include/linux/perf_event.h

   1 /*
   2  * Performance events:
   3  *
   4  *    Copyright (C) 2008-2009, Thomas Gleixner <tglx@linutronix.de>
   5  *    Copyright (C) 2008-2011, Red Hat, Inc., Ingo Molnar
   6  *    Copyright (C) 2008-2011, Red Hat, Inc., Peter Zijlstra
   7  *
   8  * Data type definitions, declarations, prototypes.
   9  *
  10  *    Started by: Thomas Gleixner and Ingo Molnar
  11  *
  12  * For licencing details see kernel-base/COPYING
  13  */
  14 #ifndef _LINUX_PERF_EVENT_H
  15 #define _LINUX_PERF_EVENT_H
  16
  17 #include <uapi/linux/perf_event.h>
  18
  19 /*
  20  * Kernel-internal data types and definitions:
  21  */
  22
  23 #ifdef CONFIG_PERF_EVENTS
  24 # include <asm/perf_event.h>
  25 # include <asm/local64.h>
  26 #endif
  27
  28 struct perf_guest_info_callbacks {
  29         int                             (*is_in_guest)(void);
  30         int                             (*is_user_mode)(void);
  31         unsigned long                   (*get_guest_ip)(void);
  32 };
  33
  34 #ifdef CONFIG_HAVE_HW_BREAKPOINT
  35 #include <asm/hw_breakpoint.h>
  36 #endif
  37
  38 #include <linux/list.h>
  39 #include <linux/mutex.h>
  40 #include <linux/rculist.h>
  41 #include <linux/rcupdate.h>
  42 #include <linux/spinlock.h>
  43 #include <linux/hrtimer.h>
  44 #include <linux/fs.h>
  45 #include <linux/pid_namespace.h>
  46 #include <linux/workqueue.h>
  47 #include <linux/ftrace.h>
  48 #include <linux/cpu.h>
  49 #include <linux/irq_work.h>
  50 #include <linux/static_key.h>
  51 #include <linux/jump_label_ratelimit.h>
  52 #include <linux/atomic.h>
  53 #include <linux/sysfs.h>
  54 #include <linux/perf_regs.h>
  55 #include <asm/local.h>
  56
  57 struct perf_callchain_entry {
  58         __u64                           nr;
  59         __u64                           ip[PERF_MAX_STACK_DEPTH];
  60 };
  61
  62 struct perf_raw_record {
  63         u32                             size;
  64         void                            *data;
  65 };
  66
  67 /*
  68  * branch stack layout:
  69  *  nr: number of taken branches stored in entries[]
  70  *
  71  * Note that nr can vary from sample to sample
  72  * branches (to, from) are stored from most recent
  73  * to least recent, i.e., entries[0] contains the most
  74  * recent branch.
  75  */
  76 struct perf_branch_stack {
  77         __u64                           nr;
  78         struct perf_branch_entry        entries[0];
  79 };
  80
  81 struct perf_regs_user {
  82         __u64           abi;
  83         struct pt_regs  *regs;
  84 };
  85
  86 struct task_struct;
  87
  88 /*
  89  * extra PMU register associated with an event
  90  */
  91 struct hw_perf_event_extra {
  92         u64             config; /* register value */
  93         unsigned int    reg;    /* register address or index */
  94         int             alloc;  /* extra register already allocated */
  95         int             idx;    /* index in shared_regs->regs[] */
  96 };
  97
  98 struct event_constraint;
  99
 100 /**
 101  * struct hw_perf_event - performance event hardware details:
 102  */
 103 struct hw_perf_event {
 104 #ifdef CONFIG_PERF_EVENTS
 105         union {
 106                 struct { /* hardware */
 107                         u64             config;
 108                         u64             last_tag;
 109                         unsigned long   config_base;
 110                         unsigned long   event_base;
 111                         int             event_base_rdpmc;
 112                         int             idx;
 113                         int             last_cpu;
 114                         int             flags;
 115
 116                         struct hw_perf_event_extra extra_reg;
 117                         struct hw_perf_event_extra branch_reg;
 118
 119                         struct event_constraint *constraint;
 120                 };
 121                 struct { /* software */
 122                         struct hrtimer  hrtimer;
 123                 };
 124                 struct { /* tracepoint */
 125                         struct task_struct      *tp_target;
 126                         /* for tp_event->class */
 127                         struct list_head        tp_list;
 128                 };
 129 #ifdef CONFIG_HAVE_HW_BREAKPOINT
 130                 struct { /* breakpoint */
 131                         /*
 132                          * Crufty hack to avoid the chicken and egg
 133                          * problem hw_breakpoint has with context
 134                          * creation and event initalization.
 135                          */
 136                         struct task_struct              *bp_target;
 137                         struct arch_hw_breakpoint       info;
 138                         struct list_head                bp_list;
 139                 };
 140 #endif
 141         };
 142         int                             state;
 143         local64_t                       prev_count;
 144         u64                             sample_period;
 145         u64                             last_period;
 146         local64_t                       period_left;
 147         u64                             interrupts_seq;
 148         u64                             interrupts;
 149
 150         u64                             freq_time_stamp;
 151         u64                             freq_count_stamp;
 152 #endif
 153 };
 154
 155 /*
 156  * hw_perf_event::state flags
 157  */
 158 #define PERF_HES_STOPPED        0x01 /* the counter is stopped */
 159 #define PERF_HES_UPTODATE       0x02 /* event->count up-to-date */
 160 #define PERF_HES_ARCH           0x04
 161
 162 struct perf_event;
 163
 164 /*
 165  * Common implementation detail of pmu::{start,commit,cancel}_txn
 166  */
 167 #define PERF_EVENT_TXN 0x1
 168
 169 /**
 170  * pmu::capabilities flags
 171  */
 172 #define PERF_PMU_CAP_NO_INTERRUPT               0x01
 173
 174 /**
 175  * struct pmu - generic performance monitoring unit
 176  */
 177 struct pmu {
 178         struct list_head                entry;
 179
 180         struct module                   *module;
 181         struct device                   *dev;
 182         const struct attribute_group    **attr_groups;
 183         const char                      *name;
 184         int                             type;
 185
 186         /*
 187          * various common per-pmu feature flags
 188          */
 189         int                             capabilities;
 190
 191         int * __percpu                  pmu_disable_count;
 192         struct perf_cpu_context * __percpu pmu_cpu_context;
 193         int                             task_ctx_nr;
 194         int                             hrtimer_interval_ms;
 195
 196         /*
 197          * Fully disable/enable this PMU, can be used to protect from the PMI
 198          * as well as for lazy/batch writing of the MSRs.
 199          */
 200         void (*pmu_enable)              (struct pmu *pmu); /* optional */
 201         void (*pmu_disable)             (struct pmu *pmu); /* optional */
 202
 203         /*
 204          * Try and initialize the event for this PMU.
 205          * Should return -ENOENT when the @event doesn't match this PMU.
 206          */
 207         int (*event_init)               (struct perf_event *event);
 208
 209 #define PERF_EF_START   0x01            /* start the counter when adding    */
 210 #define PERF_EF_RELOAD  0x02            /* reload the counter when starting */
 211 #define PERF_EF_UPDATE  0x04            /* update the counter when stopping */
 212
 213         /*
 214          * Adds/Removes a counter to/from the PMU, can be done inside
 215          * a transaction, see the ->*_txn() methods.
 216          */
 217         int  (*add)                     (struct perf_event *event, int flags);
 218         void (*del)                     (struct perf_event *event, int flags);
 219
 220         /*
 221          * Starts/Stops a counter present on the PMU. The PMI handler
 222          * should stop the counter when perf_event_overflow() returns
 223          * !0. ->start() will be used to continue.
 224          */
 225         void (*start)                   (struct perf_event *event, int flags);
 226         void (*stop)                    (struct perf_event *event, int flags);
 227
 228         /*
 229          * Updates the counter value of the event.
 230          */
 231         void (*read)                    (struct perf_event *event);
 232
 233         /*
 234          * Group events scheduling is treated as a transaction, add
 235          * group events as a whole and perform one schedulability test.
 236          * If the test fails, roll back the whole group
 237          *
 238          * Start the transaction, after this ->add() doesn't need to
 239          * do schedulability tests.
 240          */
 241         void (*start_txn)               (struct pmu *pmu); /* optional */
 242         /*
 243          * If ->start_txn() disabled the ->add() schedulability test
 244          * then ->commit_txn() is required to perform one. On success
 245          * the transaction is closed. On error the transaction is kept
 246          * open until ->cancel_txn() is called.
 247          */
 248         int  (*commit_txn)              (struct pmu *pmu); /* optional */
 249         /*
 250          * Will cancel the transaction, assumes ->del() is called
 251          * for each successful ->add() during the transaction.
 252          */
 253         void (*cancel_txn)              (struct pmu *pmu); /* optional */
 254
 255         /*
 256          * Will return the value for perf_event_mmap_page::index for this event,
 257          * if no implementation is provided it will default to: event->hw.idx + 1.
 258          */
 259         int (*event_idx)                (struct perf_event *event); /*optional */
 260
 261         /*
 262          * flush branch stack on context-switches (needed in cpu-wide mode)
 263          */
 264         void (*flush_branch_stack)      (void);
 265
 266         /*
 267          * Check period value for PERF_EVENT_IOC_PERIOD ioctl.
 268          */
 269         int (*check_period)             (struct perf_event *event, u64 value); /* optional */
 270 };
 271
 272 /**
 273  * enum perf_event_active_state - the states of a event
 274  */
 275 enum perf_event_active_state {
 276         PERF_EVENT_STATE_ERROR          = -2,
 277         PERF_EVENT_STATE_OFF            = -1,
 278         PERF_EVENT_STATE_INACTIVE       =  0,
 279         PERF_EVENT_STATE_ACTIVE         =  1,
 280 };
 281
 282 struct file;
 283 struct perf_sample_data;
 284
 285 typedef void (*perf_overflow_handler_t)(struct perf_event *,
 286                                         struct perf_sample_data *,
 287                                         struct pt_regs *regs);
 288
 289 enum perf_group_flag {
 290         PERF_GROUP_SOFTWARE             = 0x1,
 291 };
 292
 293 #define SWEVENT_HLIST_BITS              8
 294 #define SWEVENT_HLIST_SIZE              (1 << SWEVENT_HLIST_BITS)
 295
 296 struct swevent_hlist {
 297         struct hlist_head               heads[SWEVENT_HLIST_SIZE];
 298         struct rcu_head                 rcu_head;
 299 };
 300
 301 #define PERF_ATTACH_CONTEXT     0x01
 302 #define PERF_ATTACH_GROUP       0x02
 303 #define PERF_ATTACH_TASK        0x04
 304
 305 struct perf_cgroup;
 306 struct ring_buffer;
 307
 308 /**
 309  * struct perf_event - performance event kernel representation:
 310  */
 311 struct perf_event {
 312 #ifdef CONFIG_PERF_EVENTS
 313         /*
 314          * entry onto perf_event_context::event_list;
 315          *   modifications require ctx->lock
 316          *   RCU safe iterations.
 317          */
 318         struct list_head                event_entry;
 319
 320         /*
 321          * XXX: group_entry and sibling_list should be mutually exclusive;
 322          * either you're a sibling on a group, or you're the group leader.
 323          * Rework the code to always use the same list element.
 324          *
 325          * Locked for modification by both ctx->mutex and ctx->lock; holding
 326          * either sufficies for read.
 327          */
 328         struct list_head                group_entry;
 329         struct list_head                sibling_list;
 330
 331         /*
 332          * We need storage to track the entries in perf_pmu_migrate_context; we
 333          * cannot use the event_entry because of RCU and we want to keep the
 334          * group in tact which avoids us using the other two entries.
 335          */
 336         struct list_head                migrate_entry;
 337
 338         struct hlist_node               hlist_entry;
 339         struct list_head                active_entry;
 340         int                             nr_siblings;
 341         int                             group_flags;
 342         struct perf_event               *group_leader;
 343         struct pmu                      *pmu;
 344
 345         enum perf_event_active_state    state;
 346         unsigned int                    attach_state;
 347         local64_t                       count;
 348         atomic64_t                      child_count;
 349
 350         /*
 351          * These are the total time in nanoseconds that the event
 352          * has been enabled (i.e. eligible to run, and the task has
 353          * been scheduled in, if this is a per-task event)
 354          * and running (scheduled onto the CPU), respectively.
 355          *
 356          * They are computed from tstamp_enabled, tstamp_running and
 357          * tstamp_stopped when the event is in INACTIVE or ACTIVE state.
 358          */
 359         u64                             total_time_enabled;
 360         u64                             total_time_running;
 361
 362         /*
 363          * These are timestamps used for computing total_time_enabled
 364          * and total_time_running when the event is in INACTIVE or
 365          * ACTIVE state, measured in nanoseconds from an arbitrary point
 366          * in time.
 367          * tstamp_enabled: the notional time when the event was enabled
 368          * tstamp_running: the notional time when the event was scheduled on
 369          * tstamp_stopped: in INACTIVE state, the notional time when the
 370          *      event was scheduled off.
 371          */
 372         u64                             tstamp_enabled;
 373         u64                             tstamp_running;
 374         u64                             tstamp_stopped;
 375
 376         /*
 377          * timestamp shadows the actual context timing but it can
 378          * be safely used in NMI interrupt context. It reflects the
 379          * context time as it was when the event was last scheduled in.
 380          *
 381          * ctx_time already accounts for ctx->timestamp. Therefore to
 382          * compute ctx_time for a sample, simply add perf_clock().
 383          */
 384         u64                             shadow_ctx_time;
 385
 386         struct perf_event_attr          attr;
 387         u16                             header_size;
 388         u16                             id_header_size;
 389         u16                             read_size;
 390         struct hw_perf_event            hw;
 391
 392         struct perf_event_context       *ctx;
 393         atomic_long_t                   refcount;
 394
 395         /*
 396          * These accumulate total time (in nanoseconds) that children
 397          * events have been enabled and running, respectively.
 398          */
 399         atomic64_t                      child_total_time_enabled;
 400         atomic64_t                      child_total_time_running;
 401
 402         /*
 403          * Protect attach/detach and child_list:
 404          */
 405         struct mutex                    child_mutex;
 406         struct list_head                child_list;
 407         struct perf_event               *parent;
 408
 409         int                             oncpu;
 410         int                             cpu;
 411
 412         struct list_head                owner_entry;
 413         struct task_struct              *owner;
 414
 415         /* mmap bits */
 416         struct mutex                    mmap_mutex;
 417         atomic_t                        mmap_count;
 418
 419         struct ring_buffer              *rb;
 420         struct list_head                rb_entry;
 421         unsigned long                   rcu_batches;
 422         int                             rcu_pending;
 423
 424         /* poll related */
 425         wait_queue_head_t               waitq;
 426         struct fasync_struct            *fasync;
 427
 428         /* delayed work for NMIs and such */
 429         int                             pending_wakeup;
 430         int                             pending_kill;
 431         int                             pending_disable;
 432         struct irq_work                 pending;
 433
 434         atomic_t                        event_limit;
 435
 436         void (*destroy)(struct perf_event *);
 437         struct rcu_head                 rcu_head;
 438
 439         struct pid_namespace            *ns;
 440         u64                             id;
 441
 442         perf_overflow_handler_t         overflow_handler;
 443         void                            *overflow_handler_context;
 444
 445 #ifdef CONFIG_EVENT_TRACING
 446         struct ftrace_event_call        *tp_event;
 447         struct event_filter             *filter;
 448 #ifdef CONFIG_FUNCTION_TRACER
 449         struct ftrace_ops               ftrace_ops;
 450 #endif
 451 #endif
 452
 453 #ifdef CONFIG_CGROUP_PERF
 454         struct perf_cgroup              *cgrp; /* cgroup event is attach to */
 455         int                             cgrp_defer_enabled;
 456 #endif
 457
 458 #endif /* CONFIG_PERF_EVENTS */
 459 };
 460
 461 /**
 462  * struct perf_event_context - event context structure
 463  *
 464  * Used as a container for task events and CPU events as well:
 465  */
 466 struct perf_event_context {
 467         struct pmu                      *pmu;
 468         /*
 469          * Protect the states of the events in the list,
 470          * nr_active, and the list:
 471          */
 472         raw_spinlock_t                  lock;
 473         /*
 474          * Protect the list of events.  Locking either mutex or lock
 475          * is sufficient to ensure the list doesn't change; to change
 476          * the list you need to lock both the mutex and the spinlock.
 477          */
 478         struct mutex                    mutex;
 479
 480         struct list_head                pinned_groups;
 481         struct list_head                flexible_groups;
 482         struct list_head                event_list;
 483         int                             nr_events;
 484         int                             nr_active;
 485         int                             is_active;
 486         int                             nr_stat;
 487         int                             nr_freq;
 488         int                             rotate_disable;
 489         atomic_t                        refcount;
 490         struct task_struct              *task;
 491
 492         /*
 493          * Context clock, runs when context enabled.
 494          */
 495         u64                             time;
 496         u64                             timestamp;
 497
 498         /*
 499          * These fields let us detect when two contexts have both
 500          * been cloned (inherited) from a common ancestor.
 501          */
 502         struct perf_event_context       *parent_ctx;
 503         u64                             parent_gen;
 504         u64                             generation;
 505         int                             pin_count;
 506         int                             nr_cgroups;      /* cgroup evts */
 507         int                             nr_branch_stack; /* branch_stack evt */
 508         struct rcu_head                 rcu_head;
 509 };
 510
 511 /*
 512  * Number of contexts where an event can trigger:
 513  *      task, softirq, hardirq, nmi.
 514  */
 515 #define PERF_NR_CONTEXTS        4
 516
 517 /**
 518  * struct perf_event_cpu_context - per cpu event context structure
 519  */
 520 struct perf_cpu_context {
 521         struct perf_event_context       ctx;
 522         struct perf_event_context       *task_ctx;
 523         int                             active_oncpu;
 524         int                             exclusive;
 525         struct hrtimer                  hrtimer;
 526         ktime_t                         hrtimer_interval;
 527         struct list_head                rotation_list;
 528         struct pmu                      *unique_pmu;
 529         struct perf_cgroup              *cgrp;
 530 };
 531
 532 struct perf_output_handle {
 533         struct perf_event               *event;
 534         struct ring_buffer              *rb;
 535         unsigned long                   wakeup;
 536         unsigned long                   size;
 537         void                            *addr;
 538         int                             page;
 539 };
 540
 541 #ifdef CONFIG_PERF_EVENTS
 542
 543 extern int perf_pmu_register(struct pmu *pmu, const char *name, int type);
 544 extern void perf_pmu_unregister(struct pmu *pmu);
 545
 546 extern int perf_num_counters(void);
 547 extern const char *perf_pmu_name(void);
 548 extern void __perf_event_task_sched_in(struct task_struct *prev,
 549                                        struct task_struct *task);
 550 extern void __perf_event_task_sched_out(struct task_struct *prev,
 551                                         struct task_struct *next);
 552 extern int perf_event_init_task(struct task_struct *child);
 553 extern void perf_event_exit_task(struct task_struct *child);
 554 extern void perf_event_free_task(struct task_struct *task);
 555 extern void perf_event_delayed_put(struct task_struct *task);
 556 extern void perf_event_print_debug(void);
 557 extern void perf_pmu_disable(struct pmu *pmu);
 558 extern void perf_pmu_enable(struct pmu *pmu);
 559 extern int perf_event_task_disable(void);
 560 extern int perf_event_task_enable(void);
 561 extern int perf_event_refresh(struct perf_event *event, int refresh);
 562 extern void perf_event_update_userpage(struct perf_event *event);
 563 extern int perf_event_release_kernel(struct perf_event *event);
 564 extern struct perf_event *
 565 perf_event_create_kernel_counter(struct perf_event_attr *attr,
 566                                 int cpu,
 567                                 struct task_struct *task,
 568                                 perf_overflow_handler_t callback,
 569                                 void *context);
 570 extern void perf_pmu_migrate_context(struct pmu *pmu,
 571                                 int src_cpu, int dst_cpu);
 572 extern u64 perf_event_read_value(struct perf_event *event,
 573                                  u64 *enabled, u64 *running);
 574
 575
 576 struct perf_sample_data {
 577         u64                             type;
 578
 579         u64                             ip;
 580         struct {
 581                 u32     pid;
 582                 u32     tid;
 583         }                               tid_entry;
 584         u64                             time;
 585         u64                             addr;
 586         u64                             id;
 587         u64                             stream_id;
 588         struct {
 589                 u32     cpu;
 590                 u32     reserved;
 591         }                               cpu_entry;
 592         u64                             period;
 593         union  perf_mem_data_src        data_src;
 594         struct perf_callchain_entry     *callchain;
 595         struct perf_raw_record          *raw;
 596         struct perf_branch_stack        *br_stack;
 597         struct perf_regs_user           regs_user;
 598         u64                             stack_user_size;
 599         u64                             weight;
 600         /*
 601          * Transaction flags for abort events:
 602          */
 603         u64                             txn;
 604 };
 605
 606 static inline void perf_sample_data_init(struct perf_sample_data *data,
 607                                          u64 addr, u64 period)
 608 {
 609         /* remaining struct members initialized in perf_prepare_sample() */
 610         data->addr = addr;
 611         data->raw  = NULL;
 612         data->br_stack = NULL;
 613         data->period = period;
 614         data->regs_user.abi = PERF_SAMPLE_REGS_ABI_NONE;
 615         data->regs_user.regs = NULL;
 616         data->stack_user_size = 0;
 617         data->weight = 0;
 618         data->data_src.val = 0;
 619         data->txn = 0;
 620 }
 621
 622 extern void perf_output_sample(struct perf_output_handle *handle,
 623                                struct perf_event_header *header,
 624                                struct perf_sample_data *data,
 625                                struct perf_event *event);
 626 extern void perf_prepare_sample(struct perf_event_header *header,
 627                                 struct perf_sample_data *data,
 628                                 struct perf_event *event,
 629                                 struct pt_regs *regs);
 630
 631 extern int perf_event_overflow(struct perf_event *event,
 632                                  struct perf_sample_data *data,
 633                                  struct pt_regs *regs);
 634
 635 static inline bool is_sampling_event(struct perf_event *event)
 636 {
 637         return event->attr.sample_period != 0;
 638 }
 639
 640 /*
 641  * Return 1 for a software event, 0 for a hardware event
 642  */
 643 static inline int is_software_event(struct perf_event *event)
 644 {
 645         return event->pmu->task_ctx_nr == perf_sw_context;
 646 }
 647
 648 /*
 649  * Return 1 for event in sw context, 0 for event in hw context
 650  */
 651 static inline int in_software_context(struct perf_event *event)
 652 {
 653         return event->ctx->pmu->task_ctx_nr == perf_sw_context;
 654 }
 655
 656 extern struct static_key perf_swevent_enabled[PERF_COUNT_SW_MAX];
 657
 658 extern void ___perf_sw_event(u32, u64, struct pt_regs *, u64);
 659 extern void __perf_sw_event(u32, u64, struct pt_regs *, u64);
 660
 661 #ifndef perf_arch_fetch_caller_regs
 662 static inline void perf_arch_fetch_caller_regs(struct pt_regs *regs, unsigned long ip) { }
 663 #endif
 664
 665 /*
 666  * Take a snapshot of the regs. Skip ip and frame pointer to
 667  * the nth caller. We only need a few of the regs:
 668  * - ip for PERF_SAMPLE_IP
 669  * - cs for user_mode() tests
 670  * - bp for callchains
 671  * - eflags, for future purposes, just in case
 672  */
 673 static inline void perf_fetch_caller_regs(struct pt_regs *regs)
 674 {
 675         memset(regs, 0, sizeof(*regs));
 676
 677         perf_arch_fetch_caller_regs(regs, CALLER_ADDR0);
 678 }
 679
 680 static __always_inline void
 681 perf_sw_event(u32 event_id, u64 nr, struct pt_regs *regs, u64 addr)
 682 {
 683         if (static_key_false(&perf_swevent_enabled[event_id]))
 684                 __perf_sw_event(event_id, nr, regs, addr);
 685 }
 686
 687 DECLARE_PER_CPU(struct pt_regs, __perf_regs[4]);
 688
 689 /*
 690  * 'Special' version for the scheduler, it hard assumes no recursion,
 691  * which is guaranteed by us not actually scheduling inside other swevents
 692  * because those disable preemption.
 693  */
 694 static __always_inline void
 695 perf_sw_event_sched(u32 event_id, u64 nr, u64 addr)
 696 {
 697         if (static_key_false(&perf_swevent_enabled[event_id])) {
 698                 struct pt_regs *regs = this_cpu_ptr(&__perf_regs[0]);
 699
 700                 perf_fetch_caller_regs(regs);
 701                 ___perf_sw_event(event_id, nr, regs, addr);
 702         }
 703 }
 704
 705 extern struct static_key_deferred perf_sched_events;
 706
 707 static inline void perf_event_task_sched_in(struct task_struct *prev,
 708                                             struct task_struct *task)
 709 {
 710         if (static_key_false(&perf_sched_events.key))
 711                 __perf_event_task_sched_in(prev, task);
 712 }
 713
 714 static inline void perf_event_task_sched_out(struct task_struct *prev,
 715                                              struct task_struct *next)
 716 {
 717         perf_sw_event_sched(PERF_COUNT_SW_CONTEXT_SWITCHES, 1, 0);
 718
 719         if (static_key_false(&perf_sched_events.key))
 720                 __perf_event_task_sched_out(prev, next);
 721 }
 722
 723 extern void perf_event_mmap(struct vm_area_struct *vma);
 724 extern struct perf_guest_info_callbacks *perf_guest_cbs;
 725 extern int perf_register_guest_info_callbacks(struct perf_guest_info_callbacks *callbacks);
 726 extern int perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks *callbacks);
 727
 728 extern void perf_event_exec(void);
 729 extern void perf_event_comm(struct task_struct *tsk, bool exec);
 730 extern void perf_event_fork(struct task_struct *tsk);
 731
 732 /* Callchains */
 733 DECLARE_PER_CPU(struct perf_callchain_entry, perf_callchain_entry);
 734
 735 extern void perf_callchain_user(struct perf_callchain_entry *entry, struct pt_regs *regs);
 736 extern void perf_callchain_kernel(struct perf_callchain_entry *entry, struct pt_regs *regs);
 737
 738 static inline void perf_callchain_store(struct perf_callchain_entry *entry, u64 ip)
 739 {
 740         if (entry->nr < PERF_MAX_STACK_DEPTH)
 741                 entry->ip[entry->nr++] = ip;
 742 }
 743
 744 extern int sysctl_perf_event_paranoid;
 745 extern int sysctl_perf_event_mlock;
 746 extern int sysctl_perf_event_sample_rate;
 747 extern int sysctl_perf_cpu_time_max_percent;
 748
 749 extern void perf_sample_event_took(u64 sample_len_ns);
 750
 751 extern int perf_proc_update_handler(struct ctl_table *table, int write,
 752                 void __user *buffer, size_t *lenp,
 753                 loff_t *ppos);
 754 extern int perf_cpu_time_max_percent_handler(struct ctl_table *table, int write,
 755                 void __user *buffer, size_t *lenp,
 756                 loff_t *ppos);
 757
 758
 759 static inline bool perf_paranoid_tracepoint_raw(void)
 760 {
 761         return sysctl_perf_event_paranoid > -1;
 762 }
 763
 764 static inline bool perf_paranoid_cpu(void)
 765 {
 766         return sysctl_perf_event_paranoid > 0;
 767 }
 768
 769 static inline bool perf_paranoid_kernel(void)
 770 {
 771         return sysctl_perf_event_paranoid > 1;
 772 }
 773
 774 extern void perf_event_init(void);
 775 extern void perf_tp_event(u64 addr, u64 count, void *record,
 776                           int entry_size, struct pt_regs *regs,
 777                           struct hlist_head *head, int rctx,
 778                           struct task_struct *task);
 779 extern void perf_bp_event(struct perf_event *event, void *data);
 780
 781 #ifndef perf_misc_flags
 782 # define perf_misc_flags(regs) \
 783                 (user_mode(regs) ? PERF_RECORD_MISC_USER : PERF_RECORD_MISC_KERNEL)
 784 # define perf_instruction_pointer(regs) instruction_pointer(regs)
 785 #endif
 786
 787 static inline bool has_branch_stack(struct perf_event *event)
 788 {
 789         return event->attr.sample_type & PERF_SAMPLE_BRANCH_STACK;
 790 }
 791
 792 extern int perf_output_begin(struct perf_output_handle *handle,
 793                              struct perf_event *event, unsigned int size);
 794 extern void perf_output_end(struct perf_output_handle *handle);
 795 extern unsigned int perf_output_copy(struct perf_output_handle *handle,
 796                              const void *buf, unsigned int len);
 797 extern unsigned int perf_output_skip(struct perf_output_handle *handle,
 798                                      unsigned int len);
 799 extern int perf_swevent_get_recursion_context(void);
 800 extern void perf_swevent_put_recursion_context(int rctx);
 801 extern u64 perf_swevent_set_period(struct perf_event *event);
 802 extern void perf_event_enable(struct perf_event *event);
 803 extern void perf_event_disable(struct perf_event *event);
 804 extern int __perf_event_disable(void *info);
 805 extern void perf_event_task_tick(void);
 806 #else /* !CONFIG_PERF_EVENTS: */
 807 static inline void
 808 perf_event_task_sched_in(struct task_struct *prev,
 809                          struct task_struct *task)                      { }
 810 static inline void
 811 perf_event_task_sched_out(struct task_struct *prev,
 812                           struct task_struct *next)                     { }
 813 static inline int perf_event_init_task(struct task_struct *child)       { return 0; }
 814 static inline void perf_event_exit_task(struct task_struct *child)      { }
 815 static inline void perf_event_free_task(struct task_struct *task)       { }
 816 static inline void perf_event_delayed_put(struct task_struct *task)     { }
 817 static inline void perf_event_print_debug(void)                         { }
 818 static inline int perf_event_task_disable(void)                         { return -EINVAL; }
 819 static inline int perf_event_task_enable(void)                          { return -EINVAL; }
 820 static inline int perf_event_refresh(struct perf_event *event, int refresh)
 821 {
 822         return -EINVAL;
 823 }
 824
 825 static inline void
 826 perf_sw_event(u32 event_id, u64 nr, struct pt_regs *regs, u64 addr)     { }
 827 static inline void
 828 perf_sw_event_sched(u32 event_id, u64 nr, u64 addr)                     { }
 829 static inline void
 830 perf_bp_event(struct perf_event *event, void *data)                     { }
 831
 832 static inline int perf_register_guest_info_callbacks
 833 (struct perf_guest_info_callbacks *callbacks)                           { return 0; }
 834 static inline int perf_unregister_guest_info_callbacks
 835 (struct perf_guest_info_callbacks *callbacks)                           { return 0; }
 836
 837 static inline void perf_event_mmap(struct vm_area_struct *vma)          { }
 838 static inline void perf_event_exec(void)                                { }
 839 static inline void perf_event_comm(struct task_struct *tsk, bool exec)  { }
 840 static inline void perf_event_fork(struct task_struct *tsk)             { }
 841 static inline void perf_event_init(void)                                { }
 842 static inline int  perf_swevent_get_recursion_context(void)             { return -1; }
 843 static inline void perf_swevent_put_recursion_context(int rctx)         { }
 844 static inline u64 perf_swevent_set_period(struct perf_event *event)     { return 0; }
 845 static inline void perf_event_enable(struct perf_event *event)          { }
 846 static inline void perf_event_disable(struct perf_event *event)         { }
 847 static inline int __perf_event_disable(void *info)                      { return -1; }
 848 static inline void perf_event_task_tick(void)                           { }
 849 #endif
 850
 851 #if defined(CONFIG_PERF_EVENTS) && defined(CONFIG_NO_HZ_FULL)
 852 extern bool perf_event_can_stop_tick(void);
 853 #else
 854 static inline bool perf_event_can_stop_tick(void)                       { return true; }
 855 #endif
 856
 857 #if defined(CONFIG_PERF_EVENTS) && defined(CONFIG_CPU_SUP_INTEL)
 858 extern void perf_restore_debug_store(void);
 859 #else
 860 static inline void perf_restore_debug_store(void)                       { }
 861 #endif
 862
 863 #define perf_output_put(handle, x) perf_output_copy((handle), &(x), sizeof(x))
 864
 865 /*
 866  * This has to have a higher priority than migration_notifier in sched/core.c.
 867  */
 868 #define perf_cpu_notifier(fn)                                           \
 869 do {                                                                    \
 870         static struct notifier_block fn##_nb =                          \
 871                 { .notifier_call = fn, .priority = CPU_PRI_PERF };      \
 872         unsigned long cpu = smp_processor_id();                         \
 873         unsigned long flags;                                            \
 874                                                                         \
 875         cpu_notifier_register_begin();                                  \
 876         fn(&fn##_nb, (unsigned long)CPU_UP_PREPARE,                     \
 877                 (void *)(unsigned long)cpu);                            \
 878         local_irq_save(flags);                                          \
 879         fn(&fn##_nb, (unsigned long)CPU_STARTING,                       \
 880                 (void *)(unsigned long)cpu);                            \
 881         local_irq_restore(flags);                                       \
 882         fn(&fn##_nb, (unsigned long)CPU_ONLINE,                         \
 883                 (void *)(unsigned long)cpu);                            \
 884         __register_cpu_notifier(&fn##_nb);                              \
 885         cpu_notifier_register_done();                                   \
 886 } while (0)
 887
 888 /*
 889  * Bare-bones version of perf_cpu_notifier(), which doesn't invoke the
 890  * callback for already online CPUs.
 891  */
 892 #define __perf_cpu_notifier(fn)                                         \
 893 do {                                                                    \
 894         static struct notifier_block fn##_nb =                          \
 895                 { .notifier_call = fn, .priority = CPU_PRI_PERF };      \
 896                                                                         \
 897         __register_cpu_notifier(&fn##_nb);                              \
 898 } while (0)
 899
 900 struct perf_pmu_events_attr {
 901         struct device_attribute attr;
 902         u64 id;
 903         const char *event_str;
 904 };
 905
 906 #define PMU_EVENT_ATTR(_name, _var, _id, _show)                         \
 907 static struct perf_pmu_events_attr _var = {                             \
 908         .attr = __ATTR(_name, 0444, _show, NULL),                       \
 909         .id   =  _id,                                                   \
 910 };
 911
 912 #define PMU_FORMAT_ATTR(_name, _format)                                 \
 913 static ssize_t                                                          \
 914 _name##_show(struct device *dev,                                        \
 915                                struct device_attribute *attr,           \
 916                                char *page)                              \
 917 {                                                                       \
 918         BUILD_BUG_ON(sizeof(_format) >= PAGE_SIZE);                     \
 919         return sprintf(page, _format "\n");                             \
 920 }                                                                       \
 921                                                                         \
 922 static struct device_attribute format_attr_##_name = __ATTR_RO(_name)
 923
 924 #endif /* _LINUX_PERF_EVENT_H */