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>
34 #include <linux/hw_breakpoint.h>
36 #include <asm/irq_regs.h>
39 * Each CPU has a list of per CPU events:
41 static DEFINE_PER_CPU(struct perf_cpu_context, perf_cpu_context);
43 int perf_max_events __read_mostly = 1;
44 static int perf_reserved_percpu __read_mostly;
45 static int perf_overcommit __read_mostly = 1;
47 static atomic_t nr_events __read_mostly;
48 static atomic_t nr_mmap_events __read_mostly;
49 static atomic_t nr_comm_events __read_mostly;
50 static atomic_t nr_task_events __read_mostly;
53 * perf event paranoia level:
54 * -1 - not paranoid at all
55 * 0 - disallow raw tracepoint access for unpriv
56 * 1 - disallow cpu events for unpriv
57 * 2 - disallow kernel profiling for unpriv
59 int sysctl_perf_event_paranoid __read_mostly = 1;
61 int sysctl_perf_event_mlock __read_mostly = 512; /* 'free' kb per user */
64 * max perf event sample rate
66 int sysctl_perf_event_sample_rate __read_mostly = 100000;
68 static atomic64_t perf_event_id;
71 * Lock for (sysadmin-configurable) event reservations:
73 static DEFINE_SPINLOCK(perf_resource_lock);
76 * Architecture provided APIs - weak aliases:
78 extern __weak const struct pmu *hw_perf_event_init(struct perf_event *event)
83 void __weak hw_perf_disable(void) { barrier(); }
84 void __weak hw_perf_enable(void) { barrier(); }
86 void __weak perf_event_print_debug(void) { }
88 static DEFINE_PER_CPU(int, perf_disable_count);
90 void perf_disable(void)
92 if (!__get_cpu_var(perf_disable_count)++)
96 void perf_enable(void)
98 if (!--__get_cpu_var(perf_disable_count))
102 static void get_ctx(struct perf_event_context *ctx)
104 WARN_ON(!atomic_inc_not_zero(&ctx->refcount));
107 static void free_ctx(struct rcu_head *head)
109 struct perf_event_context *ctx;
111 ctx = container_of(head, struct perf_event_context, rcu_head);
115 static void put_ctx(struct perf_event_context *ctx)
117 if (atomic_dec_and_test(&ctx->refcount)) {
119 put_ctx(ctx->parent_ctx);
121 put_task_struct(ctx->task);
122 call_rcu(&ctx->rcu_head, free_ctx);
126 static void unclone_ctx(struct perf_event_context *ctx)
128 if (ctx->parent_ctx) {
129 put_ctx(ctx->parent_ctx);
130 ctx->parent_ctx = NULL;
135 * If we inherit events we want to return the parent event id
138 static u64 primary_event_id(struct perf_event *event)
143 id = event->parent->id;
149 * Get the perf_event_context for a task and lock it.
150 * This has to cope with with the fact that until it is locked,
151 * the context could get moved to another task.
153 static struct perf_event_context *
154 perf_lock_task_context(struct task_struct *task, unsigned long *flags)
156 struct perf_event_context *ctx;
160 ctx = rcu_dereference(task->perf_event_ctxp);
163 * If this context is a clone of another, it might
164 * get swapped for another underneath us by
165 * perf_event_task_sched_out, though the
166 * rcu_read_lock() protects us from any context
167 * getting freed. Lock the context and check if it
168 * got swapped before we could get the lock, and retry
169 * if so. If we locked the right context, then it
170 * can't get swapped on us any more.
172 raw_spin_lock_irqsave(&ctx->lock, *flags);
173 if (ctx != rcu_dereference(task->perf_event_ctxp)) {
174 raw_spin_unlock_irqrestore(&ctx->lock, *flags);
178 if (!atomic_inc_not_zero(&ctx->refcount)) {
179 raw_spin_unlock_irqrestore(&ctx->lock, *flags);
188 * Get the context for a task and increment its pin_count so it
189 * can't get swapped to another task. This also increments its
190 * reference count so that the context can't get freed.
192 static struct perf_event_context *perf_pin_task_context(struct task_struct *task)
194 struct perf_event_context *ctx;
197 ctx = perf_lock_task_context(task, &flags);
200 raw_spin_unlock_irqrestore(&ctx->lock, flags);
205 static void perf_unpin_context(struct perf_event_context *ctx)
209 raw_spin_lock_irqsave(&ctx->lock, flags);
211 raw_spin_unlock_irqrestore(&ctx->lock, flags);
215 static inline u64 perf_clock(void)
217 return local_clock();
221 * Update the record of the current time in a context.
223 static void update_context_time(struct perf_event_context *ctx)
225 u64 now = perf_clock();
227 ctx->time += now - ctx->timestamp;
228 ctx->timestamp = now;
232 * Update the total_time_enabled and total_time_running fields for a event.
234 static void update_event_times(struct perf_event *event)
236 struct perf_event_context *ctx = event->ctx;
239 if (event->state < PERF_EVENT_STATE_INACTIVE ||
240 event->group_leader->state < PERF_EVENT_STATE_INACTIVE)
246 run_end = event->tstamp_stopped;
248 event->total_time_enabled = run_end - event->tstamp_enabled;
250 if (event->state == PERF_EVENT_STATE_INACTIVE)
251 run_end = event->tstamp_stopped;
255 event->total_time_running = run_end - event->tstamp_running;
259 * Update total_time_enabled and total_time_running for all events in a group.
261 static void update_group_times(struct perf_event *leader)
263 struct perf_event *event;
265 update_event_times(leader);
266 list_for_each_entry(event, &leader->sibling_list, group_entry)
267 update_event_times(event);
270 static struct list_head *
271 ctx_group_list(struct perf_event *event, struct perf_event_context *ctx)
273 if (event->attr.pinned)
274 return &ctx->pinned_groups;
276 return &ctx->flexible_groups;
280 * Add a event from the lists for its context.
281 * Must be called with ctx->mutex and ctx->lock held.
284 list_add_event(struct perf_event *event, struct perf_event_context *ctx)
286 WARN_ON_ONCE(event->attach_state & PERF_ATTACH_CONTEXT);
287 event->attach_state |= PERF_ATTACH_CONTEXT;
290 * If we're a stand alone event or group leader, we go to the context
291 * list, group events are kept attached to the group so that
292 * perf_group_detach can, at all times, locate all siblings.
294 if (event->group_leader == event) {
295 struct list_head *list;
297 if (is_software_event(event))
298 event->group_flags |= PERF_GROUP_SOFTWARE;
300 list = ctx_group_list(event, ctx);
301 list_add_tail(&event->group_entry, list);
304 list_add_rcu(&event->event_entry, &ctx->event_list);
306 if (event->attr.inherit_stat)
310 static void perf_group_attach(struct perf_event *event)
312 struct perf_event *group_leader = event->group_leader;
314 WARN_ON_ONCE(event->attach_state & PERF_ATTACH_GROUP);
315 event->attach_state |= PERF_ATTACH_GROUP;
317 if (group_leader == event)
320 if (group_leader->group_flags & PERF_GROUP_SOFTWARE &&
321 !is_software_event(event))
322 group_leader->group_flags &= ~PERF_GROUP_SOFTWARE;
324 list_add_tail(&event->group_entry, &group_leader->sibling_list);
325 group_leader->nr_siblings++;
329 * Remove a event from the lists for its context.
330 * Must be called with ctx->mutex and ctx->lock held.
333 list_del_event(struct perf_event *event, struct perf_event_context *ctx)
336 * We can have double detach due to exit/hot-unplug + close.
338 if (!(event->attach_state & PERF_ATTACH_CONTEXT))
341 event->attach_state &= ~PERF_ATTACH_CONTEXT;
344 if (event->attr.inherit_stat)
347 list_del_rcu(&event->event_entry);
349 if (event->group_leader == event)
350 list_del_init(&event->group_entry);
352 update_group_times(event);
355 * If event was in error state, then keep it
356 * that way, otherwise bogus counts will be
357 * returned on read(). The only way to get out
358 * of error state is by explicit re-enabling
361 if (event->state > PERF_EVENT_STATE_OFF)
362 event->state = PERF_EVENT_STATE_OFF;
365 static void perf_group_detach(struct perf_event *event)
367 struct perf_event *sibling, *tmp;
368 struct list_head *list = NULL;
371 * We can have double detach due to exit/hot-unplug + close.
373 if (!(event->attach_state & PERF_ATTACH_GROUP))
376 event->attach_state &= ~PERF_ATTACH_GROUP;
379 * If this is a sibling, remove it from its group.
381 if (event->group_leader != event) {
382 list_del_init(&event->group_entry);
383 event->group_leader->nr_siblings--;
387 if (!list_empty(&event->group_entry))
388 list = &event->group_entry;
391 * If this was a group event with sibling events then
392 * upgrade the siblings to singleton events by adding them
393 * to whatever list we are on.
395 list_for_each_entry_safe(sibling, tmp, &event->sibling_list, group_entry) {
397 list_move_tail(&sibling->group_entry, list);
398 sibling->group_leader = sibling;
400 /* Inherit group flags from the previous leader */
401 sibling->group_flags = event->group_flags;
406 event_filter_match(struct perf_event *event)
408 return event->cpu == -1 || event->cpu == smp_processor_id();
412 event_sched_out(struct perf_event *event,
413 struct perf_cpu_context *cpuctx,
414 struct perf_event_context *ctx)
418 * An event which could not be activated because of
419 * filter mismatch still needs to have its timings
420 * maintained, otherwise bogus information is return
421 * via read() for time_enabled, time_running:
423 if (event->state == PERF_EVENT_STATE_INACTIVE
424 && !event_filter_match(event)) {
425 delta = ctx->time - event->tstamp_stopped;
426 event->tstamp_running += delta;
427 event->tstamp_stopped = ctx->time;
430 if (event->state != PERF_EVENT_STATE_ACTIVE)
433 event->state = PERF_EVENT_STATE_INACTIVE;
434 if (event->pending_disable) {
435 event->pending_disable = 0;
436 event->state = PERF_EVENT_STATE_OFF;
438 event->tstamp_stopped = ctx->time;
439 event->pmu->disable(event);
442 if (!is_software_event(event))
443 cpuctx->active_oncpu--;
445 if (event->attr.exclusive || !cpuctx->active_oncpu)
446 cpuctx->exclusive = 0;
450 group_sched_out(struct perf_event *group_event,
451 struct perf_cpu_context *cpuctx,
452 struct perf_event_context *ctx)
454 struct perf_event *event;
455 int state = group_event->state;
457 event_sched_out(group_event, cpuctx, ctx);
460 * Schedule out siblings (if any):
462 list_for_each_entry(event, &group_event->sibling_list, group_entry)
463 event_sched_out(event, cpuctx, ctx);
465 if (state == PERF_EVENT_STATE_ACTIVE && group_event->attr.exclusive)
466 cpuctx->exclusive = 0;
470 * Cross CPU call to remove a performance event
472 * We disable the event on the hardware level first. After that we
473 * remove it from the context list.
475 static void __perf_event_remove_from_context(void *info)
477 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
478 struct perf_event *event = info;
479 struct perf_event_context *ctx = event->ctx;
482 * If this is a task context, we need to check whether it is
483 * the current task context of this cpu. If not it has been
484 * scheduled out before the smp call arrived.
486 if (ctx->task && cpuctx->task_ctx != ctx)
489 raw_spin_lock(&ctx->lock);
491 * Protect the list operation against NMI by disabling the
492 * events on a global level.
496 event_sched_out(event, cpuctx, ctx);
498 list_del_event(event, ctx);
502 * Allow more per task events with respect to the
505 cpuctx->max_pertask =
506 min(perf_max_events - ctx->nr_events,
507 perf_max_events - perf_reserved_percpu);
511 raw_spin_unlock(&ctx->lock);
516 * Remove the event from a task's (or a CPU's) list of events.
518 * Must be called with ctx->mutex held.
520 * CPU events are removed with a smp call. For task events we only
521 * call when the task is on a CPU.
523 * If event->ctx is a cloned context, callers must make sure that
524 * every task struct that event->ctx->task could possibly point to
525 * remains valid. This is OK when called from perf_release since
526 * that only calls us on the top-level context, which can't be a clone.
527 * When called from perf_event_exit_task, it's OK because the
528 * context has been detached from its task.
530 static void perf_event_remove_from_context(struct perf_event *event)
532 struct perf_event_context *ctx = event->ctx;
533 struct task_struct *task = ctx->task;
537 * Per cpu events are removed via an smp call and
538 * the removal is always successful.
540 smp_call_function_single(event->cpu,
541 __perf_event_remove_from_context,
547 task_oncpu_function_call(task, __perf_event_remove_from_context,
550 raw_spin_lock_irq(&ctx->lock);
552 * If the context is active we need to retry the smp call.
554 if (ctx->nr_active && !list_empty(&event->group_entry)) {
555 raw_spin_unlock_irq(&ctx->lock);
560 * The lock prevents that this context is scheduled in so we
561 * can remove the event safely, if the call above did not
564 if (!list_empty(&event->group_entry))
565 list_del_event(event, ctx);
566 raw_spin_unlock_irq(&ctx->lock);
570 * Cross CPU call to disable a performance event
572 static void __perf_event_disable(void *info)
574 struct perf_event *event = info;
575 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
576 struct perf_event_context *ctx = event->ctx;
579 * If this is a per-task event, need to check whether this
580 * event's task is the current task on this cpu.
582 if (ctx->task && cpuctx->task_ctx != ctx)
585 raw_spin_lock(&ctx->lock);
588 * If the event is on, turn it off.
589 * If it is in error state, leave it in error state.
591 if (event->state >= PERF_EVENT_STATE_INACTIVE) {
592 update_context_time(ctx);
593 update_group_times(event);
594 if (event == event->group_leader)
595 group_sched_out(event, cpuctx, ctx);
597 event_sched_out(event, cpuctx, ctx);
598 event->state = PERF_EVENT_STATE_OFF;
601 raw_spin_unlock(&ctx->lock);
607 * If event->ctx is a cloned context, callers must make sure that
608 * every task struct that event->ctx->task could possibly point to
609 * remains valid. This condition is satisifed when called through
610 * perf_event_for_each_child or perf_event_for_each because they
611 * hold the top-level event's child_mutex, so any descendant that
612 * goes to exit will block in sync_child_event.
613 * When called from perf_pending_event it's OK because event->ctx
614 * is the current context on this CPU and preemption is disabled,
615 * hence we can't get into perf_event_task_sched_out for this context.
617 void perf_event_disable(struct perf_event *event)
619 struct perf_event_context *ctx = event->ctx;
620 struct task_struct *task = ctx->task;
624 * Disable the event on the cpu that it's on
626 smp_call_function_single(event->cpu, __perf_event_disable,
632 task_oncpu_function_call(task, __perf_event_disable, event);
634 raw_spin_lock_irq(&ctx->lock);
636 * If the event is still active, we need to retry the cross-call.
638 if (event->state == PERF_EVENT_STATE_ACTIVE) {
639 raw_spin_unlock_irq(&ctx->lock);
644 * Since we have the lock this context can't be scheduled
645 * in, so we can change the state safely.
647 if (event->state == PERF_EVENT_STATE_INACTIVE) {
648 update_group_times(event);
649 event->state = PERF_EVENT_STATE_OFF;
652 raw_spin_unlock_irq(&ctx->lock);
656 event_sched_in(struct perf_event *event,
657 struct perf_cpu_context *cpuctx,
658 struct perf_event_context *ctx)
660 if (event->state <= PERF_EVENT_STATE_OFF)
663 event->state = PERF_EVENT_STATE_ACTIVE;
664 event->oncpu = smp_processor_id();
666 * The new state must be visible before we turn it on in the hardware:
670 if (event->pmu->enable(event)) {
671 event->state = PERF_EVENT_STATE_INACTIVE;
676 event->tstamp_running += ctx->time - event->tstamp_stopped;
678 if (!is_software_event(event))
679 cpuctx->active_oncpu++;
682 if (event->attr.exclusive)
683 cpuctx->exclusive = 1;
689 group_sched_in(struct perf_event *group_event,
690 struct perf_cpu_context *cpuctx,
691 struct perf_event_context *ctx)
693 struct perf_event *event, *partial_group = NULL;
694 const struct pmu *pmu = group_event->pmu;
697 if (group_event->state == PERF_EVENT_STATE_OFF)
700 /* Check if group transaction availabe */
707 if (event_sched_in(group_event, cpuctx, ctx)) {
709 pmu->cancel_txn(pmu);
714 * Schedule in siblings as one group (if any):
716 list_for_each_entry(event, &group_event->sibling_list, group_entry) {
717 if (event_sched_in(event, cpuctx, ctx)) {
718 partial_group = event;
723 if (!txn || !pmu->commit_txn(pmu))
728 * Groups can be scheduled in as one unit only, so undo any
729 * partial group before returning:
731 list_for_each_entry(event, &group_event->sibling_list, group_entry) {
732 if (event == partial_group)
734 event_sched_out(event, cpuctx, ctx);
736 event_sched_out(group_event, cpuctx, ctx);
739 pmu->cancel_txn(pmu);
745 * Work out whether we can put this event group on the CPU now.
747 static int group_can_go_on(struct perf_event *event,
748 struct perf_cpu_context *cpuctx,
752 * Groups consisting entirely of software events can always go on.
754 if (event->group_flags & PERF_GROUP_SOFTWARE)
757 * If an exclusive group is already on, no other hardware
760 if (cpuctx->exclusive)
763 * If this group is exclusive and there are already
764 * events on the CPU, it can't go on.
766 if (event->attr.exclusive && cpuctx->active_oncpu)
769 * Otherwise, try to add it if all previous groups were able
775 static void add_event_to_ctx(struct perf_event *event,
776 struct perf_event_context *ctx)
778 list_add_event(event, ctx);
779 perf_group_attach(event);
780 event->tstamp_enabled = ctx->time;
781 event->tstamp_running = ctx->time;
782 event->tstamp_stopped = ctx->time;
786 * Cross CPU call to install and enable a performance event
788 * Must be called with ctx->mutex held
790 static void __perf_install_in_context(void *info)
792 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
793 struct perf_event *event = info;
794 struct perf_event_context *ctx = event->ctx;
795 struct perf_event *leader = event->group_leader;
799 * If this is a task context, we need to check whether it is
800 * the current task context of this cpu. If not it has been
801 * scheduled out before the smp call arrived.
802 * Or possibly this is the right context but it isn't
803 * on this cpu because it had no events.
805 if (ctx->task && cpuctx->task_ctx != ctx) {
806 if (cpuctx->task_ctx || ctx->task != current)
808 cpuctx->task_ctx = ctx;
811 raw_spin_lock(&ctx->lock);
813 update_context_time(ctx);
816 * Protect the list operation against NMI by disabling the
817 * events on a global level. NOP for non NMI based events.
821 add_event_to_ctx(event, ctx);
823 if (event->cpu != -1 && event->cpu != smp_processor_id())
827 * Don't put the event on if it is disabled or if
828 * it is in a group and the group isn't on.
830 if (event->state != PERF_EVENT_STATE_INACTIVE ||
831 (leader != event && leader->state != PERF_EVENT_STATE_ACTIVE))
835 * An exclusive event can't go on if there are already active
836 * hardware events, and no hardware event can go on if there
837 * is already an exclusive event on.
839 if (!group_can_go_on(event, cpuctx, 1))
842 err = event_sched_in(event, cpuctx, ctx);
846 * This event couldn't go on. If it is in a group
847 * then we have to pull the whole group off.
848 * If the event group is pinned then put it in error state.
851 group_sched_out(leader, cpuctx, ctx);
852 if (leader->attr.pinned) {
853 update_group_times(leader);
854 leader->state = PERF_EVENT_STATE_ERROR;
858 if (!err && !ctx->task && cpuctx->max_pertask)
859 cpuctx->max_pertask--;
864 raw_spin_unlock(&ctx->lock);
868 * Attach a performance event to a context
870 * First we add the event to the list with the hardware enable bit
871 * in event->hw_config cleared.
873 * If the event is attached to a task which is on a CPU we use a smp
874 * call to enable it in the task context. The task might have been
875 * scheduled away, but we check this in the smp call again.
877 * Must be called with ctx->mutex held.
880 perf_install_in_context(struct perf_event_context *ctx,
881 struct perf_event *event,
884 struct task_struct *task = ctx->task;
888 * Per cpu events are installed via an smp call and
889 * the install is always successful.
891 smp_call_function_single(cpu, __perf_install_in_context,
897 task_oncpu_function_call(task, __perf_install_in_context,
900 raw_spin_lock_irq(&ctx->lock);
902 * we need to retry the smp call.
904 if (ctx->is_active && list_empty(&event->group_entry)) {
905 raw_spin_unlock_irq(&ctx->lock);
910 * The lock prevents that this context is scheduled in so we
911 * can add the event safely, if it the call above did not
914 if (list_empty(&event->group_entry))
915 add_event_to_ctx(event, ctx);
916 raw_spin_unlock_irq(&ctx->lock);
920 * Put a event into inactive state and update time fields.
921 * Enabling the leader of a group effectively enables all
922 * the group members that aren't explicitly disabled, so we
923 * have to update their ->tstamp_enabled also.
924 * Note: this works for group members as well as group leaders
925 * since the non-leader members' sibling_lists will be empty.
927 static void __perf_event_mark_enabled(struct perf_event *event,
928 struct perf_event_context *ctx)
930 struct perf_event *sub;
932 event->state = PERF_EVENT_STATE_INACTIVE;
933 event->tstamp_enabled = ctx->time - event->total_time_enabled;
934 list_for_each_entry(sub, &event->sibling_list, group_entry)
935 if (sub->state >= PERF_EVENT_STATE_INACTIVE)
936 sub->tstamp_enabled =
937 ctx->time - sub->total_time_enabled;
941 * Cross CPU call to enable a performance event
943 static void __perf_event_enable(void *info)
945 struct perf_event *event = info;
946 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
947 struct perf_event_context *ctx = event->ctx;
948 struct perf_event *leader = event->group_leader;
952 * If this is a per-task event, need to check whether this
953 * event's task is the current task on this cpu.
955 if (ctx->task && cpuctx->task_ctx != ctx) {
956 if (cpuctx->task_ctx || ctx->task != current)
958 cpuctx->task_ctx = ctx;
961 raw_spin_lock(&ctx->lock);
963 update_context_time(ctx);
965 if (event->state >= PERF_EVENT_STATE_INACTIVE)
967 __perf_event_mark_enabled(event, ctx);
969 if (event->cpu != -1 && event->cpu != smp_processor_id())
973 * If the event is in a group and isn't the group leader,
974 * then don't put it on unless the group is on.
976 if (leader != event && leader->state != PERF_EVENT_STATE_ACTIVE)
979 if (!group_can_go_on(event, cpuctx, 1)) {
984 err = group_sched_in(event, cpuctx, ctx);
986 err = event_sched_in(event, cpuctx, ctx);
992 * If this event can't go on and it's part of a
993 * group, then the whole group has to come off.
996 group_sched_out(leader, cpuctx, ctx);
997 if (leader->attr.pinned) {
998 update_group_times(leader);
999 leader->state = PERF_EVENT_STATE_ERROR;
1004 raw_spin_unlock(&ctx->lock);
1010 * If event->ctx is a cloned context, callers must make sure that
1011 * every task struct that event->ctx->task could possibly point to
1012 * remains valid. This condition is satisfied when called through
1013 * perf_event_for_each_child or perf_event_for_each as described
1014 * for perf_event_disable.
1016 void perf_event_enable(struct perf_event *event)
1018 struct perf_event_context *ctx = event->ctx;
1019 struct task_struct *task = ctx->task;
1023 * Enable the event on the cpu that it's on
1025 smp_call_function_single(event->cpu, __perf_event_enable,
1030 raw_spin_lock_irq(&ctx->lock);
1031 if (event->state >= PERF_EVENT_STATE_INACTIVE)
1035 * If the event is in error state, clear that first.
1036 * That way, if we see the event in error state below, we
1037 * know that it has gone back into error state, as distinct
1038 * from the task having been scheduled away before the
1039 * cross-call arrived.
1041 if (event->state == PERF_EVENT_STATE_ERROR)
1042 event->state = PERF_EVENT_STATE_OFF;
1045 raw_spin_unlock_irq(&ctx->lock);
1046 task_oncpu_function_call(task, __perf_event_enable, event);
1048 raw_spin_lock_irq(&ctx->lock);
1051 * If the context is active and the event is still off,
1052 * we need to retry the cross-call.
1054 if (ctx->is_active && event->state == PERF_EVENT_STATE_OFF)
1058 * Since we have the lock this context can't be scheduled
1059 * in, so we can change the state safely.
1061 if (event->state == PERF_EVENT_STATE_OFF)
1062 __perf_event_mark_enabled(event, ctx);
1065 raw_spin_unlock_irq(&ctx->lock);
1068 static int perf_event_refresh(struct perf_event *event, int refresh)
1071 * not supported on inherited events
1073 if (event->attr.inherit)
1076 atomic_add(refresh, &event->event_limit);
1077 perf_event_enable(event);
1083 EVENT_FLEXIBLE = 0x1,
1085 EVENT_ALL = EVENT_FLEXIBLE | EVENT_PINNED,
1088 static void ctx_sched_out(struct perf_event_context *ctx,
1089 struct perf_cpu_context *cpuctx,
1090 enum event_type_t event_type)
1092 struct perf_event *event;
1094 raw_spin_lock(&ctx->lock);
1096 if (likely(!ctx->nr_events))
1098 update_context_time(ctx);
1101 if (!ctx->nr_active)
1104 if (event_type & EVENT_PINNED)
1105 list_for_each_entry(event, &ctx->pinned_groups, group_entry)
1106 group_sched_out(event, cpuctx, ctx);
1108 if (event_type & EVENT_FLEXIBLE)
1109 list_for_each_entry(event, &ctx->flexible_groups, group_entry)
1110 group_sched_out(event, cpuctx, ctx);
1115 raw_spin_unlock(&ctx->lock);
1119 * Test whether two contexts are equivalent, i.e. whether they
1120 * have both been cloned from the same version of the same context
1121 * and they both have the same number of enabled events.
1122 * If the number of enabled events is the same, then the set
1123 * of enabled events should be the same, because these are both
1124 * inherited contexts, therefore we can't access individual events
1125 * in them directly with an fd; we can only enable/disable all
1126 * events via prctl, or enable/disable all events in a family
1127 * via ioctl, which will have the same effect on both contexts.
1129 static int context_equiv(struct perf_event_context *ctx1,
1130 struct perf_event_context *ctx2)
1132 return ctx1->parent_ctx && ctx1->parent_ctx == ctx2->parent_ctx
1133 && ctx1->parent_gen == ctx2->parent_gen
1134 && !ctx1->pin_count && !ctx2->pin_count;
1137 static void __perf_event_sync_stat(struct perf_event *event,
1138 struct perf_event *next_event)
1142 if (!event->attr.inherit_stat)
1146 * Update the event value, we cannot use perf_event_read()
1147 * because we're in the middle of a context switch and have IRQs
1148 * disabled, which upsets smp_call_function_single(), however
1149 * we know the event must be on the current CPU, therefore we
1150 * don't need to use it.
1152 switch (event->state) {
1153 case PERF_EVENT_STATE_ACTIVE:
1154 event->pmu->read(event);
1157 case PERF_EVENT_STATE_INACTIVE:
1158 update_event_times(event);
1166 * In order to keep per-task stats reliable we need to flip the event
1167 * values when we flip the contexts.
1169 value = local64_read(&next_event->count);
1170 value = local64_xchg(&event->count, value);
1171 local64_set(&next_event->count, value);
1173 swap(event->total_time_enabled, next_event->total_time_enabled);
1174 swap(event->total_time_running, next_event->total_time_running);
1177 * Since we swizzled the values, update the user visible data too.
1179 perf_event_update_userpage(event);
1180 perf_event_update_userpage(next_event);
1183 #define list_next_entry(pos, member) \
1184 list_entry(pos->member.next, typeof(*pos), member)
1186 static void perf_event_sync_stat(struct perf_event_context *ctx,
1187 struct perf_event_context *next_ctx)
1189 struct perf_event *event, *next_event;
1194 update_context_time(ctx);
1196 event = list_first_entry(&ctx->event_list,
1197 struct perf_event, event_entry);
1199 next_event = list_first_entry(&next_ctx->event_list,
1200 struct perf_event, event_entry);
1202 while (&event->event_entry != &ctx->event_list &&
1203 &next_event->event_entry != &next_ctx->event_list) {
1205 __perf_event_sync_stat(event, next_event);
1207 event = list_next_entry(event, event_entry);
1208 next_event = list_next_entry(next_event, event_entry);
1213 * Called from scheduler to remove the events of the current task,
1214 * with interrupts disabled.
1216 * We stop each event and update the event value in event->count.
1218 * This does not protect us against NMI, but disable()
1219 * sets the disabled bit in the control field of event _before_
1220 * accessing the event control register. If a NMI hits, then it will
1221 * not restart the event.
1223 void perf_event_task_sched_out(struct task_struct *task,
1224 struct task_struct *next)
1226 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
1227 struct perf_event_context *ctx = task->perf_event_ctxp;
1228 struct perf_event_context *next_ctx;
1229 struct perf_event_context *parent;
1232 perf_sw_event(PERF_COUNT_SW_CONTEXT_SWITCHES, 1, 1, NULL, 0);
1234 if (likely(!ctx || !cpuctx->task_ctx))
1238 parent = rcu_dereference(ctx->parent_ctx);
1239 next_ctx = next->perf_event_ctxp;
1240 if (parent && next_ctx &&
1241 rcu_dereference(next_ctx->parent_ctx) == parent) {
1243 * Looks like the two contexts are clones, so we might be
1244 * able to optimize the context switch. We lock both
1245 * contexts and check that they are clones under the
1246 * lock (including re-checking that neither has been
1247 * uncloned in the meantime). It doesn't matter which
1248 * order we take the locks because no other cpu could
1249 * be trying to lock both of these tasks.
1251 raw_spin_lock(&ctx->lock);
1252 raw_spin_lock_nested(&next_ctx->lock, SINGLE_DEPTH_NESTING);
1253 if (context_equiv(ctx, next_ctx)) {
1255 * XXX do we need a memory barrier of sorts
1256 * wrt to rcu_dereference() of perf_event_ctxp
1258 task->perf_event_ctxp = next_ctx;
1259 next->perf_event_ctxp = ctx;
1261 next_ctx->task = task;
1264 perf_event_sync_stat(ctx, next_ctx);
1266 raw_spin_unlock(&next_ctx->lock);
1267 raw_spin_unlock(&ctx->lock);
1272 ctx_sched_out(ctx, cpuctx, EVENT_ALL);
1273 cpuctx->task_ctx = NULL;
1277 static void task_ctx_sched_out(struct perf_event_context *ctx,
1278 enum event_type_t event_type)
1280 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
1282 if (!cpuctx->task_ctx)
1285 if (WARN_ON_ONCE(ctx != cpuctx->task_ctx))
1288 ctx_sched_out(ctx, cpuctx, event_type);
1289 cpuctx->task_ctx = NULL;
1293 * Called with IRQs disabled
1295 static void __perf_event_task_sched_out(struct perf_event_context *ctx)
1297 task_ctx_sched_out(ctx, EVENT_ALL);
1301 * Called with IRQs disabled
1303 static void cpu_ctx_sched_out(struct perf_cpu_context *cpuctx,
1304 enum event_type_t event_type)
1306 ctx_sched_out(&cpuctx->ctx, cpuctx, event_type);
1310 ctx_pinned_sched_in(struct perf_event_context *ctx,
1311 struct perf_cpu_context *cpuctx)
1313 struct perf_event *event;
1315 list_for_each_entry(event, &ctx->pinned_groups, group_entry) {
1316 if (event->state <= PERF_EVENT_STATE_OFF)
1318 if (event->cpu != -1 && event->cpu != smp_processor_id())
1321 if (group_can_go_on(event, cpuctx, 1))
1322 group_sched_in(event, cpuctx, ctx);
1325 * If this pinned group hasn't been scheduled,
1326 * put it in error state.
1328 if (event->state == PERF_EVENT_STATE_INACTIVE) {
1329 update_group_times(event);
1330 event->state = PERF_EVENT_STATE_ERROR;
1336 ctx_flexible_sched_in(struct perf_event_context *ctx,
1337 struct perf_cpu_context *cpuctx)
1339 struct perf_event *event;
1342 list_for_each_entry(event, &ctx->flexible_groups, group_entry) {
1343 /* Ignore events in OFF or ERROR state */
1344 if (event->state <= PERF_EVENT_STATE_OFF)
1347 * Listen to the 'cpu' scheduling filter constraint
1350 if (event->cpu != -1 && event->cpu != smp_processor_id())
1353 if (group_can_go_on(event, cpuctx, can_add_hw))
1354 if (group_sched_in(event, cpuctx, ctx))
1360 ctx_sched_in(struct perf_event_context *ctx,
1361 struct perf_cpu_context *cpuctx,
1362 enum event_type_t event_type)
1364 raw_spin_lock(&ctx->lock);
1366 if (likely(!ctx->nr_events))
1369 ctx->timestamp = perf_clock();
1374 * First go through the list and put on any pinned groups
1375 * in order to give them the best chance of going on.
1377 if (event_type & EVENT_PINNED)
1378 ctx_pinned_sched_in(ctx, cpuctx);
1380 /* Then walk through the lower prio flexible groups */
1381 if (event_type & EVENT_FLEXIBLE)
1382 ctx_flexible_sched_in(ctx, cpuctx);
1386 raw_spin_unlock(&ctx->lock);
1389 static void cpu_ctx_sched_in(struct perf_cpu_context *cpuctx,
1390 enum event_type_t event_type)
1392 struct perf_event_context *ctx = &cpuctx->ctx;
1394 ctx_sched_in(ctx, cpuctx, event_type);
1397 static void task_ctx_sched_in(struct task_struct *task,
1398 enum event_type_t event_type)
1400 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
1401 struct perf_event_context *ctx = task->perf_event_ctxp;
1405 if (cpuctx->task_ctx == ctx)
1407 ctx_sched_in(ctx, cpuctx, event_type);
1408 cpuctx->task_ctx = ctx;
1411 * Called from scheduler to add the events of the current task
1412 * with interrupts disabled.
1414 * We restore the event value and then enable it.
1416 * This does not protect us against NMI, but enable()
1417 * sets the enabled bit in the control field of event _before_
1418 * accessing the event control register. If a NMI hits, then it will
1419 * keep the event running.
1421 void perf_event_task_sched_in(struct task_struct *task)
1423 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
1424 struct perf_event_context *ctx = task->perf_event_ctxp;
1429 if (cpuctx->task_ctx == ctx)
1435 * We want to keep the following priority order:
1436 * cpu pinned (that don't need to move), task pinned,
1437 * cpu flexible, task flexible.
1439 cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE);
1441 ctx_sched_in(ctx, cpuctx, EVENT_PINNED);
1442 cpu_ctx_sched_in(cpuctx, EVENT_FLEXIBLE);
1443 ctx_sched_in(ctx, cpuctx, EVENT_FLEXIBLE);
1445 cpuctx->task_ctx = ctx;
1450 #define MAX_INTERRUPTS (~0ULL)
1452 static void perf_log_throttle(struct perf_event *event, int enable);
1454 static u64 perf_calculate_period(struct perf_event *event, u64 nsec, u64 count)
1456 u64 frequency = event->attr.sample_freq;
1457 u64 sec = NSEC_PER_SEC;
1458 u64 divisor, dividend;
1460 int count_fls, nsec_fls, frequency_fls, sec_fls;
1462 count_fls = fls64(count);
1463 nsec_fls = fls64(nsec);
1464 frequency_fls = fls64(frequency);
1468 * We got @count in @nsec, with a target of sample_freq HZ
1469 * the target period becomes:
1472 * period = -------------------
1473 * @nsec * sample_freq
1478 * Reduce accuracy by one bit such that @a and @b converge
1479 * to a similar magnitude.
1481 #define REDUCE_FLS(a, b) \
1483 if (a##_fls > b##_fls) { \
1493 * Reduce accuracy until either term fits in a u64, then proceed with
1494 * the other, so that finally we can do a u64/u64 division.
1496 while (count_fls + sec_fls > 64 && nsec_fls + frequency_fls > 64) {
1497 REDUCE_FLS(nsec, frequency);
1498 REDUCE_FLS(sec, count);
1501 if (count_fls + sec_fls > 64) {
1502 divisor = nsec * frequency;
1504 while (count_fls + sec_fls > 64) {
1505 REDUCE_FLS(count, sec);
1509 dividend = count * sec;
1511 dividend = count * sec;
1513 while (nsec_fls + frequency_fls > 64) {
1514 REDUCE_FLS(nsec, frequency);
1518 divisor = nsec * frequency;
1524 return div64_u64(dividend, divisor);
1527 static void perf_event_stop(struct perf_event *event)
1529 if (!event->pmu->stop)
1530 return event->pmu->disable(event);
1532 return event->pmu->stop(event);
1535 static int perf_event_start(struct perf_event *event)
1537 if (!event->pmu->start)
1538 return event->pmu->enable(event);
1540 return event->pmu->start(event);
1543 static void perf_adjust_period(struct perf_event *event, u64 nsec, u64 count)
1545 struct hw_perf_event *hwc = &event->hw;
1546 s64 period, sample_period;
1549 period = perf_calculate_period(event, nsec, count);
1551 delta = (s64)(period - hwc->sample_period);
1552 delta = (delta + 7) / 8; /* low pass filter */
1554 sample_period = hwc->sample_period + delta;
1559 hwc->sample_period = sample_period;
1561 if (local64_read(&hwc->period_left) > 8*sample_period) {
1563 perf_event_stop(event);
1564 local64_set(&hwc->period_left, 0);
1565 perf_event_start(event);
1570 static void perf_ctx_adjust_freq(struct perf_event_context *ctx)
1572 struct perf_event *event;
1573 struct hw_perf_event *hwc;
1574 u64 interrupts, now;
1577 raw_spin_lock(&ctx->lock);
1578 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
1579 if (event->state != PERF_EVENT_STATE_ACTIVE)
1582 if (event->cpu != -1 && event->cpu != smp_processor_id())
1587 interrupts = hwc->interrupts;
1588 hwc->interrupts = 0;
1591 * unthrottle events on the tick
1593 if (interrupts == MAX_INTERRUPTS) {
1594 perf_log_throttle(event, 1);
1596 event->pmu->unthrottle(event);
1600 if (!event->attr.freq || !event->attr.sample_freq)
1604 event->pmu->read(event);
1605 now = local64_read(&event->count);
1606 delta = now - hwc->freq_count_stamp;
1607 hwc->freq_count_stamp = now;
1610 perf_adjust_period(event, TICK_NSEC, delta);
1613 raw_spin_unlock(&ctx->lock);
1617 * Round-robin a context's events:
1619 static void rotate_ctx(struct perf_event_context *ctx)
1621 raw_spin_lock(&ctx->lock);
1624 * Rotate the first entry last of non-pinned groups. Rotation might be
1625 * disabled by the inheritance code.
1627 if (!ctx->rotate_disable)
1628 list_rotate_left(&ctx->flexible_groups);
1630 raw_spin_unlock(&ctx->lock);
1633 void perf_event_task_tick(struct task_struct *curr)
1635 struct perf_cpu_context *cpuctx;
1636 struct perf_event_context *ctx;
1639 if (!atomic_read(&nr_events))
1642 cpuctx = &__get_cpu_var(perf_cpu_context);
1643 if (cpuctx->ctx.nr_events &&
1644 cpuctx->ctx.nr_events != cpuctx->ctx.nr_active)
1647 ctx = curr->perf_event_ctxp;
1648 if (ctx && ctx->nr_events && ctx->nr_events != ctx->nr_active)
1651 perf_ctx_adjust_freq(&cpuctx->ctx);
1653 perf_ctx_adjust_freq(ctx);
1659 cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE);
1661 task_ctx_sched_out(ctx, EVENT_FLEXIBLE);
1663 rotate_ctx(&cpuctx->ctx);
1667 cpu_ctx_sched_in(cpuctx, EVENT_FLEXIBLE);
1669 task_ctx_sched_in(curr, EVENT_FLEXIBLE);
1673 static int event_enable_on_exec(struct perf_event *event,
1674 struct perf_event_context *ctx)
1676 if (!event->attr.enable_on_exec)
1679 event->attr.enable_on_exec = 0;
1680 if (event->state >= PERF_EVENT_STATE_INACTIVE)
1683 __perf_event_mark_enabled(event, ctx);
1689 * Enable all of a task's events that have been marked enable-on-exec.
1690 * This expects task == current.
1692 static void perf_event_enable_on_exec(struct task_struct *task)
1694 struct perf_event_context *ctx;
1695 struct perf_event *event;
1696 unsigned long flags;
1700 local_irq_save(flags);
1701 ctx = task->perf_event_ctxp;
1702 if (!ctx || !ctx->nr_events)
1705 __perf_event_task_sched_out(ctx);
1707 raw_spin_lock(&ctx->lock);
1709 list_for_each_entry(event, &ctx->pinned_groups, group_entry) {
1710 ret = event_enable_on_exec(event, ctx);
1715 list_for_each_entry(event, &ctx->flexible_groups, group_entry) {
1716 ret = event_enable_on_exec(event, ctx);
1722 * Unclone this context if we enabled any event.
1727 raw_spin_unlock(&ctx->lock);
1729 perf_event_task_sched_in(task);
1731 local_irq_restore(flags);
1735 * Cross CPU call to read the hardware event
1737 static void __perf_event_read(void *info)
1739 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
1740 struct perf_event *event = info;
1741 struct perf_event_context *ctx = event->ctx;
1744 * If this is a task context, we need to check whether it is
1745 * the current task context of this cpu. If not it has been
1746 * scheduled out before the smp call arrived. In that case
1747 * event->count would have been updated to a recent sample
1748 * when the event was scheduled out.
1750 if (ctx->task && cpuctx->task_ctx != ctx)
1753 raw_spin_lock(&ctx->lock);
1754 update_context_time(ctx);
1755 update_event_times(event);
1756 raw_spin_unlock(&ctx->lock);
1758 event->pmu->read(event);
1761 static inline u64 perf_event_count(struct perf_event *event)
1763 return local64_read(&event->count) + atomic64_read(&event->child_count);
1766 static u64 perf_event_read(struct perf_event *event)
1769 * If event is enabled and currently active on a CPU, update the
1770 * value in the event structure:
1772 if (event->state == PERF_EVENT_STATE_ACTIVE) {
1773 smp_call_function_single(event->oncpu,
1774 __perf_event_read, event, 1);
1775 } else if (event->state == PERF_EVENT_STATE_INACTIVE) {
1776 struct perf_event_context *ctx = event->ctx;
1777 unsigned long flags;
1779 raw_spin_lock_irqsave(&ctx->lock, flags);
1781 * may read while context is not active
1782 * (e.g., thread is blocked), in that case
1783 * we cannot update context time
1786 update_context_time(ctx);
1787 update_event_times(event);
1788 raw_spin_unlock_irqrestore(&ctx->lock, flags);
1791 return perf_event_count(event);
1795 * Initialize the perf_event context in a task_struct:
1798 __perf_event_init_context(struct perf_event_context *ctx,
1799 struct task_struct *task)
1801 raw_spin_lock_init(&ctx->lock);
1802 mutex_init(&ctx->mutex);
1803 INIT_LIST_HEAD(&ctx->pinned_groups);
1804 INIT_LIST_HEAD(&ctx->flexible_groups);
1805 INIT_LIST_HEAD(&ctx->event_list);
1806 atomic_set(&ctx->refcount, 1);
1810 static struct perf_event_context *find_get_context(pid_t pid, int cpu)
1812 struct perf_event_context *ctx;
1813 struct perf_cpu_context *cpuctx;
1814 struct task_struct *task;
1815 unsigned long flags;
1818 if (pid == -1 && cpu != -1) {
1819 /* Must be root to operate on a CPU event: */
1820 if (perf_paranoid_cpu() && !capable(CAP_SYS_ADMIN))
1821 return ERR_PTR(-EACCES);
1823 if (cpu < 0 || cpu >= nr_cpumask_bits)
1824 return ERR_PTR(-EINVAL);
1827 * We could be clever and allow to attach a event to an
1828 * offline CPU and activate it when the CPU comes up, but
1831 if (!cpu_online(cpu))
1832 return ERR_PTR(-ENODEV);
1834 cpuctx = &per_cpu(perf_cpu_context, cpu);
1845 task = find_task_by_vpid(pid);
1847 get_task_struct(task);
1851 return ERR_PTR(-ESRCH);
1854 * Can't attach events to a dying task.
1857 if (task->flags & PF_EXITING)
1860 /* Reuse ptrace permission checks for now. */
1862 if (!ptrace_may_access(task, PTRACE_MODE_READ))
1866 ctx = perf_lock_task_context(task, &flags);
1869 raw_spin_unlock_irqrestore(&ctx->lock, flags);
1873 ctx = kzalloc(sizeof(struct perf_event_context), GFP_KERNEL);
1877 __perf_event_init_context(ctx, task);
1879 if (cmpxchg(&task->perf_event_ctxp, NULL, ctx)) {
1881 * We raced with some other task; use
1882 * the context they set.
1887 get_task_struct(task);
1890 put_task_struct(task);
1894 put_task_struct(task);
1895 return ERR_PTR(err);
1898 static void perf_event_free_filter(struct perf_event *event);
1900 static void free_event_rcu(struct rcu_head *head)
1902 struct perf_event *event;
1904 event = container_of(head, struct perf_event, rcu_head);
1906 put_pid_ns(event->ns);
1907 perf_event_free_filter(event);
1911 static void perf_pending_sync(struct perf_event *event);
1912 static void perf_buffer_put(struct perf_buffer *buffer);
1914 static void free_event(struct perf_event *event)
1916 perf_pending_sync(event);
1918 if (!event->parent) {
1919 atomic_dec(&nr_events);
1920 if (event->attr.mmap || event->attr.mmap_data)
1921 atomic_dec(&nr_mmap_events);
1922 if (event->attr.comm)
1923 atomic_dec(&nr_comm_events);
1924 if (event->attr.task)
1925 atomic_dec(&nr_task_events);
1928 if (event->buffer) {
1929 perf_buffer_put(event->buffer);
1930 event->buffer = NULL;
1934 event->destroy(event);
1936 put_ctx(event->ctx);
1937 call_rcu(&event->rcu_head, free_event_rcu);
1940 int perf_event_release_kernel(struct perf_event *event)
1942 struct perf_event_context *ctx = event->ctx;
1945 * Remove from the PMU, can't get re-enabled since we got
1946 * here because the last ref went.
1948 perf_event_disable(event);
1950 WARN_ON_ONCE(ctx->parent_ctx);
1952 * There are two ways this annotation is useful:
1954 * 1) there is a lock recursion from perf_event_exit_task
1955 * see the comment there.
1957 * 2) there is a lock-inversion with mmap_sem through
1958 * perf_event_read_group(), which takes faults while
1959 * holding ctx->mutex, however this is called after
1960 * the last filedesc died, so there is no possibility
1961 * to trigger the AB-BA case.
1963 mutex_lock_nested(&ctx->mutex, SINGLE_DEPTH_NESTING);
1964 raw_spin_lock_irq(&ctx->lock);
1965 perf_group_detach(event);
1966 list_del_event(event, ctx);
1967 raw_spin_unlock_irq(&ctx->lock);
1968 mutex_unlock(&ctx->mutex);
1970 mutex_lock(&event->owner->perf_event_mutex);
1971 list_del_init(&event->owner_entry);
1972 mutex_unlock(&event->owner->perf_event_mutex);
1973 put_task_struct(event->owner);
1979 EXPORT_SYMBOL_GPL(perf_event_release_kernel);
1982 * Called when the last reference to the file is gone.
1984 static int perf_release(struct inode *inode, struct file *file)
1986 struct perf_event *event = file->private_data;
1988 file->private_data = NULL;
1990 return perf_event_release_kernel(event);
1993 static int perf_event_read_size(struct perf_event *event)
1995 int entry = sizeof(u64); /* value */
1999 if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
2000 size += sizeof(u64);
2002 if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
2003 size += sizeof(u64);
2005 if (event->attr.read_format & PERF_FORMAT_ID)
2006 entry += sizeof(u64);
2008 if (event->attr.read_format & PERF_FORMAT_GROUP) {
2009 nr += event->group_leader->nr_siblings;
2010 size += sizeof(u64);
2018 u64 perf_event_read_value(struct perf_event *event, u64 *enabled, u64 *running)
2020 struct perf_event *child;
2026 mutex_lock(&event->child_mutex);
2027 total += perf_event_read(event);
2028 *enabled += event->total_time_enabled +
2029 atomic64_read(&event->child_total_time_enabled);
2030 *running += event->total_time_running +
2031 atomic64_read(&event->child_total_time_running);
2033 list_for_each_entry(child, &event->child_list, child_list) {
2034 total += perf_event_read(child);
2035 *enabled += child->total_time_enabled;
2036 *running += child->total_time_running;
2038 mutex_unlock(&event->child_mutex);
2042 EXPORT_SYMBOL_GPL(perf_event_read_value);
2044 static int perf_event_read_group(struct perf_event *event,
2045 u64 read_format, char __user *buf)
2047 struct perf_event *leader = event->group_leader, *sub;
2048 int n = 0, size = 0, ret = -EFAULT;
2049 struct perf_event_context *ctx = leader->ctx;
2051 u64 count, enabled, running;
2053 mutex_lock(&ctx->mutex);
2054 count = perf_event_read_value(leader, &enabled, &running);
2056 values[n++] = 1 + leader->nr_siblings;
2057 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
2058 values[n++] = enabled;
2059 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
2060 values[n++] = running;
2061 values[n++] = count;
2062 if (read_format & PERF_FORMAT_ID)
2063 values[n++] = primary_event_id(leader);
2065 size = n * sizeof(u64);
2067 if (copy_to_user(buf, values, size))
2072 list_for_each_entry(sub, &leader->sibling_list, group_entry) {
2075 values[n++] = perf_event_read_value(sub, &enabled, &running);
2076 if (read_format & PERF_FORMAT_ID)
2077 values[n++] = primary_event_id(sub);
2079 size = n * sizeof(u64);
2081 if (copy_to_user(buf + ret, values, size)) {
2089 mutex_unlock(&ctx->mutex);
2094 static int perf_event_read_one(struct perf_event *event,
2095 u64 read_format, char __user *buf)
2097 u64 enabled, running;
2101 values[n++] = perf_event_read_value(event, &enabled, &running);
2102 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
2103 values[n++] = enabled;
2104 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
2105 values[n++] = running;
2106 if (read_format & PERF_FORMAT_ID)
2107 values[n++] = primary_event_id(event);
2109 if (copy_to_user(buf, values, n * sizeof(u64)))
2112 return n * sizeof(u64);
2116 * Read the performance event - simple non blocking version for now
2119 perf_read_hw(struct perf_event *event, char __user *buf, size_t count)
2121 u64 read_format = event->attr.read_format;
2125 * Return end-of-file for a read on a event that is in
2126 * error state (i.e. because it was pinned but it couldn't be
2127 * scheduled on to the CPU at some point).
2129 if (event->state == PERF_EVENT_STATE_ERROR)
2132 if (count < perf_event_read_size(event))
2135 WARN_ON_ONCE(event->ctx->parent_ctx);
2136 if (read_format & PERF_FORMAT_GROUP)
2137 ret = perf_event_read_group(event, read_format, buf);
2139 ret = perf_event_read_one(event, read_format, buf);
2145 perf_read(struct file *file, char __user *buf, size_t count, loff_t *ppos)
2147 struct perf_event *event = file->private_data;
2149 return perf_read_hw(event, buf, count);
2152 static unsigned int perf_poll(struct file *file, poll_table *wait)
2154 struct perf_event *event = file->private_data;
2155 struct perf_buffer *buffer;
2156 unsigned int events = POLL_HUP;
2159 buffer = rcu_dereference(event->buffer);
2161 events = atomic_xchg(&buffer->poll, 0);
2164 poll_wait(file, &event->waitq, wait);
2169 static void perf_event_reset(struct perf_event *event)
2171 (void)perf_event_read(event);
2172 local64_set(&event->count, 0);
2173 perf_event_update_userpage(event);
2177 * Holding the top-level event's child_mutex means that any
2178 * descendant process that has inherited this event will block
2179 * in sync_child_event if it goes to exit, thus satisfying the
2180 * task existence requirements of perf_event_enable/disable.
2182 static void perf_event_for_each_child(struct perf_event *event,
2183 void (*func)(struct perf_event *))
2185 struct perf_event *child;
2187 WARN_ON_ONCE(event->ctx->parent_ctx);
2188 mutex_lock(&event->child_mutex);
2190 list_for_each_entry(child, &event->child_list, child_list)
2192 mutex_unlock(&event->child_mutex);
2195 static void perf_event_for_each(struct perf_event *event,
2196 void (*func)(struct perf_event *))
2198 struct perf_event_context *ctx = event->ctx;
2199 struct perf_event *sibling;
2201 WARN_ON_ONCE(ctx->parent_ctx);
2202 mutex_lock(&ctx->mutex);
2203 event = event->group_leader;
2205 perf_event_for_each_child(event, func);
2207 list_for_each_entry(sibling, &event->sibling_list, group_entry)
2208 perf_event_for_each_child(event, func);
2209 mutex_unlock(&ctx->mutex);
2212 static int perf_event_period(struct perf_event *event, u64 __user *arg)
2214 struct perf_event_context *ctx = event->ctx;
2218 if (!event->attr.sample_period)
2221 if (copy_from_user(&value, arg, sizeof(value)))
2227 raw_spin_lock_irq(&ctx->lock);
2228 if (event->attr.freq) {
2229 if (value > sysctl_perf_event_sample_rate) {
2234 event->attr.sample_freq = value;
2236 event->attr.sample_period = value;
2237 event->hw.sample_period = value;
2240 raw_spin_unlock_irq(&ctx->lock);
2245 static const struct file_operations perf_fops;
2247 static struct perf_event *perf_fget_light(int fd, int *fput_needed)
2251 file = fget_light(fd, fput_needed);
2253 return ERR_PTR(-EBADF);
2255 if (file->f_op != &perf_fops) {
2256 fput_light(file, *fput_needed);
2258 return ERR_PTR(-EBADF);
2261 return file->private_data;
2264 static int perf_event_set_output(struct perf_event *event,
2265 struct perf_event *output_event);
2266 static int perf_event_set_filter(struct perf_event *event, void __user *arg);
2268 static long perf_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
2270 struct perf_event *event = file->private_data;
2271 void (*func)(struct perf_event *);
2275 case PERF_EVENT_IOC_ENABLE:
2276 func = perf_event_enable;
2278 case PERF_EVENT_IOC_DISABLE:
2279 func = perf_event_disable;
2281 case PERF_EVENT_IOC_RESET:
2282 func = perf_event_reset;
2285 case PERF_EVENT_IOC_REFRESH:
2286 return perf_event_refresh(event, arg);
2288 case PERF_EVENT_IOC_PERIOD:
2289 return perf_event_period(event, (u64 __user *)arg);
2291 case PERF_EVENT_IOC_SET_OUTPUT:
2293 struct perf_event *output_event = NULL;
2294 int fput_needed = 0;
2298 output_event = perf_fget_light(arg, &fput_needed);
2299 if (IS_ERR(output_event))
2300 return PTR_ERR(output_event);
2303 ret = perf_event_set_output(event, output_event);
2305 fput_light(output_event->filp, fput_needed);
2310 case PERF_EVENT_IOC_SET_FILTER:
2311 return perf_event_set_filter(event, (void __user *)arg);
2317 if (flags & PERF_IOC_FLAG_GROUP)
2318 perf_event_for_each(event, func);
2320 perf_event_for_each_child(event, func);
2325 int perf_event_task_enable(void)
2327 struct perf_event *event;
2329 mutex_lock(¤t->perf_event_mutex);
2330 list_for_each_entry(event, ¤t->perf_event_list, owner_entry)
2331 perf_event_for_each_child(event, perf_event_enable);
2332 mutex_unlock(¤t->perf_event_mutex);
2337 int perf_event_task_disable(void)
2339 struct perf_event *event;
2341 mutex_lock(¤t->perf_event_mutex);
2342 list_for_each_entry(event, ¤t->perf_event_list, owner_entry)
2343 perf_event_for_each_child(event, perf_event_disable);
2344 mutex_unlock(¤t->perf_event_mutex);
2349 #ifndef PERF_EVENT_INDEX_OFFSET
2350 # define PERF_EVENT_INDEX_OFFSET 0
2353 static int perf_event_index(struct perf_event *event)
2355 if (event->state != PERF_EVENT_STATE_ACTIVE)
2358 return event->hw.idx + 1 - PERF_EVENT_INDEX_OFFSET;
2362 * Callers need to ensure there can be no nesting of this function, otherwise
2363 * the seqlock logic goes bad. We can not serialize this because the arch
2364 * code calls this from NMI context.
2366 void perf_event_update_userpage(struct perf_event *event)
2368 struct perf_event_mmap_page *userpg;
2369 struct perf_buffer *buffer;
2372 buffer = rcu_dereference(event->buffer);
2376 userpg = buffer->user_page;
2379 * Disable preemption so as to not let the corresponding user-space
2380 * spin too long if we get preempted.
2385 userpg->index = perf_event_index(event);
2386 userpg->offset = perf_event_count(event);
2387 if (event->state == PERF_EVENT_STATE_ACTIVE)
2388 userpg->offset -= local64_read(&event->hw.prev_count);
2390 userpg->time_enabled = event->total_time_enabled +
2391 atomic64_read(&event->child_total_time_enabled);
2393 userpg->time_running = event->total_time_running +
2394 atomic64_read(&event->child_total_time_running);
2403 static unsigned long perf_data_size(struct perf_buffer *buffer);
2406 perf_buffer_init(struct perf_buffer *buffer, long watermark, int flags)
2408 long max_size = perf_data_size(buffer);
2411 buffer->watermark = min(max_size, watermark);
2413 if (!buffer->watermark)
2414 buffer->watermark = max_size / 2;
2416 if (flags & PERF_BUFFER_WRITABLE)
2417 buffer->writable = 1;
2419 atomic_set(&buffer->refcount, 1);
2422 #ifndef CONFIG_PERF_USE_VMALLOC
2425 * Back perf_mmap() with regular GFP_KERNEL-0 pages.
2428 static struct page *
2429 perf_mmap_to_page(struct perf_buffer *buffer, unsigned long pgoff)
2431 if (pgoff > buffer->nr_pages)
2435 return virt_to_page(buffer->user_page);
2437 return virt_to_page(buffer->data_pages[pgoff - 1]);
2440 static void *perf_mmap_alloc_page(int cpu)
2445 node = (cpu == -1) ? cpu : cpu_to_node(cpu);
2446 page = alloc_pages_node(node, GFP_KERNEL | __GFP_ZERO, 0);
2450 return page_address(page);
2453 static struct perf_buffer *
2454 perf_buffer_alloc(int nr_pages, long watermark, int cpu, int flags)
2456 struct perf_buffer *buffer;
2460 size = sizeof(struct perf_buffer);
2461 size += nr_pages * sizeof(void *);
2463 buffer = kzalloc(size, GFP_KERNEL);
2467 buffer->user_page = perf_mmap_alloc_page(cpu);
2468 if (!buffer->user_page)
2469 goto fail_user_page;
2471 for (i = 0; i < nr_pages; i++) {
2472 buffer->data_pages[i] = perf_mmap_alloc_page(cpu);
2473 if (!buffer->data_pages[i])
2474 goto fail_data_pages;
2477 buffer->nr_pages = nr_pages;
2479 perf_buffer_init(buffer, watermark, flags);
2484 for (i--; i >= 0; i--)
2485 free_page((unsigned long)buffer->data_pages[i]);
2487 free_page((unsigned long)buffer->user_page);
2496 static void perf_mmap_free_page(unsigned long addr)
2498 struct page *page = virt_to_page((void *)addr);
2500 page->mapping = NULL;
2504 static void perf_buffer_free(struct perf_buffer *buffer)
2508 perf_mmap_free_page((unsigned long)buffer->user_page);
2509 for (i = 0; i < buffer->nr_pages; i++)
2510 perf_mmap_free_page((unsigned long)buffer->data_pages[i]);
2514 static inline int page_order(struct perf_buffer *buffer)
2522 * Back perf_mmap() with vmalloc memory.
2524 * Required for architectures that have d-cache aliasing issues.
2527 static inline int page_order(struct perf_buffer *buffer)
2529 return buffer->page_order;
2532 static struct page *
2533 perf_mmap_to_page(struct perf_buffer *buffer, unsigned long pgoff)
2535 if (pgoff > (1UL << page_order(buffer)))
2538 return vmalloc_to_page((void *)buffer->user_page + pgoff * PAGE_SIZE);
2541 static void perf_mmap_unmark_page(void *addr)
2543 struct page *page = vmalloc_to_page(addr);
2545 page->mapping = NULL;
2548 static void perf_buffer_free_work(struct work_struct *work)
2550 struct perf_buffer *buffer;
2554 buffer = container_of(work, struct perf_buffer, work);
2555 nr = 1 << page_order(buffer);
2557 base = buffer->user_page;
2558 for (i = 0; i < nr + 1; i++)
2559 perf_mmap_unmark_page(base + (i * PAGE_SIZE));
2565 static void perf_buffer_free(struct perf_buffer *buffer)
2567 schedule_work(&buffer->work);
2570 static struct perf_buffer *
2571 perf_buffer_alloc(int nr_pages, long watermark, int cpu, int flags)
2573 struct perf_buffer *buffer;
2577 size = sizeof(struct perf_buffer);
2578 size += sizeof(void *);
2580 buffer = kzalloc(size, GFP_KERNEL);
2584 INIT_WORK(&buffer->work, perf_buffer_free_work);
2586 all_buf = vmalloc_user((nr_pages + 1) * PAGE_SIZE);
2590 buffer->user_page = all_buf;
2591 buffer->data_pages[0] = all_buf + PAGE_SIZE;
2592 buffer->page_order = ilog2(nr_pages);
2593 buffer->nr_pages = 1;
2595 perf_buffer_init(buffer, watermark, flags);
2608 static unsigned long perf_data_size(struct perf_buffer *buffer)
2610 return buffer->nr_pages << (PAGE_SHIFT + page_order(buffer));
2613 static int perf_mmap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
2615 struct perf_event *event = vma->vm_file->private_data;
2616 struct perf_buffer *buffer;
2617 int ret = VM_FAULT_SIGBUS;
2619 if (vmf->flags & FAULT_FLAG_MKWRITE) {
2620 if (vmf->pgoff == 0)
2626 buffer = rcu_dereference(event->buffer);
2630 if (vmf->pgoff && (vmf->flags & FAULT_FLAG_WRITE))
2633 vmf->page = perf_mmap_to_page(buffer, vmf->pgoff);
2637 get_page(vmf->page);
2638 vmf->page->mapping = vma->vm_file->f_mapping;
2639 vmf->page->index = vmf->pgoff;
2648 static void perf_buffer_free_rcu(struct rcu_head *rcu_head)
2650 struct perf_buffer *buffer;
2652 buffer = container_of(rcu_head, struct perf_buffer, rcu_head);
2653 perf_buffer_free(buffer);
2656 static struct perf_buffer *perf_buffer_get(struct perf_event *event)
2658 struct perf_buffer *buffer;
2661 buffer = rcu_dereference(event->buffer);
2663 if (!atomic_inc_not_zero(&buffer->refcount))
2671 static void perf_buffer_put(struct perf_buffer *buffer)
2673 if (!atomic_dec_and_test(&buffer->refcount))
2676 call_rcu(&buffer->rcu_head, perf_buffer_free_rcu);
2679 static void perf_mmap_open(struct vm_area_struct *vma)
2681 struct perf_event *event = vma->vm_file->private_data;
2683 atomic_inc(&event->mmap_count);
2686 static void perf_mmap_close(struct vm_area_struct *vma)
2688 struct perf_event *event = vma->vm_file->private_data;
2690 if (atomic_dec_and_mutex_lock(&event->mmap_count, &event->mmap_mutex)) {
2691 unsigned long size = perf_data_size(event->buffer);
2692 struct user_struct *user = event->mmap_user;
2693 struct perf_buffer *buffer = event->buffer;
2695 atomic_long_sub((size >> PAGE_SHIFT) + 1, &user->locked_vm);
2696 vma->vm_mm->locked_vm -= event->mmap_locked;
2697 rcu_assign_pointer(event->buffer, NULL);
2698 mutex_unlock(&event->mmap_mutex);
2700 perf_buffer_put(buffer);
2705 static const struct vm_operations_struct perf_mmap_vmops = {
2706 .open = perf_mmap_open,
2707 .close = perf_mmap_close,
2708 .fault = perf_mmap_fault,
2709 .page_mkwrite = perf_mmap_fault,
2712 static int perf_mmap(struct file *file, struct vm_area_struct *vma)
2714 struct perf_event *event = file->private_data;
2715 unsigned long user_locked, user_lock_limit;
2716 struct user_struct *user = current_user();
2717 unsigned long locked, lock_limit;
2718 struct perf_buffer *buffer;
2719 unsigned long vma_size;
2720 unsigned long nr_pages;
2721 long user_extra, extra;
2722 int ret = 0, flags = 0;
2725 * Don't allow mmap() of inherited per-task counters. This would
2726 * create a performance issue due to all children writing to the
2729 if (event->cpu == -1 && event->attr.inherit)
2732 if (!(vma->vm_flags & VM_SHARED))
2735 vma_size = vma->vm_end - vma->vm_start;
2736 nr_pages = (vma_size / PAGE_SIZE) - 1;
2739 * If we have buffer pages ensure they're a power-of-two number, so we
2740 * can do bitmasks instead of modulo.
2742 if (nr_pages != 0 && !is_power_of_2(nr_pages))
2745 if (vma_size != PAGE_SIZE * (1 + nr_pages))
2748 if (vma->vm_pgoff != 0)
2751 WARN_ON_ONCE(event->ctx->parent_ctx);
2752 mutex_lock(&event->mmap_mutex);
2753 if (event->buffer) {
2754 if (event->buffer->nr_pages == nr_pages)
2755 atomic_inc(&event->buffer->refcount);
2761 user_extra = nr_pages + 1;
2762 user_lock_limit = sysctl_perf_event_mlock >> (PAGE_SHIFT - 10);
2765 * Increase the limit linearly with more CPUs:
2767 user_lock_limit *= num_online_cpus();
2769 user_locked = atomic_long_read(&user->locked_vm) + user_extra;
2772 if (user_locked > user_lock_limit)
2773 extra = user_locked - user_lock_limit;
2775 lock_limit = rlimit(RLIMIT_MEMLOCK);
2776 lock_limit >>= PAGE_SHIFT;
2777 locked = vma->vm_mm->locked_vm + extra;
2779 if ((locked > lock_limit) && perf_paranoid_tracepoint_raw() &&
2780 !capable(CAP_IPC_LOCK)) {
2785 WARN_ON(event->buffer);
2787 if (vma->vm_flags & VM_WRITE)
2788 flags |= PERF_BUFFER_WRITABLE;
2790 buffer = perf_buffer_alloc(nr_pages, event->attr.wakeup_watermark,
2796 rcu_assign_pointer(event->buffer, buffer);
2798 atomic_long_add(user_extra, &user->locked_vm);
2799 event->mmap_locked = extra;
2800 event->mmap_user = get_current_user();
2801 vma->vm_mm->locked_vm += event->mmap_locked;
2805 atomic_inc(&event->mmap_count);
2806 mutex_unlock(&event->mmap_mutex);
2808 vma->vm_flags |= VM_RESERVED;
2809 vma->vm_ops = &perf_mmap_vmops;
2814 static int perf_fasync(int fd, struct file *filp, int on)
2816 struct inode *inode = filp->f_path.dentry->d_inode;
2817 struct perf_event *event = filp->private_data;
2820 mutex_lock(&inode->i_mutex);
2821 retval = fasync_helper(fd, filp, on, &event->fasync);
2822 mutex_unlock(&inode->i_mutex);
2830 static const struct file_operations perf_fops = {
2831 .llseek = no_llseek,
2832 .release = perf_release,
2835 .unlocked_ioctl = perf_ioctl,
2836 .compat_ioctl = perf_ioctl,
2838 .fasync = perf_fasync,
2844 * If there's data, ensure we set the poll() state and publish everything
2845 * to user-space before waking everybody up.
2848 void perf_event_wakeup(struct perf_event *event)
2850 wake_up_all(&event->waitq);
2852 if (event->pending_kill) {
2853 kill_fasync(&event->fasync, SIGIO, event->pending_kill);
2854 event->pending_kill = 0;
2861 * Handle the case where we need to wakeup up from NMI (or rq->lock) context.
2863 * The NMI bit means we cannot possibly take locks. Therefore, maintain a
2864 * single linked list and use cmpxchg() to add entries lockless.
2867 static void perf_pending_event(struct perf_pending_entry *entry)
2869 struct perf_event *event = container_of(entry,
2870 struct perf_event, pending);
2872 if (event->pending_disable) {
2873 event->pending_disable = 0;
2874 __perf_event_disable(event);
2877 if (event->pending_wakeup) {
2878 event->pending_wakeup = 0;
2879 perf_event_wakeup(event);
2883 #define PENDING_TAIL ((struct perf_pending_entry *)-1UL)
2885 static DEFINE_PER_CPU(struct perf_pending_entry *, perf_pending_head) = {
2889 static void perf_pending_queue(struct perf_pending_entry *entry,
2890 void (*func)(struct perf_pending_entry *))
2892 struct perf_pending_entry **head;
2894 if (cmpxchg(&entry->next, NULL, PENDING_TAIL) != NULL)
2899 head = &get_cpu_var(perf_pending_head);
2902 entry->next = *head;
2903 } while (cmpxchg(head, entry->next, entry) != entry->next);
2905 set_perf_event_pending();
2907 put_cpu_var(perf_pending_head);
2910 static int __perf_pending_run(void)
2912 struct perf_pending_entry *list;
2915 list = xchg(&__get_cpu_var(perf_pending_head), PENDING_TAIL);
2916 while (list != PENDING_TAIL) {
2917 void (*func)(struct perf_pending_entry *);
2918 struct perf_pending_entry *entry = list;
2925 * Ensure we observe the unqueue before we issue the wakeup,
2926 * so that we won't be waiting forever.
2927 * -- see perf_not_pending().
2938 static inline int perf_not_pending(struct perf_event *event)
2941 * If we flush on whatever cpu we run, there is a chance we don't
2945 __perf_pending_run();
2949 * Ensure we see the proper queue state before going to sleep
2950 * so that we do not miss the wakeup. -- see perf_pending_handle()
2953 return event->pending.next == NULL;
2956 static void perf_pending_sync(struct perf_event *event)
2958 wait_event(event->waitq, perf_not_pending(event));
2961 void perf_event_do_pending(void)
2963 __perf_pending_run();
2967 * Callchain support -- arch specific
2970 __weak struct perf_callchain_entry *perf_callchain(struct pt_regs *regs)
2977 * We assume there is only KVM supporting the callbacks.
2978 * Later on, we might change it to a list if there is
2979 * another virtualization implementation supporting the callbacks.
2981 struct perf_guest_info_callbacks *perf_guest_cbs;
2983 int perf_register_guest_info_callbacks(struct perf_guest_info_callbacks *cbs)
2985 perf_guest_cbs = cbs;
2988 EXPORT_SYMBOL_GPL(perf_register_guest_info_callbacks);
2990 int perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks *cbs)
2992 perf_guest_cbs = NULL;
2995 EXPORT_SYMBOL_GPL(perf_unregister_guest_info_callbacks);
3000 static bool perf_output_space(struct perf_buffer *buffer, unsigned long tail,
3001 unsigned long offset, unsigned long head)
3005 if (!buffer->writable)
3008 mask = perf_data_size(buffer) - 1;
3010 offset = (offset - tail) & mask;
3011 head = (head - tail) & mask;
3013 if ((int)(head - offset) < 0)
3019 static void perf_output_wakeup(struct perf_output_handle *handle)
3021 atomic_set(&handle->buffer->poll, POLL_IN);
3024 handle->event->pending_wakeup = 1;
3025 perf_pending_queue(&handle->event->pending,
3026 perf_pending_event);
3028 perf_event_wakeup(handle->event);
3032 * We need to ensure a later event_id doesn't publish a head when a former
3033 * event isn't done writing. However since we need to deal with NMIs we
3034 * cannot fully serialize things.
3036 * We only publish the head (and generate a wakeup) when the outer-most
3039 static void perf_output_get_handle(struct perf_output_handle *handle)
3041 struct perf_buffer *buffer = handle->buffer;
3044 local_inc(&buffer->nest);
3045 handle->wakeup = local_read(&buffer->wakeup);
3048 static void perf_output_put_handle(struct perf_output_handle *handle)
3050 struct perf_buffer *buffer = handle->buffer;
3054 head = local_read(&buffer->head);
3057 * IRQ/NMI can happen here, which means we can miss a head update.
3060 if (!local_dec_and_test(&buffer->nest))
3064 * Publish the known good head. Rely on the full barrier implied
3065 * by atomic_dec_and_test() order the buffer->head read and this
3068 buffer->user_page->data_head = head;
3071 * Now check if we missed an update, rely on the (compiler)
3072 * barrier in atomic_dec_and_test() to re-read buffer->head.
3074 if (unlikely(head != local_read(&buffer->head))) {
3075 local_inc(&buffer->nest);
3079 if (handle->wakeup != local_read(&buffer->wakeup))
3080 perf_output_wakeup(handle);
3086 __always_inline void perf_output_copy(struct perf_output_handle *handle,
3087 const void *buf, unsigned int len)
3090 unsigned long size = min_t(unsigned long, handle->size, len);
3092 memcpy(handle->addr, buf, size);
3095 handle->addr += size;
3097 handle->size -= size;
3098 if (!handle->size) {
3099 struct perf_buffer *buffer = handle->buffer;
3102 handle->page &= buffer->nr_pages - 1;
3103 handle->addr = buffer->data_pages[handle->page];
3104 handle->size = PAGE_SIZE << page_order(buffer);
3109 int perf_output_begin(struct perf_output_handle *handle,
3110 struct perf_event *event, unsigned int size,
3111 int nmi, int sample)
3113 struct perf_buffer *buffer;
3114 unsigned long tail, offset, head;
3117 struct perf_event_header header;
3124 * For inherited events we send all the output towards the parent.
3127 event = event->parent;
3129 buffer = rcu_dereference(event->buffer);
3133 handle->buffer = buffer;
3134 handle->event = event;
3136 handle->sample = sample;
3138 if (!buffer->nr_pages)
3141 have_lost = local_read(&buffer->lost);
3143 size += sizeof(lost_event);
3145 perf_output_get_handle(handle);
3149 * Userspace could choose to issue a mb() before updating the
3150 * tail pointer. So that all reads will be completed before the
3153 tail = ACCESS_ONCE(buffer->user_page->data_tail);
3155 offset = head = local_read(&buffer->head);
3157 if (unlikely(!perf_output_space(buffer, tail, offset, head)))
3159 } while (local_cmpxchg(&buffer->head, offset, head) != offset);
3161 if (head - local_read(&buffer->wakeup) > buffer->watermark)
3162 local_add(buffer->watermark, &buffer->wakeup);
3164 handle->page = offset >> (PAGE_SHIFT + page_order(buffer));
3165 handle->page &= buffer->nr_pages - 1;
3166 handle->size = offset & ((PAGE_SIZE << page_order(buffer)) - 1);
3167 handle->addr = buffer->data_pages[handle->page];
3168 handle->addr += handle->size;
3169 handle->size = (PAGE_SIZE << page_order(buffer)) - handle->size;
3172 lost_event.header.type = PERF_RECORD_LOST;
3173 lost_event.header.misc = 0;
3174 lost_event.header.size = sizeof(lost_event);
3175 lost_event.id = event->id;
3176 lost_event.lost = local_xchg(&buffer->lost, 0);
3178 perf_output_put(handle, lost_event);
3184 local_inc(&buffer->lost);
3185 perf_output_put_handle(handle);
3192 void perf_output_end(struct perf_output_handle *handle)
3194 struct perf_event *event = handle->event;
3195 struct perf_buffer *buffer = handle->buffer;
3197 int wakeup_events = event->attr.wakeup_events;
3199 if (handle->sample && wakeup_events) {
3200 int events = local_inc_return(&buffer->events);
3201 if (events >= wakeup_events) {
3202 local_sub(wakeup_events, &buffer->events);
3203 local_inc(&buffer->wakeup);
3207 perf_output_put_handle(handle);
3211 static u32 perf_event_pid(struct perf_event *event, struct task_struct *p)
3214 * only top level events have the pid namespace they were created in
3217 event = event->parent;
3219 return task_tgid_nr_ns(p, event->ns);
3222 static u32 perf_event_tid(struct perf_event *event, struct task_struct *p)
3225 * only top level events have the pid namespace they were created in
3228 event = event->parent;
3230 return task_pid_nr_ns(p, event->ns);
3233 static void perf_output_read_one(struct perf_output_handle *handle,
3234 struct perf_event *event)
3236 u64 read_format = event->attr.read_format;
3240 values[n++] = perf_event_count(event);
3241 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
3242 values[n++] = event->total_time_enabled +
3243 atomic64_read(&event->child_total_time_enabled);
3245 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
3246 values[n++] = event->total_time_running +
3247 atomic64_read(&event->child_total_time_running);
3249 if (read_format & PERF_FORMAT_ID)
3250 values[n++] = primary_event_id(event);
3252 perf_output_copy(handle, values, n * sizeof(u64));
3256 * XXX PERF_FORMAT_GROUP vs inherited events seems difficult.
3258 static void perf_output_read_group(struct perf_output_handle *handle,
3259 struct perf_event *event)
3261 struct perf_event *leader = event->group_leader, *sub;
3262 u64 read_format = event->attr.read_format;
3266 values[n++] = 1 + leader->nr_siblings;
3268 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
3269 values[n++] = leader->total_time_enabled;
3271 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
3272 values[n++] = leader->total_time_running;
3274 if (leader != event)
3275 leader->pmu->read(leader);
3277 values[n++] = perf_event_count(leader);
3278 if (read_format & PERF_FORMAT_ID)
3279 values[n++] = primary_event_id(leader);
3281 perf_output_copy(handle, values, n * sizeof(u64));
3283 list_for_each_entry(sub, &leader->sibling_list, group_entry) {
3287 sub->pmu->read(sub);
3289 values[n++] = perf_event_count(sub);
3290 if (read_format & PERF_FORMAT_ID)
3291 values[n++] = primary_event_id(sub);
3293 perf_output_copy(handle, values, n * sizeof(u64));
3297 static void perf_output_read(struct perf_output_handle *handle,
3298 struct perf_event *event)
3300 if (event->attr.read_format & PERF_FORMAT_GROUP)
3301 perf_output_read_group(handle, event);
3303 perf_output_read_one(handle, event);
3306 void perf_output_sample(struct perf_output_handle *handle,
3307 struct perf_event_header *header,
3308 struct perf_sample_data *data,
3309 struct perf_event *event)
3311 u64 sample_type = data->type;
3313 perf_output_put(handle, *header);
3315 if (sample_type & PERF_SAMPLE_IP)
3316 perf_output_put(handle, data->ip);
3318 if (sample_type & PERF_SAMPLE_TID)
3319 perf_output_put(handle, data->tid_entry);
3321 if (sample_type & PERF_SAMPLE_TIME)
3322 perf_output_put(handle, data->time);
3324 if (sample_type & PERF_SAMPLE_ADDR)
3325 perf_output_put(handle, data->addr);
3327 if (sample_type & PERF_SAMPLE_ID)
3328 perf_output_put(handle, data->id);
3330 if (sample_type & PERF_SAMPLE_STREAM_ID)
3331 perf_output_put(handle, data->stream_id);
3333 if (sample_type & PERF_SAMPLE_CPU)
3334 perf_output_put(handle, data->cpu_entry);
3336 if (sample_type & PERF_SAMPLE_PERIOD)
3337 perf_output_put(handle, data->period);
3339 if (sample_type & PERF_SAMPLE_READ)
3340 perf_output_read(handle, event);
3342 if (sample_type & PERF_SAMPLE_CALLCHAIN) {
3343 if (data->callchain) {
3346 if (data->callchain)
3347 size += data->callchain->nr;
3349 size *= sizeof(u64);
3351 perf_output_copy(handle, data->callchain, size);
3354 perf_output_put(handle, nr);
3358 if (sample_type & PERF_SAMPLE_RAW) {
3360 perf_output_put(handle, data->raw->size);
3361 perf_output_copy(handle, data->raw->data,
3368 .size = sizeof(u32),
3371 perf_output_put(handle, raw);
3376 void perf_prepare_sample(struct perf_event_header *header,
3377 struct perf_sample_data *data,
3378 struct perf_event *event,
3379 struct pt_regs *regs)
3381 u64 sample_type = event->attr.sample_type;
3383 data->type = sample_type;
3385 header->type = PERF_RECORD_SAMPLE;
3386 header->size = sizeof(*header);
3389 header->misc |= perf_misc_flags(regs);
3391 if (sample_type & PERF_SAMPLE_IP) {
3392 data->ip = perf_instruction_pointer(regs);
3394 header->size += sizeof(data->ip);
3397 if (sample_type & PERF_SAMPLE_TID) {
3398 /* namespace issues */
3399 data->tid_entry.pid = perf_event_pid(event, current);
3400 data->tid_entry.tid = perf_event_tid(event, current);
3402 header->size += sizeof(data->tid_entry);
3405 if (sample_type & PERF_SAMPLE_TIME) {
3406 data->time = perf_clock();
3408 header->size += sizeof(data->time);
3411 if (sample_type & PERF_SAMPLE_ADDR)
3412 header->size += sizeof(data->addr);
3414 if (sample_type & PERF_SAMPLE_ID) {
3415 data->id = primary_event_id(event);
3417 header->size += sizeof(data->id);
3420 if (sample_type & PERF_SAMPLE_STREAM_ID) {
3421 data->stream_id = event->id;
3423 header->size += sizeof(data->stream_id);
3426 if (sample_type & PERF_SAMPLE_CPU) {
3427 data->cpu_entry.cpu = raw_smp_processor_id();
3428 data->cpu_entry.reserved = 0;
3430 header->size += sizeof(data->cpu_entry);
3433 if (sample_type & PERF_SAMPLE_PERIOD)
3434 header->size += sizeof(data->period);
3436 if (sample_type & PERF_SAMPLE_READ)
3437 header->size += perf_event_read_size(event);
3439 if (sample_type & PERF_SAMPLE_CALLCHAIN) {
3442 data->callchain = perf_callchain(regs);
3444 if (data->callchain)
3445 size += data->callchain->nr;
3447 header->size += size * sizeof(u64);
3450 if (sample_type & PERF_SAMPLE_RAW) {
3451 int size = sizeof(u32);
3454 size += data->raw->size;
3456 size += sizeof(u32);
3458 WARN_ON_ONCE(size & (sizeof(u64)-1));
3459 header->size += size;
3463 static void perf_event_output(struct perf_event *event, int nmi,
3464 struct perf_sample_data *data,
3465 struct pt_regs *regs)
3467 struct perf_output_handle handle;
3468 struct perf_event_header header;
3470 perf_prepare_sample(&header, data, event, regs);
3472 if (perf_output_begin(&handle, event, header.size, nmi, 1))
3475 perf_output_sample(&handle, &header, data, event);
3477 perf_output_end(&handle);
3484 struct perf_read_event {
3485 struct perf_event_header header;
3492 perf_event_read_event(struct perf_event *event,
3493 struct task_struct *task)
3495 struct perf_output_handle handle;
3496 struct perf_read_event read_event = {
3498 .type = PERF_RECORD_READ,
3500 .size = sizeof(read_event) + perf_event_read_size(event),
3502 .pid = perf_event_pid(event, task),
3503 .tid = perf_event_tid(event, task),
3507 ret = perf_output_begin(&handle, event, read_event.header.size, 0, 0);
3511 perf_output_put(&handle, read_event);
3512 perf_output_read(&handle, event);
3514 perf_output_end(&handle);
3518 * task tracking -- fork/exit
3520 * enabled by: attr.comm | attr.mmap | attr.mmap_data | attr.task
3523 struct perf_task_event {
3524 struct task_struct *task;
3525 struct perf_event_context *task_ctx;
3528 struct perf_event_header header;
3538 static void perf_event_task_output(struct perf_event *event,
3539 struct perf_task_event *task_event)
3541 struct perf_output_handle handle;
3542 struct task_struct *task = task_event->task;
3545 size = task_event->event_id.header.size;
3546 ret = perf_output_begin(&handle, event, size, 0, 0);
3551 task_event->event_id.pid = perf_event_pid(event, task);
3552 task_event->event_id.ppid = perf_event_pid(event, current);
3554 task_event->event_id.tid = perf_event_tid(event, task);
3555 task_event->event_id.ptid = perf_event_tid(event, current);
3557 perf_output_put(&handle, task_event->event_id);
3559 perf_output_end(&handle);
3562 static int perf_event_task_match(struct perf_event *event)
3564 if (event->state < PERF_EVENT_STATE_INACTIVE)
3567 if (event->cpu != -1 && event->cpu != smp_processor_id())
3570 if (event->attr.comm || event->attr.mmap ||
3571 event->attr.mmap_data || event->attr.task)
3577 static void perf_event_task_ctx(struct perf_event_context *ctx,
3578 struct perf_task_event *task_event)
3580 struct perf_event *event;
3582 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
3583 if (perf_event_task_match(event))
3584 perf_event_task_output(event, task_event);
3588 static void perf_event_task_event(struct perf_task_event *task_event)
3590 struct perf_cpu_context *cpuctx;
3591 struct perf_event_context *ctx = task_event->task_ctx;
3594 cpuctx = &get_cpu_var(perf_cpu_context);
3595 perf_event_task_ctx(&cpuctx->ctx, task_event);
3597 ctx = rcu_dereference(current->perf_event_ctxp);
3599 perf_event_task_ctx(ctx, task_event);
3600 put_cpu_var(perf_cpu_context);
3604 static void perf_event_task(struct task_struct *task,
3605 struct perf_event_context *task_ctx,
3608 struct perf_task_event task_event;
3610 if (!atomic_read(&nr_comm_events) &&
3611 !atomic_read(&nr_mmap_events) &&
3612 !atomic_read(&nr_task_events))
3615 task_event = (struct perf_task_event){
3617 .task_ctx = task_ctx,
3620 .type = new ? PERF_RECORD_FORK : PERF_RECORD_EXIT,
3622 .size = sizeof(task_event.event_id),
3628 .time = perf_clock(),
3632 perf_event_task_event(&task_event);
3635 void perf_event_fork(struct task_struct *task)
3637 perf_event_task(task, NULL, 1);
3644 struct perf_comm_event {
3645 struct task_struct *task;
3650 struct perf_event_header header;
3657 static void perf_event_comm_output(struct perf_event *event,
3658 struct perf_comm_event *comm_event)
3660 struct perf_output_handle handle;
3661 int size = comm_event->event_id.header.size;
3662 int ret = perf_output_begin(&handle, event, size, 0, 0);
3667 comm_event->event_id.pid = perf_event_pid(event, comm_event->task);
3668 comm_event->event_id.tid = perf_event_tid(event, comm_event->task);
3670 perf_output_put(&handle, comm_event->event_id);
3671 perf_output_copy(&handle, comm_event->comm,
3672 comm_event->comm_size);
3673 perf_output_end(&handle);
3676 static int perf_event_comm_match(struct perf_event *event)
3678 if (event->state < PERF_EVENT_STATE_INACTIVE)
3681 if (event->cpu != -1 && event->cpu != smp_processor_id())
3684 if (event->attr.comm)
3690 static void perf_event_comm_ctx(struct perf_event_context *ctx,
3691 struct perf_comm_event *comm_event)
3693 struct perf_event *event;
3695 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
3696 if (perf_event_comm_match(event))
3697 perf_event_comm_output(event, comm_event);
3701 static void perf_event_comm_event(struct perf_comm_event *comm_event)
3703 struct perf_cpu_context *cpuctx;
3704 struct perf_event_context *ctx;
3706 char comm[TASK_COMM_LEN];
3708 memset(comm, 0, sizeof(comm));
3709 strlcpy(comm, comm_event->task->comm, sizeof(comm));
3710 size = ALIGN(strlen(comm)+1, sizeof(u64));
3712 comm_event->comm = comm;
3713 comm_event->comm_size = size;
3715 comm_event->event_id.header.size = sizeof(comm_event->event_id) + size;
3718 cpuctx = &get_cpu_var(perf_cpu_context);
3719 perf_event_comm_ctx(&cpuctx->ctx, comm_event);
3720 ctx = rcu_dereference(current->perf_event_ctxp);
3722 perf_event_comm_ctx(ctx, comm_event);
3723 put_cpu_var(perf_cpu_context);
3727 void perf_event_comm(struct task_struct *task)
3729 struct perf_comm_event comm_event;
3731 if (task->perf_event_ctxp)
3732 perf_event_enable_on_exec(task);
3734 if (!atomic_read(&nr_comm_events))
3737 comm_event = (struct perf_comm_event){
3743 .type = PERF_RECORD_COMM,
3752 perf_event_comm_event(&comm_event);
3759 struct perf_mmap_event {
3760 struct vm_area_struct *vma;
3762 const char *file_name;
3766 struct perf_event_header header;
3776 static void perf_event_mmap_output(struct perf_event *event,
3777 struct perf_mmap_event *mmap_event)
3779 struct perf_output_handle handle;
3780 int size = mmap_event->event_id.header.size;
3781 int ret = perf_output_begin(&handle, event, size, 0, 0);
3786 mmap_event->event_id.pid = perf_event_pid(event, current);
3787 mmap_event->event_id.tid = perf_event_tid(event, current);
3789 perf_output_put(&handle, mmap_event->event_id);
3790 perf_output_copy(&handle, mmap_event->file_name,
3791 mmap_event->file_size);
3792 perf_output_end(&handle);
3795 static int perf_event_mmap_match(struct perf_event *event,
3796 struct perf_mmap_event *mmap_event,
3799 if (event->state < PERF_EVENT_STATE_INACTIVE)
3802 if (event->cpu != -1 && event->cpu != smp_processor_id())
3805 if ((!executable && event->attr.mmap_data) ||
3806 (executable && event->attr.mmap))
3812 static void perf_event_mmap_ctx(struct perf_event_context *ctx,
3813 struct perf_mmap_event *mmap_event,
3816 struct perf_event *event;
3818 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
3819 if (perf_event_mmap_match(event, mmap_event, executable))
3820 perf_event_mmap_output(event, mmap_event);
3824 static void perf_event_mmap_event(struct perf_mmap_event *mmap_event)
3826 struct perf_cpu_context *cpuctx;
3827 struct perf_event_context *ctx;
3828 struct vm_area_struct *vma = mmap_event->vma;
3829 struct file *file = vma->vm_file;
3835 memset(tmp, 0, sizeof(tmp));
3839 * d_path works from the end of the buffer backwards, so we
3840 * need to add enough zero bytes after the string to handle
3841 * the 64bit alignment we do later.
3843 buf = kzalloc(PATH_MAX + sizeof(u64), GFP_KERNEL);
3845 name = strncpy(tmp, "//enomem", sizeof(tmp));
3848 name = d_path(&file->f_path, buf, PATH_MAX);
3850 name = strncpy(tmp, "//toolong", sizeof(tmp));
3854 if (arch_vma_name(mmap_event->vma)) {
3855 name = strncpy(tmp, arch_vma_name(mmap_event->vma),
3861 name = strncpy(tmp, "[vdso]", sizeof(tmp));
3863 } else if (vma->vm_start <= vma->vm_mm->start_brk &&
3864 vma->vm_end >= vma->vm_mm->brk) {
3865 name = strncpy(tmp, "[heap]", sizeof(tmp));
3867 } else if (vma->vm_start <= vma->vm_mm->start_stack &&
3868 vma->vm_end >= vma->vm_mm->start_stack) {
3869 name = strncpy(tmp, "[stack]", sizeof(tmp));
3873 name = strncpy(tmp, "//anon", sizeof(tmp));
3878 size = ALIGN(strlen(name)+1, sizeof(u64));
3880 mmap_event->file_name = name;
3881 mmap_event->file_size = size;
3883 mmap_event->event_id.header.size = sizeof(mmap_event->event_id) + size;
3886 cpuctx = &get_cpu_var(perf_cpu_context);
3887 perf_event_mmap_ctx(&cpuctx->ctx, mmap_event, vma->vm_flags & VM_EXEC);
3888 ctx = rcu_dereference(current->perf_event_ctxp);
3890 perf_event_mmap_ctx(ctx, mmap_event, vma->vm_flags & VM_EXEC);
3891 put_cpu_var(perf_cpu_context);
3897 void perf_event_mmap(struct vm_area_struct *vma)
3899 struct perf_mmap_event mmap_event;
3901 if (!atomic_read(&nr_mmap_events))
3904 mmap_event = (struct perf_mmap_event){
3910 .type = PERF_RECORD_MMAP,
3911 .misc = PERF_RECORD_MISC_USER,
3916 .start = vma->vm_start,
3917 .len = vma->vm_end - vma->vm_start,
3918 .pgoff = (u64)vma->vm_pgoff << PAGE_SHIFT,
3922 perf_event_mmap_event(&mmap_event);
3926 * IRQ throttle logging
3929 static void perf_log_throttle(struct perf_event *event, int enable)
3931 struct perf_output_handle handle;
3935 struct perf_event_header header;
3939 } throttle_event = {
3941 .type = PERF_RECORD_THROTTLE,
3943 .size = sizeof(throttle_event),
3945 .time = perf_clock(),
3946 .id = primary_event_id(event),
3947 .stream_id = event->id,
3951 throttle_event.header.type = PERF_RECORD_UNTHROTTLE;
3953 ret = perf_output_begin(&handle, event, sizeof(throttle_event), 1, 0);
3957 perf_output_put(&handle, throttle_event);
3958 perf_output_end(&handle);
3962 * Generic event overflow handling, sampling.
3965 static int __perf_event_overflow(struct perf_event *event, int nmi,
3966 int throttle, struct perf_sample_data *data,
3967 struct pt_regs *regs)
3969 int events = atomic_read(&event->event_limit);
3970 struct hw_perf_event *hwc = &event->hw;
3973 throttle = (throttle && event->pmu->unthrottle != NULL);
3978 if (hwc->interrupts != MAX_INTERRUPTS) {
3980 if (HZ * hwc->interrupts >
3981 (u64)sysctl_perf_event_sample_rate) {
3982 hwc->interrupts = MAX_INTERRUPTS;
3983 perf_log_throttle(event, 0);
3988 * Keep re-disabling events even though on the previous
3989 * pass we disabled it - just in case we raced with a
3990 * sched-in and the event got enabled again:
3996 if (event->attr.freq) {
3997 u64 now = perf_clock();
3998 s64 delta = now - hwc->freq_time_stamp;
4000 hwc->freq_time_stamp = now;
4002 if (delta > 0 && delta < 2*TICK_NSEC)
4003 perf_adjust_period(event, delta, hwc->last_period);
4007 * XXX event_limit might not quite work as expected on inherited
4011 event->pending_kill = POLL_IN;
4012 if (events && atomic_dec_and_test(&event->event_limit)) {
4014 event->pending_kill = POLL_HUP;
4016 event->pending_disable = 1;
4017 perf_pending_queue(&event->pending,
4018 perf_pending_event);
4020 perf_event_disable(event);
4023 if (event->overflow_handler)
4024 event->overflow_handler(event, nmi, data, regs);
4026 perf_event_output(event, nmi, data, regs);
4031 int perf_event_overflow(struct perf_event *event, int nmi,
4032 struct perf_sample_data *data,
4033 struct pt_regs *regs)
4035 return __perf_event_overflow(event, nmi, 1, data, regs);
4039 * Generic software event infrastructure
4043 * We directly increment event->count and keep a second value in
4044 * event->hw.period_left to count intervals. This period event
4045 * is kept in the range [-sample_period, 0] so that we can use the
4049 static u64 perf_swevent_set_period(struct perf_event *event)
4051 struct hw_perf_event *hwc = &event->hw;
4052 u64 period = hwc->last_period;
4056 hwc->last_period = hwc->sample_period;
4059 old = val = local64_read(&hwc->period_left);
4063 nr = div64_u64(period + val, period);
4064 offset = nr * period;
4066 if (local64_cmpxchg(&hwc->period_left, old, val) != old)
4072 static void perf_swevent_overflow(struct perf_event *event, u64 overflow,
4073 int nmi, struct perf_sample_data *data,
4074 struct pt_regs *regs)
4076 struct hw_perf_event *hwc = &event->hw;
4079 data->period = event->hw.last_period;
4081 overflow = perf_swevent_set_period(event);
4083 if (hwc->interrupts == MAX_INTERRUPTS)
4086 for (; overflow; overflow--) {
4087 if (__perf_event_overflow(event, nmi, throttle,
4090 * We inhibit the overflow from happening when
4091 * hwc->interrupts == MAX_INTERRUPTS.
4099 static void perf_swevent_add(struct perf_event *event, u64 nr,
4100 int nmi, struct perf_sample_data *data,
4101 struct pt_regs *regs)
4103 struct hw_perf_event *hwc = &event->hw;
4105 local64_add(nr, &event->count);
4110 if (!hwc->sample_period)
4113 if (nr == 1 && hwc->sample_period == 1 && !event->attr.freq)
4114 return perf_swevent_overflow(event, 1, nmi, data, regs);
4116 if (local64_add_negative(nr, &hwc->period_left))
4119 perf_swevent_overflow(event, 0, nmi, data, regs);
4122 static int perf_exclude_event(struct perf_event *event,
4123 struct pt_regs *regs)
4126 if (event->attr.exclude_user && user_mode(regs))
4129 if (event->attr.exclude_kernel && !user_mode(regs))
4136 static int perf_swevent_match(struct perf_event *event,
4137 enum perf_type_id type,
4139 struct perf_sample_data *data,
4140 struct pt_regs *regs)
4142 if (event->attr.type != type)
4145 if (event->attr.config != event_id)
4148 if (perf_exclude_event(event, regs))
4154 static inline u64 swevent_hash(u64 type, u32 event_id)
4156 u64 val = event_id | (type << 32);
4158 return hash_64(val, SWEVENT_HLIST_BITS);
4161 static inline struct hlist_head *
4162 __find_swevent_head(struct swevent_hlist *hlist, u64 type, u32 event_id)
4164 u64 hash = swevent_hash(type, event_id);
4166 return &hlist->heads[hash];
4169 /* For the read side: events when they trigger */
4170 static inline struct hlist_head *
4171 find_swevent_head_rcu(struct perf_cpu_context *ctx, u64 type, u32 event_id)
4173 struct swevent_hlist *hlist;
4175 hlist = rcu_dereference(ctx->swevent_hlist);
4179 return __find_swevent_head(hlist, type, event_id);
4182 /* For the event head insertion and removal in the hlist */
4183 static inline struct hlist_head *
4184 find_swevent_head(struct perf_cpu_context *ctx, struct perf_event *event)
4186 struct swevent_hlist *hlist;
4187 u32 event_id = event->attr.config;
4188 u64 type = event->attr.type;
4191 * Event scheduling is always serialized against hlist allocation
4192 * and release. Which makes the protected version suitable here.
4193 * The context lock guarantees that.
4195 hlist = rcu_dereference_protected(ctx->swevent_hlist,
4196 lockdep_is_held(&event->ctx->lock));
4200 return __find_swevent_head(hlist, type, event_id);
4203 static void do_perf_sw_event(enum perf_type_id type, u32 event_id,
4205 struct perf_sample_data *data,
4206 struct pt_regs *regs)
4208 struct perf_cpu_context *cpuctx;
4209 struct perf_event *event;
4210 struct hlist_node *node;
4211 struct hlist_head *head;
4213 cpuctx = &__get_cpu_var(perf_cpu_context);
4217 head = find_swevent_head_rcu(cpuctx, type, event_id);
4222 hlist_for_each_entry_rcu(event, node, head, hlist_entry) {
4223 if (perf_swevent_match(event, type, event_id, data, regs))
4224 perf_swevent_add(event, nr, nmi, data, regs);
4230 int perf_swevent_get_recursion_context(void)
4232 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
4239 else if (in_softirq())
4244 if (cpuctx->recursion[rctx])
4247 cpuctx->recursion[rctx]++;
4252 EXPORT_SYMBOL_GPL(perf_swevent_get_recursion_context);
4254 void inline perf_swevent_put_recursion_context(int rctx)
4256 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
4258 cpuctx->recursion[rctx]--;
4261 void __perf_sw_event(u32 event_id, u64 nr, int nmi,
4262 struct pt_regs *regs, u64 addr)
4264 struct perf_sample_data data;
4267 preempt_disable_notrace();
4268 rctx = perf_swevent_get_recursion_context();
4272 perf_sample_data_init(&data, addr);
4274 do_perf_sw_event(PERF_TYPE_SOFTWARE, event_id, nr, nmi, &data, regs);
4276 perf_swevent_put_recursion_context(rctx);
4277 preempt_enable_notrace();
4280 static void perf_swevent_read(struct perf_event *event)
4284 static int perf_swevent_enable(struct perf_event *event)
4286 struct hw_perf_event *hwc = &event->hw;
4287 struct perf_cpu_context *cpuctx;
4288 struct hlist_head *head;
4290 cpuctx = &__get_cpu_var(perf_cpu_context);
4292 if (hwc->sample_period) {
4293 hwc->last_period = hwc->sample_period;
4294 perf_swevent_set_period(event);
4297 head = find_swevent_head(cpuctx, event);
4298 if (WARN_ON_ONCE(!head))
4301 hlist_add_head_rcu(&event->hlist_entry, head);
4306 static void perf_swevent_disable(struct perf_event *event)
4308 hlist_del_rcu(&event->hlist_entry);
4311 static void perf_swevent_void(struct perf_event *event)
4315 static int perf_swevent_int(struct perf_event *event)
4320 static const struct pmu perf_ops_generic = {
4321 .enable = perf_swevent_enable,
4322 .disable = perf_swevent_disable,
4323 .start = perf_swevent_int,
4324 .stop = perf_swevent_void,
4325 .read = perf_swevent_read,
4326 .unthrottle = perf_swevent_void, /* hwc->interrupts already reset */
4330 * hrtimer based swevent callback
4333 static enum hrtimer_restart perf_swevent_hrtimer(struct hrtimer *hrtimer)
4335 enum hrtimer_restart ret = HRTIMER_RESTART;
4336 struct perf_sample_data data;
4337 struct pt_regs *regs;
4338 struct perf_event *event;
4341 event = container_of(hrtimer, struct perf_event, hw.hrtimer);
4342 event->pmu->read(event);
4344 perf_sample_data_init(&data, 0);
4345 data.period = event->hw.last_period;
4346 regs = get_irq_regs();
4348 if (regs && !perf_exclude_event(event, regs)) {
4349 if (!(event->attr.exclude_idle && current->pid == 0))
4350 if (perf_event_overflow(event, 0, &data, regs))
4351 ret = HRTIMER_NORESTART;
4354 period = max_t(u64, 10000, event->hw.sample_period);
4355 hrtimer_forward_now(hrtimer, ns_to_ktime(period));
4360 static void perf_swevent_start_hrtimer(struct perf_event *event)
4362 struct hw_perf_event *hwc = &event->hw;
4364 hrtimer_init(&hwc->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
4365 hwc->hrtimer.function = perf_swevent_hrtimer;
4366 if (hwc->sample_period) {
4369 if (hwc->remaining) {
4370 if (hwc->remaining < 0)
4373 period = hwc->remaining;
4376 period = max_t(u64, 10000, hwc->sample_period);
4378 __hrtimer_start_range_ns(&hwc->hrtimer,
4379 ns_to_ktime(period), 0,
4380 HRTIMER_MODE_REL, 0);
4384 static void perf_swevent_cancel_hrtimer(struct perf_event *event)
4386 struct hw_perf_event *hwc = &event->hw;
4388 if (hwc->sample_period) {
4389 ktime_t remaining = hrtimer_get_remaining(&hwc->hrtimer);
4390 hwc->remaining = ktime_to_ns(remaining);
4392 hrtimer_cancel(&hwc->hrtimer);
4397 * Software event: cpu wall time clock
4400 static void cpu_clock_perf_event_update(struct perf_event *event)
4402 int cpu = raw_smp_processor_id();
4406 now = cpu_clock(cpu);
4407 prev = local64_xchg(&event->hw.prev_count, now);
4408 local64_add(now - prev, &event->count);
4411 static int cpu_clock_perf_event_enable(struct perf_event *event)
4413 struct hw_perf_event *hwc = &event->hw;
4414 int cpu = raw_smp_processor_id();
4416 local64_set(&hwc->prev_count, cpu_clock(cpu));
4417 perf_swevent_start_hrtimer(event);
4422 static void cpu_clock_perf_event_disable(struct perf_event *event)
4424 perf_swevent_cancel_hrtimer(event);
4425 cpu_clock_perf_event_update(event);
4428 static void cpu_clock_perf_event_read(struct perf_event *event)
4430 cpu_clock_perf_event_update(event);
4433 static const struct pmu perf_ops_cpu_clock = {
4434 .enable = cpu_clock_perf_event_enable,
4435 .disable = cpu_clock_perf_event_disable,
4436 .read = cpu_clock_perf_event_read,
4440 * Software event: task time clock
4443 static void task_clock_perf_event_update(struct perf_event *event, u64 now)
4448 prev = local64_xchg(&event->hw.prev_count, now);
4450 local64_add(delta, &event->count);
4453 static int task_clock_perf_event_enable(struct perf_event *event)
4455 struct hw_perf_event *hwc = &event->hw;
4458 now = event->ctx->time;
4460 local64_set(&hwc->prev_count, now);
4462 perf_swevent_start_hrtimer(event);
4467 static void task_clock_perf_event_disable(struct perf_event *event)
4469 perf_swevent_cancel_hrtimer(event);
4470 task_clock_perf_event_update(event, event->ctx->time);
4474 static void task_clock_perf_event_read(struct perf_event *event)
4479 update_context_time(event->ctx);
4480 time = event->ctx->time;
4482 u64 now = perf_clock();
4483 u64 delta = now - event->ctx->timestamp;
4484 time = event->ctx->time + delta;
4487 task_clock_perf_event_update(event, time);
4490 static const struct pmu perf_ops_task_clock = {
4491 .enable = task_clock_perf_event_enable,
4492 .disable = task_clock_perf_event_disable,
4493 .read = task_clock_perf_event_read,
4496 /* Deref the hlist from the update side */
4497 static inline struct swevent_hlist *
4498 swevent_hlist_deref(struct perf_cpu_context *cpuctx)
4500 return rcu_dereference_protected(cpuctx->swevent_hlist,
4501 lockdep_is_held(&cpuctx->hlist_mutex));
4504 static void swevent_hlist_release_rcu(struct rcu_head *rcu_head)
4506 struct swevent_hlist *hlist;
4508 hlist = container_of(rcu_head, struct swevent_hlist, rcu_head);
4512 static void swevent_hlist_release(struct perf_cpu_context *cpuctx)
4514 struct swevent_hlist *hlist = swevent_hlist_deref(cpuctx);
4519 rcu_assign_pointer(cpuctx->swevent_hlist, NULL);
4520 call_rcu(&hlist->rcu_head, swevent_hlist_release_rcu);
4523 static void swevent_hlist_put_cpu(struct perf_event *event, int cpu)
4525 struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
4527 mutex_lock(&cpuctx->hlist_mutex);
4529 if (!--cpuctx->hlist_refcount)
4530 swevent_hlist_release(cpuctx);
4532 mutex_unlock(&cpuctx->hlist_mutex);
4535 static void swevent_hlist_put(struct perf_event *event)
4539 if (event->cpu != -1) {
4540 swevent_hlist_put_cpu(event, event->cpu);
4544 for_each_possible_cpu(cpu)
4545 swevent_hlist_put_cpu(event, cpu);
4548 static int swevent_hlist_get_cpu(struct perf_event *event, int cpu)
4550 struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
4553 mutex_lock(&cpuctx->hlist_mutex);
4555 if (!swevent_hlist_deref(cpuctx) && cpu_online(cpu)) {
4556 struct swevent_hlist *hlist;
4558 hlist = kzalloc(sizeof(*hlist), GFP_KERNEL);
4563 rcu_assign_pointer(cpuctx->swevent_hlist, hlist);
4565 cpuctx->hlist_refcount++;
4567 mutex_unlock(&cpuctx->hlist_mutex);
4572 static int swevent_hlist_get(struct perf_event *event)
4575 int cpu, failed_cpu;
4577 if (event->cpu != -1)
4578 return swevent_hlist_get_cpu(event, event->cpu);
4581 for_each_possible_cpu(cpu) {
4582 err = swevent_hlist_get_cpu(event, cpu);
4592 for_each_possible_cpu(cpu) {
4593 if (cpu == failed_cpu)
4595 swevent_hlist_put_cpu(event, cpu);
4602 #ifdef CONFIG_EVENT_TRACING
4604 static const struct pmu perf_ops_tracepoint = {
4605 .enable = perf_trace_enable,
4606 .disable = perf_trace_disable,
4607 .start = perf_swevent_int,
4608 .stop = perf_swevent_void,
4609 .read = perf_swevent_read,
4610 .unthrottle = perf_swevent_void,
4613 static int perf_tp_filter_match(struct perf_event *event,
4614 struct perf_sample_data *data)
4616 void *record = data->raw->data;
4618 if (likely(!event->filter) || filter_match_preds(event->filter, record))
4623 static int perf_tp_event_match(struct perf_event *event,
4624 struct perf_sample_data *data,
4625 struct pt_regs *regs)
4628 * All tracepoints are from kernel-space.
4630 if (event->attr.exclude_kernel)
4633 if (!perf_tp_filter_match(event, data))
4639 void perf_tp_event(u64 addr, u64 count, void *record, int entry_size,
4640 struct pt_regs *regs, struct hlist_head *head, int rctx)
4642 struct perf_sample_data data;
4643 struct perf_event *event;
4644 struct hlist_node *node;
4646 struct perf_raw_record raw = {
4651 perf_sample_data_init(&data, addr);
4654 hlist_for_each_entry_rcu(event, node, head, hlist_entry) {
4655 if (perf_tp_event_match(event, &data, regs))
4656 perf_swevent_add(event, count, 1, &data, regs);
4659 perf_swevent_put_recursion_context(rctx);
4661 EXPORT_SYMBOL_GPL(perf_tp_event);
4663 static void tp_perf_event_destroy(struct perf_event *event)
4665 perf_trace_destroy(event);
4668 static const struct pmu *tp_perf_event_init(struct perf_event *event)
4673 * Raw tracepoint data is a severe data leak, only allow root to
4676 if ((event->attr.sample_type & PERF_SAMPLE_RAW) &&
4677 perf_paranoid_tracepoint_raw() &&
4678 !capable(CAP_SYS_ADMIN))
4679 return ERR_PTR(-EPERM);
4681 err = perf_trace_init(event);
4685 event->destroy = tp_perf_event_destroy;
4687 return &perf_ops_tracepoint;
4690 static int perf_event_set_filter(struct perf_event *event, void __user *arg)
4695 if (event->attr.type != PERF_TYPE_TRACEPOINT)
4698 filter_str = strndup_user(arg, PAGE_SIZE);
4699 if (IS_ERR(filter_str))
4700 return PTR_ERR(filter_str);
4702 ret = ftrace_profile_set_filter(event, event->attr.config, filter_str);
4708 static void perf_event_free_filter(struct perf_event *event)
4710 ftrace_profile_free_filter(event);
4715 static const struct pmu *tp_perf_event_init(struct perf_event *event)
4720 static int perf_event_set_filter(struct perf_event *event, void __user *arg)
4725 static void perf_event_free_filter(struct perf_event *event)
4729 #endif /* CONFIG_EVENT_TRACING */
4731 #ifdef CONFIG_HAVE_HW_BREAKPOINT
4732 static void bp_perf_event_destroy(struct perf_event *event)
4734 release_bp_slot(event);
4737 static const struct pmu *bp_perf_event_init(struct perf_event *bp)
4741 err = register_perf_hw_breakpoint(bp);
4743 return ERR_PTR(err);
4745 bp->destroy = bp_perf_event_destroy;
4747 return &perf_ops_bp;
4750 void perf_bp_event(struct perf_event *bp, void *data)
4752 struct perf_sample_data sample;
4753 struct pt_regs *regs = data;
4755 perf_sample_data_init(&sample, bp->attr.bp_addr);
4757 if (!perf_exclude_event(bp, regs))
4758 perf_swevent_add(bp, 1, 1, &sample, regs);
4761 static const struct pmu *bp_perf_event_init(struct perf_event *bp)
4766 void perf_bp_event(struct perf_event *bp, void *regs)
4771 atomic_t perf_swevent_enabled[PERF_COUNT_SW_MAX];
4773 static void sw_perf_event_destroy(struct perf_event *event)
4775 u64 event_id = event->attr.config;
4777 WARN_ON(event->parent);
4779 atomic_dec(&perf_swevent_enabled[event_id]);
4780 swevent_hlist_put(event);
4783 static const struct pmu *sw_perf_event_init(struct perf_event *event)
4785 const struct pmu *pmu = NULL;
4786 u64 event_id = event->attr.config;
4789 * Software events (currently) can't in general distinguish
4790 * between user, kernel and hypervisor events.
4791 * However, context switches and cpu migrations are considered
4792 * to be kernel events, and page faults are never hypervisor
4796 case PERF_COUNT_SW_CPU_CLOCK:
4797 pmu = &perf_ops_cpu_clock;
4800 case PERF_COUNT_SW_TASK_CLOCK:
4802 * If the user instantiates this as a per-cpu event,
4803 * use the cpu_clock event instead.
4805 if (event->ctx->task)
4806 pmu = &perf_ops_task_clock;
4808 pmu = &perf_ops_cpu_clock;
4811 case PERF_COUNT_SW_PAGE_FAULTS:
4812 case PERF_COUNT_SW_PAGE_FAULTS_MIN:
4813 case PERF_COUNT_SW_PAGE_FAULTS_MAJ:
4814 case PERF_COUNT_SW_CONTEXT_SWITCHES:
4815 case PERF_COUNT_SW_CPU_MIGRATIONS:
4816 case PERF_COUNT_SW_ALIGNMENT_FAULTS:
4817 case PERF_COUNT_SW_EMULATION_FAULTS:
4818 if (!event->parent) {
4821 err = swevent_hlist_get(event);
4823 return ERR_PTR(err);
4825 atomic_inc(&perf_swevent_enabled[event_id]);
4826 event->destroy = sw_perf_event_destroy;
4828 pmu = &perf_ops_generic;
4836 * Allocate and initialize a event structure
4838 static struct perf_event *
4839 perf_event_alloc(struct perf_event_attr *attr,
4841 struct perf_event_context *ctx,
4842 struct perf_event *group_leader,
4843 struct perf_event *parent_event,
4844 perf_overflow_handler_t overflow_handler,
4847 const struct pmu *pmu;
4848 struct perf_event *event;
4849 struct hw_perf_event *hwc;
4852 event = kzalloc(sizeof(*event), gfpflags);
4854 return ERR_PTR(-ENOMEM);
4857 * Single events are their own group leaders, with an
4858 * empty sibling list:
4861 group_leader = event;
4863 mutex_init(&event->child_mutex);
4864 INIT_LIST_HEAD(&event->child_list);
4866 INIT_LIST_HEAD(&event->group_entry);
4867 INIT_LIST_HEAD(&event->event_entry);
4868 INIT_LIST_HEAD(&event->sibling_list);
4869 init_waitqueue_head(&event->waitq);
4871 mutex_init(&event->mmap_mutex);
4874 event->attr = *attr;
4875 event->group_leader = group_leader;
4880 event->parent = parent_event;
4882 event->ns = get_pid_ns(current->nsproxy->pid_ns);
4883 event->id = atomic64_inc_return(&perf_event_id);
4885 event->state = PERF_EVENT_STATE_INACTIVE;
4887 if (!overflow_handler && parent_event)
4888 overflow_handler = parent_event->overflow_handler;
4890 event->overflow_handler = overflow_handler;
4893 event->state = PERF_EVENT_STATE_OFF;
4898 hwc->sample_period = attr->sample_period;
4899 if (attr->freq && attr->sample_freq)
4900 hwc->sample_period = 1;
4901 hwc->last_period = hwc->sample_period;
4903 local64_set(&hwc->period_left, hwc->sample_period);
4906 * we currently do not support PERF_FORMAT_GROUP on inherited events
4908 if (attr->inherit && (attr->read_format & PERF_FORMAT_GROUP))
4911 switch (attr->type) {
4913 case PERF_TYPE_HARDWARE:
4914 case PERF_TYPE_HW_CACHE:
4915 pmu = hw_perf_event_init(event);
4918 case PERF_TYPE_SOFTWARE:
4919 pmu = sw_perf_event_init(event);
4922 case PERF_TYPE_TRACEPOINT:
4923 pmu = tp_perf_event_init(event);
4926 case PERF_TYPE_BREAKPOINT:
4927 pmu = bp_perf_event_init(event);
4938 else if (IS_ERR(pmu))
4943 put_pid_ns(event->ns);
4945 return ERR_PTR(err);
4950 if (!event->parent) {
4951 atomic_inc(&nr_events);
4952 if (event->attr.mmap || event->attr.mmap_data)
4953 atomic_inc(&nr_mmap_events);
4954 if (event->attr.comm)
4955 atomic_inc(&nr_comm_events);
4956 if (event->attr.task)
4957 atomic_inc(&nr_task_events);
4963 static int perf_copy_attr(struct perf_event_attr __user *uattr,
4964 struct perf_event_attr *attr)
4969 if (!access_ok(VERIFY_WRITE, uattr, PERF_ATTR_SIZE_VER0))
4973 * zero the full structure, so that a short copy will be nice.
4975 memset(attr, 0, sizeof(*attr));
4977 ret = get_user(size, &uattr->size);
4981 if (size > PAGE_SIZE) /* silly large */
4984 if (!size) /* abi compat */
4985 size = PERF_ATTR_SIZE_VER0;
4987 if (size < PERF_ATTR_SIZE_VER0)
4991 * If we're handed a bigger struct than we know of,
4992 * ensure all the unknown bits are 0 - i.e. new
4993 * user-space does not rely on any kernel feature
4994 * extensions we dont know about yet.
4996 if (size > sizeof(*attr)) {
4997 unsigned char __user *addr;
4998 unsigned char __user *end;
5001 addr = (void __user *)uattr + sizeof(*attr);
5002 end = (void __user *)uattr + size;
5004 for (; addr < end; addr++) {
5005 ret = get_user(val, addr);
5011 size = sizeof(*attr);
5014 ret = copy_from_user(attr, uattr, size);
5019 * If the type exists, the corresponding creation will verify
5022 if (attr->type >= PERF_TYPE_MAX)
5025 if (attr->__reserved_1)
5028 if (attr->sample_type & ~(PERF_SAMPLE_MAX-1))
5031 if (attr->read_format & ~(PERF_FORMAT_MAX-1))
5038 put_user(sizeof(*attr), &uattr->size);
5044 perf_event_set_output(struct perf_event *event, struct perf_event *output_event)
5046 struct perf_buffer *buffer = NULL, *old_buffer = NULL;
5052 /* don't allow circular references */
5053 if (event == output_event)
5057 * Don't allow cross-cpu buffers
5059 if (output_event->cpu != event->cpu)
5063 * If its not a per-cpu buffer, it must be the same task.
5065 if (output_event->cpu == -1 && output_event->ctx != event->ctx)
5069 mutex_lock(&event->mmap_mutex);
5070 /* Can't redirect output if we've got an active mmap() */
5071 if (atomic_read(&event->mmap_count))
5075 /* get the buffer we want to redirect to */
5076 buffer = perf_buffer_get(output_event);
5081 old_buffer = event->buffer;
5082 rcu_assign_pointer(event->buffer, buffer);
5085 mutex_unlock(&event->mmap_mutex);
5088 perf_buffer_put(old_buffer);
5094 * sys_perf_event_open - open a performance event, associate it to a task/cpu
5096 * @attr_uptr: event_id type attributes for monitoring/sampling
5099 * @group_fd: group leader event fd
5101 SYSCALL_DEFINE5(perf_event_open,
5102 struct perf_event_attr __user *, attr_uptr,
5103 pid_t, pid, int, cpu, int, group_fd, unsigned long, flags)
5105 struct perf_event *event, *group_leader = NULL, *output_event = NULL;
5106 struct perf_event_attr attr;
5107 struct perf_event_context *ctx;
5108 struct file *event_file = NULL;
5109 struct file *group_file = NULL;
5111 int fput_needed = 0;
5114 /* for future expandability... */
5115 if (flags & ~(PERF_FLAG_FD_NO_GROUP | PERF_FLAG_FD_OUTPUT))
5118 err = perf_copy_attr(attr_uptr, &attr);
5122 if (!attr.exclude_kernel) {
5123 if (perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN))
5128 if (attr.sample_freq > sysctl_perf_event_sample_rate)
5132 event_fd = get_unused_fd_flags(O_RDWR);
5137 * Get the target context (task or percpu):
5139 ctx = find_get_context(pid, cpu);
5145 if (group_fd != -1) {
5146 group_leader = perf_fget_light(group_fd, &fput_needed);
5147 if (IS_ERR(group_leader)) {
5148 err = PTR_ERR(group_leader);
5149 goto err_put_context;
5151 group_file = group_leader->filp;
5152 if (flags & PERF_FLAG_FD_OUTPUT)
5153 output_event = group_leader;
5154 if (flags & PERF_FLAG_FD_NO_GROUP)
5155 group_leader = NULL;
5159 * Look up the group leader (we will attach this event to it):
5165 * Do not allow a recursive hierarchy (this new sibling
5166 * becoming part of another group-sibling):
5168 if (group_leader->group_leader != group_leader)
5169 goto err_put_context;
5171 * Do not allow to attach to a group in a different
5172 * task or CPU context:
5174 if (group_leader->ctx != ctx)
5175 goto err_put_context;
5177 * Only a group leader can be exclusive or pinned
5179 if (attr.exclusive || attr.pinned)
5180 goto err_put_context;
5183 event = perf_event_alloc(&attr, cpu, ctx, group_leader,
5184 NULL, NULL, GFP_KERNEL);
5185 if (IS_ERR(event)) {
5186 err = PTR_ERR(event);
5187 goto err_put_context;
5191 err = perf_event_set_output(event, output_event);
5193 goto err_free_put_context;
5196 event_file = anon_inode_getfile("[perf_event]", &perf_fops, event, O_RDWR);
5197 if (IS_ERR(event_file)) {
5198 err = PTR_ERR(event_file);
5199 goto err_free_put_context;
5202 event->filp = event_file;
5203 WARN_ON_ONCE(ctx->parent_ctx);
5204 mutex_lock(&ctx->mutex);
5205 perf_install_in_context(ctx, event, cpu);
5207 mutex_unlock(&ctx->mutex);
5209 event->owner = current;
5210 get_task_struct(current);
5211 mutex_lock(¤t->perf_event_mutex);
5212 list_add_tail(&event->owner_entry, ¤t->perf_event_list);
5213 mutex_unlock(¤t->perf_event_mutex);
5216 * Drop the reference on the group_event after placing the
5217 * new event on the sibling_list. This ensures destruction
5218 * of the group leader will find the pointer to itself in
5219 * perf_group_detach().
5221 fput_light(group_file, fput_needed);
5222 fd_install(event_fd, event_file);
5225 err_free_put_context:
5228 fput_light(group_file, fput_needed);
5231 put_unused_fd(event_fd);
5236 * perf_event_create_kernel_counter
5238 * @attr: attributes of the counter to create
5239 * @cpu: cpu in which the counter is bound
5240 * @pid: task to profile
5243 perf_event_create_kernel_counter(struct perf_event_attr *attr, int cpu,
5245 perf_overflow_handler_t overflow_handler)
5247 struct perf_event *event;
5248 struct perf_event_context *ctx;
5252 * Get the target context (task or percpu):
5255 ctx = find_get_context(pid, cpu);
5261 event = perf_event_alloc(attr, cpu, ctx, NULL,
5262 NULL, overflow_handler, GFP_KERNEL);
5263 if (IS_ERR(event)) {
5264 err = PTR_ERR(event);
5265 goto err_put_context;
5269 WARN_ON_ONCE(ctx->parent_ctx);
5270 mutex_lock(&ctx->mutex);
5271 perf_install_in_context(ctx, event, cpu);
5273 mutex_unlock(&ctx->mutex);
5275 event->owner = current;
5276 get_task_struct(current);
5277 mutex_lock(¤t->perf_event_mutex);
5278 list_add_tail(&event->owner_entry, ¤t->perf_event_list);
5279 mutex_unlock(¤t->perf_event_mutex);
5286 return ERR_PTR(err);
5288 EXPORT_SYMBOL_GPL(perf_event_create_kernel_counter);
5291 * inherit a event from parent task to child task:
5293 static struct perf_event *
5294 inherit_event(struct perf_event *parent_event,
5295 struct task_struct *parent,
5296 struct perf_event_context *parent_ctx,
5297 struct task_struct *child,
5298 struct perf_event *group_leader,
5299 struct perf_event_context *child_ctx)
5301 struct perf_event *child_event;
5304 * Instead of creating recursive hierarchies of events,
5305 * we link inherited events back to the original parent,
5306 * which has a filp for sure, which we use as the reference
5309 if (parent_event->parent)
5310 parent_event = parent_event->parent;
5312 child_event = perf_event_alloc(&parent_event->attr,
5313 parent_event->cpu, child_ctx,
5314 group_leader, parent_event,
5316 if (IS_ERR(child_event))
5321 * Make the child state follow the state of the parent event,
5322 * not its attr.disabled bit. We hold the parent's mutex,
5323 * so we won't race with perf_event_{en, dis}able_family.
5325 if (parent_event->state >= PERF_EVENT_STATE_INACTIVE)
5326 child_event->state = PERF_EVENT_STATE_INACTIVE;
5328 child_event->state = PERF_EVENT_STATE_OFF;
5330 if (parent_event->attr.freq) {
5331 u64 sample_period = parent_event->hw.sample_period;
5332 struct hw_perf_event *hwc = &child_event->hw;
5334 hwc->sample_period = sample_period;
5335 hwc->last_period = sample_period;
5337 local64_set(&hwc->period_left, sample_period);
5340 child_event->overflow_handler = parent_event->overflow_handler;
5343 * Link it up in the child's context:
5345 add_event_to_ctx(child_event, child_ctx);
5348 * Get a reference to the parent filp - we will fput it
5349 * when the child event exits. This is safe to do because
5350 * we are in the parent and we know that the filp still
5351 * exists and has a nonzero count:
5353 atomic_long_inc(&parent_event->filp->f_count);
5356 * Link this into the parent event's child list
5358 WARN_ON_ONCE(parent_event->ctx->parent_ctx);
5359 mutex_lock(&parent_event->child_mutex);
5360 list_add_tail(&child_event->child_list, &parent_event->child_list);
5361 mutex_unlock(&parent_event->child_mutex);
5366 static int inherit_group(struct perf_event *parent_event,
5367 struct task_struct *parent,
5368 struct perf_event_context *parent_ctx,
5369 struct task_struct *child,
5370 struct perf_event_context *child_ctx)
5372 struct perf_event *leader;
5373 struct perf_event *sub;
5374 struct perf_event *child_ctr;
5376 leader = inherit_event(parent_event, parent, parent_ctx,
5377 child, NULL, child_ctx);
5379 return PTR_ERR(leader);
5380 list_for_each_entry(sub, &parent_event->sibling_list, group_entry) {
5381 child_ctr = inherit_event(sub, parent, parent_ctx,
5382 child, leader, child_ctx);
5383 if (IS_ERR(child_ctr))
5384 return PTR_ERR(child_ctr);
5389 static void sync_child_event(struct perf_event *child_event,
5390 struct task_struct *child)
5392 struct perf_event *parent_event = child_event->parent;
5395 if (child_event->attr.inherit_stat)
5396 perf_event_read_event(child_event, child);
5398 child_val = perf_event_count(child_event);
5401 * Add back the child's count to the parent's count:
5403 atomic64_add(child_val, &parent_event->child_count);
5404 atomic64_add(child_event->total_time_enabled,
5405 &parent_event->child_total_time_enabled);
5406 atomic64_add(child_event->total_time_running,
5407 &parent_event->child_total_time_running);
5410 * Remove this event from the parent's list
5412 WARN_ON_ONCE(parent_event->ctx->parent_ctx);
5413 mutex_lock(&parent_event->child_mutex);
5414 list_del_init(&child_event->child_list);
5415 mutex_unlock(&parent_event->child_mutex);
5418 * Release the parent event, if this was the last
5421 fput(parent_event->filp);
5425 __perf_event_exit_task(struct perf_event *child_event,
5426 struct perf_event_context *child_ctx,
5427 struct task_struct *child)
5429 struct perf_event *parent_event;
5431 perf_event_remove_from_context(child_event);
5433 parent_event = child_event->parent;
5435 * It can happen that parent exits first, and has events
5436 * that are still around due to the child reference. These
5437 * events need to be zapped - but otherwise linger.
5440 sync_child_event(child_event, child);
5441 free_event(child_event);
5446 * When a child task exits, feed back event values to parent events.
5448 void perf_event_exit_task(struct task_struct *child)
5450 struct perf_event *child_event, *tmp;
5451 struct perf_event_context *child_ctx;
5452 unsigned long flags;
5454 if (likely(!child->perf_event_ctxp)) {
5455 perf_event_task(child, NULL, 0);
5459 local_irq_save(flags);
5461 * We can't reschedule here because interrupts are disabled,
5462 * and either child is current or it is a task that can't be
5463 * scheduled, so we are now safe from rescheduling changing
5466 child_ctx = child->perf_event_ctxp;
5467 __perf_event_task_sched_out(child_ctx);
5470 * Take the context lock here so that if find_get_context is
5471 * reading child->perf_event_ctxp, we wait until it has
5472 * incremented the context's refcount before we do put_ctx below.
5474 raw_spin_lock(&child_ctx->lock);
5475 child->perf_event_ctxp = NULL;
5477 * If this context is a clone; unclone it so it can't get
5478 * swapped to another process while we're removing all
5479 * the events from it.
5481 unclone_ctx(child_ctx);
5482 update_context_time(child_ctx);
5483 raw_spin_unlock_irqrestore(&child_ctx->lock, flags);
5486 * Report the task dead after unscheduling the events so that we
5487 * won't get any samples after PERF_RECORD_EXIT. We can however still
5488 * get a few PERF_RECORD_READ events.
5490 perf_event_task(child, child_ctx, 0);
5493 * We can recurse on the same lock type through:
5495 * __perf_event_exit_task()
5496 * sync_child_event()
5497 * fput(parent_event->filp)
5499 * mutex_lock(&ctx->mutex)
5501 * But since its the parent context it won't be the same instance.
5503 mutex_lock(&child_ctx->mutex);
5506 list_for_each_entry_safe(child_event, tmp, &child_ctx->pinned_groups,
5508 __perf_event_exit_task(child_event, child_ctx, child);
5510 list_for_each_entry_safe(child_event, tmp, &child_ctx->flexible_groups,
5512 __perf_event_exit_task(child_event, child_ctx, child);
5515 * If the last event was a group event, it will have appended all
5516 * its siblings to the list, but we obtained 'tmp' before that which
5517 * will still point to the list head terminating the iteration.
5519 if (!list_empty(&child_ctx->pinned_groups) ||
5520 !list_empty(&child_ctx->flexible_groups))
5523 mutex_unlock(&child_ctx->mutex);
5528 static void perf_free_event(struct perf_event *event,
5529 struct perf_event_context *ctx)
5531 struct perf_event *parent = event->parent;
5533 if (WARN_ON_ONCE(!parent))
5536 mutex_lock(&parent->child_mutex);
5537 list_del_init(&event->child_list);
5538 mutex_unlock(&parent->child_mutex);
5542 perf_group_detach(event);
5543 list_del_event(event, ctx);
5548 * free an unexposed, unused context as created by inheritance by
5549 * init_task below, used by fork() in case of fail.
5551 void perf_event_free_task(struct task_struct *task)
5553 struct perf_event_context *ctx = task->perf_event_ctxp;
5554 struct perf_event *event, *tmp;
5559 mutex_lock(&ctx->mutex);
5561 list_for_each_entry_safe(event, tmp, &ctx->pinned_groups, group_entry)
5562 perf_free_event(event, ctx);
5564 list_for_each_entry_safe(event, tmp, &ctx->flexible_groups,
5566 perf_free_event(event, ctx);
5568 if (!list_empty(&ctx->pinned_groups) ||
5569 !list_empty(&ctx->flexible_groups))
5572 mutex_unlock(&ctx->mutex);
5578 inherit_task_group(struct perf_event *event, struct task_struct *parent,
5579 struct perf_event_context *parent_ctx,
5580 struct task_struct *child,
5584 struct perf_event_context *child_ctx = child->perf_event_ctxp;
5586 if (!event->attr.inherit) {
5593 * This is executed from the parent task context, so
5594 * inherit events that have been marked for cloning.
5595 * First allocate and initialize a context for the
5599 child_ctx = kzalloc(sizeof(struct perf_event_context),
5604 __perf_event_init_context(child_ctx, child);
5605 child->perf_event_ctxp = child_ctx;
5606 get_task_struct(child);
5609 ret = inherit_group(event, parent, parent_ctx,
5620 * Initialize the perf_event context in task_struct
5622 int perf_event_init_task(struct task_struct *child)
5624 struct perf_event_context *child_ctx, *parent_ctx;
5625 struct perf_event_context *cloned_ctx;
5626 struct perf_event *event;
5627 struct task_struct *parent = current;
5628 int inherited_all = 1;
5629 unsigned long flags;
5632 child->perf_event_ctxp = NULL;
5634 mutex_init(&child->perf_event_mutex);
5635 INIT_LIST_HEAD(&child->perf_event_list);
5637 if (likely(!parent->perf_event_ctxp))
5641 * If the parent's context is a clone, pin it so it won't get
5644 parent_ctx = perf_pin_task_context(parent);
5647 * No need to check if parent_ctx != NULL here; since we saw
5648 * it non-NULL earlier, the only reason for it to become NULL
5649 * is if we exit, and since we're currently in the middle of
5650 * a fork we can't be exiting at the same time.
5654 * Lock the parent list. No need to lock the child - not PID
5655 * hashed yet and not running, so nobody can access it.
5657 mutex_lock(&parent_ctx->mutex);
5660 * We dont have to disable NMIs - we are only looking at
5661 * the list, not manipulating it:
5663 list_for_each_entry(event, &parent_ctx->pinned_groups, group_entry) {
5664 ret = inherit_task_group(event, parent, parent_ctx, child,
5671 * We can't hold ctx->lock when iterating the ->flexible_group list due
5672 * to allocations, but we need to prevent rotation because
5673 * rotate_ctx() will change the list from interrupt context.
5675 raw_spin_lock_irqsave(&parent_ctx->lock, flags);
5676 parent_ctx->rotate_disable = 1;
5677 raw_spin_unlock_irqrestore(&parent_ctx->lock, flags);
5679 list_for_each_entry(event, &parent_ctx->flexible_groups, group_entry) {
5680 ret = inherit_task_group(event, parent, parent_ctx, child,
5686 raw_spin_lock_irqsave(&parent_ctx->lock, flags);
5687 parent_ctx->rotate_disable = 0;
5688 raw_spin_unlock_irqrestore(&parent_ctx->lock, flags);
5690 child_ctx = child->perf_event_ctxp;
5692 if (child_ctx && inherited_all) {
5694 * Mark the child context as a clone of the parent
5695 * context, or of whatever the parent is a clone of.
5696 * Note that if the parent is a clone, it could get
5697 * uncloned at any point, but that doesn't matter
5698 * because the list of events and the generation
5699 * count can't have changed since we took the mutex.
5701 cloned_ctx = rcu_dereference(parent_ctx->parent_ctx);
5703 child_ctx->parent_ctx = cloned_ctx;
5704 child_ctx->parent_gen = parent_ctx->parent_gen;
5706 child_ctx->parent_ctx = parent_ctx;
5707 child_ctx->parent_gen = parent_ctx->generation;
5709 get_ctx(child_ctx->parent_ctx);
5712 mutex_unlock(&parent_ctx->mutex);
5714 perf_unpin_context(parent_ctx);
5719 static void __init perf_event_init_all_cpus(void)
5722 struct perf_cpu_context *cpuctx;
5724 for_each_possible_cpu(cpu) {
5725 cpuctx = &per_cpu(perf_cpu_context, cpu);
5726 mutex_init(&cpuctx->hlist_mutex);
5727 __perf_event_init_context(&cpuctx->ctx, NULL);
5731 static void __cpuinit perf_event_init_cpu(int cpu)
5733 struct perf_cpu_context *cpuctx;
5735 cpuctx = &per_cpu(perf_cpu_context, cpu);
5737 spin_lock(&perf_resource_lock);
5738 cpuctx->max_pertask = perf_max_events - perf_reserved_percpu;
5739 spin_unlock(&perf_resource_lock);
5741 mutex_lock(&cpuctx->hlist_mutex);
5742 if (cpuctx->hlist_refcount > 0) {
5743 struct swevent_hlist *hlist;
5745 hlist = kzalloc(sizeof(*hlist), GFP_KERNEL);
5746 WARN_ON_ONCE(!hlist);
5747 rcu_assign_pointer(cpuctx->swevent_hlist, hlist);
5749 mutex_unlock(&cpuctx->hlist_mutex);
5752 #ifdef CONFIG_HOTPLUG_CPU
5753 static void __perf_event_exit_cpu(void *info)
5755 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
5756 struct perf_event_context *ctx = &cpuctx->ctx;
5757 struct perf_event *event, *tmp;
5759 list_for_each_entry_safe(event, tmp, &ctx->pinned_groups, group_entry)
5760 __perf_event_remove_from_context(event);
5761 list_for_each_entry_safe(event, tmp, &ctx->flexible_groups, group_entry)
5762 __perf_event_remove_from_context(event);
5764 static void perf_event_exit_cpu(int cpu)
5766 struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
5767 struct perf_event_context *ctx = &cpuctx->ctx;
5769 mutex_lock(&cpuctx->hlist_mutex);
5770 swevent_hlist_release(cpuctx);
5771 mutex_unlock(&cpuctx->hlist_mutex);
5773 mutex_lock(&ctx->mutex);
5774 smp_call_function_single(cpu, __perf_event_exit_cpu, NULL, 1);
5775 mutex_unlock(&ctx->mutex);
5778 static inline void perf_event_exit_cpu(int cpu) { }
5781 static int __cpuinit
5782 perf_cpu_notify(struct notifier_block *self, unsigned long action, void *hcpu)
5784 unsigned int cpu = (long)hcpu;
5786 switch (action & ~CPU_TASKS_FROZEN) {
5788 case CPU_UP_PREPARE:
5789 case CPU_DOWN_FAILED:
5790 perf_event_init_cpu(cpu);
5793 case CPU_UP_CANCELED:
5794 case CPU_DOWN_PREPARE:
5795 perf_event_exit_cpu(cpu);
5806 * This has to have a higher priority than migration_notifier in sched.c.
5808 static struct notifier_block __cpuinitdata perf_cpu_nb = {
5809 .notifier_call = perf_cpu_notify,
5813 void __init perf_event_init(void)
5815 perf_event_init_all_cpus();
5816 perf_cpu_notify(&perf_cpu_nb, (unsigned long)CPU_UP_PREPARE,
5817 (void *)(long)smp_processor_id());
5818 perf_cpu_notify(&perf_cpu_nb, (unsigned long)CPU_ONLINE,
5819 (void *)(long)smp_processor_id());
5820 register_cpu_notifier(&perf_cpu_nb);
5823 static ssize_t perf_show_reserve_percpu(struct sysdev_class *class,
5824 struct sysdev_class_attribute *attr,
5827 return sprintf(buf, "%d\n", perf_reserved_percpu);
5831 perf_set_reserve_percpu(struct sysdev_class *class,
5832 struct sysdev_class_attribute *attr,
5836 struct perf_cpu_context *cpuctx;
5840 err = strict_strtoul(buf, 10, &val);
5843 if (val > perf_max_events)
5846 spin_lock(&perf_resource_lock);
5847 perf_reserved_percpu = val;
5848 for_each_online_cpu(cpu) {
5849 cpuctx = &per_cpu(perf_cpu_context, cpu);
5850 raw_spin_lock_irq(&cpuctx->ctx.lock);
5851 mpt = min(perf_max_events - cpuctx->ctx.nr_events,
5852 perf_max_events - perf_reserved_percpu);
5853 cpuctx->max_pertask = mpt;
5854 raw_spin_unlock_irq(&cpuctx->ctx.lock);
5856 spin_unlock(&perf_resource_lock);
5861 static ssize_t perf_show_overcommit(struct sysdev_class *class,
5862 struct sysdev_class_attribute *attr,
5865 return sprintf(buf, "%d\n", perf_overcommit);
5869 perf_set_overcommit(struct sysdev_class *class,
5870 struct sysdev_class_attribute *attr,
5871 const char *buf, size_t count)
5876 err = strict_strtoul(buf, 10, &val);
5882 spin_lock(&perf_resource_lock);
5883 perf_overcommit = val;
5884 spin_unlock(&perf_resource_lock);
5889 static SYSDEV_CLASS_ATTR(
5892 perf_show_reserve_percpu,
5893 perf_set_reserve_percpu
5896 static SYSDEV_CLASS_ATTR(
5899 perf_show_overcommit,
5903 static struct attribute *perfclass_attrs[] = {
5904 &attr_reserve_percpu.attr,
5905 &attr_overcommit.attr,
5909 static struct attribute_group perfclass_attr_group = {
5910 .attrs = perfclass_attrs,
5911 .name = "perf_events",
5914 static int __init perf_event_sysfs_init(void)
5916 return sysfs_create_group(&cpu_sysdev_class.kset.kobj,
5917 &perfclass_attr_group);
5919 device_initcall(perf_event_sysfs_init);