1 // SPDX-License-Identifier: GPL-2.0
3 * Resource Director Technology (RDT)
5 * Pseudo-locking support built on top of Cache Allocation Technology (CAT)
7 * Copyright (C) 2018 Intel Corporation
9 * Author: Reinette Chatre <reinette.chatre@intel.com>
12 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
14 #include <linux/cacheinfo.h>
15 #include <linux/cpu.h>
16 #include <linux/cpumask.h>
17 #include <linux/debugfs.h>
18 #include <linux/kthread.h>
19 #include <linux/mman.h>
20 #include <linux/perf_event.h>
21 #include <linux/pm_qos.h>
22 #include <linux/slab.h>
23 #include <linux/uaccess.h>
25 #include <asm/cacheflush.h>
26 #include <asm/intel-family.h>
27 #include <asm/resctrl_sched.h>
28 #include <asm/perf_event.h>
30 #include "../../events/perf_event.h" /* For X86_CONFIG() */
33 #define CREATE_TRACE_POINTS
34 #include "pseudo_lock_event.h"
37 * The bits needed to disable hardware prefetching varies based on the
38 * platform. During initialization we will discover which bits to use.
40 static u64 prefetch_disable_bits;
43 * Major number assigned to and shared by all devices exposing
44 * pseudo-locked regions.
46 static unsigned int pseudo_lock_major;
47 static unsigned long pseudo_lock_minor_avail = GENMASK(MINORBITS, 0);
48 static struct class *pseudo_lock_class;
51 * get_prefetch_disable_bits - prefetch disable bits of supported platforms
53 * Capture the list of platforms that have been validated to support
54 * pseudo-locking. This includes testing to ensure pseudo-locked regions
55 * with low cache miss rates can be created under variety of load conditions
56 * as well as that these pseudo-locked regions can maintain their low cache
57 * miss rates under variety of load conditions for significant lengths of time.
59 * After a platform has been validated to support pseudo-locking its
60 * hardware prefetch disable bits are included here as they are documented
63 * When adding a platform here also add support for its cache events to
64 * measure_cycles_perf_fn()
67 * If platform is supported, the bits to disable hardware prefetchers, 0
68 * if platform is not supported.
70 static u64 get_prefetch_disable_bits(void)
72 if (boot_cpu_data.x86_vendor != X86_VENDOR_INTEL ||
73 boot_cpu_data.x86 != 6)
76 switch (boot_cpu_data.x86_model) {
77 case INTEL_FAM6_BROADWELL_X:
79 * SDM defines bits of MSR_MISC_FEATURE_CONTROL register
81 * 0 L2 Hardware Prefetcher Disable (R/W)
82 * 1 L2 Adjacent Cache Line Prefetcher Disable (R/W)
83 * 2 DCU Hardware Prefetcher Disable (R/W)
84 * 3 DCU IP Prefetcher Disable (R/W)
88 case INTEL_FAM6_ATOM_GOLDMONT:
89 case INTEL_FAM6_ATOM_GOLDMONT_PLUS:
91 * SDM defines bits of MSR_MISC_FEATURE_CONTROL register
93 * 0 L2 Hardware Prefetcher Disable (R/W)
95 * 2 DCU Hardware Prefetcher Disable (R/W)
105 * pseudo_lock_minor_get - Obtain available minor number
106 * @minor: Pointer to where new minor number will be stored
108 * A bitmask is used to track available minor numbers. Here the next free
109 * minor number is marked as unavailable and returned.
111 * Return: 0 on success, <0 on failure.
113 static int pseudo_lock_minor_get(unsigned int *minor)
115 unsigned long first_bit;
117 first_bit = find_first_bit(&pseudo_lock_minor_avail, MINORBITS);
119 if (first_bit == MINORBITS)
122 __clear_bit(first_bit, &pseudo_lock_minor_avail);
129 * pseudo_lock_minor_release - Return minor number to available
130 * @minor: The minor number made available
132 static void pseudo_lock_minor_release(unsigned int minor)
134 __set_bit(minor, &pseudo_lock_minor_avail);
138 * region_find_by_minor - Locate a pseudo-lock region by inode minor number
139 * @minor: The minor number of the device representing pseudo-locked region
141 * When the character device is accessed we need to determine which
142 * pseudo-locked region it belongs to. This is done by matching the minor
143 * number of the device to the pseudo-locked region it belongs.
145 * Minor numbers are assigned at the time a pseudo-locked region is associated
146 * with a cache instance.
148 * Return: On success return pointer to resource group owning the pseudo-locked
149 * region, NULL on failure.
151 static struct rdtgroup *region_find_by_minor(unsigned int minor)
153 struct rdtgroup *rdtgrp, *rdtgrp_match = NULL;
155 list_for_each_entry(rdtgrp, &rdt_all_groups, rdtgroup_list) {
156 if (rdtgrp->plr && rdtgrp->plr->minor == minor) {
157 rdtgrp_match = rdtgrp;
165 * pseudo_lock_pm_req - A power management QoS request list entry
166 * @list: Entry within the @pm_reqs list for a pseudo-locked region
167 * @req: PM QoS request
169 struct pseudo_lock_pm_req {
170 struct list_head list;
171 struct dev_pm_qos_request req;
174 static void pseudo_lock_cstates_relax(struct pseudo_lock_region *plr)
176 struct pseudo_lock_pm_req *pm_req, *next;
178 list_for_each_entry_safe(pm_req, next, &plr->pm_reqs, list) {
179 dev_pm_qos_remove_request(&pm_req->req);
180 list_del(&pm_req->list);
186 * pseudo_lock_cstates_constrain - Restrict cores from entering C6
188 * To prevent the cache from being affected by power management entering
189 * C6 has to be avoided. This is accomplished by requesting a latency
190 * requirement lower than lowest C6 exit latency of all supported
191 * platforms as found in the cpuidle state tables in the intel_idle driver.
192 * At this time it is possible to do so with a single latency requirement
193 * for all supported platforms.
195 * Since Goldmont is supported, which is affected by X86_BUG_MONITOR,
196 * the ACPI latencies need to be considered while keeping in mind that C2
197 * may be set to map to deeper sleep states. In this case the latency
198 * requirement needs to prevent entering C2 also.
200 static int pseudo_lock_cstates_constrain(struct pseudo_lock_region *plr)
202 struct pseudo_lock_pm_req *pm_req;
206 for_each_cpu(cpu, &plr->d->cpu_mask) {
207 pm_req = kzalloc(sizeof(*pm_req), GFP_KERNEL);
209 rdt_last_cmd_puts("Failure to allocate memory for PM QoS\n");
213 ret = dev_pm_qos_add_request(get_cpu_device(cpu),
215 DEV_PM_QOS_RESUME_LATENCY,
218 rdt_last_cmd_printf("Failed to add latency req CPU%d\n",
224 list_add(&pm_req->list, &plr->pm_reqs);
230 pseudo_lock_cstates_relax(plr);
235 * pseudo_lock_region_clear - Reset pseudo-lock region data
236 * @plr: pseudo-lock region
238 * All content of the pseudo-locked region is reset - any memory allocated
243 static void pseudo_lock_region_clear(struct pseudo_lock_region *plr)
254 plr->debugfs_dir = NULL;
258 * pseudo_lock_region_init - Initialize pseudo-lock region information
259 * @plr: pseudo-lock region
261 * Called after user provided a schemata to be pseudo-locked. From the
262 * schemata the &struct pseudo_lock_region is on entry already initialized
263 * with the resource, domain, and capacity bitmask. Here the information
264 * required for pseudo-locking is deduced from this data and &struct
265 * pseudo_lock_region initialized further. This information includes:
266 * - size in bytes of the region to be pseudo-locked
267 * - cache line size to know the stride with which data needs to be accessed
268 * to be pseudo-locked
269 * - a cpu associated with the cache instance on which the pseudo-locking
270 * flow can be executed
272 * Return: 0 on success, <0 on failure. Descriptive error will be written
273 * to last_cmd_status buffer.
275 static int pseudo_lock_region_init(struct pseudo_lock_region *plr)
277 struct cpu_cacheinfo *ci;
281 /* Pick the first cpu we find that is associated with the cache. */
282 plr->cpu = cpumask_first(&plr->d->cpu_mask);
284 if (!cpu_online(plr->cpu)) {
285 rdt_last_cmd_printf("CPU %u associated with cache not online\n",
291 ci = get_cpu_cacheinfo(plr->cpu);
293 plr->size = rdtgroup_cbm_to_size(plr->r, plr->d, plr->cbm);
295 for (i = 0; i < ci->num_leaves; i++) {
296 if (ci->info_list[i].level == plr->r->cache_level) {
297 plr->line_size = ci->info_list[i].coherency_line_size;
303 rdt_last_cmd_puts("Unable to determine cache line size\n");
305 pseudo_lock_region_clear(plr);
310 * pseudo_lock_init - Initialize a pseudo-lock region
311 * @rdtgrp: resource group to which new pseudo-locked region will belong
313 * A pseudo-locked region is associated with a resource group. When this
314 * association is created the pseudo-locked region is initialized. The
315 * details of the pseudo-locked region are not known at this time so only
316 * allocation is done and association established.
318 * Return: 0 on success, <0 on failure
320 static int pseudo_lock_init(struct rdtgroup *rdtgrp)
322 struct pseudo_lock_region *plr;
324 plr = kzalloc(sizeof(*plr), GFP_KERNEL);
328 init_waitqueue_head(&plr->lock_thread_wq);
329 INIT_LIST_HEAD(&plr->pm_reqs);
335 * pseudo_lock_region_alloc - Allocate kernel memory that will be pseudo-locked
336 * @plr: pseudo-lock region
338 * Initialize the details required to set up the pseudo-locked region and
339 * allocate the contiguous memory that will be pseudo-locked to the cache.
341 * Return: 0 on success, <0 on failure. Descriptive error will be written
342 * to last_cmd_status buffer.
344 static int pseudo_lock_region_alloc(struct pseudo_lock_region *plr)
348 ret = pseudo_lock_region_init(plr);
353 * We do not yet support contiguous regions larger than
356 if (plr->size > KMALLOC_MAX_SIZE) {
357 rdt_last_cmd_puts("Requested region exceeds maximum size\n");
362 plr->kmem = kzalloc(plr->size, GFP_KERNEL);
364 rdt_last_cmd_puts("Unable to allocate memory\n");
372 pseudo_lock_region_clear(plr);
378 * pseudo_lock_free - Free a pseudo-locked region
379 * @rdtgrp: resource group to which pseudo-locked region belonged
381 * The pseudo-locked region's resources have already been released, or not
382 * yet created at this point. Now it can be freed and disassociated from the
387 static void pseudo_lock_free(struct rdtgroup *rdtgrp)
389 pseudo_lock_region_clear(rdtgrp->plr);
395 * pseudo_lock_fn - Load kernel memory into cache
396 * @_rdtgrp: resource group to which pseudo-lock region belongs
398 * This is the core pseudo-locking flow.
400 * First we ensure that the kernel memory cannot be found in the cache.
401 * Then, while taking care that there will be as little interference as
402 * possible, the memory to be loaded is accessed while core is running
403 * with class of service set to the bitmask of the pseudo-locked region.
404 * After this is complete no future CAT allocations will be allowed to
405 * overlap with this bitmask.
407 * Local register variables are utilized to ensure that the memory region
408 * to be locked is the only memory access made during the critical locking
411 * Return: 0. Waiter on waitqueue will be woken on completion.
413 static int pseudo_lock_fn(void *_rdtgrp)
415 struct rdtgroup *rdtgrp = _rdtgrp;
416 struct pseudo_lock_region *plr = rdtgrp->plr;
417 u32 rmid_p, closid_p;
421 * The registers used for local register variables are also used
422 * when KASAN is active. When KASAN is active we use a regular
423 * variable to ensure we always use a valid pointer, but the cost
424 * is that this variable will enter the cache through evicting the
425 * memory we are trying to lock into the cache. Thus expect lower
426 * pseudo-locking success rate when KASAN is active.
428 unsigned int line_size;
432 register unsigned int line_size asm("esi");
433 register unsigned int size asm("edi");
434 register void *mem_r asm(_ASM_BX);
435 #endif /* CONFIG_KASAN */
438 * Make sure none of the allocated memory is cached. If it is we
439 * will get a cache hit in below loop from outside of pseudo-locked
441 * wbinvd (as opposed to clflush/clflushopt) is required to
442 * increase likelihood that allocated cache portion will be filled
443 * with associated memory.
448 * Always called with interrupts enabled. By disabling interrupts
449 * ensure that we will not be preempted during this critical section.
454 * Call wrmsr and rdmsr as directly as possible to avoid tracing
455 * clobbering local register variables or affecting cache accesses.
457 * Disable the hardware prefetcher so that when the end of the memory
458 * being pseudo-locked is reached the hardware will not read beyond
459 * the buffer and evict pseudo-locked memory read earlier from the
462 __wrmsr(MSR_MISC_FEATURE_CONTROL, prefetch_disable_bits, 0x0);
463 closid_p = this_cpu_read(pqr_state.cur_closid);
464 rmid_p = this_cpu_read(pqr_state.cur_rmid);
467 line_size = plr->line_size;
469 * Critical section begin: start by writing the closid associated
470 * with the capacity bitmask of the cache region being
471 * pseudo-locked followed by reading of kernel memory to load it
474 __wrmsr(IA32_PQR_ASSOC, rmid_p, rdtgrp->closid);
476 * Cache was flushed earlier. Now access kernel memory to read it
477 * into cache region associated with just activated plr->closid.
478 * Loop over data twice:
479 * - In first loop the cache region is shared with the page walker
480 * as it populates the paging structure caches (including TLB).
481 * - In the second loop the paging structure caches are used and
482 * cache region is populated with the memory being referenced.
484 for (i = 0; i < size; i += PAGE_SIZE) {
486 * Add a barrier to prevent speculative execution of this
487 * loop reading beyond the end of the buffer.
490 asm volatile("mov (%0,%1,1), %%eax\n\t"
492 : "r" (mem_r), "r" (i)
495 for (i = 0; i < size; i += line_size) {
497 * Add a barrier to prevent speculative execution of this
498 * loop reading beyond the end of the buffer.
501 asm volatile("mov (%0,%1,1), %%eax\n\t"
503 : "r" (mem_r), "r" (i)
507 * Critical section end: restore closid with capacity bitmask that
508 * does not overlap with pseudo-locked region.
510 __wrmsr(IA32_PQR_ASSOC, rmid_p, closid_p);
512 /* Re-enable the hardware prefetcher(s) */
513 wrmsr(MSR_MISC_FEATURE_CONTROL, 0x0, 0x0);
516 plr->thread_done = 1;
517 wake_up_interruptible(&plr->lock_thread_wq);
522 * rdtgroup_monitor_in_progress - Test if monitoring in progress
523 * @r: resource group being queried
525 * Return: 1 if monitor groups have been created for this resource
526 * group, 0 otherwise.
528 static int rdtgroup_monitor_in_progress(struct rdtgroup *rdtgrp)
530 return !list_empty(&rdtgrp->mon.crdtgrp_list);
534 * rdtgroup_locksetup_user_restrict - Restrict user access to group
535 * @rdtgrp: resource group needing access restricted
537 * A resource group used for cache pseudo-locking cannot have cpus or tasks
538 * assigned to it. This is communicated to the user by restricting access
539 * to all the files that can be used to make such changes.
541 * Permissions restored with rdtgroup_locksetup_user_restore()
543 * Return: 0 on success, <0 on failure. If a failure occurs during the
544 * restriction of access an attempt will be made to restore permissions but
545 * the state of the mode of these files will be uncertain when a failure
548 static int rdtgroup_locksetup_user_restrict(struct rdtgroup *rdtgrp)
552 ret = rdtgroup_kn_mode_restrict(rdtgrp, "tasks");
556 ret = rdtgroup_kn_mode_restrict(rdtgrp, "cpus");
560 ret = rdtgroup_kn_mode_restrict(rdtgrp, "cpus_list");
564 if (rdt_mon_capable) {
565 ret = rdtgroup_kn_mode_restrict(rdtgrp, "mon_groups");
574 rdtgroup_kn_mode_restore(rdtgrp, "cpus_list", 0777);
576 rdtgroup_kn_mode_restore(rdtgrp, "cpus", 0777);
578 rdtgroup_kn_mode_restore(rdtgrp, "tasks", 0777);
584 * rdtgroup_locksetup_user_restore - Restore user access to group
585 * @rdtgrp: resource group needing access restored
587 * Restore all file access previously removed using
588 * rdtgroup_locksetup_user_restrict()
590 * Return: 0 on success, <0 on failure. If a failure occurs during the
591 * restoration of access an attempt will be made to restrict permissions
592 * again but the state of the mode of these files will be uncertain when
595 static int rdtgroup_locksetup_user_restore(struct rdtgroup *rdtgrp)
599 ret = rdtgroup_kn_mode_restore(rdtgrp, "tasks", 0777);
603 ret = rdtgroup_kn_mode_restore(rdtgrp, "cpus", 0777);
607 ret = rdtgroup_kn_mode_restore(rdtgrp, "cpus_list", 0777);
611 if (rdt_mon_capable) {
612 ret = rdtgroup_kn_mode_restore(rdtgrp, "mon_groups", 0777);
621 rdtgroup_kn_mode_restrict(rdtgrp, "cpus_list");
623 rdtgroup_kn_mode_restrict(rdtgrp, "cpus");
625 rdtgroup_kn_mode_restrict(rdtgrp, "tasks");
631 * rdtgroup_locksetup_enter - Resource group enters locksetup mode
632 * @rdtgrp: resource group requested to enter locksetup mode
634 * A resource group enters locksetup mode to reflect that it would be used
635 * to represent a pseudo-locked region and is in the process of being set
636 * up to do so. A resource group used for a pseudo-locked region would
637 * lose the closid associated with it so we cannot allow it to have any
638 * tasks or cpus assigned nor permit tasks or cpus to be assigned in the
639 * future. Monitoring of a pseudo-locked region is not allowed either.
641 * The above and more restrictions on a pseudo-locked region are checked
642 * for and enforced before the resource group enters the locksetup mode.
644 * Returns: 0 if the resource group successfully entered locksetup mode, <0
645 * on failure. On failure the last_cmd_status buffer is updated with text to
646 * communicate details of failure to the user.
648 int rdtgroup_locksetup_enter(struct rdtgroup *rdtgrp)
653 * The default resource group can neither be removed nor lose the
654 * default closid associated with it.
656 if (rdtgrp == &rdtgroup_default) {
657 rdt_last_cmd_puts("Cannot pseudo-lock default group\n");
662 * Cache Pseudo-locking not supported when CDP is enabled.
664 * Some things to consider if you would like to enable this
665 * support (using L3 CDP as example):
666 * - When CDP is enabled two separate resources are exposed,
667 * L3DATA and L3CODE, but they are actually on the same cache.
668 * The implication for pseudo-locking is that if a
669 * pseudo-locked region is created on a domain of one
670 * resource (eg. L3CODE), then a pseudo-locked region cannot
671 * be created on that same domain of the other resource
672 * (eg. L3DATA). This is because the creation of a
673 * pseudo-locked region involves a call to wbinvd that will
674 * affect all cache allocations on particular domain.
675 * - Considering the previous, it may be possible to only
676 * expose one of the CDP resources to pseudo-locking and
677 * hide the other. For example, we could consider to only
678 * expose L3DATA and since the L3 cache is unified it is
679 * still possible to place instructions there are execute it.
680 * - If only one region is exposed to pseudo-locking we should
681 * still keep in mind that availability of a portion of cache
682 * for pseudo-locking should take into account both resources.
683 * Similarly, if a pseudo-locked region is created in one
684 * resource, the portion of cache used by it should be made
685 * unavailable to all future allocations from both resources.
687 if (rdt_resources_all[RDT_RESOURCE_L3DATA].alloc_enabled ||
688 rdt_resources_all[RDT_RESOURCE_L2DATA].alloc_enabled) {
689 rdt_last_cmd_puts("CDP enabled\n");
694 * Not knowing the bits to disable prefetching implies that this
695 * platform does not support Cache Pseudo-Locking.
697 prefetch_disable_bits = get_prefetch_disable_bits();
698 if (prefetch_disable_bits == 0) {
699 rdt_last_cmd_puts("Pseudo-locking not supported\n");
703 if (rdtgroup_monitor_in_progress(rdtgrp)) {
704 rdt_last_cmd_puts("Monitoring in progress\n");
708 if (rdtgroup_tasks_assigned(rdtgrp)) {
709 rdt_last_cmd_puts("Tasks assigned to resource group\n");
713 if (!cpumask_empty(&rdtgrp->cpu_mask)) {
714 rdt_last_cmd_puts("CPUs assigned to resource group\n");
718 if (rdtgroup_locksetup_user_restrict(rdtgrp)) {
719 rdt_last_cmd_puts("Unable to modify resctrl permissions\n");
723 ret = pseudo_lock_init(rdtgrp);
725 rdt_last_cmd_puts("Unable to init pseudo-lock region\n");
730 * If this system is capable of monitoring a rmid would have been
731 * allocated when the control group was created. This is not needed
732 * anymore when this group would be used for pseudo-locking. This
733 * is safe to call on platforms not capable of monitoring.
735 free_rmid(rdtgrp->mon.rmid);
741 rdtgroup_locksetup_user_restore(rdtgrp);
747 * rdtgroup_locksetup_exit - resource group exist locksetup mode
748 * @rdtgrp: resource group
750 * When a resource group exits locksetup mode the earlier restrictions are
753 * Return: 0 on success, <0 on failure
755 int rdtgroup_locksetup_exit(struct rdtgroup *rdtgrp)
759 if (rdt_mon_capable) {
762 rdt_last_cmd_puts("Out of RMIDs\n");
765 rdtgrp->mon.rmid = ret;
768 ret = rdtgroup_locksetup_user_restore(rdtgrp);
770 free_rmid(rdtgrp->mon.rmid);
774 pseudo_lock_free(rdtgrp);
779 * rdtgroup_cbm_overlaps_pseudo_locked - Test if CBM or portion is pseudo-locked
783 * @d represents a cache instance and @cbm a capacity bitmask that is
784 * considered for it. Determine if @cbm overlaps with any existing
785 * pseudo-locked region on @d.
787 * @cbm is unsigned long, even if only 32 bits are used, to make the
788 * bitmap functions work correctly.
790 * Return: true if @cbm overlaps with pseudo-locked region on @d, false
793 bool rdtgroup_cbm_overlaps_pseudo_locked(struct rdt_domain *d, unsigned long cbm)
795 unsigned int cbm_len;
799 cbm_len = d->plr->r->cache.cbm_len;
801 if (bitmap_intersects(&cbm, &cbm_b, cbm_len))
808 * rdtgroup_pseudo_locked_in_hierarchy - Pseudo-locked region in cache hierarchy
809 * @d: RDT domain under test
811 * The setup of a pseudo-locked region affects all cache instances within
812 * the hierarchy of the region. It is thus essential to know if any
813 * pseudo-locked regions exist within a cache hierarchy to prevent any
814 * attempts to create new pseudo-locked regions in the same hierarchy.
816 * Return: true if a pseudo-locked region exists in the hierarchy of @d or
817 * if it is not possible to test due to memory allocation issue,
820 bool rdtgroup_pseudo_locked_in_hierarchy(struct rdt_domain *d)
822 cpumask_var_t cpu_with_psl;
823 struct rdt_resource *r;
824 struct rdt_domain *d_i;
827 if (!zalloc_cpumask_var(&cpu_with_psl, GFP_KERNEL))
831 * First determine which cpus have pseudo-locked regions
832 * associated with them.
834 for_each_alloc_enabled_rdt_resource(r) {
835 list_for_each_entry(d_i, &r->domains, list) {
837 cpumask_or(cpu_with_psl, cpu_with_psl,
843 * Next test if new pseudo-locked region would intersect with
846 if (cpumask_intersects(&d->cpu_mask, cpu_with_psl))
849 free_cpumask_var(cpu_with_psl);
854 * measure_cycles_lat_fn - Measure cycle latency to read pseudo-locked memory
855 * @_plr: pseudo-lock region to measure
857 * There is no deterministic way to test if a memory region is cached. One
858 * way is to measure how long it takes to read the memory, the speed of
859 * access is a good way to learn how close to the cpu the data was. Even
860 * more, if the prefetcher is disabled and the memory is read at a stride
861 * of half the cache line, then a cache miss will be easy to spot since the
862 * read of the first half would be significantly slower than the read of
865 * Return: 0. Waiter on waitqueue will be woken on completion.
867 static int measure_cycles_lat_fn(void *_plr)
869 struct pseudo_lock_region *plr = _plr;
876 * Disable hardware prefetchers.
878 wrmsr(MSR_MISC_FEATURE_CONTROL, prefetch_disable_bits, 0x0);
879 mem_r = READ_ONCE(plr->kmem);
881 * Dummy execute of the time measurement to load the needed
882 * instructions into the L1 instruction cache.
884 start = rdtsc_ordered();
885 for (i = 0; i < plr->size; i += 32) {
886 start = rdtsc_ordered();
887 asm volatile("mov (%0,%1,1), %%eax\n\t"
889 : "r" (mem_r), "r" (i)
891 end = rdtsc_ordered();
892 trace_pseudo_lock_mem_latency((u32)(end - start));
894 wrmsr(MSR_MISC_FEATURE_CONTROL, 0x0, 0x0);
896 plr->thread_done = 1;
897 wake_up_interruptible(&plr->lock_thread_wq);
902 * Create a perf_event_attr for the hit and miss perf events that will
903 * be used during the performance measurement. A perf_event maintains
904 * a pointer to its perf_event_attr so a unique attribute structure is
905 * created for each perf_event.
907 * The actual configuration of the event is set right before use in order
908 * to use the X86_CONFIG macro.
910 static struct perf_event_attr perf_miss_attr = {
911 .type = PERF_TYPE_RAW,
912 .size = sizeof(struct perf_event_attr),
918 static struct perf_event_attr perf_hit_attr = {
919 .type = PERF_TYPE_RAW,
920 .size = sizeof(struct perf_event_attr),
926 struct residency_counts {
927 u64 miss_before, hits_before;
928 u64 miss_after, hits_after;
931 static int measure_residency_fn(struct perf_event_attr *miss_attr,
932 struct perf_event_attr *hit_attr,
933 struct pseudo_lock_region *plr,
934 struct residency_counts *counts)
936 u64 hits_before = 0, hits_after = 0, miss_before = 0, miss_after = 0;
937 struct perf_event *miss_event, *hit_event;
938 int hit_pmcnum, miss_pmcnum;
939 unsigned int line_size;
945 miss_event = perf_event_create_kernel_counter(miss_attr, plr->cpu,
947 if (IS_ERR(miss_event))
950 hit_event = perf_event_create_kernel_counter(hit_attr, plr->cpu,
952 if (IS_ERR(hit_event))
957 * Check any possible error state of events used by performing
960 if (perf_event_read_local(miss_event, &tmp, NULL, NULL)) {
964 if (perf_event_read_local(hit_event, &tmp, NULL, NULL)) {
970 * Disable hardware prefetchers.
972 wrmsr(MSR_MISC_FEATURE_CONTROL, prefetch_disable_bits, 0x0);
974 /* Initialize rest of local variables */
976 * Performance event has been validated right before this with
977 * interrupts disabled - it is thus safe to read the counter index.
979 miss_pmcnum = x86_perf_rdpmc_index(miss_event);
980 hit_pmcnum = x86_perf_rdpmc_index(hit_event);
981 line_size = READ_ONCE(plr->line_size);
982 mem_r = READ_ONCE(plr->kmem);
983 size = READ_ONCE(plr->size);
986 * Read counter variables twice - first to load the instructions
987 * used in L1 cache, second to capture accurate value that does not
988 * include cache misses incurred because of instruction loads.
990 rdpmcl(hit_pmcnum, hits_before);
991 rdpmcl(miss_pmcnum, miss_before);
993 * From SDM: Performing back-to-back fast reads are not guaranteed
995 * Use LFENCE to ensure all previous instructions are retired
999 rdpmcl(hit_pmcnum, hits_before);
1000 rdpmcl(miss_pmcnum, miss_before);
1002 * Use LFENCE to ensure all previous instructions are retired
1003 * before proceeding.
1006 for (i = 0; i < size; i += line_size) {
1008 * Add a barrier to prevent speculative execution of this
1009 * loop reading beyond the end of the buffer.
1012 asm volatile("mov (%0,%1,1), %%eax\n\t"
1014 : "r" (mem_r), "r" (i)
1015 : "%eax", "memory");
1018 * Use LFENCE to ensure all previous instructions are retired
1019 * before proceeding.
1022 rdpmcl(hit_pmcnum, hits_after);
1023 rdpmcl(miss_pmcnum, miss_after);
1025 * Use LFENCE to ensure all previous instructions are retired
1026 * before proceeding.
1029 /* Re-enable hardware prefetchers */
1030 wrmsr(MSR_MISC_FEATURE_CONTROL, 0x0, 0x0);
1033 perf_event_release_kernel(hit_event);
1035 perf_event_release_kernel(miss_event);
1038 * All counts will be zero on failure.
1040 counts->miss_before = miss_before;
1041 counts->hits_before = hits_before;
1042 counts->miss_after = miss_after;
1043 counts->hits_after = hits_after;
1047 static int measure_l2_residency(void *_plr)
1049 struct pseudo_lock_region *plr = _plr;
1050 struct residency_counts counts = {0};
1053 * Non-architectural event for the Goldmont Microarchitecture
1054 * from Intel x86 Architecture Software Developer Manual (SDM):
1055 * MEM_LOAD_UOPS_RETIRED D1H (event number)
1060 switch (boot_cpu_data.x86_model) {
1061 case INTEL_FAM6_ATOM_GOLDMONT:
1062 case INTEL_FAM6_ATOM_GOLDMONT_PLUS:
1063 perf_miss_attr.config = X86_CONFIG(.event = 0xd1,
1065 perf_hit_attr.config = X86_CONFIG(.event = 0xd1,
1072 measure_residency_fn(&perf_miss_attr, &perf_hit_attr, plr, &counts);
1074 * If a failure prevented the measurements from succeeding
1075 * tracepoints will still be written and all counts will be zero.
1077 trace_pseudo_lock_l2(counts.hits_after - counts.hits_before,
1078 counts.miss_after - counts.miss_before);
1080 plr->thread_done = 1;
1081 wake_up_interruptible(&plr->lock_thread_wq);
1085 static int measure_l3_residency(void *_plr)
1087 struct pseudo_lock_region *plr = _plr;
1088 struct residency_counts counts = {0};
1091 * On Broadwell Microarchitecture the MEM_LOAD_UOPS_RETIRED event
1092 * has two "no fix" errata associated with it: BDM35 and BDM100. On
1093 * this platform the following events are used instead:
1094 * LONGEST_LAT_CACHE 2EH (Documented in SDM)
1099 switch (boot_cpu_data.x86_model) {
1100 case INTEL_FAM6_BROADWELL_X:
1101 /* On BDW the hit event counts references, not hits */
1102 perf_hit_attr.config = X86_CONFIG(.event = 0x2e,
1104 perf_miss_attr.config = X86_CONFIG(.event = 0x2e,
1111 measure_residency_fn(&perf_miss_attr, &perf_hit_attr, plr, &counts);
1113 * If a failure prevented the measurements from succeeding
1114 * tracepoints will still be written and all counts will be zero.
1117 counts.miss_after -= counts.miss_before;
1118 if (boot_cpu_data.x86_model == INTEL_FAM6_BROADWELL_X) {
1120 * On BDW references and misses are counted, need to adjust.
1121 * Sometimes the "hits" counter is a bit more than the
1122 * references, for example, x references but x + 1 hits.
1123 * To not report invalid hit values in this case we treat
1124 * that as misses equal to references.
1126 /* First compute the number of cache references measured */
1127 counts.hits_after -= counts.hits_before;
1128 /* Next convert references to cache hits */
1129 counts.hits_after -= min(counts.miss_after, counts.hits_after);
1131 counts.hits_after -= counts.hits_before;
1134 trace_pseudo_lock_l3(counts.hits_after, counts.miss_after);
1136 plr->thread_done = 1;
1137 wake_up_interruptible(&plr->lock_thread_wq);
1142 * pseudo_lock_measure_cycles - Trigger latency measure to pseudo-locked region
1144 * The measurement of latency to access a pseudo-locked region should be
1145 * done from a cpu that is associated with that pseudo-locked region.
1146 * Determine which cpu is associated with this region and start a thread on
1147 * that cpu to perform the measurement, wait for that thread to complete.
1149 * Return: 0 on success, <0 on failure
1151 static int pseudo_lock_measure_cycles(struct rdtgroup *rdtgrp, int sel)
1153 struct pseudo_lock_region *plr = rdtgrp->plr;
1154 struct task_struct *thread;
1159 mutex_lock(&rdtgroup_mutex);
1161 if (rdtgrp->flags & RDT_DELETED) {
1171 plr->thread_done = 0;
1172 cpu = cpumask_first(&plr->d->cpu_mask);
1173 if (!cpu_online(cpu)) {
1181 thread = kthread_create_on_node(measure_cycles_lat_fn, plr,
1183 "pseudo_lock_measure/%u",
1186 thread = kthread_create_on_node(measure_l2_residency, plr,
1188 "pseudo_lock_measure/%u",
1191 thread = kthread_create_on_node(measure_l3_residency, plr,
1193 "pseudo_lock_measure/%u",
1198 if (IS_ERR(thread)) {
1199 ret = PTR_ERR(thread);
1202 kthread_bind(thread, cpu);
1203 wake_up_process(thread);
1205 ret = wait_event_interruptible(plr->lock_thread_wq,
1206 plr->thread_done == 1);
1213 mutex_unlock(&rdtgroup_mutex);
1218 static ssize_t pseudo_lock_measure_trigger(struct file *file,
1219 const char __user *user_buf,
1220 size_t count, loff_t *ppos)
1222 struct rdtgroup *rdtgrp = file->private_data;
1228 buf_size = min(count, (sizeof(buf) - 1));
1229 if (copy_from_user(buf, user_buf, buf_size))
1232 buf[buf_size] = '\0';
1233 ret = kstrtoint(buf, 10, &sel);
1235 if (sel != 1 && sel != 2 && sel != 3)
1237 ret = debugfs_file_get(file->f_path.dentry);
1240 ret = pseudo_lock_measure_cycles(rdtgrp, sel);
1243 debugfs_file_put(file->f_path.dentry);
1249 static const struct file_operations pseudo_measure_fops = {
1250 .write = pseudo_lock_measure_trigger,
1251 .open = simple_open,
1252 .llseek = default_llseek,
1256 * rdtgroup_pseudo_lock_create - Create a pseudo-locked region
1257 * @rdtgrp: resource group to which pseudo-lock region belongs
1259 * Called when a resource group in the pseudo-locksetup mode receives a
1260 * valid schemata that should be pseudo-locked. Since the resource group is
1261 * in pseudo-locksetup mode the &struct pseudo_lock_region has already been
1262 * allocated and initialized with the essential information. If a failure
1263 * occurs the resource group remains in the pseudo-locksetup mode with the
1264 * &struct pseudo_lock_region associated with it, but cleared from all
1265 * information and ready for the user to re-attempt pseudo-locking by
1266 * writing the schemata again.
1268 * Return: 0 if the pseudo-locked region was successfully pseudo-locked, <0
1269 * on failure. Descriptive error will be written to last_cmd_status buffer.
1271 int rdtgroup_pseudo_lock_create(struct rdtgroup *rdtgrp)
1273 struct pseudo_lock_region *plr = rdtgrp->plr;
1274 struct task_struct *thread;
1275 unsigned int new_minor;
1279 ret = pseudo_lock_region_alloc(plr);
1283 ret = pseudo_lock_cstates_constrain(plr);
1289 plr->thread_done = 0;
1291 thread = kthread_create_on_node(pseudo_lock_fn, rdtgrp,
1292 cpu_to_node(plr->cpu),
1293 "pseudo_lock/%u", plr->cpu);
1294 if (IS_ERR(thread)) {
1295 ret = PTR_ERR(thread);
1296 rdt_last_cmd_printf("Locking thread returned error %d\n", ret);
1300 kthread_bind(thread, plr->cpu);
1301 wake_up_process(thread);
1303 ret = wait_event_interruptible(plr->lock_thread_wq,
1304 plr->thread_done == 1);
1307 * If the thread does not get on the CPU for whatever
1308 * reason and the process which sets up the region is
1309 * interrupted then this will leave the thread in runnable
1310 * state and once it gets on the CPU it will derefence
1311 * the cleared, but not freed, plr struct resulting in an
1312 * empty pseudo-locking loop.
1314 rdt_last_cmd_puts("Locking thread interrupted\n");
1318 ret = pseudo_lock_minor_get(&new_minor);
1320 rdt_last_cmd_puts("Unable to obtain a new minor number\n");
1325 * Unlock access but do not release the reference. The
1326 * pseudo-locked region will still be here on return.
1328 * The mutex has to be released temporarily to avoid a potential
1329 * deadlock with the mm->mmap_sem semaphore which is obtained in
1330 * the device_create() and debugfs_create_dir() callpath below
1331 * as well as before the mmap() callback is called.
1333 mutex_unlock(&rdtgroup_mutex);
1335 if (!IS_ERR_OR_NULL(debugfs_resctrl)) {
1336 plr->debugfs_dir = debugfs_create_dir(rdtgrp->kn->name,
1338 if (!IS_ERR_OR_NULL(plr->debugfs_dir))
1339 debugfs_create_file("pseudo_lock_measure", 0200,
1340 plr->debugfs_dir, rdtgrp,
1341 &pseudo_measure_fops);
1344 dev = device_create(pseudo_lock_class, NULL,
1345 MKDEV(pseudo_lock_major, new_minor),
1346 rdtgrp, "%s", rdtgrp->kn->name);
1348 mutex_lock(&rdtgroup_mutex);
1352 rdt_last_cmd_printf("Failed to create character device: %d\n",
1357 /* We released the mutex - check if group was removed while we did so */
1358 if (rdtgrp->flags & RDT_DELETED) {
1363 plr->minor = new_minor;
1365 rdtgrp->mode = RDT_MODE_PSEUDO_LOCKED;
1366 closid_free(rdtgrp->closid);
1367 rdtgroup_kn_mode_restore(rdtgrp, "cpus", 0444);
1368 rdtgroup_kn_mode_restore(rdtgrp, "cpus_list", 0444);
1374 device_destroy(pseudo_lock_class, MKDEV(pseudo_lock_major, new_minor));
1376 debugfs_remove_recursive(plr->debugfs_dir);
1377 pseudo_lock_minor_release(new_minor);
1379 pseudo_lock_cstates_relax(plr);
1381 pseudo_lock_region_clear(plr);
1387 * rdtgroup_pseudo_lock_remove - Remove a pseudo-locked region
1388 * @rdtgrp: resource group to which the pseudo-locked region belongs
1390 * The removal of a pseudo-locked region can be initiated when the resource
1391 * group is removed from user space via a "rmdir" from userspace or the
1392 * unmount of the resctrl filesystem. On removal the resource group does
1393 * not go back to pseudo-locksetup mode before it is removed, instead it is
1394 * removed directly. There is thus assymmetry with the creation where the
1395 * &struct pseudo_lock_region is removed here while it was not created in
1396 * rdtgroup_pseudo_lock_create().
1400 void rdtgroup_pseudo_lock_remove(struct rdtgroup *rdtgrp)
1402 struct pseudo_lock_region *plr = rdtgrp->plr;
1404 if (rdtgrp->mode == RDT_MODE_PSEUDO_LOCKSETUP) {
1406 * Default group cannot be a pseudo-locked region so we can
1409 closid_free(rdtgrp->closid);
1413 pseudo_lock_cstates_relax(plr);
1414 debugfs_remove_recursive(rdtgrp->plr->debugfs_dir);
1415 device_destroy(pseudo_lock_class, MKDEV(pseudo_lock_major, plr->minor));
1416 pseudo_lock_minor_release(plr->minor);
1419 pseudo_lock_free(rdtgrp);
1422 static int pseudo_lock_dev_open(struct inode *inode, struct file *filp)
1424 struct rdtgroup *rdtgrp;
1426 mutex_lock(&rdtgroup_mutex);
1428 rdtgrp = region_find_by_minor(iminor(inode));
1430 mutex_unlock(&rdtgroup_mutex);
1434 filp->private_data = rdtgrp;
1435 atomic_inc(&rdtgrp->waitcount);
1436 /* Perform a non-seekable open - llseek is not supported */
1437 filp->f_mode &= ~(FMODE_LSEEK | FMODE_PREAD | FMODE_PWRITE);
1439 mutex_unlock(&rdtgroup_mutex);
1444 static int pseudo_lock_dev_release(struct inode *inode, struct file *filp)
1446 struct rdtgroup *rdtgrp;
1448 mutex_lock(&rdtgroup_mutex);
1449 rdtgrp = filp->private_data;
1452 mutex_unlock(&rdtgroup_mutex);
1455 filp->private_data = NULL;
1456 atomic_dec(&rdtgrp->waitcount);
1457 mutex_unlock(&rdtgroup_mutex);
1461 static int pseudo_lock_dev_mremap(struct vm_area_struct *area)
1467 static const struct vm_operations_struct pseudo_mmap_ops = {
1468 .mremap = pseudo_lock_dev_mremap,
1471 static int pseudo_lock_dev_mmap(struct file *filp, struct vm_area_struct *vma)
1473 unsigned long vsize = vma->vm_end - vma->vm_start;
1474 unsigned long off = vma->vm_pgoff << PAGE_SHIFT;
1475 struct pseudo_lock_region *plr;
1476 struct rdtgroup *rdtgrp;
1477 unsigned long physical;
1478 unsigned long psize;
1480 mutex_lock(&rdtgroup_mutex);
1482 rdtgrp = filp->private_data;
1485 mutex_unlock(&rdtgroup_mutex);
1492 mutex_unlock(&rdtgroup_mutex);
1497 * Task is required to run with affinity to the cpus associated
1498 * with the pseudo-locked region. If this is not the case the task
1499 * may be scheduled elsewhere and invalidate entries in the
1500 * pseudo-locked region.
1502 if (!cpumask_subset(current->cpus_ptr, &plr->d->cpu_mask)) {
1503 mutex_unlock(&rdtgroup_mutex);
1507 physical = __pa(plr->kmem) >> PAGE_SHIFT;
1508 psize = plr->size - off;
1510 if (off > plr->size) {
1511 mutex_unlock(&rdtgroup_mutex);
1516 * Ensure changes are carried directly to the memory being mapped,
1517 * do not allow copy-on-write mapping.
1519 if (!(vma->vm_flags & VM_SHARED)) {
1520 mutex_unlock(&rdtgroup_mutex);
1524 if (vsize > psize) {
1525 mutex_unlock(&rdtgroup_mutex);
1529 memset(plr->kmem + off, 0, vsize);
1531 if (remap_pfn_range(vma, vma->vm_start, physical + vma->vm_pgoff,
1532 vsize, vma->vm_page_prot)) {
1533 mutex_unlock(&rdtgroup_mutex);
1536 vma->vm_ops = &pseudo_mmap_ops;
1537 mutex_unlock(&rdtgroup_mutex);
1541 static const struct file_operations pseudo_lock_dev_fops = {
1542 .owner = THIS_MODULE,
1543 .llseek = no_llseek,
1546 .open = pseudo_lock_dev_open,
1547 .release = pseudo_lock_dev_release,
1548 .mmap = pseudo_lock_dev_mmap,
1551 static char *pseudo_lock_devnode(struct device *dev, umode_t *mode)
1553 struct rdtgroup *rdtgrp;
1555 rdtgrp = dev_get_drvdata(dev);
1558 return kasprintf(GFP_KERNEL, "pseudo_lock/%s", rdtgrp->kn->name);
1561 int rdt_pseudo_lock_init(void)
1565 ret = register_chrdev(0, "pseudo_lock", &pseudo_lock_dev_fops);
1569 pseudo_lock_major = ret;
1571 pseudo_lock_class = class_create(THIS_MODULE, "pseudo_lock");
1572 if (IS_ERR(pseudo_lock_class)) {
1573 ret = PTR_ERR(pseudo_lock_class);
1574 unregister_chrdev(pseudo_lock_major, "pseudo_lock");
1578 pseudo_lock_class->devnode = pseudo_lock_devnode;
1582 void rdt_pseudo_lock_release(void)
1584 class_destroy(pseudo_lock_class);
1585 pseudo_lock_class = NULL;
1586 unregister_chrdev(pseudo_lock_major, "pseudo_lock");
1587 pseudo_lock_major = 0;