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 void (*hw_watchdog_set_attr)(struct perf_event_attr *attr);
237 int (*hw_config)(struct perf_event *event);
238 int (*schedule_events)(struct cpu_hw_events *cpuc, int n, int *assign);
241 u64 (*event_map)(int);
244 int num_counters_fixed;
249 struct event_constraint *
250 (*get_event_constraints)(struct cpu_hw_events *cpuc,
251 struct perf_event *event);
253 void (*put_event_constraints)(struct cpu_hw_events *cpuc,
254 struct perf_event *event);
255 struct event_constraint *event_constraints;
256 struct event_constraint *percore_constraints;
257 void (*quirks)(void);
258 int perfctr_second_write;
260 int (*cpu_prepare)(int cpu);
261 void (*cpu_starting)(int cpu);
262 void (*cpu_dying)(int cpu);
263 void (*cpu_dead)(int cpu);
266 * Intel Arch Perfmon v2+
269 union perf_capabilities intel_cap;
272 * Intel DebugStore bits
275 int bts_active, pebs_active;
276 int pebs_record_size;
277 void (*drain_pebs)(struct pt_regs *regs);
278 struct event_constraint *pebs_constraints;
283 unsigned long lbr_tos, lbr_from, lbr_to; /* MSR base regs */
284 int lbr_nr; /* hardware stack size */
287 * Extra registers for events
289 struct extra_reg *extra_regs;
292 static struct x86_pmu x86_pmu __read_mostly;
294 static DEFINE_PER_CPU(struct cpu_hw_events, cpu_hw_events) = {
298 static int x86_perf_event_set_period(struct perf_event *event);
301 * Generalized hw caching related hw_event table, filled
302 * in on a per model basis. A value of 0 means
303 * 'not supported', -1 means 'hw_event makes no sense on
304 * this CPU', any other value means the raw hw_event
308 #define C(x) PERF_COUNT_HW_CACHE_##x
310 static u64 __read_mostly hw_cache_event_ids
311 [PERF_COUNT_HW_CACHE_MAX]
312 [PERF_COUNT_HW_CACHE_OP_MAX]
313 [PERF_COUNT_HW_CACHE_RESULT_MAX];
314 static u64 __read_mostly hw_cache_extra_regs
315 [PERF_COUNT_HW_CACHE_MAX]
316 [PERF_COUNT_HW_CACHE_OP_MAX]
317 [PERF_COUNT_HW_CACHE_RESULT_MAX];
319 void hw_nmi_watchdog_set_attr(struct perf_event_attr *wd_attr)
321 if (x86_pmu.hw_watchdog_set_attr)
322 x86_pmu.hw_watchdog_set_attr(wd_attr);
326 * Propagate event elapsed time into the generic event.
327 * Can only be executed on the CPU where the event is active.
328 * Returns the delta events processed.
331 x86_perf_event_update(struct perf_event *event)
333 struct hw_perf_event *hwc = &event->hw;
334 int shift = 64 - x86_pmu.cntval_bits;
335 u64 prev_raw_count, new_raw_count;
339 if (idx == X86_PMC_IDX_FIXED_BTS)
343 * Careful: an NMI might modify the previous event value.
345 * Our tactic to handle this is to first atomically read and
346 * exchange a new raw count - then add that new-prev delta
347 * count to the generic event atomically:
350 prev_raw_count = local64_read(&hwc->prev_count);
351 rdmsrl(hwc->event_base, new_raw_count);
353 if (local64_cmpxchg(&hwc->prev_count, prev_raw_count,
354 new_raw_count) != prev_raw_count)
358 * Now we have the new raw value and have updated the prev
359 * timestamp already. We can now calculate the elapsed delta
360 * (event-)time and add that to the generic event.
362 * Careful, not all hw sign-extends above the physical width
365 delta = (new_raw_count << shift) - (prev_raw_count << shift);
368 local64_add(delta, &event->count);
369 local64_sub(delta, &hwc->period_left);
371 return new_raw_count;
374 static inline int x86_pmu_addr_offset(int index)
378 /* offset = X86_FEATURE_PERFCTR_CORE ? index << 1 : index */
379 alternative_io(ASM_NOP2,
381 X86_FEATURE_PERFCTR_CORE,
388 static inline unsigned int x86_pmu_config_addr(int index)
390 return x86_pmu.eventsel + x86_pmu_addr_offset(index);
393 static inline unsigned int x86_pmu_event_addr(int index)
395 return x86_pmu.perfctr + x86_pmu_addr_offset(index);
399 * Find and validate any extra registers to set up.
401 static int x86_pmu_extra_regs(u64 config, struct perf_event *event)
403 struct extra_reg *er;
405 event->hw.extra_reg = 0;
406 event->hw.extra_config = 0;
408 if (!x86_pmu.extra_regs)
411 for (er = x86_pmu.extra_regs; er->msr; er++) {
412 if (er->event != (config & er->config_mask))
414 if (event->attr.config1 & ~er->valid_mask)
416 event->hw.extra_reg = er->msr;
417 event->hw.extra_config = event->attr.config1;
423 static atomic_t active_events;
424 static DEFINE_MUTEX(pmc_reserve_mutex);
426 #ifdef CONFIG_X86_LOCAL_APIC
428 static bool reserve_pmc_hardware(void)
432 for (i = 0; i < x86_pmu.num_counters; i++) {
433 if (!reserve_perfctr_nmi(x86_pmu_event_addr(i)))
437 for (i = 0; i < x86_pmu.num_counters; i++) {
438 if (!reserve_evntsel_nmi(x86_pmu_config_addr(i)))
445 for (i--; i >= 0; i--)
446 release_evntsel_nmi(x86_pmu_config_addr(i));
448 i = x86_pmu.num_counters;
451 for (i--; i >= 0; i--)
452 release_perfctr_nmi(x86_pmu_event_addr(i));
457 static void release_pmc_hardware(void)
461 for (i = 0; i < x86_pmu.num_counters; i++) {
462 release_perfctr_nmi(x86_pmu_event_addr(i));
463 release_evntsel_nmi(x86_pmu_config_addr(i));
469 static bool reserve_pmc_hardware(void) { return true; }
470 static void release_pmc_hardware(void) {}
474 static bool check_hw_exists(void)
476 u64 val, val_new = 0;
480 * Check to see if the BIOS enabled any of the counters, if so
483 for (i = 0; i < x86_pmu.num_counters; i++) {
484 reg = x86_pmu_config_addr(i);
485 ret = rdmsrl_safe(reg, &val);
488 if (val & ARCH_PERFMON_EVENTSEL_ENABLE)
492 if (x86_pmu.num_counters_fixed) {
493 reg = MSR_ARCH_PERFMON_FIXED_CTR_CTRL;
494 ret = rdmsrl_safe(reg, &val);
497 for (i = 0; i < x86_pmu.num_counters_fixed; i++) {
498 if (val & (0x03 << i*4))
504 * Now write a value and read it back to see if it matches,
505 * this is needed to detect certain hardware emulators (qemu/kvm)
506 * that don't trap on the MSR access and always return 0s.
509 ret = checking_wrmsrl(x86_pmu_event_addr(0), val);
510 ret |= rdmsrl_safe(x86_pmu_event_addr(0), &val_new);
511 if (ret || val != val_new)
518 * We still allow the PMU driver to operate:
520 printk(KERN_CONT "Broken BIOS detected, complain to your hardware vendor.\n");
521 printk(KERN_ERR FW_BUG "the BIOS has corrupted hw-PMU resources (MSR %x is %Lx)\n", reg, val);
526 printk(KERN_CONT "Broken PMU hardware detected, using software events only.\n");
531 static void reserve_ds_buffers(void);
532 static void release_ds_buffers(void);
534 static void hw_perf_event_destroy(struct perf_event *event)
536 if (atomic_dec_and_mutex_lock(&active_events, &pmc_reserve_mutex)) {
537 release_pmc_hardware();
538 release_ds_buffers();
539 mutex_unlock(&pmc_reserve_mutex);
543 static inline int x86_pmu_initialized(void)
545 return x86_pmu.handle_irq != NULL;
549 set_ext_hw_attr(struct hw_perf_event *hwc, struct perf_event *event)
551 struct perf_event_attr *attr = &event->attr;
552 unsigned int cache_type, cache_op, cache_result;
555 config = attr->config;
557 cache_type = (config >> 0) & 0xff;
558 if (cache_type >= PERF_COUNT_HW_CACHE_MAX)
561 cache_op = (config >> 8) & 0xff;
562 if (cache_op >= PERF_COUNT_HW_CACHE_OP_MAX)
565 cache_result = (config >> 16) & 0xff;
566 if (cache_result >= PERF_COUNT_HW_CACHE_RESULT_MAX)
569 val = hw_cache_event_ids[cache_type][cache_op][cache_result];
578 attr->config1 = hw_cache_extra_regs[cache_type][cache_op][cache_result];
579 return x86_pmu_extra_regs(val, event);
582 static int x86_setup_perfctr(struct perf_event *event)
584 struct perf_event_attr *attr = &event->attr;
585 struct hw_perf_event *hwc = &event->hw;
588 if (!is_sampling_event(event)) {
589 hwc->sample_period = x86_pmu.max_period;
590 hwc->last_period = hwc->sample_period;
591 local64_set(&hwc->period_left, hwc->sample_period);
594 * If we have a PMU initialized but no APIC
595 * interrupts, we cannot sample hardware
596 * events (user-space has to fall back and
597 * sample via a hrtimer based software event):
604 * Do not allow config1 (extended registers) to propagate,
605 * there's no sane user-space generalization yet:
607 if (attr->type == PERF_TYPE_RAW)
610 if (attr->type == PERF_TYPE_HW_CACHE)
611 return set_ext_hw_attr(hwc, event);
613 if (attr->config >= x86_pmu.max_events)
619 config = x86_pmu.event_map(attr->config);
630 if (attr->config == PERF_COUNT_HW_BRANCH_INSTRUCTIONS &&
631 !attr->freq && hwc->sample_period == 1) {
632 /* BTS is not supported by this architecture. */
633 if (!x86_pmu.bts_active)
636 /* BTS is currently only allowed for user-mode. */
637 if (!attr->exclude_kernel)
641 hwc->config |= config;
646 static int x86_pmu_hw_config(struct perf_event *event)
648 if (event->attr.precise_ip) {
651 /* Support for constant skid */
652 if (x86_pmu.pebs_active) {
655 /* Support for IP fixup */
660 if (event->attr.precise_ip > precise)
666 * (keep 'enabled' bit clear for now)
668 event->hw.config = ARCH_PERFMON_EVENTSEL_INT;
671 * Count user and OS events unless requested not to
673 if (!event->attr.exclude_user)
674 event->hw.config |= ARCH_PERFMON_EVENTSEL_USR;
675 if (!event->attr.exclude_kernel)
676 event->hw.config |= ARCH_PERFMON_EVENTSEL_OS;
678 if (event->attr.type == PERF_TYPE_RAW)
679 event->hw.config |= event->attr.config & X86_RAW_EVENT_MASK;
681 return x86_setup_perfctr(event);
685 * Setup the hardware configuration for a given attr_type
687 static int __x86_pmu_event_init(struct perf_event *event)
691 if (!x86_pmu_initialized())
695 if (!atomic_inc_not_zero(&active_events)) {
696 mutex_lock(&pmc_reserve_mutex);
697 if (atomic_read(&active_events) == 0) {
698 if (!reserve_pmc_hardware())
701 reserve_ds_buffers();
704 atomic_inc(&active_events);
705 mutex_unlock(&pmc_reserve_mutex);
710 event->destroy = hw_perf_event_destroy;
713 event->hw.last_cpu = -1;
714 event->hw.last_tag = ~0ULL;
716 return x86_pmu.hw_config(event);
719 static void x86_pmu_disable_all(void)
721 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
724 for (idx = 0; idx < x86_pmu.num_counters; idx++) {
727 if (!test_bit(idx, cpuc->active_mask))
729 rdmsrl(x86_pmu_config_addr(idx), val);
730 if (!(val & ARCH_PERFMON_EVENTSEL_ENABLE))
732 val &= ~ARCH_PERFMON_EVENTSEL_ENABLE;
733 wrmsrl(x86_pmu_config_addr(idx), val);
737 static void x86_pmu_disable(struct pmu *pmu)
739 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
741 if (!x86_pmu_initialized())
751 x86_pmu.disable_all();
754 static inline void __x86_pmu_enable_event(struct hw_perf_event *hwc,
758 wrmsrl(hwc->extra_reg, hwc->extra_config);
759 wrmsrl(hwc->config_base, hwc->config | enable_mask);
762 static void x86_pmu_enable_all(int added)
764 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
767 for (idx = 0; idx < x86_pmu.num_counters; idx++) {
768 struct hw_perf_event *hwc = &cpuc->events[idx]->hw;
770 if (!test_bit(idx, cpuc->active_mask))
773 __x86_pmu_enable_event(hwc, ARCH_PERFMON_EVENTSEL_ENABLE);
777 static struct pmu pmu;
779 static inline int is_x86_event(struct perf_event *event)
781 return event->pmu == &pmu;
784 static int x86_schedule_events(struct cpu_hw_events *cpuc, int n, int *assign)
786 struct event_constraint *c, *constraints[X86_PMC_IDX_MAX];
787 unsigned long used_mask[BITS_TO_LONGS(X86_PMC_IDX_MAX)];
788 int i, j, w, wmax, num = 0;
789 struct hw_perf_event *hwc;
791 bitmap_zero(used_mask, X86_PMC_IDX_MAX);
793 for (i = 0; i < n; i++) {
794 c = x86_pmu.get_event_constraints(cpuc, cpuc->event_list[i]);
799 * fastpath, try to reuse previous register
801 for (i = 0; i < n; i++) {
802 hwc = &cpuc->event_list[i]->hw;
809 /* constraint still honored */
810 if (!test_bit(hwc->idx, c->idxmsk))
813 /* not already used */
814 if (test_bit(hwc->idx, used_mask))
817 __set_bit(hwc->idx, used_mask);
819 assign[i] = hwc->idx;
828 bitmap_zero(used_mask, X86_PMC_IDX_MAX);
831 * weight = number of possible counters
833 * 1 = most constrained, only works on one counter
834 * wmax = least constrained, works on any counter
836 * assign events to counters starting with most
837 * constrained events.
839 wmax = x86_pmu.num_counters;
842 * when fixed event counters are present,
843 * wmax is incremented by 1 to account
844 * for one more choice
846 if (x86_pmu.num_counters_fixed)
849 for (w = 1, num = n; num && w <= wmax; w++) {
851 for (i = 0; num && i < n; i++) {
853 hwc = &cpuc->event_list[i]->hw;
858 for_each_set_bit(j, c->idxmsk, X86_PMC_IDX_MAX) {
859 if (!test_bit(j, used_mask))
863 if (j == X86_PMC_IDX_MAX)
866 __set_bit(j, used_mask);
875 * scheduling failed or is just a simulation,
876 * free resources if necessary
878 if (!assign || num) {
879 for (i = 0; i < n; i++) {
880 if (x86_pmu.put_event_constraints)
881 x86_pmu.put_event_constraints(cpuc, cpuc->event_list[i]);
884 return num ? -ENOSPC : 0;
888 * dogrp: true if must collect siblings events (group)
889 * returns total number of events and error code
891 static int collect_events(struct cpu_hw_events *cpuc, struct perf_event *leader, bool dogrp)
893 struct perf_event *event;
896 max_count = x86_pmu.num_counters + x86_pmu.num_counters_fixed;
898 /* current number of events already accepted */
901 if (is_x86_event(leader)) {
904 cpuc->event_list[n] = leader;
910 list_for_each_entry(event, &leader->sibling_list, group_entry) {
911 if (!is_x86_event(event) ||
912 event->state <= PERF_EVENT_STATE_OFF)
918 cpuc->event_list[n] = event;
924 static inline void x86_assign_hw_event(struct perf_event *event,
925 struct cpu_hw_events *cpuc, int i)
927 struct hw_perf_event *hwc = &event->hw;
929 hwc->idx = cpuc->assign[i];
930 hwc->last_cpu = smp_processor_id();
931 hwc->last_tag = ++cpuc->tags[i];
933 if (hwc->idx == X86_PMC_IDX_FIXED_BTS) {
934 hwc->config_base = 0;
936 } else if (hwc->idx >= X86_PMC_IDX_FIXED) {
937 hwc->config_base = MSR_ARCH_PERFMON_FIXED_CTR_CTRL;
938 hwc->event_base = MSR_ARCH_PERFMON_FIXED_CTR0 + (hwc->idx - X86_PMC_IDX_FIXED);
940 hwc->config_base = x86_pmu_config_addr(hwc->idx);
941 hwc->event_base = x86_pmu_event_addr(hwc->idx);
945 static inline int match_prev_assignment(struct hw_perf_event *hwc,
946 struct cpu_hw_events *cpuc,
949 return hwc->idx == cpuc->assign[i] &&
950 hwc->last_cpu == smp_processor_id() &&
951 hwc->last_tag == cpuc->tags[i];
954 static void x86_pmu_start(struct perf_event *event, int flags);
955 static void x86_pmu_stop(struct perf_event *event, int flags);
957 static void x86_pmu_enable(struct pmu *pmu)
959 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
960 struct perf_event *event;
961 struct hw_perf_event *hwc;
962 int i, added = cpuc->n_added;
964 if (!x86_pmu_initialized())
971 int n_running = cpuc->n_events - cpuc->n_added;
973 * apply assignment obtained either from
974 * hw_perf_group_sched_in() or x86_pmu_enable()
976 * step1: save events moving to new counters
977 * step2: reprogram moved events into new counters
979 for (i = 0; i < n_running; i++) {
980 event = cpuc->event_list[i];
984 * we can avoid reprogramming counter if:
985 * - assigned same counter as last time
986 * - running on same CPU as last time
987 * - no other event has used the counter since
989 if (hwc->idx == -1 ||
990 match_prev_assignment(hwc, cpuc, i))
994 * Ensure we don't accidentally enable a stopped
995 * counter simply because we rescheduled.
997 if (hwc->state & PERF_HES_STOPPED)
998 hwc->state |= PERF_HES_ARCH;
1000 x86_pmu_stop(event, PERF_EF_UPDATE);
1003 for (i = 0; i < cpuc->n_events; i++) {
1004 event = cpuc->event_list[i];
1007 if (!match_prev_assignment(hwc, cpuc, i))
1008 x86_assign_hw_event(event, cpuc, i);
1009 else if (i < n_running)
1012 if (hwc->state & PERF_HES_ARCH)
1015 x86_pmu_start(event, PERF_EF_RELOAD);
1018 perf_events_lapic_init();
1024 x86_pmu.enable_all(added);
1027 static inline void x86_pmu_disable_event(struct perf_event *event)
1029 struct hw_perf_event *hwc = &event->hw;
1031 wrmsrl(hwc->config_base, hwc->config);
1034 static DEFINE_PER_CPU(u64 [X86_PMC_IDX_MAX], pmc_prev_left);
1037 * Set the next IRQ period, based on the hwc->period_left value.
1038 * To be called with the event disabled in hw:
1041 x86_perf_event_set_period(struct perf_event *event)
1043 struct hw_perf_event *hwc = &event->hw;
1044 s64 left = local64_read(&hwc->period_left);
1045 s64 period = hwc->sample_period;
1046 int ret = 0, idx = hwc->idx;
1048 if (idx == X86_PMC_IDX_FIXED_BTS)
1052 * If we are way outside a reasonable range then just skip forward:
1054 if (unlikely(left <= -period)) {
1056 local64_set(&hwc->period_left, left);
1057 hwc->last_period = period;
1061 if (unlikely(left <= 0)) {
1063 local64_set(&hwc->period_left, left);
1064 hwc->last_period = period;
1068 * Quirk: certain CPUs dont like it if just 1 hw_event is left:
1070 if (unlikely(left < 2))
1073 if (left > x86_pmu.max_period)
1074 left = x86_pmu.max_period;
1076 per_cpu(pmc_prev_left[idx], smp_processor_id()) = left;
1079 * The hw event starts counting from this event offset,
1080 * mark it to be able to extra future deltas:
1082 local64_set(&hwc->prev_count, (u64)-left);
1084 wrmsrl(hwc->event_base, (u64)(-left) & x86_pmu.cntval_mask);
1087 * Due to erratum on certan cpu we need
1088 * a second write to be sure the register
1089 * is updated properly
1091 if (x86_pmu.perfctr_second_write) {
1092 wrmsrl(hwc->event_base,
1093 (u64)(-left) & x86_pmu.cntval_mask);
1096 perf_event_update_userpage(event);
1101 static void x86_pmu_enable_event(struct perf_event *event)
1103 if (__this_cpu_read(cpu_hw_events.enabled))
1104 __x86_pmu_enable_event(&event->hw,
1105 ARCH_PERFMON_EVENTSEL_ENABLE);
1109 * Add a single event to the PMU.
1111 * The event is added to the group of enabled events
1112 * but only if it can be scehduled with existing events.
1114 static int x86_pmu_add(struct perf_event *event, int flags)
1116 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
1117 struct hw_perf_event *hwc;
1118 int assign[X86_PMC_IDX_MAX];
1123 perf_pmu_disable(event->pmu);
1124 n0 = cpuc->n_events;
1125 ret = n = collect_events(cpuc, event, false);
1129 hwc->state = PERF_HES_UPTODATE | PERF_HES_STOPPED;
1130 if (!(flags & PERF_EF_START))
1131 hwc->state |= PERF_HES_ARCH;
1134 * If group events scheduling transaction was started,
1135 * skip the schedulability test here, it will be performed
1136 * at commit time (->commit_txn) as a whole
1138 if (cpuc->group_flag & PERF_EVENT_TXN)
1141 ret = x86_pmu.schedule_events(cpuc, n, assign);
1145 * copy new assignment, now we know it is possible
1146 * will be used by hw_perf_enable()
1148 memcpy(cpuc->assign, assign, n*sizeof(int));
1152 cpuc->n_added += n - n0;
1153 cpuc->n_txn += n - n0;
1157 perf_pmu_enable(event->pmu);
1161 static void x86_pmu_start(struct perf_event *event, int flags)
1163 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
1164 int idx = event->hw.idx;
1166 if (WARN_ON_ONCE(!(event->hw.state & PERF_HES_STOPPED)))
1169 if (WARN_ON_ONCE(idx == -1))
1172 if (flags & PERF_EF_RELOAD) {
1173 WARN_ON_ONCE(!(event->hw.state & PERF_HES_UPTODATE));
1174 x86_perf_event_set_period(event);
1177 event->hw.state = 0;
1179 cpuc->events[idx] = event;
1180 __set_bit(idx, cpuc->active_mask);
1181 __set_bit(idx, cpuc->running);
1182 x86_pmu.enable(event);
1183 perf_event_update_userpage(event);
1186 void perf_event_print_debug(void)
1188 u64 ctrl, status, overflow, pmc_ctrl, pmc_count, prev_left, fixed;
1190 struct cpu_hw_events *cpuc;
1191 unsigned long flags;
1194 if (!x86_pmu.num_counters)
1197 local_irq_save(flags);
1199 cpu = smp_processor_id();
1200 cpuc = &per_cpu(cpu_hw_events, cpu);
1202 if (x86_pmu.version >= 2) {
1203 rdmsrl(MSR_CORE_PERF_GLOBAL_CTRL, ctrl);
1204 rdmsrl(MSR_CORE_PERF_GLOBAL_STATUS, status);
1205 rdmsrl(MSR_CORE_PERF_GLOBAL_OVF_CTRL, overflow);
1206 rdmsrl(MSR_ARCH_PERFMON_FIXED_CTR_CTRL, fixed);
1207 rdmsrl(MSR_IA32_PEBS_ENABLE, pebs);
1210 pr_info("CPU#%d: ctrl: %016llx\n", cpu, ctrl);
1211 pr_info("CPU#%d: status: %016llx\n", cpu, status);
1212 pr_info("CPU#%d: overflow: %016llx\n", cpu, overflow);
1213 pr_info("CPU#%d: fixed: %016llx\n", cpu, fixed);
1214 pr_info("CPU#%d: pebs: %016llx\n", cpu, pebs);
1216 pr_info("CPU#%d: active: %016llx\n", cpu, *(u64 *)cpuc->active_mask);
1218 for (idx = 0; idx < x86_pmu.num_counters; idx++) {
1219 rdmsrl(x86_pmu_config_addr(idx), pmc_ctrl);
1220 rdmsrl(x86_pmu_event_addr(idx), pmc_count);
1222 prev_left = per_cpu(pmc_prev_left[idx], cpu);
1224 pr_info("CPU#%d: gen-PMC%d ctrl: %016llx\n",
1225 cpu, idx, pmc_ctrl);
1226 pr_info("CPU#%d: gen-PMC%d count: %016llx\n",
1227 cpu, idx, pmc_count);
1228 pr_info("CPU#%d: gen-PMC%d left: %016llx\n",
1229 cpu, idx, prev_left);
1231 for (idx = 0; idx < x86_pmu.num_counters_fixed; idx++) {
1232 rdmsrl(MSR_ARCH_PERFMON_FIXED_CTR0 + idx, pmc_count);
1234 pr_info("CPU#%d: fixed-PMC%d count: %016llx\n",
1235 cpu, idx, pmc_count);
1237 local_irq_restore(flags);
1240 static void x86_pmu_stop(struct perf_event *event, int flags)
1242 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
1243 struct hw_perf_event *hwc = &event->hw;
1245 if (__test_and_clear_bit(hwc->idx, cpuc->active_mask)) {
1246 x86_pmu.disable(event);
1247 cpuc->events[hwc->idx] = NULL;
1248 WARN_ON_ONCE(hwc->state & PERF_HES_STOPPED);
1249 hwc->state |= PERF_HES_STOPPED;
1252 if ((flags & PERF_EF_UPDATE) && !(hwc->state & PERF_HES_UPTODATE)) {
1254 * Drain the remaining delta count out of a event
1255 * that we are disabling:
1257 x86_perf_event_update(event);
1258 hwc->state |= PERF_HES_UPTODATE;
1262 static void x86_pmu_del(struct perf_event *event, int flags)
1264 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
1268 * If we're called during a txn, we don't need to do anything.
1269 * The events never got scheduled and ->cancel_txn will truncate
1272 if (cpuc->group_flag & PERF_EVENT_TXN)
1275 x86_pmu_stop(event, PERF_EF_UPDATE);
1277 for (i = 0; i < cpuc->n_events; i++) {
1278 if (event == cpuc->event_list[i]) {
1280 if (x86_pmu.put_event_constraints)
1281 x86_pmu.put_event_constraints(cpuc, event);
1283 while (++i < cpuc->n_events)
1284 cpuc->event_list[i-1] = cpuc->event_list[i];
1290 perf_event_update_userpage(event);
1293 static int x86_pmu_handle_irq(struct pt_regs *regs)
1295 struct perf_sample_data data;
1296 struct cpu_hw_events *cpuc;
1297 struct perf_event *event;
1298 int idx, handled = 0;
1301 perf_sample_data_init(&data, 0);
1303 cpuc = &__get_cpu_var(cpu_hw_events);
1306 * Some chipsets need to unmask the LVTPC in a particular spot
1307 * inside the nmi handler. As a result, the unmasking was pushed
1308 * into all the nmi handlers.
1310 * This generic handler doesn't seem to have any issues where the
1311 * unmasking occurs so it was left at the top.
1313 apic_write(APIC_LVTPC, APIC_DM_NMI);
1315 for (idx = 0; idx < x86_pmu.num_counters; idx++) {
1316 if (!test_bit(idx, cpuc->active_mask)) {
1318 * Though we deactivated the counter some cpus
1319 * might still deliver spurious interrupts still
1320 * in flight. Catch them:
1322 if (__test_and_clear_bit(idx, cpuc->running))
1327 event = cpuc->events[idx];
1329 val = x86_perf_event_update(event);
1330 if (val & (1ULL << (x86_pmu.cntval_bits - 1)))
1337 data.period = event->hw.last_period;
1339 if (!x86_perf_event_set_period(event))
1342 if (perf_event_overflow(event, &data, regs))
1343 x86_pmu_stop(event, 0);
1347 inc_irq_stat(apic_perf_irqs);
1352 void perf_events_lapic_init(void)
1354 if (!x86_pmu.apic || !x86_pmu_initialized())
1358 * Always use NMI for PMU
1360 apic_write(APIC_LVTPC, APIC_DM_NMI);
1363 struct pmu_nmi_state {
1364 unsigned int marked;
1368 static DEFINE_PER_CPU(struct pmu_nmi_state, pmu_nmi);
1370 static int __kprobes
1371 perf_event_nmi_handler(struct notifier_block *self,
1372 unsigned long cmd, void *__args)
1374 struct die_args *args = __args;
1375 unsigned int this_nmi;
1378 if (!atomic_read(&active_events))
1384 case DIE_NMIUNKNOWN:
1385 this_nmi = percpu_read(irq_stat.__nmi_count);
1386 if (this_nmi != __this_cpu_read(pmu_nmi.marked))
1387 /* let the kernel handle the unknown nmi */
1390 * This one is a PMU back-to-back nmi. Two events
1391 * trigger 'simultaneously' raising two back-to-back
1392 * NMIs. If the first NMI handles both, the latter
1393 * will be empty and daze the CPU. So, we drop it to
1394 * avoid false-positive 'unknown nmi' messages.
1401 handled = x86_pmu.handle_irq(args->regs);
1405 this_nmi = percpu_read(irq_stat.__nmi_count);
1406 if ((handled > 1) ||
1407 /* the next nmi could be a back-to-back nmi */
1408 ((__this_cpu_read(pmu_nmi.marked) == this_nmi) &&
1409 (__this_cpu_read(pmu_nmi.handled) > 1))) {
1411 * We could have two subsequent back-to-back nmis: The
1412 * first handles more than one counter, the 2nd
1413 * handles only one counter and the 3rd handles no
1416 * This is the 2nd nmi because the previous was
1417 * handling more than one counter. We will mark the
1418 * next (3rd) and then drop it if unhandled.
1420 __this_cpu_write(pmu_nmi.marked, this_nmi + 1);
1421 __this_cpu_write(pmu_nmi.handled, handled);
1427 static __read_mostly struct notifier_block perf_event_nmi_notifier = {
1428 .notifier_call = perf_event_nmi_handler,
1430 .priority = NMI_LOCAL_LOW_PRIOR,
1433 static struct event_constraint unconstrained;
1434 static struct event_constraint emptyconstraint;
1436 static struct event_constraint *
1437 x86_get_event_constraints(struct cpu_hw_events *cpuc, struct perf_event *event)
1439 struct event_constraint *c;
1441 if (x86_pmu.event_constraints) {
1442 for_each_event_constraint(c, x86_pmu.event_constraints) {
1443 if ((event->hw.config & c->cmask) == c->code)
1448 return &unconstrained;
1451 #include "perf_event_amd.c"
1452 #include "perf_event_p6.c"
1453 #include "perf_event_p4.c"
1454 #include "perf_event_intel_lbr.c"
1455 #include "perf_event_intel_ds.c"
1456 #include "perf_event_intel.c"
1458 static int __cpuinit
1459 x86_pmu_notifier(struct notifier_block *self, unsigned long action, void *hcpu)
1461 unsigned int cpu = (long)hcpu;
1462 int ret = NOTIFY_OK;
1464 switch (action & ~CPU_TASKS_FROZEN) {
1465 case CPU_UP_PREPARE:
1466 if (x86_pmu.cpu_prepare)
1467 ret = x86_pmu.cpu_prepare(cpu);
1471 if (x86_pmu.cpu_starting)
1472 x86_pmu.cpu_starting(cpu);
1476 if (x86_pmu.cpu_dying)
1477 x86_pmu.cpu_dying(cpu);
1480 case CPU_UP_CANCELED:
1482 if (x86_pmu.cpu_dead)
1483 x86_pmu.cpu_dead(cpu);
1493 static void __init pmu_check_apic(void)
1499 pr_info("no APIC, boot with the \"lapic\" boot parameter to force-enable it.\n");
1500 pr_info("no hardware sampling interrupt available.\n");
1503 static int __init init_hw_perf_events(void)
1505 struct event_constraint *c;
1508 pr_info("Performance Events: ");
1510 switch (boot_cpu_data.x86_vendor) {
1511 case X86_VENDOR_INTEL:
1512 err = intel_pmu_init();
1514 case X86_VENDOR_AMD:
1515 err = amd_pmu_init();
1521 pr_cont("no PMU driver, software events only.\n");
1527 /* sanity check that the hardware exists or is emulated */
1528 if (!check_hw_exists())
1531 pr_cont("%s PMU driver.\n", x86_pmu.name);
1536 if (x86_pmu.num_counters > X86_PMC_MAX_GENERIC) {
1537 WARN(1, KERN_ERR "hw perf events %d > max(%d), clipping!",
1538 x86_pmu.num_counters, X86_PMC_MAX_GENERIC);
1539 x86_pmu.num_counters = X86_PMC_MAX_GENERIC;
1541 x86_pmu.intel_ctrl = (1 << x86_pmu.num_counters) - 1;
1543 if (x86_pmu.num_counters_fixed > X86_PMC_MAX_FIXED) {
1544 WARN(1, KERN_ERR "hw perf events fixed %d > max(%d), clipping!",
1545 x86_pmu.num_counters_fixed, X86_PMC_MAX_FIXED);
1546 x86_pmu.num_counters_fixed = X86_PMC_MAX_FIXED;
1549 x86_pmu.intel_ctrl |=
1550 ((1LL << x86_pmu.num_counters_fixed)-1) << X86_PMC_IDX_FIXED;
1552 perf_events_lapic_init();
1553 register_die_notifier(&perf_event_nmi_notifier);
1555 unconstrained = (struct event_constraint)
1556 __EVENT_CONSTRAINT(0, (1ULL << x86_pmu.num_counters) - 1,
1557 0, x86_pmu.num_counters);
1559 if (x86_pmu.event_constraints) {
1560 for_each_event_constraint(c, x86_pmu.event_constraints) {
1561 if (c->cmask != X86_RAW_EVENT_MASK)
1564 c->idxmsk64 |= (1ULL << x86_pmu.num_counters) - 1;
1565 c->weight += x86_pmu.num_counters;
1569 pr_info("... version: %d\n", x86_pmu.version);
1570 pr_info("... bit width: %d\n", x86_pmu.cntval_bits);
1571 pr_info("... generic registers: %d\n", x86_pmu.num_counters);
1572 pr_info("... value mask: %016Lx\n", x86_pmu.cntval_mask);
1573 pr_info("... max period: %016Lx\n", x86_pmu.max_period);
1574 pr_info("... fixed-purpose events: %d\n", x86_pmu.num_counters_fixed);
1575 pr_info("... event mask: %016Lx\n", x86_pmu.intel_ctrl);
1577 perf_pmu_register(&pmu, "cpu", PERF_TYPE_RAW);
1578 perf_cpu_notifier(x86_pmu_notifier);
1582 early_initcall(init_hw_perf_events);
1584 static inline void x86_pmu_read(struct perf_event *event)
1586 x86_perf_event_update(event);
1590 * Start group events scheduling transaction
1591 * Set the flag to make pmu::enable() not perform the
1592 * schedulability test, it will be performed at commit time
1594 static void x86_pmu_start_txn(struct pmu *pmu)
1596 perf_pmu_disable(pmu);
1597 __this_cpu_or(cpu_hw_events.group_flag, PERF_EVENT_TXN);
1598 __this_cpu_write(cpu_hw_events.n_txn, 0);
1602 * Stop group events scheduling transaction
1603 * Clear the flag and pmu::enable() will perform the
1604 * schedulability test.
1606 static void x86_pmu_cancel_txn(struct pmu *pmu)
1608 __this_cpu_and(cpu_hw_events.group_flag, ~PERF_EVENT_TXN);
1610 * Truncate the collected events.
1612 __this_cpu_sub(cpu_hw_events.n_added, __this_cpu_read(cpu_hw_events.n_txn));
1613 __this_cpu_sub(cpu_hw_events.n_events, __this_cpu_read(cpu_hw_events.n_txn));
1614 perf_pmu_enable(pmu);
1618 * Commit group events scheduling transaction
1619 * Perform the group schedulability test as a whole
1620 * Return 0 if success
1622 static int x86_pmu_commit_txn(struct pmu *pmu)
1624 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
1625 int assign[X86_PMC_IDX_MAX];
1630 if (!x86_pmu_initialized())
1633 ret = x86_pmu.schedule_events(cpuc, n, assign);
1638 * copy new assignment, now we know it is possible
1639 * will be used by hw_perf_enable()
1641 memcpy(cpuc->assign, assign, n*sizeof(int));
1643 cpuc->group_flag &= ~PERF_EVENT_TXN;
1644 perf_pmu_enable(pmu);
1649 * validate that we can schedule this event
1651 static int validate_event(struct perf_event *event)
1653 struct cpu_hw_events *fake_cpuc;
1654 struct event_constraint *c;
1657 fake_cpuc = kmalloc(sizeof(*fake_cpuc), GFP_KERNEL | __GFP_ZERO);
1661 c = x86_pmu.get_event_constraints(fake_cpuc, event);
1663 if (!c || !c->weight)
1666 if (x86_pmu.put_event_constraints)
1667 x86_pmu.put_event_constraints(fake_cpuc, event);
1675 * validate a single event group
1677 * validation include:
1678 * - check events are compatible which each other
1679 * - events do not compete for the same counter
1680 * - number of events <= number of counters
1682 * validation ensures the group can be loaded onto the
1683 * PMU if it was the only group available.
1685 static int validate_group(struct perf_event *event)
1687 struct perf_event *leader = event->group_leader;
1688 struct cpu_hw_events *fake_cpuc;
1692 fake_cpuc = kmalloc(sizeof(*fake_cpuc), GFP_KERNEL | __GFP_ZERO);
1697 * the event is not yet connected with its
1698 * siblings therefore we must first collect
1699 * existing siblings, then add the new event
1700 * before we can simulate the scheduling
1703 n = collect_events(fake_cpuc, leader, true);
1707 fake_cpuc->n_events = n;
1708 n = collect_events(fake_cpuc, event, false);
1712 fake_cpuc->n_events = n;
1714 ret = x86_pmu.schedule_events(fake_cpuc, n, NULL);
1722 static int x86_pmu_event_init(struct perf_event *event)
1727 switch (event->attr.type) {
1729 case PERF_TYPE_HARDWARE:
1730 case PERF_TYPE_HW_CACHE:
1737 err = __x86_pmu_event_init(event);
1740 * we temporarily connect event to its pmu
1741 * such that validate_group() can classify
1742 * it as an x86 event using is_x86_event()
1747 if (event->group_leader != event)
1748 err = validate_group(event);
1750 err = validate_event(event);
1756 event->destroy(event);
1762 static struct pmu pmu = {
1763 .pmu_enable = x86_pmu_enable,
1764 .pmu_disable = x86_pmu_disable,
1766 .event_init = x86_pmu_event_init,
1770 .start = x86_pmu_start,
1771 .stop = x86_pmu_stop,
1772 .read = x86_pmu_read,
1774 .start_txn = x86_pmu_start_txn,
1775 .cancel_txn = x86_pmu_cancel_txn,
1776 .commit_txn = x86_pmu_commit_txn,
1783 static int backtrace_stack(void *data, char *name)
1788 static void backtrace_address(void *data, unsigned long addr, int reliable)
1790 struct perf_callchain_entry *entry = data;
1792 perf_callchain_store(entry, addr);
1795 static const struct stacktrace_ops backtrace_ops = {
1796 .stack = backtrace_stack,
1797 .address = backtrace_address,
1798 .walk_stack = print_context_stack_bp,
1802 perf_callchain_kernel(struct perf_callchain_entry *entry, struct pt_regs *regs)
1804 if (perf_guest_cbs && perf_guest_cbs->is_in_guest()) {
1805 /* TODO: We don't support guest os callchain now */
1809 perf_callchain_store(entry, regs->ip);
1811 dump_trace(NULL, regs, NULL, 0, &backtrace_ops, entry);
1814 #ifdef CONFIG_COMPAT
1816 perf_callchain_user32(struct pt_regs *regs, struct perf_callchain_entry *entry)
1818 /* 32-bit process in 64-bit kernel. */
1819 struct stack_frame_ia32 frame;
1820 const void __user *fp;
1822 if (!test_thread_flag(TIF_IA32))
1825 fp = compat_ptr(regs->bp);
1826 while (entry->nr < PERF_MAX_STACK_DEPTH) {
1827 unsigned long bytes;
1828 frame.next_frame = 0;
1829 frame.return_address = 0;
1831 bytes = copy_from_user_nmi(&frame, fp, sizeof(frame));
1832 if (bytes != sizeof(frame))
1835 if (fp < compat_ptr(regs->sp))
1838 perf_callchain_store(entry, frame.return_address);
1839 fp = compat_ptr(frame.next_frame);
1845 perf_callchain_user32(struct pt_regs *regs, struct perf_callchain_entry *entry)
1852 perf_callchain_user(struct perf_callchain_entry *entry, struct pt_regs *regs)
1854 struct stack_frame frame;
1855 const void __user *fp;
1857 if (perf_guest_cbs && perf_guest_cbs->is_in_guest()) {
1858 /* TODO: We don't support guest os callchain now */
1862 fp = (void __user *)regs->bp;
1864 perf_callchain_store(entry, regs->ip);
1866 if (perf_callchain_user32(regs, entry))
1869 while (entry->nr < PERF_MAX_STACK_DEPTH) {
1870 unsigned long bytes;
1871 frame.next_frame = NULL;
1872 frame.return_address = 0;
1874 bytes = copy_from_user_nmi(&frame, fp, sizeof(frame));
1875 if (bytes != sizeof(frame))
1878 if ((unsigned long)fp < regs->sp)
1881 perf_callchain_store(entry, frame.return_address);
1882 fp = frame.next_frame;
1886 unsigned long perf_instruction_pointer(struct pt_regs *regs)
1890 if (perf_guest_cbs && perf_guest_cbs->is_in_guest())
1891 ip = perf_guest_cbs->get_guest_ip();
1893 ip = instruction_pointer(regs);
1898 unsigned long perf_misc_flags(struct pt_regs *regs)
1902 if (perf_guest_cbs && perf_guest_cbs->is_in_guest()) {
1903 if (perf_guest_cbs->is_user_mode())
1904 misc |= PERF_RECORD_MISC_GUEST_USER;
1906 misc |= PERF_RECORD_MISC_GUEST_KERNEL;
1908 if (user_mode(regs))
1909 misc |= PERF_RECORD_MISC_USER;
1911 misc |= PERF_RECORD_MISC_KERNEL;
1914 if (regs->flags & PERF_EFLAGS_EXACT)
1915 misc |= PERF_RECORD_MISC_EXACT_IP;