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.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
52 * @void: It takes no parameters.
54 * Capture the list of platforms that have been validated to support
55 * pseudo-locking. This includes testing to ensure pseudo-locked regions
56 * with low cache miss rates can be created under variety of load conditions
57 * as well as that these pseudo-locked regions can maintain their low cache
58 * miss rates under variety of load conditions for significant lengths of time.
60 * After a platform has been validated to support pseudo-locking its
61 * hardware prefetch disable bits are included here as they are documented
64 * When adding a platform here also add support for its cache events to
65 * measure_cycles_perf_fn()
68 * If platform is supported, the bits to disable hardware prefetchers, 0
69 * if platform is not supported.
71 static u64 get_prefetch_disable_bits(void)
73 if (boot_cpu_data.x86_vendor != X86_VENDOR_INTEL ||
74 boot_cpu_data.x86 != 6)
77 switch (boot_cpu_data.x86_model) {
78 case INTEL_FAM6_BROADWELL_X:
80 * SDM defines bits of MSR_MISC_FEATURE_CONTROL register
82 * 0 L2 Hardware Prefetcher Disable (R/W)
83 * 1 L2 Adjacent Cache Line Prefetcher Disable (R/W)
84 * 2 DCU Hardware Prefetcher Disable (R/W)
85 * 3 DCU IP Prefetcher Disable (R/W)
89 case INTEL_FAM6_ATOM_GOLDMONT:
90 case INTEL_FAM6_ATOM_GOLDMONT_PLUS:
92 * SDM defines bits of MSR_MISC_FEATURE_CONTROL register
94 * 0 L2 Hardware Prefetcher Disable (R/W)
96 * 2 DCU Hardware Prefetcher Disable (R/W)
106 * pseudo_lock_minor_get - Obtain available minor number
107 * @minor: Pointer to where new minor number will be stored
109 * A bitmask is used to track available minor numbers. Here the next free
110 * minor number is marked as unavailable and returned.
112 * Return: 0 on success, <0 on failure.
114 static int pseudo_lock_minor_get(unsigned int *minor)
116 unsigned long first_bit;
118 first_bit = find_first_bit(&pseudo_lock_minor_avail, MINORBITS);
120 if (first_bit == MINORBITS)
123 __clear_bit(first_bit, &pseudo_lock_minor_avail);
130 * pseudo_lock_minor_release - Return minor number to available
131 * @minor: The minor number made available
133 static void pseudo_lock_minor_release(unsigned int minor)
135 __set_bit(minor, &pseudo_lock_minor_avail);
139 * region_find_by_minor - Locate a pseudo-lock region by inode minor number
140 * @minor: The minor number of the device representing pseudo-locked region
142 * When the character device is accessed we need to determine which
143 * pseudo-locked region it belongs to. This is done by matching the minor
144 * number of the device to the pseudo-locked region it belongs.
146 * Minor numbers are assigned at the time a pseudo-locked region is associated
147 * with a cache instance.
149 * Return: On success return pointer to resource group owning the pseudo-locked
150 * region, NULL on failure.
152 static struct rdtgroup *region_find_by_minor(unsigned int minor)
154 struct rdtgroup *rdtgrp, *rdtgrp_match = NULL;
156 list_for_each_entry(rdtgrp, &rdt_all_groups, rdtgroup_list) {
157 if (rdtgrp->plr && rdtgrp->plr->minor == minor) {
158 rdtgrp_match = rdtgrp;
166 * struct pseudo_lock_pm_req - A power management QoS request list entry
167 * @list: Entry within the @pm_reqs list for a pseudo-locked region
168 * @req: PM QoS request
170 struct pseudo_lock_pm_req {
171 struct list_head list;
172 struct dev_pm_qos_request req;
175 static void pseudo_lock_cstates_relax(struct pseudo_lock_region *plr)
177 struct pseudo_lock_pm_req *pm_req, *next;
179 list_for_each_entry_safe(pm_req, next, &plr->pm_reqs, list) {
180 dev_pm_qos_remove_request(&pm_req->req);
181 list_del(&pm_req->list);
187 * pseudo_lock_cstates_constrain - Restrict cores from entering C6
188 * @plr: Pseudo-locked region
190 * To prevent the cache from being affected by power management entering
191 * C6 has to be avoided. This is accomplished by requesting a latency
192 * requirement lower than lowest C6 exit latency of all supported
193 * platforms as found in the cpuidle state tables in the intel_idle driver.
194 * At this time it is possible to do so with a single latency requirement
195 * for all supported platforms.
197 * Since Goldmont is supported, which is affected by X86_BUG_MONITOR,
198 * the ACPI latencies need to be considered while keeping in mind that C2
199 * may be set to map to deeper sleep states. In this case the latency
200 * requirement needs to prevent entering C2 also.
202 * Return: 0 on success, <0 on failure
204 static int pseudo_lock_cstates_constrain(struct pseudo_lock_region *plr)
206 struct pseudo_lock_pm_req *pm_req;
210 for_each_cpu(cpu, &plr->d->cpu_mask) {
211 pm_req = kzalloc(sizeof(*pm_req), GFP_KERNEL);
213 rdt_last_cmd_puts("Failure to allocate memory for PM QoS\n");
217 ret = dev_pm_qos_add_request(get_cpu_device(cpu),
219 DEV_PM_QOS_RESUME_LATENCY,
222 rdt_last_cmd_printf("Failed to add latency req CPU%d\n",
228 list_add(&pm_req->list, &plr->pm_reqs);
234 pseudo_lock_cstates_relax(plr);
239 * pseudo_lock_region_clear - Reset pseudo-lock region data
240 * @plr: pseudo-lock region
242 * All content of the pseudo-locked region is reset - any memory allocated
247 static void pseudo_lock_region_clear(struct pseudo_lock_region *plr)
258 plr->debugfs_dir = NULL;
262 * pseudo_lock_region_init - Initialize pseudo-lock region information
263 * @plr: pseudo-lock region
265 * Called after user provided a schemata to be pseudo-locked. From the
266 * schemata the &struct pseudo_lock_region is on entry already initialized
267 * with the resource, domain, and capacity bitmask. Here the information
268 * required for pseudo-locking is deduced from this data and &struct
269 * pseudo_lock_region initialized further. This information includes:
270 * - size in bytes of the region to be pseudo-locked
271 * - cache line size to know the stride with which data needs to be accessed
272 * to be pseudo-locked
273 * - a cpu associated with the cache instance on which the pseudo-locking
274 * flow can be executed
276 * Return: 0 on success, <0 on failure. Descriptive error will be written
277 * to last_cmd_status buffer.
279 static int pseudo_lock_region_init(struct pseudo_lock_region *plr)
281 struct cpu_cacheinfo *ci;
285 /* Pick the first cpu we find that is associated with the cache. */
286 plr->cpu = cpumask_first(&plr->d->cpu_mask);
288 if (!cpu_online(plr->cpu)) {
289 rdt_last_cmd_printf("CPU %u associated with cache not online\n",
295 ci = get_cpu_cacheinfo(plr->cpu);
297 plr->size = rdtgroup_cbm_to_size(plr->s->res, plr->d, plr->cbm);
299 for (i = 0; i < ci->num_leaves; i++) {
300 if (ci->info_list[i].level == plr->s->res->cache_level) {
301 plr->line_size = ci->info_list[i].coherency_line_size;
307 rdt_last_cmd_puts("Unable to determine cache line size\n");
309 pseudo_lock_region_clear(plr);
314 * pseudo_lock_init - Initialize a pseudo-lock region
315 * @rdtgrp: resource group to which new pseudo-locked region will belong
317 * A pseudo-locked region is associated with a resource group. When this
318 * association is created the pseudo-locked region is initialized. The
319 * details of the pseudo-locked region are not known at this time so only
320 * allocation is done and association established.
322 * Return: 0 on success, <0 on failure
324 static int pseudo_lock_init(struct rdtgroup *rdtgrp)
326 struct pseudo_lock_region *plr;
328 plr = kzalloc(sizeof(*plr), GFP_KERNEL);
332 init_waitqueue_head(&plr->lock_thread_wq);
333 INIT_LIST_HEAD(&plr->pm_reqs);
339 * pseudo_lock_region_alloc - Allocate kernel memory that will be pseudo-locked
340 * @plr: pseudo-lock region
342 * Initialize the details required to set up the pseudo-locked region and
343 * allocate the contiguous memory that will be pseudo-locked to the cache.
345 * Return: 0 on success, <0 on failure. Descriptive error will be written
346 * to last_cmd_status buffer.
348 static int pseudo_lock_region_alloc(struct pseudo_lock_region *plr)
352 ret = pseudo_lock_region_init(plr);
357 * We do not yet support contiguous regions larger than
360 if (plr->size > KMALLOC_MAX_SIZE) {
361 rdt_last_cmd_puts("Requested region exceeds maximum size\n");
366 plr->kmem = kzalloc(plr->size, GFP_KERNEL);
368 rdt_last_cmd_puts("Unable to allocate memory\n");
376 pseudo_lock_region_clear(plr);
382 * pseudo_lock_free - Free a pseudo-locked region
383 * @rdtgrp: resource group to which pseudo-locked region belonged
385 * The pseudo-locked region's resources have already been released, or not
386 * yet created at this point. Now it can be freed and disassociated from the
391 static void pseudo_lock_free(struct rdtgroup *rdtgrp)
393 pseudo_lock_region_clear(rdtgrp->plr);
399 * pseudo_lock_fn - Load kernel memory into cache
400 * @_rdtgrp: resource group to which pseudo-lock region belongs
402 * This is the core pseudo-locking flow.
404 * First we ensure that the kernel memory cannot be found in the cache.
405 * Then, while taking care that there will be as little interference as
406 * possible, the memory to be loaded is accessed while core is running
407 * with class of service set to the bitmask of the pseudo-locked region.
408 * After this is complete no future CAT allocations will be allowed to
409 * overlap with this bitmask.
411 * Local register variables are utilized to ensure that the memory region
412 * to be locked is the only memory access made during the critical locking
415 * Return: 0. Waiter on waitqueue will be woken on completion.
417 static int pseudo_lock_fn(void *_rdtgrp)
419 struct rdtgroup *rdtgrp = _rdtgrp;
420 struct pseudo_lock_region *plr = rdtgrp->plr;
421 u32 rmid_p, closid_p;
425 * The registers used for local register variables are also used
426 * when KASAN is active. When KASAN is active we use a regular
427 * variable to ensure we always use a valid pointer, but the cost
428 * is that this variable will enter the cache through evicting the
429 * memory we are trying to lock into the cache. Thus expect lower
430 * pseudo-locking success rate when KASAN is active.
432 unsigned int line_size;
436 register unsigned int line_size asm("esi");
437 register unsigned int size asm("edi");
438 register void *mem_r asm(_ASM_BX);
439 #endif /* CONFIG_KASAN */
442 * Make sure none of the allocated memory is cached. If it is we
443 * will get a cache hit in below loop from outside of pseudo-locked
445 * wbinvd (as opposed to clflush/clflushopt) is required to
446 * increase likelihood that allocated cache portion will be filled
447 * with associated memory.
452 * Always called with interrupts enabled. By disabling interrupts
453 * ensure that we will not be preempted during this critical section.
458 * Call wrmsr and rdmsr as directly as possible to avoid tracing
459 * clobbering local register variables or affecting cache accesses.
461 * Disable the hardware prefetcher so that when the end of the memory
462 * being pseudo-locked is reached the hardware will not read beyond
463 * the buffer and evict pseudo-locked memory read earlier from the
466 __wrmsr(MSR_MISC_FEATURE_CONTROL, prefetch_disable_bits, 0x0);
467 closid_p = this_cpu_read(pqr_state.cur_closid);
468 rmid_p = this_cpu_read(pqr_state.cur_rmid);
471 line_size = plr->line_size;
473 * Critical section begin: start by writing the closid associated
474 * with the capacity bitmask of the cache region being
475 * pseudo-locked followed by reading of kernel memory to load it
478 __wrmsr(IA32_PQR_ASSOC, rmid_p, rdtgrp->closid);
480 * Cache was flushed earlier. Now access kernel memory to read it
481 * into cache region associated with just activated plr->closid.
482 * Loop over data twice:
483 * - In first loop the cache region is shared with the page walker
484 * as it populates the paging structure caches (including TLB).
485 * - In the second loop the paging structure caches are used and
486 * cache region is populated with the memory being referenced.
488 for (i = 0; i < size; i += PAGE_SIZE) {
490 * Add a barrier to prevent speculative execution of this
491 * loop reading beyond the end of the buffer.
494 asm volatile("mov (%0,%1,1), %%eax\n\t"
496 : "r" (mem_r), "r" (i)
499 for (i = 0; i < size; i += line_size) {
501 * Add a barrier to prevent speculative execution of this
502 * loop reading beyond the end of the buffer.
505 asm volatile("mov (%0,%1,1), %%eax\n\t"
507 : "r" (mem_r), "r" (i)
511 * Critical section end: restore closid with capacity bitmask that
512 * does not overlap with pseudo-locked region.
514 __wrmsr(IA32_PQR_ASSOC, rmid_p, closid_p);
516 /* Re-enable the hardware prefetcher(s) */
517 wrmsr(MSR_MISC_FEATURE_CONTROL, 0x0, 0x0);
520 plr->thread_done = 1;
521 wake_up_interruptible(&plr->lock_thread_wq);
526 * rdtgroup_monitor_in_progress - Test if monitoring in progress
527 * @rdtgrp: resource group being queried
529 * Return: 1 if monitor groups have been created for this resource
530 * group, 0 otherwise.
532 static int rdtgroup_monitor_in_progress(struct rdtgroup *rdtgrp)
534 return !list_empty(&rdtgrp->mon.crdtgrp_list);
538 * rdtgroup_locksetup_user_restrict - Restrict user access to group
539 * @rdtgrp: resource group needing access restricted
541 * A resource group used for cache pseudo-locking cannot have cpus or tasks
542 * assigned to it. This is communicated to the user by restricting access
543 * to all the files that can be used to make such changes.
545 * Permissions restored with rdtgroup_locksetup_user_restore()
547 * Return: 0 on success, <0 on failure. If a failure occurs during the
548 * restriction of access an attempt will be made to restore permissions but
549 * the state of the mode of these files will be uncertain when a failure
552 static int rdtgroup_locksetup_user_restrict(struct rdtgroup *rdtgrp)
556 ret = rdtgroup_kn_mode_restrict(rdtgrp, "tasks");
560 ret = rdtgroup_kn_mode_restrict(rdtgrp, "cpus");
564 ret = rdtgroup_kn_mode_restrict(rdtgrp, "cpus_list");
568 if (rdt_mon_capable) {
569 ret = rdtgroup_kn_mode_restrict(rdtgrp, "mon_groups");
578 rdtgroup_kn_mode_restore(rdtgrp, "cpus_list", 0777);
580 rdtgroup_kn_mode_restore(rdtgrp, "cpus", 0777);
582 rdtgroup_kn_mode_restore(rdtgrp, "tasks", 0777);
588 * rdtgroup_locksetup_user_restore - Restore user access to group
589 * @rdtgrp: resource group needing access restored
591 * Restore all file access previously removed using
592 * rdtgroup_locksetup_user_restrict()
594 * Return: 0 on success, <0 on failure. If a failure occurs during the
595 * restoration of access an attempt will be made to restrict permissions
596 * again but the state of the mode of these files will be uncertain when
599 static int rdtgroup_locksetup_user_restore(struct rdtgroup *rdtgrp)
603 ret = rdtgroup_kn_mode_restore(rdtgrp, "tasks", 0777);
607 ret = rdtgroup_kn_mode_restore(rdtgrp, "cpus", 0777);
611 ret = rdtgroup_kn_mode_restore(rdtgrp, "cpus_list", 0777);
615 if (rdt_mon_capable) {
616 ret = rdtgroup_kn_mode_restore(rdtgrp, "mon_groups", 0777);
625 rdtgroup_kn_mode_restrict(rdtgrp, "cpus_list");
627 rdtgroup_kn_mode_restrict(rdtgrp, "cpus");
629 rdtgroup_kn_mode_restrict(rdtgrp, "tasks");
635 * rdtgroup_locksetup_enter - Resource group enters locksetup mode
636 * @rdtgrp: resource group requested to enter locksetup mode
638 * A resource group enters locksetup mode to reflect that it would be used
639 * to represent a pseudo-locked region and is in the process of being set
640 * up to do so. A resource group used for a pseudo-locked region would
641 * lose the closid associated with it so we cannot allow it to have any
642 * tasks or cpus assigned nor permit tasks or cpus to be assigned in the
643 * future. Monitoring of a pseudo-locked region is not allowed either.
645 * The above and more restrictions on a pseudo-locked region are checked
646 * for and enforced before the resource group enters the locksetup mode.
648 * Returns: 0 if the resource group successfully entered locksetup mode, <0
649 * on failure. On failure the last_cmd_status buffer is updated with text to
650 * communicate details of failure to the user.
652 int rdtgroup_locksetup_enter(struct rdtgroup *rdtgrp)
657 * The default resource group can neither be removed nor lose the
658 * default closid associated with it.
660 if (rdtgrp == &rdtgroup_default) {
661 rdt_last_cmd_puts("Cannot pseudo-lock default group\n");
666 * Cache Pseudo-locking not supported when CDP is enabled.
668 * Some things to consider if you would like to enable this
669 * support (using L3 CDP as example):
670 * - When CDP is enabled two separate resources are exposed,
671 * L3DATA and L3CODE, but they are actually on the same cache.
672 * The implication for pseudo-locking is that if a
673 * pseudo-locked region is created on a domain of one
674 * resource (eg. L3CODE), then a pseudo-locked region cannot
675 * be created on that same domain of the other resource
676 * (eg. L3DATA). This is because the creation of a
677 * pseudo-locked region involves a call to wbinvd that will
678 * affect all cache allocations on particular domain.
679 * - Considering the previous, it may be possible to only
680 * expose one of the CDP resources to pseudo-locking and
681 * hide the other. For example, we could consider to only
682 * expose L3DATA and since the L3 cache is unified it is
683 * still possible to place instructions there are execute it.
684 * - If only one region is exposed to pseudo-locking we should
685 * still keep in mind that availability of a portion of cache
686 * for pseudo-locking should take into account both resources.
687 * Similarly, if a pseudo-locked region is created in one
688 * resource, the portion of cache used by it should be made
689 * unavailable to all future allocations from both resources.
691 if (resctrl_arch_get_cdp_enabled(RDT_RESOURCE_L3) ||
692 resctrl_arch_get_cdp_enabled(RDT_RESOURCE_L2)) {
693 rdt_last_cmd_puts("CDP enabled\n");
698 * Not knowing the bits to disable prefetching implies that this
699 * platform does not support Cache Pseudo-Locking.
701 prefetch_disable_bits = get_prefetch_disable_bits();
702 if (prefetch_disable_bits == 0) {
703 rdt_last_cmd_puts("Pseudo-locking not supported\n");
707 if (rdtgroup_monitor_in_progress(rdtgrp)) {
708 rdt_last_cmd_puts("Monitoring in progress\n");
712 if (rdtgroup_tasks_assigned(rdtgrp)) {
713 rdt_last_cmd_puts("Tasks assigned to resource group\n");
717 if (!cpumask_empty(&rdtgrp->cpu_mask)) {
718 rdt_last_cmd_puts("CPUs assigned to resource group\n");
722 if (rdtgroup_locksetup_user_restrict(rdtgrp)) {
723 rdt_last_cmd_puts("Unable to modify resctrl permissions\n");
727 ret = pseudo_lock_init(rdtgrp);
729 rdt_last_cmd_puts("Unable to init pseudo-lock region\n");
734 * If this system is capable of monitoring a rmid would have been
735 * allocated when the control group was created. This is not needed
736 * anymore when this group would be used for pseudo-locking. This
737 * is safe to call on platforms not capable of monitoring.
739 free_rmid(rdtgrp->mon.rmid);
745 rdtgroup_locksetup_user_restore(rdtgrp);
751 * rdtgroup_locksetup_exit - resource group exist locksetup mode
752 * @rdtgrp: resource group
754 * When a resource group exits locksetup mode the earlier restrictions are
757 * Return: 0 on success, <0 on failure
759 int rdtgroup_locksetup_exit(struct rdtgroup *rdtgrp)
763 if (rdt_mon_capable) {
766 rdt_last_cmd_puts("Out of RMIDs\n");
769 rdtgrp->mon.rmid = ret;
772 ret = rdtgroup_locksetup_user_restore(rdtgrp);
774 free_rmid(rdtgrp->mon.rmid);
778 pseudo_lock_free(rdtgrp);
783 * rdtgroup_cbm_overlaps_pseudo_locked - Test if CBM or portion is pseudo-locked
787 * @d represents a cache instance and @cbm a capacity bitmask that is
788 * considered for it. Determine if @cbm overlaps with any existing
789 * pseudo-locked region on @d.
791 * @cbm is unsigned long, even if only 32 bits are used, to make the
792 * bitmap functions work correctly.
794 * Return: true if @cbm overlaps with pseudo-locked region on @d, false
797 bool rdtgroup_cbm_overlaps_pseudo_locked(struct rdt_domain *d, unsigned long cbm)
799 unsigned int cbm_len;
803 cbm_len = d->plr->s->res->cache.cbm_len;
805 if (bitmap_intersects(&cbm, &cbm_b, cbm_len))
812 * rdtgroup_pseudo_locked_in_hierarchy - Pseudo-locked region in cache hierarchy
813 * @d: RDT domain under test
815 * The setup of a pseudo-locked region affects all cache instances within
816 * the hierarchy of the region. It is thus essential to know if any
817 * pseudo-locked regions exist within a cache hierarchy to prevent any
818 * attempts to create new pseudo-locked regions in the same hierarchy.
820 * Return: true if a pseudo-locked region exists in the hierarchy of @d or
821 * if it is not possible to test due to memory allocation issue,
824 bool rdtgroup_pseudo_locked_in_hierarchy(struct rdt_domain *d)
826 cpumask_var_t cpu_with_psl;
827 struct rdt_resource *r;
828 struct rdt_domain *d_i;
831 if (!zalloc_cpumask_var(&cpu_with_psl, GFP_KERNEL))
835 * First determine which cpus have pseudo-locked regions
836 * associated with them.
838 for_each_alloc_enabled_rdt_resource(r) {
839 list_for_each_entry(d_i, &r->domains, list) {
841 cpumask_or(cpu_with_psl, cpu_with_psl,
847 * Next test if new pseudo-locked region would intersect with
850 if (cpumask_intersects(&d->cpu_mask, cpu_with_psl))
853 free_cpumask_var(cpu_with_psl);
858 * measure_cycles_lat_fn - Measure cycle latency to read pseudo-locked memory
859 * @_plr: pseudo-lock region to measure
861 * There is no deterministic way to test if a memory region is cached. One
862 * way is to measure how long it takes to read the memory, the speed of
863 * access is a good way to learn how close to the cpu the data was. Even
864 * more, if the prefetcher is disabled and the memory is read at a stride
865 * of half the cache line, then a cache miss will be easy to spot since the
866 * read of the first half would be significantly slower than the read of
869 * Return: 0. Waiter on waitqueue will be woken on completion.
871 static int measure_cycles_lat_fn(void *_plr)
873 struct pseudo_lock_region *plr = _plr;
880 * Disable hardware prefetchers.
882 wrmsr(MSR_MISC_FEATURE_CONTROL, prefetch_disable_bits, 0x0);
883 mem_r = READ_ONCE(plr->kmem);
885 * Dummy execute of the time measurement to load the needed
886 * instructions into the L1 instruction cache.
888 start = rdtsc_ordered();
889 for (i = 0; i < plr->size; i += 32) {
890 start = rdtsc_ordered();
891 asm volatile("mov (%0,%1,1), %%eax\n\t"
893 : "r" (mem_r), "r" (i)
895 end = rdtsc_ordered();
896 trace_pseudo_lock_mem_latency((u32)(end - start));
898 wrmsr(MSR_MISC_FEATURE_CONTROL, 0x0, 0x0);
900 plr->thread_done = 1;
901 wake_up_interruptible(&plr->lock_thread_wq);
906 * Create a perf_event_attr for the hit and miss perf events that will
907 * be used during the performance measurement. A perf_event maintains
908 * a pointer to its perf_event_attr so a unique attribute structure is
909 * created for each perf_event.
911 * The actual configuration of the event is set right before use in order
912 * to use the X86_CONFIG macro.
914 static struct perf_event_attr perf_miss_attr = {
915 .type = PERF_TYPE_RAW,
916 .size = sizeof(struct perf_event_attr),
922 static struct perf_event_attr perf_hit_attr = {
923 .type = PERF_TYPE_RAW,
924 .size = sizeof(struct perf_event_attr),
930 struct residency_counts {
931 u64 miss_before, hits_before;
932 u64 miss_after, hits_after;
935 static int measure_residency_fn(struct perf_event_attr *miss_attr,
936 struct perf_event_attr *hit_attr,
937 struct pseudo_lock_region *plr,
938 struct residency_counts *counts)
940 u64 hits_before = 0, hits_after = 0, miss_before = 0, miss_after = 0;
941 struct perf_event *miss_event, *hit_event;
942 int hit_pmcnum, miss_pmcnum;
943 unsigned int line_size;
949 miss_event = perf_event_create_kernel_counter(miss_attr, plr->cpu,
951 if (IS_ERR(miss_event))
954 hit_event = perf_event_create_kernel_counter(hit_attr, plr->cpu,
956 if (IS_ERR(hit_event))
961 * Check any possible error state of events used by performing
964 if (perf_event_read_local(miss_event, &tmp, NULL, NULL)) {
968 if (perf_event_read_local(hit_event, &tmp, NULL, NULL)) {
974 * Disable hardware prefetchers.
976 wrmsr(MSR_MISC_FEATURE_CONTROL, prefetch_disable_bits, 0x0);
978 /* Initialize rest of local variables */
980 * Performance event has been validated right before this with
981 * interrupts disabled - it is thus safe to read the counter index.
983 miss_pmcnum = x86_perf_rdpmc_index(miss_event);
984 hit_pmcnum = x86_perf_rdpmc_index(hit_event);
985 line_size = READ_ONCE(plr->line_size);
986 mem_r = READ_ONCE(plr->kmem);
987 size = READ_ONCE(plr->size);
990 * Read counter variables twice - first to load the instructions
991 * used in L1 cache, second to capture accurate value that does not
992 * include cache misses incurred because of instruction loads.
994 rdpmcl(hit_pmcnum, hits_before);
995 rdpmcl(miss_pmcnum, miss_before);
997 * From SDM: Performing back-to-back fast reads are not guaranteed
999 * Use LFENCE to ensure all previous instructions are retired
1000 * before proceeding.
1003 rdpmcl(hit_pmcnum, hits_before);
1004 rdpmcl(miss_pmcnum, miss_before);
1006 * Use LFENCE to ensure all previous instructions are retired
1007 * before proceeding.
1010 for (i = 0; i < size; i += line_size) {
1012 * Add a barrier to prevent speculative execution of this
1013 * loop reading beyond the end of the buffer.
1016 asm volatile("mov (%0,%1,1), %%eax\n\t"
1018 : "r" (mem_r), "r" (i)
1019 : "%eax", "memory");
1022 * Use LFENCE to ensure all previous instructions are retired
1023 * before proceeding.
1026 rdpmcl(hit_pmcnum, hits_after);
1027 rdpmcl(miss_pmcnum, miss_after);
1029 * Use LFENCE to ensure all previous instructions are retired
1030 * before proceeding.
1033 /* Re-enable hardware prefetchers */
1034 wrmsr(MSR_MISC_FEATURE_CONTROL, 0x0, 0x0);
1037 perf_event_release_kernel(hit_event);
1039 perf_event_release_kernel(miss_event);
1042 * All counts will be zero on failure.
1044 counts->miss_before = miss_before;
1045 counts->hits_before = hits_before;
1046 counts->miss_after = miss_after;
1047 counts->hits_after = hits_after;
1051 static int measure_l2_residency(void *_plr)
1053 struct pseudo_lock_region *plr = _plr;
1054 struct residency_counts counts = {0};
1057 * Non-architectural event for the Goldmont Microarchitecture
1058 * from Intel x86 Architecture Software Developer Manual (SDM):
1059 * MEM_LOAD_UOPS_RETIRED D1H (event number)
1064 switch (boot_cpu_data.x86_model) {
1065 case INTEL_FAM6_ATOM_GOLDMONT:
1066 case INTEL_FAM6_ATOM_GOLDMONT_PLUS:
1067 perf_miss_attr.config = X86_CONFIG(.event = 0xd1,
1069 perf_hit_attr.config = X86_CONFIG(.event = 0xd1,
1076 measure_residency_fn(&perf_miss_attr, &perf_hit_attr, plr, &counts);
1078 * If a failure prevented the measurements from succeeding
1079 * tracepoints will still be written and all counts will be zero.
1081 trace_pseudo_lock_l2(counts.hits_after - counts.hits_before,
1082 counts.miss_after - counts.miss_before);
1084 plr->thread_done = 1;
1085 wake_up_interruptible(&plr->lock_thread_wq);
1089 static int measure_l3_residency(void *_plr)
1091 struct pseudo_lock_region *plr = _plr;
1092 struct residency_counts counts = {0};
1095 * On Broadwell Microarchitecture the MEM_LOAD_UOPS_RETIRED event
1096 * has two "no fix" errata associated with it: BDM35 and BDM100. On
1097 * this platform the following events are used instead:
1098 * LONGEST_LAT_CACHE 2EH (Documented in SDM)
1103 switch (boot_cpu_data.x86_model) {
1104 case INTEL_FAM6_BROADWELL_X:
1105 /* On BDW the hit event counts references, not hits */
1106 perf_hit_attr.config = X86_CONFIG(.event = 0x2e,
1108 perf_miss_attr.config = X86_CONFIG(.event = 0x2e,
1115 measure_residency_fn(&perf_miss_attr, &perf_hit_attr, plr, &counts);
1117 * If a failure prevented the measurements from succeeding
1118 * tracepoints will still be written and all counts will be zero.
1121 counts.miss_after -= counts.miss_before;
1122 if (boot_cpu_data.x86_model == INTEL_FAM6_BROADWELL_X) {
1124 * On BDW references and misses are counted, need to adjust.
1125 * Sometimes the "hits" counter is a bit more than the
1126 * references, for example, x references but x + 1 hits.
1127 * To not report invalid hit values in this case we treat
1128 * that as misses equal to references.
1130 /* First compute the number of cache references measured */
1131 counts.hits_after -= counts.hits_before;
1132 /* Next convert references to cache hits */
1133 counts.hits_after -= min(counts.miss_after, counts.hits_after);
1135 counts.hits_after -= counts.hits_before;
1138 trace_pseudo_lock_l3(counts.hits_after, counts.miss_after);
1140 plr->thread_done = 1;
1141 wake_up_interruptible(&plr->lock_thread_wq);
1146 * pseudo_lock_measure_cycles - Trigger latency measure to pseudo-locked region
1147 * @rdtgrp: Resource group to which the pseudo-locked region belongs.
1148 * @sel: Selector of which measurement to perform on a pseudo-locked region.
1150 * The measurement of latency to access a pseudo-locked region should be
1151 * done from a cpu that is associated with that pseudo-locked region.
1152 * Determine which cpu is associated with this region and start a thread on
1153 * that cpu to perform the measurement, wait for that thread to complete.
1155 * Return: 0 on success, <0 on failure
1157 static int pseudo_lock_measure_cycles(struct rdtgroup *rdtgrp, int sel)
1159 struct pseudo_lock_region *plr = rdtgrp->plr;
1160 struct task_struct *thread;
1165 mutex_lock(&rdtgroup_mutex);
1167 if (rdtgrp->flags & RDT_DELETED) {
1177 plr->thread_done = 0;
1178 cpu = cpumask_first(&plr->d->cpu_mask);
1179 if (!cpu_online(cpu)) {
1187 thread = kthread_create_on_node(measure_cycles_lat_fn, plr,
1189 "pseudo_lock_measure/%u",
1192 thread = kthread_create_on_node(measure_l2_residency, plr,
1194 "pseudo_lock_measure/%u",
1197 thread = kthread_create_on_node(measure_l3_residency, plr,
1199 "pseudo_lock_measure/%u",
1204 if (IS_ERR(thread)) {
1205 ret = PTR_ERR(thread);
1208 kthread_bind(thread, cpu);
1209 wake_up_process(thread);
1211 ret = wait_event_interruptible(plr->lock_thread_wq,
1212 plr->thread_done == 1);
1219 mutex_unlock(&rdtgroup_mutex);
1224 static ssize_t pseudo_lock_measure_trigger(struct file *file,
1225 const char __user *user_buf,
1226 size_t count, loff_t *ppos)
1228 struct rdtgroup *rdtgrp = file->private_data;
1234 buf_size = min(count, (sizeof(buf) - 1));
1235 if (copy_from_user(buf, user_buf, buf_size))
1238 buf[buf_size] = '\0';
1239 ret = kstrtoint(buf, 10, &sel);
1241 if (sel != 1 && sel != 2 && sel != 3)
1243 ret = debugfs_file_get(file->f_path.dentry);
1246 ret = pseudo_lock_measure_cycles(rdtgrp, sel);
1249 debugfs_file_put(file->f_path.dentry);
1255 static const struct file_operations pseudo_measure_fops = {
1256 .write = pseudo_lock_measure_trigger,
1257 .open = simple_open,
1258 .llseek = default_llseek,
1262 * rdtgroup_pseudo_lock_create - Create a pseudo-locked region
1263 * @rdtgrp: resource group to which pseudo-lock region belongs
1265 * Called when a resource group in the pseudo-locksetup mode receives a
1266 * valid schemata that should be pseudo-locked. Since the resource group is
1267 * in pseudo-locksetup mode the &struct pseudo_lock_region has already been
1268 * allocated and initialized with the essential information. If a failure
1269 * occurs the resource group remains in the pseudo-locksetup mode with the
1270 * &struct pseudo_lock_region associated with it, but cleared from all
1271 * information and ready for the user to re-attempt pseudo-locking by
1272 * writing the schemata again.
1274 * Return: 0 if the pseudo-locked region was successfully pseudo-locked, <0
1275 * on failure. Descriptive error will be written to last_cmd_status buffer.
1277 int rdtgroup_pseudo_lock_create(struct rdtgroup *rdtgrp)
1279 struct pseudo_lock_region *plr = rdtgrp->plr;
1280 struct task_struct *thread;
1281 unsigned int new_minor;
1285 ret = pseudo_lock_region_alloc(plr);
1289 ret = pseudo_lock_cstates_constrain(plr);
1295 plr->thread_done = 0;
1297 thread = kthread_create_on_node(pseudo_lock_fn, rdtgrp,
1298 cpu_to_node(plr->cpu),
1299 "pseudo_lock/%u", plr->cpu);
1300 if (IS_ERR(thread)) {
1301 ret = PTR_ERR(thread);
1302 rdt_last_cmd_printf("Locking thread returned error %d\n", ret);
1306 kthread_bind(thread, plr->cpu);
1307 wake_up_process(thread);
1309 ret = wait_event_interruptible(plr->lock_thread_wq,
1310 plr->thread_done == 1);
1313 * If the thread does not get on the CPU for whatever
1314 * reason and the process which sets up the region is
1315 * interrupted then this will leave the thread in runnable
1316 * state and once it gets on the CPU it will dereference
1317 * the cleared, but not freed, plr struct resulting in an
1318 * empty pseudo-locking loop.
1320 rdt_last_cmd_puts("Locking thread interrupted\n");
1324 ret = pseudo_lock_minor_get(&new_minor);
1326 rdt_last_cmd_puts("Unable to obtain a new minor number\n");
1331 * Unlock access but do not release the reference. The
1332 * pseudo-locked region will still be here on return.
1334 * The mutex has to be released temporarily to avoid a potential
1335 * deadlock with the mm->mmap_lock which is obtained in the
1336 * device_create() and debugfs_create_dir() callpath below as well as
1337 * before the mmap() callback is called.
1339 mutex_unlock(&rdtgroup_mutex);
1341 if (!IS_ERR_OR_NULL(debugfs_resctrl)) {
1342 plr->debugfs_dir = debugfs_create_dir(rdtgrp->kn->name,
1344 if (!IS_ERR_OR_NULL(plr->debugfs_dir))
1345 debugfs_create_file("pseudo_lock_measure", 0200,
1346 plr->debugfs_dir, rdtgrp,
1347 &pseudo_measure_fops);
1350 dev = device_create(pseudo_lock_class, NULL,
1351 MKDEV(pseudo_lock_major, new_minor),
1352 rdtgrp, "%s", rdtgrp->kn->name);
1354 mutex_lock(&rdtgroup_mutex);
1358 rdt_last_cmd_printf("Failed to create character device: %d\n",
1363 /* We released the mutex - check if group was removed while we did so */
1364 if (rdtgrp->flags & RDT_DELETED) {
1369 plr->minor = new_minor;
1371 rdtgrp->mode = RDT_MODE_PSEUDO_LOCKED;
1372 closid_free(rdtgrp->closid);
1373 rdtgroup_kn_mode_restore(rdtgrp, "cpus", 0444);
1374 rdtgroup_kn_mode_restore(rdtgrp, "cpus_list", 0444);
1380 device_destroy(pseudo_lock_class, MKDEV(pseudo_lock_major, new_minor));
1382 debugfs_remove_recursive(plr->debugfs_dir);
1383 pseudo_lock_minor_release(new_minor);
1385 pseudo_lock_cstates_relax(plr);
1387 pseudo_lock_region_clear(plr);
1393 * rdtgroup_pseudo_lock_remove - Remove a pseudo-locked region
1394 * @rdtgrp: resource group to which the pseudo-locked region belongs
1396 * The removal of a pseudo-locked region can be initiated when the resource
1397 * group is removed from user space via a "rmdir" from userspace or the
1398 * unmount of the resctrl filesystem. On removal the resource group does
1399 * not go back to pseudo-locksetup mode before it is removed, instead it is
1400 * removed directly. There is thus asymmetry with the creation where the
1401 * &struct pseudo_lock_region is removed here while it was not created in
1402 * rdtgroup_pseudo_lock_create().
1406 void rdtgroup_pseudo_lock_remove(struct rdtgroup *rdtgrp)
1408 struct pseudo_lock_region *plr = rdtgrp->plr;
1410 if (rdtgrp->mode == RDT_MODE_PSEUDO_LOCKSETUP) {
1412 * Default group cannot be a pseudo-locked region so we can
1415 closid_free(rdtgrp->closid);
1419 pseudo_lock_cstates_relax(plr);
1420 debugfs_remove_recursive(rdtgrp->plr->debugfs_dir);
1421 device_destroy(pseudo_lock_class, MKDEV(pseudo_lock_major, plr->minor));
1422 pseudo_lock_minor_release(plr->minor);
1425 pseudo_lock_free(rdtgrp);
1428 static int pseudo_lock_dev_open(struct inode *inode, struct file *filp)
1430 struct rdtgroup *rdtgrp;
1432 mutex_lock(&rdtgroup_mutex);
1434 rdtgrp = region_find_by_minor(iminor(inode));
1436 mutex_unlock(&rdtgroup_mutex);
1440 filp->private_data = rdtgrp;
1441 atomic_inc(&rdtgrp->waitcount);
1442 /* Perform a non-seekable open - llseek is not supported */
1443 filp->f_mode &= ~(FMODE_LSEEK | FMODE_PREAD | FMODE_PWRITE);
1445 mutex_unlock(&rdtgroup_mutex);
1450 static int pseudo_lock_dev_release(struct inode *inode, struct file *filp)
1452 struct rdtgroup *rdtgrp;
1454 mutex_lock(&rdtgroup_mutex);
1455 rdtgrp = filp->private_data;
1458 mutex_unlock(&rdtgroup_mutex);
1461 filp->private_data = NULL;
1462 atomic_dec(&rdtgrp->waitcount);
1463 mutex_unlock(&rdtgroup_mutex);
1467 static int pseudo_lock_dev_mremap(struct vm_area_struct *area)
1473 static const struct vm_operations_struct pseudo_mmap_ops = {
1474 .mremap = pseudo_lock_dev_mremap,
1477 static int pseudo_lock_dev_mmap(struct file *filp, struct vm_area_struct *vma)
1479 unsigned long vsize = vma->vm_end - vma->vm_start;
1480 unsigned long off = vma->vm_pgoff << PAGE_SHIFT;
1481 struct pseudo_lock_region *plr;
1482 struct rdtgroup *rdtgrp;
1483 unsigned long physical;
1484 unsigned long psize;
1486 mutex_lock(&rdtgroup_mutex);
1488 rdtgrp = filp->private_data;
1491 mutex_unlock(&rdtgroup_mutex);
1498 mutex_unlock(&rdtgroup_mutex);
1503 * Task is required to run with affinity to the cpus associated
1504 * with the pseudo-locked region. If this is not the case the task
1505 * may be scheduled elsewhere and invalidate entries in the
1506 * pseudo-locked region.
1508 if (!cpumask_subset(current->cpus_ptr, &plr->d->cpu_mask)) {
1509 mutex_unlock(&rdtgroup_mutex);
1513 physical = __pa(plr->kmem) >> PAGE_SHIFT;
1514 psize = plr->size - off;
1516 if (off > plr->size) {
1517 mutex_unlock(&rdtgroup_mutex);
1522 * Ensure changes are carried directly to the memory being mapped,
1523 * do not allow copy-on-write mapping.
1525 if (!(vma->vm_flags & VM_SHARED)) {
1526 mutex_unlock(&rdtgroup_mutex);
1530 if (vsize > psize) {
1531 mutex_unlock(&rdtgroup_mutex);
1535 memset(plr->kmem + off, 0, vsize);
1537 if (remap_pfn_range(vma, vma->vm_start, physical + vma->vm_pgoff,
1538 vsize, vma->vm_page_prot)) {
1539 mutex_unlock(&rdtgroup_mutex);
1542 vma->vm_ops = &pseudo_mmap_ops;
1543 mutex_unlock(&rdtgroup_mutex);
1547 static const struct file_operations pseudo_lock_dev_fops = {
1548 .owner = THIS_MODULE,
1549 .llseek = no_llseek,
1552 .open = pseudo_lock_dev_open,
1553 .release = pseudo_lock_dev_release,
1554 .mmap = pseudo_lock_dev_mmap,
1557 static char *pseudo_lock_devnode(struct device *dev, umode_t *mode)
1559 struct rdtgroup *rdtgrp;
1561 rdtgrp = dev_get_drvdata(dev);
1564 return kasprintf(GFP_KERNEL, "pseudo_lock/%s", rdtgrp->kn->name);
1567 int rdt_pseudo_lock_init(void)
1571 ret = register_chrdev(0, "pseudo_lock", &pseudo_lock_dev_fops);
1575 pseudo_lock_major = ret;
1577 pseudo_lock_class = class_create(THIS_MODULE, "pseudo_lock");
1578 if (IS_ERR(pseudo_lock_class)) {
1579 ret = PTR_ERR(pseudo_lock_class);
1580 unregister_chrdev(pseudo_lock_major, "pseudo_lock");
1584 pseudo_lock_class->devnode = pseudo_lock_devnode;
1588 void rdt_pseudo_lock_release(void)
1590 class_destroy(pseudo_lock_class);
1591 pseudo_lock_class = NULL;
1592 unregister_chrdev(pseudo_lock_major, "pseudo_lock");
1593 pseudo_lock_major = 0;