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 static inline bool perf_paranoid_tracepoint_raw(void)
61 return sysctl_perf_event_paranoid
> -1;
64 static inline bool perf_paranoid_cpu(void)
66 return sysctl_perf_event_paranoid
> 0;
69 static inline bool perf_paranoid_kernel(void)
71 return sysctl_perf_event_paranoid
> 1;
74 int sysctl_perf_event_mlock __read_mostly
= 512; /* 'free' kb per user */
77 * max perf event sample rate
79 int sysctl_perf_event_sample_rate __read_mostly
= 100000;
81 static atomic64_t perf_event_id
;
84 * Lock for (sysadmin-configurable) event reservations:
86 static DEFINE_SPINLOCK(perf_resource_lock
);
89 * Architecture provided APIs - weak aliases:
91 extern __weak
const struct pmu
*hw_perf_event_init(struct perf_event
*event
)
96 void __weak
hw_perf_disable(void) { barrier(); }
97 void __weak
hw_perf_enable(void) { barrier(); }
99 void __weak
hw_perf_event_setup(int cpu
) { barrier(); }
100 void __weak
hw_perf_event_setup_online(int cpu
) { barrier(); }
101 void __weak
hw_perf_event_setup_offline(int cpu
) { barrier(); }
104 hw_perf_group_sched_in(struct perf_event
*group_leader
,
105 struct perf_cpu_context
*cpuctx
,
106 struct perf_event_context
*ctx
)
111 void __weak
perf_event_print_debug(void) { }
113 static DEFINE_PER_CPU(int, perf_disable_count
);
115 void __perf_disable(void)
117 __get_cpu_var(perf_disable_count
)++;
120 bool __perf_enable(void)
122 return !--__get_cpu_var(perf_disable_count
);
125 void perf_disable(void)
131 void perf_enable(void)
137 static void get_ctx(struct perf_event_context
*ctx
)
139 WARN_ON(!atomic_inc_not_zero(&ctx
->refcount
));
142 static void free_ctx(struct rcu_head
*head
)
144 struct perf_event_context
*ctx
;
146 ctx
= container_of(head
, struct perf_event_context
, rcu_head
);
150 static void put_ctx(struct perf_event_context
*ctx
)
152 if (atomic_dec_and_test(&ctx
->refcount
)) {
154 put_ctx(ctx
->parent_ctx
);
156 put_task_struct(ctx
->task
);
157 call_rcu(&ctx
->rcu_head
, free_ctx
);
161 static void unclone_ctx(struct perf_event_context
*ctx
)
163 if (ctx
->parent_ctx
) {
164 put_ctx(ctx
->parent_ctx
);
165 ctx
->parent_ctx
= NULL
;
170 * If we inherit events we want to return the parent event id
173 static u64
primary_event_id(struct perf_event
*event
)
178 id
= event
->parent
->id
;
184 * Get the perf_event_context for a task and lock it.
185 * This has to cope with with the fact that until it is locked,
186 * the context could get moved to another task.
188 static struct perf_event_context
*
189 perf_lock_task_context(struct task_struct
*task
, unsigned long *flags
)
191 struct perf_event_context
*ctx
;
195 ctx
= rcu_dereference(task
->perf_event_ctxp
);
198 * If this context is a clone of another, it might
199 * get swapped for another underneath us by
200 * perf_event_task_sched_out, though the
201 * rcu_read_lock() protects us from any context
202 * getting freed. Lock the context and check if it
203 * got swapped before we could get the lock, and retry
204 * if so. If we locked the right context, then it
205 * can't get swapped on us any more.
207 raw_spin_lock_irqsave(&ctx
->lock
, *flags
);
208 if (ctx
!= rcu_dereference(task
->perf_event_ctxp
)) {
209 raw_spin_unlock_irqrestore(&ctx
->lock
, *flags
);
213 if (!atomic_inc_not_zero(&ctx
->refcount
)) {
214 raw_spin_unlock_irqrestore(&ctx
->lock
, *flags
);
223 * Get the context for a task and increment its pin_count so it
224 * can't get swapped to another task. This also increments its
225 * reference count so that the context can't get freed.
227 static struct perf_event_context
*perf_pin_task_context(struct task_struct
*task
)
229 struct perf_event_context
*ctx
;
232 ctx
= perf_lock_task_context(task
, &flags
);
235 raw_spin_unlock_irqrestore(&ctx
->lock
, flags
);
240 static void perf_unpin_context(struct perf_event_context
*ctx
)
244 raw_spin_lock_irqsave(&ctx
->lock
, flags
);
246 raw_spin_unlock_irqrestore(&ctx
->lock
, flags
);
250 static inline u64
perf_clock(void)
252 return cpu_clock(smp_processor_id());
256 * Update the record of the current time in a context.
258 static void update_context_time(struct perf_event_context
*ctx
)
260 u64 now
= perf_clock();
262 ctx
->time
+= now
- ctx
->timestamp
;
263 ctx
->timestamp
= now
;
267 * Update the total_time_enabled and total_time_running fields for a event.
269 static void update_event_times(struct perf_event
*event
)
271 struct perf_event_context
*ctx
= event
->ctx
;
274 if (event
->state
< PERF_EVENT_STATE_INACTIVE
||
275 event
->group_leader
->state
< PERF_EVENT_STATE_INACTIVE
)
281 run_end
= event
->tstamp_stopped
;
283 event
->total_time_enabled
= run_end
- event
->tstamp_enabled
;
285 if (event
->state
== PERF_EVENT_STATE_INACTIVE
)
286 run_end
= event
->tstamp_stopped
;
290 event
->total_time_running
= run_end
- event
->tstamp_running
;
293 static struct list_head
*
294 ctx_group_list(struct perf_event
*event
, struct perf_event_context
*ctx
)
296 if (event
->attr
.pinned
)
297 return &ctx
->pinned_groups
;
299 return &ctx
->flexible_groups
;
303 * Add a event from the lists for its context.
304 * Must be called with ctx->mutex and ctx->lock held.
307 list_add_event(struct perf_event
*event
, struct perf_event_context
*ctx
)
309 struct perf_event
*group_leader
= event
->group_leader
;
312 * Depending on whether it is a standalone or sibling event,
313 * add it straight to the context's event list, or to the group
314 * leader's sibling list:
316 if (group_leader
== event
) {
317 struct list_head
*list
;
319 if (is_software_event(event
))
320 event
->group_flags
|= PERF_GROUP_SOFTWARE
;
322 list
= ctx_group_list(event
, ctx
);
323 list_add_tail(&event
->group_entry
, list
);
325 if (group_leader
->group_flags
& PERF_GROUP_SOFTWARE
&&
326 !is_software_event(event
))
327 group_leader
->group_flags
&= ~PERF_GROUP_SOFTWARE
;
329 list_add_tail(&event
->group_entry
, &group_leader
->sibling_list
);
330 group_leader
->nr_siblings
++;
333 list_add_rcu(&event
->event_entry
, &ctx
->event_list
);
335 if (event
->attr
.inherit_stat
)
340 * Remove a event from the lists for its context.
341 * Must be called with ctx->mutex and ctx->lock held.
344 list_del_event(struct perf_event
*event
, struct perf_event_context
*ctx
)
346 struct perf_event
*sibling
, *tmp
;
348 if (list_empty(&event
->group_entry
))
351 if (event
->attr
.inherit_stat
)
354 list_del_init(&event
->group_entry
);
355 list_del_rcu(&event
->event_entry
);
357 if (event
->group_leader
!= event
)
358 event
->group_leader
->nr_siblings
--;
360 update_event_times(event
);
363 * If event was in error state, then keep it
364 * that way, otherwise bogus counts will be
365 * returned on read(). The only way to get out
366 * of error state is by explicit re-enabling
369 if (event
->state
> PERF_EVENT_STATE_OFF
)
370 event
->state
= PERF_EVENT_STATE_OFF
;
373 * If this was a group event with sibling events then
374 * upgrade the siblings to singleton events by adding them
375 * to the context list directly:
377 list_for_each_entry_safe(sibling
, tmp
, &event
->sibling_list
, group_entry
) {
378 struct list_head
*list
;
380 list
= ctx_group_list(event
, ctx
);
381 list_move_tail(&sibling
->group_entry
, list
);
382 sibling
->group_leader
= sibling
;
384 /* Inherit group flags from the previous leader */
385 sibling
->group_flags
= event
->group_flags
;
390 event_sched_out(struct perf_event
*event
,
391 struct perf_cpu_context
*cpuctx
,
392 struct perf_event_context
*ctx
)
394 if (event
->state
!= PERF_EVENT_STATE_ACTIVE
)
397 event
->state
= PERF_EVENT_STATE_INACTIVE
;
398 if (event
->pending_disable
) {
399 event
->pending_disable
= 0;
400 event
->state
= PERF_EVENT_STATE_OFF
;
402 event
->tstamp_stopped
= ctx
->time
;
403 event
->pmu
->disable(event
);
406 if (!is_software_event(event
))
407 cpuctx
->active_oncpu
--;
409 if (event
->attr
.exclusive
|| !cpuctx
->active_oncpu
)
410 cpuctx
->exclusive
= 0;
414 group_sched_out(struct perf_event
*group_event
,
415 struct perf_cpu_context
*cpuctx
,
416 struct perf_event_context
*ctx
)
418 struct perf_event
*event
;
420 if (group_event
->state
!= PERF_EVENT_STATE_ACTIVE
)
423 event_sched_out(group_event
, cpuctx
, ctx
);
426 * Schedule out siblings (if any):
428 list_for_each_entry(event
, &group_event
->sibling_list
, group_entry
)
429 event_sched_out(event
, cpuctx
, ctx
);
431 if (group_event
->attr
.exclusive
)
432 cpuctx
->exclusive
= 0;
436 * Cross CPU call to remove a performance event
438 * We disable the event on the hardware level first. After that we
439 * remove it from the context list.
441 static void __perf_event_remove_from_context(void *info
)
443 struct perf_cpu_context
*cpuctx
= &__get_cpu_var(perf_cpu_context
);
444 struct perf_event
*event
= info
;
445 struct perf_event_context
*ctx
= event
->ctx
;
448 * If this is a task context, we need to check whether it is
449 * the current task context of this cpu. If not it has been
450 * scheduled out before the smp call arrived.
452 if (ctx
->task
&& cpuctx
->task_ctx
!= ctx
)
455 raw_spin_lock(&ctx
->lock
);
457 * Protect the list operation against NMI by disabling the
458 * events on a global level.
462 event_sched_out(event
, cpuctx
, ctx
);
464 list_del_event(event
, ctx
);
468 * Allow more per task events with respect to the
471 cpuctx
->max_pertask
=
472 min(perf_max_events
- ctx
->nr_events
,
473 perf_max_events
- perf_reserved_percpu
);
477 raw_spin_unlock(&ctx
->lock
);
482 * Remove the event from a task's (or a CPU's) list of events.
484 * Must be called with ctx->mutex held.
486 * CPU events are removed with a smp call. For task events we only
487 * call when the task is on a CPU.
489 * If event->ctx is a cloned context, callers must make sure that
490 * every task struct that event->ctx->task could possibly point to
491 * remains valid. This is OK when called from perf_release since
492 * that only calls us on the top-level context, which can't be a clone.
493 * When called from perf_event_exit_task, it's OK because the
494 * context has been detached from its task.
496 static void perf_event_remove_from_context(struct perf_event
*event
)
498 struct perf_event_context
*ctx
= event
->ctx
;
499 struct task_struct
*task
= ctx
->task
;
503 * Per cpu events are removed via an smp call and
504 * the removal is always successful.
506 smp_call_function_single(event
->cpu
,
507 __perf_event_remove_from_context
,
513 task_oncpu_function_call(task
, __perf_event_remove_from_context
,
516 raw_spin_lock_irq(&ctx
->lock
);
518 * If the context is active we need to retry the smp call.
520 if (ctx
->nr_active
&& !list_empty(&event
->group_entry
)) {
521 raw_spin_unlock_irq(&ctx
->lock
);
526 * The lock prevents that this context is scheduled in so we
527 * can remove the event safely, if the call above did not
530 if (!list_empty(&event
->group_entry
))
531 list_del_event(event
, ctx
);
532 raw_spin_unlock_irq(&ctx
->lock
);
536 * Update total_time_enabled and total_time_running for all events in a group.
538 static void update_group_times(struct perf_event
*leader
)
540 struct perf_event
*event
;
542 update_event_times(leader
);
543 list_for_each_entry(event
, &leader
->sibling_list
, group_entry
)
544 update_event_times(event
);
548 * Cross CPU call to disable a performance event
550 static void __perf_event_disable(void *info
)
552 struct perf_event
*event
= info
;
553 struct perf_cpu_context
*cpuctx
= &__get_cpu_var(perf_cpu_context
);
554 struct perf_event_context
*ctx
= event
->ctx
;
557 * If this is a per-task event, need to check whether this
558 * event's task is the current task on this cpu.
560 if (ctx
->task
&& cpuctx
->task_ctx
!= ctx
)
563 raw_spin_lock(&ctx
->lock
);
566 * If the event is on, turn it off.
567 * If it is in error state, leave it in error state.
569 if (event
->state
>= PERF_EVENT_STATE_INACTIVE
) {
570 update_context_time(ctx
);
571 update_group_times(event
);
572 if (event
== event
->group_leader
)
573 group_sched_out(event
, cpuctx
, ctx
);
575 event_sched_out(event
, cpuctx
, ctx
);
576 event
->state
= PERF_EVENT_STATE_OFF
;
579 raw_spin_unlock(&ctx
->lock
);
585 * If event->ctx is a cloned context, callers must make sure that
586 * every task struct that event->ctx->task could possibly point to
587 * remains valid. This condition is satisifed when called through
588 * perf_event_for_each_child or perf_event_for_each because they
589 * hold the top-level event's child_mutex, so any descendant that
590 * goes to exit will block in sync_child_event.
591 * When called from perf_pending_event it's OK because event->ctx
592 * is the current context on this CPU and preemption is disabled,
593 * hence we can't get into perf_event_task_sched_out for this context.
595 void perf_event_disable(struct perf_event
*event
)
597 struct perf_event_context
*ctx
= event
->ctx
;
598 struct task_struct
*task
= ctx
->task
;
602 * Disable the event on the cpu that it's on
604 smp_call_function_single(event
->cpu
, __perf_event_disable
,
610 task_oncpu_function_call(task
, __perf_event_disable
, event
);
612 raw_spin_lock_irq(&ctx
->lock
);
614 * If the event is still active, we need to retry the cross-call.
616 if (event
->state
== PERF_EVENT_STATE_ACTIVE
) {
617 raw_spin_unlock_irq(&ctx
->lock
);
622 * Since we have the lock this context can't be scheduled
623 * in, so we can change the state safely.
625 if (event
->state
== PERF_EVENT_STATE_INACTIVE
) {
626 update_group_times(event
);
627 event
->state
= PERF_EVENT_STATE_OFF
;
630 raw_spin_unlock_irq(&ctx
->lock
);
634 event_sched_in(struct perf_event
*event
,
635 struct perf_cpu_context
*cpuctx
,
636 struct perf_event_context
*ctx
)
638 if (event
->state
<= PERF_EVENT_STATE_OFF
)
641 event
->state
= PERF_EVENT_STATE_ACTIVE
;
642 event
->oncpu
= smp_processor_id();
644 * The new state must be visible before we turn it on in the hardware:
648 if (event
->pmu
->enable(event
)) {
649 event
->state
= PERF_EVENT_STATE_INACTIVE
;
654 event
->tstamp_running
+= ctx
->time
- event
->tstamp_stopped
;
656 if (!is_software_event(event
))
657 cpuctx
->active_oncpu
++;
660 if (event
->attr
.exclusive
)
661 cpuctx
->exclusive
= 1;
667 group_sched_in(struct perf_event
*group_event
,
668 struct perf_cpu_context
*cpuctx
,
669 struct perf_event_context
*ctx
)
671 struct perf_event
*event
, *partial_group
;
674 if (group_event
->state
== PERF_EVENT_STATE_OFF
)
677 ret
= hw_perf_group_sched_in(group_event
, cpuctx
, ctx
);
679 return ret
< 0 ? ret
: 0;
681 if (event_sched_in(group_event
, cpuctx
, ctx
))
685 * Schedule in siblings as one group (if any):
687 list_for_each_entry(event
, &group_event
->sibling_list
, group_entry
) {
688 if (event_sched_in(event
, cpuctx
, ctx
)) {
689 partial_group
= event
;
698 * Groups can be scheduled in as one unit only, so undo any
699 * partial group before returning:
701 list_for_each_entry(event
, &group_event
->sibling_list
, group_entry
) {
702 if (event
== partial_group
)
704 event_sched_out(event
, cpuctx
, ctx
);
706 event_sched_out(group_event
, cpuctx
, ctx
);
712 * Work out whether we can put this event group on the CPU now.
714 static int group_can_go_on(struct perf_event
*event
,
715 struct perf_cpu_context
*cpuctx
,
719 * Groups consisting entirely of software events can always go on.
721 if (event
->group_flags
& PERF_GROUP_SOFTWARE
)
724 * If an exclusive group is already on, no other hardware
727 if (cpuctx
->exclusive
)
730 * If this group is exclusive and there are already
731 * events on the CPU, it can't go on.
733 if (event
->attr
.exclusive
&& cpuctx
->active_oncpu
)
736 * Otherwise, try to add it if all previous groups were able
742 static void add_event_to_ctx(struct perf_event
*event
,
743 struct perf_event_context
*ctx
)
745 list_add_event(event
, ctx
);
746 event
->tstamp_enabled
= ctx
->time
;
747 event
->tstamp_running
= ctx
->time
;
748 event
->tstamp_stopped
= ctx
->time
;
752 * Cross CPU call to install and enable a performance event
754 * Must be called with ctx->mutex held
756 static void __perf_install_in_context(void *info
)
758 struct perf_cpu_context
*cpuctx
= &__get_cpu_var(perf_cpu_context
);
759 struct perf_event
*event
= info
;
760 struct perf_event_context
*ctx
= event
->ctx
;
761 struct perf_event
*leader
= event
->group_leader
;
765 * If this is a task context, we need to check whether it is
766 * the current task context of this cpu. If not it has been
767 * scheduled out before the smp call arrived.
768 * Or possibly this is the right context but it isn't
769 * on this cpu because it had no events.
771 if (ctx
->task
&& cpuctx
->task_ctx
!= ctx
) {
772 if (cpuctx
->task_ctx
|| ctx
->task
!= current
)
774 cpuctx
->task_ctx
= ctx
;
777 raw_spin_lock(&ctx
->lock
);
779 update_context_time(ctx
);
782 * Protect the list operation against NMI by disabling the
783 * events on a global level. NOP for non NMI based events.
787 add_event_to_ctx(event
, ctx
);
789 if (event
->cpu
!= -1 && event
->cpu
!= smp_processor_id())
793 * Don't put the event on if it is disabled or if
794 * it is in a group and the group isn't on.
796 if (event
->state
!= PERF_EVENT_STATE_INACTIVE
||
797 (leader
!= event
&& leader
->state
!= PERF_EVENT_STATE_ACTIVE
))
801 * An exclusive event can't go on if there are already active
802 * hardware events, and no hardware event can go on if there
803 * is already an exclusive event on.
805 if (!group_can_go_on(event
, cpuctx
, 1))
808 err
= event_sched_in(event
, cpuctx
, ctx
);
812 * This event couldn't go on. If it is in a group
813 * then we have to pull the whole group off.
814 * If the event group is pinned then put it in error state.
817 group_sched_out(leader
, cpuctx
, ctx
);
818 if (leader
->attr
.pinned
) {
819 update_group_times(leader
);
820 leader
->state
= PERF_EVENT_STATE_ERROR
;
824 if (!err
&& !ctx
->task
&& cpuctx
->max_pertask
)
825 cpuctx
->max_pertask
--;
830 raw_spin_unlock(&ctx
->lock
);
834 * Attach a performance event to a context
836 * First we add the event to the list with the hardware enable bit
837 * in event->hw_config cleared.
839 * If the event is attached to a task which is on a CPU we use a smp
840 * call to enable it in the task context. The task might have been
841 * scheduled away, but we check this in the smp call again.
843 * Must be called with ctx->mutex held.
846 perf_install_in_context(struct perf_event_context
*ctx
,
847 struct perf_event
*event
,
850 struct task_struct
*task
= ctx
->task
;
854 * Per cpu events are installed via an smp call and
855 * the install is always successful.
857 smp_call_function_single(cpu
, __perf_install_in_context
,
863 task_oncpu_function_call(task
, __perf_install_in_context
,
866 raw_spin_lock_irq(&ctx
->lock
);
868 * we need to retry the smp call.
870 if (ctx
->is_active
&& list_empty(&event
->group_entry
)) {
871 raw_spin_unlock_irq(&ctx
->lock
);
876 * The lock prevents that this context is scheduled in so we
877 * can add the event safely, if it the call above did not
880 if (list_empty(&event
->group_entry
))
881 add_event_to_ctx(event
, ctx
);
882 raw_spin_unlock_irq(&ctx
->lock
);
886 * Put a event into inactive state and update time fields.
887 * Enabling the leader of a group effectively enables all
888 * the group members that aren't explicitly disabled, so we
889 * have to update their ->tstamp_enabled also.
890 * Note: this works for group members as well as group leaders
891 * since the non-leader members' sibling_lists will be empty.
893 static void __perf_event_mark_enabled(struct perf_event
*event
,
894 struct perf_event_context
*ctx
)
896 struct perf_event
*sub
;
898 event
->state
= PERF_EVENT_STATE_INACTIVE
;
899 event
->tstamp_enabled
= ctx
->time
- event
->total_time_enabled
;
900 list_for_each_entry(sub
, &event
->sibling_list
, group_entry
)
901 if (sub
->state
>= PERF_EVENT_STATE_INACTIVE
)
902 sub
->tstamp_enabled
=
903 ctx
->time
- sub
->total_time_enabled
;
907 * Cross CPU call to enable a performance event
909 static void __perf_event_enable(void *info
)
911 struct perf_event
*event
= info
;
912 struct perf_cpu_context
*cpuctx
= &__get_cpu_var(perf_cpu_context
);
913 struct perf_event_context
*ctx
= event
->ctx
;
914 struct perf_event
*leader
= event
->group_leader
;
918 * If this is a per-task event, need to check whether this
919 * event's task is the current task on this cpu.
921 if (ctx
->task
&& cpuctx
->task_ctx
!= ctx
) {
922 if (cpuctx
->task_ctx
|| ctx
->task
!= current
)
924 cpuctx
->task_ctx
= ctx
;
927 raw_spin_lock(&ctx
->lock
);
929 update_context_time(ctx
);
931 if (event
->state
>= PERF_EVENT_STATE_INACTIVE
)
933 __perf_event_mark_enabled(event
, ctx
);
935 if (event
->cpu
!= -1 && event
->cpu
!= smp_processor_id())
939 * If the event is in a group and isn't the group leader,
940 * then don't put it on unless the group is on.
942 if (leader
!= event
&& leader
->state
!= PERF_EVENT_STATE_ACTIVE
)
945 if (!group_can_go_on(event
, cpuctx
, 1)) {
950 err
= group_sched_in(event
, cpuctx
, ctx
);
952 err
= event_sched_in(event
, cpuctx
, ctx
);
958 * If this event can't go on and it's part of a
959 * group, then the whole group has to come off.
962 group_sched_out(leader
, cpuctx
, ctx
);
963 if (leader
->attr
.pinned
) {
964 update_group_times(leader
);
965 leader
->state
= PERF_EVENT_STATE_ERROR
;
970 raw_spin_unlock(&ctx
->lock
);
976 * If event->ctx is a cloned context, callers must make sure that
977 * every task struct that event->ctx->task could possibly point to
978 * remains valid. This condition is satisfied when called through
979 * perf_event_for_each_child or perf_event_for_each as described
980 * for perf_event_disable.
982 void perf_event_enable(struct perf_event
*event
)
984 struct perf_event_context
*ctx
= event
->ctx
;
985 struct task_struct
*task
= ctx
->task
;
989 * Enable the event on the cpu that it's on
991 smp_call_function_single(event
->cpu
, __perf_event_enable
,
996 raw_spin_lock_irq(&ctx
->lock
);
997 if (event
->state
>= PERF_EVENT_STATE_INACTIVE
)
1001 * If the event is in error state, clear that first.
1002 * That way, if we see the event in error state below, we
1003 * know that it has gone back into error state, as distinct
1004 * from the task having been scheduled away before the
1005 * cross-call arrived.
1007 if (event
->state
== PERF_EVENT_STATE_ERROR
)
1008 event
->state
= PERF_EVENT_STATE_OFF
;
1011 raw_spin_unlock_irq(&ctx
->lock
);
1012 task_oncpu_function_call(task
, __perf_event_enable
, event
);
1014 raw_spin_lock_irq(&ctx
->lock
);
1017 * If the context is active and the event is still off,
1018 * we need to retry the cross-call.
1020 if (ctx
->is_active
&& event
->state
== PERF_EVENT_STATE_OFF
)
1024 * Since we have the lock this context can't be scheduled
1025 * in, so we can change the state safely.
1027 if (event
->state
== PERF_EVENT_STATE_OFF
)
1028 __perf_event_mark_enabled(event
, ctx
);
1031 raw_spin_unlock_irq(&ctx
->lock
);
1034 static int perf_event_refresh(struct perf_event
*event
, int refresh
)
1037 * not supported on inherited events
1039 if (event
->attr
.inherit
)
1042 atomic_add(refresh
, &event
->event_limit
);
1043 perf_event_enable(event
);
1049 EVENT_FLEXIBLE
= 0x1,
1051 EVENT_ALL
= EVENT_FLEXIBLE
| EVENT_PINNED
,
1054 static void ctx_sched_out(struct perf_event_context
*ctx
,
1055 struct perf_cpu_context
*cpuctx
,
1056 enum event_type_t event_type
)
1058 struct perf_event
*event
;
1060 raw_spin_lock(&ctx
->lock
);
1062 if (likely(!ctx
->nr_events
))
1064 update_context_time(ctx
);
1067 if (!ctx
->nr_active
)
1070 if (event_type
& EVENT_PINNED
)
1071 list_for_each_entry(event
, &ctx
->pinned_groups
, group_entry
)
1072 group_sched_out(event
, cpuctx
, ctx
);
1074 if (event_type
& EVENT_FLEXIBLE
)
1075 list_for_each_entry(event
, &ctx
->flexible_groups
, group_entry
)
1076 group_sched_out(event
, cpuctx
, ctx
);
1081 raw_spin_unlock(&ctx
->lock
);
1085 * Test whether two contexts are equivalent, i.e. whether they
1086 * have both been cloned from the same version of the same context
1087 * and they both have the same number of enabled events.
1088 * If the number of enabled events is the same, then the set
1089 * of enabled events should be the same, because these are both
1090 * inherited contexts, therefore we can't access individual events
1091 * in them directly with an fd; we can only enable/disable all
1092 * events via prctl, or enable/disable all events in a family
1093 * via ioctl, which will have the same effect on both contexts.
1095 static int context_equiv(struct perf_event_context
*ctx1
,
1096 struct perf_event_context
*ctx2
)
1098 return ctx1
->parent_ctx
&& ctx1
->parent_ctx
== ctx2
->parent_ctx
1099 && ctx1
->parent_gen
== ctx2
->parent_gen
1100 && !ctx1
->pin_count
&& !ctx2
->pin_count
;
1103 static void __perf_event_sync_stat(struct perf_event
*event
,
1104 struct perf_event
*next_event
)
1108 if (!event
->attr
.inherit_stat
)
1112 * Update the event value, we cannot use perf_event_read()
1113 * because we're in the middle of a context switch and have IRQs
1114 * disabled, which upsets smp_call_function_single(), however
1115 * we know the event must be on the current CPU, therefore we
1116 * don't need to use it.
1118 switch (event
->state
) {
1119 case PERF_EVENT_STATE_ACTIVE
:
1120 event
->pmu
->read(event
);
1123 case PERF_EVENT_STATE_INACTIVE
:
1124 update_event_times(event
);
1132 * In order to keep per-task stats reliable we need to flip the event
1133 * values when we flip the contexts.
1135 value
= atomic64_read(&next_event
->count
);
1136 value
= atomic64_xchg(&event
->count
, value
);
1137 atomic64_set(&next_event
->count
, value
);
1139 swap(event
->total_time_enabled
, next_event
->total_time_enabled
);
1140 swap(event
->total_time_running
, next_event
->total_time_running
);
1143 * Since we swizzled the values, update the user visible data too.
1145 perf_event_update_userpage(event
);
1146 perf_event_update_userpage(next_event
);
1149 #define list_next_entry(pos, member) \
1150 list_entry(pos->member.next, typeof(*pos), member)
1152 static void perf_event_sync_stat(struct perf_event_context
*ctx
,
1153 struct perf_event_context
*next_ctx
)
1155 struct perf_event
*event
, *next_event
;
1160 update_context_time(ctx
);
1162 event
= list_first_entry(&ctx
->event_list
,
1163 struct perf_event
, event_entry
);
1165 next_event
= list_first_entry(&next_ctx
->event_list
,
1166 struct perf_event
, event_entry
);
1168 while (&event
->event_entry
!= &ctx
->event_list
&&
1169 &next_event
->event_entry
!= &next_ctx
->event_list
) {
1171 __perf_event_sync_stat(event
, next_event
);
1173 event
= list_next_entry(event
, event_entry
);
1174 next_event
= list_next_entry(next_event
, event_entry
);
1179 * Called from scheduler to remove the events of the current task,
1180 * with interrupts disabled.
1182 * We stop each event and update the event value in event->count.
1184 * This does not protect us against NMI, but disable()
1185 * sets the disabled bit in the control field of event _before_
1186 * accessing the event control register. If a NMI hits, then it will
1187 * not restart the event.
1189 void perf_event_task_sched_out(struct task_struct
*task
,
1190 struct task_struct
*next
)
1192 struct perf_cpu_context
*cpuctx
= &__get_cpu_var(perf_cpu_context
);
1193 struct perf_event_context
*ctx
= task
->perf_event_ctxp
;
1194 struct perf_event_context
*next_ctx
;
1195 struct perf_event_context
*parent
;
1196 struct pt_regs
*regs
;
1199 regs
= task_pt_regs(task
);
1200 perf_sw_event(PERF_COUNT_SW_CONTEXT_SWITCHES
, 1, 1, regs
, 0);
1202 if (likely(!ctx
|| !cpuctx
->task_ctx
))
1206 parent
= rcu_dereference(ctx
->parent_ctx
);
1207 next_ctx
= next
->perf_event_ctxp
;
1208 if (parent
&& next_ctx
&&
1209 rcu_dereference(next_ctx
->parent_ctx
) == parent
) {
1211 * Looks like the two contexts are clones, so we might be
1212 * able to optimize the context switch. We lock both
1213 * contexts and check that they are clones under the
1214 * lock (including re-checking that neither has been
1215 * uncloned in the meantime). It doesn't matter which
1216 * order we take the locks because no other cpu could
1217 * be trying to lock both of these tasks.
1219 raw_spin_lock(&ctx
->lock
);
1220 raw_spin_lock_nested(&next_ctx
->lock
, SINGLE_DEPTH_NESTING
);
1221 if (context_equiv(ctx
, next_ctx
)) {
1223 * XXX do we need a memory barrier of sorts
1224 * wrt to rcu_dereference() of perf_event_ctxp
1226 task
->perf_event_ctxp
= next_ctx
;
1227 next
->perf_event_ctxp
= ctx
;
1229 next_ctx
->task
= task
;
1232 perf_event_sync_stat(ctx
, next_ctx
);
1234 raw_spin_unlock(&next_ctx
->lock
);
1235 raw_spin_unlock(&ctx
->lock
);
1240 ctx_sched_out(ctx
, cpuctx
, EVENT_ALL
);
1241 cpuctx
->task_ctx
= NULL
;
1245 static void task_ctx_sched_out(struct perf_event_context
*ctx
,
1246 enum event_type_t event_type
)
1248 struct perf_cpu_context
*cpuctx
= &__get_cpu_var(perf_cpu_context
);
1250 if (!cpuctx
->task_ctx
)
1253 if (WARN_ON_ONCE(ctx
!= cpuctx
->task_ctx
))
1256 ctx_sched_out(ctx
, cpuctx
, event_type
);
1257 cpuctx
->task_ctx
= NULL
;
1261 * Called with IRQs disabled
1263 static void __perf_event_task_sched_out(struct perf_event_context
*ctx
)
1265 task_ctx_sched_out(ctx
, EVENT_ALL
);
1269 * Called with IRQs disabled
1271 static void cpu_ctx_sched_out(struct perf_cpu_context
*cpuctx
,
1272 enum event_type_t event_type
)
1274 ctx_sched_out(&cpuctx
->ctx
, cpuctx
, event_type
);
1278 ctx_pinned_sched_in(struct perf_event_context
*ctx
,
1279 struct perf_cpu_context
*cpuctx
)
1281 struct perf_event
*event
;
1283 list_for_each_entry(event
, &ctx
->pinned_groups
, group_entry
) {
1284 if (event
->state
<= PERF_EVENT_STATE_OFF
)
1286 if (event
->cpu
!= -1 && event
->cpu
!= smp_processor_id())
1289 if (group_can_go_on(event
, cpuctx
, 1))
1290 group_sched_in(event
, cpuctx
, ctx
);
1293 * If this pinned group hasn't been scheduled,
1294 * put it in error state.
1296 if (event
->state
== PERF_EVENT_STATE_INACTIVE
) {
1297 update_group_times(event
);
1298 event
->state
= PERF_EVENT_STATE_ERROR
;
1304 ctx_flexible_sched_in(struct perf_event_context
*ctx
,
1305 struct perf_cpu_context
*cpuctx
)
1307 struct perf_event
*event
;
1310 list_for_each_entry(event
, &ctx
->flexible_groups
, group_entry
) {
1311 /* Ignore events in OFF or ERROR state */
1312 if (event
->state
<= PERF_EVENT_STATE_OFF
)
1315 * Listen to the 'cpu' scheduling filter constraint
1318 if (event
->cpu
!= -1 && event
->cpu
!= smp_processor_id())
1321 if (group_can_go_on(event
, cpuctx
, can_add_hw
))
1322 if (group_sched_in(event
, cpuctx
, ctx
))
1328 ctx_sched_in(struct perf_event_context
*ctx
,
1329 struct perf_cpu_context
*cpuctx
,
1330 enum event_type_t event_type
)
1332 raw_spin_lock(&ctx
->lock
);
1334 if (likely(!ctx
->nr_events
))
1337 ctx
->timestamp
= perf_clock();
1342 * First go through the list and put on any pinned groups
1343 * in order to give them the best chance of going on.
1345 if (event_type
& EVENT_PINNED
)
1346 ctx_pinned_sched_in(ctx
, cpuctx
);
1348 /* Then walk through the lower prio flexible groups */
1349 if (event_type
& EVENT_FLEXIBLE
)
1350 ctx_flexible_sched_in(ctx
, cpuctx
);
1354 raw_spin_unlock(&ctx
->lock
);
1357 static void cpu_ctx_sched_in(struct perf_cpu_context
*cpuctx
,
1358 enum event_type_t event_type
)
1360 struct perf_event_context
*ctx
= &cpuctx
->ctx
;
1362 ctx_sched_in(ctx
, cpuctx
, event_type
);
1365 static void task_ctx_sched_in(struct task_struct
*task
,
1366 enum event_type_t event_type
)
1368 struct perf_cpu_context
*cpuctx
= &__get_cpu_var(perf_cpu_context
);
1369 struct perf_event_context
*ctx
= task
->perf_event_ctxp
;
1373 if (cpuctx
->task_ctx
== ctx
)
1375 ctx_sched_in(ctx
, cpuctx
, event_type
);
1376 cpuctx
->task_ctx
= ctx
;
1379 * Called from scheduler to add the events of the current task
1380 * with interrupts disabled.
1382 * We restore the event value and then enable it.
1384 * This does not protect us against NMI, but enable()
1385 * sets the enabled bit in the control field of event _before_
1386 * accessing the event control register. If a NMI hits, then it will
1387 * keep the event running.
1389 void perf_event_task_sched_in(struct task_struct
*task
)
1391 struct perf_cpu_context
*cpuctx
= &__get_cpu_var(perf_cpu_context
);
1392 struct perf_event_context
*ctx
= task
->perf_event_ctxp
;
1397 if (cpuctx
->task_ctx
== ctx
)
1401 * We want to keep the following priority order:
1402 * cpu pinned (that don't need to move), task pinned,
1403 * cpu flexible, task flexible.
1405 cpu_ctx_sched_out(cpuctx
, EVENT_FLEXIBLE
);
1407 ctx_sched_in(ctx
, cpuctx
, EVENT_PINNED
);
1408 cpu_ctx_sched_in(cpuctx
, EVENT_FLEXIBLE
);
1409 ctx_sched_in(ctx
, cpuctx
, EVENT_FLEXIBLE
);
1411 cpuctx
->task_ctx
= ctx
;
1414 #define MAX_INTERRUPTS (~0ULL)
1416 static void perf_log_throttle(struct perf_event
*event
, int enable
);
1418 static u64
perf_calculate_period(struct perf_event
*event
, u64 nsec
, u64 count
)
1420 u64 frequency
= event
->attr
.sample_freq
;
1421 u64 sec
= NSEC_PER_SEC
;
1422 u64 divisor
, dividend
;
1424 int count_fls
, nsec_fls
, frequency_fls
, sec_fls
;
1426 count_fls
= fls64(count
);
1427 nsec_fls
= fls64(nsec
);
1428 frequency_fls
= fls64(frequency
);
1432 * We got @count in @nsec, with a target of sample_freq HZ
1433 * the target period becomes:
1436 * period = -------------------
1437 * @nsec * sample_freq
1442 * Reduce accuracy by one bit such that @a and @b converge
1443 * to a similar magnitude.
1445 #define REDUCE_FLS(a, b) \
1447 if (a##_fls > b##_fls) { \
1457 * Reduce accuracy until either term fits in a u64, then proceed with
1458 * the other, so that finally we can do a u64/u64 division.
1460 while (count_fls
+ sec_fls
> 64 && nsec_fls
+ frequency_fls
> 64) {
1461 REDUCE_FLS(nsec
, frequency
);
1462 REDUCE_FLS(sec
, count
);
1465 if (count_fls
+ sec_fls
> 64) {
1466 divisor
= nsec
* frequency
;
1468 while (count_fls
+ sec_fls
> 64) {
1469 REDUCE_FLS(count
, sec
);
1473 dividend
= count
* sec
;
1475 dividend
= count
* sec
;
1477 while (nsec_fls
+ frequency_fls
> 64) {
1478 REDUCE_FLS(nsec
, frequency
);
1482 divisor
= nsec
* frequency
;
1485 return div64_u64(dividend
, divisor
);
1488 static void perf_event_stop(struct perf_event
*event
)
1490 if (!event
->pmu
->stop
)
1491 return event
->pmu
->disable(event
);
1493 return event
->pmu
->stop(event
);
1496 static int perf_event_start(struct perf_event
*event
)
1498 if (!event
->pmu
->start
)
1499 return event
->pmu
->enable(event
);
1501 return event
->pmu
->start(event
);
1504 static void perf_adjust_period(struct perf_event
*event
, u64 nsec
, u64 count
)
1506 struct hw_perf_event
*hwc
= &event
->hw
;
1507 u64 period
, sample_period
;
1510 period
= perf_calculate_period(event
, nsec
, count
);
1512 delta
= (s64
)(period
- hwc
->sample_period
);
1513 delta
= (delta
+ 7) / 8; /* low pass filter */
1515 sample_period
= hwc
->sample_period
+ delta
;
1520 hwc
->sample_period
= sample_period
;
1522 if (atomic64_read(&hwc
->period_left
) > 8*sample_period
) {
1524 perf_event_stop(event
);
1525 atomic64_set(&hwc
->period_left
, 0);
1526 perf_event_start(event
);
1531 static void perf_ctx_adjust_freq(struct perf_event_context
*ctx
)
1533 struct perf_event
*event
;
1534 struct hw_perf_event
*hwc
;
1535 u64 interrupts
, now
;
1538 raw_spin_lock(&ctx
->lock
);
1539 list_for_each_entry_rcu(event
, &ctx
->event_list
, event_entry
) {
1540 if (event
->state
!= PERF_EVENT_STATE_ACTIVE
)
1543 if (event
->cpu
!= -1 && event
->cpu
!= smp_processor_id())
1548 interrupts
= hwc
->interrupts
;
1549 hwc
->interrupts
= 0;
1552 * unthrottle events on the tick
1554 if (interrupts
== MAX_INTERRUPTS
) {
1555 perf_log_throttle(event
, 1);
1556 event
->pmu
->unthrottle(event
);
1559 if (!event
->attr
.freq
|| !event
->attr
.sample_freq
)
1562 event
->pmu
->read(event
);
1563 now
= atomic64_read(&event
->count
);
1564 delta
= now
- hwc
->freq_count_stamp
;
1565 hwc
->freq_count_stamp
= now
;
1568 perf_adjust_period(event
, TICK_NSEC
, delta
);
1570 raw_spin_unlock(&ctx
->lock
);
1574 * Round-robin a context's events:
1576 static void rotate_ctx(struct perf_event_context
*ctx
)
1578 if (!ctx
->nr_events
)
1581 raw_spin_lock(&ctx
->lock
);
1583 /* Rotate the first entry last of non-pinned groups */
1584 list_rotate_left(&ctx
->flexible_groups
);
1586 raw_spin_unlock(&ctx
->lock
);
1589 void perf_event_task_tick(struct task_struct
*curr
)
1591 struct perf_cpu_context
*cpuctx
;
1592 struct perf_event_context
*ctx
;
1594 if (!atomic_read(&nr_events
))
1597 cpuctx
= &__get_cpu_var(perf_cpu_context
);
1598 ctx
= curr
->perf_event_ctxp
;
1602 perf_ctx_adjust_freq(&cpuctx
->ctx
);
1604 perf_ctx_adjust_freq(ctx
);
1606 cpu_ctx_sched_out(cpuctx
, EVENT_FLEXIBLE
);
1608 task_ctx_sched_out(ctx
, EVENT_FLEXIBLE
);
1610 rotate_ctx(&cpuctx
->ctx
);
1614 cpu_ctx_sched_in(cpuctx
, EVENT_FLEXIBLE
);
1616 task_ctx_sched_in(curr
, EVENT_FLEXIBLE
);
1621 static int event_enable_on_exec(struct perf_event
*event
,
1622 struct perf_event_context
*ctx
)
1624 if (!event
->attr
.enable_on_exec
)
1627 event
->attr
.enable_on_exec
= 0;
1628 if (event
->state
>= PERF_EVENT_STATE_INACTIVE
)
1631 __perf_event_mark_enabled(event
, ctx
);
1637 * Enable all of a task's events that have been marked enable-on-exec.
1638 * This expects task == current.
1640 static void perf_event_enable_on_exec(struct task_struct
*task
)
1642 struct perf_event_context
*ctx
;
1643 struct perf_event
*event
;
1644 unsigned long flags
;
1648 local_irq_save(flags
);
1649 ctx
= task
->perf_event_ctxp
;
1650 if (!ctx
|| !ctx
->nr_events
)
1653 __perf_event_task_sched_out(ctx
);
1655 raw_spin_lock(&ctx
->lock
);
1657 list_for_each_entry(event
, &ctx
->pinned_groups
, group_entry
) {
1658 ret
= event_enable_on_exec(event
, ctx
);
1663 list_for_each_entry(event
, &ctx
->flexible_groups
, group_entry
) {
1664 ret
= event_enable_on_exec(event
, ctx
);
1670 * Unclone this context if we enabled any event.
1675 raw_spin_unlock(&ctx
->lock
);
1677 perf_event_task_sched_in(task
);
1679 local_irq_restore(flags
);
1683 * Cross CPU call to read the hardware event
1685 static void __perf_event_read(void *info
)
1687 struct perf_cpu_context
*cpuctx
= &__get_cpu_var(perf_cpu_context
);
1688 struct perf_event
*event
= info
;
1689 struct perf_event_context
*ctx
= event
->ctx
;
1692 * If this is a task context, we need to check whether it is
1693 * the current task context of this cpu. If not it has been
1694 * scheduled out before the smp call arrived. In that case
1695 * event->count would have been updated to a recent sample
1696 * when the event was scheduled out.
1698 if (ctx
->task
&& cpuctx
->task_ctx
!= ctx
)
1701 raw_spin_lock(&ctx
->lock
);
1702 update_context_time(ctx
);
1703 update_event_times(event
);
1704 raw_spin_unlock(&ctx
->lock
);
1706 event
->pmu
->read(event
);
1709 static u64
perf_event_read(struct perf_event
*event
)
1712 * If event is enabled and currently active on a CPU, update the
1713 * value in the event structure:
1715 if (event
->state
== PERF_EVENT_STATE_ACTIVE
) {
1716 smp_call_function_single(event
->oncpu
,
1717 __perf_event_read
, event
, 1);
1718 } else if (event
->state
== PERF_EVENT_STATE_INACTIVE
) {
1719 struct perf_event_context
*ctx
= event
->ctx
;
1720 unsigned long flags
;
1722 raw_spin_lock_irqsave(&ctx
->lock
, flags
);
1723 update_context_time(ctx
);
1724 update_event_times(event
);
1725 raw_spin_unlock_irqrestore(&ctx
->lock
, flags
);
1728 return atomic64_read(&event
->count
);
1732 * Initialize the perf_event context in a task_struct:
1735 __perf_event_init_context(struct perf_event_context
*ctx
,
1736 struct task_struct
*task
)
1738 raw_spin_lock_init(&ctx
->lock
);
1739 mutex_init(&ctx
->mutex
);
1740 INIT_LIST_HEAD(&ctx
->pinned_groups
);
1741 INIT_LIST_HEAD(&ctx
->flexible_groups
);
1742 INIT_LIST_HEAD(&ctx
->event_list
);
1743 atomic_set(&ctx
->refcount
, 1);
1747 static struct perf_event_context
*find_get_context(pid_t pid
, int cpu
)
1749 struct perf_event_context
*ctx
;
1750 struct perf_cpu_context
*cpuctx
;
1751 struct task_struct
*task
;
1752 unsigned long flags
;
1755 if (pid
== -1 && cpu
!= -1) {
1756 /* Must be root to operate on a CPU event: */
1757 if (perf_paranoid_cpu() && !capable(CAP_SYS_ADMIN
))
1758 return ERR_PTR(-EACCES
);
1760 if (cpu
< 0 || cpu
>= nr_cpumask_bits
)
1761 return ERR_PTR(-EINVAL
);
1764 * We could be clever and allow to attach a event to an
1765 * offline CPU and activate it when the CPU comes up, but
1768 if (!cpu_online(cpu
))
1769 return ERR_PTR(-ENODEV
);
1771 cpuctx
= &per_cpu(perf_cpu_context
, cpu
);
1782 task
= find_task_by_vpid(pid
);
1784 get_task_struct(task
);
1788 return ERR_PTR(-ESRCH
);
1791 * Can't attach events to a dying task.
1794 if (task
->flags
& PF_EXITING
)
1797 /* Reuse ptrace permission checks for now. */
1799 if (!ptrace_may_access(task
, PTRACE_MODE_READ
))
1803 ctx
= perf_lock_task_context(task
, &flags
);
1806 raw_spin_unlock_irqrestore(&ctx
->lock
, flags
);
1810 ctx
= kzalloc(sizeof(struct perf_event_context
), GFP_KERNEL
);
1814 __perf_event_init_context(ctx
, task
);
1816 if (cmpxchg(&task
->perf_event_ctxp
, NULL
, ctx
)) {
1818 * We raced with some other task; use
1819 * the context they set.
1824 get_task_struct(task
);
1827 put_task_struct(task
);
1831 put_task_struct(task
);
1832 return ERR_PTR(err
);
1835 static void perf_event_free_filter(struct perf_event
*event
);
1837 static void free_event_rcu(struct rcu_head
*head
)
1839 struct perf_event
*event
;
1841 event
= container_of(head
, struct perf_event
, rcu_head
);
1843 put_pid_ns(event
->ns
);
1844 perf_event_free_filter(event
);
1848 static void perf_pending_sync(struct perf_event
*event
);
1850 static void free_event(struct perf_event
*event
)
1852 perf_pending_sync(event
);
1854 if (!event
->parent
) {
1855 atomic_dec(&nr_events
);
1856 if (event
->attr
.mmap
)
1857 atomic_dec(&nr_mmap_events
);
1858 if (event
->attr
.comm
)
1859 atomic_dec(&nr_comm_events
);
1860 if (event
->attr
.task
)
1861 atomic_dec(&nr_task_events
);
1864 if (event
->output
) {
1865 fput(event
->output
->filp
);
1866 event
->output
= NULL
;
1870 event
->destroy(event
);
1872 put_ctx(event
->ctx
);
1873 call_rcu(&event
->rcu_head
, free_event_rcu
);
1876 int perf_event_release_kernel(struct perf_event
*event
)
1878 struct perf_event_context
*ctx
= event
->ctx
;
1880 WARN_ON_ONCE(ctx
->parent_ctx
);
1881 mutex_lock(&ctx
->mutex
);
1882 perf_event_remove_from_context(event
);
1883 mutex_unlock(&ctx
->mutex
);
1885 mutex_lock(&event
->owner
->perf_event_mutex
);
1886 list_del_init(&event
->owner_entry
);
1887 mutex_unlock(&event
->owner
->perf_event_mutex
);
1888 put_task_struct(event
->owner
);
1894 EXPORT_SYMBOL_GPL(perf_event_release_kernel
);
1897 * Called when the last reference to the file is gone.
1899 static int perf_release(struct inode
*inode
, struct file
*file
)
1901 struct perf_event
*event
= file
->private_data
;
1903 file
->private_data
= NULL
;
1905 return perf_event_release_kernel(event
);
1908 static int perf_event_read_size(struct perf_event
*event
)
1910 int entry
= sizeof(u64
); /* value */
1914 if (event
->attr
.read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
)
1915 size
+= sizeof(u64
);
1917 if (event
->attr
.read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
)
1918 size
+= sizeof(u64
);
1920 if (event
->attr
.read_format
& PERF_FORMAT_ID
)
1921 entry
+= sizeof(u64
);
1923 if (event
->attr
.read_format
& PERF_FORMAT_GROUP
) {
1924 nr
+= event
->group_leader
->nr_siblings
;
1925 size
+= sizeof(u64
);
1933 u64
perf_event_read_value(struct perf_event
*event
, u64
*enabled
, u64
*running
)
1935 struct perf_event
*child
;
1941 mutex_lock(&event
->child_mutex
);
1942 total
+= perf_event_read(event
);
1943 *enabled
+= event
->total_time_enabled
+
1944 atomic64_read(&event
->child_total_time_enabled
);
1945 *running
+= event
->total_time_running
+
1946 atomic64_read(&event
->child_total_time_running
);
1948 list_for_each_entry(child
, &event
->child_list
, child_list
) {
1949 total
+= perf_event_read(child
);
1950 *enabled
+= child
->total_time_enabled
;
1951 *running
+= child
->total_time_running
;
1953 mutex_unlock(&event
->child_mutex
);
1957 EXPORT_SYMBOL_GPL(perf_event_read_value
);
1959 static int perf_event_read_group(struct perf_event
*event
,
1960 u64 read_format
, char __user
*buf
)
1962 struct perf_event
*leader
= event
->group_leader
, *sub
;
1963 int n
= 0, size
= 0, ret
= -EFAULT
;
1964 struct perf_event_context
*ctx
= leader
->ctx
;
1966 u64 count
, enabled
, running
;
1968 mutex_lock(&ctx
->mutex
);
1969 count
= perf_event_read_value(leader
, &enabled
, &running
);
1971 values
[n
++] = 1 + leader
->nr_siblings
;
1972 if (read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
)
1973 values
[n
++] = enabled
;
1974 if (read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
)
1975 values
[n
++] = running
;
1976 values
[n
++] = count
;
1977 if (read_format
& PERF_FORMAT_ID
)
1978 values
[n
++] = primary_event_id(leader
);
1980 size
= n
* sizeof(u64
);
1982 if (copy_to_user(buf
, values
, size
))
1987 list_for_each_entry(sub
, &leader
->sibling_list
, group_entry
) {
1990 values
[n
++] = perf_event_read_value(sub
, &enabled
, &running
);
1991 if (read_format
& PERF_FORMAT_ID
)
1992 values
[n
++] = primary_event_id(sub
);
1994 size
= n
* sizeof(u64
);
1996 if (copy_to_user(buf
+ ret
, values
, size
)) {
2004 mutex_unlock(&ctx
->mutex
);
2009 static int perf_event_read_one(struct perf_event
*event
,
2010 u64 read_format
, char __user
*buf
)
2012 u64 enabled
, running
;
2016 values
[n
++] = perf_event_read_value(event
, &enabled
, &running
);
2017 if (read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
)
2018 values
[n
++] = enabled
;
2019 if (read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
)
2020 values
[n
++] = running
;
2021 if (read_format
& PERF_FORMAT_ID
)
2022 values
[n
++] = primary_event_id(event
);
2024 if (copy_to_user(buf
, values
, n
* sizeof(u64
)))
2027 return n
* sizeof(u64
);
2031 * Read the performance event - simple non blocking version for now
2034 perf_read_hw(struct perf_event
*event
, char __user
*buf
, size_t count
)
2036 u64 read_format
= event
->attr
.read_format
;
2040 * Return end-of-file for a read on a event that is in
2041 * error state (i.e. because it was pinned but it couldn't be
2042 * scheduled on to the CPU at some point).
2044 if (event
->state
== PERF_EVENT_STATE_ERROR
)
2047 if (count
< perf_event_read_size(event
))
2050 WARN_ON_ONCE(event
->ctx
->parent_ctx
);
2051 if (read_format
& PERF_FORMAT_GROUP
)
2052 ret
= perf_event_read_group(event
, read_format
, buf
);
2054 ret
= perf_event_read_one(event
, read_format
, buf
);
2060 perf_read(struct file
*file
, char __user
*buf
, size_t count
, loff_t
*ppos
)
2062 struct perf_event
*event
= file
->private_data
;
2064 return perf_read_hw(event
, buf
, count
);
2067 static unsigned int perf_poll(struct file
*file
, poll_table
*wait
)
2069 struct perf_event
*event
= file
->private_data
;
2070 struct perf_mmap_data
*data
;
2071 unsigned int events
= POLL_HUP
;
2074 data
= rcu_dereference(event
->data
);
2076 events
= atomic_xchg(&data
->poll
, 0);
2079 poll_wait(file
, &event
->waitq
, wait
);
2084 static void perf_event_reset(struct perf_event
*event
)
2086 (void)perf_event_read(event
);
2087 atomic64_set(&event
->count
, 0);
2088 perf_event_update_userpage(event
);
2092 * Holding the top-level event's child_mutex means that any
2093 * descendant process that has inherited this event will block
2094 * in sync_child_event if it goes to exit, thus satisfying the
2095 * task existence requirements of perf_event_enable/disable.
2097 static void perf_event_for_each_child(struct perf_event
*event
,
2098 void (*func
)(struct perf_event
*))
2100 struct perf_event
*child
;
2102 WARN_ON_ONCE(event
->ctx
->parent_ctx
);
2103 mutex_lock(&event
->child_mutex
);
2105 list_for_each_entry(child
, &event
->child_list
, child_list
)
2107 mutex_unlock(&event
->child_mutex
);
2110 static void perf_event_for_each(struct perf_event
*event
,
2111 void (*func
)(struct perf_event
*))
2113 struct perf_event_context
*ctx
= event
->ctx
;
2114 struct perf_event
*sibling
;
2116 WARN_ON_ONCE(ctx
->parent_ctx
);
2117 mutex_lock(&ctx
->mutex
);
2118 event
= event
->group_leader
;
2120 perf_event_for_each_child(event
, func
);
2122 list_for_each_entry(sibling
, &event
->sibling_list
, group_entry
)
2123 perf_event_for_each_child(event
, func
);
2124 mutex_unlock(&ctx
->mutex
);
2127 static int perf_event_period(struct perf_event
*event
, u64 __user
*arg
)
2129 struct perf_event_context
*ctx
= event
->ctx
;
2134 if (!event
->attr
.sample_period
)
2137 size
= copy_from_user(&value
, arg
, sizeof(value
));
2138 if (size
!= sizeof(value
))
2144 raw_spin_lock_irq(&ctx
->lock
);
2145 if (event
->attr
.freq
) {
2146 if (value
> sysctl_perf_event_sample_rate
) {
2151 event
->attr
.sample_freq
= value
;
2153 event
->attr
.sample_period
= value
;
2154 event
->hw
.sample_period
= value
;
2157 raw_spin_unlock_irq(&ctx
->lock
);
2162 static int perf_event_set_output(struct perf_event
*event
, int output_fd
);
2163 static int perf_event_set_filter(struct perf_event
*event
, void __user
*arg
);
2165 static long perf_ioctl(struct file
*file
, unsigned int cmd
, unsigned long arg
)
2167 struct perf_event
*event
= file
->private_data
;
2168 void (*func
)(struct perf_event
*);
2172 case PERF_EVENT_IOC_ENABLE
:
2173 func
= perf_event_enable
;
2175 case PERF_EVENT_IOC_DISABLE
:
2176 func
= perf_event_disable
;
2178 case PERF_EVENT_IOC_RESET
:
2179 func
= perf_event_reset
;
2182 case PERF_EVENT_IOC_REFRESH
:
2183 return perf_event_refresh(event
, arg
);
2185 case PERF_EVENT_IOC_PERIOD
:
2186 return perf_event_period(event
, (u64 __user
*)arg
);
2188 case PERF_EVENT_IOC_SET_OUTPUT
:
2189 return perf_event_set_output(event
, arg
);
2191 case PERF_EVENT_IOC_SET_FILTER
:
2192 return perf_event_set_filter(event
, (void __user
*)arg
);
2198 if (flags
& PERF_IOC_FLAG_GROUP
)
2199 perf_event_for_each(event
, func
);
2201 perf_event_for_each_child(event
, func
);
2206 int perf_event_task_enable(void)
2208 struct perf_event
*event
;
2210 mutex_lock(¤t
->perf_event_mutex
);
2211 list_for_each_entry(event
, ¤t
->perf_event_list
, owner_entry
)
2212 perf_event_for_each_child(event
, perf_event_enable
);
2213 mutex_unlock(¤t
->perf_event_mutex
);
2218 int perf_event_task_disable(void)
2220 struct perf_event
*event
;
2222 mutex_lock(¤t
->perf_event_mutex
);
2223 list_for_each_entry(event
, ¤t
->perf_event_list
, owner_entry
)
2224 perf_event_for_each_child(event
, perf_event_disable
);
2225 mutex_unlock(¤t
->perf_event_mutex
);
2230 #ifndef PERF_EVENT_INDEX_OFFSET
2231 # define PERF_EVENT_INDEX_OFFSET 0
2234 static int perf_event_index(struct perf_event
*event
)
2236 if (event
->state
!= PERF_EVENT_STATE_ACTIVE
)
2239 return event
->hw
.idx
+ 1 - PERF_EVENT_INDEX_OFFSET
;
2243 * Callers need to ensure there can be no nesting of this function, otherwise
2244 * the seqlock logic goes bad. We can not serialize this because the arch
2245 * code calls this from NMI context.
2247 void perf_event_update_userpage(struct perf_event
*event
)
2249 struct perf_event_mmap_page
*userpg
;
2250 struct perf_mmap_data
*data
;
2253 data
= rcu_dereference(event
->data
);
2257 userpg
= data
->user_page
;
2260 * Disable preemption so as to not let the corresponding user-space
2261 * spin too long if we get preempted.
2266 userpg
->index
= perf_event_index(event
);
2267 userpg
->offset
= atomic64_read(&event
->count
);
2268 if (event
->state
== PERF_EVENT_STATE_ACTIVE
)
2269 userpg
->offset
-= atomic64_read(&event
->hw
.prev_count
);
2271 userpg
->time_enabled
= event
->total_time_enabled
+
2272 atomic64_read(&event
->child_total_time_enabled
);
2274 userpg
->time_running
= event
->total_time_running
+
2275 atomic64_read(&event
->child_total_time_running
);
2284 static unsigned long perf_data_size(struct perf_mmap_data
*data
)
2286 return data
->nr_pages
<< (PAGE_SHIFT
+ data
->data_order
);
2289 #ifndef CONFIG_PERF_USE_VMALLOC
2292 * Back perf_mmap() with regular GFP_KERNEL-0 pages.
2295 static struct page
*
2296 perf_mmap_to_page(struct perf_mmap_data
*data
, unsigned long pgoff
)
2298 if (pgoff
> data
->nr_pages
)
2302 return virt_to_page(data
->user_page
);
2304 return virt_to_page(data
->data_pages
[pgoff
- 1]);
2307 static struct perf_mmap_data
*
2308 perf_mmap_data_alloc(struct perf_event
*event
, int nr_pages
)
2310 struct perf_mmap_data
*data
;
2314 WARN_ON(atomic_read(&event
->mmap_count
));
2316 size
= sizeof(struct perf_mmap_data
);
2317 size
+= nr_pages
* sizeof(void *);
2319 data
= kzalloc(size
, GFP_KERNEL
);
2323 data
->user_page
= (void *)get_zeroed_page(GFP_KERNEL
);
2324 if (!data
->user_page
)
2325 goto fail_user_page
;
2327 for (i
= 0; i
< nr_pages
; i
++) {
2328 data
->data_pages
[i
] = (void *)get_zeroed_page(GFP_KERNEL
);
2329 if (!data
->data_pages
[i
])
2330 goto fail_data_pages
;
2333 data
->data_order
= 0;
2334 data
->nr_pages
= nr_pages
;
2339 for (i
--; i
>= 0; i
--)
2340 free_page((unsigned long)data
->data_pages
[i
]);
2342 free_page((unsigned long)data
->user_page
);
2351 static void perf_mmap_free_page(unsigned long addr
)
2353 struct page
*page
= virt_to_page((void *)addr
);
2355 page
->mapping
= NULL
;
2359 static void perf_mmap_data_free(struct perf_mmap_data
*data
)
2363 perf_mmap_free_page((unsigned long)data
->user_page
);
2364 for (i
= 0; i
< data
->nr_pages
; i
++)
2365 perf_mmap_free_page((unsigned long)data
->data_pages
[i
]);
2372 * Back perf_mmap() with vmalloc memory.
2374 * Required for architectures that have d-cache aliasing issues.
2377 static struct page
*
2378 perf_mmap_to_page(struct perf_mmap_data
*data
, unsigned long pgoff
)
2380 if (pgoff
> (1UL << data
->data_order
))
2383 return vmalloc_to_page((void *)data
->user_page
+ pgoff
* PAGE_SIZE
);
2386 static void perf_mmap_unmark_page(void *addr
)
2388 struct page
*page
= vmalloc_to_page(addr
);
2390 page
->mapping
= NULL
;
2393 static void perf_mmap_data_free_work(struct work_struct
*work
)
2395 struct perf_mmap_data
*data
;
2399 data
= container_of(work
, struct perf_mmap_data
, work
);
2400 nr
= 1 << data
->data_order
;
2402 base
= data
->user_page
;
2403 for (i
= 0; i
< nr
+ 1; i
++)
2404 perf_mmap_unmark_page(base
+ (i
* PAGE_SIZE
));
2410 static void perf_mmap_data_free(struct perf_mmap_data
*data
)
2412 schedule_work(&data
->work
);
2415 static struct perf_mmap_data
*
2416 perf_mmap_data_alloc(struct perf_event
*event
, int nr_pages
)
2418 struct perf_mmap_data
*data
;
2422 WARN_ON(atomic_read(&event
->mmap_count
));
2424 size
= sizeof(struct perf_mmap_data
);
2425 size
+= sizeof(void *);
2427 data
= kzalloc(size
, GFP_KERNEL
);
2431 INIT_WORK(&data
->work
, perf_mmap_data_free_work
);
2433 all_buf
= vmalloc_user((nr_pages
+ 1) * PAGE_SIZE
);
2437 data
->user_page
= all_buf
;
2438 data
->data_pages
[0] = all_buf
+ PAGE_SIZE
;
2439 data
->data_order
= ilog2(nr_pages
);
2453 static int perf_mmap_fault(struct vm_area_struct
*vma
, struct vm_fault
*vmf
)
2455 struct perf_event
*event
= vma
->vm_file
->private_data
;
2456 struct perf_mmap_data
*data
;
2457 int ret
= VM_FAULT_SIGBUS
;
2459 if (vmf
->flags
& FAULT_FLAG_MKWRITE
) {
2460 if (vmf
->pgoff
== 0)
2466 data
= rcu_dereference(event
->data
);
2470 if (vmf
->pgoff
&& (vmf
->flags
& FAULT_FLAG_WRITE
))
2473 vmf
->page
= perf_mmap_to_page(data
, vmf
->pgoff
);
2477 get_page(vmf
->page
);
2478 vmf
->page
->mapping
= vma
->vm_file
->f_mapping
;
2479 vmf
->page
->index
= vmf
->pgoff
;
2489 perf_mmap_data_init(struct perf_event
*event
, struct perf_mmap_data
*data
)
2491 long max_size
= perf_data_size(data
);
2493 atomic_set(&data
->lock
, -1);
2495 if (event
->attr
.watermark
) {
2496 data
->watermark
= min_t(long, max_size
,
2497 event
->attr
.wakeup_watermark
);
2500 if (!data
->watermark
)
2501 data
->watermark
= max_size
/ 2;
2504 rcu_assign_pointer(event
->data
, data
);
2507 static void perf_mmap_data_free_rcu(struct rcu_head
*rcu_head
)
2509 struct perf_mmap_data
*data
;
2511 data
= container_of(rcu_head
, struct perf_mmap_data
, rcu_head
);
2512 perf_mmap_data_free(data
);
2515 static void perf_mmap_data_release(struct perf_event
*event
)
2517 struct perf_mmap_data
*data
= event
->data
;
2519 WARN_ON(atomic_read(&event
->mmap_count
));
2521 rcu_assign_pointer(event
->data
, NULL
);
2522 call_rcu(&data
->rcu_head
, perf_mmap_data_free_rcu
);
2525 static void perf_mmap_open(struct vm_area_struct
*vma
)
2527 struct perf_event
*event
= vma
->vm_file
->private_data
;
2529 atomic_inc(&event
->mmap_count
);
2532 static void perf_mmap_close(struct vm_area_struct
*vma
)
2534 struct perf_event
*event
= vma
->vm_file
->private_data
;
2536 WARN_ON_ONCE(event
->ctx
->parent_ctx
);
2537 if (atomic_dec_and_mutex_lock(&event
->mmap_count
, &event
->mmap_mutex
)) {
2538 unsigned long size
= perf_data_size(event
->data
);
2539 struct user_struct
*user
= current_user();
2541 atomic_long_sub((size
>> PAGE_SHIFT
) + 1, &user
->locked_vm
);
2542 vma
->vm_mm
->locked_vm
-= event
->data
->nr_locked
;
2543 perf_mmap_data_release(event
);
2544 mutex_unlock(&event
->mmap_mutex
);
2548 static const struct vm_operations_struct perf_mmap_vmops
= {
2549 .open
= perf_mmap_open
,
2550 .close
= perf_mmap_close
,
2551 .fault
= perf_mmap_fault
,
2552 .page_mkwrite
= perf_mmap_fault
,
2555 static int perf_mmap(struct file
*file
, struct vm_area_struct
*vma
)
2557 struct perf_event
*event
= file
->private_data
;
2558 unsigned long user_locked
, user_lock_limit
;
2559 struct user_struct
*user
= current_user();
2560 unsigned long locked
, lock_limit
;
2561 struct perf_mmap_data
*data
;
2562 unsigned long vma_size
;
2563 unsigned long nr_pages
;
2564 long user_extra
, extra
;
2567 if (!(vma
->vm_flags
& VM_SHARED
))
2570 vma_size
= vma
->vm_end
- vma
->vm_start
;
2571 nr_pages
= (vma_size
/ PAGE_SIZE
) - 1;
2574 * If we have data pages ensure they're a power-of-two number, so we
2575 * can do bitmasks instead of modulo.
2577 if (nr_pages
!= 0 && !is_power_of_2(nr_pages
))
2580 if (vma_size
!= PAGE_SIZE
* (1 + nr_pages
))
2583 if (vma
->vm_pgoff
!= 0)
2586 WARN_ON_ONCE(event
->ctx
->parent_ctx
);
2587 mutex_lock(&event
->mmap_mutex
);
2588 if (event
->output
) {
2593 if (atomic_inc_not_zero(&event
->mmap_count
)) {
2594 if (nr_pages
!= event
->data
->nr_pages
)
2599 user_extra
= nr_pages
+ 1;
2600 user_lock_limit
= sysctl_perf_event_mlock
>> (PAGE_SHIFT
- 10);
2603 * Increase the limit linearly with more CPUs:
2605 user_lock_limit
*= num_online_cpus();
2607 user_locked
= atomic_long_read(&user
->locked_vm
) + user_extra
;
2610 if (user_locked
> user_lock_limit
)
2611 extra
= user_locked
- user_lock_limit
;
2613 lock_limit
= current
->signal
->rlim
[RLIMIT_MEMLOCK
].rlim_cur
;
2614 lock_limit
>>= PAGE_SHIFT
;
2615 locked
= vma
->vm_mm
->locked_vm
+ extra
;
2617 if ((locked
> lock_limit
) && perf_paranoid_tracepoint_raw() &&
2618 !capable(CAP_IPC_LOCK
)) {
2623 WARN_ON(event
->data
);
2625 data
= perf_mmap_data_alloc(event
, nr_pages
);
2631 perf_mmap_data_init(event
, data
);
2633 atomic_set(&event
->mmap_count
, 1);
2634 atomic_long_add(user_extra
, &user
->locked_vm
);
2635 vma
->vm_mm
->locked_vm
+= extra
;
2636 event
->data
->nr_locked
= extra
;
2637 if (vma
->vm_flags
& VM_WRITE
)
2638 event
->data
->writable
= 1;
2641 mutex_unlock(&event
->mmap_mutex
);
2643 vma
->vm_flags
|= VM_RESERVED
;
2644 vma
->vm_ops
= &perf_mmap_vmops
;
2649 static int perf_fasync(int fd
, struct file
*filp
, int on
)
2651 struct inode
*inode
= filp
->f_path
.dentry
->d_inode
;
2652 struct perf_event
*event
= filp
->private_data
;
2655 mutex_lock(&inode
->i_mutex
);
2656 retval
= fasync_helper(fd
, filp
, on
, &event
->fasync
);
2657 mutex_unlock(&inode
->i_mutex
);
2665 static const struct file_operations perf_fops
= {
2666 .release
= perf_release
,
2669 .unlocked_ioctl
= perf_ioctl
,
2670 .compat_ioctl
= perf_ioctl
,
2672 .fasync
= perf_fasync
,
2678 * If there's data, ensure we set the poll() state and publish everything
2679 * to user-space before waking everybody up.
2682 void perf_event_wakeup(struct perf_event
*event
)
2684 wake_up_all(&event
->waitq
);
2686 if (event
->pending_kill
) {
2687 kill_fasync(&event
->fasync
, SIGIO
, event
->pending_kill
);
2688 event
->pending_kill
= 0;
2695 * Handle the case where we need to wakeup up from NMI (or rq->lock) context.
2697 * The NMI bit means we cannot possibly take locks. Therefore, maintain a
2698 * single linked list and use cmpxchg() to add entries lockless.
2701 static void perf_pending_event(struct perf_pending_entry
*entry
)
2703 struct perf_event
*event
= container_of(entry
,
2704 struct perf_event
, pending
);
2706 if (event
->pending_disable
) {
2707 event
->pending_disable
= 0;
2708 __perf_event_disable(event
);
2711 if (event
->pending_wakeup
) {
2712 event
->pending_wakeup
= 0;
2713 perf_event_wakeup(event
);
2717 #define PENDING_TAIL ((struct perf_pending_entry *)-1UL)
2719 static DEFINE_PER_CPU(struct perf_pending_entry
*, perf_pending_head
) = {
2723 static void perf_pending_queue(struct perf_pending_entry
*entry
,
2724 void (*func
)(struct perf_pending_entry
*))
2726 struct perf_pending_entry
**head
;
2728 if (cmpxchg(&entry
->next
, NULL
, PENDING_TAIL
) != NULL
)
2733 head
= &get_cpu_var(perf_pending_head
);
2736 entry
->next
= *head
;
2737 } while (cmpxchg(head
, entry
->next
, entry
) != entry
->next
);
2739 set_perf_event_pending();
2741 put_cpu_var(perf_pending_head
);
2744 static int __perf_pending_run(void)
2746 struct perf_pending_entry
*list
;
2749 list
= xchg(&__get_cpu_var(perf_pending_head
), PENDING_TAIL
);
2750 while (list
!= PENDING_TAIL
) {
2751 void (*func
)(struct perf_pending_entry
*);
2752 struct perf_pending_entry
*entry
= list
;
2759 * Ensure we observe the unqueue before we issue the wakeup,
2760 * so that we won't be waiting forever.
2761 * -- see perf_not_pending().
2772 static inline int perf_not_pending(struct perf_event
*event
)
2775 * If we flush on whatever cpu we run, there is a chance we don't
2779 __perf_pending_run();
2783 * Ensure we see the proper queue state before going to sleep
2784 * so that we do not miss the wakeup. -- see perf_pending_handle()
2787 return event
->pending
.next
== NULL
;
2790 static void perf_pending_sync(struct perf_event
*event
)
2792 wait_event(event
->waitq
, perf_not_pending(event
));
2795 void perf_event_do_pending(void)
2797 __perf_pending_run();
2801 * Callchain support -- arch specific
2804 __weak
struct perf_callchain_entry
*perf_callchain(struct pt_regs
*regs
)
2812 static bool perf_output_space(struct perf_mmap_data
*data
, unsigned long tail
,
2813 unsigned long offset
, unsigned long head
)
2817 if (!data
->writable
)
2820 mask
= perf_data_size(data
) - 1;
2822 offset
= (offset
- tail
) & mask
;
2823 head
= (head
- tail
) & mask
;
2825 if ((int)(head
- offset
) < 0)
2831 static void perf_output_wakeup(struct perf_output_handle
*handle
)
2833 atomic_set(&handle
->data
->poll
, POLL_IN
);
2836 handle
->event
->pending_wakeup
= 1;
2837 perf_pending_queue(&handle
->event
->pending
,
2838 perf_pending_event
);
2840 perf_event_wakeup(handle
->event
);
2844 * Curious locking construct.
2846 * We need to ensure a later event_id doesn't publish a head when a former
2847 * event_id isn't done writing. However since we need to deal with NMIs we
2848 * cannot fully serialize things.
2850 * What we do is serialize between CPUs so we only have to deal with NMI
2851 * nesting on a single CPU.
2853 * We only publish the head (and generate a wakeup) when the outer-most
2854 * event_id completes.
2856 static void perf_output_lock(struct perf_output_handle
*handle
)
2858 struct perf_mmap_data
*data
= handle
->data
;
2859 int cur
, cpu
= get_cpu();
2864 cur
= atomic_cmpxchg(&data
->lock
, -1, cpu
);
2876 static void perf_output_unlock(struct perf_output_handle
*handle
)
2878 struct perf_mmap_data
*data
= handle
->data
;
2882 data
->done_head
= data
->head
;
2884 if (!handle
->locked
)
2889 * The xchg implies a full barrier that ensures all writes are done
2890 * before we publish the new head, matched by a rmb() in userspace when
2891 * reading this position.
2893 while ((head
= atomic_long_xchg(&data
->done_head
, 0)))
2894 data
->user_page
->data_head
= head
;
2897 * NMI can happen here, which means we can miss a done_head update.
2900 cpu
= atomic_xchg(&data
->lock
, -1);
2901 WARN_ON_ONCE(cpu
!= smp_processor_id());
2904 * Therefore we have to validate we did not indeed do so.
2906 if (unlikely(atomic_long_read(&data
->done_head
))) {
2908 * Since we had it locked, we can lock it again.
2910 while (atomic_cmpxchg(&data
->lock
, -1, cpu
) != -1)
2916 if (atomic_xchg(&data
->wakeup
, 0))
2917 perf_output_wakeup(handle
);
2922 void perf_output_copy(struct perf_output_handle
*handle
,
2923 const void *buf
, unsigned int len
)
2925 unsigned int pages_mask
;
2926 unsigned long offset
;
2930 offset
= handle
->offset
;
2931 pages_mask
= handle
->data
->nr_pages
- 1;
2932 pages
= handle
->data
->data_pages
;
2935 unsigned long page_offset
;
2936 unsigned long page_size
;
2939 nr
= (offset
>> PAGE_SHIFT
) & pages_mask
;
2940 page_size
= 1UL << (handle
->data
->data_order
+ PAGE_SHIFT
);
2941 page_offset
= offset
& (page_size
- 1);
2942 size
= min_t(unsigned int, page_size
- page_offset
, len
);
2944 memcpy(pages
[nr
] + page_offset
, buf
, size
);
2951 handle
->offset
= offset
;
2954 * Check we didn't copy past our reservation window, taking the
2955 * possible unsigned int wrap into account.
2957 WARN_ON_ONCE(((long)(handle
->head
- handle
->offset
)) < 0);
2960 int perf_output_begin(struct perf_output_handle
*handle
,
2961 struct perf_event
*event
, unsigned int size
,
2962 int nmi
, int sample
)
2964 struct perf_event
*output_event
;
2965 struct perf_mmap_data
*data
;
2966 unsigned long tail
, offset
, head
;
2969 struct perf_event_header header
;
2976 * For inherited events we send all the output towards the parent.
2979 event
= event
->parent
;
2981 output_event
= rcu_dereference(event
->output
);
2983 event
= output_event
;
2985 data
= rcu_dereference(event
->data
);
2989 handle
->data
= data
;
2990 handle
->event
= event
;
2992 handle
->sample
= sample
;
2994 if (!data
->nr_pages
)
2997 have_lost
= atomic_read(&data
->lost
);
2999 size
+= sizeof(lost_event
);
3001 perf_output_lock(handle
);
3005 * Userspace could choose to issue a mb() before updating the
3006 * tail pointer. So that all reads will be completed before the
3009 tail
= ACCESS_ONCE(data
->user_page
->data_tail
);
3011 offset
= head
= atomic_long_read(&data
->head
);
3013 if (unlikely(!perf_output_space(data
, tail
, offset
, head
)))
3015 } while (atomic_long_cmpxchg(&data
->head
, offset
, head
) != offset
);
3017 handle
->offset
= offset
;
3018 handle
->head
= head
;
3020 if (head
- tail
> data
->watermark
)
3021 atomic_set(&data
->wakeup
, 1);
3024 lost_event
.header
.type
= PERF_RECORD_LOST
;
3025 lost_event
.header
.misc
= 0;
3026 lost_event
.header
.size
= sizeof(lost_event
);
3027 lost_event
.id
= event
->id
;
3028 lost_event
.lost
= atomic_xchg(&data
->lost
, 0);
3030 perf_output_put(handle
, lost_event
);
3036 atomic_inc(&data
->lost
);
3037 perf_output_unlock(handle
);
3044 void perf_output_end(struct perf_output_handle
*handle
)
3046 struct perf_event
*event
= handle
->event
;
3047 struct perf_mmap_data
*data
= handle
->data
;
3049 int wakeup_events
= event
->attr
.wakeup_events
;
3051 if (handle
->sample
&& wakeup_events
) {
3052 int events
= atomic_inc_return(&data
->events
);
3053 if (events
>= wakeup_events
) {
3054 atomic_sub(wakeup_events
, &data
->events
);
3055 atomic_set(&data
->wakeup
, 1);
3059 perf_output_unlock(handle
);
3063 static u32
perf_event_pid(struct perf_event
*event
, struct task_struct
*p
)
3066 * only top level events have the pid namespace they were created in
3069 event
= event
->parent
;
3071 return task_tgid_nr_ns(p
, event
->ns
);
3074 static u32
perf_event_tid(struct perf_event
*event
, struct task_struct
*p
)
3077 * only top level events have the pid namespace they were created in
3080 event
= event
->parent
;
3082 return task_pid_nr_ns(p
, event
->ns
);
3085 static void perf_output_read_one(struct perf_output_handle
*handle
,
3086 struct perf_event
*event
)
3088 u64 read_format
= event
->attr
.read_format
;
3092 values
[n
++] = atomic64_read(&event
->count
);
3093 if (read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
) {
3094 values
[n
++] = event
->total_time_enabled
+
3095 atomic64_read(&event
->child_total_time_enabled
);
3097 if (read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
) {
3098 values
[n
++] = event
->total_time_running
+
3099 atomic64_read(&event
->child_total_time_running
);
3101 if (read_format
& PERF_FORMAT_ID
)
3102 values
[n
++] = primary_event_id(event
);
3104 perf_output_copy(handle
, values
, n
* sizeof(u64
));
3108 * XXX PERF_FORMAT_GROUP vs inherited events seems difficult.
3110 static void perf_output_read_group(struct perf_output_handle
*handle
,
3111 struct perf_event
*event
)
3113 struct perf_event
*leader
= event
->group_leader
, *sub
;
3114 u64 read_format
= event
->attr
.read_format
;
3118 values
[n
++] = 1 + leader
->nr_siblings
;
3120 if (read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
)
3121 values
[n
++] = leader
->total_time_enabled
;
3123 if (read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
)
3124 values
[n
++] = leader
->total_time_running
;
3126 if (leader
!= event
)
3127 leader
->pmu
->read(leader
);
3129 values
[n
++] = atomic64_read(&leader
->count
);
3130 if (read_format
& PERF_FORMAT_ID
)
3131 values
[n
++] = primary_event_id(leader
);
3133 perf_output_copy(handle
, values
, n
* sizeof(u64
));
3135 list_for_each_entry(sub
, &leader
->sibling_list
, group_entry
) {
3139 sub
->pmu
->read(sub
);
3141 values
[n
++] = atomic64_read(&sub
->count
);
3142 if (read_format
& PERF_FORMAT_ID
)
3143 values
[n
++] = primary_event_id(sub
);
3145 perf_output_copy(handle
, values
, n
* sizeof(u64
));
3149 static void perf_output_read(struct perf_output_handle
*handle
,
3150 struct perf_event
*event
)
3152 if (event
->attr
.read_format
& PERF_FORMAT_GROUP
)
3153 perf_output_read_group(handle
, event
);
3155 perf_output_read_one(handle
, event
);
3158 void perf_output_sample(struct perf_output_handle
*handle
,
3159 struct perf_event_header
*header
,
3160 struct perf_sample_data
*data
,
3161 struct perf_event
*event
)
3163 u64 sample_type
= data
->type
;
3165 perf_output_put(handle
, *header
);
3167 if (sample_type
& PERF_SAMPLE_IP
)
3168 perf_output_put(handle
, data
->ip
);
3170 if (sample_type
& PERF_SAMPLE_TID
)
3171 perf_output_put(handle
, data
->tid_entry
);
3173 if (sample_type
& PERF_SAMPLE_TIME
)
3174 perf_output_put(handle
, data
->time
);
3176 if (sample_type
& PERF_SAMPLE_ADDR
)
3177 perf_output_put(handle
, data
->addr
);
3179 if (sample_type
& PERF_SAMPLE_ID
)
3180 perf_output_put(handle
, data
->id
);
3182 if (sample_type
& PERF_SAMPLE_STREAM_ID
)
3183 perf_output_put(handle
, data
->stream_id
);
3185 if (sample_type
& PERF_SAMPLE_CPU
)
3186 perf_output_put(handle
, data
->cpu_entry
);
3188 if (sample_type
& PERF_SAMPLE_PERIOD
)
3189 perf_output_put(handle
, data
->period
);
3191 if (sample_type
& PERF_SAMPLE_READ
)
3192 perf_output_read(handle
, event
);
3194 if (sample_type
& PERF_SAMPLE_CALLCHAIN
) {
3195 if (data
->callchain
) {
3198 if (data
->callchain
)
3199 size
+= data
->callchain
->nr
;
3201 size
*= sizeof(u64
);
3203 perf_output_copy(handle
, data
->callchain
, size
);
3206 perf_output_put(handle
, nr
);
3210 if (sample_type
& PERF_SAMPLE_RAW
) {
3212 perf_output_put(handle
, data
->raw
->size
);
3213 perf_output_copy(handle
, data
->raw
->data
,
3220 .size
= sizeof(u32
),
3223 perf_output_put(handle
, raw
);
3228 void perf_prepare_sample(struct perf_event_header
*header
,
3229 struct perf_sample_data
*data
,
3230 struct perf_event
*event
,
3231 struct pt_regs
*regs
)
3233 u64 sample_type
= event
->attr
.sample_type
;
3235 data
->type
= sample_type
;
3237 header
->type
= PERF_RECORD_SAMPLE
;
3238 header
->size
= sizeof(*header
);
3241 header
->misc
|= perf_misc_flags(regs
);
3243 if (sample_type
& PERF_SAMPLE_IP
) {
3244 data
->ip
= perf_instruction_pointer(regs
);
3246 header
->size
+= sizeof(data
->ip
);
3249 if (sample_type
& PERF_SAMPLE_TID
) {
3250 /* namespace issues */
3251 data
->tid_entry
.pid
= perf_event_pid(event
, current
);
3252 data
->tid_entry
.tid
= perf_event_tid(event
, current
);
3254 header
->size
+= sizeof(data
->tid_entry
);
3257 if (sample_type
& PERF_SAMPLE_TIME
) {
3258 data
->time
= perf_clock();
3260 header
->size
+= sizeof(data
->time
);
3263 if (sample_type
& PERF_SAMPLE_ADDR
)
3264 header
->size
+= sizeof(data
->addr
);
3266 if (sample_type
& PERF_SAMPLE_ID
) {
3267 data
->id
= primary_event_id(event
);
3269 header
->size
+= sizeof(data
->id
);
3272 if (sample_type
& PERF_SAMPLE_STREAM_ID
) {
3273 data
->stream_id
= event
->id
;
3275 header
->size
+= sizeof(data
->stream_id
);
3278 if (sample_type
& PERF_SAMPLE_CPU
) {
3279 data
->cpu_entry
.cpu
= raw_smp_processor_id();
3280 data
->cpu_entry
.reserved
= 0;
3282 header
->size
+= sizeof(data
->cpu_entry
);
3285 if (sample_type
& PERF_SAMPLE_PERIOD
)
3286 header
->size
+= sizeof(data
->period
);
3288 if (sample_type
& PERF_SAMPLE_READ
)
3289 header
->size
+= perf_event_read_size(event
);
3291 if (sample_type
& PERF_SAMPLE_CALLCHAIN
) {
3294 data
->callchain
= perf_callchain(regs
);
3296 if (data
->callchain
)
3297 size
+= data
->callchain
->nr
;
3299 header
->size
+= size
* sizeof(u64
);
3302 if (sample_type
& PERF_SAMPLE_RAW
) {
3303 int size
= sizeof(u32
);
3306 size
+= data
->raw
->size
;
3308 size
+= sizeof(u32
);
3310 WARN_ON_ONCE(size
& (sizeof(u64
)-1));
3311 header
->size
+= size
;
3315 static void perf_event_output(struct perf_event
*event
, int nmi
,
3316 struct perf_sample_data
*data
,
3317 struct pt_regs
*regs
)
3319 struct perf_output_handle handle
;
3320 struct perf_event_header header
;
3322 perf_prepare_sample(&header
, data
, event
, regs
);
3324 if (perf_output_begin(&handle
, event
, header
.size
, nmi
, 1))
3327 perf_output_sample(&handle
, &header
, data
, event
);
3329 perf_output_end(&handle
);
3336 struct perf_read_event
{
3337 struct perf_event_header header
;
3344 perf_event_read_event(struct perf_event
*event
,
3345 struct task_struct
*task
)
3347 struct perf_output_handle handle
;
3348 struct perf_read_event read_event
= {
3350 .type
= PERF_RECORD_READ
,
3352 .size
= sizeof(read_event
) + perf_event_read_size(event
),
3354 .pid
= perf_event_pid(event
, task
),
3355 .tid
= perf_event_tid(event
, task
),
3359 ret
= perf_output_begin(&handle
, event
, read_event
.header
.size
, 0, 0);
3363 perf_output_put(&handle
, read_event
);
3364 perf_output_read(&handle
, event
);
3366 perf_output_end(&handle
);
3370 * task tracking -- fork/exit
3372 * enabled by: attr.comm | attr.mmap | attr.task
3375 struct perf_task_event
{
3376 struct task_struct
*task
;
3377 struct perf_event_context
*task_ctx
;
3380 struct perf_event_header header
;
3390 static void perf_event_task_output(struct perf_event
*event
,
3391 struct perf_task_event
*task_event
)
3393 struct perf_output_handle handle
;
3395 struct task_struct
*task
= task_event
->task
;
3398 size
= task_event
->event_id
.header
.size
;
3399 ret
= perf_output_begin(&handle
, event
, size
, 0, 0);
3404 task_event
->event_id
.pid
= perf_event_pid(event
, task
);
3405 task_event
->event_id
.ppid
= perf_event_pid(event
, current
);
3407 task_event
->event_id
.tid
= perf_event_tid(event
, task
);
3408 task_event
->event_id
.ptid
= perf_event_tid(event
, current
);
3410 task_event
->event_id
.time
= perf_clock();
3412 perf_output_put(&handle
, task_event
->event_id
);
3414 perf_output_end(&handle
);
3417 static int perf_event_task_match(struct perf_event
*event
)
3419 if (event
->state
!= PERF_EVENT_STATE_ACTIVE
)
3422 if (event
->cpu
!= -1 && event
->cpu
!= smp_processor_id())
3425 if (event
->attr
.comm
|| event
->attr
.mmap
|| event
->attr
.task
)
3431 static void perf_event_task_ctx(struct perf_event_context
*ctx
,
3432 struct perf_task_event
*task_event
)
3434 struct perf_event
*event
;
3436 list_for_each_entry_rcu(event
, &ctx
->event_list
, event_entry
) {
3437 if (perf_event_task_match(event
))
3438 perf_event_task_output(event
, task_event
);
3442 static void perf_event_task_event(struct perf_task_event
*task_event
)
3444 struct perf_cpu_context
*cpuctx
;
3445 struct perf_event_context
*ctx
= task_event
->task_ctx
;
3448 cpuctx
= &get_cpu_var(perf_cpu_context
);
3449 perf_event_task_ctx(&cpuctx
->ctx
, task_event
);
3451 ctx
= rcu_dereference(task_event
->task
->perf_event_ctxp
);
3453 perf_event_task_ctx(ctx
, task_event
);
3454 put_cpu_var(perf_cpu_context
);
3458 static void perf_event_task(struct task_struct
*task
,
3459 struct perf_event_context
*task_ctx
,
3462 struct perf_task_event task_event
;
3464 if (!atomic_read(&nr_comm_events
) &&
3465 !atomic_read(&nr_mmap_events
) &&
3466 !atomic_read(&nr_task_events
))
3469 task_event
= (struct perf_task_event
){
3471 .task_ctx
= task_ctx
,
3474 .type
= new ? PERF_RECORD_FORK
: PERF_RECORD_EXIT
,
3476 .size
= sizeof(task_event
.event_id
),
3485 perf_event_task_event(&task_event
);
3488 void perf_event_fork(struct task_struct
*task
)
3490 perf_event_task(task
, NULL
, 1);
3497 struct perf_comm_event
{
3498 struct task_struct
*task
;
3503 struct perf_event_header header
;
3510 static void perf_event_comm_output(struct perf_event
*event
,
3511 struct perf_comm_event
*comm_event
)
3513 struct perf_output_handle handle
;
3514 int size
= comm_event
->event_id
.header
.size
;
3515 int ret
= perf_output_begin(&handle
, event
, size
, 0, 0);
3520 comm_event
->event_id
.pid
= perf_event_pid(event
, comm_event
->task
);
3521 comm_event
->event_id
.tid
= perf_event_tid(event
, comm_event
->task
);
3523 perf_output_put(&handle
, comm_event
->event_id
);
3524 perf_output_copy(&handle
, comm_event
->comm
,
3525 comm_event
->comm_size
);
3526 perf_output_end(&handle
);
3529 static int perf_event_comm_match(struct perf_event
*event
)
3531 if (event
->state
!= PERF_EVENT_STATE_ACTIVE
)
3534 if (event
->cpu
!= -1 && event
->cpu
!= smp_processor_id())
3537 if (event
->attr
.comm
)
3543 static void perf_event_comm_ctx(struct perf_event_context
*ctx
,
3544 struct perf_comm_event
*comm_event
)
3546 struct perf_event
*event
;
3548 list_for_each_entry_rcu(event
, &ctx
->event_list
, event_entry
) {
3549 if (perf_event_comm_match(event
))
3550 perf_event_comm_output(event
, comm_event
);
3554 static void perf_event_comm_event(struct perf_comm_event
*comm_event
)
3556 struct perf_cpu_context
*cpuctx
;
3557 struct perf_event_context
*ctx
;
3559 char comm
[TASK_COMM_LEN
];
3561 memset(comm
, 0, sizeof(comm
));
3562 strlcpy(comm
, comm_event
->task
->comm
, sizeof(comm
));
3563 size
= ALIGN(strlen(comm
)+1, sizeof(u64
));
3565 comm_event
->comm
= comm
;
3566 comm_event
->comm_size
= size
;
3568 comm_event
->event_id
.header
.size
= sizeof(comm_event
->event_id
) + size
;
3571 cpuctx
= &get_cpu_var(perf_cpu_context
);
3572 perf_event_comm_ctx(&cpuctx
->ctx
, comm_event
);
3573 ctx
= rcu_dereference(current
->perf_event_ctxp
);
3575 perf_event_comm_ctx(ctx
, comm_event
);
3576 put_cpu_var(perf_cpu_context
);
3580 void perf_event_comm(struct task_struct
*task
)
3582 struct perf_comm_event comm_event
;
3584 if (task
->perf_event_ctxp
)
3585 perf_event_enable_on_exec(task
);
3587 if (!atomic_read(&nr_comm_events
))
3590 comm_event
= (struct perf_comm_event
){
3596 .type
= PERF_RECORD_COMM
,
3605 perf_event_comm_event(&comm_event
);
3612 struct perf_mmap_event
{
3613 struct vm_area_struct
*vma
;
3615 const char *file_name
;
3619 struct perf_event_header header
;
3629 static void perf_event_mmap_output(struct perf_event
*event
,
3630 struct perf_mmap_event
*mmap_event
)
3632 struct perf_output_handle handle
;
3633 int size
= mmap_event
->event_id
.header
.size
;
3634 int ret
= perf_output_begin(&handle
, event
, size
, 0, 0);
3639 mmap_event
->event_id
.pid
= perf_event_pid(event
, current
);
3640 mmap_event
->event_id
.tid
= perf_event_tid(event
, current
);
3642 perf_output_put(&handle
, mmap_event
->event_id
);
3643 perf_output_copy(&handle
, mmap_event
->file_name
,
3644 mmap_event
->file_size
);
3645 perf_output_end(&handle
);
3648 static int perf_event_mmap_match(struct perf_event
*event
,
3649 struct perf_mmap_event
*mmap_event
)
3651 if (event
->state
!= PERF_EVENT_STATE_ACTIVE
)
3654 if (event
->cpu
!= -1 && event
->cpu
!= smp_processor_id())
3657 if (event
->attr
.mmap
)
3663 static void perf_event_mmap_ctx(struct perf_event_context
*ctx
,
3664 struct perf_mmap_event
*mmap_event
)
3666 struct perf_event
*event
;
3668 list_for_each_entry_rcu(event
, &ctx
->event_list
, event_entry
) {
3669 if (perf_event_mmap_match(event
, mmap_event
))
3670 perf_event_mmap_output(event
, mmap_event
);
3674 static void perf_event_mmap_event(struct perf_mmap_event
*mmap_event
)
3676 struct perf_cpu_context
*cpuctx
;
3677 struct perf_event_context
*ctx
;
3678 struct vm_area_struct
*vma
= mmap_event
->vma
;
3679 struct file
*file
= vma
->vm_file
;
3685 memset(tmp
, 0, sizeof(tmp
));
3689 * d_path works from the end of the buffer backwards, so we
3690 * need to add enough zero bytes after the string to handle
3691 * the 64bit alignment we do later.
3693 buf
= kzalloc(PATH_MAX
+ sizeof(u64
), GFP_KERNEL
);
3695 name
= strncpy(tmp
, "//enomem", sizeof(tmp
));
3698 name
= d_path(&file
->f_path
, buf
, PATH_MAX
);
3700 name
= strncpy(tmp
, "//toolong", sizeof(tmp
));
3704 if (arch_vma_name(mmap_event
->vma
)) {
3705 name
= strncpy(tmp
, arch_vma_name(mmap_event
->vma
),
3711 name
= strncpy(tmp
, "[vdso]", sizeof(tmp
));
3715 name
= strncpy(tmp
, "//anon", sizeof(tmp
));
3720 size
= ALIGN(strlen(name
)+1, sizeof(u64
));
3722 mmap_event
->file_name
= name
;
3723 mmap_event
->file_size
= size
;
3725 mmap_event
->event_id
.header
.size
= sizeof(mmap_event
->event_id
) + size
;
3728 cpuctx
= &get_cpu_var(perf_cpu_context
);
3729 perf_event_mmap_ctx(&cpuctx
->ctx
, mmap_event
);
3730 ctx
= rcu_dereference(current
->perf_event_ctxp
);
3732 perf_event_mmap_ctx(ctx
, mmap_event
);
3733 put_cpu_var(perf_cpu_context
);
3739 void __perf_event_mmap(struct vm_area_struct
*vma
)
3741 struct perf_mmap_event mmap_event
;
3743 if (!atomic_read(&nr_mmap_events
))
3746 mmap_event
= (struct perf_mmap_event
){
3752 .type
= PERF_RECORD_MMAP
,
3758 .start
= vma
->vm_start
,
3759 .len
= vma
->vm_end
- vma
->vm_start
,
3760 .pgoff
= (u64
)vma
->vm_pgoff
<< PAGE_SHIFT
,
3764 perf_event_mmap_event(&mmap_event
);
3768 * IRQ throttle logging
3771 static void perf_log_throttle(struct perf_event
*event
, int enable
)
3773 struct perf_output_handle handle
;
3777 struct perf_event_header header
;
3781 } throttle_event
= {
3783 .type
= PERF_RECORD_THROTTLE
,
3785 .size
= sizeof(throttle_event
),
3787 .time
= perf_clock(),
3788 .id
= primary_event_id(event
),
3789 .stream_id
= event
->id
,
3793 throttle_event
.header
.type
= PERF_RECORD_UNTHROTTLE
;
3795 ret
= perf_output_begin(&handle
, event
, sizeof(throttle_event
), 1, 0);
3799 perf_output_put(&handle
, throttle_event
);
3800 perf_output_end(&handle
);
3804 * Generic event overflow handling, sampling.
3807 static int __perf_event_overflow(struct perf_event
*event
, int nmi
,
3808 int throttle
, struct perf_sample_data
*data
,
3809 struct pt_regs
*regs
)
3811 int events
= atomic_read(&event
->event_limit
);
3812 struct hw_perf_event
*hwc
= &event
->hw
;
3815 throttle
= (throttle
&& event
->pmu
->unthrottle
!= NULL
);
3820 if (hwc
->interrupts
!= MAX_INTERRUPTS
) {
3822 if (HZ
* hwc
->interrupts
>
3823 (u64
)sysctl_perf_event_sample_rate
) {
3824 hwc
->interrupts
= MAX_INTERRUPTS
;
3825 perf_log_throttle(event
, 0);
3830 * Keep re-disabling events even though on the previous
3831 * pass we disabled it - just in case we raced with a
3832 * sched-in and the event got enabled again:
3838 if (event
->attr
.freq
) {
3839 u64 now
= perf_clock();
3840 s64 delta
= now
- hwc
->freq_time_stamp
;
3842 hwc
->freq_time_stamp
= now
;
3844 if (delta
> 0 && delta
< 2*TICK_NSEC
)
3845 perf_adjust_period(event
, delta
, hwc
->last_period
);
3849 * XXX event_limit might not quite work as expected on inherited
3853 event
->pending_kill
= POLL_IN
;
3854 if (events
&& atomic_dec_and_test(&event
->event_limit
)) {
3856 event
->pending_kill
= POLL_HUP
;
3858 event
->pending_disable
= 1;
3859 perf_pending_queue(&event
->pending
,
3860 perf_pending_event
);
3862 perf_event_disable(event
);
3865 if (event
->overflow_handler
)
3866 event
->overflow_handler(event
, nmi
, data
, regs
);
3868 perf_event_output(event
, nmi
, data
, regs
);
3873 int perf_event_overflow(struct perf_event
*event
, int nmi
,
3874 struct perf_sample_data
*data
,
3875 struct pt_regs
*regs
)
3877 return __perf_event_overflow(event
, nmi
, 1, data
, regs
);
3881 * Generic software event infrastructure
3885 * We directly increment event->count and keep a second value in
3886 * event->hw.period_left to count intervals. This period event
3887 * is kept in the range [-sample_period, 0] so that we can use the
3891 static u64
perf_swevent_set_period(struct perf_event
*event
)
3893 struct hw_perf_event
*hwc
= &event
->hw
;
3894 u64 period
= hwc
->last_period
;
3898 hwc
->last_period
= hwc
->sample_period
;
3901 old
= val
= atomic64_read(&hwc
->period_left
);
3905 nr
= div64_u64(period
+ val
, period
);
3906 offset
= nr
* period
;
3908 if (atomic64_cmpxchg(&hwc
->period_left
, old
, val
) != old
)
3914 static void perf_swevent_overflow(struct perf_event
*event
, u64 overflow
,
3915 int nmi
, struct perf_sample_data
*data
,
3916 struct pt_regs
*regs
)
3918 struct hw_perf_event
*hwc
= &event
->hw
;
3921 data
->period
= event
->hw
.last_period
;
3923 overflow
= perf_swevent_set_period(event
);
3925 if (hwc
->interrupts
== MAX_INTERRUPTS
)
3928 for (; overflow
; overflow
--) {
3929 if (__perf_event_overflow(event
, nmi
, throttle
,
3932 * We inhibit the overflow from happening when
3933 * hwc->interrupts == MAX_INTERRUPTS.
3941 static void perf_swevent_unthrottle(struct perf_event
*event
)
3944 * Nothing to do, we already reset hwc->interrupts.
3948 static void perf_swevent_add(struct perf_event
*event
, u64 nr
,
3949 int nmi
, struct perf_sample_data
*data
,
3950 struct pt_regs
*regs
)
3952 struct hw_perf_event
*hwc
= &event
->hw
;
3954 atomic64_add(nr
, &event
->count
);
3959 if (!hwc
->sample_period
)
3962 if (nr
== 1 && hwc
->sample_period
== 1 && !event
->attr
.freq
)
3963 return perf_swevent_overflow(event
, 1, nmi
, data
, regs
);
3965 if (atomic64_add_negative(nr
, &hwc
->period_left
))
3968 perf_swevent_overflow(event
, 0, nmi
, data
, regs
);
3971 static int perf_swevent_is_counting(struct perf_event
*event
)
3974 * The event is active, we're good!
3976 if (event
->state
== PERF_EVENT_STATE_ACTIVE
)
3980 * The event is off/error, not counting.
3982 if (event
->state
!= PERF_EVENT_STATE_INACTIVE
)
3986 * The event is inactive, if the context is active
3987 * we're part of a group that didn't make it on the 'pmu',
3990 if (event
->ctx
->is_active
)
3994 * We're inactive and the context is too, this means the
3995 * task is scheduled out, we're counting events that happen
3996 * to us, like migration events.
4001 static int perf_tp_event_match(struct perf_event
*event
,
4002 struct perf_sample_data
*data
);
4004 static int perf_exclude_event(struct perf_event
*event
,
4005 struct pt_regs
*regs
)
4008 if (event
->attr
.exclude_user
&& user_mode(regs
))
4011 if (event
->attr
.exclude_kernel
&& !user_mode(regs
))
4018 static int perf_swevent_match(struct perf_event
*event
,
4019 enum perf_type_id type
,
4021 struct perf_sample_data
*data
,
4022 struct pt_regs
*regs
)
4024 if (event
->cpu
!= -1 && event
->cpu
!= smp_processor_id())
4027 if (!perf_swevent_is_counting(event
))
4030 if (event
->attr
.type
!= type
)
4033 if (event
->attr
.config
!= event_id
)
4036 if (perf_exclude_event(event
, regs
))
4039 if (event
->attr
.type
== PERF_TYPE_TRACEPOINT
&&
4040 !perf_tp_event_match(event
, data
))
4046 static void perf_swevent_ctx_event(struct perf_event_context
*ctx
,
4047 enum perf_type_id type
,
4048 u32 event_id
, u64 nr
, int nmi
,
4049 struct perf_sample_data
*data
,
4050 struct pt_regs
*regs
)
4052 struct perf_event
*event
;
4054 list_for_each_entry_rcu(event
, &ctx
->event_list
, event_entry
) {
4055 if (perf_swevent_match(event
, type
, event_id
, data
, regs
))
4056 perf_swevent_add(event
, nr
, nmi
, data
, regs
);
4060 int perf_swevent_get_recursion_context(void)
4062 struct perf_cpu_context
*cpuctx
= &get_cpu_var(perf_cpu_context
);
4069 else if (in_softirq())
4074 if (cpuctx
->recursion
[rctx
]) {
4075 put_cpu_var(perf_cpu_context
);
4079 cpuctx
->recursion
[rctx
]++;
4084 EXPORT_SYMBOL_GPL(perf_swevent_get_recursion_context
);
4086 void perf_swevent_put_recursion_context(int rctx
)
4088 struct perf_cpu_context
*cpuctx
= &__get_cpu_var(perf_cpu_context
);
4090 cpuctx
->recursion
[rctx
]--;
4091 put_cpu_var(perf_cpu_context
);
4093 EXPORT_SYMBOL_GPL(perf_swevent_put_recursion_context
);
4095 static void do_perf_sw_event(enum perf_type_id type
, u32 event_id
,
4097 struct perf_sample_data
*data
,
4098 struct pt_regs
*regs
)
4100 struct perf_cpu_context
*cpuctx
;
4101 struct perf_event_context
*ctx
;
4103 cpuctx
= &__get_cpu_var(perf_cpu_context
);
4105 perf_swevent_ctx_event(&cpuctx
->ctx
, type
, event_id
,
4106 nr
, nmi
, data
, regs
);
4108 * doesn't really matter which of the child contexts the
4109 * events ends up in.
4111 ctx
= rcu_dereference(current
->perf_event_ctxp
);
4113 perf_swevent_ctx_event(ctx
, type
, event_id
, nr
, nmi
, data
, regs
);
4117 void __perf_sw_event(u32 event_id
, u64 nr
, int nmi
,
4118 struct pt_regs
*regs
, u64 addr
)
4120 struct perf_sample_data data
;
4123 rctx
= perf_swevent_get_recursion_context();
4130 do_perf_sw_event(PERF_TYPE_SOFTWARE
, event_id
, nr
, nmi
, &data
, regs
);
4132 perf_swevent_put_recursion_context(rctx
);
4135 static void perf_swevent_read(struct perf_event
*event
)
4139 static int perf_swevent_enable(struct perf_event
*event
)
4141 struct hw_perf_event
*hwc
= &event
->hw
;
4143 if (hwc
->sample_period
) {
4144 hwc
->last_period
= hwc
->sample_period
;
4145 perf_swevent_set_period(event
);
4150 static void perf_swevent_disable(struct perf_event
*event
)
4154 static const struct pmu perf_ops_generic
= {
4155 .enable
= perf_swevent_enable
,
4156 .disable
= perf_swevent_disable
,
4157 .read
= perf_swevent_read
,
4158 .unthrottle
= perf_swevent_unthrottle
,
4162 * hrtimer based swevent callback
4165 static enum hrtimer_restart
perf_swevent_hrtimer(struct hrtimer
*hrtimer
)
4167 enum hrtimer_restart ret
= HRTIMER_RESTART
;
4168 struct perf_sample_data data
;
4169 struct pt_regs
*regs
;
4170 struct perf_event
*event
;
4173 event
= container_of(hrtimer
, struct perf_event
, hw
.hrtimer
);
4174 event
->pmu
->read(event
);
4178 data
.period
= event
->hw
.last_period
;
4179 regs
= get_irq_regs();
4181 * In case we exclude kernel IPs or are somehow not in interrupt
4182 * context, provide the next best thing, the user IP.
4184 if ((event
->attr
.exclude_kernel
|| !regs
) &&
4185 !event
->attr
.exclude_user
)
4186 regs
= task_pt_regs(current
);
4189 if (!(event
->attr
.exclude_idle
&& current
->pid
== 0))
4190 if (perf_event_overflow(event
, 0, &data
, regs
))
4191 ret
= HRTIMER_NORESTART
;
4194 period
= max_t(u64
, 10000, event
->hw
.sample_period
);
4195 hrtimer_forward_now(hrtimer
, ns_to_ktime(period
));
4200 static void perf_swevent_start_hrtimer(struct perf_event
*event
)
4202 struct hw_perf_event
*hwc
= &event
->hw
;
4204 hrtimer_init(&hwc
->hrtimer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
4205 hwc
->hrtimer
.function
= perf_swevent_hrtimer
;
4206 if (hwc
->sample_period
) {
4209 if (hwc
->remaining
) {
4210 if (hwc
->remaining
< 0)
4213 period
= hwc
->remaining
;
4216 period
= max_t(u64
, 10000, hwc
->sample_period
);
4218 __hrtimer_start_range_ns(&hwc
->hrtimer
,
4219 ns_to_ktime(period
), 0,
4220 HRTIMER_MODE_REL
, 0);
4224 static void perf_swevent_cancel_hrtimer(struct perf_event
*event
)
4226 struct hw_perf_event
*hwc
= &event
->hw
;
4228 if (hwc
->sample_period
) {
4229 ktime_t remaining
= hrtimer_get_remaining(&hwc
->hrtimer
);
4230 hwc
->remaining
= ktime_to_ns(remaining
);
4232 hrtimer_cancel(&hwc
->hrtimer
);
4237 * Software event: cpu wall time clock
4240 static void cpu_clock_perf_event_update(struct perf_event
*event
)
4242 int cpu
= raw_smp_processor_id();
4246 now
= cpu_clock(cpu
);
4247 prev
= atomic64_xchg(&event
->hw
.prev_count
, now
);
4248 atomic64_add(now
- prev
, &event
->count
);
4251 static int cpu_clock_perf_event_enable(struct perf_event
*event
)
4253 struct hw_perf_event
*hwc
= &event
->hw
;
4254 int cpu
= raw_smp_processor_id();
4256 atomic64_set(&hwc
->prev_count
, cpu_clock(cpu
));
4257 perf_swevent_start_hrtimer(event
);
4262 static void cpu_clock_perf_event_disable(struct perf_event
*event
)
4264 perf_swevent_cancel_hrtimer(event
);
4265 cpu_clock_perf_event_update(event
);
4268 static void cpu_clock_perf_event_read(struct perf_event
*event
)
4270 cpu_clock_perf_event_update(event
);
4273 static const struct pmu perf_ops_cpu_clock
= {
4274 .enable
= cpu_clock_perf_event_enable
,
4275 .disable
= cpu_clock_perf_event_disable
,
4276 .read
= cpu_clock_perf_event_read
,
4280 * Software event: task time clock
4283 static void task_clock_perf_event_update(struct perf_event
*event
, u64 now
)
4288 prev
= atomic64_xchg(&event
->hw
.prev_count
, now
);
4290 atomic64_add(delta
, &event
->count
);
4293 static int task_clock_perf_event_enable(struct perf_event
*event
)
4295 struct hw_perf_event
*hwc
= &event
->hw
;
4298 now
= event
->ctx
->time
;
4300 atomic64_set(&hwc
->prev_count
, now
);
4302 perf_swevent_start_hrtimer(event
);
4307 static void task_clock_perf_event_disable(struct perf_event
*event
)
4309 perf_swevent_cancel_hrtimer(event
);
4310 task_clock_perf_event_update(event
, event
->ctx
->time
);
4314 static void task_clock_perf_event_read(struct perf_event
*event
)
4319 update_context_time(event
->ctx
);
4320 time
= event
->ctx
->time
;
4322 u64 now
= perf_clock();
4323 u64 delta
= now
- event
->ctx
->timestamp
;
4324 time
= event
->ctx
->time
+ delta
;
4327 task_clock_perf_event_update(event
, time
);
4330 static const struct pmu perf_ops_task_clock
= {
4331 .enable
= task_clock_perf_event_enable
,
4332 .disable
= task_clock_perf_event_disable
,
4333 .read
= task_clock_perf_event_read
,
4336 #ifdef CONFIG_EVENT_TRACING
4338 void perf_tp_event(int event_id
, u64 addr
, u64 count
, void *record
,
4341 struct perf_raw_record raw
= {
4346 struct perf_sample_data data
= {
4351 struct pt_regs
*regs
= get_irq_regs();
4354 regs
= task_pt_regs(current
);
4356 /* Trace events already protected against recursion */
4357 do_perf_sw_event(PERF_TYPE_TRACEPOINT
, event_id
, count
, 1,
4360 EXPORT_SYMBOL_GPL(perf_tp_event
);
4362 static int perf_tp_event_match(struct perf_event
*event
,
4363 struct perf_sample_data
*data
)
4365 void *record
= data
->raw
->data
;
4367 if (likely(!event
->filter
) || filter_match_preds(event
->filter
, record
))
4372 static void tp_perf_event_destroy(struct perf_event
*event
)
4374 ftrace_profile_disable(event
->attr
.config
);
4377 static const struct pmu
*tp_perf_event_init(struct perf_event
*event
)
4380 * Raw tracepoint data is a severe data leak, only allow root to
4383 if ((event
->attr
.sample_type
& PERF_SAMPLE_RAW
) &&
4384 perf_paranoid_tracepoint_raw() &&
4385 !capable(CAP_SYS_ADMIN
))
4386 return ERR_PTR(-EPERM
);
4388 if (ftrace_profile_enable(event
->attr
.config
))
4391 event
->destroy
= tp_perf_event_destroy
;
4393 return &perf_ops_generic
;
4396 static int perf_event_set_filter(struct perf_event
*event
, void __user
*arg
)
4401 if (event
->attr
.type
!= PERF_TYPE_TRACEPOINT
)
4404 filter_str
= strndup_user(arg
, PAGE_SIZE
);
4405 if (IS_ERR(filter_str
))
4406 return PTR_ERR(filter_str
);
4408 ret
= ftrace_profile_set_filter(event
, event
->attr
.config
, filter_str
);
4414 static void perf_event_free_filter(struct perf_event
*event
)
4416 ftrace_profile_free_filter(event
);
4421 static int perf_tp_event_match(struct perf_event
*event
,
4422 struct perf_sample_data
*data
)
4427 static const struct pmu
*tp_perf_event_init(struct perf_event
*event
)
4432 static int perf_event_set_filter(struct perf_event
*event
, void __user
*arg
)
4437 static void perf_event_free_filter(struct perf_event
*event
)
4441 #endif /* CONFIG_EVENT_TRACING */
4443 #ifdef CONFIG_HAVE_HW_BREAKPOINT
4444 static void bp_perf_event_destroy(struct perf_event
*event
)
4446 release_bp_slot(event
);
4449 static const struct pmu
*bp_perf_event_init(struct perf_event
*bp
)
4453 err
= register_perf_hw_breakpoint(bp
);
4455 return ERR_PTR(err
);
4457 bp
->destroy
= bp_perf_event_destroy
;
4459 return &perf_ops_bp
;
4462 void perf_bp_event(struct perf_event
*bp
, void *data
)
4464 struct perf_sample_data sample
;
4465 struct pt_regs
*regs
= data
;
4468 sample
.addr
= bp
->attr
.bp_addr
;
4470 if (!perf_exclude_event(bp
, regs
))
4471 perf_swevent_add(bp
, 1, 1, &sample
, regs
);
4474 static const struct pmu
*bp_perf_event_init(struct perf_event
*bp
)
4479 void perf_bp_event(struct perf_event
*bp
, void *regs
)
4484 atomic_t perf_swevent_enabled
[PERF_COUNT_SW_MAX
];
4486 static void sw_perf_event_destroy(struct perf_event
*event
)
4488 u64 event_id
= event
->attr
.config
;
4490 WARN_ON(event
->parent
);
4492 atomic_dec(&perf_swevent_enabled
[event_id
]);
4495 static const struct pmu
*sw_perf_event_init(struct perf_event
*event
)
4497 const struct pmu
*pmu
= NULL
;
4498 u64 event_id
= event
->attr
.config
;
4501 * Software events (currently) can't in general distinguish
4502 * between user, kernel and hypervisor events.
4503 * However, context switches and cpu migrations are considered
4504 * to be kernel events, and page faults are never hypervisor
4508 case PERF_COUNT_SW_CPU_CLOCK
:
4509 pmu
= &perf_ops_cpu_clock
;
4512 case PERF_COUNT_SW_TASK_CLOCK
:
4514 * If the user instantiates this as a per-cpu event,
4515 * use the cpu_clock event instead.
4517 if (event
->ctx
->task
)
4518 pmu
= &perf_ops_task_clock
;
4520 pmu
= &perf_ops_cpu_clock
;
4523 case PERF_COUNT_SW_PAGE_FAULTS
:
4524 case PERF_COUNT_SW_PAGE_FAULTS_MIN
:
4525 case PERF_COUNT_SW_PAGE_FAULTS_MAJ
:
4526 case PERF_COUNT_SW_CONTEXT_SWITCHES
:
4527 case PERF_COUNT_SW_CPU_MIGRATIONS
:
4528 case PERF_COUNT_SW_ALIGNMENT_FAULTS
:
4529 case PERF_COUNT_SW_EMULATION_FAULTS
:
4530 if (!event
->parent
) {
4531 atomic_inc(&perf_swevent_enabled
[event_id
]);
4532 event
->destroy
= sw_perf_event_destroy
;
4534 pmu
= &perf_ops_generic
;
4542 * Allocate and initialize a event structure
4544 static struct perf_event
*
4545 perf_event_alloc(struct perf_event_attr
*attr
,
4547 struct perf_event_context
*ctx
,
4548 struct perf_event
*group_leader
,
4549 struct perf_event
*parent_event
,
4550 perf_overflow_handler_t overflow_handler
,
4553 const struct pmu
*pmu
;
4554 struct perf_event
*event
;
4555 struct hw_perf_event
*hwc
;
4558 event
= kzalloc(sizeof(*event
), gfpflags
);
4560 return ERR_PTR(-ENOMEM
);
4563 * Single events are their own group leaders, with an
4564 * empty sibling list:
4567 group_leader
= event
;
4569 mutex_init(&event
->child_mutex
);
4570 INIT_LIST_HEAD(&event
->child_list
);
4572 INIT_LIST_HEAD(&event
->group_entry
);
4573 INIT_LIST_HEAD(&event
->event_entry
);
4574 INIT_LIST_HEAD(&event
->sibling_list
);
4575 init_waitqueue_head(&event
->waitq
);
4577 mutex_init(&event
->mmap_mutex
);
4580 event
->attr
= *attr
;
4581 event
->group_leader
= group_leader
;
4586 event
->parent
= parent_event
;
4588 event
->ns
= get_pid_ns(current
->nsproxy
->pid_ns
);
4589 event
->id
= atomic64_inc_return(&perf_event_id
);
4591 event
->state
= PERF_EVENT_STATE_INACTIVE
;
4593 if (!overflow_handler
&& parent_event
)
4594 overflow_handler
= parent_event
->overflow_handler
;
4596 event
->overflow_handler
= overflow_handler
;
4599 event
->state
= PERF_EVENT_STATE_OFF
;
4604 hwc
->sample_period
= attr
->sample_period
;
4605 if (attr
->freq
&& attr
->sample_freq
)
4606 hwc
->sample_period
= 1;
4607 hwc
->last_period
= hwc
->sample_period
;
4609 atomic64_set(&hwc
->period_left
, hwc
->sample_period
);
4612 * we currently do not support PERF_FORMAT_GROUP on inherited events
4614 if (attr
->inherit
&& (attr
->read_format
& PERF_FORMAT_GROUP
))
4617 switch (attr
->type
) {
4619 case PERF_TYPE_HARDWARE
:
4620 case PERF_TYPE_HW_CACHE
:
4621 pmu
= hw_perf_event_init(event
);
4624 case PERF_TYPE_SOFTWARE
:
4625 pmu
= sw_perf_event_init(event
);
4628 case PERF_TYPE_TRACEPOINT
:
4629 pmu
= tp_perf_event_init(event
);
4632 case PERF_TYPE_BREAKPOINT
:
4633 pmu
= bp_perf_event_init(event
);
4644 else if (IS_ERR(pmu
))
4649 put_pid_ns(event
->ns
);
4651 return ERR_PTR(err
);
4656 if (!event
->parent
) {
4657 atomic_inc(&nr_events
);
4658 if (event
->attr
.mmap
)
4659 atomic_inc(&nr_mmap_events
);
4660 if (event
->attr
.comm
)
4661 atomic_inc(&nr_comm_events
);
4662 if (event
->attr
.task
)
4663 atomic_inc(&nr_task_events
);
4669 static int perf_copy_attr(struct perf_event_attr __user
*uattr
,
4670 struct perf_event_attr
*attr
)
4675 if (!access_ok(VERIFY_WRITE
, uattr
, PERF_ATTR_SIZE_VER0
))
4679 * zero the full structure, so that a short copy will be nice.
4681 memset(attr
, 0, sizeof(*attr
));
4683 ret
= get_user(size
, &uattr
->size
);
4687 if (size
> PAGE_SIZE
) /* silly large */
4690 if (!size
) /* abi compat */
4691 size
= PERF_ATTR_SIZE_VER0
;
4693 if (size
< PERF_ATTR_SIZE_VER0
)
4697 * If we're handed a bigger struct than we know of,
4698 * ensure all the unknown bits are 0 - i.e. new
4699 * user-space does not rely on any kernel feature
4700 * extensions we dont know about yet.
4702 if (size
> sizeof(*attr
)) {
4703 unsigned char __user
*addr
;
4704 unsigned char __user
*end
;
4707 addr
= (void __user
*)uattr
+ sizeof(*attr
);
4708 end
= (void __user
*)uattr
+ size
;
4710 for (; addr
< end
; addr
++) {
4711 ret
= get_user(val
, addr
);
4717 size
= sizeof(*attr
);
4720 ret
= copy_from_user(attr
, uattr
, size
);
4725 * If the type exists, the corresponding creation will verify
4728 if (attr
->type
>= PERF_TYPE_MAX
)
4731 if (attr
->__reserved_1
|| attr
->__reserved_2
)
4734 if (attr
->sample_type
& ~(PERF_SAMPLE_MAX
-1))
4737 if (attr
->read_format
& ~(PERF_FORMAT_MAX
-1))
4744 put_user(sizeof(*attr
), &uattr
->size
);
4749 static int perf_event_set_output(struct perf_event
*event
, int output_fd
)
4751 struct perf_event
*output_event
= NULL
;
4752 struct file
*output_file
= NULL
;
4753 struct perf_event
*old_output
;
4754 int fput_needed
= 0;
4760 output_file
= fget_light(output_fd
, &fput_needed
);
4764 if (output_file
->f_op
!= &perf_fops
)
4767 output_event
= output_file
->private_data
;
4769 /* Don't chain output fds */
4770 if (output_event
->output
)
4773 /* Don't set an output fd when we already have an output channel */
4777 atomic_long_inc(&output_file
->f_count
);
4780 mutex_lock(&event
->mmap_mutex
);
4781 old_output
= event
->output
;
4782 rcu_assign_pointer(event
->output
, output_event
);
4783 mutex_unlock(&event
->mmap_mutex
);
4787 * we need to make sure no existing perf_output_*()
4788 * is still referencing this event.
4791 fput(old_output
->filp
);
4796 fput_light(output_file
, fput_needed
);
4801 * sys_perf_event_open - open a performance event, associate it to a task/cpu
4803 * @attr_uptr: event_id type attributes for monitoring/sampling
4806 * @group_fd: group leader event fd
4808 SYSCALL_DEFINE5(perf_event_open
,
4809 struct perf_event_attr __user
*, attr_uptr
,
4810 pid_t
, pid
, int, cpu
, int, group_fd
, unsigned long, flags
)
4812 struct perf_event
*event
, *group_leader
;
4813 struct perf_event_attr attr
;
4814 struct perf_event_context
*ctx
;
4815 struct file
*event_file
= NULL
;
4816 struct file
*group_file
= NULL
;
4817 int fput_needed
= 0;
4818 int fput_needed2
= 0;
4821 /* for future expandability... */
4822 if (flags
& ~(PERF_FLAG_FD_NO_GROUP
| PERF_FLAG_FD_OUTPUT
))
4825 err
= perf_copy_attr(attr_uptr
, &attr
);
4829 if (!attr
.exclude_kernel
) {
4830 if (perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN
))
4835 if (attr
.sample_freq
> sysctl_perf_event_sample_rate
)
4840 * Get the target context (task or percpu):
4842 ctx
= find_get_context(pid
, cpu
);
4844 return PTR_ERR(ctx
);
4847 * Look up the group leader (we will attach this event to it):
4849 group_leader
= NULL
;
4850 if (group_fd
!= -1 && !(flags
& PERF_FLAG_FD_NO_GROUP
)) {
4852 group_file
= fget_light(group_fd
, &fput_needed
);
4854 goto err_put_context
;
4855 if (group_file
->f_op
!= &perf_fops
)
4856 goto err_put_context
;
4858 group_leader
= group_file
->private_data
;
4860 * Do not allow a recursive hierarchy (this new sibling
4861 * becoming part of another group-sibling):
4863 if (group_leader
->group_leader
!= group_leader
)
4864 goto err_put_context
;
4866 * Do not allow to attach to a group in a different
4867 * task or CPU context:
4869 if (group_leader
->ctx
!= ctx
)
4870 goto err_put_context
;
4872 * Only a group leader can be exclusive or pinned
4874 if (attr
.exclusive
|| attr
.pinned
)
4875 goto err_put_context
;
4878 event
= perf_event_alloc(&attr
, cpu
, ctx
, group_leader
,
4879 NULL
, NULL
, GFP_KERNEL
);
4880 err
= PTR_ERR(event
);
4882 goto err_put_context
;
4884 err
= anon_inode_getfd("[perf_event]", &perf_fops
, event
, O_RDWR
);
4886 goto err_free_put_context
;
4888 event_file
= fget_light(err
, &fput_needed2
);
4890 goto err_free_put_context
;
4892 if (flags
& PERF_FLAG_FD_OUTPUT
) {
4893 err
= perf_event_set_output(event
, group_fd
);
4895 goto err_fput_free_put_context
;
4898 event
->filp
= event_file
;
4899 WARN_ON_ONCE(ctx
->parent_ctx
);
4900 mutex_lock(&ctx
->mutex
);
4901 perf_install_in_context(ctx
, event
, cpu
);
4903 mutex_unlock(&ctx
->mutex
);
4905 event
->owner
= current
;
4906 get_task_struct(current
);
4907 mutex_lock(¤t
->perf_event_mutex
);
4908 list_add_tail(&event
->owner_entry
, ¤t
->perf_event_list
);
4909 mutex_unlock(¤t
->perf_event_mutex
);
4911 err_fput_free_put_context
:
4912 fput_light(event_file
, fput_needed2
);
4914 err_free_put_context
:
4922 fput_light(group_file
, fput_needed
);
4928 * perf_event_create_kernel_counter
4930 * @attr: attributes of the counter to create
4931 * @cpu: cpu in which the counter is bound
4932 * @pid: task to profile
4935 perf_event_create_kernel_counter(struct perf_event_attr
*attr
, int cpu
,
4937 perf_overflow_handler_t overflow_handler
)
4939 struct perf_event
*event
;
4940 struct perf_event_context
*ctx
;
4944 * Get the target context (task or percpu):
4947 ctx
= find_get_context(pid
, cpu
);
4953 event
= perf_event_alloc(attr
, cpu
, ctx
, NULL
,
4954 NULL
, overflow_handler
, GFP_KERNEL
);
4955 if (IS_ERR(event
)) {
4956 err
= PTR_ERR(event
);
4957 goto err_put_context
;
4961 WARN_ON_ONCE(ctx
->parent_ctx
);
4962 mutex_lock(&ctx
->mutex
);
4963 perf_install_in_context(ctx
, event
, cpu
);
4965 mutex_unlock(&ctx
->mutex
);
4967 event
->owner
= current
;
4968 get_task_struct(current
);
4969 mutex_lock(¤t
->perf_event_mutex
);
4970 list_add_tail(&event
->owner_entry
, ¤t
->perf_event_list
);
4971 mutex_unlock(¤t
->perf_event_mutex
);
4978 return ERR_PTR(err
);
4980 EXPORT_SYMBOL_GPL(perf_event_create_kernel_counter
);
4983 * inherit a event from parent task to child task:
4985 static struct perf_event
*
4986 inherit_event(struct perf_event
*parent_event
,
4987 struct task_struct
*parent
,
4988 struct perf_event_context
*parent_ctx
,
4989 struct task_struct
*child
,
4990 struct perf_event
*group_leader
,
4991 struct perf_event_context
*child_ctx
)
4993 struct perf_event
*child_event
;
4996 * Instead of creating recursive hierarchies of events,
4997 * we link inherited events back to the original parent,
4998 * which has a filp for sure, which we use as the reference
5001 if (parent_event
->parent
)
5002 parent_event
= parent_event
->parent
;
5004 child_event
= perf_event_alloc(&parent_event
->attr
,
5005 parent_event
->cpu
, child_ctx
,
5006 group_leader
, parent_event
,
5008 if (IS_ERR(child_event
))
5013 * Make the child state follow the state of the parent event,
5014 * not its attr.disabled bit. We hold the parent's mutex,
5015 * so we won't race with perf_event_{en, dis}able_family.
5017 if (parent_event
->state
>= PERF_EVENT_STATE_INACTIVE
)
5018 child_event
->state
= PERF_EVENT_STATE_INACTIVE
;
5020 child_event
->state
= PERF_EVENT_STATE_OFF
;
5022 if (parent_event
->attr
.freq
) {
5023 u64 sample_period
= parent_event
->hw
.sample_period
;
5024 struct hw_perf_event
*hwc
= &child_event
->hw
;
5026 hwc
->sample_period
= sample_period
;
5027 hwc
->last_period
= sample_period
;
5029 atomic64_set(&hwc
->period_left
, sample_period
);
5032 child_event
->overflow_handler
= parent_event
->overflow_handler
;
5035 * Link it up in the child's context:
5037 add_event_to_ctx(child_event
, child_ctx
);
5040 * Get a reference to the parent filp - we will fput it
5041 * when the child event exits. This is safe to do because
5042 * we are in the parent and we know that the filp still
5043 * exists and has a nonzero count:
5045 atomic_long_inc(&parent_event
->filp
->f_count
);
5048 * Link this into the parent event's child list
5050 WARN_ON_ONCE(parent_event
->ctx
->parent_ctx
);
5051 mutex_lock(&parent_event
->child_mutex
);
5052 list_add_tail(&child_event
->child_list
, &parent_event
->child_list
);
5053 mutex_unlock(&parent_event
->child_mutex
);
5058 static int inherit_group(struct perf_event
*parent_event
,
5059 struct task_struct
*parent
,
5060 struct perf_event_context
*parent_ctx
,
5061 struct task_struct
*child
,
5062 struct perf_event_context
*child_ctx
)
5064 struct perf_event
*leader
;
5065 struct perf_event
*sub
;
5066 struct perf_event
*child_ctr
;
5068 leader
= inherit_event(parent_event
, parent
, parent_ctx
,
5069 child
, NULL
, child_ctx
);
5071 return PTR_ERR(leader
);
5072 list_for_each_entry(sub
, &parent_event
->sibling_list
, group_entry
) {
5073 child_ctr
= inherit_event(sub
, parent
, parent_ctx
,
5074 child
, leader
, child_ctx
);
5075 if (IS_ERR(child_ctr
))
5076 return PTR_ERR(child_ctr
);
5081 static void sync_child_event(struct perf_event
*child_event
,
5082 struct task_struct
*child
)
5084 struct perf_event
*parent_event
= child_event
->parent
;
5087 if (child_event
->attr
.inherit_stat
)
5088 perf_event_read_event(child_event
, child
);
5090 child_val
= atomic64_read(&child_event
->count
);
5093 * Add back the child's count to the parent's count:
5095 atomic64_add(child_val
, &parent_event
->count
);
5096 atomic64_add(child_event
->total_time_enabled
,
5097 &parent_event
->child_total_time_enabled
);
5098 atomic64_add(child_event
->total_time_running
,
5099 &parent_event
->child_total_time_running
);
5102 * Remove this event from the parent's list
5104 WARN_ON_ONCE(parent_event
->ctx
->parent_ctx
);
5105 mutex_lock(&parent_event
->child_mutex
);
5106 list_del_init(&child_event
->child_list
);
5107 mutex_unlock(&parent_event
->child_mutex
);
5110 * Release the parent event, if this was the last
5113 fput(parent_event
->filp
);
5117 __perf_event_exit_task(struct perf_event
*child_event
,
5118 struct perf_event_context
*child_ctx
,
5119 struct task_struct
*child
)
5121 struct perf_event
*parent_event
;
5123 perf_event_remove_from_context(child_event
);
5125 parent_event
= child_event
->parent
;
5127 * It can happen that parent exits first, and has events
5128 * that are still around due to the child reference. These
5129 * events need to be zapped - but otherwise linger.
5132 sync_child_event(child_event
, child
);
5133 free_event(child_event
);
5138 * When a child task exits, feed back event values to parent events.
5140 void perf_event_exit_task(struct task_struct
*child
)
5142 struct perf_event
*child_event
, *tmp
;
5143 struct perf_event_context
*child_ctx
;
5144 unsigned long flags
;
5146 if (likely(!child
->perf_event_ctxp
)) {
5147 perf_event_task(child
, NULL
, 0);
5151 local_irq_save(flags
);
5153 * We can't reschedule here because interrupts are disabled,
5154 * and either child is current or it is a task that can't be
5155 * scheduled, so we are now safe from rescheduling changing
5158 child_ctx
= child
->perf_event_ctxp
;
5159 __perf_event_task_sched_out(child_ctx
);
5162 * Take the context lock here so that if find_get_context is
5163 * reading child->perf_event_ctxp, we wait until it has
5164 * incremented the context's refcount before we do put_ctx below.
5166 raw_spin_lock(&child_ctx
->lock
);
5167 child
->perf_event_ctxp
= NULL
;
5169 * If this context is a clone; unclone it so it can't get
5170 * swapped to another process while we're removing all
5171 * the events from it.
5173 unclone_ctx(child_ctx
);
5174 update_context_time(child_ctx
);
5175 raw_spin_unlock_irqrestore(&child_ctx
->lock
, flags
);
5178 * Report the task dead after unscheduling the events so that we
5179 * won't get any samples after PERF_RECORD_EXIT. We can however still
5180 * get a few PERF_RECORD_READ events.
5182 perf_event_task(child
, child_ctx
, 0);
5185 * We can recurse on the same lock type through:
5187 * __perf_event_exit_task()
5188 * sync_child_event()
5189 * fput(parent_event->filp)
5191 * mutex_lock(&ctx->mutex)
5193 * But since its the parent context it won't be the same instance.
5195 mutex_lock_nested(&child_ctx
->mutex
, SINGLE_DEPTH_NESTING
);
5198 list_for_each_entry_safe(child_event
, tmp
, &child_ctx
->pinned_groups
,
5200 __perf_event_exit_task(child_event
, child_ctx
, child
);
5202 list_for_each_entry_safe(child_event
, tmp
, &child_ctx
->flexible_groups
,
5204 __perf_event_exit_task(child_event
, child_ctx
, child
);
5207 * If the last event was a group event, it will have appended all
5208 * its siblings to the list, but we obtained 'tmp' before that which
5209 * will still point to the list head terminating the iteration.
5211 if (!list_empty(&child_ctx
->pinned_groups
) ||
5212 !list_empty(&child_ctx
->flexible_groups
))
5215 mutex_unlock(&child_ctx
->mutex
);
5220 static void perf_free_event(struct perf_event
*event
,
5221 struct perf_event_context
*ctx
)
5223 struct perf_event
*parent
= event
->parent
;
5225 if (WARN_ON_ONCE(!parent
))
5228 mutex_lock(&parent
->child_mutex
);
5229 list_del_init(&event
->child_list
);
5230 mutex_unlock(&parent
->child_mutex
);
5234 list_del_event(event
, ctx
);
5239 * free an unexposed, unused context as created by inheritance by
5240 * init_task below, used by fork() in case of fail.
5242 void perf_event_free_task(struct task_struct
*task
)
5244 struct perf_event_context
*ctx
= task
->perf_event_ctxp
;
5245 struct perf_event
*event
, *tmp
;
5250 mutex_lock(&ctx
->mutex
);
5252 list_for_each_entry_safe(event
, tmp
, &ctx
->pinned_groups
, group_entry
)
5253 perf_free_event(event
, ctx
);
5255 list_for_each_entry_safe(event
, tmp
, &ctx
->flexible_groups
,
5257 perf_free_event(event
, ctx
);
5259 if (!list_empty(&ctx
->pinned_groups
) ||
5260 !list_empty(&ctx
->flexible_groups
))
5263 mutex_unlock(&ctx
->mutex
);
5269 inherit_task_group(struct perf_event
*event
, struct task_struct
*parent
,
5270 struct perf_event_context
*parent_ctx
,
5271 struct task_struct
*child
,
5275 struct perf_event_context
*child_ctx
= child
->perf_event_ctxp
;
5277 if (!event
->attr
.inherit
) {
5284 * This is executed from the parent task context, so
5285 * inherit events that have been marked for cloning.
5286 * First allocate and initialize a context for the
5290 child_ctx
= kzalloc(sizeof(struct perf_event_context
),
5295 __perf_event_init_context(child_ctx
, child
);
5296 child
->perf_event_ctxp
= child_ctx
;
5297 get_task_struct(child
);
5300 ret
= inherit_group(event
, parent
, parent_ctx
,
5311 * Initialize the perf_event context in task_struct
5313 int perf_event_init_task(struct task_struct
*child
)
5315 struct perf_event_context
*child_ctx
, *parent_ctx
;
5316 struct perf_event_context
*cloned_ctx
;
5317 struct perf_event
*event
;
5318 struct task_struct
*parent
= current
;
5319 int inherited_all
= 1;
5322 child
->perf_event_ctxp
= NULL
;
5324 mutex_init(&child
->perf_event_mutex
);
5325 INIT_LIST_HEAD(&child
->perf_event_list
);
5327 if (likely(!parent
->perf_event_ctxp
))
5331 * If the parent's context is a clone, pin it so it won't get
5334 parent_ctx
= perf_pin_task_context(parent
);
5337 * No need to check if parent_ctx != NULL here; since we saw
5338 * it non-NULL earlier, the only reason for it to become NULL
5339 * is if we exit, and since we're currently in the middle of
5340 * a fork we can't be exiting at the same time.
5344 * Lock the parent list. No need to lock the child - not PID
5345 * hashed yet and not running, so nobody can access it.
5347 mutex_lock(&parent_ctx
->mutex
);
5350 * We dont have to disable NMIs - we are only looking at
5351 * the list, not manipulating it:
5353 list_for_each_entry(event
, &parent_ctx
->pinned_groups
, group_entry
) {
5354 ret
= inherit_task_group(event
, parent
, parent_ctx
, child
,
5360 list_for_each_entry(event
, &parent_ctx
->flexible_groups
, group_entry
) {
5361 ret
= inherit_task_group(event
, parent
, parent_ctx
, child
,
5367 child_ctx
= child
->perf_event_ctxp
;
5369 if (child_ctx
&& inherited_all
) {
5371 * Mark the child context as a clone of the parent
5372 * context, or of whatever the parent is a clone of.
5373 * Note that if the parent is a clone, it could get
5374 * uncloned at any point, but that doesn't matter
5375 * because the list of events and the generation
5376 * count can't have changed since we took the mutex.
5378 cloned_ctx
= rcu_dereference(parent_ctx
->parent_ctx
);
5380 child_ctx
->parent_ctx
= cloned_ctx
;
5381 child_ctx
->parent_gen
= parent_ctx
->parent_gen
;
5383 child_ctx
->parent_ctx
= parent_ctx
;
5384 child_ctx
->parent_gen
= parent_ctx
->generation
;
5386 get_ctx(child_ctx
->parent_ctx
);
5389 mutex_unlock(&parent_ctx
->mutex
);
5391 perf_unpin_context(parent_ctx
);
5396 static void __cpuinit
perf_event_init_cpu(int cpu
)
5398 struct perf_cpu_context
*cpuctx
;
5400 cpuctx
= &per_cpu(perf_cpu_context
, cpu
);
5401 __perf_event_init_context(&cpuctx
->ctx
, NULL
);
5403 spin_lock(&perf_resource_lock
);
5404 cpuctx
->max_pertask
= perf_max_events
- perf_reserved_percpu
;
5405 spin_unlock(&perf_resource_lock
);
5407 hw_perf_event_setup(cpu
);
5410 #ifdef CONFIG_HOTPLUG_CPU
5411 static void __perf_event_exit_cpu(void *info
)
5413 struct perf_cpu_context
*cpuctx
= &__get_cpu_var(perf_cpu_context
);
5414 struct perf_event_context
*ctx
= &cpuctx
->ctx
;
5415 struct perf_event
*event
, *tmp
;
5417 list_for_each_entry_safe(event
, tmp
, &ctx
->pinned_groups
, group_entry
)
5418 __perf_event_remove_from_context(event
);
5419 list_for_each_entry_safe(event
, tmp
, &ctx
->flexible_groups
, group_entry
)
5420 __perf_event_remove_from_context(event
);
5422 static void perf_event_exit_cpu(int cpu
)
5424 struct perf_cpu_context
*cpuctx
= &per_cpu(perf_cpu_context
, cpu
);
5425 struct perf_event_context
*ctx
= &cpuctx
->ctx
;
5427 mutex_lock(&ctx
->mutex
);
5428 smp_call_function_single(cpu
, __perf_event_exit_cpu
, NULL
, 1);
5429 mutex_unlock(&ctx
->mutex
);
5432 static inline void perf_event_exit_cpu(int cpu
) { }
5435 static int __cpuinit
5436 perf_cpu_notify(struct notifier_block
*self
, unsigned long action
, void *hcpu
)
5438 unsigned int cpu
= (long)hcpu
;
5442 case CPU_UP_PREPARE
:
5443 case CPU_UP_PREPARE_FROZEN
:
5444 perf_event_init_cpu(cpu
);
5448 case CPU_ONLINE_FROZEN
:
5449 hw_perf_event_setup_online(cpu
);
5452 case CPU_DOWN_PREPARE
:
5453 case CPU_DOWN_PREPARE_FROZEN
:
5454 perf_event_exit_cpu(cpu
);
5458 hw_perf_event_setup_offline(cpu
);
5469 * This has to have a higher priority than migration_notifier in sched.c.
5471 static struct notifier_block __cpuinitdata perf_cpu_nb
= {
5472 .notifier_call
= perf_cpu_notify
,
5476 void __init
perf_event_init(void)
5478 perf_cpu_notify(&perf_cpu_nb
, (unsigned long)CPU_UP_PREPARE
,
5479 (void *)(long)smp_processor_id());
5480 perf_cpu_notify(&perf_cpu_nb
, (unsigned long)CPU_ONLINE
,
5481 (void *)(long)smp_processor_id());
5482 register_cpu_notifier(&perf_cpu_nb
);
5485 static ssize_t
perf_show_reserve_percpu(struct sysdev_class
*class, char *buf
)
5487 return sprintf(buf
, "%d\n", perf_reserved_percpu
);
5491 perf_set_reserve_percpu(struct sysdev_class
*class,
5495 struct perf_cpu_context
*cpuctx
;
5499 err
= strict_strtoul(buf
, 10, &val
);
5502 if (val
> perf_max_events
)
5505 spin_lock(&perf_resource_lock
);
5506 perf_reserved_percpu
= val
;
5507 for_each_online_cpu(cpu
) {
5508 cpuctx
= &per_cpu(perf_cpu_context
, cpu
);
5509 raw_spin_lock_irq(&cpuctx
->ctx
.lock
);
5510 mpt
= min(perf_max_events
- cpuctx
->ctx
.nr_events
,
5511 perf_max_events
- perf_reserved_percpu
);
5512 cpuctx
->max_pertask
= mpt
;
5513 raw_spin_unlock_irq(&cpuctx
->ctx
.lock
);
5515 spin_unlock(&perf_resource_lock
);
5520 static ssize_t
perf_show_overcommit(struct sysdev_class
*class, char *buf
)
5522 return sprintf(buf
, "%d\n", perf_overcommit
);
5526 perf_set_overcommit(struct sysdev_class
*class, const char *buf
, size_t count
)
5531 err
= strict_strtoul(buf
, 10, &val
);
5537 spin_lock(&perf_resource_lock
);
5538 perf_overcommit
= val
;
5539 spin_unlock(&perf_resource_lock
);
5544 static SYSDEV_CLASS_ATTR(
5547 perf_show_reserve_percpu
,
5548 perf_set_reserve_percpu
5551 static SYSDEV_CLASS_ATTR(
5554 perf_show_overcommit
,
5558 static struct attribute
*perfclass_attrs
[] = {
5559 &attr_reserve_percpu
.attr
,
5560 &attr_overcommit
.attr
,
5564 static struct attribute_group perfclass_attr_group
= {
5565 .attrs
= perfclass_attrs
,
5566 .name
= "perf_events",
5569 static int __init
perf_event_sysfs_init(void)
5571 return sysfs_create_group(&cpu_sysdev_class
.kset
.kobj
,
5572 &perfclass_attr_group
);
5574 device_initcall(perf_event_sysfs_init
);