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/export.h>
21 #include <linux/init.h>
22 #include <linux/kdebug.h>
23 #include <linux/sched.h>
24 #include <linux/uaccess.h>
25 #include <linux/slab.h>
26 #include <linux/cpu.h>
27 #include <linux/bitops.h>
28 #include <linux/device.h>
31 #include <asm/stacktrace.h>
34 #include <asm/alternative.h>
35 #include <asm/mmu_context.h>
36 #include <asm/tlbflush.h>
37 #include <asm/timer.h>
41 #include "perf_event.h"
43 struct x86_pmu x86_pmu __read_mostly;
45 DEFINE_PER_CPU(struct cpu_hw_events, cpu_hw_events) = {
49 struct static_key rdpmc_always_available = STATIC_KEY_INIT_FALSE;
51 u64 __read_mostly hw_cache_event_ids
52 [PERF_COUNT_HW_CACHE_MAX]
53 [PERF_COUNT_HW_CACHE_OP_MAX]
54 [PERF_COUNT_HW_CACHE_RESULT_MAX];
55 u64 __read_mostly hw_cache_extra_regs
56 [PERF_COUNT_HW_CACHE_MAX]
57 [PERF_COUNT_HW_CACHE_OP_MAX]
58 [PERF_COUNT_HW_CACHE_RESULT_MAX];
61 * Propagate event elapsed time into the generic event.
62 * Can only be executed on the CPU where the event is active.
63 * Returns the delta events processed.
65 u64 x86_perf_event_update(struct perf_event *event)
67 struct hw_perf_event *hwc = &event->hw;
68 int shift = 64 - x86_pmu.cntval_bits;
69 u64 prev_raw_count, new_raw_count;
73 if (idx == INTEL_PMC_IDX_FIXED_BTS)
77 * Careful: an NMI might modify the previous event value.
79 * Our tactic to handle this is to first atomically read and
80 * exchange a new raw count - then add that new-prev delta
81 * count to the generic event atomically:
84 prev_raw_count = local64_read(&hwc->prev_count);
85 rdpmcl(hwc->event_base_rdpmc, new_raw_count);
87 if (local64_cmpxchg(&hwc->prev_count, prev_raw_count,
88 new_raw_count) != prev_raw_count)
92 * Now we have the new raw value and have updated the prev
93 * timestamp already. We can now calculate the elapsed delta
94 * (event-)time and add that to the generic event.
96 * Careful, not all hw sign-extends above the physical width
99 delta = (new_raw_count << shift) - (prev_raw_count << shift);
102 local64_add(delta, &event->count);
103 local64_sub(delta, &hwc->period_left);
105 return new_raw_count;
109 * Find and validate any extra registers to set up.
111 static int x86_pmu_extra_regs(u64 config, struct perf_event *event)
113 struct hw_perf_event_extra *reg;
114 struct extra_reg *er;
116 reg = &event->hw.extra_reg;
118 if (!x86_pmu.extra_regs)
121 for (er = x86_pmu.extra_regs; er->msr; er++) {
122 if (er->event != (config & er->config_mask))
124 if (event->attr.config1 & ~er->valid_mask)
126 /* Check if the extra msrs can be safely accessed*/
127 if (!er->extra_msr_access)
131 reg->config = event->attr.config1;
138 static atomic_t active_events;
139 static atomic_t pmc_refcount;
140 static DEFINE_MUTEX(pmc_reserve_mutex);
142 #ifdef CONFIG_X86_LOCAL_APIC
144 static bool reserve_pmc_hardware(void)
148 for (i = 0; i < x86_pmu.num_counters; i++) {
149 if (!reserve_perfctr_nmi(x86_pmu_event_addr(i)))
153 for (i = 0; i < x86_pmu.num_counters; i++) {
154 if (!reserve_evntsel_nmi(x86_pmu_config_addr(i)))
161 for (i--; i >= 0; i--)
162 release_evntsel_nmi(x86_pmu_config_addr(i));
164 i = x86_pmu.num_counters;
167 for (i--; i >= 0; i--)
168 release_perfctr_nmi(x86_pmu_event_addr(i));
173 static void release_pmc_hardware(void)
177 for (i = 0; i < x86_pmu.num_counters; i++) {
178 release_perfctr_nmi(x86_pmu_event_addr(i));
179 release_evntsel_nmi(x86_pmu_config_addr(i));
185 static bool reserve_pmc_hardware(void) { return true; }
186 static void release_pmc_hardware(void) {}
190 static bool check_hw_exists(void)
192 u64 val, val_fail, val_new= ~0;
193 int i, reg, reg_fail, ret = 0;
198 * Check to see if the BIOS enabled any of the counters, if so
201 for (i = 0; i < x86_pmu.num_counters; i++) {
202 reg = x86_pmu_config_addr(i);
203 ret = rdmsrl_safe(reg, &val);
206 if (val & ARCH_PERFMON_EVENTSEL_ENABLE) {
215 if (x86_pmu.num_counters_fixed) {
216 reg = MSR_ARCH_PERFMON_FIXED_CTR_CTRL;
217 ret = rdmsrl_safe(reg, &val);
220 for (i = 0; i < x86_pmu.num_counters_fixed; i++) {
221 if (val & (0x03 << i*4)) {
230 * If all the counters are enabled, the below test will always
231 * fail. The tools will also become useless in this scenario.
232 * Just fail and disable the hardware counters.
235 if (reg_safe == -1) {
241 * Read the current value, change it and read it back to see if it
242 * matches, this is needed to detect certain hardware emulators
243 * (qemu/kvm) that don't trap on the MSR access and always return 0s.
245 reg = x86_pmu_event_addr(reg_safe);
246 if (rdmsrl_safe(reg, &val))
249 ret = wrmsrl_safe(reg, val);
250 ret |= rdmsrl_safe(reg, &val_new);
251 if (ret || val != val_new)
255 * We still allow the PMU driver to operate:
258 pr_cont("Broken BIOS detected, complain to your hardware vendor.\n");
259 pr_err(FW_BUG "the BIOS has corrupted hw-PMU resources (MSR %x is %Lx)\n",
266 if (boot_cpu_has(X86_FEATURE_HYPERVISOR)) {
267 pr_cont("PMU not available due to virtualization, using software events only.\n");
269 pr_cont("Broken PMU hardware detected, using software events only.\n");
270 pr_err("Failed to access perfctr msr (MSR %x is %Lx)\n",
277 static void hw_perf_event_destroy(struct perf_event *event)
279 x86_release_hardware();
280 atomic_dec(&active_events);
283 void hw_perf_lbr_event_destroy(struct perf_event *event)
285 hw_perf_event_destroy(event);
287 /* undo the lbr/bts event accounting */
288 x86_del_exclusive(x86_lbr_exclusive_lbr);
291 static inline int x86_pmu_initialized(void)
293 return x86_pmu.handle_irq != NULL;
297 set_ext_hw_attr(struct hw_perf_event *hwc, struct perf_event *event)
299 struct perf_event_attr *attr = &event->attr;
300 unsigned int cache_type, cache_op, cache_result;
303 config = attr->config;
305 cache_type = (config >> 0) & 0xff;
306 if (cache_type >= PERF_COUNT_HW_CACHE_MAX)
309 cache_op = (config >> 8) & 0xff;
310 if (cache_op >= PERF_COUNT_HW_CACHE_OP_MAX)
313 cache_result = (config >> 16) & 0xff;
314 if (cache_result >= PERF_COUNT_HW_CACHE_RESULT_MAX)
317 val = hw_cache_event_ids[cache_type][cache_op][cache_result];
326 attr->config1 = hw_cache_extra_regs[cache_type][cache_op][cache_result];
327 return x86_pmu_extra_regs(val, event);
330 int x86_reserve_hardware(void)
334 if (!atomic_inc_not_zero(&pmc_refcount)) {
335 mutex_lock(&pmc_reserve_mutex);
336 if (atomic_read(&pmc_refcount) == 0) {
337 if (!reserve_pmc_hardware())
340 reserve_ds_buffers();
343 atomic_inc(&pmc_refcount);
344 mutex_unlock(&pmc_reserve_mutex);
350 void x86_release_hardware(void)
352 if (atomic_dec_and_mutex_lock(&pmc_refcount, &pmc_reserve_mutex)) {
353 release_pmc_hardware();
354 release_ds_buffers();
355 mutex_unlock(&pmc_reserve_mutex);
360 * Check if we can create event of a certain type (that no conflicting events
363 int x86_add_exclusive(unsigned int what)
367 if (x86_pmu.lbr_pt_coexist)
370 if (!atomic_inc_not_zero(&x86_pmu.lbr_exclusive[what])) {
371 mutex_lock(&pmc_reserve_mutex);
372 for (i = 0; i < ARRAY_SIZE(x86_pmu.lbr_exclusive); i++) {
373 if (i != what && atomic_read(&x86_pmu.lbr_exclusive[i]))
376 atomic_inc(&x86_pmu.lbr_exclusive[what]);
377 mutex_unlock(&pmc_reserve_mutex);
380 atomic_inc(&active_events);
384 mutex_unlock(&pmc_reserve_mutex);
388 void x86_del_exclusive(unsigned int what)
390 if (x86_pmu.lbr_pt_coexist)
393 atomic_dec(&x86_pmu.lbr_exclusive[what]);
394 atomic_dec(&active_events);
397 int x86_setup_perfctr(struct perf_event *event)
399 struct perf_event_attr *attr = &event->attr;
400 struct hw_perf_event *hwc = &event->hw;
403 if (!is_sampling_event(event)) {
404 hwc->sample_period = x86_pmu.max_period;
405 hwc->last_period = hwc->sample_period;
406 local64_set(&hwc->period_left, hwc->sample_period);
409 if (attr->type == PERF_TYPE_RAW)
410 return x86_pmu_extra_regs(event->attr.config, event);
412 if (attr->type == PERF_TYPE_HW_CACHE)
413 return set_ext_hw_attr(hwc, event);
415 if (attr->config >= x86_pmu.max_events)
421 config = x86_pmu.event_map(attr->config);
432 if (attr->config == PERF_COUNT_HW_BRANCH_INSTRUCTIONS &&
433 !attr->freq && hwc->sample_period == 1) {
434 /* BTS is not supported by this architecture. */
435 if (!x86_pmu.bts_active)
438 /* BTS is currently only allowed for user-mode. */
439 if (!attr->exclude_kernel)
442 /* disallow bts if conflicting events are present */
443 if (x86_add_exclusive(x86_lbr_exclusive_lbr))
446 event->destroy = hw_perf_lbr_event_destroy;
449 hwc->config |= config;
455 * check that branch_sample_type is compatible with
456 * settings needed for precise_ip > 1 which implies
457 * using the LBR to capture ALL taken branches at the
458 * priv levels of the measurement
460 static inline int precise_br_compat(struct perf_event *event)
462 u64 m = event->attr.branch_sample_type;
465 /* must capture all branches */
466 if (!(m & PERF_SAMPLE_BRANCH_ANY))
469 m &= PERF_SAMPLE_BRANCH_KERNEL | PERF_SAMPLE_BRANCH_USER;
471 if (!event->attr.exclude_user)
472 b |= PERF_SAMPLE_BRANCH_USER;
474 if (!event->attr.exclude_kernel)
475 b |= PERF_SAMPLE_BRANCH_KERNEL;
478 * ignore PERF_SAMPLE_BRANCH_HV, not supported on x86
484 int x86_pmu_hw_config(struct perf_event *event)
486 if (event->attr.precise_ip) {
489 /* Support for constant skid */
490 if (x86_pmu.pebs_active && !x86_pmu.pebs_broken) {
493 /* Support for IP fixup */
494 if (x86_pmu.lbr_nr || x86_pmu.intel_cap.pebs_format >= 2)
497 if (x86_pmu.pebs_prec_dist)
501 if (event->attr.precise_ip > precise)
505 * check that PEBS LBR correction does not conflict with
506 * whatever the user is asking with attr->branch_sample_type
508 if (event->attr.precise_ip > 1 && x86_pmu.intel_cap.pebs_format < 2) {
509 u64 *br_type = &event->attr.branch_sample_type;
511 if (has_branch_stack(event)) {
512 if (!precise_br_compat(event))
515 /* branch_sample_type is compatible */
519 * user did not specify branch_sample_type
521 * For PEBS fixups, we capture all
522 * the branches at the priv level of the
525 *br_type = PERF_SAMPLE_BRANCH_ANY;
527 if (!event->attr.exclude_user)
528 *br_type |= PERF_SAMPLE_BRANCH_USER;
530 if (!event->attr.exclude_kernel)
531 *br_type |= PERF_SAMPLE_BRANCH_KERNEL;
535 if (event->attr.branch_sample_type & PERF_SAMPLE_BRANCH_CALL_STACK)
536 event->attach_state |= PERF_ATTACH_TASK_DATA;
540 * (keep 'enabled' bit clear for now)
542 event->hw.config = ARCH_PERFMON_EVENTSEL_INT;
545 * Count user and OS events unless requested not to
547 if (!event->attr.exclude_user)
548 event->hw.config |= ARCH_PERFMON_EVENTSEL_USR;
549 if (!event->attr.exclude_kernel)
550 event->hw.config |= ARCH_PERFMON_EVENTSEL_OS;
552 if (event->attr.type == PERF_TYPE_RAW)
553 event->hw.config |= event->attr.config & X86_RAW_EVENT_MASK;
555 if (event->attr.sample_period && x86_pmu.limit_period) {
556 if (x86_pmu.limit_period(event, event->attr.sample_period) >
557 event->attr.sample_period)
561 return x86_setup_perfctr(event);
565 * Setup the hardware configuration for a given attr_type
567 static int __x86_pmu_event_init(struct perf_event *event)
571 if (!x86_pmu_initialized())
574 err = x86_reserve_hardware();
578 atomic_inc(&active_events);
579 event->destroy = hw_perf_event_destroy;
582 event->hw.last_cpu = -1;
583 event->hw.last_tag = ~0ULL;
586 event->hw.extra_reg.idx = EXTRA_REG_NONE;
587 event->hw.branch_reg.idx = EXTRA_REG_NONE;
589 return x86_pmu.hw_config(event);
592 void x86_pmu_disable_all(void)
594 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
597 for (idx = 0; idx < x86_pmu.num_counters; idx++) {
600 if (!test_bit(idx, cpuc->active_mask))
602 rdmsrl(x86_pmu_config_addr(idx), val);
603 if (!(val & ARCH_PERFMON_EVENTSEL_ENABLE))
605 val &= ~ARCH_PERFMON_EVENTSEL_ENABLE;
606 wrmsrl(x86_pmu_config_addr(idx), val);
611 * There may be PMI landing after enabled=0. The PMI hitting could be before or
614 * If PMI hits before disable_all, the PMU will be disabled in the NMI handler.
615 * It will not be re-enabled in the NMI handler again, because enabled=0. After
616 * handling the NMI, disable_all will be called, which will not change the
617 * state either. If PMI hits after disable_all, the PMU is already disabled
618 * before entering NMI handler. The NMI handler will not change the state
621 * So either situation is harmless.
623 static void x86_pmu_disable(struct pmu *pmu)
625 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
627 if (!x86_pmu_initialized())
637 x86_pmu.disable_all();
640 void x86_pmu_enable_all(int added)
642 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
645 for (idx = 0; idx < x86_pmu.num_counters; idx++) {
646 struct hw_perf_event *hwc = &cpuc->events[idx]->hw;
648 if (!test_bit(idx, cpuc->active_mask))
651 __x86_pmu_enable_event(hwc, ARCH_PERFMON_EVENTSEL_ENABLE);
655 static struct pmu pmu;
657 static inline int is_x86_event(struct perf_event *event)
659 return event->pmu == &pmu;
663 * Event scheduler state:
665 * Assign events iterating over all events and counters, beginning
666 * with events with least weights first. Keep the current iterator
667 * state in struct sched_state.
671 int event; /* event index */
672 int counter; /* counter index */
673 int unassigned; /* number of events to be assigned left */
674 int nr_gp; /* number of GP counters used */
675 unsigned long used[BITS_TO_LONGS(X86_PMC_IDX_MAX)];
678 /* Total max is X86_PMC_IDX_MAX, but we are O(n!) limited */
679 #define SCHED_STATES_MAX 2
686 struct event_constraint **constraints;
687 struct sched_state state;
688 struct sched_state saved[SCHED_STATES_MAX];
692 * Initialize interator that runs through all events and counters.
694 static void perf_sched_init(struct perf_sched *sched, struct event_constraint **constraints,
695 int num, int wmin, int wmax, int gpmax)
699 memset(sched, 0, sizeof(*sched));
700 sched->max_events = num;
701 sched->max_weight = wmax;
702 sched->max_gp = gpmax;
703 sched->constraints = constraints;
705 for (idx = 0; idx < num; idx++) {
706 if (constraints[idx]->weight == wmin)
710 sched->state.event = idx; /* start with min weight */
711 sched->state.weight = wmin;
712 sched->state.unassigned = num;
715 static void perf_sched_save_state(struct perf_sched *sched)
717 if (WARN_ON_ONCE(sched->saved_states >= SCHED_STATES_MAX))
720 sched->saved[sched->saved_states] = sched->state;
721 sched->saved_states++;
724 static bool perf_sched_restore_state(struct perf_sched *sched)
726 if (!sched->saved_states)
729 sched->saved_states--;
730 sched->state = sched->saved[sched->saved_states];
732 /* continue with next counter: */
733 clear_bit(sched->state.counter++, sched->state.used);
739 * Select a counter for the current event to schedule. Return true on
742 static bool __perf_sched_find_counter(struct perf_sched *sched)
744 struct event_constraint *c;
747 if (!sched->state.unassigned)
750 if (sched->state.event >= sched->max_events)
753 c = sched->constraints[sched->state.event];
754 /* Prefer fixed purpose counters */
755 if (c->idxmsk64 & (~0ULL << INTEL_PMC_IDX_FIXED)) {
756 idx = INTEL_PMC_IDX_FIXED;
757 for_each_set_bit_from(idx, c->idxmsk, X86_PMC_IDX_MAX) {
758 if (!__test_and_set_bit(idx, sched->state.used))
763 /* Grab the first unused counter starting with idx */
764 idx = sched->state.counter;
765 for_each_set_bit_from(idx, c->idxmsk, INTEL_PMC_IDX_FIXED) {
766 if (!__test_and_set_bit(idx, sched->state.used)) {
767 if (sched->state.nr_gp++ >= sched->max_gp)
777 sched->state.counter = idx;
780 perf_sched_save_state(sched);
785 static bool perf_sched_find_counter(struct perf_sched *sched)
787 while (!__perf_sched_find_counter(sched)) {
788 if (!perf_sched_restore_state(sched))
796 * Go through all unassigned events and find the next one to schedule.
797 * Take events with the least weight first. Return true on success.
799 static bool perf_sched_next_event(struct perf_sched *sched)
801 struct event_constraint *c;
803 if (!sched->state.unassigned || !--sched->state.unassigned)
808 sched->state.event++;
809 if (sched->state.event >= sched->max_events) {
811 sched->state.event = 0;
812 sched->state.weight++;
813 if (sched->state.weight > sched->max_weight)
816 c = sched->constraints[sched->state.event];
817 } while (c->weight != sched->state.weight);
819 sched->state.counter = 0; /* start with first counter */
825 * Assign a counter for each event.
827 int perf_assign_events(struct event_constraint **constraints, int n,
828 int wmin, int wmax, int gpmax, int *assign)
830 struct perf_sched sched;
832 perf_sched_init(&sched, constraints, n, wmin, wmax, gpmax);
835 if (!perf_sched_find_counter(&sched))
838 assign[sched.state.event] = sched.state.counter;
839 } while (perf_sched_next_event(&sched));
841 return sched.state.unassigned;
843 EXPORT_SYMBOL_GPL(perf_assign_events);
845 int x86_schedule_events(struct cpu_hw_events *cpuc, int n, int *assign)
847 struct event_constraint *c;
848 unsigned long used_mask[BITS_TO_LONGS(X86_PMC_IDX_MAX)];
849 struct perf_event *e;
850 int i, wmin, wmax, unsched = 0;
851 struct hw_perf_event *hwc;
853 bitmap_zero(used_mask, X86_PMC_IDX_MAX);
855 if (x86_pmu.start_scheduling)
856 x86_pmu.start_scheduling(cpuc);
858 for (i = 0, wmin = X86_PMC_IDX_MAX, wmax = 0; i < n; i++) {
859 cpuc->event_constraint[i] = NULL;
860 c = x86_pmu.get_event_constraints(cpuc, i, cpuc->event_list[i]);
861 cpuc->event_constraint[i] = c;
863 wmin = min(wmin, c->weight);
864 wmax = max(wmax, c->weight);
868 * fastpath, try to reuse previous register
870 for (i = 0; i < n; i++) {
871 hwc = &cpuc->event_list[i]->hw;
872 c = cpuc->event_constraint[i];
878 /* constraint still honored */
879 if (!test_bit(hwc->idx, c->idxmsk))
882 /* not already used */
883 if (test_bit(hwc->idx, used_mask))
886 __set_bit(hwc->idx, used_mask);
888 assign[i] = hwc->idx;
893 int gpmax = x86_pmu.num_counters;
896 * Do not allow scheduling of more than half the available
899 * This helps avoid counter starvation of sibling thread by
900 * ensuring at most half the counters cannot be in exclusive
901 * mode. There is no designated counters for the limits. Any
902 * N/2 counters can be used. This helps with events with
903 * specific counter constraints.
905 if (is_ht_workaround_enabled() && !cpuc->is_fake &&
906 READ_ONCE(cpuc->excl_cntrs->exclusive_present))
909 unsched = perf_assign_events(cpuc->event_constraint, n, wmin,
910 wmax, gpmax, assign);
914 * In case of success (unsched = 0), mark events as committed,
915 * so we do not put_constraint() in case new events are added
916 * and fail to be scheduled
918 * We invoke the lower level commit callback to lock the resource
920 * We do not need to do all of this in case we are called to
921 * validate an event group (assign == NULL)
923 if (!unsched && assign) {
924 for (i = 0; i < n; i++) {
925 e = cpuc->event_list[i];
926 e->hw.flags |= PERF_X86_EVENT_COMMITTED;
927 if (x86_pmu.commit_scheduling)
928 x86_pmu.commit_scheduling(cpuc, i, assign[i]);
931 for (i = 0; i < n; i++) {
932 e = cpuc->event_list[i];
934 * do not put_constraint() on comitted events,
935 * because they are good to go
937 if ((e->hw.flags & PERF_X86_EVENT_COMMITTED))
941 * release events that failed scheduling
943 if (x86_pmu.put_event_constraints)
944 x86_pmu.put_event_constraints(cpuc, e);
948 if (x86_pmu.stop_scheduling)
949 x86_pmu.stop_scheduling(cpuc);
951 return unsched ? -EINVAL : 0;
955 * dogrp: true if must collect siblings events (group)
956 * returns total number of events and error code
958 static int collect_events(struct cpu_hw_events *cpuc, struct perf_event *leader, bool dogrp)
960 struct perf_event *event;
963 max_count = x86_pmu.num_counters + x86_pmu.num_counters_fixed;
965 /* current number of events already accepted */
968 if (is_x86_event(leader)) {
971 cpuc->event_list[n] = leader;
977 list_for_each_entry(event, &leader->sibling_list, group_entry) {
978 if (!is_x86_event(event) ||
979 event->state <= PERF_EVENT_STATE_OFF)
985 cpuc->event_list[n] = event;
991 static inline void x86_assign_hw_event(struct perf_event *event,
992 struct cpu_hw_events *cpuc, int i)
994 struct hw_perf_event *hwc = &event->hw;
996 hwc->idx = cpuc->assign[i];
997 hwc->last_cpu = smp_processor_id();
998 hwc->last_tag = ++cpuc->tags[i];
1000 if (hwc->idx == INTEL_PMC_IDX_FIXED_BTS) {
1001 hwc->config_base = 0;
1002 hwc->event_base = 0;
1003 } else if (hwc->idx >= INTEL_PMC_IDX_FIXED) {
1004 hwc->config_base = MSR_ARCH_PERFMON_FIXED_CTR_CTRL;
1005 hwc->event_base = MSR_ARCH_PERFMON_FIXED_CTR0 + (hwc->idx - INTEL_PMC_IDX_FIXED);
1006 hwc->event_base_rdpmc = (hwc->idx - INTEL_PMC_IDX_FIXED) | 1<<30;
1008 hwc->config_base = x86_pmu_config_addr(hwc->idx);
1009 hwc->event_base = x86_pmu_event_addr(hwc->idx);
1010 hwc->event_base_rdpmc = x86_pmu_rdpmc_index(hwc->idx);
1014 static inline int match_prev_assignment(struct hw_perf_event *hwc,
1015 struct cpu_hw_events *cpuc,
1018 return hwc->idx == cpuc->assign[i] &&
1019 hwc->last_cpu == smp_processor_id() &&
1020 hwc->last_tag == cpuc->tags[i];
1023 static void x86_pmu_start(struct perf_event *event, int flags);
1025 static void x86_pmu_enable(struct pmu *pmu)
1027 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1028 struct perf_event *event;
1029 struct hw_perf_event *hwc;
1030 int i, added = cpuc->n_added;
1032 if (!x86_pmu_initialized())
1038 if (cpuc->n_added) {
1039 int n_running = cpuc->n_events - cpuc->n_added;
1041 * apply assignment obtained either from
1042 * hw_perf_group_sched_in() or x86_pmu_enable()
1044 * step1: save events moving to new counters
1046 for (i = 0; i < n_running; i++) {
1047 event = cpuc->event_list[i];
1051 * we can avoid reprogramming counter if:
1052 * - assigned same counter as last time
1053 * - running on same CPU as last time
1054 * - no other event has used the counter since
1056 if (hwc->idx == -1 ||
1057 match_prev_assignment(hwc, cpuc, i))
1061 * Ensure we don't accidentally enable a stopped
1062 * counter simply because we rescheduled.
1064 if (hwc->state & PERF_HES_STOPPED)
1065 hwc->state |= PERF_HES_ARCH;
1067 x86_pmu_stop(event, PERF_EF_UPDATE);
1071 * step2: reprogram moved events into new counters
1073 for (i = 0; i < cpuc->n_events; i++) {
1074 event = cpuc->event_list[i];
1077 if (!match_prev_assignment(hwc, cpuc, i))
1078 x86_assign_hw_event(event, cpuc, i);
1079 else if (i < n_running)
1082 if (hwc->state & PERF_HES_ARCH)
1085 x86_pmu_start(event, PERF_EF_RELOAD);
1088 perf_events_lapic_init();
1094 x86_pmu.enable_all(added);
1097 static DEFINE_PER_CPU(u64 [X86_PMC_IDX_MAX], pmc_prev_left);
1100 * Set the next IRQ period, based on the hwc->period_left value.
1101 * To be called with the event disabled in hw:
1103 int x86_perf_event_set_period(struct perf_event *event)
1105 struct hw_perf_event *hwc = &event->hw;
1106 s64 left = local64_read(&hwc->period_left);
1107 s64 period = hwc->sample_period;
1108 int ret = 0, idx = hwc->idx;
1110 if (idx == INTEL_PMC_IDX_FIXED_BTS)
1114 * If we are way outside a reasonable range then just skip forward:
1116 if (unlikely(left <= -period)) {
1118 local64_set(&hwc->period_left, left);
1119 hwc->last_period = period;
1123 if (unlikely(left <= 0)) {
1125 local64_set(&hwc->period_left, left);
1126 hwc->last_period = period;
1130 * Quirk: certain CPUs dont like it if just 1 hw_event is left:
1132 if (unlikely(left < 2))
1135 if (left > x86_pmu.max_period)
1136 left = x86_pmu.max_period;
1138 if (x86_pmu.limit_period)
1139 left = x86_pmu.limit_period(event, left);
1141 per_cpu(pmc_prev_left[idx], smp_processor_id()) = left;
1143 if (!(hwc->flags & PERF_X86_EVENT_AUTO_RELOAD) ||
1144 local64_read(&hwc->prev_count) != (u64)-left) {
1146 * The hw event starts counting from this event offset,
1147 * mark it to be able to extra future deltas:
1149 local64_set(&hwc->prev_count, (u64)-left);
1151 wrmsrl(hwc->event_base, (u64)(-left) & x86_pmu.cntval_mask);
1155 * Due to erratum on certan cpu we need
1156 * a second write to be sure the register
1157 * is updated properly
1159 if (x86_pmu.perfctr_second_write) {
1160 wrmsrl(hwc->event_base,
1161 (u64)(-left) & x86_pmu.cntval_mask);
1164 perf_event_update_userpage(event);
1169 void x86_pmu_enable_event(struct perf_event *event)
1171 if (__this_cpu_read(cpu_hw_events.enabled))
1172 __x86_pmu_enable_event(&event->hw,
1173 ARCH_PERFMON_EVENTSEL_ENABLE);
1177 * Add a single event to the PMU.
1179 * The event is added to the group of enabled events
1180 * but only if it can be scehduled with existing events.
1182 static int x86_pmu_add(struct perf_event *event, int flags)
1184 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1185 struct hw_perf_event *hwc;
1186 int assign[X86_PMC_IDX_MAX];
1191 n0 = cpuc->n_events;
1192 ret = n = collect_events(cpuc, event, false);
1196 hwc->state = PERF_HES_UPTODATE | PERF_HES_STOPPED;
1197 if (!(flags & PERF_EF_START))
1198 hwc->state |= PERF_HES_ARCH;
1201 * If group events scheduling transaction was started,
1202 * skip the schedulability test here, it will be performed
1203 * at commit time (->commit_txn) as a whole.
1205 if (cpuc->txn_flags & PERF_PMU_TXN_ADD)
1208 ret = x86_pmu.schedule_events(cpuc, n, assign);
1212 * copy new assignment, now we know it is possible
1213 * will be used by hw_perf_enable()
1215 memcpy(cpuc->assign, assign, n*sizeof(int));
1219 * Commit the collect_events() state. See x86_pmu_del() and
1223 cpuc->n_added += n - n0;
1224 cpuc->n_txn += n - n0;
1231 static void x86_pmu_start(struct perf_event *event, int flags)
1233 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1234 int idx = event->hw.idx;
1236 if (WARN_ON_ONCE(!(event->hw.state & PERF_HES_STOPPED)))
1239 if (WARN_ON_ONCE(idx == -1))
1242 if (flags & PERF_EF_RELOAD) {
1243 WARN_ON_ONCE(!(event->hw.state & PERF_HES_UPTODATE));
1244 x86_perf_event_set_period(event);
1247 event->hw.state = 0;
1249 cpuc->events[idx] = event;
1250 __set_bit(idx, cpuc->active_mask);
1251 __set_bit(idx, cpuc->running);
1252 x86_pmu.enable(event);
1253 perf_event_update_userpage(event);
1256 void perf_event_print_debug(void)
1258 u64 ctrl, status, overflow, pmc_ctrl, pmc_count, prev_left, fixed;
1260 struct cpu_hw_events *cpuc;
1261 unsigned long flags;
1264 if (!x86_pmu.num_counters)
1267 local_irq_save(flags);
1269 cpu = smp_processor_id();
1270 cpuc = &per_cpu(cpu_hw_events, cpu);
1272 if (x86_pmu.version >= 2) {
1273 rdmsrl(MSR_CORE_PERF_GLOBAL_CTRL, ctrl);
1274 rdmsrl(MSR_CORE_PERF_GLOBAL_STATUS, status);
1275 rdmsrl(MSR_CORE_PERF_GLOBAL_OVF_CTRL, overflow);
1276 rdmsrl(MSR_ARCH_PERFMON_FIXED_CTR_CTRL, fixed);
1279 pr_info("CPU#%d: ctrl: %016llx\n", cpu, ctrl);
1280 pr_info("CPU#%d: status: %016llx\n", cpu, status);
1281 pr_info("CPU#%d: overflow: %016llx\n", cpu, overflow);
1282 pr_info("CPU#%d: fixed: %016llx\n", cpu, fixed);
1283 if (x86_pmu.pebs_constraints) {
1284 rdmsrl(MSR_IA32_PEBS_ENABLE, pebs);
1285 pr_info("CPU#%d: pebs: %016llx\n", cpu, pebs);
1287 if (x86_pmu.lbr_nr) {
1288 rdmsrl(MSR_IA32_DEBUGCTLMSR, debugctl);
1289 pr_info("CPU#%d: debugctl: %016llx\n", cpu, debugctl);
1292 pr_info("CPU#%d: active: %016llx\n", cpu, *(u64 *)cpuc->active_mask);
1294 for (idx = 0; idx < x86_pmu.num_counters; idx++) {
1295 rdmsrl(x86_pmu_config_addr(idx), pmc_ctrl);
1296 rdmsrl(x86_pmu_event_addr(idx), pmc_count);
1298 prev_left = per_cpu(pmc_prev_left[idx], cpu);
1300 pr_info("CPU#%d: gen-PMC%d ctrl: %016llx\n",
1301 cpu, idx, pmc_ctrl);
1302 pr_info("CPU#%d: gen-PMC%d count: %016llx\n",
1303 cpu, idx, pmc_count);
1304 pr_info("CPU#%d: gen-PMC%d left: %016llx\n",
1305 cpu, idx, prev_left);
1307 for (idx = 0; idx < x86_pmu.num_counters_fixed; idx++) {
1308 rdmsrl(MSR_ARCH_PERFMON_FIXED_CTR0 + idx, pmc_count);
1310 pr_info("CPU#%d: fixed-PMC%d count: %016llx\n",
1311 cpu, idx, pmc_count);
1313 local_irq_restore(flags);
1316 void x86_pmu_stop(struct perf_event *event, int flags)
1318 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1319 struct hw_perf_event *hwc = &event->hw;
1321 if (__test_and_clear_bit(hwc->idx, cpuc->active_mask)) {
1322 x86_pmu.disable(event);
1323 cpuc->events[hwc->idx] = NULL;
1324 WARN_ON_ONCE(hwc->state & PERF_HES_STOPPED);
1325 hwc->state |= PERF_HES_STOPPED;
1328 if ((flags & PERF_EF_UPDATE) && !(hwc->state & PERF_HES_UPTODATE)) {
1330 * Drain the remaining delta count out of a event
1331 * that we are disabling:
1333 x86_perf_event_update(event);
1334 hwc->state |= PERF_HES_UPTODATE;
1338 static void x86_pmu_del(struct perf_event *event, int flags)
1340 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1344 * event is descheduled
1346 event->hw.flags &= ~PERF_X86_EVENT_COMMITTED;
1349 * If we're called during a txn, we don't need to do anything.
1350 * The events never got scheduled and ->cancel_txn will truncate
1353 * XXX assumes any ->del() called during a TXN will only be on
1354 * an event added during that same TXN.
1356 if (cpuc->txn_flags & PERF_PMU_TXN_ADD)
1360 * Not a TXN, therefore cleanup properly.
1362 x86_pmu_stop(event, PERF_EF_UPDATE);
1364 for (i = 0; i < cpuc->n_events; i++) {
1365 if (event == cpuc->event_list[i])
1369 if (WARN_ON_ONCE(i == cpuc->n_events)) /* called ->del() without ->add() ? */
1372 /* If we have a newly added event; make sure to decrease n_added. */
1373 if (i >= cpuc->n_events - cpuc->n_added)
1376 if (x86_pmu.put_event_constraints)
1377 x86_pmu.put_event_constraints(cpuc, event);
1379 /* Delete the array entry. */
1380 while (++i < cpuc->n_events) {
1381 cpuc->event_list[i-1] = cpuc->event_list[i];
1382 cpuc->event_constraint[i-1] = cpuc->event_constraint[i];
1386 perf_event_update_userpage(event);
1389 int x86_pmu_handle_irq(struct pt_regs *regs)
1391 struct perf_sample_data data;
1392 struct cpu_hw_events *cpuc;
1393 struct perf_event *event;
1394 int idx, handled = 0;
1397 cpuc = this_cpu_ptr(&cpu_hw_events);
1400 * Some chipsets need to unmask the LVTPC in a particular spot
1401 * inside the nmi handler. As a result, the unmasking was pushed
1402 * into all the nmi handlers.
1404 * This generic handler doesn't seem to have any issues where the
1405 * unmasking occurs so it was left at the top.
1407 apic_write(APIC_LVTPC, APIC_DM_NMI);
1409 for (idx = 0; idx < x86_pmu.num_counters; idx++) {
1410 if (!test_bit(idx, cpuc->active_mask)) {
1412 * Though we deactivated the counter some cpus
1413 * might still deliver spurious interrupts still
1414 * in flight. Catch them:
1416 if (__test_and_clear_bit(idx, cpuc->running))
1421 event = cpuc->events[idx];
1423 val = x86_perf_event_update(event);
1424 if (val & (1ULL << (x86_pmu.cntval_bits - 1)))
1431 perf_sample_data_init(&data, 0, event->hw.last_period);
1433 if (!x86_perf_event_set_period(event))
1436 if (perf_event_overflow(event, &data, regs))
1437 x86_pmu_stop(event, 0);
1441 inc_irq_stat(apic_perf_irqs);
1446 void perf_events_lapic_init(void)
1448 if (!x86_pmu.apic || !x86_pmu_initialized())
1452 * Always use NMI for PMU
1454 apic_write(APIC_LVTPC, APIC_DM_NMI);
1458 perf_event_nmi_handler(unsigned int cmd, struct pt_regs *regs)
1465 * All PMUs/events that share this PMI handler should make sure to
1466 * increment active_events for their events.
1468 if (!atomic_read(&active_events))
1471 start_clock = sched_clock();
1472 ret = x86_pmu.handle_irq(regs);
1473 finish_clock = sched_clock();
1475 perf_sample_event_took(finish_clock - start_clock);
1479 NOKPROBE_SYMBOL(perf_event_nmi_handler);
1481 struct event_constraint emptyconstraint;
1482 struct event_constraint unconstrained;
1484 static int x86_pmu_prepare_cpu(unsigned int cpu)
1486 struct cpu_hw_events *cpuc = &per_cpu(cpu_hw_events, cpu);
1489 for (i = 0 ; i < X86_PERF_KFREE_MAX; i++)
1490 cpuc->kfree_on_online[i] = NULL;
1491 if (x86_pmu.cpu_prepare)
1492 return x86_pmu.cpu_prepare(cpu);
1496 static int x86_pmu_dead_cpu(unsigned int cpu)
1498 if (x86_pmu.cpu_dead)
1499 x86_pmu.cpu_dead(cpu);
1503 static int x86_pmu_online_cpu(unsigned int cpu)
1505 struct cpu_hw_events *cpuc = &per_cpu(cpu_hw_events, cpu);
1508 for (i = 0 ; i < X86_PERF_KFREE_MAX; i++) {
1509 kfree(cpuc->kfree_on_online[i]);
1510 cpuc->kfree_on_online[i] = NULL;
1515 static int x86_pmu_starting_cpu(unsigned int cpu)
1517 if (x86_pmu.cpu_starting)
1518 x86_pmu.cpu_starting(cpu);
1522 static int x86_pmu_dying_cpu(unsigned int cpu)
1524 if (x86_pmu.cpu_dying)
1525 x86_pmu.cpu_dying(cpu);
1529 static void __init pmu_check_apic(void)
1531 if (boot_cpu_has(X86_FEATURE_APIC))
1535 pr_info("no APIC, boot with the \"lapic\" boot parameter to force-enable it.\n");
1536 pr_info("no hardware sampling interrupt available.\n");
1539 * If we have a PMU initialized but no APIC
1540 * interrupts, we cannot sample hardware
1541 * events (user-space has to fall back and
1542 * sample via a hrtimer based software event):
1544 pmu.capabilities |= PERF_PMU_CAP_NO_INTERRUPT;
1548 static struct attribute_group x86_pmu_format_group = {
1554 * Remove all undefined events (x86_pmu.event_map(id) == 0)
1555 * out of events_attr attributes.
1557 static void __init filter_events(struct attribute **attrs)
1559 struct device_attribute *d;
1560 struct perf_pmu_events_attr *pmu_attr;
1564 for (i = 0; attrs[i]; i++) {
1565 d = (struct device_attribute *)attrs[i];
1566 pmu_attr = container_of(d, struct perf_pmu_events_attr, attr);
1568 if (pmu_attr->event_str)
1570 if (x86_pmu.event_map(i + offset))
1573 for (j = i; attrs[j]; j++)
1574 attrs[j] = attrs[j + 1];
1576 /* Check the shifted attr. */
1580 * event_map() is index based, the attrs array is organized
1581 * by increasing event index. If we shift the events, then
1582 * we need to compensate for the event_map(), otherwise
1583 * we are looking up the wrong event in the map
1589 /* Merge two pointer arrays */
1590 __init struct attribute **merge_attr(struct attribute **a, struct attribute **b)
1592 struct attribute **new;
1595 for (j = 0; a[j]; j++)
1597 for (i = 0; b[i]; i++)
1601 new = kmalloc(sizeof(struct attribute *) * j, GFP_KERNEL);
1606 for (i = 0; a[i]; i++)
1608 for (i = 0; b[i]; i++)
1615 ssize_t events_sysfs_show(struct device *dev, struct device_attribute *attr, char *page)
1617 struct perf_pmu_events_attr *pmu_attr = \
1618 container_of(attr, struct perf_pmu_events_attr, attr);
1619 u64 config = x86_pmu.event_map(pmu_attr->id);
1621 /* string trumps id */
1622 if (pmu_attr->event_str)
1623 return sprintf(page, "%s", pmu_attr->event_str);
1625 return x86_pmu.events_sysfs_show(page, config);
1627 EXPORT_SYMBOL_GPL(events_sysfs_show);
1629 ssize_t events_ht_sysfs_show(struct device *dev, struct device_attribute *attr,
1632 struct perf_pmu_events_ht_attr *pmu_attr =
1633 container_of(attr, struct perf_pmu_events_ht_attr, attr);
1636 * Report conditional events depending on Hyper-Threading.
1638 * This is overly conservative as usually the HT special
1639 * handling is not needed if the other CPU thread is idle.
1641 * Note this does not (and cannot) handle the case when thread
1642 * siblings are invisible, for example with virtualization
1643 * if they are owned by some other guest. The user tool
1644 * has to re-read when a thread sibling gets onlined later.
1646 return sprintf(page, "%s",
1647 topology_max_smt_threads() > 1 ?
1648 pmu_attr->event_str_ht :
1649 pmu_attr->event_str_noht);
1652 EVENT_ATTR(cpu-cycles, CPU_CYCLES );
1653 EVENT_ATTR(instructions, INSTRUCTIONS );
1654 EVENT_ATTR(cache-references, CACHE_REFERENCES );
1655 EVENT_ATTR(cache-misses, CACHE_MISSES );
1656 EVENT_ATTR(branch-instructions, BRANCH_INSTRUCTIONS );
1657 EVENT_ATTR(branch-misses, BRANCH_MISSES );
1658 EVENT_ATTR(bus-cycles, BUS_CYCLES );
1659 EVENT_ATTR(stalled-cycles-frontend, STALLED_CYCLES_FRONTEND );
1660 EVENT_ATTR(stalled-cycles-backend, STALLED_CYCLES_BACKEND );
1661 EVENT_ATTR(ref-cycles, REF_CPU_CYCLES );
1663 static struct attribute *empty_attrs;
1665 static struct attribute *events_attr[] = {
1666 EVENT_PTR(CPU_CYCLES),
1667 EVENT_PTR(INSTRUCTIONS),
1668 EVENT_PTR(CACHE_REFERENCES),
1669 EVENT_PTR(CACHE_MISSES),
1670 EVENT_PTR(BRANCH_INSTRUCTIONS),
1671 EVENT_PTR(BRANCH_MISSES),
1672 EVENT_PTR(BUS_CYCLES),
1673 EVENT_PTR(STALLED_CYCLES_FRONTEND),
1674 EVENT_PTR(STALLED_CYCLES_BACKEND),
1675 EVENT_PTR(REF_CPU_CYCLES),
1679 static struct attribute_group x86_pmu_events_group = {
1681 .attrs = events_attr,
1684 ssize_t x86_event_sysfs_show(char *page, u64 config, u64 event)
1686 u64 umask = (config & ARCH_PERFMON_EVENTSEL_UMASK) >> 8;
1687 u64 cmask = (config & ARCH_PERFMON_EVENTSEL_CMASK) >> 24;
1688 bool edge = (config & ARCH_PERFMON_EVENTSEL_EDGE);
1689 bool pc = (config & ARCH_PERFMON_EVENTSEL_PIN_CONTROL);
1690 bool any = (config & ARCH_PERFMON_EVENTSEL_ANY);
1691 bool inv = (config & ARCH_PERFMON_EVENTSEL_INV);
1695 * We have whole page size to spend and just little data
1696 * to write, so we can safely use sprintf.
1698 ret = sprintf(page, "event=0x%02llx", event);
1701 ret += sprintf(page + ret, ",umask=0x%02llx", umask);
1704 ret += sprintf(page + ret, ",edge");
1707 ret += sprintf(page + ret, ",pc");
1710 ret += sprintf(page + ret, ",any");
1713 ret += sprintf(page + ret, ",inv");
1716 ret += sprintf(page + ret, ",cmask=0x%02llx", cmask);
1718 ret += sprintf(page + ret, "\n");
1723 static int __init init_hw_perf_events(void)
1725 struct x86_pmu_quirk *quirk;
1728 pr_info("Performance Events: ");
1730 switch (boot_cpu_data.x86_vendor) {
1731 case X86_VENDOR_INTEL:
1732 err = intel_pmu_init();
1734 case X86_VENDOR_AMD:
1735 err = amd_pmu_init();
1741 pr_cont("no PMU driver, software events only.\n");
1747 /* sanity check that the hardware exists or is emulated */
1748 if (!check_hw_exists())
1751 pr_cont("%s PMU driver.\n", x86_pmu.name);
1753 x86_pmu.attr_rdpmc = 1; /* enable userspace RDPMC usage by default */
1755 for (quirk = x86_pmu.quirks; quirk; quirk = quirk->next)
1758 if (!x86_pmu.intel_ctrl)
1759 x86_pmu.intel_ctrl = (1 << x86_pmu.num_counters) - 1;
1761 perf_events_lapic_init();
1762 register_nmi_handler(NMI_LOCAL, perf_event_nmi_handler, 0, "PMI");
1764 unconstrained = (struct event_constraint)
1765 __EVENT_CONSTRAINT(0, (1ULL << x86_pmu.num_counters) - 1,
1766 0, x86_pmu.num_counters, 0, 0);
1768 x86_pmu_format_group.attrs = x86_pmu.format_attrs;
1770 if (x86_pmu.event_attrs)
1771 x86_pmu_events_group.attrs = x86_pmu.event_attrs;
1773 if (!x86_pmu.events_sysfs_show)
1774 x86_pmu_events_group.attrs = &empty_attrs;
1776 filter_events(x86_pmu_events_group.attrs);
1778 if (x86_pmu.cpu_events) {
1779 struct attribute **tmp;
1781 tmp = merge_attr(x86_pmu_events_group.attrs, x86_pmu.cpu_events);
1783 x86_pmu_events_group.attrs = tmp;
1786 pr_info("... version: %d\n", x86_pmu.version);
1787 pr_info("... bit width: %d\n", x86_pmu.cntval_bits);
1788 pr_info("... generic registers: %d\n", x86_pmu.num_counters);
1789 pr_info("... value mask: %016Lx\n", x86_pmu.cntval_mask);
1790 pr_info("... max period: %016Lx\n", x86_pmu.max_period);
1791 pr_info("... fixed-purpose events: %d\n", x86_pmu.num_counters_fixed);
1792 pr_info("... event mask: %016Lx\n", x86_pmu.intel_ctrl);
1795 * Install callbacks. Core will call them for each online
1798 err = cpuhp_setup_state(CPUHP_PERF_X86_PREPARE, "PERF_X86_PREPARE",
1799 x86_pmu_prepare_cpu, x86_pmu_dead_cpu);
1803 err = cpuhp_setup_state(CPUHP_AP_PERF_X86_STARTING,
1804 "AP_PERF_X86_STARTING", x86_pmu_starting_cpu,
1809 err = cpuhp_setup_state(CPUHP_AP_PERF_X86_ONLINE, "AP_PERF_X86_ONLINE",
1810 x86_pmu_online_cpu, NULL);
1814 err = perf_pmu_register(&pmu, "cpu", PERF_TYPE_RAW);
1821 cpuhp_remove_state(CPUHP_AP_PERF_X86_ONLINE);
1823 cpuhp_remove_state(CPUHP_AP_PERF_X86_STARTING);
1825 cpuhp_remove_state(CPUHP_PERF_X86_PREPARE);
1828 early_initcall(init_hw_perf_events);
1830 static inline void x86_pmu_read(struct perf_event *event)
1832 x86_perf_event_update(event);
1836 * Start group events scheduling transaction
1837 * Set the flag to make pmu::enable() not perform the
1838 * schedulability test, it will be performed at commit time
1840 * We only support PERF_PMU_TXN_ADD transactions. Save the
1841 * transaction flags but otherwise ignore non-PERF_PMU_TXN_ADD
1844 static void x86_pmu_start_txn(struct pmu *pmu, unsigned int txn_flags)
1846 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1848 WARN_ON_ONCE(cpuc->txn_flags); /* txn already in flight */
1850 cpuc->txn_flags = txn_flags;
1851 if (txn_flags & ~PERF_PMU_TXN_ADD)
1854 perf_pmu_disable(pmu);
1855 __this_cpu_write(cpu_hw_events.n_txn, 0);
1859 * Stop group events scheduling transaction
1860 * Clear the flag and pmu::enable() will perform the
1861 * schedulability test.
1863 static void x86_pmu_cancel_txn(struct pmu *pmu)
1865 unsigned int txn_flags;
1866 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1868 WARN_ON_ONCE(!cpuc->txn_flags); /* no txn in flight */
1870 txn_flags = cpuc->txn_flags;
1871 cpuc->txn_flags = 0;
1872 if (txn_flags & ~PERF_PMU_TXN_ADD)
1876 * Truncate collected array by the number of events added in this
1877 * transaction. See x86_pmu_add() and x86_pmu_*_txn().
1879 __this_cpu_sub(cpu_hw_events.n_added, __this_cpu_read(cpu_hw_events.n_txn));
1880 __this_cpu_sub(cpu_hw_events.n_events, __this_cpu_read(cpu_hw_events.n_txn));
1881 perf_pmu_enable(pmu);
1885 * Commit group events scheduling transaction
1886 * Perform the group schedulability test as a whole
1887 * Return 0 if success
1889 * Does not cancel the transaction on failure; expects the caller to do this.
1891 static int x86_pmu_commit_txn(struct pmu *pmu)
1893 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1894 int assign[X86_PMC_IDX_MAX];
1897 WARN_ON_ONCE(!cpuc->txn_flags); /* no txn in flight */
1899 if (cpuc->txn_flags & ~PERF_PMU_TXN_ADD) {
1900 cpuc->txn_flags = 0;
1906 if (!x86_pmu_initialized())
1909 ret = x86_pmu.schedule_events(cpuc, n, assign);
1914 * copy new assignment, now we know it is possible
1915 * will be used by hw_perf_enable()
1917 memcpy(cpuc->assign, assign, n*sizeof(int));
1919 cpuc->txn_flags = 0;
1920 perf_pmu_enable(pmu);
1924 * a fake_cpuc is used to validate event groups. Due to
1925 * the extra reg logic, we need to also allocate a fake
1926 * per_core and per_cpu structure. Otherwise, group events
1927 * using extra reg may conflict without the kernel being
1928 * able to catch this when the last event gets added to
1931 static void free_fake_cpuc(struct cpu_hw_events *cpuc)
1933 kfree(cpuc->shared_regs);
1937 static struct cpu_hw_events *allocate_fake_cpuc(void)
1939 struct cpu_hw_events *cpuc;
1940 int cpu = raw_smp_processor_id();
1942 cpuc = kzalloc(sizeof(*cpuc), GFP_KERNEL);
1944 return ERR_PTR(-ENOMEM);
1946 /* only needed, if we have extra_regs */
1947 if (x86_pmu.extra_regs) {
1948 cpuc->shared_regs = allocate_shared_regs(cpu);
1949 if (!cpuc->shared_regs)
1955 free_fake_cpuc(cpuc);
1956 return ERR_PTR(-ENOMEM);
1960 * validate that we can schedule this event
1962 static int validate_event(struct perf_event *event)
1964 struct cpu_hw_events *fake_cpuc;
1965 struct event_constraint *c;
1968 fake_cpuc = allocate_fake_cpuc();
1969 if (IS_ERR(fake_cpuc))
1970 return PTR_ERR(fake_cpuc);
1972 c = x86_pmu.get_event_constraints(fake_cpuc, -1, event);
1974 if (!c || !c->weight)
1977 if (x86_pmu.put_event_constraints)
1978 x86_pmu.put_event_constraints(fake_cpuc, event);
1980 free_fake_cpuc(fake_cpuc);
1986 * validate a single event group
1988 * validation include:
1989 * - check events are compatible which each other
1990 * - events do not compete for the same counter
1991 * - number of events <= number of counters
1993 * validation ensures the group can be loaded onto the
1994 * PMU if it was the only group available.
1996 static int validate_group(struct perf_event *event)
1998 struct perf_event *leader = event->group_leader;
1999 struct cpu_hw_events *fake_cpuc;
2000 int ret = -EINVAL, n;
2002 fake_cpuc = allocate_fake_cpuc();
2003 if (IS_ERR(fake_cpuc))
2004 return PTR_ERR(fake_cpuc);
2006 * the event is not yet connected with its
2007 * siblings therefore we must first collect
2008 * existing siblings, then add the new event
2009 * before we can simulate the scheduling
2011 n = collect_events(fake_cpuc, leader, true);
2015 fake_cpuc->n_events = n;
2016 n = collect_events(fake_cpuc, event, false);
2020 fake_cpuc->n_events = n;
2022 ret = x86_pmu.schedule_events(fake_cpuc, n, NULL);
2025 free_fake_cpuc(fake_cpuc);
2029 static int x86_pmu_event_init(struct perf_event *event)
2034 switch (event->attr.type) {
2036 case PERF_TYPE_HARDWARE:
2037 case PERF_TYPE_HW_CACHE:
2044 err = __x86_pmu_event_init(event);
2047 * we temporarily connect event to its pmu
2048 * such that validate_group() can classify
2049 * it as an x86 event using is_x86_event()
2054 if (event->group_leader != event)
2055 err = validate_group(event);
2057 err = validate_event(event);
2063 event->destroy(event);
2066 if (ACCESS_ONCE(x86_pmu.attr_rdpmc))
2067 event->hw.flags |= PERF_X86_EVENT_RDPMC_ALLOWED;
2072 static void refresh_pce(void *ignored)
2075 load_mm_cr4(current->mm);
2078 static void x86_pmu_event_mapped(struct perf_event *event)
2080 if (!(event->hw.flags & PERF_X86_EVENT_RDPMC_ALLOWED))
2083 if (atomic_inc_return(¤t->mm->context.perf_rdpmc_allowed) == 1)
2084 on_each_cpu_mask(mm_cpumask(current->mm), refresh_pce, NULL, 1);
2087 static void x86_pmu_event_unmapped(struct perf_event *event)
2092 if (!(event->hw.flags & PERF_X86_EVENT_RDPMC_ALLOWED))
2095 if (atomic_dec_and_test(¤t->mm->context.perf_rdpmc_allowed))
2096 on_each_cpu_mask(mm_cpumask(current->mm), refresh_pce, NULL, 1);
2099 static int x86_pmu_event_idx(struct perf_event *event)
2101 int idx = event->hw.idx;
2103 if (!(event->hw.flags & PERF_X86_EVENT_RDPMC_ALLOWED))
2106 if (x86_pmu.num_counters_fixed && idx >= INTEL_PMC_IDX_FIXED) {
2107 idx -= INTEL_PMC_IDX_FIXED;
2114 static ssize_t get_attr_rdpmc(struct device *cdev,
2115 struct device_attribute *attr,
2118 return snprintf(buf, 40, "%d\n", x86_pmu.attr_rdpmc);
2121 static ssize_t set_attr_rdpmc(struct device *cdev,
2122 struct device_attribute *attr,
2123 const char *buf, size_t count)
2128 ret = kstrtoul(buf, 0, &val);
2135 if (x86_pmu.attr_rdpmc_broken)
2138 if ((val == 2) != (x86_pmu.attr_rdpmc == 2)) {
2140 * Changing into or out of always available, aka
2141 * perf-event-bypassing mode. This path is extremely slow,
2142 * but only root can trigger it, so it's okay.
2145 static_key_slow_inc(&rdpmc_always_available);
2147 static_key_slow_dec(&rdpmc_always_available);
2148 on_each_cpu(refresh_pce, NULL, 1);
2151 x86_pmu.attr_rdpmc = val;
2156 static DEVICE_ATTR(rdpmc, S_IRUSR | S_IWUSR, get_attr_rdpmc, set_attr_rdpmc);
2158 static struct attribute *x86_pmu_attrs[] = {
2159 &dev_attr_rdpmc.attr,
2163 static struct attribute_group x86_pmu_attr_group = {
2164 .attrs = x86_pmu_attrs,
2167 static const struct attribute_group *x86_pmu_attr_groups[] = {
2168 &x86_pmu_attr_group,
2169 &x86_pmu_format_group,
2170 &x86_pmu_events_group,
2174 static void x86_pmu_sched_task(struct perf_event_context *ctx, bool sched_in)
2176 if (x86_pmu.sched_task)
2177 x86_pmu.sched_task(ctx, sched_in);
2180 void perf_check_microcode(void)
2182 if (x86_pmu.check_microcode)
2183 x86_pmu.check_microcode();
2185 EXPORT_SYMBOL_GPL(perf_check_microcode);
2187 static struct pmu pmu = {
2188 .pmu_enable = x86_pmu_enable,
2189 .pmu_disable = x86_pmu_disable,
2191 .attr_groups = x86_pmu_attr_groups,
2193 .event_init = x86_pmu_event_init,
2195 .event_mapped = x86_pmu_event_mapped,
2196 .event_unmapped = x86_pmu_event_unmapped,
2200 .start = x86_pmu_start,
2201 .stop = x86_pmu_stop,
2202 .read = x86_pmu_read,
2204 .start_txn = x86_pmu_start_txn,
2205 .cancel_txn = x86_pmu_cancel_txn,
2206 .commit_txn = x86_pmu_commit_txn,
2208 .event_idx = x86_pmu_event_idx,
2209 .sched_task = x86_pmu_sched_task,
2210 .task_ctx_size = sizeof(struct x86_perf_task_context),
2213 void arch_perf_update_userpage(struct perf_event *event,
2214 struct perf_event_mmap_page *userpg, u64 now)
2216 struct cyc2ns_data *data;
2218 userpg->cap_user_time = 0;
2219 userpg->cap_user_time_zero = 0;
2220 userpg->cap_user_rdpmc =
2221 !!(event->hw.flags & PERF_X86_EVENT_RDPMC_ALLOWED);
2222 userpg->pmc_width = x86_pmu.cntval_bits;
2224 if (!sched_clock_stable())
2227 data = cyc2ns_read_begin();
2230 * Internal timekeeping for enabled/running/stopped times
2231 * is always in the local_clock domain.
2233 userpg->cap_user_time = 1;
2234 userpg->time_mult = data->cyc2ns_mul;
2235 userpg->time_shift = data->cyc2ns_shift;
2236 userpg->time_offset = data->cyc2ns_offset - now;
2239 * cap_user_time_zero doesn't make sense when we're using a different
2240 * time base for the records.
2242 if (!event->attr.use_clockid) {
2243 userpg->cap_user_time_zero = 1;
2244 userpg->time_zero = data->cyc2ns_offset;
2247 cyc2ns_read_end(data);
2254 static int backtrace_stack(void *data, char *name)
2259 static int backtrace_address(void *data, unsigned long addr, int reliable)
2261 struct perf_callchain_entry_ctx *entry = data;
2263 return perf_callchain_store(entry, addr);
2266 static const struct stacktrace_ops backtrace_ops = {
2267 .stack = backtrace_stack,
2268 .address = backtrace_address,
2269 .walk_stack = print_context_stack_bp,
2273 perf_callchain_kernel(struct perf_callchain_entry_ctx *entry, struct pt_regs *regs)
2275 if (perf_guest_cbs && perf_guest_cbs->is_in_guest()) {
2276 /* TODO: We don't support guest os callchain now */
2280 perf_callchain_store(entry, regs->ip);
2282 dump_trace(NULL, regs, NULL, 0, &backtrace_ops, entry);
2286 valid_user_frame(const void __user *fp, unsigned long size)
2288 return (__range_not_ok(fp, size, TASK_SIZE) == 0);
2291 static unsigned long get_segment_base(unsigned int segment)
2293 struct desc_struct *desc;
2294 int idx = segment >> 3;
2296 if ((segment & SEGMENT_TI_MASK) == SEGMENT_LDT) {
2297 #ifdef CONFIG_MODIFY_LDT_SYSCALL
2298 struct ldt_struct *ldt;
2300 if (idx > LDT_ENTRIES)
2303 /* IRQs are off, so this synchronizes with smp_store_release */
2304 ldt = lockless_dereference(current->active_mm->context.ldt);
2305 if (!ldt || idx > ldt->size)
2308 desc = &ldt->entries[idx];
2313 if (idx > GDT_ENTRIES)
2316 desc = raw_cpu_ptr(gdt_page.gdt) + idx;
2319 return get_desc_base(desc);
2322 #ifdef CONFIG_IA32_EMULATION
2324 #include <asm/compat.h>
2327 perf_callchain_user32(struct pt_regs *regs, struct perf_callchain_entry_ctx *entry)
2329 /* 32-bit process in 64-bit kernel. */
2330 unsigned long ss_base, cs_base;
2331 struct stack_frame_ia32 frame;
2332 const void __user *fp;
2334 if (!test_thread_flag(TIF_IA32))
2337 cs_base = get_segment_base(regs->cs);
2338 ss_base = get_segment_base(regs->ss);
2340 fp = compat_ptr(ss_base + regs->bp);
2341 pagefault_disable();
2342 while (entry->nr < entry->max_stack) {
2343 unsigned long bytes;
2344 frame.next_frame = 0;
2345 frame.return_address = 0;
2347 if (!access_ok(VERIFY_READ, fp, 8))
2350 bytes = __copy_from_user_nmi(&frame.next_frame, fp, 4);
2353 bytes = __copy_from_user_nmi(&frame.return_address, fp+4, 4);
2357 if (!valid_user_frame(fp, sizeof(frame)))
2360 perf_callchain_store(entry, cs_base + frame.return_address);
2361 fp = compat_ptr(ss_base + frame.next_frame);
2368 perf_callchain_user32(struct pt_regs *regs, struct perf_callchain_entry_ctx *entry)
2375 perf_callchain_user(struct perf_callchain_entry_ctx *entry, struct pt_regs *regs)
2377 struct stack_frame frame;
2378 const unsigned long __user *fp;
2380 if (perf_guest_cbs && perf_guest_cbs->is_in_guest()) {
2381 /* TODO: We don't support guest os callchain now */
2386 * We don't know what to do with VM86 stacks.. ignore them for now.
2388 if (regs->flags & (X86_VM_MASK | PERF_EFLAGS_VM))
2391 fp = (unsigned long __user *)regs->bp;
2393 perf_callchain_store(entry, regs->ip);
2398 if (perf_callchain_user32(regs, entry))
2401 pagefault_disable();
2402 while (entry->nr < entry->max_stack) {
2403 unsigned long bytes;
2405 frame.next_frame = NULL;
2406 frame.return_address = 0;
2408 if (!access_ok(VERIFY_READ, fp, sizeof(*fp) * 2))
2411 bytes = __copy_from_user_nmi(&frame.next_frame, fp, sizeof(*fp));
2414 bytes = __copy_from_user_nmi(&frame.return_address, fp + 1, sizeof(*fp));
2418 if (!valid_user_frame(fp, sizeof(frame)))
2421 perf_callchain_store(entry, frame.return_address);
2422 fp = (void __user *)frame.next_frame;
2428 * Deal with code segment offsets for the various execution modes:
2430 * VM86 - the good olde 16 bit days, where the linear address is
2431 * 20 bits and we use regs->ip + 0x10 * regs->cs.
2433 * IA32 - Where we need to look at GDT/LDT segment descriptor tables
2434 * to figure out what the 32bit base address is.
2436 * X32 - has TIF_X32 set, but is running in x86_64
2438 * X86_64 - CS,DS,SS,ES are all zero based.
2440 static unsigned long code_segment_base(struct pt_regs *regs)
2443 * For IA32 we look at the GDT/LDT segment base to convert the
2444 * effective IP to a linear address.
2447 #ifdef CONFIG_X86_32
2449 * If we are in VM86 mode, add the segment offset to convert to a
2452 if (regs->flags & X86_VM_MASK)
2453 return 0x10 * regs->cs;
2455 if (user_mode(regs) && regs->cs != __USER_CS)
2456 return get_segment_base(regs->cs);
2458 if (user_mode(regs) && !user_64bit_mode(regs) &&
2459 regs->cs != __USER32_CS)
2460 return get_segment_base(regs->cs);
2465 unsigned long perf_instruction_pointer(struct pt_regs *regs)
2467 if (perf_guest_cbs && perf_guest_cbs->is_in_guest())
2468 return perf_guest_cbs->get_guest_ip();
2470 return regs->ip + code_segment_base(regs);
2473 unsigned long perf_misc_flags(struct pt_regs *regs)
2477 if (perf_guest_cbs && perf_guest_cbs->is_in_guest()) {
2478 if (perf_guest_cbs->is_user_mode())
2479 misc |= PERF_RECORD_MISC_GUEST_USER;
2481 misc |= PERF_RECORD_MISC_GUEST_KERNEL;
2483 if (user_mode(regs))
2484 misc |= PERF_RECORD_MISC_USER;
2486 misc |= PERF_RECORD_MISC_KERNEL;
2489 if (regs->flags & PERF_EFLAGS_EXACT)
2490 misc |= PERF_RECORD_MISC_EXACT_IP;
2495 void perf_get_x86_pmu_capability(struct x86_pmu_capability *cap)
2497 cap->version = x86_pmu.version;
2498 cap->num_counters_gp = x86_pmu.num_counters;
2499 cap->num_counters_fixed = x86_pmu.num_counters_fixed;
2500 cap->bit_width_gp = x86_pmu.cntval_bits;
2501 cap->bit_width_fixed = x86_pmu.cntval_bits;
2502 cap->events_mask = (unsigned int)x86_pmu.events_maskl;
2503 cap->events_mask_len = x86_pmu.events_mask_len;
2505 EXPORT_SYMBOL_GPL(perf_get_x86_pmu_capability);