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 INTEL_UEVENT_CONSTRAINT(c, n) \
170 EVENT_CONSTRAINT(c, n, INTEL_ARCH_EVENT_MASK)
171 #define PEBS_EVENT_CONSTRAINT(c, n) \
172 INTEL_UEVENT_CONSTRAINT(c, n)
174 #define EVENT_CONSTRAINT_END \
175 EVENT_CONSTRAINT(0, 0, 0)
177 #define for_each_event_constraint(e, c) \
178 for ((e) = (c); (e)->weight; (e)++)
180 union perf_capabilities {
184 u64 pebs_arch_reg : 1;
192 * struct x86_pmu - generic x86 pmu
196 * Generic x86 PMC bits
200 int (*handle_irq)(struct pt_regs *);
201 void (*disable_all)(void);
202 void (*enable_all)(int added);
203 void (*enable)(struct perf_event *);
204 void (*disable)(struct perf_event *);
205 int (*hw_config)(struct perf_event *event);
206 int (*schedule_events)(struct cpu_hw_events *cpuc, int n, int *assign);
209 u64 (*event_map)(int);
212 int num_counters_fixed;
217 struct event_constraint *
218 (*get_event_constraints)(struct cpu_hw_events *cpuc,
219 struct perf_event *event);
221 void (*put_event_constraints)(struct cpu_hw_events *cpuc,
222 struct perf_event *event);
223 struct event_constraint *event_constraints;
224 void (*quirks)(void);
225 int perfctr_second_write;
227 int (*cpu_prepare)(int cpu);
228 void (*cpu_starting)(int cpu);
229 void (*cpu_dying)(int cpu);
230 void (*cpu_dead)(int cpu);
233 * Intel Arch Perfmon v2+
236 union perf_capabilities intel_cap;
239 * Intel DebugStore bits
242 int bts_active, pebs_active;
243 int pebs_record_size;
244 void (*drain_pebs)(struct pt_regs *regs);
245 struct event_constraint *pebs_constraints;
250 unsigned long lbr_tos, lbr_from, lbr_to; /* MSR base regs */
251 int lbr_nr; /* hardware stack size */
254 static struct x86_pmu x86_pmu __read_mostly;
256 static DEFINE_PER_CPU(struct cpu_hw_events, cpu_hw_events) = {
260 static int x86_perf_event_set_period(struct perf_event *event);
263 * Generalized hw caching related hw_event table, filled
264 * in on a per model basis. A value of 0 means
265 * 'not supported', -1 means 'hw_event makes no sense on
266 * this CPU', any other value means the raw hw_event
270 #define C(x) PERF_COUNT_HW_CACHE_##x
272 static u64 __read_mostly hw_cache_event_ids
273 [PERF_COUNT_HW_CACHE_MAX]
274 [PERF_COUNT_HW_CACHE_OP_MAX]
275 [PERF_COUNT_HW_CACHE_RESULT_MAX];
278 * Propagate event elapsed time into the generic event.
279 * Can only be executed on the CPU where the event is active.
280 * Returns the delta events processed.
283 x86_perf_event_update(struct perf_event *event)
285 struct hw_perf_event *hwc = &event->hw;
286 int shift = 64 - x86_pmu.cntval_bits;
287 u64 prev_raw_count, new_raw_count;
291 if (idx == X86_PMC_IDX_FIXED_BTS)
295 * Careful: an NMI might modify the previous event value.
297 * Our tactic to handle this is to first atomically read and
298 * exchange a new raw count - then add that new-prev delta
299 * count to the generic event atomically:
302 prev_raw_count = local64_read(&hwc->prev_count);
303 rdmsrl(hwc->event_base, new_raw_count);
305 if (local64_cmpxchg(&hwc->prev_count, prev_raw_count,
306 new_raw_count) != prev_raw_count)
310 * Now we have the new raw value and have updated the prev
311 * timestamp already. We can now calculate the elapsed delta
312 * (event-)time and add that to the generic event.
314 * Careful, not all hw sign-extends above the physical width
317 delta = (new_raw_count << shift) - (prev_raw_count << shift);
320 local64_add(delta, &event->count);
321 local64_sub(delta, &hwc->period_left);
323 return new_raw_count;
326 /* using X86_FEATURE_PERFCTR_CORE to later implement ALTERNATIVE() here */
327 static inline int x86_pmu_addr_offset(int index)
329 if (boot_cpu_has(X86_FEATURE_PERFCTR_CORE))
334 static inline unsigned int x86_pmu_config_addr(int index)
336 return x86_pmu.eventsel + x86_pmu_addr_offset(index);
339 static inline unsigned int x86_pmu_event_addr(int index)
341 return x86_pmu.perfctr + x86_pmu_addr_offset(index);
344 static atomic_t active_events;
345 static DEFINE_MUTEX(pmc_reserve_mutex);
347 #ifdef CONFIG_X86_LOCAL_APIC
349 static bool reserve_pmc_hardware(void)
353 for (i = 0; i < x86_pmu.num_counters; i++) {
354 if (!reserve_perfctr_nmi(x86_pmu_event_addr(i)))
358 for (i = 0; i < x86_pmu.num_counters; i++) {
359 if (!reserve_evntsel_nmi(x86_pmu_config_addr(i)))
366 for (i--; i >= 0; i--)
367 release_evntsel_nmi(x86_pmu_config_addr(i));
369 i = x86_pmu.num_counters;
372 for (i--; i >= 0; i--)
373 release_perfctr_nmi(x86_pmu_event_addr(i));
378 static void release_pmc_hardware(void)
382 for (i = 0; i < x86_pmu.num_counters; i++) {
383 release_perfctr_nmi(x86_pmu_event_addr(i));
384 release_evntsel_nmi(x86_pmu_config_addr(i));
390 static bool reserve_pmc_hardware(void) { return true; }
391 static void release_pmc_hardware(void) {}
395 static bool check_hw_exists(void)
397 u64 val, val_new = 0;
401 * Check to see if the BIOS enabled any of the counters, if so
404 for (i = 0; i < x86_pmu.num_counters; i++) {
405 reg = x86_pmu_config_addr(i);
406 ret = rdmsrl_safe(reg, &val);
409 if (val & ARCH_PERFMON_EVENTSEL_ENABLE)
413 if (x86_pmu.num_counters_fixed) {
414 reg = MSR_ARCH_PERFMON_FIXED_CTR_CTRL;
415 ret = rdmsrl_safe(reg, &val);
418 for (i = 0; i < x86_pmu.num_counters_fixed; i++) {
419 if (val & (0x03 << i*4))
425 * Now write a value and read it back to see if it matches,
426 * this is needed to detect certain hardware emulators (qemu/kvm)
427 * that don't trap on the MSR access and always return 0s.
430 ret = checking_wrmsrl(x86_pmu_event_addr(0), val);
431 ret |= rdmsrl_safe(x86_pmu_event_addr(0), &val_new);
432 if (ret || val != val_new)
438 printk(KERN_CONT "Broken BIOS detected, using software events only.\n");
439 printk(KERN_ERR FW_BUG "the BIOS has corrupted hw-PMU resources (MSR %x is %Lx)\n", reg, val);
443 printk(KERN_CONT "Broken PMU hardware detected, using software events only.\n");
447 static void reserve_ds_buffers(void);
448 static void release_ds_buffers(void);
450 static void hw_perf_event_destroy(struct perf_event *event)
452 if (atomic_dec_and_mutex_lock(&active_events, &pmc_reserve_mutex)) {
453 release_pmc_hardware();
454 release_ds_buffers();
455 mutex_unlock(&pmc_reserve_mutex);
459 static inline int x86_pmu_initialized(void)
461 return x86_pmu.handle_irq != NULL;
465 set_ext_hw_attr(struct hw_perf_event *hwc, struct perf_event_attr *attr)
467 unsigned int cache_type, cache_op, cache_result;
470 config = attr->config;
472 cache_type = (config >> 0) & 0xff;
473 if (cache_type >= PERF_COUNT_HW_CACHE_MAX)
476 cache_op = (config >> 8) & 0xff;
477 if (cache_op >= PERF_COUNT_HW_CACHE_OP_MAX)
480 cache_result = (config >> 16) & 0xff;
481 if (cache_result >= PERF_COUNT_HW_CACHE_RESULT_MAX)
484 val = hw_cache_event_ids[cache_type][cache_op][cache_result];
497 static int x86_setup_perfctr(struct perf_event *event)
499 struct perf_event_attr *attr = &event->attr;
500 struct hw_perf_event *hwc = &event->hw;
503 if (!is_sampling_event(event)) {
504 hwc->sample_period = x86_pmu.max_period;
505 hwc->last_period = hwc->sample_period;
506 local64_set(&hwc->period_left, hwc->sample_period);
509 * If we have a PMU initialized but no APIC
510 * interrupts, we cannot sample hardware
511 * events (user-space has to fall back and
512 * sample via a hrtimer based software event):
518 if (attr->type == PERF_TYPE_RAW)
521 if (attr->type == PERF_TYPE_HW_CACHE)
522 return set_ext_hw_attr(hwc, attr);
524 if (attr->config >= x86_pmu.max_events)
530 config = x86_pmu.event_map(attr->config);
541 if ((attr->config == PERF_COUNT_HW_BRANCH_INSTRUCTIONS) &&
542 (hwc->sample_period == 1)) {
543 /* BTS is not supported by this architecture. */
544 if (!x86_pmu.bts_active)
547 /* BTS is currently only allowed for user-mode. */
548 if (!attr->exclude_kernel)
552 hwc->config |= config;
557 static int x86_pmu_hw_config(struct perf_event *event)
559 if (event->attr.precise_ip) {
562 /* Support for constant skid */
563 if (x86_pmu.pebs_active) {
566 /* Support for IP fixup */
571 if (event->attr.precise_ip > precise)
577 * (keep 'enabled' bit clear for now)
579 event->hw.config = ARCH_PERFMON_EVENTSEL_INT;
582 * Count user and OS events unless requested not to
584 if (!event->attr.exclude_user)
585 event->hw.config |= ARCH_PERFMON_EVENTSEL_USR;
586 if (!event->attr.exclude_kernel)
587 event->hw.config |= ARCH_PERFMON_EVENTSEL_OS;
589 if (event->attr.type == PERF_TYPE_RAW)
590 event->hw.config |= event->attr.config & X86_RAW_EVENT_MASK;
592 return x86_setup_perfctr(event);
596 * Setup the hardware configuration for a given attr_type
598 static int __x86_pmu_event_init(struct perf_event *event)
602 if (!x86_pmu_initialized())
606 if (!atomic_inc_not_zero(&active_events)) {
607 mutex_lock(&pmc_reserve_mutex);
608 if (atomic_read(&active_events) == 0) {
609 if (!reserve_pmc_hardware())
612 reserve_ds_buffers();
615 atomic_inc(&active_events);
616 mutex_unlock(&pmc_reserve_mutex);
621 event->destroy = hw_perf_event_destroy;
624 event->hw.last_cpu = -1;
625 event->hw.last_tag = ~0ULL;
627 return x86_pmu.hw_config(event);
630 static void x86_pmu_disable_all(void)
632 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
635 for (idx = 0; idx < x86_pmu.num_counters; idx++) {
638 if (!test_bit(idx, cpuc->active_mask))
640 rdmsrl(x86_pmu_config_addr(idx), val);
641 if (!(val & ARCH_PERFMON_EVENTSEL_ENABLE))
643 val &= ~ARCH_PERFMON_EVENTSEL_ENABLE;
644 wrmsrl(x86_pmu_config_addr(idx), val);
648 static void x86_pmu_disable(struct pmu *pmu)
650 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
652 if (!x86_pmu_initialized())
662 x86_pmu.disable_all();
665 static inline void __x86_pmu_enable_event(struct hw_perf_event *hwc,
668 wrmsrl(hwc->config_base, hwc->config | enable_mask);
671 static void x86_pmu_enable_all(int added)
673 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
676 for (idx = 0; idx < x86_pmu.num_counters; idx++) {
677 struct hw_perf_event *hwc = &cpuc->events[idx]->hw;
679 if (!test_bit(idx, cpuc->active_mask))
682 __x86_pmu_enable_event(hwc, ARCH_PERFMON_EVENTSEL_ENABLE);
686 static struct pmu pmu;
688 static inline int is_x86_event(struct perf_event *event)
690 return event->pmu == &pmu;
693 static int x86_schedule_events(struct cpu_hw_events *cpuc, int n, int *assign)
695 struct event_constraint *c, *constraints[X86_PMC_IDX_MAX];
696 unsigned long used_mask[BITS_TO_LONGS(X86_PMC_IDX_MAX)];
697 int i, j, w, wmax, num = 0;
698 struct hw_perf_event *hwc;
700 bitmap_zero(used_mask, X86_PMC_IDX_MAX);
702 for (i = 0; i < n; i++) {
703 c = x86_pmu.get_event_constraints(cpuc, cpuc->event_list[i]);
708 * fastpath, try to reuse previous register
710 for (i = 0; i < n; i++) {
711 hwc = &cpuc->event_list[i]->hw;
718 /* constraint still honored */
719 if (!test_bit(hwc->idx, c->idxmsk))
722 /* not already used */
723 if (test_bit(hwc->idx, used_mask))
726 __set_bit(hwc->idx, used_mask);
728 assign[i] = hwc->idx;
737 bitmap_zero(used_mask, X86_PMC_IDX_MAX);
740 * weight = number of possible counters
742 * 1 = most constrained, only works on one counter
743 * wmax = least constrained, works on any counter
745 * assign events to counters starting with most
746 * constrained events.
748 wmax = x86_pmu.num_counters;
751 * when fixed event counters are present,
752 * wmax is incremented by 1 to account
753 * for one more choice
755 if (x86_pmu.num_counters_fixed)
758 for (w = 1, num = n; num && w <= wmax; w++) {
760 for (i = 0; num && i < n; i++) {
762 hwc = &cpuc->event_list[i]->hw;
767 for_each_set_bit(j, c->idxmsk, X86_PMC_IDX_MAX) {
768 if (!test_bit(j, used_mask))
772 if (j == X86_PMC_IDX_MAX)
775 __set_bit(j, used_mask);
784 * scheduling failed or is just a simulation,
785 * free resources if necessary
787 if (!assign || num) {
788 for (i = 0; i < n; i++) {
789 if (x86_pmu.put_event_constraints)
790 x86_pmu.put_event_constraints(cpuc, cpuc->event_list[i]);
793 return num ? -ENOSPC : 0;
797 * dogrp: true if must collect siblings events (group)
798 * returns total number of events and error code
800 static int collect_events(struct cpu_hw_events *cpuc, struct perf_event *leader, bool dogrp)
802 struct perf_event *event;
805 max_count = x86_pmu.num_counters + x86_pmu.num_counters_fixed;
807 /* current number of events already accepted */
810 if (is_x86_event(leader)) {
813 cpuc->event_list[n] = leader;
819 list_for_each_entry(event, &leader->sibling_list, group_entry) {
820 if (!is_x86_event(event) ||
821 event->state <= PERF_EVENT_STATE_OFF)
827 cpuc->event_list[n] = event;
833 static inline void x86_assign_hw_event(struct perf_event *event,
834 struct cpu_hw_events *cpuc, int i)
836 struct hw_perf_event *hwc = &event->hw;
838 hwc->idx = cpuc->assign[i];
839 hwc->last_cpu = smp_processor_id();
840 hwc->last_tag = ++cpuc->tags[i];
842 if (hwc->idx == X86_PMC_IDX_FIXED_BTS) {
843 hwc->config_base = 0;
845 } else if (hwc->idx >= X86_PMC_IDX_FIXED) {
846 hwc->config_base = MSR_ARCH_PERFMON_FIXED_CTR_CTRL;
847 hwc->event_base = MSR_ARCH_PERFMON_FIXED_CTR0;
849 hwc->config_base = x86_pmu_config_addr(hwc->idx);
850 hwc->event_base = x86_pmu_event_addr(hwc->idx);
854 static inline int match_prev_assignment(struct hw_perf_event *hwc,
855 struct cpu_hw_events *cpuc,
858 return hwc->idx == cpuc->assign[i] &&
859 hwc->last_cpu == smp_processor_id() &&
860 hwc->last_tag == cpuc->tags[i];
863 static void x86_pmu_start(struct perf_event *event, int flags);
864 static void x86_pmu_stop(struct perf_event *event, int flags);
866 static void x86_pmu_enable(struct pmu *pmu)
868 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
869 struct perf_event *event;
870 struct hw_perf_event *hwc;
871 int i, added = cpuc->n_added;
873 if (!x86_pmu_initialized())
880 int n_running = cpuc->n_events - cpuc->n_added;
882 * apply assignment obtained either from
883 * hw_perf_group_sched_in() or x86_pmu_enable()
885 * step1: save events moving to new counters
886 * step2: reprogram moved events into new counters
888 for (i = 0; i < n_running; i++) {
889 event = cpuc->event_list[i];
893 * we can avoid reprogramming counter if:
894 * - assigned same counter as last time
895 * - running on same CPU as last time
896 * - no other event has used the counter since
898 if (hwc->idx == -1 ||
899 match_prev_assignment(hwc, cpuc, i))
903 * Ensure we don't accidentally enable a stopped
904 * counter simply because we rescheduled.
906 if (hwc->state & PERF_HES_STOPPED)
907 hwc->state |= PERF_HES_ARCH;
909 x86_pmu_stop(event, PERF_EF_UPDATE);
912 for (i = 0; i < cpuc->n_events; i++) {
913 event = cpuc->event_list[i];
916 if (!match_prev_assignment(hwc, cpuc, i))
917 x86_assign_hw_event(event, cpuc, i);
918 else if (i < n_running)
921 if (hwc->state & PERF_HES_ARCH)
924 x86_pmu_start(event, PERF_EF_RELOAD);
927 perf_events_lapic_init();
933 x86_pmu.enable_all(added);
936 static inline void x86_pmu_disable_event(struct perf_event *event)
938 struct hw_perf_event *hwc = &event->hw;
940 wrmsrl(hwc->config_base, hwc->config);
943 static DEFINE_PER_CPU(u64 [X86_PMC_IDX_MAX], pmc_prev_left);
946 * Set the next IRQ period, based on the hwc->period_left value.
947 * To be called with the event disabled in hw:
950 x86_perf_event_set_period(struct perf_event *event)
952 struct hw_perf_event *hwc = &event->hw;
953 s64 left = local64_read(&hwc->period_left);
954 s64 period = hwc->sample_period;
955 int ret = 0, idx = hwc->idx;
957 if (idx == X86_PMC_IDX_FIXED_BTS)
961 * If we are way outside a reasonable range then just skip forward:
963 if (unlikely(left <= -period)) {
965 local64_set(&hwc->period_left, left);
966 hwc->last_period = period;
970 if (unlikely(left <= 0)) {
972 local64_set(&hwc->period_left, left);
973 hwc->last_period = period;
977 * Quirk: certain CPUs dont like it if just 1 hw_event is left:
979 if (unlikely(left < 2))
982 if (left > x86_pmu.max_period)
983 left = x86_pmu.max_period;
985 per_cpu(pmc_prev_left[idx], smp_processor_id()) = left;
988 * The hw event starts counting from this event offset,
989 * mark it to be able to extra future deltas:
991 local64_set(&hwc->prev_count, (u64)-left);
993 wrmsrl(hwc->event_base, (u64)(-left) & x86_pmu.cntval_mask);
996 * Due to erratum on certan cpu we need
997 * a second write to be sure the register
998 * is updated properly
1000 if (x86_pmu.perfctr_second_write) {
1001 wrmsrl(hwc->event_base,
1002 (u64)(-left) & x86_pmu.cntval_mask);
1005 perf_event_update_userpage(event);
1010 static void x86_pmu_enable_event(struct perf_event *event)
1012 if (__this_cpu_read(cpu_hw_events.enabled))
1013 __x86_pmu_enable_event(&event->hw,
1014 ARCH_PERFMON_EVENTSEL_ENABLE);
1018 * Add a single event to the PMU.
1020 * The event is added to the group of enabled events
1021 * but only if it can be scehduled with existing events.
1023 static int x86_pmu_add(struct perf_event *event, int flags)
1025 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
1026 struct hw_perf_event *hwc;
1027 int assign[X86_PMC_IDX_MAX];
1032 perf_pmu_disable(event->pmu);
1033 n0 = cpuc->n_events;
1034 ret = n = collect_events(cpuc, event, false);
1038 hwc->state = PERF_HES_UPTODATE | PERF_HES_STOPPED;
1039 if (!(flags & PERF_EF_START))
1040 hwc->state |= PERF_HES_ARCH;
1043 * If group events scheduling transaction was started,
1044 * skip the schedulability test here, it will be peformed
1045 * at commit time (->commit_txn) as a whole
1047 if (cpuc->group_flag & PERF_EVENT_TXN)
1050 ret = x86_pmu.schedule_events(cpuc, n, assign);
1054 * copy new assignment, now we know it is possible
1055 * will be used by hw_perf_enable()
1057 memcpy(cpuc->assign, assign, n*sizeof(int));
1061 cpuc->n_added += n - n0;
1062 cpuc->n_txn += n - n0;
1066 perf_pmu_enable(event->pmu);
1070 static void x86_pmu_start(struct perf_event *event, int flags)
1072 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
1073 int idx = event->hw.idx;
1075 if (WARN_ON_ONCE(!(event->hw.state & PERF_HES_STOPPED)))
1078 if (WARN_ON_ONCE(idx == -1))
1081 if (flags & PERF_EF_RELOAD) {
1082 WARN_ON_ONCE(!(event->hw.state & PERF_HES_UPTODATE));
1083 x86_perf_event_set_period(event);
1086 event->hw.state = 0;
1088 cpuc->events[idx] = event;
1089 __set_bit(idx, cpuc->active_mask);
1090 __set_bit(idx, cpuc->running);
1091 x86_pmu.enable(event);
1092 perf_event_update_userpage(event);
1095 void perf_event_print_debug(void)
1097 u64 ctrl, status, overflow, pmc_ctrl, pmc_count, prev_left, fixed;
1099 struct cpu_hw_events *cpuc;
1100 unsigned long flags;
1103 if (!x86_pmu.num_counters)
1106 local_irq_save(flags);
1108 cpu = smp_processor_id();
1109 cpuc = &per_cpu(cpu_hw_events, cpu);
1111 if (x86_pmu.version >= 2) {
1112 rdmsrl(MSR_CORE_PERF_GLOBAL_CTRL, ctrl);
1113 rdmsrl(MSR_CORE_PERF_GLOBAL_STATUS, status);
1114 rdmsrl(MSR_CORE_PERF_GLOBAL_OVF_CTRL, overflow);
1115 rdmsrl(MSR_ARCH_PERFMON_FIXED_CTR_CTRL, fixed);
1116 rdmsrl(MSR_IA32_PEBS_ENABLE, pebs);
1119 pr_info("CPU#%d: ctrl: %016llx\n", cpu, ctrl);
1120 pr_info("CPU#%d: status: %016llx\n", cpu, status);
1121 pr_info("CPU#%d: overflow: %016llx\n", cpu, overflow);
1122 pr_info("CPU#%d: fixed: %016llx\n", cpu, fixed);
1123 pr_info("CPU#%d: pebs: %016llx\n", cpu, pebs);
1125 pr_info("CPU#%d: active: %016llx\n", cpu, *(u64 *)cpuc->active_mask);
1127 for (idx = 0; idx < x86_pmu.num_counters; idx++) {
1128 rdmsrl(x86_pmu_config_addr(idx), pmc_ctrl);
1129 rdmsrl(x86_pmu_event_addr(idx), pmc_count);
1131 prev_left = per_cpu(pmc_prev_left[idx], cpu);
1133 pr_info("CPU#%d: gen-PMC%d ctrl: %016llx\n",
1134 cpu, idx, pmc_ctrl);
1135 pr_info("CPU#%d: gen-PMC%d count: %016llx\n",
1136 cpu, idx, pmc_count);
1137 pr_info("CPU#%d: gen-PMC%d left: %016llx\n",
1138 cpu, idx, prev_left);
1140 for (idx = 0; idx < x86_pmu.num_counters_fixed; idx++) {
1141 rdmsrl(MSR_ARCH_PERFMON_FIXED_CTR0 + idx, pmc_count);
1143 pr_info("CPU#%d: fixed-PMC%d count: %016llx\n",
1144 cpu, idx, pmc_count);
1146 local_irq_restore(flags);
1149 static void x86_pmu_stop(struct perf_event *event, int flags)
1151 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
1152 struct hw_perf_event *hwc = &event->hw;
1154 if (__test_and_clear_bit(hwc->idx, cpuc->active_mask)) {
1155 x86_pmu.disable(event);
1156 cpuc->events[hwc->idx] = NULL;
1157 WARN_ON_ONCE(hwc->state & PERF_HES_STOPPED);
1158 hwc->state |= PERF_HES_STOPPED;
1161 if ((flags & PERF_EF_UPDATE) && !(hwc->state & PERF_HES_UPTODATE)) {
1163 * Drain the remaining delta count out of a event
1164 * that we are disabling:
1166 x86_perf_event_update(event);
1167 hwc->state |= PERF_HES_UPTODATE;
1171 static void x86_pmu_del(struct perf_event *event, int flags)
1173 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
1177 * If we're called during a txn, we don't need to do anything.
1178 * The events never got scheduled and ->cancel_txn will truncate
1181 if (cpuc->group_flag & PERF_EVENT_TXN)
1184 x86_pmu_stop(event, PERF_EF_UPDATE);
1186 for (i = 0; i < cpuc->n_events; i++) {
1187 if (event == cpuc->event_list[i]) {
1189 if (x86_pmu.put_event_constraints)
1190 x86_pmu.put_event_constraints(cpuc, event);
1192 while (++i < cpuc->n_events)
1193 cpuc->event_list[i-1] = cpuc->event_list[i];
1199 perf_event_update_userpage(event);
1202 static int x86_pmu_handle_irq(struct pt_regs *regs)
1204 struct perf_sample_data data;
1205 struct cpu_hw_events *cpuc;
1206 struct perf_event *event;
1207 int idx, handled = 0;
1210 perf_sample_data_init(&data, 0);
1212 cpuc = &__get_cpu_var(cpu_hw_events);
1214 for (idx = 0; idx < x86_pmu.num_counters; idx++) {
1215 if (!test_bit(idx, cpuc->active_mask)) {
1217 * Though we deactivated the counter some cpus
1218 * might still deliver spurious interrupts still
1219 * in flight. Catch them:
1221 if (__test_and_clear_bit(idx, cpuc->running))
1226 event = cpuc->events[idx];
1228 val = x86_perf_event_update(event);
1229 if (val & (1ULL << (x86_pmu.cntval_bits - 1)))
1236 data.period = event->hw.last_period;
1238 if (!x86_perf_event_set_period(event))
1241 if (perf_event_overflow(event, 1, &data, regs))
1242 x86_pmu_stop(event, 0);
1246 inc_irq_stat(apic_perf_irqs);
1251 void perf_events_lapic_init(void)
1253 if (!x86_pmu.apic || !x86_pmu_initialized())
1257 * Always use NMI for PMU
1259 apic_write(APIC_LVTPC, APIC_DM_NMI);
1262 struct pmu_nmi_state {
1263 unsigned int marked;
1267 static DEFINE_PER_CPU(struct pmu_nmi_state, pmu_nmi);
1269 static int __kprobes
1270 perf_event_nmi_handler(struct notifier_block *self,
1271 unsigned long cmd, void *__args)
1273 struct die_args *args = __args;
1274 unsigned int this_nmi;
1277 if (!atomic_read(&active_events))
1283 case DIE_NMIUNKNOWN:
1284 this_nmi = percpu_read(irq_stat.__nmi_count);
1285 if (this_nmi != __this_cpu_read(pmu_nmi.marked))
1286 /* let the kernel handle the unknown nmi */
1289 * This one is a PMU back-to-back nmi. Two events
1290 * trigger 'simultaneously' raising two back-to-back
1291 * NMIs. If the first NMI handles both, the latter
1292 * will be empty and daze the CPU. So, we drop it to
1293 * avoid false-positive 'unknown nmi' messages.
1300 apic_write(APIC_LVTPC, APIC_DM_NMI);
1302 handled = x86_pmu.handle_irq(args->regs);
1306 this_nmi = percpu_read(irq_stat.__nmi_count);
1307 if ((handled > 1) ||
1308 /* the next nmi could be a back-to-back nmi */
1309 ((__this_cpu_read(pmu_nmi.marked) == this_nmi) &&
1310 (__this_cpu_read(pmu_nmi.handled) > 1))) {
1312 * We could have two subsequent back-to-back nmis: The
1313 * first handles more than one counter, the 2nd
1314 * handles only one counter and the 3rd handles no
1317 * This is the 2nd nmi because the previous was
1318 * handling more than one counter. We will mark the
1319 * next (3rd) and then drop it if unhandled.
1321 __this_cpu_write(pmu_nmi.marked, this_nmi + 1);
1322 __this_cpu_write(pmu_nmi.handled, handled);
1328 static __read_mostly struct notifier_block perf_event_nmi_notifier = {
1329 .notifier_call = perf_event_nmi_handler,
1331 .priority = NMI_LOCAL_LOW_PRIOR,
1334 static struct event_constraint unconstrained;
1335 static struct event_constraint emptyconstraint;
1337 static struct event_constraint *
1338 x86_get_event_constraints(struct cpu_hw_events *cpuc, struct perf_event *event)
1340 struct event_constraint *c;
1342 if (x86_pmu.event_constraints) {
1343 for_each_event_constraint(c, x86_pmu.event_constraints) {
1344 if ((event->hw.config & c->cmask) == c->code)
1349 return &unconstrained;
1352 #include "perf_event_amd.c"
1353 #include "perf_event_p6.c"
1354 #include "perf_event_p4.c"
1355 #include "perf_event_intel_lbr.c"
1356 #include "perf_event_intel_ds.c"
1357 #include "perf_event_intel.c"
1359 static int __cpuinit
1360 x86_pmu_notifier(struct notifier_block *self, unsigned long action, void *hcpu)
1362 unsigned int cpu = (long)hcpu;
1363 int ret = NOTIFY_OK;
1365 switch (action & ~CPU_TASKS_FROZEN) {
1366 case CPU_UP_PREPARE:
1367 if (x86_pmu.cpu_prepare)
1368 ret = x86_pmu.cpu_prepare(cpu);
1372 if (x86_pmu.cpu_starting)
1373 x86_pmu.cpu_starting(cpu);
1377 if (x86_pmu.cpu_dying)
1378 x86_pmu.cpu_dying(cpu);
1381 case CPU_UP_CANCELED:
1383 if (x86_pmu.cpu_dead)
1384 x86_pmu.cpu_dead(cpu);
1394 static void __init pmu_check_apic(void)
1400 pr_info("no APIC, boot with the \"lapic\" boot parameter to force-enable it.\n");
1401 pr_info("no hardware sampling interrupt available.\n");
1404 static int __init init_hw_perf_events(void)
1406 struct event_constraint *c;
1409 pr_info("Performance Events: ");
1411 switch (boot_cpu_data.x86_vendor) {
1412 case X86_VENDOR_INTEL:
1413 err = intel_pmu_init();
1415 case X86_VENDOR_AMD:
1416 err = amd_pmu_init();
1422 pr_cont("no PMU driver, software events only.\n");
1428 /* sanity check that the hardware exists or is emulated */
1429 if (!check_hw_exists())
1432 pr_cont("%s PMU driver.\n", x86_pmu.name);
1437 if (x86_pmu.num_counters > X86_PMC_MAX_GENERIC) {
1438 WARN(1, KERN_ERR "hw perf events %d > max(%d), clipping!",
1439 x86_pmu.num_counters, X86_PMC_MAX_GENERIC);
1440 x86_pmu.num_counters = X86_PMC_MAX_GENERIC;
1442 x86_pmu.intel_ctrl = (1 << x86_pmu.num_counters) - 1;
1444 if (x86_pmu.num_counters_fixed > X86_PMC_MAX_FIXED) {
1445 WARN(1, KERN_ERR "hw perf events fixed %d > max(%d), clipping!",
1446 x86_pmu.num_counters_fixed, X86_PMC_MAX_FIXED);
1447 x86_pmu.num_counters_fixed = X86_PMC_MAX_FIXED;
1450 x86_pmu.intel_ctrl |=
1451 ((1LL << x86_pmu.num_counters_fixed)-1) << X86_PMC_IDX_FIXED;
1453 perf_events_lapic_init();
1454 register_die_notifier(&perf_event_nmi_notifier);
1456 unconstrained = (struct event_constraint)
1457 __EVENT_CONSTRAINT(0, (1ULL << x86_pmu.num_counters) - 1,
1458 0, x86_pmu.num_counters);
1460 if (x86_pmu.event_constraints) {
1461 for_each_event_constraint(c, x86_pmu.event_constraints) {
1462 if (c->cmask != X86_RAW_EVENT_MASK)
1465 c->idxmsk64 |= (1ULL << x86_pmu.num_counters) - 1;
1466 c->weight += x86_pmu.num_counters;
1470 pr_info("... version: %d\n", x86_pmu.version);
1471 pr_info("... bit width: %d\n", x86_pmu.cntval_bits);
1472 pr_info("... generic registers: %d\n", x86_pmu.num_counters);
1473 pr_info("... value mask: %016Lx\n", x86_pmu.cntval_mask);
1474 pr_info("... max period: %016Lx\n", x86_pmu.max_period);
1475 pr_info("... fixed-purpose events: %d\n", x86_pmu.num_counters_fixed);
1476 pr_info("... event mask: %016Lx\n", x86_pmu.intel_ctrl);
1478 perf_pmu_register(&pmu, "cpu", PERF_TYPE_RAW);
1479 perf_cpu_notifier(x86_pmu_notifier);
1483 early_initcall(init_hw_perf_events);
1485 static inline void x86_pmu_read(struct perf_event *event)
1487 x86_perf_event_update(event);
1491 * Start group events scheduling transaction
1492 * Set the flag to make pmu::enable() not perform the
1493 * schedulability test, it will be performed at commit time
1495 static void x86_pmu_start_txn(struct pmu *pmu)
1497 perf_pmu_disable(pmu);
1498 __this_cpu_or(cpu_hw_events.group_flag, PERF_EVENT_TXN);
1499 __this_cpu_write(cpu_hw_events.n_txn, 0);
1503 * Stop group events scheduling transaction
1504 * Clear the flag and pmu::enable() will perform the
1505 * schedulability test.
1507 static void x86_pmu_cancel_txn(struct pmu *pmu)
1509 __this_cpu_and(cpu_hw_events.group_flag, ~PERF_EVENT_TXN);
1511 * Truncate the collected events.
1513 __this_cpu_sub(cpu_hw_events.n_added, __this_cpu_read(cpu_hw_events.n_txn));
1514 __this_cpu_sub(cpu_hw_events.n_events, __this_cpu_read(cpu_hw_events.n_txn));
1515 perf_pmu_enable(pmu);
1519 * Commit group events scheduling transaction
1520 * Perform the group schedulability test as a whole
1521 * Return 0 if success
1523 static int x86_pmu_commit_txn(struct pmu *pmu)
1525 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
1526 int assign[X86_PMC_IDX_MAX];
1531 if (!x86_pmu_initialized())
1534 ret = x86_pmu.schedule_events(cpuc, n, assign);
1539 * copy new assignment, now we know it is possible
1540 * will be used by hw_perf_enable()
1542 memcpy(cpuc->assign, assign, n*sizeof(int));
1544 cpuc->group_flag &= ~PERF_EVENT_TXN;
1545 perf_pmu_enable(pmu);
1550 * validate that we can schedule this event
1552 static int validate_event(struct perf_event *event)
1554 struct cpu_hw_events *fake_cpuc;
1555 struct event_constraint *c;
1558 fake_cpuc = kmalloc(sizeof(*fake_cpuc), GFP_KERNEL | __GFP_ZERO);
1562 c = x86_pmu.get_event_constraints(fake_cpuc, event);
1564 if (!c || !c->weight)
1567 if (x86_pmu.put_event_constraints)
1568 x86_pmu.put_event_constraints(fake_cpuc, event);
1576 * validate a single event group
1578 * validation include:
1579 * - check events are compatible which each other
1580 * - events do not compete for the same counter
1581 * - number of events <= number of counters
1583 * validation ensures the group can be loaded onto the
1584 * PMU if it was the only group available.
1586 static int validate_group(struct perf_event *event)
1588 struct perf_event *leader = event->group_leader;
1589 struct cpu_hw_events *fake_cpuc;
1593 fake_cpuc = kmalloc(sizeof(*fake_cpuc), GFP_KERNEL | __GFP_ZERO);
1598 * the event is not yet connected with its
1599 * siblings therefore we must first collect
1600 * existing siblings, then add the new event
1601 * before we can simulate the scheduling
1604 n = collect_events(fake_cpuc, leader, true);
1608 fake_cpuc->n_events = n;
1609 n = collect_events(fake_cpuc, event, false);
1613 fake_cpuc->n_events = n;
1615 ret = x86_pmu.schedule_events(fake_cpuc, n, NULL);
1623 static int x86_pmu_event_init(struct perf_event *event)
1628 switch (event->attr.type) {
1630 case PERF_TYPE_HARDWARE:
1631 case PERF_TYPE_HW_CACHE:
1638 err = __x86_pmu_event_init(event);
1641 * we temporarily connect event to its pmu
1642 * such that validate_group() can classify
1643 * it as an x86 event using is_x86_event()
1648 if (event->group_leader != event)
1649 err = validate_group(event);
1651 err = validate_event(event);
1657 event->destroy(event);
1663 static struct pmu pmu = {
1664 .pmu_enable = x86_pmu_enable,
1665 .pmu_disable = x86_pmu_disable,
1667 .event_init = x86_pmu_event_init,
1671 .start = x86_pmu_start,
1672 .stop = x86_pmu_stop,
1673 .read = x86_pmu_read,
1675 .start_txn = x86_pmu_start_txn,
1676 .cancel_txn = x86_pmu_cancel_txn,
1677 .commit_txn = x86_pmu_commit_txn,
1685 backtrace_warning_symbol(void *data, char *msg, unsigned long symbol)
1687 /* Ignore warnings */
1690 static void backtrace_warning(void *data, char *msg)
1692 /* Ignore warnings */
1695 static int backtrace_stack(void *data, char *name)
1700 static void backtrace_address(void *data, unsigned long addr, int reliable)
1702 struct perf_callchain_entry *entry = data;
1704 perf_callchain_store(entry, addr);
1707 static const struct stacktrace_ops backtrace_ops = {
1708 .warning = backtrace_warning,
1709 .warning_symbol = backtrace_warning_symbol,
1710 .stack = backtrace_stack,
1711 .address = backtrace_address,
1712 .walk_stack = print_context_stack_bp,
1716 perf_callchain_kernel(struct perf_callchain_entry *entry, struct pt_regs *regs)
1718 if (perf_guest_cbs && perf_guest_cbs->is_in_guest()) {
1719 /* TODO: We don't support guest os callchain now */
1723 perf_callchain_store(entry, regs->ip);
1725 dump_trace(NULL, regs, NULL, &backtrace_ops, entry);
1728 #ifdef CONFIG_COMPAT
1730 perf_callchain_user32(struct pt_regs *regs, struct perf_callchain_entry *entry)
1732 /* 32-bit process in 64-bit kernel. */
1733 struct stack_frame_ia32 frame;
1734 const void __user *fp;
1736 if (!test_thread_flag(TIF_IA32))
1739 fp = compat_ptr(regs->bp);
1740 while (entry->nr < PERF_MAX_STACK_DEPTH) {
1741 unsigned long bytes;
1742 frame.next_frame = 0;
1743 frame.return_address = 0;
1745 bytes = copy_from_user_nmi(&frame, fp, sizeof(frame));
1746 if (bytes != sizeof(frame))
1749 if (fp < compat_ptr(regs->sp))
1752 perf_callchain_store(entry, frame.return_address);
1753 fp = compat_ptr(frame.next_frame);
1759 perf_callchain_user32(struct pt_regs *regs, struct perf_callchain_entry *entry)
1766 perf_callchain_user(struct perf_callchain_entry *entry, struct pt_regs *regs)
1768 struct stack_frame frame;
1769 const void __user *fp;
1771 if (perf_guest_cbs && perf_guest_cbs->is_in_guest()) {
1772 /* TODO: We don't support guest os callchain now */
1776 fp = (void __user *)regs->bp;
1778 perf_callchain_store(entry, regs->ip);
1780 if (perf_callchain_user32(regs, entry))
1783 while (entry->nr < PERF_MAX_STACK_DEPTH) {
1784 unsigned long bytes;
1785 frame.next_frame = NULL;
1786 frame.return_address = 0;
1788 bytes = copy_from_user_nmi(&frame, fp, sizeof(frame));
1789 if (bytes != sizeof(frame))
1792 if ((unsigned long)fp < regs->sp)
1795 perf_callchain_store(entry, frame.return_address);
1796 fp = frame.next_frame;
1800 unsigned long perf_instruction_pointer(struct pt_regs *regs)
1804 if (perf_guest_cbs && perf_guest_cbs->is_in_guest())
1805 ip = perf_guest_cbs->get_guest_ip();
1807 ip = instruction_pointer(regs);
1812 unsigned long perf_misc_flags(struct pt_regs *regs)
1816 if (perf_guest_cbs && perf_guest_cbs->is_in_guest()) {
1817 if (perf_guest_cbs->is_user_mode())
1818 misc |= PERF_RECORD_MISC_GUEST_USER;
1820 misc |= PERF_RECORD_MISC_GUEST_KERNEL;
1822 if (user_mode(regs))
1823 misc |= PERF_RECORD_MISC_USER;
1825 misc |= PERF_RECORD_MISC_KERNEL;
1828 if (regs->flags & PERF_EFLAGS_EXACT)
1829 misc |= PERF_RECORD_MISC_EXACT_IP;