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>
36 #include "perf_event.h"
38 struct x86_pmu x86_pmu __read_mostly;
40 DEFINE_PER_CPU(struct cpu_hw_events, cpu_hw_events) = {
44 u64 __read_mostly hw_cache_event_ids
45 [PERF_COUNT_HW_CACHE_MAX]
46 [PERF_COUNT_HW_CACHE_OP_MAX]
47 [PERF_COUNT_HW_CACHE_RESULT_MAX];
48 u64 __read_mostly hw_cache_extra_regs
49 [PERF_COUNT_HW_CACHE_MAX]
50 [PERF_COUNT_HW_CACHE_OP_MAX]
51 [PERF_COUNT_HW_CACHE_RESULT_MAX];
54 * Propagate event elapsed time into the generic event.
55 * Can only be executed on the CPU where the event is active.
56 * Returns the delta events processed.
58 u64 x86_perf_event_update(struct perf_event *event)
60 struct hw_perf_event *hwc = &event->hw;
61 int shift = 64 - x86_pmu.cntval_bits;
62 u64 prev_raw_count, new_raw_count;
66 if (idx == INTEL_PMC_IDX_FIXED_BTS)
70 * Careful: an NMI might modify the previous event value.
72 * Our tactic to handle this is to first atomically read and
73 * exchange a new raw count - then add that new-prev delta
74 * count to the generic event atomically:
77 prev_raw_count = local64_read(&hwc->prev_count);
78 rdpmcl(hwc->event_base_rdpmc, new_raw_count);
80 if (local64_cmpxchg(&hwc->prev_count, prev_raw_count,
81 new_raw_count) != prev_raw_count)
85 * Now we have the new raw value and have updated the prev
86 * timestamp already. We can now calculate the elapsed delta
87 * (event-)time and add that to the generic event.
89 * Careful, not all hw sign-extends above the physical width
92 delta = (new_raw_count << shift) - (prev_raw_count << shift);
95 local64_add(delta, &event->count);
96 local64_sub(delta, &hwc->period_left);
102 * Find and validate any extra registers to set up.
104 static int x86_pmu_extra_regs(u64 config, struct perf_event *event)
106 struct hw_perf_event_extra *reg;
107 struct extra_reg *er;
109 reg = &event->hw.extra_reg;
111 if (!x86_pmu.extra_regs)
114 for (er = x86_pmu.extra_regs; er->msr; er++) {
115 if (er->event != (config & er->config_mask))
117 if (event->attr.config1 & ~er->valid_mask)
121 reg->config = event->attr.config1;
128 static atomic_t active_events;
129 static DEFINE_MUTEX(pmc_reserve_mutex);
131 #ifdef CONFIG_X86_LOCAL_APIC
133 static bool reserve_pmc_hardware(void)
137 for (i = 0; i < x86_pmu.num_counters; i++) {
138 if (!reserve_perfctr_nmi(x86_pmu_event_addr(i)))
142 for (i = 0; i < x86_pmu.num_counters; i++) {
143 if (!reserve_evntsel_nmi(x86_pmu_config_addr(i)))
150 for (i--; i >= 0; i--)
151 release_evntsel_nmi(x86_pmu_config_addr(i));
153 i = x86_pmu.num_counters;
156 for (i--; i >= 0; i--)
157 release_perfctr_nmi(x86_pmu_event_addr(i));
162 static void release_pmc_hardware(void)
166 for (i = 0; i < x86_pmu.num_counters; i++) {
167 release_perfctr_nmi(x86_pmu_event_addr(i));
168 release_evntsel_nmi(x86_pmu_config_addr(i));
174 static bool reserve_pmc_hardware(void) { return true; }
175 static void release_pmc_hardware(void) {}
179 static bool check_hw_exists(void)
181 u64 val, val_new = ~0;
185 * Check to see if the BIOS enabled any of the counters, if so
188 for (i = 0; i < x86_pmu.num_counters; i++) {
189 reg = x86_pmu_config_addr(i);
190 ret = rdmsrl_safe(reg, &val);
193 if (val & ARCH_PERFMON_EVENTSEL_ENABLE)
197 if (x86_pmu.num_counters_fixed) {
198 reg = MSR_ARCH_PERFMON_FIXED_CTR_CTRL;
199 ret = rdmsrl_safe(reg, &val);
202 for (i = 0; i < x86_pmu.num_counters_fixed; i++) {
203 if (val & (0x03 << i*4))
209 * Now write a value and read it back to see if it matches,
210 * this is needed to detect certain hardware emulators (qemu/kvm)
211 * that don't trap on the MSR access and always return 0s.
214 reg = x86_pmu_event_addr(0);
215 ret = wrmsrl_safe(reg, val);
216 ret |= rdmsrl_safe(reg, &val_new);
217 if (ret || val != val_new)
224 * We still allow the PMU driver to operate:
226 printk(KERN_CONT "Broken BIOS detected, complain to your hardware vendor.\n");
227 printk(KERN_ERR FW_BUG "the BIOS has corrupted hw-PMU resources (MSR %x is %Lx)\n", reg, val);
232 printk(KERN_CONT "Broken PMU hardware detected, using software events only.\n");
233 printk(KERN_ERR "Failed to access perfctr msr (MSR %x is %Lx)\n", reg, val_new);
238 static void hw_perf_event_destroy(struct perf_event *event)
240 if (atomic_dec_and_mutex_lock(&active_events, &pmc_reserve_mutex)) {
241 release_pmc_hardware();
242 release_ds_buffers();
243 mutex_unlock(&pmc_reserve_mutex);
247 static inline int x86_pmu_initialized(void)
249 return x86_pmu.handle_irq != NULL;
253 set_ext_hw_attr(struct hw_perf_event *hwc, struct perf_event *event)
255 struct perf_event_attr *attr = &event->attr;
256 unsigned int cache_type, cache_op, cache_result;
259 config = attr->config;
261 cache_type = (config >> 0) & 0xff;
262 if (cache_type >= PERF_COUNT_HW_CACHE_MAX)
265 cache_op = (config >> 8) & 0xff;
266 if (cache_op >= PERF_COUNT_HW_CACHE_OP_MAX)
269 cache_result = (config >> 16) & 0xff;
270 if (cache_result >= PERF_COUNT_HW_CACHE_RESULT_MAX)
273 val = hw_cache_event_ids[cache_type][cache_op][cache_result];
282 attr->config1 = hw_cache_extra_regs[cache_type][cache_op][cache_result];
283 return x86_pmu_extra_regs(val, event);
286 int x86_setup_perfctr(struct perf_event *event)
288 struct perf_event_attr *attr = &event->attr;
289 struct hw_perf_event *hwc = &event->hw;
292 if (!is_sampling_event(event)) {
293 hwc->sample_period = x86_pmu.max_period;
294 hwc->last_period = hwc->sample_period;
295 local64_set(&hwc->period_left, hwc->sample_period);
298 * If we have a PMU initialized but no APIC
299 * interrupts, we cannot sample hardware
300 * events (user-space has to fall back and
301 * sample via a hrtimer based software event):
307 if (attr->type == PERF_TYPE_RAW)
308 return x86_pmu_extra_regs(event->attr.config, event);
310 if (attr->type == PERF_TYPE_HW_CACHE)
311 return set_ext_hw_attr(hwc, event);
313 if (attr->config >= x86_pmu.max_events)
319 config = x86_pmu.event_map(attr->config);
330 if (attr->config == PERF_COUNT_HW_BRANCH_INSTRUCTIONS &&
331 !attr->freq && hwc->sample_period == 1) {
332 /* BTS is not supported by this architecture. */
333 if (!x86_pmu.bts_active)
336 /* BTS is currently only allowed for user-mode. */
337 if (!attr->exclude_kernel)
341 hwc->config |= config;
347 * check that branch_sample_type is compatible with
348 * settings needed for precise_ip > 1 which implies
349 * using the LBR to capture ALL taken branches at the
350 * priv levels of the measurement
352 static inline int precise_br_compat(struct perf_event *event)
354 u64 m = event->attr.branch_sample_type;
357 /* must capture all branches */
358 if (!(m & PERF_SAMPLE_BRANCH_ANY))
361 m &= PERF_SAMPLE_BRANCH_KERNEL | PERF_SAMPLE_BRANCH_USER;
363 if (!event->attr.exclude_user)
364 b |= PERF_SAMPLE_BRANCH_USER;
366 if (!event->attr.exclude_kernel)
367 b |= PERF_SAMPLE_BRANCH_KERNEL;
370 * ignore PERF_SAMPLE_BRANCH_HV, not supported on x86
376 int x86_pmu_hw_config(struct perf_event *event)
378 if (event->attr.precise_ip) {
381 /* Support for constant skid */
382 if (x86_pmu.pebs_active && !x86_pmu.pebs_broken) {
385 /* Support for IP fixup */
390 if (event->attr.precise_ip > precise)
393 * check that PEBS LBR correction does not conflict with
394 * whatever the user is asking with attr->branch_sample_type
396 if (event->attr.precise_ip > 1) {
397 u64 *br_type = &event->attr.branch_sample_type;
399 if (has_branch_stack(event)) {
400 if (!precise_br_compat(event))
403 /* branch_sample_type is compatible */
407 * user did not specify branch_sample_type
409 * For PEBS fixups, we capture all
410 * the branches at the priv level of the
413 *br_type = PERF_SAMPLE_BRANCH_ANY;
415 if (!event->attr.exclude_user)
416 *br_type |= PERF_SAMPLE_BRANCH_USER;
418 if (!event->attr.exclude_kernel)
419 *br_type |= PERF_SAMPLE_BRANCH_KERNEL;
426 * (keep 'enabled' bit clear for now)
428 event->hw.config = ARCH_PERFMON_EVENTSEL_INT;
431 * Count user and OS events unless requested not to
433 if (!event->attr.exclude_user)
434 event->hw.config |= ARCH_PERFMON_EVENTSEL_USR;
435 if (!event->attr.exclude_kernel)
436 event->hw.config |= ARCH_PERFMON_EVENTSEL_OS;
438 if (event->attr.type == PERF_TYPE_RAW)
439 event->hw.config |= event->attr.config & X86_RAW_EVENT_MASK;
441 return x86_setup_perfctr(event);
445 * Setup the hardware configuration for a given attr_type
447 static int __x86_pmu_event_init(struct perf_event *event)
451 if (!x86_pmu_initialized())
455 if (!atomic_inc_not_zero(&active_events)) {
456 mutex_lock(&pmc_reserve_mutex);
457 if (atomic_read(&active_events) == 0) {
458 if (!reserve_pmc_hardware())
461 reserve_ds_buffers();
464 atomic_inc(&active_events);
465 mutex_unlock(&pmc_reserve_mutex);
470 event->destroy = hw_perf_event_destroy;
473 event->hw.last_cpu = -1;
474 event->hw.last_tag = ~0ULL;
477 event->hw.extra_reg.idx = EXTRA_REG_NONE;
478 event->hw.branch_reg.idx = EXTRA_REG_NONE;
480 return x86_pmu.hw_config(event);
483 void x86_pmu_disable_all(void)
485 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
488 for (idx = 0; idx < x86_pmu.num_counters; idx++) {
491 if (!test_bit(idx, cpuc->active_mask))
493 rdmsrl(x86_pmu_config_addr(idx), val);
494 if (!(val & ARCH_PERFMON_EVENTSEL_ENABLE))
496 val &= ~ARCH_PERFMON_EVENTSEL_ENABLE;
497 wrmsrl(x86_pmu_config_addr(idx), val);
501 static void x86_pmu_disable(struct pmu *pmu)
503 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
505 if (!x86_pmu_initialized())
515 x86_pmu.disable_all();
518 void x86_pmu_enable_all(int added)
520 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
523 for (idx = 0; idx < x86_pmu.num_counters; idx++) {
524 struct hw_perf_event *hwc = &cpuc->events[idx]->hw;
526 if (!test_bit(idx, cpuc->active_mask))
529 __x86_pmu_enable_event(hwc, ARCH_PERFMON_EVENTSEL_ENABLE);
533 static struct pmu pmu;
535 static inline int is_x86_event(struct perf_event *event)
537 return event->pmu == &pmu;
541 * Event scheduler state:
543 * Assign events iterating over all events and counters, beginning
544 * with events with least weights first. Keep the current iterator
545 * state in struct sched_state.
549 int event; /* event index */
550 int counter; /* counter index */
551 int unassigned; /* number of events to be assigned left */
552 unsigned long used[BITS_TO_LONGS(X86_PMC_IDX_MAX)];
555 /* Total max is X86_PMC_IDX_MAX, but we are O(n!) limited */
556 #define SCHED_STATES_MAX 2
561 struct event_constraint **constraints;
562 struct sched_state state;
564 struct sched_state saved[SCHED_STATES_MAX];
568 * Initialize interator that runs through all events and counters.
570 static void perf_sched_init(struct perf_sched *sched, struct event_constraint **c,
571 int num, int wmin, int wmax)
575 memset(sched, 0, sizeof(*sched));
576 sched->max_events = num;
577 sched->max_weight = wmax;
578 sched->constraints = c;
580 for (idx = 0; idx < num; idx++) {
581 if (c[idx]->weight == wmin)
585 sched->state.event = idx; /* start with min weight */
586 sched->state.weight = wmin;
587 sched->state.unassigned = num;
590 static void perf_sched_save_state(struct perf_sched *sched)
592 if (WARN_ON_ONCE(sched->saved_states >= SCHED_STATES_MAX))
595 sched->saved[sched->saved_states] = sched->state;
596 sched->saved_states++;
599 static bool perf_sched_restore_state(struct perf_sched *sched)
601 if (!sched->saved_states)
604 sched->saved_states--;
605 sched->state = sched->saved[sched->saved_states];
607 /* continue with next counter: */
608 clear_bit(sched->state.counter++, sched->state.used);
614 * Select a counter for the current event to schedule. Return true on
617 static bool __perf_sched_find_counter(struct perf_sched *sched)
619 struct event_constraint *c;
622 if (!sched->state.unassigned)
625 if (sched->state.event >= sched->max_events)
628 c = sched->constraints[sched->state.event];
630 /* Prefer fixed purpose counters */
631 if (c->idxmsk64 & (~0ULL << INTEL_PMC_IDX_FIXED)) {
632 idx = INTEL_PMC_IDX_FIXED;
633 for_each_set_bit_from(idx, c->idxmsk, X86_PMC_IDX_MAX) {
634 if (!__test_and_set_bit(idx, sched->state.used))
638 /* Grab the first unused counter starting with idx */
639 idx = sched->state.counter;
640 for_each_set_bit_from(idx, c->idxmsk, INTEL_PMC_IDX_FIXED) {
641 if (!__test_and_set_bit(idx, sched->state.used))
648 sched->state.counter = idx;
651 perf_sched_save_state(sched);
656 static bool perf_sched_find_counter(struct perf_sched *sched)
658 while (!__perf_sched_find_counter(sched)) {
659 if (!perf_sched_restore_state(sched))
667 * Go through all unassigned events and find the next one to schedule.
668 * Take events with the least weight first. Return true on success.
670 static bool perf_sched_next_event(struct perf_sched *sched)
672 struct event_constraint *c;
674 if (!sched->state.unassigned || !--sched->state.unassigned)
679 sched->state.event++;
680 if (sched->state.event >= sched->max_events) {
682 sched->state.event = 0;
683 sched->state.weight++;
684 if (sched->state.weight > sched->max_weight)
687 c = sched->constraints[sched->state.event];
688 } while (c->weight != sched->state.weight);
690 sched->state.counter = 0; /* start with first counter */
696 * Assign a counter for each event.
698 int perf_assign_events(struct event_constraint **constraints, int n,
699 int wmin, int wmax, int *assign)
701 struct perf_sched sched;
703 perf_sched_init(&sched, constraints, n, wmin, wmax);
706 if (!perf_sched_find_counter(&sched))
709 assign[sched.state.event] = sched.state.counter;
710 } while (perf_sched_next_event(&sched));
712 return sched.state.unassigned;
715 int x86_schedule_events(struct cpu_hw_events *cpuc, int n, int *assign)
717 struct event_constraint *c, *constraints[X86_PMC_IDX_MAX];
718 unsigned long used_mask[BITS_TO_LONGS(X86_PMC_IDX_MAX)];
719 int i, wmin, wmax, num = 0;
720 struct hw_perf_event *hwc;
722 bitmap_zero(used_mask, X86_PMC_IDX_MAX);
724 for (i = 0, wmin = X86_PMC_IDX_MAX, wmax = 0; i < n; i++) {
725 c = x86_pmu.get_event_constraints(cpuc, cpuc->event_list[i]);
727 wmin = min(wmin, c->weight);
728 wmax = max(wmax, c->weight);
732 * fastpath, try to reuse previous register
734 for (i = 0; i < n; i++) {
735 hwc = &cpuc->event_list[i]->hw;
742 /* constraint still honored */
743 if (!test_bit(hwc->idx, c->idxmsk))
746 /* not already used */
747 if (test_bit(hwc->idx, used_mask))
750 __set_bit(hwc->idx, used_mask);
752 assign[i] = hwc->idx;
757 num = perf_assign_events(constraints, n, wmin, wmax, assign);
760 * scheduling failed or is just a simulation,
761 * free resources if necessary
763 if (!assign || num) {
764 for (i = 0; i < n; i++) {
765 if (x86_pmu.put_event_constraints)
766 x86_pmu.put_event_constraints(cpuc, cpuc->event_list[i]);
769 return num ? -EINVAL : 0;
773 * dogrp: true if must collect siblings events (group)
774 * returns total number of events and error code
776 static int collect_events(struct cpu_hw_events *cpuc, struct perf_event *leader, bool dogrp)
778 struct perf_event *event;
781 max_count = x86_pmu.num_counters + x86_pmu.num_counters_fixed;
783 /* current number of events already accepted */
786 if (is_x86_event(leader)) {
789 cpuc->event_list[n] = leader;
795 list_for_each_entry(event, &leader->sibling_list, group_entry) {
796 if (!is_x86_event(event) ||
797 event->state <= PERF_EVENT_STATE_OFF)
803 cpuc->event_list[n] = event;
809 static inline void x86_assign_hw_event(struct perf_event *event,
810 struct cpu_hw_events *cpuc, int i)
812 struct hw_perf_event *hwc = &event->hw;
814 hwc->idx = cpuc->assign[i];
815 hwc->last_cpu = smp_processor_id();
816 hwc->last_tag = ++cpuc->tags[i];
818 if (hwc->idx == INTEL_PMC_IDX_FIXED_BTS) {
819 hwc->config_base = 0;
821 } else if (hwc->idx >= INTEL_PMC_IDX_FIXED) {
822 hwc->config_base = MSR_ARCH_PERFMON_FIXED_CTR_CTRL;
823 hwc->event_base = MSR_ARCH_PERFMON_FIXED_CTR0 + (hwc->idx - INTEL_PMC_IDX_FIXED);
824 hwc->event_base_rdpmc = (hwc->idx - INTEL_PMC_IDX_FIXED) | 1<<30;
826 hwc->config_base = x86_pmu_config_addr(hwc->idx);
827 hwc->event_base = x86_pmu_event_addr(hwc->idx);
828 hwc->event_base_rdpmc = hwc->idx;
832 static inline int match_prev_assignment(struct hw_perf_event *hwc,
833 struct cpu_hw_events *cpuc,
836 return hwc->idx == cpuc->assign[i] &&
837 hwc->last_cpu == smp_processor_id() &&
838 hwc->last_tag == cpuc->tags[i];
841 static void x86_pmu_start(struct perf_event *event, int flags);
843 static void x86_pmu_enable(struct pmu *pmu)
845 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
846 struct perf_event *event;
847 struct hw_perf_event *hwc;
848 int i, added = cpuc->n_added;
850 if (!x86_pmu_initialized())
857 int n_running = cpuc->n_events - cpuc->n_added;
859 * apply assignment obtained either from
860 * hw_perf_group_sched_in() or x86_pmu_enable()
862 * step1: save events moving to new counters
863 * step2: reprogram moved events into new counters
865 for (i = 0; i < n_running; i++) {
866 event = cpuc->event_list[i];
870 * we can avoid reprogramming counter if:
871 * - assigned same counter as last time
872 * - running on same CPU as last time
873 * - no other event has used the counter since
875 if (hwc->idx == -1 ||
876 match_prev_assignment(hwc, cpuc, i))
880 * Ensure we don't accidentally enable a stopped
881 * counter simply because we rescheduled.
883 if (hwc->state & PERF_HES_STOPPED)
884 hwc->state |= PERF_HES_ARCH;
886 x86_pmu_stop(event, PERF_EF_UPDATE);
889 for (i = 0; i < cpuc->n_events; i++) {
890 event = cpuc->event_list[i];
893 if (!match_prev_assignment(hwc, cpuc, i))
894 x86_assign_hw_event(event, cpuc, i);
895 else if (i < n_running)
898 if (hwc->state & PERF_HES_ARCH)
901 x86_pmu_start(event, PERF_EF_RELOAD);
904 perf_events_lapic_init();
910 x86_pmu.enable_all(added);
913 static DEFINE_PER_CPU(u64 [X86_PMC_IDX_MAX], pmc_prev_left);
916 * Set the next IRQ period, based on the hwc->period_left value.
917 * To be called with the event disabled in hw:
919 int x86_perf_event_set_period(struct perf_event *event)
921 struct hw_perf_event *hwc = &event->hw;
922 s64 left = local64_read(&hwc->period_left);
923 s64 period = hwc->sample_period;
924 int ret = 0, idx = hwc->idx;
926 if (idx == INTEL_PMC_IDX_FIXED_BTS)
930 * If we are way outside a reasonable range then just skip forward:
932 if (unlikely(left <= -period)) {
934 local64_set(&hwc->period_left, left);
935 hwc->last_period = period;
939 if (unlikely(left <= 0)) {
941 local64_set(&hwc->period_left, left);
942 hwc->last_period = period;
946 * Quirk: certain CPUs dont like it if just 1 hw_event is left:
948 if (unlikely(left < 2))
951 if (left > x86_pmu.max_period)
952 left = x86_pmu.max_period;
954 per_cpu(pmc_prev_left[idx], smp_processor_id()) = left;
957 * The hw event starts counting from this event offset,
958 * mark it to be able to extra future deltas:
960 local64_set(&hwc->prev_count, (u64)-left);
962 wrmsrl(hwc->event_base, (u64)(-left) & x86_pmu.cntval_mask);
965 * Due to erratum on certan cpu we need
966 * a second write to be sure the register
967 * is updated properly
969 if (x86_pmu.perfctr_second_write) {
970 wrmsrl(hwc->event_base,
971 (u64)(-left) & x86_pmu.cntval_mask);
974 perf_event_update_userpage(event);
979 void x86_pmu_enable_event(struct perf_event *event)
981 if (__this_cpu_read(cpu_hw_events.enabled))
982 __x86_pmu_enable_event(&event->hw,
983 ARCH_PERFMON_EVENTSEL_ENABLE);
987 * Add a single event to the PMU.
989 * The event is added to the group of enabled events
990 * but only if it can be scehduled with existing events.
992 static int x86_pmu_add(struct perf_event *event, int flags)
994 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
995 struct hw_perf_event *hwc;
996 int assign[X86_PMC_IDX_MAX];
1001 perf_pmu_disable(event->pmu);
1002 n0 = cpuc->n_events;
1003 ret = n = collect_events(cpuc, event, false);
1007 hwc->state = PERF_HES_UPTODATE | PERF_HES_STOPPED;
1008 if (!(flags & PERF_EF_START))
1009 hwc->state |= PERF_HES_ARCH;
1012 * If group events scheduling transaction was started,
1013 * skip the schedulability test here, it will be performed
1014 * at commit time (->commit_txn) as a whole
1016 if (cpuc->group_flag & PERF_EVENT_TXN)
1019 ret = x86_pmu.schedule_events(cpuc, n, assign);
1023 * copy new assignment, now we know it is possible
1024 * will be used by hw_perf_enable()
1026 memcpy(cpuc->assign, assign, n*sizeof(int));
1030 cpuc->n_added += n - n0;
1031 cpuc->n_txn += n - n0;
1035 perf_pmu_enable(event->pmu);
1039 static void x86_pmu_start(struct perf_event *event, int flags)
1041 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
1042 int idx = event->hw.idx;
1044 if (WARN_ON_ONCE(!(event->hw.state & PERF_HES_STOPPED)))
1047 if (WARN_ON_ONCE(idx == -1))
1050 if (flags & PERF_EF_RELOAD) {
1051 WARN_ON_ONCE(!(event->hw.state & PERF_HES_UPTODATE));
1052 x86_perf_event_set_period(event);
1055 event->hw.state = 0;
1057 cpuc->events[idx] = event;
1058 __set_bit(idx, cpuc->active_mask);
1059 __set_bit(idx, cpuc->running);
1060 x86_pmu.enable(event);
1061 perf_event_update_userpage(event);
1064 void perf_event_print_debug(void)
1066 u64 ctrl, status, overflow, pmc_ctrl, pmc_count, prev_left, fixed;
1068 struct cpu_hw_events *cpuc;
1069 unsigned long flags;
1072 if (!x86_pmu.num_counters)
1075 local_irq_save(flags);
1077 cpu = smp_processor_id();
1078 cpuc = &per_cpu(cpu_hw_events, cpu);
1080 if (x86_pmu.version >= 2) {
1081 rdmsrl(MSR_CORE_PERF_GLOBAL_CTRL, ctrl);
1082 rdmsrl(MSR_CORE_PERF_GLOBAL_STATUS, status);
1083 rdmsrl(MSR_CORE_PERF_GLOBAL_OVF_CTRL, overflow);
1084 rdmsrl(MSR_ARCH_PERFMON_FIXED_CTR_CTRL, fixed);
1085 rdmsrl(MSR_IA32_PEBS_ENABLE, pebs);
1088 pr_info("CPU#%d: ctrl: %016llx\n", cpu, ctrl);
1089 pr_info("CPU#%d: status: %016llx\n", cpu, status);
1090 pr_info("CPU#%d: overflow: %016llx\n", cpu, overflow);
1091 pr_info("CPU#%d: fixed: %016llx\n", cpu, fixed);
1092 pr_info("CPU#%d: pebs: %016llx\n", cpu, pebs);
1094 pr_info("CPU#%d: active: %016llx\n", cpu, *(u64 *)cpuc->active_mask);
1096 for (idx = 0; idx < x86_pmu.num_counters; idx++) {
1097 rdmsrl(x86_pmu_config_addr(idx), pmc_ctrl);
1098 rdmsrl(x86_pmu_event_addr(idx), pmc_count);
1100 prev_left = per_cpu(pmc_prev_left[idx], cpu);
1102 pr_info("CPU#%d: gen-PMC%d ctrl: %016llx\n",
1103 cpu, idx, pmc_ctrl);
1104 pr_info("CPU#%d: gen-PMC%d count: %016llx\n",
1105 cpu, idx, pmc_count);
1106 pr_info("CPU#%d: gen-PMC%d left: %016llx\n",
1107 cpu, idx, prev_left);
1109 for (idx = 0; idx < x86_pmu.num_counters_fixed; idx++) {
1110 rdmsrl(MSR_ARCH_PERFMON_FIXED_CTR0 + idx, pmc_count);
1112 pr_info("CPU#%d: fixed-PMC%d count: %016llx\n",
1113 cpu, idx, pmc_count);
1115 local_irq_restore(flags);
1118 void x86_pmu_stop(struct perf_event *event, int flags)
1120 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
1121 struct hw_perf_event *hwc = &event->hw;
1123 if (__test_and_clear_bit(hwc->idx, cpuc->active_mask)) {
1124 x86_pmu.disable(event);
1125 cpuc->events[hwc->idx] = NULL;
1126 WARN_ON_ONCE(hwc->state & PERF_HES_STOPPED);
1127 hwc->state |= PERF_HES_STOPPED;
1130 if ((flags & PERF_EF_UPDATE) && !(hwc->state & PERF_HES_UPTODATE)) {
1132 * Drain the remaining delta count out of a event
1133 * that we are disabling:
1135 x86_perf_event_update(event);
1136 hwc->state |= PERF_HES_UPTODATE;
1140 static void x86_pmu_del(struct perf_event *event, int flags)
1142 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
1146 * If we're called during a txn, we don't need to do anything.
1147 * The events never got scheduled and ->cancel_txn will truncate
1150 if (cpuc->group_flag & PERF_EVENT_TXN)
1153 x86_pmu_stop(event, PERF_EF_UPDATE);
1155 for (i = 0; i < cpuc->n_events; i++) {
1156 if (event == cpuc->event_list[i]) {
1158 if (x86_pmu.put_event_constraints)
1159 x86_pmu.put_event_constraints(cpuc, event);
1161 while (++i < cpuc->n_events)
1162 cpuc->event_list[i-1] = cpuc->event_list[i];
1168 perf_event_update_userpage(event);
1171 int x86_pmu_handle_irq(struct pt_regs *regs)
1173 struct perf_sample_data data;
1174 struct cpu_hw_events *cpuc;
1175 struct perf_event *event;
1176 int idx, handled = 0;
1179 cpuc = &__get_cpu_var(cpu_hw_events);
1182 * Some chipsets need to unmask the LVTPC in a particular spot
1183 * inside the nmi handler. As a result, the unmasking was pushed
1184 * into all the nmi handlers.
1186 * This generic handler doesn't seem to have any issues where the
1187 * unmasking occurs so it was left at the top.
1189 apic_write(APIC_LVTPC, APIC_DM_NMI);
1191 for (idx = 0; idx < x86_pmu.num_counters; idx++) {
1192 if (!test_bit(idx, cpuc->active_mask)) {
1194 * Though we deactivated the counter some cpus
1195 * might still deliver spurious interrupts still
1196 * in flight. Catch them:
1198 if (__test_and_clear_bit(idx, cpuc->running))
1203 event = cpuc->events[idx];
1205 val = x86_perf_event_update(event);
1206 if (val & (1ULL << (x86_pmu.cntval_bits - 1)))
1213 perf_sample_data_init(&data, 0, event->hw.last_period);
1215 if (!x86_perf_event_set_period(event))
1218 if (perf_event_overflow(event, &data, regs))
1219 x86_pmu_stop(event, 0);
1223 inc_irq_stat(apic_perf_irqs);
1228 void perf_events_lapic_init(void)
1230 if (!x86_pmu.apic || !x86_pmu_initialized())
1234 * Always use NMI for PMU
1236 apic_write(APIC_LVTPC, APIC_DM_NMI);
1239 static int __kprobes
1240 perf_event_nmi_handler(unsigned int cmd, struct pt_regs *regs)
1242 if (!atomic_read(&active_events))
1245 return x86_pmu.handle_irq(regs);
1248 struct event_constraint emptyconstraint;
1249 struct event_constraint unconstrained;
1251 static int __cpuinit
1252 x86_pmu_notifier(struct notifier_block *self, unsigned long action, void *hcpu)
1254 unsigned int cpu = (long)hcpu;
1255 struct cpu_hw_events *cpuc = &per_cpu(cpu_hw_events, cpu);
1256 int ret = NOTIFY_OK;
1258 switch (action & ~CPU_TASKS_FROZEN) {
1259 case CPU_UP_PREPARE:
1260 cpuc->kfree_on_online = NULL;
1261 if (x86_pmu.cpu_prepare)
1262 ret = x86_pmu.cpu_prepare(cpu);
1266 if (x86_pmu.attr_rdpmc)
1267 set_in_cr4(X86_CR4_PCE);
1268 if (x86_pmu.cpu_starting)
1269 x86_pmu.cpu_starting(cpu);
1273 kfree(cpuc->kfree_on_online);
1277 if (x86_pmu.cpu_dying)
1278 x86_pmu.cpu_dying(cpu);
1281 case CPU_UP_CANCELED:
1283 if (x86_pmu.cpu_dead)
1284 x86_pmu.cpu_dead(cpu);
1294 static void __init pmu_check_apic(void)
1300 pr_info("no APIC, boot with the \"lapic\" boot parameter to force-enable it.\n");
1301 pr_info("no hardware sampling interrupt available.\n");
1304 static struct attribute_group x86_pmu_format_group = {
1309 static int __init init_hw_perf_events(void)
1311 struct x86_pmu_quirk *quirk;
1314 pr_info("Performance Events: ");
1316 switch (boot_cpu_data.x86_vendor) {
1317 case X86_VENDOR_INTEL:
1318 err = intel_pmu_init();
1320 case X86_VENDOR_AMD:
1321 err = amd_pmu_init();
1327 pr_cont("no PMU driver, software events only.\n");
1333 /* sanity check that the hardware exists or is emulated */
1334 if (!check_hw_exists())
1337 pr_cont("%s PMU driver.\n", x86_pmu.name);
1339 for (quirk = x86_pmu.quirks; quirk; quirk = quirk->next)
1342 if (!x86_pmu.intel_ctrl)
1343 x86_pmu.intel_ctrl = (1 << x86_pmu.num_counters) - 1;
1345 perf_events_lapic_init();
1346 register_nmi_handler(NMI_LOCAL, perf_event_nmi_handler, 0, "PMI");
1348 unconstrained = (struct event_constraint)
1349 __EVENT_CONSTRAINT(0, (1ULL << x86_pmu.num_counters) - 1,
1350 0, x86_pmu.num_counters, 0);
1352 x86_pmu.attr_rdpmc = 1; /* enable userspace RDPMC usage by default */
1353 x86_pmu_format_group.attrs = x86_pmu.format_attrs;
1355 pr_info("... version: %d\n", x86_pmu.version);
1356 pr_info("... bit width: %d\n", x86_pmu.cntval_bits);
1357 pr_info("... generic registers: %d\n", x86_pmu.num_counters);
1358 pr_info("... value mask: %016Lx\n", x86_pmu.cntval_mask);
1359 pr_info("... max period: %016Lx\n", x86_pmu.max_period);
1360 pr_info("... fixed-purpose events: %d\n", x86_pmu.num_counters_fixed);
1361 pr_info("... event mask: %016Lx\n", x86_pmu.intel_ctrl);
1363 perf_pmu_register(&pmu, "cpu", PERF_TYPE_RAW);
1364 perf_cpu_notifier(x86_pmu_notifier);
1368 early_initcall(init_hw_perf_events);
1370 static inline void x86_pmu_read(struct perf_event *event)
1372 x86_perf_event_update(event);
1376 * Start group events scheduling transaction
1377 * Set the flag to make pmu::enable() not perform the
1378 * schedulability test, it will be performed at commit time
1380 static void x86_pmu_start_txn(struct pmu *pmu)
1382 perf_pmu_disable(pmu);
1383 __this_cpu_or(cpu_hw_events.group_flag, PERF_EVENT_TXN);
1384 __this_cpu_write(cpu_hw_events.n_txn, 0);
1388 * Stop group events scheduling transaction
1389 * Clear the flag and pmu::enable() will perform the
1390 * schedulability test.
1392 static void x86_pmu_cancel_txn(struct pmu *pmu)
1394 __this_cpu_and(cpu_hw_events.group_flag, ~PERF_EVENT_TXN);
1396 * Truncate the collected events.
1398 __this_cpu_sub(cpu_hw_events.n_added, __this_cpu_read(cpu_hw_events.n_txn));
1399 __this_cpu_sub(cpu_hw_events.n_events, __this_cpu_read(cpu_hw_events.n_txn));
1400 perf_pmu_enable(pmu);
1404 * Commit group events scheduling transaction
1405 * Perform the group schedulability test as a whole
1406 * Return 0 if success
1408 static int x86_pmu_commit_txn(struct pmu *pmu)
1410 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
1411 int assign[X86_PMC_IDX_MAX];
1416 if (!x86_pmu_initialized())
1419 ret = x86_pmu.schedule_events(cpuc, n, assign);
1424 * copy new assignment, now we know it is possible
1425 * will be used by hw_perf_enable()
1427 memcpy(cpuc->assign, assign, n*sizeof(int));
1429 cpuc->group_flag &= ~PERF_EVENT_TXN;
1430 perf_pmu_enable(pmu);
1434 * a fake_cpuc is used to validate event groups. Due to
1435 * the extra reg logic, we need to also allocate a fake
1436 * per_core and per_cpu structure. Otherwise, group events
1437 * using extra reg may conflict without the kernel being
1438 * able to catch this when the last event gets added to
1441 static void free_fake_cpuc(struct cpu_hw_events *cpuc)
1443 kfree(cpuc->shared_regs);
1447 static struct cpu_hw_events *allocate_fake_cpuc(void)
1449 struct cpu_hw_events *cpuc;
1450 int cpu = raw_smp_processor_id();
1452 cpuc = kzalloc(sizeof(*cpuc), GFP_KERNEL);
1454 return ERR_PTR(-ENOMEM);
1456 /* only needed, if we have extra_regs */
1457 if (x86_pmu.extra_regs) {
1458 cpuc->shared_regs = allocate_shared_regs(cpu);
1459 if (!cpuc->shared_regs)
1465 free_fake_cpuc(cpuc);
1466 return ERR_PTR(-ENOMEM);
1470 * validate that we can schedule this event
1472 static int validate_event(struct perf_event *event)
1474 struct cpu_hw_events *fake_cpuc;
1475 struct event_constraint *c;
1478 fake_cpuc = allocate_fake_cpuc();
1479 if (IS_ERR(fake_cpuc))
1480 return PTR_ERR(fake_cpuc);
1482 c = x86_pmu.get_event_constraints(fake_cpuc, event);
1484 if (!c || !c->weight)
1487 if (x86_pmu.put_event_constraints)
1488 x86_pmu.put_event_constraints(fake_cpuc, event);
1490 free_fake_cpuc(fake_cpuc);
1496 * validate a single event group
1498 * validation include:
1499 * - check events are compatible which each other
1500 * - events do not compete for the same counter
1501 * - number of events <= number of counters
1503 * validation ensures the group can be loaded onto the
1504 * PMU if it was the only group available.
1506 static int validate_group(struct perf_event *event)
1508 struct perf_event *leader = event->group_leader;
1509 struct cpu_hw_events *fake_cpuc;
1510 int ret = -EINVAL, n;
1512 fake_cpuc = allocate_fake_cpuc();
1513 if (IS_ERR(fake_cpuc))
1514 return PTR_ERR(fake_cpuc);
1516 * the event is not yet connected with its
1517 * siblings therefore we must first collect
1518 * existing siblings, then add the new event
1519 * before we can simulate the scheduling
1521 n = collect_events(fake_cpuc, leader, true);
1525 fake_cpuc->n_events = n;
1526 n = collect_events(fake_cpuc, event, false);
1530 fake_cpuc->n_events = n;
1532 ret = x86_pmu.schedule_events(fake_cpuc, n, NULL);
1535 free_fake_cpuc(fake_cpuc);
1539 static int x86_pmu_event_init(struct perf_event *event)
1544 switch (event->attr.type) {
1546 case PERF_TYPE_HARDWARE:
1547 case PERF_TYPE_HW_CACHE:
1554 err = __x86_pmu_event_init(event);
1557 * we temporarily connect event to its pmu
1558 * such that validate_group() can classify
1559 * it as an x86 event using is_x86_event()
1564 if (event->group_leader != event)
1565 err = validate_group(event);
1567 err = validate_event(event);
1573 event->destroy(event);
1579 static int x86_pmu_event_idx(struct perf_event *event)
1581 int idx = event->hw.idx;
1583 if (!x86_pmu.attr_rdpmc)
1586 if (x86_pmu.num_counters_fixed && idx >= INTEL_PMC_IDX_FIXED) {
1587 idx -= INTEL_PMC_IDX_FIXED;
1594 static ssize_t get_attr_rdpmc(struct device *cdev,
1595 struct device_attribute *attr,
1598 return snprintf(buf, 40, "%d\n", x86_pmu.attr_rdpmc);
1601 static void change_rdpmc(void *info)
1603 bool enable = !!(unsigned long)info;
1606 set_in_cr4(X86_CR4_PCE);
1608 clear_in_cr4(X86_CR4_PCE);
1611 static ssize_t set_attr_rdpmc(struct device *cdev,
1612 struct device_attribute *attr,
1613 const char *buf, size_t count)
1618 ret = kstrtoul(buf, 0, &val);
1622 if (!!val != !!x86_pmu.attr_rdpmc) {
1623 x86_pmu.attr_rdpmc = !!val;
1624 smp_call_function(change_rdpmc, (void *)val, 1);
1630 static DEVICE_ATTR(rdpmc, S_IRUSR | S_IWUSR, get_attr_rdpmc, set_attr_rdpmc);
1632 static struct attribute *x86_pmu_attrs[] = {
1633 &dev_attr_rdpmc.attr,
1637 static struct attribute_group x86_pmu_attr_group = {
1638 .attrs = x86_pmu_attrs,
1641 static const struct attribute_group *x86_pmu_attr_groups[] = {
1642 &x86_pmu_attr_group,
1643 &x86_pmu_format_group,
1647 static void x86_pmu_flush_branch_stack(void)
1649 if (x86_pmu.flush_branch_stack)
1650 x86_pmu.flush_branch_stack();
1653 void perf_check_microcode(void)
1655 if (x86_pmu.check_microcode)
1656 x86_pmu.check_microcode();
1658 EXPORT_SYMBOL_GPL(perf_check_microcode);
1660 static struct pmu pmu = {
1661 .pmu_enable = x86_pmu_enable,
1662 .pmu_disable = x86_pmu_disable,
1664 .attr_groups = x86_pmu_attr_groups,
1666 .event_init = x86_pmu_event_init,
1670 .start = x86_pmu_start,
1671 .stop = x86_pmu_stop,
1672 .read = x86_pmu_read,
1674 .start_txn = x86_pmu_start_txn,
1675 .cancel_txn = x86_pmu_cancel_txn,
1676 .commit_txn = x86_pmu_commit_txn,
1678 .event_idx = x86_pmu_event_idx,
1679 .flush_branch_stack = x86_pmu_flush_branch_stack,
1682 void arch_perf_update_userpage(struct perf_event_mmap_page *userpg, u64 now)
1684 userpg->cap_usr_time = 0;
1685 userpg->cap_usr_rdpmc = x86_pmu.attr_rdpmc;
1686 userpg->pmc_width = x86_pmu.cntval_bits;
1688 if (!boot_cpu_has(X86_FEATURE_CONSTANT_TSC))
1691 if (!boot_cpu_has(X86_FEATURE_NONSTOP_TSC))
1694 userpg->cap_usr_time = 1;
1695 userpg->time_mult = this_cpu_read(cyc2ns);
1696 userpg->time_shift = CYC2NS_SCALE_FACTOR;
1697 userpg->time_offset = this_cpu_read(cyc2ns_offset) - now;
1704 static int backtrace_stack(void *data, char *name)
1709 static void backtrace_address(void *data, unsigned long addr, int reliable)
1711 struct perf_callchain_entry *entry = data;
1713 perf_callchain_store(entry, addr);
1716 static const struct stacktrace_ops backtrace_ops = {
1717 .stack = backtrace_stack,
1718 .address = backtrace_address,
1719 .walk_stack = print_context_stack_bp,
1723 perf_callchain_kernel(struct perf_callchain_entry *entry, struct pt_regs *regs)
1725 if (perf_guest_cbs && perf_guest_cbs->is_in_guest()) {
1726 /* TODO: We don't support guest os callchain now */
1730 perf_callchain_store(entry, regs->ip);
1732 dump_trace(NULL, regs, NULL, 0, &backtrace_ops, entry);
1736 valid_user_frame(const void __user *fp, unsigned long size)
1738 return (__range_not_ok(fp, size, TASK_SIZE) == 0);
1741 #ifdef CONFIG_COMPAT
1743 #include <asm/compat.h>
1746 perf_callchain_user32(struct pt_regs *regs, struct perf_callchain_entry *entry)
1748 /* 32-bit process in 64-bit kernel. */
1749 struct stack_frame_ia32 frame;
1750 const void __user *fp;
1752 if (!test_thread_flag(TIF_IA32))
1755 fp = compat_ptr(regs->bp);
1756 while (entry->nr < PERF_MAX_STACK_DEPTH) {
1757 unsigned long bytes;
1758 frame.next_frame = 0;
1759 frame.return_address = 0;
1761 bytes = copy_from_user_nmi(&frame, fp, sizeof(frame));
1762 if (bytes != sizeof(frame))
1765 if (!valid_user_frame(fp, sizeof(frame)))
1768 perf_callchain_store(entry, frame.return_address);
1769 fp = compat_ptr(frame.next_frame);
1775 perf_callchain_user32(struct pt_regs *regs, struct perf_callchain_entry *entry)
1782 perf_callchain_user(struct perf_callchain_entry *entry, struct pt_regs *regs)
1784 struct stack_frame frame;
1785 const void __user *fp;
1787 if (perf_guest_cbs && perf_guest_cbs->is_in_guest()) {
1788 /* TODO: We don't support guest os callchain now */
1792 fp = (void __user *)regs->bp;
1794 perf_callchain_store(entry, regs->ip);
1799 if (perf_callchain_user32(regs, entry))
1802 while (entry->nr < PERF_MAX_STACK_DEPTH) {
1803 unsigned long bytes;
1804 frame.next_frame = NULL;
1805 frame.return_address = 0;
1807 bytes = copy_from_user_nmi(&frame, fp, sizeof(frame));
1808 if (bytes != sizeof(frame))
1811 if (!valid_user_frame(fp, sizeof(frame)))
1814 perf_callchain_store(entry, frame.return_address);
1815 fp = frame.next_frame;
1819 unsigned long perf_instruction_pointer(struct pt_regs *regs)
1823 if (perf_guest_cbs && perf_guest_cbs->is_in_guest())
1824 ip = perf_guest_cbs->get_guest_ip();
1826 ip = instruction_pointer(regs);
1831 unsigned long perf_misc_flags(struct pt_regs *regs)
1835 if (perf_guest_cbs && perf_guest_cbs->is_in_guest()) {
1836 if (perf_guest_cbs->is_user_mode())
1837 misc |= PERF_RECORD_MISC_GUEST_USER;
1839 misc |= PERF_RECORD_MISC_GUEST_KERNEL;
1841 if (!kernel_ip(regs->ip))
1842 misc |= PERF_RECORD_MISC_USER;
1844 misc |= PERF_RECORD_MISC_KERNEL;
1847 if (regs->flags & PERF_EFLAGS_EXACT)
1848 misc |= PERF_RECORD_MISC_EXACT_IP;
1853 void perf_get_x86_pmu_capability(struct x86_pmu_capability *cap)
1855 cap->version = x86_pmu.version;
1856 cap->num_counters_gp = x86_pmu.num_counters;
1857 cap->num_counters_fixed = x86_pmu.num_counters_fixed;
1858 cap->bit_width_gp = x86_pmu.cntval_bits;
1859 cap->bit_width_fixed = x86_pmu.cntval_bits;
1860 cap->events_mask = (unsigned int)x86_pmu.events_maskl;
1861 cap->events_mask_len = x86_pmu.events_mask_len;
1863 EXPORT_SYMBOL_GPL(perf_get_x86_pmu_capability);