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
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/cpu.h>
26 #include <linux/bitops.h>
27 #include <linux/device.h>
30 #include <asm/stacktrace.h>
33 #include <asm/alternative.h>
34 #include <asm/mmu_context.h>
35 #include <asm/tlbflush.h>
36 #include <asm/timer.h>
40 #include "perf_event.h"
42 struct x86_pmu x86_pmu __read_mostly;
44 DEFINE_PER_CPU(struct cpu_hw_events, cpu_hw_events) = {
48 struct static_key rdpmc_always_available = STATIC_KEY_INIT_FALSE;
50 u64 __read_mostly hw_cache_event_ids
51 [PERF_COUNT_HW_CACHE_MAX]
52 [PERF_COUNT_HW_CACHE_OP_MAX]
53 [PERF_COUNT_HW_CACHE_RESULT_MAX];
54 u64 __read_mostly hw_cache_extra_regs
55 [PERF_COUNT_HW_CACHE_MAX]
56 [PERF_COUNT_HW_CACHE_OP_MAX]
57 [PERF_COUNT_HW_CACHE_RESULT_MAX];
60 * Propagate event elapsed time into the generic event.
61 * Can only be executed on the CPU where the event is active.
62 * Returns the delta events processed.
64 u64 x86_perf_event_update(struct perf_event *event)
66 struct hw_perf_event *hwc = &event->hw;
67 int shift = 64 - x86_pmu.cntval_bits;
68 u64 prev_raw_count, new_raw_count;
72 if (idx == INTEL_PMC_IDX_FIXED_BTS)
76 * Careful: an NMI might modify the previous event value.
78 * Our tactic to handle this is to first atomically read and
79 * exchange a new raw count - then add that new-prev delta
80 * count to the generic event atomically:
83 prev_raw_count = local64_read(&hwc->prev_count);
84 rdpmcl(hwc->event_base_rdpmc, new_raw_count);
86 if (local64_cmpxchg(&hwc->prev_count, prev_raw_count,
87 new_raw_count) != prev_raw_count)
91 * Now we have the new raw value and have updated the prev
92 * timestamp already. We can now calculate the elapsed delta
93 * (event-)time and add that to the generic event.
95 * Careful, not all hw sign-extends above the physical width
98 delta = (new_raw_count << shift) - (prev_raw_count << shift);
101 local64_add(delta, &event->count);
102 local64_sub(delta, &hwc->period_left);
104 return new_raw_count;
108 * Find and validate any extra registers to set up.
110 static int x86_pmu_extra_regs(u64 config, struct perf_event *event)
112 struct hw_perf_event_extra *reg;
113 struct extra_reg *er;
115 reg = &event->hw.extra_reg;
117 if (!x86_pmu.extra_regs)
120 for (er = x86_pmu.extra_regs; er->msr; er++) {
121 if (er->event != (config & er->config_mask))
123 if (event->attr.config1 & ~er->valid_mask)
125 /* Check if the extra msrs can be safely accessed*/
126 if (!er->extra_msr_access)
130 reg->config = event->attr.config1;
137 static atomic_t active_events;
138 static atomic_t pmc_refcount;
139 static DEFINE_MUTEX(pmc_reserve_mutex);
141 #ifdef CONFIG_X86_LOCAL_APIC
143 static bool reserve_pmc_hardware(void)
147 for (i = 0; i < x86_pmu.num_counters; i++) {
148 if (!reserve_perfctr_nmi(x86_pmu_event_addr(i)))
152 for (i = 0; i < x86_pmu.num_counters; i++) {
153 if (!reserve_evntsel_nmi(x86_pmu_config_addr(i)))
160 for (i--; i >= 0; i--)
161 release_evntsel_nmi(x86_pmu_config_addr(i));
163 i = x86_pmu.num_counters;
166 for (i--; i >= 0; i--)
167 release_perfctr_nmi(x86_pmu_event_addr(i));
172 static void release_pmc_hardware(void)
176 for (i = 0; i < x86_pmu.num_counters; i++) {
177 release_perfctr_nmi(x86_pmu_event_addr(i));
178 release_evntsel_nmi(x86_pmu_config_addr(i));
184 static bool reserve_pmc_hardware(void) { return true; }
185 static void release_pmc_hardware(void) {}
189 static bool check_hw_exists(void)
191 u64 val, val_fail, val_new= ~0;
192 int i, reg, reg_fail, ret = 0;
197 * Check to see if the BIOS enabled any of the counters, if so
200 for (i = 0; i < x86_pmu.num_counters; i++) {
201 reg = x86_pmu_config_addr(i);
202 ret = rdmsrl_safe(reg, &val);
205 if (val & ARCH_PERFMON_EVENTSEL_ENABLE) {
214 if (x86_pmu.num_counters_fixed) {
215 reg = MSR_ARCH_PERFMON_FIXED_CTR_CTRL;
216 ret = rdmsrl_safe(reg, &val);
219 for (i = 0; i < x86_pmu.num_counters_fixed; i++) {
220 if (val & (0x03 << i*4)) {
229 * If all the counters are enabled, the below test will always
230 * fail. The tools will also become useless in this scenario.
231 * Just fail and disable the hardware counters.
234 if (reg_safe == -1) {
240 * Read the current value, change it and read it back to see if it
241 * matches, this is needed to detect certain hardware emulators
242 * (qemu/kvm) that don't trap on the MSR access and always return 0s.
244 reg = x86_pmu_event_addr(reg_safe);
245 if (rdmsrl_safe(reg, &val))
248 ret = wrmsrl_safe(reg, val);
249 ret |= rdmsrl_safe(reg, &val_new);
250 if (ret || val != val_new)
254 * We still allow the PMU driver to operate:
257 pr_cont("Broken BIOS detected, complain to your hardware vendor.\n");
258 pr_err(FW_BUG "the BIOS has corrupted hw-PMU resources (MSR %x is %Lx)\n",
265 if (boot_cpu_has(X86_FEATURE_HYPERVISOR)) {
266 pr_cont("PMU not available due to virtualization, using software events only.\n");
268 pr_cont("Broken PMU hardware detected, using software events only.\n");
269 pr_err("Failed to access perfctr msr (MSR %x is %Lx)\n",
276 static void hw_perf_event_destroy(struct perf_event *event)
278 x86_release_hardware();
279 atomic_dec(&active_events);
282 void hw_perf_lbr_event_destroy(struct perf_event *event)
284 hw_perf_event_destroy(event);
286 /* undo the lbr/bts event accounting */
287 x86_del_exclusive(x86_lbr_exclusive_lbr);
290 static inline int x86_pmu_initialized(void)
292 return x86_pmu.handle_irq != NULL;
296 set_ext_hw_attr(struct hw_perf_event *hwc, struct perf_event *event)
298 struct perf_event_attr *attr = &event->attr;
299 unsigned int cache_type, cache_op, cache_result;
302 config = attr->config;
304 cache_type = (config >> 0) & 0xff;
305 if (cache_type >= PERF_COUNT_HW_CACHE_MAX)
308 cache_op = (config >> 8) & 0xff;
309 if (cache_op >= PERF_COUNT_HW_CACHE_OP_MAX)
312 cache_result = (config >> 16) & 0xff;
313 if (cache_result >= PERF_COUNT_HW_CACHE_RESULT_MAX)
316 val = hw_cache_event_ids[cache_type][cache_op][cache_result];
325 attr->config1 = hw_cache_extra_regs[cache_type][cache_op][cache_result];
326 return x86_pmu_extra_regs(val, event);
329 int x86_reserve_hardware(void)
333 if (!atomic_inc_not_zero(&pmc_refcount)) {
334 mutex_lock(&pmc_reserve_mutex);
335 if (atomic_read(&pmc_refcount) == 0) {
336 if (!reserve_pmc_hardware())
339 reserve_ds_buffers();
342 atomic_inc(&pmc_refcount);
343 mutex_unlock(&pmc_reserve_mutex);
349 void x86_release_hardware(void)
351 if (atomic_dec_and_mutex_lock(&pmc_refcount, &pmc_reserve_mutex)) {
352 release_pmc_hardware();
353 release_ds_buffers();
354 mutex_unlock(&pmc_reserve_mutex);
359 * Check if we can create event of a certain type (that no conflicting events
362 int x86_add_exclusive(unsigned int what)
366 if (x86_pmu.lbr_pt_coexist)
369 if (!atomic_inc_not_zero(&x86_pmu.lbr_exclusive[what])) {
370 mutex_lock(&pmc_reserve_mutex);
371 for (i = 0; i < ARRAY_SIZE(x86_pmu.lbr_exclusive); i++) {
372 if (i != what && atomic_read(&x86_pmu.lbr_exclusive[i]))
375 atomic_inc(&x86_pmu.lbr_exclusive[what]);
376 mutex_unlock(&pmc_reserve_mutex);
379 atomic_inc(&active_events);
383 mutex_unlock(&pmc_reserve_mutex);
387 void x86_del_exclusive(unsigned int what)
389 if (x86_pmu.lbr_pt_coexist)
392 atomic_dec(&x86_pmu.lbr_exclusive[what]);
393 atomic_dec(&active_events);
396 int x86_setup_perfctr(struct perf_event *event)
398 struct perf_event_attr *attr = &event->attr;
399 struct hw_perf_event *hwc = &event->hw;
402 if (!is_sampling_event(event)) {
403 hwc->sample_period = x86_pmu.max_period;
404 hwc->last_period = hwc->sample_period;
405 local64_set(&hwc->period_left, hwc->sample_period);
408 if (attr->type == PERF_TYPE_RAW)
409 return x86_pmu_extra_regs(event->attr.config, event);
411 if (attr->type == PERF_TYPE_HW_CACHE)
412 return set_ext_hw_attr(hwc, event);
414 if (attr->config >= x86_pmu.max_events)
420 config = x86_pmu.event_map(attr->config);
431 if (attr->config == PERF_COUNT_HW_BRANCH_INSTRUCTIONS &&
432 !attr->freq && hwc->sample_period == 1) {
433 /* BTS is not supported by this architecture. */
434 if (!x86_pmu.bts_active)
437 /* BTS is currently only allowed for user-mode. */
438 if (!attr->exclude_kernel)
441 /* disallow bts if conflicting events are present */
442 if (x86_add_exclusive(x86_lbr_exclusive_lbr))
445 event->destroy = hw_perf_lbr_event_destroy;
448 hwc->config |= config;
454 * check that branch_sample_type is compatible with
455 * settings needed for precise_ip > 1 which implies
456 * using the LBR to capture ALL taken branches at the
457 * priv levels of the measurement
459 static inline int precise_br_compat(struct perf_event *event)
461 u64 m = event->attr.branch_sample_type;
464 /* must capture all branches */
465 if (!(m & PERF_SAMPLE_BRANCH_ANY))
468 m &= PERF_SAMPLE_BRANCH_KERNEL | PERF_SAMPLE_BRANCH_USER;
470 if (!event->attr.exclude_user)
471 b |= PERF_SAMPLE_BRANCH_USER;
473 if (!event->attr.exclude_kernel)
474 b |= PERF_SAMPLE_BRANCH_KERNEL;
477 * ignore PERF_SAMPLE_BRANCH_HV, not supported on x86
483 int x86_pmu_hw_config(struct perf_event *event)
485 if (event->attr.precise_ip) {
488 /* Support for constant skid */
489 if (x86_pmu.pebs_active && !x86_pmu.pebs_broken) {
492 /* Support for IP fixup */
493 if (x86_pmu.lbr_nr || x86_pmu.intel_cap.pebs_format >= 2)
496 if (x86_pmu.pebs_prec_dist)
500 if (event->attr.precise_ip > precise)
504 * check that PEBS LBR correction does not conflict with
505 * whatever the user is asking with attr->branch_sample_type
507 if (event->attr.precise_ip > 1 && x86_pmu.intel_cap.pebs_format < 2) {
508 u64 *br_type = &event->attr.branch_sample_type;
510 if (has_branch_stack(event)) {
511 if (!precise_br_compat(event))
514 /* branch_sample_type is compatible */
518 * user did not specify branch_sample_type
520 * For PEBS fixups, we capture all
521 * the branches at the priv level of the
524 *br_type = PERF_SAMPLE_BRANCH_ANY;
526 if (!event->attr.exclude_user)
527 *br_type |= PERF_SAMPLE_BRANCH_USER;
529 if (!event->attr.exclude_kernel)
530 *br_type |= PERF_SAMPLE_BRANCH_KERNEL;
534 if (event->attr.branch_sample_type & PERF_SAMPLE_BRANCH_CALL_STACK)
535 event->attach_state |= PERF_ATTACH_TASK_DATA;
539 * (keep 'enabled' bit clear for now)
541 event->hw.config = ARCH_PERFMON_EVENTSEL_INT;
544 * Count user and OS events unless requested not to
546 if (!event->attr.exclude_user)
547 event->hw.config |= ARCH_PERFMON_EVENTSEL_USR;
548 if (!event->attr.exclude_kernel)
549 event->hw.config |= ARCH_PERFMON_EVENTSEL_OS;
551 if (event->attr.type == PERF_TYPE_RAW)
552 event->hw.config |= event->attr.config & X86_RAW_EVENT_MASK;
554 if (event->attr.sample_period && x86_pmu.limit_period) {
555 if (x86_pmu.limit_period(event, event->attr.sample_period) >
556 event->attr.sample_period)
560 return x86_setup_perfctr(event);
564 * Setup the hardware configuration for a given attr_type
566 static int __x86_pmu_event_init(struct perf_event *event)
570 if (!x86_pmu_initialized())
573 err = x86_reserve_hardware();
577 atomic_inc(&active_events);
578 event->destroy = hw_perf_event_destroy;
581 event->hw.last_cpu = -1;
582 event->hw.last_tag = ~0ULL;
585 event->hw.extra_reg.idx = EXTRA_REG_NONE;
586 event->hw.branch_reg.idx = EXTRA_REG_NONE;
588 return x86_pmu.hw_config(event);
591 void x86_pmu_disable_all(void)
593 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
596 for (idx = 0; idx < x86_pmu.num_counters; idx++) {
599 if (!test_bit(idx, cpuc->active_mask))
601 rdmsrl(x86_pmu_config_addr(idx), val);
602 if (!(val & ARCH_PERFMON_EVENTSEL_ENABLE))
604 val &= ~ARCH_PERFMON_EVENTSEL_ENABLE;
605 wrmsrl(x86_pmu_config_addr(idx), val);
610 * There may be PMI landing after enabled=0. The PMI hitting could be before or
613 * If PMI hits before disable_all, the PMU will be disabled in the NMI handler.
614 * It will not be re-enabled in the NMI handler again, because enabled=0. After
615 * handling the NMI, disable_all will be called, which will not change the
616 * state either. If PMI hits after disable_all, the PMU is already disabled
617 * before entering NMI handler. The NMI handler will not change the state
620 * So either situation is harmless.
622 static void x86_pmu_disable(struct pmu *pmu)
624 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
626 if (!x86_pmu_initialized())
636 x86_pmu.disable_all();
639 void x86_pmu_enable_all(int added)
641 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
644 for (idx = 0; idx < x86_pmu.num_counters; idx++) {
645 struct hw_perf_event *hwc = &cpuc->events[idx]->hw;
647 if (!test_bit(idx, cpuc->active_mask))
650 __x86_pmu_enable_event(hwc, ARCH_PERFMON_EVENTSEL_ENABLE);
654 static struct pmu pmu;
656 static inline int is_x86_event(struct perf_event *event)
658 return event->pmu == &pmu;
662 * Event scheduler state:
664 * Assign events iterating over all events and counters, beginning
665 * with events with least weights first. Keep the current iterator
666 * state in struct sched_state.
670 int event; /* event index */
671 int counter; /* counter index */
672 int unassigned; /* number of events to be assigned left */
673 int nr_gp; /* number of GP counters used */
674 unsigned long used[BITS_TO_LONGS(X86_PMC_IDX_MAX)];
677 /* Total max is X86_PMC_IDX_MAX, but we are O(n!) limited */
678 #define SCHED_STATES_MAX 2
685 struct event_constraint **constraints;
686 struct sched_state state;
687 struct sched_state saved[SCHED_STATES_MAX];
691 * Initialize interator that runs through all events and counters.
693 static void perf_sched_init(struct perf_sched *sched, struct event_constraint **constraints,
694 int num, int wmin, int wmax, int gpmax)
698 memset(sched, 0, sizeof(*sched));
699 sched->max_events = num;
700 sched->max_weight = wmax;
701 sched->max_gp = gpmax;
702 sched->constraints = constraints;
704 for (idx = 0; idx < num; idx++) {
705 if (constraints[idx]->weight == wmin)
709 sched->state.event = idx; /* start with min weight */
710 sched->state.weight = wmin;
711 sched->state.unassigned = num;
714 static void perf_sched_save_state(struct perf_sched *sched)
716 if (WARN_ON_ONCE(sched->saved_states >= SCHED_STATES_MAX))
719 sched->saved[sched->saved_states] = sched->state;
720 sched->saved_states++;
723 static bool perf_sched_restore_state(struct perf_sched *sched)
725 if (!sched->saved_states)
728 sched->saved_states--;
729 sched->state = sched->saved[sched->saved_states];
731 /* continue with next counter: */
732 clear_bit(sched->state.counter++, sched->state.used);
738 * Select a counter for the current event to schedule. Return true on
741 static bool __perf_sched_find_counter(struct perf_sched *sched)
743 struct event_constraint *c;
746 if (!sched->state.unassigned)
749 if (sched->state.event >= sched->max_events)
752 c = sched->constraints[sched->state.event];
753 /* Prefer fixed purpose counters */
754 if (c->idxmsk64 & (~0ULL << INTEL_PMC_IDX_FIXED)) {
755 idx = INTEL_PMC_IDX_FIXED;
756 for_each_set_bit_from(idx, c->idxmsk, X86_PMC_IDX_MAX) {
757 if (!__test_and_set_bit(idx, sched->state.used))
762 /* Grab the first unused counter starting with idx */
763 idx = sched->state.counter;
764 for_each_set_bit_from(idx, c->idxmsk, INTEL_PMC_IDX_FIXED) {
765 if (!__test_and_set_bit(idx, sched->state.used)) {
766 if (sched->state.nr_gp++ >= sched->max_gp)
776 sched->state.counter = idx;
779 perf_sched_save_state(sched);
784 static bool perf_sched_find_counter(struct perf_sched *sched)
786 while (!__perf_sched_find_counter(sched)) {
787 if (!perf_sched_restore_state(sched))
795 * Go through all unassigned events and find the next one to schedule.
796 * Take events with the least weight first. Return true on success.
798 static bool perf_sched_next_event(struct perf_sched *sched)
800 struct event_constraint *c;
802 if (!sched->state.unassigned || !--sched->state.unassigned)
807 sched->state.event++;
808 if (sched->state.event >= sched->max_events) {
810 sched->state.event = 0;
811 sched->state.weight++;
812 if (sched->state.weight > sched->max_weight)
815 c = sched->constraints[sched->state.event];
816 } while (c->weight != sched->state.weight);
818 sched->state.counter = 0; /* start with first counter */
824 * Assign a counter for each event.
826 int perf_assign_events(struct event_constraint **constraints, int n,
827 int wmin, int wmax, int gpmax, int *assign)
829 struct perf_sched sched;
831 perf_sched_init(&sched, constraints, n, wmin, wmax, gpmax);
834 if (!perf_sched_find_counter(&sched))
837 assign[sched.state.event] = sched.state.counter;
838 } while (perf_sched_next_event(&sched));
840 return sched.state.unassigned;
842 EXPORT_SYMBOL_GPL(perf_assign_events);
844 int x86_schedule_events(struct cpu_hw_events *cpuc, int n, int *assign)
846 struct event_constraint *c;
847 unsigned long used_mask[BITS_TO_LONGS(X86_PMC_IDX_MAX)];
848 struct perf_event *e;
849 int i, wmin, wmax, unsched = 0;
850 struct hw_perf_event *hwc;
852 bitmap_zero(used_mask, X86_PMC_IDX_MAX);
854 if (x86_pmu.start_scheduling)
855 x86_pmu.start_scheduling(cpuc);
857 for (i = 0, wmin = X86_PMC_IDX_MAX, wmax = 0; i < n; i++) {
858 cpuc->event_constraint[i] = NULL;
859 c = x86_pmu.get_event_constraints(cpuc, i, cpuc->event_list[i]);
860 cpuc->event_constraint[i] = c;
862 wmin = min(wmin, c->weight);
863 wmax = max(wmax, c->weight);
867 * fastpath, try to reuse previous register
869 for (i = 0; i < n; i++) {
870 hwc = &cpuc->event_list[i]->hw;
871 c = cpuc->event_constraint[i];
877 /* constraint still honored */
878 if (!test_bit(hwc->idx, c->idxmsk))
881 /* not already used */
882 if (test_bit(hwc->idx, used_mask))
885 __set_bit(hwc->idx, used_mask);
887 assign[i] = hwc->idx;
892 int gpmax = x86_pmu.num_counters;
895 * Do not allow scheduling of more than half the available
898 * This helps avoid counter starvation of sibling thread by
899 * ensuring at most half the counters cannot be in exclusive
900 * mode. There is no designated counters for the limits. Any
901 * N/2 counters can be used. This helps with events with
902 * specific counter constraints.
904 if (is_ht_workaround_enabled() && !cpuc->is_fake &&
905 READ_ONCE(cpuc->excl_cntrs->exclusive_present))
908 unsched = perf_assign_events(cpuc->event_constraint, n, wmin,
909 wmax, gpmax, assign);
913 * In case of success (unsched = 0), mark events as committed,
914 * so we do not put_constraint() in case new events are added
915 * and fail to be scheduled
917 * We invoke the lower level commit callback to lock the resource
919 * We do not need to do all of this in case we are called to
920 * validate an event group (assign == NULL)
922 if (!unsched && assign) {
923 for (i = 0; i < n; i++) {
924 e = cpuc->event_list[i];
925 e->hw.flags |= PERF_X86_EVENT_COMMITTED;
926 if (x86_pmu.commit_scheduling)
927 x86_pmu.commit_scheduling(cpuc, i, assign[i]);
930 for (i = 0; i < n; i++) {
931 e = cpuc->event_list[i];
933 * do not put_constraint() on comitted events,
934 * because they are good to go
936 if ((e->hw.flags & PERF_X86_EVENT_COMMITTED))
940 * release events that failed scheduling
942 if (x86_pmu.put_event_constraints)
943 x86_pmu.put_event_constraints(cpuc, e);
947 if (x86_pmu.stop_scheduling)
948 x86_pmu.stop_scheduling(cpuc);
950 return unsched ? -EINVAL : 0;
954 * dogrp: true if must collect siblings events (group)
955 * returns total number of events and error code
957 static int collect_events(struct cpu_hw_events *cpuc, struct perf_event *leader, bool dogrp)
959 struct perf_event *event;
962 max_count = x86_pmu.num_counters + x86_pmu.num_counters_fixed;
964 /* current number of events already accepted */
967 if (is_x86_event(leader)) {
970 cpuc->event_list[n] = leader;
976 list_for_each_entry(event, &leader->sibling_list, group_entry) {
977 if (!is_x86_event(event) ||
978 event->state <= PERF_EVENT_STATE_OFF)
984 cpuc->event_list[n] = event;
990 static inline void x86_assign_hw_event(struct perf_event *event,
991 struct cpu_hw_events *cpuc, int i)
993 struct hw_perf_event *hwc = &event->hw;
995 hwc->idx = cpuc->assign[i];
996 hwc->last_cpu = smp_processor_id();
997 hwc->last_tag = ++cpuc->tags[i];
999 if (hwc->idx == INTEL_PMC_IDX_FIXED_BTS) {
1000 hwc->config_base = 0;
1001 hwc->event_base = 0;
1002 } else if (hwc->idx >= INTEL_PMC_IDX_FIXED) {
1003 hwc->config_base = MSR_ARCH_PERFMON_FIXED_CTR_CTRL;
1004 hwc->event_base = MSR_ARCH_PERFMON_FIXED_CTR0 + (hwc->idx - INTEL_PMC_IDX_FIXED);
1005 hwc->event_base_rdpmc = (hwc->idx - INTEL_PMC_IDX_FIXED) | 1<<30;
1007 hwc->config_base = x86_pmu_config_addr(hwc->idx);
1008 hwc->event_base = x86_pmu_event_addr(hwc->idx);
1009 hwc->event_base_rdpmc = x86_pmu_rdpmc_index(hwc->idx);
1013 static inline int match_prev_assignment(struct hw_perf_event *hwc,
1014 struct cpu_hw_events *cpuc,
1017 return hwc->idx == cpuc->assign[i] &&
1018 hwc->last_cpu == smp_processor_id() &&
1019 hwc->last_tag == cpuc->tags[i];
1022 static void x86_pmu_start(struct perf_event *event, int flags);
1024 static void x86_pmu_enable(struct pmu *pmu)
1026 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1027 struct perf_event *event;
1028 struct hw_perf_event *hwc;
1029 int i, added = cpuc->n_added;
1031 if (!x86_pmu_initialized())
1037 if (cpuc->n_added) {
1038 int n_running = cpuc->n_events - cpuc->n_added;
1040 * apply assignment obtained either from
1041 * hw_perf_group_sched_in() or x86_pmu_enable()
1043 * step1: save events moving to new counters
1045 for (i = 0; i < n_running; i++) {
1046 event = cpuc->event_list[i];
1050 * we can avoid reprogramming counter if:
1051 * - assigned same counter as last time
1052 * - running on same CPU as last time
1053 * - no other event has used the counter since
1055 if (hwc->idx == -1 ||
1056 match_prev_assignment(hwc, cpuc, i))
1060 * Ensure we don't accidentally enable a stopped
1061 * counter simply because we rescheduled.
1063 if (hwc->state & PERF_HES_STOPPED)
1064 hwc->state |= PERF_HES_ARCH;
1066 x86_pmu_stop(event, PERF_EF_UPDATE);
1070 * step2: reprogram moved events into new counters
1072 for (i = 0; i < cpuc->n_events; i++) {
1073 event = cpuc->event_list[i];
1076 if (!match_prev_assignment(hwc, cpuc, i))
1077 x86_assign_hw_event(event, cpuc, i);
1078 else if (i < n_running)
1081 if (hwc->state & PERF_HES_ARCH)
1084 x86_pmu_start(event, PERF_EF_RELOAD);
1087 perf_events_lapic_init();
1093 x86_pmu.enable_all(added);
1096 static DEFINE_PER_CPU(u64 [X86_PMC_IDX_MAX], pmc_prev_left);
1099 * Set the next IRQ period, based on the hwc->period_left value.
1100 * To be called with the event disabled in hw:
1102 int x86_perf_event_set_period(struct perf_event *event)
1104 struct hw_perf_event *hwc = &event->hw;
1105 s64 left = local64_read(&hwc->period_left);
1106 s64 period = hwc->sample_period;
1107 int ret = 0, idx = hwc->idx;
1109 if (idx == INTEL_PMC_IDX_FIXED_BTS)
1113 * If we are way outside a reasonable range then just skip forward:
1115 if (unlikely(left <= -period)) {
1117 local64_set(&hwc->period_left, left);
1118 hwc->last_period = period;
1122 if (unlikely(left <= 0)) {
1124 local64_set(&hwc->period_left, left);
1125 hwc->last_period = period;
1129 * Quirk: certain CPUs dont like it if just 1 hw_event is left:
1131 if (unlikely(left < 2))
1134 if (left > x86_pmu.max_period)
1135 left = x86_pmu.max_period;
1137 if (x86_pmu.limit_period)
1138 left = x86_pmu.limit_period(event, left);
1140 per_cpu(pmc_prev_left[idx], smp_processor_id()) = left;
1142 if (!(hwc->flags & PERF_X86_EVENT_AUTO_RELOAD) ||
1143 local64_read(&hwc->prev_count) != (u64)-left) {
1145 * The hw event starts counting from this event offset,
1146 * mark it to be able to extra future deltas:
1148 local64_set(&hwc->prev_count, (u64)-left);
1150 wrmsrl(hwc->event_base, (u64)(-left) & x86_pmu.cntval_mask);
1154 * Due to erratum on certan cpu we need
1155 * a second write to be sure the register
1156 * is updated properly
1158 if (x86_pmu.perfctr_second_write) {
1159 wrmsrl(hwc->event_base,
1160 (u64)(-left) & x86_pmu.cntval_mask);
1163 perf_event_update_userpage(event);
1168 void x86_pmu_enable_event(struct perf_event *event)
1170 if (__this_cpu_read(cpu_hw_events.enabled))
1171 __x86_pmu_enable_event(&event->hw,
1172 ARCH_PERFMON_EVENTSEL_ENABLE);
1176 * Add a single event to the PMU.
1178 * The event is added to the group of enabled events
1179 * but only if it can be scehduled with existing events.
1181 static int x86_pmu_add(struct perf_event *event, int flags)
1183 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1184 struct hw_perf_event *hwc;
1185 int assign[X86_PMC_IDX_MAX];
1190 n0 = cpuc->n_events;
1191 ret = n = collect_events(cpuc, event, false);
1195 hwc->state = PERF_HES_UPTODATE | PERF_HES_STOPPED;
1196 if (!(flags & PERF_EF_START))
1197 hwc->state |= PERF_HES_ARCH;
1200 * If group events scheduling transaction was started,
1201 * skip the schedulability test here, it will be performed
1202 * at commit time (->commit_txn) as a whole.
1204 if (cpuc->txn_flags & PERF_PMU_TXN_ADD)
1207 ret = x86_pmu.schedule_events(cpuc, n, assign);
1211 * copy new assignment, now we know it is possible
1212 * will be used by hw_perf_enable()
1214 memcpy(cpuc->assign, assign, n*sizeof(int));
1218 * Commit the collect_events() state. See x86_pmu_del() and
1222 cpuc->n_added += n - n0;
1223 cpuc->n_txn += n - n0;
1230 static void x86_pmu_start(struct perf_event *event, int flags)
1232 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1233 int idx = event->hw.idx;
1235 if (WARN_ON_ONCE(!(event->hw.state & PERF_HES_STOPPED)))
1238 if (WARN_ON_ONCE(idx == -1))
1241 if (flags & PERF_EF_RELOAD) {
1242 WARN_ON_ONCE(!(event->hw.state & PERF_HES_UPTODATE));
1243 x86_perf_event_set_period(event);
1246 event->hw.state = 0;
1248 cpuc->events[idx] = event;
1249 __set_bit(idx, cpuc->active_mask);
1250 __set_bit(idx, cpuc->running);
1251 x86_pmu.enable(event);
1252 perf_event_update_userpage(event);
1255 void perf_event_print_debug(void)
1257 u64 ctrl, status, overflow, pmc_ctrl, pmc_count, prev_left, fixed;
1259 struct cpu_hw_events *cpuc;
1260 unsigned long flags;
1263 if (!x86_pmu.num_counters)
1266 local_irq_save(flags);
1268 cpu = smp_processor_id();
1269 cpuc = &per_cpu(cpu_hw_events, cpu);
1271 if (x86_pmu.version >= 2) {
1272 rdmsrl(MSR_CORE_PERF_GLOBAL_CTRL, ctrl);
1273 rdmsrl(MSR_CORE_PERF_GLOBAL_STATUS, status);
1274 rdmsrl(MSR_CORE_PERF_GLOBAL_OVF_CTRL, overflow);
1275 rdmsrl(MSR_ARCH_PERFMON_FIXED_CTR_CTRL, fixed);
1278 pr_info("CPU#%d: ctrl: %016llx\n", cpu, ctrl);
1279 pr_info("CPU#%d: status: %016llx\n", cpu, status);
1280 pr_info("CPU#%d: overflow: %016llx\n", cpu, overflow);
1281 pr_info("CPU#%d: fixed: %016llx\n", cpu, fixed);
1282 if (x86_pmu.pebs_constraints) {
1283 rdmsrl(MSR_IA32_PEBS_ENABLE, pebs);
1284 pr_info("CPU#%d: pebs: %016llx\n", cpu, pebs);
1286 if (x86_pmu.lbr_nr) {
1287 rdmsrl(MSR_IA32_DEBUGCTLMSR, debugctl);
1288 pr_info("CPU#%d: debugctl: %016llx\n", cpu, debugctl);
1291 pr_info("CPU#%d: active: %016llx\n", cpu, *(u64 *)cpuc->active_mask);
1293 for (idx = 0; idx < x86_pmu.num_counters; idx++) {
1294 rdmsrl(x86_pmu_config_addr(idx), pmc_ctrl);
1295 rdmsrl(x86_pmu_event_addr(idx), pmc_count);
1297 prev_left = per_cpu(pmc_prev_left[idx], cpu);
1299 pr_info("CPU#%d: gen-PMC%d ctrl: %016llx\n",
1300 cpu, idx, pmc_ctrl);
1301 pr_info("CPU#%d: gen-PMC%d count: %016llx\n",
1302 cpu, idx, pmc_count);
1303 pr_info("CPU#%d: gen-PMC%d left: %016llx\n",
1304 cpu, idx, prev_left);
1306 for (idx = 0; idx < x86_pmu.num_counters_fixed; idx++) {
1307 rdmsrl(MSR_ARCH_PERFMON_FIXED_CTR0 + idx, pmc_count);
1309 pr_info("CPU#%d: fixed-PMC%d count: %016llx\n",
1310 cpu, idx, pmc_count);
1312 local_irq_restore(flags);
1315 void x86_pmu_stop(struct perf_event *event, int flags)
1317 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1318 struct hw_perf_event *hwc = &event->hw;
1320 if (__test_and_clear_bit(hwc->idx, cpuc->active_mask)) {
1321 x86_pmu.disable(event);
1322 cpuc->events[hwc->idx] = NULL;
1323 WARN_ON_ONCE(hwc->state & PERF_HES_STOPPED);
1324 hwc->state |= PERF_HES_STOPPED;
1327 if ((flags & PERF_EF_UPDATE) && !(hwc->state & PERF_HES_UPTODATE)) {
1329 * Drain the remaining delta count out of a event
1330 * that we are disabling:
1332 x86_perf_event_update(event);
1333 hwc->state |= PERF_HES_UPTODATE;
1337 static void x86_pmu_del(struct perf_event *event, int flags)
1339 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1343 * event is descheduled
1345 event->hw.flags &= ~PERF_X86_EVENT_COMMITTED;
1348 * If we're called during a txn, we don't need to do anything.
1349 * The events never got scheduled and ->cancel_txn will truncate
1352 * XXX assumes any ->del() called during a TXN will only be on
1353 * an event added during that same TXN.
1355 if (cpuc->txn_flags & PERF_PMU_TXN_ADD)
1359 * Not a TXN, therefore cleanup properly.
1361 x86_pmu_stop(event, PERF_EF_UPDATE);
1363 for (i = 0; i < cpuc->n_events; i++) {
1364 if (event == cpuc->event_list[i])
1368 if (WARN_ON_ONCE(i == cpuc->n_events)) /* called ->del() without ->add() ? */
1371 /* If we have a newly added event; make sure to decrease n_added. */
1372 if (i >= cpuc->n_events - cpuc->n_added)
1375 if (x86_pmu.put_event_constraints)
1376 x86_pmu.put_event_constraints(cpuc, event);
1378 /* Delete the array entry. */
1379 while (++i < cpuc->n_events) {
1380 cpuc->event_list[i-1] = cpuc->event_list[i];
1381 cpuc->event_constraint[i-1] = cpuc->event_constraint[i];
1385 perf_event_update_userpage(event);
1388 int x86_pmu_handle_irq(struct pt_regs *regs)
1390 struct perf_sample_data data;
1391 struct cpu_hw_events *cpuc;
1392 struct perf_event *event;
1393 int idx, handled = 0;
1396 cpuc = this_cpu_ptr(&cpu_hw_events);
1399 * Some chipsets need to unmask the LVTPC in a particular spot
1400 * inside the nmi handler. As a result, the unmasking was pushed
1401 * into all the nmi handlers.
1403 * This generic handler doesn't seem to have any issues where the
1404 * unmasking occurs so it was left at the top.
1406 apic_write(APIC_LVTPC, APIC_DM_NMI);
1408 for (idx = 0; idx < x86_pmu.num_counters; idx++) {
1409 if (!test_bit(idx, cpuc->active_mask)) {
1411 * Though we deactivated the counter some cpus
1412 * might still deliver spurious interrupts still
1413 * in flight. Catch them:
1415 if (__test_and_clear_bit(idx, cpuc->running))
1420 event = cpuc->events[idx];
1422 val = x86_perf_event_update(event);
1423 if (val & (1ULL << (x86_pmu.cntval_bits - 1)))
1430 perf_sample_data_init(&data, 0, event->hw.last_period);
1432 if (!x86_perf_event_set_period(event))
1435 if (perf_event_overflow(event, &data, regs))
1436 x86_pmu_stop(event, 0);
1440 inc_irq_stat(apic_perf_irqs);
1445 void perf_events_lapic_init(void)
1447 if (!x86_pmu.apic || !x86_pmu_initialized())
1451 * Always use NMI for PMU
1453 apic_write(APIC_LVTPC, APIC_DM_NMI);
1457 perf_event_nmi_handler(unsigned int cmd, struct pt_regs *regs)
1464 * All PMUs/events that share this PMI handler should make sure to
1465 * increment active_events for their events.
1467 if (!atomic_read(&active_events))
1470 start_clock = sched_clock();
1471 ret = x86_pmu.handle_irq(regs);
1472 finish_clock = sched_clock();
1474 perf_sample_event_took(finish_clock - start_clock);
1478 NOKPROBE_SYMBOL(perf_event_nmi_handler);
1480 struct event_constraint emptyconstraint;
1481 struct event_constraint unconstrained;
1483 static int x86_pmu_prepare_cpu(unsigned int cpu)
1485 struct cpu_hw_events *cpuc = &per_cpu(cpu_hw_events, cpu);
1488 for (i = 0 ; i < X86_PERF_KFREE_MAX; i++)
1489 cpuc->kfree_on_online[i] = NULL;
1490 if (x86_pmu.cpu_prepare)
1491 return x86_pmu.cpu_prepare(cpu);
1495 static int x86_pmu_dead_cpu(unsigned int cpu)
1497 if (x86_pmu.cpu_dead)
1498 x86_pmu.cpu_dead(cpu);
1502 static int x86_pmu_online_cpu(unsigned int cpu)
1504 struct cpu_hw_events *cpuc = &per_cpu(cpu_hw_events, cpu);
1507 for (i = 0 ; i < X86_PERF_KFREE_MAX; i++) {
1508 kfree(cpuc->kfree_on_online[i]);
1509 cpuc->kfree_on_online[i] = NULL;
1514 static int x86_pmu_starting_cpu(unsigned int cpu)
1516 if (x86_pmu.cpu_starting)
1517 x86_pmu.cpu_starting(cpu);
1521 static int x86_pmu_dying_cpu(unsigned int cpu)
1523 if (x86_pmu.cpu_dying)
1524 x86_pmu.cpu_dying(cpu);
1528 static void __init pmu_check_apic(void)
1530 if (boot_cpu_has(X86_FEATURE_APIC))
1534 pr_info("no APIC, boot with the \"lapic\" boot parameter to force-enable it.\n");
1535 pr_info("no hardware sampling interrupt available.\n");
1538 * If we have a PMU initialized but no APIC
1539 * interrupts, we cannot sample hardware
1540 * events (user-space has to fall back and
1541 * sample via a hrtimer based software event):
1543 pmu.capabilities |= PERF_PMU_CAP_NO_INTERRUPT;
1547 static struct attribute_group x86_pmu_format_group = {
1553 * Remove all undefined events (x86_pmu.event_map(id) == 0)
1554 * out of events_attr attributes.
1556 static void __init filter_events(struct attribute **attrs)
1558 struct device_attribute *d;
1559 struct perf_pmu_events_attr *pmu_attr;
1563 for (i = 0; attrs[i]; i++) {
1564 d = (struct device_attribute *)attrs[i];
1565 pmu_attr = container_of(d, struct perf_pmu_events_attr, attr);
1567 if (pmu_attr->event_str)
1569 if (x86_pmu.event_map(i + offset))
1572 for (j = i; attrs[j]; j++)
1573 attrs[j] = attrs[j + 1];
1575 /* Check the shifted attr. */
1579 * event_map() is index based, the attrs array is organized
1580 * by increasing event index. If we shift the events, then
1581 * we need to compensate for the event_map(), otherwise
1582 * we are looking up the wrong event in the map
1588 /* Merge two pointer arrays */
1589 __init struct attribute **merge_attr(struct attribute **a, struct attribute **b)
1591 struct attribute **new;
1594 for (j = 0; a[j]; j++)
1596 for (i = 0; b[i]; i++)
1600 new = kmalloc(sizeof(struct attribute *) * j, GFP_KERNEL);
1605 for (i = 0; a[i]; i++)
1607 for (i = 0; b[i]; i++)
1614 ssize_t events_sysfs_show(struct device *dev, struct device_attribute *attr, char *page)
1616 struct perf_pmu_events_attr *pmu_attr = \
1617 container_of(attr, struct perf_pmu_events_attr, attr);
1618 u64 config = x86_pmu.event_map(pmu_attr->id);
1620 /* string trumps id */
1621 if (pmu_attr->event_str)
1622 return sprintf(page, "%s", pmu_attr->event_str);
1624 return x86_pmu.events_sysfs_show(page, config);
1626 EXPORT_SYMBOL_GPL(events_sysfs_show);
1628 ssize_t events_ht_sysfs_show(struct device *dev, struct device_attribute *attr,
1631 struct perf_pmu_events_ht_attr *pmu_attr =
1632 container_of(attr, struct perf_pmu_events_ht_attr, attr);
1635 * Report conditional events depending on Hyper-Threading.
1637 * This is overly conservative as usually the HT special
1638 * handling is not needed if the other CPU thread is idle.
1640 * Note this does not (and cannot) handle the case when thread
1641 * siblings are invisible, for example with virtualization
1642 * if they are owned by some other guest. The user tool
1643 * has to re-read when a thread sibling gets onlined later.
1645 return sprintf(page, "%s",
1646 topology_max_smt_threads() > 1 ?
1647 pmu_attr->event_str_ht :
1648 pmu_attr->event_str_noht);
1651 EVENT_ATTR(cpu-cycles, CPU_CYCLES );
1652 EVENT_ATTR(instructions, INSTRUCTIONS );
1653 EVENT_ATTR(cache-references, CACHE_REFERENCES );
1654 EVENT_ATTR(cache-misses, CACHE_MISSES );
1655 EVENT_ATTR(branch-instructions, BRANCH_INSTRUCTIONS );
1656 EVENT_ATTR(branch-misses, BRANCH_MISSES );
1657 EVENT_ATTR(bus-cycles, BUS_CYCLES );
1658 EVENT_ATTR(stalled-cycles-frontend, STALLED_CYCLES_FRONTEND );
1659 EVENT_ATTR(stalled-cycles-backend, STALLED_CYCLES_BACKEND );
1660 EVENT_ATTR(ref-cycles, REF_CPU_CYCLES );
1662 static struct attribute *empty_attrs;
1664 static struct attribute *events_attr[] = {
1665 EVENT_PTR(CPU_CYCLES),
1666 EVENT_PTR(INSTRUCTIONS),
1667 EVENT_PTR(CACHE_REFERENCES),
1668 EVENT_PTR(CACHE_MISSES),
1669 EVENT_PTR(BRANCH_INSTRUCTIONS),
1670 EVENT_PTR(BRANCH_MISSES),
1671 EVENT_PTR(BUS_CYCLES),
1672 EVENT_PTR(STALLED_CYCLES_FRONTEND),
1673 EVENT_PTR(STALLED_CYCLES_BACKEND),
1674 EVENT_PTR(REF_CPU_CYCLES),
1678 static struct attribute_group x86_pmu_events_group = {
1680 .attrs = events_attr,
1683 ssize_t x86_event_sysfs_show(char *page, u64 config, u64 event)
1685 u64 umask = (config & ARCH_PERFMON_EVENTSEL_UMASK) >> 8;
1686 u64 cmask = (config & ARCH_PERFMON_EVENTSEL_CMASK) >> 24;
1687 bool edge = (config & ARCH_PERFMON_EVENTSEL_EDGE);
1688 bool pc = (config & ARCH_PERFMON_EVENTSEL_PIN_CONTROL);
1689 bool any = (config & ARCH_PERFMON_EVENTSEL_ANY);
1690 bool inv = (config & ARCH_PERFMON_EVENTSEL_INV);
1694 * We have whole page size to spend and just little data
1695 * to write, so we can safely use sprintf.
1697 ret = sprintf(page, "event=0x%02llx", event);
1700 ret += sprintf(page + ret, ",umask=0x%02llx", umask);
1703 ret += sprintf(page + ret, ",edge");
1706 ret += sprintf(page + ret, ",pc");
1709 ret += sprintf(page + ret, ",any");
1712 ret += sprintf(page + ret, ",inv");
1715 ret += sprintf(page + ret, ",cmask=0x%02llx", cmask);
1717 ret += sprintf(page + ret, "\n");
1722 static int __init init_hw_perf_events(void)
1724 struct x86_pmu_quirk *quirk;
1727 pr_info("Performance Events: ");
1729 switch (boot_cpu_data.x86_vendor) {
1730 case X86_VENDOR_INTEL:
1731 err = intel_pmu_init();
1733 case X86_VENDOR_AMD:
1734 err = amd_pmu_init();
1740 pr_cont("no PMU driver, software events only.\n");
1746 /* sanity check that the hardware exists or is emulated */
1747 if (!check_hw_exists())
1750 pr_cont("%s PMU driver.\n", x86_pmu.name);
1752 x86_pmu.attr_rdpmc = 1; /* enable userspace RDPMC usage by default */
1754 for (quirk = x86_pmu.quirks; quirk; quirk = quirk->next)
1757 if (!x86_pmu.intel_ctrl)
1758 x86_pmu.intel_ctrl = (1 << x86_pmu.num_counters) - 1;
1760 perf_events_lapic_init();
1761 register_nmi_handler(NMI_LOCAL, perf_event_nmi_handler, 0, "PMI");
1763 unconstrained = (struct event_constraint)
1764 __EVENT_CONSTRAINT(0, (1ULL << x86_pmu.num_counters) - 1,
1765 0, x86_pmu.num_counters, 0, 0);
1767 x86_pmu_format_group.attrs = x86_pmu.format_attrs;
1769 if (x86_pmu.event_attrs)
1770 x86_pmu_events_group.attrs = x86_pmu.event_attrs;
1772 if (!x86_pmu.events_sysfs_show)
1773 x86_pmu_events_group.attrs = &empty_attrs;
1775 filter_events(x86_pmu_events_group.attrs);
1777 if (x86_pmu.cpu_events) {
1778 struct attribute **tmp;
1780 tmp = merge_attr(x86_pmu_events_group.attrs, x86_pmu.cpu_events);
1782 x86_pmu_events_group.attrs = tmp;
1785 pr_info("... version: %d\n", x86_pmu.version);
1786 pr_info("... bit width: %d\n", x86_pmu.cntval_bits);
1787 pr_info("... generic registers: %d\n", x86_pmu.num_counters);
1788 pr_info("... value mask: %016Lx\n", x86_pmu.cntval_mask);
1789 pr_info("... max period: %016Lx\n", x86_pmu.max_period);
1790 pr_info("... fixed-purpose events: %d\n", x86_pmu.num_counters_fixed);
1791 pr_info("... event mask: %016Lx\n", x86_pmu.intel_ctrl);
1794 * Install callbacks. Core will call them for each online
1797 err = cpuhp_setup_state(CPUHP_PERF_X86_PREPARE, "PERF_X86_PREPARE",
1798 x86_pmu_prepare_cpu, x86_pmu_dead_cpu);
1802 err = cpuhp_setup_state(CPUHP_AP_PERF_X86_STARTING,
1803 "AP_PERF_X86_STARTING", x86_pmu_starting_cpu,
1808 err = cpuhp_setup_state(CPUHP_AP_PERF_X86_ONLINE, "AP_PERF_X86_ONLINE",
1809 x86_pmu_online_cpu, NULL);
1813 err = perf_pmu_register(&pmu, "cpu", PERF_TYPE_RAW);
1820 cpuhp_remove_state(CPUHP_AP_PERF_X86_ONLINE);
1822 cpuhp_remove_state(CPUHP_AP_PERF_X86_STARTING);
1824 cpuhp_remove_state(CPUHP_PERF_X86_PREPARE);
1827 early_initcall(init_hw_perf_events);
1829 static inline void x86_pmu_read(struct perf_event *event)
1831 x86_perf_event_update(event);
1835 * Start group events scheduling transaction
1836 * Set the flag to make pmu::enable() not perform the
1837 * schedulability test, it will be performed at commit time
1839 * We only support PERF_PMU_TXN_ADD transactions. Save the
1840 * transaction flags but otherwise ignore non-PERF_PMU_TXN_ADD
1843 static void x86_pmu_start_txn(struct pmu *pmu, unsigned int txn_flags)
1845 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1847 WARN_ON_ONCE(cpuc->txn_flags); /* txn already in flight */
1849 cpuc->txn_flags = txn_flags;
1850 if (txn_flags & ~PERF_PMU_TXN_ADD)
1853 perf_pmu_disable(pmu);
1854 __this_cpu_write(cpu_hw_events.n_txn, 0);
1858 * Stop group events scheduling transaction
1859 * Clear the flag and pmu::enable() will perform the
1860 * schedulability test.
1862 static void x86_pmu_cancel_txn(struct pmu *pmu)
1864 unsigned int txn_flags;
1865 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1867 WARN_ON_ONCE(!cpuc->txn_flags); /* no txn in flight */
1869 txn_flags = cpuc->txn_flags;
1870 cpuc->txn_flags = 0;
1871 if (txn_flags & ~PERF_PMU_TXN_ADD)
1875 * Truncate collected array by the number of events added in this
1876 * transaction. See x86_pmu_add() and x86_pmu_*_txn().
1878 __this_cpu_sub(cpu_hw_events.n_added, __this_cpu_read(cpu_hw_events.n_txn));
1879 __this_cpu_sub(cpu_hw_events.n_events, __this_cpu_read(cpu_hw_events.n_txn));
1880 perf_pmu_enable(pmu);
1884 * Commit group events scheduling transaction
1885 * Perform the group schedulability test as a whole
1886 * Return 0 if success
1888 * Does not cancel the transaction on failure; expects the caller to do this.
1890 static int x86_pmu_commit_txn(struct pmu *pmu)
1892 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1893 int assign[X86_PMC_IDX_MAX];
1896 WARN_ON_ONCE(!cpuc->txn_flags); /* no txn in flight */
1898 if (cpuc->txn_flags & ~PERF_PMU_TXN_ADD) {
1899 cpuc->txn_flags = 0;
1905 if (!x86_pmu_initialized())
1908 ret = x86_pmu.schedule_events(cpuc, n, assign);
1913 * copy new assignment, now we know it is possible
1914 * will be used by hw_perf_enable()
1916 memcpy(cpuc->assign, assign, n*sizeof(int));
1918 cpuc->txn_flags = 0;
1919 perf_pmu_enable(pmu);
1923 * a fake_cpuc is used to validate event groups. Due to
1924 * the extra reg logic, we need to also allocate a fake
1925 * per_core and per_cpu structure. Otherwise, group events
1926 * using extra reg may conflict without the kernel being
1927 * able to catch this when the last event gets added to
1930 static void free_fake_cpuc(struct cpu_hw_events *cpuc)
1932 kfree(cpuc->shared_regs);
1936 static struct cpu_hw_events *allocate_fake_cpuc(void)
1938 struct cpu_hw_events *cpuc;
1939 int cpu = raw_smp_processor_id();
1941 cpuc = kzalloc(sizeof(*cpuc), GFP_KERNEL);
1943 return ERR_PTR(-ENOMEM);
1945 /* only needed, if we have extra_regs */
1946 if (x86_pmu.extra_regs) {
1947 cpuc->shared_regs = allocate_shared_regs(cpu);
1948 if (!cpuc->shared_regs)
1954 free_fake_cpuc(cpuc);
1955 return ERR_PTR(-ENOMEM);
1959 * validate that we can schedule this event
1961 static int validate_event(struct perf_event *event)
1963 struct cpu_hw_events *fake_cpuc;
1964 struct event_constraint *c;
1967 fake_cpuc = allocate_fake_cpuc();
1968 if (IS_ERR(fake_cpuc))
1969 return PTR_ERR(fake_cpuc);
1971 c = x86_pmu.get_event_constraints(fake_cpuc, -1, event);
1973 if (!c || !c->weight)
1976 if (x86_pmu.put_event_constraints)
1977 x86_pmu.put_event_constraints(fake_cpuc, event);
1979 free_fake_cpuc(fake_cpuc);
1985 * validate a single event group
1987 * validation include:
1988 * - check events are compatible which each other
1989 * - events do not compete for the same counter
1990 * - number of events <= number of counters
1992 * validation ensures the group can be loaded onto the
1993 * PMU if it was the only group available.
1995 static int validate_group(struct perf_event *event)
1997 struct perf_event *leader = event->group_leader;
1998 struct cpu_hw_events *fake_cpuc;
1999 int ret = -EINVAL, n;
2001 fake_cpuc = allocate_fake_cpuc();
2002 if (IS_ERR(fake_cpuc))
2003 return PTR_ERR(fake_cpuc);
2005 * the event is not yet connected with its
2006 * siblings therefore we must first collect
2007 * existing siblings, then add the new event
2008 * before we can simulate the scheduling
2010 n = collect_events(fake_cpuc, leader, true);
2014 fake_cpuc->n_events = n;
2015 n = collect_events(fake_cpuc, event, false);
2019 fake_cpuc->n_events = n;
2021 ret = x86_pmu.schedule_events(fake_cpuc, n, NULL);
2024 free_fake_cpuc(fake_cpuc);
2028 static int x86_pmu_event_init(struct perf_event *event)
2033 switch (event->attr.type) {
2035 case PERF_TYPE_HARDWARE:
2036 case PERF_TYPE_HW_CACHE:
2043 err = __x86_pmu_event_init(event);
2046 * we temporarily connect event to its pmu
2047 * such that validate_group() can classify
2048 * it as an x86 event using is_x86_event()
2053 if (event->group_leader != event)
2054 err = validate_group(event);
2056 err = validate_event(event);
2062 event->destroy(event);
2065 if (ACCESS_ONCE(x86_pmu.attr_rdpmc))
2066 event->hw.flags |= PERF_X86_EVENT_RDPMC_ALLOWED;
2071 static void refresh_pce(void *ignored)
2074 load_mm_cr4(current->mm);
2077 static void x86_pmu_event_mapped(struct perf_event *event)
2079 if (!(event->hw.flags & PERF_X86_EVENT_RDPMC_ALLOWED))
2082 if (atomic_inc_return(¤t->mm->context.perf_rdpmc_allowed) == 1)
2083 on_each_cpu_mask(mm_cpumask(current->mm), refresh_pce, NULL, 1);
2086 static void x86_pmu_event_unmapped(struct perf_event *event)
2091 if (!(event->hw.flags & PERF_X86_EVENT_RDPMC_ALLOWED))
2094 if (atomic_dec_and_test(¤t->mm->context.perf_rdpmc_allowed))
2095 on_each_cpu_mask(mm_cpumask(current->mm), refresh_pce, NULL, 1);
2098 static int x86_pmu_event_idx(struct perf_event *event)
2100 int idx = event->hw.idx;
2102 if (!(event->hw.flags & PERF_X86_EVENT_RDPMC_ALLOWED))
2105 if (x86_pmu.num_counters_fixed && idx >= INTEL_PMC_IDX_FIXED) {
2106 idx -= INTEL_PMC_IDX_FIXED;
2113 static ssize_t get_attr_rdpmc(struct device *cdev,
2114 struct device_attribute *attr,
2117 return snprintf(buf, 40, "%d\n", x86_pmu.attr_rdpmc);
2120 static ssize_t set_attr_rdpmc(struct device *cdev,
2121 struct device_attribute *attr,
2122 const char *buf, size_t count)
2127 ret = kstrtoul(buf, 0, &val);
2134 if (x86_pmu.attr_rdpmc_broken)
2137 if ((val == 2) != (x86_pmu.attr_rdpmc == 2)) {
2139 * Changing into or out of always available, aka
2140 * perf-event-bypassing mode. This path is extremely slow,
2141 * but only root can trigger it, so it's okay.
2144 static_key_slow_inc(&rdpmc_always_available);
2146 static_key_slow_dec(&rdpmc_always_available);
2147 on_each_cpu(refresh_pce, NULL, 1);
2150 x86_pmu.attr_rdpmc = val;
2155 static DEVICE_ATTR(rdpmc, S_IRUSR | S_IWUSR, get_attr_rdpmc, set_attr_rdpmc);
2157 static struct attribute *x86_pmu_attrs[] = {
2158 &dev_attr_rdpmc.attr,
2162 static struct attribute_group x86_pmu_attr_group = {
2163 .attrs = x86_pmu_attrs,
2166 static const struct attribute_group *x86_pmu_attr_groups[] = {
2167 &x86_pmu_attr_group,
2168 &x86_pmu_format_group,
2169 &x86_pmu_events_group,
2173 static void x86_pmu_sched_task(struct perf_event_context *ctx, bool sched_in)
2175 if (x86_pmu.sched_task)
2176 x86_pmu.sched_task(ctx, sched_in);
2179 void perf_check_microcode(void)
2181 if (x86_pmu.check_microcode)
2182 x86_pmu.check_microcode();
2184 EXPORT_SYMBOL_GPL(perf_check_microcode);
2186 static struct pmu pmu = {
2187 .pmu_enable = x86_pmu_enable,
2188 .pmu_disable = x86_pmu_disable,
2190 .attr_groups = x86_pmu_attr_groups,
2192 .event_init = x86_pmu_event_init,
2194 .event_mapped = x86_pmu_event_mapped,
2195 .event_unmapped = x86_pmu_event_unmapped,
2199 .start = x86_pmu_start,
2200 .stop = x86_pmu_stop,
2201 .read = x86_pmu_read,
2203 .start_txn = x86_pmu_start_txn,
2204 .cancel_txn = x86_pmu_cancel_txn,
2205 .commit_txn = x86_pmu_commit_txn,
2207 .event_idx = x86_pmu_event_idx,
2208 .sched_task = x86_pmu_sched_task,
2209 .task_ctx_size = sizeof(struct x86_perf_task_context),
2212 void arch_perf_update_userpage(struct perf_event *event,
2213 struct perf_event_mmap_page *userpg, u64 now)
2215 struct cyc2ns_data *data;
2217 userpg->cap_user_time = 0;
2218 userpg->cap_user_time_zero = 0;
2219 userpg->cap_user_rdpmc =
2220 !!(event->hw.flags & PERF_X86_EVENT_RDPMC_ALLOWED);
2221 userpg->pmc_width = x86_pmu.cntval_bits;
2223 if (!sched_clock_stable())
2226 data = cyc2ns_read_begin();
2229 * Internal timekeeping for enabled/running/stopped times
2230 * is always in the local_clock domain.
2232 userpg->cap_user_time = 1;
2233 userpg->time_mult = data->cyc2ns_mul;
2234 userpg->time_shift = data->cyc2ns_shift;
2235 userpg->time_offset = data->cyc2ns_offset - now;
2238 * cap_user_time_zero doesn't make sense when we're using a different
2239 * time base for the records.
2241 if (!event->attr.use_clockid) {
2242 userpg->cap_user_time_zero = 1;
2243 userpg->time_zero = data->cyc2ns_offset;
2246 cyc2ns_read_end(data);
2253 static int backtrace_stack(void *data, char *name)
2258 static int backtrace_address(void *data, unsigned long addr, int reliable)
2260 struct perf_callchain_entry_ctx *entry = data;
2262 return perf_callchain_store(entry, addr);
2265 static const struct stacktrace_ops backtrace_ops = {
2266 .stack = backtrace_stack,
2267 .address = backtrace_address,
2268 .walk_stack = print_context_stack_bp,
2272 perf_callchain_kernel(struct perf_callchain_entry_ctx *entry, struct pt_regs *regs)
2274 if (perf_guest_cbs && perf_guest_cbs->is_in_guest()) {
2275 /* TODO: We don't support guest os callchain now */
2279 perf_callchain_store(entry, regs->ip);
2281 dump_trace(NULL, regs, NULL, 0, &backtrace_ops, entry);
2285 valid_user_frame(const void __user *fp, unsigned long size)
2287 return (__range_not_ok(fp, size, TASK_SIZE) == 0);
2290 static unsigned long get_segment_base(unsigned int segment)
2292 struct desc_struct *desc;
2293 int idx = segment >> 3;
2295 if ((segment & SEGMENT_TI_MASK) == SEGMENT_LDT) {
2296 #ifdef CONFIG_MODIFY_LDT_SYSCALL
2297 struct ldt_struct *ldt;
2299 if (idx > LDT_ENTRIES)
2302 /* IRQs are off, so this synchronizes with smp_store_release */
2303 ldt = lockless_dereference(current->active_mm->context.ldt);
2304 if (!ldt || idx > ldt->size)
2307 desc = &ldt->entries[idx];
2312 if (idx > GDT_ENTRIES)
2315 desc = raw_cpu_ptr(gdt_page.gdt) + idx;
2318 return get_desc_base(desc);
2321 #ifdef CONFIG_IA32_EMULATION
2323 #include <asm/compat.h>
2326 perf_callchain_user32(struct pt_regs *regs, struct perf_callchain_entry_ctx *entry)
2328 /* 32-bit process in 64-bit kernel. */
2329 unsigned long ss_base, cs_base;
2330 struct stack_frame_ia32 frame;
2331 const void __user *fp;
2333 if (!test_thread_flag(TIF_IA32))
2336 cs_base = get_segment_base(regs->cs);
2337 ss_base = get_segment_base(regs->ss);
2339 fp = compat_ptr(ss_base + regs->bp);
2340 pagefault_disable();
2341 while (entry->nr < entry->max_stack) {
2342 unsigned long bytes;
2343 frame.next_frame = 0;
2344 frame.return_address = 0;
2346 if (!access_ok(VERIFY_READ, fp, 8))
2349 bytes = __copy_from_user_nmi(&frame.next_frame, fp, 4);
2352 bytes = __copy_from_user_nmi(&frame.return_address, fp+4, 4);
2356 if (!valid_user_frame(fp, sizeof(frame)))
2359 perf_callchain_store(entry, cs_base + frame.return_address);
2360 fp = compat_ptr(ss_base + frame.next_frame);
2367 perf_callchain_user32(struct pt_regs *regs, struct perf_callchain_entry_ctx *entry)
2374 perf_callchain_user(struct perf_callchain_entry_ctx *entry, struct pt_regs *regs)
2376 struct stack_frame frame;
2377 const unsigned long __user *fp;
2379 if (perf_guest_cbs && perf_guest_cbs->is_in_guest()) {
2380 /* TODO: We don't support guest os callchain now */
2385 * We don't know what to do with VM86 stacks.. ignore them for now.
2387 if (regs->flags & (X86_VM_MASK | PERF_EFLAGS_VM))
2390 fp = (unsigned long __user *)regs->bp;
2392 perf_callchain_store(entry, regs->ip);
2397 if (perf_callchain_user32(regs, entry))
2400 pagefault_disable();
2401 while (entry->nr < entry->max_stack) {
2402 unsigned long bytes;
2404 frame.next_frame = NULL;
2405 frame.return_address = 0;
2407 if (!access_ok(VERIFY_READ, fp, sizeof(*fp) * 2))
2410 bytes = __copy_from_user_nmi(&frame.next_frame, fp, sizeof(*fp));
2413 bytes = __copy_from_user_nmi(&frame.return_address, fp + 1, sizeof(*fp));
2417 if (!valid_user_frame(fp, sizeof(frame)))
2420 perf_callchain_store(entry, frame.return_address);
2421 fp = (void __user *)frame.next_frame;
2427 * Deal with code segment offsets for the various execution modes:
2429 * VM86 - the good olde 16 bit days, where the linear address is
2430 * 20 bits and we use regs->ip + 0x10 * regs->cs.
2432 * IA32 - Where we need to look at GDT/LDT segment descriptor tables
2433 * to figure out what the 32bit base address is.
2435 * X32 - has TIF_X32 set, but is running in x86_64
2437 * X86_64 - CS,DS,SS,ES are all zero based.
2439 static unsigned long code_segment_base(struct pt_regs *regs)
2442 * For IA32 we look at the GDT/LDT segment base to convert the
2443 * effective IP to a linear address.
2446 #ifdef CONFIG_X86_32
2448 * If we are in VM86 mode, add the segment offset to convert to a
2451 if (regs->flags & X86_VM_MASK)
2452 return 0x10 * regs->cs;
2454 if (user_mode(regs) && regs->cs != __USER_CS)
2455 return get_segment_base(regs->cs);
2457 if (user_mode(regs) && !user_64bit_mode(regs) &&
2458 regs->cs != __USER32_CS)
2459 return get_segment_base(regs->cs);
2464 unsigned long perf_instruction_pointer(struct pt_regs *regs)
2466 if (perf_guest_cbs && perf_guest_cbs->is_in_guest())
2467 return perf_guest_cbs->get_guest_ip();
2469 return regs->ip + code_segment_base(regs);
2472 unsigned long perf_misc_flags(struct pt_regs *regs)
2476 if (perf_guest_cbs && perf_guest_cbs->is_in_guest()) {
2477 if (perf_guest_cbs->is_user_mode())
2478 misc |= PERF_RECORD_MISC_GUEST_USER;
2480 misc |= PERF_RECORD_MISC_GUEST_KERNEL;
2482 if (user_mode(regs))
2483 misc |= PERF_RECORD_MISC_USER;
2485 misc |= PERF_RECORD_MISC_KERNEL;
2488 if (regs->flags & PERF_EFLAGS_EXACT)
2489 misc |= PERF_RECORD_MISC_EXACT_IP;
2494 void perf_get_x86_pmu_capability(struct x86_pmu_capability *cap)
2496 cap->version = x86_pmu.version;
2497 cap->num_counters_gp = x86_pmu.num_counters;
2498 cap->num_counters_fixed = x86_pmu.num_counters_fixed;
2499 cap->bit_width_gp = x86_pmu.cntval_bits;
2500 cap->bit_width_fixed = x86_pmu.cntval_bits;
2501 cap->events_mask = (unsigned int)x86_pmu.events_maskl;
2502 cap->events_mask_len = x86_pmu.events_mask_len;
2504 EXPORT_SYMBOL_GPL(perf_get_x86_pmu_capability);