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/sysfs.h>
19 #include <linux/dcache.h>
20 #include <linux/percpu.h>
21 #include <linux/ptrace.h>
22 #include <linux/vmstat.h>
23 #include <linux/vmalloc.h>
24 #include <linux/hardirq.h>
25 #include <linux/rculist.h>
26 #include <linux/uaccess.h>
27 #include <linux/syscalls.h>
28 #include <linux/anon_inodes.h>
29 #include <linux/kernel_stat.h>
30 #include <linux/perf_event.h>
31 #include <linux/ftrace_event.h>
32 #include <linux/hw_breakpoint.h>
34 #include <asm/irq_regs.h>
37 * Each CPU has a list of per CPU events:
39 static DEFINE_PER_CPU(struct perf_cpu_context, perf_cpu_context);
41 int perf_max_events __read_mostly = 1;
42 static int perf_reserved_percpu __read_mostly;
43 static int perf_overcommit __read_mostly = 1;
45 static atomic_t nr_events __read_mostly;
46 static atomic_t nr_mmap_events __read_mostly;
47 static atomic_t nr_comm_events __read_mostly;
48 static atomic_t nr_task_events __read_mostly;
51 * perf event paranoia level:
52 * -1 - not paranoid at all
53 * 0 - disallow raw tracepoint access for unpriv
54 * 1 - disallow cpu events for unpriv
55 * 2 - disallow kernel profiling for unpriv
57 int sysctl_perf_event_paranoid __read_mostly = 1;
59 int sysctl_perf_event_mlock __read_mostly = 512; /* 'free' kb per user */
62 * max perf event sample rate
64 int sysctl_perf_event_sample_rate __read_mostly = 100000;
66 static atomic64_t perf_event_id;
69 * Lock for (sysadmin-configurable) event reservations:
71 static DEFINE_SPINLOCK(perf_resource_lock);
74 * Architecture provided APIs - weak aliases:
76 extern __weak const struct pmu *hw_perf_event_init(struct perf_event *event)
81 void __weak hw_perf_disable(void) { barrier(); }
82 void __weak hw_perf_enable(void) { barrier(); }
85 hw_perf_group_sched_in(struct perf_event *group_leader,
86 struct perf_cpu_context *cpuctx,
87 struct perf_event_context *ctx)
92 void __weak perf_event_print_debug(void) { }
94 static DEFINE_PER_CPU(int, perf_disable_count);
96 void perf_disable(void)
98 if (!__get_cpu_var(perf_disable_count)++)
102 void perf_enable(void)
104 if (!--__get_cpu_var(perf_disable_count))
108 static void get_ctx(struct perf_event_context *ctx)
110 WARN_ON(!atomic_inc_not_zero(&ctx->refcount));
113 static void free_ctx(struct rcu_head *head)
115 struct perf_event_context *ctx;
117 ctx = container_of(head, struct perf_event_context, rcu_head);
121 static void put_ctx(struct perf_event_context *ctx)
123 if (atomic_dec_and_test(&ctx->refcount)) {
125 put_ctx(ctx->parent_ctx);
127 put_task_struct(ctx->task);
128 call_rcu(&ctx->rcu_head, free_ctx);
132 static void unclone_ctx(struct perf_event_context *ctx)
134 if (ctx->parent_ctx) {
135 put_ctx(ctx->parent_ctx);
136 ctx->parent_ctx = NULL;
141 * If we inherit events we want to return the parent event id
144 static u64 primary_event_id(struct perf_event *event)
149 id = event->parent->id;
155 * Get the perf_event_context for a task and lock it.
156 * This has to cope with with the fact that until it is locked,
157 * the context could get moved to another task.
159 static struct perf_event_context *
160 perf_lock_task_context(struct task_struct *task, unsigned long *flags)
162 struct perf_event_context *ctx;
166 ctx = rcu_dereference(task->perf_event_ctxp);
169 * If this context is a clone of another, it might
170 * get swapped for another underneath us by
171 * perf_event_task_sched_out, though the
172 * rcu_read_lock() protects us from any context
173 * getting freed. Lock the context and check if it
174 * got swapped before we could get the lock, and retry
175 * if so. If we locked the right context, then it
176 * can't get swapped on us any more.
178 raw_spin_lock_irqsave(&ctx->lock, *flags);
179 if (ctx != rcu_dereference(task->perf_event_ctxp)) {
180 raw_spin_unlock_irqrestore(&ctx->lock, *flags);
184 if (!atomic_inc_not_zero(&ctx->refcount)) {
185 raw_spin_unlock_irqrestore(&ctx->lock, *flags);
194 * Get the context for a task and increment its pin_count so it
195 * can't get swapped to another task. This also increments its
196 * reference count so that the context can't get freed.
198 static struct perf_event_context *perf_pin_task_context(struct task_struct *task)
200 struct perf_event_context *ctx;
203 ctx = perf_lock_task_context(task, &flags);
206 raw_spin_unlock_irqrestore(&ctx->lock, flags);
211 static void perf_unpin_context(struct perf_event_context *ctx)
215 raw_spin_lock_irqsave(&ctx->lock, flags);
217 raw_spin_unlock_irqrestore(&ctx->lock, flags);
221 static inline u64 perf_clock(void)
223 return cpu_clock(raw_smp_processor_id());
227 * Update the record of the current time in a context.
229 static void update_context_time(struct perf_event_context *ctx)
231 u64 now = perf_clock();
233 ctx->time += now - ctx->timestamp;
234 ctx->timestamp = now;
238 * Update the total_time_enabled and total_time_running fields for a event.
240 static void update_event_times(struct perf_event *event)
242 struct perf_event_context *ctx = event->ctx;
245 if (event->state < PERF_EVENT_STATE_INACTIVE ||
246 event->group_leader->state < PERF_EVENT_STATE_INACTIVE)
252 run_end = event->tstamp_stopped;
254 event->total_time_enabled = run_end - event->tstamp_enabled;
256 if (event->state == PERF_EVENT_STATE_INACTIVE)
257 run_end = event->tstamp_stopped;
261 event->total_time_running = run_end - event->tstamp_running;
264 static struct list_head *
265 ctx_group_list(struct perf_event *event, struct perf_event_context *ctx)
267 if (event->attr.pinned)
268 return &ctx->pinned_groups;
270 return &ctx->flexible_groups;
274 * Add a event from the lists for its context.
275 * Must be called with ctx->mutex and ctx->lock held.
278 list_add_event(struct perf_event *event, struct perf_event_context *ctx)
280 struct perf_event *group_leader = event->group_leader;
283 * Depending on whether it is a standalone or sibling event,
284 * add it straight to the context's event list, or to the group
285 * leader's sibling list:
287 if (group_leader == event) {
288 struct list_head *list;
290 if (is_software_event(event))
291 event->group_flags |= PERF_GROUP_SOFTWARE;
293 list = ctx_group_list(event, ctx);
294 list_add_tail(&event->group_entry, list);
296 if (group_leader->group_flags & PERF_GROUP_SOFTWARE &&
297 !is_software_event(event))
298 group_leader->group_flags &= ~PERF_GROUP_SOFTWARE;
300 list_add_tail(&event->group_entry, &group_leader->sibling_list);
301 group_leader->nr_siblings++;
304 list_add_rcu(&event->event_entry, &ctx->event_list);
306 if (event->attr.inherit_stat)
311 * Remove a event from the lists for its context.
312 * Must be called with ctx->mutex and ctx->lock held.
315 list_del_event(struct perf_event *event, struct perf_event_context *ctx)
317 struct perf_event *sibling, *tmp;
319 if (list_empty(&event->group_entry))
322 if (event->attr.inherit_stat)
325 list_del_init(&event->group_entry);
326 list_del_rcu(&event->event_entry);
328 if (event->group_leader != event)
329 event->group_leader->nr_siblings--;
331 update_event_times(event);
334 * If event was in error state, then keep it
335 * that way, otherwise bogus counts will be
336 * returned on read(). The only way to get out
337 * of error state is by explicit re-enabling
340 if (event->state > PERF_EVENT_STATE_OFF)
341 event->state = PERF_EVENT_STATE_OFF;
344 * If this was a group event with sibling events then
345 * upgrade the siblings to singleton events by adding them
346 * to the context list directly:
348 list_for_each_entry_safe(sibling, tmp, &event->sibling_list, group_entry) {
349 struct list_head *list;
351 list = ctx_group_list(event, ctx);
352 list_move_tail(&sibling->group_entry, list);
353 sibling->group_leader = sibling;
355 /* Inherit group flags from the previous leader */
356 sibling->group_flags = event->group_flags;
361 event_sched_out(struct perf_event *event,
362 struct perf_cpu_context *cpuctx,
363 struct perf_event_context *ctx)
365 if (event->state != PERF_EVENT_STATE_ACTIVE)
368 event->state = PERF_EVENT_STATE_INACTIVE;
369 if (event->pending_disable) {
370 event->pending_disable = 0;
371 event->state = PERF_EVENT_STATE_OFF;
373 event->tstamp_stopped = ctx->time;
374 event->pmu->disable(event);
377 if (!is_software_event(event))
378 cpuctx->active_oncpu--;
380 if (event->attr.exclusive || !cpuctx->active_oncpu)
381 cpuctx->exclusive = 0;
385 group_sched_out(struct perf_event *group_event,
386 struct perf_cpu_context *cpuctx,
387 struct perf_event_context *ctx)
389 struct perf_event *event;
391 if (group_event->state != PERF_EVENT_STATE_ACTIVE)
394 event_sched_out(group_event, cpuctx, ctx);
397 * Schedule out siblings (if any):
399 list_for_each_entry(event, &group_event->sibling_list, group_entry)
400 event_sched_out(event, cpuctx, ctx);
402 if (group_event->attr.exclusive)
403 cpuctx->exclusive = 0;
407 * Cross CPU call to remove a performance event
409 * We disable the event on the hardware level first. After that we
410 * remove it from the context list.
412 static void __perf_event_remove_from_context(void *info)
414 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
415 struct perf_event *event = info;
416 struct perf_event_context *ctx = event->ctx;
419 * If this is a task context, we need to check whether it is
420 * the current task context of this cpu. If not it has been
421 * scheduled out before the smp call arrived.
423 if (ctx->task && cpuctx->task_ctx != ctx)
426 raw_spin_lock(&ctx->lock);
428 * Protect the list operation against NMI by disabling the
429 * events on a global level.
433 event_sched_out(event, cpuctx, ctx);
435 list_del_event(event, ctx);
439 * Allow more per task events with respect to the
442 cpuctx->max_pertask =
443 min(perf_max_events - ctx->nr_events,
444 perf_max_events - perf_reserved_percpu);
448 raw_spin_unlock(&ctx->lock);
453 * Remove the event from a task's (or a CPU's) list of events.
455 * Must be called with ctx->mutex held.
457 * CPU events are removed with a smp call. For task events we only
458 * call when the task is on a CPU.
460 * If event->ctx is a cloned context, callers must make sure that
461 * every task struct that event->ctx->task could possibly point to
462 * remains valid. This is OK when called from perf_release since
463 * that only calls us on the top-level context, which can't be a clone.
464 * When called from perf_event_exit_task, it's OK because the
465 * context has been detached from its task.
467 static void perf_event_remove_from_context(struct perf_event *event)
469 struct perf_event_context *ctx = event->ctx;
470 struct task_struct *task = ctx->task;
474 * Per cpu events are removed via an smp call and
475 * the removal is always successful.
477 smp_call_function_single(event->cpu,
478 __perf_event_remove_from_context,
484 task_oncpu_function_call(task, __perf_event_remove_from_context,
487 raw_spin_lock_irq(&ctx->lock);
489 * If the context is active we need to retry the smp call.
491 if (ctx->nr_active && !list_empty(&event->group_entry)) {
492 raw_spin_unlock_irq(&ctx->lock);
497 * The lock prevents that this context is scheduled in so we
498 * can remove the event safely, if the call above did not
501 if (!list_empty(&event->group_entry))
502 list_del_event(event, ctx);
503 raw_spin_unlock_irq(&ctx->lock);
507 * Update total_time_enabled and total_time_running for all events in a group.
509 static void update_group_times(struct perf_event *leader)
511 struct perf_event *event;
513 update_event_times(leader);
514 list_for_each_entry(event, &leader->sibling_list, group_entry)
515 update_event_times(event);
519 * Cross CPU call to disable a performance event
521 static void __perf_event_disable(void *info)
523 struct perf_event *event = info;
524 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
525 struct perf_event_context *ctx = event->ctx;
528 * If this is a per-task event, need to check whether this
529 * event's task is the current task on this cpu.
531 if (ctx->task && cpuctx->task_ctx != ctx)
534 raw_spin_lock(&ctx->lock);
537 * If the event is on, turn it off.
538 * If it is in error state, leave it in error state.
540 if (event->state >= PERF_EVENT_STATE_INACTIVE) {
541 update_context_time(ctx);
542 update_group_times(event);
543 if (event == event->group_leader)
544 group_sched_out(event, cpuctx, ctx);
546 event_sched_out(event, cpuctx, ctx);
547 event->state = PERF_EVENT_STATE_OFF;
550 raw_spin_unlock(&ctx->lock);
556 * If event->ctx is a cloned context, callers must make sure that
557 * every task struct that event->ctx->task could possibly point to
558 * remains valid. This condition is satisifed when called through
559 * perf_event_for_each_child or perf_event_for_each because they
560 * hold the top-level event's child_mutex, so any descendant that
561 * goes to exit will block in sync_child_event.
562 * When called from perf_pending_event it's OK because event->ctx
563 * is the current context on this CPU and preemption is disabled,
564 * hence we can't get into perf_event_task_sched_out for this context.
566 void perf_event_disable(struct perf_event *event)
568 struct perf_event_context *ctx = event->ctx;
569 struct task_struct *task = ctx->task;
573 * Disable the event on the cpu that it's on
575 smp_call_function_single(event->cpu, __perf_event_disable,
581 task_oncpu_function_call(task, __perf_event_disable, event);
583 raw_spin_lock_irq(&ctx->lock);
585 * If the event is still active, we need to retry the cross-call.
587 if (event->state == PERF_EVENT_STATE_ACTIVE) {
588 raw_spin_unlock_irq(&ctx->lock);
593 * Since we have the lock this context can't be scheduled
594 * in, so we can change the state safely.
596 if (event->state == PERF_EVENT_STATE_INACTIVE) {
597 update_group_times(event);
598 event->state = PERF_EVENT_STATE_OFF;
601 raw_spin_unlock_irq(&ctx->lock);
605 event_sched_in(struct perf_event *event,
606 struct perf_cpu_context *cpuctx,
607 struct perf_event_context *ctx)
609 if (event->state <= PERF_EVENT_STATE_OFF)
612 event->state = PERF_EVENT_STATE_ACTIVE;
613 event->oncpu = smp_processor_id();
615 * The new state must be visible before we turn it on in the hardware:
619 if (event->pmu->enable(event)) {
620 event->state = PERF_EVENT_STATE_INACTIVE;
625 event->tstamp_running += ctx->time - event->tstamp_stopped;
627 if (!is_software_event(event))
628 cpuctx->active_oncpu++;
631 if (event->attr.exclusive)
632 cpuctx->exclusive = 1;
638 group_sched_in(struct perf_event *group_event,
639 struct perf_cpu_context *cpuctx,
640 struct perf_event_context *ctx)
642 struct perf_event *event, *partial_group;
645 if (group_event->state == PERF_EVENT_STATE_OFF)
648 ret = hw_perf_group_sched_in(group_event, cpuctx, ctx);
650 return ret < 0 ? ret : 0;
652 if (event_sched_in(group_event, cpuctx, ctx))
656 * Schedule in siblings as one group (if any):
658 list_for_each_entry(event, &group_event->sibling_list, group_entry) {
659 if (event_sched_in(event, cpuctx, ctx)) {
660 partial_group = event;
669 * Groups can be scheduled in as one unit only, so undo any
670 * partial group before returning:
672 list_for_each_entry(event, &group_event->sibling_list, group_entry) {
673 if (event == partial_group)
675 event_sched_out(event, cpuctx, ctx);
677 event_sched_out(group_event, cpuctx, ctx);
683 * Work out whether we can put this event group on the CPU now.
685 static int group_can_go_on(struct perf_event *event,
686 struct perf_cpu_context *cpuctx,
690 * Groups consisting entirely of software events can always go on.
692 if (event->group_flags & PERF_GROUP_SOFTWARE)
695 * If an exclusive group is already on, no other hardware
698 if (cpuctx->exclusive)
701 * If this group is exclusive and there are already
702 * events on the CPU, it can't go on.
704 if (event->attr.exclusive && cpuctx->active_oncpu)
707 * Otherwise, try to add it if all previous groups were able
713 static void add_event_to_ctx(struct perf_event *event,
714 struct perf_event_context *ctx)
716 list_add_event(event, ctx);
717 event->tstamp_enabled = ctx->time;
718 event->tstamp_running = ctx->time;
719 event->tstamp_stopped = ctx->time;
723 * Cross CPU call to install and enable a performance event
725 * Must be called with ctx->mutex held
727 static void __perf_install_in_context(void *info)
729 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
730 struct perf_event *event = info;
731 struct perf_event_context *ctx = event->ctx;
732 struct perf_event *leader = event->group_leader;
736 * If this is a task context, we need to check whether it is
737 * the current task context of this cpu. If not it has been
738 * scheduled out before the smp call arrived.
739 * Or possibly this is the right context but it isn't
740 * on this cpu because it had no events.
742 if (ctx->task && cpuctx->task_ctx != ctx) {
743 if (cpuctx->task_ctx || ctx->task != current)
745 cpuctx->task_ctx = ctx;
748 raw_spin_lock(&ctx->lock);
750 update_context_time(ctx);
753 * Protect the list operation against NMI by disabling the
754 * events on a global level. NOP for non NMI based events.
758 add_event_to_ctx(event, ctx);
760 if (event->cpu != -1 && event->cpu != smp_processor_id())
764 * Don't put the event on if it is disabled or if
765 * it is in a group and the group isn't on.
767 if (event->state != PERF_EVENT_STATE_INACTIVE ||
768 (leader != event && leader->state != PERF_EVENT_STATE_ACTIVE))
772 * An exclusive event can't go on if there are already active
773 * hardware events, and no hardware event can go on if there
774 * is already an exclusive event on.
776 if (!group_can_go_on(event, cpuctx, 1))
779 err = event_sched_in(event, cpuctx, ctx);
783 * This event couldn't go on. If it is in a group
784 * then we have to pull the whole group off.
785 * If the event group is pinned then put it in error state.
788 group_sched_out(leader, cpuctx, ctx);
789 if (leader->attr.pinned) {
790 update_group_times(leader);
791 leader->state = PERF_EVENT_STATE_ERROR;
795 if (!err && !ctx->task && cpuctx->max_pertask)
796 cpuctx->max_pertask--;
801 raw_spin_unlock(&ctx->lock);
805 * Attach a performance event to a context
807 * First we add the event to the list with the hardware enable bit
808 * in event->hw_config cleared.
810 * If the event is attached to a task which is on a CPU we use a smp
811 * call to enable it in the task context. The task might have been
812 * scheduled away, but we check this in the smp call again.
814 * Must be called with ctx->mutex held.
817 perf_install_in_context(struct perf_event_context *ctx,
818 struct perf_event *event,
821 struct task_struct *task = ctx->task;
825 * Per cpu events are installed via an smp call and
826 * the install is always successful.
828 smp_call_function_single(cpu, __perf_install_in_context,
834 task_oncpu_function_call(task, __perf_install_in_context,
837 raw_spin_lock_irq(&ctx->lock);
839 * we need to retry the smp call.
841 if (ctx->is_active && list_empty(&event->group_entry)) {
842 raw_spin_unlock_irq(&ctx->lock);
847 * The lock prevents that this context is scheduled in so we
848 * can add the event safely, if it the call above did not
851 if (list_empty(&event->group_entry))
852 add_event_to_ctx(event, ctx);
853 raw_spin_unlock_irq(&ctx->lock);
857 * Put a event into inactive state and update time fields.
858 * Enabling the leader of a group effectively enables all
859 * the group members that aren't explicitly disabled, so we
860 * have to update their ->tstamp_enabled also.
861 * Note: this works for group members as well as group leaders
862 * since the non-leader members' sibling_lists will be empty.
864 static void __perf_event_mark_enabled(struct perf_event *event,
865 struct perf_event_context *ctx)
867 struct perf_event *sub;
869 event->state = PERF_EVENT_STATE_INACTIVE;
870 event->tstamp_enabled = ctx->time - event->total_time_enabled;
871 list_for_each_entry(sub, &event->sibling_list, group_entry)
872 if (sub->state >= PERF_EVENT_STATE_INACTIVE)
873 sub->tstamp_enabled =
874 ctx->time - sub->total_time_enabled;
878 * Cross CPU call to enable a performance event
880 static void __perf_event_enable(void *info)
882 struct perf_event *event = info;
883 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
884 struct perf_event_context *ctx = event->ctx;
885 struct perf_event *leader = event->group_leader;
889 * If this is a per-task event, need to check whether this
890 * event's task is the current task on this cpu.
892 if (ctx->task && cpuctx->task_ctx != ctx) {
893 if (cpuctx->task_ctx || ctx->task != current)
895 cpuctx->task_ctx = ctx;
898 raw_spin_lock(&ctx->lock);
900 update_context_time(ctx);
902 if (event->state >= PERF_EVENT_STATE_INACTIVE)
904 __perf_event_mark_enabled(event, ctx);
906 if (event->cpu != -1 && event->cpu != smp_processor_id())
910 * If the event is in a group and isn't the group leader,
911 * then don't put it on unless the group is on.
913 if (leader != event && leader->state != PERF_EVENT_STATE_ACTIVE)
916 if (!group_can_go_on(event, cpuctx, 1)) {
921 err = group_sched_in(event, cpuctx, ctx);
923 err = event_sched_in(event, cpuctx, ctx);
929 * If this event can't go on and it's part of a
930 * group, then the whole group has to come off.
933 group_sched_out(leader, cpuctx, ctx);
934 if (leader->attr.pinned) {
935 update_group_times(leader);
936 leader->state = PERF_EVENT_STATE_ERROR;
941 raw_spin_unlock(&ctx->lock);
947 * If event->ctx is a cloned context, callers must make sure that
948 * every task struct that event->ctx->task could possibly point to
949 * remains valid. This condition is satisfied when called through
950 * perf_event_for_each_child or perf_event_for_each as described
951 * for perf_event_disable.
953 void perf_event_enable(struct perf_event *event)
955 struct perf_event_context *ctx = event->ctx;
956 struct task_struct *task = ctx->task;
960 * Enable the event on the cpu that it's on
962 smp_call_function_single(event->cpu, __perf_event_enable,
967 raw_spin_lock_irq(&ctx->lock);
968 if (event->state >= PERF_EVENT_STATE_INACTIVE)
972 * If the event is in error state, clear that first.
973 * That way, if we see the event in error state below, we
974 * know that it has gone back into error state, as distinct
975 * from the task having been scheduled away before the
976 * cross-call arrived.
978 if (event->state == PERF_EVENT_STATE_ERROR)
979 event->state = PERF_EVENT_STATE_OFF;
982 raw_spin_unlock_irq(&ctx->lock);
983 task_oncpu_function_call(task, __perf_event_enable, event);
985 raw_spin_lock_irq(&ctx->lock);
988 * If the context is active and the event is still off,
989 * we need to retry the cross-call.
991 if (ctx->is_active && event->state == PERF_EVENT_STATE_OFF)
995 * Since we have the lock this context can't be scheduled
996 * in, so we can change the state safely.
998 if (event->state == PERF_EVENT_STATE_OFF)
999 __perf_event_mark_enabled(event, ctx);
1002 raw_spin_unlock_irq(&ctx->lock);
1005 static int perf_event_refresh(struct perf_event *event, int refresh)
1008 * not supported on inherited events
1010 if (event->attr.inherit)
1013 atomic_add(refresh, &event->event_limit);
1014 perf_event_enable(event);
1020 EVENT_FLEXIBLE = 0x1,
1022 EVENT_ALL = EVENT_FLEXIBLE | EVENT_PINNED,
1025 static void ctx_sched_out(struct perf_event_context *ctx,
1026 struct perf_cpu_context *cpuctx,
1027 enum event_type_t event_type)
1029 struct perf_event *event;
1031 raw_spin_lock(&ctx->lock);
1033 if (likely(!ctx->nr_events))
1035 update_context_time(ctx);
1038 if (!ctx->nr_active)
1041 if (event_type & EVENT_PINNED)
1042 list_for_each_entry(event, &ctx->pinned_groups, group_entry)
1043 group_sched_out(event, cpuctx, ctx);
1045 if (event_type & EVENT_FLEXIBLE)
1046 list_for_each_entry(event, &ctx->flexible_groups, group_entry)
1047 group_sched_out(event, cpuctx, ctx);
1052 raw_spin_unlock(&ctx->lock);
1056 * Test whether two contexts are equivalent, i.e. whether they
1057 * have both been cloned from the same version of the same context
1058 * and they both have the same number of enabled events.
1059 * If the number of enabled events is the same, then the set
1060 * of enabled events should be the same, because these are both
1061 * inherited contexts, therefore we can't access individual events
1062 * in them directly with an fd; we can only enable/disable all
1063 * events via prctl, or enable/disable all events in a family
1064 * via ioctl, which will have the same effect on both contexts.
1066 static int context_equiv(struct perf_event_context *ctx1,
1067 struct perf_event_context *ctx2)
1069 return ctx1->parent_ctx && ctx1->parent_ctx == ctx2->parent_ctx
1070 && ctx1->parent_gen == ctx2->parent_gen
1071 && !ctx1->pin_count && !ctx2->pin_count;
1074 static void __perf_event_sync_stat(struct perf_event *event,
1075 struct perf_event *next_event)
1079 if (!event->attr.inherit_stat)
1083 * Update the event value, we cannot use perf_event_read()
1084 * because we're in the middle of a context switch and have IRQs
1085 * disabled, which upsets smp_call_function_single(), however
1086 * we know the event must be on the current CPU, therefore we
1087 * don't need to use it.
1089 switch (event->state) {
1090 case PERF_EVENT_STATE_ACTIVE:
1091 event->pmu->read(event);
1094 case PERF_EVENT_STATE_INACTIVE:
1095 update_event_times(event);
1103 * In order to keep per-task stats reliable we need to flip the event
1104 * values when we flip the contexts.
1106 value = atomic64_read(&next_event->count);
1107 value = atomic64_xchg(&event->count, value);
1108 atomic64_set(&next_event->count, value);
1110 swap(event->total_time_enabled, next_event->total_time_enabled);
1111 swap(event->total_time_running, next_event->total_time_running);
1114 * Since we swizzled the values, update the user visible data too.
1116 perf_event_update_userpage(event);
1117 perf_event_update_userpage(next_event);
1120 #define list_next_entry(pos, member) \
1121 list_entry(pos->member.next, typeof(*pos), member)
1123 static void perf_event_sync_stat(struct perf_event_context *ctx,
1124 struct perf_event_context *next_ctx)
1126 struct perf_event *event, *next_event;
1131 update_context_time(ctx);
1133 event = list_first_entry(&ctx->event_list,
1134 struct perf_event, event_entry);
1136 next_event = list_first_entry(&next_ctx->event_list,
1137 struct perf_event, event_entry);
1139 while (&event->event_entry != &ctx->event_list &&
1140 &next_event->event_entry != &next_ctx->event_list) {
1142 __perf_event_sync_stat(event, next_event);
1144 event = list_next_entry(event, event_entry);
1145 next_event = list_next_entry(next_event, event_entry);
1150 * Called from scheduler to remove the events of the current task,
1151 * with interrupts disabled.
1153 * We stop each event and update the event value in event->count.
1155 * This does not protect us against NMI, but disable()
1156 * sets the disabled bit in the control field of event _before_
1157 * accessing the event control register. If a NMI hits, then it will
1158 * not restart the event.
1160 void perf_event_task_sched_out(struct task_struct *task,
1161 struct task_struct *next)
1163 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
1164 struct perf_event_context *ctx = task->perf_event_ctxp;
1165 struct perf_event_context *next_ctx;
1166 struct perf_event_context *parent;
1167 struct pt_regs *regs;
1170 regs = task_pt_regs(task);
1171 perf_sw_event(PERF_COUNT_SW_CONTEXT_SWITCHES, 1, 1, regs, 0);
1173 if (likely(!ctx || !cpuctx->task_ctx))
1177 parent = rcu_dereference(ctx->parent_ctx);
1178 next_ctx = next->perf_event_ctxp;
1179 if (parent && next_ctx &&
1180 rcu_dereference(next_ctx->parent_ctx) == parent) {
1182 * Looks like the two contexts are clones, so we might be
1183 * able to optimize the context switch. We lock both
1184 * contexts and check that they are clones under the
1185 * lock (including re-checking that neither has been
1186 * uncloned in the meantime). It doesn't matter which
1187 * order we take the locks because no other cpu could
1188 * be trying to lock both of these tasks.
1190 raw_spin_lock(&ctx->lock);
1191 raw_spin_lock_nested(&next_ctx->lock, SINGLE_DEPTH_NESTING);
1192 if (context_equiv(ctx, next_ctx)) {
1194 * XXX do we need a memory barrier of sorts
1195 * wrt to rcu_dereference() of perf_event_ctxp
1197 task->perf_event_ctxp = next_ctx;
1198 next->perf_event_ctxp = ctx;
1200 next_ctx->task = task;
1203 perf_event_sync_stat(ctx, next_ctx);
1205 raw_spin_unlock(&next_ctx->lock);
1206 raw_spin_unlock(&ctx->lock);
1211 ctx_sched_out(ctx, cpuctx, EVENT_ALL);
1212 cpuctx->task_ctx = NULL;
1216 static void task_ctx_sched_out(struct perf_event_context *ctx,
1217 enum event_type_t event_type)
1219 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
1221 if (!cpuctx->task_ctx)
1224 if (WARN_ON_ONCE(ctx != cpuctx->task_ctx))
1227 ctx_sched_out(ctx, cpuctx, event_type);
1228 cpuctx->task_ctx = NULL;
1232 * Called with IRQs disabled
1234 static void __perf_event_task_sched_out(struct perf_event_context *ctx)
1236 task_ctx_sched_out(ctx, EVENT_ALL);
1240 * Called with IRQs disabled
1242 static void cpu_ctx_sched_out(struct perf_cpu_context *cpuctx,
1243 enum event_type_t event_type)
1245 ctx_sched_out(&cpuctx->ctx, cpuctx, event_type);
1249 ctx_pinned_sched_in(struct perf_event_context *ctx,
1250 struct perf_cpu_context *cpuctx)
1252 struct perf_event *event;
1254 list_for_each_entry(event, &ctx->pinned_groups, group_entry) {
1255 if (event->state <= PERF_EVENT_STATE_OFF)
1257 if (event->cpu != -1 && event->cpu != smp_processor_id())
1260 if (group_can_go_on(event, cpuctx, 1))
1261 group_sched_in(event, cpuctx, ctx);
1264 * If this pinned group hasn't been scheduled,
1265 * put it in error state.
1267 if (event->state == PERF_EVENT_STATE_INACTIVE) {
1268 update_group_times(event);
1269 event->state = PERF_EVENT_STATE_ERROR;
1275 ctx_flexible_sched_in(struct perf_event_context *ctx,
1276 struct perf_cpu_context *cpuctx)
1278 struct perf_event *event;
1281 list_for_each_entry(event, &ctx->flexible_groups, group_entry) {
1282 /* Ignore events in OFF or ERROR state */
1283 if (event->state <= PERF_EVENT_STATE_OFF)
1286 * Listen to the 'cpu' scheduling filter constraint
1289 if (event->cpu != -1 && event->cpu != smp_processor_id())
1292 if (group_can_go_on(event, cpuctx, can_add_hw))
1293 if (group_sched_in(event, cpuctx, ctx))
1299 ctx_sched_in(struct perf_event_context *ctx,
1300 struct perf_cpu_context *cpuctx,
1301 enum event_type_t event_type)
1303 raw_spin_lock(&ctx->lock);
1305 if (likely(!ctx->nr_events))
1308 ctx->timestamp = perf_clock();
1313 * First go through the list and put on any pinned groups
1314 * in order to give them the best chance of going on.
1316 if (event_type & EVENT_PINNED)
1317 ctx_pinned_sched_in(ctx, cpuctx);
1319 /* Then walk through the lower prio flexible groups */
1320 if (event_type & EVENT_FLEXIBLE)
1321 ctx_flexible_sched_in(ctx, cpuctx);
1325 raw_spin_unlock(&ctx->lock);
1328 static void cpu_ctx_sched_in(struct perf_cpu_context *cpuctx,
1329 enum event_type_t event_type)
1331 struct perf_event_context *ctx = &cpuctx->ctx;
1333 ctx_sched_in(ctx, cpuctx, event_type);
1336 static void task_ctx_sched_in(struct task_struct *task,
1337 enum event_type_t event_type)
1339 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
1340 struct perf_event_context *ctx = task->perf_event_ctxp;
1344 if (cpuctx->task_ctx == ctx)
1346 ctx_sched_in(ctx, cpuctx, event_type);
1347 cpuctx->task_ctx = ctx;
1350 * Called from scheduler to add the events of the current task
1351 * with interrupts disabled.
1353 * We restore the event value and then enable it.
1355 * This does not protect us against NMI, but enable()
1356 * sets the enabled bit in the control field of event _before_
1357 * accessing the event control register. If a NMI hits, then it will
1358 * keep the event running.
1360 void perf_event_task_sched_in(struct task_struct *task)
1362 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
1363 struct perf_event_context *ctx = task->perf_event_ctxp;
1368 if (cpuctx->task_ctx == ctx)
1372 * We want to keep the following priority order:
1373 * cpu pinned (that don't need to move), task pinned,
1374 * cpu flexible, task flexible.
1376 cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE);
1378 ctx_sched_in(ctx, cpuctx, EVENT_PINNED);
1379 cpu_ctx_sched_in(cpuctx, EVENT_FLEXIBLE);
1380 ctx_sched_in(ctx, cpuctx, EVENT_FLEXIBLE);
1382 cpuctx->task_ctx = ctx;
1385 #define MAX_INTERRUPTS (~0ULL)
1387 static void perf_log_throttle(struct perf_event *event, int enable);
1389 static u64 perf_calculate_period(struct perf_event *event, u64 nsec, u64 count)
1391 u64 frequency = event->attr.sample_freq;
1392 u64 sec = NSEC_PER_SEC;
1393 u64 divisor, dividend;
1395 int count_fls, nsec_fls, frequency_fls, sec_fls;
1397 count_fls = fls64(count);
1398 nsec_fls = fls64(nsec);
1399 frequency_fls = fls64(frequency);
1403 * We got @count in @nsec, with a target of sample_freq HZ
1404 * the target period becomes:
1407 * period = -------------------
1408 * @nsec * sample_freq
1413 * Reduce accuracy by one bit such that @a and @b converge
1414 * to a similar magnitude.
1416 #define REDUCE_FLS(a, b) \
1418 if (a##_fls > b##_fls) { \
1428 * Reduce accuracy until either term fits in a u64, then proceed with
1429 * the other, so that finally we can do a u64/u64 division.
1431 while (count_fls + sec_fls > 64 && nsec_fls + frequency_fls > 64) {
1432 REDUCE_FLS(nsec, frequency);
1433 REDUCE_FLS(sec, count);
1436 if (count_fls + sec_fls > 64) {
1437 divisor = nsec * frequency;
1439 while (count_fls + sec_fls > 64) {
1440 REDUCE_FLS(count, sec);
1444 dividend = count * sec;
1446 dividend = count * sec;
1448 while (nsec_fls + frequency_fls > 64) {
1449 REDUCE_FLS(nsec, frequency);
1453 divisor = nsec * frequency;
1456 return div64_u64(dividend, divisor);
1459 static void perf_event_stop(struct perf_event *event)
1461 if (!event->pmu->stop)
1462 return event->pmu->disable(event);
1464 return event->pmu->stop(event);
1467 static int perf_event_start(struct perf_event *event)
1469 if (!event->pmu->start)
1470 return event->pmu->enable(event);
1472 return event->pmu->start(event);
1475 static void perf_adjust_period(struct perf_event *event, u64 nsec, u64 count)
1477 struct hw_perf_event *hwc = &event->hw;
1478 u64 period, sample_period;
1481 period = perf_calculate_period(event, nsec, count);
1483 delta = (s64)(period - hwc->sample_period);
1484 delta = (delta + 7) / 8; /* low pass filter */
1486 sample_period = hwc->sample_period + delta;
1491 hwc->sample_period = sample_period;
1493 if (atomic64_read(&hwc->period_left) > 8*sample_period) {
1495 perf_event_stop(event);
1496 atomic64_set(&hwc->period_left, 0);
1497 perf_event_start(event);
1502 static void perf_ctx_adjust_freq(struct perf_event_context *ctx)
1504 struct perf_event *event;
1505 struct hw_perf_event *hwc;
1506 u64 interrupts, now;
1509 raw_spin_lock(&ctx->lock);
1510 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
1511 if (event->state != PERF_EVENT_STATE_ACTIVE)
1514 if (event->cpu != -1 && event->cpu != smp_processor_id())
1519 interrupts = hwc->interrupts;
1520 hwc->interrupts = 0;
1523 * unthrottle events on the tick
1525 if (interrupts == MAX_INTERRUPTS) {
1526 perf_log_throttle(event, 1);
1528 event->pmu->unthrottle(event);
1532 if (!event->attr.freq || !event->attr.sample_freq)
1536 event->pmu->read(event);
1537 now = atomic64_read(&event->count);
1538 delta = now - hwc->freq_count_stamp;
1539 hwc->freq_count_stamp = now;
1542 perf_adjust_period(event, TICK_NSEC, delta);
1545 raw_spin_unlock(&ctx->lock);
1549 * Round-robin a context's events:
1551 static void rotate_ctx(struct perf_event_context *ctx)
1553 raw_spin_lock(&ctx->lock);
1555 /* Rotate the first entry last of non-pinned groups */
1556 list_rotate_left(&ctx->flexible_groups);
1558 raw_spin_unlock(&ctx->lock);
1561 void perf_event_task_tick(struct task_struct *curr)
1563 struct perf_cpu_context *cpuctx;
1564 struct perf_event_context *ctx;
1567 if (!atomic_read(&nr_events))
1570 cpuctx = &__get_cpu_var(perf_cpu_context);
1571 if (cpuctx->ctx.nr_events &&
1572 cpuctx->ctx.nr_events != cpuctx->ctx.nr_active)
1575 ctx = curr->perf_event_ctxp;
1576 if (ctx && ctx->nr_events && ctx->nr_events != ctx->nr_active)
1579 perf_ctx_adjust_freq(&cpuctx->ctx);
1581 perf_ctx_adjust_freq(ctx);
1587 cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE);
1589 task_ctx_sched_out(ctx, EVENT_FLEXIBLE);
1591 rotate_ctx(&cpuctx->ctx);
1595 cpu_ctx_sched_in(cpuctx, EVENT_FLEXIBLE);
1597 task_ctx_sched_in(curr, EVENT_FLEXIBLE);
1601 static int event_enable_on_exec(struct perf_event *event,
1602 struct perf_event_context *ctx)
1604 if (!event->attr.enable_on_exec)
1607 event->attr.enable_on_exec = 0;
1608 if (event->state >= PERF_EVENT_STATE_INACTIVE)
1611 __perf_event_mark_enabled(event, ctx);
1617 * Enable all of a task's events that have been marked enable-on-exec.
1618 * This expects task == current.
1620 static void perf_event_enable_on_exec(struct task_struct *task)
1622 struct perf_event_context *ctx;
1623 struct perf_event *event;
1624 unsigned long flags;
1628 local_irq_save(flags);
1629 ctx = task->perf_event_ctxp;
1630 if (!ctx || !ctx->nr_events)
1633 __perf_event_task_sched_out(ctx);
1635 raw_spin_lock(&ctx->lock);
1637 list_for_each_entry(event, &ctx->pinned_groups, group_entry) {
1638 ret = event_enable_on_exec(event, ctx);
1643 list_for_each_entry(event, &ctx->flexible_groups, group_entry) {
1644 ret = event_enable_on_exec(event, ctx);
1650 * Unclone this context if we enabled any event.
1655 raw_spin_unlock(&ctx->lock);
1657 perf_event_task_sched_in(task);
1659 local_irq_restore(flags);
1663 * Cross CPU call to read the hardware event
1665 static void __perf_event_read(void *info)
1667 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
1668 struct perf_event *event = info;
1669 struct perf_event_context *ctx = event->ctx;
1672 * If this is a task context, we need to check whether it is
1673 * the current task context of this cpu. If not it has been
1674 * scheduled out before the smp call arrived. In that case
1675 * event->count would have been updated to a recent sample
1676 * when the event was scheduled out.
1678 if (ctx->task && cpuctx->task_ctx != ctx)
1681 raw_spin_lock(&ctx->lock);
1682 update_context_time(ctx);
1683 update_event_times(event);
1684 raw_spin_unlock(&ctx->lock);
1686 event->pmu->read(event);
1689 static u64 perf_event_read(struct perf_event *event)
1692 * If event is enabled and currently active on a CPU, update the
1693 * value in the event structure:
1695 if (event->state == PERF_EVENT_STATE_ACTIVE) {
1696 smp_call_function_single(event->oncpu,
1697 __perf_event_read, event, 1);
1698 } else if (event->state == PERF_EVENT_STATE_INACTIVE) {
1699 struct perf_event_context *ctx = event->ctx;
1700 unsigned long flags;
1702 raw_spin_lock_irqsave(&ctx->lock, flags);
1703 update_context_time(ctx);
1704 update_event_times(event);
1705 raw_spin_unlock_irqrestore(&ctx->lock, flags);
1708 return atomic64_read(&event->count);
1712 * Initialize the perf_event context in a task_struct:
1715 __perf_event_init_context(struct perf_event_context *ctx,
1716 struct task_struct *task)
1718 raw_spin_lock_init(&ctx->lock);
1719 mutex_init(&ctx->mutex);
1720 INIT_LIST_HEAD(&ctx->pinned_groups);
1721 INIT_LIST_HEAD(&ctx->flexible_groups);
1722 INIT_LIST_HEAD(&ctx->event_list);
1723 atomic_set(&ctx->refcount, 1);
1727 static struct perf_event_context *find_get_context(pid_t pid, int cpu)
1729 struct perf_event_context *ctx;
1730 struct perf_cpu_context *cpuctx;
1731 struct task_struct *task;
1732 unsigned long flags;
1735 if (pid == -1 && cpu != -1) {
1736 /* Must be root to operate on a CPU event: */
1737 if (perf_paranoid_cpu() && !capable(CAP_SYS_ADMIN))
1738 return ERR_PTR(-EACCES);
1740 if (cpu < 0 || cpu >= nr_cpumask_bits)
1741 return ERR_PTR(-EINVAL);
1744 * We could be clever and allow to attach a event to an
1745 * offline CPU and activate it when the CPU comes up, but
1748 if (!cpu_online(cpu))
1749 return ERR_PTR(-ENODEV);
1751 cpuctx = &per_cpu(perf_cpu_context, cpu);
1762 task = find_task_by_vpid(pid);
1764 get_task_struct(task);
1768 return ERR_PTR(-ESRCH);
1771 * Can't attach events to a dying task.
1774 if (task->flags & PF_EXITING)
1777 /* Reuse ptrace permission checks for now. */
1779 if (!ptrace_may_access(task, PTRACE_MODE_READ))
1783 ctx = perf_lock_task_context(task, &flags);
1786 raw_spin_unlock_irqrestore(&ctx->lock, flags);
1790 ctx = kzalloc(sizeof(struct perf_event_context), GFP_KERNEL);
1794 __perf_event_init_context(ctx, task);
1796 if (cmpxchg(&task->perf_event_ctxp, NULL, ctx)) {
1798 * We raced with some other task; use
1799 * the context they set.
1804 get_task_struct(task);
1807 put_task_struct(task);
1811 put_task_struct(task);
1812 return ERR_PTR(err);
1815 static void perf_event_free_filter(struct perf_event *event);
1817 static void free_event_rcu(struct rcu_head *head)
1819 struct perf_event *event;
1821 event = container_of(head, struct perf_event, rcu_head);
1823 put_pid_ns(event->ns);
1824 perf_event_free_filter(event);
1828 static void perf_pending_sync(struct perf_event *event);
1830 static void free_event(struct perf_event *event)
1832 perf_pending_sync(event);
1834 if (!event->parent) {
1835 atomic_dec(&nr_events);
1836 if (event->attr.mmap)
1837 atomic_dec(&nr_mmap_events);
1838 if (event->attr.comm)
1839 atomic_dec(&nr_comm_events);
1840 if (event->attr.task)
1841 atomic_dec(&nr_task_events);
1844 if (event->output) {
1845 fput(event->output->filp);
1846 event->output = NULL;
1850 event->destroy(event);
1852 put_ctx(event->ctx);
1853 call_rcu(&event->rcu_head, free_event_rcu);
1856 int perf_event_release_kernel(struct perf_event *event)
1858 struct perf_event_context *ctx = event->ctx;
1860 WARN_ON_ONCE(ctx->parent_ctx);
1861 mutex_lock(&ctx->mutex);
1862 perf_event_remove_from_context(event);
1863 mutex_unlock(&ctx->mutex);
1865 mutex_lock(&event->owner->perf_event_mutex);
1866 list_del_init(&event->owner_entry);
1867 mutex_unlock(&event->owner->perf_event_mutex);
1868 put_task_struct(event->owner);
1874 EXPORT_SYMBOL_GPL(perf_event_release_kernel);
1877 * Called when the last reference to the file is gone.
1879 static int perf_release(struct inode *inode, struct file *file)
1881 struct perf_event *event = file->private_data;
1883 file->private_data = NULL;
1885 return perf_event_release_kernel(event);
1888 static int perf_event_read_size(struct perf_event *event)
1890 int entry = sizeof(u64); /* value */
1894 if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1895 size += sizeof(u64);
1897 if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1898 size += sizeof(u64);
1900 if (event->attr.read_format & PERF_FORMAT_ID)
1901 entry += sizeof(u64);
1903 if (event->attr.read_format & PERF_FORMAT_GROUP) {
1904 nr += event->group_leader->nr_siblings;
1905 size += sizeof(u64);
1913 u64 perf_event_read_value(struct perf_event *event, u64 *enabled, u64 *running)
1915 struct perf_event *child;
1921 mutex_lock(&event->child_mutex);
1922 total += perf_event_read(event);
1923 *enabled += event->total_time_enabled +
1924 atomic64_read(&event->child_total_time_enabled);
1925 *running += event->total_time_running +
1926 atomic64_read(&event->child_total_time_running);
1928 list_for_each_entry(child, &event->child_list, child_list) {
1929 total += perf_event_read(child);
1930 *enabled += child->total_time_enabled;
1931 *running += child->total_time_running;
1933 mutex_unlock(&event->child_mutex);
1937 EXPORT_SYMBOL_GPL(perf_event_read_value);
1939 static int perf_event_read_group(struct perf_event *event,
1940 u64 read_format, char __user *buf)
1942 struct perf_event *leader = event->group_leader, *sub;
1943 int n = 0, size = 0, ret = -EFAULT;
1944 struct perf_event_context *ctx = leader->ctx;
1946 u64 count, enabled, running;
1948 mutex_lock(&ctx->mutex);
1949 count = perf_event_read_value(leader, &enabled, &running);
1951 values[n++] = 1 + leader->nr_siblings;
1952 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1953 values[n++] = enabled;
1954 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1955 values[n++] = running;
1956 values[n++] = count;
1957 if (read_format & PERF_FORMAT_ID)
1958 values[n++] = primary_event_id(leader);
1960 size = n * sizeof(u64);
1962 if (copy_to_user(buf, values, size))
1967 list_for_each_entry(sub, &leader->sibling_list, group_entry) {
1970 values[n++] = perf_event_read_value(sub, &enabled, &running);
1971 if (read_format & PERF_FORMAT_ID)
1972 values[n++] = primary_event_id(sub);
1974 size = n * sizeof(u64);
1976 if (copy_to_user(buf + ret, values, size)) {
1984 mutex_unlock(&ctx->mutex);
1989 static int perf_event_read_one(struct perf_event *event,
1990 u64 read_format, char __user *buf)
1992 u64 enabled, running;
1996 values[n++] = perf_event_read_value(event, &enabled, &running);
1997 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1998 values[n++] = enabled;
1999 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
2000 values[n++] = running;
2001 if (read_format & PERF_FORMAT_ID)
2002 values[n++] = primary_event_id(event);
2004 if (copy_to_user(buf, values, n * sizeof(u64)))
2007 return n * sizeof(u64);
2011 * Read the performance event - simple non blocking version for now
2014 perf_read_hw(struct perf_event *event, char __user *buf, size_t count)
2016 u64 read_format = event->attr.read_format;
2020 * Return end-of-file for a read on a event that is in
2021 * error state (i.e. because it was pinned but it couldn't be
2022 * scheduled on to the CPU at some point).
2024 if (event->state == PERF_EVENT_STATE_ERROR)
2027 if (count < perf_event_read_size(event))
2030 WARN_ON_ONCE(event->ctx->parent_ctx);
2031 if (read_format & PERF_FORMAT_GROUP)
2032 ret = perf_event_read_group(event, read_format, buf);
2034 ret = perf_event_read_one(event, read_format, buf);
2040 perf_read(struct file *file, char __user *buf, size_t count, loff_t *ppos)
2042 struct perf_event *event = file->private_data;
2044 return perf_read_hw(event, buf, count);
2047 static unsigned int perf_poll(struct file *file, poll_table *wait)
2049 struct perf_event *event = file->private_data;
2050 struct perf_mmap_data *data;
2051 unsigned int events = POLL_HUP;
2054 data = rcu_dereference(event->data);
2056 events = atomic_xchg(&data->poll, 0);
2059 poll_wait(file, &event->waitq, wait);
2064 static void perf_event_reset(struct perf_event *event)
2066 (void)perf_event_read(event);
2067 atomic64_set(&event->count, 0);
2068 perf_event_update_userpage(event);
2072 * Holding the top-level event's child_mutex means that any
2073 * descendant process that has inherited this event will block
2074 * in sync_child_event if it goes to exit, thus satisfying the
2075 * task existence requirements of perf_event_enable/disable.
2077 static void perf_event_for_each_child(struct perf_event *event,
2078 void (*func)(struct perf_event *))
2080 struct perf_event *child;
2082 WARN_ON_ONCE(event->ctx->parent_ctx);
2083 mutex_lock(&event->child_mutex);
2085 list_for_each_entry(child, &event->child_list, child_list)
2087 mutex_unlock(&event->child_mutex);
2090 static void perf_event_for_each(struct perf_event *event,
2091 void (*func)(struct perf_event *))
2093 struct perf_event_context *ctx = event->ctx;
2094 struct perf_event *sibling;
2096 WARN_ON_ONCE(ctx->parent_ctx);
2097 mutex_lock(&ctx->mutex);
2098 event = event->group_leader;
2100 perf_event_for_each_child(event, func);
2102 list_for_each_entry(sibling, &event->sibling_list, group_entry)
2103 perf_event_for_each_child(event, func);
2104 mutex_unlock(&ctx->mutex);
2107 static int perf_event_period(struct perf_event *event, u64 __user *arg)
2109 struct perf_event_context *ctx = event->ctx;
2114 if (!event->attr.sample_period)
2117 size = copy_from_user(&value, arg, sizeof(value));
2118 if (size != sizeof(value))
2124 raw_spin_lock_irq(&ctx->lock);
2125 if (event->attr.freq) {
2126 if (value > sysctl_perf_event_sample_rate) {
2131 event->attr.sample_freq = value;
2133 event->attr.sample_period = value;
2134 event->hw.sample_period = value;
2137 raw_spin_unlock_irq(&ctx->lock);
2142 static int perf_event_set_output(struct perf_event *event, int output_fd);
2143 static int perf_event_set_filter(struct perf_event *event, void __user *arg);
2145 static long perf_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
2147 struct perf_event *event = file->private_data;
2148 void (*func)(struct perf_event *);
2152 case PERF_EVENT_IOC_ENABLE:
2153 func = perf_event_enable;
2155 case PERF_EVENT_IOC_DISABLE:
2156 func = perf_event_disable;
2158 case PERF_EVENT_IOC_RESET:
2159 func = perf_event_reset;
2162 case PERF_EVENT_IOC_REFRESH:
2163 return perf_event_refresh(event, arg);
2165 case PERF_EVENT_IOC_PERIOD:
2166 return perf_event_period(event, (u64 __user *)arg);
2168 case PERF_EVENT_IOC_SET_OUTPUT:
2169 return perf_event_set_output(event, arg);
2171 case PERF_EVENT_IOC_SET_FILTER:
2172 return perf_event_set_filter(event, (void __user *)arg);
2178 if (flags & PERF_IOC_FLAG_GROUP)
2179 perf_event_for_each(event, func);
2181 perf_event_for_each_child(event, func);
2186 int perf_event_task_enable(void)
2188 struct perf_event *event;
2190 mutex_lock(¤t->perf_event_mutex);
2191 list_for_each_entry(event, ¤t->perf_event_list, owner_entry)
2192 perf_event_for_each_child(event, perf_event_enable);
2193 mutex_unlock(¤t->perf_event_mutex);
2198 int perf_event_task_disable(void)
2200 struct perf_event *event;
2202 mutex_lock(¤t->perf_event_mutex);
2203 list_for_each_entry(event, ¤t->perf_event_list, owner_entry)
2204 perf_event_for_each_child(event, perf_event_disable);
2205 mutex_unlock(¤t->perf_event_mutex);
2210 #ifndef PERF_EVENT_INDEX_OFFSET
2211 # define PERF_EVENT_INDEX_OFFSET 0
2214 static int perf_event_index(struct perf_event *event)
2216 if (event->state != PERF_EVENT_STATE_ACTIVE)
2219 return event->hw.idx + 1 - PERF_EVENT_INDEX_OFFSET;
2223 * Callers need to ensure there can be no nesting of this function, otherwise
2224 * the seqlock logic goes bad. We can not serialize this because the arch
2225 * code calls this from NMI context.
2227 void perf_event_update_userpage(struct perf_event *event)
2229 struct perf_event_mmap_page *userpg;
2230 struct perf_mmap_data *data;
2233 data = rcu_dereference(event->data);
2237 userpg = data->user_page;
2240 * Disable preemption so as to not let the corresponding user-space
2241 * spin too long if we get preempted.
2246 userpg->index = perf_event_index(event);
2247 userpg->offset = atomic64_read(&event->count);
2248 if (event->state == PERF_EVENT_STATE_ACTIVE)
2249 userpg->offset -= atomic64_read(&event->hw.prev_count);
2251 userpg->time_enabled = event->total_time_enabled +
2252 atomic64_read(&event->child_total_time_enabled);
2254 userpg->time_running = event->total_time_running +
2255 atomic64_read(&event->child_total_time_running);
2264 static unsigned long perf_data_size(struct perf_mmap_data *data)
2266 return data->nr_pages << (PAGE_SHIFT + data->data_order);
2269 #ifndef CONFIG_PERF_USE_VMALLOC
2272 * Back perf_mmap() with regular GFP_KERNEL-0 pages.
2275 static struct page *
2276 perf_mmap_to_page(struct perf_mmap_data *data, unsigned long pgoff)
2278 if (pgoff > data->nr_pages)
2282 return virt_to_page(data->user_page);
2284 return virt_to_page(data->data_pages[pgoff - 1]);
2287 static struct perf_mmap_data *
2288 perf_mmap_data_alloc(struct perf_event *event, int nr_pages)
2290 struct perf_mmap_data *data;
2294 WARN_ON(atomic_read(&event->mmap_count));
2296 size = sizeof(struct perf_mmap_data);
2297 size += nr_pages * sizeof(void *);
2299 data = kzalloc(size, GFP_KERNEL);
2303 data->user_page = (void *)get_zeroed_page(GFP_KERNEL);
2304 if (!data->user_page)
2305 goto fail_user_page;
2307 for (i = 0; i < nr_pages; i++) {
2308 data->data_pages[i] = (void *)get_zeroed_page(GFP_KERNEL);
2309 if (!data->data_pages[i])
2310 goto fail_data_pages;
2313 data->data_order = 0;
2314 data->nr_pages = nr_pages;
2319 for (i--; i >= 0; i--)
2320 free_page((unsigned long)data->data_pages[i]);
2322 free_page((unsigned long)data->user_page);
2331 static void perf_mmap_free_page(unsigned long addr)
2333 struct page *page = virt_to_page((void *)addr);
2335 page->mapping = NULL;
2339 static void perf_mmap_data_free(struct perf_mmap_data *data)
2343 perf_mmap_free_page((unsigned long)data->user_page);
2344 for (i = 0; i < data->nr_pages; i++)
2345 perf_mmap_free_page((unsigned long)data->data_pages[i]);
2352 * Back perf_mmap() with vmalloc memory.
2354 * Required for architectures that have d-cache aliasing issues.
2357 static struct page *
2358 perf_mmap_to_page(struct perf_mmap_data *data, unsigned long pgoff)
2360 if (pgoff > (1UL << data->data_order))
2363 return vmalloc_to_page((void *)data->user_page + pgoff * PAGE_SIZE);
2366 static void perf_mmap_unmark_page(void *addr)
2368 struct page *page = vmalloc_to_page(addr);
2370 page->mapping = NULL;
2373 static void perf_mmap_data_free_work(struct work_struct *work)
2375 struct perf_mmap_data *data;
2379 data = container_of(work, struct perf_mmap_data, work);
2380 nr = 1 << data->data_order;
2382 base = data->user_page;
2383 for (i = 0; i < nr + 1; i++)
2384 perf_mmap_unmark_page(base + (i * PAGE_SIZE));
2390 static void perf_mmap_data_free(struct perf_mmap_data *data)
2392 schedule_work(&data->work);
2395 static struct perf_mmap_data *
2396 perf_mmap_data_alloc(struct perf_event *event, int nr_pages)
2398 struct perf_mmap_data *data;
2402 WARN_ON(atomic_read(&event->mmap_count));
2404 size = sizeof(struct perf_mmap_data);
2405 size += sizeof(void *);
2407 data = kzalloc(size, GFP_KERNEL);
2411 INIT_WORK(&data->work, perf_mmap_data_free_work);
2413 all_buf = vmalloc_user((nr_pages + 1) * PAGE_SIZE);
2417 data->user_page = all_buf;
2418 data->data_pages[0] = all_buf + PAGE_SIZE;
2419 data->data_order = ilog2(nr_pages);
2433 static int perf_mmap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
2435 struct perf_event *event = vma->vm_file->private_data;
2436 struct perf_mmap_data *data;
2437 int ret = VM_FAULT_SIGBUS;
2439 if (vmf->flags & FAULT_FLAG_MKWRITE) {
2440 if (vmf->pgoff == 0)
2446 data = rcu_dereference(event->data);
2450 if (vmf->pgoff && (vmf->flags & FAULT_FLAG_WRITE))
2453 vmf->page = perf_mmap_to_page(data, vmf->pgoff);
2457 get_page(vmf->page);
2458 vmf->page->mapping = vma->vm_file->f_mapping;
2459 vmf->page->index = vmf->pgoff;
2469 perf_mmap_data_init(struct perf_event *event, struct perf_mmap_data *data)
2471 long max_size = perf_data_size(data);
2473 atomic_set(&data->lock, -1);
2475 if (event->attr.watermark) {
2476 data->watermark = min_t(long, max_size,
2477 event->attr.wakeup_watermark);
2480 if (!data->watermark)
2481 data->watermark = max_size / 2;
2484 rcu_assign_pointer(event->data, data);
2487 static void perf_mmap_data_free_rcu(struct rcu_head *rcu_head)
2489 struct perf_mmap_data *data;
2491 data = container_of(rcu_head, struct perf_mmap_data, rcu_head);
2492 perf_mmap_data_free(data);
2495 static void perf_mmap_data_release(struct perf_event *event)
2497 struct perf_mmap_data *data = event->data;
2499 WARN_ON(atomic_read(&event->mmap_count));
2501 rcu_assign_pointer(event->data, NULL);
2502 call_rcu(&data->rcu_head, perf_mmap_data_free_rcu);
2505 static void perf_mmap_open(struct vm_area_struct *vma)
2507 struct perf_event *event = vma->vm_file->private_data;
2509 atomic_inc(&event->mmap_count);
2512 static void perf_mmap_close(struct vm_area_struct *vma)
2514 struct perf_event *event = vma->vm_file->private_data;
2516 WARN_ON_ONCE(event->ctx->parent_ctx);
2517 if (atomic_dec_and_mutex_lock(&event->mmap_count, &event->mmap_mutex)) {
2518 unsigned long size = perf_data_size(event->data);
2519 struct user_struct *user = current_user();
2521 atomic_long_sub((size >> PAGE_SHIFT) + 1, &user->locked_vm);
2522 vma->vm_mm->locked_vm -= event->data->nr_locked;
2523 perf_mmap_data_release(event);
2524 mutex_unlock(&event->mmap_mutex);
2528 static const struct vm_operations_struct perf_mmap_vmops = {
2529 .open = perf_mmap_open,
2530 .close = perf_mmap_close,
2531 .fault = perf_mmap_fault,
2532 .page_mkwrite = perf_mmap_fault,
2535 static int perf_mmap(struct file *file, struct vm_area_struct *vma)
2537 struct perf_event *event = file->private_data;
2538 unsigned long user_locked, user_lock_limit;
2539 struct user_struct *user = current_user();
2540 unsigned long locked, lock_limit;
2541 struct perf_mmap_data *data;
2542 unsigned long vma_size;
2543 unsigned long nr_pages;
2544 long user_extra, extra;
2547 if (!(vma->vm_flags & VM_SHARED))
2550 vma_size = vma->vm_end - vma->vm_start;
2551 nr_pages = (vma_size / PAGE_SIZE) - 1;
2554 * If we have data pages ensure they're a power-of-two number, so we
2555 * can do bitmasks instead of modulo.
2557 if (nr_pages != 0 && !is_power_of_2(nr_pages))
2560 if (vma_size != PAGE_SIZE * (1 + nr_pages))
2563 if (vma->vm_pgoff != 0)
2566 WARN_ON_ONCE(event->ctx->parent_ctx);
2567 mutex_lock(&event->mmap_mutex);
2568 if (event->output) {
2573 if (atomic_inc_not_zero(&event->mmap_count)) {
2574 if (nr_pages != event->data->nr_pages)
2579 user_extra = nr_pages + 1;
2580 user_lock_limit = sysctl_perf_event_mlock >> (PAGE_SHIFT - 10);
2583 * Increase the limit linearly with more CPUs:
2585 user_lock_limit *= num_online_cpus();
2587 user_locked = atomic_long_read(&user->locked_vm) + user_extra;
2590 if (user_locked > user_lock_limit)
2591 extra = user_locked - user_lock_limit;
2593 lock_limit = rlimit(RLIMIT_MEMLOCK);
2594 lock_limit >>= PAGE_SHIFT;
2595 locked = vma->vm_mm->locked_vm + extra;
2597 if ((locked > lock_limit) && perf_paranoid_tracepoint_raw() &&
2598 !capable(CAP_IPC_LOCK)) {
2603 WARN_ON(event->data);
2605 data = perf_mmap_data_alloc(event, nr_pages);
2611 perf_mmap_data_init(event, data);
2613 atomic_set(&event->mmap_count, 1);
2614 atomic_long_add(user_extra, &user->locked_vm);
2615 vma->vm_mm->locked_vm += extra;
2616 event->data->nr_locked = extra;
2617 if (vma->vm_flags & VM_WRITE)
2618 event->data->writable = 1;
2621 mutex_unlock(&event->mmap_mutex);
2623 vma->vm_flags |= VM_RESERVED;
2624 vma->vm_ops = &perf_mmap_vmops;
2629 static int perf_fasync(int fd, struct file *filp, int on)
2631 struct inode *inode = filp->f_path.dentry->d_inode;
2632 struct perf_event *event = filp->private_data;
2635 mutex_lock(&inode->i_mutex);
2636 retval = fasync_helper(fd, filp, on, &event->fasync);
2637 mutex_unlock(&inode->i_mutex);
2645 static const struct file_operations perf_fops = {
2646 .release = perf_release,
2649 .unlocked_ioctl = perf_ioctl,
2650 .compat_ioctl = perf_ioctl,
2652 .fasync = perf_fasync,
2658 * If there's data, ensure we set the poll() state and publish everything
2659 * to user-space before waking everybody up.
2662 void perf_event_wakeup(struct perf_event *event)
2664 wake_up_all(&event->waitq);
2666 if (event->pending_kill) {
2667 kill_fasync(&event->fasync, SIGIO, event->pending_kill);
2668 event->pending_kill = 0;
2675 * Handle the case where we need to wakeup up from NMI (or rq->lock) context.
2677 * The NMI bit means we cannot possibly take locks. Therefore, maintain a
2678 * single linked list and use cmpxchg() to add entries lockless.
2681 static void perf_pending_event(struct perf_pending_entry *entry)
2683 struct perf_event *event = container_of(entry,
2684 struct perf_event, pending);
2686 if (event->pending_disable) {
2687 event->pending_disable = 0;
2688 __perf_event_disable(event);
2691 if (event->pending_wakeup) {
2692 event->pending_wakeup = 0;
2693 perf_event_wakeup(event);
2697 #define PENDING_TAIL ((struct perf_pending_entry *)-1UL)
2699 static DEFINE_PER_CPU(struct perf_pending_entry *, perf_pending_head) = {
2703 static void perf_pending_queue(struct perf_pending_entry *entry,
2704 void (*func)(struct perf_pending_entry *))
2706 struct perf_pending_entry **head;
2708 if (cmpxchg(&entry->next, NULL, PENDING_TAIL) != NULL)
2713 head = &get_cpu_var(perf_pending_head);
2716 entry->next = *head;
2717 } while (cmpxchg(head, entry->next, entry) != entry->next);
2719 set_perf_event_pending();
2721 put_cpu_var(perf_pending_head);
2724 static int __perf_pending_run(void)
2726 struct perf_pending_entry *list;
2729 list = xchg(&__get_cpu_var(perf_pending_head), PENDING_TAIL);
2730 while (list != PENDING_TAIL) {
2731 void (*func)(struct perf_pending_entry *);
2732 struct perf_pending_entry *entry = list;
2739 * Ensure we observe the unqueue before we issue the wakeup,
2740 * so that we won't be waiting forever.
2741 * -- see perf_not_pending().
2752 static inline int perf_not_pending(struct perf_event *event)
2755 * If we flush on whatever cpu we run, there is a chance we don't
2759 __perf_pending_run();
2763 * Ensure we see the proper queue state before going to sleep
2764 * so that we do not miss the wakeup. -- see perf_pending_handle()
2767 return event->pending.next == NULL;
2770 static void perf_pending_sync(struct perf_event *event)
2772 wait_event(event->waitq, perf_not_pending(event));
2775 void perf_event_do_pending(void)
2777 __perf_pending_run();
2781 * Callchain support -- arch specific
2784 __weak struct perf_callchain_entry *perf_callchain(struct pt_regs *regs)
2792 static bool perf_output_space(struct perf_mmap_data *data, unsigned long tail,
2793 unsigned long offset, unsigned long head)
2797 if (!data->writable)
2800 mask = perf_data_size(data) - 1;
2802 offset = (offset - tail) & mask;
2803 head = (head - tail) & mask;
2805 if ((int)(head - offset) < 0)
2811 static void perf_output_wakeup(struct perf_output_handle *handle)
2813 atomic_set(&handle->data->poll, POLL_IN);
2816 handle->event->pending_wakeup = 1;
2817 perf_pending_queue(&handle->event->pending,
2818 perf_pending_event);
2820 perf_event_wakeup(handle->event);
2824 * Curious locking construct.
2826 * We need to ensure a later event_id doesn't publish a head when a former
2827 * event_id isn't done writing. However since we need to deal with NMIs we
2828 * cannot fully serialize things.
2830 * What we do is serialize between CPUs so we only have to deal with NMI
2831 * nesting on a single CPU.
2833 * We only publish the head (and generate a wakeup) when the outer-most
2834 * event_id completes.
2836 static void perf_output_lock(struct perf_output_handle *handle)
2838 struct perf_mmap_data *data = handle->data;
2839 int cur, cpu = get_cpu();
2844 cur = atomic_cmpxchg(&data->lock, -1, cpu);
2856 static void perf_output_unlock(struct perf_output_handle *handle)
2858 struct perf_mmap_data *data = handle->data;
2862 data->done_head = data->head;
2864 if (!handle->locked)
2869 * The xchg implies a full barrier that ensures all writes are done
2870 * before we publish the new head, matched by a rmb() in userspace when
2871 * reading this position.
2873 while ((head = atomic_long_xchg(&data->done_head, 0)))
2874 data->user_page->data_head = head;
2877 * NMI can happen here, which means we can miss a done_head update.
2880 cpu = atomic_xchg(&data->lock, -1);
2881 WARN_ON_ONCE(cpu != smp_processor_id());
2884 * Therefore we have to validate we did not indeed do so.
2886 if (unlikely(atomic_long_read(&data->done_head))) {
2888 * Since we had it locked, we can lock it again.
2890 while (atomic_cmpxchg(&data->lock, -1, cpu) != -1)
2896 if (atomic_xchg(&data->wakeup, 0))
2897 perf_output_wakeup(handle);
2902 void perf_output_copy(struct perf_output_handle *handle,
2903 const void *buf, unsigned int len)
2905 unsigned int pages_mask;
2906 unsigned long offset;
2910 offset = handle->offset;
2911 pages_mask = handle->data->nr_pages - 1;
2912 pages = handle->data->data_pages;
2915 unsigned long page_offset;
2916 unsigned long page_size;
2919 nr = (offset >> PAGE_SHIFT) & pages_mask;
2920 page_size = 1UL << (handle->data->data_order + PAGE_SHIFT);
2921 page_offset = offset & (page_size - 1);
2922 size = min_t(unsigned int, page_size - page_offset, len);
2924 memcpy(pages[nr] + page_offset, buf, size);
2931 handle->offset = offset;
2934 * Check we didn't copy past our reservation window, taking the
2935 * possible unsigned int wrap into account.
2937 WARN_ON_ONCE(((long)(handle->head - handle->offset)) < 0);
2940 int perf_output_begin(struct perf_output_handle *handle,
2941 struct perf_event *event, unsigned int size,
2942 int nmi, int sample)
2944 struct perf_event *output_event;
2945 struct perf_mmap_data *data;
2946 unsigned long tail, offset, head;
2949 struct perf_event_header header;
2956 * For inherited events we send all the output towards the parent.
2959 event = event->parent;
2961 output_event = rcu_dereference(event->output);
2963 event = output_event;
2965 data = rcu_dereference(event->data);
2969 handle->data = data;
2970 handle->event = event;
2972 handle->sample = sample;
2974 if (!data->nr_pages)
2977 have_lost = atomic_read(&data->lost);
2979 size += sizeof(lost_event);
2981 perf_output_lock(handle);
2985 * Userspace could choose to issue a mb() before updating the
2986 * tail pointer. So that all reads will be completed before the
2989 tail = ACCESS_ONCE(data->user_page->data_tail);
2991 offset = head = atomic_long_read(&data->head);
2993 if (unlikely(!perf_output_space(data, tail, offset, head)))
2995 } while (atomic_long_cmpxchg(&data->head, offset, head) != offset);
2997 handle->offset = offset;
2998 handle->head = head;
3000 if (head - tail > data->watermark)
3001 atomic_set(&data->wakeup, 1);
3004 lost_event.header.type = PERF_RECORD_LOST;
3005 lost_event.header.misc = 0;
3006 lost_event.header.size = sizeof(lost_event);
3007 lost_event.id = event->id;
3008 lost_event.lost = atomic_xchg(&data->lost, 0);
3010 perf_output_put(handle, lost_event);
3016 atomic_inc(&data->lost);
3017 perf_output_unlock(handle);
3024 void perf_output_end(struct perf_output_handle *handle)
3026 struct perf_event *event = handle->event;
3027 struct perf_mmap_data *data = handle->data;
3029 int wakeup_events = event->attr.wakeup_events;
3031 if (handle->sample && wakeup_events) {
3032 int events = atomic_inc_return(&data->events);
3033 if (events >= wakeup_events) {
3034 atomic_sub(wakeup_events, &data->events);
3035 atomic_set(&data->wakeup, 1);
3039 perf_output_unlock(handle);
3043 static u32 perf_event_pid(struct perf_event *event, struct task_struct *p)
3046 * only top level events have the pid namespace they were created in
3049 event = event->parent;
3051 return task_tgid_nr_ns(p, event->ns);
3054 static u32 perf_event_tid(struct perf_event *event, struct task_struct *p)
3057 * only top level events have the pid namespace they were created in
3060 event = event->parent;
3062 return task_pid_nr_ns(p, event->ns);
3065 static void perf_output_read_one(struct perf_output_handle *handle,
3066 struct perf_event *event)
3068 u64 read_format = event->attr.read_format;
3072 values[n++] = atomic64_read(&event->count);
3073 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
3074 values[n++] = event->total_time_enabled +
3075 atomic64_read(&event->child_total_time_enabled);
3077 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
3078 values[n++] = event->total_time_running +
3079 atomic64_read(&event->child_total_time_running);
3081 if (read_format & PERF_FORMAT_ID)
3082 values[n++] = primary_event_id(event);
3084 perf_output_copy(handle, values, n * sizeof(u64));
3088 * XXX PERF_FORMAT_GROUP vs inherited events seems difficult.
3090 static void perf_output_read_group(struct perf_output_handle *handle,
3091 struct perf_event *event)
3093 struct perf_event *leader = event->group_leader, *sub;
3094 u64 read_format = event->attr.read_format;
3098 values[n++] = 1 + leader->nr_siblings;
3100 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
3101 values[n++] = leader->total_time_enabled;
3103 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
3104 values[n++] = leader->total_time_running;
3106 if (leader != event)
3107 leader->pmu->read(leader);
3109 values[n++] = atomic64_read(&leader->count);
3110 if (read_format & PERF_FORMAT_ID)
3111 values[n++] = primary_event_id(leader);
3113 perf_output_copy(handle, values, n * sizeof(u64));
3115 list_for_each_entry(sub, &leader->sibling_list, group_entry) {
3119 sub->pmu->read(sub);
3121 values[n++] = atomic64_read(&sub->count);
3122 if (read_format & PERF_FORMAT_ID)
3123 values[n++] = primary_event_id(sub);
3125 perf_output_copy(handle, values, n * sizeof(u64));
3129 static void perf_output_read(struct perf_output_handle *handle,
3130 struct perf_event *event)
3132 if (event->attr.read_format & PERF_FORMAT_GROUP)
3133 perf_output_read_group(handle, event);
3135 perf_output_read_one(handle, event);
3138 void perf_output_sample(struct perf_output_handle *handle,
3139 struct perf_event_header *header,
3140 struct perf_sample_data *data,
3141 struct perf_event *event)
3143 u64 sample_type = data->type;
3145 perf_output_put(handle, *header);
3147 if (sample_type & PERF_SAMPLE_IP)
3148 perf_output_put(handle, data->ip);
3150 if (sample_type & PERF_SAMPLE_TID)
3151 perf_output_put(handle, data->tid_entry);
3153 if (sample_type & PERF_SAMPLE_TIME)
3154 perf_output_put(handle, data->time);
3156 if (sample_type & PERF_SAMPLE_ADDR)
3157 perf_output_put(handle, data->addr);
3159 if (sample_type & PERF_SAMPLE_ID)
3160 perf_output_put(handle, data->id);
3162 if (sample_type & PERF_SAMPLE_STREAM_ID)
3163 perf_output_put(handle, data->stream_id);
3165 if (sample_type & PERF_SAMPLE_CPU)
3166 perf_output_put(handle, data->cpu_entry);
3168 if (sample_type & PERF_SAMPLE_PERIOD)
3169 perf_output_put(handle, data->period);
3171 if (sample_type & PERF_SAMPLE_READ)
3172 perf_output_read(handle, event);
3174 if (sample_type & PERF_SAMPLE_CALLCHAIN) {
3175 if (data->callchain) {
3178 if (data->callchain)
3179 size += data->callchain->nr;
3181 size *= sizeof(u64);
3183 perf_output_copy(handle, data->callchain, size);
3186 perf_output_put(handle, nr);
3190 if (sample_type & PERF_SAMPLE_RAW) {
3192 perf_output_put(handle, data->raw->size);
3193 perf_output_copy(handle, data->raw->data,
3200 .size = sizeof(u32),
3203 perf_output_put(handle, raw);
3208 void perf_prepare_sample(struct perf_event_header *header,
3209 struct perf_sample_data *data,
3210 struct perf_event *event,
3211 struct pt_regs *regs)
3213 u64 sample_type = event->attr.sample_type;
3215 data->type = sample_type;
3217 header->type = PERF_RECORD_SAMPLE;
3218 header->size = sizeof(*header);
3221 header->misc |= perf_misc_flags(regs);
3223 if (sample_type & PERF_SAMPLE_IP) {
3224 data->ip = perf_instruction_pointer(regs);
3226 header->size += sizeof(data->ip);
3229 if (sample_type & PERF_SAMPLE_TID) {
3230 /* namespace issues */
3231 data->tid_entry.pid = perf_event_pid(event, current);
3232 data->tid_entry.tid = perf_event_tid(event, current);
3234 header->size += sizeof(data->tid_entry);
3237 if (sample_type & PERF_SAMPLE_TIME) {
3238 data->time = perf_clock();
3240 header->size += sizeof(data->time);
3243 if (sample_type & PERF_SAMPLE_ADDR)
3244 header->size += sizeof(data->addr);
3246 if (sample_type & PERF_SAMPLE_ID) {
3247 data->id = primary_event_id(event);
3249 header->size += sizeof(data->id);
3252 if (sample_type & PERF_SAMPLE_STREAM_ID) {
3253 data->stream_id = event->id;
3255 header->size += sizeof(data->stream_id);
3258 if (sample_type & PERF_SAMPLE_CPU) {
3259 data->cpu_entry.cpu = raw_smp_processor_id();
3260 data->cpu_entry.reserved = 0;
3262 header->size += sizeof(data->cpu_entry);
3265 if (sample_type & PERF_SAMPLE_PERIOD)
3266 header->size += sizeof(data->period);
3268 if (sample_type & PERF_SAMPLE_READ)
3269 header->size += perf_event_read_size(event);
3271 if (sample_type & PERF_SAMPLE_CALLCHAIN) {
3274 data->callchain = perf_callchain(regs);
3276 if (data->callchain)
3277 size += data->callchain->nr;
3279 header->size += size * sizeof(u64);
3282 if (sample_type & PERF_SAMPLE_RAW) {
3283 int size = sizeof(u32);
3286 size += data->raw->size;
3288 size += sizeof(u32);
3290 WARN_ON_ONCE(size & (sizeof(u64)-1));
3291 header->size += size;
3295 static void perf_event_output(struct perf_event *event, int nmi,
3296 struct perf_sample_data *data,
3297 struct pt_regs *regs)
3299 struct perf_output_handle handle;
3300 struct perf_event_header header;
3302 perf_prepare_sample(&header, data, event, regs);
3304 if (perf_output_begin(&handle, event, header.size, nmi, 1))
3307 perf_output_sample(&handle, &header, data, event);
3309 perf_output_end(&handle);
3316 struct perf_read_event {
3317 struct perf_event_header header;
3324 perf_event_read_event(struct perf_event *event,
3325 struct task_struct *task)
3327 struct perf_output_handle handle;
3328 struct perf_read_event read_event = {
3330 .type = PERF_RECORD_READ,
3332 .size = sizeof(read_event) + perf_event_read_size(event),
3334 .pid = perf_event_pid(event, task),
3335 .tid = perf_event_tid(event, task),
3339 ret = perf_output_begin(&handle, event, read_event.header.size, 0, 0);
3343 perf_output_put(&handle, read_event);
3344 perf_output_read(&handle, event);
3346 perf_output_end(&handle);
3350 * task tracking -- fork/exit
3352 * enabled by: attr.comm | attr.mmap | attr.task
3355 struct perf_task_event {
3356 struct task_struct *task;
3357 struct perf_event_context *task_ctx;
3360 struct perf_event_header header;
3370 static void perf_event_task_output(struct perf_event *event,
3371 struct perf_task_event *task_event)
3373 struct perf_output_handle handle;
3375 struct task_struct *task = task_event->task;
3378 size = task_event->event_id.header.size;
3379 ret = perf_output_begin(&handle, event, size, 0, 0);
3384 task_event->event_id.pid = perf_event_pid(event, task);
3385 task_event->event_id.ppid = perf_event_pid(event, current);
3387 task_event->event_id.tid = perf_event_tid(event, task);
3388 task_event->event_id.ptid = perf_event_tid(event, current);
3390 perf_output_put(&handle, task_event->event_id);
3392 perf_output_end(&handle);
3395 static int perf_event_task_match(struct perf_event *event)
3397 if (event->state < PERF_EVENT_STATE_INACTIVE)
3400 if (event->cpu != -1 && event->cpu != smp_processor_id())
3403 if (event->attr.comm || event->attr.mmap || event->attr.task)
3409 static void perf_event_task_ctx(struct perf_event_context *ctx,
3410 struct perf_task_event *task_event)
3412 struct perf_event *event;
3414 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
3415 if (perf_event_task_match(event))
3416 perf_event_task_output(event, task_event);
3420 static void perf_event_task_event(struct perf_task_event *task_event)
3422 struct perf_cpu_context *cpuctx;
3423 struct perf_event_context *ctx = task_event->task_ctx;
3426 cpuctx = &get_cpu_var(perf_cpu_context);
3427 perf_event_task_ctx(&cpuctx->ctx, task_event);
3429 ctx = rcu_dereference(current->perf_event_ctxp);
3431 perf_event_task_ctx(ctx, task_event);
3432 put_cpu_var(perf_cpu_context);
3436 static void perf_event_task(struct task_struct *task,
3437 struct perf_event_context *task_ctx,
3440 struct perf_task_event task_event;
3442 if (!atomic_read(&nr_comm_events) &&
3443 !atomic_read(&nr_mmap_events) &&
3444 !atomic_read(&nr_task_events))
3447 task_event = (struct perf_task_event){
3449 .task_ctx = task_ctx,
3452 .type = new ? PERF_RECORD_FORK : PERF_RECORD_EXIT,
3454 .size = sizeof(task_event.event_id),
3460 .time = perf_clock(),
3464 perf_event_task_event(&task_event);
3467 void perf_event_fork(struct task_struct *task)
3469 perf_event_task(task, NULL, 1);
3476 struct perf_comm_event {
3477 struct task_struct *task;
3482 struct perf_event_header header;
3489 static void perf_event_comm_output(struct perf_event *event,
3490 struct perf_comm_event *comm_event)
3492 struct perf_output_handle handle;
3493 int size = comm_event->event_id.header.size;
3494 int ret = perf_output_begin(&handle, event, size, 0, 0);
3499 comm_event->event_id.pid = perf_event_pid(event, comm_event->task);
3500 comm_event->event_id.tid = perf_event_tid(event, comm_event->task);
3502 perf_output_put(&handle, comm_event->event_id);
3503 perf_output_copy(&handle, comm_event->comm,
3504 comm_event->comm_size);
3505 perf_output_end(&handle);
3508 static int perf_event_comm_match(struct perf_event *event)
3510 if (event->state < PERF_EVENT_STATE_INACTIVE)
3513 if (event->cpu != -1 && event->cpu != smp_processor_id())
3516 if (event->attr.comm)
3522 static void perf_event_comm_ctx(struct perf_event_context *ctx,
3523 struct perf_comm_event *comm_event)
3525 struct perf_event *event;
3527 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
3528 if (perf_event_comm_match(event))
3529 perf_event_comm_output(event, comm_event);
3533 static void perf_event_comm_event(struct perf_comm_event *comm_event)
3535 struct perf_cpu_context *cpuctx;
3536 struct perf_event_context *ctx;
3538 char comm[TASK_COMM_LEN];
3540 memset(comm, 0, sizeof(comm));
3541 strlcpy(comm, comm_event->task->comm, sizeof(comm));
3542 size = ALIGN(strlen(comm)+1, sizeof(u64));
3544 comm_event->comm = comm;
3545 comm_event->comm_size = size;
3547 comm_event->event_id.header.size = sizeof(comm_event->event_id) + size;
3550 cpuctx = &get_cpu_var(perf_cpu_context);
3551 perf_event_comm_ctx(&cpuctx->ctx, comm_event);
3552 ctx = rcu_dereference(current->perf_event_ctxp);
3554 perf_event_comm_ctx(ctx, comm_event);
3555 put_cpu_var(perf_cpu_context);
3559 void perf_event_comm(struct task_struct *task)
3561 struct perf_comm_event comm_event;
3563 if (task->perf_event_ctxp)
3564 perf_event_enable_on_exec(task);
3566 if (!atomic_read(&nr_comm_events))
3569 comm_event = (struct perf_comm_event){
3575 .type = PERF_RECORD_COMM,
3584 perf_event_comm_event(&comm_event);
3591 struct perf_mmap_event {
3592 struct vm_area_struct *vma;
3594 const char *file_name;
3598 struct perf_event_header header;
3608 static void perf_event_mmap_output(struct perf_event *event,
3609 struct perf_mmap_event *mmap_event)
3611 struct perf_output_handle handle;
3612 int size = mmap_event->event_id.header.size;
3613 int ret = perf_output_begin(&handle, event, size, 0, 0);
3618 mmap_event->event_id.pid = perf_event_pid(event, current);
3619 mmap_event->event_id.tid = perf_event_tid(event, current);
3621 perf_output_put(&handle, mmap_event->event_id);
3622 perf_output_copy(&handle, mmap_event->file_name,
3623 mmap_event->file_size);
3624 perf_output_end(&handle);
3627 static int perf_event_mmap_match(struct perf_event *event,
3628 struct perf_mmap_event *mmap_event)
3630 if (event->state < PERF_EVENT_STATE_INACTIVE)
3633 if (event->cpu != -1 && event->cpu != smp_processor_id())
3636 if (event->attr.mmap)
3642 static void perf_event_mmap_ctx(struct perf_event_context *ctx,
3643 struct perf_mmap_event *mmap_event)
3645 struct perf_event *event;
3647 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
3648 if (perf_event_mmap_match(event, mmap_event))
3649 perf_event_mmap_output(event, mmap_event);
3653 static void perf_event_mmap_event(struct perf_mmap_event *mmap_event)
3655 struct perf_cpu_context *cpuctx;
3656 struct perf_event_context *ctx;
3657 struct vm_area_struct *vma = mmap_event->vma;
3658 struct file *file = vma->vm_file;
3664 memset(tmp, 0, sizeof(tmp));
3668 * d_path works from the end of the buffer backwards, so we
3669 * need to add enough zero bytes after the string to handle
3670 * the 64bit alignment we do later.
3672 buf = kzalloc(PATH_MAX + sizeof(u64), GFP_KERNEL);
3674 name = strncpy(tmp, "//enomem", sizeof(tmp));
3677 name = d_path(&file->f_path, buf, PATH_MAX);
3679 name = strncpy(tmp, "//toolong", sizeof(tmp));
3683 if (arch_vma_name(mmap_event->vma)) {
3684 name = strncpy(tmp, arch_vma_name(mmap_event->vma),
3690 name = strncpy(tmp, "[vdso]", sizeof(tmp));
3694 name = strncpy(tmp, "//anon", sizeof(tmp));
3699 size = ALIGN(strlen(name)+1, sizeof(u64));
3701 mmap_event->file_name = name;
3702 mmap_event->file_size = size;
3704 mmap_event->event_id.header.size = sizeof(mmap_event->event_id) + size;
3707 cpuctx = &get_cpu_var(perf_cpu_context);
3708 perf_event_mmap_ctx(&cpuctx->ctx, mmap_event);
3709 ctx = rcu_dereference(current->perf_event_ctxp);
3711 perf_event_mmap_ctx(ctx, mmap_event);
3712 put_cpu_var(perf_cpu_context);
3718 void __perf_event_mmap(struct vm_area_struct *vma)
3720 struct perf_mmap_event mmap_event;
3722 if (!atomic_read(&nr_mmap_events))
3725 mmap_event = (struct perf_mmap_event){
3731 .type = PERF_RECORD_MMAP,
3737 .start = vma->vm_start,
3738 .len = vma->vm_end - vma->vm_start,
3739 .pgoff = (u64)vma->vm_pgoff << PAGE_SHIFT,
3743 perf_event_mmap_event(&mmap_event);
3747 * IRQ throttle logging
3750 static void perf_log_throttle(struct perf_event *event, int enable)
3752 struct perf_output_handle handle;
3756 struct perf_event_header header;
3760 } throttle_event = {
3762 .type = PERF_RECORD_THROTTLE,
3764 .size = sizeof(throttle_event),
3766 .time = perf_clock(),
3767 .id = primary_event_id(event),
3768 .stream_id = event->id,
3772 throttle_event.header.type = PERF_RECORD_UNTHROTTLE;
3774 ret = perf_output_begin(&handle, event, sizeof(throttle_event), 1, 0);
3778 perf_output_put(&handle, throttle_event);
3779 perf_output_end(&handle);
3783 * Generic event overflow handling, sampling.
3786 static int __perf_event_overflow(struct perf_event *event, int nmi,
3787 int throttle, struct perf_sample_data *data,
3788 struct pt_regs *regs)
3790 int events = atomic_read(&event->event_limit);
3791 struct hw_perf_event *hwc = &event->hw;
3794 throttle = (throttle && event->pmu->unthrottle != NULL);
3799 if (hwc->interrupts != MAX_INTERRUPTS) {
3801 if (HZ * hwc->interrupts >
3802 (u64)sysctl_perf_event_sample_rate) {
3803 hwc->interrupts = MAX_INTERRUPTS;
3804 perf_log_throttle(event, 0);
3809 * Keep re-disabling events even though on the previous
3810 * pass we disabled it - just in case we raced with a
3811 * sched-in and the event got enabled again:
3817 if (event->attr.freq) {
3818 u64 now = perf_clock();
3819 s64 delta = now - hwc->freq_time_stamp;
3821 hwc->freq_time_stamp = now;
3823 if (delta > 0 && delta < 2*TICK_NSEC)
3824 perf_adjust_period(event, delta, hwc->last_period);
3828 * XXX event_limit might not quite work as expected on inherited
3832 event->pending_kill = POLL_IN;
3833 if (events && atomic_dec_and_test(&event->event_limit)) {
3835 event->pending_kill = POLL_HUP;
3837 event->pending_disable = 1;
3838 perf_pending_queue(&event->pending,
3839 perf_pending_event);
3841 perf_event_disable(event);
3844 if (event->overflow_handler)
3845 event->overflow_handler(event, nmi, data, regs);
3847 perf_event_output(event, nmi, data, regs);
3852 int perf_event_overflow(struct perf_event *event, int nmi,
3853 struct perf_sample_data *data,
3854 struct pt_regs *regs)
3856 return __perf_event_overflow(event, nmi, 1, data, regs);
3860 * Generic software event infrastructure
3864 * We directly increment event->count and keep a second value in
3865 * event->hw.period_left to count intervals. This period event
3866 * is kept in the range [-sample_period, 0] so that we can use the
3870 static u64 perf_swevent_set_period(struct perf_event *event)
3872 struct hw_perf_event *hwc = &event->hw;
3873 u64 period = hwc->last_period;
3877 hwc->last_period = hwc->sample_period;
3880 old = val = atomic64_read(&hwc->period_left);
3884 nr = div64_u64(period + val, period);
3885 offset = nr * period;
3887 if (atomic64_cmpxchg(&hwc->period_left, old, val) != old)
3893 static void perf_swevent_overflow(struct perf_event *event, u64 overflow,
3894 int nmi, struct perf_sample_data *data,
3895 struct pt_regs *regs)
3897 struct hw_perf_event *hwc = &event->hw;
3900 data->period = event->hw.last_period;
3902 overflow = perf_swevent_set_period(event);
3904 if (hwc->interrupts == MAX_INTERRUPTS)
3907 for (; overflow; overflow--) {
3908 if (__perf_event_overflow(event, nmi, throttle,
3911 * We inhibit the overflow from happening when
3912 * hwc->interrupts == MAX_INTERRUPTS.
3920 static void perf_swevent_unthrottle(struct perf_event *event)
3923 * Nothing to do, we already reset hwc->interrupts.
3927 static void perf_swevent_add(struct perf_event *event, u64 nr,
3928 int nmi, struct perf_sample_data *data,
3929 struct pt_regs *regs)
3931 struct hw_perf_event *hwc = &event->hw;
3933 atomic64_add(nr, &event->count);
3938 if (!hwc->sample_period)
3941 if (nr == 1 && hwc->sample_period == 1 && !event->attr.freq)
3942 return perf_swevent_overflow(event, 1, nmi, data, regs);
3944 if (atomic64_add_negative(nr, &hwc->period_left))
3947 perf_swevent_overflow(event, 0, nmi, data, regs);
3950 static int perf_swevent_is_counting(struct perf_event *event)
3953 * The event is active, we're good!
3955 if (event->state == PERF_EVENT_STATE_ACTIVE)
3959 * The event is off/error, not counting.
3961 if (event->state != PERF_EVENT_STATE_INACTIVE)
3965 * The event is inactive, if the context is active
3966 * we're part of a group that didn't make it on the 'pmu',
3969 if (event->ctx->is_active)
3973 * We're inactive and the context is too, this means the
3974 * task is scheduled out, we're counting events that happen
3975 * to us, like migration events.
3980 static int perf_tp_event_match(struct perf_event *event,
3981 struct perf_sample_data *data);
3983 static int perf_exclude_event(struct perf_event *event,
3984 struct pt_regs *regs)
3987 if (event->attr.exclude_user && user_mode(regs))
3990 if (event->attr.exclude_kernel && !user_mode(regs))
3997 static int perf_swevent_match(struct perf_event *event,
3998 enum perf_type_id type,
4000 struct perf_sample_data *data,
4001 struct pt_regs *regs)
4003 if (event->cpu != -1 && event->cpu != smp_processor_id())
4006 if (!perf_swevent_is_counting(event))
4009 if (event->attr.type != type)
4012 if (event->attr.config != event_id)
4015 if (perf_exclude_event(event, regs))
4018 if (event->attr.type == PERF_TYPE_TRACEPOINT &&
4019 !perf_tp_event_match(event, data))
4025 static void perf_swevent_ctx_event(struct perf_event_context *ctx,
4026 enum perf_type_id type,
4027 u32 event_id, u64 nr, int nmi,
4028 struct perf_sample_data *data,
4029 struct pt_regs *regs)
4031 struct perf_event *event;
4033 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
4034 if (perf_swevent_match(event, type, event_id, data, regs))
4035 perf_swevent_add(event, nr, nmi, data, regs);
4039 int perf_swevent_get_recursion_context(void)
4041 struct perf_cpu_context *cpuctx = &get_cpu_var(perf_cpu_context);
4048 else if (in_softirq())
4053 if (cpuctx->recursion[rctx]) {
4054 put_cpu_var(perf_cpu_context);
4058 cpuctx->recursion[rctx]++;
4063 EXPORT_SYMBOL_GPL(perf_swevent_get_recursion_context);
4065 void perf_swevent_put_recursion_context(int rctx)
4067 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
4069 cpuctx->recursion[rctx]--;
4070 put_cpu_var(perf_cpu_context);
4072 EXPORT_SYMBOL_GPL(perf_swevent_put_recursion_context);
4074 static void do_perf_sw_event(enum perf_type_id type, u32 event_id,
4076 struct perf_sample_data *data,
4077 struct pt_regs *regs)
4079 struct perf_cpu_context *cpuctx;
4080 struct perf_event_context *ctx;
4082 cpuctx = &__get_cpu_var(perf_cpu_context);
4084 perf_swevent_ctx_event(&cpuctx->ctx, type, event_id,
4085 nr, nmi, data, regs);
4087 * doesn't really matter which of the child contexts the
4088 * events ends up in.
4090 ctx = rcu_dereference(current->perf_event_ctxp);
4092 perf_swevent_ctx_event(ctx, type, event_id, nr, nmi, data, regs);
4096 void __perf_sw_event(u32 event_id, u64 nr, int nmi,
4097 struct pt_regs *regs, u64 addr)
4099 struct perf_sample_data data;
4102 rctx = perf_swevent_get_recursion_context();
4106 perf_sample_data_init(&data, addr);
4108 do_perf_sw_event(PERF_TYPE_SOFTWARE, event_id, nr, nmi, &data, regs);
4110 perf_swevent_put_recursion_context(rctx);
4113 static void perf_swevent_read(struct perf_event *event)
4117 static int perf_swevent_enable(struct perf_event *event)
4119 struct hw_perf_event *hwc = &event->hw;
4121 if (hwc->sample_period) {
4122 hwc->last_period = hwc->sample_period;
4123 perf_swevent_set_period(event);
4128 static void perf_swevent_disable(struct perf_event *event)
4132 static const struct pmu perf_ops_generic = {
4133 .enable = perf_swevent_enable,
4134 .disable = perf_swevent_disable,
4135 .read = perf_swevent_read,
4136 .unthrottle = perf_swevent_unthrottle,
4140 * hrtimer based swevent callback
4143 static enum hrtimer_restart perf_swevent_hrtimer(struct hrtimer *hrtimer)
4145 enum hrtimer_restart ret = HRTIMER_RESTART;
4146 struct perf_sample_data data;
4147 struct pt_regs *regs;
4148 struct perf_event *event;
4151 event = container_of(hrtimer, struct perf_event, hw.hrtimer);
4152 event->pmu->read(event);
4154 perf_sample_data_init(&data, 0);
4155 data.period = event->hw.last_period;
4156 regs = get_irq_regs();
4158 * In case we exclude kernel IPs or are somehow not in interrupt
4159 * context, provide the next best thing, the user IP.
4161 if ((event->attr.exclude_kernel || !regs) &&
4162 !event->attr.exclude_user)
4163 regs = task_pt_regs(current);
4166 if (!(event->attr.exclude_idle && current->pid == 0))
4167 if (perf_event_overflow(event, 0, &data, regs))
4168 ret = HRTIMER_NORESTART;
4171 period = max_t(u64, 10000, event->hw.sample_period);
4172 hrtimer_forward_now(hrtimer, ns_to_ktime(period));
4177 static void perf_swevent_start_hrtimer(struct perf_event *event)
4179 struct hw_perf_event *hwc = &event->hw;
4181 hrtimer_init(&hwc->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
4182 hwc->hrtimer.function = perf_swevent_hrtimer;
4183 if (hwc->sample_period) {
4186 if (hwc->remaining) {
4187 if (hwc->remaining < 0)
4190 period = hwc->remaining;
4193 period = max_t(u64, 10000, hwc->sample_period);
4195 __hrtimer_start_range_ns(&hwc->hrtimer,
4196 ns_to_ktime(period), 0,
4197 HRTIMER_MODE_REL, 0);
4201 static void perf_swevent_cancel_hrtimer(struct perf_event *event)
4203 struct hw_perf_event *hwc = &event->hw;
4205 if (hwc->sample_period) {
4206 ktime_t remaining = hrtimer_get_remaining(&hwc->hrtimer);
4207 hwc->remaining = ktime_to_ns(remaining);
4209 hrtimer_cancel(&hwc->hrtimer);
4214 * Software event: cpu wall time clock
4217 static void cpu_clock_perf_event_update(struct perf_event *event)
4219 int cpu = raw_smp_processor_id();
4223 now = cpu_clock(cpu);
4224 prev = atomic64_xchg(&event->hw.prev_count, now);
4225 atomic64_add(now - prev, &event->count);
4228 static int cpu_clock_perf_event_enable(struct perf_event *event)
4230 struct hw_perf_event *hwc = &event->hw;
4231 int cpu = raw_smp_processor_id();
4233 atomic64_set(&hwc->prev_count, cpu_clock(cpu));
4234 perf_swevent_start_hrtimer(event);
4239 static void cpu_clock_perf_event_disable(struct perf_event *event)
4241 perf_swevent_cancel_hrtimer(event);
4242 cpu_clock_perf_event_update(event);
4245 static void cpu_clock_perf_event_read(struct perf_event *event)
4247 cpu_clock_perf_event_update(event);
4250 static const struct pmu perf_ops_cpu_clock = {
4251 .enable = cpu_clock_perf_event_enable,
4252 .disable = cpu_clock_perf_event_disable,
4253 .read = cpu_clock_perf_event_read,
4257 * Software event: task time clock
4260 static void task_clock_perf_event_update(struct perf_event *event, u64 now)
4265 prev = atomic64_xchg(&event->hw.prev_count, now);
4267 atomic64_add(delta, &event->count);
4270 static int task_clock_perf_event_enable(struct perf_event *event)
4272 struct hw_perf_event *hwc = &event->hw;
4275 now = event->ctx->time;
4277 atomic64_set(&hwc->prev_count, now);
4279 perf_swevent_start_hrtimer(event);
4284 static void task_clock_perf_event_disable(struct perf_event *event)
4286 perf_swevent_cancel_hrtimer(event);
4287 task_clock_perf_event_update(event, event->ctx->time);
4291 static void task_clock_perf_event_read(struct perf_event *event)
4296 update_context_time(event->ctx);
4297 time = event->ctx->time;
4299 u64 now = perf_clock();
4300 u64 delta = now - event->ctx->timestamp;
4301 time = event->ctx->time + delta;
4304 task_clock_perf_event_update(event, time);
4307 static const struct pmu perf_ops_task_clock = {
4308 .enable = task_clock_perf_event_enable,
4309 .disable = task_clock_perf_event_disable,
4310 .read = task_clock_perf_event_read,
4313 #ifdef CONFIG_EVENT_TRACING
4315 void perf_tp_event(int event_id, u64 addr, u64 count, void *record,
4318 struct pt_regs *regs = get_irq_regs();
4319 struct perf_sample_data data;
4320 struct perf_raw_record raw = {
4325 perf_sample_data_init(&data, addr);
4329 regs = task_pt_regs(current);
4331 /* Trace events already protected against recursion */
4332 do_perf_sw_event(PERF_TYPE_TRACEPOINT, event_id, count, 1,
4335 EXPORT_SYMBOL_GPL(perf_tp_event);
4337 static int perf_tp_event_match(struct perf_event *event,
4338 struct perf_sample_data *data)
4340 void *record = data->raw->data;
4342 if (likely(!event->filter) || filter_match_preds(event->filter, record))
4347 static void tp_perf_event_destroy(struct perf_event *event)
4349 ftrace_profile_disable(event->attr.config);
4352 static const struct pmu *tp_perf_event_init(struct perf_event *event)
4355 * Raw tracepoint data is a severe data leak, only allow root to
4358 if ((event->attr.sample_type & PERF_SAMPLE_RAW) &&
4359 perf_paranoid_tracepoint_raw() &&
4360 !capable(CAP_SYS_ADMIN))
4361 return ERR_PTR(-EPERM);
4363 if (ftrace_profile_enable(event->attr.config))
4366 event->destroy = tp_perf_event_destroy;
4368 return &perf_ops_generic;
4371 static int perf_event_set_filter(struct perf_event *event, void __user *arg)
4376 if (event->attr.type != PERF_TYPE_TRACEPOINT)
4379 filter_str = strndup_user(arg, PAGE_SIZE);
4380 if (IS_ERR(filter_str))
4381 return PTR_ERR(filter_str);
4383 ret = ftrace_profile_set_filter(event, event->attr.config, filter_str);
4389 static void perf_event_free_filter(struct perf_event *event)
4391 ftrace_profile_free_filter(event);
4396 static int perf_tp_event_match(struct perf_event *event,
4397 struct perf_sample_data *data)
4402 static const struct pmu *tp_perf_event_init(struct perf_event *event)
4407 static int perf_event_set_filter(struct perf_event *event, void __user *arg)
4412 static void perf_event_free_filter(struct perf_event *event)
4416 #endif /* CONFIG_EVENT_TRACING */
4418 #ifdef CONFIG_HAVE_HW_BREAKPOINT
4419 static void bp_perf_event_destroy(struct perf_event *event)
4421 release_bp_slot(event);
4424 static const struct pmu *bp_perf_event_init(struct perf_event *bp)
4428 err = register_perf_hw_breakpoint(bp);
4430 return ERR_PTR(err);
4432 bp->destroy = bp_perf_event_destroy;
4434 return &perf_ops_bp;
4437 void perf_bp_event(struct perf_event *bp, void *data)
4439 struct perf_sample_data sample;
4440 struct pt_regs *regs = data;
4442 perf_sample_data_init(&sample, bp->attr.bp_addr);
4444 if (!perf_exclude_event(bp, regs))
4445 perf_swevent_add(bp, 1, 1, &sample, regs);
4448 static const struct pmu *bp_perf_event_init(struct perf_event *bp)
4453 void perf_bp_event(struct perf_event *bp, void *regs)
4458 atomic_t perf_swevent_enabled[PERF_COUNT_SW_MAX];
4460 static void sw_perf_event_destroy(struct perf_event *event)
4462 u64 event_id = event->attr.config;
4464 WARN_ON(event->parent);
4466 atomic_dec(&perf_swevent_enabled[event_id]);
4469 static const struct pmu *sw_perf_event_init(struct perf_event *event)
4471 const struct pmu *pmu = NULL;
4472 u64 event_id = event->attr.config;
4475 * Software events (currently) can't in general distinguish
4476 * between user, kernel and hypervisor events.
4477 * However, context switches and cpu migrations are considered
4478 * to be kernel events, and page faults are never hypervisor
4482 case PERF_COUNT_SW_CPU_CLOCK:
4483 pmu = &perf_ops_cpu_clock;
4486 case PERF_COUNT_SW_TASK_CLOCK:
4488 * If the user instantiates this as a per-cpu event,
4489 * use the cpu_clock event instead.
4491 if (event->ctx->task)
4492 pmu = &perf_ops_task_clock;
4494 pmu = &perf_ops_cpu_clock;
4497 case PERF_COUNT_SW_PAGE_FAULTS:
4498 case PERF_COUNT_SW_PAGE_FAULTS_MIN:
4499 case PERF_COUNT_SW_PAGE_FAULTS_MAJ:
4500 case PERF_COUNT_SW_CONTEXT_SWITCHES:
4501 case PERF_COUNT_SW_CPU_MIGRATIONS:
4502 case PERF_COUNT_SW_ALIGNMENT_FAULTS:
4503 case PERF_COUNT_SW_EMULATION_FAULTS:
4504 if (!event->parent) {
4505 atomic_inc(&perf_swevent_enabled[event_id]);
4506 event->destroy = sw_perf_event_destroy;
4508 pmu = &perf_ops_generic;
4516 * Allocate and initialize a event structure
4518 static struct perf_event *
4519 perf_event_alloc(struct perf_event_attr *attr,
4521 struct perf_event_context *ctx,
4522 struct perf_event *group_leader,
4523 struct perf_event *parent_event,
4524 perf_overflow_handler_t overflow_handler,
4527 const struct pmu *pmu;
4528 struct perf_event *event;
4529 struct hw_perf_event *hwc;
4532 event = kzalloc(sizeof(*event), gfpflags);
4534 return ERR_PTR(-ENOMEM);
4537 * Single events are their own group leaders, with an
4538 * empty sibling list:
4541 group_leader = event;
4543 mutex_init(&event->child_mutex);
4544 INIT_LIST_HEAD(&event->child_list);
4546 INIT_LIST_HEAD(&event->group_entry);
4547 INIT_LIST_HEAD(&event->event_entry);
4548 INIT_LIST_HEAD(&event->sibling_list);
4549 init_waitqueue_head(&event->waitq);
4551 mutex_init(&event->mmap_mutex);
4554 event->attr = *attr;
4555 event->group_leader = group_leader;
4560 event->parent = parent_event;
4562 event->ns = get_pid_ns(current->nsproxy->pid_ns);
4563 event->id = atomic64_inc_return(&perf_event_id);
4565 event->state = PERF_EVENT_STATE_INACTIVE;
4567 if (!overflow_handler && parent_event)
4568 overflow_handler = parent_event->overflow_handler;
4570 event->overflow_handler = overflow_handler;
4573 event->state = PERF_EVENT_STATE_OFF;
4578 hwc->sample_period = attr->sample_period;
4579 if (attr->freq && attr->sample_freq)
4580 hwc->sample_period = 1;
4581 hwc->last_period = hwc->sample_period;
4583 atomic64_set(&hwc->period_left, hwc->sample_period);
4586 * we currently do not support PERF_FORMAT_GROUP on inherited events
4588 if (attr->inherit && (attr->read_format & PERF_FORMAT_GROUP))
4591 switch (attr->type) {
4593 case PERF_TYPE_HARDWARE:
4594 case PERF_TYPE_HW_CACHE:
4595 pmu = hw_perf_event_init(event);
4598 case PERF_TYPE_SOFTWARE:
4599 pmu = sw_perf_event_init(event);
4602 case PERF_TYPE_TRACEPOINT:
4603 pmu = tp_perf_event_init(event);
4606 case PERF_TYPE_BREAKPOINT:
4607 pmu = bp_perf_event_init(event);
4618 else if (IS_ERR(pmu))
4623 put_pid_ns(event->ns);
4625 return ERR_PTR(err);
4630 if (!event->parent) {
4631 atomic_inc(&nr_events);
4632 if (event->attr.mmap)
4633 atomic_inc(&nr_mmap_events);
4634 if (event->attr.comm)
4635 atomic_inc(&nr_comm_events);
4636 if (event->attr.task)
4637 atomic_inc(&nr_task_events);
4643 static int perf_copy_attr(struct perf_event_attr __user *uattr,
4644 struct perf_event_attr *attr)
4649 if (!access_ok(VERIFY_WRITE, uattr, PERF_ATTR_SIZE_VER0))
4653 * zero the full structure, so that a short copy will be nice.
4655 memset(attr, 0, sizeof(*attr));
4657 ret = get_user(size, &uattr->size);
4661 if (size > PAGE_SIZE) /* silly large */
4664 if (!size) /* abi compat */
4665 size = PERF_ATTR_SIZE_VER0;
4667 if (size < PERF_ATTR_SIZE_VER0)
4671 * If we're handed a bigger struct than we know of,
4672 * ensure all the unknown bits are 0 - i.e. new
4673 * user-space does not rely on any kernel feature
4674 * extensions we dont know about yet.
4676 if (size > sizeof(*attr)) {
4677 unsigned char __user *addr;
4678 unsigned char __user *end;
4681 addr = (void __user *)uattr + sizeof(*attr);
4682 end = (void __user *)uattr + size;
4684 for (; addr < end; addr++) {
4685 ret = get_user(val, addr);
4691 size = sizeof(*attr);
4694 ret = copy_from_user(attr, uattr, size);
4699 * If the type exists, the corresponding creation will verify
4702 if (attr->type >= PERF_TYPE_MAX)
4705 if (attr->__reserved_1)
4708 if (attr->sample_type & ~(PERF_SAMPLE_MAX-1))
4711 if (attr->read_format & ~(PERF_FORMAT_MAX-1))
4718 put_user(sizeof(*attr), &uattr->size);
4723 static int perf_event_set_output(struct perf_event *event, int output_fd)
4725 struct perf_event *output_event = NULL;
4726 struct file *output_file = NULL;
4727 struct perf_event *old_output;
4728 int fput_needed = 0;
4734 output_file = fget_light(output_fd, &fput_needed);
4738 if (output_file->f_op != &perf_fops)
4741 output_event = output_file->private_data;
4743 /* Don't chain output fds */
4744 if (output_event->output)
4747 /* Don't set an output fd when we already have an output channel */
4751 atomic_long_inc(&output_file->f_count);
4754 mutex_lock(&event->mmap_mutex);
4755 old_output = event->output;
4756 rcu_assign_pointer(event->output, output_event);
4757 mutex_unlock(&event->mmap_mutex);
4761 * we need to make sure no existing perf_output_*()
4762 * is still referencing this event.
4765 fput(old_output->filp);
4770 fput_light(output_file, fput_needed);
4775 * sys_perf_event_open - open a performance event, associate it to a task/cpu
4777 * @attr_uptr: event_id type attributes for monitoring/sampling
4780 * @group_fd: group leader event fd
4782 SYSCALL_DEFINE5(perf_event_open,
4783 struct perf_event_attr __user *, attr_uptr,
4784 pid_t, pid, int, cpu, int, group_fd, unsigned long, flags)
4786 struct perf_event *event, *group_leader;
4787 struct perf_event_attr attr;
4788 struct perf_event_context *ctx;
4789 struct file *event_file = NULL;
4790 struct file *group_file = NULL;
4791 int fput_needed = 0;
4792 int fput_needed2 = 0;
4795 /* for future expandability... */
4796 if (flags & ~(PERF_FLAG_FD_NO_GROUP | PERF_FLAG_FD_OUTPUT))
4799 err = perf_copy_attr(attr_uptr, &attr);
4803 if (!attr.exclude_kernel) {
4804 if (perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN))
4809 if (attr.sample_freq > sysctl_perf_event_sample_rate)
4814 * Get the target context (task or percpu):
4816 ctx = find_get_context(pid, cpu);
4818 return PTR_ERR(ctx);
4821 * Look up the group leader (we will attach this event to it):
4823 group_leader = NULL;
4824 if (group_fd != -1 && !(flags & PERF_FLAG_FD_NO_GROUP)) {
4826 group_file = fget_light(group_fd, &fput_needed);
4828 goto err_put_context;
4829 if (group_file->f_op != &perf_fops)
4830 goto err_put_context;
4832 group_leader = group_file->private_data;
4834 * Do not allow a recursive hierarchy (this new sibling
4835 * becoming part of another group-sibling):
4837 if (group_leader->group_leader != group_leader)
4838 goto err_put_context;
4840 * Do not allow to attach to a group in a different
4841 * task or CPU context:
4843 if (group_leader->ctx != ctx)
4844 goto err_put_context;
4846 * Only a group leader can be exclusive or pinned
4848 if (attr.exclusive || attr.pinned)
4849 goto err_put_context;
4852 event = perf_event_alloc(&attr, cpu, ctx, group_leader,
4853 NULL, NULL, GFP_KERNEL);
4854 err = PTR_ERR(event);
4856 goto err_put_context;
4858 err = anon_inode_getfd("[perf_event]", &perf_fops, event, O_RDWR);
4860 goto err_free_put_context;
4862 event_file = fget_light(err, &fput_needed2);
4864 goto err_free_put_context;
4866 if (flags & PERF_FLAG_FD_OUTPUT) {
4867 err = perf_event_set_output(event, group_fd);
4869 goto err_fput_free_put_context;
4872 event->filp = event_file;
4873 WARN_ON_ONCE(ctx->parent_ctx);
4874 mutex_lock(&ctx->mutex);
4875 perf_install_in_context(ctx, event, cpu);
4877 mutex_unlock(&ctx->mutex);
4879 event->owner = current;
4880 get_task_struct(current);
4881 mutex_lock(¤t->perf_event_mutex);
4882 list_add_tail(&event->owner_entry, ¤t->perf_event_list);
4883 mutex_unlock(¤t->perf_event_mutex);
4885 err_fput_free_put_context:
4886 fput_light(event_file, fput_needed2);
4888 err_free_put_context:
4896 fput_light(group_file, fput_needed);
4902 * perf_event_create_kernel_counter
4904 * @attr: attributes of the counter to create
4905 * @cpu: cpu in which the counter is bound
4906 * @pid: task to profile
4909 perf_event_create_kernel_counter(struct perf_event_attr *attr, int cpu,
4911 perf_overflow_handler_t overflow_handler)
4913 struct perf_event *event;
4914 struct perf_event_context *ctx;
4918 * Get the target context (task or percpu):
4921 ctx = find_get_context(pid, cpu);
4927 event = perf_event_alloc(attr, cpu, ctx, NULL,
4928 NULL, overflow_handler, GFP_KERNEL);
4929 if (IS_ERR(event)) {
4930 err = PTR_ERR(event);
4931 goto err_put_context;
4935 WARN_ON_ONCE(ctx->parent_ctx);
4936 mutex_lock(&ctx->mutex);
4937 perf_install_in_context(ctx, event, cpu);
4939 mutex_unlock(&ctx->mutex);
4941 event->owner = current;
4942 get_task_struct(current);
4943 mutex_lock(¤t->perf_event_mutex);
4944 list_add_tail(&event->owner_entry, ¤t->perf_event_list);
4945 mutex_unlock(¤t->perf_event_mutex);
4952 return ERR_PTR(err);
4954 EXPORT_SYMBOL_GPL(perf_event_create_kernel_counter);
4957 * inherit a event from parent task to child task:
4959 static struct perf_event *
4960 inherit_event(struct perf_event *parent_event,
4961 struct task_struct *parent,
4962 struct perf_event_context *parent_ctx,
4963 struct task_struct *child,
4964 struct perf_event *group_leader,
4965 struct perf_event_context *child_ctx)
4967 struct perf_event *child_event;
4970 * Instead of creating recursive hierarchies of events,
4971 * we link inherited events back to the original parent,
4972 * which has a filp for sure, which we use as the reference
4975 if (parent_event->parent)
4976 parent_event = parent_event->parent;
4978 child_event = perf_event_alloc(&parent_event->attr,
4979 parent_event->cpu, child_ctx,
4980 group_leader, parent_event,
4982 if (IS_ERR(child_event))
4987 * Make the child state follow the state of the parent event,
4988 * not its attr.disabled bit. We hold the parent's mutex,
4989 * so we won't race with perf_event_{en, dis}able_family.
4991 if (parent_event->state >= PERF_EVENT_STATE_INACTIVE)
4992 child_event->state = PERF_EVENT_STATE_INACTIVE;
4994 child_event->state = PERF_EVENT_STATE_OFF;
4996 if (parent_event->attr.freq) {
4997 u64 sample_period = parent_event->hw.sample_period;
4998 struct hw_perf_event *hwc = &child_event->hw;
5000 hwc->sample_period = sample_period;
5001 hwc->last_period = sample_period;
5003 atomic64_set(&hwc->period_left, sample_period);
5006 child_event->overflow_handler = parent_event->overflow_handler;
5009 * Link it up in the child's context:
5011 add_event_to_ctx(child_event, child_ctx);
5014 * Get a reference to the parent filp - we will fput it
5015 * when the child event exits. This is safe to do because
5016 * we are in the parent and we know that the filp still
5017 * exists and has a nonzero count:
5019 atomic_long_inc(&parent_event->filp->f_count);
5022 * Link this into the parent event's child list
5024 WARN_ON_ONCE(parent_event->ctx->parent_ctx);
5025 mutex_lock(&parent_event->child_mutex);
5026 list_add_tail(&child_event->child_list, &parent_event->child_list);
5027 mutex_unlock(&parent_event->child_mutex);
5032 static int inherit_group(struct perf_event *parent_event,
5033 struct task_struct *parent,
5034 struct perf_event_context *parent_ctx,
5035 struct task_struct *child,
5036 struct perf_event_context *child_ctx)
5038 struct perf_event *leader;
5039 struct perf_event *sub;
5040 struct perf_event *child_ctr;
5042 leader = inherit_event(parent_event, parent, parent_ctx,
5043 child, NULL, child_ctx);
5045 return PTR_ERR(leader);
5046 list_for_each_entry(sub, &parent_event->sibling_list, group_entry) {
5047 child_ctr = inherit_event(sub, parent, parent_ctx,
5048 child, leader, child_ctx);
5049 if (IS_ERR(child_ctr))
5050 return PTR_ERR(child_ctr);
5055 static void sync_child_event(struct perf_event *child_event,
5056 struct task_struct *child)
5058 struct perf_event *parent_event = child_event->parent;
5061 if (child_event->attr.inherit_stat)
5062 perf_event_read_event(child_event, child);
5064 child_val = atomic64_read(&child_event->count);
5067 * Add back the child's count to the parent's count:
5069 atomic64_add(child_val, &parent_event->count);
5070 atomic64_add(child_event->total_time_enabled,
5071 &parent_event->child_total_time_enabled);
5072 atomic64_add(child_event->total_time_running,
5073 &parent_event->child_total_time_running);
5076 * Remove this event from the parent's list
5078 WARN_ON_ONCE(parent_event->ctx->parent_ctx);
5079 mutex_lock(&parent_event->child_mutex);
5080 list_del_init(&child_event->child_list);
5081 mutex_unlock(&parent_event->child_mutex);
5084 * Release the parent event, if this was the last
5087 fput(parent_event->filp);
5091 __perf_event_exit_task(struct perf_event *child_event,
5092 struct perf_event_context *child_ctx,
5093 struct task_struct *child)
5095 struct perf_event *parent_event;
5097 perf_event_remove_from_context(child_event);
5099 parent_event = child_event->parent;
5101 * It can happen that parent exits first, and has events
5102 * that are still around due to the child reference. These
5103 * events need to be zapped - but otherwise linger.
5106 sync_child_event(child_event, child);
5107 free_event(child_event);
5112 * When a child task exits, feed back event values to parent events.
5114 void perf_event_exit_task(struct task_struct *child)
5116 struct perf_event *child_event, *tmp;
5117 struct perf_event_context *child_ctx;
5118 unsigned long flags;
5120 if (likely(!child->perf_event_ctxp)) {
5121 perf_event_task(child, NULL, 0);
5125 local_irq_save(flags);
5127 * We can't reschedule here because interrupts are disabled,
5128 * and either child is current or it is a task that can't be
5129 * scheduled, so we are now safe from rescheduling changing
5132 child_ctx = child->perf_event_ctxp;
5133 __perf_event_task_sched_out(child_ctx);
5136 * Take the context lock here so that if find_get_context is
5137 * reading child->perf_event_ctxp, we wait until it has
5138 * incremented the context's refcount before we do put_ctx below.
5140 raw_spin_lock(&child_ctx->lock);
5141 child->perf_event_ctxp = NULL;
5143 * If this context is a clone; unclone it so it can't get
5144 * swapped to another process while we're removing all
5145 * the events from it.
5147 unclone_ctx(child_ctx);
5148 update_context_time(child_ctx);
5149 raw_spin_unlock_irqrestore(&child_ctx->lock, flags);
5152 * Report the task dead after unscheduling the events so that we
5153 * won't get any samples after PERF_RECORD_EXIT. We can however still
5154 * get a few PERF_RECORD_READ events.
5156 perf_event_task(child, child_ctx, 0);
5159 * We can recurse on the same lock type through:
5161 * __perf_event_exit_task()
5162 * sync_child_event()
5163 * fput(parent_event->filp)
5165 * mutex_lock(&ctx->mutex)
5167 * But since its the parent context it won't be the same instance.
5169 mutex_lock_nested(&child_ctx->mutex, SINGLE_DEPTH_NESTING);
5172 list_for_each_entry_safe(child_event, tmp, &child_ctx->pinned_groups,
5174 __perf_event_exit_task(child_event, child_ctx, child);
5176 list_for_each_entry_safe(child_event, tmp, &child_ctx->flexible_groups,
5178 __perf_event_exit_task(child_event, child_ctx, child);
5181 * If the last event was a group event, it will have appended all
5182 * its siblings to the list, but we obtained 'tmp' before that which
5183 * will still point to the list head terminating the iteration.
5185 if (!list_empty(&child_ctx->pinned_groups) ||
5186 !list_empty(&child_ctx->flexible_groups))
5189 mutex_unlock(&child_ctx->mutex);
5194 static void perf_free_event(struct perf_event *event,
5195 struct perf_event_context *ctx)
5197 struct perf_event *parent = event->parent;
5199 if (WARN_ON_ONCE(!parent))
5202 mutex_lock(&parent->child_mutex);
5203 list_del_init(&event->child_list);
5204 mutex_unlock(&parent->child_mutex);
5208 list_del_event(event, ctx);
5213 * free an unexposed, unused context as created by inheritance by
5214 * init_task below, used by fork() in case of fail.
5216 void perf_event_free_task(struct task_struct *task)
5218 struct perf_event_context *ctx = task->perf_event_ctxp;
5219 struct perf_event *event, *tmp;
5224 mutex_lock(&ctx->mutex);
5226 list_for_each_entry_safe(event, tmp, &ctx->pinned_groups, group_entry)
5227 perf_free_event(event, ctx);
5229 list_for_each_entry_safe(event, tmp, &ctx->flexible_groups,
5231 perf_free_event(event, ctx);
5233 if (!list_empty(&ctx->pinned_groups) ||
5234 !list_empty(&ctx->flexible_groups))
5237 mutex_unlock(&ctx->mutex);
5243 inherit_task_group(struct perf_event *event, struct task_struct *parent,
5244 struct perf_event_context *parent_ctx,
5245 struct task_struct *child,
5249 struct perf_event_context *child_ctx = child->perf_event_ctxp;
5251 if (!event->attr.inherit) {
5258 * This is executed from the parent task context, so
5259 * inherit events that have been marked for cloning.
5260 * First allocate and initialize a context for the
5264 child_ctx = kzalloc(sizeof(struct perf_event_context),
5269 __perf_event_init_context(child_ctx, child);
5270 child->perf_event_ctxp = child_ctx;
5271 get_task_struct(child);
5274 ret = inherit_group(event, parent, parent_ctx,
5285 * Initialize the perf_event context in task_struct
5287 int perf_event_init_task(struct task_struct *child)
5289 struct perf_event_context *child_ctx, *parent_ctx;
5290 struct perf_event_context *cloned_ctx;
5291 struct perf_event *event;
5292 struct task_struct *parent = current;
5293 int inherited_all = 1;
5296 child->perf_event_ctxp = NULL;
5298 mutex_init(&child->perf_event_mutex);
5299 INIT_LIST_HEAD(&child->perf_event_list);
5301 if (likely(!parent->perf_event_ctxp))
5305 * If the parent's context is a clone, pin it so it won't get
5308 parent_ctx = perf_pin_task_context(parent);
5311 * No need to check if parent_ctx != NULL here; since we saw
5312 * it non-NULL earlier, the only reason for it to become NULL
5313 * is if we exit, and since we're currently in the middle of
5314 * a fork we can't be exiting at the same time.
5318 * Lock the parent list. No need to lock the child - not PID
5319 * hashed yet and not running, so nobody can access it.
5321 mutex_lock(&parent_ctx->mutex);
5324 * We dont have to disable NMIs - we are only looking at
5325 * the list, not manipulating it:
5327 list_for_each_entry(event, &parent_ctx->pinned_groups, group_entry) {
5328 ret = inherit_task_group(event, parent, parent_ctx, child,
5334 list_for_each_entry(event, &parent_ctx->flexible_groups, group_entry) {
5335 ret = inherit_task_group(event, parent, parent_ctx, child,
5341 child_ctx = child->perf_event_ctxp;
5343 if (child_ctx && inherited_all) {
5345 * Mark the child context as a clone of the parent
5346 * context, or of whatever the parent is a clone of.
5347 * Note that if the parent is a clone, it could get
5348 * uncloned at any point, but that doesn't matter
5349 * because the list of events and the generation
5350 * count can't have changed since we took the mutex.
5352 cloned_ctx = rcu_dereference(parent_ctx->parent_ctx);
5354 child_ctx->parent_ctx = cloned_ctx;
5355 child_ctx->parent_gen = parent_ctx->parent_gen;
5357 child_ctx->parent_ctx = parent_ctx;
5358 child_ctx->parent_gen = parent_ctx->generation;
5360 get_ctx(child_ctx->parent_ctx);
5363 mutex_unlock(&parent_ctx->mutex);
5365 perf_unpin_context(parent_ctx);
5370 static void __cpuinit perf_event_init_cpu(int cpu)
5372 struct perf_cpu_context *cpuctx;
5374 cpuctx = &per_cpu(perf_cpu_context, cpu);
5375 __perf_event_init_context(&cpuctx->ctx, NULL);
5377 spin_lock(&perf_resource_lock);
5378 cpuctx->max_pertask = perf_max_events - perf_reserved_percpu;
5379 spin_unlock(&perf_resource_lock);
5382 #ifdef CONFIG_HOTPLUG_CPU
5383 static void __perf_event_exit_cpu(void *info)
5385 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
5386 struct perf_event_context *ctx = &cpuctx->ctx;
5387 struct perf_event *event, *tmp;
5389 list_for_each_entry_safe(event, tmp, &ctx->pinned_groups, group_entry)
5390 __perf_event_remove_from_context(event);
5391 list_for_each_entry_safe(event, tmp, &ctx->flexible_groups, group_entry)
5392 __perf_event_remove_from_context(event);
5394 static void perf_event_exit_cpu(int cpu)
5396 struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
5397 struct perf_event_context *ctx = &cpuctx->ctx;
5399 mutex_lock(&ctx->mutex);
5400 smp_call_function_single(cpu, __perf_event_exit_cpu, NULL, 1);
5401 mutex_unlock(&ctx->mutex);
5404 static inline void perf_event_exit_cpu(int cpu) { }
5407 static int __cpuinit
5408 perf_cpu_notify(struct notifier_block *self, unsigned long action, void *hcpu)
5410 unsigned int cpu = (long)hcpu;
5414 case CPU_UP_PREPARE:
5415 case CPU_UP_PREPARE_FROZEN:
5416 perf_event_init_cpu(cpu);
5419 case CPU_DOWN_PREPARE:
5420 case CPU_DOWN_PREPARE_FROZEN:
5421 perf_event_exit_cpu(cpu);
5432 * This has to have a higher priority than migration_notifier in sched.c.
5434 static struct notifier_block __cpuinitdata perf_cpu_nb = {
5435 .notifier_call = perf_cpu_notify,
5439 void __init perf_event_init(void)
5441 perf_cpu_notify(&perf_cpu_nb, (unsigned long)CPU_UP_PREPARE,
5442 (void *)(long)smp_processor_id());
5443 perf_cpu_notify(&perf_cpu_nb, (unsigned long)CPU_ONLINE,
5444 (void *)(long)smp_processor_id());
5445 register_cpu_notifier(&perf_cpu_nb);
5448 static ssize_t perf_show_reserve_percpu(struct sysdev_class *class,
5449 struct sysdev_class_attribute *attr,
5452 return sprintf(buf, "%d\n", perf_reserved_percpu);
5456 perf_set_reserve_percpu(struct sysdev_class *class,
5457 struct sysdev_class_attribute *attr,
5461 struct perf_cpu_context *cpuctx;
5465 err = strict_strtoul(buf, 10, &val);
5468 if (val > perf_max_events)
5471 spin_lock(&perf_resource_lock);
5472 perf_reserved_percpu = val;
5473 for_each_online_cpu(cpu) {
5474 cpuctx = &per_cpu(perf_cpu_context, cpu);
5475 raw_spin_lock_irq(&cpuctx->ctx.lock);
5476 mpt = min(perf_max_events - cpuctx->ctx.nr_events,
5477 perf_max_events - perf_reserved_percpu);
5478 cpuctx->max_pertask = mpt;
5479 raw_spin_unlock_irq(&cpuctx->ctx.lock);
5481 spin_unlock(&perf_resource_lock);
5486 static ssize_t perf_show_overcommit(struct sysdev_class *class,
5487 struct sysdev_class_attribute *attr,
5490 return sprintf(buf, "%d\n", perf_overcommit);
5494 perf_set_overcommit(struct sysdev_class *class,
5495 struct sysdev_class_attribute *attr,
5496 const char *buf, size_t count)
5501 err = strict_strtoul(buf, 10, &val);
5507 spin_lock(&perf_resource_lock);
5508 perf_overcommit = val;
5509 spin_unlock(&perf_resource_lock);
5514 static SYSDEV_CLASS_ATTR(
5517 perf_show_reserve_percpu,
5518 perf_set_reserve_percpu
5521 static SYSDEV_CLASS_ATTR(
5524 perf_show_overcommit,
5528 static struct attribute *perfclass_attrs[] = {
5529 &attr_reserve_percpu.attr,
5530 &attr_overcommit.attr,
5534 static struct attribute_group perfclass_attr_group = {
5535 .attrs = perfclass_attrs,
5536 .name = "perf_events",
5539 static int __init perf_event_sysfs_init(void)
5541 return sysfs_create_group(&cpu_sysdev_class.kset.kobj,
5542 &perfclass_attr_group);
5544 device_initcall(perf_event_sysfs_init);