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/highmem.h>
26 #include <linux/cpu.h>
27 #include <linux/bitops.h>
30 #include <asm/stacktrace.h>
32 #include <asm/compat.h>
36 #define wrmsrl(msr, val) \
38 trace_printk("wrmsrl(%lx, %lx)\n", (unsigned long)(msr),\
39 (unsigned long)(val)); \
40 native_write_msr((msr), (u32)((u64)(val)), \
41 (u32)((u64)(val) >> 32)); \
46 * best effort, GUP based copy_from_user() that assumes IRQ or NMI context
49 copy_from_user_nmi(void *to, const void __user *from, unsigned long n)
51 unsigned long offset, addr = (unsigned long)from;
52 unsigned long size, len = 0;
58 ret = __get_user_pages_fast(addr, 1, 0, &page);
62 offset = addr & (PAGE_SIZE - 1);
63 size = min(PAGE_SIZE - offset, n - len);
65 map = kmap_atomic(page);
66 memcpy(to, map+offset, size);
79 struct event_constraint {
81 unsigned long idxmsk[BITS_TO_LONGS(X86_PMC_IDX_MAX)];
90 int nb_id; /* NorthBridge id */
91 int refcnt; /* reference count */
92 struct perf_event *owners[X86_PMC_IDX_MAX];
93 struct event_constraint event_constraints[X86_PMC_IDX_MAX];
96 #define MAX_LBR_ENTRIES 16
98 struct cpu_hw_events {
100 * Generic x86 PMC bits
102 struct perf_event *events[X86_PMC_IDX_MAX]; /* in counter order */
103 unsigned long active_mask[BITS_TO_LONGS(X86_PMC_IDX_MAX)];
104 unsigned long running[BITS_TO_LONGS(X86_PMC_IDX_MAX)];
110 int assign[X86_PMC_IDX_MAX]; /* event to counter assignment */
111 u64 tags[X86_PMC_IDX_MAX];
112 struct perf_event *event_list[X86_PMC_IDX_MAX]; /* in enabled order */
114 unsigned int group_flag;
117 * Intel DebugStore bits
119 struct debug_store *ds;
127 struct perf_branch_stack lbr_stack;
128 struct perf_branch_entry lbr_entries[MAX_LBR_ENTRIES];
133 struct amd_nb *amd_nb;
136 #define __EVENT_CONSTRAINT(c, n, m, w) {\
137 { .idxmsk64 = (n) }, \
143 #define EVENT_CONSTRAINT(c, n, m) \
144 __EVENT_CONSTRAINT(c, n, m, HWEIGHT(n))
147 * Constraint on the Event code.
149 #define INTEL_EVENT_CONSTRAINT(c, n) \
150 EVENT_CONSTRAINT(c, n, ARCH_PERFMON_EVENTSEL_EVENT)
153 * Constraint on the Event code + UMask + fixed-mask
155 * filter mask to validate fixed counter events.
156 * the following filters disqualify for fixed counters:
160 * The other filters are supported by fixed counters.
161 * The any-thread option is supported starting with v3.
163 #define FIXED_EVENT_CONSTRAINT(c, n) \
164 EVENT_CONSTRAINT(c, (1ULL << (32+n)), X86_RAW_EVENT_MASK)
167 * Constraint on the Event code + UMask
169 #define PEBS_EVENT_CONSTRAINT(c, n) \
170 EVENT_CONSTRAINT(c, n, INTEL_ARCH_EVENT_MASK)
172 #define EVENT_CONSTRAINT_END \
173 EVENT_CONSTRAINT(0, 0, 0)
175 #define for_each_event_constraint(e, c) \
176 for ((e) = (c); (e)->weight; (e)++)
178 union perf_capabilities {
182 u64 pebs_arch_reg : 1;
190 * struct x86_pmu - generic x86 pmu
194 * Generic x86 PMC bits
198 int (*handle_irq)(struct pt_regs *);
199 void (*disable_all)(void);
200 void (*enable_all)(int added);
201 void (*enable)(struct perf_event *);
202 void (*disable)(struct perf_event *);
203 int (*hw_config)(struct perf_event *event);
204 int (*schedule_events)(struct cpu_hw_events *cpuc, int n, int *assign);
207 u64 (*event_map)(int);
210 int num_counters_fixed;
215 struct event_constraint *
216 (*get_event_constraints)(struct cpu_hw_events *cpuc,
217 struct perf_event *event);
219 void (*put_event_constraints)(struct cpu_hw_events *cpuc,
220 struct perf_event *event);
221 struct event_constraint *event_constraints;
222 void (*quirks)(void);
223 int perfctr_second_write;
225 int (*cpu_prepare)(int cpu);
226 void (*cpu_starting)(int cpu);
227 void (*cpu_dying)(int cpu);
228 void (*cpu_dead)(int cpu);
231 * Intel Arch Perfmon v2+
234 union perf_capabilities intel_cap;
237 * Intel DebugStore bits
240 int bts_active, pebs_active;
241 int pebs_record_size;
242 void (*drain_pebs)(struct pt_regs *regs);
243 struct event_constraint *pebs_constraints;
248 unsigned long lbr_tos, lbr_from, lbr_to; /* MSR base regs */
249 int lbr_nr; /* hardware stack size */
252 static struct x86_pmu x86_pmu __read_mostly;
254 static DEFINE_PER_CPU(struct cpu_hw_events, cpu_hw_events) = {
258 static int x86_perf_event_set_period(struct perf_event *event);
261 * Generalized hw caching related hw_event table, filled
262 * in on a per model basis. A value of 0 means
263 * 'not supported', -1 means 'hw_event makes no sense on
264 * this CPU', any other value means the raw hw_event
268 #define C(x) PERF_COUNT_HW_CACHE_##x
270 static u64 __read_mostly hw_cache_event_ids
271 [PERF_COUNT_HW_CACHE_MAX]
272 [PERF_COUNT_HW_CACHE_OP_MAX]
273 [PERF_COUNT_HW_CACHE_RESULT_MAX];
276 * Propagate event elapsed time into the generic event.
277 * Can only be executed on the CPU where the event is active.
278 * Returns the delta events processed.
281 x86_perf_event_update(struct perf_event *event)
283 struct hw_perf_event *hwc = &event->hw;
284 int shift = 64 - x86_pmu.cntval_bits;
285 u64 prev_raw_count, new_raw_count;
289 if (idx == X86_PMC_IDX_FIXED_BTS)
293 * Careful: an NMI might modify the previous event value.
295 * Our tactic to handle this is to first atomically read and
296 * exchange a new raw count - then add that new-prev delta
297 * count to the generic event atomically:
300 prev_raw_count = local64_read(&hwc->prev_count);
301 rdmsrl(hwc->event_base + idx, new_raw_count);
303 if (local64_cmpxchg(&hwc->prev_count, prev_raw_count,
304 new_raw_count) != prev_raw_count)
308 * Now we have the new raw value and have updated the prev
309 * timestamp already. We can now calculate the elapsed delta
310 * (event-)time and add that to the generic event.
312 * Careful, not all hw sign-extends above the physical width
315 delta = (new_raw_count << shift) - (prev_raw_count << shift);
318 local64_add(delta, &event->count);
319 local64_sub(delta, &hwc->period_left);
321 return new_raw_count;
324 static atomic_t active_events;
325 static DEFINE_MUTEX(pmc_reserve_mutex);
327 #ifdef CONFIG_X86_LOCAL_APIC
329 static bool reserve_pmc_hardware(void)
333 for (i = 0; i < x86_pmu.num_counters; i++) {
334 if (!reserve_perfctr_nmi(x86_pmu.perfctr + i))
338 for (i = 0; i < x86_pmu.num_counters; i++) {
339 if (!reserve_evntsel_nmi(x86_pmu.eventsel + i))
346 for (i--; i >= 0; i--)
347 release_evntsel_nmi(x86_pmu.eventsel + i);
349 i = x86_pmu.num_counters;
352 for (i--; i >= 0; i--)
353 release_perfctr_nmi(x86_pmu.perfctr + i);
358 static void release_pmc_hardware(void)
362 for (i = 0; i < x86_pmu.num_counters; i++) {
363 release_perfctr_nmi(x86_pmu.perfctr + i);
364 release_evntsel_nmi(x86_pmu.eventsel + i);
370 static bool reserve_pmc_hardware(void) { return true; }
371 static void release_pmc_hardware(void) {}
375 static bool check_hw_exists(void)
377 u64 val, val_new = 0;
381 * Check to see if the BIOS enabled any of the counters, if so
384 for (i = 0; i < x86_pmu.num_counters; i++) {
385 reg = x86_pmu.eventsel + i;
386 ret = rdmsrl_safe(reg, &val);
389 if (val & ARCH_PERFMON_EVENTSEL_ENABLE)
393 if (x86_pmu.num_counters_fixed) {
394 reg = MSR_ARCH_PERFMON_FIXED_CTR_CTRL;
395 ret = rdmsrl_safe(reg, &val);
398 for (i = 0; i < x86_pmu.num_counters_fixed; i++) {
399 if (val & (0x03 << i*4))
405 * Now write a value and read it back to see if it matches,
406 * this is needed to detect certain hardware emulators (qemu/kvm)
407 * that don't trap on the MSR access and always return 0s.
410 ret = checking_wrmsrl(x86_pmu.perfctr, val);
411 ret |= rdmsrl_safe(x86_pmu.perfctr, &val_new);
412 if (ret || val != val_new)
418 printk(KERN_CONT "Broken BIOS detected, using software events only.\n");
419 printk(KERN_ERR FW_BUG "the BIOS has corrupted hw-PMU resources (MSR %x is %Lx)\n", reg, val);
423 printk(KERN_CONT "Broken PMU hardware detected, using software events only.\n");
427 static void reserve_ds_buffers(void);
428 static void release_ds_buffers(void);
430 static void hw_perf_event_destroy(struct perf_event *event)
432 if (atomic_dec_and_mutex_lock(&active_events, &pmc_reserve_mutex)) {
433 release_pmc_hardware();
434 release_ds_buffers();
435 mutex_unlock(&pmc_reserve_mutex);
439 static inline int x86_pmu_initialized(void)
441 return x86_pmu.handle_irq != NULL;
445 set_ext_hw_attr(struct hw_perf_event *hwc, struct perf_event_attr *attr)
447 unsigned int cache_type, cache_op, cache_result;
450 config = attr->config;
452 cache_type = (config >> 0) & 0xff;
453 if (cache_type >= PERF_COUNT_HW_CACHE_MAX)
456 cache_op = (config >> 8) & 0xff;
457 if (cache_op >= PERF_COUNT_HW_CACHE_OP_MAX)
460 cache_result = (config >> 16) & 0xff;
461 if (cache_result >= PERF_COUNT_HW_CACHE_RESULT_MAX)
464 val = hw_cache_event_ids[cache_type][cache_op][cache_result];
477 static int x86_setup_perfctr(struct perf_event *event)
479 struct perf_event_attr *attr = &event->attr;
480 struct hw_perf_event *hwc = &event->hw;
483 if (!is_sampling_event(event)) {
484 hwc->sample_period = x86_pmu.max_period;
485 hwc->last_period = hwc->sample_period;
486 local64_set(&hwc->period_left, hwc->sample_period);
489 * If we have a PMU initialized but no APIC
490 * interrupts, we cannot sample hardware
491 * events (user-space has to fall back and
492 * sample via a hrtimer based software event):
498 if (attr->type == PERF_TYPE_RAW)
501 if (attr->type == PERF_TYPE_HW_CACHE)
502 return set_ext_hw_attr(hwc, attr);
504 if (attr->config >= x86_pmu.max_events)
510 config = x86_pmu.event_map(attr->config);
521 if ((attr->config == PERF_COUNT_HW_BRANCH_INSTRUCTIONS) &&
522 (hwc->sample_period == 1)) {
523 /* BTS is not supported by this architecture. */
524 if (!x86_pmu.bts_active)
527 /* BTS is currently only allowed for user-mode. */
528 if (!attr->exclude_kernel)
532 hwc->config |= config;
537 static int x86_pmu_hw_config(struct perf_event *event)
539 if (event->attr.precise_ip) {
542 /* Support for constant skid */
543 if (x86_pmu.pebs_active) {
546 /* Support for IP fixup */
551 if (event->attr.precise_ip > precise)
557 * (keep 'enabled' bit clear for now)
559 event->hw.config = ARCH_PERFMON_EVENTSEL_INT;
562 * Count user and OS events unless requested not to
564 if (!event->attr.exclude_user)
565 event->hw.config |= ARCH_PERFMON_EVENTSEL_USR;
566 if (!event->attr.exclude_kernel)
567 event->hw.config |= ARCH_PERFMON_EVENTSEL_OS;
569 if (event->attr.type == PERF_TYPE_RAW)
570 event->hw.config |= event->attr.config & X86_RAW_EVENT_MASK;
572 return x86_setup_perfctr(event);
576 * Setup the hardware configuration for a given attr_type
578 static int __x86_pmu_event_init(struct perf_event *event)
582 if (!x86_pmu_initialized())
586 if (!atomic_inc_not_zero(&active_events)) {
587 mutex_lock(&pmc_reserve_mutex);
588 if (atomic_read(&active_events) == 0) {
589 if (!reserve_pmc_hardware())
592 reserve_ds_buffers();
595 atomic_inc(&active_events);
596 mutex_unlock(&pmc_reserve_mutex);
601 event->destroy = hw_perf_event_destroy;
604 event->hw.last_cpu = -1;
605 event->hw.last_tag = ~0ULL;
607 return x86_pmu.hw_config(event);
610 static void x86_pmu_disable_all(void)
612 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
615 for (idx = 0; idx < x86_pmu.num_counters; idx++) {
618 if (!test_bit(idx, cpuc->active_mask))
620 rdmsrl(x86_pmu.eventsel + idx, val);
621 if (!(val & ARCH_PERFMON_EVENTSEL_ENABLE))
623 val &= ~ARCH_PERFMON_EVENTSEL_ENABLE;
624 wrmsrl(x86_pmu.eventsel + idx, val);
628 static void x86_pmu_disable(struct pmu *pmu)
630 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
632 if (!x86_pmu_initialized())
642 x86_pmu.disable_all();
645 static inline void __x86_pmu_enable_event(struct hw_perf_event *hwc,
648 wrmsrl(hwc->config_base + hwc->idx, hwc->config | enable_mask);
651 static void x86_pmu_enable_all(int added)
653 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
656 for (idx = 0; idx < x86_pmu.num_counters; idx++) {
657 struct hw_perf_event *hwc = &cpuc->events[idx]->hw;
659 if (!test_bit(idx, cpuc->active_mask))
662 __x86_pmu_enable_event(hwc, ARCH_PERFMON_EVENTSEL_ENABLE);
666 static struct pmu pmu;
668 static inline int is_x86_event(struct perf_event *event)
670 return event->pmu == &pmu;
673 static int x86_schedule_events(struct cpu_hw_events *cpuc, int n, int *assign)
675 struct event_constraint *c, *constraints[X86_PMC_IDX_MAX];
676 unsigned long used_mask[BITS_TO_LONGS(X86_PMC_IDX_MAX)];
677 int i, j, w, wmax, num = 0;
678 struct hw_perf_event *hwc;
680 bitmap_zero(used_mask, X86_PMC_IDX_MAX);
682 for (i = 0; i < n; i++) {
683 c = x86_pmu.get_event_constraints(cpuc, cpuc->event_list[i]);
688 * fastpath, try to reuse previous register
690 for (i = 0; i < n; i++) {
691 hwc = &cpuc->event_list[i]->hw;
698 /* constraint still honored */
699 if (!test_bit(hwc->idx, c->idxmsk))
702 /* not already used */
703 if (test_bit(hwc->idx, used_mask))
706 __set_bit(hwc->idx, used_mask);
708 assign[i] = hwc->idx;
717 bitmap_zero(used_mask, X86_PMC_IDX_MAX);
720 * weight = number of possible counters
722 * 1 = most constrained, only works on one counter
723 * wmax = least constrained, works on any counter
725 * assign events to counters starting with most
726 * constrained events.
728 wmax = x86_pmu.num_counters;
731 * when fixed event counters are present,
732 * wmax is incremented by 1 to account
733 * for one more choice
735 if (x86_pmu.num_counters_fixed)
738 for (w = 1, num = n; num && w <= wmax; w++) {
740 for (i = 0; num && i < n; i++) {
742 hwc = &cpuc->event_list[i]->hw;
747 for_each_set_bit(j, c->idxmsk, X86_PMC_IDX_MAX) {
748 if (!test_bit(j, used_mask))
752 if (j == X86_PMC_IDX_MAX)
755 __set_bit(j, used_mask);
764 * scheduling failed or is just a simulation,
765 * free resources if necessary
767 if (!assign || num) {
768 for (i = 0; i < n; i++) {
769 if (x86_pmu.put_event_constraints)
770 x86_pmu.put_event_constraints(cpuc, cpuc->event_list[i]);
773 return num ? -ENOSPC : 0;
777 * dogrp: true if must collect siblings events (group)
778 * returns total number of events and error code
780 static int collect_events(struct cpu_hw_events *cpuc, struct perf_event *leader, bool dogrp)
782 struct perf_event *event;
785 max_count = x86_pmu.num_counters + x86_pmu.num_counters_fixed;
787 /* current number of events already accepted */
790 if (is_x86_event(leader)) {
793 cpuc->event_list[n] = leader;
799 list_for_each_entry(event, &leader->sibling_list, group_entry) {
800 if (!is_x86_event(event) ||
801 event->state <= PERF_EVENT_STATE_OFF)
807 cpuc->event_list[n] = event;
813 static inline void x86_assign_hw_event(struct perf_event *event,
814 struct cpu_hw_events *cpuc, int i)
816 struct hw_perf_event *hwc = &event->hw;
818 hwc->idx = cpuc->assign[i];
819 hwc->last_cpu = smp_processor_id();
820 hwc->last_tag = ++cpuc->tags[i];
822 if (hwc->idx == X86_PMC_IDX_FIXED_BTS) {
823 hwc->config_base = 0;
825 } else if (hwc->idx >= X86_PMC_IDX_FIXED) {
826 hwc->config_base = MSR_ARCH_PERFMON_FIXED_CTR_CTRL;
828 * We set it so that event_base + idx in wrmsr/rdmsr maps to
829 * MSR_ARCH_PERFMON_FIXED_CTR0 ... CTR2:
832 MSR_ARCH_PERFMON_FIXED_CTR0 - X86_PMC_IDX_FIXED;
834 hwc->config_base = x86_pmu.eventsel;
835 hwc->event_base = x86_pmu.perfctr;
839 static inline int match_prev_assignment(struct hw_perf_event *hwc,
840 struct cpu_hw_events *cpuc,
843 return hwc->idx == cpuc->assign[i] &&
844 hwc->last_cpu == smp_processor_id() &&
845 hwc->last_tag == cpuc->tags[i];
848 static void x86_pmu_start(struct perf_event *event, int flags);
849 static void x86_pmu_stop(struct perf_event *event, int flags);
851 static void x86_pmu_enable(struct pmu *pmu)
853 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
854 struct perf_event *event;
855 struct hw_perf_event *hwc;
856 int i, added = cpuc->n_added;
858 if (!x86_pmu_initialized())
865 int n_running = cpuc->n_events - cpuc->n_added;
867 * apply assignment obtained either from
868 * hw_perf_group_sched_in() or x86_pmu_enable()
870 * step1: save events moving to new counters
871 * step2: reprogram moved events into new counters
873 for (i = 0; i < n_running; i++) {
874 event = cpuc->event_list[i];
878 * we can avoid reprogramming counter if:
879 * - assigned same counter as last time
880 * - running on same CPU as last time
881 * - no other event has used the counter since
883 if (hwc->idx == -1 ||
884 match_prev_assignment(hwc, cpuc, i))
888 * Ensure we don't accidentally enable a stopped
889 * counter simply because we rescheduled.
891 if (hwc->state & PERF_HES_STOPPED)
892 hwc->state |= PERF_HES_ARCH;
894 x86_pmu_stop(event, PERF_EF_UPDATE);
897 for (i = 0; i < cpuc->n_events; i++) {
898 event = cpuc->event_list[i];
901 if (!match_prev_assignment(hwc, cpuc, i))
902 x86_assign_hw_event(event, cpuc, i);
903 else if (i < n_running)
906 if (hwc->state & PERF_HES_ARCH)
909 x86_pmu_start(event, PERF_EF_RELOAD);
912 perf_events_lapic_init();
918 x86_pmu.enable_all(added);
921 static inline void x86_pmu_disable_event(struct perf_event *event)
923 struct hw_perf_event *hwc = &event->hw;
925 wrmsrl(hwc->config_base + hwc->idx, hwc->config);
928 static DEFINE_PER_CPU(u64 [X86_PMC_IDX_MAX], pmc_prev_left);
931 * Set the next IRQ period, based on the hwc->period_left value.
932 * To be called with the event disabled in hw:
935 x86_perf_event_set_period(struct perf_event *event)
937 struct hw_perf_event *hwc = &event->hw;
938 s64 left = local64_read(&hwc->period_left);
939 s64 period = hwc->sample_period;
940 int ret = 0, idx = hwc->idx;
942 if (idx == X86_PMC_IDX_FIXED_BTS)
946 * If we are way outside a reasonable range then just skip forward:
948 if (unlikely(left <= -period)) {
950 local64_set(&hwc->period_left, left);
951 hwc->last_period = period;
955 if (unlikely(left <= 0)) {
957 local64_set(&hwc->period_left, left);
958 hwc->last_period = period;
962 * Quirk: certain CPUs dont like it if just 1 hw_event is left:
964 if (unlikely(left < 2))
967 if (left > x86_pmu.max_period)
968 left = x86_pmu.max_period;
970 per_cpu(pmc_prev_left[idx], smp_processor_id()) = left;
973 * The hw event starts counting from this event offset,
974 * mark it to be able to extra future deltas:
976 local64_set(&hwc->prev_count, (u64)-left);
978 wrmsrl(hwc->event_base + idx, (u64)(-left) & x86_pmu.cntval_mask);
981 * Due to erratum on certan cpu we need
982 * a second write to be sure the register
983 * is updated properly
985 if (x86_pmu.perfctr_second_write) {
986 wrmsrl(hwc->event_base + idx,
987 (u64)(-left) & x86_pmu.cntval_mask);
990 perf_event_update_userpage(event);
995 static void x86_pmu_enable_event(struct perf_event *event)
997 if (__this_cpu_read(cpu_hw_events.enabled))
998 __x86_pmu_enable_event(&event->hw,
999 ARCH_PERFMON_EVENTSEL_ENABLE);
1003 * Add a single event to the PMU.
1005 * The event is added to the group of enabled events
1006 * but only if it can be scehduled with existing events.
1008 static int x86_pmu_add(struct perf_event *event, int flags)
1010 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
1011 struct hw_perf_event *hwc;
1012 int assign[X86_PMC_IDX_MAX];
1017 perf_pmu_disable(event->pmu);
1018 n0 = cpuc->n_events;
1019 ret = n = collect_events(cpuc, event, false);
1023 hwc->state = PERF_HES_UPTODATE | PERF_HES_STOPPED;
1024 if (!(flags & PERF_EF_START))
1025 hwc->state |= PERF_HES_ARCH;
1028 * If group events scheduling transaction was started,
1029 * skip the schedulability test here, it will be peformed
1030 * at commit time (->commit_txn) as a whole
1032 if (cpuc->group_flag & PERF_EVENT_TXN)
1035 ret = x86_pmu.schedule_events(cpuc, n, assign);
1039 * copy new assignment, now we know it is possible
1040 * will be used by hw_perf_enable()
1042 memcpy(cpuc->assign, assign, n*sizeof(int));
1046 cpuc->n_added += n - n0;
1047 cpuc->n_txn += n - n0;
1051 perf_pmu_enable(event->pmu);
1055 static void x86_pmu_start(struct perf_event *event, int flags)
1057 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
1058 int idx = event->hw.idx;
1060 if (WARN_ON_ONCE(!(event->hw.state & PERF_HES_STOPPED)))
1063 if (WARN_ON_ONCE(idx == -1))
1066 if (flags & PERF_EF_RELOAD) {
1067 WARN_ON_ONCE(!(event->hw.state & PERF_HES_UPTODATE));
1068 x86_perf_event_set_period(event);
1071 event->hw.state = 0;
1073 cpuc->events[idx] = event;
1074 __set_bit(idx, cpuc->active_mask);
1075 __set_bit(idx, cpuc->running);
1076 x86_pmu.enable(event);
1077 perf_event_update_userpage(event);
1080 void perf_event_print_debug(void)
1082 u64 ctrl, status, overflow, pmc_ctrl, pmc_count, prev_left, fixed;
1084 struct cpu_hw_events *cpuc;
1085 unsigned long flags;
1088 if (!x86_pmu.num_counters)
1091 local_irq_save(flags);
1093 cpu = smp_processor_id();
1094 cpuc = &per_cpu(cpu_hw_events, cpu);
1096 if (x86_pmu.version >= 2) {
1097 rdmsrl(MSR_CORE_PERF_GLOBAL_CTRL, ctrl);
1098 rdmsrl(MSR_CORE_PERF_GLOBAL_STATUS, status);
1099 rdmsrl(MSR_CORE_PERF_GLOBAL_OVF_CTRL, overflow);
1100 rdmsrl(MSR_ARCH_PERFMON_FIXED_CTR_CTRL, fixed);
1101 rdmsrl(MSR_IA32_PEBS_ENABLE, pebs);
1104 pr_info("CPU#%d: ctrl: %016llx\n", cpu, ctrl);
1105 pr_info("CPU#%d: status: %016llx\n", cpu, status);
1106 pr_info("CPU#%d: overflow: %016llx\n", cpu, overflow);
1107 pr_info("CPU#%d: fixed: %016llx\n", cpu, fixed);
1108 pr_info("CPU#%d: pebs: %016llx\n", cpu, pebs);
1110 pr_info("CPU#%d: active: %016llx\n", cpu, *(u64 *)cpuc->active_mask);
1112 for (idx = 0; idx < x86_pmu.num_counters; idx++) {
1113 rdmsrl(x86_pmu.eventsel + idx, pmc_ctrl);
1114 rdmsrl(x86_pmu.perfctr + idx, pmc_count);
1116 prev_left = per_cpu(pmc_prev_left[idx], cpu);
1118 pr_info("CPU#%d: gen-PMC%d ctrl: %016llx\n",
1119 cpu, idx, pmc_ctrl);
1120 pr_info("CPU#%d: gen-PMC%d count: %016llx\n",
1121 cpu, idx, pmc_count);
1122 pr_info("CPU#%d: gen-PMC%d left: %016llx\n",
1123 cpu, idx, prev_left);
1125 for (idx = 0; idx < x86_pmu.num_counters_fixed; idx++) {
1126 rdmsrl(MSR_ARCH_PERFMON_FIXED_CTR0 + idx, pmc_count);
1128 pr_info("CPU#%d: fixed-PMC%d count: %016llx\n",
1129 cpu, idx, pmc_count);
1131 local_irq_restore(flags);
1134 static void x86_pmu_stop(struct perf_event *event, int flags)
1136 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
1137 struct hw_perf_event *hwc = &event->hw;
1139 if (__test_and_clear_bit(hwc->idx, cpuc->active_mask)) {
1140 x86_pmu.disable(event);
1141 cpuc->events[hwc->idx] = NULL;
1142 WARN_ON_ONCE(hwc->state & PERF_HES_STOPPED);
1143 hwc->state |= PERF_HES_STOPPED;
1146 if ((flags & PERF_EF_UPDATE) && !(hwc->state & PERF_HES_UPTODATE)) {
1148 * Drain the remaining delta count out of a event
1149 * that we are disabling:
1151 x86_perf_event_update(event);
1152 hwc->state |= PERF_HES_UPTODATE;
1156 static void x86_pmu_del(struct perf_event *event, int flags)
1158 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
1162 * If we're called during a txn, we don't need to do anything.
1163 * The events never got scheduled and ->cancel_txn will truncate
1166 if (cpuc->group_flag & PERF_EVENT_TXN)
1169 x86_pmu_stop(event, PERF_EF_UPDATE);
1171 for (i = 0; i < cpuc->n_events; i++) {
1172 if (event == cpuc->event_list[i]) {
1174 if (x86_pmu.put_event_constraints)
1175 x86_pmu.put_event_constraints(cpuc, event);
1177 while (++i < cpuc->n_events)
1178 cpuc->event_list[i-1] = cpuc->event_list[i];
1184 perf_event_update_userpage(event);
1187 static int x86_pmu_handle_irq(struct pt_regs *regs)
1189 struct perf_sample_data data;
1190 struct cpu_hw_events *cpuc;
1191 struct perf_event *event;
1192 int idx, handled = 0;
1195 perf_sample_data_init(&data, 0);
1197 cpuc = &__get_cpu_var(cpu_hw_events);
1199 for (idx = 0; idx < x86_pmu.num_counters; idx++) {
1200 if (!test_bit(idx, cpuc->active_mask)) {
1202 * Though we deactivated the counter some cpus
1203 * might still deliver spurious interrupts still
1204 * in flight. Catch them:
1206 if (__test_and_clear_bit(idx, cpuc->running))
1211 event = cpuc->events[idx];
1213 val = x86_perf_event_update(event);
1214 if (val & (1ULL << (x86_pmu.cntval_bits - 1)))
1221 data.period = event->hw.last_period;
1223 if (!x86_perf_event_set_period(event))
1226 if (perf_event_overflow(event, 1, &data, regs))
1227 x86_pmu_stop(event, 0);
1231 inc_irq_stat(apic_perf_irqs);
1236 void perf_events_lapic_init(void)
1238 if (!x86_pmu.apic || !x86_pmu_initialized())
1242 * Always use NMI for PMU
1244 apic_write(APIC_LVTPC, APIC_DM_NMI);
1247 struct pmu_nmi_state {
1248 unsigned int marked;
1252 static DEFINE_PER_CPU(struct pmu_nmi_state, pmu_nmi);
1254 static int __kprobes
1255 perf_event_nmi_handler(struct notifier_block *self,
1256 unsigned long cmd, void *__args)
1258 struct die_args *args = __args;
1259 unsigned int this_nmi;
1262 if (!atomic_read(&active_events))
1268 case DIE_NMIUNKNOWN:
1269 this_nmi = percpu_read(irq_stat.__nmi_count);
1270 if (this_nmi != __this_cpu_read(pmu_nmi.marked))
1271 /* let the kernel handle the unknown nmi */
1274 * This one is a PMU back-to-back nmi. Two events
1275 * trigger 'simultaneously' raising two back-to-back
1276 * NMIs. If the first NMI handles both, the latter
1277 * will be empty and daze the CPU. So, we drop it to
1278 * avoid false-positive 'unknown nmi' messages.
1285 apic_write(APIC_LVTPC, APIC_DM_NMI);
1287 handled = x86_pmu.handle_irq(args->regs);
1291 this_nmi = percpu_read(irq_stat.__nmi_count);
1292 if ((handled > 1) ||
1293 /* the next nmi could be a back-to-back nmi */
1294 ((__this_cpu_read(pmu_nmi.marked) == this_nmi) &&
1295 (__this_cpu_read(pmu_nmi.handled) > 1))) {
1297 * We could have two subsequent back-to-back nmis: The
1298 * first handles more than one counter, the 2nd
1299 * handles only one counter and the 3rd handles no
1302 * This is the 2nd nmi because the previous was
1303 * handling more than one counter. We will mark the
1304 * next (3rd) and then drop it if unhandled.
1306 __this_cpu_write(pmu_nmi.marked, this_nmi + 1);
1307 __this_cpu_write(pmu_nmi.handled, handled);
1313 static __read_mostly struct notifier_block perf_event_nmi_notifier = {
1314 .notifier_call = perf_event_nmi_handler,
1316 .priority = NMI_LOCAL_LOW_PRIOR,
1319 static struct event_constraint unconstrained;
1320 static struct event_constraint emptyconstraint;
1322 static struct event_constraint *
1323 x86_get_event_constraints(struct cpu_hw_events *cpuc, struct perf_event *event)
1325 struct event_constraint *c;
1327 if (x86_pmu.event_constraints) {
1328 for_each_event_constraint(c, x86_pmu.event_constraints) {
1329 if ((event->hw.config & c->cmask) == c->code)
1334 return &unconstrained;
1337 #include "perf_event_amd.c"
1338 #include "perf_event_p6.c"
1339 #include "perf_event_p4.c"
1340 #include "perf_event_intel_lbr.c"
1341 #include "perf_event_intel_ds.c"
1342 #include "perf_event_intel.c"
1344 static int __cpuinit
1345 x86_pmu_notifier(struct notifier_block *self, unsigned long action, void *hcpu)
1347 unsigned int cpu = (long)hcpu;
1348 int ret = NOTIFY_OK;
1350 switch (action & ~CPU_TASKS_FROZEN) {
1351 case CPU_UP_PREPARE:
1352 if (x86_pmu.cpu_prepare)
1353 ret = x86_pmu.cpu_prepare(cpu);
1357 if (x86_pmu.cpu_starting)
1358 x86_pmu.cpu_starting(cpu);
1362 if (x86_pmu.cpu_dying)
1363 x86_pmu.cpu_dying(cpu);
1366 case CPU_UP_CANCELED:
1368 if (x86_pmu.cpu_dead)
1369 x86_pmu.cpu_dead(cpu);
1379 static void __init pmu_check_apic(void)
1385 pr_info("no APIC, boot with the \"lapic\" boot parameter to force-enable it.\n");
1386 pr_info("no hardware sampling interrupt available.\n");
1389 static int __init init_hw_perf_events(void)
1391 struct event_constraint *c;
1394 pr_info("Performance Events: ");
1396 switch (boot_cpu_data.x86_vendor) {
1397 case X86_VENDOR_INTEL:
1398 err = intel_pmu_init();
1400 case X86_VENDOR_AMD:
1401 err = amd_pmu_init();
1407 pr_cont("no PMU driver, software events only.\n");
1413 /* sanity check that the hardware exists or is emulated */
1414 if (!check_hw_exists())
1417 pr_cont("%s PMU driver.\n", x86_pmu.name);
1422 if (x86_pmu.num_counters > X86_PMC_MAX_GENERIC) {
1423 WARN(1, KERN_ERR "hw perf events %d > max(%d), clipping!",
1424 x86_pmu.num_counters, X86_PMC_MAX_GENERIC);
1425 x86_pmu.num_counters = X86_PMC_MAX_GENERIC;
1427 x86_pmu.intel_ctrl = (1 << x86_pmu.num_counters) - 1;
1429 if (x86_pmu.num_counters_fixed > X86_PMC_MAX_FIXED) {
1430 WARN(1, KERN_ERR "hw perf events fixed %d > max(%d), clipping!",
1431 x86_pmu.num_counters_fixed, X86_PMC_MAX_FIXED);
1432 x86_pmu.num_counters_fixed = X86_PMC_MAX_FIXED;
1435 x86_pmu.intel_ctrl |=
1436 ((1LL << x86_pmu.num_counters_fixed)-1) << X86_PMC_IDX_FIXED;
1438 perf_events_lapic_init();
1439 register_die_notifier(&perf_event_nmi_notifier);
1441 unconstrained = (struct event_constraint)
1442 __EVENT_CONSTRAINT(0, (1ULL << x86_pmu.num_counters) - 1,
1443 0, x86_pmu.num_counters);
1445 if (x86_pmu.event_constraints) {
1446 for_each_event_constraint(c, x86_pmu.event_constraints) {
1447 if (c->cmask != X86_RAW_EVENT_MASK)
1450 c->idxmsk64 |= (1ULL << x86_pmu.num_counters) - 1;
1451 c->weight += x86_pmu.num_counters;
1455 pr_info("... version: %d\n", x86_pmu.version);
1456 pr_info("... bit width: %d\n", x86_pmu.cntval_bits);
1457 pr_info("... generic registers: %d\n", x86_pmu.num_counters);
1458 pr_info("... value mask: %016Lx\n", x86_pmu.cntval_mask);
1459 pr_info("... max period: %016Lx\n", x86_pmu.max_period);
1460 pr_info("... fixed-purpose events: %d\n", x86_pmu.num_counters_fixed);
1461 pr_info("... event mask: %016Lx\n", x86_pmu.intel_ctrl);
1463 perf_pmu_register(&pmu, "cpu", PERF_TYPE_RAW);
1464 perf_cpu_notifier(x86_pmu_notifier);
1468 early_initcall(init_hw_perf_events);
1470 static inline void x86_pmu_read(struct perf_event *event)
1472 x86_perf_event_update(event);
1476 * Start group events scheduling transaction
1477 * Set the flag to make pmu::enable() not perform the
1478 * schedulability test, it will be performed at commit time
1480 static void x86_pmu_start_txn(struct pmu *pmu)
1482 perf_pmu_disable(pmu);
1483 __this_cpu_or(cpu_hw_events.group_flag, PERF_EVENT_TXN);
1484 __this_cpu_write(cpu_hw_events.n_txn, 0);
1488 * Stop group events scheduling transaction
1489 * Clear the flag and pmu::enable() will perform the
1490 * schedulability test.
1492 static void x86_pmu_cancel_txn(struct pmu *pmu)
1494 __this_cpu_and(cpu_hw_events.group_flag, ~PERF_EVENT_TXN);
1496 * Truncate the collected events.
1498 __this_cpu_sub(cpu_hw_events.n_added, __this_cpu_read(cpu_hw_events.n_txn));
1499 __this_cpu_sub(cpu_hw_events.n_events, __this_cpu_read(cpu_hw_events.n_txn));
1500 perf_pmu_enable(pmu);
1504 * Commit group events scheduling transaction
1505 * Perform the group schedulability test as a whole
1506 * Return 0 if success
1508 static int x86_pmu_commit_txn(struct pmu *pmu)
1510 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
1511 int assign[X86_PMC_IDX_MAX];
1516 if (!x86_pmu_initialized())
1519 ret = x86_pmu.schedule_events(cpuc, n, assign);
1524 * copy new assignment, now we know it is possible
1525 * will be used by hw_perf_enable()
1527 memcpy(cpuc->assign, assign, n*sizeof(int));
1529 cpuc->group_flag &= ~PERF_EVENT_TXN;
1530 perf_pmu_enable(pmu);
1535 * validate that we can schedule this event
1537 static int validate_event(struct perf_event *event)
1539 struct cpu_hw_events *fake_cpuc;
1540 struct event_constraint *c;
1543 fake_cpuc = kmalloc(sizeof(*fake_cpuc), GFP_KERNEL | __GFP_ZERO);
1547 c = x86_pmu.get_event_constraints(fake_cpuc, event);
1549 if (!c || !c->weight)
1552 if (x86_pmu.put_event_constraints)
1553 x86_pmu.put_event_constraints(fake_cpuc, event);
1561 * validate a single event group
1563 * validation include:
1564 * - check events are compatible which each other
1565 * - events do not compete for the same counter
1566 * - number of events <= number of counters
1568 * validation ensures the group can be loaded onto the
1569 * PMU if it was the only group available.
1571 static int validate_group(struct perf_event *event)
1573 struct perf_event *leader = event->group_leader;
1574 struct cpu_hw_events *fake_cpuc;
1578 fake_cpuc = kmalloc(sizeof(*fake_cpuc), GFP_KERNEL | __GFP_ZERO);
1583 * the event is not yet connected with its
1584 * siblings therefore we must first collect
1585 * existing siblings, then add the new event
1586 * before we can simulate the scheduling
1589 n = collect_events(fake_cpuc, leader, true);
1593 fake_cpuc->n_events = n;
1594 n = collect_events(fake_cpuc, event, false);
1598 fake_cpuc->n_events = n;
1600 ret = x86_pmu.schedule_events(fake_cpuc, n, NULL);
1608 static int x86_pmu_event_init(struct perf_event *event)
1613 switch (event->attr.type) {
1615 case PERF_TYPE_HARDWARE:
1616 case PERF_TYPE_HW_CACHE:
1623 err = __x86_pmu_event_init(event);
1626 * we temporarily connect event to its pmu
1627 * such that validate_group() can classify
1628 * it as an x86 event using is_x86_event()
1633 if (event->group_leader != event)
1634 err = validate_group(event);
1636 err = validate_event(event);
1642 event->destroy(event);
1648 static struct pmu pmu = {
1649 .pmu_enable = x86_pmu_enable,
1650 .pmu_disable = x86_pmu_disable,
1652 .event_init = x86_pmu_event_init,
1656 .start = x86_pmu_start,
1657 .stop = x86_pmu_stop,
1658 .read = x86_pmu_read,
1660 .start_txn = x86_pmu_start_txn,
1661 .cancel_txn = x86_pmu_cancel_txn,
1662 .commit_txn = x86_pmu_commit_txn,
1670 backtrace_warning_symbol(void *data, char *msg, unsigned long symbol)
1672 /* Ignore warnings */
1675 static void backtrace_warning(void *data, char *msg)
1677 /* Ignore warnings */
1680 static int backtrace_stack(void *data, char *name)
1685 static void backtrace_address(void *data, unsigned long addr, int reliable)
1687 struct perf_callchain_entry *entry = data;
1689 perf_callchain_store(entry, addr);
1692 static const struct stacktrace_ops backtrace_ops = {
1693 .warning = backtrace_warning,
1694 .warning_symbol = backtrace_warning_symbol,
1695 .stack = backtrace_stack,
1696 .address = backtrace_address,
1697 .walk_stack = print_context_stack_bp,
1701 perf_callchain_kernel(struct perf_callchain_entry *entry, struct pt_regs *regs)
1703 if (perf_guest_cbs && perf_guest_cbs->is_in_guest()) {
1704 /* TODO: We don't support guest os callchain now */
1708 perf_callchain_store(entry, regs->ip);
1710 dump_trace(NULL, regs, NULL, &backtrace_ops, entry);
1713 #ifdef CONFIG_COMPAT
1715 perf_callchain_user32(struct pt_regs *regs, struct perf_callchain_entry *entry)
1717 /* 32-bit process in 64-bit kernel. */
1718 struct stack_frame_ia32 frame;
1719 const void __user *fp;
1721 if (!test_thread_flag(TIF_IA32))
1724 fp = compat_ptr(regs->bp);
1725 while (entry->nr < PERF_MAX_STACK_DEPTH) {
1726 unsigned long bytes;
1727 frame.next_frame = 0;
1728 frame.return_address = 0;
1730 bytes = copy_from_user_nmi(&frame, fp, sizeof(frame));
1731 if (bytes != sizeof(frame))
1734 if (fp < compat_ptr(regs->sp))
1737 perf_callchain_store(entry, frame.return_address);
1738 fp = compat_ptr(frame.next_frame);
1744 perf_callchain_user32(struct pt_regs *regs, struct perf_callchain_entry *entry)
1751 perf_callchain_user(struct perf_callchain_entry *entry, struct pt_regs *regs)
1753 struct stack_frame frame;
1754 const void __user *fp;
1756 if (perf_guest_cbs && perf_guest_cbs->is_in_guest()) {
1757 /* TODO: We don't support guest os callchain now */
1761 fp = (void __user *)regs->bp;
1763 perf_callchain_store(entry, regs->ip);
1765 if (perf_callchain_user32(regs, entry))
1768 while (entry->nr < PERF_MAX_STACK_DEPTH) {
1769 unsigned long bytes;
1770 frame.next_frame = NULL;
1771 frame.return_address = 0;
1773 bytes = copy_from_user_nmi(&frame, fp, sizeof(frame));
1774 if (bytes != sizeof(frame))
1777 if ((unsigned long)fp < regs->sp)
1780 perf_callchain_store(entry, frame.return_address);
1781 fp = frame.next_frame;
1785 unsigned long perf_instruction_pointer(struct pt_regs *regs)
1789 if (perf_guest_cbs && perf_guest_cbs->is_in_guest())
1790 ip = perf_guest_cbs->get_guest_ip();
1792 ip = instruction_pointer(regs);
1797 unsigned long perf_misc_flags(struct pt_regs *regs)
1801 if (perf_guest_cbs && perf_guest_cbs->is_in_guest()) {
1802 if (perf_guest_cbs->is_user_mode())
1803 misc |= PERF_RECORD_MISC_GUEST_USER;
1805 misc |= PERF_RECORD_MISC_GUEST_KERNEL;
1807 if (user_mode(regs))
1808 misc |= PERF_RECORD_MISC_USER;
1810 misc |= PERF_RECORD_MISC_KERNEL;
1813 if (regs->flags & PERF_EFLAGS_EXACT)
1814 misc |= PERF_RECORD_MISC_EXACT_IP;