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 void reserve_ds_buffers(void);
376 static void release_ds_buffers(void);
378 static void hw_perf_event_destroy(struct perf_event *event)
380 if (atomic_dec_and_mutex_lock(&active_events, &pmc_reserve_mutex)) {
381 release_pmc_hardware();
382 release_ds_buffers();
383 mutex_unlock(&pmc_reserve_mutex);
387 static inline int x86_pmu_initialized(void)
389 return x86_pmu.handle_irq != NULL;
393 set_ext_hw_attr(struct hw_perf_event *hwc, struct perf_event_attr *attr)
395 unsigned int cache_type, cache_op, cache_result;
398 config = attr->config;
400 cache_type = (config >> 0) & 0xff;
401 if (cache_type >= PERF_COUNT_HW_CACHE_MAX)
404 cache_op = (config >> 8) & 0xff;
405 if (cache_op >= PERF_COUNT_HW_CACHE_OP_MAX)
408 cache_result = (config >> 16) & 0xff;
409 if (cache_result >= PERF_COUNT_HW_CACHE_RESULT_MAX)
412 val = hw_cache_event_ids[cache_type][cache_op][cache_result];
425 static int x86_setup_perfctr(struct perf_event *event)
427 struct perf_event_attr *attr = &event->attr;
428 struct hw_perf_event *hwc = &event->hw;
431 if (!hwc->sample_period) {
432 hwc->sample_period = x86_pmu.max_period;
433 hwc->last_period = hwc->sample_period;
434 local64_set(&hwc->period_left, hwc->sample_period);
437 * If we have a PMU initialized but no APIC
438 * interrupts, we cannot sample hardware
439 * events (user-space has to fall back and
440 * sample via a hrtimer based software event):
446 if (attr->type == PERF_TYPE_RAW)
449 if (attr->type == PERF_TYPE_HW_CACHE)
450 return set_ext_hw_attr(hwc, attr);
452 if (attr->config >= x86_pmu.max_events)
458 config = x86_pmu.event_map(attr->config);
469 if ((attr->config == PERF_COUNT_HW_BRANCH_INSTRUCTIONS) &&
470 (hwc->sample_period == 1)) {
471 /* BTS is not supported by this architecture. */
472 if (!x86_pmu.bts_active)
475 /* BTS is currently only allowed for user-mode. */
476 if (!attr->exclude_kernel)
480 hwc->config |= config;
485 static int x86_pmu_hw_config(struct perf_event *event)
487 if (event->attr.precise_ip) {
490 /* Support for constant skid */
491 if (x86_pmu.pebs_active) {
494 /* Support for IP fixup */
499 if (event->attr.precise_ip > precise)
505 * (keep 'enabled' bit clear for now)
507 event->hw.config = ARCH_PERFMON_EVENTSEL_INT;
510 * Count user and OS events unless requested not to
512 if (!event->attr.exclude_user)
513 event->hw.config |= ARCH_PERFMON_EVENTSEL_USR;
514 if (!event->attr.exclude_kernel)
515 event->hw.config |= ARCH_PERFMON_EVENTSEL_OS;
517 if (event->attr.type == PERF_TYPE_RAW)
518 event->hw.config |= event->attr.config & X86_RAW_EVENT_MASK;
520 return x86_setup_perfctr(event);
524 * Setup the hardware configuration for a given attr_type
526 static int __x86_pmu_event_init(struct perf_event *event)
530 if (!x86_pmu_initialized())
534 if (!atomic_inc_not_zero(&active_events)) {
535 mutex_lock(&pmc_reserve_mutex);
536 if (atomic_read(&active_events) == 0) {
537 if (!reserve_pmc_hardware())
540 reserve_ds_buffers();
543 atomic_inc(&active_events);
544 mutex_unlock(&pmc_reserve_mutex);
549 event->destroy = hw_perf_event_destroy;
552 event->hw.last_cpu = -1;
553 event->hw.last_tag = ~0ULL;
555 return x86_pmu.hw_config(event);
558 static void x86_pmu_disable_all(void)
560 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
563 for (idx = 0; idx < x86_pmu.num_counters; idx++) {
566 if (!test_bit(idx, cpuc->active_mask))
568 rdmsrl(x86_pmu.eventsel + idx, val);
569 if (!(val & ARCH_PERFMON_EVENTSEL_ENABLE))
571 val &= ~ARCH_PERFMON_EVENTSEL_ENABLE;
572 wrmsrl(x86_pmu.eventsel + idx, val);
576 static void x86_pmu_disable(struct pmu *pmu)
578 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
580 if (!x86_pmu_initialized())
590 x86_pmu.disable_all();
593 static void x86_pmu_enable_all(int added)
595 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
598 for (idx = 0; idx < x86_pmu.num_counters; idx++) {
599 struct perf_event *event = cpuc->events[idx];
602 if (!test_bit(idx, cpuc->active_mask))
605 val = event->hw.config;
606 val |= ARCH_PERFMON_EVENTSEL_ENABLE;
607 wrmsrl(x86_pmu.eventsel + idx, val);
611 static struct pmu pmu;
613 static inline int is_x86_event(struct perf_event *event)
615 return event->pmu == &pmu;
618 static int x86_schedule_events(struct cpu_hw_events *cpuc, int n, int *assign)
620 struct event_constraint *c, *constraints[X86_PMC_IDX_MAX];
621 unsigned long used_mask[BITS_TO_LONGS(X86_PMC_IDX_MAX)];
622 int i, j, w, wmax, num = 0;
623 struct hw_perf_event *hwc;
625 bitmap_zero(used_mask, X86_PMC_IDX_MAX);
627 for (i = 0; i < n; i++) {
628 c = x86_pmu.get_event_constraints(cpuc, cpuc->event_list[i]);
633 * fastpath, try to reuse previous register
635 for (i = 0; i < n; i++) {
636 hwc = &cpuc->event_list[i]->hw;
643 /* constraint still honored */
644 if (!test_bit(hwc->idx, c->idxmsk))
647 /* not already used */
648 if (test_bit(hwc->idx, used_mask))
651 __set_bit(hwc->idx, used_mask);
653 assign[i] = hwc->idx;
662 bitmap_zero(used_mask, X86_PMC_IDX_MAX);
665 * weight = number of possible counters
667 * 1 = most constrained, only works on one counter
668 * wmax = least constrained, works on any counter
670 * assign events to counters starting with most
671 * constrained events.
673 wmax = x86_pmu.num_counters;
676 * when fixed event counters are present,
677 * wmax is incremented by 1 to account
678 * for one more choice
680 if (x86_pmu.num_counters_fixed)
683 for (w = 1, num = n; num && w <= wmax; w++) {
685 for (i = 0; num && i < n; i++) {
687 hwc = &cpuc->event_list[i]->hw;
692 for_each_set_bit(j, c->idxmsk, X86_PMC_IDX_MAX) {
693 if (!test_bit(j, used_mask))
697 if (j == X86_PMC_IDX_MAX)
700 __set_bit(j, used_mask);
709 * scheduling failed or is just a simulation,
710 * free resources if necessary
712 if (!assign || num) {
713 for (i = 0; i < n; i++) {
714 if (x86_pmu.put_event_constraints)
715 x86_pmu.put_event_constraints(cpuc, cpuc->event_list[i]);
718 return num ? -ENOSPC : 0;
722 * dogrp: true if must collect siblings events (group)
723 * returns total number of events and error code
725 static int collect_events(struct cpu_hw_events *cpuc, struct perf_event *leader, bool dogrp)
727 struct perf_event *event;
730 max_count = x86_pmu.num_counters + x86_pmu.num_counters_fixed;
732 /* current number of events already accepted */
735 if (is_x86_event(leader)) {
738 cpuc->event_list[n] = leader;
744 list_for_each_entry(event, &leader->sibling_list, group_entry) {
745 if (!is_x86_event(event) ||
746 event->state <= PERF_EVENT_STATE_OFF)
752 cpuc->event_list[n] = event;
758 static inline void x86_assign_hw_event(struct perf_event *event,
759 struct cpu_hw_events *cpuc, int i)
761 struct hw_perf_event *hwc = &event->hw;
763 hwc->idx = cpuc->assign[i];
764 hwc->last_cpu = smp_processor_id();
765 hwc->last_tag = ++cpuc->tags[i];
767 if (hwc->idx == X86_PMC_IDX_FIXED_BTS) {
768 hwc->config_base = 0;
770 } else if (hwc->idx >= X86_PMC_IDX_FIXED) {
771 hwc->config_base = MSR_ARCH_PERFMON_FIXED_CTR_CTRL;
773 * We set it so that event_base + idx in wrmsr/rdmsr maps to
774 * MSR_ARCH_PERFMON_FIXED_CTR0 ... CTR2:
777 MSR_ARCH_PERFMON_FIXED_CTR0 - X86_PMC_IDX_FIXED;
779 hwc->config_base = x86_pmu.eventsel;
780 hwc->event_base = x86_pmu.perfctr;
784 static inline int match_prev_assignment(struct hw_perf_event *hwc,
785 struct cpu_hw_events *cpuc,
788 return hwc->idx == cpuc->assign[i] &&
789 hwc->last_cpu == smp_processor_id() &&
790 hwc->last_tag == cpuc->tags[i];
793 static void x86_pmu_start(struct perf_event *event, int flags);
794 static void x86_pmu_stop(struct perf_event *event, int flags);
796 static void x86_pmu_enable(struct pmu *pmu)
798 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
799 struct perf_event *event;
800 struct hw_perf_event *hwc;
801 int i, added = cpuc->n_added;
803 if (!x86_pmu_initialized())
810 int n_running = cpuc->n_events - cpuc->n_added;
812 * apply assignment obtained either from
813 * hw_perf_group_sched_in() or x86_pmu_enable()
815 * step1: save events moving to new counters
816 * step2: reprogram moved events into new counters
818 for (i = 0; i < n_running; i++) {
819 event = cpuc->event_list[i];
823 * we can avoid reprogramming counter if:
824 * - assigned same counter as last time
825 * - running on same CPU as last time
826 * - no other event has used the counter since
828 if (hwc->idx == -1 ||
829 match_prev_assignment(hwc, cpuc, i))
833 * Ensure we don't accidentally enable a stopped
834 * counter simply because we rescheduled.
836 if (hwc->state & PERF_HES_STOPPED)
837 hwc->state |= PERF_HES_ARCH;
839 x86_pmu_stop(event, PERF_EF_UPDATE);
842 for (i = 0; i < cpuc->n_events; i++) {
843 event = cpuc->event_list[i];
846 if (!match_prev_assignment(hwc, cpuc, i))
847 x86_assign_hw_event(event, cpuc, i);
848 else if (i < n_running)
851 if (hwc->state & PERF_HES_ARCH)
854 x86_pmu_start(event, PERF_EF_RELOAD);
857 perf_events_lapic_init();
863 x86_pmu.enable_all(added);
866 static inline void __x86_pmu_enable_event(struct hw_perf_event *hwc,
869 wrmsrl(hwc->config_base + hwc->idx, hwc->config | enable_mask);
872 static inline void x86_pmu_disable_event(struct perf_event *event)
874 struct hw_perf_event *hwc = &event->hw;
876 wrmsrl(hwc->config_base + hwc->idx, hwc->config);
879 static DEFINE_PER_CPU(u64 [X86_PMC_IDX_MAX], pmc_prev_left);
882 * Set the next IRQ period, based on the hwc->period_left value.
883 * To be called with the event disabled in hw:
886 x86_perf_event_set_period(struct perf_event *event)
888 struct hw_perf_event *hwc = &event->hw;
889 s64 left = local64_read(&hwc->period_left);
890 s64 period = hwc->sample_period;
891 int ret = 0, idx = hwc->idx;
893 if (idx == X86_PMC_IDX_FIXED_BTS)
897 * If we are way outside a reasonable range then just skip forward:
899 if (unlikely(left <= -period)) {
901 local64_set(&hwc->period_left, left);
902 hwc->last_period = period;
906 if (unlikely(left <= 0)) {
908 local64_set(&hwc->period_left, left);
909 hwc->last_period = period;
913 * Quirk: certain CPUs dont like it if just 1 hw_event is left:
915 if (unlikely(left < 2))
918 if (left > x86_pmu.max_period)
919 left = x86_pmu.max_period;
921 per_cpu(pmc_prev_left[idx], smp_processor_id()) = left;
924 * The hw event starts counting from this event offset,
925 * mark it to be able to extra future deltas:
927 local64_set(&hwc->prev_count, (u64)-left);
929 wrmsrl(hwc->event_base + idx, (u64)(-left) & x86_pmu.cntval_mask);
932 * Due to erratum on certan cpu we need
933 * a second write to be sure the register
934 * is updated properly
936 if (x86_pmu.perfctr_second_write) {
937 wrmsrl(hwc->event_base + idx,
938 (u64)(-left) & x86_pmu.cntval_mask);
941 perf_event_update_userpage(event);
946 static void x86_pmu_enable_event(struct perf_event *event)
948 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
950 __x86_pmu_enable_event(&event->hw,
951 ARCH_PERFMON_EVENTSEL_ENABLE);
955 * Add a single event to the PMU.
957 * The event is added to the group of enabled events
958 * but only if it can be scehduled with existing events.
960 static int x86_pmu_add(struct perf_event *event, int flags)
962 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
963 struct hw_perf_event *hwc;
964 int assign[X86_PMC_IDX_MAX];
969 perf_pmu_disable(event->pmu);
971 ret = n = collect_events(cpuc, event, false);
975 hwc->state = PERF_HES_UPTODATE | PERF_HES_STOPPED;
976 if (!(flags & PERF_EF_START))
977 hwc->state |= PERF_HES_ARCH;
980 * If group events scheduling transaction was started,
981 * skip the schedulability test here, it will be peformed
982 * at commit time (->commit_txn) as a whole
984 if (cpuc->group_flag & PERF_EVENT_TXN)
987 ret = x86_pmu.schedule_events(cpuc, n, assign);
991 * copy new assignment, now we know it is possible
992 * will be used by hw_perf_enable()
994 memcpy(cpuc->assign, assign, n*sizeof(int));
998 cpuc->n_added += n - n0;
999 cpuc->n_txn += n - n0;
1003 perf_pmu_enable(event->pmu);
1007 static void x86_pmu_start(struct perf_event *event, int flags)
1009 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
1010 int idx = event->hw.idx;
1012 if (WARN_ON_ONCE(!(event->hw.state & PERF_HES_STOPPED)))
1015 if (WARN_ON_ONCE(idx == -1))
1018 if (flags & PERF_EF_RELOAD) {
1019 WARN_ON_ONCE(!(event->hw.state & PERF_HES_UPTODATE));
1020 x86_perf_event_set_period(event);
1023 event->hw.state = 0;
1025 cpuc->events[idx] = event;
1026 __set_bit(idx, cpuc->active_mask);
1027 __set_bit(idx, cpuc->running);
1028 x86_pmu.enable(event);
1029 perf_event_update_userpage(event);
1032 void perf_event_print_debug(void)
1034 u64 ctrl, status, overflow, pmc_ctrl, pmc_count, prev_left, fixed;
1036 struct cpu_hw_events *cpuc;
1037 unsigned long flags;
1040 if (!x86_pmu.num_counters)
1043 local_irq_save(flags);
1045 cpu = smp_processor_id();
1046 cpuc = &per_cpu(cpu_hw_events, cpu);
1048 if (x86_pmu.version >= 2) {
1049 rdmsrl(MSR_CORE_PERF_GLOBAL_CTRL, ctrl);
1050 rdmsrl(MSR_CORE_PERF_GLOBAL_STATUS, status);
1051 rdmsrl(MSR_CORE_PERF_GLOBAL_OVF_CTRL, overflow);
1052 rdmsrl(MSR_ARCH_PERFMON_FIXED_CTR_CTRL, fixed);
1053 rdmsrl(MSR_IA32_PEBS_ENABLE, pebs);
1056 pr_info("CPU#%d: ctrl: %016llx\n", cpu, ctrl);
1057 pr_info("CPU#%d: status: %016llx\n", cpu, status);
1058 pr_info("CPU#%d: overflow: %016llx\n", cpu, overflow);
1059 pr_info("CPU#%d: fixed: %016llx\n", cpu, fixed);
1060 pr_info("CPU#%d: pebs: %016llx\n", cpu, pebs);
1062 pr_info("CPU#%d: active: %016llx\n", cpu, *(u64 *)cpuc->active_mask);
1064 for (idx = 0; idx < x86_pmu.num_counters; idx++) {
1065 rdmsrl(x86_pmu.eventsel + idx, pmc_ctrl);
1066 rdmsrl(x86_pmu.perfctr + idx, pmc_count);
1068 prev_left = per_cpu(pmc_prev_left[idx], cpu);
1070 pr_info("CPU#%d: gen-PMC%d ctrl: %016llx\n",
1071 cpu, idx, pmc_ctrl);
1072 pr_info("CPU#%d: gen-PMC%d count: %016llx\n",
1073 cpu, idx, pmc_count);
1074 pr_info("CPU#%d: gen-PMC%d left: %016llx\n",
1075 cpu, idx, prev_left);
1077 for (idx = 0; idx < x86_pmu.num_counters_fixed; idx++) {
1078 rdmsrl(MSR_ARCH_PERFMON_FIXED_CTR0 + idx, pmc_count);
1080 pr_info("CPU#%d: fixed-PMC%d count: %016llx\n",
1081 cpu, idx, pmc_count);
1083 local_irq_restore(flags);
1086 static void x86_pmu_stop(struct perf_event *event, int flags)
1088 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
1089 struct hw_perf_event *hwc = &event->hw;
1091 if (__test_and_clear_bit(hwc->idx, cpuc->active_mask)) {
1092 x86_pmu.disable(event);
1093 cpuc->events[hwc->idx] = NULL;
1094 WARN_ON_ONCE(hwc->state & PERF_HES_STOPPED);
1095 hwc->state |= PERF_HES_STOPPED;
1098 if ((flags & PERF_EF_UPDATE) && !(hwc->state & PERF_HES_UPTODATE)) {
1100 * Drain the remaining delta count out of a event
1101 * that we are disabling:
1103 x86_perf_event_update(event);
1104 hwc->state |= PERF_HES_UPTODATE;
1108 static void x86_pmu_del(struct perf_event *event, int flags)
1110 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
1114 * If we're called during a txn, we don't need to do anything.
1115 * The events never got scheduled and ->cancel_txn will truncate
1118 if (cpuc->group_flag & PERF_EVENT_TXN)
1121 x86_pmu_stop(event, PERF_EF_UPDATE);
1123 for (i = 0; i < cpuc->n_events; i++) {
1124 if (event == cpuc->event_list[i]) {
1126 if (x86_pmu.put_event_constraints)
1127 x86_pmu.put_event_constraints(cpuc, event);
1129 while (++i < cpuc->n_events)
1130 cpuc->event_list[i-1] = cpuc->event_list[i];
1136 perf_event_update_userpage(event);
1139 static int x86_pmu_handle_irq(struct pt_regs *regs)
1141 struct perf_sample_data data;
1142 struct cpu_hw_events *cpuc;
1143 struct perf_event *event;
1144 int idx, handled = 0;
1147 perf_sample_data_init(&data, 0);
1149 cpuc = &__get_cpu_var(cpu_hw_events);
1151 for (idx = 0; idx < x86_pmu.num_counters; idx++) {
1152 if (!test_bit(idx, cpuc->active_mask)) {
1154 * Though we deactivated the counter some cpus
1155 * might still deliver spurious interrupts still
1156 * in flight. Catch them:
1158 if (__test_and_clear_bit(idx, cpuc->running))
1163 event = cpuc->events[idx];
1165 val = x86_perf_event_update(event);
1166 if (val & (1ULL << (x86_pmu.cntval_bits - 1)))
1173 data.period = event->hw.last_period;
1175 if (!x86_perf_event_set_period(event))
1178 if (perf_event_overflow(event, 1, &data, regs))
1179 x86_pmu_stop(event, 0);
1183 inc_irq_stat(apic_perf_irqs);
1188 void perf_events_lapic_init(void)
1190 if (!x86_pmu.apic || !x86_pmu_initialized())
1194 * Always use NMI for PMU
1196 apic_write(APIC_LVTPC, APIC_DM_NMI);
1199 struct pmu_nmi_state {
1200 unsigned int marked;
1204 static DEFINE_PER_CPU(struct pmu_nmi_state, pmu_nmi);
1206 static int __kprobes
1207 perf_event_nmi_handler(struct notifier_block *self,
1208 unsigned long cmd, void *__args)
1210 struct die_args *args = __args;
1211 unsigned int this_nmi;
1214 if (!atomic_read(&active_events))
1221 case DIE_NMIUNKNOWN:
1222 this_nmi = percpu_read(irq_stat.__nmi_count);
1223 if (this_nmi != __get_cpu_var(pmu_nmi).marked)
1224 /* let the kernel handle the unknown nmi */
1227 * This one is a PMU back-to-back nmi. Two events
1228 * trigger 'simultaneously' raising two back-to-back
1229 * NMIs. If the first NMI handles both, the latter
1230 * will be empty and daze the CPU. So, we drop it to
1231 * avoid false-positive 'unknown nmi' messages.
1238 apic_write(APIC_LVTPC, APIC_DM_NMI);
1240 handled = x86_pmu.handle_irq(args->regs);
1244 this_nmi = percpu_read(irq_stat.__nmi_count);
1245 if ((handled > 1) ||
1246 /* the next nmi could be a back-to-back nmi */
1247 ((__get_cpu_var(pmu_nmi).marked == this_nmi) &&
1248 (__get_cpu_var(pmu_nmi).handled > 1))) {
1250 * We could have two subsequent back-to-back nmis: The
1251 * first handles more than one counter, the 2nd
1252 * handles only one counter and the 3rd handles no
1255 * This is the 2nd nmi because the previous was
1256 * handling more than one counter. We will mark the
1257 * next (3rd) and then drop it if unhandled.
1259 __get_cpu_var(pmu_nmi).marked = this_nmi + 1;
1260 __get_cpu_var(pmu_nmi).handled = handled;
1266 static __read_mostly struct notifier_block perf_event_nmi_notifier = {
1267 .notifier_call = perf_event_nmi_handler,
1272 static struct event_constraint unconstrained;
1273 static struct event_constraint emptyconstraint;
1275 static struct event_constraint *
1276 x86_get_event_constraints(struct cpu_hw_events *cpuc, struct perf_event *event)
1278 struct event_constraint *c;
1280 if (x86_pmu.event_constraints) {
1281 for_each_event_constraint(c, x86_pmu.event_constraints) {
1282 if ((event->hw.config & c->cmask) == c->code)
1287 return &unconstrained;
1290 #include "perf_event_amd.c"
1291 #include "perf_event_p6.c"
1292 #include "perf_event_p4.c"
1293 #include "perf_event_intel_lbr.c"
1294 #include "perf_event_intel_ds.c"
1295 #include "perf_event_intel.c"
1297 static int __cpuinit
1298 x86_pmu_notifier(struct notifier_block *self, unsigned long action, void *hcpu)
1300 unsigned int cpu = (long)hcpu;
1301 int ret = NOTIFY_OK;
1303 switch (action & ~CPU_TASKS_FROZEN) {
1304 case CPU_UP_PREPARE:
1305 if (x86_pmu.cpu_prepare)
1306 ret = x86_pmu.cpu_prepare(cpu);
1310 if (x86_pmu.cpu_starting)
1311 x86_pmu.cpu_starting(cpu);
1315 if (x86_pmu.cpu_dying)
1316 x86_pmu.cpu_dying(cpu);
1319 case CPU_UP_CANCELED:
1321 if (x86_pmu.cpu_dead)
1322 x86_pmu.cpu_dead(cpu);
1332 static void __init pmu_check_apic(void)
1338 pr_info("no APIC, boot with the \"lapic\" boot parameter to force-enable it.\n");
1339 pr_info("no hardware sampling interrupt available.\n");
1342 void __init init_hw_perf_events(void)
1344 struct event_constraint *c;
1347 pr_info("Performance Events: ");
1349 switch (boot_cpu_data.x86_vendor) {
1350 case X86_VENDOR_INTEL:
1351 err = intel_pmu_init();
1353 case X86_VENDOR_AMD:
1354 err = amd_pmu_init();
1360 pr_cont("no PMU driver, software events only.\n");
1366 pr_cont("%s PMU driver.\n", x86_pmu.name);
1371 if (x86_pmu.num_counters > X86_PMC_MAX_GENERIC) {
1372 WARN(1, KERN_ERR "hw perf events %d > max(%d), clipping!",
1373 x86_pmu.num_counters, X86_PMC_MAX_GENERIC);
1374 x86_pmu.num_counters = X86_PMC_MAX_GENERIC;
1376 x86_pmu.intel_ctrl = (1 << x86_pmu.num_counters) - 1;
1378 if (x86_pmu.num_counters_fixed > X86_PMC_MAX_FIXED) {
1379 WARN(1, KERN_ERR "hw perf events fixed %d > max(%d), clipping!",
1380 x86_pmu.num_counters_fixed, X86_PMC_MAX_FIXED);
1381 x86_pmu.num_counters_fixed = X86_PMC_MAX_FIXED;
1384 x86_pmu.intel_ctrl |=
1385 ((1LL << x86_pmu.num_counters_fixed)-1) << X86_PMC_IDX_FIXED;
1387 perf_events_lapic_init();
1388 register_die_notifier(&perf_event_nmi_notifier);
1390 unconstrained = (struct event_constraint)
1391 __EVENT_CONSTRAINT(0, (1ULL << x86_pmu.num_counters) - 1,
1392 0, x86_pmu.num_counters);
1394 if (x86_pmu.event_constraints) {
1395 for_each_event_constraint(c, x86_pmu.event_constraints) {
1396 if (c->cmask != X86_RAW_EVENT_MASK)
1399 c->idxmsk64 |= (1ULL << x86_pmu.num_counters) - 1;
1400 c->weight += x86_pmu.num_counters;
1404 pr_info("... version: %d\n", x86_pmu.version);
1405 pr_info("... bit width: %d\n", x86_pmu.cntval_bits);
1406 pr_info("... generic registers: %d\n", x86_pmu.num_counters);
1407 pr_info("... value mask: %016Lx\n", x86_pmu.cntval_mask);
1408 pr_info("... max period: %016Lx\n", x86_pmu.max_period);
1409 pr_info("... fixed-purpose events: %d\n", x86_pmu.num_counters_fixed);
1410 pr_info("... event mask: %016Lx\n", x86_pmu.intel_ctrl);
1412 perf_pmu_register(&pmu);
1413 perf_cpu_notifier(x86_pmu_notifier);
1416 static inline void x86_pmu_read(struct perf_event *event)
1418 x86_perf_event_update(event);
1422 * Start group events scheduling transaction
1423 * Set the flag to make pmu::enable() not perform the
1424 * schedulability test, it will be performed at commit time
1426 static void x86_pmu_start_txn(struct pmu *pmu)
1428 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
1430 perf_pmu_disable(pmu);
1431 cpuc->group_flag |= PERF_EVENT_TXN;
1436 * Stop group events scheduling transaction
1437 * Clear the flag and pmu::enable() will perform the
1438 * schedulability test.
1440 static void x86_pmu_cancel_txn(struct pmu *pmu)
1442 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
1444 cpuc->group_flag &= ~PERF_EVENT_TXN;
1446 * Truncate the collected events.
1448 cpuc->n_added -= cpuc->n_txn;
1449 cpuc->n_events -= cpuc->n_txn;
1450 perf_pmu_enable(pmu);
1454 * Commit group events scheduling transaction
1455 * Perform the group schedulability test as a whole
1456 * Return 0 if success
1458 static int x86_pmu_commit_txn(struct pmu *pmu)
1460 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
1461 int assign[X86_PMC_IDX_MAX];
1466 if (!x86_pmu_initialized())
1469 ret = x86_pmu.schedule_events(cpuc, n, assign);
1474 * copy new assignment, now we know it is possible
1475 * will be used by hw_perf_enable()
1477 memcpy(cpuc->assign, assign, n*sizeof(int));
1479 cpuc->group_flag &= ~PERF_EVENT_TXN;
1480 perf_pmu_enable(pmu);
1485 * validate that we can schedule this event
1487 static int validate_event(struct perf_event *event)
1489 struct cpu_hw_events *fake_cpuc;
1490 struct event_constraint *c;
1493 fake_cpuc = kmalloc(sizeof(*fake_cpuc), GFP_KERNEL | __GFP_ZERO);
1497 c = x86_pmu.get_event_constraints(fake_cpuc, event);
1499 if (!c || !c->weight)
1502 if (x86_pmu.put_event_constraints)
1503 x86_pmu.put_event_constraints(fake_cpuc, event);
1511 * validate a single event group
1513 * validation include:
1514 * - check events are compatible which each other
1515 * - events do not compete for the same counter
1516 * - number of events <= number of counters
1518 * validation ensures the group can be loaded onto the
1519 * PMU if it was the only group available.
1521 static int validate_group(struct perf_event *event)
1523 struct perf_event *leader = event->group_leader;
1524 struct cpu_hw_events *fake_cpuc;
1528 fake_cpuc = kmalloc(sizeof(*fake_cpuc), GFP_KERNEL | __GFP_ZERO);
1533 * the event is not yet connected with its
1534 * siblings therefore we must first collect
1535 * existing siblings, then add the new event
1536 * before we can simulate the scheduling
1539 n = collect_events(fake_cpuc, leader, true);
1543 fake_cpuc->n_events = n;
1544 n = collect_events(fake_cpuc, event, false);
1548 fake_cpuc->n_events = n;
1550 ret = x86_pmu.schedule_events(fake_cpuc, n, NULL);
1558 int x86_pmu_event_init(struct perf_event *event)
1563 switch (event->attr.type) {
1565 case PERF_TYPE_HARDWARE:
1566 case PERF_TYPE_HW_CACHE:
1573 err = __x86_pmu_event_init(event);
1576 * we temporarily connect event to its pmu
1577 * such that validate_group() can classify
1578 * it as an x86 event using is_x86_event()
1583 if (event->group_leader != event)
1584 err = validate_group(event);
1586 err = validate_event(event);
1592 event->destroy(event);
1598 static struct pmu pmu = {
1599 .pmu_enable = x86_pmu_enable,
1600 .pmu_disable = x86_pmu_disable,
1602 .event_init = x86_pmu_event_init,
1606 .start = x86_pmu_start,
1607 .stop = x86_pmu_stop,
1608 .read = x86_pmu_read,
1610 .start_txn = x86_pmu_start_txn,
1611 .cancel_txn = x86_pmu_cancel_txn,
1612 .commit_txn = x86_pmu_commit_txn,
1620 backtrace_warning_symbol(void *data, char *msg, unsigned long symbol)
1622 /* Ignore warnings */
1625 static void backtrace_warning(void *data, char *msg)
1627 /* Ignore warnings */
1630 static int backtrace_stack(void *data, char *name)
1635 static void backtrace_address(void *data, unsigned long addr, int reliable)
1637 struct perf_callchain_entry *entry = data;
1639 perf_callchain_store(entry, addr);
1642 static const struct stacktrace_ops backtrace_ops = {
1643 .warning = backtrace_warning,
1644 .warning_symbol = backtrace_warning_symbol,
1645 .stack = backtrace_stack,
1646 .address = backtrace_address,
1647 .walk_stack = print_context_stack_bp,
1651 perf_callchain_kernel(struct perf_callchain_entry *entry, struct pt_regs *regs)
1653 if (perf_guest_cbs && perf_guest_cbs->is_in_guest()) {
1654 /* TODO: We don't support guest os callchain now */
1658 perf_callchain_store(entry, regs->ip);
1660 dump_trace(NULL, regs, NULL, &backtrace_ops, entry);
1663 #ifdef CONFIG_COMPAT
1665 perf_callchain_user32(struct pt_regs *regs, struct perf_callchain_entry *entry)
1667 /* 32-bit process in 64-bit kernel. */
1668 struct stack_frame_ia32 frame;
1669 const void __user *fp;
1671 if (!test_thread_flag(TIF_IA32))
1674 fp = compat_ptr(regs->bp);
1675 while (entry->nr < PERF_MAX_STACK_DEPTH) {
1676 unsigned long bytes;
1677 frame.next_frame = 0;
1678 frame.return_address = 0;
1680 bytes = copy_from_user_nmi(&frame, fp, sizeof(frame));
1681 if (bytes != sizeof(frame))
1684 if (fp < compat_ptr(regs->sp))
1687 perf_callchain_store(entry, frame.return_address);
1688 fp = compat_ptr(frame.next_frame);
1694 perf_callchain_user32(struct pt_regs *regs, struct perf_callchain_entry *entry)
1701 perf_callchain_user(struct perf_callchain_entry *entry, struct pt_regs *regs)
1703 struct stack_frame frame;
1704 const void __user *fp;
1706 if (perf_guest_cbs && perf_guest_cbs->is_in_guest()) {
1707 /* TODO: We don't support guest os callchain now */
1711 fp = (void __user *)regs->bp;
1713 perf_callchain_store(entry, regs->ip);
1715 if (perf_callchain_user32(regs, entry))
1718 while (entry->nr < PERF_MAX_STACK_DEPTH) {
1719 unsigned long bytes;
1720 frame.next_frame = NULL;
1721 frame.return_address = 0;
1723 bytes = copy_from_user_nmi(&frame, fp, sizeof(frame));
1724 if (bytes != sizeof(frame))
1727 if ((unsigned long)fp < regs->sp)
1730 perf_callchain_store(entry, frame.return_address);
1731 fp = frame.next_frame;
1735 unsigned long perf_instruction_pointer(struct pt_regs *regs)
1739 if (perf_guest_cbs && perf_guest_cbs->is_in_guest())
1740 ip = perf_guest_cbs->get_guest_ip();
1742 ip = instruction_pointer(regs);
1747 unsigned long perf_misc_flags(struct pt_regs *regs)
1751 if (perf_guest_cbs && perf_guest_cbs->is_in_guest()) {
1752 if (perf_guest_cbs->is_user_mode())
1753 misc |= PERF_RECORD_MISC_GUEST_USER;
1755 misc |= PERF_RECORD_MISC_GUEST_KERNEL;
1757 if (user_mode(regs))
1758 misc |= PERF_RECORD_MISC_USER;
1760 misc |= PERF_RECORD_MISC_KERNEL;
1763 if (regs->flags & PERF_EFLAGS_EXACT)
1764 misc |= PERF_RECORD_MISC_EXACT_IP;