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 __get_cpu_var(perf_disable_count)++;
101 bool __perf_enable(void)
103 return !--__get_cpu_var(perf_disable_count);
106 void perf_disable(void)
112 void perf_enable(void)
118 static void get_ctx(struct perf_event_context *ctx)
120 WARN_ON(!atomic_inc_not_zero(&ctx->refcount));
123 static void free_ctx(struct rcu_head *head)
125 struct perf_event_context *ctx;
127 ctx = container_of(head, struct perf_event_context, rcu_head);
131 static void put_ctx(struct perf_event_context *ctx)
133 if (atomic_dec_and_test(&ctx->refcount)) {
135 put_ctx(ctx->parent_ctx);
137 put_task_struct(ctx->task);
138 call_rcu(&ctx->rcu_head, free_ctx);
142 static void unclone_ctx(struct perf_event_context *ctx)
144 if (ctx->parent_ctx) {
145 put_ctx(ctx->parent_ctx);
146 ctx->parent_ctx = NULL;
151 * If we inherit events we want to return the parent event id
154 static u64 primary_event_id(struct perf_event *event)
159 id = event->parent->id;
165 * Get the perf_event_context for a task and lock it.
166 * This has to cope with with the fact that until it is locked,
167 * the context could get moved to another task.
169 static struct perf_event_context *
170 perf_lock_task_context(struct task_struct *task, unsigned long *flags)
172 struct perf_event_context *ctx;
176 ctx = rcu_dereference(task->perf_event_ctxp);
179 * If this context is a clone of another, it might
180 * get swapped for another underneath us by
181 * perf_event_task_sched_out, though the
182 * rcu_read_lock() protects us from any context
183 * getting freed. Lock the context and check if it
184 * got swapped before we could get the lock, and retry
185 * if so. If we locked the right context, then it
186 * can't get swapped on us any more.
188 raw_spin_lock_irqsave(&ctx->lock, *flags);
189 if (ctx != rcu_dereference(task->perf_event_ctxp)) {
190 raw_spin_unlock_irqrestore(&ctx->lock, *flags);
194 if (!atomic_inc_not_zero(&ctx->refcount)) {
195 raw_spin_unlock_irqrestore(&ctx->lock, *flags);
204 * Get the context for a task and increment its pin_count so it
205 * can't get swapped to another task. This also increments its
206 * reference count so that the context can't get freed.
208 static struct perf_event_context *perf_pin_task_context(struct task_struct *task)
210 struct perf_event_context *ctx;
213 ctx = perf_lock_task_context(task, &flags);
216 raw_spin_unlock_irqrestore(&ctx->lock, flags);
221 static void perf_unpin_context(struct perf_event_context *ctx)
225 raw_spin_lock_irqsave(&ctx->lock, flags);
227 raw_spin_unlock_irqrestore(&ctx->lock, flags);
231 static inline u64 perf_clock(void)
233 return cpu_clock(raw_smp_processor_id());
237 * Update the record of the current time in a context.
239 static void update_context_time(struct perf_event_context *ctx)
241 u64 now = perf_clock();
243 ctx->time += now - ctx->timestamp;
244 ctx->timestamp = now;
248 * Update the total_time_enabled and total_time_running fields for a event.
250 static void update_event_times(struct perf_event *event)
252 struct perf_event_context *ctx = event->ctx;
255 if (event->state < PERF_EVENT_STATE_INACTIVE ||
256 event->group_leader->state < PERF_EVENT_STATE_INACTIVE)
262 run_end = event->tstamp_stopped;
264 event->total_time_enabled = run_end - event->tstamp_enabled;
266 if (event->state == PERF_EVENT_STATE_INACTIVE)
267 run_end = event->tstamp_stopped;
271 event->total_time_running = run_end - event->tstamp_running;
274 static struct list_head *
275 ctx_group_list(struct perf_event *event, struct perf_event_context *ctx)
277 if (event->attr.pinned)
278 return &ctx->pinned_groups;
280 return &ctx->flexible_groups;
284 * Add a event from the lists for its context.
285 * Must be called with ctx->mutex and ctx->lock held.
288 list_add_event(struct perf_event *event, struct perf_event_context *ctx)
290 struct perf_event *group_leader = event->group_leader;
293 * Depending on whether it is a standalone or sibling event,
294 * add it straight to the context's event list, or to the group
295 * leader's sibling list:
297 if (group_leader == event) {
298 struct list_head *list;
300 if (is_software_event(event))
301 event->group_flags |= PERF_GROUP_SOFTWARE;
303 list = ctx_group_list(event, ctx);
304 list_add_tail(&event->group_entry, list);
306 if (group_leader->group_flags & PERF_GROUP_SOFTWARE &&
307 !is_software_event(event))
308 group_leader->group_flags &= ~PERF_GROUP_SOFTWARE;
310 list_add_tail(&event->group_entry, &group_leader->sibling_list);
311 group_leader->nr_siblings++;
314 list_add_rcu(&event->event_entry, &ctx->event_list);
316 if (event->attr.inherit_stat)
321 * Remove a event from the lists for its context.
322 * Must be called with ctx->mutex and ctx->lock held.
325 list_del_event(struct perf_event *event, struct perf_event_context *ctx)
327 struct perf_event *sibling, *tmp;
329 if (list_empty(&event->group_entry))
332 if (event->attr.inherit_stat)
335 list_del_init(&event->group_entry);
336 list_del_rcu(&event->event_entry);
338 if (event->group_leader != event)
339 event->group_leader->nr_siblings--;
341 update_event_times(event);
344 * If event was in error state, then keep it
345 * that way, otherwise bogus counts will be
346 * returned on read(). The only way to get out
347 * of error state is by explicit re-enabling
350 if (event->state > PERF_EVENT_STATE_OFF)
351 event->state = PERF_EVENT_STATE_OFF;
354 * If this was a group event with sibling events then
355 * upgrade the siblings to singleton events by adding them
356 * to the context list directly:
358 list_for_each_entry_safe(sibling, tmp, &event->sibling_list, group_entry) {
359 struct list_head *list;
361 list = ctx_group_list(event, ctx);
362 list_move_tail(&sibling->group_entry, list);
363 sibling->group_leader = sibling;
365 /* Inherit group flags from the previous leader */
366 sibling->group_flags = event->group_flags;
371 event_sched_out(struct perf_event *event,
372 struct perf_cpu_context *cpuctx,
373 struct perf_event_context *ctx)
375 if (event->state != PERF_EVENT_STATE_ACTIVE)
378 event->state = PERF_EVENT_STATE_INACTIVE;
379 if (event->pending_disable) {
380 event->pending_disable = 0;
381 event->state = PERF_EVENT_STATE_OFF;
383 event->tstamp_stopped = ctx->time;
384 event->pmu->disable(event);
387 if (!is_software_event(event))
388 cpuctx->active_oncpu--;
390 if (event->attr.exclusive || !cpuctx->active_oncpu)
391 cpuctx->exclusive = 0;
395 group_sched_out(struct perf_event *group_event,
396 struct perf_cpu_context *cpuctx,
397 struct perf_event_context *ctx)
399 struct perf_event *event;
401 if (group_event->state != PERF_EVENT_STATE_ACTIVE)
404 event_sched_out(group_event, cpuctx, ctx);
407 * Schedule out siblings (if any):
409 list_for_each_entry(event, &group_event->sibling_list, group_entry)
410 event_sched_out(event, cpuctx, ctx);
412 if (group_event->attr.exclusive)
413 cpuctx->exclusive = 0;
417 * Cross CPU call to remove a performance event
419 * We disable the event on the hardware level first. After that we
420 * remove it from the context list.
422 static void __perf_event_remove_from_context(void *info)
424 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
425 struct perf_event *event = info;
426 struct perf_event_context *ctx = event->ctx;
429 * If this is a task context, we need to check whether it is
430 * the current task context of this cpu. If not it has been
431 * scheduled out before the smp call arrived.
433 if (ctx->task && cpuctx->task_ctx != ctx)
436 raw_spin_lock(&ctx->lock);
438 * Protect the list operation against NMI by disabling the
439 * events on a global level.
443 event_sched_out(event, cpuctx, ctx);
445 list_del_event(event, ctx);
449 * Allow more per task events with respect to the
452 cpuctx->max_pertask =
453 min(perf_max_events - ctx->nr_events,
454 perf_max_events - perf_reserved_percpu);
458 raw_spin_unlock(&ctx->lock);
463 * Remove the event from a task's (or a CPU's) list of events.
465 * Must be called with ctx->mutex held.
467 * CPU events are removed with a smp call. For task events we only
468 * call when the task is on a CPU.
470 * If event->ctx is a cloned context, callers must make sure that
471 * every task struct that event->ctx->task could possibly point to
472 * remains valid. This is OK when called from perf_release since
473 * that only calls us on the top-level context, which can't be a clone.
474 * When called from perf_event_exit_task, it's OK because the
475 * context has been detached from its task.
477 static void perf_event_remove_from_context(struct perf_event *event)
479 struct perf_event_context *ctx = event->ctx;
480 struct task_struct *task = ctx->task;
484 * Per cpu events are removed via an smp call and
485 * the removal is always successful.
487 smp_call_function_single(event->cpu,
488 __perf_event_remove_from_context,
494 task_oncpu_function_call(task, __perf_event_remove_from_context,
497 raw_spin_lock_irq(&ctx->lock);
499 * If the context is active we need to retry the smp call.
501 if (ctx->nr_active && !list_empty(&event->group_entry)) {
502 raw_spin_unlock_irq(&ctx->lock);
507 * The lock prevents that this context is scheduled in so we
508 * can remove the event safely, if the call above did not
511 if (!list_empty(&event->group_entry))
512 list_del_event(event, ctx);
513 raw_spin_unlock_irq(&ctx->lock);
517 * Update total_time_enabled and total_time_running for all events in a group.
519 static void update_group_times(struct perf_event *leader)
521 struct perf_event *event;
523 update_event_times(leader);
524 list_for_each_entry(event, &leader->sibling_list, group_entry)
525 update_event_times(event);
529 * Cross CPU call to disable a performance event
531 static void __perf_event_disable(void *info)
533 struct perf_event *event = info;
534 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
535 struct perf_event_context *ctx = event->ctx;
538 * If this is a per-task event, need to check whether this
539 * event's task is the current task on this cpu.
541 if (ctx->task && cpuctx->task_ctx != ctx)
544 raw_spin_lock(&ctx->lock);
547 * If the event is on, turn it off.
548 * If it is in error state, leave it in error state.
550 if (event->state >= PERF_EVENT_STATE_INACTIVE) {
551 update_context_time(ctx);
552 update_group_times(event);
553 if (event == event->group_leader)
554 group_sched_out(event, cpuctx, ctx);
556 event_sched_out(event, cpuctx, ctx);
557 event->state = PERF_EVENT_STATE_OFF;
560 raw_spin_unlock(&ctx->lock);
566 * If event->ctx is a cloned context, callers must make sure that
567 * every task struct that event->ctx->task could possibly point to
568 * remains valid. This condition is satisifed when called through
569 * perf_event_for_each_child or perf_event_for_each because they
570 * hold the top-level event's child_mutex, so any descendant that
571 * goes to exit will block in sync_child_event.
572 * When called from perf_pending_event it's OK because event->ctx
573 * is the current context on this CPU and preemption is disabled,
574 * hence we can't get into perf_event_task_sched_out for this context.
576 void perf_event_disable(struct perf_event *event)
578 struct perf_event_context *ctx = event->ctx;
579 struct task_struct *task = ctx->task;
583 * Disable the event on the cpu that it's on
585 smp_call_function_single(event->cpu, __perf_event_disable,
591 task_oncpu_function_call(task, __perf_event_disable, event);
593 raw_spin_lock_irq(&ctx->lock);
595 * If the event is still active, we need to retry the cross-call.
597 if (event->state == PERF_EVENT_STATE_ACTIVE) {
598 raw_spin_unlock_irq(&ctx->lock);
603 * Since we have the lock this context can't be scheduled
604 * in, so we can change the state safely.
606 if (event->state == PERF_EVENT_STATE_INACTIVE) {
607 update_group_times(event);
608 event->state = PERF_EVENT_STATE_OFF;
611 raw_spin_unlock_irq(&ctx->lock);
615 event_sched_in(struct perf_event *event,
616 struct perf_cpu_context *cpuctx,
617 struct perf_event_context *ctx)
619 if (event->state <= PERF_EVENT_STATE_OFF)
622 event->state = PERF_EVENT_STATE_ACTIVE;
623 event->oncpu = smp_processor_id();
625 * The new state must be visible before we turn it on in the hardware:
629 if (event->pmu->enable(event)) {
630 event->state = PERF_EVENT_STATE_INACTIVE;
635 event->tstamp_running += ctx->time - event->tstamp_stopped;
637 if (!is_software_event(event))
638 cpuctx->active_oncpu++;
641 if (event->attr.exclusive)
642 cpuctx->exclusive = 1;
648 group_sched_in(struct perf_event *group_event,
649 struct perf_cpu_context *cpuctx,
650 struct perf_event_context *ctx)
652 struct perf_event *event, *partial_group;
655 if (group_event->state == PERF_EVENT_STATE_OFF)
658 ret = hw_perf_group_sched_in(group_event, cpuctx, ctx);
660 return ret < 0 ? ret : 0;
662 if (event_sched_in(group_event, cpuctx, ctx))
666 * Schedule in siblings as one group (if any):
668 list_for_each_entry(event, &group_event->sibling_list, group_entry) {
669 if (event_sched_in(event, cpuctx, ctx)) {
670 partial_group = event;
679 * Groups can be scheduled in as one unit only, so undo any
680 * partial group before returning:
682 list_for_each_entry(event, &group_event->sibling_list, group_entry) {
683 if (event == partial_group)
685 event_sched_out(event, cpuctx, ctx);
687 event_sched_out(group_event, cpuctx, ctx);
693 * Work out whether we can put this event group on the CPU now.
695 static int group_can_go_on(struct perf_event *event,
696 struct perf_cpu_context *cpuctx,
700 * Groups consisting entirely of software events can always go on.
702 if (event->group_flags & PERF_GROUP_SOFTWARE)
705 * If an exclusive group is already on, no other hardware
708 if (cpuctx->exclusive)
711 * If this group is exclusive and there are already
712 * events on the CPU, it can't go on.
714 if (event->attr.exclusive && cpuctx->active_oncpu)
717 * Otherwise, try to add it if all previous groups were able
723 static void add_event_to_ctx(struct perf_event *event,
724 struct perf_event_context *ctx)
726 list_add_event(event, ctx);
727 event->tstamp_enabled = ctx->time;
728 event->tstamp_running = ctx->time;
729 event->tstamp_stopped = ctx->time;
733 * Cross CPU call to install and enable a performance event
735 * Must be called with ctx->mutex held
737 static void __perf_install_in_context(void *info)
739 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
740 struct perf_event *event = info;
741 struct perf_event_context *ctx = event->ctx;
742 struct perf_event *leader = event->group_leader;
746 * If this is a task context, we need to check whether it is
747 * the current task context of this cpu. If not it has been
748 * scheduled out before the smp call arrived.
749 * Or possibly this is the right context but it isn't
750 * on this cpu because it had no events.
752 if (ctx->task && cpuctx->task_ctx != ctx) {
753 if (cpuctx->task_ctx || ctx->task != current)
755 cpuctx->task_ctx = ctx;
758 raw_spin_lock(&ctx->lock);
760 update_context_time(ctx);
763 * Protect the list operation against NMI by disabling the
764 * events on a global level. NOP for non NMI based events.
768 add_event_to_ctx(event, ctx);
770 if (event->cpu != -1 && event->cpu != smp_processor_id())
774 * Don't put the event on if it is disabled or if
775 * it is in a group and the group isn't on.
777 if (event->state != PERF_EVENT_STATE_INACTIVE ||
778 (leader != event && leader->state != PERF_EVENT_STATE_ACTIVE))
782 * An exclusive event can't go on if there are already active
783 * hardware events, and no hardware event can go on if there
784 * is already an exclusive event on.
786 if (!group_can_go_on(event, cpuctx, 1))
789 err = event_sched_in(event, cpuctx, ctx);
793 * This event couldn't go on. If it is in a group
794 * then we have to pull the whole group off.
795 * If the event group is pinned then put it in error state.
798 group_sched_out(leader, cpuctx, ctx);
799 if (leader->attr.pinned) {
800 update_group_times(leader);
801 leader->state = PERF_EVENT_STATE_ERROR;
805 if (!err && !ctx->task && cpuctx->max_pertask)
806 cpuctx->max_pertask--;
811 raw_spin_unlock(&ctx->lock);
815 * Attach a performance event to a context
817 * First we add the event to the list with the hardware enable bit
818 * in event->hw_config cleared.
820 * If the event is attached to a task which is on a CPU we use a smp
821 * call to enable it in the task context. The task might have been
822 * scheduled away, but we check this in the smp call again.
824 * Must be called with ctx->mutex held.
827 perf_install_in_context(struct perf_event_context *ctx,
828 struct perf_event *event,
831 struct task_struct *task = ctx->task;
835 * Per cpu events are installed via an smp call and
836 * the install is always successful.
838 smp_call_function_single(cpu, __perf_install_in_context,
844 task_oncpu_function_call(task, __perf_install_in_context,
847 raw_spin_lock_irq(&ctx->lock);
849 * we need to retry the smp call.
851 if (ctx->is_active && list_empty(&event->group_entry)) {
852 raw_spin_unlock_irq(&ctx->lock);
857 * The lock prevents that this context is scheduled in so we
858 * can add the event safely, if it the call above did not
861 if (list_empty(&event->group_entry))
862 add_event_to_ctx(event, ctx);
863 raw_spin_unlock_irq(&ctx->lock);
867 * Put a event into inactive state and update time fields.
868 * Enabling the leader of a group effectively enables all
869 * the group members that aren't explicitly disabled, so we
870 * have to update their ->tstamp_enabled also.
871 * Note: this works for group members as well as group leaders
872 * since the non-leader members' sibling_lists will be empty.
874 static void __perf_event_mark_enabled(struct perf_event *event,
875 struct perf_event_context *ctx)
877 struct perf_event *sub;
879 event->state = PERF_EVENT_STATE_INACTIVE;
880 event->tstamp_enabled = ctx->time - event->total_time_enabled;
881 list_for_each_entry(sub, &event->sibling_list, group_entry)
882 if (sub->state >= PERF_EVENT_STATE_INACTIVE)
883 sub->tstamp_enabled =
884 ctx->time - sub->total_time_enabled;
888 * Cross CPU call to enable a performance event
890 static void __perf_event_enable(void *info)
892 struct perf_event *event = info;
893 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
894 struct perf_event_context *ctx = event->ctx;
895 struct perf_event *leader = event->group_leader;
899 * If this is a per-task event, need to check whether this
900 * event's task is the current task on this cpu.
902 if (ctx->task && cpuctx->task_ctx != ctx) {
903 if (cpuctx->task_ctx || ctx->task != current)
905 cpuctx->task_ctx = ctx;
908 raw_spin_lock(&ctx->lock);
910 update_context_time(ctx);
912 if (event->state >= PERF_EVENT_STATE_INACTIVE)
914 __perf_event_mark_enabled(event, ctx);
916 if (event->cpu != -1 && event->cpu != smp_processor_id())
920 * If the event is in a group and isn't the group leader,
921 * then don't put it on unless the group is on.
923 if (leader != event && leader->state != PERF_EVENT_STATE_ACTIVE)
926 if (!group_can_go_on(event, cpuctx, 1)) {
931 err = group_sched_in(event, cpuctx, ctx);
933 err = event_sched_in(event, cpuctx, ctx);
939 * If this event can't go on and it's part of a
940 * group, then the whole group has to come off.
943 group_sched_out(leader, cpuctx, ctx);
944 if (leader->attr.pinned) {
945 update_group_times(leader);
946 leader->state = PERF_EVENT_STATE_ERROR;
951 raw_spin_unlock(&ctx->lock);
957 * If event->ctx is a cloned context, callers must make sure that
958 * every task struct that event->ctx->task could possibly point to
959 * remains valid. This condition is satisfied when called through
960 * perf_event_for_each_child or perf_event_for_each as described
961 * for perf_event_disable.
963 void perf_event_enable(struct perf_event *event)
965 struct perf_event_context *ctx = event->ctx;
966 struct task_struct *task = ctx->task;
970 * Enable the event on the cpu that it's on
972 smp_call_function_single(event->cpu, __perf_event_enable,
977 raw_spin_lock_irq(&ctx->lock);
978 if (event->state >= PERF_EVENT_STATE_INACTIVE)
982 * If the event is in error state, clear that first.
983 * That way, if we see the event in error state below, we
984 * know that it has gone back into error state, as distinct
985 * from the task having been scheduled away before the
986 * cross-call arrived.
988 if (event->state == PERF_EVENT_STATE_ERROR)
989 event->state = PERF_EVENT_STATE_OFF;
992 raw_spin_unlock_irq(&ctx->lock);
993 task_oncpu_function_call(task, __perf_event_enable, event);
995 raw_spin_lock_irq(&ctx->lock);
998 * If the context is active and the event is still off,
999 * we need to retry the cross-call.
1001 if (ctx->is_active && event->state == PERF_EVENT_STATE_OFF)
1005 * Since we have the lock this context can't be scheduled
1006 * in, so we can change the state safely.
1008 if (event->state == PERF_EVENT_STATE_OFF)
1009 __perf_event_mark_enabled(event, ctx);
1012 raw_spin_unlock_irq(&ctx->lock);
1015 static int perf_event_refresh(struct perf_event *event, int refresh)
1018 * not supported on inherited events
1020 if (event->attr.inherit)
1023 atomic_add(refresh, &event->event_limit);
1024 perf_event_enable(event);
1030 EVENT_FLEXIBLE = 0x1,
1032 EVENT_ALL = EVENT_FLEXIBLE | EVENT_PINNED,
1035 static void ctx_sched_out(struct perf_event_context *ctx,
1036 struct perf_cpu_context *cpuctx,
1037 enum event_type_t event_type)
1039 struct perf_event *event;
1041 raw_spin_lock(&ctx->lock);
1043 if (likely(!ctx->nr_events))
1045 update_context_time(ctx);
1048 if (!ctx->nr_active)
1051 if (event_type & EVENT_PINNED)
1052 list_for_each_entry(event, &ctx->pinned_groups, group_entry)
1053 group_sched_out(event, cpuctx, ctx);
1055 if (event_type & EVENT_FLEXIBLE)
1056 list_for_each_entry(event, &ctx->flexible_groups, group_entry)
1057 group_sched_out(event, cpuctx, ctx);
1062 raw_spin_unlock(&ctx->lock);
1066 * Test whether two contexts are equivalent, i.e. whether they
1067 * have both been cloned from the same version of the same context
1068 * and they both have the same number of enabled events.
1069 * If the number of enabled events is the same, then the set
1070 * of enabled events should be the same, because these are both
1071 * inherited contexts, therefore we can't access individual events
1072 * in them directly with an fd; we can only enable/disable all
1073 * events via prctl, or enable/disable all events in a family
1074 * via ioctl, which will have the same effect on both contexts.
1076 static int context_equiv(struct perf_event_context *ctx1,
1077 struct perf_event_context *ctx2)
1079 return ctx1->parent_ctx && ctx1->parent_ctx == ctx2->parent_ctx
1080 && ctx1->parent_gen == ctx2->parent_gen
1081 && !ctx1->pin_count && !ctx2->pin_count;
1084 static void __perf_event_sync_stat(struct perf_event *event,
1085 struct perf_event *next_event)
1089 if (!event->attr.inherit_stat)
1093 * Update the event value, we cannot use perf_event_read()
1094 * because we're in the middle of a context switch and have IRQs
1095 * disabled, which upsets smp_call_function_single(), however
1096 * we know the event must be on the current CPU, therefore we
1097 * don't need to use it.
1099 switch (event->state) {
1100 case PERF_EVENT_STATE_ACTIVE:
1101 event->pmu->read(event);
1104 case PERF_EVENT_STATE_INACTIVE:
1105 update_event_times(event);
1113 * In order to keep per-task stats reliable we need to flip the event
1114 * values when we flip the contexts.
1116 value = atomic64_read(&next_event->count);
1117 value = atomic64_xchg(&event->count, value);
1118 atomic64_set(&next_event->count, value);
1120 swap(event->total_time_enabled, next_event->total_time_enabled);
1121 swap(event->total_time_running, next_event->total_time_running);
1124 * Since we swizzled the values, update the user visible data too.
1126 perf_event_update_userpage(event);
1127 perf_event_update_userpage(next_event);
1130 #define list_next_entry(pos, member) \
1131 list_entry(pos->member.next, typeof(*pos), member)
1133 static void perf_event_sync_stat(struct perf_event_context *ctx,
1134 struct perf_event_context *next_ctx)
1136 struct perf_event *event, *next_event;
1141 update_context_time(ctx);
1143 event = list_first_entry(&ctx->event_list,
1144 struct perf_event, event_entry);
1146 next_event = list_first_entry(&next_ctx->event_list,
1147 struct perf_event, event_entry);
1149 while (&event->event_entry != &ctx->event_list &&
1150 &next_event->event_entry != &next_ctx->event_list) {
1152 __perf_event_sync_stat(event, next_event);
1154 event = list_next_entry(event, event_entry);
1155 next_event = list_next_entry(next_event, event_entry);
1160 * Called from scheduler to remove the events of the current task,
1161 * with interrupts disabled.
1163 * We stop each event and update the event value in event->count.
1165 * This does not protect us against NMI, but disable()
1166 * sets the disabled bit in the control field of event _before_
1167 * accessing the event control register. If a NMI hits, then it will
1168 * not restart the event.
1170 void perf_event_task_sched_out(struct task_struct *task,
1171 struct task_struct *next)
1173 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
1174 struct perf_event_context *ctx = task->perf_event_ctxp;
1175 struct perf_event_context *next_ctx;
1176 struct perf_event_context *parent;
1177 struct pt_regs *regs;
1180 regs = task_pt_regs(task);
1181 perf_sw_event(PERF_COUNT_SW_CONTEXT_SWITCHES, 1, 1, regs, 0);
1183 if (likely(!ctx || !cpuctx->task_ctx))
1187 parent = rcu_dereference(ctx->parent_ctx);
1188 next_ctx = next->perf_event_ctxp;
1189 if (parent && next_ctx &&
1190 rcu_dereference(next_ctx->parent_ctx) == parent) {
1192 * Looks like the two contexts are clones, so we might be
1193 * able to optimize the context switch. We lock both
1194 * contexts and check that they are clones under the
1195 * lock (including re-checking that neither has been
1196 * uncloned in the meantime). It doesn't matter which
1197 * order we take the locks because no other cpu could
1198 * be trying to lock both of these tasks.
1200 raw_spin_lock(&ctx->lock);
1201 raw_spin_lock_nested(&next_ctx->lock, SINGLE_DEPTH_NESTING);
1202 if (context_equiv(ctx, next_ctx)) {
1204 * XXX do we need a memory barrier of sorts
1205 * wrt to rcu_dereference() of perf_event_ctxp
1207 task->perf_event_ctxp = next_ctx;
1208 next->perf_event_ctxp = ctx;
1210 next_ctx->task = task;
1213 perf_event_sync_stat(ctx, next_ctx);
1215 raw_spin_unlock(&next_ctx->lock);
1216 raw_spin_unlock(&ctx->lock);
1221 ctx_sched_out(ctx, cpuctx, EVENT_ALL);
1222 cpuctx->task_ctx = NULL;
1226 static void task_ctx_sched_out(struct perf_event_context *ctx,
1227 enum event_type_t event_type)
1229 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
1231 if (!cpuctx->task_ctx)
1234 if (WARN_ON_ONCE(ctx != cpuctx->task_ctx))
1237 ctx_sched_out(ctx, cpuctx, event_type);
1238 cpuctx->task_ctx = NULL;
1242 * Called with IRQs disabled
1244 static void __perf_event_task_sched_out(struct perf_event_context *ctx)
1246 task_ctx_sched_out(ctx, EVENT_ALL);
1250 * Called with IRQs disabled
1252 static void cpu_ctx_sched_out(struct perf_cpu_context *cpuctx,
1253 enum event_type_t event_type)
1255 ctx_sched_out(&cpuctx->ctx, cpuctx, event_type);
1259 ctx_pinned_sched_in(struct perf_event_context *ctx,
1260 struct perf_cpu_context *cpuctx)
1262 struct perf_event *event;
1264 list_for_each_entry(event, &ctx->pinned_groups, group_entry) {
1265 if (event->state <= PERF_EVENT_STATE_OFF)
1267 if (event->cpu != -1 && event->cpu != smp_processor_id())
1270 if (group_can_go_on(event, cpuctx, 1))
1271 group_sched_in(event, cpuctx, ctx);
1274 * If this pinned group hasn't been scheduled,
1275 * put it in error state.
1277 if (event->state == PERF_EVENT_STATE_INACTIVE) {
1278 update_group_times(event);
1279 event->state = PERF_EVENT_STATE_ERROR;
1285 ctx_flexible_sched_in(struct perf_event_context *ctx,
1286 struct perf_cpu_context *cpuctx)
1288 struct perf_event *event;
1291 list_for_each_entry(event, &ctx->flexible_groups, group_entry) {
1292 /* Ignore events in OFF or ERROR state */
1293 if (event->state <= PERF_EVENT_STATE_OFF)
1296 * Listen to the 'cpu' scheduling filter constraint
1299 if (event->cpu != -1 && event->cpu != smp_processor_id())
1302 if (group_can_go_on(event, cpuctx, can_add_hw))
1303 if (group_sched_in(event, cpuctx, ctx))
1309 ctx_sched_in(struct perf_event_context *ctx,
1310 struct perf_cpu_context *cpuctx,
1311 enum event_type_t event_type)
1313 raw_spin_lock(&ctx->lock);
1315 if (likely(!ctx->nr_events))
1318 ctx->timestamp = perf_clock();
1323 * First go through the list and put on any pinned groups
1324 * in order to give them the best chance of going on.
1326 if (event_type & EVENT_PINNED)
1327 ctx_pinned_sched_in(ctx, cpuctx);
1329 /* Then walk through the lower prio flexible groups */
1330 if (event_type & EVENT_FLEXIBLE)
1331 ctx_flexible_sched_in(ctx, cpuctx);
1335 raw_spin_unlock(&ctx->lock);
1338 static void cpu_ctx_sched_in(struct perf_cpu_context *cpuctx,
1339 enum event_type_t event_type)
1341 struct perf_event_context *ctx = &cpuctx->ctx;
1343 ctx_sched_in(ctx, cpuctx, event_type);
1346 static void task_ctx_sched_in(struct task_struct *task,
1347 enum event_type_t event_type)
1349 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
1350 struct perf_event_context *ctx = task->perf_event_ctxp;
1354 if (cpuctx->task_ctx == ctx)
1356 ctx_sched_in(ctx, cpuctx, event_type);
1357 cpuctx->task_ctx = ctx;
1360 * Called from scheduler to add the events of the current task
1361 * with interrupts disabled.
1363 * We restore the event value and then enable it.
1365 * This does not protect us against NMI, but enable()
1366 * sets the enabled bit in the control field of event _before_
1367 * accessing the event control register. If a NMI hits, then it will
1368 * keep the event running.
1370 void perf_event_task_sched_in(struct task_struct *task)
1372 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
1373 struct perf_event_context *ctx = task->perf_event_ctxp;
1378 if (cpuctx->task_ctx == ctx)
1382 * We want to keep the following priority order:
1383 * cpu pinned (that don't need to move), task pinned,
1384 * cpu flexible, task flexible.
1386 cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE);
1388 ctx_sched_in(ctx, cpuctx, EVENT_PINNED);
1389 cpu_ctx_sched_in(cpuctx, EVENT_FLEXIBLE);
1390 ctx_sched_in(ctx, cpuctx, EVENT_FLEXIBLE);
1392 cpuctx->task_ctx = ctx;
1395 #define MAX_INTERRUPTS (~0ULL)
1397 static void perf_log_throttle(struct perf_event *event, int enable);
1399 static u64 perf_calculate_period(struct perf_event *event, u64 nsec, u64 count)
1401 u64 frequency = event->attr.sample_freq;
1402 u64 sec = NSEC_PER_SEC;
1403 u64 divisor, dividend;
1405 int count_fls, nsec_fls, frequency_fls, sec_fls;
1407 count_fls = fls64(count);
1408 nsec_fls = fls64(nsec);
1409 frequency_fls = fls64(frequency);
1413 * We got @count in @nsec, with a target of sample_freq HZ
1414 * the target period becomes:
1417 * period = -------------------
1418 * @nsec * sample_freq
1423 * Reduce accuracy by one bit such that @a and @b converge
1424 * to a similar magnitude.
1426 #define REDUCE_FLS(a, b) \
1428 if (a##_fls > b##_fls) { \
1438 * Reduce accuracy until either term fits in a u64, then proceed with
1439 * the other, so that finally we can do a u64/u64 division.
1441 while (count_fls + sec_fls > 64 && nsec_fls + frequency_fls > 64) {
1442 REDUCE_FLS(nsec, frequency);
1443 REDUCE_FLS(sec, count);
1446 if (count_fls + sec_fls > 64) {
1447 divisor = nsec * frequency;
1449 while (count_fls + sec_fls > 64) {
1450 REDUCE_FLS(count, sec);
1454 dividend = count * sec;
1456 dividend = count * sec;
1458 while (nsec_fls + frequency_fls > 64) {
1459 REDUCE_FLS(nsec, frequency);
1463 divisor = nsec * frequency;
1466 return div64_u64(dividend, divisor);
1469 static void perf_event_stop(struct perf_event *event)
1471 if (!event->pmu->stop)
1472 return event->pmu->disable(event);
1474 return event->pmu->stop(event);
1477 static int perf_event_start(struct perf_event *event)
1479 if (!event->pmu->start)
1480 return event->pmu->enable(event);
1482 return event->pmu->start(event);
1485 static void perf_adjust_period(struct perf_event *event, u64 nsec, u64 count)
1487 struct hw_perf_event *hwc = &event->hw;
1488 u64 period, sample_period;
1491 period = perf_calculate_period(event, nsec, count);
1493 delta = (s64)(period - hwc->sample_period);
1494 delta = (delta + 7) / 8; /* low pass filter */
1496 sample_period = hwc->sample_period + delta;
1501 hwc->sample_period = sample_period;
1503 if (atomic64_read(&hwc->period_left) > 8*sample_period) {
1505 perf_event_stop(event);
1506 atomic64_set(&hwc->period_left, 0);
1507 perf_event_start(event);
1512 static void perf_ctx_adjust_freq(struct perf_event_context *ctx)
1514 struct perf_event *event;
1515 struct hw_perf_event *hwc;
1516 u64 interrupts, now;
1519 raw_spin_lock(&ctx->lock);
1520 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
1521 if (event->state != PERF_EVENT_STATE_ACTIVE)
1524 if (event->cpu != -1 && event->cpu != smp_processor_id())
1529 interrupts = hwc->interrupts;
1530 hwc->interrupts = 0;
1533 * unthrottle events on the tick
1535 if (interrupts == MAX_INTERRUPTS) {
1536 perf_log_throttle(event, 1);
1537 event->pmu->unthrottle(event);
1540 if (!event->attr.freq || !event->attr.sample_freq)
1543 event->pmu->read(event);
1544 now = atomic64_read(&event->count);
1545 delta = now - hwc->freq_count_stamp;
1546 hwc->freq_count_stamp = now;
1549 perf_adjust_period(event, TICK_NSEC, delta);
1551 raw_spin_unlock(&ctx->lock);
1555 * Round-robin a context's events:
1557 static void rotate_ctx(struct perf_event_context *ctx)
1559 if (!ctx->nr_events)
1562 raw_spin_lock(&ctx->lock);
1564 /* Rotate the first entry last of non-pinned groups */
1565 list_rotate_left(&ctx->flexible_groups);
1567 raw_spin_unlock(&ctx->lock);
1570 void perf_event_task_tick(struct task_struct *curr)
1572 struct perf_cpu_context *cpuctx;
1573 struct perf_event_context *ctx;
1575 if (!atomic_read(&nr_events))
1578 cpuctx = &__get_cpu_var(perf_cpu_context);
1579 ctx = curr->perf_event_ctxp;
1583 perf_ctx_adjust_freq(&cpuctx->ctx);
1585 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);
1602 static int event_enable_on_exec(struct perf_event *event,
1603 struct perf_event_context *ctx)
1605 if (!event->attr.enable_on_exec)
1608 event->attr.enable_on_exec = 0;
1609 if (event->state >= PERF_EVENT_STATE_INACTIVE)
1612 __perf_event_mark_enabled(event, ctx);
1618 * Enable all of a task's events that have been marked enable-on-exec.
1619 * This expects task == current.
1621 static void perf_event_enable_on_exec(struct task_struct *task)
1623 struct perf_event_context *ctx;
1624 struct perf_event *event;
1625 unsigned long flags;
1629 local_irq_save(flags);
1630 ctx = task->perf_event_ctxp;
1631 if (!ctx || !ctx->nr_events)
1634 __perf_event_task_sched_out(ctx);
1636 raw_spin_lock(&ctx->lock);
1638 list_for_each_entry(event, &ctx->pinned_groups, group_entry) {
1639 ret = event_enable_on_exec(event, ctx);
1644 list_for_each_entry(event, &ctx->flexible_groups, group_entry) {
1645 ret = event_enable_on_exec(event, ctx);
1651 * Unclone this context if we enabled any event.
1656 raw_spin_unlock(&ctx->lock);
1658 perf_event_task_sched_in(task);
1660 local_irq_restore(flags);
1664 * Cross CPU call to read the hardware event
1666 static void __perf_event_read(void *info)
1668 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
1669 struct perf_event *event = info;
1670 struct perf_event_context *ctx = event->ctx;
1673 * If this is a task context, we need to check whether it is
1674 * the current task context of this cpu. If not it has been
1675 * scheduled out before the smp call arrived. In that case
1676 * event->count would have been updated to a recent sample
1677 * when the event was scheduled out.
1679 if (ctx->task && cpuctx->task_ctx != ctx)
1682 raw_spin_lock(&ctx->lock);
1683 update_context_time(ctx);
1684 update_event_times(event);
1685 raw_spin_unlock(&ctx->lock);
1687 event->pmu->read(event);
1690 static u64 perf_event_read(struct perf_event *event)
1693 * If event is enabled and currently active on a CPU, update the
1694 * value in the event structure:
1696 if (event->state == PERF_EVENT_STATE_ACTIVE) {
1697 smp_call_function_single(event->oncpu,
1698 __perf_event_read, event, 1);
1699 } else if (event->state == PERF_EVENT_STATE_INACTIVE) {
1700 struct perf_event_context *ctx = event->ctx;
1701 unsigned long flags;
1703 raw_spin_lock_irqsave(&ctx->lock, flags);
1704 update_context_time(ctx);
1705 update_event_times(event);
1706 raw_spin_unlock_irqrestore(&ctx->lock, flags);
1709 return atomic64_read(&event->count);
1713 * Initialize the perf_event context in a task_struct:
1716 __perf_event_init_context(struct perf_event_context *ctx,
1717 struct task_struct *task)
1719 raw_spin_lock_init(&ctx->lock);
1720 mutex_init(&ctx->mutex);
1721 INIT_LIST_HEAD(&ctx->pinned_groups);
1722 INIT_LIST_HEAD(&ctx->flexible_groups);
1723 INIT_LIST_HEAD(&ctx->event_list);
1724 atomic_set(&ctx->refcount, 1);
1728 static struct perf_event_context *find_get_context(pid_t pid, int cpu)
1730 struct perf_event_context *ctx;
1731 struct perf_cpu_context *cpuctx;
1732 struct task_struct *task;
1733 unsigned long flags;
1736 if (pid == -1 && cpu != -1) {
1737 /* Must be root to operate on a CPU event: */
1738 if (perf_paranoid_cpu() && !capable(CAP_SYS_ADMIN))
1739 return ERR_PTR(-EACCES);
1741 if (cpu < 0 || cpu >= nr_cpumask_bits)
1742 return ERR_PTR(-EINVAL);
1745 * We could be clever and allow to attach a event to an
1746 * offline CPU and activate it when the CPU comes up, but
1749 if (!cpu_online(cpu))
1750 return ERR_PTR(-ENODEV);
1752 cpuctx = &per_cpu(perf_cpu_context, cpu);
1763 task = find_task_by_vpid(pid);
1765 get_task_struct(task);
1769 return ERR_PTR(-ESRCH);
1772 * Can't attach events to a dying task.
1775 if (task->flags & PF_EXITING)
1778 /* Reuse ptrace permission checks for now. */
1780 if (!ptrace_may_access(task, PTRACE_MODE_READ))
1784 ctx = perf_lock_task_context(task, &flags);
1787 raw_spin_unlock_irqrestore(&ctx->lock, flags);
1791 ctx = kzalloc(sizeof(struct perf_event_context), GFP_KERNEL);
1795 __perf_event_init_context(ctx, task);
1797 if (cmpxchg(&task->perf_event_ctxp, NULL, ctx)) {
1799 * We raced with some other task; use
1800 * the context they set.
1805 get_task_struct(task);
1808 put_task_struct(task);
1812 put_task_struct(task);
1813 return ERR_PTR(err);
1816 static void perf_event_free_filter(struct perf_event *event);
1818 static void free_event_rcu(struct rcu_head *head)
1820 struct perf_event *event;
1822 event = container_of(head, struct perf_event, rcu_head);
1824 put_pid_ns(event->ns);
1825 perf_event_free_filter(event);
1829 static void perf_pending_sync(struct perf_event *event);
1831 static void free_event(struct perf_event *event)
1833 perf_pending_sync(event);
1835 if (!event->parent) {
1836 atomic_dec(&nr_events);
1837 if (event->attr.mmap)
1838 atomic_dec(&nr_mmap_events);
1839 if (event->attr.comm)
1840 atomic_dec(&nr_comm_events);
1841 if (event->attr.task)
1842 atomic_dec(&nr_task_events);
1845 if (event->output) {
1846 fput(event->output->filp);
1847 event->output = NULL;
1851 event->destroy(event);
1853 put_ctx(event->ctx);
1854 call_rcu(&event->rcu_head, free_event_rcu);
1857 int perf_event_release_kernel(struct perf_event *event)
1859 struct perf_event_context *ctx = event->ctx;
1861 WARN_ON_ONCE(ctx->parent_ctx);
1862 mutex_lock(&ctx->mutex);
1863 perf_event_remove_from_context(event);
1864 mutex_unlock(&ctx->mutex);
1866 mutex_lock(&event->owner->perf_event_mutex);
1867 list_del_init(&event->owner_entry);
1868 mutex_unlock(&event->owner->perf_event_mutex);
1869 put_task_struct(event->owner);
1875 EXPORT_SYMBOL_GPL(perf_event_release_kernel);
1878 * Called when the last reference to the file is gone.
1880 static int perf_release(struct inode *inode, struct file *file)
1882 struct perf_event *event = file->private_data;
1884 file->private_data = NULL;
1886 return perf_event_release_kernel(event);
1889 static int perf_event_read_size(struct perf_event *event)
1891 int entry = sizeof(u64); /* value */
1895 if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1896 size += sizeof(u64);
1898 if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1899 size += sizeof(u64);
1901 if (event->attr.read_format & PERF_FORMAT_ID)
1902 entry += sizeof(u64);
1904 if (event->attr.read_format & PERF_FORMAT_GROUP) {
1905 nr += event->group_leader->nr_siblings;
1906 size += sizeof(u64);
1914 u64 perf_event_read_value(struct perf_event *event, u64 *enabled, u64 *running)
1916 struct perf_event *child;
1922 mutex_lock(&event->child_mutex);
1923 total += perf_event_read(event);
1924 *enabled += event->total_time_enabled +
1925 atomic64_read(&event->child_total_time_enabled);
1926 *running += event->total_time_running +
1927 atomic64_read(&event->child_total_time_running);
1929 list_for_each_entry(child, &event->child_list, child_list) {
1930 total += perf_event_read(child);
1931 *enabled += child->total_time_enabled;
1932 *running += child->total_time_running;
1934 mutex_unlock(&event->child_mutex);
1938 EXPORT_SYMBOL_GPL(perf_event_read_value);
1940 static int perf_event_read_group(struct perf_event *event,
1941 u64 read_format, char __user *buf)
1943 struct perf_event *leader = event->group_leader, *sub;
1944 int n = 0, size = 0, ret = -EFAULT;
1945 struct perf_event_context *ctx = leader->ctx;
1947 u64 count, enabled, running;
1949 mutex_lock(&ctx->mutex);
1950 count = perf_event_read_value(leader, &enabled, &running);
1952 values[n++] = 1 + leader->nr_siblings;
1953 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1954 values[n++] = enabled;
1955 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1956 values[n++] = running;
1957 values[n++] = count;
1958 if (read_format & PERF_FORMAT_ID)
1959 values[n++] = primary_event_id(leader);
1961 size = n * sizeof(u64);
1963 if (copy_to_user(buf, values, size))
1968 list_for_each_entry(sub, &leader->sibling_list, group_entry) {
1971 values[n++] = perf_event_read_value(sub, &enabled, &running);
1972 if (read_format & PERF_FORMAT_ID)
1973 values[n++] = primary_event_id(sub);
1975 size = n * sizeof(u64);
1977 if (copy_to_user(buf + ret, values, size)) {
1985 mutex_unlock(&ctx->mutex);
1990 static int perf_event_read_one(struct perf_event *event,
1991 u64 read_format, char __user *buf)
1993 u64 enabled, running;
1997 values[n++] = perf_event_read_value(event, &enabled, &running);
1998 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1999 values[n++] = enabled;
2000 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
2001 values[n++] = running;
2002 if (read_format & PERF_FORMAT_ID)
2003 values[n++] = primary_event_id(event);
2005 if (copy_to_user(buf, values, n * sizeof(u64)))
2008 return n * sizeof(u64);
2012 * Read the performance event - simple non blocking version for now
2015 perf_read_hw(struct perf_event *event, char __user *buf, size_t count)
2017 u64 read_format = event->attr.read_format;
2021 * Return end-of-file for a read on a event that is in
2022 * error state (i.e. because it was pinned but it couldn't be
2023 * scheduled on to the CPU at some point).
2025 if (event->state == PERF_EVENT_STATE_ERROR)
2028 if (count < perf_event_read_size(event))
2031 WARN_ON_ONCE(event->ctx->parent_ctx);
2032 if (read_format & PERF_FORMAT_GROUP)
2033 ret = perf_event_read_group(event, read_format, buf);
2035 ret = perf_event_read_one(event, read_format, buf);
2041 perf_read(struct file *file, char __user *buf, size_t count, loff_t *ppos)
2043 struct perf_event *event = file->private_data;
2045 return perf_read_hw(event, buf, count);
2048 static unsigned int perf_poll(struct file *file, poll_table *wait)
2050 struct perf_event *event = file->private_data;
2051 struct perf_mmap_data *data;
2052 unsigned int events = POLL_HUP;
2055 data = rcu_dereference(event->data);
2057 events = atomic_xchg(&data->poll, 0);
2060 poll_wait(file, &event->waitq, wait);
2065 static void perf_event_reset(struct perf_event *event)
2067 (void)perf_event_read(event);
2068 atomic64_set(&event->count, 0);
2069 perf_event_update_userpage(event);
2073 * Holding the top-level event's child_mutex means that any
2074 * descendant process that has inherited this event will block
2075 * in sync_child_event if it goes to exit, thus satisfying the
2076 * task existence requirements of perf_event_enable/disable.
2078 static void perf_event_for_each_child(struct perf_event *event,
2079 void (*func)(struct perf_event *))
2081 struct perf_event *child;
2083 WARN_ON_ONCE(event->ctx->parent_ctx);
2084 mutex_lock(&event->child_mutex);
2086 list_for_each_entry(child, &event->child_list, child_list)
2088 mutex_unlock(&event->child_mutex);
2091 static void perf_event_for_each(struct perf_event *event,
2092 void (*func)(struct perf_event *))
2094 struct perf_event_context *ctx = event->ctx;
2095 struct perf_event *sibling;
2097 WARN_ON_ONCE(ctx->parent_ctx);
2098 mutex_lock(&ctx->mutex);
2099 event = event->group_leader;
2101 perf_event_for_each_child(event, func);
2103 list_for_each_entry(sibling, &event->sibling_list, group_entry)
2104 perf_event_for_each_child(event, func);
2105 mutex_unlock(&ctx->mutex);
2108 static int perf_event_period(struct perf_event *event, u64 __user *arg)
2110 struct perf_event_context *ctx = event->ctx;
2115 if (!event->attr.sample_period)
2118 size = copy_from_user(&value, arg, sizeof(value));
2119 if (size != sizeof(value))
2125 raw_spin_lock_irq(&ctx->lock);
2126 if (event->attr.freq) {
2127 if (value > sysctl_perf_event_sample_rate) {
2132 event->attr.sample_freq = value;
2134 event->attr.sample_period = value;
2135 event->hw.sample_period = value;
2138 raw_spin_unlock_irq(&ctx->lock);
2143 static int perf_event_set_output(struct perf_event *event, int output_fd);
2144 static int perf_event_set_filter(struct perf_event *event, void __user *arg);
2146 static long perf_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
2148 struct perf_event *event = file->private_data;
2149 void (*func)(struct perf_event *);
2153 case PERF_EVENT_IOC_ENABLE:
2154 func = perf_event_enable;
2156 case PERF_EVENT_IOC_DISABLE:
2157 func = perf_event_disable;
2159 case PERF_EVENT_IOC_RESET:
2160 func = perf_event_reset;
2163 case PERF_EVENT_IOC_REFRESH:
2164 return perf_event_refresh(event, arg);
2166 case PERF_EVENT_IOC_PERIOD:
2167 return perf_event_period(event, (u64 __user *)arg);
2169 case PERF_EVENT_IOC_SET_OUTPUT:
2170 return perf_event_set_output(event, arg);
2172 case PERF_EVENT_IOC_SET_FILTER:
2173 return perf_event_set_filter(event, (void __user *)arg);
2179 if (flags & PERF_IOC_FLAG_GROUP)
2180 perf_event_for_each(event, func);
2182 perf_event_for_each_child(event, func);
2187 int perf_event_task_enable(void)
2189 struct perf_event *event;
2191 mutex_lock(¤t->perf_event_mutex);
2192 list_for_each_entry(event, ¤t->perf_event_list, owner_entry)
2193 perf_event_for_each_child(event, perf_event_enable);
2194 mutex_unlock(¤t->perf_event_mutex);
2199 int perf_event_task_disable(void)
2201 struct perf_event *event;
2203 mutex_lock(¤t->perf_event_mutex);
2204 list_for_each_entry(event, ¤t->perf_event_list, owner_entry)
2205 perf_event_for_each_child(event, perf_event_disable);
2206 mutex_unlock(¤t->perf_event_mutex);
2211 #ifndef PERF_EVENT_INDEX_OFFSET
2212 # define PERF_EVENT_INDEX_OFFSET 0
2215 static int perf_event_index(struct perf_event *event)
2217 if (event->state != PERF_EVENT_STATE_ACTIVE)
2220 return event->hw.idx + 1 - PERF_EVENT_INDEX_OFFSET;
2224 * Callers need to ensure there can be no nesting of this function, otherwise
2225 * the seqlock logic goes bad. We can not serialize this because the arch
2226 * code calls this from NMI context.
2228 void perf_event_update_userpage(struct perf_event *event)
2230 struct perf_event_mmap_page *userpg;
2231 struct perf_mmap_data *data;
2234 data = rcu_dereference(event->data);
2238 userpg = data->user_page;
2241 * Disable preemption so as to not let the corresponding user-space
2242 * spin too long if we get preempted.
2247 userpg->index = perf_event_index(event);
2248 userpg->offset = atomic64_read(&event->count);
2249 if (event->state == PERF_EVENT_STATE_ACTIVE)
2250 userpg->offset -= atomic64_read(&event->hw.prev_count);
2252 userpg->time_enabled = event->total_time_enabled +
2253 atomic64_read(&event->child_total_time_enabled);
2255 userpg->time_running = event->total_time_running +
2256 atomic64_read(&event->child_total_time_running);
2265 static unsigned long perf_data_size(struct perf_mmap_data *data)
2267 return data->nr_pages << (PAGE_SHIFT + data->data_order);
2270 #ifndef CONFIG_PERF_USE_VMALLOC
2273 * Back perf_mmap() with regular GFP_KERNEL-0 pages.
2276 static struct page *
2277 perf_mmap_to_page(struct perf_mmap_data *data, unsigned long pgoff)
2279 if (pgoff > data->nr_pages)
2283 return virt_to_page(data->user_page);
2285 return virt_to_page(data->data_pages[pgoff - 1]);
2288 static struct perf_mmap_data *
2289 perf_mmap_data_alloc(struct perf_event *event, int nr_pages)
2291 struct perf_mmap_data *data;
2295 WARN_ON(atomic_read(&event->mmap_count));
2297 size = sizeof(struct perf_mmap_data);
2298 size += nr_pages * sizeof(void *);
2300 data = kzalloc(size, GFP_KERNEL);
2304 data->user_page = (void *)get_zeroed_page(GFP_KERNEL);
2305 if (!data->user_page)
2306 goto fail_user_page;
2308 for (i = 0; i < nr_pages; i++) {
2309 data->data_pages[i] = (void *)get_zeroed_page(GFP_KERNEL);
2310 if (!data->data_pages[i])
2311 goto fail_data_pages;
2314 data->data_order = 0;
2315 data->nr_pages = nr_pages;
2320 for (i--; i >= 0; i--)
2321 free_page((unsigned long)data->data_pages[i]);
2323 free_page((unsigned long)data->user_page);
2332 static void perf_mmap_free_page(unsigned long addr)
2334 struct page *page = virt_to_page((void *)addr);
2336 page->mapping = NULL;
2340 static void perf_mmap_data_free(struct perf_mmap_data *data)
2344 perf_mmap_free_page((unsigned long)data->user_page);
2345 for (i = 0; i < data->nr_pages; i++)
2346 perf_mmap_free_page((unsigned long)data->data_pages[i]);
2353 * Back perf_mmap() with vmalloc memory.
2355 * Required for architectures that have d-cache aliasing issues.
2358 static struct page *
2359 perf_mmap_to_page(struct perf_mmap_data *data, unsigned long pgoff)
2361 if (pgoff > (1UL << data->data_order))
2364 return vmalloc_to_page((void *)data->user_page + pgoff * PAGE_SIZE);
2367 static void perf_mmap_unmark_page(void *addr)
2369 struct page *page = vmalloc_to_page(addr);
2371 page->mapping = NULL;
2374 static void perf_mmap_data_free_work(struct work_struct *work)
2376 struct perf_mmap_data *data;
2380 data = container_of(work, struct perf_mmap_data, work);
2381 nr = 1 << data->data_order;
2383 base = data->user_page;
2384 for (i = 0; i < nr + 1; i++)
2385 perf_mmap_unmark_page(base + (i * PAGE_SIZE));
2391 static void perf_mmap_data_free(struct perf_mmap_data *data)
2393 schedule_work(&data->work);
2396 static struct perf_mmap_data *
2397 perf_mmap_data_alloc(struct perf_event *event, int nr_pages)
2399 struct perf_mmap_data *data;
2403 WARN_ON(atomic_read(&event->mmap_count));
2405 size = sizeof(struct perf_mmap_data);
2406 size += sizeof(void *);
2408 data = kzalloc(size, GFP_KERNEL);
2412 INIT_WORK(&data->work, perf_mmap_data_free_work);
2414 all_buf = vmalloc_user((nr_pages + 1) * PAGE_SIZE);
2418 data->user_page = all_buf;
2419 data->data_pages[0] = all_buf + PAGE_SIZE;
2420 data->data_order = ilog2(nr_pages);
2434 static int perf_mmap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
2436 struct perf_event *event = vma->vm_file->private_data;
2437 struct perf_mmap_data *data;
2438 int ret = VM_FAULT_SIGBUS;
2440 if (vmf->flags & FAULT_FLAG_MKWRITE) {
2441 if (vmf->pgoff == 0)
2447 data = rcu_dereference(event->data);
2451 if (vmf->pgoff && (vmf->flags & FAULT_FLAG_WRITE))
2454 vmf->page = perf_mmap_to_page(data, vmf->pgoff);
2458 get_page(vmf->page);
2459 vmf->page->mapping = vma->vm_file->f_mapping;
2460 vmf->page->index = vmf->pgoff;
2470 perf_mmap_data_init(struct perf_event *event, struct perf_mmap_data *data)
2472 long max_size = perf_data_size(data);
2474 atomic_set(&data->lock, -1);
2476 if (event->attr.watermark) {
2477 data->watermark = min_t(long, max_size,
2478 event->attr.wakeup_watermark);
2481 if (!data->watermark)
2482 data->watermark = max_size / 2;
2485 rcu_assign_pointer(event->data, data);
2488 static void perf_mmap_data_free_rcu(struct rcu_head *rcu_head)
2490 struct perf_mmap_data *data;
2492 data = container_of(rcu_head, struct perf_mmap_data, rcu_head);
2493 perf_mmap_data_free(data);
2496 static void perf_mmap_data_release(struct perf_event *event)
2498 struct perf_mmap_data *data = event->data;
2500 WARN_ON(atomic_read(&event->mmap_count));
2502 rcu_assign_pointer(event->data, NULL);
2503 call_rcu(&data->rcu_head, perf_mmap_data_free_rcu);
2506 static void perf_mmap_open(struct vm_area_struct *vma)
2508 struct perf_event *event = vma->vm_file->private_data;
2510 atomic_inc(&event->mmap_count);
2513 static void perf_mmap_close(struct vm_area_struct *vma)
2515 struct perf_event *event = vma->vm_file->private_data;
2517 WARN_ON_ONCE(event->ctx->parent_ctx);
2518 if (atomic_dec_and_mutex_lock(&event->mmap_count, &event->mmap_mutex)) {
2519 unsigned long size = perf_data_size(event->data);
2520 struct user_struct *user = current_user();
2522 atomic_long_sub((size >> PAGE_SHIFT) + 1, &user->locked_vm);
2523 vma->vm_mm->locked_vm -= event->data->nr_locked;
2524 perf_mmap_data_release(event);
2525 mutex_unlock(&event->mmap_mutex);
2529 static const struct vm_operations_struct perf_mmap_vmops = {
2530 .open = perf_mmap_open,
2531 .close = perf_mmap_close,
2532 .fault = perf_mmap_fault,
2533 .page_mkwrite = perf_mmap_fault,
2536 static int perf_mmap(struct file *file, struct vm_area_struct *vma)
2538 struct perf_event *event = file->private_data;
2539 unsigned long user_locked, user_lock_limit;
2540 struct user_struct *user = current_user();
2541 unsigned long locked, lock_limit;
2542 struct perf_mmap_data *data;
2543 unsigned long vma_size;
2544 unsigned long nr_pages;
2545 long user_extra, extra;
2548 if (!(vma->vm_flags & VM_SHARED))
2551 vma_size = vma->vm_end - vma->vm_start;
2552 nr_pages = (vma_size / PAGE_SIZE) - 1;
2555 * If we have data pages ensure they're a power-of-two number, so we
2556 * can do bitmasks instead of modulo.
2558 if (nr_pages != 0 && !is_power_of_2(nr_pages))
2561 if (vma_size != PAGE_SIZE * (1 + nr_pages))
2564 if (vma->vm_pgoff != 0)
2567 WARN_ON_ONCE(event->ctx->parent_ctx);
2568 mutex_lock(&event->mmap_mutex);
2569 if (event->output) {
2574 if (atomic_inc_not_zero(&event->mmap_count)) {
2575 if (nr_pages != event->data->nr_pages)
2580 user_extra = nr_pages + 1;
2581 user_lock_limit = sysctl_perf_event_mlock >> (PAGE_SHIFT - 10);
2584 * Increase the limit linearly with more CPUs:
2586 user_lock_limit *= num_online_cpus();
2588 user_locked = atomic_long_read(&user->locked_vm) + user_extra;
2591 if (user_locked > user_lock_limit)
2592 extra = user_locked - user_lock_limit;
2594 lock_limit = rlimit(RLIMIT_MEMLOCK);
2595 lock_limit >>= PAGE_SHIFT;
2596 locked = vma->vm_mm->locked_vm + extra;
2598 if ((locked > lock_limit) && perf_paranoid_tracepoint_raw() &&
2599 !capable(CAP_IPC_LOCK)) {
2604 WARN_ON(event->data);
2606 data = perf_mmap_data_alloc(event, nr_pages);
2612 perf_mmap_data_init(event, data);
2614 atomic_set(&event->mmap_count, 1);
2615 atomic_long_add(user_extra, &user->locked_vm);
2616 vma->vm_mm->locked_vm += extra;
2617 event->data->nr_locked = extra;
2618 if (vma->vm_flags & VM_WRITE)
2619 event->data->writable = 1;
2622 mutex_unlock(&event->mmap_mutex);
2624 vma->vm_flags |= VM_RESERVED;
2625 vma->vm_ops = &perf_mmap_vmops;
2630 static int perf_fasync(int fd, struct file *filp, int on)
2632 struct inode *inode = filp->f_path.dentry->d_inode;
2633 struct perf_event *event = filp->private_data;
2636 mutex_lock(&inode->i_mutex);
2637 retval = fasync_helper(fd, filp, on, &event->fasync);
2638 mutex_unlock(&inode->i_mutex);
2646 static const struct file_operations perf_fops = {
2647 .release = perf_release,
2650 .unlocked_ioctl = perf_ioctl,
2651 .compat_ioctl = perf_ioctl,
2653 .fasync = perf_fasync,
2659 * If there's data, ensure we set the poll() state and publish everything
2660 * to user-space before waking everybody up.
2663 void perf_event_wakeup(struct perf_event *event)
2665 wake_up_all(&event->waitq);
2667 if (event->pending_kill) {
2668 kill_fasync(&event->fasync, SIGIO, event->pending_kill);
2669 event->pending_kill = 0;
2676 * Handle the case where we need to wakeup up from NMI (or rq->lock) context.
2678 * The NMI bit means we cannot possibly take locks. Therefore, maintain a
2679 * single linked list and use cmpxchg() to add entries lockless.
2682 static void perf_pending_event(struct perf_pending_entry *entry)
2684 struct perf_event *event = container_of(entry,
2685 struct perf_event, pending);
2687 if (event->pending_disable) {
2688 event->pending_disable = 0;
2689 __perf_event_disable(event);
2692 if (event->pending_wakeup) {
2693 event->pending_wakeup = 0;
2694 perf_event_wakeup(event);
2698 #define PENDING_TAIL ((struct perf_pending_entry *)-1UL)
2700 static DEFINE_PER_CPU(struct perf_pending_entry *, perf_pending_head) = {
2704 static void perf_pending_queue(struct perf_pending_entry *entry,
2705 void (*func)(struct perf_pending_entry *))
2707 struct perf_pending_entry **head;
2709 if (cmpxchg(&entry->next, NULL, PENDING_TAIL) != NULL)
2714 head = &get_cpu_var(perf_pending_head);
2717 entry->next = *head;
2718 } while (cmpxchg(head, entry->next, entry) != entry->next);
2720 set_perf_event_pending();
2722 put_cpu_var(perf_pending_head);
2725 static int __perf_pending_run(void)
2727 struct perf_pending_entry *list;
2730 list = xchg(&__get_cpu_var(perf_pending_head), PENDING_TAIL);
2731 while (list != PENDING_TAIL) {
2732 void (*func)(struct perf_pending_entry *);
2733 struct perf_pending_entry *entry = list;
2740 * Ensure we observe the unqueue before we issue the wakeup,
2741 * so that we won't be waiting forever.
2742 * -- see perf_not_pending().
2753 static inline int perf_not_pending(struct perf_event *event)
2756 * If we flush on whatever cpu we run, there is a chance we don't
2760 __perf_pending_run();
2764 * Ensure we see the proper queue state before going to sleep
2765 * so that we do not miss the wakeup. -- see perf_pending_handle()
2768 return event->pending.next == NULL;
2771 static void perf_pending_sync(struct perf_event *event)
2773 wait_event(event->waitq, perf_not_pending(event));
2776 void perf_event_do_pending(void)
2778 __perf_pending_run();
2782 * Callchain support -- arch specific
2785 __weak struct perf_callchain_entry *perf_callchain(struct pt_regs *regs)
2793 static bool perf_output_space(struct perf_mmap_data *data, unsigned long tail,
2794 unsigned long offset, unsigned long head)
2798 if (!data->writable)
2801 mask = perf_data_size(data) - 1;
2803 offset = (offset - tail) & mask;
2804 head = (head - tail) & mask;
2806 if ((int)(head - offset) < 0)
2812 static void perf_output_wakeup(struct perf_output_handle *handle)
2814 atomic_set(&handle->data->poll, POLL_IN);
2817 handle->event->pending_wakeup = 1;
2818 perf_pending_queue(&handle->event->pending,
2819 perf_pending_event);
2821 perf_event_wakeup(handle->event);
2825 * Curious locking construct.
2827 * We need to ensure a later event_id doesn't publish a head when a former
2828 * event_id isn't done writing. However since we need to deal with NMIs we
2829 * cannot fully serialize things.
2831 * What we do is serialize between CPUs so we only have to deal with NMI
2832 * nesting on a single CPU.
2834 * We only publish the head (and generate a wakeup) when the outer-most
2835 * event_id completes.
2837 static void perf_output_lock(struct perf_output_handle *handle)
2839 struct perf_mmap_data *data = handle->data;
2840 int cur, cpu = get_cpu();
2845 cur = atomic_cmpxchg(&data->lock, -1, cpu);
2857 static void perf_output_unlock(struct perf_output_handle *handle)
2859 struct perf_mmap_data *data = handle->data;
2863 data->done_head = data->head;
2865 if (!handle->locked)
2870 * The xchg implies a full barrier that ensures all writes are done
2871 * before we publish the new head, matched by a rmb() in userspace when
2872 * reading this position.
2874 while ((head = atomic_long_xchg(&data->done_head, 0)))
2875 data->user_page->data_head = head;
2878 * NMI can happen here, which means we can miss a done_head update.
2881 cpu = atomic_xchg(&data->lock, -1);
2882 WARN_ON_ONCE(cpu != smp_processor_id());
2885 * Therefore we have to validate we did not indeed do so.
2887 if (unlikely(atomic_long_read(&data->done_head))) {
2889 * Since we had it locked, we can lock it again.
2891 while (atomic_cmpxchg(&data->lock, -1, cpu) != -1)
2897 if (atomic_xchg(&data->wakeup, 0))
2898 perf_output_wakeup(handle);
2903 void perf_output_copy(struct perf_output_handle *handle,
2904 const void *buf, unsigned int len)
2906 unsigned int pages_mask;
2907 unsigned long offset;
2911 offset = handle->offset;
2912 pages_mask = handle->data->nr_pages - 1;
2913 pages = handle->data->data_pages;
2916 unsigned long page_offset;
2917 unsigned long page_size;
2920 nr = (offset >> PAGE_SHIFT) & pages_mask;
2921 page_size = 1UL << (handle->data->data_order + PAGE_SHIFT);
2922 page_offset = offset & (page_size - 1);
2923 size = min_t(unsigned int, page_size - page_offset, len);
2925 memcpy(pages[nr] + page_offset, buf, size);
2932 handle->offset = offset;
2935 * Check we didn't copy past our reservation window, taking the
2936 * possible unsigned int wrap into account.
2938 WARN_ON_ONCE(((long)(handle->head - handle->offset)) < 0);
2941 int perf_output_begin(struct perf_output_handle *handle,
2942 struct perf_event *event, unsigned int size,
2943 int nmi, int sample)
2945 struct perf_event *output_event;
2946 struct perf_mmap_data *data;
2947 unsigned long tail, offset, head;
2950 struct perf_event_header header;
2957 * For inherited events we send all the output towards the parent.
2960 event = event->parent;
2962 output_event = rcu_dereference(event->output);
2964 event = output_event;
2966 data = rcu_dereference(event->data);
2970 handle->data = data;
2971 handle->event = event;
2973 handle->sample = sample;
2975 if (!data->nr_pages)
2978 have_lost = atomic_read(&data->lost);
2980 size += sizeof(lost_event);
2982 perf_output_lock(handle);
2986 * Userspace could choose to issue a mb() before updating the
2987 * tail pointer. So that all reads will be completed before the
2990 tail = ACCESS_ONCE(data->user_page->data_tail);
2992 offset = head = atomic_long_read(&data->head);
2994 if (unlikely(!perf_output_space(data, tail, offset, head)))
2996 } while (atomic_long_cmpxchg(&data->head, offset, head) != offset);
2998 handle->offset = offset;
2999 handle->head = head;
3001 if (head - tail > data->watermark)
3002 atomic_set(&data->wakeup, 1);
3005 lost_event.header.type = PERF_RECORD_LOST;
3006 lost_event.header.misc = 0;
3007 lost_event.header.size = sizeof(lost_event);
3008 lost_event.id = event->id;
3009 lost_event.lost = atomic_xchg(&data->lost, 0);
3011 perf_output_put(handle, lost_event);
3017 atomic_inc(&data->lost);
3018 perf_output_unlock(handle);
3025 void perf_output_end(struct perf_output_handle *handle)
3027 struct perf_event *event = handle->event;
3028 struct perf_mmap_data *data = handle->data;
3030 int wakeup_events = event->attr.wakeup_events;
3032 if (handle->sample && wakeup_events) {
3033 int events = atomic_inc_return(&data->events);
3034 if (events >= wakeup_events) {
3035 atomic_sub(wakeup_events, &data->events);
3036 atomic_set(&data->wakeup, 1);
3040 perf_output_unlock(handle);
3044 static u32 perf_event_pid(struct perf_event *event, struct task_struct *p)
3047 * only top level events have the pid namespace they were created in
3050 event = event->parent;
3052 return task_tgid_nr_ns(p, event->ns);
3055 static u32 perf_event_tid(struct perf_event *event, struct task_struct *p)
3058 * only top level events have the pid namespace they were created in
3061 event = event->parent;
3063 return task_pid_nr_ns(p, event->ns);
3066 static void perf_output_read_one(struct perf_output_handle *handle,
3067 struct perf_event *event)
3069 u64 read_format = event->attr.read_format;
3073 values[n++] = atomic64_read(&event->count);
3074 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
3075 values[n++] = event->total_time_enabled +
3076 atomic64_read(&event->child_total_time_enabled);
3078 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
3079 values[n++] = event->total_time_running +
3080 atomic64_read(&event->child_total_time_running);
3082 if (read_format & PERF_FORMAT_ID)
3083 values[n++] = primary_event_id(event);
3085 perf_output_copy(handle, values, n * sizeof(u64));
3089 * XXX PERF_FORMAT_GROUP vs inherited events seems difficult.
3091 static void perf_output_read_group(struct perf_output_handle *handle,
3092 struct perf_event *event)
3094 struct perf_event *leader = event->group_leader, *sub;
3095 u64 read_format = event->attr.read_format;
3099 values[n++] = 1 + leader->nr_siblings;
3101 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
3102 values[n++] = leader->total_time_enabled;
3104 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
3105 values[n++] = leader->total_time_running;
3107 if (leader != event)
3108 leader->pmu->read(leader);
3110 values[n++] = atomic64_read(&leader->count);
3111 if (read_format & PERF_FORMAT_ID)
3112 values[n++] = primary_event_id(leader);
3114 perf_output_copy(handle, values, n * sizeof(u64));
3116 list_for_each_entry(sub, &leader->sibling_list, group_entry) {
3120 sub->pmu->read(sub);
3122 values[n++] = atomic64_read(&sub->count);
3123 if (read_format & PERF_FORMAT_ID)
3124 values[n++] = primary_event_id(sub);
3126 perf_output_copy(handle, values, n * sizeof(u64));
3130 static void perf_output_read(struct perf_output_handle *handle,
3131 struct perf_event *event)
3133 if (event->attr.read_format & PERF_FORMAT_GROUP)
3134 perf_output_read_group(handle, event);
3136 perf_output_read_one(handle, event);
3139 void perf_output_sample(struct perf_output_handle *handle,
3140 struct perf_event_header *header,
3141 struct perf_sample_data *data,
3142 struct perf_event *event)
3144 u64 sample_type = data->type;
3146 perf_output_put(handle, *header);
3148 if (sample_type & PERF_SAMPLE_IP)
3149 perf_output_put(handle, data->ip);
3151 if (sample_type & PERF_SAMPLE_TID)
3152 perf_output_put(handle, data->tid_entry);
3154 if (sample_type & PERF_SAMPLE_TIME)
3155 perf_output_put(handle, data->time);
3157 if (sample_type & PERF_SAMPLE_ADDR)
3158 perf_output_put(handle, data->addr);
3160 if (sample_type & PERF_SAMPLE_ID)
3161 perf_output_put(handle, data->id);
3163 if (sample_type & PERF_SAMPLE_STREAM_ID)
3164 perf_output_put(handle, data->stream_id);
3166 if (sample_type & PERF_SAMPLE_CPU)
3167 perf_output_put(handle, data->cpu_entry);
3169 if (sample_type & PERF_SAMPLE_PERIOD)
3170 perf_output_put(handle, data->period);
3172 if (sample_type & PERF_SAMPLE_READ)
3173 perf_output_read(handle, event);
3175 if (sample_type & PERF_SAMPLE_CALLCHAIN) {
3176 if (data->callchain) {
3179 if (data->callchain)
3180 size += data->callchain->nr;
3182 size *= sizeof(u64);
3184 perf_output_copy(handle, data->callchain, size);
3187 perf_output_put(handle, nr);
3191 if (sample_type & PERF_SAMPLE_RAW) {
3193 perf_output_put(handle, data->raw->size);
3194 perf_output_copy(handle, data->raw->data,
3201 .size = sizeof(u32),
3204 perf_output_put(handle, raw);
3209 void perf_prepare_sample(struct perf_event_header *header,
3210 struct perf_sample_data *data,
3211 struct perf_event *event,
3212 struct pt_regs *regs)
3214 u64 sample_type = event->attr.sample_type;
3216 data->type = sample_type;
3218 header->type = PERF_RECORD_SAMPLE;
3219 header->size = sizeof(*header);
3222 header->misc |= perf_misc_flags(regs);
3224 if (sample_type & PERF_SAMPLE_IP) {
3225 data->ip = perf_instruction_pointer(regs);
3227 header->size += sizeof(data->ip);
3230 if (sample_type & PERF_SAMPLE_TID) {
3231 /* namespace issues */
3232 data->tid_entry.pid = perf_event_pid(event, current);
3233 data->tid_entry.tid = perf_event_tid(event, current);
3235 header->size += sizeof(data->tid_entry);
3238 if (sample_type & PERF_SAMPLE_TIME) {
3239 data->time = perf_clock();
3241 header->size += sizeof(data->time);
3244 if (sample_type & PERF_SAMPLE_ADDR)
3245 header->size += sizeof(data->addr);
3247 if (sample_type & PERF_SAMPLE_ID) {
3248 data->id = primary_event_id(event);
3250 header->size += sizeof(data->id);
3253 if (sample_type & PERF_SAMPLE_STREAM_ID) {
3254 data->stream_id = event->id;
3256 header->size += sizeof(data->stream_id);
3259 if (sample_type & PERF_SAMPLE_CPU) {
3260 data->cpu_entry.cpu = raw_smp_processor_id();
3261 data->cpu_entry.reserved = 0;
3263 header->size += sizeof(data->cpu_entry);
3266 if (sample_type & PERF_SAMPLE_PERIOD)
3267 header->size += sizeof(data->period);
3269 if (sample_type & PERF_SAMPLE_READ)
3270 header->size += perf_event_read_size(event);
3272 if (sample_type & PERF_SAMPLE_CALLCHAIN) {
3275 data->callchain = perf_callchain(regs);
3277 if (data->callchain)
3278 size += data->callchain->nr;
3280 header->size += size * sizeof(u64);
3283 if (sample_type & PERF_SAMPLE_RAW) {
3284 int size = sizeof(u32);
3287 size += data->raw->size;
3289 size += sizeof(u32);
3291 WARN_ON_ONCE(size & (sizeof(u64)-1));
3292 header->size += size;
3296 static void perf_event_output(struct perf_event *event, int nmi,
3297 struct perf_sample_data *data,
3298 struct pt_regs *regs)
3300 struct perf_output_handle handle;
3301 struct perf_event_header header;
3303 perf_prepare_sample(&header, data, event, regs);
3305 if (perf_output_begin(&handle, event, header.size, nmi, 1))
3308 perf_output_sample(&handle, &header, data, event);
3310 perf_output_end(&handle);
3317 struct perf_read_event {
3318 struct perf_event_header header;
3325 perf_event_read_event(struct perf_event *event,
3326 struct task_struct *task)
3328 struct perf_output_handle handle;
3329 struct perf_read_event read_event = {
3331 .type = PERF_RECORD_READ,
3333 .size = sizeof(read_event) + perf_event_read_size(event),
3335 .pid = perf_event_pid(event, task),
3336 .tid = perf_event_tid(event, task),
3340 ret = perf_output_begin(&handle, event, read_event.header.size, 0, 0);
3344 perf_output_put(&handle, read_event);
3345 perf_output_read(&handle, event);
3347 perf_output_end(&handle);
3351 * task tracking -- fork/exit
3353 * enabled by: attr.comm | attr.mmap | attr.task
3356 struct perf_task_event {
3357 struct task_struct *task;
3358 struct perf_event_context *task_ctx;
3361 struct perf_event_header header;
3371 static void perf_event_task_output(struct perf_event *event,
3372 struct perf_task_event *task_event)
3374 struct perf_output_handle handle;
3376 struct task_struct *task = task_event->task;
3379 size = task_event->event_id.header.size;
3380 ret = perf_output_begin(&handle, event, size, 0, 0);
3385 task_event->event_id.pid = perf_event_pid(event, task);
3386 task_event->event_id.ppid = perf_event_pid(event, current);
3388 task_event->event_id.tid = perf_event_tid(event, task);
3389 task_event->event_id.ptid = perf_event_tid(event, current);
3391 perf_output_put(&handle, task_event->event_id);
3393 perf_output_end(&handle);
3396 static int perf_event_task_match(struct perf_event *event)
3398 if (event->state < PERF_EVENT_STATE_INACTIVE)
3401 if (event->cpu != -1 && event->cpu != smp_processor_id())
3404 if (event->attr.comm || event->attr.mmap || event->attr.task)
3410 static void perf_event_task_ctx(struct perf_event_context *ctx,
3411 struct perf_task_event *task_event)
3413 struct perf_event *event;
3415 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
3416 if (perf_event_task_match(event))
3417 perf_event_task_output(event, task_event);
3421 static void perf_event_task_event(struct perf_task_event *task_event)
3423 struct perf_cpu_context *cpuctx;
3424 struct perf_event_context *ctx = task_event->task_ctx;
3427 cpuctx = &get_cpu_var(perf_cpu_context);
3428 perf_event_task_ctx(&cpuctx->ctx, task_event);
3430 ctx = rcu_dereference(current->perf_event_ctxp);
3432 perf_event_task_ctx(ctx, task_event);
3433 put_cpu_var(perf_cpu_context);
3437 static void perf_event_task(struct task_struct *task,
3438 struct perf_event_context *task_ctx,
3441 struct perf_task_event task_event;
3443 if (!atomic_read(&nr_comm_events) &&
3444 !atomic_read(&nr_mmap_events) &&
3445 !atomic_read(&nr_task_events))
3448 task_event = (struct perf_task_event){
3450 .task_ctx = task_ctx,
3453 .type = new ? PERF_RECORD_FORK : PERF_RECORD_EXIT,
3455 .size = sizeof(task_event.event_id),
3461 .time = perf_clock(),
3465 perf_event_task_event(&task_event);
3468 void perf_event_fork(struct task_struct *task)
3470 perf_event_task(task, NULL, 1);
3477 struct perf_comm_event {
3478 struct task_struct *task;
3483 struct perf_event_header header;
3490 static void perf_event_comm_output(struct perf_event *event,
3491 struct perf_comm_event *comm_event)
3493 struct perf_output_handle handle;
3494 int size = comm_event->event_id.header.size;
3495 int ret = perf_output_begin(&handle, event, size, 0, 0);
3500 comm_event->event_id.pid = perf_event_pid(event, comm_event->task);
3501 comm_event->event_id.tid = perf_event_tid(event, comm_event->task);
3503 perf_output_put(&handle, comm_event->event_id);
3504 perf_output_copy(&handle, comm_event->comm,
3505 comm_event->comm_size);
3506 perf_output_end(&handle);
3509 static int perf_event_comm_match(struct perf_event *event)
3511 if (event->state < PERF_EVENT_STATE_INACTIVE)
3514 if (event->cpu != -1 && event->cpu != smp_processor_id())
3517 if (event->attr.comm)
3523 static void perf_event_comm_ctx(struct perf_event_context *ctx,
3524 struct perf_comm_event *comm_event)
3526 struct perf_event *event;
3528 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
3529 if (perf_event_comm_match(event))
3530 perf_event_comm_output(event, comm_event);
3534 static void perf_event_comm_event(struct perf_comm_event *comm_event)
3536 struct perf_cpu_context *cpuctx;
3537 struct perf_event_context *ctx;
3539 char comm[TASK_COMM_LEN];
3541 memset(comm, 0, sizeof(comm));
3542 strlcpy(comm, comm_event->task->comm, sizeof(comm));
3543 size = ALIGN(strlen(comm)+1, sizeof(u64));
3545 comm_event->comm = comm;
3546 comm_event->comm_size = size;
3548 comm_event->event_id.header.size = sizeof(comm_event->event_id) + size;
3551 cpuctx = &get_cpu_var(perf_cpu_context);
3552 perf_event_comm_ctx(&cpuctx->ctx, comm_event);
3553 ctx = rcu_dereference(current->perf_event_ctxp);
3555 perf_event_comm_ctx(ctx, comm_event);
3556 put_cpu_var(perf_cpu_context);
3560 void perf_event_comm(struct task_struct *task)
3562 struct perf_comm_event comm_event;
3564 if (task->perf_event_ctxp)
3565 perf_event_enable_on_exec(task);
3567 if (!atomic_read(&nr_comm_events))
3570 comm_event = (struct perf_comm_event){
3576 .type = PERF_RECORD_COMM,
3585 perf_event_comm_event(&comm_event);
3592 struct perf_mmap_event {
3593 struct vm_area_struct *vma;
3595 const char *file_name;
3599 struct perf_event_header header;
3609 static void perf_event_mmap_output(struct perf_event *event,
3610 struct perf_mmap_event *mmap_event)
3612 struct perf_output_handle handle;
3613 int size = mmap_event->event_id.header.size;
3614 int ret = perf_output_begin(&handle, event, size, 0, 0);
3619 mmap_event->event_id.pid = perf_event_pid(event, current);
3620 mmap_event->event_id.tid = perf_event_tid(event, current);
3622 perf_output_put(&handle, mmap_event->event_id);
3623 perf_output_copy(&handle, mmap_event->file_name,
3624 mmap_event->file_size);
3625 perf_output_end(&handle);
3628 static int perf_event_mmap_match(struct perf_event *event,
3629 struct perf_mmap_event *mmap_event)
3631 if (event->state < PERF_EVENT_STATE_INACTIVE)
3634 if (event->cpu != -1 && event->cpu != smp_processor_id())
3637 if (event->attr.mmap)
3643 static void perf_event_mmap_ctx(struct perf_event_context *ctx,
3644 struct perf_mmap_event *mmap_event)
3646 struct perf_event *event;
3648 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
3649 if (perf_event_mmap_match(event, mmap_event))
3650 perf_event_mmap_output(event, mmap_event);
3654 static void perf_event_mmap_event(struct perf_mmap_event *mmap_event)
3656 struct perf_cpu_context *cpuctx;
3657 struct perf_event_context *ctx;
3658 struct vm_area_struct *vma = mmap_event->vma;
3659 struct file *file = vma->vm_file;
3665 memset(tmp, 0, sizeof(tmp));
3669 * d_path works from the end of the buffer backwards, so we
3670 * need to add enough zero bytes after the string to handle
3671 * the 64bit alignment we do later.
3673 buf = kzalloc(PATH_MAX + sizeof(u64), GFP_KERNEL);
3675 name = strncpy(tmp, "//enomem", sizeof(tmp));
3678 name = d_path(&file->f_path, buf, PATH_MAX);
3680 name = strncpy(tmp, "//toolong", sizeof(tmp));
3684 if (arch_vma_name(mmap_event->vma)) {
3685 name = strncpy(tmp, arch_vma_name(mmap_event->vma),
3691 name = strncpy(tmp, "[vdso]", sizeof(tmp));
3695 name = strncpy(tmp, "//anon", sizeof(tmp));
3700 size = ALIGN(strlen(name)+1, sizeof(u64));
3702 mmap_event->file_name = name;
3703 mmap_event->file_size = size;
3705 mmap_event->event_id.header.size = sizeof(mmap_event->event_id) + size;
3708 cpuctx = &get_cpu_var(perf_cpu_context);
3709 perf_event_mmap_ctx(&cpuctx->ctx, mmap_event);
3710 ctx = rcu_dereference(current->perf_event_ctxp);
3712 perf_event_mmap_ctx(ctx, mmap_event);
3713 put_cpu_var(perf_cpu_context);
3719 void __perf_event_mmap(struct vm_area_struct *vma)
3721 struct perf_mmap_event mmap_event;
3723 if (!atomic_read(&nr_mmap_events))
3726 mmap_event = (struct perf_mmap_event){
3732 .type = PERF_RECORD_MMAP,
3738 .start = vma->vm_start,
3739 .len = vma->vm_end - vma->vm_start,
3740 .pgoff = (u64)vma->vm_pgoff << PAGE_SHIFT,
3744 perf_event_mmap_event(&mmap_event);
3748 * IRQ throttle logging
3751 static void perf_log_throttle(struct perf_event *event, int enable)
3753 struct perf_output_handle handle;
3757 struct perf_event_header header;
3761 } throttle_event = {
3763 .type = PERF_RECORD_THROTTLE,
3765 .size = sizeof(throttle_event),
3767 .time = perf_clock(),
3768 .id = primary_event_id(event),
3769 .stream_id = event->id,
3773 throttle_event.header.type = PERF_RECORD_UNTHROTTLE;
3775 ret = perf_output_begin(&handle, event, sizeof(throttle_event), 1, 0);
3779 perf_output_put(&handle, throttle_event);
3780 perf_output_end(&handle);
3784 * Generic event overflow handling, sampling.
3787 static int __perf_event_overflow(struct perf_event *event, int nmi,
3788 int throttle, struct perf_sample_data *data,
3789 struct pt_regs *regs)
3791 int events = atomic_read(&event->event_limit);
3792 struct hw_perf_event *hwc = &event->hw;
3795 throttle = (throttle && event->pmu->unthrottle != NULL);
3800 if (hwc->interrupts != MAX_INTERRUPTS) {
3802 if (HZ * hwc->interrupts >
3803 (u64)sysctl_perf_event_sample_rate) {
3804 hwc->interrupts = MAX_INTERRUPTS;
3805 perf_log_throttle(event, 0);
3810 * Keep re-disabling events even though on the previous
3811 * pass we disabled it - just in case we raced with a
3812 * sched-in and the event got enabled again:
3818 if (event->attr.freq) {
3819 u64 now = perf_clock();
3820 s64 delta = now - hwc->freq_time_stamp;
3822 hwc->freq_time_stamp = now;
3824 if (delta > 0 && delta < 2*TICK_NSEC)
3825 perf_adjust_period(event, delta, hwc->last_period);
3829 * XXX event_limit might not quite work as expected on inherited
3833 event->pending_kill = POLL_IN;
3834 if (events && atomic_dec_and_test(&event->event_limit)) {
3836 event->pending_kill = POLL_HUP;
3838 event->pending_disable = 1;
3839 perf_pending_queue(&event->pending,
3840 perf_pending_event);
3842 perf_event_disable(event);
3845 if (event->overflow_handler)
3846 event->overflow_handler(event, nmi, data, regs);
3848 perf_event_output(event, nmi, data, regs);
3853 int perf_event_overflow(struct perf_event *event, int nmi,
3854 struct perf_sample_data *data,
3855 struct pt_regs *regs)
3857 return __perf_event_overflow(event, nmi, 1, data, regs);
3861 * Generic software event infrastructure
3865 * We directly increment event->count and keep a second value in
3866 * event->hw.period_left to count intervals. This period event
3867 * is kept in the range [-sample_period, 0] so that we can use the
3871 static u64 perf_swevent_set_period(struct perf_event *event)
3873 struct hw_perf_event *hwc = &event->hw;
3874 u64 period = hwc->last_period;
3878 hwc->last_period = hwc->sample_period;
3881 old = val = atomic64_read(&hwc->period_left);
3885 nr = div64_u64(period + val, period);
3886 offset = nr * period;
3888 if (atomic64_cmpxchg(&hwc->period_left, old, val) != old)
3894 static void perf_swevent_overflow(struct perf_event *event, u64 overflow,
3895 int nmi, struct perf_sample_data *data,
3896 struct pt_regs *regs)
3898 struct hw_perf_event *hwc = &event->hw;
3901 data->period = event->hw.last_period;
3903 overflow = perf_swevent_set_period(event);
3905 if (hwc->interrupts == MAX_INTERRUPTS)
3908 for (; overflow; overflow--) {
3909 if (__perf_event_overflow(event, nmi, throttle,
3912 * We inhibit the overflow from happening when
3913 * hwc->interrupts == MAX_INTERRUPTS.
3921 static void perf_swevent_unthrottle(struct perf_event *event)
3924 * Nothing to do, we already reset hwc->interrupts.
3928 static void perf_swevent_add(struct perf_event *event, u64 nr,
3929 int nmi, struct perf_sample_data *data,
3930 struct pt_regs *regs)
3932 struct hw_perf_event *hwc = &event->hw;
3934 atomic64_add(nr, &event->count);
3939 if (!hwc->sample_period)
3942 if (nr == 1 && hwc->sample_period == 1 && !event->attr.freq)
3943 return perf_swevent_overflow(event, 1, nmi, data, regs);
3945 if (atomic64_add_negative(nr, &hwc->period_left))
3948 perf_swevent_overflow(event, 0, nmi, data, regs);
3951 static int perf_swevent_is_counting(struct perf_event *event)
3954 * The event is active, we're good!
3956 if (event->state == PERF_EVENT_STATE_ACTIVE)
3960 * The event is off/error, not counting.
3962 if (event->state != PERF_EVENT_STATE_INACTIVE)
3966 * The event is inactive, if the context is active
3967 * we're part of a group that didn't make it on the 'pmu',
3970 if (event->ctx->is_active)
3974 * We're inactive and the context is too, this means the
3975 * task is scheduled out, we're counting events that happen
3976 * to us, like migration events.
3981 static int perf_tp_event_match(struct perf_event *event,
3982 struct perf_sample_data *data);
3984 static int perf_exclude_event(struct perf_event *event,
3985 struct pt_regs *regs)
3988 if (event->attr.exclude_user && user_mode(regs))
3991 if (event->attr.exclude_kernel && !user_mode(regs))
3998 static int perf_swevent_match(struct perf_event *event,
3999 enum perf_type_id type,
4001 struct perf_sample_data *data,
4002 struct pt_regs *regs)
4004 if (event->cpu != -1 && event->cpu != smp_processor_id())
4007 if (!perf_swevent_is_counting(event))
4010 if (event->attr.type != type)
4013 if (event->attr.config != event_id)
4016 if (perf_exclude_event(event, regs))
4019 if (event->attr.type == PERF_TYPE_TRACEPOINT &&
4020 !perf_tp_event_match(event, data))
4026 static void perf_swevent_ctx_event(struct perf_event_context *ctx,
4027 enum perf_type_id type,
4028 u32 event_id, u64 nr, int nmi,
4029 struct perf_sample_data *data,
4030 struct pt_regs *regs)
4032 struct perf_event *event;
4034 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
4035 if (perf_swevent_match(event, type, event_id, data, regs))
4036 perf_swevent_add(event, nr, nmi, data, regs);
4040 int perf_swevent_get_recursion_context(void)
4042 struct perf_cpu_context *cpuctx = &get_cpu_var(perf_cpu_context);
4049 else if (in_softirq())
4054 if (cpuctx->recursion[rctx]) {
4055 put_cpu_var(perf_cpu_context);
4059 cpuctx->recursion[rctx]++;
4064 EXPORT_SYMBOL_GPL(perf_swevent_get_recursion_context);
4066 void perf_swevent_put_recursion_context(int rctx)
4068 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
4070 cpuctx->recursion[rctx]--;
4071 put_cpu_var(perf_cpu_context);
4073 EXPORT_SYMBOL_GPL(perf_swevent_put_recursion_context);
4075 static void do_perf_sw_event(enum perf_type_id type, u32 event_id,
4077 struct perf_sample_data *data,
4078 struct pt_regs *regs)
4080 struct perf_cpu_context *cpuctx;
4081 struct perf_event_context *ctx;
4083 cpuctx = &__get_cpu_var(perf_cpu_context);
4085 perf_swevent_ctx_event(&cpuctx->ctx, type, event_id,
4086 nr, nmi, data, regs);
4088 * doesn't really matter which of the child contexts the
4089 * events ends up in.
4091 ctx = rcu_dereference(current->perf_event_ctxp);
4093 perf_swevent_ctx_event(ctx, type, event_id, nr, nmi, data, regs);
4097 void __perf_sw_event(u32 event_id, u64 nr, int nmi,
4098 struct pt_regs *regs, u64 addr)
4100 struct perf_sample_data data;
4103 rctx = perf_swevent_get_recursion_context();
4107 perf_sample_data_init(&data, addr);
4109 do_perf_sw_event(PERF_TYPE_SOFTWARE, event_id, nr, nmi, &data, regs);
4111 perf_swevent_put_recursion_context(rctx);
4114 static void perf_swevent_read(struct perf_event *event)
4118 static int perf_swevent_enable(struct perf_event *event)
4120 struct hw_perf_event *hwc = &event->hw;
4122 if (hwc->sample_period) {
4123 hwc->last_period = hwc->sample_period;
4124 perf_swevent_set_period(event);
4129 static void perf_swevent_disable(struct perf_event *event)
4133 static const struct pmu perf_ops_generic = {
4134 .enable = perf_swevent_enable,
4135 .disable = perf_swevent_disable,
4136 .read = perf_swevent_read,
4137 .unthrottle = perf_swevent_unthrottle,
4141 * hrtimer based swevent callback
4144 static enum hrtimer_restart perf_swevent_hrtimer(struct hrtimer *hrtimer)
4146 enum hrtimer_restart ret = HRTIMER_RESTART;
4147 struct perf_sample_data data;
4148 struct pt_regs *regs;
4149 struct perf_event *event;
4152 event = container_of(hrtimer, struct perf_event, hw.hrtimer);
4153 event->pmu->read(event);
4155 perf_sample_data_init(&data, 0);
4156 data.period = event->hw.last_period;
4157 regs = get_irq_regs();
4159 * In case we exclude kernel IPs or are somehow not in interrupt
4160 * context, provide the next best thing, the user IP.
4162 if ((event->attr.exclude_kernel || !regs) &&
4163 !event->attr.exclude_user)
4164 regs = task_pt_regs(current);
4167 if (!(event->attr.exclude_idle && current->pid == 0))
4168 if (perf_event_overflow(event, 0, &data, regs))
4169 ret = HRTIMER_NORESTART;
4172 period = max_t(u64, 10000, event->hw.sample_period);
4173 hrtimer_forward_now(hrtimer, ns_to_ktime(period));
4178 static void perf_swevent_start_hrtimer(struct perf_event *event)
4180 struct hw_perf_event *hwc = &event->hw;
4182 hrtimer_init(&hwc->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
4183 hwc->hrtimer.function = perf_swevent_hrtimer;
4184 if (hwc->sample_period) {
4187 if (hwc->remaining) {
4188 if (hwc->remaining < 0)
4191 period = hwc->remaining;
4194 period = max_t(u64, 10000, hwc->sample_period);
4196 __hrtimer_start_range_ns(&hwc->hrtimer,
4197 ns_to_ktime(period), 0,
4198 HRTIMER_MODE_REL, 0);
4202 static void perf_swevent_cancel_hrtimer(struct perf_event *event)
4204 struct hw_perf_event *hwc = &event->hw;
4206 if (hwc->sample_period) {
4207 ktime_t remaining = hrtimer_get_remaining(&hwc->hrtimer);
4208 hwc->remaining = ktime_to_ns(remaining);
4210 hrtimer_cancel(&hwc->hrtimer);
4215 * Software event: cpu wall time clock
4218 static void cpu_clock_perf_event_update(struct perf_event *event)
4220 int cpu = raw_smp_processor_id();
4224 now = cpu_clock(cpu);
4225 prev = atomic64_xchg(&event->hw.prev_count, now);
4226 atomic64_add(now - prev, &event->count);
4229 static int cpu_clock_perf_event_enable(struct perf_event *event)
4231 struct hw_perf_event *hwc = &event->hw;
4232 int cpu = raw_smp_processor_id();
4234 atomic64_set(&hwc->prev_count, cpu_clock(cpu));
4235 perf_swevent_start_hrtimer(event);
4240 static void cpu_clock_perf_event_disable(struct perf_event *event)
4242 perf_swevent_cancel_hrtimer(event);
4243 cpu_clock_perf_event_update(event);
4246 static void cpu_clock_perf_event_read(struct perf_event *event)
4248 cpu_clock_perf_event_update(event);
4251 static const struct pmu perf_ops_cpu_clock = {
4252 .enable = cpu_clock_perf_event_enable,
4253 .disable = cpu_clock_perf_event_disable,
4254 .read = cpu_clock_perf_event_read,
4258 * Software event: task time clock
4261 static void task_clock_perf_event_update(struct perf_event *event, u64 now)
4266 prev = atomic64_xchg(&event->hw.prev_count, now);
4268 atomic64_add(delta, &event->count);
4271 static int task_clock_perf_event_enable(struct perf_event *event)
4273 struct hw_perf_event *hwc = &event->hw;
4276 now = event->ctx->time;
4278 atomic64_set(&hwc->prev_count, now);
4280 perf_swevent_start_hrtimer(event);
4285 static void task_clock_perf_event_disable(struct perf_event *event)
4287 perf_swevent_cancel_hrtimer(event);
4288 task_clock_perf_event_update(event, event->ctx->time);
4292 static void task_clock_perf_event_read(struct perf_event *event)
4297 update_context_time(event->ctx);
4298 time = event->ctx->time;
4300 u64 now = perf_clock();
4301 u64 delta = now - event->ctx->timestamp;
4302 time = event->ctx->time + delta;
4305 task_clock_perf_event_update(event, time);
4308 static const struct pmu perf_ops_task_clock = {
4309 .enable = task_clock_perf_event_enable,
4310 .disable = task_clock_perf_event_disable,
4311 .read = task_clock_perf_event_read,
4314 #ifdef CONFIG_EVENT_TRACING
4316 void perf_tp_event(int event_id, u64 addr, u64 count, void *record,
4319 struct pt_regs *regs = get_irq_regs();
4320 struct perf_sample_data data;
4321 struct perf_raw_record raw = {
4326 perf_sample_data_init(&data, addr);
4330 regs = task_pt_regs(current);
4332 /* Trace events already protected against recursion */
4333 do_perf_sw_event(PERF_TYPE_TRACEPOINT, event_id, count, 1,
4336 EXPORT_SYMBOL_GPL(perf_tp_event);
4338 static int perf_tp_event_match(struct perf_event *event,
4339 struct perf_sample_data *data)
4341 void *record = data->raw->data;
4343 if (likely(!event->filter) || filter_match_preds(event->filter, record))
4348 static void tp_perf_event_destroy(struct perf_event *event)
4350 ftrace_profile_disable(event->attr.config);
4353 static const struct pmu *tp_perf_event_init(struct perf_event *event)
4356 * Raw tracepoint data is a severe data leak, only allow root to
4359 if ((event->attr.sample_type & PERF_SAMPLE_RAW) &&
4360 perf_paranoid_tracepoint_raw() &&
4361 !capable(CAP_SYS_ADMIN))
4362 return ERR_PTR(-EPERM);
4364 if (ftrace_profile_enable(event->attr.config))
4367 event->destroy = tp_perf_event_destroy;
4369 return &perf_ops_generic;
4372 static int perf_event_set_filter(struct perf_event *event, void __user *arg)
4377 if (event->attr.type != PERF_TYPE_TRACEPOINT)
4380 filter_str = strndup_user(arg, PAGE_SIZE);
4381 if (IS_ERR(filter_str))
4382 return PTR_ERR(filter_str);
4384 ret = ftrace_profile_set_filter(event, event->attr.config, filter_str);
4390 static void perf_event_free_filter(struct perf_event *event)
4392 ftrace_profile_free_filter(event);
4397 static int perf_tp_event_match(struct perf_event *event,
4398 struct perf_sample_data *data)
4403 static const struct pmu *tp_perf_event_init(struct perf_event *event)
4408 static int perf_event_set_filter(struct perf_event *event, void __user *arg)
4413 static void perf_event_free_filter(struct perf_event *event)
4417 #endif /* CONFIG_EVENT_TRACING */
4419 #ifdef CONFIG_HAVE_HW_BREAKPOINT
4420 static void bp_perf_event_destroy(struct perf_event *event)
4422 release_bp_slot(event);
4425 static const struct pmu *bp_perf_event_init(struct perf_event *bp)
4429 err = register_perf_hw_breakpoint(bp);
4431 return ERR_PTR(err);
4433 bp->destroy = bp_perf_event_destroy;
4435 return &perf_ops_bp;
4438 void perf_bp_event(struct perf_event *bp, void *data)
4440 struct perf_sample_data sample;
4441 struct pt_regs *regs = data;
4443 perf_sample_data_init(&sample, bp->attr.bp_addr);
4445 if (!perf_exclude_event(bp, regs))
4446 perf_swevent_add(bp, 1, 1, &sample, regs);
4449 static const struct pmu *bp_perf_event_init(struct perf_event *bp)
4454 void perf_bp_event(struct perf_event *bp, void *regs)
4459 atomic_t perf_swevent_enabled[PERF_COUNT_SW_MAX];
4461 static void sw_perf_event_destroy(struct perf_event *event)
4463 u64 event_id = event->attr.config;
4465 WARN_ON(event->parent);
4467 atomic_dec(&perf_swevent_enabled[event_id]);
4470 static const struct pmu *sw_perf_event_init(struct perf_event *event)
4472 const struct pmu *pmu = NULL;
4473 u64 event_id = event->attr.config;
4476 * Software events (currently) can't in general distinguish
4477 * between user, kernel and hypervisor events.
4478 * However, context switches and cpu migrations are considered
4479 * to be kernel events, and page faults are never hypervisor
4483 case PERF_COUNT_SW_CPU_CLOCK:
4484 pmu = &perf_ops_cpu_clock;
4487 case PERF_COUNT_SW_TASK_CLOCK:
4489 * If the user instantiates this as a per-cpu event,
4490 * use the cpu_clock event instead.
4492 if (event->ctx->task)
4493 pmu = &perf_ops_task_clock;
4495 pmu = &perf_ops_cpu_clock;
4498 case PERF_COUNT_SW_PAGE_FAULTS:
4499 case PERF_COUNT_SW_PAGE_FAULTS_MIN:
4500 case PERF_COUNT_SW_PAGE_FAULTS_MAJ:
4501 case PERF_COUNT_SW_CONTEXT_SWITCHES:
4502 case PERF_COUNT_SW_CPU_MIGRATIONS:
4503 case PERF_COUNT_SW_ALIGNMENT_FAULTS:
4504 case PERF_COUNT_SW_EMULATION_FAULTS:
4505 if (!event->parent) {
4506 atomic_inc(&perf_swevent_enabled[event_id]);
4507 event->destroy = sw_perf_event_destroy;
4509 pmu = &perf_ops_generic;
4517 * Allocate and initialize a event structure
4519 static struct perf_event *
4520 perf_event_alloc(struct perf_event_attr *attr,
4522 struct perf_event_context *ctx,
4523 struct perf_event *group_leader,
4524 struct perf_event *parent_event,
4525 perf_overflow_handler_t overflow_handler,
4528 const struct pmu *pmu;
4529 struct perf_event *event;
4530 struct hw_perf_event *hwc;
4533 event = kzalloc(sizeof(*event), gfpflags);
4535 return ERR_PTR(-ENOMEM);
4538 * Single events are their own group leaders, with an
4539 * empty sibling list:
4542 group_leader = event;
4544 mutex_init(&event->child_mutex);
4545 INIT_LIST_HEAD(&event->child_list);
4547 INIT_LIST_HEAD(&event->group_entry);
4548 INIT_LIST_HEAD(&event->event_entry);
4549 INIT_LIST_HEAD(&event->sibling_list);
4550 init_waitqueue_head(&event->waitq);
4552 mutex_init(&event->mmap_mutex);
4555 event->attr = *attr;
4556 event->group_leader = group_leader;
4561 event->parent = parent_event;
4563 event->ns = get_pid_ns(current->nsproxy->pid_ns);
4564 event->id = atomic64_inc_return(&perf_event_id);
4566 event->state = PERF_EVENT_STATE_INACTIVE;
4568 if (!overflow_handler && parent_event)
4569 overflow_handler = parent_event->overflow_handler;
4571 event->overflow_handler = overflow_handler;
4574 event->state = PERF_EVENT_STATE_OFF;
4579 hwc->sample_period = attr->sample_period;
4580 if (attr->freq && attr->sample_freq)
4581 hwc->sample_period = 1;
4582 hwc->last_period = hwc->sample_period;
4584 atomic64_set(&hwc->period_left, hwc->sample_period);
4587 * we currently do not support PERF_FORMAT_GROUP on inherited events
4589 if (attr->inherit && (attr->read_format & PERF_FORMAT_GROUP))
4592 switch (attr->type) {
4594 case PERF_TYPE_HARDWARE:
4595 case PERF_TYPE_HW_CACHE:
4596 pmu = hw_perf_event_init(event);
4599 case PERF_TYPE_SOFTWARE:
4600 pmu = sw_perf_event_init(event);
4603 case PERF_TYPE_TRACEPOINT:
4604 pmu = tp_perf_event_init(event);
4607 case PERF_TYPE_BREAKPOINT:
4608 pmu = bp_perf_event_init(event);
4619 else if (IS_ERR(pmu))
4624 put_pid_ns(event->ns);
4626 return ERR_PTR(err);
4631 if (!event->parent) {
4632 atomic_inc(&nr_events);
4633 if (event->attr.mmap)
4634 atomic_inc(&nr_mmap_events);
4635 if (event->attr.comm)
4636 atomic_inc(&nr_comm_events);
4637 if (event->attr.task)
4638 atomic_inc(&nr_task_events);
4644 static int perf_copy_attr(struct perf_event_attr __user *uattr,
4645 struct perf_event_attr *attr)
4650 if (!access_ok(VERIFY_WRITE, uattr, PERF_ATTR_SIZE_VER0))
4654 * zero the full structure, so that a short copy will be nice.
4656 memset(attr, 0, sizeof(*attr));
4658 ret = get_user(size, &uattr->size);
4662 if (size > PAGE_SIZE) /* silly large */
4665 if (!size) /* abi compat */
4666 size = PERF_ATTR_SIZE_VER0;
4668 if (size < PERF_ATTR_SIZE_VER0)
4672 * If we're handed a bigger struct than we know of,
4673 * ensure all the unknown bits are 0 - i.e. new
4674 * user-space does not rely on any kernel feature
4675 * extensions we dont know about yet.
4677 if (size > sizeof(*attr)) {
4678 unsigned char __user *addr;
4679 unsigned char __user *end;
4682 addr = (void __user *)uattr + sizeof(*attr);
4683 end = (void __user *)uattr + size;
4685 for (; addr < end; addr++) {
4686 ret = get_user(val, addr);
4692 size = sizeof(*attr);
4695 ret = copy_from_user(attr, uattr, size);
4700 * If the type exists, the corresponding creation will verify
4703 if (attr->type >= PERF_TYPE_MAX)
4706 if (attr->__reserved_1)
4709 if (attr->sample_type & ~(PERF_SAMPLE_MAX-1))
4712 if (attr->read_format & ~(PERF_FORMAT_MAX-1))
4719 put_user(sizeof(*attr), &uattr->size);
4724 static int perf_event_set_output(struct perf_event *event, int output_fd)
4726 struct perf_event *output_event = NULL;
4727 struct file *output_file = NULL;
4728 struct perf_event *old_output;
4729 int fput_needed = 0;
4735 output_file = fget_light(output_fd, &fput_needed);
4739 if (output_file->f_op != &perf_fops)
4742 output_event = output_file->private_data;
4744 /* Don't chain output fds */
4745 if (output_event->output)
4748 /* Don't set an output fd when we already have an output channel */
4752 atomic_long_inc(&output_file->f_count);
4755 mutex_lock(&event->mmap_mutex);
4756 old_output = event->output;
4757 rcu_assign_pointer(event->output, output_event);
4758 mutex_unlock(&event->mmap_mutex);
4762 * we need to make sure no existing perf_output_*()
4763 * is still referencing this event.
4766 fput(old_output->filp);
4771 fput_light(output_file, fput_needed);
4776 * sys_perf_event_open - open a performance event, associate it to a task/cpu
4778 * @attr_uptr: event_id type attributes for monitoring/sampling
4781 * @group_fd: group leader event fd
4783 SYSCALL_DEFINE5(perf_event_open,
4784 struct perf_event_attr __user *, attr_uptr,
4785 pid_t, pid, int, cpu, int, group_fd, unsigned long, flags)
4787 struct perf_event *event, *group_leader;
4788 struct perf_event_attr attr;
4789 struct perf_event_context *ctx;
4790 struct file *event_file = NULL;
4791 struct file *group_file = NULL;
4792 int fput_needed = 0;
4793 int fput_needed2 = 0;
4796 /* for future expandability... */
4797 if (flags & ~(PERF_FLAG_FD_NO_GROUP | PERF_FLAG_FD_OUTPUT))
4800 err = perf_copy_attr(attr_uptr, &attr);
4804 if (!attr.exclude_kernel) {
4805 if (perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN))
4810 if (attr.sample_freq > sysctl_perf_event_sample_rate)
4815 * Get the target context (task or percpu):
4817 ctx = find_get_context(pid, cpu);
4819 return PTR_ERR(ctx);
4822 * Look up the group leader (we will attach this event to it):
4824 group_leader = NULL;
4825 if (group_fd != -1 && !(flags & PERF_FLAG_FD_NO_GROUP)) {
4827 group_file = fget_light(group_fd, &fput_needed);
4829 goto err_put_context;
4830 if (group_file->f_op != &perf_fops)
4831 goto err_put_context;
4833 group_leader = group_file->private_data;
4835 * Do not allow a recursive hierarchy (this new sibling
4836 * becoming part of another group-sibling):
4838 if (group_leader->group_leader != group_leader)
4839 goto err_put_context;
4841 * Do not allow to attach to a group in a different
4842 * task or CPU context:
4844 if (group_leader->ctx != ctx)
4845 goto err_put_context;
4847 * Only a group leader can be exclusive or pinned
4849 if (attr.exclusive || attr.pinned)
4850 goto err_put_context;
4853 event = perf_event_alloc(&attr, cpu, ctx, group_leader,
4854 NULL, NULL, GFP_KERNEL);
4855 err = PTR_ERR(event);
4857 goto err_put_context;
4859 err = anon_inode_getfd("[perf_event]", &perf_fops, event, O_RDWR);
4861 goto err_free_put_context;
4863 event_file = fget_light(err, &fput_needed2);
4865 goto err_free_put_context;
4867 if (flags & PERF_FLAG_FD_OUTPUT) {
4868 err = perf_event_set_output(event, group_fd);
4870 goto err_fput_free_put_context;
4873 event->filp = event_file;
4874 WARN_ON_ONCE(ctx->parent_ctx);
4875 mutex_lock(&ctx->mutex);
4876 perf_install_in_context(ctx, event, cpu);
4878 mutex_unlock(&ctx->mutex);
4880 event->owner = current;
4881 get_task_struct(current);
4882 mutex_lock(¤t->perf_event_mutex);
4883 list_add_tail(&event->owner_entry, ¤t->perf_event_list);
4884 mutex_unlock(¤t->perf_event_mutex);
4886 err_fput_free_put_context:
4887 fput_light(event_file, fput_needed2);
4889 err_free_put_context:
4897 fput_light(group_file, fput_needed);
4903 * perf_event_create_kernel_counter
4905 * @attr: attributes of the counter to create
4906 * @cpu: cpu in which the counter is bound
4907 * @pid: task to profile
4910 perf_event_create_kernel_counter(struct perf_event_attr *attr, int cpu,
4912 perf_overflow_handler_t overflow_handler)
4914 struct perf_event *event;
4915 struct perf_event_context *ctx;
4919 * Get the target context (task or percpu):
4922 ctx = find_get_context(pid, cpu);
4928 event = perf_event_alloc(attr, cpu, ctx, NULL,
4929 NULL, overflow_handler, GFP_KERNEL);
4930 if (IS_ERR(event)) {
4931 err = PTR_ERR(event);
4932 goto err_put_context;
4936 WARN_ON_ONCE(ctx->parent_ctx);
4937 mutex_lock(&ctx->mutex);
4938 perf_install_in_context(ctx, event, cpu);
4940 mutex_unlock(&ctx->mutex);
4942 event->owner = current;
4943 get_task_struct(current);
4944 mutex_lock(¤t->perf_event_mutex);
4945 list_add_tail(&event->owner_entry, ¤t->perf_event_list);
4946 mutex_unlock(¤t->perf_event_mutex);
4953 return ERR_PTR(err);
4955 EXPORT_SYMBOL_GPL(perf_event_create_kernel_counter);
4958 * inherit a event from parent task to child task:
4960 static struct perf_event *
4961 inherit_event(struct perf_event *parent_event,
4962 struct task_struct *parent,
4963 struct perf_event_context *parent_ctx,
4964 struct task_struct *child,
4965 struct perf_event *group_leader,
4966 struct perf_event_context *child_ctx)
4968 struct perf_event *child_event;
4971 * Instead of creating recursive hierarchies of events,
4972 * we link inherited events back to the original parent,
4973 * which has a filp for sure, which we use as the reference
4976 if (parent_event->parent)
4977 parent_event = parent_event->parent;
4979 child_event = perf_event_alloc(&parent_event->attr,
4980 parent_event->cpu, child_ctx,
4981 group_leader, parent_event,
4983 if (IS_ERR(child_event))
4988 * Make the child state follow the state of the parent event,
4989 * not its attr.disabled bit. We hold the parent's mutex,
4990 * so we won't race with perf_event_{en, dis}able_family.
4992 if (parent_event->state >= PERF_EVENT_STATE_INACTIVE)
4993 child_event->state = PERF_EVENT_STATE_INACTIVE;
4995 child_event->state = PERF_EVENT_STATE_OFF;
4997 if (parent_event->attr.freq) {
4998 u64 sample_period = parent_event->hw.sample_period;
4999 struct hw_perf_event *hwc = &child_event->hw;
5001 hwc->sample_period = sample_period;
5002 hwc->last_period = sample_period;
5004 atomic64_set(&hwc->period_left, sample_period);
5007 child_event->overflow_handler = parent_event->overflow_handler;
5010 * Link it up in the child's context:
5012 add_event_to_ctx(child_event, child_ctx);
5015 * Get a reference to the parent filp - we will fput it
5016 * when the child event exits. This is safe to do because
5017 * we are in the parent and we know that the filp still
5018 * exists and has a nonzero count:
5020 atomic_long_inc(&parent_event->filp->f_count);
5023 * Link this into the parent event's child list
5025 WARN_ON_ONCE(parent_event->ctx->parent_ctx);
5026 mutex_lock(&parent_event->child_mutex);
5027 list_add_tail(&child_event->child_list, &parent_event->child_list);
5028 mutex_unlock(&parent_event->child_mutex);
5033 static int inherit_group(struct perf_event *parent_event,
5034 struct task_struct *parent,
5035 struct perf_event_context *parent_ctx,
5036 struct task_struct *child,
5037 struct perf_event_context *child_ctx)
5039 struct perf_event *leader;
5040 struct perf_event *sub;
5041 struct perf_event *child_ctr;
5043 leader = inherit_event(parent_event, parent, parent_ctx,
5044 child, NULL, child_ctx);
5046 return PTR_ERR(leader);
5047 list_for_each_entry(sub, &parent_event->sibling_list, group_entry) {
5048 child_ctr = inherit_event(sub, parent, parent_ctx,
5049 child, leader, child_ctx);
5050 if (IS_ERR(child_ctr))
5051 return PTR_ERR(child_ctr);
5056 static void sync_child_event(struct perf_event *child_event,
5057 struct task_struct *child)
5059 struct perf_event *parent_event = child_event->parent;
5062 if (child_event->attr.inherit_stat)
5063 perf_event_read_event(child_event, child);
5065 child_val = atomic64_read(&child_event->count);
5068 * Add back the child's count to the parent's count:
5070 atomic64_add(child_val, &parent_event->count);
5071 atomic64_add(child_event->total_time_enabled,
5072 &parent_event->child_total_time_enabled);
5073 atomic64_add(child_event->total_time_running,
5074 &parent_event->child_total_time_running);
5077 * Remove this event from the parent's list
5079 WARN_ON_ONCE(parent_event->ctx->parent_ctx);
5080 mutex_lock(&parent_event->child_mutex);
5081 list_del_init(&child_event->child_list);
5082 mutex_unlock(&parent_event->child_mutex);
5085 * Release the parent event, if this was the last
5088 fput(parent_event->filp);
5092 __perf_event_exit_task(struct perf_event *child_event,
5093 struct perf_event_context *child_ctx,
5094 struct task_struct *child)
5096 struct perf_event *parent_event;
5098 perf_event_remove_from_context(child_event);
5100 parent_event = child_event->parent;
5102 * It can happen that parent exits first, and has events
5103 * that are still around due to the child reference. These
5104 * events need to be zapped - but otherwise linger.
5107 sync_child_event(child_event, child);
5108 free_event(child_event);
5113 * When a child task exits, feed back event values to parent events.
5115 void perf_event_exit_task(struct task_struct *child)
5117 struct perf_event *child_event, *tmp;
5118 struct perf_event_context *child_ctx;
5119 unsigned long flags;
5121 if (likely(!child->perf_event_ctxp)) {
5122 perf_event_task(child, NULL, 0);
5126 local_irq_save(flags);
5128 * We can't reschedule here because interrupts are disabled,
5129 * and either child is current or it is a task that can't be
5130 * scheduled, so we are now safe from rescheduling changing
5133 child_ctx = child->perf_event_ctxp;
5134 __perf_event_task_sched_out(child_ctx);
5137 * Take the context lock here so that if find_get_context is
5138 * reading child->perf_event_ctxp, we wait until it has
5139 * incremented the context's refcount before we do put_ctx below.
5141 raw_spin_lock(&child_ctx->lock);
5142 child->perf_event_ctxp = NULL;
5144 * If this context is a clone; unclone it so it can't get
5145 * swapped to another process while we're removing all
5146 * the events from it.
5148 unclone_ctx(child_ctx);
5149 update_context_time(child_ctx);
5150 raw_spin_unlock_irqrestore(&child_ctx->lock, flags);
5153 * Report the task dead after unscheduling the events so that we
5154 * won't get any samples after PERF_RECORD_EXIT. We can however still
5155 * get a few PERF_RECORD_READ events.
5157 perf_event_task(child, child_ctx, 0);
5160 * We can recurse on the same lock type through:
5162 * __perf_event_exit_task()
5163 * sync_child_event()
5164 * fput(parent_event->filp)
5166 * mutex_lock(&ctx->mutex)
5168 * But since its the parent context it won't be the same instance.
5170 mutex_lock_nested(&child_ctx->mutex, SINGLE_DEPTH_NESTING);
5173 list_for_each_entry_safe(child_event, tmp, &child_ctx->pinned_groups,
5175 __perf_event_exit_task(child_event, child_ctx, child);
5177 list_for_each_entry_safe(child_event, tmp, &child_ctx->flexible_groups,
5179 __perf_event_exit_task(child_event, child_ctx, child);
5182 * If the last event was a group event, it will have appended all
5183 * its siblings to the list, but we obtained 'tmp' before that which
5184 * will still point to the list head terminating the iteration.
5186 if (!list_empty(&child_ctx->pinned_groups) ||
5187 !list_empty(&child_ctx->flexible_groups))
5190 mutex_unlock(&child_ctx->mutex);
5195 static void perf_free_event(struct perf_event *event,
5196 struct perf_event_context *ctx)
5198 struct perf_event *parent = event->parent;
5200 if (WARN_ON_ONCE(!parent))
5203 mutex_lock(&parent->child_mutex);
5204 list_del_init(&event->child_list);
5205 mutex_unlock(&parent->child_mutex);
5209 list_del_event(event, ctx);
5214 * free an unexposed, unused context as created by inheritance by
5215 * init_task below, used by fork() in case of fail.
5217 void perf_event_free_task(struct task_struct *task)
5219 struct perf_event_context *ctx = task->perf_event_ctxp;
5220 struct perf_event *event, *tmp;
5225 mutex_lock(&ctx->mutex);
5227 list_for_each_entry_safe(event, tmp, &ctx->pinned_groups, group_entry)
5228 perf_free_event(event, ctx);
5230 list_for_each_entry_safe(event, tmp, &ctx->flexible_groups,
5232 perf_free_event(event, ctx);
5234 if (!list_empty(&ctx->pinned_groups) ||
5235 !list_empty(&ctx->flexible_groups))
5238 mutex_unlock(&ctx->mutex);
5244 inherit_task_group(struct perf_event *event, struct task_struct *parent,
5245 struct perf_event_context *parent_ctx,
5246 struct task_struct *child,
5250 struct perf_event_context *child_ctx = child->perf_event_ctxp;
5252 if (!event->attr.inherit) {
5259 * This is executed from the parent task context, so
5260 * inherit events that have been marked for cloning.
5261 * First allocate and initialize a context for the
5265 child_ctx = kzalloc(sizeof(struct perf_event_context),
5270 __perf_event_init_context(child_ctx, child);
5271 child->perf_event_ctxp = child_ctx;
5272 get_task_struct(child);
5275 ret = inherit_group(event, parent, parent_ctx,
5286 * Initialize the perf_event context in task_struct
5288 int perf_event_init_task(struct task_struct *child)
5290 struct perf_event_context *child_ctx, *parent_ctx;
5291 struct perf_event_context *cloned_ctx;
5292 struct perf_event *event;
5293 struct task_struct *parent = current;
5294 int inherited_all = 1;
5297 child->perf_event_ctxp = NULL;
5299 mutex_init(&child->perf_event_mutex);
5300 INIT_LIST_HEAD(&child->perf_event_list);
5302 if (likely(!parent->perf_event_ctxp))
5306 * If the parent's context is a clone, pin it so it won't get
5309 parent_ctx = perf_pin_task_context(parent);
5312 * No need to check if parent_ctx != NULL here; since we saw
5313 * it non-NULL earlier, the only reason for it to become NULL
5314 * is if we exit, and since we're currently in the middle of
5315 * a fork we can't be exiting at the same time.
5319 * Lock the parent list. No need to lock the child - not PID
5320 * hashed yet and not running, so nobody can access it.
5322 mutex_lock(&parent_ctx->mutex);
5325 * We dont have to disable NMIs - we are only looking at
5326 * the list, not manipulating it:
5328 list_for_each_entry(event, &parent_ctx->pinned_groups, group_entry) {
5329 ret = inherit_task_group(event, parent, parent_ctx, child,
5335 list_for_each_entry(event, &parent_ctx->flexible_groups, group_entry) {
5336 ret = inherit_task_group(event, parent, parent_ctx, child,
5342 child_ctx = child->perf_event_ctxp;
5344 if (child_ctx && inherited_all) {
5346 * Mark the child context as a clone of the parent
5347 * context, or of whatever the parent is a clone of.
5348 * Note that if the parent is a clone, it could get
5349 * uncloned at any point, but that doesn't matter
5350 * because the list of events and the generation
5351 * count can't have changed since we took the mutex.
5353 cloned_ctx = rcu_dereference(parent_ctx->parent_ctx);
5355 child_ctx->parent_ctx = cloned_ctx;
5356 child_ctx->parent_gen = parent_ctx->parent_gen;
5358 child_ctx->parent_ctx = parent_ctx;
5359 child_ctx->parent_gen = parent_ctx->generation;
5361 get_ctx(child_ctx->parent_ctx);
5364 mutex_unlock(&parent_ctx->mutex);
5366 perf_unpin_context(parent_ctx);
5371 static void __cpuinit perf_event_init_cpu(int cpu)
5373 struct perf_cpu_context *cpuctx;
5375 cpuctx = &per_cpu(perf_cpu_context, cpu);
5376 __perf_event_init_context(&cpuctx->ctx, NULL);
5378 spin_lock(&perf_resource_lock);
5379 cpuctx->max_pertask = perf_max_events - perf_reserved_percpu;
5380 spin_unlock(&perf_resource_lock);
5383 #ifdef CONFIG_HOTPLUG_CPU
5384 static void __perf_event_exit_cpu(void *info)
5386 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
5387 struct perf_event_context *ctx = &cpuctx->ctx;
5388 struct perf_event *event, *tmp;
5390 list_for_each_entry_safe(event, tmp, &ctx->pinned_groups, group_entry)
5391 __perf_event_remove_from_context(event);
5392 list_for_each_entry_safe(event, tmp, &ctx->flexible_groups, group_entry)
5393 __perf_event_remove_from_context(event);
5395 static void perf_event_exit_cpu(int cpu)
5397 struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
5398 struct perf_event_context *ctx = &cpuctx->ctx;
5400 mutex_lock(&ctx->mutex);
5401 smp_call_function_single(cpu, __perf_event_exit_cpu, NULL, 1);
5402 mutex_unlock(&ctx->mutex);
5405 static inline void perf_event_exit_cpu(int cpu) { }
5408 static int __cpuinit
5409 perf_cpu_notify(struct notifier_block *self, unsigned long action, void *hcpu)
5411 unsigned int cpu = (long)hcpu;
5415 case CPU_UP_PREPARE:
5416 case CPU_UP_PREPARE_FROZEN:
5417 perf_event_init_cpu(cpu);
5420 case CPU_DOWN_PREPARE:
5421 case CPU_DOWN_PREPARE_FROZEN:
5422 perf_event_exit_cpu(cpu);
5433 * This has to have a higher priority than migration_notifier in sched.c.
5435 static struct notifier_block __cpuinitdata perf_cpu_nb = {
5436 .notifier_call = perf_cpu_notify,
5440 void __init perf_event_init(void)
5442 perf_cpu_notify(&perf_cpu_nb, (unsigned long)CPU_UP_PREPARE,
5443 (void *)(long)smp_processor_id());
5444 perf_cpu_notify(&perf_cpu_nb, (unsigned long)CPU_ONLINE,
5445 (void *)(long)smp_processor_id());
5446 register_cpu_notifier(&perf_cpu_nb);
5449 static ssize_t perf_show_reserve_percpu(struct sysdev_class *class,
5450 struct sysdev_class_attribute *attr,
5453 return sprintf(buf, "%d\n", perf_reserved_percpu);
5457 perf_set_reserve_percpu(struct sysdev_class *class,
5458 struct sysdev_class_attribute *attr,
5462 struct perf_cpu_context *cpuctx;
5466 err = strict_strtoul(buf, 10, &val);
5469 if (val > perf_max_events)
5472 spin_lock(&perf_resource_lock);
5473 perf_reserved_percpu = val;
5474 for_each_online_cpu(cpu) {
5475 cpuctx = &per_cpu(perf_cpu_context, cpu);
5476 raw_spin_lock_irq(&cpuctx->ctx.lock);
5477 mpt = min(perf_max_events - cpuctx->ctx.nr_events,
5478 perf_max_events - perf_reserved_percpu);
5479 cpuctx->max_pertask = mpt;
5480 raw_spin_unlock_irq(&cpuctx->ctx.lock);
5482 spin_unlock(&perf_resource_lock);
5487 static ssize_t perf_show_overcommit(struct sysdev_class *class,
5488 struct sysdev_class_attribute *attr,
5491 return sprintf(buf, "%d\n", perf_overcommit);
5495 perf_set_overcommit(struct sysdev_class *class,
5496 struct sysdev_class_attribute *attr,
5497 const char *buf, size_t count)
5502 err = strict_strtoul(buf, 10, &val);
5508 spin_lock(&perf_resource_lock);
5509 perf_overcommit = val;
5510 spin_unlock(&perf_resource_lock);
5515 static SYSDEV_CLASS_ATTR(
5518 perf_show_reserve_percpu,
5519 perf_set_reserve_percpu
5522 static SYSDEV_CLASS_ATTR(
5525 perf_show_overcommit,
5529 static struct attribute *perfclass_attrs[] = {
5530 &attr_reserve_percpu.attr,
5531 &attr_overcommit.attr,
5535 static struct attribute_group perfclass_attr_group = {
5536 .attrs = perfclass_attrs,
5537 .name = "perf_events",
5540 static int __init perf_event_sysfs_init(void)
5542 return sysfs_create_group(&cpu_sysdev_class.kset.kobj,
5543 &perfclass_attr_group);
5545 device_initcall(perf_event_sysfs_init);