2 * Performance events x86 architecture code
4 * Copyright (C) 2008 Thomas Gleixner <tglx@linutronix.de>
5 * Copyright (C) 2008-2009 Red Hat, Inc., Ingo Molnar
6 * Copyright (C) 2009 Jaswinder Singh Rajput
7 * Copyright (C) 2009 Advanced Micro Devices, Inc., Robert Richter
8 * Copyright (C) 2008-2009 Red Hat, Inc., Peter Zijlstra <pzijlstr@redhat.com>
9 * Copyright (C) 2009 Intel Corporation, <markus.t.metzger@intel.com>
10 * Copyright (C) 2009 Google, Inc., Stephane Eranian
12 * For licencing details see kernel-base/COPYING
15 #include <linux/perf_event.h>
16 #include <linux/capability.h>
17 #include <linux/notifier.h>
18 #include <linux/hardirq.h>
19 #include <linux/kprobes.h>
20 #include <linux/module.h>
21 #include <linux/kdebug.h>
22 #include <linux/sched.h>
23 #include <linux/uaccess.h>
24 #include <linux/slab.h>
25 #include <linux/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 DEFINE_MUTEX(pmc_reserve_mutex);
140 #ifdef CONFIG_X86_LOCAL_APIC
142 static bool reserve_pmc_hardware(void)
146 for (i = 0; i < x86_pmu.num_counters; i++) {
147 if (!reserve_perfctr_nmi(x86_pmu_event_addr(i)))
151 for (i = 0; i < x86_pmu.num_counters; i++) {
152 if (!reserve_evntsel_nmi(x86_pmu_config_addr(i)))
159 for (i--; i >= 0; i--)
160 release_evntsel_nmi(x86_pmu_config_addr(i));
162 i = x86_pmu.num_counters;
165 for (i--; i >= 0; i--)
166 release_perfctr_nmi(x86_pmu_event_addr(i));
171 static void release_pmc_hardware(void)
175 for (i = 0; i < x86_pmu.num_counters; i++) {
176 release_perfctr_nmi(x86_pmu_event_addr(i));
177 release_evntsel_nmi(x86_pmu_config_addr(i));
183 static bool reserve_pmc_hardware(void) { return true; }
184 static void release_pmc_hardware(void) {}
188 static bool check_hw_exists(void)
190 u64 val, val_fail, val_new= ~0;
191 int i, reg, reg_fail, ret = 0;
196 * Check to see if the BIOS enabled any of the counters, if so
199 for (i = 0; i < x86_pmu.num_counters; i++) {
200 reg = x86_pmu_config_addr(i);
201 ret = rdmsrl_safe(reg, &val);
204 if (val & ARCH_PERFMON_EVENTSEL_ENABLE) {
213 if (x86_pmu.num_counters_fixed) {
214 reg = MSR_ARCH_PERFMON_FIXED_CTR_CTRL;
215 ret = rdmsrl_safe(reg, &val);
218 for (i = 0; i < x86_pmu.num_counters_fixed; i++) {
219 if (val & (0x03 << i*4)) {
228 * If all the counters are enabled, the below test will always
229 * fail. The tools will also become useless in this scenario.
230 * Just fail and disable the hardware counters.
233 if (reg_safe == -1) {
239 * Read the current value, change it and read it back to see if it
240 * matches, this is needed to detect certain hardware emulators
241 * (qemu/kvm) that don't trap on the MSR access and always return 0s.
243 reg = x86_pmu_event_addr(reg_safe);
244 if (rdmsrl_safe(reg, &val))
247 ret = wrmsrl_safe(reg, val);
248 ret |= rdmsrl_safe(reg, &val_new);
249 if (ret || val != val_new)
253 * We still allow the PMU driver to operate:
256 printk(KERN_CONT "Broken BIOS detected, complain to your hardware vendor.\n");
257 printk(KERN_ERR FW_BUG "the BIOS has corrupted hw-PMU resources (MSR %x is %Lx)\n", reg_fail, val_fail);
263 printk(KERN_CONT "Broken PMU hardware detected, using software events only.\n");
264 printk("%sFailed to access perfctr msr (MSR %x is %Lx)\n",
265 boot_cpu_has(X86_FEATURE_HYPERVISOR) ? KERN_INFO : KERN_ERR,
271 static void hw_perf_event_destroy(struct perf_event *event)
273 if (atomic_dec_and_mutex_lock(&active_events, &pmc_reserve_mutex)) {
274 release_pmc_hardware();
275 release_ds_buffers();
276 mutex_unlock(&pmc_reserve_mutex);
280 void hw_perf_lbr_event_destroy(struct perf_event *event)
282 hw_perf_event_destroy(event);
284 /* undo the lbr/bts event accounting */
285 x86_del_exclusive(x86_lbr_exclusive_lbr);
288 static inline int x86_pmu_initialized(void)
290 return x86_pmu.handle_irq != NULL;
294 set_ext_hw_attr(struct hw_perf_event *hwc, struct perf_event *event)
296 struct perf_event_attr *attr = &event->attr;
297 unsigned int cache_type, cache_op, cache_result;
300 config = attr->config;
302 cache_type = (config >> 0) & 0xff;
303 if (cache_type >= PERF_COUNT_HW_CACHE_MAX)
306 cache_op = (config >> 8) & 0xff;
307 if (cache_op >= PERF_COUNT_HW_CACHE_OP_MAX)
310 cache_result = (config >> 16) & 0xff;
311 if (cache_result >= PERF_COUNT_HW_CACHE_RESULT_MAX)
314 val = hw_cache_event_ids[cache_type][cache_op][cache_result];
323 attr->config1 = hw_cache_extra_regs[cache_type][cache_op][cache_result];
324 return x86_pmu_extra_regs(val, event);
328 * Check if we can create event of a certain type (that no conflicting events
331 int x86_add_exclusive(unsigned int what)
335 if (atomic_inc_not_zero(&x86_pmu.lbr_exclusive[what]))
338 mutex_lock(&pmc_reserve_mutex);
339 for (i = 0; i < ARRAY_SIZE(x86_pmu.lbr_exclusive); i++)
340 if (i != what && atomic_read(&x86_pmu.lbr_exclusive[i]))
343 atomic_inc(&x86_pmu.lbr_exclusive[what]);
347 mutex_unlock(&pmc_reserve_mutex);
351 void x86_del_exclusive(unsigned int what)
353 atomic_dec(&x86_pmu.lbr_exclusive[what]);
356 int x86_setup_perfctr(struct perf_event *event)
358 struct perf_event_attr *attr = &event->attr;
359 struct hw_perf_event *hwc = &event->hw;
362 if (!is_sampling_event(event)) {
363 hwc->sample_period = x86_pmu.max_period;
364 hwc->last_period = hwc->sample_period;
365 local64_set(&hwc->period_left, hwc->sample_period);
368 if (attr->type == PERF_TYPE_RAW)
369 return x86_pmu_extra_regs(event->attr.config, event);
371 if (attr->type == PERF_TYPE_HW_CACHE)
372 return set_ext_hw_attr(hwc, event);
374 if (attr->config >= x86_pmu.max_events)
380 config = x86_pmu.event_map(attr->config);
391 if (attr->config == PERF_COUNT_HW_BRANCH_INSTRUCTIONS &&
392 !attr->freq && hwc->sample_period == 1) {
393 /* BTS is not supported by this architecture. */
394 if (!x86_pmu.bts_active)
397 /* BTS is currently only allowed for user-mode. */
398 if (!attr->exclude_kernel)
401 /* disallow bts if conflicting events are present */
402 if (x86_add_exclusive(x86_lbr_exclusive_lbr))
405 event->destroy = hw_perf_lbr_event_destroy;
408 hwc->config |= config;
414 * check that branch_sample_type is compatible with
415 * settings needed for precise_ip > 1 which implies
416 * using the LBR to capture ALL taken branches at the
417 * priv levels of the measurement
419 static inline int precise_br_compat(struct perf_event *event)
421 u64 m = event->attr.branch_sample_type;
424 /* must capture all branches */
425 if (!(m & PERF_SAMPLE_BRANCH_ANY))
428 m &= PERF_SAMPLE_BRANCH_KERNEL | PERF_SAMPLE_BRANCH_USER;
430 if (!event->attr.exclude_user)
431 b |= PERF_SAMPLE_BRANCH_USER;
433 if (!event->attr.exclude_kernel)
434 b |= PERF_SAMPLE_BRANCH_KERNEL;
437 * ignore PERF_SAMPLE_BRANCH_HV, not supported on x86
443 int x86_pmu_hw_config(struct perf_event *event)
445 if (event->attr.precise_ip) {
448 /* Support for constant skid */
449 if (x86_pmu.pebs_active && !x86_pmu.pebs_broken) {
452 /* Support for IP fixup */
453 if (x86_pmu.lbr_nr || x86_pmu.intel_cap.pebs_format >= 2)
457 if (event->attr.precise_ip > precise)
461 * check that PEBS LBR correction does not conflict with
462 * whatever the user is asking with attr->branch_sample_type
464 if (event->attr.precise_ip > 1 && x86_pmu.intel_cap.pebs_format < 2) {
465 u64 *br_type = &event->attr.branch_sample_type;
467 if (has_branch_stack(event)) {
468 if (!precise_br_compat(event))
471 /* branch_sample_type is compatible */
475 * user did not specify branch_sample_type
477 * For PEBS fixups, we capture all
478 * the branches at the priv level of the
481 *br_type = PERF_SAMPLE_BRANCH_ANY;
483 if (!event->attr.exclude_user)
484 *br_type |= PERF_SAMPLE_BRANCH_USER;
486 if (!event->attr.exclude_kernel)
487 *br_type |= PERF_SAMPLE_BRANCH_KERNEL;
491 if (event->attr.branch_sample_type & PERF_SAMPLE_BRANCH_CALL_STACK)
492 event->attach_state |= PERF_ATTACH_TASK_DATA;
496 * (keep 'enabled' bit clear for now)
498 event->hw.config = ARCH_PERFMON_EVENTSEL_INT;
501 * Count user and OS events unless requested not to
503 if (!event->attr.exclude_user)
504 event->hw.config |= ARCH_PERFMON_EVENTSEL_USR;
505 if (!event->attr.exclude_kernel)
506 event->hw.config |= ARCH_PERFMON_EVENTSEL_OS;
508 if (event->attr.type == PERF_TYPE_RAW)
509 event->hw.config |= event->attr.config & X86_RAW_EVENT_MASK;
511 if (event->attr.sample_period && x86_pmu.limit_period) {
512 if (x86_pmu.limit_period(event, event->attr.sample_period) >
513 event->attr.sample_period)
517 return x86_setup_perfctr(event);
521 * Setup the hardware configuration for a given attr_type
523 static int __x86_pmu_event_init(struct perf_event *event)
527 if (!x86_pmu_initialized())
531 if (!atomic_inc_not_zero(&active_events)) {
532 mutex_lock(&pmc_reserve_mutex);
533 if (atomic_read(&active_events) == 0) {
534 if (!reserve_pmc_hardware())
537 reserve_ds_buffers();
540 atomic_inc(&active_events);
541 mutex_unlock(&pmc_reserve_mutex);
546 event->destroy = hw_perf_event_destroy;
549 event->hw.last_cpu = -1;
550 event->hw.last_tag = ~0ULL;
553 event->hw.extra_reg.idx = EXTRA_REG_NONE;
554 event->hw.branch_reg.idx = EXTRA_REG_NONE;
556 return x86_pmu.hw_config(event);
559 void x86_pmu_disable_all(void)
561 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
564 for (idx = 0; idx < x86_pmu.num_counters; idx++) {
567 if (!test_bit(idx, cpuc->active_mask))
569 rdmsrl(x86_pmu_config_addr(idx), val);
570 if (!(val & ARCH_PERFMON_EVENTSEL_ENABLE))
572 val &= ~ARCH_PERFMON_EVENTSEL_ENABLE;
573 wrmsrl(x86_pmu_config_addr(idx), val);
577 static void x86_pmu_disable(struct pmu *pmu)
579 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
581 if (!x86_pmu_initialized())
591 x86_pmu.disable_all();
594 void x86_pmu_enable_all(int added)
596 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
599 for (idx = 0; idx < x86_pmu.num_counters; idx++) {
600 struct hw_perf_event *hwc = &cpuc->events[idx]->hw;
602 if (!test_bit(idx, cpuc->active_mask))
605 __x86_pmu_enable_event(hwc, ARCH_PERFMON_EVENTSEL_ENABLE);
609 static struct pmu pmu;
611 static inline int is_x86_event(struct perf_event *event)
613 return event->pmu == &pmu;
617 * Event scheduler state:
619 * Assign events iterating over all events and counters, beginning
620 * with events with least weights first. Keep the current iterator
621 * state in struct sched_state.
625 int event; /* event index */
626 int counter; /* counter index */
627 int unassigned; /* number of events to be assigned left */
628 int nr_gp; /* number of GP counters used */
629 unsigned long used[BITS_TO_LONGS(X86_PMC_IDX_MAX)];
632 /* Total max is X86_PMC_IDX_MAX, but we are O(n!) limited */
633 #define SCHED_STATES_MAX 2
640 struct event_constraint **constraints;
641 struct sched_state state;
642 struct sched_state saved[SCHED_STATES_MAX];
646 * Initialize interator that runs through all events and counters.
648 static void perf_sched_init(struct perf_sched *sched, struct event_constraint **constraints,
649 int num, int wmin, int wmax, int gpmax)
653 memset(sched, 0, sizeof(*sched));
654 sched->max_events = num;
655 sched->max_weight = wmax;
656 sched->max_gp = gpmax;
657 sched->constraints = constraints;
659 for (idx = 0; idx < num; idx++) {
660 if (constraints[idx]->weight == wmin)
664 sched->state.event = idx; /* start with min weight */
665 sched->state.weight = wmin;
666 sched->state.unassigned = num;
669 static void perf_sched_save_state(struct perf_sched *sched)
671 if (WARN_ON_ONCE(sched->saved_states >= SCHED_STATES_MAX))
674 sched->saved[sched->saved_states] = sched->state;
675 sched->saved_states++;
678 static bool perf_sched_restore_state(struct perf_sched *sched)
680 if (!sched->saved_states)
683 sched->saved_states--;
684 sched->state = sched->saved[sched->saved_states];
686 /* continue with next counter: */
687 clear_bit(sched->state.counter++, sched->state.used);
693 * Select a counter for the current event to schedule. Return true on
696 static bool __perf_sched_find_counter(struct perf_sched *sched)
698 struct event_constraint *c;
701 if (!sched->state.unassigned)
704 if (sched->state.event >= sched->max_events)
707 c = sched->constraints[sched->state.event];
708 /* Prefer fixed purpose counters */
709 if (c->idxmsk64 & (~0ULL << INTEL_PMC_IDX_FIXED)) {
710 idx = INTEL_PMC_IDX_FIXED;
711 for_each_set_bit_from(idx, c->idxmsk, X86_PMC_IDX_MAX) {
712 if (!__test_and_set_bit(idx, sched->state.used))
717 /* Grab the first unused counter starting with idx */
718 idx = sched->state.counter;
719 for_each_set_bit_from(idx, c->idxmsk, INTEL_PMC_IDX_FIXED) {
720 if (!__test_and_set_bit(idx, sched->state.used)) {
721 if (sched->state.nr_gp++ >= sched->max_gp)
731 sched->state.counter = idx;
734 perf_sched_save_state(sched);
739 static bool perf_sched_find_counter(struct perf_sched *sched)
741 while (!__perf_sched_find_counter(sched)) {
742 if (!perf_sched_restore_state(sched))
750 * Go through all unassigned events and find the next one to schedule.
751 * Take events with the least weight first. Return true on success.
753 static bool perf_sched_next_event(struct perf_sched *sched)
755 struct event_constraint *c;
757 if (!sched->state.unassigned || !--sched->state.unassigned)
762 sched->state.event++;
763 if (sched->state.event >= sched->max_events) {
765 sched->state.event = 0;
766 sched->state.weight++;
767 if (sched->state.weight > sched->max_weight)
770 c = sched->constraints[sched->state.event];
771 } while (c->weight != sched->state.weight);
773 sched->state.counter = 0; /* start with first counter */
779 * Assign a counter for each event.
781 int perf_assign_events(struct event_constraint **constraints, int n,
782 int wmin, int wmax, int gpmax, int *assign)
784 struct perf_sched sched;
786 perf_sched_init(&sched, constraints, n, wmin, wmax, gpmax);
789 if (!perf_sched_find_counter(&sched))
792 assign[sched.state.event] = sched.state.counter;
793 } while (perf_sched_next_event(&sched));
795 return sched.state.unassigned;
797 EXPORT_SYMBOL_GPL(perf_assign_events);
799 int x86_schedule_events(struct cpu_hw_events *cpuc, int n, int *assign)
801 struct event_constraint *c;
802 unsigned long used_mask[BITS_TO_LONGS(X86_PMC_IDX_MAX)];
803 struct perf_event *e;
804 int i, wmin, wmax, unsched = 0;
805 struct hw_perf_event *hwc;
807 bitmap_zero(used_mask, X86_PMC_IDX_MAX);
809 if (x86_pmu.start_scheduling)
810 x86_pmu.start_scheduling(cpuc);
812 for (i = 0, wmin = X86_PMC_IDX_MAX, wmax = 0; i < n; i++) {
813 cpuc->event_constraint[i] = NULL;
814 c = x86_pmu.get_event_constraints(cpuc, i, cpuc->event_list[i]);
815 cpuc->event_constraint[i] = c;
817 wmin = min(wmin, c->weight);
818 wmax = max(wmax, c->weight);
822 * fastpath, try to reuse previous register
824 for (i = 0; i < n; i++) {
825 hwc = &cpuc->event_list[i]->hw;
826 c = cpuc->event_constraint[i];
832 /* constraint still honored */
833 if (!test_bit(hwc->idx, c->idxmsk))
836 /* not already used */
837 if (test_bit(hwc->idx, used_mask))
840 __set_bit(hwc->idx, used_mask);
842 assign[i] = hwc->idx;
847 int gpmax = x86_pmu.num_counters;
850 * Do not allow scheduling of more than half the available
853 * This helps avoid counter starvation of sibling thread by
854 * ensuring at most half the counters cannot be in exclusive
855 * mode. There is no designated counters for the limits. Any
856 * N/2 counters can be used. This helps with events with
857 * specific counter constraints.
859 if (is_ht_workaround_enabled() && !cpuc->is_fake &&
860 READ_ONCE(cpuc->excl_cntrs->exclusive_present))
863 unsched = perf_assign_events(cpuc->event_constraint, n, wmin,
864 wmax, gpmax, assign);
868 * In case of success (unsched = 0), mark events as committed,
869 * so we do not put_constraint() in case new events are added
870 * and fail to be scheduled
872 * We invoke the lower level commit callback to lock the resource
874 * We do not need to do all of this in case we are called to
875 * validate an event group (assign == NULL)
877 if (!unsched && assign) {
878 for (i = 0; i < n; i++) {
879 e = cpuc->event_list[i];
880 e->hw.flags |= PERF_X86_EVENT_COMMITTED;
881 if (x86_pmu.commit_scheduling)
882 x86_pmu.commit_scheduling(cpuc, i, assign[i]);
886 if (!assign || unsched) {
888 for (i = 0; i < n; i++) {
889 e = cpuc->event_list[i];
891 * do not put_constraint() on comitted events,
892 * because they are good to go
894 if ((e->hw.flags & PERF_X86_EVENT_COMMITTED))
898 * release events that failed scheduling
900 if (x86_pmu.put_event_constraints)
901 x86_pmu.put_event_constraints(cpuc, e);
905 if (x86_pmu.stop_scheduling)
906 x86_pmu.stop_scheduling(cpuc);
908 return unsched ? -EINVAL : 0;
912 * dogrp: true if must collect siblings events (group)
913 * returns total number of events and error code
915 static int collect_events(struct cpu_hw_events *cpuc, struct perf_event *leader, bool dogrp)
917 struct perf_event *event;
920 max_count = x86_pmu.num_counters + x86_pmu.num_counters_fixed;
922 /* current number of events already accepted */
925 if (is_x86_event(leader)) {
928 cpuc->event_list[n] = leader;
934 list_for_each_entry(event, &leader->sibling_list, group_entry) {
935 if (!is_x86_event(event) ||
936 event->state <= PERF_EVENT_STATE_OFF)
942 cpuc->event_list[n] = event;
948 static inline void x86_assign_hw_event(struct perf_event *event,
949 struct cpu_hw_events *cpuc, int i)
951 struct hw_perf_event *hwc = &event->hw;
953 hwc->idx = cpuc->assign[i];
954 hwc->last_cpu = smp_processor_id();
955 hwc->last_tag = ++cpuc->tags[i];
957 if (hwc->idx == INTEL_PMC_IDX_FIXED_BTS) {
958 hwc->config_base = 0;
960 } else if (hwc->idx >= INTEL_PMC_IDX_FIXED) {
961 hwc->config_base = MSR_ARCH_PERFMON_FIXED_CTR_CTRL;
962 hwc->event_base = MSR_ARCH_PERFMON_FIXED_CTR0 + (hwc->idx - INTEL_PMC_IDX_FIXED);
963 hwc->event_base_rdpmc = (hwc->idx - INTEL_PMC_IDX_FIXED) | 1<<30;
965 hwc->config_base = x86_pmu_config_addr(hwc->idx);
966 hwc->event_base = x86_pmu_event_addr(hwc->idx);
967 hwc->event_base_rdpmc = x86_pmu_rdpmc_index(hwc->idx);
971 static inline int match_prev_assignment(struct hw_perf_event *hwc,
972 struct cpu_hw_events *cpuc,
975 return hwc->idx == cpuc->assign[i] &&
976 hwc->last_cpu == smp_processor_id() &&
977 hwc->last_tag == cpuc->tags[i];
980 static void x86_pmu_start(struct perf_event *event, int flags);
982 static void x86_pmu_enable(struct pmu *pmu)
984 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
985 struct perf_event *event;
986 struct hw_perf_event *hwc;
987 int i, added = cpuc->n_added;
989 if (!x86_pmu_initialized())
996 int n_running = cpuc->n_events - cpuc->n_added;
998 * apply assignment obtained either from
999 * hw_perf_group_sched_in() or x86_pmu_enable()
1001 * step1: save events moving to new counters
1003 for (i = 0; i < n_running; i++) {
1004 event = cpuc->event_list[i];
1008 * we can avoid reprogramming counter if:
1009 * - assigned same counter as last time
1010 * - running on same CPU as last time
1011 * - no other event has used the counter since
1013 if (hwc->idx == -1 ||
1014 match_prev_assignment(hwc, cpuc, i))
1018 * Ensure we don't accidentally enable a stopped
1019 * counter simply because we rescheduled.
1021 if (hwc->state & PERF_HES_STOPPED)
1022 hwc->state |= PERF_HES_ARCH;
1024 x86_pmu_stop(event, PERF_EF_UPDATE);
1028 * step2: reprogram moved events into new counters
1030 for (i = 0; i < cpuc->n_events; i++) {
1031 event = cpuc->event_list[i];
1034 if (!match_prev_assignment(hwc, cpuc, i))
1035 x86_assign_hw_event(event, cpuc, i);
1036 else if (i < n_running)
1039 if (hwc->state & PERF_HES_ARCH)
1042 x86_pmu_start(event, PERF_EF_RELOAD);
1045 perf_events_lapic_init();
1051 x86_pmu.enable_all(added);
1054 static DEFINE_PER_CPU(u64 [X86_PMC_IDX_MAX], pmc_prev_left);
1057 * Set the next IRQ period, based on the hwc->period_left value.
1058 * To be called with the event disabled in hw:
1060 int x86_perf_event_set_period(struct perf_event *event)
1062 struct hw_perf_event *hwc = &event->hw;
1063 s64 left = local64_read(&hwc->period_left);
1064 s64 period = hwc->sample_period;
1065 int ret = 0, idx = hwc->idx;
1067 if (idx == INTEL_PMC_IDX_FIXED_BTS)
1071 * If we are way outside a reasonable range then just skip forward:
1073 if (unlikely(left <= -period)) {
1075 local64_set(&hwc->period_left, left);
1076 hwc->last_period = period;
1080 if (unlikely(left <= 0)) {
1082 local64_set(&hwc->period_left, left);
1083 hwc->last_period = period;
1087 * Quirk: certain CPUs dont like it if just 1 hw_event is left:
1089 if (unlikely(left < 2))
1092 if (left > x86_pmu.max_period)
1093 left = x86_pmu.max_period;
1095 if (x86_pmu.limit_period)
1096 left = x86_pmu.limit_period(event, left);
1098 per_cpu(pmc_prev_left[idx], smp_processor_id()) = left;
1101 * The hw event starts counting from this event offset,
1102 * mark it to be able to extra future deltas:
1104 local64_set(&hwc->prev_count, (u64)-left);
1106 wrmsrl(hwc->event_base, (u64)(-left) & x86_pmu.cntval_mask);
1109 * Due to erratum on certan cpu we need
1110 * a second write to be sure the register
1111 * is updated properly
1113 if (x86_pmu.perfctr_second_write) {
1114 wrmsrl(hwc->event_base,
1115 (u64)(-left) & x86_pmu.cntval_mask);
1118 perf_event_update_userpage(event);
1123 void x86_pmu_enable_event(struct perf_event *event)
1125 if (__this_cpu_read(cpu_hw_events.enabled))
1126 __x86_pmu_enable_event(&event->hw,
1127 ARCH_PERFMON_EVENTSEL_ENABLE);
1131 * Add a single event to the PMU.
1133 * The event is added to the group of enabled events
1134 * but only if it can be scehduled with existing events.
1136 static int x86_pmu_add(struct perf_event *event, int flags)
1138 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1139 struct hw_perf_event *hwc;
1140 int assign[X86_PMC_IDX_MAX];
1145 n0 = cpuc->n_events;
1146 ret = n = collect_events(cpuc, event, false);
1150 hwc->state = PERF_HES_UPTODATE | PERF_HES_STOPPED;
1151 if (!(flags & PERF_EF_START))
1152 hwc->state |= PERF_HES_ARCH;
1155 * If group events scheduling transaction was started,
1156 * skip the schedulability test here, it will be performed
1157 * at commit time (->commit_txn) as a whole.
1159 if (cpuc->group_flag & PERF_EVENT_TXN)
1162 ret = x86_pmu.schedule_events(cpuc, n, assign);
1166 * copy new assignment, now we know it is possible
1167 * will be used by hw_perf_enable()
1169 memcpy(cpuc->assign, assign, n*sizeof(int));
1173 * Commit the collect_events() state. See x86_pmu_del() and
1177 cpuc->n_added += n - n0;
1178 cpuc->n_txn += n - n0;
1185 static void x86_pmu_start(struct perf_event *event, int flags)
1187 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1188 int idx = event->hw.idx;
1190 if (WARN_ON_ONCE(!(event->hw.state & PERF_HES_STOPPED)))
1193 if (WARN_ON_ONCE(idx == -1))
1196 if (flags & PERF_EF_RELOAD) {
1197 WARN_ON_ONCE(!(event->hw.state & PERF_HES_UPTODATE));
1198 x86_perf_event_set_period(event);
1201 event->hw.state = 0;
1203 cpuc->events[idx] = event;
1204 __set_bit(idx, cpuc->active_mask);
1205 __set_bit(idx, cpuc->running);
1206 x86_pmu.enable(event);
1207 perf_event_update_userpage(event);
1210 void perf_event_print_debug(void)
1212 u64 ctrl, status, overflow, pmc_ctrl, pmc_count, prev_left, fixed;
1214 struct cpu_hw_events *cpuc;
1215 unsigned long flags;
1218 if (!x86_pmu.num_counters)
1221 local_irq_save(flags);
1223 cpu = smp_processor_id();
1224 cpuc = &per_cpu(cpu_hw_events, cpu);
1226 if (x86_pmu.version >= 2) {
1227 rdmsrl(MSR_CORE_PERF_GLOBAL_CTRL, ctrl);
1228 rdmsrl(MSR_CORE_PERF_GLOBAL_STATUS, status);
1229 rdmsrl(MSR_CORE_PERF_GLOBAL_OVF_CTRL, overflow);
1230 rdmsrl(MSR_ARCH_PERFMON_FIXED_CTR_CTRL, fixed);
1233 pr_info("CPU#%d: ctrl: %016llx\n", cpu, ctrl);
1234 pr_info("CPU#%d: status: %016llx\n", cpu, status);
1235 pr_info("CPU#%d: overflow: %016llx\n", cpu, overflow);
1236 pr_info("CPU#%d: fixed: %016llx\n", cpu, fixed);
1237 if (x86_pmu.pebs_constraints) {
1238 rdmsrl(MSR_IA32_PEBS_ENABLE, pebs);
1239 pr_info("CPU#%d: pebs: %016llx\n", cpu, pebs);
1241 if (x86_pmu.lbr_nr) {
1242 rdmsrl(MSR_IA32_DEBUGCTLMSR, debugctl);
1243 pr_info("CPU#%d: debugctl: %016llx\n", cpu, debugctl);
1246 pr_info("CPU#%d: active: %016llx\n", cpu, *(u64 *)cpuc->active_mask);
1248 for (idx = 0; idx < x86_pmu.num_counters; idx++) {
1249 rdmsrl(x86_pmu_config_addr(idx), pmc_ctrl);
1250 rdmsrl(x86_pmu_event_addr(idx), pmc_count);
1252 prev_left = per_cpu(pmc_prev_left[idx], cpu);
1254 pr_info("CPU#%d: gen-PMC%d ctrl: %016llx\n",
1255 cpu, idx, pmc_ctrl);
1256 pr_info("CPU#%d: gen-PMC%d count: %016llx\n",
1257 cpu, idx, pmc_count);
1258 pr_info("CPU#%d: gen-PMC%d left: %016llx\n",
1259 cpu, idx, prev_left);
1261 for (idx = 0; idx < x86_pmu.num_counters_fixed; idx++) {
1262 rdmsrl(MSR_ARCH_PERFMON_FIXED_CTR0 + idx, pmc_count);
1264 pr_info("CPU#%d: fixed-PMC%d count: %016llx\n",
1265 cpu, idx, pmc_count);
1267 local_irq_restore(flags);
1270 void x86_pmu_stop(struct perf_event *event, int flags)
1272 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1273 struct hw_perf_event *hwc = &event->hw;
1275 if (__test_and_clear_bit(hwc->idx, cpuc->active_mask)) {
1276 x86_pmu.disable(event);
1277 cpuc->events[hwc->idx] = NULL;
1278 WARN_ON_ONCE(hwc->state & PERF_HES_STOPPED);
1279 hwc->state |= PERF_HES_STOPPED;
1282 if ((flags & PERF_EF_UPDATE) && !(hwc->state & PERF_HES_UPTODATE)) {
1284 * Drain the remaining delta count out of a event
1285 * that we are disabling:
1287 x86_perf_event_update(event);
1288 hwc->state |= PERF_HES_UPTODATE;
1292 static void x86_pmu_del(struct perf_event *event, int flags)
1294 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1298 * event is descheduled
1300 event->hw.flags &= ~PERF_X86_EVENT_COMMITTED;
1303 * If we're called during a txn, we don't need to do anything.
1304 * The events never got scheduled and ->cancel_txn will truncate
1307 * XXX assumes any ->del() called during a TXN will only be on
1308 * an event added during that same TXN.
1310 if (cpuc->group_flag & PERF_EVENT_TXN)
1314 * Not a TXN, therefore cleanup properly.
1316 x86_pmu_stop(event, PERF_EF_UPDATE);
1318 for (i = 0; i < cpuc->n_events; i++) {
1319 if (event == cpuc->event_list[i])
1323 if (WARN_ON_ONCE(i == cpuc->n_events)) /* called ->del() without ->add() ? */
1326 /* If we have a newly added event; make sure to decrease n_added. */
1327 if (i >= cpuc->n_events - cpuc->n_added)
1330 if (x86_pmu.put_event_constraints)
1331 x86_pmu.put_event_constraints(cpuc, event);
1333 /* Delete the array entry. */
1334 while (++i < cpuc->n_events) {
1335 cpuc->event_list[i-1] = cpuc->event_list[i];
1336 cpuc->event_constraint[i-1] = cpuc->event_constraint[i];
1340 perf_event_update_userpage(event);
1343 int x86_pmu_handle_irq(struct pt_regs *regs)
1345 struct perf_sample_data data;
1346 struct cpu_hw_events *cpuc;
1347 struct perf_event *event;
1348 int idx, handled = 0;
1351 cpuc = this_cpu_ptr(&cpu_hw_events);
1354 * Some chipsets need to unmask the LVTPC in a particular spot
1355 * inside the nmi handler. As a result, the unmasking was pushed
1356 * into all the nmi handlers.
1358 * This generic handler doesn't seem to have any issues where the
1359 * unmasking occurs so it was left at the top.
1361 apic_write(APIC_LVTPC, APIC_DM_NMI);
1363 for (idx = 0; idx < x86_pmu.num_counters; idx++) {
1364 if (!test_bit(idx, cpuc->active_mask)) {
1366 * Though we deactivated the counter some cpus
1367 * might still deliver spurious interrupts still
1368 * in flight. Catch them:
1370 if (__test_and_clear_bit(idx, cpuc->running))
1375 event = cpuc->events[idx];
1377 val = x86_perf_event_update(event);
1378 if (val & (1ULL << (x86_pmu.cntval_bits - 1)))
1385 perf_sample_data_init(&data, 0, event->hw.last_period);
1387 if (!x86_perf_event_set_period(event))
1390 if (perf_event_overflow(event, &data, regs))
1391 x86_pmu_stop(event, 0);
1395 inc_irq_stat(apic_perf_irqs);
1400 void perf_events_lapic_init(void)
1402 if (!x86_pmu.apic || !x86_pmu_initialized())
1406 * Always use NMI for PMU
1408 apic_write(APIC_LVTPC, APIC_DM_NMI);
1412 perf_event_nmi_handler(unsigned int cmd, struct pt_regs *regs)
1418 if (!atomic_read(&active_events))
1421 start_clock = sched_clock();
1422 ret = x86_pmu.handle_irq(regs);
1423 finish_clock = sched_clock();
1425 perf_sample_event_took(finish_clock - start_clock);
1429 NOKPROBE_SYMBOL(perf_event_nmi_handler);
1431 struct event_constraint emptyconstraint;
1432 struct event_constraint unconstrained;
1435 x86_pmu_notifier(struct notifier_block *self, unsigned long action, void *hcpu)
1437 unsigned int cpu = (long)hcpu;
1438 struct cpu_hw_events *cpuc = &per_cpu(cpu_hw_events, cpu);
1439 int i, ret = NOTIFY_OK;
1441 switch (action & ~CPU_TASKS_FROZEN) {
1442 case CPU_UP_PREPARE:
1443 for (i = 0 ; i < X86_PERF_KFREE_MAX; i++)
1444 cpuc->kfree_on_online[i] = NULL;
1445 if (x86_pmu.cpu_prepare)
1446 ret = x86_pmu.cpu_prepare(cpu);
1450 if (x86_pmu.cpu_starting)
1451 x86_pmu.cpu_starting(cpu);
1455 for (i = 0 ; i < X86_PERF_KFREE_MAX; i++) {
1456 kfree(cpuc->kfree_on_online[i]);
1457 cpuc->kfree_on_online[i] = NULL;
1462 if (x86_pmu.cpu_dying)
1463 x86_pmu.cpu_dying(cpu);
1466 case CPU_UP_CANCELED:
1468 if (x86_pmu.cpu_dead)
1469 x86_pmu.cpu_dead(cpu);
1479 static void __init pmu_check_apic(void)
1485 pr_info("no APIC, boot with the \"lapic\" boot parameter to force-enable it.\n");
1486 pr_info("no hardware sampling interrupt available.\n");
1489 * If we have a PMU initialized but no APIC
1490 * interrupts, we cannot sample hardware
1491 * events (user-space has to fall back and
1492 * sample via a hrtimer based software event):
1494 pmu.capabilities |= PERF_PMU_CAP_NO_INTERRUPT;
1498 static struct attribute_group x86_pmu_format_group = {
1504 * Remove all undefined events (x86_pmu.event_map(id) == 0)
1505 * out of events_attr attributes.
1507 static void __init filter_events(struct attribute **attrs)
1509 struct device_attribute *d;
1510 struct perf_pmu_events_attr *pmu_attr;
1513 for (i = 0; attrs[i]; i++) {
1514 d = (struct device_attribute *)attrs[i];
1515 pmu_attr = container_of(d, struct perf_pmu_events_attr, attr);
1517 if (pmu_attr->event_str)
1519 if (x86_pmu.event_map(i))
1522 for (j = i; attrs[j]; j++)
1523 attrs[j] = attrs[j + 1];
1525 /* Check the shifted attr. */
1530 /* Merge two pointer arrays */
1531 static __init struct attribute **merge_attr(struct attribute **a, struct attribute **b)
1533 struct attribute **new;
1536 for (j = 0; a[j]; j++)
1538 for (i = 0; b[i]; i++)
1542 new = kmalloc(sizeof(struct attribute *) * j, GFP_KERNEL);
1547 for (i = 0; a[i]; i++)
1549 for (i = 0; b[i]; i++)
1556 ssize_t events_sysfs_show(struct device *dev, struct device_attribute *attr,
1559 struct perf_pmu_events_attr *pmu_attr = \
1560 container_of(attr, struct perf_pmu_events_attr, attr);
1561 u64 config = x86_pmu.event_map(pmu_attr->id);
1563 /* string trumps id */
1564 if (pmu_attr->event_str)
1565 return sprintf(page, "%s", pmu_attr->event_str);
1567 return x86_pmu.events_sysfs_show(page, config);
1570 EVENT_ATTR(cpu-cycles, CPU_CYCLES );
1571 EVENT_ATTR(instructions, INSTRUCTIONS );
1572 EVENT_ATTR(cache-references, CACHE_REFERENCES );
1573 EVENT_ATTR(cache-misses, CACHE_MISSES );
1574 EVENT_ATTR(branch-instructions, BRANCH_INSTRUCTIONS );
1575 EVENT_ATTR(branch-misses, BRANCH_MISSES );
1576 EVENT_ATTR(bus-cycles, BUS_CYCLES );
1577 EVENT_ATTR(stalled-cycles-frontend, STALLED_CYCLES_FRONTEND );
1578 EVENT_ATTR(stalled-cycles-backend, STALLED_CYCLES_BACKEND );
1579 EVENT_ATTR(ref-cycles, REF_CPU_CYCLES );
1581 static struct attribute *empty_attrs;
1583 static struct attribute *events_attr[] = {
1584 EVENT_PTR(CPU_CYCLES),
1585 EVENT_PTR(INSTRUCTIONS),
1586 EVENT_PTR(CACHE_REFERENCES),
1587 EVENT_PTR(CACHE_MISSES),
1588 EVENT_PTR(BRANCH_INSTRUCTIONS),
1589 EVENT_PTR(BRANCH_MISSES),
1590 EVENT_PTR(BUS_CYCLES),
1591 EVENT_PTR(STALLED_CYCLES_FRONTEND),
1592 EVENT_PTR(STALLED_CYCLES_BACKEND),
1593 EVENT_PTR(REF_CPU_CYCLES),
1597 static struct attribute_group x86_pmu_events_group = {
1599 .attrs = events_attr,
1602 ssize_t x86_event_sysfs_show(char *page, u64 config, u64 event)
1604 u64 umask = (config & ARCH_PERFMON_EVENTSEL_UMASK) >> 8;
1605 u64 cmask = (config & ARCH_PERFMON_EVENTSEL_CMASK) >> 24;
1606 bool edge = (config & ARCH_PERFMON_EVENTSEL_EDGE);
1607 bool pc = (config & ARCH_PERFMON_EVENTSEL_PIN_CONTROL);
1608 bool any = (config & ARCH_PERFMON_EVENTSEL_ANY);
1609 bool inv = (config & ARCH_PERFMON_EVENTSEL_INV);
1613 * We have whole page size to spend and just little data
1614 * to write, so we can safely use sprintf.
1616 ret = sprintf(page, "event=0x%02llx", event);
1619 ret += sprintf(page + ret, ",umask=0x%02llx", umask);
1622 ret += sprintf(page + ret, ",edge");
1625 ret += sprintf(page + ret, ",pc");
1628 ret += sprintf(page + ret, ",any");
1631 ret += sprintf(page + ret, ",inv");
1634 ret += sprintf(page + ret, ",cmask=0x%02llx", cmask);
1636 ret += sprintf(page + ret, "\n");
1641 static int __init init_hw_perf_events(void)
1643 struct x86_pmu_quirk *quirk;
1646 pr_info("Performance Events: ");
1648 switch (boot_cpu_data.x86_vendor) {
1649 case X86_VENDOR_INTEL:
1650 err = intel_pmu_init();
1652 case X86_VENDOR_AMD:
1653 err = amd_pmu_init();
1659 pr_cont("no PMU driver, software events only.\n");
1665 /* sanity check that the hardware exists or is emulated */
1666 if (!check_hw_exists())
1669 pr_cont("%s PMU driver.\n", x86_pmu.name);
1671 x86_pmu.attr_rdpmc = 1; /* enable userspace RDPMC usage by default */
1673 for (quirk = x86_pmu.quirks; quirk; quirk = quirk->next)
1676 if (!x86_pmu.intel_ctrl)
1677 x86_pmu.intel_ctrl = (1 << x86_pmu.num_counters) - 1;
1679 perf_events_lapic_init();
1680 register_nmi_handler(NMI_LOCAL, perf_event_nmi_handler, 0, "PMI");
1682 unconstrained = (struct event_constraint)
1683 __EVENT_CONSTRAINT(0, (1ULL << x86_pmu.num_counters) - 1,
1684 0, x86_pmu.num_counters, 0, 0);
1686 x86_pmu_format_group.attrs = x86_pmu.format_attrs;
1688 if (x86_pmu.event_attrs)
1689 x86_pmu_events_group.attrs = x86_pmu.event_attrs;
1691 if (!x86_pmu.events_sysfs_show)
1692 x86_pmu_events_group.attrs = &empty_attrs;
1694 filter_events(x86_pmu_events_group.attrs);
1696 if (x86_pmu.cpu_events) {
1697 struct attribute **tmp;
1699 tmp = merge_attr(x86_pmu_events_group.attrs, x86_pmu.cpu_events);
1701 x86_pmu_events_group.attrs = tmp;
1704 pr_info("... version: %d\n", x86_pmu.version);
1705 pr_info("... bit width: %d\n", x86_pmu.cntval_bits);
1706 pr_info("... generic registers: %d\n", x86_pmu.num_counters);
1707 pr_info("... value mask: %016Lx\n", x86_pmu.cntval_mask);
1708 pr_info("... max period: %016Lx\n", x86_pmu.max_period);
1709 pr_info("... fixed-purpose events: %d\n", x86_pmu.num_counters_fixed);
1710 pr_info("... event mask: %016Lx\n", x86_pmu.intel_ctrl);
1712 perf_pmu_register(&pmu, "cpu", PERF_TYPE_RAW);
1713 perf_cpu_notifier(x86_pmu_notifier);
1717 early_initcall(init_hw_perf_events);
1719 static inline void x86_pmu_read(struct perf_event *event)
1721 x86_perf_event_update(event);
1725 * Start group events scheduling transaction
1726 * Set the flag to make pmu::enable() not perform the
1727 * schedulability test, it will be performed at commit time
1729 static void x86_pmu_start_txn(struct pmu *pmu)
1731 perf_pmu_disable(pmu);
1732 __this_cpu_or(cpu_hw_events.group_flag, PERF_EVENT_TXN);
1733 __this_cpu_write(cpu_hw_events.n_txn, 0);
1737 * Stop group events scheduling transaction
1738 * Clear the flag and pmu::enable() will perform the
1739 * schedulability test.
1741 static void x86_pmu_cancel_txn(struct pmu *pmu)
1743 __this_cpu_and(cpu_hw_events.group_flag, ~PERF_EVENT_TXN);
1745 * Truncate collected array by the number of events added in this
1746 * transaction. See x86_pmu_add() and x86_pmu_*_txn().
1748 __this_cpu_sub(cpu_hw_events.n_added, __this_cpu_read(cpu_hw_events.n_txn));
1749 __this_cpu_sub(cpu_hw_events.n_events, __this_cpu_read(cpu_hw_events.n_txn));
1750 perf_pmu_enable(pmu);
1754 * Commit group events scheduling transaction
1755 * Perform the group schedulability test as a whole
1756 * Return 0 if success
1758 * Does not cancel the transaction on failure; expects the caller to do this.
1760 static int x86_pmu_commit_txn(struct pmu *pmu)
1762 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1763 int assign[X86_PMC_IDX_MAX];
1768 if (!x86_pmu_initialized())
1771 ret = x86_pmu.schedule_events(cpuc, n, assign);
1776 * copy new assignment, now we know it is possible
1777 * will be used by hw_perf_enable()
1779 memcpy(cpuc->assign, assign, n*sizeof(int));
1781 cpuc->group_flag &= ~PERF_EVENT_TXN;
1782 perf_pmu_enable(pmu);
1786 * a fake_cpuc is used to validate event groups. Due to
1787 * the extra reg logic, we need to also allocate a fake
1788 * per_core and per_cpu structure. Otherwise, group events
1789 * using extra reg may conflict without the kernel being
1790 * able to catch this when the last event gets added to
1793 static void free_fake_cpuc(struct cpu_hw_events *cpuc)
1795 kfree(cpuc->shared_regs);
1799 static struct cpu_hw_events *allocate_fake_cpuc(void)
1801 struct cpu_hw_events *cpuc;
1802 int cpu = raw_smp_processor_id();
1804 cpuc = kzalloc(sizeof(*cpuc), GFP_KERNEL);
1806 return ERR_PTR(-ENOMEM);
1808 /* only needed, if we have extra_regs */
1809 if (x86_pmu.extra_regs) {
1810 cpuc->shared_regs = allocate_shared_regs(cpu);
1811 if (!cpuc->shared_regs)
1817 free_fake_cpuc(cpuc);
1818 return ERR_PTR(-ENOMEM);
1822 * validate that we can schedule this event
1824 static int validate_event(struct perf_event *event)
1826 struct cpu_hw_events *fake_cpuc;
1827 struct event_constraint *c;
1830 fake_cpuc = allocate_fake_cpuc();
1831 if (IS_ERR(fake_cpuc))
1832 return PTR_ERR(fake_cpuc);
1834 c = x86_pmu.get_event_constraints(fake_cpuc, -1, event);
1836 if (!c || !c->weight)
1839 if (x86_pmu.put_event_constraints)
1840 x86_pmu.put_event_constraints(fake_cpuc, event);
1842 free_fake_cpuc(fake_cpuc);
1848 * validate a single event group
1850 * validation include:
1851 * - check events are compatible which each other
1852 * - events do not compete for the same counter
1853 * - number of events <= number of counters
1855 * validation ensures the group can be loaded onto the
1856 * PMU if it was the only group available.
1858 static int validate_group(struct perf_event *event)
1860 struct perf_event *leader = event->group_leader;
1861 struct cpu_hw_events *fake_cpuc;
1862 int ret = -EINVAL, n;
1864 fake_cpuc = allocate_fake_cpuc();
1865 if (IS_ERR(fake_cpuc))
1866 return PTR_ERR(fake_cpuc);
1868 * the event is not yet connected with its
1869 * siblings therefore we must first collect
1870 * existing siblings, then add the new event
1871 * before we can simulate the scheduling
1873 n = collect_events(fake_cpuc, leader, true);
1877 fake_cpuc->n_events = n;
1878 n = collect_events(fake_cpuc, event, false);
1882 fake_cpuc->n_events = n;
1884 ret = x86_pmu.schedule_events(fake_cpuc, n, NULL);
1887 free_fake_cpuc(fake_cpuc);
1891 static int x86_pmu_event_init(struct perf_event *event)
1896 switch (event->attr.type) {
1898 case PERF_TYPE_HARDWARE:
1899 case PERF_TYPE_HW_CACHE:
1906 err = __x86_pmu_event_init(event);
1909 * we temporarily connect event to its pmu
1910 * such that validate_group() can classify
1911 * it as an x86 event using is_x86_event()
1916 if (event->group_leader != event)
1917 err = validate_group(event);
1919 err = validate_event(event);
1925 event->destroy(event);
1928 if (ACCESS_ONCE(x86_pmu.attr_rdpmc))
1929 event->hw.flags |= PERF_X86_EVENT_RDPMC_ALLOWED;
1934 static void refresh_pce(void *ignored)
1937 load_mm_cr4(current->mm);
1940 static void x86_pmu_event_mapped(struct perf_event *event)
1942 if (!(event->hw.flags & PERF_X86_EVENT_RDPMC_ALLOWED))
1945 if (atomic_inc_return(¤t->mm->context.perf_rdpmc_allowed) == 1)
1946 on_each_cpu_mask(mm_cpumask(current->mm), refresh_pce, NULL, 1);
1949 static void x86_pmu_event_unmapped(struct perf_event *event)
1954 if (!(event->hw.flags & PERF_X86_EVENT_RDPMC_ALLOWED))
1957 if (atomic_dec_and_test(¤t->mm->context.perf_rdpmc_allowed))
1958 on_each_cpu_mask(mm_cpumask(current->mm), refresh_pce, NULL, 1);
1961 static int x86_pmu_event_idx(struct perf_event *event)
1963 int idx = event->hw.idx;
1965 if (!(event->hw.flags & PERF_X86_EVENT_RDPMC_ALLOWED))
1968 if (x86_pmu.num_counters_fixed && idx >= INTEL_PMC_IDX_FIXED) {
1969 idx -= INTEL_PMC_IDX_FIXED;
1976 static ssize_t get_attr_rdpmc(struct device *cdev,
1977 struct device_attribute *attr,
1980 return snprintf(buf, 40, "%d\n", x86_pmu.attr_rdpmc);
1983 static ssize_t set_attr_rdpmc(struct device *cdev,
1984 struct device_attribute *attr,
1985 const char *buf, size_t count)
1990 ret = kstrtoul(buf, 0, &val);
1997 if (x86_pmu.attr_rdpmc_broken)
2000 if ((val == 2) != (x86_pmu.attr_rdpmc == 2)) {
2002 * Changing into or out of always available, aka
2003 * perf-event-bypassing mode. This path is extremely slow,
2004 * but only root can trigger it, so it's okay.
2007 static_key_slow_inc(&rdpmc_always_available);
2009 static_key_slow_dec(&rdpmc_always_available);
2010 on_each_cpu(refresh_pce, NULL, 1);
2013 x86_pmu.attr_rdpmc = val;
2018 static DEVICE_ATTR(rdpmc, S_IRUSR | S_IWUSR, get_attr_rdpmc, set_attr_rdpmc);
2020 static struct attribute *x86_pmu_attrs[] = {
2021 &dev_attr_rdpmc.attr,
2025 static struct attribute_group x86_pmu_attr_group = {
2026 .attrs = x86_pmu_attrs,
2029 static const struct attribute_group *x86_pmu_attr_groups[] = {
2030 &x86_pmu_attr_group,
2031 &x86_pmu_format_group,
2032 &x86_pmu_events_group,
2036 static void x86_pmu_sched_task(struct perf_event_context *ctx, bool sched_in)
2038 if (x86_pmu.sched_task)
2039 x86_pmu.sched_task(ctx, sched_in);
2042 void perf_check_microcode(void)
2044 if (x86_pmu.check_microcode)
2045 x86_pmu.check_microcode();
2047 EXPORT_SYMBOL_GPL(perf_check_microcode);
2049 static struct pmu pmu = {
2050 .pmu_enable = x86_pmu_enable,
2051 .pmu_disable = x86_pmu_disable,
2053 .attr_groups = x86_pmu_attr_groups,
2055 .event_init = x86_pmu_event_init,
2057 .event_mapped = x86_pmu_event_mapped,
2058 .event_unmapped = x86_pmu_event_unmapped,
2062 .start = x86_pmu_start,
2063 .stop = x86_pmu_stop,
2064 .read = x86_pmu_read,
2066 .start_txn = x86_pmu_start_txn,
2067 .cancel_txn = x86_pmu_cancel_txn,
2068 .commit_txn = x86_pmu_commit_txn,
2070 .event_idx = x86_pmu_event_idx,
2071 .sched_task = x86_pmu_sched_task,
2072 .task_ctx_size = sizeof(struct x86_perf_task_context),
2075 void arch_perf_update_userpage(struct perf_event *event,
2076 struct perf_event_mmap_page *userpg, u64 now)
2078 struct cyc2ns_data *data;
2080 userpg->cap_user_time = 0;
2081 userpg->cap_user_time_zero = 0;
2082 userpg->cap_user_rdpmc =
2083 !!(event->hw.flags & PERF_X86_EVENT_RDPMC_ALLOWED);
2084 userpg->pmc_width = x86_pmu.cntval_bits;
2086 if (!sched_clock_stable())
2089 data = cyc2ns_read_begin();
2092 * Internal timekeeping for enabled/running/stopped times
2093 * is always in the local_clock domain.
2095 userpg->cap_user_time = 1;
2096 userpg->time_mult = data->cyc2ns_mul;
2097 userpg->time_shift = data->cyc2ns_shift;
2098 userpg->time_offset = data->cyc2ns_offset - now;
2101 * cap_user_time_zero doesn't make sense when we're using a different
2102 * time base for the records.
2104 if (event->clock == &local_clock) {
2105 userpg->cap_user_time_zero = 1;
2106 userpg->time_zero = data->cyc2ns_offset;
2109 cyc2ns_read_end(data);
2116 static int backtrace_stack(void *data, char *name)
2121 static void backtrace_address(void *data, unsigned long addr, int reliable)
2123 struct perf_callchain_entry *entry = data;
2125 perf_callchain_store(entry, addr);
2128 static const struct stacktrace_ops backtrace_ops = {
2129 .stack = backtrace_stack,
2130 .address = backtrace_address,
2131 .walk_stack = print_context_stack_bp,
2135 perf_callchain_kernel(struct perf_callchain_entry *entry, struct pt_regs *regs)
2137 if (perf_guest_cbs && perf_guest_cbs->is_in_guest()) {
2138 /* TODO: We don't support guest os callchain now */
2142 perf_callchain_store(entry, regs->ip);
2144 dump_trace(NULL, regs, NULL, 0, &backtrace_ops, entry);
2148 valid_user_frame(const void __user *fp, unsigned long size)
2150 return (__range_not_ok(fp, size, TASK_SIZE) == 0);
2153 static unsigned long get_segment_base(unsigned int segment)
2155 struct desc_struct *desc;
2156 int idx = segment >> 3;
2158 if ((segment & SEGMENT_TI_MASK) == SEGMENT_LDT) {
2159 if (idx > LDT_ENTRIES)
2162 if (idx > current->active_mm->context.size)
2165 desc = current->active_mm->context.ldt;
2167 if (idx > GDT_ENTRIES)
2170 desc = raw_cpu_ptr(gdt_page.gdt);
2173 return get_desc_base(desc + idx);
2176 #ifdef CONFIG_COMPAT
2178 #include <asm/compat.h>
2181 perf_callchain_user32(struct pt_regs *regs, struct perf_callchain_entry *entry)
2183 /* 32-bit process in 64-bit kernel. */
2184 unsigned long ss_base, cs_base;
2185 struct stack_frame_ia32 frame;
2186 const void __user *fp;
2188 if (!test_thread_flag(TIF_IA32))
2191 cs_base = get_segment_base(regs->cs);
2192 ss_base = get_segment_base(regs->ss);
2194 fp = compat_ptr(ss_base + regs->bp);
2195 while (entry->nr < PERF_MAX_STACK_DEPTH) {
2196 unsigned long bytes;
2197 frame.next_frame = 0;
2198 frame.return_address = 0;
2200 bytes = copy_from_user_nmi(&frame, fp, sizeof(frame));
2204 if (!valid_user_frame(fp, sizeof(frame)))
2207 perf_callchain_store(entry, cs_base + frame.return_address);
2208 fp = compat_ptr(ss_base + frame.next_frame);
2214 perf_callchain_user32(struct pt_regs *regs, struct perf_callchain_entry *entry)
2221 perf_callchain_user(struct perf_callchain_entry *entry, struct pt_regs *regs)
2223 struct stack_frame frame;
2224 const void __user *fp;
2226 if (perf_guest_cbs && perf_guest_cbs->is_in_guest()) {
2227 /* TODO: We don't support guest os callchain now */
2232 * We don't know what to do with VM86 stacks.. ignore them for now.
2234 if (regs->flags & (X86_VM_MASK | PERF_EFLAGS_VM))
2237 fp = (void __user *)regs->bp;
2239 perf_callchain_store(entry, regs->ip);
2244 if (perf_callchain_user32(regs, entry))
2247 while (entry->nr < PERF_MAX_STACK_DEPTH) {
2248 unsigned long bytes;
2249 frame.next_frame = NULL;
2250 frame.return_address = 0;
2252 bytes = copy_from_user_nmi(&frame, fp, sizeof(frame));
2256 if (!valid_user_frame(fp, sizeof(frame)))
2259 perf_callchain_store(entry, frame.return_address);
2260 fp = frame.next_frame;
2265 * Deal with code segment offsets for the various execution modes:
2267 * VM86 - the good olde 16 bit days, where the linear address is
2268 * 20 bits and we use regs->ip + 0x10 * regs->cs.
2270 * IA32 - Where we need to look at GDT/LDT segment descriptor tables
2271 * to figure out what the 32bit base address is.
2273 * X32 - has TIF_X32 set, but is running in x86_64
2275 * X86_64 - CS,DS,SS,ES are all zero based.
2277 static unsigned long code_segment_base(struct pt_regs *regs)
2280 * For IA32 we look at the GDT/LDT segment base to convert the
2281 * effective IP to a linear address.
2284 #ifdef CONFIG_X86_32
2286 * If we are in VM86 mode, add the segment offset to convert to a
2289 if (regs->flags & X86_VM_MASK)
2290 return 0x10 * regs->cs;
2292 if (user_mode(regs) && regs->cs != __USER_CS)
2293 return get_segment_base(regs->cs);
2295 if (user_mode(regs) && !user_64bit_mode(regs) &&
2296 regs->cs != __USER32_CS)
2297 return get_segment_base(regs->cs);
2302 unsigned long perf_instruction_pointer(struct pt_regs *regs)
2304 if (perf_guest_cbs && perf_guest_cbs->is_in_guest())
2305 return perf_guest_cbs->get_guest_ip();
2307 return regs->ip + code_segment_base(regs);
2310 unsigned long perf_misc_flags(struct pt_regs *regs)
2314 if (perf_guest_cbs && perf_guest_cbs->is_in_guest()) {
2315 if (perf_guest_cbs->is_user_mode())
2316 misc |= PERF_RECORD_MISC_GUEST_USER;
2318 misc |= PERF_RECORD_MISC_GUEST_KERNEL;
2320 if (user_mode(regs))
2321 misc |= PERF_RECORD_MISC_USER;
2323 misc |= PERF_RECORD_MISC_KERNEL;
2326 if (regs->flags & PERF_EFLAGS_EXACT)
2327 misc |= PERF_RECORD_MISC_EXACT_IP;
2332 void perf_get_x86_pmu_capability(struct x86_pmu_capability *cap)
2334 cap->version = x86_pmu.version;
2335 cap->num_counters_gp = x86_pmu.num_counters;
2336 cap->num_counters_fixed = x86_pmu.num_counters_fixed;
2337 cap->bit_width_gp = x86_pmu.cntval_bits;
2338 cap->bit_width_fixed = x86_pmu.cntval_bits;
2339 cap->events_mask = (unsigned int)x86_pmu.events_maskl;
2340 cap->events_mask_len = x86_pmu.events_mask_len;
2342 EXPORT_SYMBOL_GPL(perf_get_x86_pmu_capability);