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 if (event->attr.sample_period && x86_pmu.limit_period) {
449 if (x86_pmu.limit_period(event, event->attr.sample_period) >
450 event->attr.sample_period)
454 return x86_setup_perfctr(event);
458 * Setup the hardware configuration for a given attr_type
460 static int __x86_pmu_event_init(struct perf_event *event)
464 if (!x86_pmu_initialized())
468 if (!atomic_inc_not_zero(&active_events)) {
469 mutex_lock(&pmc_reserve_mutex);
470 if (atomic_read(&active_events) == 0) {
471 if (!reserve_pmc_hardware())
474 reserve_ds_buffers();
477 atomic_inc(&active_events);
478 mutex_unlock(&pmc_reserve_mutex);
483 event->destroy = hw_perf_event_destroy;
486 event->hw.last_cpu = -1;
487 event->hw.last_tag = ~0ULL;
490 event->hw.extra_reg.idx = EXTRA_REG_NONE;
491 event->hw.branch_reg.idx = EXTRA_REG_NONE;
493 return x86_pmu.hw_config(event);
496 void x86_pmu_disable_all(void)
498 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
501 for (idx = 0; idx < x86_pmu.num_counters; idx++) {
504 if (!test_bit(idx, cpuc->active_mask))
506 rdmsrl(x86_pmu_config_addr(idx), val);
507 if (!(val & ARCH_PERFMON_EVENTSEL_ENABLE))
509 val &= ~ARCH_PERFMON_EVENTSEL_ENABLE;
510 wrmsrl(x86_pmu_config_addr(idx), val);
514 static void x86_pmu_disable(struct pmu *pmu)
516 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
518 if (!x86_pmu_initialized())
528 x86_pmu.disable_all();
531 void x86_pmu_enable_all(int added)
533 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
536 for (idx = 0; idx < x86_pmu.num_counters; idx++) {
537 struct hw_perf_event *hwc = &cpuc->events[idx]->hw;
539 if (!test_bit(idx, cpuc->active_mask))
542 __x86_pmu_enable_event(hwc, ARCH_PERFMON_EVENTSEL_ENABLE);
546 static struct pmu pmu;
548 static inline int is_x86_event(struct perf_event *event)
550 return event->pmu == &pmu;
554 * Event scheduler state:
556 * Assign events iterating over all events and counters, beginning
557 * with events with least weights first. Keep the current iterator
558 * state in struct sched_state.
562 int event; /* event index */
563 int counter; /* counter index */
564 int unassigned; /* number of events to be assigned left */
565 unsigned long used[BITS_TO_LONGS(X86_PMC_IDX_MAX)];
568 /* Total max is X86_PMC_IDX_MAX, but we are O(n!) limited */
569 #define SCHED_STATES_MAX 2
574 struct perf_event **events;
575 struct sched_state state;
577 struct sched_state saved[SCHED_STATES_MAX];
581 * Initialize interator that runs through all events and counters.
583 static void perf_sched_init(struct perf_sched *sched, struct perf_event **events,
584 int num, int wmin, int wmax)
588 memset(sched, 0, sizeof(*sched));
589 sched->max_events = num;
590 sched->max_weight = wmax;
591 sched->events = events;
593 for (idx = 0; idx < num; idx++) {
594 if (events[idx]->hw.constraint->weight == wmin)
598 sched->state.event = idx; /* start with min weight */
599 sched->state.weight = wmin;
600 sched->state.unassigned = num;
603 static void perf_sched_save_state(struct perf_sched *sched)
605 if (WARN_ON_ONCE(sched->saved_states >= SCHED_STATES_MAX))
608 sched->saved[sched->saved_states] = sched->state;
609 sched->saved_states++;
612 static bool perf_sched_restore_state(struct perf_sched *sched)
614 if (!sched->saved_states)
617 sched->saved_states--;
618 sched->state = sched->saved[sched->saved_states];
620 /* continue with next counter: */
621 clear_bit(sched->state.counter++, sched->state.used);
627 * Select a counter for the current event to schedule. Return true on
630 static bool __perf_sched_find_counter(struct perf_sched *sched)
632 struct event_constraint *c;
635 if (!sched->state.unassigned)
638 if (sched->state.event >= sched->max_events)
641 c = sched->events[sched->state.event]->hw.constraint;
642 /* Prefer fixed purpose counters */
643 if (c->idxmsk64 & (~0ULL << INTEL_PMC_IDX_FIXED)) {
644 idx = INTEL_PMC_IDX_FIXED;
645 for_each_set_bit_from(idx, c->idxmsk, X86_PMC_IDX_MAX) {
646 if (!__test_and_set_bit(idx, sched->state.used))
650 /* Grab the first unused counter starting with idx */
651 idx = sched->state.counter;
652 for_each_set_bit_from(idx, c->idxmsk, INTEL_PMC_IDX_FIXED) {
653 if (!__test_and_set_bit(idx, sched->state.used))
660 sched->state.counter = idx;
663 perf_sched_save_state(sched);
668 static bool perf_sched_find_counter(struct perf_sched *sched)
670 while (!__perf_sched_find_counter(sched)) {
671 if (!perf_sched_restore_state(sched))
679 * Go through all unassigned events and find the next one to schedule.
680 * Take events with the least weight first. Return true on success.
682 static bool perf_sched_next_event(struct perf_sched *sched)
684 struct event_constraint *c;
686 if (!sched->state.unassigned || !--sched->state.unassigned)
691 sched->state.event++;
692 if (sched->state.event >= sched->max_events) {
694 sched->state.event = 0;
695 sched->state.weight++;
696 if (sched->state.weight > sched->max_weight)
699 c = sched->events[sched->state.event]->hw.constraint;
700 } while (c->weight != sched->state.weight);
702 sched->state.counter = 0; /* start with first counter */
708 * Assign a counter for each event.
710 int perf_assign_events(struct perf_event **events, int n,
711 int wmin, int wmax, int *assign)
713 struct perf_sched sched;
715 perf_sched_init(&sched, events, n, wmin, wmax);
718 if (!perf_sched_find_counter(&sched))
721 assign[sched.state.event] = sched.state.counter;
722 } while (perf_sched_next_event(&sched));
724 return sched.state.unassigned;
726 EXPORT_SYMBOL_GPL(perf_assign_events);
728 int x86_schedule_events(struct cpu_hw_events *cpuc, int n, int *assign)
730 struct event_constraint *c;
731 unsigned long used_mask[BITS_TO_LONGS(X86_PMC_IDX_MAX)];
732 struct perf_event *e;
733 int i, wmin, wmax, num = 0;
734 struct hw_perf_event *hwc;
736 bitmap_zero(used_mask, X86_PMC_IDX_MAX);
738 for (i = 0, wmin = X86_PMC_IDX_MAX, wmax = 0; i < n; i++) {
739 hwc = &cpuc->event_list[i]->hw;
740 c = x86_pmu.get_event_constraints(cpuc, cpuc->event_list[i]);
743 wmin = min(wmin, c->weight);
744 wmax = max(wmax, c->weight);
748 * fastpath, try to reuse previous register
750 for (i = 0; i < n; i++) {
751 hwc = &cpuc->event_list[i]->hw;
758 /* constraint still honored */
759 if (!test_bit(hwc->idx, c->idxmsk))
762 /* not already used */
763 if (test_bit(hwc->idx, used_mask))
766 __set_bit(hwc->idx, used_mask);
768 assign[i] = hwc->idx;
773 num = perf_assign_events(cpuc->event_list, n, wmin,
777 * Mark the event as committed, so we do not put_constraint()
778 * in case new events are added and fail scheduling.
780 if (!num && assign) {
781 for (i = 0; i < n; i++) {
782 e = cpuc->event_list[i];
783 e->hw.flags |= PERF_X86_EVENT_COMMITTED;
787 * scheduling failed or is just a simulation,
788 * free resources if necessary
790 if (!assign || num) {
791 for (i = 0; i < n; i++) {
792 e = cpuc->event_list[i];
794 * do not put_constraint() on comitted events,
795 * because they are good to go
797 if ((e->hw.flags & PERF_X86_EVENT_COMMITTED))
800 if (x86_pmu.put_event_constraints)
801 x86_pmu.put_event_constraints(cpuc, e);
804 return num ? -EINVAL : 0;
808 * dogrp: true if must collect siblings events (group)
809 * returns total number of events and error code
811 static int collect_events(struct cpu_hw_events *cpuc, struct perf_event *leader, bool dogrp)
813 struct perf_event *event;
816 max_count = x86_pmu.num_counters + x86_pmu.num_counters_fixed;
818 /* current number of events already accepted */
821 if (is_x86_event(leader)) {
824 cpuc->event_list[n] = leader;
830 list_for_each_entry(event, &leader->sibling_list, group_entry) {
831 if (!is_x86_event(event) ||
832 event->state <= PERF_EVENT_STATE_OFF)
838 cpuc->event_list[n] = event;
844 static inline void x86_assign_hw_event(struct perf_event *event,
845 struct cpu_hw_events *cpuc, int i)
847 struct hw_perf_event *hwc = &event->hw;
849 hwc->idx = cpuc->assign[i];
850 hwc->last_cpu = smp_processor_id();
851 hwc->last_tag = ++cpuc->tags[i];
853 if (hwc->idx == INTEL_PMC_IDX_FIXED_BTS) {
854 hwc->config_base = 0;
856 } else if (hwc->idx >= INTEL_PMC_IDX_FIXED) {
857 hwc->config_base = MSR_ARCH_PERFMON_FIXED_CTR_CTRL;
858 hwc->event_base = MSR_ARCH_PERFMON_FIXED_CTR0 + (hwc->idx - INTEL_PMC_IDX_FIXED);
859 hwc->event_base_rdpmc = (hwc->idx - INTEL_PMC_IDX_FIXED) | 1<<30;
861 hwc->config_base = x86_pmu_config_addr(hwc->idx);
862 hwc->event_base = x86_pmu_event_addr(hwc->idx);
863 hwc->event_base_rdpmc = x86_pmu_rdpmc_index(hwc->idx);
867 static inline int match_prev_assignment(struct hw_perf_event *hwc,
868 struct cpu_hw_events *cpuc,
871 return hwc->idx == cpuc->assign[i] &&
872 hwc->last_cpu == smp_processor_id() &&
873 hwc->last_tag == cpuc->tags[i];
876 static void x86_pmu_start(struct perf_event *event, int flags);
878 static void x86_pmu_enable(struct pmu *pmu)
880 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
881 struct perf_event *event;
882 struct hw_perf_event *hwc;
883 int i, added = cpuc->n_added;
885 if (!x86_pmu_initialized())
892 int n_running = cpuc->n_events - cpuc->n_added;
894 * apply assignment obtained either from
895 * hw_perf_group_sched_in() or x86_pmu_enable()
897 * step1: save events moving to new counters
899 for (i = 0; i < n_running; i++) {
900 event = cpuc->event_list[i];
904 * we can avoid reprogramming counter if:
905 * - assigned same counter as last time
906 * - running on same CPU as last time
907 * - no other event has used the counter since
909 if (hwc->idx == -1 ||
910 match_prev_assignment(hwc, cpuc, i))
914 * Ensure we don't accidentally enable a stopped
915 * counter simply because we rescheduled.
917 if (hwc->state & PERF_HES_STOPPED)
918 hwc->state |= PERF_HES_ARCH;
920 x86_pmu_stop(event, PERF_EF_UPDATE);
924 * step2: reprogram moved events into new counters
926 for (i = 0; i < cpuc->n_events; i++) {
927 event = cpuc->event_list[i];
930 if (!match_prev_assignment(hwc, cpuc, i))
931 x86_assign_hw_event(event, cpuc, i);
932 else if (i < n_running)
935 if (hwc->state & PERF_HES_ARCH)
938 x86_pmu_start(event, PERF_EF_RELOAD);
941 perf_events_lapic_init();
947 x86_pmu.enable_all(added);
950 static DEFINE_PER_CPU(u64 [X86_PMC_IDX_MAX], pmc_prev_left);
953 * Set the next IRQ period, based on the hwc->period_left value.
954 * To be called with the event disabled in hw:
956 int x86_perf_event_set_period(struct perf_event *event)
958 struct hw_perf_event *hwc = &event->hw;
959 s64 left = local64_read(&hwc->period_left);
960 s64 period = hwc->sample_period;
961 int ret = 0, idx = hwc->idx;
963 if (idx == INTEL_PMC_IDX_FIXED_BTS)
967 * If we are way outside a reasonable range then just skip forward:
969 if (unlikely(left <= -period)) {
971 local64_set(&hwc->period_left, left);
972 hwc->last_period = period;
976 if (unlikely(left <= 0)) {
978 local64_set(&hwc->period_left, left);
979 hwc->last_period = period;
983 * Quirk: certain CPUs dont like it if just 1 hw_event is left:
985 if (unlikely(left < 2))
988 if (left > x86_pmu.max_period)
989 left = x86_pmu.max_period;
991 if (x86_pmu.limit_period)
992 left = x86_pmu.limit_period(event, left);
994 per_cpu(pmc_prev_left[idx], smp_processor_id()) = left;
997 * The hw event starts counting from this event offset,
998 * mark it to be able to extra future deltas:
1000 local64_set(&hwc->prev_count, (u64)-left);
1002 wrmsrl(hwc->event_base, (u64)(-left) & x86_pmu.cntval_mask);
1005 * Due to erratum on certan cpu we need
1006 * a second write to be sure the register
1007 * is updated properly
1009 if (x86_pmu.perfctr_second_write) {
1010 wrmsrl(hwc->event_base,
1011 (u64)(-left) & x86_pmu.cntval_mask);
1014 perf_event_update_userpage(event);
1019 void x86_pmu_enable_event(struct perf_event *event)
1021 if (__this_cpu_read(cpu_hw_events.enabled))
1022 __x86_pmu_enable_event(&event->hw,
1023 ARCH_PERFMON_EVENTSEL_ENABLE);
1027 * Add a single event to the PMU.
1029 * The event is added to the group of enabled events
1030 * but only if it can be scehduled with existing events.
1032 static int x86_pmu_add(struct perf_event *event, int flags)
1034 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1035 struct hw_perf_event *hwc;
1036 int assign[X86_PMC_IDX_MAX];
1041 perf_pmu_disable(event->pmu);
1042 n0 = cpuc->n_events;
1043 ret = n = collect_events(cpuc, event, false);
1047 hwc->state = PERF_HES_UPTODATE | PERF_HES_STOPPED;
1048 if (!(flags & PERF_EF_START))
1049 hwc->state |= PERF_HES_ARCH;
1052 * If group events scheduling transaction was started,
1053 * skip the schedulability test here, it will be performed
1054 * at commit time (->commit_txn) as a whole.
1056 if (cpuc->group_flag & PERF_EVENT_TXN)
1059 ret = x86_pmu.schedule_events(cpuc, n, assign);
1063 * copy new assignment, now we know it is possible
1064 * will be used by hw_perf_enable()
1066 memcpy(cpuc->assign, assign, n*sizeof(int));
1070 * Commit the collect_events() state. See x86_pmu_del() and
1074 cpuc->n_added += n - n0;
1075 cpuc->n_txn += n - n0;
1079 perf_pmu_enable(event->pmu);
1083 static void x86_pmu_start(struct perf_event *event, int flags)
1085 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1086 int idx = event->hw.idx;
1088 if (WARN_ON_ONCE(!(event->hw.state & PERF_HES_STOPPED)))
1091 if (WARN_ON_ONCE(idx == -1))
1094 if (flags & PERF_EF_RELOAD) {
1095 WARN_ON_ONCE(!(event->hw.state & PERF_HES_UPTODATE));
1096 x86_perf_event_set_period(event);
1099 event->hw.state = 0;
1101 cpuc->events[idx] = event;
1102 __set_bit(idx, cpuc->active_mask);
1103 __set_bit(idx, cpuc->running);
1104 x86_pmu.enable(event);
1105 perf_event_update_userpage(event);
1108 void perf_event_print_debug(void)
1110 u64 ctrl, status, overflow, pmc_ctrl, pmc_count, prev_left, fixed;
1112 struct cpu_hw_events *cpuc;
1113 unsigned long flags;
1116 if (!x86_pmu.num_counters)
1119 local_irq_save(flags);
1121 cpu = smp_processor_id();
1122 cpuc = &per_cpu(cpu_hw_events, cpu);
1124 if (x86_pmu.version >= 2) {
1125 rdmsrl(MSR_CORE_PERF_GLOBAL_CTRL, ctrl);
1126 rdmsrl(MSR_CORE_PERF_GLOBAL_STATUS, status);
1127 rdmsrl(MSR_CORE_PERF_GLOBAL_OVF_CTRL, overflow);
1128 rdmsrl(MSR_ARCH_PERFMON_FIXED_CTR_CTRL, fixed);
1129 rdmsrl(MSR_IA32_PEBS_ENABLE, pebs);
1132 pr_info("CPU#%d: ctrl: %016llx\n", cpu, ctrl);
1133 pr_info("CPU#%d: status: %016llx\n", cpu, status);
1134 pr_info("CPU#%d: overflow: %016llx\n", cpu, overflow);
1135 pr_info("CPU#%d: fixed: %016llx\n", cpu, fixed);
1136 pr_info("CPU#%d: pebs: %016llx\n", cpu, pebs);
1138 pr_info("CPU#%d: active: %016llx\n", cpu, *(u64 *)cpuc->active_mask);
1140 for (idx = 0; idx < x86_pmu.num_counters; idx++) {
1141 rdmsrl(x86_pmu_config_addr(idx), pmc_ctrl);
1142 rdmsrl(x86_pmu_event_addr(idx), pmc_count);
1144 prev_left = per_cpu(pmc_prev_left[idx], cpu);
1146 pr_info("CPU#%d: gen-PMC%d ctrl: %016llx\n",
1147 cpu, idx, pmc_ctrl);
1148 pr_info("CPU#%d: gen-PMC%d count: %016llx\n",
1149 cpu, idx, pmc_count);
1150 pr_info("CPU#%d: gen-PMC%d left: %016llx\n",
1151 cpu, idx, prev_left);
1153 for (idx = 0; idx < x86_pmu.num_counters_fixed; idx++) {
1154 rdmsrl(MSR_ARCH_PERFMON_FIXED_CTR0 + idx, pmc_count);
1156 pr_info("CPU#%d: fixed-PMC%d count: %016llx\n",
1157 cpu, idx, pmc_count);
1159 local_irq_restore(flags);
1162 void x86_pmu_stop(struct perf_event *event, int flags)
1164 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1165 struct hw_perf_event *hwc = &event->hw;
1167 if (__test_and_clear_bit(hwc->idx, cpuc->active_mask)) {
1168 x86_pmu.disable(event);
1169 cpuc->events[hwc->idx] = NULL;
1170 WARN_ON_ONCE(hwc->state & PERF_HES_STOPPED);
1171 hwc->state |= PERF_HES_STOPPED;
1174 if ((flags & PERF_EF_UPDATE) && !(hwc->state & PERF_HES_UPTODATE)) {
1176 * Drain the remaining delta count out of a event
1177 * that we are disabling:
1179 x86_perf_event_update(event);
1180 hwc->state |= PERF_HES_UPTODATE;
1184 static void x86_pmu_del(struct perf_event *event, int flags)
1186 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1190 * event is descheduled
1192 event->hw.flags &= ~PERF_X86_EVENT_COMMITTED;
1195 * If we're called during a txn, we don't need to do anything.
1196 * The events never got scheduled and ->cancel_txn will truncate
1199 * XXX assumes any ->del() called during a TXN will only be on
1200 * an event added during that same TXN.
1202 if (cpuc->group_flag & PERF_EVENT_TXN)
1206 * Not a TXN, therefore cleanup properly.
1208 x86_pmu_stop(event, PERF_EF_UPDATE);
1210 for (i = 0; i < cpuc->n_events; i++) {
1211 if (event == cpuc->event_list[i])
1215 if (WARN_ON_ONCE(i == cpuc->n_events)) /* called ->del() without ->add() ? */
1218 /* If we have a newly added event; make sure to decrease n_added. */
1219 if (i >= cpuc->n_events - cpuc->n_added)
1222 if (x86_pmu.put_event_constraints)
1223 x86_pmu.put_event_constraints(cpuc, event);
1225 /* Delete the array entry. */
1226 while (++i < cpuc->n_events)
1227 cpuc->event_list[i-1] = cpuc->event_list[i];
1230 perf_event_update_userpage(event);
1233 int x86_pmu_handle_irq(struct pt_regs *regs)
1235 struct perf_sample_data data;
1236 struct cpu_hw_events *cpuc;
1237 struct perf_event *event;
1238 int idx, handled = 0;
1241 cpuc = this_cpu_ptr(&cpu_hw_events);
1244 * Some chipsets need to unmask the LVTPC in a particular spot
1245 * inside the nmi handler. As a result, the unmasking was pushed
1246 * into all the nmi handlers.
1248 * This generic handler doesn't seem to have any issues where the
1249 * unmasking occurs so it was left at the top.
1251 apic_write(APIC_LVTPC, APIC_DM_NMI);
1253 for (idx = 0; idx < x86_pmu.num_counters; idx++) {
1254 if (!test_bit(idx, cpuc->active_mask)) {
1256 * Though we deactivated the counter some cpus
1257 * might still deliver spurious interrupts still
1258 * in flight. Catch them:
1260 if (__test_and_clear_bit(idx, cpuc->running))
1265 event = cpuc->events[idx];
1267 val = x86_perf_event_update(event);
1268 if (val & (1ULL << (x86_pmu.cntval_bits - 1)))
1275 perf_sample_data_init(&data, 0, event->hw.last_period);
1277 if (!x86_perf_event_set_period(event))
1280 if (perf_event_overflow(event, &data, regs))
1281 x86_pmu_stop(event, 0);
1285 inc_irq_stat(apic_perf_irqs);
1290 void perf_events_lapic_init(void)
1292 if (!x86_pmu.apic || !x86_pmu_initialized())
1296 * Always use NMI for PMU
1298 apic_write(APIC_LVTPC, APIC_DM_NMI);
1302 perf_event_nmi_handler(unsigned int cmd, struct pt_regs *regs)
1308 if (!atomic_read(&active_events))
1311 start_clock = sched_clock();
1312 ret = x86_pmu.handle_irq(regs);
1313 finish_clock = sched_clock();
1315 perf_sample_event_took(finish_clock - start_clock);
1319 NOKPROBE_SYMBOL(perf_event_nmi_handler);
1321 struct event_constraint emptyconstraint;
1322 struct event_constraint unconstrained;
1325 x86_pmu_notifier(struct notifier_block *self, unsigned long action, void *hcpu)
1327 unsigned int cpu = (long)hcpu;
1328 struct cpu_hw_events *cpuc = &per_cpu(cpu_hw_events, cpu);
1329 int ret = NOTIFY_OK;
1331 switch (action & ~CPU_TASKS_FROZEN) {
1332 case CPU_UP_PREPARE:
1333 cpuc->kfree_on_online = NULL;
1334 if (x86_pmu.cpu_prepare)
1335 ret = x86_pmu.cpu_prepare(cpu);
1339 if (x86_pmu.attr_rdpmc)
1340 set_in_cr4(X86_CR4_PCE);
1341 if (x86_pmu.cpu_starting)
1342 x86_pmu.cpu_starting(cpu);
1346 kfree(cpuc->kfree_on_online);
1350 if (x86_pmu.cpu_dying)
1351 x86_pmu.cpu_dying(cpu);
1354 case CPU_UP_CANCELED:
1356 if (x86_pmu.cpu_dead)
1357 x86_pmu.cpu_dead(cpu);
1367 static void __init pmu_check_apic(void)
1373 pr_info("no APIC, boot with the \"lapic\" boot parameter to force-enable it.\n");
1374 pr_info("no hardware sampling interrupt available.\n");
1377 * If we have a PMU initialized but no APIC
1378 * interrupts, we cannot sample hardware
1379 * events (user-space has to fall back and
1380 * sample via a hrtimer based software event):
1382 pmu.capabilities |= PERF_PMU_CAP_NO_INTERRUPT;
1386 static struct attribute_group x86_pmu_format_group = {
1392 * Remove all undefined events (x86_pmu.event_map(id) == 0)
1393 * out of events_attr attributes.
1395 static void __init filter_events(struct attribute **attrs)
1397 struct device_attribute *d;
1398 struct perf_pmu_events_attr *pmu_attr;
1401 for (i = 0; attrs[i]; i++) {
1402 d = (struct device_attribute *)attrs[i];
1403 pmu_attr = container_of(d, struct perf_pmu_events_attr, attr);
1405 if (pmu_attr->event_str)
1407 if (x86_pmu.event_map(i))
1410 for (j = i; attrs[j]; j++)
1411 attrs[j] = attrs[j + 1];
1413 /* Check the shifted attr. */
1418 /* Merge two pointer arrays */
1419 static __init struct attribute **merge_attr(struct attribute **a, struct attribute **b)
1421 struct attribute **new;
1424 for (j = 0; a[j]; j++)
1426 for (i = 0; b[i]; i++)
1430 new = kmalloc(sizeof(struct attribute *) * j, GFP_KERNEL);
1435 for (i = 0; a[i]; i++)
1437 for (i = 0; b[i]; i++)
1444 ssize_t events_sysfs_show(struct device *dev, struct device_attribute *attr,
1447 struct perf_pmu_events_attr *pmu_attr = \
1448 container_of(attr, struct perf_pmu_events_attr, attr);
1449 u64 config = x86_pmu.event_map(pmu_attr->id);
1451 /* string trumps id */
1452 if (pmu_attr->event_str)
1453 return sprintf(page, "%s", pmu_attr->event_str);
1455 return x86_pmu.events_sysfs_show(page, config);
1458 EVENT_ATTR(cpu-cycles, CPU_CYCLES );
1459 EVENT_ATTR(instructions, INSTRUCTIONS );
1460 EVENT_ATTR(cache-references, CACHE_REFERENCES );
1461 EVENT_ATTR(cache-misses, CACHE_MISSES );
1462 EVENT_ATTR(branch-instructions, BRANCH_INSTRUCTIONS );
1463 EVENT_ATTR(branch-misses, BRANCH_MISSES );
1464 EVENT_ATTR(bus-cycles, BUS_CYCLES );
1465 EVENT_ATTR(stalled-cycles-frontend, STALLED_CYCLES_FRONTEND );
1466 EVENT_ATTR(stalled-cycles-backend, STALLED_CYCLES_BACKEND );
1467 EVENT_ATTR(ref-cycles, REF_CPU_CYCLES );
1469 static struct attribute *empty_attrs;
1471 static struct attribute *events_attr[] = {
1472 EVENT_PTR(CPU_CYCLES),
1473 EVENT_PTR(INSTRUCTIONS),
1474 EVENT_PTR(CACHE_REFERENCES),
1475 EVENT_PTR(CACHE_MISSES),
1476 EVENT_PTR(BRANCH_INSTRUCTIONS),
1477 EVENT_PTR(BRANCH_MISSES),
1478 EVENT_PTR(BUS_CYCLES),
1479 EVENT_PTR(STALLED_CYCLES_FRONTEND),
1480 EVENT_PTR(STALLED_CYCLES_BACKEND),
1481 EVENT_PTR(REF_CPU_CYCLES),
1485 static struct attribute_group x86_pmu_events_group = {
1487 .attrs = events_attr,
1490 ssize_t x86_event_sysfs_show(char *page, u64 config, u64 event)
1492 u64 umask = (config & ARCH_PERFMON_EVENTSEL_UMASK) >> 8;
1493 u64 cmask = (config & ARCH_PERFMON_EVENTSEL_CMASK) >> 24;
1494 bool edge = (config & ARCH_PERFMON_EVENTSEL_EDGE);
1495 bool pc = (config & ARCH_PERFMON_EVENTSEL_PIN_CONTROL);
1496 bool any = (config & ARCH_PERFMON_EVENTSEL_ANY);
1497 bool inv = (config & ARCH_PERFMON_EVENTSEL_INV);
1501 * We have whole page size to spend and just little data
1502 * to write, so we can safely use sprintf.
1504 ret = sprintf(page, "event=0x%02llx", event);
1507 ret += sprintf(page + ret, ",umask=0x%02llx", umask);
1510 ret += sprintf(page + ret, ",edge");
1513 ret += sprintf(page + ret, ",pc");
1516 ret += sprintf(page + ret, ",any");
1519 ret += sprintf(page + ret, ",inv");
1522 ret += sprintf(page + ret, ",cmask=0x%02llx", cmask);
1524 ret += sprintf(page + ret, "\n");
1529 static int __init init_hw_perf_events(void)
1531 struct x86_pmu_quirk *quirk;
1534 pr_info("Performance Events: ");
1536 switch (boot_cpu_data.x86_vendor) {
1537 case X86_VENDOR_INTEL:
1538 err = intel_pmu_init();
1540 case X86_VENDOR_AMD:
1541 err = amd_pmu_init();
1547 pr_cont("no PMU driver, software events only.\n");
1553 /* sanity check that the hardware exists or is emulated */
1554 if (!check_hw_exists())
1557 pr_cont("%s PMU driver.\n", x86_pmu.name);
1559 x86_pmu.attr_rdpmc = 1; /* enable userspace RDPMC usage by default */
1561 for (quirk = x86_pmu.quirks; quirk; quirk = quirk->next)
1564 if (!x86_pmu.intel_ctrl)
1565 x86_pmu.intel_ctrl = (1 << x86_pmu.num_counters) - 1;
1567 perf_events_lapic_init();
1568 register_nmi_handler(NMI_LOCAL, perf_event_nmi_handler, 0, "PMI");
1570 unconstrained = (struct event_constraint)
1571 __EVENT_CONSTRAINT(0, (1ULL << x86_pmu.num_counters) - 1,
1572 0, x86_pmu.num_counters, 0, 0);
1574 x86_pmu_format_group.attrs = x86_pmu.format_attrs;
1576 if (x86_pmu.event_attrs)
1577 x86_pmu_events_group.attrs = x86_pmu.event_attrs;
1579 if (!x86_pmu.events_sysfs_show)
1580 x86_pmu_events_group.attrs = &empty_attrs;
1582 filter_events(x86_pmu_events_group.attrs);
1584 if (x86_pmu.cpu_events) {
1585 struct attribute **tmp;
1587 tmp = merge_attr(x86_pmu_events_group.attrs, x86_pmu.cpu_events);
1589 x86_pmu_events_group.attrs = tmp;
1592 pr_info("... version: %d\n", x86_pmu.version);
1593 pr_info("... bit width: %d\n", x86_pmu.cntval_bits);
1594 pr_info("... generic registers: %d\n", x86_pmu.num_counters);
1595 pr_info("... value mask: %016Lx\n", x86_pmu.cntval_mask);
1596 pr_info("... max period: %016Lx\n", x86_pmu.max_period);
1597 pr_info("... fixed-purpose events: %d\n", x86_pmu.num_counters_fixed);
1598 pr_info("... event mask: %016Lx\n", x86_pmu.intel_ctrl);
1600 perf_pmu_register(&pmu, "cpu", PERF_TYPE_RAW);
1601 perf_cpu_notifier(x86_pmu_notifier);
1605 early_initcall(init_hw_perf_events);
1607 static inline void x86_pmu_read(struct perf_event *event)
1609 x86_perf_event_update(event);
1613 * Start group events scheduling transaction
1614 * Set the flag to make pmu::enable() not perform the
1615 * schedulability test, it will be performed at commit time
1617 static void x86_pmu_start_txn(struct pmu *pmu)
1619 perf_pmu_disable(pmu);
1620 __this_cpu_or(cpu_hw_events.group_flag, PERF_EVENT_TXN);
1621 __this_cpu_write(cpu_hw_events.n_txn, 0);
1625 * Stop group events scheduling transaction
1626 * Clear the flag and pmu::enable() will perform the
1627 * schedulability test.
1629 static void x86_pmu_cancel_txn(struct pmu *pmu)
1631 __this_cpu_and(cpu_hw_events.group_flag, ~PERF_EVENT_TXN);
1633 * Truncate collected array by the number of events added in this
1634 * transaction. See x86_pmu_add() and x86_pmu_*_txn().
1636 __this_cpu_sub(cpu_hw_events.n_added, __this_cpu_read(cpu_hw_events.n_txn));
1637 __this_cpu_sub(cpu_hw_events.n_events, __this_cpu_read(cpu_hw_events.n_txn));
1638 perf_pmu_enable(pmu);
1642 * Commit group events scheduling transaction
1643 * Perform the group schedulability test as a whole
1644 * Return 0 if success
1646 * Does not cancel the transaction on failure; expects the caller to do this.
1648 static int x86_pmu_commit_txn(struct pmu *pmu)
1650 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1651 int assign[X86_PMC_IDX_MAX];
1656 if (!x86_pmu_initialized())
1659 ret = x86_pmu.schedule_events(cpuc, n, assign);
1664 * copy new assignment, now we know it is possible
1665 * will be used by hw_perf_enable()
1667 memcpy(cpuc->assign, assign, n*sizeof(int));
1669 cpuc->group_flag &= ~PERF_EVENT_TXN;
1670 perf_pmu_enable(pmu);
1674 * a fake_cpuc is used to validate event groups. Due to
1675 * the extra reg logic, we need to also allocate a fake
1676 * per_core and per_cpu structure. Otherwise, group events
1677 * using extra reg may conflict without the kernel being
1678 * able to catch this when the last event gets added to
1681 static void free_fake_cpuc(struct cpu_hw_events *cpuc)
1683 kfree(cpuc->shared_regs);
1687 static struct cpu_hw_events *allocate_fake_cpuc(void)
1689 struct cpu_hw_events *cpuc;
1690 int cpu = raw_smp_processor_id();
1692 cpuc = kzalloc(sizeof(*cpuc), GFP_KERNEL);
1694 return ERR_PTR(-ENOMEM);
1696 /* only needed, if we have extra_regs */
1697 if (x86_pmu.extra_regs) {
1698 cpuc->shared_regs = allocate_shared_regs(cpu);
1699 if (!cpuc->shared_regs)
1705 free_fake_cpuc(cpuc);
1706 return ERR_PTR(-ENOMEM);
1710 * validate that we can schedule this event
1712 static int validate_event(struct perf_event *event)
1714 struct cpu_hw_events *fake_cpuc;
1715 struct event_constraint *c;
1718 fake_cpuc = allocate_fake_cpuc();
1719 if (IS_ERR(fake_cpuc))
1720 return PTR_ERR(fake_cpuc);
1722 c = x86_pmu.get_event_constraints(fake_cpuc, event);
1724 if (!c || !c->weight)
1727 if (x86_pmu.put_event_constraints)
1728 x86_pmu.put_event_constraints(fake_cpuc, event);
1730 free_fake_cpuc(fake_cpuc);
1736 * validate a single event group
1738 * validation include:
1739 * - check events are compatible which each other
1740 * - events do not compete for the same counter
1741 * - number of events <= number of counters
1743 * validation ensures the group can be loaded onto the
1744 * PMU if it was the only group available.
1746 static int validate_group(struct perf_event *event)
1748 struct perf_event *leader = event->group_leader;
1749 struct cpu_hw_events *fake_cpuc;
1750 int ret = -EINVAL, n;
1752 fake_cpuc = allocate_fake_cpuc();
1753 if (IS_ERR(fake_cpuc))
1754 return PTR_ERR(fake_cpuc);
1756 * the event is not yet connected with its
1757 * siblings therefore we must first collect
1758 * existing siblings, then add the new event
1759 * before we can simulate the scheduling
1761 n = collect_events(fake_cpuc, leader, true);
1765 fake_cpuc->n_events = n;
1766 n = collect_events(fake_cpuc, event, false);
1770 fake_cpuc->n_events = n;
1772 ret = x86_pmu.schedule_events(fake_cpuc, n, NULL);
1775 free_fake_cpuc(fake_cpuc);
1779 static int x86_pmu_event_init(struct perf_event *event)
1784 switch (event->attr.type) {
1786 case PERF_TYPE_HARDWARE:
1787 case PERF_TYPE_HW_CACHE:
1794 err = __x86_pmu_event_init(event);
1797 * we temporarily connect event to its pmu
1798 * such that validate_group() can classify
1799 * it as an x86 event using is_x86_event()
1804 if (event->group_leader != event)
1805 err = validate_group(event);
1807 err = validate_event(event);
1813 event->destroy(event);
1819 static int x86_pmu_event_idx(struct perf_event *event)
1821 int idx = event->hw.idx;
1823 if (!x86_pmu.attr_rdpmc)
1826 if (x86_pmu.num_counters_fixed && idx >= INTEL_PMC_IDX_FIXED) {
1827 idx -= INTEL_PMC_IDX_FIXED;
1834 static ssize_t get_attr_rdpmc(struct device *cdev,
1835 struct device_attribute *attr,
1838 return snprintf(buf, 40, "%d\n", x86_pmu.attr_rdpmc);
1841 static void change_rdpmc(void *info)
1843 bool enable = !!(unsigned long)info;
1846 set_in_cr4(X86_CR4_PCE);
1848 clear_in_cr4(X86_CR4_PCE);
1851 static ssize_t set_attr_rdpmc(struct device *cdev,
1852 struct device_attribute *attr,
1853 const char *buf, size_t count)
1858 ret = kstrtoul(buf, 0, &val);
1862 if (x86_pmu.attr_rdpmc_broken)
1865 if (!!val != !!x86_pmu.attr_rdpmc) {
1866 x86_pmu.attr_rdpmc = !!val;
1867 on_each_cpu(change_rdpmc, (void *)val, 1);
1873 static DEVICE_ATTR(rdpmc, S_IRUSR | S_IWUSR, get_attr_rdpmc, set_attr_rdpmc);
1875 static struct attribute *x86_pmu_attrs[] = {
1876 &dev_attr_rdpmc.attr,
1880 static struct attribute_group x86_pmu_attr_group = {
1881 .attrs = x86_pmu_attrs,
1884 static const struct attribute_group *x86_pmu_attr_groups[] = {
1885 &x86_pmu_attr_group,
1886 &x86_pmu_format_group,
1887 &x86_pmu_events_group,
1891 static void x86_pmu_flush_branch_stack(void)
1893 if (x86_pmu.flush_branch_stack)
1894 x86_pmu.flush_branch_stack();
1897 void perf_check_microcode(void)
1899 if (x86_pmu.check_microcode)
1900 x86_pmu.check_microcode();
1902 EXPORT_SYMBOL_GPL(perf_check_microcode);
1904 static struct pmu pmu = {
1905 .pmu_enable = x86_pmu_enable,
1906 .pmu_disable = x86_pmu_disable,
1908 .attr_groups = x86_pmu_attr_groups,
1910 .event_init = x86_pmu_event_init,
1914 .start = x86_pmu_start,
1915 .stop = x86_pmu_stop,
1916 .read = x86_pmu_read,
1918 .start_txn = x86_pmu_start_txn,
1919 .cancel_txn = x86_pmu_cancel_txn,
1920 .commit_txn = x86_pmu_commit_txn,
1922 .event_idx = x86_pmu_event_idx,
1923 .flush_branch_stack = x86_pmu_flush_branch_stack,
1926 void arch_perf_update_userpage(struct perf_event_mmap_page *userpg, u64 now)
1928 struct cyc2ns_data *data;
1930 userpg->cap_user_time = 0;
1931 userpg->cap_user_time_zero = 0;
1932 userpg->cap_user_rdpmc = x86_pmu.attr_rdpmc;
1933 userpg->pmc_width = x86_pmu.cntval_bits;
1935 if (!sched_clock_stable())
1938 data = cyc2ns_read_begin();
1940 userpg->cap_user_time = 1;
1941 userpg->time_mult = data->cyc2ns_mul;
1942 userpg->time_shift = data->cyc2ns_shift;
1943 userpg->time_offset = data->cyc2ns_offset - now;
1945 userpg->cap_user_time_zero = 1;
1946 userpg->time_zero = data->cyc2ns_offset;
1948 cyc2ns_read_end(data);
1955 static int backtrace_stack(void *data, char *name)
1960 static void backtrace_address(void *data, unsigned long addr, int reliable)
1962 struct perf_callchain_entry *entry = data;
1964 perf_callchain_store(entry, addr);
1967 static const struct stacktrace_ops backtrace_ops = {
1968 .stack = backtrace_stack,
1969 .address = backtrace_address,
1970 .walk_stack = print_context_stack_bp,
1974 perf_callchain_kernel(struct perf_callchain_entry *entry, struct pt_regs *regs)
1976 if (perf_guest_cbs && perf_guest_cbs->is_in_guest()) {
1977 /* TODO: We don't support guest os callchain now */
1981 perf_callchain_store(entry, regs->ip);
1983 dump_trace(NULL, regs, NULL, 0, &backtrace_ops, entry);
1987 valid_user_frame(const void __user *fp, unsigned long size)
1989 return (__range_not_ok(fp, size, TASK_SIZE) == 0);
1992 static unsigned long get_segment_base(unsigned int segment)
1994 struct desc_struct *desc;
1995 int idx = segment >> 3;
1997 if ((segment & SEGMENT_TI_MASK) == SEGMENT_LDT) {
1998 if (idx > LDT_ENTRIES)
2001 if (idx > current->active_mm->context.size)
2004 desc = current->active_mm->context.ldt;
2006 if (idx > GDT_ENTRIES)
2009 desc = raw_cpu_ptr(gdt_page.gdt);
2012 return get_desc_base(desc + idx);
2015 #ifdef CONFIG_COMPAT
2017 #include <asm/compat.h>
2020 perf_callchain_user32(struct pt_regs *regs, struct perf_callchain_entry *entry)
2022 /* 32-bit process in 64-bit kernel. */
2023 unsigned long ss_base, cs_base;
2024 struct stack_frame_ia32 frame;
2025 const void __user *fp;
2027 if (!test_thread_flag(TIF_IA32))
2030 cs_base = get_segment_base(regs->cs);
2031 ss_base = get_segment_base(regs->ss);
2033 fp = compat_ptr(ss_base + regs->bp);
2034 while (entry->nr < PERF_MAX_STACK_DEPTH) {
2035 unsigned long bytes;
2036 frame.next_frame = 0;
2037 frame.return_address = 0;
2039 bytes = copy_from_user_nmi(&frame, fp, sizeof(frame));
2043 if (!valid_user_frame(fp, sizeof(frame)))
2046 perf_callchain_store(entry, cs_base + frame.return_address);
2047 fp = compat_ptr(ss_base + frame.next_frame);
2053 perf_callchain_user32(struct pt_regs *regs, struct perf_callchain_entry *entry)
2060 perf_callchain_user(struct perf_callchain_entry *entry, struct pt_regs *regs)
2062 struct stack_frame frame;
2063 const void __user *fp;
2065 if (perf_guest_cbs && perf_guest_cbs->is_in_guest()) {
2066 /* TODO: We don't support guest os callchain now */
2071 * We don't know what to do with VM86 stacks.. ignore them for now.
2073 if (regs->flags & (X86_VM_MASK | PERF_EFLAGS_VM))
2076 fp = (void __user *)regs->bp;
2078 perf_callchain_store(entry, regs->ip);
2083 if (perf_callchain_user32(regs, entry))
2086 while (entry->nr < PERF_MAX_STACK_DEPTH) {
2087 unsigned long bytes;
2088 frame.next_frame = NULL;
2089 frame.return_address = 0;
2091 bytes = copy_from_user_nmi(&frame, fp, sizeof(frame));
2095 if (!valid_user_frame(fp, sizeof(frame)))
2098 perf_callchain_store(entry, frame.return_address);
2099 fp = frame.next_frame;
2104 * Deal with code segment offsets for the various execution modes:
2106 * VM86 - the good olde 16 bit days, where the linear address is
2107 * 20 bits and we use regs->ip + 0x10 * regs->cs.
2109 * IA32 - Where we need to look at GDT/LDT segment descriptor tables
2110 * to figure out what the 32bit base address is.
2112 * X32 - has TIF_X32 set, but is running in x86_64
2114 * X86_64 - CS,DS,SS,ES are all zero based.
2116 static unsigned long code_segment_base(struct pt_regs *regs)
2119 * If we are in VM86 mode, add the segment offset to convert to a
2122 if (regs->flags & X86_VM_MASK)
2123 return 0x10 * regs->cs;
2126 * For IA32 we look at the GDT/LDT segment base to convert the
2127 * effective IP to a linear address.
2129 #ifdef CONFIG_X86_32
2130 if (user_mode(regs) && regs->cs != __USER_CS)
2131 return get_segment_base(regs->cs);
2133 if (test_thread_flag(TIF_IA32)) {
2134 if (user_mode(regs) && regs->cs != __USER32_CS)
2135 return get_segment_base(regs->cs);
2141 unsigned long perf_instruction_pointer(struct pt_regs *regs)
2143 if (perf_guest_cbs && perf_guest_cbs->is_in_guest())
2144 return perf_guest_cbs->get_guest_ip();
2146 return regs->ip + code_segment_base(regs);
2149 unsigned long perf_misc_flags(struct pt_regs *regs)
2153 if (perf_guest_cbs && perf_guest_cbs->is_in_guest()) {
2154 if (perf_guest_cbs->is_user_mode())
2155 misc |= PERF_RECORD_MISC_GUEST_USER;
2157 misc |= PERF_RECORD_MISC_GUEST_KERNEL;
2159 if (user_mode(regs))
2160 misc |= PERF_RECORD_MISC_USER;
2162 misc |= PERF_RECORD_MISC_KERNEL;
2165 if (regs->flags & PERF_EFLAGS_EXACT)
2166 misc |= PERF_RECORD_MISC_EXACT_IP;
2171 void perf_get_x86_pmu_capability(struct x86_pmu_capability *cap)
2173 cap->version = x86_pmu.version;
2174 cap->num_counters_gp = x86_pmu.num_counters;
2175 cap->num_counters_fixed = x86_pmu.num_counters_fixed;
2176 cap->bit_width_gp = x86_pmu.cntval_bits;
2177 cap->bit_width_fixed = x86_pmu.cntval_bits;
2178 cap->events_mask = (unsigned int)x86_pmu.events_maskl;
2179 cap->events_mask_len = x86_pmu.events_mask_len;
2181 EXPORT_SYMBOL_GPL(perf_get_x86_pmu_capability);