) {
+               buf[0] = '0';
+               buf[1] = '\0';
+               return strnlen(buf, PAGE_SIZE);
+       }
+       return 0;
+}
+
+static ssize_t pfm_controls_store(struct kobject *kobj, struct kobj_attribute *attr,
+                                 const char *buf, size_t count)
+{
+       int i;
+       size_t d;
+
+       if (sscanf(buf, "%zu", &d) != 1)
+               goto skip;
+
+       if (is_attr_name(attr, "debug"))
+               pfm_controls.debug = d;
+
+       if (is_attr_name(attr, "task_group"))
+               pfm_controls.task_group = d;
+
+       if (is_attr_name(attr, "sys_group"))
+               pfm_controls.sys_group = d;
+
+       if (is_attr_name(attr, "mode"))
+                pfm_controls.flags = d ? PFM_CTRL_FL_RW_EXPERT : 0;
+
+       if (is_attr_name(attr, "arg_mem_max")) {
+               /*
+                * we impose a page as the minimum.
+                *
+                * This limit may be smaller than the stack buffer
+                * available and that is fine.
+                */
+               if (d >= PAGE_SIZE)
+                       pfm_controls.arg_mem_max = d;
+       }
+       if (is_attr_name(attr, "reset_stats")) {
+               for_each_online_cpu(i) {
+                       pfm_reset_stats(i);
+               }
+       }
+
+       if (is_attr_name(attr, "smpl_buffer_mem_max")) {
+               if (d >= PAGE_SIZE)
+                       pfm_controls.smpl_buffer_mem_max = d;
+       }
+skip:
+       return count;
+}
+
+/*
+ * /sys/kernel/perfmon attributes
+ */
+static PFM_RO_ATTR(version, pfm_controls_show);
+static PFM_RO_ATTR(task_sessions_count, pfm_controls_show);
+static PFM_RO_ATTR(syscall, pfm_controls_show);
+static PFM_RO_ATTR(sys_sessions_count, pfm_controls_show);
+static PFM_RO_ATTR(smpl_buffer_mem_cur, pfm_controls_show);
+
+static PFM_RW_ATTR(debug, pfm_controls_show, pfm_controls_store);
+static PFM_RW_ATTR(task_group, pfm_controls_show, pfm_controls_store);
+static PFM_RW_ATTR(mode, pfm_controls_show, pfm_controls_store);
+static PFM_RW_ATTR(sys_group, pfm_controls_show, pfm_controls_store);
+static PFM_RW_ATTR(arg_mem_max, pfm_controls_show, pfm_controls_store);
+static PFM_RW_ATTR(smpl_buffer_mem_max, pfm_controls_show, pfm_controls_store);
+static PFM_RW_ATTR(reset_stats, pfm_controls_show, pfm_controls_store);
+
+static struct attribute *pfm_kernel_attrs[] = {
+       &attr_version.attr,
+       &attr_syscall.attr,
+       &attr_task_sessions_count.attr,
+       &attr_sys_sessions_count.attr,
+       &attr_smpl_buffer_mem_cur.attr,
+       &attr_debug.attr,
+       &attr_reset_stats.attr,
+       &attr_sys_group.attr,
+       &attr_task_group.attr,
+        &attr_mode.attr,
+       &attr_smpl_buffer_mem_max.attr,
+       &attr_arg_mem_max.attr,
+       NULL
+};
+
+static struct attribute_group pfm_kernel_attr_group = {
+       .attrs = pfm_kernel_attrs,
+};
+
+/*
+ * per-reg attributes
+ */
+static ssize_t pfm_reg_show(void *data, struct pfm_attribute *attr, char *buf)
+{
+       struct pfm_regmap_desc *reg;
+       int w;
+
+       reg = data;
+
+       if (is_attr_name(attr, "name"))
+               return snprintf(buf, PAGE_SIZE, "%s\n", reg->desc);
+
+       if (is_attr_name(attr, "dfl_val"))
+               return snprintf(buf, PAGE_SIZE, "0x%llx\n",
+                               (unsigned long long)reg->dfl_val);
+
+       if (is_attr_name(attr, "width")) {
+               w = (reg->type & PFM_REG_C64) ?
+                   pfm_pmu_conf->counter_width : 64;
+               return snprintf(buf, PAGE_SIZE, "%d\n", w);
+       }
+
+       if (is_attr_name(attr, "rsvd_msk"))
+               return snprintf(buf, PAGE_SIZE, "0x%llx\n",
+                               (unsigned long long)reg->rsvd_msk);
+
+       if (is_attr_name(attr, "addr"))
+               return snprintf(buf, PAGE_SIZE, "0x%lx\n", reg->hw_addr);
+
+       return 0;
+}
+
+static PFM_ROS_ATTR(name, pfm_reg_show);
+static PFM_ROS_ATTR(dfl_val, pfm_reg_show);
+static PFM_ROS_ATTR(rsvd_msk, pfm_reg_show);
+static PFM_ROS_ATTR(width, pfm_reg_show);
+static PFM_ROS_ATTR(addr, pfm_reg_show);
+
+static struct attribute *pfm_reg_attrs[] = {
+       &attr_name.attr,
+       &attr_dfl_val.attr,
+       &attr_rsvd_msk.attr,
+       &attr_width.attr,
+       &attr_addr.attr,
+       NULL
+};
+
+static struct attribute_group pfm_reg_attr_group = {
+       .attrs = pfm_reg_attrs,
+};
+
+static ssize_t pfm_pmu_show(struct kobject *kobj, struct kobj_attribute *attr, char *buf)
+{
+       if (is_attr_name(attr, "model"))
+               return snprintf(buf, PAGE_SIZE, "%s\n", pfm_pmu_conf->pmu_name);
+       return 0;
+}
+static PFM_RO_ATTR(model, pfm_pmu_show);
+
+static struct attribute *pfm_pmu_desc_attrs[] = {
+       &attr_model.attr,
+       NULL
+};
+
+static struct attribute_group pfm_pmu_desc_attr_group = {
+       .attrs = pfm_pmu_desc_attrs,
+};
+
+static int pfm_sysfs_add_pmu_regs(struct pfm_pmu_config *pmu)
+{
+       struct pfm_regmap_desc *reg;
+       unsigned int i, k;
+       int ret;
+
+       reg = pmu->pmc_desc;
+       for (i = 0; i < pmu->num_pmc_entries; i++, reg++) {
+
+               if (!(reg->type & PFM_REG_I))
+                       continue;
+
+               ret = kobject_init_and_add(&reg->kobj, &pfm_regs_ktype,
+                                          pfm_pmu_kobj, "pmc%u", i);
+               if (ret)
+                       goto undo_pmcs;
+
+               ret = sysfs_create_group(&reg->kobj, &pfm_reg_attr_group);
+               if (ret) {
+                       kobject_del(&reg->kobj);
+                       goto undo_pmcs;
+               }
+       }
+
+       reg = pmu->pmd_desc;
+       for (i = 0; i < pmu->num_pmd_entries; i++, reg++) {
+
+               if (!(reg->type & PFM_REG_I))
+                       continue;
+
+               ret = kobject_init_and_add(&reg->kobj, &pfm_regs_ktype,
+                                          pfm_pmu_kobj, "pmd%u", i);
+               if (ret)
+                       goto undo_pmds;
+
+               ret = sysfs_create_group(&reg->kobj, &pfm_reg_attr_group);
+               if (ret) {
+                       kobject_del(&reg->kobj);
+                       goto undo_pmds;
+               }
+       }
+       return 0;
+undo_pmds:
+       reg = pmu->pmd_desc;
+       for (k = 0; k < i; k++, reg++) {
+               if (!(reg->type & PFM_REG_I))
+                       continue;
+               sysfs_remove_group(&reg->kobj, &pfm_reg_attr_group);
+               kobject_del(&reg->kobj);
+       }
+       i = pmu->num_pmc_entries;
+       /* fall through */
+undo_pmcs:
+       reg = pmu->pmc_desc;
+       for (k = 0; k < i; k++, reg++) {
+               if (!(reg->type & PFM_REG_I))
+                       continue;
+               sysfs_remove_group(&reg->kobj, &pfm_reg_attr_group);
+               kobject_del(&reg->kobj);
+       }
+       return ret;
+}
+
+static int pfm_sysfs_del_pmu_regs(struct pfm_pmu_config *pmu)
+{
+       struct pfm_regmap_desc *reg;
+       unsigned int i;
+
+       reg = pmu->pmc_desc;
+       for (i = 0; i < pmu->num_pmc_entries; i++, reg++) {
+
+               if (!(reg->type & PFM_REG_I))
+                       continue;
+
+               sysfs_remove_group(&reg->kobj, &pfm_reg_attr_group);
+               kobject_del(&reg->kobj);
+       }
+
+       reg = pmu->pmd_desc;
+       for (i = 0; i < pmu->num_pmd_entries; i++, reg++) {
+
+               if (!(reg->type & PFM_REG_I))
+                       continue;
+
+               sysfs_remove_group(&reg->kobj, &pfm_reg_attr_group);
+               kobject_del(&reg->kobj);
+       }
+       return 0;
+}
+
+/*
+ * when a PMU description module is inserted, we create
+ * a pmu_desc subdir in sysfs and we populate it with
+ * PMU specific information, such as register mappings
+ */
+int pfm_sysfs_add_pmu(struct pfm_pmu_config *pmu)
+{
+       int ret;
+
+       pfm_pmu_kobj = kobject_create_and_add("pmu_desc", pfm_kernel_kobj);
+       if (!pfm_pmu_kobj)
+               return -ENOMEM;
+
+       ret = sysfs_create_group(pfm_pmu_kobj, &pfm_pmu_desc_attr_group);
+       if (ret) {
+               /* will release pfm_pmu_kobj */
+               kobject_put(pfm_pmu_kobj);
+               return ret;
+       }
+
+       ret = pfm_sysfs_add_pmu_regs(pmu);
+       if (ret) {
+               sysfs_remove_group(pfm_pmu_kobj, &pfm_pmu_desc_attr_group);
+               /* will release pfm_pmu_kobj */
+               kobject_put(pfm_pmu_kobj);
+       } else
+               kobject_uevent(pfm_pmu_kobj, KOBJ_ADD);
+
+       return ret;
+}
+
+/*
+ * when a PMU description module is removed, we also remove
+ * all its information from sysfs, i.e., the pmu_desc subdir
+ * disappears
+ */
+int pfm_sysfs_remove_pmu(struct pfm_pmu_config *pmu)
+{
+       pfm_sysfs_del_pmu_regs(pmu);
+       sysfs_remove_group(pfm_pmu_kobj, &pfm_pmu_desc_attr_group);
+       kobject_uevent(pfm_pmu_kobj, KOBJ_REMOVE);
+       kobject_put(pfm_pmu_kobj);
+       pfm_pmu_kobj = NULL;
+       return 0;
+}
+
+static ssize_t pfm_fmt_show(void *data, struct pfm_attribute *attr, char *buf)
+{
+       struct pfm_smpl_fmt *fmt = data;
+
+       if (is_attr_name(attr, "version"))
+               return snprintf(buf, PAGE_SIZE, "%u.%u\n",
+                       fmt->fmt_version >> 16 & 0xffff,
+                       fmt->fmt_version & 0xffff);
+       return 0;
+}
+
+/*
+ * do not use predefined macros because of name conflict
+ * with /sys/kernel/perfmon/version
+ */
+struct pfm_attribute attr_fmt_version = {
+       .attr   = { .name = "version", .mode = 0444 },
+       .show   = pfm_fmt_show,
+};
+
+static struct attribute *pfm_fmt_attrs[] = {
+       &attr_fmt_version.attr,
+       NULL
+};
+
+static struct attribute_group pfm_fmt_attr_group = {
+       .attrs = pfm_fmt_attrs,
+};
+
+/*
+ * when a sampling format module is inserted, we populate
+ * sysfs with some information
+ */
+int pfm_sysfs_add_fmt(struct pfm_smpl_fmt *fmt)
+{
+       int ret;
+
+       ret = kobject_init_and_add(&fmt->kobj, &pfm_fmt_ktype,
+                                  pfm_fmt_kobj, fmt->fmt_name);
+       if (ret)
+               return ret;
+
+       ret = sysfs_create_group(&fmt->kobj, &pfm_fmt_attr_group);
+       if (ret)
+               kobject_del(&fmt->kobj);
+       else
+               kobject_uevent(&fmt->kobj, KOBJ_ADD);
+
+       return ret;
+}
+
+/*
+ * when a sampling format module is removed, its information
+ * must also be removed from sysfs
+ */
+void pfm_sysfs_remove_fmt(struct pfm_smpl_fmt *fmt)
+{
+       sysfs_remove_group(&fmt->kobj, &pfm_fmt_attr_group);
+       kobject_uevent(&fmt->kobj, KOBJ_REMOVE);
+       kobject_del(&fmt->kobj);
+}
+
+int __init pfm_init_sysfs(void)
+{
+       int ret;
+
+       pfm_kernel_kobj = kobject_create_and_add("perfmon", kernel_kobj);
+       if (!pfm_kernel_kobj) {
+               PFM_ERR("cannot add kernel object: /sys/kernel/perfmon");
+               return -ENOMEM;
+       }
+
+       ret = sysfs_create_group(pfm_kernel_kobj, &pfm_kernel_attr_group);
+       if (ret) {
+               kobject_put(pfm_kernel_kobj);
+               return ret;
+       }
+
+       pfm_fmt_kobj = kobject_create_and_add("formats", pfm_kernel_kobj);
+       if (ret) {
+               PFM_ERR("cannot add fmt object: %d", ret);
+               goto error_fmt;
+       }
+       if (pfm_pmu_conf)
+               pfm_sysfs_add_pmu(pfm_pmu_conf);
+
+       pfm_sysfs_builtin_fmt_add();
+
+       return 0;
+
+error_fmt:
+       kobject_del(pfm_kernel_kobj);
+       return ret;
+}