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>
34 #include <asm/alternative.h>
38 #define wrmsrl(msr, val) \
40 trace_printk("wrmsrl(%lx, %lx)\n", (unsigned long)(msr),\
41 (unsigned long)(val)); \
42 native_write_msr((msr), (u32)((u64)(val)), \
43 (u32)((u64)(val) >> 32)); \
48 * best effort, GUP based copy_from_user() that assumes IRQ or NMI context
51 copy_from_user_nmi(void *to, const void __user *from, unsigned long n)
53 unsigned long offset, addr = (unsigned long)from;
54 unsigned long size, len = 0;
60 ret = __get_user_pages_fast(addr, 1, 0, &page);
64 offset = addr & (PAGE_SIZE - 1);
65 size = min(PAGE_SIZE - offset, n - len);
67 map = kmap_atomic(page);
68 memcpy(to, map+offset, size);
81 struct event_constraint {
83 unsigned long idxmsk[BITS_TO_LONGS(X86_PMC_IDX_MAX)];
92 int nb_id; /* NorthBridge id */
93 int refcnt; /* reference count */
94 struct perf_event *owners[X86_PMC_IDX_MAX];
95 struct event_constraint event_constraints[X86_PMC_IDX_MAX];
100 #define MAX_LBR_ENTRIES 16
102 struct cpu_hw_events {
104 * Generic x86 PMC bits
106 struct perf_event *events[X86_PMC_IDX_MAX]; /* in counter order */
107 unsigned long active_mask[BITS_TO_LONGS(X86_PMC_IDX_MAX)];
108 unsigned long running[BITS_TO_LONGS(X86_PMC_IDX_MAX)];
114 int assign[X86_PMC_IDX_MAX]; /* event to counter assignment */
115 u64 tags[X86_PMC_IDX_MAX];
116 struct perf_event *event_list[X86_PMC_IDX_MAX]; /* in enabled order */
118 unsigned int group_flag;
121 * Intel DebugStore bits
123 struct debug_store *ds;
131 struct perf_branch_stack lbr_stack;
132 struct perf_branch_entry lbr_entries[MAX_LBR_ENTRIES];
135 * Intel percore register state.
136 * Coordinate shared resources between HT threads.
138 int percore_used; /* Used by this CPU? */
139 struct intel_percore *per_core;
144 struct amd_nb *amd_nb;
147 #define __EVENT_CONSTRAINT(c, n, m, w) {\
148 { .idxmsk64 = (n) }, \
154 #define EVENT_CONSTRAINT(c, n, m) \
155 __EVENT_CONSTRAINT(c, n, m, HWEIGHT(n))
158 * Constraint on the Event code.
160 #define INTEL_EVENT_CONSTRAINT(c, n) \
161 EVENT_CONSTRAINT(c, n, ARCH_PERFMON_EVENTSEL_EVENT)
164 * Constraint on the Event code + UMask + fixed-mask
166 * filter mask to validate fixed counter events.
167 * the following filters disqualify for fixed counters:
171 * The other filters are supported by fixed counters.
172 * The any-thread option is supported starting with v3.
174 #define FIXED_EVENT_CONSTRAINT(c, n) \
175 EVENT_CONSTRAINT(c, (1ULL << (32+n)), X86_RAW_EVENT_MASK)
178 * Constraint on the Event code + UMask
180 #define INTEL_UEVENT_CONSTRAINT(c, n) \
181 EVENT_CONSTRAINT(c, n, INTEL_ARCH_EVENT_MASK)
183 #define EVENT_CONSTRAINT_END \
184 EVENT_CONSTRAINT(0, 0, 0)
186 #define for_each_event_constraint(e, c) \
187 for ((e) = (c); (e)->weight; (e)++)
190 * Extra registers for specific events.
191 * Some events need large masks and require external MSRs.
192 * Define a mapping to these extra registers.
201 #define EVENT_EXTRA_REG(e, ms, m, vm) { \
204 .config_mask = (m), \
205 .valid_mask = (vm), \
207 #define INTEL_EVENT_EXTRA_REG(event, msr, vm) \
208 EVENT_EXTRA_REG(event, msr, ARCH_PERFMON_EVENTSEL_EVENT, vm)
209 #define EVENT_EXTRA_END EVENT_EXTRA_REG(0, 0, 0, 0)
211 union perf_capabilities {
215 u64 pebs_arch_reg : 1;
223 * struct x86_pmu - generic x86 pmu
227 * Generic x86 PMC bits
231 int (*handle_irq)(struct pt_regs *);
232 void (*disable_all)(void);
233 void (*enable_all)(int added);
234 void (*enable)(struct perf_event *);
235 void (*disable)(struct perf_event *);
236 int (*hw_config)(struct perf_event *event);
237 int (*schedule_events)(struct cpu_hw_events *cpuc, int n, int *assign);
240 u64 (*event_map)(int);
243 int num_counters_fixed;
248 struct event_constraint *
249 (*get_event_constraints)(struct cpu_hw_events *cpuc,
250 struct perf_event *event);
252 void (*put_event_constraints)(struct cpu_hw_events *cpuc,
253 struct perf_event *event);
254 struct event_constraint *event_constraints;
255 struct event_constraint *percore_constraints;
256 void (*quirks)(void);
257 int perfctr_second_write;
259 int (*cpu_prepare)(int cpu);
260 void (*cpu_starting)(int cpu);
261 void (*cpu_dying)(int cpu);
262 void (*cpu_dead)(int cpu);
265 * Intel Arch Perfmon v2+
268 union perf_capabilities intel_cap;
271 * Intel DebugStore bits
274 int bts_active, pebs_active;
275 int pebs_record_size;
276 void (*drain_pebs)(struct pt_regs *regs);
277 struct event_constraint *pebs_constraints;
282 unsigned long lbr_tos, lbr_from, lbr_to; /* MSR base regs */
283 int lbr_nr; /* hardware stack size */
286 * Extra registers for events
288 struct extra_reg *extra_regs;
291 static struct x86_pmu x86_pmu __read_mostly;
293 static DEFINE_PER_CPU(struct cpu_hw_events, cpu_hw_events) = {
297 static int x86_perf_event_set_period(struct perf_event *event);
300 * Generalized hw caching related hw_event table, filled
301 * in on a per model basis. A value of 0 means
302 * 'not supported', -1 means 'hw_event makes no sense on
303 * this CPU', any other value means the raw hw_event
307 #define C(x) PERF_COUNT_HW_CACHE_##x
309 static u64 __read_mostly hw_cache_event_ids
310 [PERF_COUNT_HW_CACHE_MAX]
311 [PERF_COUNT_HW_CACHE_OP_MAX]
312 [PERF_COUNT_HW_CACHE_RESULT_MAX];
313 static u64 __read_mostly hw_cache_extra_regs
314 [PERF_COUNT_HW_CACHE_MAX]
315 [PERF_COUNT_HW_CACHE_OP_MAX]
316 [PERF_COUNT_HW_CACHE_RESULT_MAX];
319 * Propagate event elapsed time into the generic event.
320 * Can only be executed on the CPU where the event is active.
321 * Returns the delta events processed.
324 x86_perf_event_update(struct perf_event *event)
326 struct hw_perf_event *hwc = &event->hw;
327 int shift = 64 - x86_pmu.cntval_bits;
328 u64 prev_raw_count, new_raw_count;
332 if (idx == X86_PMC_IDX_FIXED_BTS)
336 * Careful: an NMI might modify the previous event value.
338 * Our tactic to handle this is to first atomically read and
339 * exchange a new raw count - then add that new-prev delta
340 * count to the generic event atomically:
343 prev_raw_count = local64_read(&hwc->prev_count);
344 rdmsrl(hwc->event_base, new_raw_count);
346 if (local64_cmpxchg(&hwc->prev_count, prev_raw_count,
347 new_raw_count) != prev_raw_count)
351 * Now we have the new raw value and have updated the prev
352 * timestamp already. We can now calculate the elapsed delta
353 * (event-)time and add that to the generic event.
355 * Careful, not all hw sign-extends above the physical width
358 delta = (new_raw_count << shift) - (prev_raw_count << shift);
361 local64_add(delta, &event->count);
362 local64_sub(delta, &hwc->period_left);
364 return new_raw_count;
367 static inline int x86_pmu_addr_offset(int index)
371 /* offset = X86_FEATURE_PERFCTR_CORE ? index << 1 : index */
372 alternative_io(ASM_NOP2,
374 X86_FEATURE_PERFCTR_CORE,
381 static inline unsigned int x86_pmu_config_addr(int index)
383 return x86_pmu.eventsel + x86_pmu_addr_offset(index);
386 static inline unsigned int x86_pmu_event_addr(int index)
388 return x86_pmu.perfctr + x86_pmu_addr_offset(index);
392 * Find and validate any extra registers to set up.
394 static int x86_pmu_extra_regs(u64 config, struct perf_event *event)
396 struct extra_reg *er;
398 event->hw.extra_reg = 0;
399 event->hw.extra_config = 0;
401 if (!x86_pmu.extra_regs)
404 for (er = x86_pmu.extra_regs; er->msr; er++) {
405 if (er->event != (config & er->config_mask))
407 if (event->attr.config1 & ~er->valid_mask)
409 event->hw.extra_reg = er->msr;
410 event->hw.extra_config = event->attr.config1;
416 static atomic_t active_events;
417 static DEFINE_MUTEX(pmc_reserve_mutex);
419 #ifdef CONFIG_X86_LOCAL_APIC
421 static bool reserve_pmc_hardware(void)
425 for (i = 0; i < x86_pmu.num_counters; i++) {
426 if (!reserve_perfctr_nmi(x86_pmu_event_addr(i)))
430 for (i = 0; i < x86_pmu.num_counters; i++) {
431 if (!reserve_evntsel_nmi(x86_pmu_config_addr(i)))
438 for (i--; i >= 0; i--)
439 release_evntsel_nmi(x86_pmu_config_addr(i));
441 i = x86_pmu.num_counters;
444 for (i--; i >= 0; i--)
445 release_perfctr_nmi(x86_pmu_event_addr(i));
450 static void release_pmc_hardware(void)
454 for (i = 0; i < x86_pmu.num_counters; i++) {
455 release_perfctr_nmi(x86_pmu_event_addr(i));
456 release_evntsel_nmi(x86_pmu_config_addr(i));
462 static bool reserve_pmc_hardware(void) { return true; }
463 static void release_pmc_hardware(void) {}
467 static bool check_hw_exists(void)
469 u64 val, val_new = 0;
473 * Check to see if the BIOS enabled any of the counters, if so
476 for (i = 0; i < x86_pmu.num_counters; i++) {
477 reg = x86_pmu_config_addr(i);
478 ret = rdmsrl_safe(reg, &val);
481 if (val & ARCH_PERFMON_EVENTSEL_ENABLE)
485 if (x86_pmu.num_counters_fixed) {
486 reg = MSR_ARCH_PERFMON_FIXED_CTR_CTRL;
487 ret = rdmsrl_safe(reg, &val);
490 for (i = 0; i < x86_pmu.num_counters_fixed; i++) {
491 if (val & (0x03 << i*4))
497 * Now write a value and read it back to see if it matches,
498 * this is needed to detect certain hardware emulators (qemu/kvm)
499 * that don't trap on the MSR access and always return 0s.
502 ret = checking_wrmsrl(x86_pmu_event_addr(0), val);
503 ret |= rdmsrl_safe(x86_pmu_event_addr(0), &val_new);
504 if (ret || val != val_new)
511 * We still allow the PMU driver to operate:
513 printk(KERN_CONT "Broken BIOS detected, complain to your hardware vendor.\n");
514 printk(KERN_ERR FW_BUG "the BIOS has corrupted hw-PMU resources (MSR %x is %Lx)\n", reg, val);
519 printk(KERN_CONT "Broken PMU hardware detected, using software events only.\n");
524 static void reserve_ds_buffers(void);
525 static void release_ds_buffers(void);
527 static void hw_perf_event_destroy(struct perf_event *event)
529 if (atomic_dec_and_mutex_lock(&active_events, &pmc_reserve_mutex)) {
530 release_pmc_hardware();
531 release_ds_buffers();
532 mutex_unlock(&pmc_reserve_mutex);
536 static inline int x86_pmu_initialized(void)
538 return x86_pmu.handle_irq != NULL;
542 set_ext_hw_attr(struct hw_perf_event *hwc, struct perf_event *event)
544 struct perf_event_attr *attr = &event->attr;
545 unsigned int cache_type, cache_op, cache_result;
548 config = attr->config;
550 cache_type = (config >> 0) & 0xff;
551 if (cache_type >= PERF_COUNT_HW_CACHE_MAX)
554 cache_op = (config >> 8) & 0xff;
555 if (cache_op >= PERF_COUNT_HW_CACHE_OP_MAX)
558 cache_result = (config >> 16) & 0xff;
559 if (cache_result >= PERF_COUNT_HW_CACHE_RESULT_MAX)
562 val = hw_cache_event_ids[cache_type][cache_op][cache_result];
571 attr->config1 = hw_cache_extra_regs[cache_type][cache_op][cache_result];
572 return x86_pmu_extra_regs(val, event);
575 static int x86_setup_perfctr(struct perf_event *event)
577 struct perf_event_attr *attr = &event->attr;
578 struct hw_perf_event *hwc = &event->hw;
581 if (!is_sampling_event(event)) {
582 hwc->sample_period = x86_pmu.max_period;
583 hwc->last_period = hwc->sample_period;
584 local64_set(&hwc->period_left, hwc->sample_period);
587 * If we have a PMU initialized but no APIC
588 * interrupts, we cannot sample hardware
589 * events (user-space has to fall back and
590 * sample via a hrtimer based software event):
596 if (attr->type == PERF_TYPE_RAW)
597 return x86_pmu_extra_regs(event->attr.config, event);
599 if (attr->type == PERF_TYPE_HW_CACHE)
600 return set_ext_hw_attr(hwc, event);
602 if (attr->config >= x86_pmu.max_events)
608 config = x86_pmu.event_map(attr->config);
619 if ((attr->config == PERF_COUNT_HW_BRANCH_INSTRUCTIONS) &&
620 (hwc->sample_period == 1)) {
621 /* BTS is not supported by this architecture. */
622 if (!x86_pmu.bts_active)
625 /* BTS is currently only allowed for user-mode. */
626 if (!attr->exclude_kernel)
630 hwc->config |= config;
635 static int x86_pmu_hw_config(struct perf_event *event)
637 if (event->attr.precise_ip) {
640 /* Support for constant skid */
641 if (x86_pmu.pebs_active) {
644 /* Support for IP fixup */
649 if (event->attr.precise_ip > precise)
655 * (keep 'enabled' bit clear for now)
657 event->hw.config = ARCH_PERFMON_EVENTSEL_INT;
660 * Count user and OS events unless requested not to
662 if (!event->attr.exclude_user)
663 event->hw.config |= ARCH_PERFMON_EVENTSEL_USR;
664 if (!event->attr.exclude_kernel)
665 event->hw.config |= ARCH_PERFMON_EVENTSEL_OS;
667 if (event->attr.type == PERF_TYPE_RAW)
668 event->hw.config |= event->attr.config & X86_RAW_EVENT_MASK;
670 return x86_setup_perfctr(event);
674 * Setup the hardware configuration for a given attr_type
676 static int __x86_pmu_event_init(struct perf_event *event)
680 if (!x86_pmu_initialized())
684 if (!atomic_inc_not_zero(&active_events)) {
685 mutex_lock(&pmc_reserve_mutex);
686 if (atomic_read(&active_events) == 0) {
687 if (!reserve_pmc_hardware())
690 reserve_ds_buffers();
693 atomic_inc(&active_events);
694 mutex_unlock(&pmc_reserve_mutex);
699 event->destroy = hw_perf_event_destroy;
702 event->hw.last_cpu = -1;
703 event->hw.last_tag = ~0ULL;
705 return x86_pmu.hw_config(event);
708 static void x86_pmu_disable_all(void)
710 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
713 for (idx = 0; idx < x86_pmu.num_counters; idx++) {
716 if (!test_bit(idx, cpuc->active_mask))
718 rdmsrl(x86_pmu_config_addr(idx), val);
719 if (!(val & ARCH_PERFMON_EVENTSEL_ENABLE))
721 val &= ~ARCH_PERFMON_EVENTSEL_ENABLE;
722 wrmsrl(x86_pmu_config_addr(idx), val);
726 static void x86_pmu_disable(struct pmu *pmu)
728 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
730 if (!x86_pmu_initialized())
740 x86_pmu.disable_all();
743 static inline void __x86_pmu_enable_event(struct hw_perf_event *hwc,
747 wrmsrl(hwc->extra_reg, hwc->extra_config);
748 wrmsrl(hwc->config_base, hwc->config | enable_mask);
751 static void x86_pmu_enable_all(int added)
753 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
756 for (idx = 0; idx < x86_pmu.num_counters; idx++) {
757 struct hw_perf_event *hwc = &cpuc->events[idx]->hw;
759 if (!test_bit(idx, cpuc->active_mask))
762 __x86_pmu_enable_event(hwc, ARCH_PERFMON_EVENTSEL_ENABLE);
766 static struct pmu pmu;
768 static inline int is_x86_event(struct perf_event *event)
770 return event->pmu == &pmu;
773 static int x86_schedule_events(struct cpu_hw_events *cpuc, int n, int *assign)
775 struct event_constraint *c, *constraints[X86_PMC_IDX_MAX];
776 unsigned long used_mask[BITS_TO_LONGS(X86_PMC_IDX_MAX)];
777 int i, j, w, wmax, num = 0;
778 struct hw_perf_event *hwc;
780 bitmap_zero(used_mask, X86_PMC_IDX_MAX);
782 for (i = 0; i < n; i++) {
783 c = x86_pmu.get_event_constraints(cpuc, cpuc->event_list[i]);
788 * fastpath, try to reuse previous register
790 for (i = 0; i < n; i++) {
791 hwc = &cpuc->event_list[i]->hw;
798 /* constraint still honored */
799 if (!test_bit(hwc->idx, c->idxmsk))
802 /* not already used */
803 if (test_bit(hwc->idx, used_mask))
806 __set_bit(hwc->idx, used_mask);
808 assign[i] = hwc->idx;
817 bitmap_zero(used_mask, X86_PMC_IDX_MAX);
820 * weight = number of possible counters
822 * 1 = most constrained, only works on one counter
823 * wmax = least constrained, works on any counter
825 * assign events to counters starting with most
826 * constrained events.
828 wmax = x86_pmu.num_counters;
831 * when fixed event counters are present,
832 * wmax is incremented by 1 to account
833 * for one more choice
835 if (x86_pmu.num_counters_fixed)
838 for (w = 1, num = n; num && w <= wmax; w++) {
840 for (i = 0; num && i < n; i++) {
842 hwc = &cpuc->event_list[i]->hw;
847 for_each_set_bit(j, c->idxmsk, X86_PMC_IDX_MAX) {
848 if (!test_bit(j, used_mask))
852 if (j == X86_PMC_IDX_MAX)
855 __set_bit(j, used_mask);
864 * scheduling failed or is just a simulation,
865 * free resources if necessary
867 if (!assign || num) {
868 for (i = 0; i < n; i++) {
869 if (x86_pmu.put_event_constraints)
870 x86_pmu.put_event_constraints(cpuc, cpuc->event_list[i]);
873 return num ? -ENOSPC : 0;
877 * dogrp: true if must collect siblings events (group)
878 * returns total number of events and error code
880 static int collect_events(struct cpu_hw_events *cpuc, struct perf_event *leader, bool dogrp)
882 struct perf_event *event;
885 max_count = x86_pmu.num_counters + x86_pmu.num_counters_fixed;
887 /* current number of events already accepted */
890 if (is_x86_event(leader)) {
893 cpuc->event_list[n] = leader;
899 list_for_each_entry(event, &leader->sibling_list, group_entry) {
900 if (!is_x86_event(event) ||
901 event->state <= PERF_EVENT_STATE_OFF)
907 cpuc->event_list[n] = event;
913 static inline void x86_assign_hw_event(struct perf_event *event,
914 struct cpu_hw_events *cpuc, int i)
916 struct hw_perf_event *hwc = &event->hw;
918 hwc->idx = cpuc->assign[i];
919 hwc->last_cpu = smp_processor_id();
920 hwc->last_tag = ++cpuc->tags[i];
922 if (hwc->idx == X86_PMC_IDX_FIXED_BTS) {
923 hwc->config_base = 0;
925 } else if (hwc->idx >= X86_PMC_IDX_FIXED) {
926 hwc->config_base = MSR_ARCH_PERFMON_FIXED_CTR_CTRL;
927 hwc->event_base = MSR_ARCH_PERFMON_FIXED_CTR0 + (hwc->idx - X86_PMC_IDX_FIXED);
929 hwc->config_base = x86_pmu_config_addr(hwc->idx);
930 hwc->event_base = x86_pmu_event_addr(hwc->idx);
934 static inline int match_prev_assignment(struct hw_perf_event *hwc,
935 struct cpu_hw_events *cpuc,
938 return hwc->idx == cpuc->assign[i] &&
939 hwc->last_cpu == smp_processor_id() &&
940 hwc->last_tag == cpuc->tags[i];
943 static void x86_pmu_start(struct perf_event *event, int flags);
944 static void x86_pmu_stop(struct perf_event *event, int flags);
946 static void x86_pmu_enable(struct pmu *pmu)
948 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
949 struct perf_event *event;
950 struct hw_perf_event *hwc;
951 int i, added = cpuc->n_added;
953 if (!x86_pmu_initialized())
960 int n_running = cpuc->n_events - cpuc->n_added;
962 * apply assignment obtained either from
963 * hw_perf_group_sched_in() or x86_pmu_enable()
965 * step1: save events moving to new counters
966 * step2: reprogram moved events into new counters
968 for (i = 0; i < n_running; i++) {
969 event = cpuc->event_list[i];
973 * we can avoid reprogramming counter if:
974 * - assigned same counter as last time
975 * - running on same CPU as last time
976 * - no other event has used the counter since
978 if (hwc->idx == -1 ||
979 match_prev_assignment(hwc, cpuc, i))
983 * Ensure we don't accidentally enable a stopped
984 * counter simply because we rescheduled.
986 if (hwc->state & PERF_HES_STOPPED)
987 hwc->state |= PERF_HES_ARCH;
989 x86_pmu_stop(event, PERF_EF_UPDATE);
992 for (i = 0; i < cpuc->n_events; i++) {
993 event = cpuc->event_list[i];
996 if (!match_prev_assignment(hwc, cpuc, i))
997 x86_assign_hw_event(event, cpuc, i);
998 else if (i < n_running)
1001 if (hwc->state & PERF_HES_ARCH)
1004 x86_pmu_start(event, PERF_EF_RELOAD);
1007 perf_events_lapic_init();
1013 x86_pmu.enable_all(added);
1016 static inline void x86_pmu_disable_event(struct perf_event *event)
1018 struct hw_perf_event *hwc = &event->hw;
1020 wrmsrl(hwc->config_base, hwc->config);
1023 static DEFINE_PER_CPU(u64 [X86_PMC_IDX_MAX], pmc_prev_left);
1026 * Set the next IRQ period, based on the hwc->period_left value.
1027 * To be called with the event disabled in hw:
1030 x86_perf_event_set_period(struct perf_event *event)
1032 struct hw_perf_event *hwc = &event->hw;
1033 s64 left = local64_read(&hwc->period_left);
1034 s64 period = hwc->sample_period;
1035 int ret = 0, idx = hwc->idx;
1037 if (idx == X86_PMC_IDX_FIXED_BTS)
1041 * If we are way outside a reasonable range then just skip forward:
1043 if (unlikely(left <= -period)) {
1045 local64_set(&hwc->period_left, left);
1046 hwc->last_period = period;
1050 if (unlikely(left <= 0)) {
1052 local64_set(&hwc->period_left, left);
1053 hwc->last_period = period;
1057 * Quirk: certain CPUs dont like it if just 1 hw_event is left:
1059 if (unlikely(left < 2))
1062 if (left > x86_pmu.max_period)
1063 left = x86_pmu.max_period;
1065 per_cpu(pmc_prev_left[idx], smp_processor_id()) = left;
1068 * The hw event starts counting from this event offset,
1069 * mark it to be able to extra future deltas:
1071 local64_set(&hwc->prev_count, (u64)-left);
1073 wrmsrl(hwc->event_base, (u64)(-left) & x86_pmu.cntval_mask);
1076 * Due to erratum on certan cpu we need
1077 * a second write to be sure the register
1078 * is updated properly
1080 if (x86_pmu.perfctr_second_write) {
1081 wrmsrl(hwc->event_base,
1082 (u64)(-left) & x86_pmu.cntval_mask);
1085 perf_event_update_userpage(event);
1090 static void x86_pmu_enable_event(struct perf_event *event)
1092 if (__this_cpu_read(cpu_hw_events.enabled))
1093 __x86_pmu_enable_event(&event->hw,
1094 ARCH_PERFMON_EVENTSEL_ENABLE);
1098 * Add a single event to the PMU.
1100 * The event is added to the group of enabled events
1101 * but only if it can be scehduled with existing events.
1103 static int x86_pmu_add(struct perf_event *event, int flags)
1105 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
1106 struct hw_perf_event *hwc;
1107 int assign[X86_PMC_IDX_MAX];
1112 perf_pmu_disable(event->pmu);
1113 n0 = cpuc->n_events;
1114 ret = n = collect_events(cpuc, event, false);
1118 hwc->state = PERF_HES_UPTODATE | PERF_HES_STOPPED;
1119 if (!(flags & PERF_EF_START))
1120 hwc->state |= PERF_HES_ARCH;
1123 * If group events scheduling transaction was started,
1124 * skip the schedulability test here, it will be performed
1125 * at commit time (->commit_txn) as a whole
1127 if (cpuc->group_flag & PERF_EVENT_TXN)
1130 ret = x86_pmu.schedule_events(cpuc, n, assign);
1134 * copy new assignment, now we know it is possible
1135 * will be used by hw_perf_enable()
1137 memcpy(cpuc->assign, assign, n*sizeof(int));
1141 cpuc->n_added += n - n0;
1142 cpuc->n_txn += n - n0;
1146 perf_pmu_enable(event->pmu);
1150 static void x86_pmu_start(struct perf_event *event, int flags)
1152 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
1153 int idx = event->hw.idx;
1155 if (WARN_ON_ONCE(!(event->hw.state & PERF_HES_STOPPED)))
1158 if (WARN_ON_ONCE(idx == -1))
1161 if (flags & PERF_EF_RELOAD) {
1162 WARN_ON_ONCE(!(event->hw.state & PERF_HES_UPTODATE));
1163 x86_perf_event_set_period(event);
1166 event->hw.state = 0;
1168 cpuc->events[idx] = event;
1169 __set_bit(idx, cpuc->active_mask);
1170 __set_bit(idx, cpuc->running);
1171 x86_pmu.enable(event);
1172 perf_event_update_userpage(event);
1175 void perf_event_print_debug(void)
1177 u64 ctrl, status, overflow, pmc_ctrl, pmc_count, prev_left, fixed;
1179 struct cpu_hw_events *cpuc;
1180 unsigned long flags;
1183 if (!x86_pmu.num_counters)
1186 local_irq_save(flags);
1188 cpu = smp_processor_id();
1189 cpuc = &per_cpu(cpu_hw_events, cpu);
1191 if (x86_pmu.version >= 2) {
1192 rdmsrl(MSR_CORE_PERF_GLOBAL_CTRL, ctrl);
1193 rdmsrl(MSR_CORE_PERF_GLOBAL_STATUS, status);
1194 rdmsrl(MSR_CORE_PERF_GLOBAL_OVF_CTRL, overflow);
1195 rdmsrl(MSR_ARCH_PERFMON_FIXED_CTR_CTRL, fixed);
1196 rdmsrl(MSR_IA32_PEBS_ENABLE, pebs);
1199 pr_info("CPU#%d: ctrl: %016llx\n", cpu, ctrl);
1200 pr_info("CPU#%d: status: %016llx\n", cpu, status);
1201 pr_info("CPU#%d: overflow: %016llx\n", cpu, overflow);
1202 pr_info("CPU#%d: fixed: %016llx\n", cpu, fixed);
1203 pr_info("CPU#%d: pebs: %016llx\n", cpu, pebs);
1205 pr_info("CPU#%d: active: %016llx\n", cpu, *(u64 *)cpuc->active_mask);
1207 for (idx = 0; idx < x86_pmu.num_counters; idx++) {
1208 rdmsrl(x86_pmu_config_addr(idx), pmc_ctrl);
1209 rdmsrl(x86_pmu_event_addr(idx), pmc_count);
1211 prev_left = per_cpu(pmc_prev_left[idx], cpu);
1213 pr_info("CPU#%d: gen-PMC%d ctrl: %016llx\n",
1214 cpu, idx, pmc_ctrl);
1215 pr_info("CPU#%d: gen-PMC%d count: %016llx\n",
1216 cpu, idx, pmc_count);
1217 pr_info("CPU#%d: gen-PMC%d left: %016llx\n",
1218 cpu, idx, prev_left);
1220 for (idx = 0; idx < x86_pmu.num_counters_fixed; idx++) {
1221 rdmsrl(MSR_ARCH_PERFMON_FIXED_CTR0 + idx, pmc_count);
1223 pr_info("CPU#%d: fixed-PMC%d count: %016llx\n",
1224 cpu, idx, pmc_count);
1226 local_irq_restore(flags);
1229 static void x86_pmu_stop(struct perf_event *event, int flags)
1231 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
1232 struct hw_perf_event *hwc = &event->hw;
1234 if (__test_and_clear_bit(hwc->idx, cpuc->active_mask)) {
1235 x86_pmu.disable(event);
1236 cpuc->events[hwc->idx] = NULL;
1237 WARN_ON_ONCE(hwc->state & PERF_HES_STOPPED);
1238 hwc->state |= PERF_HES_STOPPED;
1241 if ((flags & PERF_EF_UPDATE) && !(hwc->state & PERF_HES_UPTODATE)) {
1243 * Drain the remaining delta count out of a event
1244 * that we are disabling:
1246 x86_perf_event_update(event);
1247 hwc->state |= PERF_HES_UPTODATE;
1251 static void x86_pmu_del(struct perf_event *event, int flags)
1253 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
1257 * If we're called during a txn, we don't need to do anything.
1258 * The events never got scheduled and ->cancel_txn will truncate
1261 if (cpuc->group_flag & PERF_EVENT_TXN)
1264 x86_pmu_stop(event, PERF_EF_UPDATE);
1266 for (i = 0; i < cpuc->n_events; i++) {
1267 if (event == cpuc->event_list[i]) {
1269 if (x86_pmu.put_event_constraints)
1270 x86_pmu.put_event_constraints(cpuc, event);
1272 while (++i < cpuc->n_events)
1273 cpuc->event_list[i-1] = cpuc->event_list[i];
1279 perf_event_update_userpage(event);
1282 static int x86_pmu_handle_irq(struct pt_regs *regs)
1284 struct perf_sample_data data;
1285 struct cpu_hw_events *cpuc;
1286 struct perf_event *event;
1287 int idx, handled = 0;
1290 perf_sample_data_init(&data, 0);
1292 cpuc = &__get_cpu_var(cpu_hw_events);
1294 for (idx = 0; idx < x86_pmu.num_counters; idx++) {
1295 if (!test_bit(idx, cpuc->active_mask)) {
1297 * Though we deactivated the counter some cpus
1298 * might still deliver spurious interrupts still
1299 * in flight. Catch them:
1301 if (__test_and_clear_bit(idx, cpuc->running))
1306 event = cpuc->events[idx];
1308 val = x86_perf_event_update(event);
1309 if (val & (1ULL << (x86_pmu.cntval_bits - 1)))
1316 data.period = event->hw.last_period;
1318 if (!x86_perf_event_set_period(event))
1321 if (perf_event_overflow(event, 1, &data, regs))
1322 x86_pmu_stop(event, 0);
1326 inc_irq_stat(apic_perf_irqs);
1331 void perf_events_lapic_init(void)
1333 if (!x86_pmu.apic || !x86_pmu_initialized())
1337 * Always use NMI for PMU
1339 apic_write(APIC_LVTPC, APIC_DM_NMI);
1342 struct pmu_nmi_state {
1343 unsigned int marked;
1347 static DEFINE_PER_CPU(struct pmu_nmi_state, pmu_nmi);
1349 static int __kprobes
1350 perf_event_nmi_handler(struct notifier_block *self,
1351 unsigned long cmd, void *__args)
1353 struct die_args *args = __args;
1354 unsigned int this_nmi;
1357 if (!atomic_read(&active_events))
1363 case DIE_NMIUNKNOWN:
1364 this_nmi = percpu_read(irq_stat.__nmi_count);
1365 if (this_nmi != __this_cpu_read(pmu_nmi.marked))
1366 /* let the kernel handle the unknown nmi */
1369 * This one is a PMU back-to-back nmi. Two events
1370 * trigger 'simultaneously' raising two back-to-back
1371 * NMIs. If the first NMI handles both, the latter
1372 * will be empty and daze the CPU. So, we drop it to
1373 * avoid false-positive 'unknown nmi' messages.
1380 apic_write(APIC_LVTPC, APIC_DM_NMI);
1382 handled = x86_pmu.handle_irq(args->regs);
1386 this_nmi = percpu_read(irq_stat.__nmi_count);
1387 if ((handled > 1) ||
1388 /* the next nmi could be a back-to-back nmi */
1389 ((__this_cpu_read(pmu_nmi.marked) == this_nmi) &&
1390 (__this_cpu_read(pmu_nmi.handled) > 1))) {
1392 * We could have two subsequent back-to-back nmis: The
1393 * first handles more than one counter, the 2nd
1394 * handles only one counter and the 3rd handles no
1397 * This is the 2nd nmi because the previous was
1398 * handling more than one counter. We will mark the
1399 * next (3rd) and then drop it if unhandled.
1401 __this_cpu_write(pmu_nmi.marked, this_nmi + 1);
1402 __this_cpu_write(pmu_nmi.handled, handled);
1408 static __read_mostly struct notifier_block perf_event_nmi_notifier = {
1409 .notifier_call = perf_event_nmi_handler,
1411 .priority = NMI_LOCAL_LOW_PRIOR,
1414 static struct event_constraint unconstrained;
1415 static struct event_constraint emptyconstraint;
1417 static struct event_constraint *
1418 x86_get_event_constraints(struct cpu_hw_events *cpuc, struct perf_event *event)
1420 struct event_constraint *c;
1422 if (x86_pmu.event_constraints) {
1423 for_each_event_constraint(c, x86_pmu.event_constraints) {
1424 if ((event->hw.config & c->cmask) == c->code)
1429 return &unconstrained;
1432 #include "perf_event_amd.c"
1433 #include "perf_event_p6.c"
1434 #include "perf_event_p4.c"
1435 #include "perf_event_intel_lbr.c"
1436 #include "perf_event_intel_ds.c"
1437 #include "perf_event_intel.c"
1439 static int __cpuinit
1440 x86_pmu_notifier(struct notifier_block *self, unsigned long action, void *hcpu)
1442 unsigned int cpu = (long)hcpu;
1443 int ret = NOTIFY_OK;
1445 switch (action & ~CPU_TASKS_FROZEN) {
1446 case CPU_UP_PREPARE:
1447 if (x86_pmu.cpu_prepare)
1448 ret = x86_pmu.cpu_prepare(cpu);
1452 if (x86_pmu.cpu_starting)
1453 x86_pmu.cpu_starting(cpu);
1457 if (x86_pmu.cpu_dying)
1458 x86_pmu.cpu_dying(cpu);
1461 case CPU_UP_CANCELED:
1463 if (x86_pmu.cpu_dead)
1464 x86_pmu.cpu_dead(cpu);
1474 static void __init pmu_check_apic(void)
1480 pr_info("no APIC, boot with the \"lapic\" boot parameter to force-enable it.\n");
1481 pr_info("no hardware sampling interrupt available.\n");
1484 static int __init init_hw_perf_events(void)
1486 struct event_constraint *c;
1489 pr_info("Performance Events: ");
1491 switch (boot_cpu_data.x86_vendor) {
1492 case X86_VENDOR_INTEL:
1493 err = intel_pmu_init();
1495 case X86_VENDOR_AMD:
1496 err = amd_pmu_init();
1502 pr_cont("no PMU driver, software events only.\n");
1508 /* sanity check that the hardware exists or is emulated */
1509 if (!check_hw_exists())
1512 pr_cont("%s PMU driver.\n", x86_pmu.name);
1517 if (x86_pmu.num_counters > X86_PMC_MAX_GENERIC) {
1518 WARN(1, KERN_ERR "hw perf events %d > max(%d), clipping!",
1519 x86_pmu.num_counters, X86_PMC_MAX_GENERIC);
1520 x86_pmu.num_counters = X86_PMC_MAX_GENERIC;
1522 x86_pmu.intel_ctrl = (1 << x86_pmu.num_counters) - 1;
1524 if (x86_pmu.num_counters_fixed > X86_PMC_MAX_FIXED) {
1525 WARN(1, KERN_ERR "hw perf events fixed %d > max(%d), clipping!",
1526 x86_pmu.num_counters_fixed, X86_PMC_MAX_FIXED);
1527 x86_pmu.num_counters_fixed = X86_PMC_MAX_FIXED;
1530 x86_pmu.intel_ctrl |=
1531 ((1LL << x86_pmu.num_counters_fixed)-1) << X86_PMC_IDX_FIXED;
1533 perf_events_lapic_init();
1534 register_die_notifier(&perf_event_nmi_notifier);
1536 unconstrained = (struct event_constraint)
1537 __EVENT_CONSTRAINT(0, (1ULL << x86_pmu.num_counters) - 1,
1538 0, x86_pmu.num_counters);
1540 if (x86_pmu.event_constraints) {
1541 for_each_event_constraint(c, x86_pmu.event_constraints) {
1542 if (c->cmask != X86_RAW_EVENT_MASK)
1545 c->idxmsk64 |= (1ULL << x86_pmu.num_counters) - 1;
1546 c->weight += x86_pmu.num_counters;
1550 pr_info("... version: %d\n", x86_pmu.version);
1551 pr_info("... bit width: %d\n", x86_pmu.cntval_bits);
1552 pr_info("... generic registers: %d\n", x86_pmu.num_counters);
1553 pr_info("... value mask: %016Lx\n", x86_pmu.cntval_mask);
1554 pr_info("... max period: %016Lx\n", x86_pmu.max_period);
1555 pr_info("... fixed-purpose events: %d\n", x86_pmu.num_counters_fixed);
1556 pr_info("... event mask: %016Lx\n", x86_pmu.intel_ctrl);
1558 perf_pmu_register(&pmu, "cpu", PERF_TYPE_RAW);
1559 perf_cpu_notifier(x86_pmu_notifier);
1563 early_initcall(init_hw_perf_events);
1565 static inline void x86_pmu_read(struct perf_event *event)
1567 x86_perf_event_update(event);
1571 * Start group events scheduling transaction
1572 * Set the flag to make pmu::enable() not perform the
1573 * schedulability test, it will be performed at commit time
1575 static void x86_pmu_start_txn(struct pmu *pmu)
1577 perf_pmu_disable(pmu);
1578 __this_cpu_or(cpu_hw_events.group_flag, PERF_EVENT_TXN);
1579 __this_cpu_write(cpu_hw_events.n_txn, 0);
1583 * Stop group events scheduling transaction
1584 * Clear the flag and pmu::enable() will perform the
1585 * schedulability test.
1587 static void x86_pmu_cancel_txn(struct pmu *pmu)
1589 __this_cpu_and(cpu_hw_events.group_flag, ~PERF_EVENT_TXN);
1591 * Truncate the collected events.
1593 __this_cpu_sub(cpu_hw_events.n_added, __this_cpu_read(cpu_hw_events.n_txn));
1594 __this_cpu_sub(cpu_hw_events.n_events, __this_cpu_read(cpu_hw_events.n_txn));
1595 perf_pmu_enable(pmu);
1599 * Commit group events scheduling transaction
1600 * Perform the group schedulability test as a whole
1601 * Return 0 if success
1603 static int x86_pmu_commit_txn(struct pmu *pmu)
1605 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
1606 int assign[X86_PMC_IDX_MAX];
1611 if (!x86_pmu_initialized())
1614 ret = x86_pmu.schedule_events(cpuc, n, assign);
1619 * copy new assignment, now we know it is possible
1620 * will be used by hw_perf_enable()
1622 memcpy(cpuc->assign, assign, n*sizeof(int));
1624 cpuc->group_flag &= ~PERF_EVENT_TXN;
1625 perf_pmu_enable(pmu);
1630 * validate that we can schedule this event
1632 static int validate_event(struct perf_event *event)
1634 struct cpu_hw_events *fake_cpuc;
1635 struct event_constraint *c;
1638 fake_cpuc = kmalloc(sizeof(*fake_cpuc), GFP_KERNEL | __GFP_ZERO);
1642 c = x86_pmu.get_event_constraints(fake_cpuc, event);
1644 if (!c || !c->weight)
1647 if (x86_pmu.put_event_constraints)
1648 x86_pmu.put_event_constraints(fake_cpuc, event);
1656 * validate a single event group
1658 * validation include:
1659 * - check events are compatible which each other
1660 * - events do not compete for the same counter
1661 * - number of events <= number of counters
1663 * validation ensures the group can be loaded onto the
1664 * PMU if it was the only group available.
1666 static int validate_group(struct perf_event *event)
1668 struct perf_event *leader = event->group_leader;
1669 struct cpu_hw_events *fake_cpuc;
1673 fake_cpuc = kmalloc(sizeof(*fake_cpuc), GFP_KERNEL | __GFP_ZERO);
1678 * the event is not yet connected with its
1679 * siblings therefore we must first collect
1680 * existing siblings, then add the new event
1681 * before we can simulate the scheduling
1684 n = collect_events(fake_cpuc, leader, true);
1688 fake_cpuc->n_events = n;
1689 n = collect_events(fake_cpuc, event, false);
1693 fake_cpuc->n_events = n;
1695 ret = x86_pmu.schedule_events(fake_cpuc, n, NULL);
1703 static int x86_pmu_event_init(struct perf_event *event)
1708 switch (event->attr.type) {
1710 case PERF_TYPE_HARDWARE:
1711 case PERF_TYPE_HW_CACHE:
1718 err = __x86_pmu_event_init(event);
1721 * we temporarily connect event to its pmu
1722 * such that validate_group() can classify
1723 * it as an x86 event using is_x86_event()
1728 if (event->group_leader != event)
1729 err = validate_group(event);
1731 err = validate_event(event);
1737 event->destroy(event);
1743 static struct pmu pmu = {
1744 .pmu_enable = x86_pmu_enable,
1745 .pmu_disable = x86_pmu_disable,
1747 .event_init = x86_pmu_event_init,
1751 .start = x86_pmu_start,
1752 .stop = x86_pmu_stop,
1753 .read = x86_pmu_read,
1755 .start_txn = x86_pmu_start_txn,
1756 .cancel_txn = x86_pmu_cancel_txn,
1757 .commit_txn = x86_pmu_commit_txn,
1765 backtrace_warning_symbol(void *data, char *msg, unsigned long symbol)
1767 /* Ignore warnings */
1770 static void backtrace_warning(void *data, char *msg)
1772 /* Ignore warnings */
1775 static int backtrace_stack(void *data, char *name)
1780 static void backtrace_address(void *data, unsigned long addr, int reliable)
1782 struct perf_callchain_entry *entry = data;
1784 perf_callchain_store(entry, addr);
1787 static const struct stacktrace_ops backtrace_ops = {
1788 .warning = backtrace_warning,
1789 .warning_symbol = backtrace_warning_symbol,
1790 .stack = backtrace_stack,
1791 .address = backtrace_address,
1792 .walk_stack = print_context_stack_bp,
1796 perf_callchain_kernel(struct perf_callchain_entry *entry, struct pt_regs *regs)
1798 if (perf_guest_cbs && perf_guest_cbs->is_in_guest()) {
1799 /* TODO: We don't support guest os callchain now */
1803 perf_callchain_store(entry, regs->ip);
1805 dump_trace(NULL, regs, NULL, 0, &backtrace_ops, entry);
1808 #ifdef CONFIG_COMPAT
1810 perf_callchain_user32(struct pt_regs *regs, struct perf_callchain_entry *entry)
1812 /* 32-bit process in 64-bit kernel. */
1813 struct stack_frame_ia32 frame;
1814 const void __user *fp;
1816 if (!test_thread_flag(TIF_IA32))
1819 fp = compat_ptr(regs->bp);
1820 while (entry->nr < PERF_MAX_STACK_DEPTH) {
1821 unsigned long bytes;
1822 frame.next_frame = 0;
1823 frame.return_address = 0;
1825 bytes = copy_from_user_nmi(&frame, fp, sizeof(frame));
1826 if (bytes != sizeof(frame))
1829 if (fp < compat_ptr(regs->sp))
1832 perf_callchain_store(entry, frame.return_address);
1833 fp = compat_ptr(frame.next_frame);
1839 perf_callchain_user32(struct pt_regs *regs, struct perf_callchain_entry *entry)
1846 perf_callchain_user(struct perf_callchain_entry *entry, struct pt_regs *regs)
1848 struct stack_frame frame;
1849 const void __user *fp;
1851 if (perf_guest_cbs && perf_guest_cbs->is_in_guest()) {
1852 /* TODO: We don't support guest os callchain now */
1856 fp = (void __user *)regs->bp;
1858 perf_callchain_store(entry, regs->ip);
1860 if (perf_callchain_user32(regs, entry))
1863 while (entry->nr < PERF_MAX_STACK_DEPTH) {
1864 unsigned long bytes;
1865 frame.next_frame = NULL;
1866 frame.return_address = 0;
1868 bytes = copy_from_user_nmi(&frame, fp, sizeof(frame));
1869 if (bytes != sizeof(frame))
1872 if ((unsigned long)fp < regs->sp)
1875 perf_callchain_store(entry, frame.return_address);
1876 fp = frame.next_frame;
1880 unsigned long perf_instruction_pointer(struct pt_regs *regs)
1884 if (perf_guest_cbs && perf_guest_cbs->is_in_guest())
1885 ip = perf_guest_cbs->get_guest_ip();
1887 ip = instruction_pointer(regs);
1892 unsigned long perf_misc_flags(struct pt_regs *regs)
1896 if (perf_guest_cbs && perf_guest_cbs->is_in_guest()) {
1897 if (perf_guest_cbs->is_user_mode())
1898 misc |= PERF_RECORD_MISC_GUEST_USER;
1900 misc |= PERF_RECORD_MISC_GUEST_KERNEL;
1902 if (user_mode(regs))
1903 misc |= PERF_RECORD_MISC_USER;
1905 misc |= PERF_RECORD_MISC_KERNEL;
1908 if (regs->flags & PERF_EFLAGS_EXACT)
1909 misc |= PERF_RECORD_MISC_EXACT_IP;