2 * Performance events core code:
4 * Copyright (C) 2008 Thomas Gleixner <tglx@linutronix.de>
5 * Copyright (C) 2008-2009 Red Hat, Inc., Ingo Molnar
6 * Copyright (C) 2008-2009 Red Hat, Inc., Peter Zijlstra <pzijlstr@redhat.com>
7 * Copyright © 2009 Paul Mackerras, IBM Corp. <paulus@au1.ibm.com>
9 * For licensing details see kernel-base/COPYING
14 #include <linux/cpu.h>
15 #include <linux/smp.h>
16 #include <linux/file.h>
17 #include <linux/poll.h>
18 #include <linux/slab.h>
19 #include <linux/hash.h>
20 #include <linux/sysfs.h>
21 #include <linux/dcache.h>
22 #include <linux/percpu.h>
23 #include <linux/ptrace.h>
24 #include <linux/vmstat.h>
25 #include <linux/vmalloc.h>
26 #include <linux/hardirq.h>
27 #include <linux/rculist.h>
28 #include <linux/uaccess.h>
29 #include <linux/syscalls.h>
30 #include <linux/anon_inodes.h>
31 #include <linux/kernel_stat.h>
32 #include <linux/perf_event.h>
33 #include <linux/ftrace_event.h>
35 #include <asm/irq_regs.h>
38 * Each CPU has a list of per CPU events:
40 static DEFINE_PER_CPU(struct perf_cpu_context, perf_cpu_context);
42 int perf_max_events __read_mostly = 1;
43 static int perf_reserved_percpu __read_mostly;
44 static int perf_overcommit __read_mostly = 1;
46 static atomic_t nr_events __read_mostly;
47 static atomic_t nr_mmap_events __read_mostly;
48 static atomic_t nr_comm_events __read_mostly;
49 static atomic_t nr_task_events __read_mostly;
52 * perf event paranoia level:
53 * -1 - not paranoid at all
54 * 0 - disallow raw tracepoint access for unpriv
55 * 1 - disallow cpu events for unpriv
56 * 2 - disallow kernel profiling for unpriv
58 int sysctl_perf_event_paranoid __read_mostly = 1;
60 int sysctl_perf_event_mlock __read_mostly = 512; /* 'free' kb per user */
63 * max perf event sample rate
65 int sysctl_perf_event_sample_rate __read_mostly = 100000;
67 static atomic64_t perf_event_id;
70 * Lock for (sysadmin-configurable) event reservations:
72 static DEFINE_SPINLOCK(perf_resource_lock);
74 void __weak perf_event_print_debug(void) { }
76 void perf_pmu_disable(struct pmu *pmu)
78 int *count = this_cpu_ptr(pmu->pmu_disable_count);
80 pmu->pmu_disable(pmu);
83 void perf_pmu_enable(struct pmu *pmu)
85 int *count = this_cpu_ptr(pmu->pmu_disable_count);
90 static void get_ctx(struct perf_event_context *ctx)
92 WARN_ON(!atomic_inc_not_zero(&ctx->refcount));
95 static void free_ctx(struct rcu_head *head)
97 struct perf_event_context *ctx;
99 ctx = container_of(head, struct perf_event_context, rcu_head);
103 static void put_ctx(struct perf_event_context *ctx)
105 if (atomic_dec_and_test(&ctx->refcount)) {
107 put_ctx(ctx->parent_ctx);
109 put_task_struct(ctx->task);
110 call_rcu(&ctx->rcu_head, free_ctx);
114 static void unclone_ctx(struct perf_event_context *ctx)
116 if (ctx->parent_ctx) {
117 put_ctx(ctx->parent_ctx);
118 ctx->parent_ctx = NULL;
123 * If we inherit events we want to return the parent event id
126 static u64 primary_event_id(struct perf_event *event)
131 id = event->parent->id;
137 * Get the perf_event_context for a task and lock it.
138 * This has to cope with with the fact that until it is locked,
139 * the context could get moved to another task.
141 static struct perf_event_context *
142 perf_lock_task_context(struct task_struct *task, unsigned long *flags)
144 struct perf_event_context *ctx;
148 ctx = rcu_dereference(task->perf_event_ctxp);
151 * If this context is a clone of another, it might
152 * get swapped for another underneath us by
153 * perf_event_task_sched_out, though the
154 * rcu_read_lock() protects us from any context
155 * getting freed. Lock the context and check if it
156 * got swapped before we could get the lock, and retry
157 * if so. If we locked the right context, then it
158 * can't get swapped on us any more.
160 raw_spin_lock_irqsave(&ctx->lock, *flags);
161 if (ctx != rcu_dereference(task->perf_event_ctxp)) {
162 raw_spin_unlock_irqrestore(&ctx->lock, *flags);
166 if (!atomic_inc_not_zero(&ctx->refcount)) {
167 raw_spin_unlock_irqrestore(&ctx->lock, *flags);
176 * Get the context for a task and increment its pin_count so it
177 * can't get swapped to another task. This also increments its
178 * reference count so that the context can't get freed.
180 static struct perf_event_context *perf_pin_task_context(struct task_struct *task)
182 struct perf_event_context *ctx;
185 ctx = perf_lock_task_context(task, &flags);
188 raw_spin_unlock_irqrestore(&ctx->lock, flags);
193 static void perf_unpin_context(struct perf_event_context *ctx)
197 raw_spin_lock_irqsave(&ctx->lock, flags);
199 raw_spin_unlock_irqrestore(&ctx->lock, flags);
203 static inline u64 perf_clock(void)
205 return local_clock();
209 * Update the record of the current time in a context.
211 static void update_context_time(struct perf_event_context *ctx)
213 u64 now = perf_clock();
215 ctx->time += now - ctx->timestamp;
216 ctx->timestamp = now;
220 * Update the total_time_enabled and total_time_running fields for a event.
222 static void update_event_times(struct perf_event *event)
224 struct perf_event_context *ctx = event->ctx;
227 if (event->state < PERF_EVENT_STATE_INACTIVE ||
228 event->group_leader->state < PERF_EVENT_STATE_INACTIVE)
234 run_end = event->tstamp_stopped;
236 event->total_time_enabled = run_end - event->tstamp_enabled;
238 if (event->state == PERF_EVENT_STATE_INACTIVE)
239 run_end = event->tstamp_stopped;
243 event->total_time_running = run_end - event->tstamp_running;
247 * Update total_time_enabled and total_time_running for all events in a group.
249 static void update_group_times(struct perf_event *leader)
251 struct perf_event *event;
253 update_event_times(leader);
254 list_for_each_entry(event, &leader->sibling_list, group_entry)
255 update_event_times(event);
258 static struct list_head *
259 ctx_group_list(struct perf_event *event, struct perf_event_context *ctx)
261 if (event->attr.pinned)
262 return &ctx->pinned_groups;
264 return &ctx->flexible_groups;
268 * Add a event from the lists for its context.
269 * Must be called with ctx->mutex and ctx->lock held.
272 list_add_event(struct perf_event *event, struct perf_event_context *ctx)
274 WARN_ON_ONCE(event->attach_state & PERF_ATTACH_CONTEXT);
275 event->attach_state |= PERF_ATTACH_CONTEXT;
278 * If we're a stand alone event or group leader, we go to the context
279 * list, group events are kept attached to the group so that
280 * perf_group_detach can, at all times, locate all siblings.
282 if (event->group_leader == event) {
283 struct list_head *list;
285 if (is_software_event(event))
286 event->group_flags |= PERF_GROUP_SOFTWARE;
288 list = ctx_group_list(event, ctx);
289 list_add_tail(&event->group_entry, list);
292 list_add_rcu(&event->event_entry, &ctx->event_list);
294 if (event->attr.inherit_stat)
298 static void perf_group_attach(struct perf_event *event)
300 struct perf_event *group_leader = event->group_leader;
302 WARN_ON_ONCE(event->attach_state & PERF_ATTACH_GROUP);
303 event->attach_state |= PERF_ATTACH_GROUP;
305 if (group_leader == event)
308 if (group_leader->group_flags & PERF_GROUP_SOFTWARE &&
309 !is_software_event(event))
310 group_leader->group_flags &= ~PERF_GROUP_SOFTWARE;
312 list_add_tail(&event->group_entry, &group_leader->sibling_list);
313 group_leader->nr_siblings++;
317 * Remove a event from the lists for its context.
318 * Must be called with ctx->mutex and ctx->lock held.
321 list_del_event(struct perf_event *event, struct perf_event_context *ctx)
324 * We can have double detach due to exit/hot-unplug + close.
326 if (!(event->attach_state & PERF_ATTACH_CONTEXT))
329 event->attach_state &= ~PERF_ATTACH_CONTEXT;
332 if (event->attr.inherit_stat)
335 list_del_rcu(&event->event_entry);
337 if (event->group_leader == event)
338 list_del_init(&event->group_entry);
340 update_group_times(event);
343 * If event was in error state, then keep it
344 * that way, otherwise bogus counts will be
345 * returned on read(). The only way to get out
346 * of error state is by explicit re-enabling
349 if (event->state > PERF_EVENT_STATE_OFF)
350 event->state = PERF_EVENT_STATE_OFF;
353 static void perf_group_detach(struct perf_event *event)
355 struct perf_event *sibling, *tmp;
356 struct list_head *list = NULL;
359 * We can have double detach due to exit/hot-unplug + close.
361 if (!(event->attach_state & PERF_ATTACH_GROUP))
364 event->attach_state &= ~PERF_ATTACH_GROUP;
367 * If this is a sibling, remove it from its group.
369 if (event->group_leader != event) {
370 list_del_init(&event->group_entry);
371 event->group_leader->nr_siblings--;
375 if (!list_empty(&event->group_entry))
376 list = &event->group_entry;
379 * If this was a group event with sibling events then
380 * upgrade the siblings to singleton events by adding them
381 * to whatever list we are on.
383 list_for_each_entry_safe(sibling, tmp, &event->sibling_list, group_entry) {
385 list_move_tail(&sibling->group_entry, list);
386 sibling->group_leader = sibling;
388 /* Inherit group flags from the previous leader */
389 sibling->group_flags = event->group_flags;
394 event_filter_match(struct perf_event *event)
396 return event->cpu == -1 || event->cpu == smp_processor_id();
400 event_sched_out(struct perf_event *event,
401 struct perf_cpu_context *cpuctx,
402 struct perf_event_context *ctx)
406 * An event which could not be activated because of
407 * filter mismatch still needs to have its timings
408 * maintained, otherwise bogus information is return
409 * via read() for time_enabled, time_running:
411 if (event->state == PERF_EVENT_STATE_INACTIVE
412 && !event_filter_match(event)) {
413 delta = ctx->time - event->tstamp_stopped;
414 event->tstamp_running += delta;
415 event->tstamp_stopped = ctx->time;
418 if (event->state != PERF_EVENT_STATE_ACTIVE)
421 event->state = PERF_EVENT_STATE_INACTIVE;
422 if (event->pending_disable) {
423 event->pending_disable = 0;
424 event->state = PERF_EVENT_STATE_OFF;
426 event->tstamp_stopped = ctx->time;
427 event->pmu->del(event, 0);
430 if (!is_software_event(event))
431 cpuctx->active_oncpu--;
433 if (event->attr.exclusive || !cpuctx->active_oncpu)
434 cpuctx->exclusive = 0;
438 group_sched_out(struct perf_event *group_event,
439 struct perf_cpu_context *cpuctx,
440 struct perf_event_context *ctx)
442 struct perf_event *event;
443 int state = group_event->state;
445 event_sched_out(group_event, cpuctx, ctx);
448 * Schedule out siblings (if any):
450 list_for_each_entry(event, &group_event->sibling_list, group_entry)
451 event_sched_out(event, cpuctx, ctx);
453 if (state == PERF_EVENT_STATE_ACTIVE && group_event->attr.exclusive)
454 cpuctx->exclusive = 0;
458 * Cross CPU call to remove a performance event
460 * We disable the event on the hardware level first. After that we
461 * remove it from the context list.
463 static void __perf_event_remove_from_context(void *info)
465 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
466 struct perf_event *event = info;
467 struct perf_event_context *ctx = event->ctx;
470 * If this is a task context, we need to check whether it is
471 * the current task context of this cpu. If not it has been
472 * scheduled out before the smp call arrived.
474 if (ctx->task && cpuctx->task_ctx != ctx)
477 raw_spin_lock(&ctx->lock);
479 event_sched_out(event, cpuctx, ctx);
481 list_del_event(event, ctx);
485 * Allow more per task events with respect to the
488 cpuctx->max_pertask =
489 min(perf_max_events - ctx->nr_events,
490 perf_max_events - perf_reserved_percpu);
493 raw_spin_unlock(&ctx->lock);
498 * Remove the event from a task's (or a CPU's) list of events.
500 * Must be called with ctx->mutex held.
502 * CPU events are removed with a smp call. For task events we only
503 * call when the task is on a CPU.
505 * If event->ctx is a cloned context, callers must make sure that
506 * every task struct that event->ctx->task could possibly point to
507 * remains valid. This is OK when called from perf_release since
508 * that only calls us on the top-level context, which can't be a clone.
509 * When called from perf_event_exit_task, it's OK because the
510 * context has been detached from its task.
512 static void perf_event_remove_from_context(struct perf_event *event)
514 struct perf_event_context *ctx = event->ctx;
515 struct task_struct *task = ctx->task;
519 * Per cpu events are removed via an smp call and
520 * the removal is always successful.
522 smp_call_function_single(event->cpu,
523 __perf_event_remove_from_context,
529 task_oncpu_function_call(task, __perf_event_remove_from_context,
532 raw_spin_lock_irq(&ctx->lock);
534 * If the context is active we need to retry the smp call.
536 if (ctx->nr_active && !list_empty(&event->group_entry)) {
537 raw_spin_unlock_irq(&ctx->lock);
542 * The lock prevents that this context is scheduled in so we
543 * can remove the event safely, if the call above did not
546 if (!list_empty(&event->group_entry))
547 list_del_event(event, ctx);
548 raw_spin_unlock_irq(&ctx->lock);
552 * Cross CPU call to disable a performance event
554 static void __perf_event_disable(void *info)
556 struct perf_event *event = info;
557 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
558 struct perf_event_context *ctx = event->ctx;
561 * If this is a per-task event, need to check whether this
562 * event's task is the current task on this cpu.
564 if (ctx->task && cpuctx->task_ctx != ctx)
567 raw_spin_lock(&ctx->lock);
570 * If the event is on, turn it off.
571 * If it is in error state, leave it in error state.
573 if (event->state >= PERF_EVENT_STATE_INACTIVE) {
574 update_context_time(ctx);
575 update_group_times(event);
576 if (event == event->group_leader)
577 group_sched_out(event, cpuctx, ctx);
579 event_sched_out(event, cpuctx, ctx);
580 event->state = PERF_EVENT_STATE_OFF;
583 raw_spin_unlock(&ctx->lock);
589 * If event->ctx is a cloned context, callers must make sure that
590 * every task struct that event->ctx->task could possibly point to
591 * remains valid. This condition is satisifed when called through
592 * perf_event_for_each_child or perf_event_for_each because they
593 * hold the top-level event's child_mutex, so any descendant that
594 * goes to exit will block in sync_child_event.
595 * When called from perf_pending_event it's OK because event->ctx
596 * is the current context on this CPU and preemption is disabled,
597 * hence we can't get into perf_event_task_sched_out for this context.
599 void perf_event_disable(struct perf_event *event)
601 struct perf_event_context *ctx = event->ctx;
602 struct task_struct *task = ctx->task;
606 * Disable the event on the cpu that it's on
608 smp_call_function_single(event->cpu, __perf_event_disable,
614 task_oncpu_function_call(task, __perf_event_disable, event);
616 raw_spin_lock_irq(&ctx->lock);
618 * If the event is still active, we need to retry the cross-call.
620 if (event->state == PERF_EVENT_STATE_ACTIVE) {
621 raw_spin_unlock_irq(&ctx->lock);
626 * Since we have the lock this context can't be scheduled
627 * in, so we can change the state safely.
629 if (event->state == PERF_EVENT_STATE_INACTIVE) {
630 update_group_times(event);
631 event->state = PERF_EVENT_STATE_OFF;
634 raw_spin_unlock_irq(&ctx->lock);
638 event_sched_in(struct perf_event *event,
639 struct perf_cpu_context *cpuctx,
640 struct perf_event_context *ctx)
642 if (event->state <= PERF_EVENT_STATE_OFF)
645 event->state = PERF_EVENT_STATE_ACTIVE;
646 event->oncpu = smp_processor_id();
648 * The new state must be visible before we turn it on in the hardware:
652 if (event->pmu->add(event, PERF_EF_START)) {
653 event->state = PERF_EVENT_STATE_INACTIVE;
658 event->tstamp_running += ctx->time - event->tstamp_stopped;
660 if (!is_software_event(event))
661 cpuctx->active_oncpu++;
664 if (event->attr.exclusive)
665 cpuctx->exclusive = 1;
671 group_sched_in(struct perf_event *group_event,
672 struct perf_cpu_context *cpuctx,
673 struct perf_event_context *ctx)
675 struct perf_event *event, *partial_group = NULL;
676 struct pmu *pmu = group_event->pmu;
678 if (group_event->state == PERF_EVENT_STATE_OFF)
683 if (event_sched_in(group_event, cpuctx, ctx)) {
684 pmu->cancel_txn(pmu);
689 * Schedule in siblings as one group (if any):
691 list_for_each_entry(event, &group_event->sibling_list, group_entry) {
692 if (event_sched_in(event, cpuctx, ctx)) {
693 partial_group = event;
698 if (!pmu->commit_txn(pmu))
703 * Groups can be scheduled in as one unit only, so undo any
704 * partial group before returning:
706 list_for_each_entry(event, &group_event->sibling_list, group_entry) {
707 if (event == partial_group)
709 event_sched_out(event, cpuctx, ctx);
711 event_sched_out(group_event, cpuctx, ctx);
713 pmu->cancel_txn(pmu);
719 * Work out whether we can put this event group on the CPU now.
721 static int group_can_go_on(struct perf_event *event,
722 struct perf_cpu_context *cpuctx,
726 * Groups consisting entirely of software events can always go on.
728 if (event->group_flags & PERF_GROUP_SOFTWARE)
731 * If an exclusive group is already on, no other hardware
734 if (cpuctx->exclusive)
737 * If this group is exclusive and there are already
738 * events on the CPU, it can't go on.
740 if (event->attr.exclusive && cpuctx->active_oncpu)
743 * Otherwise, try to add it if all previous groups were able
749 static void add_event_to_ctx(struct perf_event *event,
750 struct perf_event_context *ctx)
752 list_add_event(event, ctx);
753 perf_group_attach(event);
754 event->tstamp_enabled = ctx->time;
755 event->tstamp_running = ctx->time;
756 event->tstamp_stopped = ctx->time;
760 * Cross CPU call to install and enable a performance event
762 * Must be called with ctx->mutex held
764 static void __perf_install_in_context(void *info)
766 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
767 struct perf_event *event = info;
768 struct perf_event_context *ctx = event->ctx;
769 struct perf_event *leader = event->group_leader;
773 * If this is a task context, we need to check whether it is
774 * the current task context of this cpu. If not it has been
775 * scheduled out before the smp call arrived.
776 * Or possibly this is the right context but it isn't
777 * on this cpu because it had no events.
779 if (ctx->task && cpuctx->task_ctx != ctx) {
780 if (cpuctx->task_ctx || ctx->task != current)
782 cpuctx->task_ctx = ctx;
785 raw_spin_lock(&ctx->lock);
787 update_context_time(ctx);
789 add_event_to_ctx(event, ctx);
791 if (event->cpu != -1 && event->cpu != smp_processor_id())
795 * Don't put the event on if it is disabled or if
796 * it is in a group and the group isn't on.
798 if (event->state != PERF_EVENT_STATE_INACTIVE ||
799 (leader != event && leader->state != PERF_EVENT_STATE_ACTIVE))
803 * An exclusive event can't go on if there are already active
804 * hardware events, and no hardware event can go on if there
805 * is already an exclusive event on.
807 if (!group_can_go_on(event, cpuctx, 1))
810 err = event_sched_in(event, cpuctx, ctx);
814 * This event couldn't go on. If it is in a group
815 * then we have to pull the whole group off.
816 * If the event group is pinned then put it in error state.
819 group_sched_out(leader, cpuctx, ctx);
820 if (leader->attr.pinned) {
821 update_group_times(leader);
822 leader->state = PERF_EVENT_STATE_ERROR;
826 if (!err && !ctx->task && cpuctx->max_pertask)
827 cpuctx->max_pertask--;
830 raw_spin_unlock(&ctx->lock);
834 * Attach a performance event to a context
836 * First we add the event to the list with the hardware enable bit
837 * in event->hw_config cleared.
839 * If the event is attached to a task which is on a CPU we use a smp
840 * call to enable it in the task context. The task might have been
841 * scheduled away, but we check this in the smp call again.
843 * Must be called with ctx->mutex held.
846 perf_install_in_context(struct perf_event_context *ctx,
847 struct perf_event *event,
850 struct task_struct *task = ctx->task;
854 * Per cpu events are installed via an smp call and
855 * the install is always successful.
857 smp_call_function_single(cpu, __perf_install_in_context,
863 task_oncpu_function_call(task, __perf_install_in_context,
866 raw_spin_lock_irq(&ctx->lock);
868 * we need to retry the smp call.
870 if (ctx->is_active && list_empty(&event->group_entry)) {
871 raw_spin_unlock_irq(&ctx->lock);
876 * The lock prevents that this context is scheduled in so we
877 * can add the event safely, if it the call above did not
880 if (list_empty(&event->group_entry))
881 add_event_to_ctx(event, ctx);
882 raw_spin_unlock_irq(&ctx->lock);
886 * Put a event into inactive state and update time fields.
887 * Enabling the leader of a group effectively enables all
888 * the group members that aren't explicitly disabled, so we
889 * have to update their ->tstamp_enabled also.
890 * Note: this works for group members as well as group leaders
891 * since the non-leader members' sibling_lists will be empty.
893 static void __perf_event_mark_enabled(struct perf_event *event,
894 struct perf_event_context *ctx)
896 struct perf_event *sub;
898 event->state = PERF_EVENT_STATE_INACTIVE;
899 event->tstamp_enabled = ctx->time - event->total_time_enabled;
900 list_for_each_entry(sub, &event->sibling_list, group_entry) {
901 if (sub->state >= PERF_EVENT_STATE_INACTIVE) {
902 sub->tstamp_enabled =
903 ctx->time - sub->total_time_enabled;
909 * Cross CPU call to enable a performance event
911 static void __perf_event_enable(void *info)
913 struct perf_event *event = info;
914 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
915 struct perf_event_context *ctx = event->ctx;
916 struct perf_event *leader = event->group_leader;
920 * If this is a per-task event, need to check whether this
921 * event's task is the current task on this cpu.
923 if (ctx->task && cpuctx->task_ctx != ctx) {
924 if (cpuctx->task_ctx || ctx->task != current)
926 cpuctx->task_ctx = ctx;
929 raw_spin_lock(&ctx->lock);
931 update_context_time(ctx);
933 if (event->state >= PERF_EVENT_STATE_INACTIVE)
935 __perf_event_mark_enabled(event, ctx);
937 if (event->cpu != -1 && event->cpu != smp_processor_id())
941 * If the event is in a group and isn't the group leader,
942 * then don't put it on unless the group is on.
944 if (leader != event && leader->state != PERF_EVENT_STATE_ACTIVE)
947 if (!group_can_go_on(event, cpuctx, 1)) {
951 err = group_sched_in(event, cpuctx, ctx);
953 err = event_sched_in(event, cpuctx, ctx);
958 * If this event can't go on and it's part of a
959 * group, then the whole group has to come off.
962 group_sched_out(leader, cpuctx, ctx);
963 if (leader->attr.pinned) {
964 update_group_times(leader);
965 leader->state = PERF_EVENT_STATE_ERROR;
970 raw_spin_unlock(&ctx->lock);
976 * If event->ctx is a cloned context, callers must make sure that
977 * every task struct that event->ctx->task could possibly point to
978 * remains valid. This condition is satisfied when called through
979 * perf_event_for_each_child or perf_event_for_each as described
980 * for perf_event_disable.
982 void perf_event_enable(struct perf_event *event)
984 struct perf_event_context *ctx = event->ctx;
985 struct task_struct *task = ctx->task;
989 * Enable the event on the cpu that it's on
991 smp_call_function_single(event->cpu, __perf_event_enable,
996 raw_spin_lock_irq(&ctx->lock);
997 if (event->state >= PERF_EVENT_STATE_INACTIVE)
1001 * If the event is in error state, clear that first.
1002 * That way, if we see the event in error state below, we
1003 * know that it has gone back into error state, as distinct
1004 * from the task having been scheduled away before the
1005 * cross-call arrived.
1007 if (event->state == PERF_EVENT_STATE_ERROR)
1008 event->state = PERF_EVENT_STATE_OFF;
1011 raw_spin_unlock_irq(&ctx->lock);
1012 task_oncpu_function_call(task, __perf_event_enable, event);
1014 raw_spin_lock_irq(&ctx->lock);
1017 * If the context is active and the event is still off,
1018 * we need to retry the cross-call.
1020 if (ctx->is_active && event->state == PERF_EVENT_STATE_OFF)
1024 * Since we have the lock this context can't be scheduled
1025 * in, so we can change the state safely.
1027 if (event->state == PERF_EVENT_STATE_OFF)
1028 __perf_event_mark_enabled(event, ctx);
1031 raw_spin_unlock_irq(&ctx->lock);
1034 static int perf_event_refresh(struct perf_event *event, int refresh)
1037 * not supported on inherited events
1039 if (event->attr.inherit)
1042 atomic_add(refresh, &event->event_limit);
1043 perf_event_enable(event);
1049 EVENT_FLEXIBLE = 0x1,
1051 EVENT_ALL = EVENT_FLEXIBLE | EVENT_PINNED,
1054 static void ctx_sched_out(struct perf_event_context *ctx,
1055 struct perf_cpu_context *cpuctx,
1056 enum event_type_t event_type)
1058 struct perf_event *event;
1060 raw_spin_lock(&ctx->lock);
1062 if (likely(!ctx->nr_events))
1064 update_context_time(ctx);
1066 if (!ctx->nr_active)
1069 if (event_type & EVENT_PINNED) {
1070 list_for_each_entry(event, &ctx->pinned_groups, group_entry)
1071 group_sched_out(event, cpuctx, ctx);
1074 if (event_type & EVENT_FLEXIBLE) {
1075 list_for_each_entry(event, &ctx->flexible_groups, group_entry)
1076 group_sched_out(event, cpuctx, ctx);
1079 raw_spin_unlock(&ctx->lock);
1083 * Test whether two contexts are equivalent, i.e. whether they
1084 * have both been cloned from the same version of the same context
1085 * and they both have the same number of enabled events.
1086 * If the number of enabled events is the same, then the set
1087 * of enabled events should be the same, because these are both
1088 * inherited contexts, therefore we can't access individual events
1089 * in them directly with an fd; we can only enable/disable all
1090 * events via prctl, or enable/disable all events in a family
1091 * via ioctl, which will have the same effect on both contexts.
1093 static int context_equiv(struct perf_event_context *ctx1,
1094 struct perf_event_context *ctx2)
1096 return ctx1->parent_ctx && ctx1->parent_ctx == ctx2->parent_ctx
1097 && ctx1->parent_gen == ctx2->parent_gen
1098 && !ctx1->pin_count && !ctx2->pin_count;
1101 static void __perf_event_sync_stat(struct perf_event *event,
1102 struct perf_event *next_event)
1106 if (!event->attr.inherit_stat)
1110 * Update the event value, we cannot use perf_event_read()
1111 * because we're in the middle of a context switch and have IRQs
1112 * disabled, which upsets smp_call_function_single(), however
1113 * we know the event must be on the current CPU, therefore we
1114 * don't need to use it.
1116 switch (event->state) {
1117 case PERF_EVENT_STATE_ACTIVE:
1118 event->pmu->read(event);
1121 case PERF_EVENT_STATE_INACTIVE:
1122 update_event_times(event);
1130 * In order to keep per-task stats reliable we need to flip the event
1131 * values when we flip the contexts.
1133 value = local64_read(&next_event->count);
1134 value = local64_xchg(&event->count, value);
1135 local64_set(&next_event->count, value);
1137 swap(event->total_time_enabled, next_event->total_time_enabled);
1138 swap(event->total_time_running, next_event->total_time_running);
1141 * Since we swizzled the values, update the user visible data too.
1143 perf_event_update_userpage(event);
1144 perf_event_update_userpage(next_event);
1147 #define list_next_entry(pos, member) \
1148 list_entry(pos->member.next, typeof(*pos), member)
1150 static void perf_event_sync_stat(struct perf_event_context *ctx,
1151 struct perf_event_context *next_ctx)
1153 struct perf_event *event, *next_event;
1158 update_context_time(ctx);
1160 event = list_first_entry(&ctx->event_list,
1161 struct perf_event, event_entry);
1163 next_event = list_first_entry(&next_ctx->event_list,
1164 struct perf_event, event_entry);
1166 while (&event->event_entry != &ctx->event_list &&
1167 &next_event->event_entry != &next_ctx->event_list) {
1169 __perf_event_sync_stat(event, next_event);
1171 event = list_next_entry(event, event_entry);
1172 next_event = list_next_entry(next_event, event_entry);
1177 * Called from scheduler to remove the events of the current task,
1178 * with interrupts disabled.
1180 * We stop each event and update the event value in event->count.
1182 * This does not protect us against NMI, but disable()
1183 * sets the disabled bit in the control field of event _before_
1184 * accessing the event control register. If a NMI hits, then it will
1185 * not restart the event.
1187 void perf_event_task_sched_out(struct task_struct *task,
1188 struct task_struct *next)
1190 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
1191 struct perf_event_context *ctx = task->perf_event_ctxp;
1192 struct perf_event_context *next_ctx;
1193 struct perf_event_context *parent;
1196 perf_sw_event(PERF_COUNT_SW_CONTEXT_SWITCHES, 1, 1, NULL, 0);
1198 if (likely(!ctx || !cpuctx->task_ctx))
1202 parent = rcu_dereference(ctx->parent_ctx);
1203 next_ctx = next->perf_event_ctxp;
1204 if (parent && next_ctx &&
1205 rcu_dereference(next_ctx->parent_ctx) == parent) {
1207 * Looks like the two contexts are clones, so we might be
1208 * able to optimize the context switch. We lock both
1209 * contexts and check that they are clones under the
1210 * lock (including re-checking that neither has been
1211 * uncloned in the meantime). It doesn't matter which
1212 * order we take the locks because no other cpu could
1213 * be trying to lock both of these tasks.
1215 raw_spin_lock(&ctx->lock);
1216 raw_spin_lock_nested(&next_ctx->lock, SINGLE_DEPTH_NESTING);
1217 if (context_equiv(ctx, next_ctx)) {
1219 * XXX do we need a memory barrier of sorts
1220 * wrt to rcu_dereference() of perf_event_ctxp
1222 task->perf_event_ctxp = next_ctx;
1223 next->perf_event_ctxp = ctx;
1225 next_ctx->task = task;
1228 perf_event_sync_stat(ctx, next_ctx);
1230 raw_spin_unlock(&next_ctx->lock);
1231 raw_spin_unlock(&ctx->lock);
1236 ctx_sched_out(ctx, cpuctx, EVENT_ALL);
1237 cpuctx->task_ctx = NULL;
1241 static void task_ctx_sched_out(struct perf_event_context *ctx,
1242 enum event_type_t event_type)
1244 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
1246 if (!cpuctx->task_ctx)
1249 if (WARN_ON_ONCE(ctx != cpuctx->task_ctx))
1252 ctx_sched_out(ctx, cpuctx, event_type);
1253 cpuctx->task_ctx = NULL;
1257 * Called with IRQs disabled
1259 static void __perf_event_task_sched_out(struct perf_event_context *ctx)
1261 task_ctx_sched_out(ctx, EVENT_ALL);
1265 * Called with IRQs disabled
1267 static void cpu_ctx_sched_out(struct perf_cpu_context *cpuctx,
1268 enum event_type_t event_type)
1270 ctx_sched_out(&cpuctx->ctx, cpuctx, event_type);
1274 ctx_pinned_sched_in(struct perf_event_context *ctx,
1275 struct perf_cpu_context *cpuctx)
1277 struct perf_event *event;
1279 list_for_each_entry(event, &ctx->pinned_groups, group_entry) {
1280 if (event->state <= PERF_EVENT_STATE_OFF)
1282 if (event->cpu != -1 && event->cpu != smp_processor_id())
1285 if (group_can_go_on(event, cpuctx, 1))
1286 group_sched_in(event, cpuctx, ctx);
1289 * If this pinned group hasn't been scheduled,
1290 * put it in error state.
1292 if (event->state == PERF_EVENT_STATE_INACTIVE) {
1293 update_group_times(event);
1294 event->state = PERF_EVENT_STATE_ERROR;
1300 ctx_flexible_sched_in(struct perf_event_context *ctx,
1301 struct perf_cpu_context *cpuctx)
1303 struct perf_event *event;
1306 list_for_each_entry(event, &ctx->flexible_groups, group_entry) {
1307 /* Ignore events in OFF or ERROR state */
1308 if (event->state <= PERF_EVENT_STATE_OFF)
1311 * Listen to the 'cpu' scheduling filter constraint
1314 if (event->cpu != -1 && event->cpu != smp_processor_id())
1317 if (group_can_go_on(event, cpuctx, can_add_hw)) {
1318 if (group_sched_in(event, cpuctx, ctx))
1325 ctx_sched_in(struct perf_event_context *ctx,
1326 struct perf_cpu_context *cpuctx,
1327 enum event_type_t event_type)
1329 raw_spin_lock(&ctx->lock);
1331 if (likely(!ctx->nr_events))
1334 ctx->timestamp = perf_clock();
1337 * First go through the list and put on any pinned groups
1338 * in order to give them the best chance of going on.
1340 if (event_type & EVENT_PINNED)
1341 ctx_pinned_sched_in(ctx, cpuctx);
1343 /* Then walk through the lower prio flexible groups */
1344 if (event_type & EVENT_FLEXIBLE)
1345 ctx_flexible_sched_in(ctx, cpuctx);
1348 raw_spin_unlock(&ctx->lock);
1351 static void cpu_ctx_sched_in(struct perf_cpu_context *cpuctx,
1352 enum event_type_t event_type)
1354 struct perf_event_context *ctx = &cpuctx->ctx;
1356 ctx_sched_in(ctx, cpuctx, event_type);
1359 static void task_ctx_sched_in(struct task_struct *task,
1360 enum event_type_t event_type)
1362 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
1363 struct perf_event_context *ctx = task->perf_event_ctxp;
1367 if (cpuctx->task_ctx == ctx)
1369 ctx_sched_in(ctx, cpuctx, event_type);
1370 cpuctx->task_ctx = ctx;
1373 * Called from scheduler to add the events of the current task
1374 * with interrupts disabled.
1376 * We restore the event value and then enable it.
1378 * This does not protect us against NMI, but enable()
1379 * sets the enabled bit in the control field of event _before_
1380 * accessing the event control register. If a NMI hits, then it will
1381 * keep the event running.
1383 void perf_event_task_sched_in(struct task_struct *task)
1385 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
1386 struct perf_event_context *ctx = task->perf_event_ctxp;
1391 if (cpuctx->task_ctx == ctx)
1395 * We want to keep the following priority order:
1396 * cpu pinned (that don't need to move), task pinned,
1397 * cpu flexible, task flexible.
1399 cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE);
1401 ctx_sched_in(ctx, cpuctx, EVENT_PINNED);
1402 cpu_ctx_sched_in(cpuctx, EVENT_FLEXIBLE);
1403 ctx_sched_in(ctx, cpuctx, EVENT_FLEXIBLE);
1405 cpuctx->task_ctx = ctx;
1408 #define MAX_INTERRUPTS (~0ULL)
1410 static void perf_log_throttle(struct perf_event *event, int enable);
1412 static u64 perf_calculate_period(struct perf_event *event, u64 nsec, u64 count)
1414 u64 frequency = event->attr.sample_freq;
1415 u64 sec = NSEC_PER_SEC;
1416 u64 divisor, dividend;
1418 int count_fls, nsec_fls, frequency_fls, sec_fls;
1420 count_fls = fls64(count);
1421 nsec_fls = fls64(nsec);
1422 frequency_fls = fls64(frequency);
1426 * We got @count in @nsec, with a target of sample_freq HZ
1427 * the target period becomes:
1430 * period = -------------------
1431 * @nsec * sample_freq
1436 * Reduce accuracy by one bit such that @a and @b converge
1437 * to a similar magnitude.
1439 #define REDUCE_FLS(a, b) \
1441 if (a##_fls > b##_fls) { \
1451 * Reduce accuracy until either term fits in a u64, then proceed with
1452 * the other, so that finally we can do a u64/u64 division.
1454 while (count_fls + sec_fls > 64 && nsec_fls + frequency_fls > 64) {
1455 REDUCE_FLS(nsec, frequency);
1456 REDUCE_FLS(sec, count);
1459 if (count_fls + sec_fls > 64) {
1460 divisor = nsec * frequency;
1462 while (count_fls + sec_fls > 64) {
1463 REDUCE_FLS(count, sec);
1467 dividend = count * sec;
1469 dividend = count * sec;
1471 while (nsec_fls + frequency_fls > 64) {
1472 REDUCE_FLS(nsec, frequency);
1476 divisor = nsec * frequency;
1482 return div64_u64(dividend, divisor);
1485 static void perf_adjust_period(struct perf_event *event, u64 nsec, u64 count)
1487 struct hw_perf_event *hwc = &event->hw;
1488 s64 period, sample_period;
1491 period = perf_calculate_period(event, nsec, count);
1493 delta = (s64)(period - hwc->sample_period);
1494 delta = (delta + 7) / 8; /* low pass filter */
1496 sample_period = hwc->sample_period + delta;
1501 hwc->sample_period = sample_period;
1503 if (local64_read(&hwc->period_left) > 8*sample_period) {
1504 event->pmu->stop(event, PERF_EF_UPDATE);
1505 local64_set(&hwc->period_left, 0);
1506 event->pmu->start(event, PERF_EF_RELOAD);
1510 static void perf_ctx_adjust_freq(struct perf_event_context *ctx)
1512 struct perf_event *event;
1513 struct hw_perf_event *hwc;
1514 u64 interrupts, now;
1517 raw_spin_lock(&ctx->lock);
1518 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
1519 if (event->state != PERF_EVENT_STATE_ACTIVE)
1522 if (event->cpu != -1 && event->cpu != smp_processor_id())
1527 interrupts = hwc->interrupts;
1528 hwc->interrupts = 0;
1531 * unthrottle events on the tick
1533 if (interrupts == MAX_INTERRUPTS) {
1534 perf_log_throttle(event, 1);
1535 event->pmu->start(event, 0);
1538 if (!event->attr.freq || !event->attr.sample_freq)
1541 event->pmu->read(event);
1542 now = local64_read(&event->count);
1543 delta = now - hwc->freq_count_stamp;
1544 hwc->freq_count_stamp = now;
1547 perf_adjust_period(event, TICK_NSEC, delta);
1549 raw_spin_unlock(&ctx->lock);
1553 * Round-robin a context's events:
1555 static void rotate_ctx(struct perf_event_context *ctx)
1557 raw_spin_lock(&ctx->lock);
1559 /* Rotate the first entry last of non-pinned groups */
1560 list_rotate_left(&ctx->flexible_groups);
1562 raw_spin_unlock(&ctx->lock);
1565 void perf_event_task_tick(struct task_struct *curr)
1567 struct perf_cpu_context *cpuctx;
1568 struct perf_event_context *ctx;
1571 if (!atomic_read(&nr_events))
1574 cpuctx = &__get_cpu_var(perf_cpu_context);
1575 if (cpuctx->ctx.nr_events &&
1576 cpuctx->ctx.nr_events != cpuctx->ctx.nr_active)
1579 ctx = curr->perf_event_ctxp;
1580 if (ctx && ctx->nr_events && ctx->nr_events != ctx->nr_active)
1583 perf_ctx_adjust_freq(&cpuctx->ctx);
1585 perf_ctx_adjust_freq(ctx);
1590 cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE);
1592 task_ctx_sched_out(ctx, EVENT_FLEXIBLE);
1594 rotate_ctx(&cpuctx->ctx);
1598 cpu_ctx_sched_in(cpuctx, EVENT_FLEXIBLE);
1600 task_ctx_sched_in(curr, EVENT_FLEXIBLE);
1603 static int event_enable_on_exec(struct perf_event *event,
1604 struct perf_event_context *ctx)
1606 if (!event->attr.enable_on_exec)
1609 event->attr.enable_on_exec = 0;
1610 if (event->state >= PERF_EVENT_STATE_INACTIVE)
1613 __perf_event_mark_enabled(event, ctx);
1619 * Enable all of a task's events that have been marked enable-on-exec.
1620 * This expects task == current.
1622 static void perf_event_enable_on_exec(struct task_struct *task)
1624 struct perf_event_context *ctx;
1625 struct perf_event *event;
1626 unsigned long flags;
1630 local_irq_save(flags);
1631 ctx = task->perf_event_ctxp;
1632 if (!ctx || !ctx->nr_events)
1635 __perf_event_task_sched_out(ctx);
1637 raw_spin_lock(&ctx->lock);
1639 list_for_each_entry(event, &ctx->pinned_groups, group_entry) {
1640 ret = event_enable_on_exec(event, ctx);
1645 list_for_each_entry(event, &ctx->flexible_groups, group_entry) {
1646 ret = event_enable_on_exec(event, ctx);
1652 * Unclone this context if we enabled any event.
1657 raw_spin_unlock(&ctx->lock);
1659 perf_event_task_sched_in(task);
1661 local_irq_restore(flags);
1665 * Cross CPU call to read the hardware event
1667 static void __perf_event_read(void *info)
1669 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
1670 struct perf_event *event = info;
1671 struct perf_event_context *ctx = event->ctx;
1674 * If this is a task context, we need to check whether it is
1675 * the current task context of this cpu. If not it has been
1676 * scheduled out before the smp call arrived. In that case
1677 * event->count would have been updated to a recent sample
1678 * when the event was scheduled out.
1680 if (ctx->task && cpuctx->task_ctx != ctx)
1683 raw_spin_lock(&ctx->lock);
1684 update_context_time(ctx);
1685 update_event_times(event);
1686 raw_spin_unlock(&ctx->lock);
1688 event->pmu->read(event);
1691 static inline u64 perf_event_count(struct perf_event *event)
1693 return local64_read(&event->count) + atomic64_read(&event->child_count);
1696 static u64 perf_event_read(struct perf_event *event)
1699 * If event is enabled and currently active on a CPU, update the
1700 * value in the event structure:
1702 if (event->state == PERF_EVENT_STATE_ACTIVE) {
1703 smp_call_function_single(event->oncpu,
1704 __perf_event_read, event, 1);
1705 } else if (event->state == PERF_EVENT_STATE_INACTIVE) {
1706 struct perf_event_context *ctx = event->ctx;
1707 unsigned long flags;
1709 raw_spin_lock_irqsave(&ctx->lock, flags);
1710 update_context_time(ctx);
1711 update_event_times(event);
1712 raw_spin_unlock_irqrestore(&ctx->lock, flags);
1715 return perf_event_count(event);
1722 struct callchain_cpus_entries {
1723 struct rcu_head rcu_head;
1724 struct perf_callchain_entry *cpu_entries[0];
1727 static DEFINE_PER_CPU(int, callchain_recursion[PERF_NR_CONTEXTS]);
1728 static atomic_t nr_callchain_events;
1729 static DEFINE_MUTEX(callchain_mutex);
1730 struct callchain_cpus_entries *callchain_cpus_entries;
1733 __weak void perf_callchain_kernel(struct perf_callchain_entry *entry,
1734 struct pt_regs *regs)
1738 __weak void perf_callchain_user(struct perf_callchain_entry *entry,
1739 struct pt_regs *regs)
1743 static void release_callchain_buffers_rcu(struct rcu_head *head)
1745 struct callchain_cpus_entries *entries;
1748 entries = container_of(head, struct callchain_cpus_entries, rcu_head);
1750 for_each_possible_cpu(cpu)
1751 kfree(entries->cpu_entries[cpu]);
1756 static void release_callchain_buffers(void)
1758 struct callchain_cpus_entries *entries;
1760 entries = callchain_cpus_entries;
1761 rcu_assign_pointer(callchain_cpus_entries, NULL);
1762 call_rcu(&entries->rcu_head, release_callchain_buffers_rcu);
1765 static int alloc_callchain_buffers(void)
1769 struct callchain_cpus_entries *entries;
1772 * We can't use the percpu allocation API for data that can be
1773 * accessed from NMI. Use a temporary manual per cpu allocation
1774 * until that gets sorted out.
1776 size = sizeof(*entries) + sizeof(struct perf_callchain_entry *) *
1777 num_possible_cpus();
1779 entries = kzalloc(size, GFP_KERNEL);
1783 size = sizeof(struct perf_callchain_entry) * PERF_NR_CONTEXTS;
1785 for_each_possible_cpu(cpu) {
1786 entries->cpu_entries[cpu] = kmalloc_node(size, GFP_KERNEL,
1788 if (!entries->cpu_entries[cpu])
1792 rcu_assign_pointer(callchain_cpus_entries, entries);
1797 for_each_possible_cpu(cpu)
1798 kfree(entries->cpu_entries[cpu]);
1804 static int get_callchain_buffers(void)
1809 mutex_lock(&callchain_mutex);
1811 count = atomic_inc_return(&nr_callchain_events);
1812 if (WARN_ON_ONCE(count < 1)) {
1818 /* If the allocation failed, give up */
1819 if (!callchain_cpus_entries)
1824 err = alloc_callchain_buffers();
1826 release_callchain_buffers();
1828 mutex_unlock(&callchain_mutex);
1833 static void put_callchain_buffers(void)
1835 if (atomic_dec_and_mutex_lock(&nr_callchain_events, &callchain_mutex)) {
1836 release_callchain_buffers();
1837 mutex_unlock(&callchain_mutex);
1841 static int get_recursion_context(int *recursion)
1849 else if (in_softirq())
1854 if (recursion[rctx])
1863 static inline void put_recursion_context(int *recursion, int rctx)
1869 static struct perf_callchain_entry *get_callchain_entry(int *rctx)
1872 struct callchain_cpus_entries *entries;
1874 *rctx = get_recursion_context(__get_cpu_var(callchain_recursion));
1878 entries = rcu_dereference(callchain_cpus_entries);
1882 cpu = smp_processor_id();
1884 return &entries->cpu_entries[cpu][*rctx];
1888 put_callchain_entry(int rctx)
1890 put_recursion_context(__get_cpu_var(callchain_recursion), rctx);
1893 static struct perf_callchain_entry *perf_callchain(struct pt_regs *regs)
1896 struct perf_callchain_entry *entry;
1899 entry = get_callchain_entry(&rctx);
1908 if (!user_mode(regs)) {
1909 perf_callchain_store(entry, PERF_CONTEXT_KERNEL);
1910 perf_callchain_kernel(entry, regs);
1912 regs = task_pt_regs(current);
1918 perf_callchain_store(entry, PERF_CONTEXT_USER);
1919 perf_callchain_user(entry, regs);
1923 put_callchain_entry(rctx);
1929 * Initialize the perf_event context in a task_struct:
1932 __perf_event_init_context(struct perf_event_context *ctx,
1933 struct task_struct *task)
1935 raw_spin_lock_init(&ctx->lock);
1936 mutex_init(&ctx->mutex);
1937 INIT_LIST_HEAD(&ctx->pinned_groups);
1938 INIT_LIST_HEAD(&ctx->flexible_groups);
1939 INIT_LIST_HEAD(&ctx->event_list);
1940 atomic_set(&ctx->refcount, 1);
1944 static struct perf_event_context *find_get_context(pid_t pid, int cpu)
1946 struct perf_event_context *ctx;
1947 struct perf_cpu_context *cpuctx;
1948 struct task_struct *task;
1949 unsigned long flags;
1952 if (pid == -1 && cpu != -1) {
1953 /* Must be root to operate on a CPU event: */
1954 if (perf_paranoid_cpu() && !capable(CAP_SYS_ADMIN))
1955 return ERR_PTR(-EACCES);
1957 if (cpu < 0 || cpu >= nr_cpumask_bits)
1958 return ERR_PTR(-EINVAL);
1961 * We could be clever and allow to attach a event to an
1962 * offline CPU and activate it when the CPU comes up, but
1965 if (!cpu_online(cpu))
1966 return ERR_PTR(-ENODEV);
1968 cpuctx = &per_cpu(perf_cpu_context, cpu);
1979 task = find_task_by_vpid(pid);
1981 get_task_struct(task);
1985 return ERR_PTR(-ESRCH);
1988 * Can't attach events to a dying task.
1991 if (task->flags & PF_EXITING)
1994 /* Reuse ptrace permission checks for now. */
1996 if (!ptrace_may_access(task, PTRACE_MODE_READ))
2000 ctx = perf_lock_task_context(task, &flags);
2003 raw_spin_unlock_irqrestore(&ctx->lock, flags);
2007 ctx = kzalloc(sizeof(struct perf_event_context), GFP_KERNEL);
2011 __perf_event_init_context(ctx, task);
2013 if (cmpxchg(&task->perf_event_ctxp, NULL, ctx)) {
2015 * We raced with some other task; use
2016 * the context they set.
2021 get_task_struct(task);
2024 put_task_struct(task);
2028 put_task_struct(task);
2029 return ERR_PTR(err);
2032 static void perf_event_free_filter(struct perf_event *event);
2034 static void free_event_rcu(struct rcu_head *head)
2036 struct perf_event *event;
2038 event = container_of(head, struct perf_event, rcu_head);
2040 put_pid_ns(event->ns);
2041 perf_event_free_filter(event);
2045 static void perf_pending_sync(struct perf_event *event);
2046 static void perf_buffer_put(struct perf_buffer *buffer);
2048 static void free_event(struct perf_event *event)
2050 perf_pending_sync(event);
2052 if (!event->parent) {
2053 atomic_dec(&nr_events);
2054 if (event->attr.mmap || event->attr.mmap_data)
2055 atomic_dec(&nr_mmap_events);
2056 if (event->attr.comm)
2057 atomic_dec(&nr_comm_events);
2058 if (event->attr.task)
2059 atomic_dec(&nr_task_events);
2060 if (event->attr.sample_type & PERF_SAMPLE_CALLCHAIN)
2061 put_callchain_buffers();
2064 if (event->buffer) {
2065 perf_buffer_put(event->buffer);
2066 event->buffer = NULL;
2070 event->destroy(event);
2072 put_ctx(event->ctx);
2073 call_rcu(&event->rcu_head, free_event_rcu);
2076 int perf_event_release_kernel(struct perf_event *event)
2078 struct perf_event_context *ctx = event->ctx;
2081 * Remove from the PMU, can't get re-enabled since we got
2082 * here because the last ref went.
2084 perf_event_disable(event);
2086 WARN_ON_ONCE(ctx->parent_ctx);
2088 * There are two ways this annotation is useful:
2090 * 1) there is a lock recursion from perf_event_exit_task
2091 * see the comment there.
2093 * 2) there is a lock-inversion with mmap_sem through
2094 * perf_event_read_group(), which takes faults while
2095 * holding ctx->mutex, however this is called after
2096 * the last filedesc died, so there is no possibility
2097 * to trigger the AB-BA case.
2099 mutex_lock_nested(&ctx->mutex, SINGLE_DEPTH_NESTING);
2100 raw_spin_lock_irq(&ctx->lock);
2101 perf_group_detach(event);
2102 list_del_event(event, ctx);
2103 raw_spin_unlock_irq(&ctx->lock);
2104 mutex_unlock(&ctx->mutex);
2106 mutex_lock(&event->owner->perf_event_mutex);
2107 list_del_init(&event->owner_entry);
2108 mutex_unlock(&event->owner->perf_event_mutex);
2109 put_task_struct(event->owner);
2115 EXPORT_SYMBOL_GPL(perf_event_release_kernel);
2118 * Called when the last reference to the file is gone.
2120 static int perf_release(struct inode *inode, struct file *file)
2122 struct perf_event *event = file->private_data;
2124 file->private_data = NULL;
2126 return perf_event_release_kernel(event);
2129 static int perf_event_read_size(struct perf_event *event)
2131 int entry = sizeof(u64); /* value */
2135 if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
2136 size += sizeof(u64);
2138 if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
2139 size += sizeof(u64);
2141 if (event->attr.read_format & PERF_FORMAT_ID)
2142 entry += sizeof(u64);
2144 if (event->attr.read_format & PERF_FORMAT_GROUP) {
2145 nr += event->group_leader->nr_siblings;
2146 size += sizeof(u64);
2154 u64 perf_event_read_value(struct perf_event *event, u64 *enabled, u64 *running)
2156 struct perf_event *child;
2162 mutex_lock(&event->child_mutex);
2163 total += perf_event_read(event);
2164 *enabled += event->total_time_enabled +
2165 atomic64_read(&event->child_total_time_enabled);
2166 *running += event->total_time_running +
2167 atomic64_read(&event->child_total_time_running);
2169 list_for_each_entry(child, &event->child_list, child_list) {
2170 total += perf_event_read(child);
2171 *enabled += child->total_time_enabled;
2172 *running += child->total_time_running;
2174 mutex_unlock(&event->child_mutex);
2178 EXPORT_SYMBOL_GPL(perf_event_read_value);
2180 static int perf_event_read_group(struct perf_event *event,
2181 u64 read_format, char __user *buf)
2183 struct perf_event *leader = event->group_leader, *sub;
2184 int n = 0, size = 0, ret = -EFAULT;
2185 struct perf_event_context *ctx = leader->ctx;
2187 u64 count, enabled, running;
2189 mutex_lock(&ctx->mutex);
2190 count = perf_event_read_value(leader, &enabled, &running);
2192 values[n++] = 1 + leader->nr_siblings;
2193 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
2194 values[n++] = enabled;
2195 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
2196 values[n++] = running;
2197 values[n++] = count;
2198 if (read_format & PERF_FORMAT_ID)
2199 values[n++] = primary_event_id(leader);
2201 size = n * sizeof(u64);
2203 if (copy_to_user(buf, values, size))
2208 list_for_each_entry(sub, &leader->sibling_list, group_entry) {
2211 values[n++] = perf_event_read_value(sub, &enabled, &running);
2212 if (read_format & PERF_FORMAT_ID)
2213 values[n++] = primary_event_id(sub);
2215 size = n * sizeof(u64);
2217 if (copy_to_user(buf + ret, values, size)) {
2225 mutex_unlock(&ctx->mutex);
2230 static int perf_event_read_one(struct perf_event *event,
2231 u64 read_format, char __user *buf)
2233 u64 enabled, running;
2237 values[n++] = perf_event_read_value(event, &enabled, &running);
2238 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
2239 values[n++] = enabled;
2240 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
2241 values[n++] = running;
2242 if (read_format & PERF_FORMAT_ID)
2243 values[n++] = primary_event_id(event);
2245 if (copy_to_user(buf, values, n * sizeof(u64)))
2248 return n * sizeof(u64);
2252 * Read the performance event - simple non blocking version for now
2255 perf_read_hw(struct perf_event *event, char __user *buf, size_t count)
2257 u64 read_format = event->attr.read_format;
2261 * Return end-of-file for a read on a event that is in
2262 * error state (i.e. because it was pinned but it couldn't be
2263 * scheduled on to the CPU at some point).
2265 if (event->state == PERF_EVENT_STATE_ERROR)
2268 if (count < perf_event_read_size(event))
2271 WARN_ON_ONCE(event->ctx->parent_ctx);
2272 if (read_format & PERF_FORMAT_GROUP)
2273 ret = perf_event_read_group(event, read_format, buf);
2275 ret = perf_event_read_one(event, read_format, buf);
2281 perf_read(struct file *file, char __user *buf, size_t count, loff_t *ppos)
2283 struct perf_event *event = file->private_data;
2285 return perf_read_hw(event, buf, count);
2288 static unsigned int perf_poll(struct file *file, poll_table *wait)
2290 struct perf_event *event = file->private_data;
2291 struct perf_buffer *buffer;
2292 unsigned int events = POLL_HUP;
2295 buffer = rcu_dereference(event->buffer);
2297 events = atomic_xchg(&buffer->poll, 0);
2300 poll_wait(file, &event->waitq, wait);
2305 static void perf_event_reset(struct perf_event *event)
2307 (void)perf_event_read(event);
2308 local64_set(&event->count, 0);
2309 perf_event_update_userpage(event);
2313 * Holding the top-level event's child_mutex means that any
2314 * descendant process that has inherited this event will block
2315 * in sync_child_event if it goes to exit, thus satisfying the
2316 * task existence requirements of perf_event_enable/disable.
2318 static void perf_event_for_each_child(struct perf_event *event,
2319 void (*func)(struct perf_event *))
2321 struct perf_event *child;
2323 WARN_ON_ONCE(event->ctx->parent_ctx);
2324 mutex_lock(&event->child_mutex);
2326 list_for_each_entry(child, &event->child_list, child_list)
2328 mutex_unlock(&event->child_mutex);
2331 static void perf_event_for_each(struct perf_event *event,
2332 void (*func)(struct perf_event *))
2334 struct perf_event_context *ctx = event->ctx;
2335 struct perf_event *sibling;
2337 WARN_ON_ONCE(ctx->parent_ctx);
2338 mutex_lock(&ctx->mutex);
2339 event = event->group_leader;
2341 perf_event_for_each_child(event, func);
2343 list_for_each_entry(sibling, &event->sibling_list, group_entry)
2344 perf_event_for_each_child(event, func);
2345 mutex_unlock(&ctx->mutex);
2348 static int perf_event_period(struct perf_event *event, u64 __user *arg)
2350 struct perf_event_context *ctx = event->ctx;
2355 if (!event->attr.sample_period)
2358 size = copy_from_user(&value, arg, sizeof(value));
2359 if (size != sizeof(value))
2365 raw_spin_lock_irq(&ctx->lock);
2366 if (event->attr.freq) {
2367 if (value > sysctl_perf_event_sample_rate) {
2372 event->attr.sample_freq = value;
2374 event->attr.sample_period = value;
2375 event->hw.sample_period = value;
2378 raw_spin_unlock_irq(&ctx->lock);
2383 static const struct file_operations perf_fops;
2385 static struct perf_event *perf_fget_light(int fd, int *fput_needed)
2389 file = fget_light(fd, fput_needed);
2391 return ERR_PTR(-EBADF);
2393 if (file->f_op != &perf_fops) {
2394 fput_light(file, *fput_needed);
2396 return ERR_PTR(-EBADF);
2399 return file->private_data;
2402 static int perf_event_set_output(struct perf_event *event,
2403 struct perf_event *output_event);
2404 static int perf_event_set_filter(struct perf_event *event, void __user *arg);
2406 static long perf_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
2408 struct perf_event *event = file->private_data;
2409 void (*func)(struct perf_event *);
2413 case PERF_EVENT_IOC_ENABLE:
2414 func = perf_event_enable;
2416 case PERF_EVENT_IOC_DISABLE:
2417 func = perf_event_disable;
2419 case PERF_EVENT_IOC_RESET:
2420 func = perf_event_reset;
2423 case PERF_EVENT_IOC_REFRESH:
2424 return perf_event_refresh(event, arg);
2426 case PERF_EVENT_IOC_PERIOD:
2427 return perf_event_period(event, (u64 __user *)arg);
2429 case PERF_EVENT_IOC_SET_OUTPUT:
2431 struct perf_event *output_event = NULL;
2432 int fput_needed = 0;
2436 output_event = perf_fget_light(arg, &fput_needed);
2437 if (IS_ERR(output_event))
2438 return PTR_ERR(output_event);
2441 ret = perf_event_set_output(event, output_event);
2443 fput_light(output_event->filp, fput_needed);
2448 case PERF_EVENT_IOC_SET_FILTER:
2449 return perf_event_set_filter(event, (void __user *)arg);
2455 if (flags & PERF_IOC_FLAG_GROUP)
2456 perf_event_for_each(event, func);
2458 perf_event_for_each_child(event, func);
2463 int perf_event_task_enable(void)
2465 struct perf_event *event;
2467 mutex_lock(¤t->perf_event_mutex);
2468 list_for_each_entry(event, ¤t->perf_event_list, owner_entry)
2469 perf_event_for_each_child(event, perf_event_enable);
2470 mutex_unlock(¤t->perf_event_mutex);
2475 int perf_event_task_disable(void)
2477 struct perf_event *event;
2479 mutex_lock(¤t->perf_event_mutex);
2480 list_for_each_entry(event, ¤t->perf_event_list, owner_entry)
2481 perf_event_for_each_child(event, perf_event_disable);
2482 mutex_unlock(¤t->perf_event_mutex);
2487 #ifndef PERF_EVENT_INDEX_OFFSET
2488 # define PERF_EVENT_INDEX_OFFSET 0
2491 static int perf_event_index(struct perf_event *event)
2493 if (event->hw.state & PERF_HES_STOPPED)
2496 if (event->state != PERF_EVENT_STATE_ACTIVE)
2499 return event->hw.idx + 1 - PERF_EVENT_INDEX_OFFSET;
2503 * Callers need to ensure there can be no nesting of this function, otherwise
2504 * the seqlock logic goes bad. We can not serialize this because the arch
2505 * code calls this from NMI context.
2507 void perf_event_update_userpage(struct perf_event *event)
2509 struct perf_event_mmap_page *userpg;
2510 struct perf_buffer *buffer;
2513 buffer = rcu_dereference(event->buffer);
2517 userpg = buffer->user_page;
2520 * Disable preemption so as to not let the corresponding user-space
2521 * spin too long if we get preempted.
2526 userpg->index = perf_event_index(event);
2527 userpg->offset = perf_event_count(event);
2528 if (event->state == PERF_EVENT_STATE_ACTIVE)
2529 userpg->offset -= local64_read(&event->hw.prev_count);
2531 userpg->time_enabled = event->total_time_enabled +
2532 atomic64_read(&event->child_total_time_enabled);
2534 userpg->time_running = event->total_time_running +
2535 atomic64_read(&event->child_total_time_running);
2544 static unsigned long perf_data_size(struct perf_buffer *buffer);
2547 perf_buffer_init(struct perf_buffer *buffer, long watermark, int flags)
2549 long max_size = perf_data_size(buffer);
2552 buffer->watermark = min(max_size, watermark);
2554 if (!buffer->watermark)
2555 buffer->watermark = max_size / 2;
2557 if (flags & PERF_BUFFER_WRITABLE)
2558 buffer->writable = 1;
2560 atomic_set(&buffer->refcount, 1);
2563 #ifndef CONFIG_PERF_USE_VMALLOC
2566 * Back perf_mmap() with regular GFP_KERNEL-0 pages.
2569 static struct page *
2570 perf_mmap_to_page(struct perf_buffer *buffer, unsigned long pgoff)
2572 if (pgoff > buffer->nr_pages)
2576 return virt_to_page(buffer->user_page);
2578 return virt_to_page(buffer->data_pages[pgoff - 1]);
2581 static void *perf_mmap_alloc_page(int cpu)
2586 node = (cpu == -1) ? cpu : cpu_to_node(cpu);
2587 page = alloc_pages_node(node, GFP_KERNEL | __GFP_ZERO, 0);
2591 return page_address(page);
2594 static struct perf_buffer *
2595 perf_buffer_alloc(int nr_pages, long watermark, int cpu, int flags)
2597 struct perf_buffer *buffer;
2601 size = sizeof(struct perf_buffer);
2602 size += nr_pages * sizeof(void *);
2604 buffer = kzalloc(size, GFP_KERNEL);
2608 buffer->user_page = perf_mmap_alloc_page(cpu);
2609 if (!buffer->user_page)
2610 goto fail_user_page;
2612 for (i = 0; i < nr_pages; i++) {
2613 buffer->data_pages[i] = perf_mmap_alloc_page(cpu);
2614 if (!buffer->data_pages[i])
2615 goto fail_data_pages;
2618 buffer->nr_pages = nr_pages;
2620 perf_buffer_init(buffer, watermark, flags);
2625 for (i--; i >= 0; i--)
2626 free_page((unsigned long)buffer->data_pages[i]);
2628 free_page((unsigned long)buffer->user_page);
2637 static void perf_mmap_free_page(unsigned long addr)
2639 struct page *page = virt_to_page((void *)addr);
2641 page->mapping = NULL;
2645 static void perf_buffer_free(struct perf_buffer *buffer)
2649 perf_mmap_free_page((unsigned long)buffer->user_page);
2650 for (i = 0; i < buffer->nr_pages; i++)
2651 perf_mmap_free_page((unsigned long)buffer->data_pages[i]);
2655 static inline int page_order(struct perf_buffer *buffer)
2663 * Back perf_mmap() with vmalloc memory.
2665 * Required for architectures that have d-cache aliasing issues.
2668 static inline int page_order(struct perf_buffer *buffer)
2670 return buffer->page_order;
2673 static struct page *
2674 perf_mmap_to_page(struct perf_buffer *buffer, unsigned long pgoff)
2676 if (pgoff > (1UL << page_order(buffer)))
2679 return vmalloc_to_page((void *)buffer->user_page + pgoff * PAGE_SIZE);
2682 static void perf_mmap_unmark_page(void *addr)
2684 struct page *page = vmalloc_to_page(addr);
2686 page->mapping = NULL;
2689 static void perf_buffer_free_work(struct work_struct *work)
2691 struct perf_buffer *buffer;
2695 buffer = container_of(work, struct perf_buffer, work);
2696 nr = 1 << page_order(buffer);
2698 base = buffer->user_page;
2699 for (i = 0; i < nr + 1; i++)
2700 perf_mmap_unmark_page(base + (i * PAGE_SIZE));
2706 static void perf_buffer_free(struct perf_buffer *buffer)
2708 schedule_work(&buffer->work);
2711 static struct perf_buffer *
2712 perf_buffer_alloc(int nr_pages, long watermark, int cpu, int flags)
2714 struct perf_buffer *buffer;
2718 size = sizeof(struct perf_buffer);
2719 size += sizeof(void *);
2721 buffer = kzalloc(size, GFP_KERNEL);
2725 INIT_WORK(&buffer->work, perf_buffer_free_work);
2727 all_buf = vmalloc_user((nr_pages + 1) * PAGE_SIZE);
2731 buffer->user_page = all_buf;
2732 buffer->data_pages[0] = all_buf + PAGE_SIZE;
2733 buffer->page_order = ilog2(nr_pages);
2734 buffer->nr_pages = 1;
2736 perf_buffer_init(buffer, watermark, flags);
2749 static unsigned long perf_data_size(struct perf_buffer *buffer)
2751 return buffer->nr_pages << (PAGE_SHIFT + page_order(buffer));
2754 static int perf_mmap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
2756 struct perf_event *event = vma->vm_file->private_data;
2757 struct perf_buffer *buffer;
2758 int ret = VM_FAULT_SIGBUS;
2760 if (vmf->flags & FAULT_FLAG_MKWRITE) {
2761 if (vmf->pgoff == 0)
2767 buffer = rcu_dereference(event->buffer);
2771 if (vmf->pgoff && (vmf->flags & FAULT_FLAG_WRITE))
2774 vmf->page = perf_mmap_to_page(buffer, vmf->pgoff);
2778 get_page(vmf->page);
2779 vmf->page->mapping = vma->vm_file->f_mapping;
2780 vmf->page->index = vmf->pgoff;
2789 static void perf_buffer_free_rcu(struct rcu_head *rcu_head)
2791 struct perf_buffer *buffer;
2793 buffer = container_of(rcu_head, struct perf_buffer, rcu_head);
2794 perf_buffer_free(buffer);
2797 static struct perf_buffer *perf_buffer_get(struct perf_event *event)
2799 struct perf_buffer *buffer;
2802 buffer = rcu_dereference(event->buffer);
2804 if (!atomic_inc_not_zero(&buffer->refcount))
2812 static void perf_buffer_put(struct perf_buffer *buffer)
2814 if (!atomic_dec_and_test(&buffer->refcount))
2817 call_rcu(&buffer->rcu_head, perf_buffer_free_rcu);
2820 static void perf_mmap_open(struct vm_area_struct *vma)
2822 struct perf_event *event = vma->vm_file->private_data;
2824 atomic_inc(&event->mmap_count);
2827 static void perf_mmap_close(struct vm_area_struct *vma)
2829 struct perf_event *event = vma->vm_file->private_data;
2831 if (atomic_dec_and_mutex_lock(&event->mmap_count, &event->mmap_mutex)) {
2832 unsigned long size = perf_data_size(event->buffer);
2833 struct user_struct *user = event->mmap_user;
2834 struct perf_buffer *buffer = event->buffer;
2836 atomic_long_sub((size >> PAGE_SHIFT) + 1, &user->locked_vm);
2837 vma->vm_mm->locked_vm -= event->mmap_locked;
2838 rcu_assign_pointer(event->buffer, NULL);
2839 mutex_unlock(&event->mmap_mutex);
2841 perf_buffer_put(buffer);
2846 static const struct vm_operations_struct perf_mmap_vmops = {
2847 .open = perf_mmap_open,
2848 .close = perf_mmap_close,
2849 .fault = perf_mmap_fault,
2850 .page_mkwrite = perf_mmap_fault,
2853 static int perf_mmap(struct file *file, struct vm_area_struct *vma)
2855 struct perf_event *event = file->private_data;
2856 unsigned long user_locked, user_lock_limit;
2857 struct user_struct *user = current_user();
2858 unsigned long locked, lock_limit;
2859 struct perf_buffer *buffer;
2860 unsigned long vma_size;
2861 unsigned long nr_pages;
2862 long user_extra, extra;
2863 int ret = 0, flags = 0;
2866 * Don't allow mmap() of inherited per-task counters. This would
2867 * create a performance issue due to all children writing to the
2870 if (event->cpu == -1 && event->attr.inherit)
2873 if (!(vma->vm_flags & VM_SHARED))
2876 vma_size = vma->vm_end - vma->vm_start;
2877 nr_pages = (vma_size / PAGE_SIZE) - 1;
2880 * If we have buffer pages ensure they're a power-of-two number, so we
2881 * can do bitmasks instead of modulo.
2883 if (nr_pages != 0 && !is_power_of_2(nr_pages))
2886 if (vma_size != PAGE_SIZE * (1 + nr_pages))
2889 if (vma->vm_pgoff != 0)
2892 WARN_ON_ONCE(event->ctx->parent_ctx);
2893 mutex_lock(&event->mmap_mutex);
2894 if (event->buffer) {
2895 if (event->buffer->nr_pages == nr_pages)
2896 atomic_inc(&event->buffer->refcount);
2902 user_extra = nr_pages + 1;
2903 user_lock_limit = sysctl_perf_event_mlock >> (PAGE_SHIFT - 10);
2906 * Increase the limit linearly with more CPUs:
2908 user_lock_limit *= num_online_cpus();
2910 user_locked = atomic_long_read(&user->locked_vm) + user_extra;
2913 if (user_locked > user_lock_limit)
2914 extra = user_locked - user_lock_limit;
2916 lock_limit = rlimit(RLIMIT_MEMLOCK);
2917 lock_limit >>= PAGE_SHIFT;
2918 locked = vma->vm_mm->locked_vm + extra;
2920 if ((locked > lock_limit) && perf_paranoid_tracepoint_raw() &&
2921 !capable(CAP_IPC_LOCK)) {
2926 WARN_ON(event->buffer);
2928 if (vma->vm_flags & VM_WRITE)
2929 flags |= PERF_BUFFER_WRITABLE;
2931 buffer = perf_buffer_alloc(nr_pages, event->attr.wakeup_watermark,
2937 rcu_assign_pointer(event->buffer, buffer);
2939 atomic_long_add(user_extra, &user->locked_vm);
2940 event->mmap_locked = extra;
2941 event->mmap_user = get_current_user();
2942 vma->vm_mm->locked_vm += event->mmap_locked;
2946 atomic_inc(&event->mmap_count);
2947 mutex_unlock(&event->mmap_mutex);
2949 vma->vm_flags |= VM_RESERVED;
2950 vma->vm_ops = &perf_mmap_vmops;
2955 static int perf_fasync(int fd, struct file *filp, int on)
2957 struct inode *inode = filp->f_path.dentry->d_inode;
2958 struct perf_event *event = filp->private_data;
2961 mutex_lock(&inode->i_mutex);
2962 retval = fasync_helper(fd, filp, on, &event->fasync);
2963 mutex_unlock(&inode->i_mutex);
2971 static const struct file_operations perf_fops = {
2972 .llseek = no_llseek,
2973 .release = perf_release,
2976 .unlocked_ioctl = perf_ioctl,
2977 .compat_ioctl = perf_ioctl,
2979 .fasync = perf_fasync,
2985 * If there's data, ensure we set the poll() state and publish everything
2986 * to user-space before waking everybody up.
2989 void perf_event_wakeup(struct perf_event *event)
2991 wake_up_all(&event->waitq);
2993 if (event->pending_kill) {
2994 kill_fasync(&event->fasync, SIGIO, event->pending_kill);
2995 event->pending_kill = 0;
3002 * Handle the case where we need to wakeup up from NMI (or rq->lock) context.
3004 * The NMI bit means we cannot possibly take locks. Therefore, maintain a
3005 * single linked list and use cmpxchg() to add entries lockless.
3008 static void perf_pending_event(struct perf_pending_entry *entry)
3010 struct perf_event *event = container_of(entry,
3011 struct perf_event, pending);
3013 if (event->pending_disable) {
3014 event->pending_disable = 0;
3015 __perf_event_disable(event);
3018 if (event->pending_wakeup) {
3019 event->pending_wakeup = 0;
3020 perf_event_wakeup(event);
3024 #define PENDING_TAIL ((struct perf_pending_entry *)-1UL)
3026 static DEFINE_PER_CPU(struct perf_pending_entry *, perf_pending_head) = {
3030 static void perf_pending_queue(struct perf_pending_entry *entry,
3031 void (*func)(struct perf_pending_entry *))
3033 struct perf_pending_entry **head;
3035 if (cmpxchg(&entry->next, NULL, PENDING_TAIL) != NULL)
3040 head = &get_cpu_var(perf_pending_head);
3043 entry->next = *head;
3044 } while (cmpxchg(head, entry->next, entry) != entry->next);
3046 set_perf_event_pending();
3048 put_cpu_var(perf_pending_head);
3051 static int __perf_pending_run(void)
3053 struct perf_pending_entry *list;
3056 list = xchg(&__get_cpu_var(perf_pending_head), PENDING_TAIL);
3057 while (list != PENDING_TAIL) {
3058 void (*func)(struct perf_pending_entry *);
3059 struct perf_pending_entry *entry = list;
3066 * Ensure we observe the unqueue before we issue the wakeup,
3067 * so that we won't be waiting forever.
3068 * -- see perf_not_pending().
3079 static inline int perf_not_pending(struct perf_event *event)
3082 * If we flush on whatever cpu we run, there is a chance we don't
3086 __perf_pending_run();
3090 * Ensure we see the proper queue state before going to sleep
3091 * so that we do not miss the wakeup. -- see perf_pending_handle()
3094 return event->pending.next == NULL;
3097 static void perf_pending_sync(struct perf_event *event)
3099 wait_event(event->waitq, perf_not_pending(event));
3102 void perf_event_do_pending(void)
3104 __perf_pending_run();
3108 * We assume there is only KVM supporting the callbacks.
3109 * Later on, we might change it to a list if there is
3110 * another virtualization implementation supporting the callbacks.
3112 struct perf_guest_info_callbacks *perf_guest_cbs;
3114 int perf_register_guest_info_callbacks(struct perf_guest_info_callbacks *cbs)
3116 perf_guest_cbs = cbs;
3119 EXPORT_SYMBOL_GPL(perf_register_guest_info_callbacks);
3121 int perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks *cbs)
3123 perf_guest_cbs = NULL;
3126 EXPORT_SYMBOL_GPL(perf_unregister_guest_info_callbacks);
3131 static bool perf_output_space(struct perf_buffer *buffer, unsigned long tail,
3132 unsigned long offset, unsigned long head)
3136 if (!buffer->writable)
3139 mask = perf_data_size(buffer) - 1;
3141 offset = (offset - tail) & mask;
3142 head = (head - tail) & mask;
3144 if ((int)(head - offset) < 0)
3150 static void perf_output_wakeup(struct perf_output_handle *handle)
3152 atomic_set(&handle->buffer->poll, POLL_IN);
3155 handle->event->pending_wakeup = 1;
3156 perf_pending_queue(&handle->event->pending,
3157 perf_pending_event);
3159 perf_event_wakeup(handle->event);
3163 * We need to ensure a later event_id doesn't publish a head when a former
3164 * event isn't done writing. However since we need to deal with NMIs we
3165 * cannot fully serialize things.
3167 * We only publish the head (and generate a wakeup) when the outer-most
3170 static void perf_output_get_handle(struct perf_output_handle *handle)
3172 struct perf_buffer *buffer = handle->buffer;
3175 local_inc(&buffer->nest);
3176 handle->wakeup = local_read(&buffer->wakeup);
3179 static void perf_output_put_handle(struct perf_output_handle *handle)
3181 struct perf_buffer *buffer = handle->buffer;
3185 head = local_read(&buffer->head);
3188 * IRQ/NMI can happen here, which means we can miss a head update.
3191 if (!local_dec_and_test(&buffer->nest))
3195 * Publish the known good head. Rely on the full barrier implied
3196 * by atomic_dec_and_test() order the buffer->head read and this
3199 buffer->user_page->data_head = head;
3202 * Now check if we missed an update, rely on the (compiler)
3203 * barrier in atomic_dec_and_test() to re-read buffer->head.
3205 if (unlikely(head != local_read(&buffer->head))) {
3206 local_inc(&buffer->nest);
3210 if (handle->wakeup != local_read(&buffer->wakeup))
3211 perf_output_wakeup(handle);
3217 __always_inline void perf_output_copy(struct perf_output_handle *handle,
3218 const void *buf, unsigned int len)
3221 unsigned long size = min_t(unsigned long, handle->size, len);
3223 memcpy(handle->addr, buf, size);
3226 handle->addr += size;
3228 handle->size -= size;
3229 if (!handle->size) {
3230 struct perf_buffer *buffer = handle->buffer;
3233 handle->page &= buffer->nr_pages - 1;
3234 handle->addr = buffer->data_pages[handle->page];
3235 handle->size = PAGE_SIZE << page_order(buffer);
3240 int perf_output_begin(struct perf_output_handle *handle,
3241 struct perf_event *event, unsigned int size,
3242 int nmi, int sample)
3244 struct perf_buffer *buffer;
3245 unsigned long tail, offset, head;
3248 struct perf_event_header header;
3255 * For inherited events we send all the output towards the parent.
3258 event = event->parent;
3260 buffer = rcu_dereference(event->buffer);
3264 handle->buffer = buffer;
3265 handle->event = event;
3267 handle->sample = sample;
3269 if (!buffer->nr_pages)
3272 have_lost = local_read(&buffer->lost);
3274 size += sizeof(lost_event);
3276 perf_output_get_handle(handle);
3280 * Userspace could choose to issue a mb() before updating the
3281 * tail pointer. So that all reads will be completed before the
3284 tail = ACCESS_ONCE(buffer->user_page->data_tail);
3286 offset = head = local_read(&buffer->head);
3288 if (unlikely(!perf_output_space(buffer, tail, offset, head)))
3290 } while (local_cmpxchg(&buffer->head, offset, head) != offset);
3292 if (head - local_read(&buffer->wakeup) > buffer->watermark)
3293 local_add(buffer->watermark, &buffer->wakeup);
3295 handle->page = offset >> (PAGE_SHIFT + page_order(buffer));
3296 handle->page &= buffer->nr_pages - 1;
3297 handle->size = offset & ((PAGE_SIZE << page_order(buffer)) - 1);
3298 handle->addr = buffer->data_pages[handle->page];
3299 handle->addr += handle->size;
3300 handle->size = (PAGE_SIZE << page_order(buffer)) - handle->size;
3303 lost_event.header.type = PERF_RECORD_LOST;
3304 lost_event.header.misc = 0;
3305 lost_event.header.size = sizeof(lost_event);
3306 lost_event.id = event->id;
3307 lost_event.lost = local_xchg(&buffer->lost, 0);
3309 perf_output_put(handle, lost_event);
3315 local_inc(&buffer->lost);
3316 perf_output_put_handle(handle);
3323 void perf_output_end(struct perf_output_handle *handle)
3325 struct perf_event *event = handle->event;
3326 struct perf_buffer *buffer = handle->buffer;
3328 int wakeup_events = event->attr.wakeup_events;
3330 if (handle->sample && wakeup_events) {
3331 int events = local_inc_return(&buffer->events);
3332 if (events >= wakeup_events) {
3333 local_sub(wakeup_events, &buffer->events);
3334 local_inc(&buffer->wakeup);
3338 perf_output_put_handle(handle);
3342 static u32 perf_event_pid(struct perf_event *event, struct task_struct *p)
3345 * only top level events have the pid namespace they were created in
3348 event = event->parent;
3350 return task_tgid_nr_ns(p, event->ns);
3353 static u32 perf_event_tid(struct perf_event *event, struct task_struct *p)
3356 * only top level events have the pid namespace they were created in
3359 event = event->parent;
3361 return task_pid_nr_ns(p, event->ns);
3364 static void perf_output_read_one(struct perf_output_handle *handle,
3365 struct perf_event *event)
3367 u64 read_format = event->attr.read_format;
3371 values[n++] = perf_event_count(event);
3372 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
3373 values[n++] = event->total_time_enabled +
3374 atomic64_read(&event->child_total_time_enabled);
3376 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
3377 values[n++] = event->total_time_running +
3378 atomic64_read(&event->child_total_time_running);
3380 if (read_format & PERF_FORMAT_ID)
3381 values[n++] = primary_event_id(event);
3383 perf_output_copy(handle, values, n * sizeof(u64));
3387 * XXX PERF_FORMAT_GROUP vs inherited events seems difficult.
3389 static void perf_output_read_group(struct perf_output_handle *handle,
3390 struct perf_event *event)
3392 struct perf_event *leader = event->group_leader, *sub;
3393 u64 read_format = event->attr.read_format;
3397 values[n++] = 1 + leader->nr_siblings;
3399 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
3400 values[n++] = leader->total_time_enabled;
3402 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
3403 values[n++] = leader->total_time_running;
3405 if (leader != event)
3406 leader->pmu->read(leader);
3408 values[n++] = perf_event_count(leader);
3409 if (read_format & PERF_FORMAT_ID)
3410 values[n++] = primary_event_id(leader);
3412 perf_output_copy(handle, values, n * sizeof(u64));
3414 list_for_each_entry(sub, &leader->sibling_list, group_entry) {
3418 sub->pmu->read(sub);
3420 values[n++] = perf_event_count(sub);
3421 if (read_format & PERF_FORMAT_ID)
3422 values[n++] = primary_event_id(sub);
3424 perf_output_copy(handle, values, n * sizeof(u64));
3428 static void perf_output_read(struct perf_output_handle *handle,
3429 struct perf_event *event)
3431 if (event->attr.read_format & PERF_FORMAT_GROUP)
3432 perf_output_read_group(handle, event);
3434 perf_output_read_one(handle, event);
3437 void perf_output_sample(struct perf_output_handle *handle,
3438 struct perf_event_header *header,
3439 struct perf_sample_data *data,
3440 struct perf_event *event)
3442 u64 sample_type = data->type;
3444 perf_output_put(handle, *header);
3446 if (sample_type & PERF_SAMPLE_IP)
3447 perf_output_put(handle, data->ip);
3449 if (sample_type & PERF_SAMPLE_TID)
3450 perf_output_put(handle, data->tid_entry);
3452 if (sample_type & PERF_SAMPLE_TIME)
3453 perf_output_put(handle, data->time);
3455 if (sample_type & PERF_SAMPLE_ADDR)
3456 perf_output_put(handle, data->addr);
3458 if (sample_type & PERF_SAMPLE_ID)
3459 perf_output_put(handle, data->id);
3461 if (sample_type & PERF_SAMPLE_STREAM_ID)
3462 perf_output_put(handle, data->stream_id);
3464 if (sample_type & PERF_SAMPLE_CPU)
3465 perf_output_put(handle, data->cpu_entry);
3467 if (sample_type & PERF_SAMPLE_PERIOD)
3468 perf_output_put(handle, data->period);
3470 if (sample_type & PERF_SAMPLE_READ)
3471 perf_output_read(handle, event);
3473 if (sample_type & PERF_SAMPLE_CALLCHAIN) {
3474 if (data->callchain) {
3477 if (data->callchain)
3478 size += data->callchain->nr;
3480 size *= sizeof(u64);
3482 perf_output_copy(handle, data->callchain, size);
3485 perf_output_put(handle, nr);
3489 if (sample_type & PERF_SAMPLE_RAW) {
3491 perf_output_put(handle, data->raw->size);
3492 perf_output_copy(handle, data->raw->data,
3499 .size = sizeof(u32),
3502 perf_output_put(handle, raw);
3507 void perf_prepare_sample(struct perf_event_header *header,
3508 struct perf_sample_data *data,
3509 struct perf_event *event,
3510 struct pt_regs *regs)
3512 u64 sample_type = event->attr.sample_type;
3514 data->type = sample_type;
3516 header->type = PERF_RECORD_SAMPLE;
3517 header->size = sizeof(*header);
3520 header->misc |= perf_misc_flags(regs);
3522 if (sample_type & PERF_SAMPLE_IP) {
3523 data->ip = perf_instruction_pointer(regs);
3525 header->size += sizeof(data->ip);
3528 if (sample_type & PERF_SAMPLE_TID) {
3529 /* namespace issues */
3530 data->tid_entry.pid = perf_event_pid(event, current);
3531 data->tid_entry.tid = perf_event_tid(event, current);
3533 header->size += sizeof(data->tid_entry);
3536 if (sample_type & PERF_SAMPLE_TIME) {
3537 data->time = perf_clock();
3539 header->size += sizeof(data->time);
3542 if (sample_type & PERF_SAMPLE_ADDR)
3543 header->size += sizeof(data->addr);
3545 if (sample_type & PERF_SAMPLE_ID) {
3546 data->id = primary_event_id(event);
3548 header->size += sizeof(data->id);
3551 if (sample_type & PERF_SAMPLE_STREAM_ID) {
3552 data->stream_id = event->id;
3554 header->size += sizeof(data->stream_id);
3557 if (sample_type & PERF_SAMPLE_CPU) {
3558 data->cpu_entry.cpu = raw_smp_processor_id();
3559 data->cpu_entry.reserved = 0;
3561 header->size += sizeof(data->cpu_entry);
3564 if (sample_type & PERF_SAMPLE_PERIOD)
3565 header->size += sizeof(data->period);
3567 if (sample_type & PERF_SAMPLE_READ)
3568 header->size += perf_event_read_size(event);
3570 if (sample_type & PERF_SAMPLE_CALLCHAIN) {
3573 data->callchain = perf_callchain(regs);
3575 if (data->callchain)
3576 size += data->callchain->nr;
3578 header->size += size * sizeof(u64);
3581 if (sample_type & PERF_SAMPLE_RAW) {
3582 int size = sizeof(u32);
3585 size += data->raw->size;
3587 size += sizeof(u32);
3589 WARN_ON_ONCE(size & (sizeof(u64)-1));
3590 header->size += size;
3594 static void perf_event_output(struct perf_event *event, int nmi,
3595 struct perf_sample_data *data,
3596 struct pt_regs *regs)
3598 struct perf_output_handle handle;
3599 struct perf_event_header header;
3601 /* protect the callchain buffers */
3604 perf_prepare_sample(&header, data, event, regs);
3606 if (perf_output_begin(&handle, event, header.size, nmi, 1))
3609 perf_output_sample(&handle, &header, data, event);
3611 perf_output_end(&handle);
3621 struct perf_read_event {
3622 struct perf_event_header header;
3629 perf_event_read_event(struct perf_event *event,
3630 struct task_struct *task)
3632 struct perf_output_handle handle;
3633 struct perf_read_event read_event = {
3635 .type = PERF_RECORD_READ,
3637 .size = sizeof(read_event) + perf_event_read_size(event),
3639 .pid = perf_event_pid(event, task),
3640 .tid = perf_event_tid(event, task),
3644 ret = perf_output_begin(&handle, event, read_event.header.size, 0, 0);
3648 perf_output_put(&handle, read_event);
3649 perf_output_read(&handle, event);
3651 perf_output_end(&handle);
3655 * task tracking -- fork/exit
3657 * enabled by: attr.comm | attr.mmap | attr.mmap_data | attr.task
3660 struct perf_task_event {
3661 struct task_struct *task;
3662 struct perf_event_context *task_ctx;
3665 struct perf_event_header header;
3675 static void perf_event_task_output(struct perf_event *event,
3676 struct perf_task_event *task_event)
3678 struct perf_output_handle handle;
3679 struct task_struct *task = task_event->task;
3682 size = task_event->event_id.header.size;
3683 ret = perf_output_begin(&handle, event, size, 0, 0);
3688 task_event->event_id.pid = perf_event_pid(event, task);
3689 task_event->event_id.ppid = perf_event_pid(event, current);
3691 task_event->event_id.tid = perf_event_tid(event, task);
3692 task_event->event_id.ptid = perf_event_tid(event, current);
3694 perf_output_put(&handle, task_event->event_id);
3696 perf_output_end(&handle);
3699 static int perf_event_task_match(struct perf_event *event)
3701 if (event->state < PERF_EVENT_STATE_INACTIVE)
3704 if (event->cpu != -1 && event->cpu != smp_processor_id())
3707 if (event->attr.comm || event->attr.mmap ||
3708 event->attr.mmap_data || event->attr.task)
3714 static void perf_event_task_ctx(struct perf_event_context *ctx,
3715 struct perf_task_event *task_event)
3717 struct perf_event *event;
3719 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
3720 if (perf_event_task_match(event))
3721 perf_event_task_output(event, task_event);
3725 static void perf_event_task_event(struct perf_task_event *task_event)
3727 struct perf_cpu_context *cpuctx;
3728 struct perf_event_context *ctx = task_event->task_ctx;
3731 cpuctx = &get_cpu_var(perf_cpu_context);
3732 perf_event_task_ctx(&cpuctx->ctx, task_event);
3734 ctx = rcu_dereference(current->perf_event_ctxp);
3736 perf_event_task_ctx(ctx, task_event);
3737 put_cpu_var(perf_cpu_context);
3741 static void perf_event_task(struct task_struct *task,
3742 struct perf_event_context *task_ctx,
3745 struct perf_task_event task_event;
3747 if (!atomic_read(&nr_comm_events) &&
3748 !atomic_read(&nr_mmap_events) &&
3749 !atomic_read(&nr_task_events))
3752 task_event = (struct perf_task_event){
3754 .task_ctx = task_ctx,
3757 .type = new ? PERF_RECORD_FORK : PERF_RECORD_EXIT,
3759 .size = sizeof(task_event.event_id),
3765 .time = perf_clock(),
3769 perf_event_task_event(&task_event);
3772 void perf_event_fork(struct task_struct *task)
3774 perf_event_task(task, NULL, 1);
3781 struct perf_comm_event {
3782 struct task_struct *task;
3787 struct perf_event_header header;
3794 static void perf_event_comm_output(struct perf_event *event,
3795 struct perf_comm_event *comm_event)
3797 struct perf_output_handle handle;
3798 int size = comm_event->event_id.header.size;
3799 int ret = perf_output_begin(&handle, event, size, 0, 0);
3804 comm_event->event_id.pid = perf_event_pid(event, comm_event->task);
3805 comm_event->event_id.tid = perf_event_tid(event, comm_event->task);
3807 perf_output_put(&handle, comm_event->event_id);
3808 perf_output_copy(&handle, comm_event->comm,
3809 comm_event->comm_size);
3810 perf_output_end(&handle);
3813 static int perf_event_comm_match(struct perf_event *event)
3815 if (event->state < PERF_EVENT_STATE_INACTIVE)
3818 if (event->cpu != -1 && event->cpu != smp_processor_id())
3821 if (event->attr.comm)
3827 static void perf_event_comm_ctx(struct perf_event_context *ctx,
3828 struct perf_comm_event *comm_event)
3830 struct perf_event *event;
3832 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
3833 if (perf_event_comm_match(event))
3834 perf_event_comm_output(event, comm_event);
3838 static void perf_event_comm_event(struct perf_comm_event *comm_event)
3840 struct perf_cpu_context *cpuctx;
3841 struct perf_event_context *ctx;
3843 char comm[TASK_COMM_LEN];
3845 memset(comm, 0, sizeof(comm));
3846 strlcpy(comm, comm_event->task->comm, sizeof(comm));
3847 size = ALIGN(strlen(comm)+1, sizeof(u64));
3849 comm_event->comm = comm;
3850 comm_event->comm_size = size;
3852 comm_event->event_id.header.size = sizeof(comm_event->event_id) + size;
3855 cpuctx = &get_cpu_var(perf_cpu_context);
3856 perf_event_comm_ctx(&cpuctx->ctx, comm_event);
3857 ctx = rcu_dereference(current->perf_event_ctxp);
3859 perf_event_comm_ctx(ctx, comm_event);
3860 put_cpu_var(perf_cpu_context);
3864 void perf_event_comm(struct task_struct *task)
3866 struct perf_comm_event comm_event;
3868 if (task->perf_event_ctxp)
3869 perf_event_enable_on_exec(task);
3871 if (!atomic_read(&nr_comm_events))
3874 comm_event = (struct perf_comm_event){
3880 .type = PERF_RECORD_COMM,
3889 perf_event_comm_event(&comm_event);
3896 struct perf_mmap_event {
3897 struct vm_area_struct *vma;
3899 const char *file_name;
3903 struct perf_event_header header;
3913 static void perf_event_mmap_output(struct perf_event *event,
3914 struct perf_mmap_event *mmap_event)
3916 struct perf_output_handle handle;
3917 int size = mmap_event->event_id.header.size;
3918 int ret = perf_output_begin(&handle, event, size, 0, 0);
3923 mmap_event->event_id.pid = perf_event_pid(event, current);
3924 mmap_event->event_id.tid = perf_event_tid(event, current);
3926 perf_output_put(&handle, mmap_event->event_id);
3927 perf_output_copy(&handle, mmap_event->file_name,
3928 mmap_event->file_size);
3929 perf_output_end(&handle);
3932 static int perf_event_mmap_match(struct perf_event *event,
3933 struct perf_mmap_event *mmap_event,
3936 if (event->state < PERF_EVENT_STATE_INACTIVE)
3939 if (event->cpu != -1 && event->cpu != smp_processor_id())
3942 if ((!executable && event->attr.mmap_data) ||
3943 (executable && event->attr.mmap))
3949 static void perf_event_mmap_ctx(struct perf_event_context *ctx,
3950 struct perf_mmap_event *mmap_event,
3953 struct perf_event *event;
3955 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
3956 if (perf_event_mmap_match(event, mmap_event, executable))
3957 perf_event_mmap_output(event, mmap_event);
3961 static void perf_event_mmap_event(struct perf_mmap_event *mmap_event)
3963 struct perf_cpu_context *cpuctx;
3964 struct perf_event_context *ctx;
3965 struct vm_area_struct *vma = mmap_event->vma;
3966 struct file *file = vma->vm_file;
3972 memset(tmp, 0, sizeof(tmp));
3976 * d_path works from the end of the buffer backwards, so we
3977 * need to add enough zero bytes after the string to handle
3978 * the 64bit alignment we do later.
3980 buf = kzalloc(PATH_MAX + sizeof(u64), GFP_KERNEL);
3982 name = strncpy(tmp, "//enomem", sizeof(tmp));
3985 name = d_path(&file->f_path, buf, PATH_MAX);
3987 name = strncpy(tmp, "//toolong", sizeof(tmp));
3991 if (arch_vma_name(mmap_event->vma)) {
3992 name = strncpy(tmp, arch_vma_name(mmap_event->vma),
3998 name = strncpy(tmp, "[vdso]", sizeof(tmp));
4000 } else if (vma->vm_start <= vma->vm_mm->start_brk &&
4001 vma->vm_end >= vma->vm_mm->brk) {
4002 name = strncpy(tmp, "[heap]", sizeof(tmp));
4004 } else if (vma->vm_start <= vma->vm_mm->start_stack &&
4005 vma->vm_end >= vma->vm_mm->start_stack) {
4006 name = strncpy(tmp, "[stack]", sizeof(tmp));
4010 name = strncpy(tmp, "//anon", sizeof(tmp));
4015 size = ALIGN(strlen(name)+1, sizeof(u64));
4017 mmap_event->file_name = name;
4018 mmap_event->file_size = size;
4020 mmap_event->event_id.header.size = sizeof(mmap_event->event_id) + size;
4023 cpuctx = &get_cpu_var(perf_cpu_context);
4024 perf_event_mmap_ctx(&cpuctx->ctx, mmap_event, vma->vm_flags & VM_EXEC);
4025 ctx = rcu_dereference(current->perf_event_ctxp);
4027 perf_event_mmap_ctx(ctx, mmap_event, vma->vm_flags & VM_EXEC);
4028 put_cpu_var(perf_cpu_context);
4034 void perf_event_mmap(struct vm_area_struct *vma)
4036 struct perf_mmap_event mmap_event;
4038 if (!atomic_read(&nr_mmap_events))
4041 mmap_event = (struct perf_mmap_event){
4047 .type = PERF_RECORD_MMAP,
4048 .misc = PERF_RECORD_MISC_USER,
4053 .start = vma->vm_start,
4054 .len = vma->vm_end - vma->vm_start,
4055 .pgoff = (u64)vma->vm_pgoff << PAGE_SHIFT,
4059 perf_event_mmap_event(&mmap_event);
4063 * IRQ throttle logging
4066 static void perf_log_throttle(struct perf_event *event, int enable)
4068 struct perf_output_handle handle;
4072 struct perf_event_header header;
4076 } throttle_event = {
4078 .type = PERF_RECORD_THROTTLE,
4080 .size = sizeof(throttle_event),
4082 .time = perf_clock(),
4083 .id = primary_event_id(event),
4084 .stream_id = event->id,
4088 throttle_event.header.type = PERF_RECORD_UNTHROTTLE;
4090 ret = perf_output_begin(&handle, event, sizeof(throttle_event), 1, 0);
4094 perf_output_put(&handle, throttle_event);
4095 perf_output_end(&handle);
4099 * Generic event overflow handling, sampling.
4102 static int __perf_event_overflow(struct perf_event *event, int nmi,
4103 int throttle, struct perf_sample_data *data,
4104 struct pt_regs *regs)
4106 int events = atomic_read(&event->event_limit);
4107 struct hw_perf_event *hwc = &event->hw;
4113 if (hwc->interrupts != MAX_INTERRUPTS) {
4115 if (HZ * hwc->interrupts >
4116 (u64)sysctl_perf_event_sample_rate) {
4117 hwc->interrupts = MAX_INTERRUPTS;
4118 perf_log_throttle(event, 0);
4123 * Keep re-disabling events even though on the previous
4124 * pass we disabled it - just in case we raced with a
4125 * sched-in and the event got enabled again:
4131 if (event->attr.freq) {
4132 u64 now = perf_clock();
4133 s64 delta = now - hwc->freq_time_stamp;
4135 hwc->freq_time_stamp = now;
4137 if (delta > 0 && delta < 2*TICK_NSEC)
4138 perf_adjust_period(event, delta, hwc->last_period);
4142 * XXX event_limit might not quite work as expected on inherited
4146 event->pending_kill = POLL_IN;
4147 if (events && atomic_dec_and_test(&event->event_limit)) {
4149 event->pending_kill = POLL_HUP;
4151 event->pending_disable = 1;
4152 perf_pending_queue(&event->pending,
4153 perf_pending_event);
4155 perf_event_disable(event);
4158 if (event->overflow_handler)
4159 event->overflow_handler(event, nmi, data, regs);
4161 perf_event_output(event, nmi, data, regs);
4166 int perf_event_overflow(struct perf_event *event, int nmi,
4167 struct perf_sample_data *data,
4168 struct pt_regs *regs)
4170 return __perf_event_overflow(event, nmi, 1, data, regs);
4174 * Generic software event infrastructure
4178 * We directly increment event->count and keep a second value in
4179 * event->hw.period_left to count intervals. This period event
4180 * is kept in the range [-sample_period, 0] so that we can use the
4184 static u64 perf_swevent_set_period(struct perf_event *event)
4186 struct hw_perf_event *hwc = &event->hw;
4187 u64 period = hwc->last_period;
4191 hwc->last_period = hwc->sample_period;
4194 old = val = local64_read(&hwc->period_left);
4198 nr = div64_u64(period + val, period);
4199 offset = nr * period;
4201 if (local64_cmpxchg(&hwc->period_left, old, val) != old)
4207 static void perf_swevent_overflow(struct perf_event *event, u64 overflow,
4208 int nmi, struct perf_sample_data *data,
4209 struct pt_regs *regs)
4211 struct hw_perf_event *hwc = &event->hw;
4214 data->period = event->hw.last_period;
4216 overflow = perf_swevent_set_period(event);
4218 if (hwc->interrupts == MAX_INTERRUPTS)
4221 for (; overflow; overflow--) {
4222 if (__perf_event_overflow(event, nmi, throttle,
4225 * We inhibit the overflow from happening when
4226 * hwc->interrupts == MAX_INTERRUPTS.
4234 static void perf_swevent_event(struct perf_event *event, u64 nr,
4235 int nmi, struct perf_sample_data *data,
4236 struct pt_regs *regs)
4238 struct hw_perf_event *hwc = &event->hw;
4240 local64_add(nr, &event->count);
4245 if (!hwc->sample_period)
4248 if (nr == 1 && hwc->sample_period == 1 && !event->attr.freq)
4249 return perf_swevent_overflow(event, 1, nmi, data, regs);
4251 if (local64_add_negative(nr, &hwc->period_left))
4254 perf_swevent_overflow(event, 0, nmi, data, regs);
4257 static int perf_exclude_event(struct perf_event *event,
4258 struct pt_regs *regs)
4260 if (event->hw.state & PERF_HES_STOPPED)
4264 if (event->attr.exclude_user && user_mode(regs))
4267 if (event->attr.exclude_kernel && !user_mode(regs))
4274 static int perf_swevent_match(struct perf_event *event,
4275 enum perf_type_id type,
4277 struct perf_sample_data *data,
4278 struct pt_regs *regs)
4280 if (event->attr.type != type)
4283 if (event->attr.config != event_id)
4286 if (perf_exclude_event(event, regs))
4292 static inline u64 swevent_hash(u64 type, u32 event_id)
4294 u64 val = event_id | (type << 32);
4296 return hash_64(val, SWEVENT_HLIST_BITS);
4299 static inline struct hlist_head *
4300 __find_swevent_head(struct swevent_hlist *hlist, u64 type, u32 event_id)
4302 u64 hash = swevent_hash(type, event_id);
4304 return &hlist->heads[hash];
4307 /* For the read side: events when they trigger */
4308 static inline struct hlist_head *
4309 find_swevent_head_rcu(struct perf_cpu_context *ctx, u64 type, u32 event_id)
4311 struct swevent_hlist *hlist;
4313 hlist = rcu_dereference(ctx->swevent_hlist);
4317 return __find_swevent_head(hlist, type, event_id);
4320 /* For the event head insertion and removal in the hlist */
4321 static inline struct hlist_head *
4322 find_swevent_head(struct perf_cpu_context *ctx, struct perf_event *event)
4324 struct swevent_hlist *hlist;
4325 u32 event_id = event->attr.config;
4326 u64 type = event->attr.type;
4329 * Event scheduling is always serialized against hlist allocation
4330 * and release. Which makes the protected version suitable here.
4331 * The context lock guarantees that.
4333 hlist = rcu_dereference_protected(ctx->swevent_hlist,
4334 lockdep_is_held(&event->ctx->lock));
4338 return __find_swevent_head(hlist, type, event_id);
4341 static void do_perf_sw_event(enum perf_type_id type, u32 event_id,
4343 struct perf_sample_data *data,
4344 struct pt_regs *regs)
4346 struct perf_cpu_context *cpuctx;
4347 struct perf_event *event;
4348 struct hlist_node *node;
4349 struct hlist_head *head;
4351 cpuctx = &__get_cpu_var(perf_cpu_context);
4355 head = find_swevent_head_rcu(cpuctx, type, event_id);
4360 hlist_for_each_entry_rcu(event, node, head, hlist_entry) {
4361 if (perf_swevent_match(event, type, event_id, data, regs))
4362 perf_swevent_event(event, nr, nmi, data, regs);
4368 int perf_swevent_get_recursion_context(void)
4370 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
4372 return get_recursion_context(cpuctx->recursion);
4374 EXPORT_SYMBOL_GPL(perf_swevent_get_recursion_context);
4376 void inline perf_swevent_put_recursion_context(int rctx)
4378 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
4380 put_recursion_context(cpuctx->recursion, rctx);
4383 void __perf_sw_event(u32 event_id, u64 nr, int nmi,
4384 struct pt_regs *regs, u64 addr)
4386 struct perf_sample_data data;
4389 preempt_disable_notrace();
4390 rctx = perf_swevent_get_recursion_context();
4394 perf_sample_data_init(&data, addr);
4396 do_perf_sw_event(PERF_TYPE_SOFTWARE, event_id, nr, nmi, &data, regs);
4398 perf_swevent_put_recursion_context(rctx);
4399 preempt_enable_notrace();
4402 static void perf_swevent_read(struct perf_event *event)
4406 static int perf_swevent_add(struct perf_event *event, int flags)
4408 struct hw_perf_event *hwc = &event->hw;
4409 struct perf_cpu_context *cpuctx;
4410 struct hlist_head *head;
4412 cpuctx = &__get_cpu_var(perf_cpu_context);
4414 if (hwc->sample_period) {
4415 hwc->last_period = hwc->sample_period;
4416 perf_swevent_set_period(event);
4419 hwc->state = !(flags & PERF_EF_START);
4421 head = find_swevent_head(cpuctx, event);
4422 if (WARN_ON_ONCE(!head))
4425 hlist_add_head_rcu(&event->hlist_entry, head);
4430 static void perf_swevent_del(struct perf_event *event, int flags)
4432 hlist_del_rcu(&event->hlist_entry);
4435 static void perf_swevent_start(struct perf_event *event, int flags)
4437 event->hw.state = 0;
4440 static void perf_swevent_stop(struct perf_event *event, int flags)
4442 event->hw.state = PERF_HES_STOPPED;
4445 /* Deref the hlist from the update side */
4446 static inline struct swevent_hlist *
4447 swevent_hlist_deref(struct perf_cpu_context *cpuctx)
4449 return rcu_dereference_protected(cpuctx->swevent_hlist,
4450 lockdep_is_held(&cpuctx->hlist_mutex));
4453 static void swevent_hlist_release_rcu(struct rcu_head *rcu_head)
4455 struct swevent_hlist *hlist;
4457 hlist = container_of(rcu_head, struct swevent_hlist, rcu_head);
4461 static void swevent_hlist_release(struct perf_cpu_context *cpuctx)
4463 struct swevent_hlist *hlist = swevent_hlist_deref(cpuctx);
4468 rcu_assign_pointer(cpuctx->swevent_hlist, NULL);
4469 call_rcu(&hlist->rcu_head, swevent_hlist_release_rcu);
4472 static void swevent_hlist_put_cpu(struct perf_event *event, int cpu)
4474 struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
4476 mutex_lock(&cpuctx->hlist_mutex);
4478 if (!--cpuctx->hlist_refcount)
4479 swevent_hlist_release(cpuctx);
4481 mutex_unlock(&cpuctx->hlist_mutex);
4484 static void swevent_hlist_put(struct perf_event *event)
4488 if (event->cpu != -1) {
4489 swevent_hlist_put_cpu(event, event->cpu);
4493 for_each_possible_cpu(cpu)
4494 swevent_hlist_put_cpu(event, cpu);
4497 static int swevent_hlist_get_cpu(struct perf_event *event, int cpu)
4499 struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
4502 mutex_lock(&cpuctx->hlist_mutex);
4504 if (!swevent_hlist_deref(cpuctx) && cpu_online(cpu)) {
4505 struct swevent_hlist *hlist;
4507 hlist = kzalloc(sizeof(*hlist), GFP_KERNEL);
4512 rcu_assign_pointer(cpuctx->swevent_hlist, hlist);
4514 cpuctx->hlist_refcount++;
4516 mutex_unlock(&cpuctx->hlist_mutex);
4521 static int swevent_hlist_get(struct perf_event *event)
4524 int cpu, failed_cpu;
4526 if (event->cpu != -1)
4527 return swevent_hlist_get_cpu(event, event->cpu);
4530 for_each_possible_cpu(cpu) {
4531 err = swevent_hlist_get_cpu(event, cpu);
4541 for_each_possible_cpu(cpu) {
4542 if (cpu == failed_cpu)
4544 swevent_hlist_put_cpu(event, cpu);
4551 atomic_t perf_swevent_enabled[PERF_COUNT_SW_MAX];
4553 static void sw_perf_event_destroy(struct perf_event *event)
4555 u64 event_id = event->attr.config;
4557 WARN_ON(event->parent);
4559 atomic_dec(&perf_swevent_enabled[event_id]);
4560 swevent_hlist_put(event);
4563 static int perf_swevent_init(struct perf_event *event)
4565 int event_id = event->attr.config;
4567 if (event->attr.type != PERF_TYPE_SOFTWARE)
4571 case PERF_COUNT_SW_CPU_CLOCK:
4572 case PERF_COUNT_SW_TASK_CLOCK:
4579 if (event_id > PERF_COUNT_SW_MAX)
4582 if (!event->parent) {
4585 err = swevent_hlist_get(event);
4589 atomic_inc(&perf_swevent_enabled[event_id]);
4590 event->destroy = sw_perf_event_destroy;
4596 static struct pmu perf_swevent = {
4597 .event_init = perf_swevent_init,
4598 .add = perf_swevent_add,
4599 .del = perf_swevent_del,
4600 .start = perf_swevent_start,
4601 .stop = perf_swevent_stop,
4602 .read = perf_swevent_read,
4605 #ifdef CONFIG_EVENT_TRACING
4607 static int perf_tp_filter_match(struct perf_event *event,
4608 struct perf_sample_data *data)
4610 void *record = data->raw->data;
4612 if (likely(!event->filter) || filter_match_preds(event->filter, record))
4617 static int perf_tp_event_match(struct perf_event *event,
4618 struct perf_sample_data *data,
4619 struct pt_regs *regs)
4622 * All tracepoints are from kernel-space.
4624 if (event->attr.exclude_kernel)
4627 if (!perf_tp_filter_match(event, data))
4633 void perf_tp_event(u64 addr, u64 count, void *record, int entry_size,
4634 struct pt_regs *regs, struct hlist_head *head, int rctx)
4636 struct perf_sample_data data;
4637 struct perf_event *event;
4638 struct hlist_node *node;
4640 struct perf_raw_record raw = {
4645 perf_sample_data_init(&data, addr);
4648 hlist_for_each_entry_rcu(event, node, head, hlist_entry) {
4649 if (perf_tp_event_match(event, &data, regs))
4650 perf_swevent_event(event, count, 1, &data, regs);
4653 perf_swevent_put_recursion_context(rctx);
4655 EXPORT_SYMBOL_GPL(perf_tp_event);
4657 static void tp_perf_event_destroy(struct perf_event *event)
4659 perf_trace_destroy(event);
4662 static int perf_tp_event_init(struct perf_event *event)
4666 if (event->attr.type != PERF_TYPE_TRACEPOINT)
4670 * Raw tracepoint data is a severe data leak, only allow root to
4673 if ((event->attr.sample_type & PERF_SAMPLE_RAW) &&
4674 perf_paranoid_tracepoint_raw() &&
4675 !capable(CAP_SYS_ADMIN))
4678 err = perf_trace_init(event);
4682 event->destroy = tp_perf_event_destroy;
4687 static struct pmu perf_tracepoint = {
4688 .event_init = perf_tp_event_init,
4689 .add = perf_trace_add,
4690 .del = perf_trace_del,
4691 .start = perf_swevent_start,
4692 .stop = perf_swevent_stop,
4693 .read = perf_swevent_read,
4696 static inline void perf_tp_register(void)
4698 perf_pmu_register(&perf_tracepoint);
4701 static int perf_event_set_filter(struct perf_event *event, void __user *arg)
4706 if (event->attr.type != PERF_TYPE_TRACEPOINT)
4709 filter_str = strndup_user(arg, PAGE_SIZE);
4710 if (IS_ERR(filter_str))
4711 return PTR_ERR(filter_str);
4713 ret = ftrace_profile_set_filter(event, event->attr.config, filter_str);
4719 static void perf_event_free_filter(struct perf_event *event)
4721 ftrace_profile_free_filter(event);
4726 static inline void perf_tp_register(void)
4730 static int perf_event_set_filter(struct perf_event *event, void __user *arg)
4735 static void perf_event_free_filter(struct perf_event *event)
4739 #endif /* CONFIG_EVENT_TRACING */
4741 #ifdef CONFIG_HAVE_HW_BREAKPOINT
4742 void perf_bp_event(struct perf_event *bp, void *data)
4744 struct perf_sample_data sample;
4745 struct pt_regs *regs = data;
4747 perf_sample_data_init(&sample, bp->attr.bp_addr);
4749 if (!bp->hw.state && !perf_exclude_event(bp, regs))
4750 perf_swevent_event(bp, 1, 1, &sample, regs);
4755 * hrtimer based swevent callback
4758 static enum hrtimer_restart perf_swevent_hrtimer(struct hrtimer *hrtimer)
4760 enum hrtimer_restart ret = HRTIMER_RESTART;
4761 struct perf_sample_data data;
4762 struct pt_regs *regs;
4763 struct perf_event *event;
4766 event = container_of(hrtimer, struct perf_event, hw.hrtimer);
4767 event->pmu->read(event);
4769 perf_sample_data_init(&data, 0);
4770 data.period = event->hw.last_period;
4771 regs = get_irq_regs();
4773 if (regs && !perf_exclude_event(event, regs)) {
4774 if (!(event->attr.exclude_idle && current->pid == 0))
4775 if (perf_event_overflow(event, 0, &data, regs))
4776 ret = HRTIMER_NORESTART;
4779 period = max_t(u64, 10000, event->hw.sample_period);
4780 hrtimer_forward_now(hrtimer, ns_to_ktime(period));
4785 static void perf_swevent_start_hrtimer(struct perf_event *event)
4787 struct hw_perf_event *hwc = &event->hw;
4789 hrtimer_init(&hwc->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
4790 hwc->hrtimer.function = perf_swevent_hrtimer;
4791 if (hwc->sample_period) {
4792 s64 period = local64_read(&hwc->period_left);
4798 local64_set(&hwc->period_left, 0);
4800 period = max_t(u64, 10000, hwc->sample_period);
4802 __hrtimer_start_range_ns(&hwc->hrtimer,
4803 ns_to_ktime(period), 0,
4804 HRTIMER_MODE_REL, 0);
4808 static void perf_swevent_cancel_hrtimer(struct perf_event *event)
4810 struct hw_perf_event *hwc = &event->hw;
4812 if (hwc->sample_period) {
4813 ktime_t remaining = hrtimer_get_remaining(&hwc->hrtimer);
4814 local64_set(&hwc->period_left, ktime_to_ns(remaining));
4816 hrtimer_cancel(&hwc->hrtimer);
4821 * Software event: cpu wall time clock
4824 static void cpu_clock_event_update(struct perf_event *event)
4829 now = local_clock();
4830 prev = local64_xchg(&event->hw.prev_count, now);
4831 local64_add(now - prev, &event->count);
4834 static void cpu_clock_event_start(struct perf_event *event, int flags)
4836 local64_set(&event->hw.prev_count, local_clock());
4837 perf_swevent_start_hrtimer(event);
4840 static void cpu_clock_event_stop(struct perf_event *event, int flags)
4842 perf_swevent_cancel_hrtimer(event);
4843 cpu_clock_event_update(event);
4846 static int cpu_clock_event_add(struct perf_event *event, int flags)
4848 if (flags & PERF_EF_START)
4849 cpu_clock_event_start(event, flags);
4854 static void cpu_clock_event_del(struct perf_event *event, int flags)
4856 cpu_clock_event_stop(event, flags);
4859 static void cpu_clock_event_read(struct perf_event *event)
4861 cpu_clock_event_update(event);
4864 static int cpu_clock_event_init(struct perf_event *event)
4866 if (event->attr.type != PERF_TYPE_SOFTWARE)
4869 if (event->attr.config != PERF_COUNT_SW_CPU_CLOCK)
4875 static struct pmu perf_cpu_clock = {
4876 .event_init = cpu_clock_event_init,
4877 .add = cpu_clock_event_add,
4878 .del = cpu_clock_event_del,
4879 .start = cpu_clock_event_start,
4880 .stop = cpu_clock_event_stop,
4881 .read = cpu_clock_event_read,
4885 * Software event: task time clock
4888 static void task_clock_event_update(struct perf_event *event, u64 now)
4893 prev = local64_xchg(&event->hw.prev_count, now);
4895 local64_add(delta, &event->count);
4898 static void task_clock_event_start(struct perf_event *event, int flags)
4900 local64_set(&event->hw.prev_count, event->ctx->time);
4901 perf_swevent_start_hrtimer(event);
4904 static void task_clock_event_stop(struct perf_event *event, int flags)
4906 perf_swevent_cancel_hrtimer(event);
4907 task_clock_event_update(event, event->ctx->time);
4910 static int task_clock_event_add(struct perf_event *event, int flags)
4912 if (flags & PERF_EF_START)
4913 task_clock_event_start(event, flags);
4918 static void task_clock_event_del(struct perf_event *event, int flags)
4920 task_clock_event_stop(event, PERF_EF_UPDATE);
4923 static void task_clock_event_read(struct perf_event *event)
4928 update_context_time(event->ctx);
4929 time = event->ctx->time;
4931 u64 now = perf_clock();
4932 u64 delta = now - event->ctx->timestamp;
4933 time = event->ctx->time + delta;
4936 task_clock_event_update(event, time);
4939 static int task_clock_event_init(struct perf_event *event)
4941 if (event->attr.type != PERF_TYPE_SOFTWARE)
4944 if (event->attr.config != PERF_COUNT_SW_TASK_CLOCK)
4950 static struct pmu perf_task_clock = {
4951 .event_init = task_clock_event_init,
4952 .add = task_clock_event_add,
4953 .del = task_clock_event_del,
4954 .start = task_clock_event_start,
4955 .stop = task_clock_event_stop,
4956 .read = task_clock_event_read,
4959 static LIST_HEAD(pmus);
4960 static DEFINE_MUTEX(pmus_lock);
4961 static struct srcu_struct pmus_srcu;
4963 static void perf_pmu_nop_void(struct pmu *pmu)
4967 static int perf_pmu_nop_int(struct pmu *pmu)
4972 static void perf_pmu_start_txn(struct pmu *pmu)
4974 perf_pmu_disable(pmu);
4977 static int perf_pmu_commit_txn(struct pmu *pmu)
4979 perf_pmu_enable(pmu);
4983 static void perf_pmu_cancel_txn(struct pmu *pmu)
4985 perf_pmu_enable(pmu);
4988 int perf_pmu_register(struct pmu *pmu)
4992 mutex_lock(&pmus_lock);
4994 pmu->pmu_disable_count = alloc_percpu(int);
4995 if (!pmu->pmu_disable_count)
4998 if (!pmu->start_txn) {
4999 if (pmu->pmu_enable) {
5001 * If we have pmu_enable/pmu_disable calls, install
5002 * transaction stubs that use that to try and batch
5003 * hardware accesses.
5005 pmu->start_txn = perf_pmu_start_txn;
5006 pmu->commit_txn = perf_pmu_commit_txn;
5007 pmu->cancel_txn = perf_pmu_cancel_txn;
5009 pmu->start_txn = perf_pmu_nop_void;
5010 pmu->commit_txn = perf_pmu_nop_int;
5011 pmu->cancel_txn = perf_pmu_nop_void;
5015 if (!pmu->pmu_enable) {
5016 pmu->pmu_enable = perf_pmu_nop_void;
5017 pmu->pmu_disable = perf_pmu_nop_void;
5020 list_add_rcu(&pmu->entry, &pmus);
5023 mutex_unlock(&pmus_lock);
5028 void perf_pmu_unregister(struct pmu *pmu)
5030 mutex_lock(&pmus_lock);
5031 list_del_rcu(&pmu->entry);
5032 mutex_unlock(&pmus_lock);
5034 synchronize_srcu(&pmus_srcu);
5036 free_percpu(pmu->pmu_disable_count);
5039 struct pmu *perf_init_event(struct perf_event *event)
5041 struct pmu *pmu = NULL;
5044 idx = srcu_read_lock(&pmus_srcu);
5045 list_for_each_entry_rcu(pmu, &pmus, entry) {
5046 int ret = pmu->event_init(event);
5049 if (ret != -ENOENT) {
5054 srcu_read_unlock(&pmus_srcu, idx);
5060 * Allocate and initialize a event structure
5062 static struct perf_event *
5063 perf_event_alloc(struct perf_event_attr *attr,
5065 struct perf_event_context *ctx,
5066 struct perf_event *group_leader,
5067 struct perf_event *parent_event,
5068 perf_overflow_handler_t overflow_handler,
5072 struct perf_event *event;
5073 struct hw_perf_event *hwc;
5076 event = kzalloc(sizeof(*event), gfpflags);
5078 return ERR_PTR(-ENOMEM);
5081 * Single events are their own group leaders, with an
5082 * empty sibling list:
5085 group_leader = event;
5087 mutex_init(&event->child_mutex);
5088 INIT_LIST_HEAD(&event->child_list);
5090 INIT_LIST_HEAD(&event->group_entry);
5091 INIT_LIST_HEAD(&event->event_entry);
5092 INIT_LIST_HEAD(&event->sibling_list);
5093 init_waitqueue_head(&event->waitq);
5095 mutex_init(&event->mmap_mutex);
5098 event->attr = *attr;
5099 event->group_leader = group_leader;
5104 event->parent = parent_event;
5106 event->ns = get_pid_ns(current->nsproxy->pid_ns);
5107 event->id = atomic64_inc_return(&perf_event_id);
5109 event->state = PERF_EVENT_STATE_INACTIVE;
5111 if (!overflow_handler && parent_event)
5112 overflow_handler = parent_event->overflow_handler;
5114 event->overflow_handler = overflow_handler;
5117 event->state = PERF_EVENT_STATE_OFF;
5122 hwc->sample_period = attr->sample_period;
5123 if (attr->freq && attr->sample_freq)
5124 hwc->sample_period = 1;
5125 hwc->last_period = hwc->sample_period;
5127 local64_set(&hwc->period_left, hwc->sample_period);
5130 * we currently do not support PERF_FORMAT_GROUP on inherited events
5132 if (attr->inherit && (attr->read_format & PERF_FORMAT_GROUP))
5135 pmu = perf_init_event(event);
5141 else if (IS_ERR(pmu))
5146 put_pid_ns(event->ns);
5148 return ERR_PTR(err);
5153 if (!event->parent) {
5154 atomic_inc(&nr_events);
5155 if (event->attr.mmap || event->attr.mmap_data)
5156 atomic_inc(&nr_mmap_events);
5157 if (event->attr.comm)
5158 atomic_inc(&nr_comm_events);
5159 if (event->attr.task)
5160 atomic_inc(&nr_task_events);
5161 if (event->attr.sample_type & PERF_SAMPLE_CALLCHAIN) {
5162 err = get_callchain_buffers();
5165 return ERR_PTR(err);
5173 static int perf_copy_attr(struct perf_event_attr __user *uattr,
5174 struct perf_event_attr *attr)
5179 if (!access_ok(VERIFY_WRITE, uattr, PERF_ATTR_SIZE_VER0))
5183 * zero the full structure, so that a short copy will be nice.
5185 memset(attr, 0, sizeof(*attr));
5187 ret = get_user(size, &uattr->size);
5191 if (size > PAGE_SIZE) /* silly large */
5194 if (!size) /* abi compat */
5195 size = PERF_ATTR_SIZE_VER0;
5197 if (size < PERF_ATTR_SIZE_VER0)
5201 * If we're handed a bigger struct than we know of,
5202 * ensure all the unknown bits are 0 - i.e. new
5203 * user-space does not rely on any kernel feature
5204 * extensions we dont know about yet.
5206 if (size > sizeof(*attr)) {
5207 unsigned char __user *addr;
5208 unsigned char __user *end;
5211 addr = (void __user *)uattr + sizeof(*attr);
5212 end = (void __user *)uattr + size;
5214 for (; addr < end; addr++) {
5215 ret = get_user(val, addr);
5221 size = sizeof(*attr);
5224 ret = copy_from_user(attr, uattr, size);
5229 * If the type exists, the corresponding creation will verify
5232 if (attr->type >= PERF_TYPE_MAX)
5235 if (attr->__reserved_1)
5238 if (attr->sample_type & ~(PERF_SAMPLE_MAX-1))
5241 if (attr->read_format & ~(PERF_FORMAT_MAX-1))
5248 put_user(sizeof(*attr), &uattr->size);
5254 perf_event_set_output(struct perf_event *event, struct perf_event *output_event)
5256 struct perf_buffer *buffer = NULL, *old_buffer = NULL;
5262 /* don't allow circular references */
5263 if (event == output_event)
5267 * Don't allow cross-cpu buffers
5269 if (output_event->cpu != event->cpu)
5273 * If its not a per-cpu buffer, it must be the same task.
5275 if (output_event->cpu == -1 && output_event->ctx != event->ctx)
5279 mutex_lock(&event->mmap_mutex);
5280 /* Can't redirect output if we've got an active mmap() */
5281 if (atomic_read(&event->mmap_count))
5285 /* get the buffer we want to redirect to */
5286 buffer = perf_buffer_get(output_event);
5291 old_buffer = event->buffer;
5292 rcu_assign_pointer(event->buffer, buffer);
5295 mutex_unlock(&event->mmap_mutex);
5298 perf_buffer_put(old_buffer);
5304 * sys_perf_event_open - open a performance event, associate it to a task/cpu
5306 * @attr_uptr: event_id type attributes for monitoring/sampling
5309 * @group_fd: group leader event fd
5311 SYSCALL_DEFINE5(perf_event_open,
5312 struct perf_event_attr __user *, attr_uptr,
5313 pid_t, pid, int, cpu, int, group_fd, unsigned long, flags)
5315 struct perf_event *event, *group_leader = NULL, *output_event = NULL;
5316 struct perf_event_attr attr;
5317 struct perf_event_context *ctx;
5318 struct file *event_file = NULL;
5319 struct file *group_file = NULL;
5321 int fput_needed = 0;
5324 /* for future expandability... */
5325 if (flags & ~(PERF_FLAG_FD_NO_GROUP | PERF_FLAG_FD_OUTPUT))
5328 err = perf_copy_attr(attr_uptr, &attr);
5332 if (!attr.exclude_kernel) {
5333 if (perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN))
5338 if (attr.sample_freq > sysctl_perf_event_sample_rate)
5342 event_fd = get_unused_fd_flags(O_RDWR);
5347 * Get the target context (task or percpu):
5349 ctx = find_get_context(pid, cpu);
5355 if (group_fd != -1) {
5356 group_leader = perf_fget_light(group_fd, &fput_needed);
5357 if (IS_ERR(group_leader)) {
5358 err = PTR_ERR(group_leader);
5359 goto err_put_context;
5361 group_file = group_leader->filp;
5362 if (flags & PERF_FLAG_FD_OUTPUT)
5363 output_event = group_leader;
5364 if (flags & PERF_FLAG_FD_NO_GROUP)
5365 group_leader = NULL;
5369 * Look up the group leader (we will attach this event to it):
5375 * Do not allow a recursive hierarchy (this new sibling
5376 * becoming part of another group-sibling):
5378 if (group_leader->group_leader != group_leader)
5379 goto err_put_context;
5381 * Do not allow to attach to a group in a different
5382 * task or CPU context:
5384 if (group_leader->ctx != ctx)
5385 goto err_put_context;
5387 * Only a group leader can be exclusive or pinned
5389 if (attr.exclusive || attr.pinned)
5390 goto err_put_context;
5393 event = perf_event_alloc(&attr, cpu, ctx, group_leader,
5394 NULL, NULL, GFP_KERNEL);
5395 if (IS_ERR(event)) {
5396 err = PTR_ERR(event);
5397 goto err_put_context;
5401 err = perf_event_set_output(event, output_event);
5403 goto err_free_put_context;
5406 event_file = anon_inode_getfile("[perf_event]", &perf_fops, event, O_RDWR);
5407 if (IS_ERR(event_file)) {
5408 err = PTR_ERR(event_file);
5409 goto err_free_put_context;
5412 event->filp = event_file;
5413 WARN_ON_ONCE(ctx->parent_ctx);
5414 mutex_lock(&ctx->mutex);
5415 perf_install_in_context(ctx, event, cpu);
5417 mutex_unlock(&ctx->mutex);
5419 event->owner = current;
5420 get_task_struct(current);
5421 mutex_lock(¤t->perf_event_mutex);
5422 list_add_tail(&event->owner_entry, ¤t->perf_event_list);
5423 mutex_unlock(¤t->perf_event_mutex);
5426 * Drop the reference on the group_event after placing the
5427 * new event on the sibling_list. This ensures destruction
5428 * of the group leader will find the pointer to itself in
5429 * perf_group_detach().
5431 fput_light(group_file, fput_needed);
5432 fd_install(event_fd, event_file);
5435 err_free_put_context:
5438 fput_light(group_file, fput_needed);
5441 put_unused_fd(event_fd);
5446 * perf_event_create_kernel_counter
5448 * @attr: attributes of the counter to create
5449 * @cpu: cpu in which the counter is bound
5450 * @pid: task to profile
5453 perf_event_create_kernel_counter(struct perf_event_attr *attr, int cpu,
5455 perf_overflow_handler_t overflow_handler)
5457 struct perf_event *event;
5458 struct perf_event_context *ctx;
5462 * Get the target context (task or percpu):
5465 ctx = find_get_context(pid, cpu);
5471 event = perf_event_alloc(attr, cpu, ctx, NULL,
5472 NULL, overflow_handler, GFP_KERNEL);
5473 if (IS_ERR(event)) {
5474 err = PTR_ERR(event);
5475 goto err_put_context;
5479 WARN_ON_ONCE(ctx->parent_ctx);
5480 mutex_lock(&ctx->mutex);
5481 perf_install_in_context(ctx, event, cpu);
5483 mutex_unlock(&ctx->mutex);
5485 event->owner = current;
5486 get_task_struct(current);
5487 mutex_lock(¤t->perf_event_mutex);
5488 list_add_tail(&event->owner_entry, ¤t->perf_event_list);
5489 mutex_unlock(¤t->perf_event_mutex);
5496 return ERR_PTR(err);
5498 EXPORT_SYMBOL_GPL(perf_event_create_kernel_counter);
5501 * inherit a event from parent task to child task:
5503 static struct perf_event *
5504 inherit_event(struct perf_event *parent_event,
5505 struct task_struct *parent,
5506 struct perf_event_context *parent_ctx,
5507 struct task_struct *child,
5508 struct perf_event *group_leader,
5509 struct perf_event_context *child_ctx)
5511 struct perf_event *child_event;
5514 * Instead of creating recursive hierarchies of events,
5515 * we link inherited events back to the original parent,
5516 * which has a filp for sure, which we use as the reference
5519 if (parent_event->parent)
5520 parent_event = parent_event->parent;
5522 child_event = perf_event_alloc(&parent_event->attr,
5523 parent_event->cpu, child_ctx,
5524 group_leader, parent_event,
5526 if (IS_ERR(child_event))
5531 * Make the child state follow the state of the parent event,
5532 * not its attr.disabled bit. We hold the parent's mutex,
5533 * so we won't race with perf_event_{en, dis}able_family.
5535 if (parent_event->state >= PERF_EVENT_STATE_INACTIVE)
5536 child_event->state = PERF_EVENT_STATE_INACTIVE;
5538 child_event->state = PERF_EVENT_STATE_OFF;
5540 if (parent_event->attr.freq) {
5541 u64 sample_period = parent_event->hw.sample_period;
5542 struct hw_perf_event *hwc = &child_event->hw;
5544 hwc->sample_period = sample_period;
5545 hwc->last_period = sample_period;
5547 local64_set(&hwc->period_left, sample_period);
5550 child_event->overflow_handler = parent_event->overflow_handler;
5553 * Link it up in the child's context:
5555 add_event_to_ctx(child_event, child_ctx);
5558 * Get a reference to the parent filp - we will fput it
5559 * when the child event exits. This is safe to do because
5560 * we are in the parent and we know that the filp still
5561 * exists and has a nonzero count:
5563 atomic_long_inc(&parent_event->filp->f_count);
5566 * Link this into the parent event's child list
5568 WARN_ON_ONCE(parent_event->ctx->parent_ctx);
5569 mutex_lock(&parent_event->child_mutex);
5570 list_add_tail(&child_event->child_list, &parent_event->child_list);
5571 mutex_unlock(&parent_event->child_mutex);
5576 static int inherit_group(struct perf_event *parent_event,
5577 struct task_struct *parent,
5578 struct perf_event_context *parent_ctx,
5579 struct task_struct *child,
5580 struct perf_event_context *child_ctx)
5582 struct perf_event *leader;
5583 struct perf_event *sub;
5584 struct perf_event *child_ctr;
5586 leader = inherit_event(parent_event, parent, parent_ctx,
5587 child, NULL, child_ctx);
5589 return PTR_ERR(leader);
5590 list_for_each_entry(sub, &parent_event->sibling_list, group_entry) {
5591 child_ctr = inherit_event(sub, parent, parent_ctx,
5592 child, leader, child_ctx);
5593 if (IS_ERR(child_ctr))
5594 return PTR_ERR(child_ctr);
5599 static void sync_child_event(struct perf_event *child_event,
5600 struct task_struct *child)
5602 struct perf_event *parent_event = child_event->parent;
5605 if (child_event->attr.inherit_stat)
5606 perf_event_read_event(child_event, child);
5608 child_val = perf_event_count(child_event);
5611 * Add back the child's count to the parent's count:
5613 atomic64_add(child_val, &parent_event->child_count);
5614 atomic64_add(child_event->total_time_enabled,
5615 &parent_event->child_total_time_enabled);
5616 atomic64_add(child_event->total_time_running,
5617 &parent_event->child_total_time_running);
5620 * Remove this event from the parent's list
5622 WARN_ON_ONCE(parent_event->ctx->parent_ctx);
5623 mutex_lock(&parent_event->child_mutex);
5624 list_del_init(&child_event->child_list);
5625 mutex_unlock(&parent_event->child_mutex);
5628 * Release the parent event, if this was the last
5631 fput(parent_event->filp);
5635 __perf_event_exit_task(struct perf_event *child_event,
5636 struct perf_event_context *child_ctx,
5637 struct task_struct *child)
5639 struct perf_event *parent_event;
5641 perf_event_remove_from_context(child_event);
5643 parent_event = child_event->parent;
5645 * It can happen that parent exits first, and has events
5646 * that are still around due to the child reference. These
5647 * events need to be zapped - but otherwise linger.
5650 sync_child_event(child_event, child);
5651 free_event(child_event);
5656 * When a child task exits, feed back event values to parent events.
5658 void perf_event_exit_task(struct task_struct *child)
5660 struct perf_event *child_event, *tmp;
5661 struct perf_event_context *child_ctx;
5662 unsigned long flags;
5664 if (likely(!child->perf_event_ctxp)) {
5665 perf_event_task(child, NULL, 0);
5669 local_irq_save(flags);
5671 * We can't reschedule here because interrupts are disabled,
5672 * and either child is current or it is a task that can't be
5673 * scheduled, so we are now safe from rescheduling changing
5676 child_ctx = child->perf_event_ctxp;
5677 __perf_event_task_sched_out(child_ctx);
5680 * Take the context lock here so that if find_get_context is
5681 * reading child->perf_event_ctxp, we wait until it has
5682 * incremented the context's refcount before we do put_ctx below.
5684 raw_spin_lock(&child_ctx->lock);
5685 child->perf_event_ctxp = NULL;
5687 * If this context is a clone; unclone it so it can't get
5688 * swapped to another process while we're removing all
5689 * the events from it.
5691 unclone_ctx(child_ctx);
5692 update_context_time(child_ctx);
5693 raw_spin_unlock_irqrestore(&child_ctx->lock, flags);
5696 * Report the task dead after unscheduling the events so that we
5697 * won't get any samples after PERF_RECORD_EXIT. We can however still
5698 * get a few PERF_RECORD_READ events.
5700 perf_event_task(child, child_ctx, 0);
5703 * We can recurse on the same lock type through:
5705 * __perf_event_exit_task()
5706 * sync_child_event()
5707 * fput(parent_event->filp)
5709 * mutex_lock(&ctx->mutex)
5711 * But since its the parent context it won't be the same instance.
5713 mutex_lock(&child_ctx->mutex);
5716 list_for_each_entry_safe(child_event, tmp, &child_ctx->pinned_groups,
5718 __perf_event_exit_task(child_event, child_ctx, child);
5720 list_for_each_entry_safe(child_event, tmp, &child_ctx->flexible_groups,
5722 __perf_event_exit_task(child_event, child_ctx, child);
5725 * If the last event was a group event, it will have appended all
5726 * its siblings to the list, but we obtained 'tmp' before that which
5727 * will still point to the list head terminating the iteration.
5729 if (!list_empty(&child_ctx->pinned_groups) ||
5730 !list_empty(&child_ctx->flexible_groups))
5733 mutex_unlock(&child_ctx->mutex);
5738 static void perf_free_event(struct perf_event *event,
5739 struct perf_event_context *ctx)
5741 struct perf_event *parent = event->parent;
5743 if (WARN_ON_ONCE(!parent))
5746 mutex_lock(&parent->child_mutex);
5747 list_del_init(&event->child_list);
5748 mutex_unlock(&parent->child_mutex);
5752 perf_group_detach(event);
5753 list_del_event(event, ctx);
5758 * free an unexposed, unused context as created by inheritance by
5759 * init_task below, used by fork() in case of fail.
5761 void perf_event_free_task(struct task_struct *task)
5763 struct perf_event_context *ctx = task->perf_event_ctxp;
5764 struct perf_event *event, *tmp;
5769 mutex_lock(&ctx->mutex);
5771 list_for_each_entry_safe(event, tmp, &ctx->pinned_groups, group_entry)
5772 perf_free_event(event, ctx);
5774 list_for_each_entry_safe(event, tmp, &ctx->flexible_groups,
5776 perf_free_event(event, ctx);
5778 if (!list_empty(&ctx->pinned_groups) ||
5779 !list_empty(&ctx->flexible_groups))
5782 mutex_unlock(&ctx->mutex);
5788 inherit_task_group(struct perf_event *event, struct task_struct *parent,
5789 struct perf_event_context *parent_ctx,
5790 struct task_struct *child,
5794 struct perf_event_context *child_ctx = child->perf_event_ctxp;
5796 if (!event->attr.inherit) {
5803 * This is executed from the parent task context, so
5804 * inherit events that have been marked for cloning.
5805 * First allocate and initialize a context for the
5809 child_ctx = kzalloc(sizeof(struct perf_event_context),
5814 __perf_event_init_context(child_ctx, child);
5815 child->perf_event_ctxp = child_ctx;
5816 get_task_struct(child);
5819 ret = inherit_group(event, parent, parent_ctx,
5830 * Initialize the perf_event context in task_struct
5832 int perf_event_init_task(struct task_struct *child)
5834 struct perf_event_context *child_ctx, *parent_ctx;
5835 struct perf_event_context *cloned_ctx;
5836 struct perf_event *event;
5837 struct task_struct *parent = current;
5838 int inherited_all = 1;
5841 child->perf_event_ctxp = NULL;
5843 mutex_init(&child->perf_event_mutex);
5844 INIT_LIST_HEAD(&child->perf_event_list);
5846 if (likely(!parent->perf_event_ctxp))
5850 * If the parent's context is a clone, pin it so it won't get
5853 parent_ctx = perf_pin_task_context(parent);
5856 * No need to check if parent_ctx != NULL here; since we saw
5857 * it non-NULL earlier, the only reason for it to become NULL
5858 * is if we exit, and since we're currently in the middle of
5859 * a fork we can't be exiting at the same time.
5863 * Lock the parent list. No need to lock the child - not PID
5864 * hashed yet and not running, so nobody can access it.
5866 mutex_lock(&parent_ctx->mutex);
5869 * We dont have to disable NMIs - we are only looking at
5870 * the list, not manipulating it:
5872 list_for_each_entry(event, &parent_ctx->pinned_groups, group_entry) {
5873 ret = inherit_task_group(event, parent, parent_ctx, child,
5879 list_for_each_entry(event, &parent_ctx->flexible_groups, group_entry) {
5880 ret = inherit_task_group(event, parent, parent_ctx, child,
5886 child_ctx = child->perf_event_ctxp;
5888 if (child_ctx && inherited_all) {
5890 * Mark the child context as a clone of the parent
5891 * context, or of whatever the parent is a clone of.
5892 * Note that if the parent is a clone, it could get
5893 * uncloned at any point, but that doesn't matter
5894 * because the list of events and the generation
5895 * count can't have changed since we took the mutex.
5897 cloned_ctx = rcu_dereference(parent_ctx->parent_ctx);
5899 child_ctx->parent_ctx = cloned_ctx;
5900 child_ctx->parent_gen = parent_ctx->parent_gen;
5902 child_ctx->parent_ctx = parent_ctx;
5903 child_ctx->parent_gen = parent_ctx->generation;
5905 get_ctx(child_ctx->parent_ctx);
5908 mutex_unlock(&parent_ctx->mutex);
5910 perf_unpin_context(parent_ctx);
5915 static void __init perf_event_init_all_cpus(void)
5918 struct perf_cpu_context *cpuctx;
5920 for_each_possible_cpu(cpu) {
5921 cpuctx = &per_cpu(perf_cpu_context, cpu);
5922 mutex_init(&cpuctx->hlist_mutex);
5923 __perf_event_init_context(&cpuctx->ctx, NULL);
5927 static void __cpuinit perf_event_init_cpu(int cpu)
5929 struct perf_cpu_context *cpuctx;
5931 cpuctx = &per_cpu(perf_cpu_context, cpu);
5933 spin_lock(&perf_resource_lock);
5934 cpuctx->max_pertask = perf_max_events - perf_reserved_percpu;
5935 spin_unlock(&perf_resource_lock);
5937 mutex_lock(&cpuctx->hlist_mutex);
5938 if (cpuctx->hlist_refcount > 0) {
5939 struct swevent_hlist *hlist;
5941 hlist = kzalloc(sizeof(*hlist), GFP_KERNEL);
5942 WARN_ON_ONCE(!hlist);
5943 rcu_assign_pointer(cpuctx->swevent_hlist, hlist);
5945 mutex_unlock(&cpuctx->hlist_mutex);
5948 #ifdef CONFIG_HOTPLUG_CPU
5949 static void __perf_event_exit_cpu(void *info)
5951 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
5952 struct perf_event_context *ctx = &cpuctx->ctx;
5953 struct perf_event *event, *tmp;
5955 list_for_each_entry_safe(event, tmp, &ctx->pinned_groups, group_entry)
5956 __perf_event_remove_from_context(event);
5957 list_for_each_entry_safe(event, tmp, &ctx->flexible_groups, group_entry)
5958 __perf_event_remove_from_context(event);
5960 static void perf_event_exit_cpu(int cpu)
5962 struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
5963 struct perf_event_context *ctx = &cpuctx->ctx;
5965 mutex_lock(&cpuctx->hlist_mutex);
5966 swevent_hlist_release(cpuctx);
5967 mutex_unlock(&cpuctx->hlist_mutex);
5969 mutex_lock(&ctx->mutex);
5970 smp_call_function_single(cpu, __perf_event_exit_cpu, NULL, 1);
5971 mutex_unlock(&ctx->mutex);
5974 static inline void perf_event_exit_cpu(int cpu) { }
5977 static int __cpuinit
5978 perf_cpu_notify(struct notifier_block *self, unsigned long action, void *hcpu)
5980 unsigned int cpu = (long)hcpu;
5982 switch (action & ~CPU_TASKS_FROZEN) {
5984 case CPU_UP_PREPARE:
5985 case CPU_DOWN_FAILED:
5986 perf_event_init_cpu(cpu);
5989 case CPU_UP_CANCELED:
5990 case CPU_DOWN_PREPARE:
5991 perf_event_exit_cpu(cpu);
6001 void __init perf_event_init(void)
6003 perf_event_init_all_cpus();
6004 init_srcu_struct(&pmus_srcu);
6005 perf_pmu_register(&perf_swevent);
6006 perf_pmu_register(&perf_cpu_clock);
6007 perf_pmu_register(&perf_task_clock);
6009 perf_cpu_notifier(perf_cpu_notify);
6012 static ssize_t perf_show_reserve_percpu(struct sysdev_class *class,
6013 struct sysdev_class_attribute *attr,
6016 return sprintf(buf, "%d\n", perf_reserved_percpu);
6020 perf_set_reserve_percpu(struct sysdev_class *class,
6021 struct sysdev_class_attribute *attr,
6025 struct perf_cpu_context *cpuctx;
6029 err = strict_strtoul(buf, 10, &val);
6032 if (val > perf_max_events)
6035 spin_lock(&perf_resource_lock);
6036 perf_reserved_percpu = val;
6037 for_each_online_cpu(cpu) {
6038 cpuctx = &per_cpu(perf_cpu_context, cpu);
6039 raw_spin_lock_irq(&cpuctx->ctx.lock);
6040 mpt = min(perf_max_events - cpuctx->ctx.nr_events,
6041 perf_max_events - perf_reserved_percpu);
6042 cpuctx->max_pertask = mpt;
6043 raw_spin_unlock_irq(&cpuctx->ctx.lock);
6045 spin_unlock(&perf_resource_lock);
6050 static ssize_t perf_show_overcommit(struct sysdev_class *class,
6051 struct sysdev_class_attribute *attr,
6054 return sprintf(buf, "%d\n", perf_overcommit);
6058 perf_set_overcommit(struct sysdev_class *class,
6059 struct sysdev_class_attribute *attr,
6060 const char *buf, size_t count)
6065 err = strict_strtoul(buf, 10, &val);
6071 spin_lock(&perf_resource_lock);
6072 perf_overcommit = val;
6073 spin_unlock(&perf_resource_lock);
6078 static SYSDEV_CLASS_ATTR(
6081 perf_show_reserve_percpu,
6082 perf_set_reserve_percpu
6085 static SYSDEV_CLASS_ATTR(
6088 perf_show_overcommit,
6092 static struct attribute *perfclass_attrs[] = {
6093 &attr_reserve_percpu.attr,
6094 &attr_overcommit.attr,
6098 static struct attribute_group perfclass_attr_group = {
6099 .attrs = perfclass_attrs,
6100 .name = "perf_events",
6103 static int __init perf_event_sysfs_init(void)
6105 return sysfs_create_group(&cpu_sysdev_class.kset.kobj,
6106 &perfclass_attr_group);
6108 device_initcall(perf_event_sysfs_init);