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 int type = in_nmi() ? KM_NMI : KM_IRQ0;
53 unsigned long size, len = 0;
59 ret = __get_user_pages_fast(addr, 1, 0, &page);
63 offset = addr & (PAGE_SIZE - 1);
64 size = min(PAGE_SIZE - offset, n - len);
66 map = kmap_atomic(page, type);
67 memcpy(to, map+offset, size);
68 kunmap_atomic(map, type);
80 struct event_constraint {
82 unsigned long idxmsk[BITS_TO_LONGS(X86_PMC_IDX_MAX)];
91 int nb_id; /* NorthBridge id */
92 int refcnt; /* reference count */
93 struct perf_event *owners[X86_PMC_IDX_MAX];
94 struct event_constraint event_constraints[X86_PMC_IDX_MAX];
97 #define MAX_LBR_ENTRIES 16
99 struct cpu_hw_events {
101 * Generic x86 PMC bits
103 struct perf_event *events[X86_PMC_IDX_MAX]; /* in counter order */
104 unsigned long active_mask[BITS_TO_LONGS(X86_PMC_IDX_MAX)];
105 unsigned long running[BITS_TO_LONGS(X86_PMC_IDX_MAX)];
111 int assign[X86_PMC_IDX_MAX]; /* event to counter assignment */
112 u64 tags[X86_PMC_IDX_MAX];
113 struct perf_event *event_list[X86_PMC_IDX_MAX]; /* in enabled order */
115 unsigned int group_flag;
118 * Intel DebugStore bits
120 struct debug_store *ds;
128 struct perf_branch_stack lbr_stack;
129 struct perf_branch_entry lbr_entries[MAX_LBR_ENTRIES];
134 struct amd_nb *amd_nb;
137 #define __EVENT_CONSTRAINT(c, n, m, w) {\
138 { .idxmsk64 = (n) }, \
144 #define EVENT_CONSTRAINT(c, n, m) \
145 __EVENT_CONSTRAINT(c, n, m, HWEIGHT(n))
148 * Constraint on the Event code.
150 #define INTEL_EVENT_CONSTRAINT(c, n) \
151 EVENT_CONSTRAINT(c, n, ARCH_PERFMON_EVENTSEL_EVENT)
154 * Constraint on the Event code + UMask + fixed-mask
156 * filter mask to validate fixed counter events.
157 * the following filters disqualify for fixed counters:
161 * The other filters are supported by fixed counters.
162 * The any-thread option is supported starting with v3.
164 #define FIXED_EVENT_CONSTRAINT(c, n) \
165 EVENT_CONSTRAINT(c, (1ULL << (32+n)), X86_RAW_EVENT_MASK)
168 * Constraint on the Event code + UMask
170 #define PEBS_EVENT_CONSTRAINT(c, n) \
171 EVENT_CONSTRAINT(c, n, INTEL_ARCH_EVENT_MASK)
173 #define EVENT_CONSTRAINT_END \
174 EVENT_CONSTRAINT(0, 0, 0)
176 #define for_each_event_constraint(e, c) \
177 for ((e) = (c); (e)->weight; (e)++)
179 union perf_capabilities {
183 u64 pebs_arch_reg : 1;
191 * struct x86_pmu - generic x86 pmu
195 * Generic x86 PMC bits
199 int (*handle_irq)(struct pt_regs *);
200 void (*disable_all)(void);
201 void (*enable_all)(int added);
202 void (*enable)(struct perf_event *);
203 void (*disable)(struct perf_event *);
204 int (*hw_config)(struct perf_event *event);
205 int (*schedule_events)(struct cpu_hw_events *cpuc, int n, int *assign);
208 u64 (*event_map)(int);
211 int num_counters_fixed;
216 struct event_constraint *
217 (*get_event_constraints)(struct cpu_hw_events *cpuc,
218 struct perf_event *event);
220 void (*put_event_constraints)(struct cpu_hw_events *cpuc,
221 struct perf_event *event);
222 struct event_constraint *event_constraints;
223 void (*quirks)(void);
224 int perfctr_second_write;
226 int (*cpu_prepare)(int cpu);
227 void (*cpu_starting)(int cpu);
228 void (*cpu_dying)(int cpu);
229 void (*cpu_dead)(int cpu);
232 * Intel Arch Perfmon v2+
235 union perf_capabilities intel_cap;
238 * Intel DebugStore bits
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 if (nmi_watchdog == NMI_LOCAL_APIC)
334 disable_lapic_nmi_watchdog();
336 for (i = 0; i < x86_pmu.num_counters; i++) {
337 if (!reserve_perfctr_nmi(x86_pmu.perfctr + i))
341 for (i = 0; i < x86_pmu.num_counters; i++) {
342 if (!reserve_evntsel_nmi(x86_pmu.eventsel + i))
349 for (i--; i >= 0; i--)
350 release_evntsel_nmi(x86_pmu.eventsel + i);
352 i = x86_pmu.num_counters;
355 for (i--; i >= 0; i--)
356 release_perfctr_nmi(x86_pmu.perfctr + i);
358 if (nmi_watchdog == NMI_LOCAL_APIC)
359 enable_lapic_nmi_watchdog();
364 static void release_pmc_hardware(void)
368 for (i = 0; i < x86_pmu.num_counters; i++) {
369 release_perfctr_nmi(x86_pmu.perfctr + i);
370 release_evntsel_nmi(x86_pmu.eventsel + i);
373 if (nmi_watchdog == NMI_LOCAL_APIC)
374 enable_lapic_nmi_watchdog();
379 static bool reserve_pmc_hardware(void) { return true; }
380 static void release_pmc_hardware(void) {}
384 static int reserve_ds_buffers(void);
385 static void release_ds_buffers(void);
387 static void hw_perf_event_destroy(struct perf_event *event)
389 if (atomic_dec_and_mutex_lock(&active_events, &pmc_reserve_mutex)) {
390 release_pmc_hardware();
391 release_ds_buffers();
392 mutex_unlock(&pmc_reserve_mutex);
396 static inline int x86_pmu_initialized(void)
398 return x86_pmu.handle_irq != NULL;
402 set_ext_hw_attr(struct hw_perf_event *hwc, struct perf_event_attr *attr)
404 unsigned int cache_type, cache_op, cache_result;
407 config = attr->config;
409 cache_type = (config >> 0) & 0xff;
410 if (cache_type >= PERF_COUNT_HW_CACHE_MAX)
413 cache_op = (config >> 8) & 0xff;
414 if (cache_op >= PERF_COUNT_HW_CACHE_OP_MAX)
417 cache_result = (config >> 16) & 0xff;
418 if (cache_result >= PERF_COUNT_HW_CACHE_RESULT_MAX)
421 val = hw_cache_event_ids[cache_type][cache_op][cache_result];
434 static int x86_setup_perfctr(struct perf_event *event)
436 struct perf_event_attr *attr = &event->attr;
437 struct hw_perf_event *hwc = &event->hw;
440 if (!hwc->sample_period) {
441 hwc->sample_period = x86_pmu.max_period;
442 hwc->last_period = hwc->sample_period;
443 local64_set(&hwc->period_left, hwc->sample_period);
446 * If we have a PMU initialized but no APIC
447 * interrupts, we cannot sample hardware
448 * events (user-space has to fall back and
449 * sample via a hrtimer based software event):
455 if (attr->type == PERF_TYPE_RAW)
458 if (attr->type == PERF_TYPE_HW_CACHE)
459 return set_ext_hw_attr(hwc, attr);
461 if (attr->config >= x86_pmu.max_events)
467 config = x86_pmu.event_map(attr->config);
478 if ((attr->config == PERF_COUNT_HW_BRANCH_INSTRUCTIONS) &&
479 (hwc->sample_period == 1)) {
480 /* BTS is not supported by this architecture. */
484 /* BTS is currently only allowed for user-mode. */
485 if (!attr->exclude_kernel)
489 hwc->config |= config;
494 static int x86_pmu_hw_config(struct perf_event *event)
496 if (event->attr.precise_ip) {
499 /* Support for constant skid */
503 /* Support for IP fixup */
508 if (event->attr.precise_ip > precise)
514 * (keep 'enabled' bit clear for now)
516 event->hw.config = ARCH_PERFMON_EVENTSEL_INT;
519 * Count user and OS events unless requested not to
521 if (!event->attr.exclude_user)
522 event->hw.config |= ARCH_PERFMON_EVENTSEL_USR;
523 if (!event->attr.exclude_kernel)
524 event->hw.config |= ARCH_PERFMON_EVENTSEL_OS;
526 if (event->attr.type == PERF_TYPE_RAW)
527 event->hw.config |= event->attr.config & X86_RAW_EVENT_MASK;
529 return x86_setup_perfctr(event);
533 * Setup the hardware configuration for a given attr_type
535 static int __x86_pmu_event_init(struct perf_event *event)
539 if (!x86_pmu_initialized())
543 if (!atomic_inc_not_zero(&active_events)) {
544 mutex_lock(&pmc_reserve_mutex);
545 if (atomic_read(&active_events) == 0) {
546 if (!reserve_pmc_hardware())
549 err = reserve_ds_buffers();
551 release_pmc_hardware();
555 atomic_inc(&active_events);
556 mutex_unlock(&pmc_reserve_mutex);
561 event->destroy = hw_perf_event_destroy;
564 event->hw.last_cpu = -1;
565 event->hw.last_tag = ~0ULL;
567 return x86_pmu.hw_config(event);
570 static void x86_pmu_disable_all(void)
572 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
575 for (idx = 0; idx < x86_pmu.num_counters; idx++) {
578 if (!test_bit(idx, cpuc->active_mask))
580 rdmsrl(x86_pmu.eventsel + idx, val);
581 if (!(val & ARCH_PERFMON_EVENTSEL_ENABLE))
583 val &= ~ARCH_PERFMON_EVENTSEL_ENABLE;
584 wrmsrl(x86_pmu.eventsel + idx, val);
588 static void x86_pmu_disable(struct pmu *pmu)
590 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
592 if (!x86_pmu_initialized())
602 x86_pmu.disable_all();
605 static void x86_pmu_enable_all(int added)
607 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
610 for (idx = 0; idx < x86_pmu.num_counters; idx++) {
611 struct perf_event *event = cpuc->events[idx];
614 if (!test_bit(idx, cpuc->active_mask))
617 val = event->hw.config;
618 val |= ARCH_PERFMON_EVENTSEL_ENABLE;
619 wrmsrl(x86_pmu.eventsel + idx, val);
623 static struct pmu pmu;
625 static inline int is_x86_event(struct perf_event *event)
627 return event->pmu == &pmu;
630 static int x86_schedule_events(struct cpu_hw_events *cpuc, int n, int *assign)
632 struct event_constraint *c, *constraints[X86_PMC_IDX_MAX];
633 unsigned long used_mask[BITS_TO_LONGS(X86_PMC_IDX_MAX)];
634 int i, j, w, wmax, num = 0;
635 struct hw_perf_event *hwc;
637 bitmap_zero(used_mask, X86_PMC_IDX_MAX);
639 for (i = 0; i < n; i++) {
640 c = x86_pmu.get_event_constraints(cpuc, cpuc->event_list[i]);
645 * fastpath, try to reuse previous register
647 for (i = 0; i < n; i++) {
648 hwc = &cpuc->event_list[i]->hw;
655 /* constraint still honored */
656 if (!test_bit(hwc->idx, c->idxmsk))
659 /* not already used */
660 if (test_bit(hwc->idx, used_mask))
663 __set_bit(hwc->idx, used_mask);
665 assign[i] = hwc->idx;
674 bitmap_zero(used_mask, X86_PMC_IDX_MAX);
677 * weight = number of possible counters
679 * 1 = most constrained, only works on one counter
680 * wmax = least constrained, works on any counter
682 * assign events to counters starting with most
683 * constrained events.
685 wmax = x86_pmu.num_counters;
688 * when fixed event counters are present,
689 * wmax is incremented by 1 to account
690 * for one more choice
692 if (x86_pmu.num_counters_fixed)
695 for (w = 1, num = n; num && w <= wmax; w++) {
697 for (i = 0; num && i < n; i++) {
699 hwc = &cpuc->event_list[i]->hw;
704 for_each_set_bit(j, c->idxmsk, X86_PMC_IDX_MAX) {
705 if (!test_bit(j, used_mask))
709 if (j == X86_PMC_IDX_MAX)
712 __set_bit(j, used_mask);
721 * scheduling failed or is just a simulation,
722 * free resources if necessary
724 if (!assign || num) {
725 for (i = 0; i < n; i++) {
726 if (x86_pmu.put_event_constraints)
727 x86_pmu.put_event_constraints(cpuc, cpuc->event_list[i]);
730 return num ? -ENOSPC : 0;
734 * dogrp: true if must collect siblings events (group)
735 * returns total number of events and error code
737 static int collect_events(struct cpu_hw_events *cpuc, struct perf_event *leader, bool dogrp)
739 struct perf_event *event;
742 max_count = x86_pmu.num_counters + x86_pmu.num_counters_fixed;
744 /* current number of events already accepted */
747 if (is_x86_event(leader)) {
750 cpuc->event_list[n] = leader;
756 list_for_each_entry(event, &leader->sibling_list, group_entry) {
757 if (!is_x86_event(event) ||
758 event->state <= PERF_EVENT_STATE_OFF)
764 cpuc->event_list[n] = event;
770 static inline void x86_assign_hw_event(struct perf_event *event,
771 struct cpu_hw_events *cpuc, int i)
773 struct hw_perf_event *hwc = &event->hw;
775 hwc->idx = cpuc->assign[i];
776 hwc->last_cpu = smp_processor_id();
777 hwc->last_tag = ++cpuc->tags[i];
779 if (hwc->idx == X86_PMC_IDX_FIXED_BTS) {
780 hwc->config_base = 0;
782 } else if (hwc->idx >= X86_PMC_IDX_FIXED) {
783 hwc->config_base = MSR_ARCH_PERFMON_FIXED_CTR_CTRL;
785 * We set it so that event_base + idx in wrmsr/rdmsr maps to
786 * MSR_ARCH_PERFMON_FIXED_CTR0 ... CTR2:
789 MSR_ARCH_PERFMON_FIXED_CTR0 - X86_PMC_IDX_FIXED;
791 hwc->config_base = x86_pmu.eventsel;
792 hwc->event_base = x86_pmu.perfctr;
796 static inline int match_prev_assignment(struct hw_perf_event *hwc,
797 struct cpu_hw_events *cpuc,
800 return hwc->idx == cpuc->assign[i] &&
801 hwc->last_cpu == smp_processor_id() &&
802 hwc->last_tag == cpuc->tags[i];
805 static void x86_pmu_start(struct perf_event *event, int flags);
806 static void x86_pmu_stop(struct perf_event *event, int flags);
808 static void x86_pmu_enable(struct pmu *pmu)
810 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
811 struct perf_event *event;
812 struct hw_perf_event *hwc;
813 int i, added = cpuc->n_added;
815 if (!x86_pmu_initialized())
822 int n_running = cpuc->n_events - cpuc->n_added;
824 * apply assignment obtained either from
825 * hw_perf_group_sched_in() or x86_pmu_enable()
827 * step1: save events moving to new counters
828 * step2: reprogram moved events into new counters
830 for (i = 0; i < n_running; i++) {
831 event = cpuc->event_list[i];
835 * we can avoid reprogramming counter if:
836 * - assigned same counter as last time
837 * - running on same CPU as last time
838 * - no other event has used the counter since
840 if (hwc->idx == -1 ||
841 match_prev_assignment(hwc, cpuc, i))
845 * Ensure we don't accidentally enable a stopped
846 * counter simply because we rescheduled.
848 if (hwc->state & PERF_HES_STOPPED)
849 hwc->state |= PERF_HES_ARCH;
851 x86_pmu_stop(event, PERF_EF_UPDATE);
854 for (i = 0; i < cpuc->n_events; i++) {
855 event = cpuc->event_list[i];
858 if (!match_prev_assignment(hwc, cpuc, i))
859 x86_assign_hw_event(event, cpuc, i);
860 else if (i < n_running)
863 if (hwc->state & PERF_HES_ARCH)
866 x86_pmu_start(event, PERF_EF_RELOAD);
869 perf_events_lapic_init();
875 x86_pmu.enable_all(added);
878 static inline void __x86_pmu_enable_event(struct hw_perf_event *hwc,
881 wrmsrl(hwc->config_base + hwc->idx, hwc->config | enable_mask);
884 static inline void x86_pmu_disable_event(struct perf_event *event)
886 struct hw_perf_event *hwc = &event->hw;
888 wrmsrl(hwc->config_base + hwc->idx, hwc->config);
891 static DEFINE_PER_CPU(u64 [X86_PMC_IDX_MAX], pmc_prev_left);
894 * Set the next IRQ period, based on the hwc->period_left value.
895 * To be called with the event disabled in hw:
898 x86_perf_event_set_period(struct perf_event *event)
900 struct hw_perf_event *hwc = &event->hw;
901 s64 left = local64_read(&hwc->period_left);
902 s64 period = hwc->sample_period;
903 int ret = 0, idx = hwc->idx;
905 if (idx == X86_PMC_IDX_FIXED_BTS)
909 * If we are way outside a reasonable range then just skip forward:
911 if (unlikely(left <= -period)) {
913 local64_set(&hwc->period_left, left);
914 hwc->last_period = period;
918 if (unlikely(left <= 0)) {
920 local64_set(&hwc->period_left, left);
921 hwc->last_period = period;
925 * Quirk: certain CPUs dont like it if just 1 hw_event is left:
927 if (unlikely(left < 2))
930 if (left > x86_pmu.max_period)
931 left = x86_pmu.max_period;
933 per_cpu(pmc_prev_left[idx], smp_processor_id()) = left;
936 * The hw event starts counting from this event offset,
937 * mark it to be able to extra future deltas:
939 local64_set(&hwc->prev_count, (u64)-left);
941 wrmsrl(hwc->event_base + idx, (u64)(-left) & x86_pmu.cntval_mask);
944 * Due to erratum on certan cpu we need
945 * a second write to be sure the register
946 * is updated properly
948 if (x86_pmu.perfctr_second_write) {
949 wrmsrl(hwc->event_base + idx,
950 (u64)(-left) & x86_pmu.cntval_mask);
953 perf_event_update_userpage(event);
958 static void x86_pmu_enable_event(struct perf_event *event)
960 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
962 __x86_pmu_enable_event(&event->hw,
963 ARCH_PERFMON_EVENTSEL_ENABLE);
967 * Add a single event to the PMU.
969 * The event is added to the group of enabled events
970 * but only if it can be scehduled with existing events.
972 static int x86_pmu_add(struct perf_event *event, int flags)
974 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
975 struct hw_perf_event *hwc;
976 int assign[X86_PMC_IDX_MAX];
981 perf_pmu_disable(event->pmu);
983 ret = n = collect_events(cpuc, event, false);
987 hwc->state = PERF_HES_UPTODATE | PERF_HES_STOPPED;
988 if (!(flags & PERF_EF_START))
989 hwc->state |= PERF_HES_ARCH;
992 * If group events scheduling transaction was started,
993 * skip the schedulability test here, it will be peformed
994 * at commit time (->commit_txn) as a whole
996 if (cpuc->group_flag & PERF_EVENT_TXN)
999 ret = x86_pmu.schedule_events(cpuc, n, assign);
1003 * copy new assignment, now we know it is possible
1004 * will be used by hw_perf_enable()
1006 memcpy(cpuc->assign, assign, n*sizeof(int));
1010 cpuc->n_added += n - n0;
1011 cpuc->n_txn += n - n0;
1015 perf_pmu_enable(event->pmu);
1019 static void x86_pmu_start(struct perf_event *event, int flags)
1021 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
1022 int idx = event->hw.idx;
1024 if (WARN_ON_ONCE(!(event->hw.state & PERF_HES_STOPPED)))
1027 if (WARN_ON_ONCE(idx == -1))
1030 if (flags & PERF_EF_RELOAD) {
1031 WARN_ON_ONCE(!(event->hw.state & PERF_HES_UPTODATE));
1032 x86_perf_event_set_period(event);
1035 event->hw.state = 0;
1037 cpuc->events[idx] = event;
1038 __set_bit(idx, cpuc->active_mask);
1039 __set_bit(idx, cpuc->running);
1040 x86_pmu.enable(event);
1041 perf_event_update_userpage(event);
1044 void perf_event_print_debug(void)
1046 u64 ctrl, status, overflow, pmc_ctrl, pmc_count, prev_left, fixed;
1048 struct cpu_hw_events *cpuc;
1049 unsigned long flags;
1052 if (!x86_pmu.num_counters)
1055 local_irq_save(flags);
1057 cpu = smp_processor_id();
1058 cpuc = &per_cpu(cpu_hw_events, cpu);
1060 if (x86_pmu.version >= 2) {
1061 rdmsrl(MSR_CORE_PERF_GLOBAL_CTRL, ctrl);
1062 rdmsrl(MSR_CORE_PERF_GLOBAL_STATUS, status);
1063 rdmsrl(MSR_CORE_PERF_GLOBAL_OVF_CTRL, overflow);
1064 rdmsrl(MSR_ARCH_PERFMON_FIXED_CTR_CTRL, fixed);
1065 rdmsrl(MSR_IA32_PEBS_ENABLE, pebs);
1068 pr_info("CPU#%d: ctrl: %016llx\n", cpu, ctrl);
1069 pr_info("CPU#%d: status: %016llx\n", cpu, status);
1070 pr_info("CPU#%d: overflow: %016llx\n", cpu, overflow);
1071 pr_info("CPU#%d: fixed: %016llx\n", cpu, fixed);
1072 pr_info("CPU#%d: pebs: %016llx\n", cpu, pebs);
1074 pr_info("CPU#%d: active: %016llx\n", cpu, *(u64 *)cpuc->active_mask);
1076 for (idx = 0; idx < x86_pmu.num_counters; idx++) {
1077 rdmsrl(x86_pmu.eventsel + idx, pmc_ctrl);
1078 rdmsrl(x86_pmu.perfctr + idx, pmc_count);
1080 prev_left = per_cpu(pmc_prev_left[idx], cpu);
1082 pr_info("CPU#%d: gen-PMC%d ctrl: %016llx\n",
1083 cpu, idx, pmc_ctrl);
1084 pr_info("CPU#%d: gen-PMC%d count: %016llx\n",
1085 cpu, idx, pmc_count);
1086 pr_info("CPU#%d: gen-PMC%d left: %016llx\n",
1087 cpu, idx, prev_left);
1089 for (idx = 0; idx < x86_pmu.num_counters_fixed; idx++) {
1090 rdmsrl(MSR_ARCH_PERFMON_FIXED_CTR0 + idx, pmc_count);
1092 pr_info("CPU#%d: fixed-PMC%d count: %016llx\n",
1093 cpu, idx, pmc_count);
1095 local_irq_restore(flags);
1098 static void x86_pmu_stop(struct perf_event *event, int flags)
1100 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
1101 struct hw_perf_event *hwc = &event->hw;
1103 if (__test_and_clear_bit(hwc->idx, cpuc->active_mask)) {
1104 x86_pmu.disable(event);
1105 cpuc->events[hwc->idx] = NULL;
1106 WARN_ON_ONCE(hwc->state & PERF_HES_STOPPED);
1107 hwc->state |= PERF_HES_STOPPED;
1110 if ((flags & PERF_EF_UPDATE) && !(hwc->state & PERF_HES_UPTODATE)) {
1112 * Drain the remaining delta count out of a event
1113 * that we are disabling:
1115 x86_perf_event_update(event);
1116 hwc->state |= PERF_HES_UPTODATE;
1120 static void x86_pmu_del(struct perf_event *event, int flags)
1122 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
1126 * If we're called during a txn, we don't need to do anything.
1127 * The events never got scheduled and ->cancel_txn will truncate
1130 if (cpuc->group_flag & PERF_EVENT_TXN)
1133 x86_pmu_stop(event, PERF_EF_UPDATE);
1135 for (i = 0; i < cpuc->n_events; i++) {
1136 if (event == cpuc->event_list[i]) {
1138 if (x86_pmu.put_event_constraints)
1139 x86_pmu.put_event_constraints(cpuc, event);
1141 while (++i < cpuc->n_events)
1142 cpuc->event_list[i-1] = cpuc->event_list[i];
1148 perf_event_update_userpage(event);
1151 static int x86_pmu_handle_irq(struct pt_regs *regs)
1153 struct perf_sample_data data;
1154 struct cpu_hw_events *cpuc;
1155 struct perf_event *event;
1156 int idx, handled = 0;
1159 perf_sample_data_init(&data, 0);
1161 cpuc = &__get_cpu_var(cpu_hw_events);
1163 for (idx = 0; idx < x86_pmu.num_counters; idx++) {
1164 if (!test_bit(idx, cpuc->active_mask)) {
1166 * Though we deactivated the counter some cpus
1167 * might still deliver spurious interrupts still
1168 * in flight. Catch them:
1170 if (__test_and_clear_bit(idx, cpuc->running))
1175 event = cpuc->events[idx];
1177 val = x86_perf_event_update(event);
1178 if (val & (1ULL << (x86_pmu.cntval_bits - 1)))
1185 data.period = event->hw.last_period;
1187 if (!x86_perf_event_set_period(event))
1190 if (perf_event_overflow(event, 1, &data, regs))
1191 x86_pmu_stop(event, 0);
1195 inc_irq_stat(apic_perf_irqs);
1200 void perf_events_lapic_init(void)
1202 if (!x86_pmu.apic || !x86_pmu_initialized())
1206 * Always use NMI for PMU
1208 apic_write(APIC_LVTPC, APIC_DM_NMI);
1211 struct pmu_nmi_state {
1212 unsigned int marked;
1216 static DEFINE_PER_CPU(struct pmu_nmi_state, pmu_nmi);
1218 static int __kprobes
1219 perf_event_nmi_handler(struct notifier_block *self,
1220 unsigned long cmd, void *__args)
1222 struct die_args *args = __args;
1223 unsigned int this_nmi;
1226 if (!atomic_read(&active_events))
1233 case DIE_NMIUNKNOWN:
1234 this_nmi = percpu_read(irq_stat.__nmi_count);
1235 if (this_nmi != __get_cpu_var(pmu_nmi).marked)
1236 /* let the kernel handle the unknown nmi */
1239 * This one is a PMU back-to-back nmi. Two events
1240 * trigger 'simultaneously' raising two back-to-back
1241 * NMIs. If the first NMI handles both, the latter
1242 * will be empty and daze the CPU. So, we drop it to
1243 * avoid false-positive 'unknown nmi' messages.
1250 apic_write(APIC_LVTPC, APIC_DM_NMI);
1252 handled = x86_pmu.handle_irq(args->regs);
1256 this_nmi = percpu_read(irq_stat.__nmi_count);
1257 if ((handled > 1) ||
1258 /* the next nmi could be a back-to-back nmi */
1259 ((__get_cpu_var(pmu_nmi).marked == this_nmi) &&
1260 (__get_cpu_var(pmu_nmi).handled > 1))) {
1262 * We could have two subsequent back-to-back nmis: The
1263 * first handles more than one counter, the 2nd
1264 * handles only one counter and the 3rd handles no
1267 * This is the 2nd nmi because the previous was
1268 * handling more than one counter. We will mark the
1269 * next (3rd) and then drop it if unhandled.
1271 __get_cpu_var(pmu_nmi).marked = this_nmi + 1;
1272 __get_cpu_var(pmu_nmi).handled = handled;
1278 static __read_mostly struct notifier_block perf_event_nmi_notifier = {
1279 .notifier_call = perf_event_nmi_handler,
1284 static struct event_constraint unconstrained;
1285 static struct event_constraint emptyconstraint;
1287 static struct event_constraint *
1288 x86_get_event_constraints(struct cpu_hw_events *cpuc, struct perf_event *event)
1290 struct event_constraint *c;
1292 if (x86_pmu.event_constraints) {
1293 for_each_event_constraint(c, x86_pmu.event_constraints) {
1294 if ((event->hw.config & c->cmask) == c->code)
1299 return &unconstrained;
1302 #include "perf_event_amd.c"
1303 #include "perf_event_p6.c"
1304 #include "perf_event_p4.c"
1305 #include "perf_event_intel_lbr.c"
1306 #include "perf_event_intel_ds.c"
1307 #include "perf_event_intel.c"
1309 static int __cpuinit
1310 x86_pmu_notifier(struct notifier_block *self, unsigned long action, void *hcpu)
1312 unsigned int cpu = (long)hcpu;
1313 int ret = NOTIFY_OK;
1315 switch (action & ~CPU_TASKS_FROZEN) {
1316 case CPU_UP_PREPARE:
1317 if (x86_pmu.cpu_prepare)
1318 ret = x86_pmu.cpu_prepare(cpu);
1322 if (x86_pmu.cpu_starting)
1323 x86_pmu.cpu_starting(cpu);
1327 if (x86_pmu.cpu_dying)
1328 x86_pmu.cpu_dying(cpu);
1331 case CPU_UP_CANCELED:
1333 if (x86_pmu.cpu_dead)
1334 x86_pmu.cpu_dead(cpu);
1344 static void __init pmu_check_apic(void)
1350 pr_info("no APIC, boot with the \"lapic\" boot parameter to force-enable it.\n");
1351 pr_info("no hardware sampling interrupt available.\n");
1354 void __init init_hw_perf_events(void)
1356 struct event_constraint *c;
1359 pr_info("Performance Events: ");
1361 switch (boot_cpu_data.x86_vendor) {
1362 case X86_VENDOR_INTEL:
1363 err = intel_pmu_init();
1365 case X86_VENDOR_AMD:
1366 err = amd_pmu_init();
1372 pr_cont("no PMU driver, software events only.\n");
1378 pr_cont("%s PMU driver.\n", x86_pmu.name);
1383 if (x86_pmu.num_counters > X86_PMC_MAX_GENERIC) {
1384 WARN(1, KERN_ERR "hw perf events %d > max(%d), clipping!",
1385 x86_pmu.num_counters, X86_PMC_MAX_GENERIC);
1386 x86_pmu.num_counters = X86_PMC_MAX_GENERIC;
1388 x86_pmu.intel_ctrl = (1 << x86_pmu.num_counters) - 1;
1390 if (x86_pmu.num_counters_fixed > X86_PMC_MAX_FIXED) {
1391 WARN(1, KERN_ERR "hw perf events fixed %d > max(%d), clipping!",
1392 x86_pmu.num_counters_fixed, X86_PMC_MAX_FIXED);
1393 x86_pmu.num_counters_fixed = X86_PMC_MAX_FIXED;
1396 x86_pmu.intel_ctrl |=
1397 ((1LL << x86_pmu.num_counters_fixed)-1) << X86_PMC_IDX_FIXED;
1399 perf_events_lapic_init();
1400 register_die_notifier(&perf_event_nmi_notifier);
1402 unconstrained = (struct event_constraint)
1403 __EVENT_CONSTRAINT(0, (1ULL << x86_pmu.num_counters) - 1,
1404 0, x86_pmu.num_counters);
1406 if (x86_pmu.event_constraints) {
1407 for_each_event_constraint(c, x86_pmu.event_constraints) {
1408 if (c->cmask != X86_RAW_EVENT_MASK)
1411 c->idxmsk64 |= (1ULL << x86_pmu.num_counters) - 1;
1412 c->weight += x86_pmu.num_counters;
1416 pr_info("... version: %d\n", x86_pmu.version);
1417 pr_info("... bit width: %d\n", x86_pmu.cntval_bits);
1418 pr_info("... generic registers: %d\n", x86_pmu.num_counters);
1419 pr_info("... value mask: %016Lx\n", x86_pmu.cntval_mask);
1420 pr_info("... max period: %016Lx\n", x86_pmu.max_period);
1421 pr_info("... fixed-purpose events: %d\n", x86_pmu.num_counters_fixed);
1422 pr_info("... event mask: %016Lx\n", x86_pmu.intel_ctrl);
1424 perf_pmu_register(&pmu);
1425 perf_cpu_notifier(x86_pmu_notifier);
1428 static inline void x86_pmu_read(struct perf_event *event)
1430 x86_perf_event_update(event);
1434 * Start group events scheduling transaction
1435 * Set the flag to make pmu::enable() not perform the
1436 * schedulability test, it will be performed at commit time
1438 static void x86_pmu_start_txn(struct pmu *pmu)
1440 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
1442 perf_pmu_disable(pmu);
1443 cpuc->group_flag |= PERF_EVENT_TXN;
1448 * Stop group events scheduling transaction
1449 * Clear the flag and pmu::enable() will perform the
1450 * schedulability test.
1452 static void x86_pmu_cancel_txn(struct pmu *pmu)
1454 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
1456 cpuc->group_flag &= ~PERF_EVENT_TXN;
1458 * Truncate the collected events.
1460 cpuc->n_added -= cpuc->n_txn;
1461 cpuc->n_events -= cpuc->n_txn;
1462 perf_pmu_enable(pmu);
1466 * Commit group events scheduling transaction
1467 * Perform the group schedulability test as a whole
1468 * Return 0 if success
1470 static int x86_pmu_commit_txn(struct pmu *pmu)
1472 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
1473 int assign[X86_PMC_IDX_MAX];
1478 if (!x86_pmu_initialized())
1481 ret = x86_pmu.schedule_events(cpuc, n, assign);
1486 * copy new assignment, now we know it is possible
1487 * will be used by hw_perf_enable()
1489 memcpy(cpuc->assign, assign, n*sizeof(int));
1491 cpuc->group_flag &= ~PERF_EVENT_TXN;
1492 perf_pmu_enable(pmu);
1497 * validate that we can schedule this event
1499 static int validate_event(struct perf_event *event)
1501 struct cpu_hw_events *fake_cpuc;
1502 struct event_constraint *c;
1505 fake_cpuc = kmalloc(sizeof(*fake_cpuc), GFP_KERNEL | __GFP_ZERO);
1509 c = x86_pmu.get_event_constraints(fake_cpuc, event);
1511 if (!c || !c->weight)
1514 if (x86_pmu.put_event_constraints)
1515 x86_pmu.put_event_constraints(fake_cpuc, event);
1523 * validate a single event group
1525 * validation include:
1526 * - check events are compatible which each other
1527 * - events do not compete for the same counter
1528 * - number of events <= number of counters
1530 * validation ensures the group can be loaded onto the
1531 * PMU if it was the only group available.
1533 static int validate_group(struct perf_event *event)
1535 struct perf_event *leader = event->group_leader;
1536 struct cpu_hw_events *fake_cpuc;
1540 fake_cpuc = kmalloc(sizeof(*fake_cpuc), GFP_KERNEL | __GFP_ZERO);
1545 * the event is not yet connected with its
1546 * siblings therefore we must first collect
1547 * existing siblings, then add the new event
1548 * before we can simulate the scheduling
1551 n = collect_events(fake_cpuc, leader, true);
1555 fake_cpuc->n_events = n;
1556 n = collect_events(fake_cpuc, event, false);
1560 fake_cpuc->n_events = n;
1562 ret = x86_pmu.schedule_events(fake_cpuc, n, NULL);
1570 int x86_pmu_event_init(struct perf_event *event)
1575 switch (event->attr.type) {
1577 case PERF_TYPE_HARDWARE:
1578 case PERF_TYPE_HW_CACHE:
1585 err = __x86_pmu_event_init(event);
1588 * we temporarily connect event to its pmu
1589 * such that validate_group() can classify
1590 * it as an x86 event using is_x86_event()
1595 if (event->group_leader != event)
1596 err = validate_group(event);
1598 err = validate_event(event);
1604 event->destroy(event);
1610 static struct pmu pmu = {
1611 .pmu_enable = x86_pmu_enable,
1612 .pmu_disable = x86_pmu_disable,
1614 .event_init = x86_pmu_event_init,
1618 .start = x86_pmu_start,
1619 .stop = x86_pmu_stop,
1620 .read = x86_pmu_read,
1622 .start_txn = x86_pmu_start_txn,
1623 .cancel_txn = x86_pmu_cancel_txn,
1624 .commit_txn = x86_pmu_commit_txn,
1632 backtrace_warning_symbol(void *data, char *msg, unsigned long symbol)
1634 /* Ignore warnings */
1637 static void backtrace_warning(void *data, char *msg)
1639 /* Ignore warnings */
1642 static int backtrace_stack(void *data, char *name)
1647 static void backtrace_address(void *data, unsigned long addr, int reliable)
1649 struct perf_callchain_entry *entry = data;
1651 perf_callchain_store(entry, addr);
1654 static const struct stacktrace_ops backtrace_ops = {
1655 .warning = backtrace_warning,
1656 .warning_symbol = backtrace_warning_symbol,
1657 .stack = backtrace_stack,
1658 .address = backtrace_address,
1659 .walk_stack = print_context_stack_bp,
1663 perf_callchain_kernel(struct perf_callchain_entry *entry, struct pt_regs *regs)
1665 if (perf_guest_cbs && perf_guest_cbs->is_in_guest()) {
1666 /* TODO: We don't support guest os callchain now */
1670 perf_callchain_store(entry, regs->ip);
1672 dump_trace(NULL, regs, NULL, regs->bp, &backtrace_ops, entry);
1675 #ifdef CONFIG_COMPAT
1677 perf_callchain_user32(struct pt_regs *regs, struct perf_callchain_entry *entry)
1679 /* 32-bit process in 64-bit kernel. */
1680 struct stack_frame_ia32 frame;
1681 const void __user *fp;
1683 if (!test_thread_flag(TIF_IA32))
1686 fp = compat_ptr(regs->bp);
1687 while (entry->nr < PERF_MAX_STACK_DEPTH) {
1688 unsigned long bytes;
1689 frame.next_frame = 0;
1690 frame.return_address = 0;
1692 bytes = copy_from_user_nmi(&frame, fp, sizeof(frame));
1693 if (bytes != sizeof(frame))
1696 if (fp < compat_ptr(regs->sp))
1699 perf_callchain_store(entry, frame.return_address);
1700 fp = compat_ptr(frame.next_frame);
1706 perf_callchain_user32(struct pt_regs *regs, struct perf_callchain_entry *entry)
1713 perf_callchain_user(struct perf_callchain_entry *entry, struct pt_regs *regs)
1715 struct stack_frame frame;
1716 const void __user *fp;
1718 if (perf_guest_cbs && perf_guest_cbs->is_in_guest()) {
1719 /* TODO: We don't support guest os callchain now */
1723 fp = (void __user *)regs->bp;
1725 perf_callchain_store(entry, regs->ip);
1727 if (perf_callchain_user32(regs, entry))
1730 while (entry->nr < PERF_MAX_STACK_DEPTH) {
1731 unsigned long bytes;
1732 frame.next_frame = NULL;
1733 frame.return_address = 0;
1735 bytes = copy_from_user_nmi(&frame, fp, sizeof(frame));
1736 if (bytes != sizeof(frame))
1739 if ((unsigned long)fp < regs->sp)
1742 perf_callchain_store(entry, frame.return_address);
1743 fp = frame.next_frame;
1747 unsigned long perf_instruction_pointer(struct pt_regs *regs)
1751 if (perf_guest_cbs && perf_guest_cbs->is_in_guest())
1752 ip = perf_guest_cbs->get_guest_ip();
1754 ip = instruction_pointer(regs);
1759 unsigned long perf_misc_flags(struct pt_regs *regs)
1763 if (perf_guest_cbs && perf_guest_cbs->is_in_guest()) {
1764 if (perf_guest_cbs->is_user_mode())
1765 misc |= PERF_RECORD_MISC_GUEST_USER;
1767 misc |= PERF_RECORD_MISC_GUEST_KERNEL;
1769 if (user_mode(regs))
1770 misc |= PERF_RECORD_MISC_USER;
1772 misc |= PERF_RECORD_MISC_KERNEL;
1775 if (regs->flags & PERF_EFLAGS_EXACT)
1776 misc |= PERF_RECORD_MISC_EXACT_IP;