2 * Performance events core code:
4 * Copyright (C) 2008 Thomas Gleixner <tglx@linutronix.de>
5 * Copyright (C) 2008-2009 Red Hat, Inc., Ingo Molnar
6 * Copyright (C) 2008-2009 Red Hat, Inc., Peter Zijlstra <pzijlstr@redhat.com>
7 * Copyright © 2009 Paul Mackerras, IBM Corp. <paulus@au1.ibm.com>
9 * For licensing details see kernel-base/COPYING
14 #include <linux/cpu.h>
15 #include <linux/smp.h>
16 #include <linux/file.h>
17 #include <linux/poll.h>
18 #include <linux/sysfs.h>
19 #include <linux/dcache.h>
20 #include <linux/percpu.h>
21 #include <linux/ptrace.h>
22 #include <linux/vmstat.h>
23 #include <linux/vmalloc.h>
24 #include <linux/hardirq.h>
25 #include <linux/rculist.h>
26 #include <linux/uaccess.h>
27 #include <linux/syscalls.h>
28 #include <linux/anon_inodes.h>
29 #include <linux/kernel_stat.h>
30 #include <linux/perf_event.h>
31 #include <linux/ftrace_event.h>
32 #include <linux/hw_breakpoint.h>
34 #include <asm/irq_regs.h>
37 * Each CPU has a list of per CPU events:
39 static DEFINE_PER_CPU(struct perf_cpu_context
, perf_cpu_context
);
41 int perf_max_events __read_mostly
= 1;
42 static int perf_reserved_percpu __read_mostly
;
43 static int perf_overcommit __read_mostly
= 1;
45 static atomic_t nr_events __read_mostly
;
46 static atomic_t nr_mmap_events __read_mostly
;
47 static atomic_t nr_comm_events __read_mostly
;
48 static atomic_t nr_task_events __read_mostly
;
51 * perf event paranoia level:
52 * -1 - not paranoid at all
53 * 0 - disallow raw tracepoint access for unpriv
54 * 1 - disallow cpu events for unpriv
55 * 2 - disallow kernel profiling for unpriv
57 int sysctl_perf_event_paranoid __read_mostly
= 1;
59 int sysctl_perf_event_mlock __read_mostly
= 512; /* 'free' kb per user */
62 * max perf event sample rate
64 int sysctl_perf_event_sample_rate __read_mostly
= 100000;
66 static atomic64_t perf_event_id
;
69 * Lock for (sysadmin-configurable) event reservations:
71 static DEFINE_SPINLOCK(perf_resource_lock
);
74 * Architecture provided APIs - weak aliases:
76 extern __weak
const struct pmu
*hw_perf_event_init(struct perf_event
*event
)
81 void __weak
hw_perf_disable(void) { barrier(); }
82 void __weak
hw_perf_enable(void) { barrier(); }
85 hw_perf_group_sched_in(struct perf_event
*group_leader
,
86 struct perf_cpu_context
*cpuctx
,
87 struct perf_event_context
*ctx
)
92 void __weak
perf_event_print_debug(void) { }
94 static DEFINE_PER_CPU(int, perf_disable_count
);
96 void perf_disable(void)
98 if (!__get_cpu_var(perf_disable_count
)++)
102 void perf_enable(void)
104 if (!--__get_cpu_var(perf_disable_count
))
108 static void get_ctx(struct perf_event_context
*ctx
)
110 WARN_ON(!atomic_inc_not_zero(&ctx
->refcount
));
113 static void free_ctx(struct rcu_head
*head
)
115 struct perf_event_context
*ctx
;
117 ctx
= container_of(head
, struct perf_event_context
, rcu_head
);
121 static void put_ctx(struct perf_event_context
*ctx
)
123 if (atomic_dec_and_test(&ctx
->refcount
)) {
125 put_ctx(ctx
->parent_ctx
);
127 put_task_struct(ctx
->task
);
128 call_rcu(&ctx
->rcu_head
, free_ctx
);
132 static void unclone_ctx(struct perf_event_context
*ctx
)
134 if (ctx
->parent_ctx
) {
135 put_ctx(ctx
->parent_ctx
);
136 ctx
->parent_ctx
= NULL
;
141 * If we inherit events we want to return the parent event id
144 static u64
primary_event_id(struct perf_event
*event
)
149 id
= event
->parent
->id
;
155 * Get the perf_event_context for a task and lock it.
156 * This has to cope with with the fact that until it is locked,
157 * the context could get moved to another task.
159 static struct perf_event_context
*
160 perf_lock_task_context(struct task_struct
*task
, unsigned long *flags
)
162 struct perf_event_context
*ctx
;
166 ctx
= rcu_dereference(task
->perf_event_ctxp
);
169 * If this context is a clone of another, it might
170 * get swapped for another underneath us by
171 * perf_event_task_sched_out, though the
172 * rcu_read_lock() protects us from any context
173 * getting freed. Lock the context and check if it
174 * got swapped before we could get the lock, and retry
175 * if so. If we locked the right context, then it
176 * can't get swapped on us any more.
178 raw_spin_lock_irqsave(&ctx
->lock
, *flags
);
179 if (ctx
!= rcu_dereference(task
->perf_event_ctxp
)) {
180 raw_spin_unlock_irqrestore(&ctx
->lock
, *flags
);
184 if (!atomic_inc_not_zero(&ctx
->refcount
)) {
185 raw_spin_unlock_irqrestore(&ctx
->lock
, *flags
);
194 * Get the context for a task and increment its pin_count so it
195 * can't get swapped to another task. This also increments its
196 * reference count so that the context can't get freed.
198 static struct perf_event_context
*perf_pin_task_context(struct task_struct
*task
)
200 struct perf_event_context
*ctx
;
203 ctx
= perf_lock_task_context(task
, &flags
);
206 raw_spin_unlock_irqrestore(&ctx
->lock
, flags
);
211 static void perf_unpin_context(struct perf_event_context
*ctx
)
215 raw_spin_lock_irqsave(&ctx
->lock
, flags
);
217 raw_spin_unlock_irqrestore(&ctx
->lock
, flags
);
221 static inline u64
perf_clock(void)
223 return cpu_clock(raw_smp_processor_id());
227 * Update the record of the current time in a context.
229 static void update_context_time(struct perf_event_context
*ctx
)
231 u64 now
= perf_clock();
233 ctx
->time
+= now
- ctx
->timestamp
;
234 ctx
->timestamp
= now
;
238 * Update the total_time_enabled and total_time_running fields for a event.
240 static void update_event_times(struct perf_event
*event
)
242 struct perf_event_context
*ctx
= event
->ctx
;
245 if (event
->state
< PERF_EVENT_STATE_INACTIVE
||
246 event
->group_leader
->state
< PERF_EVENT_STATE_INACTIVE
)
252 run_end
= event
->tstamp_stopped
;
254 event
->total_time_enabled
= run_end
- event
->tstamp_enabled
;
256 if (event
->state
== PERF_EVENT_STATE_INACTIVE
)
257 run_end
= event
->tstamp_stopped
;
261 event
->total_time_running
= run_end
- event
->tstamp_running
;
264 static struct list_head
*
265 ctx_group_list(struct perf_event
*event
, struct perf_event_context
*ctx
)
267 if (event
->attr
.pinned
)
268 return &ctx
->pinned_groups
;
270 return &ctx
->flexible_groups
;
274 * Add a event from the lists for its context.
275 * Must be called with ctx->mutex and ctx->lock held.
278 list_add_event(struct perf_event
*event
, struct perf_event_context
*ctx
)
280 struct perf_event
*group_leader
= event
->group_leader
;
283 * Depending on whether it is a standalone or sibling event,
284 * add it straight to the context's event list, or to the group
285 * leader's sibling list:
287 if (group_leader
== event
) {
288 struct list_head
*list
;
290 if (is_software_event(event
))
291 event
->group_flags
|= PERF_GROUP_SOFTWARE
;
293 list
= ctx_group_list(event
, ctx
);
294 list_add_tail(&event
->group_entry
, list
);
296 if (group_leader
->group_flags
& PERF_GROUP_SOFTWARE
&&
297 !is_software_event(event
))
298 group_leader
->group_flags
&= ~PERF_GROUP_SOFTWARE
;
300 list_add_tail(&event
->group_entry
, &group_leader
->sibling_list
);
301 group_leader
->nr_siblings
++;
304 list_add_rcu(&event
->event_entry
, &ctx
->event_list
);
306 if (event
->attr
.inherit_stat
)
311 * Remove a event from the lists for its context.
312 * Must be called with ctx->mutex and ctx->lock held.
315 list_del_event(struct perf_event
*event
, struct perf_event_context
*ctx
)
317 struct perf_event
*sibling
, *tmp
;
319 if (list_empty(&event
->group_entry
))
322 if (event
->attr
.inherit_stat
)
325 list_del_init(&event
->group_entry
);
326 list_del_rcu(&event
->event_entry
);
328 if (event
->group_leader
!= event
)
329 event
->group_leader
->nr_siblings
--;
331 update_event_times(event
);
334 * If event was in error state, then keep it
335 * that way, otherwise bogus counts will be
336 * returned on read(). The only way to get out
337 * of error state is by explicit re-enabling
340 if (event
->state
> PERF_EVENT_STATE_OFF
)
341 event
->state
= PERF_EVENT_STATE_OFF
;
344 * If this was a group event with sibling events then
345 * upgrade the siblings to singleton events by adding them
346 * to the context list directly:
348 list_for_each_entry_safe(sibling
, tmp
, &event
->sibling_list
, group_entry
) {
349 struct list_head
*list
;
351 list
= ctx_group_list(event
, ctx
);
352 list_move_tail(&sibling
->group_entry
, list
);
353 sibling
->group_leader
= sibling
;
355 /* Inherit group flags from the previous leader */
356 sibling
->group_flags
= event
->group_flags
;
361 event_sched_out(struct perf_event
*event
,
362 struct perf_cpu_context
*cpuctx
,
363 struct perf_event_context
*ctx
)
365 if (event
->state
!= PERF_EVENT_STATE_ACTIVE
)
368 event
->state
= PERF_EVENT_STATE_INACTIVE
;
369 if (event
->pending_disable
) {
370 event
->pending_disable
= 0;
371 event
->state
= PERF_EVENT_STATE_OFF
;
373 event
->tstamp_stopped
= ctx
->time
;
374 event
->pmu
->disable(event
);
377 if (!is_software_event(event
))
378 cpuctx
->active_oncpu
--;
380 if (event
->attr
.exclusive
|| !cpuctx
->active_oncpu
)
381 cpuctx
->exclusive
= 0;
385 group_sched_out(struct perf_event
*group_event
,
386 struct perf_cpu_context
*cpuctx
,
387 struct perf_event_context
*ctx
)
389 struct perf_event
*event
;
391 if (group_event
->state
!= PERF_EVENT_STATE_ACTIVE
)
394 event_sched_out(group_event
, cpuctx
, ctx
);
397 * Schedule out siblings (if any):
399 list_for_each_entry(event
, &group_event
->sibling_list
, group_entry
)
400 event_sched_out(event
, cpuctx
, ctx
);
402 if (group_event
->attr
.exclusive
)
403 cpuctx
->exclusive
= 0;
407 * Cross CPU call to remove a performance event
409 * We disable the event on the hardware level first. After that we
410 * remove it from the context list.
412 static void __perf_event_remove_from_context(void *info
)
414 struct perf_cpu_context
*cpuctx
= &__get_cpu_var(perf_cpu_context
);
415 struct perf_event
*event
= info
;
416 struct perf_event_context
*ctx
= event
->ctx
;
419 * If this is a task context, we need to check whether it is
420 * the current task context of this cpu. If not it has been
421 * scheduled out before the smp call arrived.
423 if (ctx
->task
&& cpuctx
->task_ctx
!= ctx
)
426 raw_spin_lock(&ctx
->lock
);
428 * Protect the list operation against NMI by disabling the
429 * events on a global level.
433 event_sched_out(event
, cpuctx
, ctx
);
435 list_del_event(event
, ctx
);
439 * Allow more per task events with respect to the
442 cpuctx
->max_pertask
=
443 min(perf_max_events
- ctx
->nr_events
,
444 perf_max_events
- perf_reserved_percpu
);
448 raw_spin_unlock(&ctx
->lock
);
453 * Remove the event from a task's (or a CPU's) list of events.
455 * Must be called with ctx->mutex held.
457 * CPU events are removed with a smp call. For task events we only
458 * call when the task is on a CPU.
460 * If event->ctx is a cloned context, callers must make sure that
461 * every task struct that event->ctx->task could possibly point to
462 * remains valid. This is OK when called from perf_release since
463 * that only calls us on the top-level context, which can't be a clone.
464 * When called from perf_event_exit_task, it's OK because the
465 * context has been detached from its task.
467 static void perf_event_remove_from_context(struct perf_event
*event
)
469 struct perf_event_context
*ctx
= event
->ctx
;
470 struct task_struct
*task
= ctx
->task
;
474 * Per cpu events are removed via an smp call and
475 * the removal is always successful.
477 smp_call_function_single(event
->cpu
,
478 __perf_event_remove_from_context
,
484 task_oncpu_function_call(task
, __perf_event_remove_from_context
,
487 raw_spin_lock_irq(&ctx
->lock
);
489 * If the context is active we need to retry the smp call.
491 if (ctx
->nr_active
&& !list_empty(&event
->group_entry
)) {
492 raw_spin_unlock_irq(&ctx
->lock
);
497 * The lock prevents that this context is scheduled in so we
498 * can remove the event safely, if the call above did not
501 if (!list_empty(&event
->group_entry
))
502 list_del_event(event
, ctx
);
503 raw_spin_unlock_irq(&ctx
->lock
);
507 * Update total_time_enabled and total_time_running for all events in a group.
509 static void update_group_times(struct perf_event
*leader
)
511 struct perf_event
*event
;
513 update_event_times(leader
);
514 list_for_each_entry(event
, &leader
->sibling_list
, group_entry
)
515 update_event_times(event
);
519 * Cross CPU call to disable a performance event
521 static void __perf_event_disable(void *info
)
523 struct perf_event
*event
= info
;
524 struct perf_cpu_context
*cpuctx
= &__get_cpu_var(perf_cpu_context
);
525 struct perf_event_context
*ctx
= event
->ctx
;
528 * If this is a per-task event, need to check whether this
529 * event's task is the current task on this cpu.
531 if (ctx
->task
&& cpuctx
->task_ctx
!= ctx
)
534 raw_spin_lock(&ctx
->lock
);
537 * If the event is on, turn it off.
538 * If it is in error state, leave it in error state.
540 if (event
->state
>= PERF_EVENT_STATE_INACTIVE
) {
541 update_context_time(ctx
);
542 update_group_times(event
);
543 if (event
== event
->group_leader
)
544 group_sched_out(event
, cpuctx
, ctx
);
546 event_sched_out(event
, cpuctx
, ctx
);
547 event
->state
= PERF_EVENT_STATE_OFF
;
550 raw_spin_unlock(&ctx
->lock
);
556 * If event->ctx is a cloned context, callers must make sure that
557 * every task struct that event->ctx->task could possibly point to
558 * remains valid. This condition is satisifed when called through
559 * perf_event_for_each_child or perf_event_for_each because they
560 * hold the top-level event's child_mutex, so any descendant that
561 * goes to exit will block in sync_child_event.
562 * When called from perf_pending_event it's OK because event->ctx
563 * is the current context on this CPU and preemption is disabled,
564 * hence we can't get into perf_event_task_sched_out for this context.
566 void perf_event_disable(struct perf_event
*event
)
568 struct perf_event_context
*ctx
= event
->ctx
;
569 struct task_struct
*task
= ctx
->task
;
573 * Disable the event on the cpu that it's on
575 smp_call_function_single(event
->cpu
, __perf_event_disable
,
581 task_oncpu_function_call(task
, __perf_event_disable
, event
);
583 raw_spin_lock_irq(&ctx
->lock
);
585 * If the event is still active, we need to retry the cross-call.
587 if (event
->state
== PERF_EVENT_STATE_ACTIVE
) {
588 raw_spin_unlock_irq(&ctx
->lock
);
593 * Since we have the lock this context can't be scheduled
594 * in, so we can change the state safely.
596 if (event
->state
== PERF_EVENT_STATE_INACTIVE
) {
597 update_group_times(event
);
598 event
->state
= PERF_EVENT_STATE_OFF
;
601 raw_spin_unlock_irq(&ctx
->lock
);
605 event_sched_in(struct perf_event
*event
,
606 struct perf_cpu_context
*cpuctx
,
607 struct perf_event_context
*ctx
)
609 if (event
->state
<= PERF_EVENT_STATE_OFF
)
612 event
->state
= PERF_EVENT_STATE_ACTIVE
;
613 event
->oncpu
= smp_processor_id();
615 * The new state must be visible before we turn it on in the hardware:
619 if (event
->pmu
->enable(event
)) {
620 event
->state
= PERF_EVENT_STATE_INACTIVE
;
625 event
->tstamp_running
+= ctx
->time
- event
->tstamp_stopped
;
627 if (!is_software_event(event
))
628 cpuctx
->active_oncpu
++;
631 if (event
->attr
.exclusive
)
632 cpuctx
->exclusive
= 1;
638 group_sched_in(struct perf_event
*group_event
,
639 struct perf_cpu_context
*cpuctx
,
640 struct perf_event_context
*ctx
)
642 struct perf_event
*event
, *partial_group
;
645 if (group_event
->state
== PERF_EVENT_STATE_OFF
)
648 ret
= hw_perf_group_sched_in(group_event
, cpuctx
, ctx
);
650 return ret
< 0 ? ret
: 0;
652 if (event_sched_in(group_event
, cpuctx
, ctx
))
656 * Schedule in siblings as one group (if any):
658 list_for_each_entry(event
, &group_event
->sibling_list
, group_entry
) {
659 if (event_sched_in(event
, cpuctx
, ctx
)) {
660 partial_group
= event
;
669 * Groups can be scheduled in as one unit only, so undo any
670 * partial group before returning:
672 list_for_each_entry(event
, &group_event
->sibling_list
, group_entry
) {
673 if (event
== partial_group
)
675 event_sched_out(event
, cpuctx
, ctx
);
677 event_sched_out(group_event
, cpuctx
, ctx
);
683 * Work out whether we can put this event group on the CPU now.
685 static int group_can_go_on(struct perf_event
*event
,
686 struct perf_cpu_context
*cpuctx
,
690 * Groups consisting entirely of software events can always go on.
692 if (event
->group_flags
& PERF_GROUP_SOFTWARE
)
695 * If an exclusive group is already on, no other hardware
698 if (cpuctx
->exclusive
)
701 * If this group is exclusive and there are already
702 * events on the CPU, it can't go on.
704 if (event
->attr
.exclusive
&& cpuctx
->active_oncpu
)
707 * Otherwise, try to add it if all previous groups were able
713 static void add_event_to_ctx(struct perf_event
*event
,
714 struct perf_event_context
*ctx
)
716 list_add_event(event
, ctx
);
717 event
->tstamp_enabled
= ctx
->time
;
718 event
->tstamp_running
= ctx
->time
;
719 event
->tstamp_stopped
= ctx
->time
;
723 * Cross CPU call to install and enable a performance event
725 * Must be called with ctx->mutex held
727 static void __perf_install_in_context(void *info
)
729 struct perf_cpu_context
*cpuctx
= &__get_cpu_var(perf_cpu_context
);
730 struct perf_event
*event
= info
;
731 struct perf_event_context
*ctx
= event
->ctx
;
732 struct perf_event
*leader
= event
->group_leader
;
736 * If this is a task context, we need to check whether it is
737 * the current task context of this cpu. If not it has been
738 * scheduled out before the smp call arrived.
739 * Or possibly this is the right context but it isn't
740 * on this cpu because it had no events.
742 if (ctx
->task
&& cpuctx
->task_ctx
!= ctx
) {
743 if (cpuctx
->task_ctx
|| ctx
->task
!= current
)
745 cpuctx
->task_ctx
= ctx
;
748 raw_spin_lock(&ctx
->lock
);
750 update_context_time(ctx
);
753 * Protect the list operation against NMI by disabling the
754 * events on a global level. NOP for non NMI based events.
758 add_event_to_ctx(event
, ctx
);
760 if (event
->cpu
!= -1 && event
->cpu
!= smp_processor_id())
764 * Don't put the event on if it is disabled or if
765 * it is in a group and the group isn't on.
767 if (event
->state
!= PERF_EVENT_STATE_INACTIVE
||
768 (leader
!= event
&& leader
->state
!= PERF_EVENT_STATE_ACTIVE
))
772 * An exclusive event can't go on if there are already active
773 * hardware events, and no hardware event can go on if there
774 * is already an exclusive event on.
776 if (!group_can_go_on(event
, cpuctx
, 1))
779 err
= event_sched_in(event
, cpuctx
, ctx
);
783 * This event couldn't go on. If it is in a group
784 * then we have to pull the whole group off.
785 * If the event group is pinned then put it in error state.
788 group_sched_out(leader
, cpuctx
, ctx
);
789 if (leader
->attr
.pinned
) {
790 update_group_times(leader
);
791 leader
->state
= PERF_EVENT_STATE_ERROR
;
795 if (!err
&& !ctx
->task
&& cpuctx
->max_pertask
)
796 cpuctx
->max_pertask
--;
801 raw_spin_unlock(&ctx
->lock
);
805 * Attach a performance event to a context
807 * First we add the event to the list with the hardware enable bit
808 * in event->hw_config cleared.
810 * If the event is attached to a task which is on a CPU we use a smp
811 * call to enable it in the task context. The task might have been
812 * scheduled away, but we check this in the smp call again.
814 * Must be called with ctx->mutex held.
817 perf_install_in_context(struct perf_event_context
*ctx
,
818 struct perf_event
*event
,
821 struct task_struct
*task
= ctx
->task
;
825 * Per cpu events are installed via an smp call and
826 * the install is always successful.
828 smp_call_function_single(cpu
, __perf_install_in_context
,
834 task_oncpu_function_call(task
, __perf_install_in_context
,
837 raw_spin_lock_irq(&ctx
->lock
);
839 * we need to retry the smp call.
841 if (ctx
->is_active
&& list_empty(&event
->group_entry
)) {
842 raw_spin_unlock_irq(&ctx
->lock
);
847 * The lock prevents that this context is scheduled in so we
848 * can add the event safely, if it the call above did not
851 if (list_empty(&event
->group_entry
))
852 add_event_to_ctx(event
, ctx
);
853 raw_spin_unlock_irq(&ctx
->lock
);
857 * Put a event into inactive state and update time fields.
858 * Enabling the leader of a group effectively enables all
859 * the group members that aren't explicitly disabled, so we
860 * have to update their ->tstamp_enabled also.
861 * Note: this works for group members as well as group leaders
862 * since the non-leader members' sibling_lists will be empty.
864 static void __perf_event_mark_enabled(struct perf_event
*event
,
865 struct perf_event_context
*ctx
)
867 struct perf_event
*sub
;
869 event
->state
= PERF_EVENT_STATE_INACTIVE
;
870 event
->tstamp_enabled
= ctx
->time
- event
->total_time_enabled
;
871 list_for_each_entry(sub
, &event
->sibling_list
, group_entry
)
872 if (sub
->state
>= PERF_EVENT_STATE_INACTIVE
)
873 sub
->tstamp_enabled
=
874 ctx
->time
- sub
->total_time_enabled
;
878 * Cross CPU call to enable a performance event
880 static void __perf_event_enable(void *info
)
882 struct perf_event
*event
= info
;
883 struct perf_cpu_context
*cpuctx
= &__get_cpu_var(perf_cpu_context
);
884 struct perf_event_context
*ctx
= event
->ctx
;
885 struct perf_event
*leader
= event
->group_leader
;
889 * If this is a per-task event, need to check whether this
890 * event's task is the current task on this cpu.
892 if (ctx
->task
&& cpuctx
->task_ctx
!= ctx
) {
893 if (cpuctx
->task_ctx
|| ctx
->task
!= current
)
895 cpuctx
->task_ctx
= ctx
;
898 raw_spin_lock(&ctx
->lock
);
900 update_context_time(ctx
);
902 if (event
->state
>= PERF_EVENT_STATE_INACTIVE
)
904 __perf_event_mark_enabled(event
, ctx
);
906 if (event
->cpu
!= -1 && event
->cpu
!= smp_processor_id())
910 * If the event is in a group and isn't the group leader,
911 * then don't put it on unless the group is on.
913 if (leader
!= event
&& leader
->state
!= PERF_EVENT_STATE_ACTIVE
)
916 if (!group_can_go_on(event
, cpuctx
, 1)) {
921 err
= group_sched_in(event
, cpuctx
, ctx
);
923 err
= event_sched_in(event
, cpuctx
, ctx
);
929 * If this event can't go on and it's part of a
930 * group, then the whole group has to come off.
933 group_sched_out(leader
, cpuctx
, ctx
);
934 if (leader
->attr
.pinned
) {
935 update_group_times(leader
);
936 leader
->state
= PERF_EVENT_STATE_ERROR
;
941 raw_spin_unlock(&ctx
->lock
);
947 * If event->ctx is a cloned context, callers must make sure that
948 * every task struct that event->ctx->task could possibly point to
949 * remains valid. This condition is satisfied when called through
950 * perf_event_for_each_child or perf_event_for_each as described
951 * for perf_event_disable.
953 void perf_event_enable(struct perf_event
*event
)
955 struct perf_event_context
*ctx
= event
->ctx
;
956 struct task_struct
*task
= ctx
->task
;
960 * Enable the event on the cpu that it's on
962 smp_call_function_single(event
->cpu
, __perf_event_enable
,
967 raw_spin_lock_irq(&ctx
->lock
);
968 if (event
->state
>= PERF_EVENT_STATE_INACTIVE
)
972 * If the event is in error state, clear that first.
973 * That way, if we see the event in error state below, we
974 * know that it has gone back into error state, as distinct
975 * from the task having been scheduled away before the
976 * cross-call arrived.
978 if (event
->state
== PERF_EVENT_STATE_ERROR
)
979 event
->state
= PERF_EVENT_STATE_OFF
;
982 raw_spin_unlock_irq(&ctx
->lock
);
983 task_oncpu_function_call(task
, __perf_event_enable
, event
);
985 raw_spin_lock_irq(&ctx
->lock
);
988 * If the context is active and the event is still off,
989 * we need to retry the cross-call.
991 if (ctx
->is_active
&& event
->state
== PERF_EVENT_STATE_OFF
)
995 * Since we have the lock this context can't be scheduled
996 * in, so we can change the state safely.
998 if (event
->state
== PERF_EVENT_STATE_OFF
)
999 __perf_event_mark_enabled(event
, ctx
);
1002 raw_spin_unlock_irq(&ctx
->lock
);
1005 static int perf_event_refresh(struct perf_event
*event
, int refresh
)
1008 * not supported on inherited events
1010 if (event
->attr
.inherit
)
1013 atomic_add(refresh
, &event
->event_limit
);
1014 perf_event_enable(event
);
1020 EVENT_FLEXIBLE
= 0x1,
1022 EVENT_ALL
= EVENT_FLEXIBLE
| EVENT_PINNED
,
1025 static void ctx_sched_out(struct perf_event_context
*ctx
,
1026 struct perf_cpu_context
*cpuctx
,
1027 enum event_type_t event_type
)
1029 struct perf_event
*event
;
1031 raw_spin_lock(&ctx
->lock
);
1033 if (likely(!ctx
->nr_events
))
1035 update_context_time(ctx
);
1038 if (!ctx
->nr_active
)
1041 if (event_type
& EVENT_PINNED
)
1042 list_for_each_entry(event
, &ctx
->pinned_groups
, group_entry
)
1043 group_sched_out(event
, cpuctx
, ctx
);
1045 if (event_type
& EVENT_FLEXIBLE
)
1046 list_for_each_entry(event
, &ctx
->flexible_groups
, group_entry
)
1047 group_sched_out(event
, cpuctx
, ctx
);
1052 raw_spin_unlock(&ctx
->lock
);
1056 * Test whether two contexts are equivalent, i.e. whether they
1057 * have both been cloned from the same version of the same context
1058 * and they both have the same number of enabled events.
1059 * If the number of enabled events is the same, then the set
1060 * of enabled events should be the same, because these are both
1061 * inherited contexts, therefore we can't access individual events
1062 * in them directly with an fd; we can only enable/disable all
1063 * events via prctl, or enable/disable all events in a family
1064 * via ioctl, which will have the same effect on both contexts.
1066 static int context_equiv(struct perf_event_context
*ctx1
,
1067 struct perf_event_context
*ctx2
)
1069 return ctx1
->parent_ctx
&& ctx1
->parent_ctx
== ctx2
->parent_ctx
1070 && ctx1
->parent_gen
== ctx2
->parent_gen
1071 && !ctx1
->pin_count
&& !ctx2
->pin_count
;
1074 static void __perf_event_sync_stat(struct perf_event
*event
,
1075 struct perf_event
*next_event
)
1079 if (!event
->attr
.inherit_stat
)
1083 * Update the event value, we cannot use perf_event_read()
1084 * because we're in the middle of a context switch and have IRQs
1085 * disabled, which upsets smp_call_function_single(), however
1086 * we know the event must be on the current CPU, therefore we
1087 * don't need to use it.
1089 switch (event
->state
) {
1090 case PERF_EVENT_STATE_ACTIVE
:
1091 event
->pmu
->read(event
);
1094 case PERF_EVENT_STATE_INACTIVE
:
1095 update_event_times(event
);
1103 * In order to keep per-task stats reliable we need to flip the event
1104 * values when we flip the contexts.
1106 value
= atomic64_read(&next_event
->count
);
1107 value
= atomic64_xchg(&event
->count
, value
);
1108 atomic64_set(&next_event
->count
, value
);
1110 swap(event
->total_time_enabled
, next_event
->total_time_enabled
);
1111 swap(event
->total_time_running
, next_event
->total_time_running
);
1114 * Since we swizzled the values, update the user visible data too.
1116 perf_event_update_userpage(event
);
1117 perf_event_update_userpage(next_event
);
1120 #define list_next_entry(pos, member) \
1121 list_entry(pos->member.next, typeof(*pos), member)
1123 static void perf_event_sync_stat(struct perf_event_context
*ctx
,
1124 struct perf_event_context
*next_ctx
)
1126 struct perf_event
*event
, *next_event
;
1131 update_context_time(ctx
);
1133 event
= list_first_entry(&ctx
->event_list
,
1134 struct perf_event
, event_entry
);
1136 next_event
= list_first_entry(&next_ctx
->event_list
,
1137 struct perf_event
, event_entry
);
1139 while (&event
->event_entry
!= &ctx
->event_list
&&
1140 &next_event
->event_entry
!= &next_ctx
->event_list
) {
1142 __perf_event_sync_stat(event
, next_event
);
1144 event
= list_next_entry(event
, event_entry
);
1145 next_event
= list_next_entry(next_event
, event_entry
);
1150 * Called from scheduler to remove the events of the current task,
1151 * with interrupts disabled.
1153 * We stop each event and update the event value in event->count.
1155 * This does not protect us against NMI, but disable()
1156 * sets the disabled bit in the control field of event _before_
1157 * accessing the event control register. If a NMI hits, then it will
1158 * not restart the event.
1160 void perf_event_task_sched_out(struct task_struct
*task
,
1161 struct task_struct
*next
)
1163 struct perf_cpu_context
*cpuctx
= &__get_cpu_var(perf_cpu_context
);
1164 struct perf_event_context
*ctx
= task
->perf_event_ctxp
;
1165 struct perf_event_context
*next_ctx
;
1166 struct perf_event_context
*parent
;
1169 perf_sw_event(PERF_COUNT_SW_CONTEXT_SWITCHES
, 1, 1, NULL
, 0);
1171 if (likely(!ctx
|| !cpuctx
->task_ctx
))
1175 parent
= rcu_dereference(ctx
->parent_ctx
);
1176 next_ctx
= next
->perf_event_ctxp
;
1177 if (parent
&& next_ctx
&&
1178 rcu_dereference(next_ctx
->parent_ctx
) == parent
) {
1180 * Looks like the two contexts are clones, so we might be
1181 * able to optimize the context switch. We lock both
1182 * contexts and check that they are clones under the
1183 * lock (including re-checking that neither has been
1184 * uncloned in the meantime). It doesn't matter which
1185 * order we take the locks because no other cpu could
1186 * be trying to lock both of these tasks.
1188 raw_spin_lock(&ctx
->lock
);
1189 raw_spin_lock_nested(&next_ctx
->lock
, SINGLE_DEPTH_NESTING
);
1190 if (context_equiv(ctx
, next_ctx
)) {
1192 * XXX do we need a memory barrier of sorts
1193 * wrt to rcu_dereference() of perf_event_ctxp
1195 task
->perf_event_ctxp
= next_ctx
;
1196 next
->perf_event_ctxp
= ctx
;
1198 next_ctx
->task
= task
;
1201 perf_event_sync_stat(ctx
, next_ctx
);
1203 raw_spin_unlock(&next_ctx
->lock
);
1204 raw_spin_unlock(&ctx
->lock
);
1209 ctx_sched_out(ctx
, cpuctx
, EVENT_ALL
);
1210 cpuctx
->task_ctx
= NULL
;
1214 static void task_ctx_sched_out(struct perf_event_context
*ctx
,
1215 enum event_type_t event_type
)
1217 struct perf_cpu_context
*cpuctx
= &__get_cpu_var(perf_cpu_context
);
1219 if (!cpuctx
->task_ctx
)
1222 if (WARN_ON_ONCE(ctx
!= cpuctx
->task_ctx
))
1225 ctx_sched_out(ctx
, cpuctx
, event_type
);
1226 cpuctx
->task_ctx
= NULL
;
1230 * Called with IRQs disabled
1232 static void __perf_event_task_sched_out(struct perf_event_context
*ctx
)
1234 task_ctx_sched_out(ctx
, EVENT_ALL
);
1238 * Called with IRQs disabled
1240 static void cpu_ctx_sched_out(struct perf_cpu_context
*cpuctx
,
1241 enum event_type_t event_type
)
1243 ctx_sched_out(&cpuctx
->ctx
, cpuctx
, event_type
);
1247 ctx_pinned_sched_in(struct perf_event_context
*ctx
,
1248 struct perf_cpu_context
*cpuctx
)
1250 struct perf_event
*event
;
1252 list_for_each_entry(event
, &ctx
->pinned_groups
, group_entry
) {
1253 if (event
->state
<= PERF_EVENT_STATE_OFF
)
1255 if (event
->cpu
!= -1 && event
->cpu
!= smp_processor_id())
1258 if (group_can_go_on(event
, cpuctx
, 1))
1259 group_sched_in(event
, cpuctx
, ctx
);
1262 * If this pinned group hasn't been scheduled,
1263 * put it in error state.
1265 if (event
->state
== PERF_EVENT_STATE_INACTIVE
) {
1266 update_group_times(event
);
1267 event
->state
= PERF_EVENT_STATE_ERROR
;
1273 ctx_flexible_sched_in(struct perf_event_context
*ctx
,
1274 struct perf_cpu_context
*cpuctx
)
1276 struct perf_event
*event
;
1279 list_for_each_entry(event
, &ctx
->flexible_groups
, group_entry
) {
1280 /* Ignore events in OFF or ERROR state */
1281 if (event
->state
<= PERF_EVENT_STATE_OFF
)
1284 * Listen to the 'cpu' scheduling filter constraint
1287 if (event
->cpu
!= -1 && event
->cpu
!= smp_processor_id())
1290 if (group_can_go_on(event
, cpuctx
, can_add_hw
))
1291 if (group_sched_in(event
, cpuctx
, ctx
))
1297 ctx_sched_in(struct perf_event_context
*ctx
,
1298 struct perf_cpu_context
*cpuctx
,
1299 enum event_type_t event_type
)
1301 raw_spin_lock(&ctx
->lock
);
1303 if (likely(!ctx
->nr_events
))
1306 ctx
->timestamp
= perf_clock();
1311 * First go through the list and put on any pinned groups
1312 * in order to give them the best chance of going on.
1314 if (event_type
& EVENT_PINNED
)
1315 ctx_pinned_sched_in(ctx
, cpuctx
);
1317 /* Then walk through the lower prio flexible groups */
1318 if (event_type
& EVENT_FLEXIBLE
)
1319 ctx_flexible_sched_in(ctx
, cpuctx
);
1323 raw_spin_unlock(&ctx
->lock
);
1326 static void cpu_ctx_sched_in(struct perf_cpu_context
*cpuctx
,
1327 enum event_type_t event_type
)
1329 struct perf_event_context
*ctx
= &cpuctx
->ctx
;
1331 ctx_sched_in(ctx
, cpuctx
, event_type
);
1334 static void task_ctx_sched_in(struct task_struct
*task
,
1335 enum event_type_t event_type
)
1337 struct perf_cpu_context
*cpuctx
= &__get_cpu_var(perf_cpu_context
);
1338 struct perf_event_context
*ctx
= task
->perf_event_ctxp
;
1342 if (cpuctx
->task_ctx
== ctx
)
1344 ctx_sched_in(ctx
, cpuctx
, event_type
);
1345 cpuctx
->task_ctx
= ctx
;
1348 * Called from scheduler to add the events of the current task
1349 * with interrupts disabled.
1351 * We restore the event value and then enable it.
1353 * This does not protect us against NMI, but enable()
1354 * sets the enabled bit in the control field of event _before_
1355 * accessing the event control register. If a NMI hits, then it will
1356 * keep the event running.
1358 void perf_event_task_sched_in(struct task_struct
*task
)
1360 struct perf_cpu_context
*cpuctx
= &__get_cpu_var(perf_cpu_context
);
1361 struct perf_event_context
*ctx
= task
->perf_event_ctxp
;
1366 if (cpuctx
->task_ctx
== ctx
)
1372 * We want to keep the following priority order:
1373 * cpu pinned (that don't need to move), task pinned,
1374 * cpu flexible, task flexible.
1376 cpu_ctx_sched_out(cpuctx
, EVENT_FLEXIBLE
);
1378 ctx_sched_in(ctx
, cpuctx
, EVENT_PINNED
);
1379 cpu_ctx_sched_in(cpuctx
, EVENT_FLEXIBLE
);
1380 ctx_sched_in(ctx
, cpuctx
, EVENT_FLEXIBLE
);
1382 cpuctx
->task_ctx
= ctx
;
1387 #define MAX_INTERRUPTS (~0ULL)
1389 static void perf_log_throttle(struct perf_event
*event
, int enable
);
1391 static u64
perf_calculate_period(struct perf_event
*event
, u64 nsec
, u64 count
)
1393 u64 frequency
= event
->attr
.sample_freq
;
1394 u64 sec
= NSEC_PER_SEC
;
1395 u64 divisor
, dividend
;
1397 int count_fls
, nsec_fls
, frequency_fls
, sec_fls
;
1399 count_fls
= fls64(count
);
1400 nsec_fls
= fls64(nsec
);
1401 frequency_fls
= fls64(frequency
);
1405 * We got @count in @nsec, with a target of sample_freq HZ
1406 * the target period becomes:
1409 * period = -------------------
1410 * @nsec * sample_freq
1415 * Reduce accuracy by one bit such that @a and @b converge
1416 * to a similar magnitude.
1418 #define REDUCE_FLS(a, b) \
1420 if (a##_fls > b##_fls) { \
1430 * Reduce accuracy until either term fits in a u64, then proceed with
1431 * the other, so that finally we can do a u64/u64 division.
1433 while (count_fls
+ sec_fls
> 64 && nsec_fls
+ frequency_fls
> 64) {
1434 REDUCE_FLS(nsec
, frequency
);
1435 REDUCE_FLS(sec
, count
);
1438 if (count_fls
+ sec_fls
> 64) {
1439 divisor
= nsec
* frequency
;
1441 while (count_fls
+ sec_fls
> 64) {
1442 REDUCE_FLS(count
, sec
);
1446 dividend
= count
* sec
;
1448 dividend
= count
* sec
;
1450 while (nsec_fls
+ frequency_fls
> 64) {
1451 REDUCE_FLS(nsec
, frequency
);
1455 divisor
= nsec
* frequency
;
1458 return div64_u64(dividend
, divisor
);
1461 static void perf_event_stop(struct perf_event
*event
)
1463 if (!event
->pmu
->stop
)
1464 return event
->pmu
->disable(event
);
1466 return event
->pmu
->stop(event
);
1469 static int perf_event_start(struct perf_event
*event
)
1471 if (!event
->pmu
->start
)
1472 return event
->pmu
->enable(event
);
1474 return event
->pmu
->start(event
);
1477 static void perf_adjust_period(struct perf_event
*event
, u64 nsec
, u64 count
)
1479 struct hw_perf_event
*hwc
= &event
->hw
;
1480 u64 period
, sample_period
;
1483 period
= perf_calculate_period(event
, nsec
, count
);
1485 delta
= (s64
)(period
- hwc
->sample_period
);
1486 delta
= (delta
+ 7) / 8; /* low pass filter */
1488 sample_period
= hwc
->sample_period
+ delta
;
1493 hwc
->sample_period
= sample_period
;
1495 if (atomic64_read(&hwc
->period_left
) > 8*sample_period
) {
1497 perf_event_stop(event
);
1498 atomic64_set(&hwc
->period_left
, 0);
1499 perf_event_start(event
);
1504 static void perf_ctx_adjust_freq(struct perf_event_context
*ctx
)
1506 struct perf_event
*event
;
1507 struct hw_perf_event
*hwc
;
1508 u64 interrupts
, now
;
1511 raw_spin_lock(&ctx
->lock
);
1512 list_for_each_entry_rcu(event
, &ctx
->event_list
, event_entry
) {
1513 if (event
->state
!= PERF_EVENT_STATE_ACTIVE
)
1516 if (event
->cpu
!= -1 && event
->cpu
!= smp_processor_id())
1521 interrupts
= hwc
->interrupts
;
1522 hwc
->interrupts
= 0;
1525 * unthrottle events on the tick
1527 if (interrupts
== MAX_INTERRUPTS
) {
1528 perf_log_throttle(event
, 1);
1530 event
->pmu
->unthrottle(event
);
1534 if (!event
->attr
.freq
|| !event
->attr
.sample_freq
)
1538 event
->pmu
->read(event
);
1539 now
= atomic64_read(&event
->count
);
1540 delta
= now
- hwc
->freq_count_stamp
;
1541 hwc
->freq_count_stamp
= now
;
1544 perf_adjust_period(event
, TICK_NSEC
, delta
);
1547 raw_spin_unlock(&ctx
->lock
);
1551 * Round-robin a context's events:
1553 static void rotate_ctx(struct perf_event_context
*ctx
)
1555 raw_spin_lock(&ctx
->lock
);
1557 /* Rotate the first entry last of non-pinned groups */
1558 list_rotate_left(&ctx
->flexible_groups
);
1560 raw_spin_unlock(&ctx
->lock
);
1563 void perf_event_task_tick(struct task_struct
*curr
)
1565 struct perf_cpu_context
*cpuctx
;
1566 struct perf_event_context
*ctx
;
1569 if (!atomic_read(&nr_events
))
1572 cpuctx
= &__get_cpu_var(perf_cpu_context
);
1573 if (cpuctx
->ctx
.nr_events
&&
1574 cpuctx
->ctx
.nr_events
!= cpuctx
->ctx
.nr_active
)
1577 ctx
= curr
->perf_event_ctxp
;
1578 if (ctx
&& ctx
->nr_events
&& ctx
->nr_events
!= ctx
->nr_active
)
1581 perf_ctx_adjust_freq(&cpuctx
->ctx
);
1583 perf_ctx_adjust_freq(ctx
);
1589 cpu_ctx_sched_out(cpuctx
, EVENT_FLEXIBLE
);
1591 task_ctx_sched_out(ctx
, EVENT_FLEXIBLE
);
1593 rotate_ctx(&cpuctx
->ctx
);
1597 cpu_ctx_sched_in(cpuctx
, EVENT_FLEXIBLE
);
1599 task_ctx_sched_in(curr
, EVENT_FLEXIBLE
);
1603 static int event_enable_on_exec(struct perf_event
*event
,
1604 struct perf_event_context
*ctx
)
1606 if (!event
->attr
.enable_on_exec
)
1609 event
->attr
.enable_on_exec
= 0;
1610 if (event
->state
>= PERF_EVENT_STATE_INACTIVE
)
1613 __perf_event_mark_enabled(event
, ctx
);
1619 * Enable all of a task's events that have been marked enable-on-exec.
1620 * This expects task == current.
1622 static void perf_event_enable_on_exec(struct task_struct
*task
)
1624 struct perf_event_context
*ctx
;
1625 struct perf_event
*event
;
1626 unsigned long flags
;
1630 local_irq_save(flags
);
1631 ctx
= task
->perf_event_ctxp
;
1632 if (!ctx
|| !ctx
->nr_events
)
1635 __perf_event_task_sched_out(ctx
);
1637 raw_spin_lock(&ctx
->lock
);
1639 list_for_each_entry(event
, &ctx
->pinned_groups
, group_entry
) {
1640 ret
= event_enable_on_exec(event
, ctx
);
1645 list_for_each_entry(event
, &ctx
->flexible_groups
, group_entry
) {
1646 ret
= event_enable_on_exec(event
, ctx
);
1652 * Unclone this context if we enabled any event.
1657 raw_spin_unlock(&ctx
->lock
);
1659 perf_event_task_sched_in(task
);
1661 local_irq_restore(flags
);
1665 * Cross CPU call to read the hardware event
1667 static void __perf_event_read(void *info
)
1669 struct perf_cpu_context
*cpuctx
= &__get_cpu_var(perf_cpu_context
);
1670 struct perf_event
*event
= info
;
1671 struct perf_event_context
*ctx
= event
->ctx
;
1674 * If this is a task context, we need to check whether it is
1675 * the current task context of this cpu. If not it has been
1676 * scheduled out before the smp call arrived. In that case
1677 * event->count would have been updated to a recent sample
1678 * when the event was scheduled out.
1680 if (ctx
->task
&& cpuctx
->task_ctx
!= ctx
)
1683 raw_spin_lock(&ctx
->lock
);
1684 update_context_time(ctx
);
1685 update_event_times(event
);
1686 raw_spin_unlock(&ctx
->lock
);
1688 event
->pmu
->read(event
);
1691 static u64
perf_event_read(struct perf_event
*event
)
1694 * If event is enabled and currently active on a CPU, update the
1695 * value in the event structure:
1697 if (event
->state
== PERF_EVENT_STATE_ACTIVE
) {
1698 smp_call_function_single(event
->oncpu
,
1699 __perf_event_read
, event
, 1);
1700 } else if (event
->state
== PERF_EVENT_STATE_INACTIVE
) {
1701 struct perf_event_context
*ctx
= event
->ctx
;
1702 unsigned long flags
;
1704 raw_spin_lock_irqsave(&ctx
->lock
, flags
);
1705 update_context_time(ctx
);
1706 update_event_times(event
);
1707 raw_spin_unlock_irqrestore(&ctx
->lock
, flags
);
1710 return atomic64_read(&event
->count
);
1714 * Initialize the perf_event context in a task_struct:
1717 __perf_event_init_context(struct perf_event_context
*ctx
,
1718 struct task_struct
*task
)
1720 raw_spin_lock_init(&ctx
->lock
);
1721 mutex_init(&ctx
->mutex
);
1722 INIT_LIST_HEAD(&ctx
->pinned_groups
);
1723 INIT_LIST_HEAD(&ctx
->flexible_groups
);
1724 INIT_LIST_HEAD(&ctx
->event_list
);
1725 atomic_set(&ctx
->refcount
, 1);
1729 static struct perf_event_context
*find_get_context(pid_t pid
, int cpu
)
1731 struct perf_event_context
*ctx
;
1732 struct perf_cpu_context
*cpuctx
;
1733 struct task_struct
*task
;
1734 unsigned long flags
;
1737 if (pid
== -1 && cpu
!= -1) {
1738 /* Must be root to operate on a CPU event: */
1739 if (perf_paranoid_cpu() && !capable(CAP_SYS_ADMIN
))
1740 return ERR_PTR(-EACCES
);
1742 if (cpu
< 0 || cpu
>= nr_cpumask_bits
)
1743 return ERR_PTR(-EINVAL
);
1746 * We could be clever and allow to attach a event to an
1747 * offline CPU and activate it when the CPU comes up, but
1750 if (!cpu_online(cpu
))
1751 return ERR_PTR(-ENODEV
);
1753 cpuctx
= &per_cpu(perf_cpu_context
, cpu
);
1764 task
= find_task_by_vpid(pid
);
1766 get_task_struct(task
);
1770 return ERR_PTR(-ESRCH
);
1773 * Can't attach events to a dying task.
1776 if (task
->flags
& PF_EXITING
)
1779 /* Reuse ptrace permission checks for now. */
1781 if (!ptrace_may_access(task
, PTRACE_MODE_READ
))
1785 ctx
= perf_lock_task_context(task
, &flags
);
1788 raw_spin_unlock_irqrestore(&ctx
->lock
, flags
);
1792 ctx
= kzalloc(sizeof(struct perf_event_context
), GFP_KERNEL
);
1796 __perf_event_init_context(ctx
, task
);
1798 if (cmpxchg(&task
->perf_event_ctxp
, NULL
, ctx
)) {
1800 * We raced with some other task; use
1801 * the context they set.
1806 get_task_struct(task
);
1809 put_task_struct(task
);
1813 put_task_struct(task
);
1814 return ERR_PTR(err
);
1817 static void perf_event_free_filter(struct perf_event
*event
);
1819 static void free_event_rcu(struct rcu_head
*head
)
1821 struct perf_event
*event
;
1823 event
= container_of(head
, struct perf_event
, rcu_head
);
1825 put_pid_ns(event
->ns
);
1826 perf_event_free_filter(event
);
1830 static void perf_pending_sync(struct perf_event
*event
);
1832 static void free_event(struct perf_event
*event
)
1834 perf_pending_sync(event
);
1836 if (!event
->parent
) {
1837 atomic_dec(&nr_events
);
1838 if (event
->attr
.mmap
)
1839 atomic_dec(&nr_mmap_events
);
1840 if (event
->attr
.comm
)
1841 atomic_dec(&nr_comm_events
);
1842 if (event
->attr
.task
)
1843 atomic_dec(&nr_task_events
);
1846 if (event
->output
) {
1847 fput(event
->output
->filp
);
1848 event
->output
= NULL
;
1852 event
->destroy(event
);
1854 put_ctx(event
->ctx
);
1855 call_rcu(&event
->rcu_head
, free_event_rcu
);
1858 int perf_event_release_kernel(struct perf_event
*event
)
1860 struct perf_event_context
*ctx
= event
->ctx
;
1862 WARN_ON_ONCE(ctx
->parent_ctx
);
1863 mutex_lock(&ctx
->mutex
);
1864 perf_event_remove_from_context(event
);
1865 mutex_unlock(&ctx
->mutex
);
1867 mutex_lock(&event
->owner
->perf_event_mutex
);
1868 list_del_init(&event
->owner_entry
);
1869 mutex_unlock(&event
->owner
->perf_event_mutex
);
1870 put_task_struct(event
->owner
);
1876 EXPORT_SYMBOL_GPL(perf_event_release_kernel
);
1879 * Called when the last reference to the file is gone.
1881 static int perf_release(struct inode
*inode
, struct file
*file
)
1883 struct perf_event
*event
= file
->private_data
;
1885 file
->private_data
= NULL
;
1887 return perf_event_release_kernel(event
);
1890 static int perf_event_read_size(struct perf_event
*event
)
1892 int entry
= sizeof(u64
); /* value */
1896 if (event
->attr
.read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
)
1897 size
+= sizeof(u64
);
1899 if (event
->attr
.read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
)
1900 size
+= sizeof(u64
);
1902 if (event
->attr
.read_format
& PERF_FORMAT_ID
)
1903 entry
+= sizeof(u64
);
1905 if (event
->attr
.read_format
& PERF_FORMAT_GROUP
) {
1906 nr
+= event
->group_leader
->nr_siblings
;
1907 size
+= sizeof(u64
);
1915 u64
perf_event_read_value(struct perf_event
*event
, u64
*enabled
, u64
*running
)
1917 struct perf_event
*child
;
1923 mutex_lock(&event
->child_mutex
);
1924 total
+= perf_event_read(event
);
1925 *enabled
+= event
->total_time_enabled
+
1926 atomic64_read(&event
->child_total_time_enabled
);
1927 *running
+= event
->total_time_running
+
1928 atomic64_read(&event
->child_total_time_running
);
1930 list_for_each_entry(child
, &event
->child_list
, child_list
) {
1931 total
+= perf_event_read(child
);
1932 *enabled
+= child
->total_time_enabled
;
1933 *running
+= child
->total_time_running
;
1935 mutex_unlock(&event
->child_mutex
);
1939 EXPORT_SYMBOL_GPL(perf_event_read_value
);
1941 static int perf_event_read_group(struct perf_event
*event
,
1942 u64 read_format
, char __user
*buf
)
1944 struct perf_event
*leader
= event
->group_leader
, *sub
;
1945 int n
= 0, size
= 0, ret
= -EFAULT
;
1946 struct perf_event_context
*ctx
= leader
->ctx
;
1948 u64 count
, enabled
, running
;
1950 mutex_lock(&ctx
->mutex
);
1951 count
= perf_event_read_value(leader
, &enabled
, &running
);
1953 values
[n
++] = 1 + leader
->nr_siblings
;
1954 if (read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
)
1955 values
[n
++] = enabled
;
1956 if (read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
)
1957 values
[n
++] = running
;
1958 values
[n
++] = count
;
1959 if (read_format
& PERF_FORMAT_ID
)
1960 values
[n
++] = primary_event_id(leader
);
1962 size
= n
* sizeof(u64
);
1964 if (copy_to_user(buf
, values
, size
))
1969 list_for_each_entry(sub
, &leader
->sibling_list
, group_entry
) {
1972 values
[n
++] = perf_event_read_value(sub
, &enabled
, &running
);
1973 if (read_format
& PERF_FORMAT_ID
)
1974 values
[n
++] = primary_event_id(sub
);
1976 size
= n
* sizeof(u64
);
1978 if (copy_to_user(buf
+ ret
, values
, size
)) {
1986 mutex_unlock(&ctx
->mutex
);
1991 static int perf_event_read_one(struct perf_event
*event
,
1992 u64 read_format
, char __user
*buf
)
1994 u64 enabled
, running
;
1998 values
[n
++] = perf_event_read_value(event
, &enabled
, &running
);
1999 if (read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
)
2000 values
[n
++] = enabled
;
2001 if (read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
)
2002 values
[n
++] = running
;
2003 if (read_format
& PERF_FORMAT_ID
)
2004 values
[n
++] = primary_event_id(event
);
2006 if (copy_to_user(buf
, values
, n
* sizeof(u64
)))
2009 return n
* sizeof(u64
);
2013 * Read the performance event - simple non blocking version for now
2016 perf_read_hw(struct perf_event
*event
, char __user
*buf
, size_t count
)
2018 u64 read_format
= event
->attr
.read_format
;
2022 * Return end-of-file for a read on a event that is in
2023 * error state (i.e. because it was pinned but it couldn't be
2024 * scheduled on to the CPU at some point).
2026 if (event
->state
== PERF_EVENT_STATE_ERROR
)
2029 if (count
< perf_event_read_size(event
))
2032 WARN_ON_ONCE(event
->ctx
->parent_ctx
);
2033 if (read_format
& PERF_FORMAT_GROUP
)
2034 ret
= perf_event_read_group(event
, read_format
, buf
);
2036 ret
= perf_event_read_one(event
, read_format
, buf
);
2042 perf_read(struct file
*file
, char __user
*buf
, size_t count
, loff_t
*ppos
)
2044 struct perf_event
*event
= file
->private_data
;
2046 return perf_read_hw(event
, buf
, count
);
2049 static unsigned int perf_poll(struct file
*file
, poll_table
*wait
)
2051 struct perf_event
*event
= file
->private_data
;
2052 struct perf_mmap_data
*data
;
2053 unsigned int events
= POLL_HUP
;
2056 data
= rcu_dereference(event
->data
);
2058 events
= atomic_xchg(&data
->poll
, 0);
2061 poll_wait(file
, &event
->waitq
, wait
);
2066 static void perf_event_reset(struct perf_event
*event
)
2068 (void)perf_event_read(event
);
2069 atomic64_set(&event
->count
, 0);
2070 perf_event_update_userpage(event
);
2074 * Holding the top-level event's child_mutex means that any
2075 * descendant process that has inherited this event will block
2076 * in sync_child_event if it goes to exit, thus satisfying the
2077 * task existence requirements of perf_event_enable/disable.
2079 static void perf_event_for_each_child(struct perf_event
*event
,
2080 void (*func
)(struct perf_event
*))
2082 struct perf_event
*child
;
2084 WARN_ON_ONCE(event
->ctx
->parent_ctx
);
2085 mutex_lock(&event
->child_mutex
);
2087 list_for_each_entry(child
, &event
->child_list
, child_list
)
2089 mutex_unlock(&event
->child_mutex
);
2092 static void perf_event_for_each(struct perf_event
*event
,
2093 void (*func
)(struct perf_event
*))
2095 struct perf_event_context
*ctx
= event
->ctx
;
2096 struct perf_event
*sibling
;
2098 WARN_ON_ONCE(ctx
->parent_ctx
);
2099 mutex_lock(&ctx
->mutex
);
2100 event
= event
->group_leader
;
2102 perf_event_for_each_child(event
, func
);
2104 list_for_each_entry(sibling
, &event
->sibling_list
, group_entry
)
2105 perf_event_for_each_child(event
, func
);
2106 mutex_unlock(&ctx
->mutex
);
2109 static int perf_event_period(struct perf_event
*event
, u64 __user
*arg
)
2111 struct perf_event_context
*ctx
= event
->ctx
;
2116 if (!event
->attr
.sample_period
)
2119 size
= copy_from_user(&value
, arg
, sizeof(value
));
2120 if (size
!= sizeof(value
))
2126 raw_spin_lock_irq(&ctx
->lock
);
2127 if (event
->attr
.freq
) {
2128 if (value
> sysctl_perf_event_sample_rate
) {
2133 event
->attr
.sample_freq
= value
;
2135 event
->attr
.sample_period
= value
;
2136 event
->hw
.sample_period
= value
;
2139 raw_spin_unlock_irq(&ctx
->lock
);
2144 static int perf_event_set_output(struct perf_event
*event
, int output_fd
);
2145 static int perf_event_set_filter(struct perf_event
*event
, void __user
*arg
);
2147 static long perf_ioctl(struct file
*file
, unsigned int cmd
, unsigned long arg
)
2149 struct perf_event
*event
= file
->private_data
;
2150 void (*func
)(struct perf_event
*);
2154 case PERF_EVENT_IOC_ENABLE
:
2155 func
= perf_event_enable
;
2157 case PERF_EVENT_IOC_DISABLE
:
2158 func
= perf_event_disable
;
2160 case PERF_EVENT_IOC_RESET
:
2161 func
= perf_event_reset
;
2164 case PERF_EVENT_IOC_REFRESH
:
2165 return perf_event_refresh(event
, arg
);
2167 case PERF_EVENT_IOC_PERIOD
:
2168 return perf_event_period(event
, (u64 __user
*)arg
);
2170 case PERF_EVENT_IOC_SET_OUTPUT
:
2171 return perf_event_set_output(event
, arg
);
2173 case PERF_EVENT_IOC_SET_FILTER
:
2174 return perf_event_set_filter(event
, (void __user
*)arg
);
2180 if (flags
& PERF_IOC_FLAG_GROUP
)
2181 perf_event_for_each(event
, func
);
2183 perf_event_for_each_child(event
, func
);
2188 int perf_event_task_enable(void)
2190 struct perf_event
*event
;
2192 mutex_lock(¤t
->perf_event_mutex
);
2193 list_for_each_entry(event
, ¤t
->perf_event_list
, owner_entry
)
2194 perf_event_for_each_child(event
, perf_event_enable
);
2195 mutex_unlock(¤t
->perf_event_mutex
);
2200 int perf_event_task_disable(void)
2202 struct perf_event
*event
;
2204 mutex_lock(¤t
->perf_event_mutex
);
2205 list_for_each_entry(event
, ¤t
->perf_event_list
, owner_entry
)
2206 perf_event_for_each_child(event
, perf_event_disable
);
2207 mutex_unlock(¤t
->perf_event_mutex
);
2212 #ifndef PERF_EVENT_INDEX_OFFSET
2213 # define PERF_EVENT_INDEX_OFFSET 0
2216 static int perf_event_index(struct perf_event
*event
)
2218 if (event
->state
!= PERF_EVENT_STATE_ACTIVE
)
2221 return event
->hw
.idx
+ 1 - PERF_EVENT_INDEX_OFFSET
;
2225 * Callers need to ensure there can be no nesting of this function, otherwise
2226 * the seqlock logic goes bad. We can not serialize this because the arch
2227 * code calls this from NMI context.
2229 void perf_event_update_userpage(struct perf_event
*event
)
2231 struct perf_event_mmap_page
*userpg
;
2232 struct perf_mmap_data
*data
;
2235 data
= rcu_dereference(event
->data
);
2239 userpg
= data
->user_page
;
2242 * Disable preemption so as to not let the corresponding user-space
2243 * spin too long if we get preempted.
2248 userpg
->index
= perf_event_index(event
);
2249 userpg
->offset
= atomic64_read(&event
->count
);
2250 if (event
->state
== PERF_EVENT_STATE_ACTIVE
)
2251 userpg
->offset
-= atomic64_read(&event
->hw
.prev_count
);
2253 userpg
->time_enabled
= event
->total_time_enabled
+
2254 atomic64_read(&event
->child_total_time_enabled
);
2256 userpg
->time_running
= event
->total_time_running
+
2257 atomic64_read(&event
->child_total_time_running
);
2266 static unsigned long perf_data_size(struct perf_mmap_data
*data
)
2268 return data
->nr_pages
<< (PAGE_SHIFT
+ data
->data_order
);
2271 #ifndef CONFIG_PERF_USE_VMALLOC
2274 * Back perf_mmap() with regular GFP_KERNEL-0 pages.
2277 static struct page
*
2278 perf_mmap_to_page(struct perf_mmap_data
*data
, unsigned long pgoff
)
2280 if (pgoff
> data
->nr_pages
)
2284 return virt_to_page(data
->user_page
);
2286 return virt_to_page(data
->data_pages
[pgoff
- 1]);
2289 static struct perf_mmap_data
*
2290 perf_mmap_data_alloc(struct perf_event
*event
, int nr_pages
)
2292 struct perf_mmap_data
*data
;
2296 WARN_ON(atomic_read(&event
->mmap_count
));
2298 size
= sizeof(struct perf_mmap_data
);
2299 size
+= nr_pages
* sizeof(void *);
2301 data
= kzalloc(size
, GFP_KERNEL
);
2305 data
->user_page
= (void *)get_zeroed_page(GFP_KERNEL
);
2306 if (!data
->user_page
)
2307 goto fail_user_page
;
2309 for (i
= 0; i
< nr_pages
; i
++) {
2310 data
->data_pages
[i
] = (void *)get_zeroed_page(GFP_KERNEL
);
2311 if (!data
->data_pages
[i
])
2312 goto fail_data_pages
;
2315 data
->data_order
= 0;
2316 data
->nr_pages
= nr_pages
;
2321 for (i
--; i
>= 0; i
--)
2322 free_page((unsigned long)data
->data_pages
[i
]);
2324 free_page((unsigned long)data
->user_page
);
2333 static void perf_mmap_free_page(unsigned long addr
)
2335 struct page
*page
= virt_to_page((void *)addr
);
2337 page
->mapping
= NULL
;
2341 static void perf_mmap_data_free(struct perf_mmap_data
*data
)
2345 perf_mmap_free_page((unsigned long)data
->user_page
);
2346 for (i
= 0; i
< data
->nr_pages
; i
++)
2347 perf_mmap_free_page((unsigned long)data
->data_pages
[i
]);
2354 * Back perf_mmap() with vmalloc memory.
2356 * Required for architectures that have d-cache aliasing issues.
2359 static struct page
*
2360 perf_mmap_to_page(struct perf_mmap_data
*data
, unsigned long pgoff
)
2362 if (pgoff
> (1UL << data
->data_order
))
2365 return vmalloc_to_page((void *)data
->user_page
+ pgoff
* PAGE_SIZE
);
2368 static void perf_mmap_unmark_page(void *addr
)
2370 struct page
*page
= vmalloc_to_page(addr
);
2372 page
->mapping
= NULL
;
2375 static void perf_mmap_data_free_work(struct work_struct
*work
)
2377 struct perf_mmap_data
*data
;
2381 data
= container_of(work
, struct perf_mmap_data
, work
);
2382 nr
= 1 << data
->data_order
;
2384 base
= data
->user_page
;
2385 for (i
= 0; i
< nr
+ 1; i
++)
2386 perf_mmap_unmark_page(base
+ (i
* PAGE_SIZE
));
2392 static void perf_mmap_data_free(struct perf_mmap_data
*data
)
2394 schedule_work(&data
->work
);
2397 static struct perf_mmap_data
*
2398 perf_mmap_data_alloc(struct perf_event
*event
, int nr_pages
)
2400 struct perf_mmap_data
*data
;
2404 WARN_ON(atomic_read(&event
->mmap_count
));
2406 size
= sizeof(struct perf_mmap_data
);
2407 size
+= sizeof(void *);
2409 data
= kzalloc(size
, GFP_KERNEL
);
2413 INIT_WORK(&data
->work
, perf_mmap_data_free_work
);
2415 all_buf
= vmalloc_user((nr_pages
+ 1) * PAGE_SIZE
);
2419 data
->user_page
= all_buf
;
2420 data
->data_pages
[0] = all_buf
+ PAGE_SIZE
;
2421 data
->data_order
= ilog2(nr_pages
);
2435 static int perf_mmap_fault(struct vm_area_struct
*vma
, struct vm_fault
*vmf
)
2437 struct perf_event
*event
= vma
->vm_file
->private_data
;
2438 struct perf_mmap_data
*data
;
2439 int ret
= VM_FAULT_SIGBUS
;
2441 if (vmf
->flags
& FAULT_FLAG_MKWRITE
) {
2442 if (vmf
->pgoff
== 0)
2448 data
= rcu_dereference(event
->data
);
2452 if (vmf
->pgoff
&& (vmf
->flags
& FAULT_FLAG_WRITE
))
2455 vmf
->page
= perf_mmap_to_page(data
, vmf
->pgoff
);
2459 get_page(vmf
->page
);
2460 vmf
->page
->mapping
= vma
->vm_file
->f_mapping
;
2461 vmf
->page
->index
= vmf
->pgoff
;
2471 perf_mmap_data_init(struct perf_event
*event
, struct perf_mmap_data
*data
)
2473 long max_size
= perf_data_size(data
);
2475 atomic_set(&data
->lock
, -1);
2477 if (event
->attr
.watermark
) {
2478 data
->watermark
= min_t(long, max_size
,
2479 event
->attr
.wakeup_watermark
);
2482 if (!data
->watermark
)
2483 data
->watermark
= max_size
/ 2;
2486 rcu_assign_pointer(event
->data
, data
);
2489 static void perf_mmap_data_free_rcu(struct rcu_head
*rcu_head
)
2491 struct perf_mmap_data
*data
;
2493 data
= container_of(rcu_head
, struct perf_mmap_data
, rcu_head
);
2494 perf_mmap_data_free(data
);
2497 static void perf_mmap_data_release(struct perf_event
*event
)
2499 struct perf_mmap_data
*data
= event
->data
;
2501 WARN_ON(atomic_read(&event
->mmap_count
));
2503 rcu_assign_pointer(event
->data
, NULL
);
2504 call_rcu(&data
->rcu_head
, perf_mmap_data_free_rcu
);
2507 static void perf_mmap_open(struct vm_area_struct
*vma
)
2509 struct perf_event
*event
= vma
->vm_file
->private_data
;
2511 atomic_inc(&event
->mmap_count
);
2514 static void perf_mmap_close(struct vm_area_struct
*vma
)
2516 struct perf_event
*event
= vma
->vm_file
->private_data
;
2518 WARN_ON_ONCE(event
->ctx
->parent_ctx
);
2519 if (atomic_dec_and_mutex_lock(&event
->mmap_count
, &event
->mmap_mutex
)) {
2520 unsigned long size
= perf_data_size(event
->data
);
2521 struct user_struct
*user
= current_user();
2523 atomic_long_sub((size
>> PAGE_SHIFT
) + 1, &user
->locked_vm
);
2524 vma
->vm_mm
->locked_vm
-= event
->data
->nr_locked
;
2525 perf_mmap_data_release(event
);
2526 mutex_unlock(&event
->mmap_mutex
);
2530 static const struct vm_operations_struct perf_mmap_vmops
= {
2531 .open
= perf_mmap_open
,
2532 .close
= perf_mmap_close
,
2533 .fault
= perf_mmap_fault
,
2534 .page_mkwrite
= perf_mmap_fault
,
2537 static int perf_mmap(struct file
*file
, struct vm_area_struct
*vma
)
2539 struct perf_event
*event
= file
->private_data
;
2540 unsigned long user_locked
, user_lock_limit
;
2541 struct user_struct
*user
= current_user();
2542 unsigned long locked
, lock_limit
;
2543 struct perf_mmap_data
*data
;
2544 unsigned long vma_size
;
2545 unsigned long nr_pages
;
2546 long user_extra
, extra
;
2549 if (!(vma
->vm_flags
& VM_SHARED
))
2552 vma_size
= vma
->vm_end
- vma
->vm_start
;
2553 nr_pages
= (vma_size
/ PAGE_SIZE
) - 1;
2556 * If we have data pages ensure they're a power-of-two number, so we
2557 * can do bitmasks instead of modulo.
2559 if (nr_pages
!= 0 && !is_power_of_2(nr_pages
))
2562 if (vma_size
!= PAGE_SIZE
* (1 + nr_pages
))
2565 if (vma
->vm_pgoff
!= 0)
2568 WARN_ON_ONCE(event
->ctx
->parent_ctx
);
2569 mutex_lock(&event
->mmap_mutex
);
2570 if (event
->output
) {
2575 if (atomic_inc_not_zero(&event
->mmap_count
)) {
2576 if (nr_pages
!= event
->data
->nr_pages
)
2581 user_extra
= nr_pages
+ 1;
2582 user_lock_limit
= sysctl_perf_event_mlock
>> (PAGE_SHIFT
- 10);
2585 * Increase the limit linearly with more CPUs:
2587 user_lock_limit
*= num_online_cpus();
2589 user_locked
= atomic_long_read(&user
->locked_vm
) + user_extra
;
2592 if (user_locked
> user_lock_limit
)
2593 extra
= user_locked
- user_lock_limit
;
2595 lock_limit
= rlimit(RLIMIT_MEMLOCK
);
2596 lock_limit
>>= PAGE_SHIFT
;
2597 locked
= vma
->vm_mm
->locked_vm
+ extra
;
2599 if ((locked
> lock_limit
) && perf_paranoid_tracepoint_raw() &&
2600 !capable(CAP_IPC_LOCK
)) {
2605 WARN_ON(event
->data
);
2607 data
= perf_mmap_data_alloc(event
, nr_pages
);
2613 perf_mmap_data_init(event
, data
);
2615 atomic_set(&event
->mmap_count
, 1);
2616 atomic_long_add(user_extra
, &user
->locked_vm
);
2617 vma
->vm_mm
->locked_vm
+= extra
;
2618 event
->data
->nr_locked
= extra
;
2619 if (vma
->vm_flags
& VM_WRITE
)
2620 event
->data
->writable
= 1;
2623 mutex_unlock(&event
->mmap_mutex
);
2625 vma
->vm_flags
|= VM_RESERVED
;
2626 vma
->vm_ops
= &perf_mmap_vmops
;
2631 static int perf_fasync(int fd
, struct file
*filp
, int on
)
2633 struct inode
*inode
= filp
->f_path
.dentry
->d_inode
;
2634 struct perf_event
*event
= filp
->private_data
;
2637 mutex_lock(&inode
->i_mutex
);
2638 retval
= fasync_helper(fd
, filp
, on
, &event
->fasync
);
2639 mutex_unlock(&inode
->i_mutex
);
2647 static const struct file_operations perf_fops
= {
2648 .release
= perf_release
,
2651 .unlocked_ioctl
= perf_ioctl
,
2652 .compat_ioctl
= perf_ioctl
,
2654 .fasync
= perf_fasync
,
2660 * If there's data, ensure we set the poll() state and publish everything
2661 * to user-space before waking everybody up.
2664 void perf_event_wakeup(struct perf_event
*event
)
2666 wake_up_all(&event
->waitq
);
2668 if (event
->pending_kill
) {
2669 kill_fasync(&event
->fasync
, SIGIO
, event
->pending_kill
);
2670 event
->pending_kill
= 0;
2677 * Handle the case where we need to wakeup up from NMI (or rq->lock) context.
2679 * The NMI bit means we cannot possibly take locks. Therefore, maintain a
2680 * single linked list and use cmpxchg() to add entries lockless.
2683 static void perf_pending_event(struct perf_pending_entry
*entry
)
2685 struct perf_event
*event
= container_of(entry
,
2686 struct perf_event
, pending
);
2688 if (event
->pending_disable
) {
2689 event
->pending_disable
= 0;
2690 __perf_event_disable(event
);
2693 if (event
->pending_wakeup
) {
2694 event
->pending_wakeup
= 0;
2695 perf_event_wakeup(event
);
2699 #define PENDING_TAIL ((struct perf_pending_entry *)-1UL)
2701 static DEFINE_PER_CPU(struct perf_pending_entry
*, perf_pending_head
) = {
2705 static void perf_pending_queue(struct perf_pending_entry
*entry
,
2706 void (*func
)(struct perf_pending_entry
*))
2708 struct perf_pending_entry
**head
;
2710 if (cmpxchg(&entry
->next
, NULL
, PENDING_TAIL
) != NULL
)
2715 head
= &get_cpu_var(perf_pending_head
);
2718 entry
->next
= *head
;
2719 } while (cmpxchg(head
, entry
->next
, entry
) != entry
->next
);
2721 set_perf_event_pending();
2723 put_cpu_var(perf_pending_head
);
2726 static int __perf_pending_run(void)
2728 struct perf_pending_entry
*list
;
2731 list
= xchg(&__get_cpu_var(perf_pending_head
), PENDING_TAIL
);
2732 while (list
!= PENDING_TAIL
) {
2733 void (*func
)(struct perf_pending_entry
*);
2734 struct perf_pending_entry
*entry
= list
;
2741 * Ensure we observe the unqueue before we issue the wakeup,
2742 * so that we won't be waiting forever.
2743 * -- see perf_not_pending().
2754 static inline int perf_not_pending(struct perf_event
*event
)
2757 * If we flush on whatever cpu we run, there is a chance we don't
2761 __perf_pending_run();
2765 * Ensure we see the proper queue state before going to sleep
2766 * so that we do not miss the wakeup. -- see perf_pending_handle()
2769 return event
->pending
.next
== NULL
;
2772 static void perf_pending_sync(struct perf_event
*event
)
2774 wait_event(event
->waitq
, perf_not_pending(event
));
2777 void perf_event_do_pending(void)
2779 __perf_pending_run();
2783 * Callchain support -- arch specific
2786 __weak
struct perf_callchain_entry
*perf_callchain(struct pt_regs
*regs
)
2792 void perf_arch_fetch_caller_regs(struct pt_regs
*regs
, unsigned long ip
, int skip
)
2800 static bool perf_output_space(struct perf_mmap_data
*data
, unsigned long tail
,
2801 unsigned long offset
, unsigned long head
)
2805 if (!data
->writable
)
2808 mask
= perf_data_size(data
) - 1;
2810 offset
= (offset
- tail
) & mask
;
2811 head
= (head
- tail
) & mask
;
2813 if ((int)(head
- offset
) < 0)
2819 static void perf_output_wakeup(struct perf_output_handle
*handle
)
2821 atomic_set(&handle
->data
->poll
, POLL_IN
);
2824 handle
->event
->pending_wakeup
= 1;
2825 perf_pending_queue(&handle
->event
->pending
,
2826 perf_pending_event
);
2828 perf_event_wakeup(handle
->event
);
2832 * Curious locking construct.
2834 * We need to ensure a later event_id doesn't publish a head when a former
2835 * event_id isn't done writing. However since we need to deal with NMIs we
2836 * cannot fully serialize things.
2838 * What we do is serialize between CPUs so we only have to deal with NMI
2839 * nesting on a single CPU.
2841 * We only publish the head (and generate a wakeup) when the outer-most
2842 * event_id completes.
2844 static void perf_output_lock(struct perf_output_handle
*handle
)
2846 struct perf_mmap_data
*data
= handle
->data
;
2847 int cur
, cpu
= get_cpu();
2852 cur
= atomic_cmpxchg(&data
->lock
, -1, cpu
);
2864 static void perf_output_unlock(struct perf_output_handle
*handle
)
2866 struct perf_mmap_data
*data
= handle
->data
;
2870 data
->done_head
= data
->head
;
2872 if (!handle
->locked
)
2877 * The xchg implies a full barrier that ensures all writes are done
2878 * before we publish the new head, matched by a rmb() in userspace when
2879 * reading this position.
2881 while ((head
= atomic_long_xchg(&data
->done_head
, 0)))
2882 data
->user_page
->data_head
= head
;
2885 * NMI can happen here, which means we can miss a done_head update.
2888 cpu
= atomic_xchg(&data
->lock
, -1);
2889 WARN_ON_ONCE(cpu
!= smp_processor_id());
2892 * Therefore we have to validate we did not indeed do so.
2894 if (unlikely(atomic_long_read(&data
->done_head
))) {
2896 * Since we had it locked, we can lock it again.
2898 while (atomic_cmpxchg(&data
->lock
, -1, cpu
) != -1)
2904 if (atomic_xchg(&data
->wakeup
, 0))
2905 perf_output_wakeup(handle
);
2910 void perf_output_copy(struct perf_output_handle
*handle
,
2911 const void *buf
, unsigned int len
)
2913 unsigned int pages_mask
;
2914 unsigned long offset
;
2918 offset
= handle
->offset
;
2919 pages_mask
= handle
->data
->nr_pages
- 1;
2920 pages
= handle
->data
->data_pages
;
2923 unsigned long page_offset
;
2924 unsigned long page_size
;
2927 nr
= (offset
>> PAGE_SHIFT
) & pages_mask
;
2928 page_size
= 1UL << (handle
->data
->data_order
+ PAGE_SHIFT
);
2929 page_offset
= offset
& (page_size
- 1);
2930 size
= min_t(unsigned int, page_size
- page_offset
, len
);
2932 memcpy(pages
[nr
] + page_offset
, buf
, size
);
2939 handle
->offset
= offset
;
2942 * Check we didn't copy past our reservation window, taking the
2943 * possible unsigned int wrap into account.
2945 WARN_ON_ONCE(((long)(handle
->head
- handle
->offset
)) < 0);
2948 int perf_output_begin(struct perf_output_handle
*handle
,
2949 struct perf_event
*event
, unsigned int size
,
2950 int nmi
, int sample
)
2952 struct perf_event
*output_event
;
2953 struct perf_mmap_data
*data
;
2954 unsigned long tail
, offset
, head
;
2957 struct perf_event_header header
;
2964 * For inherited events we send all the output towards the parent.
2967 event
= event
->parent
;
2969 output_event
= rcu_dereference(event
->output
);
2971 event
= output_event
;
2973 data
= rcu_dereference(event
->data
);
2977 handle
->data
= data
;
2978 handle
->event
= event
;
2980 handle
->sample
= sample
;
2982 if (!data
->nr_pages
)
2985 have_lost
= atomic_read(&data
->lost
);
2987 size
+= sizeof(lost_event
);
2989 perf_output_lock(handle
);
2993 * Userspace could choose to issue a mb() before updating the
2994 * tail pointer. So that all reads will be completed before the
2997 tail
= ACCESS_ONCE(data
->user_page
->data_tail
);
2999 offset
= head
= atomic_long_read(&data
->head
);
3001 if (unlikely(!perf_output_space(data
, tail
, offset
, head
)))
3003 } while (atomic_long_cmpxchg(&data
->head
, offset
, head
) != offset
);
3005 handle
->offset
= offset
;
3006 handle
->head
= head
;
3008 if (head
- tail
> data
->watermark
)
3009 atomic_set(&data
->wakeup
, 1);
3012 lost_event
.header
.type
= PERF_RECORD_LOST
;
3013 lost_event
.header
.misc
= 0;
3014 lost_event
.header
.size
= sizeof(lost_event
);
3015 lost_event
.id
= event
->id
;
3016 lost_event
.lost
= atomic_xchg(&data
->lost
, 0);
3018 perf_output_put(handle
, lost_event
);
3024 atomic_inc(&data
->lost
);
3025 perf_output_unlock(handle
);
3032 void perf_output_end(struct perf_output_handle
*handle
)
3034 struct perf_event
*event
= handle
->event
;
3035 struct perf_mmap_data
*data
= handle
->data
;
3037 int wakeup_events
= event
->attr
.wakeup_events
;
3039 if (handle
->sample
&& wakeup_events
) {
3040 int events
= atomic_inc_return(&data
->events
);
3041 if (events
>= wakeup_events
) {
3042 atomic_sub(wakeup_events
, &data
->events
);
3043 atomic_set(&data
->wakeup
, 1);
3047 perf_output_unlock(handle
);
3051 static u32
perf_event_pid(struct perf_event
*event
, struct task_struct
*p
)
3054 * only top level events have the pid namespace they were created in
3057 event
= event
->parent
;
3059 return task_tgid_nr_ns(p
, event
->ns
);
3062 static u32
perf_event_tid(struct perf_event
*event
, struct task_struct
*p
)
3065 * only top level events have the pid namespace they were created in
3068 event
= event
->parent
;
3070 return task_pid_nr_ns(p
, event
->ns
);
3073 static void perf_output_read_one(struct perf_output_handle
*handle
,
3074 struct perf_event
*event
)
3076 u64 read_format
= event
->attr
.read_format
;
3080 values
[n
++] = atomic64_read(&event
->count
);
3081 if (read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
) {
3082 values
[n
++] = event
->total_time_enabled
+
3083 atomic64_read(&event
->child_total_time_enabled
);
3085 if (read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
) {
3086 values
[n
++] = event
->total_time_running
+
3087 atomic64_read(&event
->child_total_time_running
);
3089 if (read_format
& PERF_FORMAT_ID
)
3090 values
[n
++] = primary_event_id(event
);
3092 perf_output_copy(handle
, values
, n
* sizeof(u64
));
3096 * XXX PERF_FORMAT_GROUP vs inherited events seems difficult.
3098 static void perf_output_read_group(struct perf_output_handle
*handle
,
3099 struct perf_event
*event
)
3101 struct perf_event
*leader
= event
->group_leader
, *sub
;
3102 u64 read_format
= event
->attr
.read_format
;
3106 values
[n
++] = 1 + leader
->nr_siblings
;
3108 if (read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
)
3109 values
[n
++] = leader
->total_time_enabled
;
3111 if (read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
)
3112 values
[n
++] = leader
->total_time_running
;
3114 if (leader
!= event
)
3115 leader
->pmu
->read(leader
);
3117 values
[n
++] = atomic64_read(&leader
->count
);
3118 if (read_format
& PERF_FORMAT_ID
)
3119 values
[n
++] = primary_event_id(leader
);
3121 perf_output_copy(handle
, values
, n
* sizeof(u64
));
3123 list_for_each_entry(sub
, &leader
->sibling_list
, group_entry
) {
3127 sub
->pmu
->read(sub
);
3129 values
[n
++] = atomic64_read(&sub
->count
);
3130 if (read_format
& PERF_FORMAT_ID
)
3131 values
[n
++] = primary_event_id(sub
);
3133 perf_output_copy(handle
, values
, n
* sizeof(u64
));
3137 static void perf_output_read(struct perf_output_handle
*handle
,
3138 struct perf_event
*event
)
3140 if (event
->attr
.read_format
& PERF_FORMAT_GROUP
)
3141 perf_output_read_group(handle
, event
);
3143 perf_output_read_one(handle
, event
);
3146 void perf_output_sample(struct perf_output_handle
*handle
,
3147 struct perf_event_header
*header
,
3148 struct perf_sample_data
*data
,
3149 struct perf_event
*event
)
3151 u64 sample_type
= data
->type
;
3153 perf_output_put(handle
, *header
);
3155 if (sample_type
& PERF_SAMPLE_IP
)
3156 perf_output_put(handle
, data
->ip
);
3158 if (sample_type
& PERF_SAMPLE_TID
)
3159 perf_output_put(handle
, data
->tid_entry
);
3161 if (sample_type
& PERF_SAMPLE_TIME
)
3162 perf_output_put(handle
, data
->time
);
3164 if (sample_type
& PERF_SAMPLE_ADDR
)
3165 perf_output_put(handle
, data
->addr
);
3167 if (sample_type
& PERF_SAMPLE_ID
)
3168 perf_output_put(handle
, data
->id
);
3170 if (sample_type
& PERF_SAMPLE_STREAM_ID
)
3171 perf_output_put(handle
, data
->stream_id
);
3173 if (sample_type
& PERF_SAMPLE_CPU
)
3174 perf_output_put(handle
, data
->cpu_entry
);
3176 if (sample_type
& PERF_SAMPLE_PERIOD
)
3177 perf_output_put(handle
, data
->period
);
3179 if (sample_type
& PERF_SAMPLE_READ
)
3180 perf_output_read(handle
, event
);
3182 if (sample_type
& PERF_SAMPLE_CALLCHAIN
) {
3183 if (data
->callchain
) {
3186 if (data
->callchain
)
3187 size
+= data
->callchain
->nr
;
3189 size
*= sizeof(u64
);
3191 perf_output_copy(handle
, data
->callchain
, size
);
3194 perf_output_put(handle
, nr
);
3198 if (sample_type
& PERF_SAMPLE_RAW
) {
3200 perf_output_put(handle
, data
->raw
->size
);
3201 perf_output_copy(handle
, data
->raw
->data
,
3208 .size
= sizeof(u32
),
3211 perf_output_put(handle
, raw
);
3216 void perf_prepare_sample(struct perf_event_header
*header
,
3217 struct perf_sample_data
*data
,
3218 struct perf_event
*event
,
3219 struct pt_regs
*regs
)
3221 u64 sample_type
= event
->attr
.sample_type
;
3223 data
->type
= sample_type
;
3225 header
->type
= PERF_RECORD_SAMPLE
;
3226 header
->size
= sizeof(*header
);
3229 header
->misc
|= perf_misc_flags(regs
);
3231 if (sample_type
& PERF_SAMPLE_IP
) {
3232 data
->ip
= perf_instruction_pointer(regs
);
3234 header
->size
+= sizeof(data
->ip
);
3237 if (sample_type
& PERF_SAMPLE_TID
) {
3238 /* namespace issues */
3239 data
->tid_entry
.pid
= perf_event_pid(event
, current
);
3240 data
->tid_entry
.tid
= perf_event_tid(event
, current
);
3242 header
->size
+= sizeof(data
->tid_entry
);
3245 if (sample_type
& PERF_SAMPLE_TIME
) {
3246 data
->time
= perf_clock();
3248 header
->size
+= sizeof(data
->time
);
3251 if (sample_type
& PERF_SAMPLE_ADDR
)
3252 header
->size
+= sizeof(data
->addr
);
3254 if (sample_type
& PERF_SAMPLE_ID
) {
3255 data
->id
= primary_event_id(event
);
3257 header
->size
+= sizeof(data
->id
);
3260 if (sample_type
& PERF_SAMPLE_STREAM_ID
) {
3261 data
->stream_id
= event
->id
;
3263 header
->size
+= sizeof(data
->stream_id
);
3266 if (sample_type
& PERF_SAMPLE_CPU
) {
3267 data
->cpu_entry
.cpu
= raw_smp_processor_id();
3268 data
->cpu_entry
.reserved
= 0;
3270 header
->size
+= sizeof(data
->cpu_entry
);
3273 if (sample_type
& PERF_SAMPLE_PERIOD
)
3274 header
->size
+= sizeof(data
->period
);
3276 if (sample_type
& PERF_SAMPLE_READ
)
3277 header
->size
+= perf_event_read_size(event
);
3279 if (sample_type
& PERF_SAMPLE_CALLCHAIN
) {
3282 data
->callchain
= perf_callchain(regs
);
3284 if (data
->callchain
)
3285 size
+= data
->callchain
->nr
;
3287 header
->size
+= size
* sizeof(u64
);
3290 if (sample_type
& PERF_SAMPLE_RAW
) {
3291 int size
= sizeof(u32
);
3294 size
+= data
->raw
->size
;
3296 size
+= sizeof(u32
);
3298 WARN_ON_ONCE(size
& (sizeof(u64
)-1));
3299 header
->size
+= size
;
3303 static void perf_event_output(struct perf_event
*event
, int nmi
,
3304 struct perf_sample_data
*data
,
3305 struct pt_regs
*regs
)
3307 struct perf_output_handle handle
;
3308 struct perf_event_header header
;
3310 perf_prepare_sample(&header
, data
, event
, regs
);
3312 if (perf_output_begin(&handle
, event
, header
.size
, nmi
, 1))
3315 perf_output_sample(&handle
, &header
, data
, event
);
3317 perf_output_end(&handle
);
3324 struct perf_read_event
{
3325 struct perf_event_header header
;
3332 perf_event_read_event(struct perf_event
*event
,
3333 struct task_struct
*task
)
3335 struct perf_output_handle handle
;
3336 struct perf_read_event read_event
= {
3338 .type
= PERF_RECORD_READ
,
3340 .size
= sizeof(read_event
) + perf_event_read_size(event
),
3342 .pid
= perf_event_pid(event
, task
),
3343 .tid
= perf_event_tid(event
, task
),
3347 ret
= perf_output_begin(&handle
, event
, read_event
.header
.size
, 0, 0);
3351 perf_output_put(&handle
, read_event
);
3352 perf_output_read(&handle
, event
);
3354 perf_output_end(&handle
);
3358 * task tracking -- fork/exit
3360 * enabled by: attr.comm | attr.mmap | attr.task
3363 struct perf_task_event
{
3364 struct task_struct
*task
;
3365 struct perf_event_context
*task_ctx
;
3368 struct perf_event_header header
;
3378 static void perf_event_task_output(struct perf_event
*event
,
3379 struct perf_task_event
*task_event
)
3381 struct perf_output_handle handle
;
3382 struct task_struct
*task
= task_event
->task
;
3383 unsigned long flags
;
3387 * If this CPU attempts to acquire an rq lock held by a CPU spinning
3388 * in perf_output_lock() from interrupt context, it's game over.
3390 local_irq_save(flags
);
3392 size
= task_event
->event_id
.header
.size
;
3393 ret
= perf_output_begin(&handle
, event
, size
, 0, 0);
3396 local_irq_restore(flags
);
3400 task_event
->event_id
.pid
= perf_event_pid(event
, task
);
3401 task_event
->event_id
.ppid
= perf_event_pid(event
, current
);
3403 task_event
->event_id
.tid
= perf_event_tid(event
, task
);
3404 task_event
->event_id
.ptid
= perf_event_tid(event
, current
);
3406 perf_output_put(&handle
, task_event
->event_id
);
3408 perf_output_end(&handle
);
3409 local_irq_restore(flags
);
3412 static int perf_event_task_match(struct perf_event
*event
)
3414 if (event
->state
< PERF_EVENT_STATE_INACTIVE
)
3417 if (event
->cpu
!= -1 && event
->cpu
!= smp_processor_id())
3420 if (event
->attr
.comm
|| event
->attr
.mmap
|| event
->attr
.task
)
3426 static void perf_event_task_ctx(struct perf_event_context
*ctx
,
3427 struct perf_task_event
*task_event
)
3429 struct perf_event
*event
;
3431 list_for_each_entry_rcu(event
, &ctx
->event_list
, event_entry
) {
3432 if (perf_event_task_match(event
))
3433 perf_event_task_output(event
, task_event
);
3437 static void perf_event_task_event(struct perf_task_event
*task_event
)
3439 struct perf_cpu_context
*cpuctx
;
3440 struct perf_event_context
*ctx
= task_event
->task_ctx
;
3443 cpuctx
= &get_cpu_var(perf_cpu_context
);
3444 perf_event_task_ctx(&cpuctx
->ctx
, task_event
);
3446 ctx
= rcu_dereference(current
->perf_event_ctxp
);
3448 perf_event_task_ctx(ctx
, task_event
);
3449 put_cpu_var(perf_cpu_context
);
3453 static void perf_event_task(struct task_struct
*task
,
3454 struct perf_event_context
*task_ctx
,
3457 struct perf_task_event task_event
;
3459 if (!atomic_read(&nr_comm_events
) &&
3460 !atomic_read(&nr_mmap_events
) &&
3461 !atomic_read(&nr_task_events
))
3464 task_event
= (struct perf_task_event
){
3466 .task_ctx
= task_ctx
,
3469 .type
= new ? PERF_RECORD_FORK
: PERF_RECORD_EXIT
,
3471 .size
= sizeof(task_event
.event_id
),
3477 .time
= perf_clock(),
3481 perf_event_task_event(&task_event
);
3484 void perf_event_fork(struct task_struct
*task
)
3486 perf_event_task(task
, NULL
, 1);
3493 struct perf_comm_event
{
3494 struct task_struct
*task
;
3499 struct perf_event_header header
;
3506 static void perf_event_comm_output(struct perf_event
*event
,
3507 struct perf_comm_event
*comm_event
)
3509 struct perf_output_handle handle
;
3510 int size
= comm_event
->event_id
.header
.size
;
3511 int ret
= perf_output_begin(&handle
, event
, size
, 0, 0);
3516 comm_event
->event_id
.pid
= perf_event_pid(event
, comm_event
->task
);
3517 comm_event
->event_id
.tid
= perf_event_tid(event
, comm_event
->task
);
3519 perf_output_put(&handle
, comm_event
->event_id
);
3520 perf_output_copy(&handle
, comm_event
->comm
,
3521 comm_event
->comm_size
);
3522 perf_output_end(&handle
);
3525 static int perf_event_comm_match(struct perf_event
*event
)
3527 if (event
->state
< PERF_EVENT_STATE_INACTIVE
)
3530 if (event
->cpu
!= -1 && event
->cpu
!= smp_processor_id())
3533 if (event
->attr
.comm
)
3539 static void perf_event_comm_ctx(struct perf_event_context
*ctx
,
3540 struct perf_comm_event
*comm_event
)
3542 struct perf_event
*event
;
3544 list_for_each_entry_rcu(event
, &ctx
->event_list
, event_entry
) {
3545 if (perf_event_comm_match(event
))
3546 perf_event_comm_output(event
, comm_event
);
3550 static void perf_event_comm_event(struct perf_comm_event
*comm_event
)
3552 struct perf_cpu_context
*cpuctx
;
3553 struct perf_event_context
*ctx
;
3555 char comm
[TASK_COMM_LEN
];
3557 memset(comm
, 0, sizeof(comm
));
3558 strlcpy(comm
, comm_event
->task
->comm
, sizeof(comm
));
3559 size
= ALIGN(strlen(comm
)+1, sizeof(u64
));
3561 comm_event
->comm
= comm
;
3562 comm_event
->comm_size
= size
;
3564 comm_event
->event_id
.header
.size
= sizeof(comm_event
->event_id
) + size
;
3567 cpuctx
= &get_cpu_var(perf_cpu_context
);
3568 perf_event_comm_ctx(&cpuctx
->ctx
, comm_event
);
3569 ctx
= rcu_dereference(current
->perf_event_ctxp
);
3571 perf_event_comm_ctx(ctx
, comm_event
);
3572 put_cpu_var(perf_cpu_context
);
3576 void perf_event_comm(struct task_struct
*task
)
3578 struct perf_comm_event comm_event
;
3580 if (task
->perf_event_ctxp
)
3581 perf_event_enable_on_exec(task
);
3583 if (!atomic_read(&nr_comm_events
))
3586 comm_event
= (struct perf_comm_event
){
3592 .type
= PERF_RECORD_COMM
,
3601 perf_event_comm_event(&comm_event
);
3608 struct perf_mmap_event
{
3609 struct vm_area_struct
*vma
;
3611 const char *file_name
;
3615 struct perf_event_header header
;
3625 static void perf_event_mmap_output(struct perf_event
*event
,
3626 struct perf_mmap_event
*mmap_event
)
3628 struct perf_output_handle handle
;
3629 int size
= mmap_event
->event_id
.header
.size
;
3630 int ret
= perf_output_begin(&handle
, event
, size
, 0, 0);
3635 mmap_event
->event_id
.pid
= perf_event_pid(event
, current
);
3636 mmap_event
->event_id
.tid
= perf_event_tid(event
, current
);
3638 perf_output_put(&handle
, mmap_event
->event_id
);
3639 perf_output_copy(&handle
, mmap_event
->file_name
,
3640 mmap_event
->file_size
);
3641 perf_output_end(&handle
);
3644 static int perf_event_mmap_match(struct perf_event
*event
,
3645 struct perf_mmap_event
*mmap_event
)
3647 if (event
->state
< PERF_EVENT_STATE_INACTIVE
)
3650 if (event
->cpu
!= -1 && event
->cpu
!= smp_processor_id())
3653 if (event
->attr
.mmap
)
3659 static void perf_event_mmap_ctx(struct perf_event_context
*ctx
,
3660 struct perf_mmap_event
*mmap_event
)
3662 struct perf_event
*event
;
3664 list_for_each_entry_rcu(event
, &ctx
->event_list
, event_entry
) {
3665 if (perf_event_mmap_match(event
, mmap_event
))
3666 perf_event_mmap_output(event
, mmap_event
);
3670 static void perf_event_mmap_event(struct perf_mmap_event
*mmap_event
)
3672 struct perf_cpu_context
*cpuctx
;
3673 struct perf_event_context
*ctx
;
3674 struct vm_area_struct
*vma
= mmap_event
->vma
;
3675 struct file
*file
= vma
->vm_file
;
3681 memset(tmp
, 0, sizeof(tmp
));
3685 * d_path works from the end of the buffer backwards, so we
3686 * need to add enough zero bytes after the string to handle
3687 * the 64bit alignment we do later.
3689 buf
= kzalloc(PATH_MAX
+ sizeof(u64
), GFP_KERNEL
);
3691 name
= strncpy(tmp
, "//enomem", sizeof(tmp
));
3694 name
= d_path(&file
->f_path
, buf
, PATH_MAX
);
3696 name
= strncpy(tmp
, "//toolong", sizeof(tmp
));
3700 if (arch_vma_name(mmap_event
->vma
)) {
3701 name
= strncpy(tmp
, arch_vma_name(mmap_event
->vma
),
3707 name
= strncpy(tmp
, "[vdso]", sizeof(tmp
));
3711 name
= strncpy(tmp
, "//anon", sizeof(tmp
));
3716 size
= ALIGN(strlen(name
)+1, sizeof(u64
));
3718 mmap_event
->file_name
= name
;
3719 mmap_event
->file_size
= size
;
3721 mmap_event
->event_id
.header
.size
= sizeof(mmap_event
->event_id
) + size
;
3724 cpuctx
= &get_cpu_var(perf_cpu_context
);
3725 perf_event_mmap_ctx(&cpuctx
->ctx
, mmap_event
);
3726 ctx
= rcu_dereference(current
->perf_event_ctxp
);
3728 perf_event_mmap_ctx(ctx
, mmap_event
);
3729 put_cpu_var(perf_cpu_context
);
3735 void __perf_event_mmap(struct vm_area_struct
*vma
)
3737 struct perf_mmap_event mmap_event
;
3739 if (!atomic_read(&nr_mmap_events
))
3742 mmap_event
= (struct perf_mmap_event
){
3748 .type
= PERF_RECORD_MMAP
,
3754 .start
= vma
->vm_start
,
3755 .len
= vma
->vm_end
- vma
->vm_start
,
3756 .pgoff
= (u64
)vma
->vm_pgoff
<< PAGE_SHIFT
,
3760 perf_event_mmap_event(&mmap_event
);
3764 * IRQ throttle logging
3767 static void perf_log_throttle(struct perf_event
*event
, int enable
)
3769 struct perf_output_handle handle
;
3773 struct perf_event_header header
;
3777 } throttle_event
= {
3779 .type
= PERF_RECORD_THROTTLE
,
3781 .size
= sizeof(throttle_event
),
3783 .time
= perf_clock(),
3784 .id
= primary_event_id(event
),
3785 .stream_id
= event
->id
,
3789 throttle_event
.header
.type
= PERF_RECORD_UNTHROTTLE
;
3791 ret
= perf_output_begin(&handle
, event
, sizeof(throttle_event
), 1, 0);
3795 perf_output_put(&handle
, throttle_event
);
3796 perf_output_end(&handle
);
3800 * Generic event overflow handling, sampling.
3803 static int __perf_event_overflow(struct perf_event
*event
, int nmi
,
3804 int throttle
, struct perf_sample_data
*data
,
3805 struct pt_regs
*regs
)
3807 int events
= atomic_read(&event
->event_limit
);
3808 struct hw_perf_event
*hwc
= &event
->hw
;
3811 throttle
= (throttle
&& event
->pmu
->unthrottle
!= NULL
);
3816 if (hwc
->interrupts
!= MAX_INTERRUPTS
) {
3818 if (HZ
* hwc
->interrupts
>
3819 (u64
)sysctl_perf_event_sample_rate
) {
3820 hwc
->interrupts
= MAX_INTERRUPTS
;
3821 perf_log_throttle(event
, 0);
3826 * Keep re-disabling events even though on the previous
3827 * pass we disabled it - just in case we raced with a
3828 * sched-in and the event got enabled again:
3834 if (event
->attr
.freq
) {
3835 u64 now
= perf_clock();
3836 s64 delta
= now
- hwc
->freq_time_stamp
;
3838 hwc
->freq_time_stamp
= now
;
3840 if (delta
> 0 && delta
< 2*TICK_NSEC
)
3841 perf_adjust_period(event
, delta
, hwc
->last_period
);
3845 * XXX event_limit might not quite work as expected on inherited
3849 event
->pending_kill
= POLL_IN
;
3850 if (events
&& atomic_dec_and_test(&event
->event_limit
)) {
3852 event
->pending_kill
= POLL_HUP
;
3854 event
->pending_disable
= 1;
3855 perf_pending_queue(&event
->pending
,
3856 perf_pending_event
);
3858 perf_event_disable(event
);
3861 if (event
->overflow_handler
)
3862 event
->overflow_handler(event
, nmi
, data
, regs
);
3864 perf_event_output(event
, nmi
, data
, regs
);
3869 int perf_event_overflow(struct perf_event
*event
, int nmi
,
3870 struct perf_sample_data
*data
,
3871 struct pt_regs
*regs
)
3873 return __perf_event_overflow(event
, nmi
, 1, data
, regs
);
3877 * Generic software event infrastructure
3881 * We directly increment event->count and keep a second value in
3882 * event->hw.period_left to count intervals. This period event
3883 * is kept in the range [-sample_period, 0] so that we can use the
3887 static u64
perf_swevent_set_period(struct perf_event
*event
)
3889 struct hw_perf_event
*hwc
= &event
->hw
;
3890 u64 period
= hwc
->last_period
;
3894 hwc
->last_period
= hwc
->sample_period
;
3897 old
= val
= atomic64_read(&hwc
->period_left
);
3901 nr
= div64_u64(period
+ val
, period
);
3902 offset
= nr
* period
;
3904 if (atomic64_cmpxchg(&hwc
->period_left
, old
, val
) != old
)
3910 static void perf_swevent_overflow(struct perf_event
*event
, u64 overflow
,
3911 int nmi
, struct perf_sample_data
*data
,
3912 struct pt_regs
*regs
)
3914 struct hw_perf_event
*hwc
= &event
->hw
;
3917 data
->period
= event
->hw
.last_period
;
3919 overflow
= perf_swevent_set_period(event
);
3921 if (hwc
->interrupts
== MAX_INTERRUPTS
)
3924 for (; overflow
; overflow
--) {
3925 if (__perf_event_overflow(event
, nmi
, throttle
,
3928 * We inhibit the overflow from happening when
3929 * hwc->interrupts == MAX_INTERRUPTS.
3937 static void perf_swevent_unthrottle(struct perf_event
*event
)
3940 * Nothing to do, we already reset hwc->interrupts.
3944 static void perf_swevent_add(struct perf_event
*event
, u64 nr
,
3945 int nmi
, struct perf_sample_data
*data
,
3946 struct pt_regs
*regs
)
3948 struct hw_perf_event
*hwc
= &event
->hw
;
3950 atomic64_add(nr
, &event
->count
);
3955 if (!hwc
->sample_period
)
3958 if (nr
== 1 && hwc
->sample_period
== 1 && !event
->attr
.freq
)
3959 return perf_swevent_overflow(event
, 1, nmi
, data
, regs
);
3961 if (atomic64_add_negative(nr
, &hwc
->period_left
))
3964 perf_swevent_overflow(event
, 0, nmi
, data
, regs
);
3967 static int perf_swevent_is_counting(struct perf_event
*event
)
3970 * The event is active, we're good!
3972 if (event
->state
== PERF_EVENT_STATE_ACTIVE
)
3976 * The event is off/error, not counting.
3978 if (event
->state
!= PERF_EVENT_STATE_INACTIVE
)
3982 * The event is inactive, if the context is active
3983 * we're part of a group that didn't make it on the 'pmu',
3986 if (event
->ctx
->is_active
)
3990 * We're inactive and the context is too, this means the
3991 * task is scheduled out, we're counting events that happen
3992 * to us, like migration events.
3997 static int perf_tp_event_match(struct perf_event
*event
,
3998 struct perf_sample_data
*data
);
4000 static int perf_exclude_event(struct perf_event
*event
,
4001 struct pt_regs
*regs
)
4004 if (event
->attr
.exclude_user
&& user_mode(regs
))
4007 if (event
->attr
.exclude_kernel
&& !user_mode(regs
))
4014 static int perf_swevent_match(struct perf_event
*event
,
4015 enum perf_type_id type
,
4017 struct perf_sample_data
*data
,
4018 struct pt_regs
*regs
)
4020 if (event
->cpu
!= -1 && event
->cpu
!= smp_processor_id())
4023 if (!perf_swevent_is_counting(event
))
4026 if (event
->attr
.type
!= type
)
4029 if (event
->attr
.config
!= event_id
)
4032 if (perf_exclude_event(event
, regs
))
4035 if (event
->attr
.type
== PERF_TYPE_TRACEPOINT
&&
4036 !perf_tp_event_match(event
, data
))
4042 static void perf_swevent_ctx_event(struct perf_event_context
*ctx
,
4043 enum perf_type_id type
,
4044 u32 event_id
, u64 nr
, int nmi
,
4045 struct perf_sample_data
*data
,
4046 struct pt_regs
*regs
)
4048 struct perf_event
*event
;
4050 list_for_each_entry_rcu(event
, &ctx
->event_list
, event_entry
) {
4051 if (perf_swevent_match(event
, type
, event_id
, data
, regs
))
4052 perf_swevent_add(event
, nr
, nmi
, data
, regs
);
4056 int perf_swevent_get_recursion_context(void)
4058 struct perf_cpu_context
*cpuctx
= &get_cpu_var(perf_cpu_context
);
4065 else if (in_softirq())
4070 if (cpuctx
->recursion
[rctx
]) {
4071 put_cpu_var(perf_cpu_context
);
4075 cpuctx
->recursion
[rctx
]++;
4080 EXPORT_SYMBOL_GPL(perf_swevent_get_recursion_context
);
4082 void perf_swevent_put_recursion_context(int rctx
)
4084 struct perf_cpu_context
*cpuctx
= &__get_cpu_var(perf_cpu_context
);
4086 cpuctx
->recursion
[rctx
]--;
4087 put_cpu_var(perf_cpu_context
);
4089 EXPORT_SYMBOL_GPL(perf_swevent_put_recursion_context
);
4091 static void do_perf_sw_event(enum perf_type_id type
, u32 event_id
,
4093 struct perf_sample_data
*data
,
4094 struct pt_regs
*regs
)
4096 struct perf_cpu_context
*cpuctx
;
4097 struct perf_event_context
*ctx
;
4099 cpuctx
= &__get_cpu_var(perf_cpu_context
);
4101 perf_swevent_ctx_event(&cpuctx
->ctx
, type
, event_id
,
4102 nr
, nmi
, data
, regs
);
4104 * doesn't really matter which of the child contexts the
4105 * events ends up in.
4107 ctx
= rcu_dereference(current
->perf_event_ctxp
);
4109 perf_swevent_ctx_event(ctx
, type
, event_id
, nr
, nmi
, data
, regs
);
4113 void __perf_sw_event(u32 event_id
, u64 nr
, int nmi
,
4114 struct pt_regs
*regs
, u64 addr
)
4116 struct perf_sample_data data
;
4119 rctx
= perf_swevent_get_recursion_context();
4123 perf_sample_data_init(&data
, addr
);
4125 do_perf_sw_event(PERF_TYPE_SOFTWARE
, event_id
, nr
, nmi
, &data
, regs
);
4127 perf_swevent_put_recursion_context(rctx
);
4130 static void perf_swevent_read(struct perf_event
*event
)
4134 static int perf_swevent_enable(struct perf_event
*event
)
4136 struct hw_perf_event
*hwc
= &event
->hw
;
4138 if (hwc
->sample_period
) {
4139 hwc
->last_period
= hwc
->sample_period
;
4140 perf_swevent_set_period(event
);
4145 static void perf_swevent_disable(struct perf_event
*event
)
4149 static const struct pmu perf_ops_generic
= {
4150 .enable
= perf_swevent_enable
,
4151 .disable
= perf_swevent_disable
,
4152 .read
= perf_swevent_read
,
4153 .unthrottle
= perf_swevent_unthrottle
,
4157 * hrtimer based swevent callback
4160 static enum hrtimer_restart
perf_swevent_hrtimer(struct hrtimer
*hrtimer
)
4162 enum hrtimer_restart ret
= HRTIMER_RESTART
;
4163 struct perf_sample_data data
;
4164 struct pt_regs
*regs
;
4165 struct perf_event
*event
;
4168 event
= container_of(hrtimer
, struct perf_event
, hw
.hrtimer
);
4169 event
->pmu
->read(event
);
4171 perf_sample_data_init(&data
, 0);
4172 data
.period
= event
->hw
.last_period
;
4173 regs
= get_irq_regs();
4175 * In case we exclude kernel IPs or are somehow not in interrupt
4176 * context, provide the next best thing, the user IP.
4178 if ((event
->attr
.exclude_kernel
|| !regs
) &&
4179 !event
->attr
.exclude_user
)
4180 regs
= task_pt_regs(current
);
4183 if (!(event
->attr
.exclude_idle
&& current
->pid
== 0))
4184 if (perf_event_overflow(event
, 0, &data
, regs
))
4185 ret
= HRTIMER_NORESTART
;
4188 period
= max_t(u64
, 10000, event
->hw
.sample_period
);
4189 hrtimer_forward_now(hrtimer
, ns_to_ktime(period
));
4194 static void perf_swevent_start_hrtimer(struct perf_event
*event
)
4196 struct hw_perf_event
*hwc
= &event
->hw
;
4198 hrtimer_init(&hwc
->hrtimer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
4199 hwc
->hrtimer
.function
= perf_swevent_hrtimer
;
4200 if (hwc
->sample_period
) {
4203 if (hwc
->remaining
) {
4204 if (hwc
->remaining
< 0)
4207 period
= hwc
->remaining
;
4210 period
= max_t(u64
, 10000, hwc
->sample_period
);
4212 __hrtimer_start_range_ns(&hwc
->hrtimer
,
4213 ns_to_ktime(period
), 0,
4214 HRTIMER_MODE_REL
, 0);
4218 static void perf_swevent_cancel_hrtimer(struct perf_event
*event
)
4220 struct hw_perf_event
*hwc
= &event
->hw
;
4222 if (hwc
->sample_period
) {
4223 ktime_t remaining
= hrtimer_get_remaining(&hwc
->hrtimer
);
4224 hwc
->remaining
= ktime_to_ns(remaining
);
4226 hrtimer_cancel(&hwc
->hrtimer
);
4231 * Software event: cpu wall time clock
4234 static void cpu_clock_perf_event_update(struct perf_event
*event
)
4236 int cpu
= raw_smp_processor_id();
4240 now
= cpu_clock(cpu
);
4241 prev
= atomic64_xchg(&event
->hw
.prev_count
, now
);
4242 atomic64_add(now
- prev
, &event
->count
);
4245 static int cpu_clock_perf_event_enable(struct perf_event
*event
)
4247 struct hw_perf_event
*hwc
= &event
->hw
;
4248 int cpu
= raw_smp_processor_id();
4250 atomic64_set(&hwc
->prev_count
, cpu_clock(cpu
));
4251 perf_swevent_start_hrtimer(event
);
4256 static void cpu_clock_perf_event_disable(struct perf_event
*event
)
4258 perf_swevent_cancel_hrtimer(event
);
4259 cpu_clock_perf_event_update(event
);
4262 static void cpu_clock_perf_event_read(struct perf_event
*event
)
4264 cpu_clock_perf_event_update(event
);
4267 static const struct pmu perf_ops_cpu_clock
= {
4268 .enable
= cpu_clock_perf_event_enable
,
4269 .disable
= cpu_clock_perf_event_disable
,
4270 .read
= cpu_clock_perf_event_read
,
4274 * Software event: task time clock
4277 static void task_clock_perf_event_update(struct perf_event
*event
, u64 now
)
4282 prev
= atomic64_xchg(&event
->hw
.prev_count
, now
);
4284 atomic64_add(delta
, &event
->count
);
4287 static int task_clock_perf_event_enable(struct perf_event
*event
)
4289 struct hw_perf_event
*hwc
= &event
->hw
;
4292 now
= event
->ctx
->time
;
4294 atomic64_set(&hwc
->prev_count
, now
);
4296 perf_swevent_start_hrtimer(event
);
4301 static void task_clock_perf_event_disable(struct perf_event
*event
)
4303 perf_swevent_cancel_hrtimer(event
);
4304 task_clock_perf_event_update(event
, event
->ctx
->time
);
4308 static void task_clock_perf_event_read(struct perf_event
*event
)
4313 update_context_time(event
->ctx
);
4314 time
= event
->ctx
->time
;
4316 u64 now
= perf_clock();
4317 u64 delta
= now
- event
->ctx
->timestamp
;
4318 time
= event
->ctx
->time
+ delta
;
4321 task_clock_perf_event_update(event
, time
);
4324 static const struct pmu perf_ops_task_clock
= {
4325 .enable
= task_clock_perf_event_enable
,
4326 .disable
= task_clock_perf_event_disable
,
4327 .read
= task_clock_perf_event_read
,
4330 #ifdef CONFIG_EVENT_TRACING
4332 void perf_tp_event(int event_id
, u64 addr
, u64 count
, void *record
,
4333 int entry_size
, struct pt_regs
*regs
)
4335 struct perf_sample_data data
;
4336 struct perf_raw_record raw
= {
4341 perf_sample_data_init(&data
, addr
);
4344 /* Trace events already protected against recursion */
4345 do_perf_sw_event(PERF_TYPE_TRACEPOINT
, event_id
, count
, 1,
4348 EXPORT_SYMBOL_GPL(perf_tp_event
);
4350 static int perf_tp_event_match(struct perf_event
*event
,
4351 struct perf_sample_data
*data
)
4353 void *record
= data
->raw
->data
;
4355 if (likely(!event
->filter
) || filter_match_preds(event
->filter
, record
))
4360 static void tp_perf_event_destroy(struct perf_event
*event
)
4362 perf_trace_disable(event
->attr
.config
);
4365 static const struct pmu
*tp_perf_event_init(struct perf_event
*event
)
4368 * Raw tracepoint data is a severe data leak, only allow root to
4371 if ((event
->attr
.sample_type
& PERF_SAMPLE_RAW
) &&
4372 perf_paranoid_tracepoint_raw() &&
4373 !capable(CAP_SYS_ADMIN
))
4374 return ERR_PTR(-EPERM
);
4376 if (perf_trace_enable(event
->attr
.config
))
4379 event
->destroy
= tp_perf_event_destroy
;
4381 return &perf_ops_generic
;
4384 static int perf_event_set_filter(struct perf_event
*event
, void __user
*arg
)
4389 if (event
->attr
.type
!= PERF_TYPE_TRACEPOINT
)
4392 filter_str
= strndup_user(arg
, PAGE_SIZE
);
4393 if (IS_ERR(filter_str
))
4394 return PTR_ERR(filter_str
);
4396 ret
= ftrace_profile_set_filter(event
, event
->attr
.config
, filter_str
);
4402 static void perf_event_free_filter(struct perf_event
*event
)
4404 ftrace_profile_free_filter(event
);
4409 static int perf_tp_event_match(struct perf_event
*event
,
4410 struct perf_sample_data
*data
)
4415 static const struct pmu
*tp_perf_event_init(struct perf_event
*event
)
4420 static int perf_event_set_filter(struct perf_event
*event
, void __user
*arg
)
4425 static void perf_event_free_filter(struct perf_event
*event
)
4429 #endif /* CONFIG_EVENT_TRACING */
4431 #ifdef CONFIG_HAVE_HW_BREAKPOINT
4432 static void bp_perf_event_destroy(struct perf_event
*event
)
4434 release_bp_slot(event
);
4437 static const struct pmu
*bp_perf_event_init(struct perf_event
*bp
)
4441 err
= register_perf_hw_breakpoint(bp
);
4443 return ERR_PTR(err
);
4445 bp
->destroy
= bp_perf_event_destroy
;
4447 return &perf_ops_bp
;
4450 void perf_bp_event(struct perf_event
*bp
, void *data
)
4452 struct perf_sample_data sample
;
4453 struct pt_regs
*regs
= data
;
4455 perf_sample_data_init(&sample
, bp
->attr
.bp_addr
);
4457 if (!perf_exclude_event(bp
, regs
))
4458 perf_swevent_add(bp
, 1, 1, &sample
, regs
);
4461 static const struct pmu
*bp_perf_event_init(struct perf_event
*bp
)
4466 void perf_bp_event(struct perf_event
*bp
, void *regs
)
4471 atomic_t perf_swevent_enabled
[PERF_COUNT_SW_MAX
];
4473 static void sw_perf_event_destroy(struct perf_event
*event
)
4475 u64 event_id
= event
->attr
.config
;
4477 WARN_ON(event
->parent
);
4479 atomic_dec(&perf_swevent_enabled
[event_id
]);
4482 static const struct pmu
*sw_perf_event_init(struct perf_event
*event
)
4484 const struct pmu
*pmu
= NULL
;
4485 u64 event_id
= event
->attr
.config
;
4488 * Software events (currently) can't in general distinguish
4489 * between user, kernel and hypervisor events.
4490 * However, context switches and cpu migrations are considered
4491 * to be kernel events, and page faults are never hypervisor
4495 case PERF_COUNT_SW_CPU_CLOCK
:
4496 pmu
= &perf_ops_cpu_clock
;
4499 case PERF_COUNT_SW_TASK_CLOCK
:
4501 * If the user instantiates this as a per-cpu event,
4502 * use the cpu_clock event instead.
4504 if (event
->ctx
->task
)
4505 pmu
= &perf_ops_task_clock
;
4507 pmu
= &perf_ops_cpu_clock
;
4510 case PERF_COUNT_SW_PAGE_FAULTS
:
4511 case PERF_COUNT_SW_PAGE_FAULTS_MIN
:
4512 case PERF_COUNT_SW_PAGE_FAULTS_MAJ
:
4513 case PERF_COUNT_SW_CONTEXT_SWITCHES
:
4514 case PERF_COUNT_SW_CPU_MIGRATIONS
:
4515 case PERF_COUNT_SW_ALIGNMENT_FAULTS
:
4516 case PERF_COUNT_SW_EMULATION_FAULTS
:
4517 if (!event
->parent
) {
4518 atomic_inc(&perf_swevent_enabled
[event_id
]);
4519 event
->destroy
= sw_perf_event_destroy
;
4521 pmu
= &perf_ops_generic
;
4529 * Allocate and initialize a event structure
4531 static struct perf_event
*
4532 perf_event_alloc(struct perf_event_attr
*attr
,
4534 struct perf_event_context
*ctx
,
4535 struct perf_event
*group_leader
,
4536 struct perf_event
*parent_event
,
4537 perf_overflow_handler_t overflow_handler
,
4540 const struct pmu
*pmu
;
4541 struct perf_event
*event
;
4542 struct hw_perf_event
*hwc
;
4545 event
= kzalloc(sizeof(*event
), gfpflags
);
4547 return ERR_PTR(-ENOMEM
);
4550 * Single events are their own group leaders, with an
4551 * empty sibling list:
4554 group_leader
= event
;
4556 mutex_init(&event
->child_mutex
);
4557 INIT_LIST_HEAD(&event
->child_list
);
4559 INIT_LIST_HEAD(&event
->group_entry
);
4560 INIT_LIST_HEAD(&event
->event_entry
);
4561 INIT_LIST_HEAD(&event
->sibling_list
);
4562 init_waitqueue_head(&event
->waitq
);
4564 mutex_init(&event
->mmap_mutex
);
4567 event
->attr
= *attr
;
4568 event
->group_leader
= group_leader
;
4573 event
->parent
= parent_event
;
4575 event
->ns
= get_pid_ns(current
->nsproxy
->pid_ns
);
4576 event
->id
= atomic64_inc_return(&perf_event_id
);
4578 event
->state
= PERF_EVENT_STATE_INACTIVE
;
4580 if (!overflow_handler
&& parent_event
)
4581 overflow_handler
= parent_event
->overflow_handler
;
4583 event
->overflow_handler
= overflow_handler
;
4586 event
->state
= PERF_EVENT_STATE_OFF
;
4591 hwc
->sample_period
= attr
->sample_period
;
4592 if (attr
->freq
&& attr
->sample_freq
)
4593 hwc
->sample_period
= 1;
4594 hwc
->last_period
= hwc
->sample_period
;
4596 atomic64_set(&hwc
->period_left
, hwc
->sample_period
);
4599 * we currently do not support PERF_FORMAT_GROUP on inherited events
4601 if (attr
->inherit
&& (attr
->read_format
& PERF_FORMAT_GROUP
))
4604 switch (attr
->type
) {
4606 case PERF_TYPE_HARDWARE
:
4607 case PERF_TYPE_HW_CACHE
:
4608 pmu
= hw_perf_event_init(event
);
4611 case PERF_TYPE_SOFTWARE
:
4612 pmu
= sw_perf_event_init(event
);
4615 case PERF_TYPE_TRACEPOINT
:
4616 pmu
= tp_perf_event_init(event
);
4619 case PERF_TYPE_BREAKPOINT
:
4620 pmu
= bp_perf_event_init(event
);
4631 else if (IS_ERR(pmu
))
4636 put_pid_ns(event
->ns
);
4638 return ERR_PTR(err
);
4643 if (!event
->parent
) {
4644 atomic_inc(&nr_events
);
4645 if (event
->attr
.mmap
)
4646 atomic_inc(&nr_mmap_events
);
4647 if (event
->attr
.comm
)
4648 atomic_inc(&nr_comm_events
);
4649 if (event
->attr
.task
)
4650 atomic_inc(&nr_task_events
);
4656 static int perf_copy_attr(struct perf_event_attr __user
*uattr
,
4657 struct perf_event_attr
*attr
)
4662 if (!access_ok(VERIFY_WRITE
, uattr
, PERF_ATTR_SIZE_VER0
))
4666 * zero the full structure, so that a short copy will be nice.
4668 memset(attr
, 0, sizeof(*attr
));
4670 ret
= get_user(size
, &uattr
->size
);
4674 if (size
> PAGE_SIZE
) /* silly large */
4677 if (!size
) /* abi compat */
4678 size
= PERF_ATTR_SIZE_VER0
;
4680 if (size
< PERF_ATTR_SIZE_VER0
)
4684 * If we're handed a bigger struct than we know of,
4685 * ensure all the unknown bits are 0 - i.e. new
4686 * user-space does not rely on any kernel feature
4687 * extensions we dont know about yet.
4689 if (size
> sizeof(*attr
)) {
4690 unsigned char __user
*addr
;
4691 unsigned char __user
*end
;
4694 addr
= (void __user
*)uattr
+ sizeof(*attr
);
4695 end
= (void __user
*)uattr
+ size
;
4697 for (; addr
< end
; addr
++) {
4698 ret
= get_user(val
, addr
);
4704 size
= sizeof(*attr
);
4707 ret
= copy_from_user(attr
, uattr
, size
);
4712 * If the type exists, the corresponding creation will verify
4715 if (attr
->type
>= PERF_TYPE_MAX
)
4718 if (attr
->__reserved_1
)
4721 if (attr
->sample_type
& ~(PERF_SAMPLE_MAX
-1))
4724 if (attr
->read_format
& ~(PERF_FORMAT_MAX
-1))
4731 put_user(sizeof(*attr
), &uattr
->size
);
4736 static int perf_event_set_output(struct perf_event
*event
, int output_fd
)
4738 struct perf_event
*output_event
= NULL
;
4739 struct file
*output_file
= NULL
;
4740 struct perf_event
*old_output
;
4741 int fput_needed
= 0;
4747 output_file
= fget_light(output_fd
, &fput_needed
);
4751 if (output_file
->f_op
!= &perf_fops
)
4754 output_event
= output_file
->private_data
;
4756 /* Don't chain output fds */
4757 if (output_event
->output
)
4760 /* Don't set an output fd when we already have an output channel */
4764 atomic_long_inc(&output_file
->f_count
);
4767 mutex_lock(&event
->mmap_mutex
);
4768 old_output
= event
->output
;
4769 rcu_assign_pointer(event
->output
, output_event
);
4770 mutex_unlock(&event
->mmap_mutex
);
4774 * we need to make sure no existing perf_output_*()
4775 * is still referencing this event.
4778 fput(old_output
->filp
);
4783 fput_light(output_file
, fput_needed
);
4788 * sys_perf_event_open - open a performance event, associate it to a task/cpu
4790 * @attr_uptr: event_id type attributes for monitoring/sampling
4793 * @group_fd: group leader event fd
4795 SYSCALL_DEFINE5(perf_event_open
,
4796 struct perf_event_attr __user
*, attr_uptr
,
4797 pid_t
, pid
, int, cpu
, int, group_fd
, unsigned long, flags
)
4799 struct perf_event
*event
, *group_leader
;
4800 struct perf_event_attr attr
;
4801 struct perf_event_context
*ctx
;
4802 struct file
*event_file
= NULL
;
4803 struct file
*group_file
= NULL
;
4804 int fput_needed
= 0;
4805 int fput_needed2
= 0;
4808 /* for future expandability... */
4809 if (flags
& ~(PERF_FLAG_FD_NO_GROUP
| PERF_FLAG_FD_OUTPUT
))
4812 err
= perf_copy_attr(attr_uptr
, &attr
);
4816 if (!attr
.exclude_kernel
) {
4817 if (perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN
))
4822 if (attr
.sample_freq
> sysctl_perf_event_sample_rate
)
4827 * Get the target context (task or percpu):
4829 ctx
= find_get_context(pid
, cpu
);
4831 return PTR_ERR(ctx
);
4834 * Look up the group leader (we will attach this event to it):
4836 group_leader
= NULL
;
4837 if (group_fd
!= -1 && !(flags
& PERF_FLAG_FD_NO_GROUP
)) {
4839 group_file
= fget_light(group_fd
, &fput_needed
);
4841 goto err_put_context
;
4842 if (group_file
->f_op
!= &perf_fops
)
4843 goto err_put_context
;
4845 group_leader
= group_file
->private_data
;
4847 * Do not allow a recursive hierarchy (this new sibling
4848 * becoming part of another group-sibling):
4850 if (group_leader
->group_leader
!= group_leader
)
4851 goto err_put_context
;
4853 * Do not allow to attach to a group in a different
4854 * task or CPU context:
4856 if (group_leader
->ctx
!= ctx
)
4857 goto err_put_context
;
4859 * Only a group leader can be exclusive or pinned
4861 if (attr
.exclusive
|| attr
.pinned
)
4862 goto err_put_context
;
4865 event
= perf_event_alloc(&attr
, cpu
, ctx
, group_leader
,
4866 NULL
, NULL
, GFP_KERNEL
);
4867 err
= PTR_ERR(event
);
4869 goto err_put_context
;
4871 err
= anon_inode_getfd("[perf_event]", &perf_fops
, event
, O_RDWR
);
4873 goto err_free_put_context
;
4875 event_file
= fget_light(err
, &fput_needed2
);
4877 goto err_free_put_context
;
4879 if (flags
& PERF_FLAG_FD_OUTPUT
) {
4880 err
= perf_event_set_output(event
, group_fd
);
4882 goto err_fput_free_put_context
;
4885 event
->filp
= event_file
;
4886 WARN_ON_ONCE(ctx
->parent_ctx
);
4887 mutex_lock(&ctx
->mutex
);
4888 perf_install_in_context(ctx
, event
, cpu
);
4890 mutex_unlock(&ctx
->mutex
);
4892 event
->owner
= current
;
4893 get_task_struct(current
);
4894 mutex_lock(¤t
->perf_event_mutex
);
4895 list_add_tail(&event
->owner_entry
, ¤t
->perf_event_list
);
4896 mutex_unlock(¤t
->perf_event_mutex
);
4898 err_fput_free_put_context
:
4899 fput_light(event_file
, fput_needed2
);
4901 err_free_put_context
:
4909 fput_light(group_file
, fput_needed
);
4915 * perf_event_create_kernel_counter
4917 * @attr: attributes of the counter to create
4918 * @cpu: cpu in which the counter is bound
4919 * @pid: task to profile
4922 perf_event_create_kernel_counter(struct perf_event_attr
*attr
, int cpu
,
4924 perf_overflow_handler_t overflow_handler
)
4926 struct perf_event
*event
;
4927 struct perf_event_context
*ctx
;
4931 * Get the target context (task or percpu):
4934 ctx
= find_get_context(pid
, cpu
);
4940 event
= perf_event_alloc(attr
, cpu
, ctx
, NULL
,
4941 NULL
, overflow_handler
, GFP_KERNEL
);
4942 if (IS_ERR(event
)) {
4943 err
= PTR_ERR(event
);
4944 goto err_put_context
;
4948 WARN_ON_ONCE(ctx
->parent_ctx
);
4949 mutex_lock(&ctx
->mutex
);
4950 perf_install_in_context(ctx
, event
, cpu
);
4952 mutex_unlock(&ctx
->mutex
);
4954 event
->owner
= current
;
4955 get_task_struct(current
);
4956 mutex_lock(¤t
->perf_event_mutex
);
4957 list_add_tail(&event
->owner_entry
, ¤t
->perf_event_list
);
4958 mutex_unlock(¤t
->perf_event_mutex
);
4965 return ERR_PTR(err
);
4967 EXPORT_SYMBOL_GPL(perf_event_create_kernel_counter
);
4970 * inherit a event from parent task to child task:
4972 static struct perf_event
*
4973 inherit_event(struct perf_event
*parent_event
,
4974 struct task_struct
*parent
,
4975 struct perf_event_context
*parent_ctx
,
4976 struct task_struct
*child
,
4977 struct perf_event
*group_leader
,
4978 struct perf_event_context
*child_ctx
)
4980 struct perf_event
*child_event
;
4983 * Instead of creating recursive hierarchies of events,
4984 * we link inherited events back to the original parent,
4985 * which has a filp for sure, which we use as the reference
4988 if (parent_event
->parent
)
4989 parent_event
= parent_event
->parent
;
4991 child_event
= perf_event_alloc(&parent_event
->attr
,
4992 parent_event
->cpu
, child_ctx
,
4993 group_leader
, parent_event
,
4995 if (IS_ERR(child_event
))
5000 * Make the child state follow the state of the parent event,
5001 * not its attr.disabled bit. We hold the parent's mutex,
5002 * so we won't race with perf_event_{en, dis}able_family.
5004 if (parent_event
->state
>= PERF_EVENT_STATE_INACTIVE
)
5005 child_event
->state
= PERF_EVENT_STATE_INACTIVE
;
5007 child_event
->state
= PERF_EVENT_STATE_OFF
;
5009 if (parent_event
->attr
.freq
) {
5010 u64 sample_period
= parent_event
->hw
.sample_period
;
5011 struct hw_perf_event
*hwc
= &child_event
->hw
;
5013 hwc
->sample_period
= sample_period
;
5014 hwc
->last_period
= sample_period
;
5016 atomic64_set(&hwc
->period_left
, sample_period
);
5019 child_event
->overflow_handler
= parent_event
->overflow_handler
;
5022 * Link it up in the child's context:
5024 add_event_to_ctx(child_event
, child_ctx
);
5027 * Get a reference to the parent filp - we will fput it
5028 * when the child event exits. This is safe to do because
5029 * we are in the parent and we know that the filp still
5030 * exists and has a nonzero count:
5032 atomic_long_inc(&parent_event
->filp
->f_count
);
5035 * Link this into the parent event's child list
5037 WARN_ON_ONCE(parent_event
->ctx
->parent_ctx
);
5038 mutex_lock(&parent_event
->child_mutex
);
5039 list_add_tail(&child_event
->child_list
, &parent_event
->child_list
);
5040 mutex_unlock(&parent_event
->child_mutex
);
5045 static int inherit_group(struct perf_event
*parent_event
,
5046 struct task_struct
*parent
,
5047 struct perf_event_context
*parent_ctx
,
5048 struct task_struct
*child
,
5049 struct perf_event_context
*child_ctx
)
5051 struct perf_event
*leader
;
5052 struct perf_event
*sub
;
5053 struct perf_event
*child_ctr
;
5055 leader
= inherit_event(parent_event
, parent
, parent_ctx
,
5056 child
, NULL
, child_ctx
);
5058 return PTR_ERR(leader
);
5059 list_for_each_entry(sub
, &parent_event
->sibling_list
, group_entry
) {
5060 child_ctr
= inherit_event(sub
, parent
, parent_ctx
,
5061 child
, leader
, child_ctx
);
5062 if (IS_ERR(child_ctr
))
5063 return PTR_ERR(child_ctr
);
5068 static void sync_child_event(struct perf_event
*child_event
,
5069 struct task_struct
*child
)
5071 struct perf_event
*parent_event
= child_event
->parent
;
5074 if (child_event
->attr
.inherit_stat
)
5075 perf_event_read_event(child_event
, child
);
5077 child_val
= atomic64_read(&child_event
->count
);
5080 * Add back the child's count to the parent's count:
5082 atomic64_add(child_val
, &parent_event
->count
);
5083 atomic64_add(child_event
->total_time_enabled
,
5084 &parent_event
->child_total_time_enabled
);
5085 atomic64_add(child_event
->total_time_running
,
5086 &parent_event
->child_total_time_running
);
5089 * Remove this event from the parent's list
5091 WARN_ON_ONCE(parent_event
->ctx
->parent_ctx
);
5092 mutex_lock(&parent_event
->child_mutex
);
5093 list_del_init(&child_event
->child_list
);
5094 mutex_unlock(&parent_event
->child_mutex
);
5097 * Release the parent event, if this was the last
5100 fput(parent_event
->filp
);
5104 __perf_event_exit_task(struct perf_event
*child_event
,
5105 struct perf_event_context
*child_ctx
,
5106 struct task_struct
*child
)
5108 struct perf_event
*parent_event
;
5110 perf_event_remove_from_context(child_event
);
5112 parent_event
= child_event
->parent
;
5114 * It can happen that parent exits first, and has events
5115 * that are still around due to the child reference. These
5116 * events need to be zapped - but otherwise linger.
5119 sync_child_event(child_event
, child
);
5120 free_event(child_event
);
5125 * When a child task exits, feed back event values to parent events.
5127 void perf_event_exit_task(struct task_struct
*child
)
5129 struct perf_event
*child_event
, *tmp
;
5130 struct perf_event_context
*child_ctx
;
5131 unsigned long flags
;
5133 if (likely(!child
->perf_event_ctxp
)) {
5134 perf_event_task(child
, NULL
, 0);
5138 local_irq_save(flags
);
5140 * We can't reschedule here because interrupts are disabled,
5141 * and either child is current or it is a task that can't be
5142 * scheduled, so we are now safe from rescheduling changing
5145 child_ctx
= child
->perf_event_ctxp
;
5146 __perf_event_task_sched_out(child_ctx
);
5149 * Take the context lock here so that if find_get_context is
5150 * reading child->perf_event_ctxp, we wait until it has
5151 * incremented the context's refcount before we do put_ctx below.
5153 raw_spin_lock(&child_ctx
->lock
);
5154 child
->perf_event_ctxp
= NULL
;
5156 * If this context is a clone; unclone it so it can't get
5157 * swapped to another process while we're removing all
5158 * the events from it.
5160 unclone_ctx(child_ctx
);
5161 update_context_time(child_ctx
);
5162 raw_spin_unlock_irqrestore(&child_ctx
->lock
, flags
);
5165 * Report the task dead after unscheduling the events so that we
5166 * won't get any samples after PERF_RECORD_EXIT. We can however still
5167 * get a few PERF_RECORD_READ events.
5169 perf_event_task(child
, child_ctx
, 0);
5172 * We can recurse on the same lock type through:
5174 * __perf_event_exit_task()
5175 * sync_child_event()
5176 * fput(parent_event->filp)
5178 * mutex_lock(&ctx->mutex)
5180 * But since its the parent context it won't be the same instance.
5182 mutex_lock_nested(&child_ctx
->mutex
, SINGLE_DEPTH_NESTING
);
5185 list_for_each_entry_safe(child_event
, tmp
, &child_ctx
->pinned_groups
,
5187 __perf_event_exit_task(child_event
, child_ctx
, child
);
5189 list_for_each_entry_safe(child_event
, tmp
, &child_ctx
->flexible_groups
,
5191 __perf_event_exit_task(child_event
, child_ctx
, child
);
5194 * If the last event was a group event, it will have appended all
5195 * its siblings to the list, but we obtained 'tmp' before that which
5196 * will still point to the list head terminating the iteration.
5198 if (!list_empty(&child_ctx
->pinned_groups
) ||
5199 !list_empty(&child_ctx
->flexible_groups
))
5202 mutex_unlock(&child_ctx
->mutex
);
5207 static void perf_free_event(struct perf_event
*event
,
5208 struct perf_event_context
*ctx
)
5210 struct perf_event
*parent
= event
->parent
;
5212 if (WARN_ON_ONCE(!parent
))
5215 mutex_lock(&parent
->child_mutex
);
5216 list_del_init(&event
->child_list
);
5217 mutex_unlock(&parent
->child_mutex
);
5221 list_del_event(event
, ctx
);
5226 * free an unexposed, unused context as created by inheritance by
5227 * init_task below, used by fork() in case of fail.
5229 void perf_event_free_task(struct task_struct
*task
)
5231 struct perf_event_context
*ctx
= task
->perf_event_ctxp
;
5232 struct perf_event
*event
, *tmp
;
5237 mutex_lock(&ctx
->mutex
);
5239 list_for_each_entry_safe(event
, tmp
, &ctx
->pinned_groups
, group_entry
)
5240 perf_free_event(event
, ctx
);
5242 list_for_each_entry_safe(event
, tmp
, &ctx
->flexible_groups
,
5244 perf_free_event(event
, ctx
);
5246 if (!list_empty(&ctx
->pinned_groups
) ||
5247 !list_empty(&ctx
->flexible_groups
))
5250 mutex_unlock(&ctx
->mutex
);
5256 inherit_task_group(struct perf_event
*event
, struct task_struct
*parent
,
5257 struct perf_event_context
*parent_ctx
,
5258 struct task_struct
*child
,
5262 struct perf_event_context
*child_ctx
= child
->perf_event_ctxp
;
5264 if (!event
->attr
.inherit
) {
5271 * This is executed from the parent task context, so
5272 * inherit events that have been marked for cloning.
5273 * First allocate and initialize a context for the
5277 child_ctx
= kzalloc(sizeof(struct perf_event_context
),
5282 __perf_event_init_context(child_ctx
, child
);
5283 child
->perf_event_ctxp
= child_ctx
;
5284 get_task_struct(child
);
5287 ret
= inherit_group(event
, parent
, parent_ctx
,
5298 * Initialize the perf_event context in task_struct
5300 int perf_event_init_task(struct task_struct
*child
)
5302 struct perf_event_context
*child_ctx
, *parent_ctx
;
5303 struct perf_event_context
*cloned_ctx
;
5304 struct perf_event
*event
;
5305 struct task_struct
*parent
= current
;
5306 int inherited_all
= 1;
5309 child
->perf_event_ctxp
= NULL
;
5311 mutex_init(&child
->perf_event_mutex
);
5312 INIT_LIST_HEAD(&child
->perf_event_list
);
5314 if (likely(!parent
->perf_event_ctxp
))
5318 * If the parent's context is a clone, pin it so it won't get
5321 parent_ctx
= perf_pin_task_context(parent
);
5324 * No need to check if parent_ctx != NULL here; since we saw
5325 * it non-NULL earlier, the only reason for it to become NULL
5326 * is if we exit, and since we're currently in the middle of
5327 * a fork we can't be exiting at the same time.
5331 * Lock the parent list. No need to lock the child - not PID
5332 * hashed yet and not running, so nobody can access it.
5334 mutex_lock(&parent_ctx
->mutex
);
5337 * We dont have to disable NMIs - we are only looking at
5338 * the list, not manipulating it:
5340 list_for_each_entry(event
, &parent_ctx
->pinned_groups
, group_entry
) {
5341 ret
= inherit_task_group(event
, parent
, parent_ctx
, child
,
5347 list_for_each_entry(event
, &parent_ctx
->flexible_groups
, group_entry
) {
5348 ret
= inherit_task_group(event
, parent
, parent_ctx
, child
,
5354 child_ctx
= child
->perf_event_ctxp
;
5356 if (child_ctx
&& inherited_all
) {
5358 * Mark the child context as a clone of the parent
5359 * context, or of whatever the parent is a clone of.
5360 * Note that if the parent is a clone, it could get
5361 * uncloned at any point, but that doesn't matter
5362 * because the list of events and the generation
5363 * count can't have changed since we took the mutex.
5365 cloned_ctx
= rcu_dereference(parent_ctx
->parent_ctx
);
5367 child_ctx
->parent_ctx
= cloned_ctx
;
5368 child_ctx
->parent_gen
= parent_ctx
->parent_gen
;
5370 child_ctx
->parent_ctx
= parent_ctx
;
5371 child_ctx
->parent_gen
= parent_ctx
->generation
;
5373 get_ctx(child_ctx
->parent_ctx
);
5376 mutex_unlock(&parent_ctx
->mutex
);
5378 perf_unpin_context(parent_ctx
);
5383 static void __init
perf_event_init_all_cpus(void)
5386 struct perf_cpu_context
*cpuctx
;
5388 for_each_possible_cpu(cpu
) {
5389 cpuctx
= &per_cpu(perf_cpu_context
, cpu
);
5390 __perf_event_init_context(&cpuctx
->ctx
, NULL
);
5394 static void __cpuinit
perf_event_init_cpu(int cpu
)
5396 struct perf_cpu_context
*cpuctx
;
5398 cpuctx
= &per_cpu(perf_cpu_context
, cpu
);
5400 spin_lock(&perf_resource_lock
);
5401 cpuctx
->max_pertask
= perf_max_events
- perf_reserved_percpu
;
5402 spin_unlock(&perf_resource_lock
);
5405 #ifdef CONFIG_HOTPLUG_CPU
5406 static void __perf_event_exit_cpu(void *info
)
5408 struct perf_cpu_context
*cpuctx
= &__get_cpu_var(perf_cpu_context
);
5409 struct perf_event_context
*ctx
= &cpuctx
->ctx
;
5410 struct perf_event
*event
, *tmp
;
5412 list_for_each_entry_safe(event
, tmp
, &ctx
->pinned_groups
, group_entry
)
5413 __perf_event_remove_from_context(event
);
5414 list_for_each_entry_safe(event
, tmp
, &ctx
->flexible_groups
, group_entry
)
5415 __perf_event_remove_from_context(event
);
5417 static void perf_event_exit_cpu(int cpu
)
5419 struct perf_cpu_context
*cpuctx
= &per_cpu(perf_cpu_context
, cpu
);
5420 struct perf_event_context
*ctx
= &cpuctx
->ctx
;
5422 mutex_lock(&ctx
->mutex
);
5423 smp_call_function_single(cpu
, __perf_event_exit_cpu
, NULL
, 1);
5424 mutex_unlock(&ctx
->mutex
);
5427 static inline void perf_event_exit_cpu(int cpu
) { }
5430 static int __cpuinit
5431 perf_cpu_notify(struct notifier_block
*self
, unsigned long action
, void *hcpu
)
5433 unsigned int cpu
= (long)hcpu
;
5437 case CPU_UP_PREPARE
:
5438 case CPU_UP_PREPARE_FROZEN
:
5439 perf_event_init_cpu(cpu
);
5442 case CPU_DOWN_PREPARE
:
5443 case CPU_DOWN_PREPARE_FROZEN
:
5444 perf_event_exit_cpu(cpu
);
5455 * This has to have a higher priority than migration_notifier in sched.c.
5457 static struct notifier_block __cpuinitdata perf_cpu_nb
= {
5458 .notifier_call
= perf_cpu_notify
,
5462 void __init
perf_event_init(void)
5464 perf_event_init_all_cpus();
5465 perf_cpu_notify(&perf_cpu_nb
, (unsigned long)CPU_UP_PREPARE
,
5466 (void *)(long)smp_processor_id());
5467 perf_cpu_notify(&perf_cpu_nb
, (unsigned long)CPU_ONLINE
,
5468 (void *)(long)smp_processor_id());
5469 register_cpu_notifier(&perf_cpu_nb
);
5472 static ssize_t
perf_show_reserve_percpu(struct sysdev_class
*class,
5473 struct sysdev_class_attribute
*attr
,
5476 return sprintf(buf
, "%d\n", perf_reserved_percpu
);
5480 perf_set_reserve_percpu(struct sysdev_class
*class,
5481 struct sysdev_class_attribute
*attr
,
5485 struct perf_cpu_context
*cpuctx
;
5489 err
= strict_strtoul(buf
, 10, &val
);
5492 if (val
> perf_max_events
)
5495 spin_lock(&perf_resource_lock
);
5496 perf_reserved_percpu
= val
;
5497 for_each_online_cpu(cpu
) {
5498 cpuctx
= &per_cpu(perf_cpu_context
, cpu
);
5499 raw_spin_lock_irq(&cpuctx
->ctx
.lock
);
5500 mpt
= min(perf_max_events
- cpuctx
->ctx
.nr_events
,
5501 perf_max_events
- perf_reserved_percpu
);
5502 cpuctx
->max_pertask
= mpt
;
5503 raw_spin_unlock_irq(&cpuctx
->ctx
.lock
);
5505 spin_unlock(&perf_resource_lock
);
5510 static ssize_t
perf_show_overcommit(struct sysdev_class
*class,
5511 struct sysdev_class_attribute
*attr
,
5514 return sprintf(buf
, "%d\n", perf_overcommit
);
5518 perf_set_overcommit(struct sysdev_class
*class,
5519 struct sysdev_class_attribute
*attr
,
5520 const char *buf
, size_t count
)
5525 err
= strict_strtoul(buf
, 10, &val
);
5531 spin_lock(&perf_resource_lock
);
5532 perf_overcommit
= val
;
5533 spin_unlock(&perf_resource_lock
);
5538 static SYSDEV_CLASS_ATTR(
5541 perf_show_reserve_percpu
,
5542 perf_set_reserve_percpu
5545 static SYSDEV_CLASS_ATTR(
5548 perf_show_overcommit
,
5552 static struct attribute
*perfclass_attrs
[] = {
5553 &attr_reserve_percpu
.attr
,
5554 &attr_overcommit
.attr
,
5558 static struct attribute_group perfclass_attr_group
= {
5559 .attrs
= perfclass_attrs
,
5560 .name
= "perf_events",
5563 static int __init
perf_event_sysfs_init(void)
5565 return sysfs_create_group(&cpu_sysdev_class
.kset
.kobj
,
5566 &perfclass_attr_group
);
5568 device_initcall(perf_event_sysfs_init
);