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/tlbflush.h>
35 #include <asm/timer.h>
39 #include "perf_event.h"
41 struct x86_pmu x86_pmu __read_mostly;
43 DEFINE_PER_CPU(struct cpu_hw_events, cpu_hw_events) = {
47 u64 __read_mostly hw_cache_event_ids
48 [PERF_COUNT_HW_CACHE_MAX]
49 [PERF_COUNT_HW_CACHE_OP_MAX]
50 [PERF_COUNT_HW_CACHE_RESULT_MAX];
51 u64 __read_mostly hw_cache_extra_regs
52 [PERF_COUNT_HW_CACHE_MAX]
53 [PERF_COUNT_HW_CACHE_OP_MAX]
54 [PERF_COUNT_HW_CACHE_RESULT_MAX];
57 * Propagate event elapsed time into the generic event.
58 * Can only be executed on the CPU where the event is active.
59 * Returns the delta events processed.
61 u64 x86_perf_event_update(struct perf_event *event)
63 struct hw_perf_event *hwc = &event->hw;
64 int shift = 64 - x86_pmu.cntval_bits;
65 u64 prev_raw_count, new_raw_count;
69 if (idx == INTEL_PMC_IDX_FIXED_BTS)
73 * Careful: an NMI might modify the previous event value.
75 * Our tactic to handle this is to first atomically read and
76 * exchange a new raw count - then add that new-prev delta
77 * count to the generic event atomically:
80 prev_raw_count = local64_read(&hwc->prev_count);
81 rdpmcl(hwc->event_base_rdpmc, new_raw_count);
83 if (local64_cmpxchg(&hwc->prev_count, prev_raw_count,
84 new_raw_count) != prev_raw_count)
88 * Now we have the new raw value and have updated the prev
89 * timestamp already. We can now calculate the elapsed delta
90 * (event-)time and add that to the generic event.
92 * Careful, not all hw sign-extends above the physical width
95 delta = (new_raw_count << shift) - (prev_raw_count << shift);
98 local64_add(delta, &event->count);
99 local64_sub(delta, &hwc->period_left);
101 return new_raw_count;
105 * Find and validate any extra registers to set up.
107 static int x86_pmu_extra_regs(u64 config, struct perf_event *event)
109 struct hw_perf_event_extra *reg;
110 struct extra_reg *er;
112 reg = &event->hw.extra_reg;
114 if (!x86_pmu.extra_regs)
117 for (er = x86_pmu.extra_regs; er->msr; er++) {
118 if (er->event != (config & er->config_mask))
120 if (event->attr.config1 & ~er->valid_mask)
122 /* Check if the extra msrs can be safely accessed*/
123 if (!er->extra_msr_access)
127 reg->config = event->attr.config1;
134 static atomic_t active_events;
135 static DEFINE_MUTEX(pmc_reserve_mutex);
137 #ifdef CONFIG_X86_LOCAL_APIC
139 static bool reserve_pmc_hardware(void)
143 for (i = 0; i < x86_pmu.num_counters; i++) {
144 if (!reserve_perfctr_nmi(x86_pmu_event_addr(i)))
148 for (i = 0; i < x86_pmu.num_counters; i++) {
149 if (!reserve_evntsel_nmi(x86_pmu_config_addr(i)))
156 for (i--; i >= 0; i--)
157 release_evntsel_nmi(x86_pmu_config_addr(i));
159 i = x86_pmu.num_counters;
162 for (i--; i >= 0; i--)
163 release_perfctr_nmi(x86_pmu_event_addr(i));
168 static void release_pmc_hardware(void)
172 for (i = 0; i < x86_pmu.num_counters; i++) {
173 release_perfctr_nmi(x86_pmu_event_addr(i));
174 release_evntsel_nmi(x86_pmu_config_addr(i));
180 static bool reserve_pmc_hardware(void) { return true; }
181 static void release_pmc_hardware(void) {}
185 static bool check_hw_exists(void)
187 u64 val, val_fail, val_new= ~0;
188 int i, reg, reg_fail, ret = 0;
192 * Check to see if the BIOS enabled any of the counters, if so
195 for (i = 0; i < x86_pmu.num_counters; i++) {
196 reg = x86_pmu_config_addr(i);
197 ret = rdmsrl_safe(reg, &val);
200 if (val & ARCH_PERFMON_EVENTSEL_ENABLE) {
207 if (x86_pmu.num_counters_fixed) {
208 reg = MSR_ARCH_PERFMON_FIXED_CTR_CTRL;
209 ret = rdmsrl_safe(reg, &val);
212 for (i = 0; i < x86_pmu.num_counters_fixed; i++) {
213 if (val & (0x03 << i*4)) {
222 * Read the current value, change it and read it back to see if it
223 * matches, this is needed to detect certain hardware emulators
224 * (qemu/kvm) that don't trap on the MSR access and always return 0s.
226 reg = x86_pmu_event_addr(0);
227 if (rdmsrl_safe(reg, &val))
230 ret = wrmsrl_safe(reg, val);
231 ret |= rdmsrl_safe(reg, &val_new);
232 if (ret || val != val_new)
236 * We still allow the PMU driver to operate:
239 printk(KERN_CONT "Broken BIOS detected, complain to your hardware vendor.\n");
240 printk(KERN_ERR FW_BUG "the BIOS has corrupted hw-PMU resources (MSR %x is %Lx)\n", reg_fail, val_fail);
246 printk(KERN_CONT "Broken PMU hardware detected, using software events only.\n");
247 printk("%sFailed to access perfctr msr (MSR %x is %Lx)\n",
248 boot_cpu_has(X86_FEATURE_HYPERVISOR) ? KERN_INFO : KERN_ERR,
254 static void hw_perf_event_destroy(struct perf_event *event)
256 if (atomic_dec_and_mutex_lock(&active_events, &pmc_reserve_mutex)) {
257 release_pmc_hardware();
258 release_ds_buffers();
259 mutex_unlock(&pmc_reserve_mutex);
263 static inline int x86_pmu_initialized(void)
265 return x86_pmu.handle_irq != NULL;
269 set_ext_hw_attr(struct hw_perf_event *hwc, struct perf_event *event)
271 struct perf_event_attr *attr = &event->attr;
272 unsigned int cache_type, cache_op, cache_result;
275 config = attr->config;
277 cache_type = (config >> 0) & 0xff;
278 if (cache_type >= PERF_COUNT_HW_CACHE_MAX)
281 cache_op = (config >> 8) & 0xff;
282 if (cache_op >= PERF_COUNT_HW_CACHE_OP_MAX)
285 cache_result = (config >> 16) & 0xff;
286 if (cache_result >= PERF_COUNT_HW_CACHE_RESULT_MAX)
289 val = hw_cache_event_ids[cache_type][cache_op][cache_result];
298 attr->config1 = hw_cache_extra_regs[cache_type][cache_op][cache_result];
299 return x86_pmu_extra_regs(val, event);
302 int x86_setup_perfctr(struct perf_event *event)
304 struct perf_event_attr *attr = &event->attr;
305 struct hw_perf_event *hwc = &event->hw;
308 if (!is_sampling_event(event)) {
309 hwc->sample_period = x86_pmu.max_period;
310 hwc->last_period = hwc->sample_period;
311 local64_set(&hwc->period_left, hwc->sample_period);
314 if (attr->type == PERF_TYPE_RAW)
315 return x86_pmu_extra_regs(event->attr.config, event);
317 if (attr->type == PERF_TYPE_HW_CACHE)
318 return set_ext_hw_attr(hwc, event);
320 if (attr->config >= x86_pmu.max_events)
326 config = x86_pmu.event_map(attr->config);
337 if (attr->config == PERF_COUNT_HW_BRANCH_INSTRUCTIONS &&
338 !attr->freq && hwc->sample_period == 1) {
339 /* BTS is not supported by this architecture. */
340 if (!x86_pmu.bts_active)
343 /* BTS is currently only allowed for user-mode. */
344 if (!attr->exclude_kernel)
348 hwc->config |= config;
354 * check that branch_sample_type is compatible with
355 * settings needed for precise_ip > 1 which implies
356 * using the LBR to capture ALL taken branches at the
357 * priv levels of the measurement
359 static inline int precise_br_compat(struct perf_event *event)
361 u64 m = event->attr.branch_sample_type;
364 /* must capture all branches */
365 if (!(m & PERF_SAMPLE_BRANCH_ANY))
368 m &= PERF_SAMPLE_BRANCH_KERNEL | PERF_SAMPLE_BRANCH_USER;
370 if (!event->attr.exclude_user)
371 b |= PERF_SAMPLE_BRANCH_USER;
373 if (!event->attr.exclude_kernel)
374 b |= PERF_SAMPLE_BRANCH_KERNEL;
377 * ignore PERF_SAMPLE_BRANCH_HV, not supported on x86
383 int x86_pmu_hw_config(struct perf_event *event)
385 if (event->attr.precise_ip) {
388 /* Support for constant skid */
389 if (x86_pmu.pebs_active && !x86_pmu.pebs_broken) {
392 /* Support for IP fixup */
393 if (x86_pmu.lbr_nr || x86_pmu.intel_cap.pebs_format >= 2)
397 if (event->attr.precise_ip > precise)
400 * check that PEBS LBR correction does not conflict with
401 * whatever the user is asking with attr->branch_sample_type
403 if (event->attr.precise_ip > 1 &&
404 x86_pmu.intel_cap.pebs_format < 2) {
405 u64 *br_type = &event->attr.branch_sample_type;
407 if (has_branch_stack(event)) {
408 if (!precise_br_compat(event))
411 /* branch_sample_type is compatible */
415 * user did not specify branch_sample_type
417 * For PEBS fixups, we capture all
418 * the branches at the priv level of the
421 *br_type = PERF_SAMPLE_BRANCH_ANY;
423 if (!event->attr.exclude_user)
424 *br_type |= PERF_SAMPLE_BRANCH_USER;
426 if (!event->attr.exclude_kernel)
427 *br_type |= PERF_SAMPLE_BRANCH_KERNEL;
434 * (keep 'enabled' bit clear for now)
436 event->hw.config = ARCH_PERFMON_EVENTSEL_INT;
439 * Count user and OS events unless requested not to
441 if (!event->attr.exclude_user)
442 event->hw.config |= ARCH_PERFMON_EVENTSEL_USR;
443 if (!event->attr.exclude_kernel)
444 event->hw.config |= ARCH_PERFMON_EVENTSEL_OS;
446 if (event->attr.type == PERF_TYPE_RAW)
447 event->hw.config |= event->attr.config & X86_RAW_EVENT_MASK;
449 return x86_setup_perfctr(event);
453 * Setup the hardware configuration for a given attr_type
455 static int __x86_pmu_event_init(struct perf_event *event)
459 if (!x86_pmu_initialized())
463 if (!atomic_inc_not_zero(&active_events)) {
464 mutex_lock(&pmc_reserve_mutex);
465 if (atomic_read(&active_events) == 0) {
466 if (!reserve_pmc_hardware())
469 reserve_ds_buffers();
472 atomic_inc(&active_events);
473 mutex_unlock(&pmc_reserve_mutex);
478 event->destroy = hw_perf_event_destroy;
481 event->hw.last_cpu = -1;
482 event->hw.last_tag = ~0ULL;
485 event->hw.extra_reg.idx = EXTRA_REG_NONE;
486 event->hw.branch_reg.idx = EXTRA_REG_NONE;
488 return x86_pmu.hw_config(event);
491 void x86_pmu_disable_all(void)
493 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
496 for (idx = 0; idx < x86_pmu.num_counters; idx++) {
499 if (!test_bit(idx, cpuc->active_mask))
501 rdmsrl(x86_pmu_config_addr(idx), val);
502 if (!(val & ARCH_PERFMON_EVENTSEL_ENABLE))
504 val &= ~ARCH_PERFMON_EVENTSEL_ENABLE;
505 wrmsrl(x86_pmu_config_addr(idx), val);
509 static void x86_pmu_disable(struct pmu *pmu)
511 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
513 if (!x86_pmu_initialized())
523 x86_pmu.disable_all();
526 void x86_pmu_enable_all(int added)
528 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
531 for (idx = 0; idx < x86_pmu.num_counters; idx++) {
532 struct hw_perf_event *hwc = &cpuc->events[idx]->hw;
534 if (!test_bit(idx, cpuc->active_mask))
537 __x86_pmu_enable_event(hwc, ARCH_PERFMON_EVENTSEL_ENABLE);
541 static struct pmu pmu;
543 static inline int is_x86_event(struct perf_event *event)
545 return event->pmu == &pmu;
549 * Event scheduler state:
551 * Assign events iterating over all events and counters, beginning
552 * with events with least weights first. Keep the current iterator
553 * state in struct sched_state.
557 int event; /* event index */
558 int counter; /* counter index */
559 int unassigned; /* number of events to be assigned left */
560 unsigned long used[BITS_TO_LONGS(X86_PMC_IDX_MAX)];
563 /* Total max is X86_PMC_IDX_MAX, but we are O(n!) limited */
564 #define SCHED_STATES_MAX 2
569 struct perf_event **events;
570 struct sched_state state;
572 struct sched_state saved[SCHED_STATES_MAX];
576 * Initialize interator that runs through all events and counters.
578 static void perf_sched_init(struct perf_sched *sched, struct perf_event **events,
579 int num, int wmin, int wmax)
583 memset(sched, 0, sizeof(*sched));
584 sched->max_events = num;
585 sched->max_weight = wmax;
586 sched->events = events;
588 for (idx = 0; idx < num; idx++) {
589 if (events[idx]->hw.constraint->weight == wmin)
593 sched->state.event = idx; /* start with min weight */
594 sched->state.weight = wmin;
595 sched->state.unassigned = num;
598 static void perf_sched_save_state(struct perf_sched *sched)
600 if (WARN_ON_ONCE(sched->saved_states >= SCHED_STATES_MAX))
603 sched->saved[sched->saved_states] = sched->state;
604 sched->saved_states++;
607 static bool perf_sched_restore_state(struct perf_sched *sched)
609 if (!sched->saved_states)
612 sched->saved_states--;
613 sched->state = sched->saved[sched->saved_states];
615 /* continue with next counter: */
616 clear_bit(sched->state.counter++, sched->state.used);
622 * Select a counter for the current event to schedule. Return true on
625 static bool __perf_sched_find_counter(struct perf_sched *sched)
627 struct event_constraint *c;
630 if (!sched->state.unassigned)
633 if (sched->state.event >= sched->max_events)
636 c = sched->events[sched->state.event]->hw.constraint;
637 /* Prefer fixed purpose counters */
638 if (c->idxmsk64 & (~0ULL << INTEL_PMC_IDX_FIXED)) {
639 idx = INTEL_PMC_IDX_FIXED;
640 for_each_set_bit_from(idx, c->idxmsk, X86_PMC_IDX_MAX) {
641 if (!__test_and_set_bit(idx, sched->state.used))
645 /* Grab the first unused counter starting with idx */
646 idx = sched->state.counter;
647 for_each_set_bit_from(idx, c->idxmsk, INTEL_PMC_IDX_FIXED) {
648 if (!__test_and_set_bit(idx, sched->state.used))
655 sched->state.counter = idx;
658 perf_sched_save_state(sched);
663 static bool perf_sched_find_counter(struct perf_sched *sched)
665 while (!__perf_sched_find_counter(sched)) {
666 if (!perf_sched_restore_state(sched))
674 * Go through all unassigned events and find the next one to schedule.
675 * Take events with the least weight first. Return true on success.
677 static bool perf_sched_next_event(struct perf_sched *sched)
679 struct event_constraint *c;
681 if (!sched->state.unassigned || !--sched->state.unassigned)
686 sched->state.event++;
687 if (sched->state.event >= sched->max_events) {
689 sched->state.event = 0;
690 sched->state.weight++;
691 if (sched->state.weight > sched->max_weight)
694 c = sched->events[sched->state.event]->hw.constraint;
695 } while (c->weight != sched->state.weight);
697 sched->state.counter = 0; /* start with first counter */
703 * Assign a counter for each event.
705 int perf_assign_events(struct perf_event **events, int n,
706 int wmin, int wmax, int *assign)
708 struct perf_sched sched;
710 perf_sched_init(&sched, events, n, wmin, wmax);
713 if (!perf_sched_find_counter(&sched))
716 assign[sched.state.event] = sched.state.counter;
717 } while (perf_sched_next_event(&sched));
719 return sched.state.unassigned;
721 EXPORT_SYMBOL_GPL(perf_assign_events);
723 int x86_schedule_events(struct cpu_hw_events *cpuc, int n, int *assign)
725 struct event_constraint *c;
726 unsigned long used_mask[BITS_TO_LONGS(X86_PMC_IDX_MAX)];
727 struct perf_event *e;
728 int i, wmin, wmax, num = 0;
729 struct hw_perf_event *hwc;
731 bitmap_zero(used_mask, X86_PMC_IDX_MAX);
733 for (i = 0, wmin = X86_PMC_IDX_MAX, wmax = 0; i < n; i++) {
734 hwc = &cpuc->event_list[i]->hw;
735 c = x86_pmu.get_event_constraints(cpuc, cpuc->event_list[i]);
738 wmin = min(wmin, c->weight);
739 wmax = max(wmax, c->weight);
743 * fastpath, try to reuse previous register
745 for (i = 0; i < n; i++) {
746 hwc = &cpuc->event_list[i]->hw;
753 /* constraint still honored */
754 if (!test_bit(hwc->idx, c->idxmsk))
757 /* not already used */
758 if (test_bit(hwc->idx, used_mask))
761 __set_bit(hwc->idx, used_mask);
763 assign[i] = hwc->idx;
768 num = perf_assign_events(cpuc->event_list, n, wmin,
772 * Mark the event as committed, so we do not put_constraint()
773 * in case new events are added and fail scheduling.
775 if (!num && assign) {
776 for (i = 0; i < n; i++) {
777 e = cpuc->event_list[i];
778 e->hw.flags |= PERF_X86_EVENT_COMMITTED;
782 * scheduling failed or is just a simulation,
783 * free resources if necessary
785 if (!assign || num) {
786 for (i = 0; i < n; i++) {
787 e = cpuc->event_list[i];
789 * do not put_constraint() on comitted events,
790 * because they are good to go
792 if ((e->hw.flags & PERF_X86_EVENT_COMMITTED))
795 if (x86_pmu.put_event_constraints)
796 x86_pmu.put_event_constraints(cpuc, e);
799 return num ? -EINVAL : 0;
803 * dogrp: true if must collect siblings events (group)
804 * returns total number of events and error code
806 static int collect_events(struct cpu_hw_events *cpuc, struct perf_event *leader, bool dogrp)
808 struct perf_event *event;
811 max_count = x86_pmu.num_counters + x86_pmu.num_counters_fixed;
813 /* current number of events already accepted */
816 if (is_x86_event(leader)) {
819 cpuc->event_list[n] = leader;
825 list_for_each_entry(event, &leader->sibling_list, group_entry) {
826 if (!is_x86_event(event) ||
827 event->state <= PERF_EVENT_STATE_OFF)
833 cpuc->event_list[n] = event;
839 static inline void x86_assign_hw_event(struct perf_event *event,
840 struct cpu_hw_events *cpuc, int i)
842 struct hw_perf_event *hwc = &event->hw;
844 hwc->idx = cpuc->assign[i];
845 hwc->last_cpu = smp_processor_id();
846 hwc->last_tag = ++cpuc->tags[i];
848 if (hwc->idx == INTEL_PMC_IDX_FIXED_BTS) {
849 hwc->config_base = 0;
851 } else if (hwc->idx >= INTEL_PMC_IDX_FIXED) {
852 hwc->config_base = MSR_ARCH_PERFMON_FIXED_CTR_CTRL;
853 hwc->event_base = MSR_ARCH_PERFMON_FIXED_CTR0 + (hwc->idx - INTEL_PMC_IDX_FIXED);
854 hwc->event_base_rdpmc = (hwc->idx - INTEL_PMC_IDX_FIXED) | 1<<30;
856 hwc->config_base = x86_pmu_config_addr(hwc->idx);
857 hwc->event_base = x86_pmu_event_addr(hwc->idx);
858 hwc->event_base_rdpmc = x86_pmu_rdpmc_index(hwc->idx);
862 static inline int match_prev_assignment(struct hw_perf_event *hwc,
863 struct cpu_hw_events *cpuc,
866 return hwc->idx == cpuc->assign[i] &&
867 hwc->last_cpu == smp_processor_id() &&
868 hwc->last_tag == cpuc->tags[i];
871 static void x86_pmu_start(struct perf_event *event, int flags);
873 static void x86_pmu_enable(struct pmu *pmu)
875 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
876 struct perf_event *event;
877 struct hw_perf_event *hwc;
878 int i, added = cpuc->n_added;
880 if (!x86_pmu_initialized())
887 int n_running = cpuc->n_events - cpuc->n_added;
889 * apply assignment obtained either from
890 * hw_perf_group_sched_in() or x86_pmu_enable()
892 * step1: save events moving to new counters
894 for (i = 0; i < n_running; i++) {
895 event = cpuc->event_list[i];
899 * we can avoid reprogramming counter if:
900 * - assigned same counter as last time
901 * - running on same CPU as last time
902 * - no other event has used the counter since
904 if (hwc->idx == -1 ||
905 match_prev_assignment(hwc, cpuc, i))
909 * Ensure we don't accidentally enable a stopped
910 * counter simply because we rescheduled.
912 if (hwc->state & PERF_HES_STOPPED)
913 hwc->state |= PERF_HES_ARCH;
915 x86_pmu_stop(event, PERF_EF_UPDATE);
919 * step2: reprogram moved events into new counters
921 for (i = 0; i < cpuc->n_events; i++) {
922 event = cpuc->event_list[i];
925 if (!match_prev_assignment(hwc, cpuc, i))
926 x86_assign_hw_event(event, cpuc, i);
927 else if (i < n_running)
930 if (hwc->state & PERF_HES_ARCH)
933 x86_pmu_start(event, PERF_EF_RELOAD);
936 perf_events_lapic_init();
942 x86_pmu.enable_all(added);
945 static DEFINE_PER_CPU(u64 [X86_PMC_IDX_MAX], pmc_prev_left);
948 * Set the next IRQ period, based on the hwc->period_left value.
949 * To be called with the event disabled in hw:
951 int x86_perf_event_set_period(struct perf_event *event)
953 struct hw_perf_event *hwc = &event->hw;
954 s64 left = local64_read(&hwc->period_left);
955 s64 period = hwc->sample_period;
956 int ret = 0, idx = hwc->idx;
958 if (idx == INTEL_PMC_IDX_FIXED_BTS)
962 * If we are way outside a reasonable range then just skip forward:
964 if (unlikely(left <= -period)) {
966 local64_set(&hwc->period_left, left);
967 hwc->last_period = period;
971 if (unlikely(left <= 0)) {
973 local64_set(&hwc->period_left, left);
974 hwc->last_period = period;
978 * Quirk: certain CPUs dont like it if just 1 hw_event is left:
980 if (unlikely(left < 2))
983 if (left > x86_pmu.max_period)
984 left = x86_pmu.max_period;
986 per_cpu(pmc_prev_left[idx], smp_processor_id()) = left;
989 * The hw event starts counting from this event offset,
990 * mark it to be able to extra future deltas:
992 local64_set(&hwc->prev_count, (u64)-left);
994 wrmsrl(hwc->event_base, (u64)(-left) & x86_pmu.cntval_mask);
997 * Due to erratum on certan cpu we need
998 * a second write to be sure the register
999 * is updated properly
1001 if (x86_pmu.perfctr_second_write) {
1002 wrmsrl(hwc->event_base,
1003 (u64)(-left) & x86_pmu.cntval_mask);
1006 perf_event_update_userpage(event);
1011 void x86_pmu_enable_event(struct perf_event *event)
1013 if (__this_cpu_read(cpu_hw_events.enabled))
1014 __x86_pmu_enable_event(&event->hw,
1015 ARCH_PERFMON_EVENTSEL_ENABLE);
1019 * Add a single event to the PMU.
1021 * The event is added to the group of enabled events
1022 * but only if it can be scehduled with existing events.
1024 static int x86_pmu_add(struct perf_event *event, int flags)
1026 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1027 struct hw_perf_event *hwc;
1028 int assign[X86_PMC_IDX_MAX];
1033 perf_pmu_disable(event->pmu);
1034 n0 = cpuc->n_events;
1035 ret = n = collect_events(cpuc, event, false);
1039 hwc->state = PERF_HES_UPTODATE | PERF_HES_STOPPED;
1040 if (!(flags & PERF_EF_START))
1041 hwc->state |= PERF_HES_ARCH;
1044 * If group events scheduling transaction was started,
1045 * skip the schedulability test here, it will be performed
1046 * at commit time (->commit_txn) as a whole.
1048 if (cpuc->group_flag & PERF_EVENT_TXN)
1051 ret = x86_pmu.schedule_events(cpuc, n, assign);
1055 * copy new assignment, now we know it is possible
1056 * will be used by hw_perf_enable()
1058 memcpy(cpuc->assign, assign, n*sizeof(int));
1062 * Commit the collect_events() state. See x86_pmu_del() and
1066 cpuc->n_added += n - n0;
1067 cpuc->n_txn += n - n0;
1071 perf_pmu_enable(event->pmu);
1075 static void x86_pmu_start(struct perf_event *event, int flags)
1077 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1078 int idx = event->hw.idx;
1080 if (WARN_ON_ONCE(!(event->hw.state & PERF_HES_STOPPED)))
1083 if (WARN_ON_ONCE(idx == -1))
1086 if (flags & PERF_EF_RELOAD) {
1087 WARN_ON_ONCE(!(event->hw.state & PERF_HES_UPTODATE));
1088 x86_perf_event_set_period(event);
1091 event->hw.state = 0;
1093 cpuc->events[idx] = event;
1094 __set_bit(idx, cpuc->active_mask);
1095 __set_bit(idx, cpuc->running);
1096 x86_pmu.enable(event);
1097 perf_event_update_userpage(event);
1100 void perf_event_print_debug(void)
1102 u64 ctrl, status, overflow, pmc_ctrl, pmc_count, prev_left, fixed;
1104 struct cpu_hw_events *cpuc;
1105 unsigned long flags;
1108 if (!x86_pmu.num_counters)
1111 local_irq_save(flags);
1113 cpu = smp_processor_id();
1114 cpuc = &per_cpu(cpu_hw_events, cpu);
1116 if (x86_pmu.version >= 2) {
1117 rdmsrl(MSR_CORE_PERF_GLOBAL_CTRL, ctrl);
1118 rdmsrl(MSR_CORE_PERF_GLOBAL_STATUS, status);
1119 rdmsrl(MSR_CORE_PERF_GLOBAL_OVF_CTRL, overflow);
1120 rdmsrl(MSR_ARCH_PERFMON_FIXED_CTR_CTRL, fixed);
1121 rdmsrl(MSR_IA32_PEBS_ENABLE, pebs);
1124 pr_info("CPU#%d: ctrl: %016llx\n", cpu, ctrl);
1125 pr_info("CPU#%d: status: %016llx\n", cpu, status);
1126 pr_info("CPU#%d: overflow: %016llx\n", cpu, overflow);
1127 pr_info("CPU#%d: fixed: %016llx\n", cpu, fixed);
1128 pr_info("CPU#%d: pebs: %016llx\n", cpu, pebs);
1130 pr_info("CPU#%d: active: %016llx\n", cpu, *(u64 *)cpuc->active_mask);
1132 for (idx = 0; idx < x86_pmu.num_counters; idx++) {
1133 rdmsrl(x86_pmu_config_addr(idx), pmc_ctrl);
1134 rdmsrl(x86_pmu_event_addr(idx), pmc_count);
1136 prev_left = per_cpu(pmc_prev_left[idx], cpu);
1138 pr_info("CPU#%d: gen-PMC%d ctrl: %016llx\n",
1139 cpu, idx, pmc_ctrl);
1140 pr_info("CPU#%d: gen-PMC%d count: %016llx\n",
1141 cpu, idx, pmc_count);
1142 pr_info("CPU#%d: gen-PMC%d left: %016llx\n",
1143 cpu, idx, prev_left);
1145 for (idx = 0; idx < x86_pmu.num_counters_fixed; idx++) {
1146 rdmsrl(MSR_ARCH_PERFMON_FIXED_CTR0 + idx, pmc_count);
1148 pr_info("CPU#%d: fixed-PMC%d count: %016llx\n",
1149 cpu, idx, pmc_count);
1151 local_irq_restore(flags);
1154 void x86_pmu_stop(struct perf_event *event, int flags)
1156 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1157 struct hw_perf_event *hwc = &event->hw;
1159 if (__test_and_clear_bit(hwc->idx, cpuc->active_mask)) {
1160 x86_pmu.disable(event);
1161 cpuc->events[hwc->idx] = NULL;
1162 WARN_ON_ONCE(hwc->state & PERF_HES_STOPPED);
1163 hwc->state |= PERF_HES_STOPPED;
1166 if ((flags & PERF_EF_UPDATE) && !(hwc->state & PERF_HES_UPTODATE)) {
1168 * Drain the remaining delta count out of a event
1169 * that we are disabling:
1171 x86_perf_event_update(event);
1172 hwc->state |= PERF_HES_UPTODATE;
1176 static void x86_pmu_del(struct perf_event *event, int flags)
1178 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1182 * event is descheduled
1184 event->hw.flags &= ~PERF_X86_EVENT_COMMITTED;
1187 * If we're called during a txn, we don't need to do anything.
1188 * The events never got scheduled and ->cancel_txn will truncate
1191 * XXX assumes any ->del() called during a TXN will only be on
1192 * an event added during that same TXN.
1194 if (cpuc->group_flag & PERF_EVENT_TXN)
1198 * Not a TXN, therefore cleanup properly.
1200 x86_pmu_stop(event, PERF_EF_UPDATE);
1202 for (i = 0; i < cpuc->n_events; i++) {
1203 if (event == cpuc->event_list[i])
1207 if (WARN_ON_ONCE(i == cpuc->n_events)) /* called ->del() without ->add() ? */
1210 /* If we have a newly added event; make sure to decrease n_added. */
1211 if (i >= cpuc->n_events - cpuc->n_added)
1214 if (x86_pmu.put_event_constraints)
1215 x86_pmu.put_event_constraints(cpuc, event);
1217 /* Delete the array entry. */
1218 while (++i < cpuc->n_events)
1219 cpuc->event_list[i-1] = cpuc->event_list[i];
1222 perf_event_update_userpage(event);
1225 int x86_pmu_handle_irq(struct pt_regs *regs)
1227 struct perf_sample_data data;
1228 struct cpu_hw_events *cpuc;
1229 struct perf_event *event;
1230 int idx, handled = 0;
1233 cpuc = this_cpu_ptr(&cpu_hw_events);
1236 * Some chipsets need to unmask the LVTPC in a particular spot
1237 * inside the nmi handler. As a result, the unmasking was pushed
1238 * into all the nmi handlers.
1240 * This generic handler doesn't seem to have any issues where the
1241 * unmasking occurs so it was left at the top.
1243 apic_write(APIC_LVTPC, APIC_DM_NMI);
1245 for (idx = 0; idx < x86_pmu.num_counters; idx++) {
1246 if (!test_bit(idx, cpuc->active_mask)) {
1248 * Though we deactivated the counter some cpus
1249 * might still deliver spurious interrupts still
1250 * in flight. Catch them:
1252 if (__test_and_clear_bit(idx, cpuc->running))
1257 event = cpuc->events[idx];
1259 val = x86_perf_event_update(event);
1260 if (val & (1ULL << (x86_pmu.cntval_bits - 1)))
1267 perf_sample_data_init(&data, 0, event->hw.last_period);
1269 if (!x86_perf_event_set_period(event))
1272 if (perf_event_overflow(event, &data, regs))
1273 x86_pmu_stop(event, 0);
1277 inc_irq_stat(apic_perf_irqs);
1282 void perf_events_lapic_init(void)
1284 if (!x86_pmu.apic || !x86_pmu_initialized())
1288 * Always use NMI for PMU
1290 apic_write(APIC_LVTPC, APIC_DM_NMI);
1294 perf_event_nmi_handler(unsigned int cmd, struct pt_regs *regs)
1300 if (!atomic_read(&active_events))
1303 start_clock = sched_clock();
1304 ret = x86_pmu.handle_irq(regs);
1305 finish_clock = sched_clock();
1307 perf_sample_event_took(finish_clock - start_clock);
1311 NOKPROBE_SYMBOL(perf_event_nmi_handler);
1313 struct event_constraint emptyconstraint;
1314 struct event_constraint unconstrained;
1317 x86_pmu_notifier(struct notifier_block *self, unsigned long action, void *hcpu)
1319 unsigned int cpu = (long)hcpu;
1320 struct cpu_hw_events *cpuc = &per_cpu(cpu_hw_events, cpu);
1321 int ret = NOTIFY_OK;
1323 switch (action & ~CPU_TASKS_FROZEN) {
1324 case CPU_UP_PREPARE:
1325 cpuc->kfree_on_online = NULL;
1326 if (x86_pmu.cpu_prepare)
1327 ret = x86_pmu.cpu_prepare(cpu);
1331 if (x86_pmu.attr_rdpmc)
1332 cr4_set_bits(X86_CR4_PCE);
1333 if (x86_pmu.cpu_starting)
1334 x86_pmu.cpu_starting(cpu);
1338 kfree(cpuc->kfree_on_online);
1342 if (x86_pmu.cpu_dying)
1343 x86_pmu.cpu_dying(cpu);
1346 case CPU_UP_CANCELED:
1348 if (x86_pmu.cpu_dead)
1349 x86_pmu.cpu_dead(cpu);
1359 static void __init pmu_check_apic(void)
1365 pr_info("no APIC, boot with the \"lapic\" boot parameter to force-enable it.\n");
1366 pr_info("no hardware sampling interrupt available.\n");
1369 * If we have a PMU initialized but no APIC
1370 * interrupts, we cannot sample hardware
1371 * events (user-space has to fall back and
1372 * sample via a hrtimer based software event):
1374 pmu.capabilities |= PERF_PMU_CAP_NO_INTERRUPT;
1378 static struct attribute_group x86_pmu_format_group = {
1384 * Remove all undefined events (x86_pmu.event_map(id) == 0)
1385 * out of events_attr attributes.
1387 static void __init filter_events(struct attribute **attrs)
1389 struct device_attribute *d;
1390 struct perf_pmu_events_attr *pmu_attr;
1393 for (i = 0; attrs[i]; i++) {
1394 d = (struct device_attribute *)attrs[i];
1395 pmu_attr = container_of(d, struct perf_pmu_events_attr, attr);
1397 if (pmu_attr->event_str)
1399 if (x86_pmu.event_map(i))
1402 for (j = i; attrs[j]; j++)
1403 attrs[j] = attrs[j + 1];
1405 /* Check the shifted attr. */
1410 /* Merge two pointer arrays */
1411 static __init struct attribute **merge_attr(struct attribute **a, struct attribute **b)
1413 struct attribute **new;
1416 for (j = 0; a[j]; j++)
1418 for (i = 0; b[i]; i++)
1422 new = kmalloc(sizeof(struct attribute *) * j, GFP_KERNEL);
1427 for (i = 0; a[i]; i++)
1429 for (i = 0; b[i]; i++)
1436 ssize_t events_sysfs_show(struct device *dev, struct device_attribute *attr,
1439 struct perf_pmu_events_attr *pmu_attr = \
1440 container_of(attr, struct perf_pmu_events_attr, attr);
1441 u64 config = x86_pmu.event_map(pmu_attr->id);
1443 /* string trumps id */
1444 if (pmu_attr->event_str)
1445 return sprintf(page, "%s", pmu_attr->event_str);
1447 return x86_pmu.events_sysfs_show(page, config);
1450 EVENT_ATTR(cpu-cycles, CPU_CYCLES );
1451 EVENT_ATTR(instructions, INSTRUCTIONS );
1452 EVENT_ATTR(cache-references, CACHE_REFERENCES );
1453 EVENT_ATTR(cache-misses, CACHE_MISSES );
1454 EVENT_ATTR(branch-instructions, BRANCH_INSTRUCTIONS );
1455 EVENT_ATTR(branch-misses, BRANCH_MISSES );
1456 EVENT_ATTR(bus-cycles, BUS_CYCLES );
1457 EVENT_ATTR(stalled-cycles-frontend, STALLED_CYCLES_FRONTEND );
1458 EVENT_ATTR(stalled-cycles-backend, STALLED_CYCLES_BACKEND );
1459 EVENT_ATTR(ref-cycles, REF_CPU_CYCLES );
1461 static struct attribute *empty_attrs;
1463 static struct attribute *events_attr[] = {
1464 EVENT_PTR(CPU_CYCLES),
1465 EVENT_PTR(INSTRUCTIONS),
1466 EVENT_PTR(CACHE_REFERENCES),
1467 EVENT_PTR(CACHE_MISSES),
1468 EVENT_PTR(BRANCH_INSTRUCTIONS),
1469 EVENT_PTR(BRANCH_MISSES),
1470 EVENT_PTR(BUS_CYCLES),
1471 EVENT_PTR(STALLED_CYCLES_FRONTEND),
1472 EVENT_PTR(STALLED_CYCLES_BACKEND),
1473 EVENT_PTR(REF_CPU_CYCLES),
1477 static struct attribute_group x86_pmu_events_group = {
1479 .attrs = events_attr,
1482 ssize_t x86_event_sysfs_show(char *page, u64 config, u64 event)
1484 u64 umask = (config & ARCH_PERFMON_EVENTSEL_UMASK) >> 8;
1485 u64 cmask = (config & ARCH_PERFMON_EVENTSEL_CMASK) >> 24;
1486 bool edge = (config & ARCH_PERFMON_EVENTSEL_EDGE);
1487 bool pc = (config & ARCH_PERFMON_EVENTSEL_PIN_CONTROL);
1488 bool any = (config & ARCH_PERFMON_EVENTSEL_ANY);
1489 bool inv = (config & ARCH_PERFMON_EVENTSEL_INV);
1493 * We have whole page size to spend and just little data
1494 * to write, so we can safely use sprintf.
1496 ret = sprintf(page, "event=0x%02llx", event);
1499 ret += sprintf(page + ret, ",umask=0x%02llx", umask);
1502 ret += sprintf(page + ret, ",edge");
1505 ret += sprintf(page + ret, ",pc");
1508 ret += sprintf(page + ret, ",any");
1511 ret += sprintf(page + ret, ",inv");
1514 ret += sprintf(page + ret, ",cmask=0x%02llx", cmask);
1516 ret += sprintf(page + ret, "\n");
1521 static int __init init_hw_perf_events(void)
1523 struct x86_pmu_quirk *quirk;
1526 pr_info("Performance Events: ");
1528 switch (boot_cpu_data.x86_vendor) {
1529 case X86_VENDOR_INTEL:
1530 err = intel_pmu_init();
1532 case X86_VENDOR_AMD:
1533 err = amd_pmu_init();
1539 pr_cont("no PMU driver, software events only.\n");
1545 /* sanity check that the hardware exists or is emulated */
1546 if (!check_hw_exists())
1549 pr_cont("%s PMU driver.\n", x86_pmu.name);
1551 x86_pmu.attr_rdpmc = 1; /* enable userspace RDPMC usage by default */
1553 for (quirk = x86_pmu.quirks; quirk; quirk = quirk->next)
1556 if (!x86_pmu.intel_ctrl)
1557 x86_pmu.intel_ctrl = (1 << x86_pmu.num_counters) - 1;
1559 perf_events_lapic_init();
1560 register_nmi_handler(NMI_LOCAL, perf_event_nmi_handler, 0, "PMI");
1562 unconstrained = (struct event_constraint)
1563 __EVENT_CONSTRAINT(0, (1ULL << x86_pmu.num_counters) - 1,
1564 0, x86_pmu.num_counters, 0, 0);
1566 x86_pmu_format_group.attrs = x86_pmu.format_attrs;
1568 if (x86_pmu.event_attrs)
1569 x86_pmu_events_group.attrs = x86_pmu.event_attrs;
1571 if (!x86_pmu.events_sysfs_show)
1572 x86_pmu_events_group.attrs = &empty_attrs;
1574 filter_events(x86_pmu_events_group.attrs);
1576 if (x86_pmu.cpu_events) {
1577 struct attribute **tmp;
1579 tmp = merge_attr(x86_pmu_events_group.attrs, x86_pmu.cpu_events);
1581 x86_pmu_events_group.attrs = tmp;
1584 pr_info("... version: %d\n", x86_pmu.version);
1585 pr_info("... bit width: %d\n", x86_pmu.cntval_bits);
1586 pr_info("... generic registers: %d\n", x86_pmu.num_counters);
1587 pr_info("... value mask: %016Lx\n", x86_pmu.cntval_mask);
1588 pr_info("... max period: %016Lx\n", x86_pmu.max_period);
1589 pr_info("... fixed-purpose events: %d\n", x86_pmu.num_counters_fixed);
1590 pr_info("... event mask: %016Lx\n", x86_pmu.intel_ctrl);
1592 perf_pmu_register(&pmu, "cpu", PERF_TYPE_RAW);
1593 perf_cpu_notifier(x86_pmu_notifier);
1597 early_initcall(init_hw_perf_events);
1599 static inline void x86_pmu_read(struct perf_event *event)
1601 x86_perf_event_update(event);
1605 * Start group events scheduling transaction
1606 * Set the flag to make pmu::enable() not perform the
1607 * schedulability test, it will be performed at commit time
1609 static void x86_pmu_start_txn(struct pmu *pmu)
1611 perf_pmu_disable(pmu);
1612 __this_cpu_or(cpu_hw_events.group_flag, PERF_EVENT_TXN);
1613 __this_cpu_write(cpu_hw_events.n_txn, 0);
1617 * Stop group events scheduling transaction
1618 * Clear the flag and pmu::enable() will perform the
1619 * schedulability test.
1621 static void x86_pmu_cancel_txn(struct pmu *pmu)
1623 __this_cpu_and(cpu_hw_events.group_flag, ~PERF_EVENT_TXN);
1625 * Truncate collected array by the number of events added in this
1626 * transaction. See x86_pmu_add() and x86_pmu_*_txn().
1628 __this_cpu_sub(cpu_hw_events.n_added, __this_cpu_read(cpu_hw_events.n_txn));
1629 __this_cpu_sub(cpu_hw_events.n_events, __this_cpu_read(cpu_hw_events.n_txn));
1630 perf_pmu_enable(pmu);
1634 * Commit group events scheduling transaction
1635 * Perform the group schedulability test as a whole
1636 * Return 0 if success
1638 * Does not cancel the transaction on failure; expects the caller to do this.
1640 static int x86_pmu_commit_txn(struct pmu *pmu)
1642 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1643 int assign[X86_PMC_IDX_MAX];
1648 if (!x86_pmu_initialized())
1651 ret = x86_pmu.schedule_events(cpuc, n, assign);
1656 * copy new assignment, now we know it is possible
1657 * will be used by hw_perf_enable()
1659 memcpy(cpuc->assign, assign, n*sizeof(int));
1661 cpuc->group_flag &= ~PERF_EVENT_TXN;
1662 perf_pmu_enable(pmu);
1666 * a fake_cpuc is used to validate event groups. Due to
1667 * the extra reg logic, we need to also allocate a fake
1668 * per_core and per_cpu structure. Otherwise, group events
1669 * using extra reg may conflict without the kernel being
1670 * able to catch this when the last event gets added to
1673 static void free_fake_cpuc(struct cpu_hw_events *cpuc)
1675 kfree(cpuc->shared_regs);
1679 static struct cpu_hw_events *allocate_fake_cpuc(void)
1681 struct cpu_hw_events *cpuc;
1682 int cpu = raw_smp_processor_id();
1684 cpuc = kzalloc(sizeof(*cpuc), GFP_KERNEL);
1686 return ERR_PTR(-ENOMEM);
1688 /* only needed, if we have extra_regs */
1689 if (x86_pmu.extra_regs) {
1690 cpuc->shared_regs = allocate_shared_regs(cpu);
1691 if (!cpuc->shared_regs)
1697 free_fake_cpuc(cpuc);
1698 return ERR_PTR(-ENOMEM);
1702 * validate that we can schedule this event
1704 static int validate_event(struct perf_event *event)
1706 struct cpu_hw_events *fake_cpuc;
1707 struct event_constraint *c;
1710 fake_cpuc = allocate_fake_cpuc();
1711 if (IS_ERR(fake_cpuc))
1712 return PTR_ERR(fake_cpuc);
1714 c = x86_pmu.get_event_constraints(fake_cpuc, event);
1716 if (!c || !c->weight)
1719 if (x86_pmu.put_event_constraints)
1720 x86_pmu.put_event_constraints(fake_cpuc, event);
1722 free_fake_cpuc(fake_cpuc);
1728 * validate a single event group
1730 * validation include:
1731 * - check events are compatible which each other
1732 * - events do not compete for the same counter
1733 * - number of events <= number of counters
1735 * validation ensures the group can be loaded onto the
1736 * PMU if it was the only group available.
1738 static int validate_group(struct perf_event *event)
1740 struct perf_event *leader = event->group_leader;
1741 struct cpu_hw_events *fake_cpuc;
1742 int ret = -EINVAL, n;
1744 fake_cpuc = allocate_fake_cpuc();
1745 if (IS_ERR(fake_cpuc))
1746 return PTR_ERR(fake_cpuc);
1748 * the event is not yet connected with its
1749 * siblings therefore we must first collect
1750 * existing siblings, then add the new event
1751 * before we can simulate the scheduling
1753 n = collect_events(fake_cpuc, leader, true);
1757 fake_cpuc->n_events = n;
1758 n = collect_events(fake_cpuc, event, false);
1762 fake_cpuc->n_events = n;
1764 ret = x86_pmu.schedule_events(fake_cpuc, n, NULL);
1767 free_fake_cpuc(fake_cpuc);
1771 static int x86_pmu_event_init(struct perf_event *event)
1776 switch (event->attr.type) {
1778 case PERF_TYPE_HARDWARE:
1779 case PERF_TYPE_HW_CACHE:
1786 err = __x86_pmu_event_init(event);
1789 * we temporarily connect event to its pmu
1790 * such that validate_group() can classify
1791 * it as an x86 event using is_x86_event()
1796 if (event->group_leader != event)
1797 err = validate_group(event);
1799 err = validate_event(event);
1805 event->destroy(event);
1811 static int x86_pmu_event_idx(struct perf_event *event)
1813 int idx = event->hw.idx;
1815 if (!x86_pmu.attr_rdpmc)
1818 if (x86_pmu.num_counters_fixed && idx >= INTEL_PMC_IDX_FIXED) {
1819 idx -= INTEL_PMC_IDX_FIXED;
1826 static ssize_t get_attr_rdpmc(struct device *cdev,
1827 struct device_attribute *attr,
1830 return snprintf(buf, 40, "%d\n", x86_pmu.attr_rdpmc);
1833 static void change_rdpmc(void *info)
1835 bool enable = !!(unsigned long)info;
1838 cr4_set_bits(X86_CR4_PCE);
1840 cr4_clear_bits(X86_CR4_PCE);
1843 static ssize_t set_attr_rdpmc(struct device *cdev,
1844 struct device_attribute *attr,
1845 const char *buf, size_t count)
1850 ret = kstrtoul(buf, 0, &val);
1854 if (x86_pmu.attr_rdpmc_broken)
1857 if (!!val != !!x86_pmu.attr_rdpmc) {
1858 x86_pmu.attr_rdpmc = !!val;
1859 on_each_cpu(change_rdpmc, (void *)val, 1);
1865 static DEVICE_ATTR(rdpmc, S_IRUSR | S_IWUSR, get_attr_rdpmc, set_attr_rdpmc);
1867 static struct attribute *x86_pmu_attrs[] = {
1868 &dev_attr_rdpmc.attr,
1872 static struct attribute_group x86_pmu_attr_group = {
1873 .attrs = x86_pmu_attrs,
1876 static const struct attribute_group *x86_pmu_attr_groups[] = {
1877 &x86_pmu_attr_group,
1878 &x86_pmu_format_group,
1879 &x86_pmu_events_group,
1883 static void x86_pmu_flush_branch_stack(void)
1885 if (x86_pmu.flush_branch_stack)
1886 x86_pmu.flush_branch_stack();
1889 void perf_check_microcode(void)
1891 if (x86_pmu.check_microcode)
1892 x86_pmu.check_microcode();
1894 EXPORT_SYMBOL_GPL(perf_check_microcode);
1896 static struct pmu pmu = {
1897 .pmu_enable = x86_pmu_enable,
1898 .pmu_disable = x86_pmu_disable,
1900 .attr_groups = x86_pmu_attr_groups,
1902 .event_init = x86_pmu_event_init,
1906 .start = x86_pmu_start,
1907 .stop = x86_pmu_stop,
1908 .read = x86_pmu_read,
1910 .start_txn = x86_pmu_start_txn,
1911 .cancel_txn = x86_pmu_cancel_txn,
1912 .commit_txn = x86_pmu_commit_txn,
1914 .event_idx = x86_pmu_event_idx,
1915 .flush_branch_stack = x86_pmu_flush_branch_stack,
1918 void arch_perf_update_userpage(struct perf_event_mmap_page *userpg, u64 now)
1920 struct cyc2ns_data *data;
1922 userpg->cap_user_time = 0;
1923 userpg->cap_user_time_zero = 0;
1924 userpg->cap_user_rdpmc = x86_pmu.attr_rdpmc;
1925 userpg->pmc_width = x86_pmu.cntval_bits;
1927 if (!sched_clock_stable())
1930 data = cyc2ns_read_begin();
1932 userpg->cap_user_time = 1;
1933 userpg->time_mult = data->cyc2ns_mul;
1934 userpg->time_shift = data->cyc2ns_shift;
1935 userpg->time_offset = data->cyc2ns_offset - now;
1937 userpg->cap_user_time_zero = 1;
1938 userpg->time_zero = data->cyc2ns_offset;
1940 cyc2ns_read_end(data);
1947 static int backtrace_stack(void *data, char *name)
1952 static void backtrace_address(void *data, unsigned long addr, int reliable)
1954 struct perf_callchain_entry *entry = data;
1956 perf_callchain_store(entry, addr);
1959 static const struct stacktrace_ops backtrace_ops = {
1960 .stack = backtrace_stack,
1961 .address = backtrace_address,
1962 .walk_stack = print_context_stack_bp,
1966 perf_callchain_kernel(struct perf_callchain_entry *entry, struct pt_regs *regs)
1968 if (perf_guest_cbs && perf_guest_cbs->is_in_guest()) {
1969 /* TODO: We don't support guest os callchain now */
1973 perf_callchain_store(entry, regs->ip);
1975 dump_trace(NULL, regs, NULL, 0, &backtrace_ops, entry);
1979 valid_user_frame(const void __user *fp, unsigned long size)
1981 return (__range_not_ok(fp, size, TASK_SIZE) == 0);
1984 static unsigned long get_segment_base(unsigned int segment)
1986 struct desc_struct *desc;
1987 int idx = segment >> 3;
1989 if ((segment & SEGMENT_TI_MASK) == SEGMENT_LDT) {
1990 if (idx > LDT_ENTRIES)
1993 if (idx > current->active_mm->context.size)
1996 desc = current->active_mm->context.ldt;
1998 if (idx > GDT_ENTRIES)
2001 desc = raw_cpu_ptr(gdt_page.gdt);
2004 return get_desc_base(desc + idx);
2007 #ifdef CONFIG_COMPAT
2009 #include <asm/compat.h>
2012 perf_callchain_user32(struct pt_regs *regs, struct perf_callchain_entry *entry)
2014 /* 32-bit process in 64-bit kernel. */
2015 unsigned long ss_base, cs_base;
2016 struct stack_frame_ia32 frame;
2017 const void __user *fp;
2019 if (!test_thread_flag(TIF_IA32))
2022 cs_base = get_segment_base(regs->cs);
2023 ss_base = get_segment_base(regs->ss);
2025 fp = compat_ptr(ss_base + regs->bp);
2026 while (entry->nr < PERF_MAX_STACK_DEPTH) {
2027 unsigned long bytes;
2028 frame.next_frame = 0;
2029 frame.return_address = 0;
2031 bytes = copy_from_user_nmi(&frame, fp, sizeof(frame));
2035 if (!valid_user_frame(fp, sizeof(frame)))
2038 perf_callchain_store(entry, cs_base + frame.return_address);
2039 fp = compat_ptr(ss_base + frame.next_frame);
2045 perf_callchain_user32(struct pt_regs *regs, struct perf_callchain_entry *entry)
2052 perf_callchain_user(struct perf_callchain_entry *entry, struct pt_regs *regs)
2054 struct stack_frame frame;
2055 const void __user *fp;
2057 if (perf_guest_cbs && perf_guest_cbs->is_in_guest()) {
2058 /* TODO: We don't support guest os callchain now */
2063 * We don't know what to do with VM86 stacks.. ignore them for now.
2065 if (regs->flags & (X86_VM_MASK | PERF_EFLAGS_VM))
2068 fp = (void __user *)regs->bp;
2070 perf_callchain_store(entry, regs->ip);
2075 if (perf_callchain_user32(regs, entry))
2078 while (entry->nr < PERF_MAX_STACK_DEPTH) {
2079 unsigned long bytes;
2080 frame.next_frame = NULL;
2081 frame.return_address = 0;
2083 bytes = copy_from_user_nmi(&frame, fp, sizeof(frame));
2087 if (!valid_user_frame(fp, sizeof(frame)))
2090 perf_callchain_store(entry, frame.return_address);
2091 fp = frame.next_frame;
2096 * Deal with code segment offsets for the various execution modes:
2098 * VM86 - the good olde 16 bit days, where the linear address is
2099 * 20 bits and we use regs->ip + 0x10 * regs->cs.
2101 * IA32 - Where we need to look at GDT/LDT segment descriptor tables
2102 * to figure out what the 32bit base address is.
2104 * X32 - has TIF_X32 set, but is running in x86_64
2106 * X86_64 - CS,DS,SS,ES are all zero based.
2108 static unsigned long code_segment_base(struct pt_regs *regs)
2111 * If we are in VM86 mode, add the segment offset to convert to a
2114 if (regs->flags & X86_VM_MASK)
2115 return 0x10 * regs->cs;
2118 * For IA32 we look at the GDT/LDT segment base to convert the
2119 * effective IP to a linear address.
2121 #ifdef CONFIG_X86_32
2122 if (user_mode(regs) && regs->cs != __USER_CS)
2123 return get_segment_base(regs->cs);
2125 if (test_thread_flag(TIF_IA32)) {
2126 if (user_mode(regs) && regs->cs != __USER32_CS)
2127 return get_segment_base(regs->cs);
2133 unsigned long perf_instruction_pointer(struct pt_regs *regs)
2135 if (perf_guest_cbs && perf_guest_cbs->is_in_guest())
2136 return perf_guest_cbs->get_guest_ip();
2138 return regs->ip + code_segment_base(regs);
2141 unsigned long perf_misc_flags(struct pt_regs *regs)
2145 if (perf_guest_cbs && perf_guest_cbs->is_in_guest()) {
2146 if (perf_guest_cbs->is_user_mode())
2147 misc |= PERF_RECORD_MISC_GUEST_USER;
2149 misc |= PERF_RECORD_MISC_GUEST_KERNEL;
2151 if (user_mode(regs))
2152 misc |= PERF_RECORD_MISC_USER;
2154 misc |= PERF_RECORD_MISC_KERNEL;
2157 if (regs->flags & PERF_EFLAGS_EXACT)
2158 misc |= PERF_RECORD_MISC_EXACT_IP;
2163 void perf_get_x86_pmu_capability(struct x86_pmu_capability *cap)
2165 cap->version = x86_pmu.version;
2166 cap->num_counters_gp = x86_pmu.num_counters;
2167 cap->num_counters_fixed = x86_pmu.num_counters_fixed;
2168 cap->bit_width_gp = x86_pmu.cntval_bits;
2169 cap->bit_width_fixed = x86_pmu.cntval_bits;
2170 cap->events_mask = (unsigned int)x86_pmu.events_maskl;
2171 cap->events_mask_len = x86_pmu.events_mask_len;
2173 EXPORT_SYMBOL_GPL(perf_get_x86_pmu_capability);