2 * Performance events x86 architecture code
4 * Copyright (C) 2008 Thomas Gleixner <tglx@linutronix.de>
5 * Copyright (C) 2008-2009 Red Hat, Inc., Ingo Molnar
6 * Copyright (C) 2009 Jaswinder Singh Rajput
7 * Copyright (C) 2009 Advanced Micro Devices, Inc., Robert Richter
8 * Copyright (C) 2008-2009 Red Hat, Inc., Peter Zijlstra <pzijlstr@redhat.com>
9 * Copyright (C) 2009 Intel Corporation, <markus.t.metzger@intel.com>
10 * Copyright (C) 2009 Google, Inc., Stephane Eranian
12 * For licencing details see kernel-base/COPYING
15 #include <linux/perf_event.h>
16 #include <linux/capability.h>
17 #include <linux/notifier.h>
18 #include <linux/hardirq.h>
19 #include <linux/kprobes.h>
20 #include <linux/module.h>
21 #include <linux/kdebug.h>
22 #include <linux/sched.h>
23 #include <linux/uaccess.h>
24 #include <linux/slab.h>
25 #include <linux/cpu.h>
26 #include <linux/bitops.h>
27 #include <linux/device.h>
30 #include <asm/stacktrace.h>
33 #include <asm/alternative.h>
34 #include <asm/timer.h>
38 #include "perf_event.h"
40 struct x86_pmu x86_pmu __read_mostly;
42 DEFINE_PER_CPU(struct cpu_hw_events, cpu_hw_events) = {
46 u64 __read_mostly hw_cache_event_ids
47 [PERF_COUNT_HW_CACHE_MAX]
48 [PERF_COUNT_HW_CACHE_OP_MAX]
49 [PERF_COUNT_HW_CACHE_RESULT_MAX];
50 u64 __read_mostly hw_cache_extra_regs
51 [PERF_COUNT_HW_CACHE_MAX]
52 [PERF_COUNT_HW_CACHE_OP_MAX]
53 [PERF_COUNT_HW_CACHE_RESULT_MAX];
56 * Propagate event elapsed time into the generic event.
57 * Can only be executed on the CPU where the event is active.
58 * Returns the delta events processed.
60 u64 x86_perf_event_update(struct perf_event *event)
62 struct hw_perf_event *hwc = &event->hw;
63 int shift = 64 - x86_pmu.cntval_bits;
64 u64 prev_raw_count, new_raw_count;
68 if (idx == INTEL_PMC_IDX_FIXED_BTS)
72 * Careful: an NMI might modify the previous event value.
74 * Our tactic to handle this is to first atomically read and
75 * exchange a new raw count - then add that new-prev delta
76 * count to the generic event atomically:
79 prev_raw_count = local64_read(&hwc->prev_count);
80 rdpmcl(hwc->event_base_rdpmc, new_raw_count);
82 if (local64_cmpxchg(&hwc->prev_count, prev_raw_count,
83 new_raw_count) != prev_raw_count)
87 * Now we have the new raw value and have updated the prev
88 * timestamp already. We can now calculate the elapsed delta
89 * (event-)time and add that to the generic event.
91 * Careful, not all hw sign-extends above the physical width
94 delta = (new_raw_count << shift) - (prev_raw_count << shift);
97 local64_add(delta, &event->count);
98 local64_sub(delta, &hwc->period_left);
100 return new_raw_count;
104 * Find and validate any extra registers to set up.
106 static int x86_pmu_extra_regs(u64 config, struct perf_event *event)
108 struct hw_perf_event_extra *reg;
109 struct extra_reg *er;
111 reg = &event->hw.extra_reg;
113 if (!x86_pmu.extra_regs)
116 for (er = x86_pmu.extra_regs; er->msr; er++) {
117 if (er->event != (config & er->config_mask))
119 if (event->attr.config1 & ~er->valid_mask)
121 /* Check if the extra msrs can be safely accessed*/
122 if (!er->extra_msr_access)
126 reg->config = event->attr.config1;
133 static atomic_t active_events;
134 static DEFINE_MUTEX(pmc_reserve_mutex);
136 #ifdef CONFIG_X86_LOCAL_APIC
138 static bool reserve_pmc_hardware(void)
142 for (i = 0; i < x86_pmu.num_counters; i++) {
143 if (!reserve_perfctr_nmi(x86_pmu_event_addr(i)))
147 for (i = 0; i < x86_pmu.num_counters; i++) {
148 if (!reserve_evntsel_nmi(x86_pmu_config_addr(i)))
155 for (i--; i >= 0; i--)
156 release_evntsel_nmi(x86_pmu_config_addr(i));
158 i = x86_pmu.num_counters;
161 for (i--; i >= 0; i--)
162 release_perfctr_nmi(x86_pmu_event_addr(i));
167 static void release_pmc_hardware(void)
171 for (i = 0; i < x86_pmu.num_counters; i++) {
172 release_perfctr_nmi(x86_pmu_event_addr(i));
173 release_evntsel_nmi(x86_pmu_config_addr(i));
179 static bool reserve_pmc_hardware(void) { return true; }
180 static void release_pmc_hardware(void) {}
184 static bool check_hw_exists(void)
186 u64 val, val_fail, val_new= ~0;
187 int i, reg, reg_fail, ret = 0;
191 * Check to see if the BIOS enabled any of the counters, if so
194 for (i = 0; i < x86_pmu.num_counters; i++) {
195 reg = x86_pmu_config_addr(i);
196 ret = rdmsrl_safe(reg, &val);
199 if (val & ARCH_PERFMON_EVENTSEL_ENABLE) {
206 if (x86_pmu.num_counters_fixed) {
207 reg = MSR_ARCH_PERFMON_FIXED_CTR_CTRL;
208 ret = rdmsrl_safe(reg, &val);
211 for (i = 0; i < x86_pmu.num_counters_fixed; i++) {
212 if (val & (0x03 << i*4)) {
221 * Read the current value, change it and read it back to see if it
222 * matches, this is needed to detect certain hardware emulators
223 * (qemu/kvm) that don't trap on the MSR access and always return 0s.
225 reg = x86_pmu_event_addr(0);
226 if (rdmsrl_safe(reg, &val))
229 ret = wrmsrl_safe(reg, val);
230 ret |= rdmsrl_safe(reg, &val_new);
231 if (ret || val != val_new)
235 * We still allow the PMU driver to operate:
238 printk(KERN_CONT "Broken BIOS detected, complain to your hardware vendor.\n");
239 printk(KERN_ERR FW_BUG "the BIOS has corrupted hw-PMU resources (MSR %x is %Lx)\n", reg_fail, val_fail);
245 printk(KERN_CONT "Broken PMU hardware detected, using software events only.\n");
246 printk("%sFailed to access perfctr msr (MSR %x is %Lx)\n",
247 boot_cpu_has(X86_FEATURE_HYPERVISOR) ? KERN_INFO : KERN_ERR,
253 static void hw_perf_event_destroy(struct perf_event *event)
255 if (atomic_dec_and_mutex_lock(&active_events, &pmc_reserve_mutex)) {
256 release_pmc_hardware();
257 release_ds_buffers();
258 mutex_unlock(&pmc_reserve_mutex);
262 static inline int x86_pmu_initialized(void)
264 return x86_pmu.handle_irq != NULL;
268 set_ext_hw_attr(struct hw_perf_event *hwc, struct perf_event *event)
270 struct perf_event_attr *attr = &event->attr;
271 unsigned int cache_type, cache_op, cache_result;
274 config = attr->config;
276 cache_type = (config >> 0) & 0xff;
277 if (cache_type >= PERF_COUNT_HW_CACHE_MAX)
280 cache_op = (config >> 8) & 0xff;
281 if (cache_op >= PERF_COUNT_HW_CACHE_OP_MAX)
284 cache_result = (config >> 16) & 0xff;
285 if (cache_result >= PERF_COUNT_HW_CACHE_RESULT_MAX)
288 val = hw_cache_event_ids[cache_type][cache_op][cache_result];
297 attr->config1 = hw_cache_extra_regs[cache_type][cache_op][cache_result];
298 return x86_pmu_extra_regs(val, event);
301 int x86_setup_perfctr(struct perf_event *event)
303 struct perf_event_attr *attr = &event->attr;
304 struct hw_perf_event *hwc = &event->hw;
307 if (!is_sampling_event(event)) {
308 hwc->sample_period = x86_pmu.max_period;
309 hwc->last_period = hwc->sample_period;
310 local64_set(&hwc->period_left, hwc->sample_period);
313 if (attr->type == PERF_TYPE_RAW)
314 return x86_pmu_extra_regs(event->attr.config, event);
316 if (attr->type == PERF_TYPE_HW_CACHE)
317 return set_ext_hw_attr(hwc, event);
319 if (attr->config >= x86_pmu.max_events)
325 config = x86_pmu.event_map(attr->config);
336 if (attr->config == PERF_COUNT_HW_BRANCH_INSTRUCTIONS &&
337 !attr->freq && hwc->sample_period == 1) {
338 /* BTS is not supported by this architecture. */
339 if (!x86_pmu.bts_active)
342 /* BTS is currently only allowed for user-mode. */
343 if (!attr->exclude_kernel)
347 hwc->config |= config;
353 * check that branch_sample_type is compatible with
354 * settings needed for precise_ip > 1 which implies
355 * using the LBR to capture ALL taken branches at the
356 * priv levels of the measurement
358 static inline int precise_br_compat(struct perf_event *event)
360 u64 m = event->attr.branch_sample_type;
363 /* must capture all branches */
364 if (!(m & PERF_SAMPLE_BRANCH_ANY))
367 m &= PERF_SAMPLE_BRANCH_KERNEL | PERF_SAMPLE_BRANCH_USER;
369 if (!event->attr.exclude_user)
370 b |= PERF_SAMPLE_BRANCH_USER;
372 if (!event->attr.exclude_kernel)
373 b |= PERF_SAMPLE_BRANCH_KERNEL;
376 * ignore PERF_SAMPLE_BRANCH_HV, not supported on x86
382 int x86_pmu_hw_config(struct perf_event *event)
384 if (event->attr.precise_ip) {
387 /* Support for constant skid */
388 if (x86_pmu.pebs_active && !x86_pmu.pebs_broken) {
391 /* Support for IP fixup */
392 if (x86_pmu.lbr_nr || x86_pmu.intel_cap.pebs_format >= 2)
396 if (event->attr.precise_ip > precise)
399 * check that PEBS LBR correction does not conflict with
400 * whatever the user is asking with attr->branch_sample_type
402 if (event->attr.precise_ip > 1 &&
403 x86_pmu.intel_cap.pebs_format < 2) {
404 u64 *br_type = &event->attr.branch_sample_type;
406 if (has_branch_stack(event)) {
407 if (!precise_br_compat(event))
410 /* branch_sample_type is compatible */
414 * user did not specify branch_sample_type
416 * For PEBS fixups, we capture all
417 * the branches at the priv level of the
420 *br_type = PERF_SAMPLE_BRANCH_ANY;
422 if (!event->attr.exclude_user)
423 *br_type |= PERF_SAMPLE_BRANCH_USER;
425 if (!event->attr.exclude_kernel)
426 *br_type |= PERF_SAMPLE_BRANCH_KERNEL;
433 * (keep 'enabled' bit clear for now)
435 event->hw.config = ARCH_PERFMON_EVENTSEL_INT;
438 * Count user and OS events unless requested not to
440 if (!event->attr.exclude_user)
441 event->hw.config |= ARCH_PERFMON_EVENTSEL_USR;
442 if (!event->attr.exclude_kernel)
443 event->hw.config |= ARCH_PERFMON_EVENTSEL_OS;
445 if (event->attr.type == PERF_TYPE_RAW)
446 event->hw.config |= event->attr.config & X86_RAW_EVENT_MASK;
448 return x86_setup_perfctr(event);
452 * Setup the hardware configuration for a given attr_type
454 static int __x86_pmu_event_init(struct perf_event *event)
458 if (!x86_pmu_initialized())
462 if (!atomic_inc_not_zero(&active_events)) {
463 mutex_lock(&pmc_reserve_mutex);
464 if (atomic_read(&active_events) == 0) {
465 if (!reserve_pmc_hardware())
468 reserve_ds_buffers();
471 atomic_inc(&active_events);
472 mutex_unlock(&pmc_reserve_mutex);
477 event->destroy = hw_perf_event_destroy;
480 event->hw.last_cpu = -1;
481 event->hw.last_tag = ~0ULL;
484 event->hw.extra_reg.idx = EXTRA_REG_NONE;
485 event->hw.branch_reg.idx = EXTRA_REG_NONE;
487 return x86_pmu.hw_config(event);
490 void x86_pmu_disable_all(void)
492 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
495 for (idx = 0; idx < x86_pmu.num_counters; idx++) {
498 if (!test_bit(idx, cpuc->active_mask))
500 rdmsrl(x86_pmu_config_addr(idx), val);
501 if (!(val & ARCH_PERFMON_EVENTSEL_ENABLE))
503 val &= ~ARCH_PERFMON_EVENTSEL_ENABLE;
504 wrmsrl(x86_pmu_config_addr(idx), val);
508 static void x86_pmu_disable(struct pmu *pmu)
510 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
512 if (!x86_pmu_initialized())
522 x86_pmu.disable_all();
525 void x86_pmu_enable_all(int added)
527 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
530 for (idx = 0; idx < x86_pmu.num_counters; idx++) {
531 struct hw_perf_event *hwc = &cpuc->events[idx]->hw;
533 if (!test_bit(idx, cpuc->active_mask))
536 __x86_pmu_enable_event(hwc, ARCH_PERFMON_EVENTSEL_ENABLE);
540 static struct pmu pmu;
542 static inline int is_x86_event(struct perf_event *event)
544 return event->pmu == &pmu;
548 * Event scheduler state:
550 * Assign events iterating over all events and counters, beginning
551 * with events with least weights first. Keep the current iterator
552 * state in struct sched_state.
556 int event; /* event index */
557 int counter; /* counter index */
558 int unassigned; /* number of events to be assigned left */
559 unsigned long used[BITS_TO_LONGS(X86_PMC_IDX_MAX)];
562 /* Total max is X86_PMC_IDX_MAX, but we are O(n!) limited */
563 #define SCHED_STATES_MAX 2
568 struct perf_event **events;
569 struct sched_state state;
571 struct sched_state saved[SCHED_STATES_MAX];
575 * Initialize interator that runs through all events and counters.
577 static void perf_sched_init(struct perf_sched *sched, struct perf_event **events,
578 int num, int wmin, int wmax)
582 memset(sched, 0, sizeof(*sched));
583 sched->max_events = num;
584 sched->max_weight = wmax;
585 sched->events = events;
587 for (idx = 0; idx < num; idx++) {
588 if (events[idx]->hw.constraint->weight == wmin)
592 sched->state.event = idx; /* start with min weight */
593 sched->state.weight = wmin;
594 sched->state.unassigned = num;
597 static void perf_sched_save_state(struct perf_sched *sched)
599 if (WARN_ON_ONCE(sched->saved_states >= SCHED_STATES_MAX))
602 sched->saved[sched->saved_states] = sched->state;
603 sched->saved_states++;
606 static bool perf_sched_restore_state(struct perf_sched *sched)
608 if (!sched->saved_states)
611 sched->saved_states--;
612 sched->state = sched->saved[sched->saved_states];
614 /* continue with next counter: */
615 clear_bit(sched->state.counter++, sched->state.used);
621 * Select a counter for the current event to schedule. Return true on
624 static bool __perf_sched_find_counter(struct perf_sched *sched)
626 struct event_constraint *c;
629 if (!sched->state.unassigned)
632 if (sched->state.event >= sched->max_events)
635 c = sched->events[sched->state.event]->hw.constraint;
636 /* Prefer fixed purpose counters */
637 if (c->idxmsk64 & (~0ULL << INTEL_PMC_IDX_FIXED)) {
638 idx = INTEL_PMC_IDX_FIXED;
639 for_each_set_bit_from(idx, c->idxmsk, X86_PMC_IDX_MAX) {
640 if (!__test_and_set_bit(idx, sched->state.used))
644 /* Grab the first unused counter starting with idx */
645 idx = sched->state.counter;
646 for_each_set_bit_from(idx, c->idxmsk, INTEL_PMC_IDX_FIXED) {
647 if (!__test_and_set_bit(idx, sched->state.used))
654 sched->state.counter = idx;
657 perf_sched_save_state(sched);
662 static bool perf_sched_find_counter(struct perf_sched *sched)
664 while (!__perf_sched_find_counter(sched)) {
665 if (!perf_sched_restore_state(sched))
673 * Go through all unassigned events and find the next one to schedule.
674 * Take events with the least weight first. Return true on success.
676 static bool perf_sched_next_event(struct perf_sched *sched)
678 struct event_constraint *c;
680 if (!sched->state.unassigned || !--sched->state.unassigned)
685 sched->state.event++;
686 if (sched->state.event >= sched->max_events) {
688 sched->state.event = 0;
689 sched->state.weight++;
690 if (sched->state.weight > sched->max_weight)
693 c = sched->events[sched->state.event]->hw.constraint;
694 } while (c->weight != sched->state.weight);
696 sched->state.counter = 0; /* start with first counter */
702 * Assign a counter for each event.
704 int perf_assign_events(struct perf_event **events, int n,
705 int wmin, int wmax, int *assign)
707 struct perf_sched sched;
709 perf_sched_init(&sched, events, n, wmin, wmax);
712 if (!perf_sched_find_counter(&sched))
715 assign[sched.state.event] = sched.state.counter;
716 } while (perf_sched_next_event(&sched));
718 return sched.state.unassigned;
720 EXPORT_SYMBOL_GPL(perf_assign_events);
722 int x86_schedule_events(struct cpu_hw_events *cpuc, int n, int *assign)
724 struct event_constraint *c;
725 unsigned long used_mask[BITS_TO_LONGS(X86_PMC_IDX_MAX)];
726 struct perf_event *e;
727 int i, wmin, wmax, num = 0;
728 struct hw_perf_event *hwc;
730 bitmap_zero(used_mask, X86_PMC_IDX_MAX);
732 for (i = 0, wmin = X86_PMC_IDX_MAX, wmax = 0; i < n; i++) {
733 hwc = &cpuc->event_list[i]->hw;
734 c = x86_pmu.get_event_constraints(cpuc, cpuc->event_list[i]);
737 wmin = min(wmin, c->weight);
738 wmax = max(wmax, c->weight);
742 * fastpath, try to reuse previous register
744 for (i = 0; i < n; i++) {
745 hwc = &cpuc->event_list[i]->hw;
752 /* constraint still honored */
753 if (!test_bit(hwc->idx, c->idxmsk))
756 /* not already used */
757 if (test_bit(hwc->idx, used_mask))
760 __set_bit(hwc->idx, used_mask);
762 assign[i] = hwc->idx;
767 num = perf_assign_events(cpuc->event_list, n, wmin,
771 * Mark the event as committed, so we do not put_constraint()
772 * in case new events are added and fail scheduling.
774 if (!num && assign) {
775 for (i = 0; i < n; i++) {
776 e = cpuc->event_list[i];
777 e->hw.flags |= PERF_X86_EVENT_COMMITTED;
781 * scheduling failed or is just a simulation,
782 * free resources if necessary
784 if (!assign || num) {
785 for (i = 0; i < n; i++) {
786 e = cpuc->event_list[i];
788 * do not put_constraint() on comitted events,
789 * because they are good to go
791 if ((e->hw.flags & PERF_X86_EVENT_COMMITTED))
794 if (x86_pmu.put_event_constraints)
795 x86_pmu.put_event_constraints(cpuc, e);
798 return num ? -EINVAL : 0;
802 * dogrp: true if must collect siblings events (group)
803 * returns total number of events and error code
805 static int collect_events(struct cpu_hw_events *cpuc, struct perf_event *leader, bool dogrp)
807 struct perf_event *event;
810 max_count = x86_pmu.num_counters + x86_pmu.num_counters_fixed;
812 /* current number of events already accepted */
815 if (is_x86_event(leader)) {
818 cpuc->event_list[n] = leader;
824 list_for_each_entry(event, &leader->sibling_list, group_entry) {
825 if (!is_x86_event(event) ||
826 event->state <= PERF_EVENT_STATE_OFF)
832 cpuc->event_list[n] = event;
838 static inline void x86_assign_hw_event(struct perf_event *event,
839 struct cpu_hw_events *cpuc, int i)
841 struct hw_perf_event *hwc = &event->hw;
843 hwc->idx = cpuc->assign[i];
844 hwc->last_cpu = smp_processor_id();
845 hwc->last_tag = ++cpuc->tags[i];
847 if (hwc->idx == INTEL_PMC_IDX_FIXED_BTS) {
848 hwc->config_base = 0;
850 } else if (hwc->idx >= INTEL_PMC_IDX_FIXED) {
851 hwc->config_base = MSR_ARCH_PERFMON_FIXED_CTR_CTRL;
852 hwc->event_base = MSR_ARCH_PERFMON_FIXED_CTR0 + (hwc->idx - INTEL_PMC_IDX_FIXED);
853 hwc->event_base_rdpmc = (hwc->idx - INTEL_PMC_IDX_FIXED) | 1<<30;
855 hwc->config_base = x86_pmu_config_addr(hwc->idx);
856 hwc->event_base = x86_pmu_event_addr(hwc->idx);
857 hwc->event_base_rdpmc = x86_pmu_rdpmc_index(hwc->idx);
861 static inline int match_prev_assignment(struct hw_perf_event *hwc,
862 struct cpu_hw_events *cpuc,
865 return hwc->idx == cpuc->assign[i] &&
866 hwc->last_cpu == smp_processor_id() &&
867 hwc->last_tag == cpuc->tags[i];
870 static void x86_pmu_start(struct perf_event *event, int flags);
872 static void x86_pmu_enable(struct pmu *pmu)
874 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
875 struct perf_event *event;
876 struct hw_perf_event *hwc;
877 int i, added = cpuc->n_added;
879 if (!x86_pmu_initialized())
886 int n_running = cpuc->n_events - cpuc->n_added;
888 * apply assignment obtained either from
889 * hw_perf_group_sched_in() or x86_pmu_enable()
891 * step1: save events moving to new counters
893 for (i = 0; i < n_running; i++) {
894 event = cpuc->event_list[i];
898 * we can avoid reprogramming counter if:
899 * - assigned same counter as last time
900 * - running on same CPU as last time
901 * - no other event has used the counter since
903 if (hwc->idx == -1 ||
904 match_prev_assignment(hwc, cpuc, i))
908 * Ensure we don't accidentally enable a stopped
909 * counter simply because we rescheduled.
911 if (hwc->state & PERF_HES_STOPPED)
912 hwc->state |= PERF_HES_ARCH;
914 x86_pmu_stop(event, PERF_EF_UPDATE);
918 * step2: reprogram moved events into new counters
920 for (i = 0; i < cpuc->n_events; i++) {
921 event = cpuc->event_list[i];
924 if (!match_prev_assignment(hwc, cpuc, i))
925 x86_assign_hw_event(event, cpuc, i);
926 else if (i < n_running)
929 if (hwc->state & PERF_HES_ARCH)
932 x86_pmu_start(event, PERF_EF_RELOAD);
935 perf_events_lapic_init();
941 x86_pmu.enable_all(added);
944 static DEFINE_PER_CPU(u64 [X86_PMC_IDX_MAX], pmc_prev_left);
947 * Set the next IRQ period, based on the hwc->period_left value.
948 * To be called with the event disabled in hw:
950 int x86_perf_event_set_period(struct perf_event *event)
952 struct hw_perf_event *hwc = &event->hw;
953 s64 left = local64_read(&hwc->period_left);
954 s64 period = hwc->sample_period;
955 int ret = 0, idx = hwc->idx;
957 if (idx == INTEL_PMC_IDX_FIXED_BTS)
961 * If we are way outside a reasonable range then just skip forward:
963 if (unlikely(left <= -period)) {
965 local64_set(&hwc->period_left, left);
966 hwc->last_period = period;
970 if (unlikely(left <= 0)) {
972 local64_set(&hwc->period_left, left);
973 hwc->last_period = period;
977 * Quirk: certain CPUs dont like it if just 1 hw_event is left:
979 if (unlikely(left < 2))
982 if (left > x86_pmu.max_period)
983 left = x86_pmu.max_period;
985 per_cpu(pmc_prev_left[idx], smp_processor_id()) = left;
988 * The hw event starts counting from this event offset,
989 * mark it to be able to extra future deltas:
991 local64_set(&hwc->prev_count, (u64)-left);
993 wrmsrl(hwc->event_base, (u64)(-left) & x86_pmu.cntval_mask);
996 * Due to erratum on certan cpu we need
997 * a second write to be sure the register
998 * is updated properly
1000 if (x86_pmu.perfctr_second_write) {
1001 wrmsrl(hwc->event_base,
1002 (u64)(-left) & x86_pmu.cntval_mask);
1005 perf_event_update_userpage(event);
1010 void x86_pmu_enable_event(struct perf_event *event)
1012 if (__this_cpu_read(cpu_hw_events.enabled))
1013 __x86_pmu_enable_event(&event->hw,
1014 ARCH_PERFMON_EVENTSEL_ENABLE);
1018 * Add a single event to the PMU.
1020 * The event is added to the group of enabled events
1021 * but only if it can be scehduled with existing events.
1023 static int x86_pmu_add(struct perf_event *event, int flags)
1025 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1026 struct hw_perf_event *hwc;
1027 int assign[X86_PMC_IDX_MAX];
1032 perf_pmu_disable(event->pmu);
1033 n0 = cpuc->n_events;
1034 ret = n = collect_events(cpuc, event, false);
1038 hwc->state = PERF_HES_UPTODATE | PERF_HES_STOPPED;
1039 if (!(flags & PERF_EF_START))
1040 hwc->state |= PERF_HES_ARCH;
1043 * If group events scheduling transaction was started,
1044 * skip the schedulability test here, it will be performed
1045 * at commit time (->commit_txn) as a whole.
1047 if (cpuc->group_flag & PERF_EVENT_TXN)
1050 ret = x86_pmu.schedule_events(cpuc, n, assign);
1054 * copy new assignment, now we know it is possible
1055 * will be used by hw_perf_enable()
1057 memcpy(cpuc->assign, assign, n*sizeof(int));
1061 * Commit the collect_events() state. See x86_pmu_del() and
1065 cpuc->n_added += n - n0;
1066 cpuc->n_txn += n - n0;
1070 perf_pmu_enable(event->pmu);
1074 static void x86_pmu_start(struct perf_event *event, int flags)
1076 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1077 int idx = event->hw.idx;
1079 if (WARN_ON_ONCE(!(event->hw.state & PERF_HES_STOPPED)))
1082 if (WARN_ON_ONCE(idx == -1))
1085 if (flags & PERF_EF_RELOAD) {
1086 WARN_ON_ONCE(!(event->hw.state & PERF_HES_UPTODATE));
1087 x86_perf_event_set_period(event);
1090 event->hw.state = 0;
1092 cpuc->events[idx] = event;
1093 __set_bit(idx, cpuc->active_mask);
1094 __set_bit(idx, cpuc->running);
1095 x86_pmu.enable(event);
1096 perf_event_update_userpage(event);
1099 void perf_event_print_debug(void)
1101 u64 ctrl, status, overflow, pmc_ctrl, pmc_count, prev_left, fixed;
1103 struct cpu_hw_events *cpuc;
1104 unsigned long flags;
1107 if (!x86_pmu.num_counters)
1110 local_irq_save(flags);
1112 cpu = smp_processor_id();
1113 cpuc = &per_cpu(cpu_hw_events, cpu);
1115 if (x86_pmu.version >= 2) {
1116 rdmsrl(MSR_CORE_PERF_GLOBAL_CTRL, ctrl);
1117 rdmsrl(MSR_CORE_PERF_GLOBAL_STATUS, status);
1118 rdmsrl(MSR_CORE_PERF_GLOBAL_OVF_CTRL, overflow);
1119 rdmsrl(MSR_ARCH_PERFMON_FIXED_CTR_CTRL, fixed);
1120 rdmsrl(MSR_IA32_PEBS_ENABLE, pebs);
1123 pr_info("CPU#%d: ctrl: %016llx\n", cpu, ctrl);
1124 pr_info("CPU#%d: status: %016llx\n", cpu, status);
1125 pr_info("CPU#%d: overflow: %016llx\n", cpu, overflow);
1126 pr_info("CPU#%d: fixed: %016llx\n", cpu, fixed);
1127 pr_info("CPU#%d: pebs: %016llx\n", cpu, pebs);
1129 pr_info("CPU#%d: active: %016llx\n", cpu, *(u64 *)cpuc->active_mask);
1131 for (idx = 0; idx < x86_pmu.num_counters; idx++) {
1132 rdmsrl(x86_pmu_config_addr(idx), pmc_ctrl);
1133 rdmsrl(x86_pmu_event_addr(idx), pmc_count);
1135 prev_left = per_cpu(pmc_prev_left[idx], cpu);
1137 pr_info("CPU#%d: gen-PMC%d ctrl: %016llx\n",
1138 cpu, idx, pmc_ctrl);
1139 pr_info("CPU#%d: gen-PMC%d count: %016llx\n",
1140 cpu, idx, pmc_count);
1141 pr_info("CPU#%d: gen-PMC%d left: %016llx\n",
1142 cpu, idx, prev_left);
1144 for (idx = 0; idx < x86_pmu.num_counters_fixed; idx++) {
1145 rdmsrl(MSR_ARCH_PERFMON_FIXED_CTR0 + idx, pmc_count);
1147 pr_info("CPU#%d: fixed-PMC%d count: %016llx\n",
1148 cpu, idx, pmc_count);
1150 local_irq_restore(flags);
1153 void x86_pmu_stop(struct perf_event *event, int flags)
1155 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1156 struct hw_perf_event *hwc = &event->hw;
1158 if (__test_and_clear_bit(hwc->idx, cpuc->active_mask)) {
1159 x86_pmu.disable(event);
1160 cpuc->events[hwc->idx] = NULL;
1161 WARN_ON_ONCE(hwc->state & PERF_HES_STOPPED);
1162 hwc->state |= PERF_HES_STOPPED;
1165 if ((flags & PERF_EF_UPDATE) && !(hwc->state & PERF_HES_UPTODATE)) {
1167 * Drain the remaining delta count out of a event
1168 * that we are disabling:
1170 x86_perf_event_update(event);
1171 hwc->state |= PERF_HES_UPTODATE;
1175 static void x86_pmu_del(struct perf_event *event, int flags)
1177 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1181 * event is descheduled
1183 event->hw.flags &= ~PERF_X86_EVENT_COMMITTED;
1186 * If we're called during a txn, we don't need to do anything.
1187 * The events never got scheduled and ->cancel_txn will truncate
1190 * XXX assumes any ->del() called during a TXN will only be on
1191 * an event added during that same TXN.
1193 if (cpuc->group_flag & PERF_EVENT_TXN)
1197 * Not a TXN, therefore cleanup properly.
1199 x86_pmu_stop(event, PERF_EF_UPDATE);
1201 for (i = 0; i < cpuc->n_events; i++) {
1202 if (event == cpuc->event_list[i])
1206 if (WARN_ON_ONCE(i == cpuc->n_events)) /* called ->del() without ->add() ? */
1209 /* If we have a newly added event; make sure to decrease n_added. */
1210 if (i >= cpuc->n_events - cpuc->n_added)
1213 if (x86_pmu.put_event_constraints)
1214 x86_pmu.put_event_constraints(cpuc, event);
1216 /* Delete the array entry. */
1217 while (++i < cpuc->n_events)
1218 cpuc->event_list[i-1] = cpuc->event_list[i];
1221 perf_event_update_userpage(event);
1224 int x86_pmu_handle_irq(struct pt_regs *regs)
1226 struct perf_sample_data data;
1227 struct cpu_hw_events *cpuc;
1228 struct perf_event *event;
1229 int idx, handled = 0;
1232 cpuc = this_cpu_ptr(&cpu_hw_events);
1235 * Some chipsets need to unmask the LVTPC in a particular spot
1236 * inside the nmi handler. As a result, the unmasking was pushed
1237 * into all the nmi handlers.
1239 * This generic handler doesn't seem to have any issues where the
1240 * unmasking occurs so it was left at the top.
1242 apic_write(APIC_LVTPC, APIC_DM_NMI);
1244 for (idx = 0; idx < x86_pmu.num_counters; idx++) {
1245 if (!test_bit(idx, cpuc->active_mask)) {
1247 * Though we deactivated the counter some cpus
1248 * might still deliver spurious interrupts still
1249 * in flight. Catch them:
1251 if (__test_and_clear_bit(idx, cpuc->running))
1256 event = cpuc->events[idx];
1258 val = x86_perf_event_update(event);
1259 if (val & (1ULL << (x86_pmu.cntval_bits - 1)))
1266 perf_sample_data_init(&data, 0, event->hw.last_period);
1268 if (!x86_perf_event_set_period(event))
1271 if (perf_event_overflow(event, &data, regs))
1272 x86_pmu_stop(event, 0);
1276 inc_irq_stat(apic_perf_irqs);
1281 void perf_events_lapic_init(void)
1283 if (!x86_pmu.apic || !x86_pmu_initialized())
1287 * Always use NMI for PMU
1289 apic_write(APIC_LVTPC, APIC_DM_NMI);
1293 perf_event_nmi_handler(unsigned int cmd, struct pt_regs *regs)
1299 if (!atomic_read(&active_events))
1302 start_clock = sched_clock();
1303 ret = x86_pmu.handle_irq(regs);
1304 finish_clock = sched_clock();
1306 perf_sample_event_took(finish_clock - start_clock);
1310 NOKPROBE_SYMBOL(perf_event_nmi_handler);
1312 struct event_constraint emptyconstraint;
1313 struct event_constraint unconstrained;
1316 x86_pmu_notifier(struct notifier_block *self, unsigned long action, void *hcpu)
1318 unsigned int cpu = (long)hcpu;
1319 struct cpu_hw_events *cpuc = &per_cpu(cpu_hw_events, cpu);
1320 int ret = NOTIFY_OK;
1322 switch (action & ~CPU_TASKS_FROZEN) {
1323 case CPU_UP_PREPARE:
1324 cpuc->kfree_on_online = NULL;
1325 if (x86_pmu.cpu_prepare)
1326 ret = x86_pmu.cpu_prepare(cpu);
1330 if (x86_pmu.attr_rdpmc)
1331 set_in_cr4(X86_CR4_PCE);
1332 if (x86_pmu.cpu_starting)
1333 x86_pmu.cpu_starting(cpu);
1337 kfree(cpuc->kfree_on_online);
1341 if (x86_pmu.cpu_dying)
1342 x86_pmu.cpu_dying(cpu);
1345 case CPU_UP_CANCELED:
1347 if (x86_pmu.cpu_dead)
1348 x86_pmu.cpu_dead(cpu);
1358 static void __init pmu_check_apic(void)
1364 pr_info("no APIC, boot with the \"lapic\" boot parameter to force-enable it.\n");
1365 pr_info("no hardware sampling interrupt available.\n");
1368 * If we have a PMU initialized but no APIC
1369 * interrupts, we cannot sample hardware
1370 * events (user-space has to fall back and
1371 * sample via a hrtimer based software event):
1373 pmu.capabilities |= PERF_PMU_CAP_NO_INTERRUPT;
1377 static struct attribute_group x86_pmu_format_group = {
1383 * Remove all undefined events (x86_pmu.event_map(id) == 0)
1384 * out of events_attr attributes.
1386 static void __init filter_events(struct attribute **attrs)
1388 struct device_attribute *d;
1389 struct perf_pmu_events_attr *pmu_attr;
1392 for (i = 0; attrs[i]; i++) {
1393 d = (struct device_attribute *)attrs[i];
1394 pmu_attr = container_of(d, struct perf_pmu_events_attr, attr);
1396 if (pmu_attr->event_str)
1398 if (x86_pmu.event_map(i))
1401 for (j = i; attrs[j]; j++)
1402 attrs[j] = attrs[j + 1];
1404 /* Check the shifted attr. */
1409 /* Merge two pointer arrays */
1410 static __init struct attribute **merge_attr(struct attribute **a, struct attribute **b)
1412 struct attribute **new;
1415 for (j = 0; a[j]; j++)
1417 for (i = 0; b[i]; i++)
1421 new = kmalloc(sizeof(struct attribute *) * j, GFP_KERNEL);
1426 for (i = 0; a[i]; i++)
1428 for (i = 0; b[i]; i++)
1435 ssize_t events_sysfs_show(struct device *dev, struct device_attribute *attr,
1438 struct perf_pmu_events_attr *pmu_attr = \
1439 container_of(attr, struct perf_pmu_events_attr, attr);
1440 u64 config = x86_pmu.event_map(pmu_attr->id);
1442 /* string trumps id */
1443 if (pmu_attr->event_str)
1444 return sprintf(page, "%s", pmu_attr->event_str);
1446 return x86_pmu.events_sysfs_show(page, config);
1449 EVENT_ATTR(cpu-cycles, CPU_CYCLES );
1450 EVENT_ATTR(instructions, INSTRUCTIONS );
1451 EVENT_ATTR(cache-references, CACHE_REFERENCES );
1452 EVENT_ATTR(cache-misses, CACHE_MISSES );
1453 EVENT_ATTR(branch-instructions, BRANCH_INSTRUCTIONS );
1454 EVENT_ATTR(branch-misses, BRANCH_MISSES );
1455 EVENT_ATTR(bus-cycles, BUS_CYCLES );
1456 EVENT_ATTR(stalled-cycles-frontend, STALLED_CYCLES_FRONTEND );
1457 EVENT_ATTR(stalled-cycles-backend, STALLED_CYCLES_BACKEND );
1458 EVENT_ATTR(ref-cycles, REF_CPU_CYCLES );
1460 static struct attribute *empty_attrs;
1462 static struct attribute *events_attr[] = {
1463 EVENT_PTR(CPU_CYCLES),
1464 EVENT_PTR(INSTRUCTIONS),
1465 EVENT_PTR(CACHE_REFERENCES),
1466 EVENT_PTR(CACHE_MISSES),
1467 EVENT_PTR(BRANCH_INSTRUCTIONS),
1468 EVENT_PTR(BRANCH_MISSES),
1469 EVENT_PTR(BUS_CYCLES),
1470 EVENT_PTR(STALLED_CYCLES_FRONTEND),
1471 EVENT_PTR(STALLED_CYCLES_BACKEND),
1472 EVENT_PTR(REF_CPU_CYCLES),
1476 static struct attribute_group x86_pmu_events_group = {
1478 .attrs = events_attr,
1481 ssize_t x86_event_sysfs_show(char *page, u64 config, u64 event)
1483 u64 umask = (config & ARCH_PERFMON_EVENTSEL_UMASK) >> 8;
1484 u64 cmask = (config & ARCH_PERFMON_EVENTSEL_CMASK) >> 24;
1485 bool edge = (config & ARCH_PERFMON_EVENTSEL_EDGE);
1486 bool pc = (config & ARCH_PERFMON_EVENTSEL_PIN_CONTROL);
1487 bool any = (config & ARCH_PERFMON_EVENTSEL_ANY);
1488 bool inv = (config & ARCH_PERFMON_EVENTSEL_INV);
1492 * We have whole page size to spend and just little data
1493 * to write, so we can safely use sprintf.
1495 ret = sprintf(page, "event=0x%02llx", event);
1498 ret += sprintf(page + ret, ",umask=0x%02llx", umask);
1501 ret += sprintf(page + ret, ",edge");
1504 ret += sprintf(page + ret, ",pc");
1507 ret += sprintf(page + ret, ",any");
1510 ret += sprintf(page + ret, ",inv");
1513 ret += sprintf(page + ret, ",cmask=0x%02llx", cmask);
1515 ret += sprintf(page + ret, "\n");
1520 static int __init init_hw_perf_events(void)
1522 struct x86_pmu_quirk *quirk;
1525 pr_info("Performance Events: ");
1527 switch (boot_cpu_data.x86_vendor) {
1528 case X86_VENDOR_INTEL:
1529 err = intel_pmu_init();
1531 case X86_VENDOR_AMD:
1532 err = amd_pmu_init();
1538 pr_cont("no PMU driver, software events only.\n");
1544 /* sanity check that the hardware exists or is emulated */
1545 if (!check_hw_exists())
1548 pr_cont("%s PMU driver.\n", x86_pmu.name);
1550 x86_pmu.attr_rdpmc = 1; /* enable userspace RDPMC usage by default */
1552 for (quirk = x86_pmu.quirks; quirk; quirk = quirk->next)
1555 if (!x86_pmu.intel_ctrl)
1556 x86_pmu.intel_ctrl = (1 << x86_pmu.num_counters) - 1;
1558 perf_events_lapic_init();
1559 register_nmi_handler(NMI_LOCAL, perf_event_nmi_handler, 0, "PMI");
1561 unconstrained = (struct event_constraint)
1562 __EVENT_CONSTRAINT(0, (1ULL << x86_pmu.num_counters) - 1,
1563 0, x86_pmu.num_counters, 0, 0);
1565 x86_pmu_format_group.attrs = x86_pmu.format_attrs;
1567 if (x86_pmu.event_attrs)
1568 x86_pmu_events_group.attrs = x86_pmu.event_attrs;
1570 if (!x86_pmu.events_sysfs_show)
1571 x86_pmu_events_group.attrs = &empty_attrs;
1573 filter_events(x86_pmu_events_group.attrs);
1575 if (x86_pmu.cpu_events) {
1576 struct attribute **tmp;
1578 tmp = merge_attr(x86_pmu_events_group.attrs, x86_pmu.cpu_events);
1580 x86_pmu_events_group.attrs = tmp;
1583 pr_info("... version: %d\n", x86_pmu.version);
1584 pr_info("... bit width: %d\n", x86_pmu.cntval_bits);
1585 pr_info("... generic registers: %d\n", x86_pmu.num_counters);
1586 pr_info("... value mask: %016Lx\n", x86_pmu.cntval_mask);
1587 pr_info("... max period: %016Lx\n", x86_pmu.max_period);
1588 pr_info("... fixed-purpose events: %d\n", x86_pmu.num_counters_fixed);
1589 pr_info("... event mask: %016Lx\n", x86_pmu.intel_ctrl);
1591 perf_pmu_register(&pmu, "cpu", PERF_TYPE_RAW);
1592 perf_cpu_notifier(x86_pmu_notifier);
1596 early_initcall(init_hw_perf_events);
1598 static inline void x86_pmu_read(struct perf_event *event)
1600 x86_perf_event_update(event);
1604 * Start group events scheduling transaction
1605 * Set the flag to make pmu::enable() not perform the
1606 * schedulability test, it will be performed at commit time
1608 static void x86_pmu_start_txn(struct pmu *pmu)
1610 perf_pmu_disable(pmu);
1611 __this_cpu_or(cpu_hw_events.group_flag, PERF_EVENT_TXN);
1612 __this_cpu_write(cpu_hw_events.n_txn, 0);
1616 * Stop group events scheduling transaction
1617 * Clear the flag and pmu::enable() will perform the
1618 * schedulability test.
1620 static void x86_pmu_cancel_txn(struct pmu *pmu)
1622 __this_cpu_and(cpu_hw_events.group_flag, ~PERF_EVENT_TXN);
1624 * Truncate collected array by the number of events added in this
1625 * transaction. See x86_pmu_add() and x86_pmu_*_txn().
1627 __this_cpu_sub(cpu_hw_events.n_added, __this_cpu_read(cpu_hw_events.n_txn));
1628 __this_cpu_sub(cpu_hw_events.n_events, __this_cpu_read(cpu_hw_events.n_txn));
1629 perf_pmu_enable(pmu);
1633 * Commit group events scheduling transaction
1634 * Perform the group schedulability test as a whole
1635 * Return 0 if success
1637 * Does not cancel the transaction on failure; expects the caller to do this.
1639 static int x86_pmu_commit_txn(struct pmu *pmu)
1641 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1642 int assign[X86_PMC_IDX_MAX];
1647 if (!x86_pmu_initialized())
1650 ret = x86_pmu.schedule_events(cpuc, n, assign);
1655 * copy new assignment, now we know it is possible
1656 * will be used by hw_perf_enable()
1658 memcpy(cpuc->assign, assign, n*sizeof(int));
1660 cpuc->group_flag &= ~PERF_EVENT_TXN;
1661 perf_pmu_enable(pmu);
1665 * a fake_cpuc is used to validate event groups. Due to
1666 * the extra reg logic, we need to also allocate a fake
1667 * per_core and per_cpu structure. Otherwise, group events
1668 * using extra reg may conflict without the kernel being
1669 * able to catch this when the last event gets added to
1672 static void free_fake_cpuc(struct cpu_hw_events *cpuc)
1674 kfree(cpuc->shared_regs);
1678 static struct cpu_hw_events *allocate_fake_cpuc(void)
1680 struct cpu_hw_events *cpuc;
1681 int cpu = raw_smp_processor_id();
1683 cpuc = kzalloc(sizeof(*cpuc), GFP_KERNEL);
1685 return ERR_PTR(-ENOMEM);
1687 /* only needed, if we have extra_regs */
1688 if (x86_pmu.extra_regs) {
1689 cpuc->shared_regs = allocate_shared_regs(cpu);
1690 if (!cpuc->shared_regs)
1696 free_fake_cpuc(cpuc);
1697 return ERR_PTR(-ENOMEM);
1701 * validate that we can schedule this event
1703 static int validate_event(struct perf_event *event)
1705 struct cpu_hw_events *fake_cpuc;
1706 struct event_constraint *c;
1709 fake_cpuc = allocate_fake_cpuc();
1710 if (IS_ERR(fake_cpuc))
1711 return PTR_ERR(fake_cpuc);
1713 c = x86_pmu.get_event_constraints(fake_cpuc, event);
1715 if (!c || !c->weight)
1718 if (x86_pmu.put_event_constraints)
1719 x86_pmu.put_event_constraints(fake_cpuc, event);
1721 free_fake_cpuc(fake_cpuc);
1727 * validate a single event group
1729 * validation include:
1730 * - check events are compatible which each other
1731 * - events do not compete for the same counter
1732 * - number of events <= number of counters
1734 * validation ensures the group can be loaded onto the
1735 * PMU if it was the only group available.
1737 static int validate_group(struct perf_event *event)
1739 struct perf_event *leader = event->group_leader;
1740 struct cpu_hw_events *fake_cpuc;
1741 int ret = -EINVAL, n;
1743 fake_cpuc = allocate_fake_cpuc();
1744 if (IS_ERR(fake_cpuc))
1745 return PTR_ERR(fake_cpuc);
1747 * the event is not yet connected with its
1748 * siblings therefore we must first collect
1749 * existing siblings, then add the new event
1750 * before we can simulate the scheduling
1752 n = collect_events(fake_cpuc, leader, true);
1756 fake_cpuc->n_events = n;
1757 n = collect_events(fake_cpuc, event, false);
1761 fake_cpuc->n_events = n;
1763 ret = x86_pmu.schedule_events(fake_cpuc, n, NULL);
1766 free_fake_cpuc(fake_cpuc);
1770 static int x86_pmu_event_init(struct perf_event *event)
1775 switch (event->attr.type) {
1777 case PERF_TYPE_HARDWARE:
1778 case PERF_TYPE_HW_CACHE:
1785 err = __x86_pmu_event_init(event);
1788 * we temporarily connect event to its pmu
1789 * such that validate_group() can classify
1790 * it as an x86 event using is_x86_event()
1795 if (event->group_leader != event)
1796 err = validate_group(event);
1798 err = validate_event(event);
1804 event->destroy(event);
1810 static int x86_pmu_event_idx(struct perf_event *event)
1812 int idx = event->hw.idx;
1814 if (!x86_pmu.attr_rdpmc)
1817 if (x86_pmu.num_counters_fixed && idx >= INTEL_PMC_IDX_FIXED) {
1818 idx -= INTEL_PMC_IDX_FIXED;
1825 static ssize_t get_attr_rdpmc(struct device *cdev,
1826 struct device_attribute *attr,
1829 return snprintf(buf, 40, "%d\n", x86_pmu.attr_rdpmc);
1832 static void change_rdpmc(void *info)
1834 bool enable = !!(unsigned long)info;
1837 set_in_cr4(X86_CR4_PCE);
1839 clear_in_cr4(X86_CR4_PCE);
1842 static ssize_t set_attr_rdpmc(struct device *cdev,
1843 struct device_attribute *attr,
1844 const char *buf, size_t count)
1849 ret = kstrtoul(buf, 0, &val);
1853 if (x86_pmu.attr_rdpmc_broken)
1856 if (!!val != !!x86_pmu.attr_rdpmc) {
1857 x86_pmu.attr_rdpmc = !!val;
1858 on_each_cpu(change_rdpmc, (void *)val, 1);
1864 static DEVICE_ATTR(rdpmc, S_IRUSR | S_IWUSR, get_attr_rdpmc, set_attr_rdpmc);
1866 static struct attribute *x86_pmu_attrs[] = {
1867 &dev_attr_rdpmc.attr,
1871 static struct attribute_group x86_pmu_attr_group = {
1872 .attrs = x86_pmu_attrs,
1875 static const struct attribute_group *x86_pmu_attr_groups[] = {
1876 &x86_pmu_attr_group,
1877 &x86_pmu_format_group,
1878 &x86_pmu_events_group,
1882 static void x86_pmu_flush_branch_stack(void)
1884 if (x86_pmu.flush_branch_stack)
1885 x86_pmu.flush_branch_stack();
1888 void perf_check_microcode(void)
1890 if (x86_pmu.check_microcode)
1891 x86_pmu.check_microcode();
1893 EXPORT_SYMBOL_GPL(perf_check_microcode);
1895 static struct pmu pmu = {
1896 .pmu_enable = x86_pmu_enable,
1897 .pmu_disable = x86_pmu_disable,
1899 .attr_groups = x86_pmu_attr_groups,
1901 .event_init = x86_pmu_event_init,
1905 .start = x86_pmu_start,
1906 .stop = x86_pmu_stop,
1907 .read = x86_pmu_read,
1909 .start_txn = x86_pmu_start_txn,
1910 .cancel_txn = x86_pmu_cancel_txn,
1911 .commit_txn = x86_pmu_commit_txn,
1913 .event_idx = x86_pmu_event_idx,
1914 .flush_branch_stack = x86_pmu_flush_branch_stack,
1917 void arch_perf_update_userpage(struct perf_event_mmap_page *userpg, u64 now)
1919 struct cyc2ns_data *data;
1921 userpg->cap_user_time = 0;
1922 userpg->cap_user_time_zero = 0;
1923 userpg->cap_user_rdpmc = x86_pmu.attr_rdpmc;
1924 userpg->pmc_width = x86_pmu.cntval_bits;
1926 if (!sched_clock_stable())
1929 data = cyc2ns_read_begin();
1931 userpg->cap_user_time = 1;
1932 userpg->time_mult = data->cyc2ns_mul;
1933 userpg->time_shift = data->cyc2ns_shift;
1934 userpg->time_offset = data->cyc2ns_offset - now;
1936 userpg->cap_user_time_zero = 1;
1937 userpg->time_zero = data->cyc2ns_offset;
1939 cyc2ns_read_end(data);
1946 static int backtrace_stack(void *data, char *name)
1951 static void backtrace_address(void *data, unsigned long addr, int reliable)
1953 struct perf_callchain_entry *entry = data;
1955 perf_callchain_store(entry, addr);
1958 static const struct stacktrace_ops backtrace_ops = {
1959 .stack = backtrace_stack,
1960 .address = backtrace_address,
1961 .walk_stack = print_context_stack_bp,
1965 perf_callchain_kernel(struct perf_callchain_entry *entry, struct pt_regs *regs)
1967 if (perf_guest_cbs && perf_guest_cbs->is_in_guest()) {
1968 /* TODO: We don't support guest os callchain now */
1972 perf_callchain_store(entry, regs->ip);
1974 dump_trace(NULL, regs, NULL, 0, &backtrace_ops, entry);
1978 valid_user_frame(const void __user *fp, unsigned long size)
1980 return (__range_not_ok(fp, size, TASK_SIZE) == 0);
1983 static unsigned long get_segment_base(unsigned int segment)
1985 struct desc_struct *desc;
1986 int idx = segment >> 3;
1988 if ((segment & SEGMENT_TI_MASK) == SEGMENT_LDT) {
1989 if (idx > LDT_ENTRIES)
1992 if (idx > current->active_mm->context.size)
1995 desc = current->active_mm->context.ldt;
1997 if (idx > GDT_ENTRIES)
2000 desc = raw_cpu_ptr(gdt_page.gdt);
2003 return get_desc_base(desc + idx);
2006 #ifdef CONFIG_COMPAT
2008 #include <asm/compat.h>
2011 perf_callchain_user32(struct pt_regs *regs, struct perf_callchain_entry *entry)
2013 /* 32-bit process in 64-bit kernel. */
2014 unsigned long ss_base, cs_base;
2015 struct stack_frame_ia32 frame;
2016 const void __user *fp;
2018 if (!test_thread_flag(TIF_IA32))
2021 cs_base = get_segment_base(regs->cs);
2022 ss_base = get_segment_base(regs->ss);
2024 fp = compat_ptr(ss_base + regs->bp);
2025 while (entry->nr < PERF_MAX_STACK_DEPTH) {
2026 unsigned long bytes;
2027 frame.next_frame = 0;
2028 frame.return_address = 0;
2030 bytes = copy_from_user_nmi(&frame, fp, sizeof(frame));
2034 if (!valid_user_frame(fp, sizeof(frame)))
2037 perf_callchain_store(entry, cs_base + frame.return_address);
2038 fp = compat_ptr(ss_base + frame.next_frame);
2044 perf_callchain_user32(struct pt_regs *regs, struct perf_callchain_entry *entry)
2051 perf_callchain_user(struct perf_callchain_entry *entry, struct pt_regs *regs)
2053 struct stack_frame frame;
2054 const void __user *fp;
2056 if (perf_guest_cbs && perf_guest_cbs->is_in_guest()) {
2057 /* TODO: We don't support guest os callchain now */
2062 * We don't know what to do with VM86 stacks.. ignore them for now.
2064 if (regs->flags & (X86_VM_MASK | PERF_EFLAGS_VM))
2067 fp = (void __user *)regs->bp;
2069 perf_callchain_store(entry, regs->ip);
2074 if (perf_callchain_user32(regs, entry))
2077 while (entry->nr < PERF_MAX_STACK_DEPTH) {
2078 unsigned long bytes;
2079 frame.next_frame = NULL;
2080 frame.return_address = 0;
2082 bytes = copy_from_user_nmi(&frame, fp, sizeof(frame));
2086 if (!valid_user_frame(fp, sizeof(frame)))
2089 perf_callchain_store(entry, frame.return_address);
2090 fp = frame.next_frame;
2095 * Deal with code segment offsets for the various execution modes:
2097 * VM86 - the good olde 16 bit days, where the linear address is
2098 * 20 bits and we use regs->ip + 0x10 * regs->cs.
2100 * IA32 - Where we need to look at GDT/LDT segment descriptor tables
2101 * to figure out what the 32bit base address is.
2103 * X32 - has TIF_X32 set, but is running in x86_64
2105 * X86_64 - CS,DS,SS,ES are all zero based.
2107 static unsigned long code_segment_base(struct pt_regs *regs)
2110 * If we are in VM86 mode, add the segment offset to convert to a
2113 if (regs->flags & X86_VM_MASK)
2114 return 0x10 * regs->cs;
2117 * For IA32 we look at the GDT/LDT segment base to convert the
2118 * effective IP to a linear address.
2120 #ifdef CONFIG_X86_32
2121 if (user_mode(regs) && regs->cs != __USER_CS)
2122 return get_segment_base(regs->cs);
2124 if (test_thread_flag(TIF_IA32)) {
2125 if (user_mode(regs) && regs->cs != __USER32_CS)
2126 return get_segment_base(regs->cs);
2132 unsigned long perf_instruction_pointer(struct pt_regs *regs)
2134 if (perf_guest_cbs && perf_guest_cbs->is_in_guest())
2135 return perf_guest_cbs->get_guest_ip();
2137 return regs->ip + code_segment_base(regs);
2140 unsigned long perf_misc_flags(struct pt_regs *regs)
2144 if (perf_guest_cbs && perf_guest_cbs->is_in_guest()) {
2145 if (perf_guest_cbs->is_user_mode())
2146 misc |= PERF_RECORD_MISC_GUEST_USER;
2148 misc |= PERF_RECORD_MISC_GUEST_KERNEL;
2150 if (user_mode(regs))
2151 misc |= PERF_RECORD_MISC_USER;
2153 misc |= PERF_RECORD_MISC_KERNEL;
2156 if (regs->flags & PERF_EFLAGS_EXACT)
2157 misc |= PERF_RECORD_MISC_EXACT_IP;
2162 void perf_get_x86_pmu_capability(struct x86_pmu_capability *cap)
2164 cap->version = x86_pmu.version;
2165 cap->num_counters_gp = x86_pmu.num_counters;
2166 cap->num_counters_fixed = x86_pmu.num_counters_fixed;
2167 cap->bit_width_gp = x86_pmu.cntval_bits;
2168 cap->bit_width_fixed = x86_pmu.cntval_bits;
2169 cap->events_mask = (unsigned int)x86_pmu.events_maskl;
2170 cap->events_mask_len = x86_pmu.events_mask_len;
2172 EXPORT_SYMBOL_GPL(perf_get_x86_pmu_capability);