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/highmem.h>
25 #include <linux/cpu.h>
26 #include <linux/bitops.h>
29 #include <asm/stacktrace.h>
32 static u64 perf_event_mask __read_mostly;
34 /* The maximal number of PEBS events: */
35 #define MAX_PEBS_EVENTS 4
37 /* The size of a BTS record in bytes: */
38 #define BTS_RECORD_SIZE 24
40 /* The size of a per-cpu BTS buffer in bytes: */
41 #define BTS_BUFFER_SIZE (BTS_RECORD_SIZE * 2048)
43 /* The BTS overflow threshold in bytes from the end of the buffer: */
44 #define BTS_OVFL_TH (BTS_RECORD_SIZE * 128)
48 * Bits in the debugctlmsr controlling branch tracing.
50 #define X86_DEBUGCTL_TR (1 << 6)
51 #define X86_DEBUGCTL_BTS (1 << 7)
52 #define X86_DEBUGCTL_BTINT (1 << 8)
53 #define X86_DEBUGCTL_BTS_OFF_OS (1 << 9)
54 #define X86_DEBUGCTL_BTS_OFF_USR (1 << 10)
57 * A debug store configuration.
59 * We only support architectures that use 64bit fields.
64 u64 bts_absolute_maximum;
65 u64 bts_interrupt_threshold;
68 u64 pebs_absolute_maximum;
69 u64 pebs_interrupt_threshold;
70 u64 pebs_event_reset[MAX_PEBS_EVENTS];
73 struct event_constraint {
75 unsigned long idxmsk[BITS_TO_LONGS(X86_PMC_IDX_MAX)];
84 int nb_id; /* NorthBridge id */
85 int refcnt; /* reference count */
86 struct perf_event *owners[X86_PMC_IDX_MAX];
87 struct event_constraint event_constraints[X86_PMC_IDX_MAX];
90 struct cpu_hw_events {
91 struct perf_event *events[X86_PMC_IDX_MAX]; /* in counter order */
92 unsigned long active_mask[BITS_TO_LONGS(X86_PMC_IDX_MAX)];
93 unsigned long interrupts;
95 struct debug_store *ds;
99 int assign[X86_PMC_IDX_MAX]; /* event to counter assignment */
100 u64 tags[X86_PMC_IDX_MAX];
101 struct perf_event *event_list[X86_PMC_IDX_MAX]; /* in enabled order */
102 struct amd_nb *amd_nb;
105 #define __EVENT_CONSTRAINT(c, n, m, w) {\
106 { .idxmsk64 = (n) }, \
112 #define EVENT_CONSTRAINT(c, n, m) \
113 __EVENT_CONSTRAINT(c, n, m, HWEIGHT(n))
115 #define INTEL_EVENT_CONSTRAINT(c, n) \
116 EVENT_CONSTRAINT(c, n, INTEL_ARCH_EVTSEL_MASK)
118 #define FIXED_EVENT_CONSTRAINT(c, n) \
119 EVENT_CONSTRAINT(c, (1ULL << (32+n)), INTEL_ARCH_FIXED_MASK)
121 #define EVENT_CONSTRAINT_END \
122 EVENT_CONSTRAINT(0, 0, 0)
124 #define for_each_event_constraint(e, c) \
125 for ((e) = (c); (e)->cmask; (e)++)
128 * struct x86_pmu - generic x86 pmu
133 int (*handle_irq)(struct pt_regs *);
134 void (*disable_all)(void);
135 void (*enable_all)(void);
136 void (*enable)(struct hw_perf_event *, int);
137 void (*disable)(struct hw_perf_event *, int);
140 u64 (*event_map)(int);
141 u64 (*raw_event)(u64);
144 int num_events_fixed;
150 void (*enable_bts)(u64 config);
151 void (*disable_bts)(void);
153 struct event_constraint *
154 (*get_event_constraints)(struct cpu_hw_events *cpuc,
155 struct perf_event *event);
157 void (*put_event_constraints)(struct cpu_hw_events *cpuc,
158 struct perf_event *event);
159 struct event_constraint *event_constraints;
161 void (*cpu_prepare)(int cpu);
162 void (*cpu_starting)(int cpu);
163 void (*cpu_dying)(int cpu);
164 void (*cpu_dead)(int cpu);
167 static struct x86_pmu x86_pmu __read_mostly;
169 static DEFINE_PER_CPU(struct cpu_hw_events, cpu_hw_events) = {
173 static int x86_perf_event_set_period(struct perf_event *event);
176 * Generalized hw caching related hw_event table, filled
177 * in on a per model basis. A value of 0 means
178 * 'not supported', -1 means 'hw_event makes no sense on
179 * this CPU', any other value means the raw hw_event
183 #define C(x) PERF_COUNT_HW_CACHE_##x
185 static u64 __read_mostly hw_cache_event_ids
186 [PERF_COUNT_HW_CACHE_MAX]
187 [PERF_COUNT_HW_CACHE_OP_MAX]
188 [PERF_COUNT_HW_CACHE_RESULT_MAX];
191 * Propagate event elapsed time into the generic event.
192 * Can only be executed on the CPU where the event is active.
193 * Returns the delta events processed.
196 x86_perf_event_update(struct perf_event *event)
198 struct hw_perf_event *hwc = &event->hw;
199 int shift = 64 - x86_pmu.event_bits;
200 u64 prev_raw_count, new_raw_count;
204 if (idx == X86_PMC_IDX_FIXED_BTS)
208 * Careful: an NMI might modify the previous event value.
210 * Our tactic to handle this is to first atomically read and
211 * exchange a new raw count - then add that new-prev delta
212 * count to the generic event atomically:
215 prev_raw_count = atomic64_read(&hwc->prev_count);
216 rdmsrl(hwc->event_base + idx, new_raw_count);
218 if (atomic64_cmpxchg(&hwc->prev_count, prev_raw_count,
219 new_raw_count) != prev_raw_count)
223 * Now we have the new raw value and have updated the prev
224 * timestamp already. We can now calculate the elapsed delta
225 * (event-)time and add that to the generic event.
227 * Careful, not all hw sign-extends above the physical width
230 delta = (new_raw_count << shift) - (prev_raw_count << shift);
233 atomic64_add(delta, &event->count);
234 atomic64_sub(delta, &hwc->period_left);
236 return new_raw_count;
239 static atomic_t active_events;
240 static DEFINE_MUTEX(pmc_reserve_mutex);
242 static bool reserve_pmc_hardware(void)
244 #ifdef CONFIG_X86_LOCAL_APIC
247 if (nmi_watchdog == NMI_LOCAL_APIC)
248 disable_lapic_nmi_watchdog();
250 for (i = 0; i < x86_pmu.num_events; i++) {
251 if (!reserve_perfctr_nmi(x86_pmu.perfctr + i))
255 for (i = 0; i < x86_pmu.num_events; i++) {
256 if (!reserve_evntsel_nmi(x86_pmu.eventsel + i))
263 #ifdef CONFIG_X86_LOCAL_APIC
265 for (i--; i >= 0; i--)
266 release_evntsel_nmi(x86_pmu.eventsel + i);
268 i = x86_pmu.num_events;
271 for (i--; i >= 0; i--)
272 release_perfctr_nmi(x86_pmu.perfctr + i);
274 if (nmi_watchdog == NMI_LOCAL_APIC)
275 enable_lapic_nmi_watchdog();
281 static void release_pmc_hardware(void)
283 #ifdef CONFIG_X86_LOCAL_APIC
286 for (i = 0; i < x86_pmu.num_events; i++) {
287 release_perfctr_nmi(x86_pmu.perfctr + i);
288 release_evntsel_nmi(x86_pmu.eventsel + i);
291 if (nmi_watchdog == NMI_LOCAL_APIC)
292 enable_lapic_nmi_watchdog();
296 static inline bool bts_available(void)
298 return x86_pmu.enable_bts != NULL;
301 static void init_debug_store_on_cpu(int cpu)
303 struct debug_store *ds = per_cpu(cpu_hw_events, cpu).ds;
308 wrmsr_on_cpu(cpu, MSR_IA32_DS_AREA,
309 (u32)((u64)(unsigned long)ds),
310 (u32)((u64)(unsigned long)ds >> 32));
313 static void fini_debug_store_on_cpu(int cpu)
315 if (!per_cpu(cpu_hw_events, cpu).ds)
318 wrmsr_on_cpu(cpu, MSR_IA32_DS_AREA, 0, 0);
321 static void release_bts_hardware(void)
325 if (!bts_available())
330 for_each_online_cpu(cpu)
331 fini_debug_store_on_cpu(cpu);
333 for_each_possible_cpu(cpu) {
334 struct debug_store *ds = per_cpu(cpu_hw_events, cpu).ds;
339 per_cpu(cpu_hw_events, cpu).ds = NULL;
341 kfree((void *)(unsigned long)ds->bts_buffer_base);
348 static int reserve_bts_hardware(void)
352 if (!bts_available())
357 for_each_possible_cpu(cpu) {
358 struct debug_store *ds;
362 buffer = kzalloc(BTS_BUFFER_SIZE, GFP_KERNEL);
363 if (unlikely(!buffer))
366 ds = kzalloc(sizeof(*ds), GFP_KERNEL);
372 ds->bts_buffer_base = (u64)(unsigned long)buffer;
373 ds->bts_index = ds->bts_buffer_base;
374 ds->bts_absolute_maximum =
375 ds->bts_buffer_base + BTS_BUFFER_SIZE;
376 ds->bts_interrupt_threshold =
377 ds->bts_absolute_maximum - BTS_OVFL_TH;
379 per_cpu(cpu_hw_events, cpu).ds = ds;
384 release_bts_hardware();
386 for_each_online_cpu(cpu)
387 init_debug_store_on_cpu(cpu);
395 static void hw_perf_event_destroy(struct perf_event *event)
397 if (atomic_dec_and_mutex_lock(&active_events, &pmc_reserve_mutex)) {
398 release_pmc_hardware();
399 release_bts_hardware();
400 mutex_unlock(&pmc_reserve_mutex);
404 static inline int x86_pmu_initialized(void)
406 return x86_pmu.handle_irq != NULL;
410 set_ext_hw_attr(struct hw_perf_event *hwc, struct perf_event_attr *attr)
412 unsigned int cache_type, cache_op, cache_result;
415 config = attr->config;
417 cache_type = (config >> 0) & 0xff;
418 if (cache_type >= PERF_COUNT_HW_CACHE_MAX)
421 cache_op = (config >> 8) & 0xff;
422 if (cache_op >= PERF_COUNT_HW_CACHE_OP_MAX)
425 cache_result = (config >> 16) & 0xff;
426 if (cache_result >= PERF_COUNT_HW_CACHE_RESULT_MAX)
429 val = hw_cache_event_ids[cache_type][cache_op][cache_result];
443 * Setup the hardware configuration for a given attr_type
445 static int __hw_perf_event_init(struct perf_event *event)
447 struct perf_event_attr *attr = &event->attr;
448 struct hw_perf_event *hwc = &event->hw;
452 if (!x86_pmu_initialized())
456 if (!atomic_inc_not_zero(&active_events)) {
457 mutex_lock(&pmc_reserve_mutex);
458 if (atomic_read(&active_events) == 0) {
459 if (!reserve_pmc_hardware())
462 err = reserve_bts_hardware();
465 atomic_inc(&active_events);
466 mutex_unlock(&pmc_reserve_mutex);
471 event->destroy = hw_perf_event_destroy;
475 * (keep 'enabled' bit clear for now)
477 hwc->config = ARCH_PERFMON_EVENTSEL_INT;
481 hwc->last_tag = ~0ULL;
484 * Count user and OS events unless requested not to.
486 if (!attr->exclude_user)
487 hwc->config |= ARCH_PERFMON_EVENTSEL_USR;
488 if (!attr->exclude_kernel)
489 hwc->config |= ARCH_PERFMON_EVENTSEL_OS;
491 if (!hwc->sample_period) {
492 hwc->sample_period = x86_pmu.max_period;
493 hwc->last_period = hwc->sample_period;
494 atomic64_set(&hwc->period_left, hwc->sample_period);
497 * If we have a PMU initialized but no APIC
498 * interrupts, we cannot sample hardware
499 * events (user-space has to fall back and
500 * sample via a hrtimer based software event):
507 * Raw hw_event type provide the config in the hw_event structure
509 if (attr->type == PERF_TYPE_RAW) {
510 hwc->config |= x86_pmu.raw_event(attr->config);
511 if ((hwc->config & ARCH_PERFMON_EVENTSEL_ANY) &&
512 perf_paranoid_cpu() && !capable(CAP_SYS_ADMIN))
517 if (attr->type == PERF_TYPE_HW_CACHE)
518 return set_ext_hw_attr(hwc, attr);
520 if (attr->config >= x86_pmu.max_events)
526 config = x86_pmu.event_map(attr->config);
537 if ((attr->config == PERF_COUNT_HW_BRANCH_INSTRUCTIONS) &&
538 (hwc->sample_period == 1)) {
539 /* BTS is not supported by this architecture. */
540 if (!bts_available())
543 /* BTS is currently only allowed for user-mode. */
544 if (hwc->config & ARCH_PERFMON_EVENTSEL_OS)
548 hwc->config |= config;
553 static void x86_pmu_disable_all(void)
555 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
558 for (idx = 0; idx < x86_pmu.num_events; idx++) {
561 if (!test_bit(idx, cpuc->active_mask))
563 rdmsrl(x86_pmu.eventsel + idx, val);
564 if (!(val & ARCH_PERFMON_EVENTSEL_ENABLE))
566 val &= ~ARCH_PERFMON_EVENTSEL_ENABLE;
567 wrmsrl(x86_pmu.eventsel + idx, val);
571 void hw_perf_disable(void)
573 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
575 if (!x86_pmu_initialized())
585 x86_pmu.disable_all();
588 static void x86_pmu_enable_all(void)
590 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
593 for (idx = 0; idx < x86_pmu.num_events; idx++) {
594 struct perf_event *event = cpuc->events[idx];
597 if (!test_bit(idx, cpuc->active_mask))
600 val = event->hw.config;
601 val |= ARCH_PERFMON_EVENTSEL_ENABLE;
602 wrmsrl(x86_pmu.eventsel + idx, val);
606 static const struct pmu pmu;
608 static inline int is_x86_event(struct perf_event *event)
610 return event->pmu == &pmu;
613 static int x86_schedule_events(struct cpu_hw_events *cpuc, int n, int *assign)
615 struct event_constraint *c, *constraints[X86_PMC_IDX_MAX];
616 unsigned long used_mask[BITS_TO_LONGS(X86_PMC_IDX_MAX)];
617 int i, j, w, wmax, num = 0;
618 struct hw_perf_event *hwc;
620 bitmap_zero(used_mask, X86_PMC_IDX_MAX);
622 for (i = 0; i < n; i++) {
623 c = x86_pmu.get_event_constraints(cpuc, cpuc->event_list[i]);
628 * fastpath, try to reuse previous register
630 for (i = 0; i < n; i++) {
631 hwc = &cpuc->event_list[i]->hw;
638 /* constraint still honored */
639 if (!test_bit(hwc->idx, c->idxmsk))
642 /* not already used */
643 if (test_bit(hwc->idx, used_mask))
646 set_bit(hwc->idx, used_mask);
648 assign[i] = hwc->idx;
657 bitmap_zero(used_mask, X86_PMC_IDX_MAX);
660 * weight = number of possible counters
662 * 1 = most constrained, only works on one counter
663 * wmax = least constrained, works on any counter
665 * assign events to counters starting with most
666 * constrained events.
668 wmax = x86_pmu.num_events;
671 * when fixed event counters are present,
672 * wmax is incremented by 1 to account
673 * for one more choice
675 if (x86_pmu.num_events_fixed)
678 for (w = 1, num = n; num && w <= wmax; w++) {
680 for (i = 0; num && i < n; i++) {
682 hwc = &cpuc->event_list[i]->hw;
687 for_each_set_bit(j, c->idxmsk, X86_PMC_IDX_MAX) {
688 if (!test_bit(j, used_mask))
692 if (j == X86_PMC_IDX_MAX)
695 set_bit(j, used_mask);
704 * scheduling failed or is just a simulation,
705 * free resources if necessary
707 if (!assign || num) {
708 for (i = 0; i < n; i++) {
709 if (x86_pmu.put_event_constraints)
710 x86_pmu.put_event_constraints(cpuc, cpuc->event_list[i]);
713 return num ? -ENOSPC : 0;
717 * dogrp: true if must collect siblings events (group)
718 * returns total number of events and error code
720 static int collect_events(struct cpu_hw_events *cpuc, struct perf_event *leader, bool dogrp)
722 struct perf_event *event;
725 max_count = x86_pmu.num_events + x86_pmu.num_events_fixed;
727 /* current number of events already accepted */
730 if (is_x86_event(leader)) {
733 cpuc->event_list[n] = leader;
739 list_for_each_entry(event, &leader->sibling_list, group_entry) {
740 if (!is_x86_event(event) ||
741 event->state <= PERF_EVENT_STATE_OFF)
747 cpuc->event_list[n] = event;
753 static inline void x86_assign_hw_event(struct perf_event *event,
754 struct cpu_hw_events *cpuc, int i)
756 struct hw_perf_event *hwc = &event->hw;
758 hwc->idx = cpuc->assign[i];
759 hwc->last_cpu = smp_processor_id();
760 hwc->last_tag = ++cpuc->tags[i];
762 if (hwc->idx == X86_PMC_IDX_FIXED_BTS) {
763 hwc->config_base = 0;
765 } else if (hwc->idx >= X86_PMC_IDX_FIXED) {
766 hwc->config_base = MSR_ARCH_PERFMON_FIXED_CTR_CTRL;
768 * We set it so that event_base + idx in wrmsr/rdmsr maps to
769 * MSR_ARCH_PERFMON_FIXED_CTR0 ... CTR2:
772 MSR_ARCH_PERFMON_FIXED_CTR0 - X86_PMC_IDX_FIXED;
774 hwc->config_base = x86_pmu.eventsel;
775 hwc->event_base = x86_pmu.perfctr;
779 static inline int match_prev_assignment(struct hw_perf_event *hwc,
780 struct cpu_hw_events *cpuc,
783 return hwc->idx == cpuc->assign[i] &&
784 hwc->last_cpu == smp_processor_id() &&
785 hwc->last_tag == cpuc->tags[i];
788 static void x86_pmu_stop(struct perf_event *event);
790 void hw_perf_enable(void)
792 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
793 struct perf_event *event;
794 struct hw_perf_event *hwc;
797 if (!x86_pmu_initialized())
805 * apply assignment obtained either from
806 * hw_perf_group_sched_in() or x86_pmu_enable()
808 * step1: save events moving to new counters
809 * step2: reprogram moved events into new counters
811 for (i = 0; i < cpuc->n_events; i++) {
813 event = cpuc->event_list[i];
817 * we can avoid reprogramming counter if:
818 * - assigned same counter as last time
819 * - running on same CPU as last time
820 * - no other event has used the counter since
822 if (hwc->idx == -1 ||
823 match_prev_assignment(hwc, cpuc, i))
831 for (i = 0; i < cpuc->n_events; i++) {
833 event = cpuc->event_list[i];
836 if (hwc->idx == -1) {
837 x86_assign_hw_event(event, cpuc, i);
838 x86_perf_event_set_period(event);
841 * need to mark as active because x86_pmu_disable()
842 * clear active_mask and events[] yet it preserves
845 set_bit(hwc->idx, cpuc->active_mask);
846 cpuc->events[hwc->idx] = event;
848 x86_pmu.enable(hwc, hwc->idx);
849 perf_event_update_userpage(event);
852 perf_events_lapic_init();
858 x86_pmu.enable_all();
861 static inline void __x86_pmu_enable_event(struct hw_perf_event *hwc, int idx)
863 (void)checking_wrmsrl(hwc->config_base + idx,
864 hwc->config | ARCH_PERFMON_EVENTSEL_ENABLE);
867 static inline void x86_pmu_disable_event(struct hw_perf_event *hwc, int idx)
869 (void)checking_wrmsrl(hwc->config_base + idx, hwc->config);
872 static DEFINE_PER_CPU(u64 [X86_PMC_IDX_MAX], pmc_prev_left);
875 * Set the next IRQ period, based on the hwc->period_left value.
876 * To be called with the event disabled in hw:
879 x86_perf_event_set_period(struct perf_event *event)
881 struct hw_perf_event *hwc = &event->hw;
882 s64 left = atomic64_read(&hwc->period_left);
883 s64 period = hwc->sample_period;
884 int err, ret = 0, idx = hwc->idx;
886 if (idx == X86_PMC_IDX_FIXED_BTS)
890 * If we are way outside a reasonable range then just skip forward:
892 if (unlikely(left <= -period)) {
894 atomic64_set(&hwc->period_left, left);
895 hwc->last_period = period;
899 if (unlikely(left <= 0)) {
901 atomic64_set(&hwc->period_left, left);
902 hwc->last_period = period;
906 * Quirk: certain CPUs dont like it if just 1 hw_event is left:
908 if (unlikely(left < 2))
911 if (left > x86_pmu.max_period)
912 left = x86_pmu.max_period;
914 per_cpu(pmc_prev_left[idx], smp_processor_id()) = left;
917 * The hw event starts counting from this event offset,
918 * mark it to be able to extra future deltas:
920 atomic64_set(&hwc->prev_count, (u64)-left);
922 err = checking_wrmsrl(hwc->event_base + idx,
923 (u64)(-left) & x86_pmu.event_mask);
925 perf_event_update_userpage(event);
930 static void x86_pmu_enable_event(struct hw_perf_event *hwc, int idx)
932 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
934 __x86_pmu_enable_event(hwc, idx);
938 * activate a single event
940 * The event is added to the group of enabled events
941 * but only if it can be scehduled with existing events.
943 * Called with PMU disabled. If successful and return value 1,
944 * then guaranteed to call perf_enable() and hw_perf_enable()
946 static int x86_pmu_enable(struct perf_event *event)
948 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
949 struct hw_perf_event *hwc;
950 int assign[X86_PMC_IDX_MAX];
956 n = collect_events(cpuc, event, false);
960 ret = x86_schedule_events(cpuc, n, assign);
964 * copy new assignment, now we know it is possible
965 * will be used by hw_perf_enable()
967 memcpy(cpuc->assign, assign, n*sizeof(int));
970 cpuc->n_added = n - n0;
975 static int x86_pmu_start(struct perf_event *event)
977 struct hw_perf_event *hwc = &event->hw;
982 x86_perf_event_set_period(event);
983 x86_pmu.enable(hwc, hwc->idx);
988 static void x86_pmu_unthrottle(struct perf_event *event)
990 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
991 struct hw_perf_event *hwc = &event->hw;
993 if (WARN_ON_ONCE(hwc->idx >= X86_PMC_IDX_MAX ||
994 cpuc->events[hwc->idx] != event))
997 x86_pmu.enable(hwc, hwc->idx);
1000 void perf_event_print_debug(void)
1002 u64 ctrl, status, overflow, pmc_ctrl, pmc_count, prev_left, fixed;
1003 struct cpu_hw_events *cpuc;
1004 unsigned long flags;
1007 if (!x86_pmu.num_events)
1010 local_irq_save(flags);
1012 cpu = smp_processor_id();
1013 cpuc = &per_cpu(cpu_hw_events, cpu);
1015 if (x86_pmu.version >= 2) {
1016 rdmsrl(MSR_CORE_PERF_GLOBAL_CTRL, ctrl);
1017 rdmsrl(MSR_CORE_PERF_GLOBAL_STATUS, status);
1018 rdmsrl(MSR_CORE_PERF_GLOBAL_OVF_CTRL, overflow);
1019 rdmsrl(MSR_ARCH_PERFMON_FIXED_CTR_CTRL, fixed);
1022 pr_info("CPU#%d: ctrl: %016llx\n", cpu, ctrl);
1023 pr_info("CPU#%d: status: %016llx\n", cpu, status);
1024 pr_info("CPU#%d: overflow: %016llx\n", cpu, overflow);
1025 pr_info("CPU#%d: fixed: %016llx\n", cpu, fixed);
1027 pr_info("CPU#%d: active: %016llx\n", cpu, *(u64 *)cpuc->active_mask);
1029 for (idx = 0; idx < x86_pmu.num_events; idx++) {
1030 rdmsrl(x86_pmu.eventsel + idx, pmc_ctrl);
1031 rdmsrl(x86_pmu.perfctr + idx, pmc_count);
1033 prev_left = per_cpu(pmc_prev_left[idx], cpu);
1035 pr_info("CPU#%d: gen-PMC%d ctrl: %016llx\n",
1036 cpu, idx, pmc_ctrl);
1037 pr_info("CPU#%d: gen-PMC%d count: %016llx\n",
1038 cpu, idx, pmc_count);
1039 pr_info("CPU#%d: gen-PMC%d left: %016llx\n",
1040 cpu, idx, prev_left);
1042 for (idx = 0; idx < x86_pmu.num_events_fixed; idx++) {
1043 rdmsrl(MSR_ARCH_PERFMON_FIXED_CTR0 + idx, pmc_count);
1045 pr_info("CPU#%d: fixed-PMC%d count: %016llx\n",
1046 cpu, idx, pmc_count);
1048 local_irq_restore(flags);
1051 static void x86_pmu_stop(struct perf_event *event)
1053 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
1054 struct hw_perf_event *hwc = &event->hw;
1058 * Must be done before we disable, otherwise the nmi handler
1059 * could reenable again:
1061 clear_bit(idx, cpuc->active_mask);
1062 x86_pmu.disable(hwc, idx);
1065 * Drain the remaining delta count out of a event
1066 * that we are disabling:
1068 x86_perf_event_update(event);
1070 cpuc->events[idx] = NULL;
1073 static void x86_pmu_disable(struct perf_event *event)
1075 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
1078 x86_pmu_stop(event);
1080 for (i = 0; i < cpuc->n_events; i++) {
1081 if (event == cpuc->event_list[i]) {
1083 if (x86_pmu.put_event_constraints)
1084 x86_pmu.put_event_constraints(cpuc, event);
1086 while (++i < cpuc->n_events)
1087 cpuc->event_list[i-1] = cpuc->event_list[i];
1093 perf_event_update_userpage(event);
1096 static int x86_pmu_handle_irq(struct pt_regs *regs)
1098 struct perf_sample_data data;
1099 struct cpu_hw_events *cpuc;
1100 struct perf_event *event;
1101 struct hw_perf_event *hwc;
1102 int idx, handled = 0;
1105 perf_sample_data_init(&data, 0);
1107 cpuc = &__get_cpu_var(cpu_hw_events);
1109 for (idx = 0; idx < x86_pmu.num_events; idx++) {
1110 if (!test_bit(idx, cpuc->active_mask))
1113 event = cpuc->events[idx];
1116 val = x86_perf_event_update(event);
1117 if (val & (1ULL << (x86_pmu.event_bits - 1)))
1124 data.period = event->hw.last_period;
1126 if (!x86_perf_event_set_period(event))
1129 if (perf_event_overflow(event, 1, &data, regs))
1130 x86_pmu.disable(hwc, idx);
1134 inc_irq_stat(apic_perf_irqs);
1139 void smp_perf_pending_interrupt(struct pt_regs *regs)
1143 inc_irq_stat(apic_pending_irqs);
1144 perf_event_do_pending();
1148 void set_perf_event_pending(void)
1150 #ifdef CONFIG_X86_LOCAL_APIC
1151 if (!x86_pmu.apic || !x86_pmu_initialized())
1154 apic->send_IPI_self(LOCAL_PENDING_VECTOR);
1158 void perf_events_lapic_init(void)
1160 #ifdef CONFIG_X86_LOCAL_APIC
1161 if (!x86_pmu.apic || !x86_pmu_initialized())
1165 * Always use NMI for PMU
1167 apic_write(APIC_LVTPC, APIC_DM_NMI);
1171 static int __kprobes
1172 perf_event_nmi_handler(struct notifier_block *self,
1173 unsigned long cmd, void *__args)
1175 struct die_args *args = __args;
1176 struct pt_regs *regs;
1178 if (!atomic_read(&active_events))
1192 #ifdef CONFIG_X86_LOCAL_APIC
1193 apic_write(APIC_LVTPC, APIC_DM_NMI);
1196 * Can't rely on the handled return value to say it was our NMI, two
1197 * events could trigger 'simultaneously' raising two back-to-back NMIs.
1199 * If the first NMI handles both, the latter will be empty and daze
1202 x86_pmu.handle_irq(regs);
1207 static __read_mostly struct notifier_block perf_event_nmi_notifier = {
1208 .notifier_call = perf_event_nmi_handler,
1213 static struct event_constraint unconstrained;
1214 static struct event_constraint emptyconstraint;
1216 static struct event_constraint *
1217 x86_get_event_constraints(struct cpu_hw_events *cpuc, struct perf_event *event)
1219 struct event_constraint *c;
1221 if (x86_pmu.event_constraints) {
1222 for_each_event_constraint(c, x86_pmu.event_constraints) {
1223 if ((event->hw.config & c->cmask) == c->code)
1228 return &unconstrained;
1231 static int x86_event_sched_in(struct perf_event *event,
1232 struct perf_cpu_context *cpuctx)
1236 event->state = PERF_EVENT_STATE_ACTIVE;
1237 event->oncpu = smp_processor_id();
1238 event->tstamp_running += event->ctx->time - event->tstamp_stopped;
1240 if (!is_x86_event(event))
1241 ret = event->pmu->enable(event);
1243 if (!ret && !is_software_event(event))
1244 cpuctx->active_oncpu++;
1246 if (!ret && event->attr.exclusive)
1247 cpuctx->exclusive = 1;
1252 static void x86_event_sched_out(struct perf_event *event,
1253 struct perf_cpu_context *cpuctx)
1255 event->state = PERF_EVENT_STATE_INACTIVE;
1258 if (!is_x86_event(event))
1259 event->pmu->disable(event);
1261 event->tstamp_running -= event->ctx->time - event->tstamp_stopped;
1263 if (!is_software_event(event))
1264 cpuctx->active_oncpu--;
1266 if (event->attr.exclusive || !cpuctx->active_oncpu)
1267 cpuctx->exclusive = 0;
1271 * Called to enable a whole group of events.
1272 * Returns 1 if the group was enabled, or -EAGAIN if it could not be.
1273 * Assumes the caller has disabled interrupts and has
1274 * frozen the PMU with hw_perf_save_disable.
1276 * called with PMU disabled. If successful and return value 1,
1277 * then guaranteed to call perf_enable() and hw_perf_enable()
1279 int hw_perf_group_sched_in(struct perf_event *leader,
1280 struct perf_cpu_context *cpuctx,
1281 struct perf_event_context *ctx)
1283 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
1284 struct perf_event *sub;
1285 int assign[X86_PMC_IDX_MAX];
1288 /* n0 = total number of events */
1289 n0 = collect_events(cpuc, leader, true);
1293 ret = x86_schedule_events(cpuc, n0, assign);
1297 ret = x86_event_sched_in(leader, cpuctx);
1302 list_for_each_entry(sub, &leader->sibling_list, group_entry) {
1303 if (sub->state > PERF_EVENT_STATE_OFF) {
1304 ret = x86_event_sched_in(sub, cpuctx);
1311 * copy new assignment, now we know it is possible
1312 * will be used by hw_perf_enable()
1314 memcpy(cpuc->assign, assign, n0*sizeof(int));
1316 cpuc->n_events = n0;
1318 ctx->nr_active += n1;
1321 * 1 means successful and events are active
1322 * This is not quite true because we defer
1323 * actual activation until hw_perf_enable() but
1324 * this way we* ensure caller won't try to enable
1329 x86_event_sched_out(leader, cpuctx);
1331 list_for_each_entry(sub, &leader->sibling_list, group_entry) {
1332 if (sub->state == PERF_EVENT_STATE_ACTIVE) {
1333 x86_event_sched_out(sub, cpuctx);
1341 #include "perf_event_amd.c"
1342 #include "perf_event_p6.c"
1343 #include "perf_event_intel.c"
1345 static int __cpuinit
1346 x86_pmu_notifier(struct notifier_block *self, unsigned long action, void *hcpu)
1348 unsigned int cpu = (long)hcpu;
1350 switch (action & ~CPU_TASKS_FROZEN) {
1351 case CPU_UP_PREPARE:
1352 if (x86_pmu.cpu_prepare)
1353 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);
1367 if (x86_pmu.cpu_dead)
1368 x86_pmu.cpu_dead(cpu);
1378 static void __init pmu_check_apic(void)
1384 pr_info("no APIC, boot with the \"lapic\" boot parameter to force-enable it.\n");
1385 pr_info("no hardware sampling interrupt available.\n");
1388 void __init init_hw_perf_events(void)
1390 struct event_constraint *c;
1393 pr_info("Performance Events: ");
1395 switch (boot_cpu_data.x86_vendor) {
1396 case X86_VENDOR_INTEL:
1397 err = intel_pmu_init();
1399 case X86_VENDOR_AMD:
1400 err = amd_pmu_init();
1406 pr_cont("no PMU driver, software events only.\n");
1412 pr_cont("%s PMU driver.\n", x86_pmu.name);
1414 if (x86_pmu.num_events > X86_PMC_MAX_GENERIC) {
1415 WARN(1, KERN_ERR "hw perf events %d > max(%d), clipping!",
1416 x86_pmu.num_events, X86_PMC_MAX_GENERIC);
1417 x86_pmu.num_events = X86_PMC_MAX_GENERIC;
1419 perf_event_mask = (1 << x86_pmu.num_events) - 1;
1420 perf_max_events = x86_pmu.num_events;
1422 if (x86_pmu.num_events_fixed > X86_PMC_MAX_FIXED) {
1423 WARN(1, KERN_ERR "hw perf events fixed %d > max(%d), clipping!",
1424 x86_pmu.num_events_fixed, X86_PMC_MAX_FIXED);
1425 x86_pmu.num_events_fixed = X86_PMC_MAX_FIXED;
1429 ((1LL << x86_pmu.num_events_fixed)-1) << X86_PMC_IDX_FIXED;
1430 x86_pmu.intel_ctrl = perf_event_mask;
1432 perf_events_lapic_init();
1433 register_die_notifier(&perf_event_nmi_notifier);
1435 unconstrained = (struct event_constraint)
1436 __EVENT_CONSTRAINT(0, (1ULL << x86_pmu.num_events) - 1,
1437 0, x86_pmu.num_events);
1439 if (x86_pmu.event_constraints) {
1440 for_each_event_constraint(c, x86_pmu.event_constraints) {
1441 if (c->cmask != INTEL_ARCH_FIXED_MASK)
1444 c->idxmsk64 |= (1ULL << x86_pmu.num_events) - 1;
1445 c->weight += x86_pmu.num_events;
1449 pr_info("... version: %d\n", x86_pmu.version);
1450 pr_info("... bit width: %d\n", x86_pmu.event_bits);
1451 pr_info("... generic registers: %d\n", x86_pmu.num_events);
1452 pr_info("... value mask: %016Lx\n", x86_pmu.event_mask);
1453 pr_info("... max period: %016Lx\n", x86_pmu.max_period);
1454 pr_info("... fixed-purpose events: %d\n", x86_pmu.num_events_fixed);
1455 pr_info("... event mask: %016Lx\n", perf_event_mask);
1457 perf_cpu_notifier(x86_pmu_notifier);
1460 static inline void x86_pmu_read(struct perf_event *event)
1462 x86_perf_event_update(event);
1465 static const struct pmu pmu = {
1466 .enable = x86_pmu_enable,
1467 .disable = x86_pmu_disable,
1468 .start = x86_pmu_start,
1469 .stop = x86_pmu_stop,
1470 .read = x86_pmu_read,
1471 .unthrottle = x86_pmu_unthrottle,
1475 * validate a single event group
1477 * validation include:
1478 * - check events are compatible which each other
1479 * - events do not compete for the same counter
1480 * - number of events <= number of counters
1482 * validation ensures the group can be loaded onto the
1483 * PMU if it was the only group available.
1485 static int validate_group(struct perf_event *event)
1487 struct perf_event *leader = event->group_leader;
1488 struct cpu_hw_events *fake_cpuc;
1492 fake_cpuc = kmalloc(sizeof(*fake_cpuc), GFP_KERNEL | __GFP_ZERO);
1497 * the event is not yet connected with its
1498 * siblings therefore we must first collect
1499 * existing siblings, then add the new event
1500 * before we can simulate the scheduling
1503 n = collect_events(fake_cpuc, leader, true);
1507 fake_cpuc->n_events = n;
1508 n = collect_events(fake_cpuc, event, false);
1512 fake_cpuc->n_events = n;
1514 ret = x86_schedule_events(fake_cpuc, n, NULL);
1522 const struct pmu *hw_perf_event_init(struct perf_event *event)
1524 const struct pmu *tmp;
1527 err = __hw_perf_event_init(event);
1530 * we temporarily connect event to its pmu
1531 * such that validate_group() can classify
1532 * it as an x86 event using is_x86_event()
1537 if (event->group_leader != event)
1538 err = validate_group(event);
1544 event->destroy(event);
1545 return ERR_PTR(err);
1556 void callchain_store(struct perf_callchain_entry *entry, u64 ip)
1558 if (entry->nr < PERF_MAX_STACK_DEPTH)
1559 entry->ip[entry->nr++] = ip;
1562 static DEFINE_PER_CPU(struct perf_callchain_entry, pmc_irq_entry);
1563 static DEFINE_PER_CPU(struct perf_callchain_entry, pmc_nmi_entry);
1567 backtrace_warning_symbol(void *data, char *msg, unsigned long symbol)
1569 /* Ignore warnings */
1572 static void backtrace_warning(void *data, char *msg)
1574 /* Ignore warnings */
1577 static int backtrace_stack(void *data, char *name)
1582 static void backtrace_address(void *data, unsigned long addr, int reliable)
1584 struct perf_callchain_entry *entry = data;
1587 callchain_store(entry, addr);
1590 static const struct stacktrace_ops backtrace_ops = {
1591 .warning = backtrace_warning,
1592 .warning_symbol = backtrace_warning_symbol,
1593 .stack = backtrace_stack,
1594 .address = backtrace_address,
1595 .walk_stack = print_context_stack_bp,
1598 #include "../dumpstack.h"
1601 perf_callchain_kernel(struct pt_regs *regs, struct perf_callchain_entry *entry)
1603 callchain_store(entry, PERF_CONTEXT_KERNEL);
1604 callchain_store(entry, regs->ip);
1606 dump_trace(NULL, regs, NULL, regs->bp, &backtrace_ops, entry);
1610 * best effort, GUP based copy_from_user() that assumes IRQ or NMI context
1612 static unsigned long
1613 copy_from_user_nmi(void *to, const void __user *from, unsigned long n)
1615 unsigned long offset, addr = (unsigned long)from;
1616 int type = in_nmi() ? KM_NMI : KM_IRQ0;
1617 unsigned long size, len = 0;
1623 ret = __get_user_pages_fast(addr, 1, 0, &page);
1627 offset = addr & (PAGE_SIZE - 1);
1628 size = min(PAGE_SIZE - offset, n - len);
1630 map = kmap_atomic(page, type);
1631 memcpy(to, map+offset, size);
1632 kunmap_atomic(map, type);
1644 static int copy_stack_frame(const void __user *fp, struct stack_frame *frame)
1646 unsigned long bytes;
1648 bytes = copy_from_user_nmi(frame, fp, sizeof(*frame));
1650 return bytes == sizeof(*frame);
1654 perf_callchain_user(struct pt_regs *regs, struct perf_callchain_entry *entry)
1656 struct stack_frame frame;
1657 const void __user *fp;
1659 if (!user_mode(regs))
1660 regs = task_pt_regs(current);
1662 fp = (void __user *)regs->bp;
1664 callchain_store(entry, PERF_CONTEXT_USER);
1665 callchain_store(entry, regs->ip);
1667 while (entry->nr < PERF_MAX_STACK_DEPTH) {
1668 frame.next_frame = NULL;
1669 frame.return_address = 0;
1671 if (!copy_stack_frame(fp, &frame))
1674 if ((unsigned long)fp < regs->sp)
1677 callchain_store(entry, frame.return_address);
1678 fp = frame.next_frame;
1683 perf_do_callchain(struct pt_regs *regs, struct perf_callchain_entry *entry)
1690 is_user = user_mode(regs);
1692 if (is_user && current->state != TASK_RUNNING)
1696 perf_callchain_kernel(regs, entry);
1699 perf_callchain_user(regs, entry);
1702 struct perf_callchain_entry *perf_callchain(struct pt_regs *regs)
1704 struct perf_callchain_entry *entry;
1707 entry = &__get_cpu_var(pmc_nmi_entry);
1709 entry = &__get_cpu_var(pmc_irq_entry);
1713 perf_do_callchain(regs, entry);