2 * Performance events core code:
4 * Copyright (C) 2008 Thomas Gleixner <tglx@linutronix.de>
5 * Copyright (C) 2008-2009 Red Hat, Inc., Ingo Molnar
6 * Copyright (C) 2008-2009 Red Hat, Inc., Peter Zijlstra <pzijlstr@redhat.com>
7 * Copyright © 2009 Paul Mackerras, IBM Corp. <paulus@au1.ibm.com>
9 * For licensing details see kernel-base/COPYING
14 #include <linux/cpu.h>
15 #include <linux/smp.h>
16 #include <linux/file.h>
17 #include <linux/poll.h>
18 #include <linux/slab.h>
19 #include <linux/hash.h>
20 #include <linux/sysfs.h>
21 #include <linux/dcache.h>
22 #include <linux/percpu.h>
23 #include <linux/ptrace.h>
24 #include <linux/vmstat.h>
25 #include <linux/vmalloc.h>
26 #include <linux/hardirq.h>
27 #include <linux/rculist.h>
28 #include <linux/uaccess.h>
29 #include <linux/syscalls.h>
30 #include <linux/anon_inodes.h>
31 #include <linux/kernel_stat.h>
32 #include <linux/perf_event.h>
33 #include <linux/ftrace_event.h>
34 #include <linux/hw_breakpoint.h>
36 #include <asm/irq_regs.h>
39 * Each CPU has a list of per CPU events:
41 static DEFINE_PER_CPU(struct perf_cpu_context
, perf_cpu_context
);
43 int perf_max_events __read_mostly
= 1;
44 static int perf_reserved_percpu __read_mostly
;
45 static int perf_overcommit __read_mostly
= 1;
47 static atomic_t nr_events __read_mostly
;
48 static atomic_t nr_mmap_events __read_mostly
;
49 static atomic_t nr_comm_events __read_mostly
;
50 static atomic_t nr_task_events __read_mostly
;
53 * perf event paranoia level:
54 * -1 - not paranoid at all
55 * 0 - disallow raw tracepoint access for unpriv
56 * 1 - disallow cpu events for unpriv
57 * 2 - disallow kernel profiling for unpriv
59 int sysctl_perf_event_paranoid __read_mostly
= 1;
61 int sysctl_perf_event_mlock __read_mostly
= 512; /* 'free' kb per user */
64 * max perf event sample rate
66 int sysctl_perf_event_sample_rate __read_mostly
= 100000;
68 static atomic64_t perf_event_id
;
71 * Lock for (sysadmin-configurable) event reservations:
73 static DEFINE_SPINLOCK(perf_resource_lock
);
76 * Architecture provided APIs - weak aliases:
78 extern __weak
const struct pmu
*hw_perf_event_init(struct perf_event
*event
)
83 void __weak
hw_perf_disable(void) { barrier(); }
84 void __weak
hw_perf_enable(void) { barrier(); }
86 void __weak
perf_event_print_debug(void) { }
88 static DEFINE_PER_CPU(int, perf_disable_count
);
90 void perf_disable(void)
92 if (!__get_cpu_var(perf_disable_count
)++)
96 void perf_enable(void)
98 if (!--__get_cpu_var(perf_disable_count
))
102 static void get_ctx(struct perf_event_context
*ctx
)
104 WARN_ON(!atomic_inc_not_zero(&ctx
->refcount
));
107 static void free_ctx(struct rcu_head
*head
)
109 struct perf_event_context
*ctx
;
111 ctx
= container_of(head
, struct perf_event_context
, rcu_head
);
115 static void put_ctx(struct perf_event_context
*ctx
)
117 if (atomic_dec_and_test(&ctx
->refcount
)) {
119 put_ctx(ctx
->parent_ctx
);
121 put_task_struct(ctx
->task
);
122 call_rcu(&ctx
->rcu_head
, free_ctx
);
126 static void unclone_ctx(struct perf_event_context
*ctx
)
128 if (ctx
->parent_ctx
) {
129 put_ctx(ctx
->parent_ctx
);
130 ctx
->parent_ctx
= NULL
;
135 * If we inherit events we want to return the parent event id
138 static u64
primary_event_id(struct perf_event
*event
)
143 id
= event
->parent
->id
;
149 * Get the perf_event_context for a task and lock it.
150 * This has to cope with with the fact that until it is locked,
151 * the context could get moved to another task.
153 static struct perf_event_context
*
154 perf_lock_task_context(struct task_struct
*task
, unsigned long *flags
)
156 struct perf_event_context
*ctx
;
160 ctx
= rcu_dereference(task
->perf_event_ctxp
);
163 * If this context is a clone of another, it might
164 * get swapped for another underneath us by
165 * perf_event_task_sched_out, though the
166 * rcu_read_lock() protects us from any context
167 * getting freed. Lock the context and check if it
168 * got swapped before we could get the lock, and retry
169 * if so. If we locked the right context, then it
170 * can't get swapped on us any more.
172 raw_spin_lock_irqsave(&ctx
->lock
, *flags
);
173 if (ctx
!= rcu_dereference(task
->perf_event_ctxp
)) {
174 raw_spin_unlock_irqrestore(&ctx
->lock
, *flags
);
178 if (!atomic_inc_not_zero(&ctx
->refcount
)) {
179 raw_spin_unlock_irqrestore(&ctx
->lock
, *flags
);
188 * Get the context for a task and increment its pin_count so it
189 * can't get swapped to another task. This also increments its
190 * reference count so that the context can't get freed.
192 static struct perf_event_context
*perf_pin_task_context(struct task_struct
*task
)
194 struct perf_event_context
*ctx
;
197 ctx
= perf_lock_task_context(task
, &flags
);
200 raw_spin_unlock_irqrestore(&ctx
->lock
, flags
);
205 static void perf_unpin_context(struct perf_event_context
*ctx
)
209 raw_spin_lock_irqsave(&ctx
->lock
, flags
);
211 raw_spin_unlock_irqrestore(&ctx
->lock
, flags
);
215 static inline u64
perf_clock(void)
217 return local_clock();
221 * Update the record of the current time in a context.
223 static void update_context_time(struct perf_event_context
*ctx
)
225 u64 now
= perf_clock();
227 ctx
->time
+= now
- ctx
->timestamp
;
228 ctx
->timestamp
= now
;
232 * Update the total_time_enabled and total_time_running fields for a event.
234 static void update_event_times(struct perf_event
*event
)
236 struct perf_event_context
*ctx
= event
->ctx
;
239 if (event
->state
< PERF_EVENT_STATE_INACTIVE
||
240 event
->group_leader
->state
< PERF_EVENT_STATE_INACTIVE
)
246 run_end
= event
->tstamp_stopped
;
248 event
->total_time_enabled
= run_end
- event
->tstamp_enabled
;
250 if (event
->state
== PERF_EVENT_STATE_INACTIVE
)
251 run_end
= event
->tstamp_stopped
;
255 event
->total_time_running
= run_end
- event
->tstamp_running
;
259 * Update total_time_enabled and total_time_running for all events in a group.
261 static void update_group_times(struct perf_event
*leader
)
263 struct perf_event
*event
;
265 update_event_times(leader
);
266 list_for_each_entry(event
, &leader
->sibling_list
, group_entry
)
267 update_event_times(event
);
270 static struct list_head
*
271 ctx_group_list(struct perf_event
*event
, struct perf_event_context
*ctx
)
273 if (event
->attr
.pinned
)
274 return &ctx
->pinned_groups
;
276 return &ctx
->flexible_groups
;
280 * Add a event from the lists for its context.
281 * Must be called with ctx->mutex and ctx->lock held.
284 list_add_event(struct perf_event
*event
, struct perf_event_context
*ctx
)
286 WARN_ON_ONCE(event
->attach_state
& PERF_ATTACH_CONTEXT
);
287 event
->attach_state
|= PERF_ATTACH_CONTEXT
;
290 * If we're a stand alone event or group leader, we go to the context
291 * list, group events are kept attached to the group so that
292 * perf_group_detach can, at all times, locate all siblings.
294 if (event
->group_leader
== event
) {
295 struct list_head
*list
;
297 if (is_software_event(event
))
298 event
->group_flags
|= PERF_GROUP_SOFTWARE
;
300 list
= ctx_group_list(event
, ctx
);
301 list_add_tail(&event
->group_entry
, list
);
304 list_add_rcu(&event
->event_entry
, &ctx
->event_list
);
306 if (event
->attr
.inherit_stat
)
310 static void perf_group_attach(struct perf_event
*event
)
312 struct perf_event
*group_leader
= event
->group_leader
;
314 WARN_ON_ONCE(event
->attach_state
& PERF_ATTACH_GROUP
);
315 event
->attach_state
|= PERF_ATTACH_GROUP
;
317 if (group_leader
== event
)
320 if (group_leader
->group_flags
& PERF_GROUP_SOFTWARE
&&
321 !is_software_event(event
))
322 group_leader
->group_flags
&= ~PERF_GROUP_SOFTWARE
;
324 list_add_tail(&event
->group_entry
, &group_leader
->sibling_list
);
325 group_leader
->nr_siblings
++;
329 * Remove a event from the lists for its context.
330 * Must be called with ctx->mutex and ctx->lock held.
333 list_del_event(struct perf_event
*event
, struct perf_event_context
*ctx
)
336 * We can have double detach due to exit/hot-unplug + close.
338 if (!(event
->attach_state
& PERF_ATTACH_CONTEXT
))
341 event
->attach_state
&= ~PERF_ATTACH_CONTEXT
;
344 if (event
->attr
.inherit_stat
)
347 list_del_rcu(&event
->event_entry
);
349 if (event
->group_leader
== event
)
350 list_del_init(&event
->group_entry
);
352 update_group_times(event
);
355 * If event was in error state, then keep it
356 * that way, otherwise bogus counts will be
357 * returned on read(). The only way to get out
358 * of error state is by explicit re-enabling
361 if (event
->state
> PERF_EVENT_STATE_OFF
)
362 event
->state
= PERF_EVENT_STATE_OFF
;
365 static void perf_group_detach(struct perf_event
*event
)
367 struct perf_event
*sibling
, *tmp
;
368 struct list_head
*list
= NULL
;
371 * We can have double detach due to exit/hot-unplug + close.
373 if (!(event
->attach_state
& PERF_ATTACH_GROUP
))
376 event
->attach_state
&= ~PERF_ATTACH_GROUP
;
379 * If this is a sibling, remove it from its group.
381 if (event
->group_leader
!= event
) {
382 list_del_init(&event
->group_entry
);
383 event
->group_leader
->nr_siblings
--;
387 if (!list_empty(&event
->group_entry
))
388 list
= &event
->group_entry
;
391 * If this was a group event with sibling events then
392 * upgrade the siblings to singleton events by adding them
393 * to whatever list we are on.
395 list_for_each_entry_safe(sibling
, tmp
, &event
->sibling_list
, group_entry
) {
397 list_move_tail(&sibling
->group_entry
, list
);
398 sibling
->group_leader
= sibling
;
400 /* Inherit group flags from the previous leader */
401 sibling
->group_flags
= event
->group_flags
;
406 event_filter_match(struct perf_event
*event
)
408 return event
->cpu
== -1 || event
->cpu
== smp_processor_id();
412 event_sched_out(struct perf_event
*event
,
413 struct perf_cpu_context
*cpuctx
,
414 struct perf_event_context
*ctx
)
418 * An event which could not be activated because of
419 * filter mismatch still needs to have its timings
420 * maintained, otherwise bogus information is return
421 * via read() for time_enabled, time_running:
423 if (event
->state
== PERF_EVENT_STATE_INACTIVE
424 && !event_filter_match(event
)) {
425 delta
= ctx
->time
- event
->tstamp_stopped
;
426 event
->tstamp_running
+= delta
;
427 event
->tstamp_stopped
= ctx
->time
;
430 if (event
->state
!= PERF_EVENT_STATE_ACTIVE
)
433 event
->state
= PERF_EVENT_STATE_INACTIVE
;
434 if (event
->pending_disable
) {
435 event
->pending_disable
= 0;
436 event
->state
= PERF_EVENT_STATE_OFF
;
438 event
->tstamp_stopped
= ctx
->time
;
439 event
->pmu
->disable(event
);
442 if (!is_software_event(event
))
443 cpuctx
->active_oncpu
--;
445 if (event
->attr
.exclusive
|| !cpuctx
->active_oncpu
)
446 cpuctx
->exclusive
= 0;
450 group_sched_out(struct perf_event
*group_event
,
451 struct perf_cpu_context
*cpuctx
,
452 struct perf_event_context
*ctx
)
454 struct perf_event
*event
;
455 int state
= group_event
->state
;
457 event_sched_out(group_event
, cpuctx
, ctx
);
460 * Schedule out siblings (if any):
462 list_for_each_entry(event
, &group_event
->sibling_list
, group_entry
)
463 event_sched_out(event
, cpuctx
, ctx
);
465 if (state
== PERF_EVENT_STATE_ACTIVE
&& group_event
->attr
.exclusive
)
466 cpuctx
->exclusive
= 0;
470 * Cross CPU call to remove a performance event
472 * We disable the event on the hardware level first. After that we
473 * remove it from the context list.
475 static void __perf_event_remove_from_context(void *info
)
477 struct perf_cpu_context
*cpuctx
= &__get_cpu_var(perf_cpu_context
);
478 struct perf_event
*event
= info
;
479 struct perf_event_context
*ctx
= event
->ctx
;
482 * If this is a task context, we need to check whether it is
483 * the current task context of this cpu. If not it has been
484 * scheduled out before the smp call arrived.
486 if (ctx
->task
&& cpuctx
->task_ctx
!= ctx
)
489 raw_spin_lock(&ctx
->lock
);
491 * Protect the list operation against NMI by disabling the
492 * events on a global level.
496 event_sched_out(event
, cpuctx
, ctx
);
498 list_del_event(event
, ctx
);
502 * Allow more per task events with respect to the
505 cpuctx
->max_pertask
=
506 min(perf_max_events
- ctx
->nr_events
,
507 perf_max_events
- perf_reserved_percpu
);
511 raw_spin_unlock(&ctx
->lock
);
516 * Remove the event from a task's (or a CPU's) list of events.
518 * Must be called with ctx->mutex held.
520 * CPU events are removed with a smp call. For task events we only
521 * call when the task is on a CPU.
523 * If event->ctx is a cloned context, callers must make sure that
524 * every task struct that event->ctx->task could possibly point to
525 * remains valid. This is OK when called from perf_release since
526 * that only calls us on the top-level context, which can't be a clone.
527 * When called from perf_event_exit_task, it's OK because the
528 * context has been detached from its task.
530 static void perf_event_remove_from_context(struct perf_event
*event
)
532 struct perf_event_context
*ctx
= event
->ctx
;
533 struct task_struct
*task
= ctx
->task
;
537 * Per cpu events are removed via an smp call and
538 * the removal is always successful.
540 smp_call_function_single(event
->cpu
,
541 __perf_event_remove_from_context
,
547 task_oncpu_function_call(task
, __perf_event_remove_from_context
,
550 raw_spin_lock_irq(&ctx
->lock
);
552 * If the context is active we need to retry the smp call.
554 if (ctx
->nr_active
&& !list_empty(&event
->group_entry
)) {
555 raw_spin_unlock_irq(&ctx
->lock
);
560 * The lock prevents that this context is scheduled in so we
561 * can remove the event safely, if the call above did not
564 if (!list_empty(&event
->group_entry
))
565 list_del_event(event
, ctx
);
566 raw_spin_unlock_irq(&ctx
->lock
);
570 * Cross CPU call to disable a performance event
572 static void __perf_event_disable(void *info
)
574 struct perf_event
*event
= info
;
575 struct perf_cpu_context
*cpuctx
= &__get_cpu_var(perf_cpu_context
);
576 struct perf_event_context
*ctx
= event
->ctx
;
579 * If this is a per-task event, need to check whether this
580 * event's task is the current task on this cpu.
582 if (ctx
->task
&& cpuctx
->task_ctx
!= ctx
)
585 raw_spin_lock(&ctx
->lock
);
588 * If the event is on, turn it off.
589 * If it is in error state, leave it in error state.
591 if (event
->state
>= PERF_EVENT_STATE_INACTIVE
) {
592 update_context_time(ctx
);
593 update_group_times(event
);
594 if (event
== event
->group_leader
)
595 group_sched_out(event
, cpuctx
, ctx
);
597 event_sched_out(event
, cpuctx
, ctx
);
598 event
->state
= PERF_EVENT_STATE_OFF
;
601 raw_spin_unlock(&ctx
->lock
);
607 * If event->ctx is a cloned context, callers must make sure that
608 * every task struct that event->ctx->task could possibly point to
609 * remains valid. This condition is satisifed when called through
610 * perf_event_for_each_child or perf_event_for_each because they
611 * hold the top-level event's child_mutex, so any descendant that
612 * goes to exit will block in sync_child_event.
613 * When called from perf_pending_event it's OK because event->ctx
614 * is the current context on this CPU and preemption is disabled,
615 * hence we can't get into perf_event_task_sched_out for this context.
617 void perf_event_disable(struct perf_event
*event
)
619 struct perf_event_context
*ctx
= event
->ctx
;
620 struct task_struct
*task
= ctx
->task
;
624 * Disable the event on the cpu that it's on
626 smp_call_function_single(event
->cpu
, __perf_event_disable
,
632 task_oncpu_function_call(task
, __perf_event_disable
, event
);
634 raw_spin_lock_irq(&ctx
->lock
);
636 * If the event is still active, we need to retry the cross-call.
638 if (event
->state
== PERF_EVENT_STATE_ACTIVE
) {
639 raw_spin_unlock_irq(&ctx
->lock
);
644 * Since we have the lock this context can't be scheduled
645 * in, so we can change the state safely.
647 if (event
->state
== PERF_EVENT_STATE_INACTIVE
) {
648 update_group_times(event
);
649 event
->state
= PERF_EVENT_STATE_OFF
;
652 raw_spin_unlock_irq(&ctx
->lock
);
656 event_sched_in(struct perf_event
*event
,
657 struct perf_cpu_context
*cpuctx
,
658 struct perf_event_context
*ctx
)
660 if (event
->state
<= PERF_EVENT_STATE_OFF
)
663 event
->state
= PERF_EVENT_STATE_ACTIVE
;
664 event
->oncpu
= smp_processor_id();
666 * The new state must be visible before we turn it on in the hardware:
670 if (event
->pmu
->enable(event
)) {
671 event
->state
= PERF_EVENT_STATE_INACTIVE
;
676 event
->tstamp_running
+= ctx
->time
- event
->tstamp_stopped
;
678 if (!is_software_event(event
))
679 cpuctx
->active_oncpu
++;
682 if (event
->attr
.exclusive
)
683 cpuctx
->exclusive
= 1;
689 group_sched_in(struct perf_event
*group_event
,
690 struct perf_cpu_context
*cpuctx
,
691 struct perf_event_context
*ctx
)
693 struct perf_event
*event
, *partial_group
= NULL
;
694 const struct pmu
*pmu
= group_event
->pmu
;
697 if (group_event
->state
== PERF_EVENT_STATE_OFF
)
700 /* Check if group transaction availabe */
707 if (event_sched_in(group_event
, cpuctx
, ctx
)) {
709 pmu
->cancel_txn(pmu
);
714 * Schedule in siblings as one group (if any):
716 list_for_each_entry(event
, &group_event
->sibling_list
, group_entry
) {
717 if (event_sched_in(event
, cpuctx
, ctx
)) {
718 partial_group
= event
;
723 if (!txn
|| !pmu
->commit_txn(pmu
))
728 * Groups can be scheduled in as one unit only, so undo any
729 * partial group before returning:
731 list_for_each_entry(event
, &group_event
->sibling_list
, group_entry
) {
732 if (event
== partial_group
)
734 event_sched_out(event
, cpuctx
, ctx
);
736 event_sched_out(group_event
, cpuctx
, ctx
);
739 pmu
->cancel_txn(pmu
);
745 * Work out whether we can put this event group on the CPU now.
747 static int group_can_go_on(struct perf_event
*event
,
748 struct perf_cpu_context
*cpuctx
,
752 * Groups consisting entirely of software events can always go on.
754 if (event
->group_flags
& PERF_GROUP_SOFTWARE
)
757 * If an exclusive group is already on, no other hardware
760 if (cpuctx
->exclusive
)
763 * If this group is exclusive and there are already
764 * events on the CPU, it can't go on.
766 if (event
->attr
.exclusive
&& cpuctx
->active_oncpu
)
769 * Otherwise, try to add it if all previous groups were able
775 static void add_event_to_ctx(struct perf_event
*event
,
776 struct perf_event_context
*ctx
)
778 list_add_event(event
, ctx
);
779 perf_group_attach(event
);
780 event
->tstamp_enabled
= ctx
->time
;
781 event
->tstamp_running
= ctx
->time
;
782 event
->tstamp_stopped
= ctx
->time
;
786 * Cross CPU call to install and enable a performance event
788 * Must be called with ctx->mutex held
790 static void __perf_install_in_context(void *info
)
792 struct perf_cpu_context
*cpuctx
= &__get_cpu_var(perf_cpu_context
);
793 struct perf_event
*event
= info
;
794 struct perf_event_context
*ctx
= event
->ctx
;
795 struct perf_event
*leader
= event
->group_leader
;
799 * If this is a task context, we need to check whether it is
800 * the current task context of this cpu. If not it has been
801 * scheduled out before the smp call arrived.
802 * Or possibly this is the right context but it isn't
803 * on this cpu because it had no events.
805 if (ctx
->task
&& cpuctx
->task_ctx
!= ctx
) {
806 if (cpuctx
->task_ctx
|| ctx
->task
!= current
)
808 cpuctx
->task_ctx
= ctx
;
811 raw_spin_lock(&ctx
->lock
);
813 update_context_time(ctx
);
816 * Protect the list operation against NMI by disabling the
817 * events on a global level. NOP for non NMI based events.
821 add_event_to_ctx(event
, ctx
);
823 if (event
->cpu
!= -1 && event
->cpu
!= smp_processor_id())
827 * Don't put the event on if it is disabled or if
828 * it is in a group and the group isn't on.
830 if (event
->state
!= PERF_EVENT_STATE_INACTIVE
||
831 (leader
!= event
&& leader
->state
!= PERF_EVENT_STATE_ACTIVE
))
835 * An exclusive event can't go on if there are already active
836 * hardware events, and no hardware event can go on if there
837 * is already an exclusive event on.
839 if (!group_can_go_on(event
, cpuctx
, 1))
842 err
= event_sched_in(event
, cpuctx
, ctx
);
846 * This event couldn't go on. If it is in a group
847 * then we have to pull the whole group off.
848 * If the event group is pinned then put it in error state.
851 group_sched_out(leader
, cpuctx
, ctx
);
852 if (leader
->attr
.pinned
) {
853 update_group_times(leader
);
854 leader
->state
= PERF_EVENT_STATE_ERROR
;
858 if (!err
&& !ctx
->task
&& cpuctx
->max_pertask
)
859 cpuctx
->max_pertask
--;
864 raw_spin_unlock(&ctx
->lock
);
868 * Attach a performance event to a context
870 * First we add the event to the list with the hardware enable bit
871 * in event->hw_config cleared.
873 * If the event is attached to a task which is on a CPU we use a smp
874 * call to enable it in the task context. The task might have been
875 * scheduled away, but we check this in the smp call again.
877 * Must be called with ctx->mutex held.
880 perf_install_in_context(struct perf_event_context
*ctx
,
881 struct perf_event
*event
,
884 struct task_struct
*task
= ctx
->task
;
888 * Per cpu events are installed via an smp call and
889 * the install is always successful.
891 smp_call_function_single(cpu
, __perf_install_in_context
,
897 task_oncpu_function_call(task
, __perf_install_in_context
,
900 raw_spin_lock_irq(&ctx
->lock
);
902 * we need to retry the smp call.
904 if (ctx
->is_active
&& list_empty(&event
->group_entry
)) {
905 raw_spin_unlock_irq(&ctx
->lock
);
910 * The lock prevents that this context is scheduled in so we
911 * can add the event safely, if it the call above did not
914 if (list_empty(&event
->group_entry
))
915 add_event_to_ctx(event
, ctx
);
916 raw_spin_unlock_irq(&ctx
->lock
);
920 * Put a event into inactive state and update time fields.
921 * Enabling the leader of a group effectively enables all
922 * the group members that aren't explicitly disabled, so we
923 * have to update their ->tstamp_enabled also.
924 * Note: this works for group members as well as group leaders
925 * since the non-leader members' sibling_lists will be empty.
927 static void __perf_event_mark_enabled(struct perf_event
*event
,
928 struct perf_event_context
*ctx
)
930 struct perf_event
*sub
;
932 event
->state
= PERF_EVENT_STATE_INACTIVE
;
933 event
->tstamp_enabled
= ctx
->time
- event
->total_time_enabled
;
934 list_for_each_entry(sub
, &event
->sibling_list
, group_entry
)
935 if (sub
->state
>= PERF_EVENT_STATE_INACTIVE
)
936 sub
->tstamp_enabled
=
937 ctx
->time
- sub
->total_time_enabled
;
941 * Cross CPU call to enable a performance event
943 static void __perf_event_enable(void *info
)
945 struct perf_event
*event
= info
;
946 struct perf_cpu_context
*cpuctx
= &__get_cpu_var(perf_cpu_context
);
947 struct perf_event_context
*ctx
= event
->ctx
;
948 struct perf_event
*leader
= event
->group_leader
;
952 * If this is a per-task event, need to check whether this
953 * event's task is the current task on this cpu.
955 if (ctx
->task
&& cpuctx
->task_ctx
!= ctx
) {
956 if (cpuctx
->task_ctx
|| ctx
->task
!= current
)
958 cpuctx
->task_ctx
= ctx
;
961 raw_spin_lock(&ctx
->lock
);
963 update_context_time(ctx
);
965 if (event
->state
>= PERF_EVENT_STATE_INACTIVE
)
967 __perf_event_mark_enabled(event
, ctx
);
969 if (event
->cpu
!= -1 && event
->cpu
!= smp_processor_id())
973 * If the event is in a group and isn't the group leader,
974 * then don't put it on unless the group is on.
976 if (leader
!= event
&& leader
->state
!= PERF_EVENT_STATE_ACTIVE
)
979 if (!group_can_go_on(event
, cpuctx
, 1)) {
984 err
= group_sched_in(event
, cpuctx
, ctx
);
986 err
= event_sched_in(event
, cpuctx
, ctx
);
992 * If this event can't go on and it's part of a
993 * group, then the whole group has to come off.
996 group_sched_out(leader
, cpuctx
, ctx
);
997 if (leader
->attr
.pinned
) {
998 update_group_times(leader
);
999 leader
->state
= PERF_EVENT_STATE_ERROR
;
1004 raw_spin_unlock(&ctx
->lock
);
1010 * If event->ctx is a cloned context, callers must make sure that
1011 * every task struct that event->ctx->task could possibly point to
1012 * remains valid. This condition is satisfied when called through
1013 * perf_event_for_each_child or perf_event_for_each as described
1014 * for perf_event_disable.
1016 void perf_event_enable(struct perf_event
*event
)
1018 struct perf_event_context
*ctx
= event
->ctx
;
1019 struct task_struct
*task
= ctx
->task
;
1023 * Enable the event on the cpu that it's on
1025 smp_call_function_single(event
->cpu
, __perf_event_enable
,
1030 raw_spin_lock_irq(&ctx
->lock
);
1031 if (event
->state
>= PERF_EVENT_STATE_INACTIVE
)
1035 * If the event is in error state, clear that first.
1036 * That way, if we see the event in error state below, we
1037 * know that it has gone back into error state, as distinct
1038 * from the task having been scheduled away before the
1039 * cross-call arrived.
1041 if (event
->state
== PERF_EVENT_STATE_ERROR
)
1042 event
->state
= PERF_EVENT_STATE_OFF
;
1045 raw_spin_unlock_irq(&ctx
->lock
);
1046 task_oncpu_function_call(task
, __perf_event_enable
, event
);
1048 raw_spin_lock_irq(&ctx
->lock
);
1051 * If the context is active and the event is still off,
1052 * we need to retry the cross-call.
1054 if (ctx
->is_active
&& event
->state
== PERF_EVENT_STATE_OFF
)
1058 * Since we have the lock this context can't be scheduled
1059 * in, so we can change the state safely.
1061 if (event
->state
== PERF_EVENT_STATE_OFF
)
1062 __perf_event_mark_enabled(event
, ctx
);
1065 raw_spin_unlock_irq(&ctx
->lock
);
1068 static int perf_event_refresh(struct perf_event
*event
, int refresh
)
1071 * not supported on inherited events
1073 if (event
->attr
.inherit
)
1076 atomic_add(refresh
, &event
->event_limit
);
1077 perf_event_enable(event
);
1083 EVENT_FLEXIBLE
= 0x1,
1085 EVENT_ALL
= EVENT_FLEXIBLE
| EVENT_PINNED
,
1088 static void ctx_sched_out(struct perf_event_context
*ctx
,
1089 struct perf_cpu_context
*cpuctx
,
1090 enum event_type_t event_type
)
1092 struct perf_event
*event
;
1094 raw_spin_lock(&ctx
->lock
);
1096 if (likely(!ctx
->nr_events
))
1098 update_context_time(ctx
);
1101 if (!ctx
->nr_active
)
1104 if (event_type
& EVENT_PINNED
)
1105 list_for_each_entry(event
, &ctx
->pinned_groups
, group_entry
)
1106 group_sched_out(event
, cpuctx
, ctx
);
1108 if (event_type
& EVENT_FLEXIBLE
)
1109 list_for_each_entry(event
, &ctx
->flexible_groups
, group_entry
)
1110 group_sched_out(event
, cpuctx
, ctx
);
1115 raw_spin_unlock(&ctx
->lock
);
1119 * Test whether two contexts are equivalent, i.e. whether they
1120 * have both been cloned from the same version of the same context
1121 * and they both have the same number of enabled events.
1122 * If the number of enabled events is the same, then the set
1123 * of enabled events should be the same, because these are both
1124 * inherited contexts, therefore we can't access individual events
1125 * in them directly with an fd; we can only enable/disable all
1126 * events via prctl, or enable/disable all events in a family
1127 * via ioctl, which will have the same effect on both contexts.
1129 static int context_equiv(struct perf_event_context
*ctx1
,
1130 struct perf_event_context
*ctx2
)
1132 return ctx1
->parent_ctx
&& ctx1
->parent_ctx
== ctx2
->parent_ctx
1133 && ctx1
->parent_gen
== ctx2
->parent_gen
1134 && !ctx1
->pin_count
&& !ctx2
->pin_count
;
1137 static void __perf_event_sync_stat(struct perf_event
*event
,
1138 struct perf_event
*next_event
)
1142 if (!event
->attr
.inherit_stat
)
1146 * Update the event value, we cannot use perf_event_read()
1147 * because we're in the middle of a context switch and have IRQs
1148 * disabled, which upsets smp_call_function_single(), however
1149 * we know the event must be on the current CPU, therefore we
1150 * don't need to use it.
1152 switch (event
->state
) {
1153 case PERF_EVENT_STATE_ACTIVE
:
1154 event
->pmu
->read(event
);
1157 case PERF_EVENT_STATE_INACTIVE
:
1158 update_event_times(event
);
1166 * In order to keep per-task stats reliable we need to flip the event
1167 * values when we flip the contexts.
1169 value
= local64_read(&next_event
->count
);
1170 value
= local64_xchg(&event
->count
, value
);
1171 local64_set(&next_event
->count
, value
);
1173 swap(event
->total_time_enabled
, next_event
->total_time_enabled
);
1174 swap(event
->total_time_running
, next_event
->total_time_running
);
1177 * Since we swizzled the values, update the user visible data too.
1179 perf_event_update_userpage(event
);
1180 perf_event_update_userpage(next_event
);
1183 #define list_next_entry(pos, member) \
1184 list_entry(pos->member.next, typeof(*pos), member)
1186 static void perf_event_sync_stat(struct perf_event_context
*ctx
,
1187 struct perf_event_context
*next_ctx
)
1189 struct perf_event
*event
, *next_event
;
1194 update_context_time(ctx
);
1196 event
= list_first_entry(&ctx
->event_list
,
1197 struct perf_event
, event_entry
);
1199 next_event
= list_first_entry(&next_ctx
->event_list
,
1200 struct perf_event
, event_entry
);
1202 while (&event
->event_entry
!= &ctx
->event_list
&&
1203 &next_event
->event_entry
!= &next_ctx
->event_list
) {
1205 __perf_event_sync_stat(event
, next_event
);
1207 event
= list_next_entry(event
, event_entry
);
1208 next_event
= list_next_entry(next_event
, event_entry
);
1213 * Called from scheduler to remove the events of the current task,
1214 * with interrupts disabled.
1216 * We stop each event and update the event value in event->count.
1218 * This does not protect us against NMI, but disable()
1219 * sets the disabled bit in the control field of event _before_
1220 * accessing the event control register. If a NMI hits, then it will
1221 * not restart the event.
1223 void perf_event_task_sched_out(struct task_struct
*task
,
1224 struct task_struct
*next
)
1226 struct perf_cpu_context
*cpuctx
= &__get_cpu_var(perf_cpu_context
);
1227 struct perf_event_context
*ctx
= task
->perf_event_ctxp
;
1228 struct perf_event_context
*next_ctx
;
1229 struct perf_event_context
*parent
;
1232 perf_sw_event(PERF_COUNT_SW_CONTEXT_SWITCHES
, 1, 1, NULL
, 0);
1234 if (likely(!ctx
|| !cpuctx
->task_ctx
))
1238 parent
= rcu_dereference(ctx
->parent_ctx
);
1239 next_ctx
= next
->perf_event_ctxp
;
1240 if (parent
&& next_ctx
&&
1241 rcu_dereference(next_ctx
->parent_ctx
) == parent
) {
1243 * Looks like the two contexts are clones, so we might be
1244 * able to optimize the context switch. We lock both
1245 * contexts and check that they are clones under the
1246 * lock (including re-checking that neither has been
1247 * uncloned in the meantime). It doesn't matter which
1248 * order we take the locks because no other cpu could
1249 * be trying to lock both of these tasks.
1251 raw_spin_lock(&ctx
->lock
);
1252 raw_spin_lock_nested(&next_ctx
->lock
, SINGLE_DEPTH_NESTING
);
1253 if (context_equiv(ctx
, next_ctx
)) {
1254 task
->perf_event_ctxp
= next_ctx
;
1255 next
->perf_event_ctxp
= ctx
;
1257 next_ctx
->task
= task
;
1260 perf_event_sync_stat(ctx
, next_ctx
);
1262 raw_spin_unlock(&next_ctx
->lock
);
1263 raw_spin_unlock(&ctx
->lock
);
1268 ctx_sched_out(ctx
, cpuctx
, EVENT_ALL
);
1269 cpuctx
->task_ctx
= NULL
;
1273 static void task_ctx_sched_out(struct perf_event_context
*ctx
,
1274 enum event_type_t event_type
)
1276 struct perf_cpu_context
*cpuctx
= &__get_cpu_var(perf_cpu_context
);
1278 if (!cpuctx
->task_ctx
)
1281 if (WARN_ON_ONCE(ctx
!= cpuctx
->task_ctx
))
1284 ctx_sched_out(ctx
, cpuctx
, event_type
);
1285 cpuctx
->task_ctx
= NULL
;
1289 * Called with IRQs disabled
1291 static void __perf_event_task_sched_out(struct perf_event_context
*ctx
)
1293 task_ctx_sched_out(ctx
, EVENT_ALL
);
1297 * Called with IRQs disabled
1299 static void cpu_ctx_sched_out(struct perf_cpu_context
*cpuctx
,
1300 enum event_type_t event_type
)
1302 ctx_sched_out(&cpuctx
->ctx
, cpuctx
, event_type
);
1306 ctx_pinned_sched_in(struct perf_event_context
*ctx
,
1307 struct perf_cpu_context
*cpuctx
)
1309 struct perf_event
*event
;
1311 list_for_each_entry(event
, &ctx
->pinned_groups
, group_entry
) {
1312 if (event
->state
<= PERF_EVENT_STATE_OFF
)
1314 if (event
->cpu
!= -1 && event
->cpu
!= smp_processor_id())
1317 if (group_can_go_on(event
, cpuctx
, 1))
1318 group_sched_in(event
, cpuctx
, ctx
);
1321 * If this pinned group hasn't been scheduled,
1322 * put it in error state.
1324 if (event
->state
== PERF_EVENT_STATE_INACTIVE
) {
1325 update_group_times(event
);
1326 event
->state
= PERF_EVENT_STATE_ERROR
;
1332 ctx_flexible_sched_in(struct perf_event_context
*ctx
,
1333 struct perf_cpu_context
*cpuctx
)
1335 struct perf_event
*event
;
1338 list_for_each_entry(event
, &ctx
->flexible_groups
, group_entry
) {
1339 /* Ignore events in OFF or ERROR state */
1340 if (event
->state
<= PERF_EVENT_STATE_OFF
)
1343 * Listen to the 'cpu' scheduling filter constraint
1346 if (event
->cpu
!= -1 && event
->cpu
!= smp_processor_id())
1349 if (group_can_go_on(event
, cpuctx
, can_add_hw
))
1350 if (group_sched_in(event
, cpuctx
, ctx
))
1356 ctx_sched_in(struct perf_event_context
*ctx
,
1357 struct perf_cpu_context
*cpuctx
,
1358 enum event_type_t event_type
)
1360 raw_spin_lock(&ctx
->lock
);
1362 if (likely(!ctx
->nr_events
))
1365 ctx
->timestamp
= perf_clock();
1370 * First go through the list and put on any pinned groups
1371 * in order to give them the best chance of going on.
1373 if (event_type
& EVENT_PINNED
)
1374 ctx_pinned_sched_in(ctx
, cpuctx
);
1376 /* Then walk through the lower prio flexible groups */
1377 if (event_type
& EVENT_FLEXIBLE
)
1378 ctx_flexible_sched_in(ctx
, cpuctx
);
1382 raw_spin_unlock(&ctx
->lock
);
1385 static void cpu_ctx_sched_in(struct perf_cpu_context
*cpuctx
,
1386 enum event_type_t event_type
)
1388 struct perf_event_context
*ctx
= &cpuctx
->ctx
;
1390 ctx_sched_in(ctx
, cpuctx
, event_type
);
1393 static void task_ctx_sched_in(struct task_struct
*task
,
1394 enum event_type_t event_type
)
1396 struct perf_cpu_context
*cpuctx
= &__get_cpu_var(perf_cpu_context
);
1397 struct perf_event_context
*ctx
= task
->perf_event_ctxp
;
1401 if (cpuctx
->task_ctx
== ctx
)
1403 ctx_sched_in(ctx
, cpuctx
, event_type
);
1404 cpuctx
->task_ctx
= ctx
;
1407 * Called from scheduler to add the events of the current task
1408 * with interrupts disabled.
1410 * We restore the event value and then enable it.
1412 * This does not protect us against NMI, but enable()
1413 * sets the enabled bit in the control field of event _before_
1414 * accessing the event control register. If a NMI hits, then it will
1415 * keep the event running.
1417 void perf_event_task_sched_in(struct task_struct
*task
)
1419 struct perf_cpu_context
*cpuctx
= &__get_cpu_var(perf_cpu_context
);
1420 struct perf_event_context
*ctx
= task
->perf_event_ctxp
;
1425 if (cpuctx
->task_ctx
== ctx
)
1431 * We want to keep the following priority order:
1432 * cpu pinned (that don't need to move), task pinned,
1433 * cpu flexible, task flexible.
1435 cpu_ctx_sched_out(cpuctx
, EVENT_FLEXIBLE
);
1437 ctx_sched_in(ctx
, cpuctx
, EVENT_PINNED
);
1438 cpu_ctx_sched_in(cpuctx
, EVENT_FLEXIBLE
);
1439 ctx_sched_in(ctx
, cpuctx
, EVENT_FLEXIBLE
);
1441 cpuctx
->task_ctx
= ctx
;
1446 #define MAX_INTERRUPTS (~0ULL)
1448 static void perf_log_throttle(struct perf_event
*event
, int enable
);
1450 static u64
perf_calculate_period(struct perf_event
*event
, u64 nsec
, u64 count
)
1452 u64 frequency
= event
->attr
.sample_freq
;
1453 u64 sec
= NSEC_PER_SEC
;
1454 u64 divisor
, dividend
;
1456 int count_fls
, nsec_fls
, frequency_fls
, sec_fls
;
1458 count_fls
= fls64(count
);
1459 nsec_fls
= fls64(nsec
);
1460 frequency_fls
= fls64(frequency
);
1464 * We got @count in @nsec, with a target of sample_freq HZ
1465 * the target period becomes:
1468 * period = -------------------
1469 * @nsec * sample_freq
1474 * Reduce accuracy by one bit such that @a and @b converge
1475 * to a similar magnitude.
1477 #define REDUCE_FLS(a, b) \
1479 if (a##_fls > b##_fls) { \
1489 * Reduce accuracy until either term fits in a u64, then proceed with
1490 * the other, so that finally we can do a u64/u64 division.
1492 while (count_fls
+ sec_fls
> 64 && nsec_fls
+ frequency_fls
> 64) {
1493 REDUCE_FLS(nsec
, frequency
);
1494 REDUCE_FLS(sec
, count
);
1497 if (count_fls
+ sec_fls
> 64) {
1498 divisor
= nsec
* frequency
;
1500 while (count_fls
+ sec_fls
> 64) {
1501 REDUCE_FLS(count
, sec
);
1505 dividend
= count
* sec
;
1507 dividend
= count
* sec
;
1509 while (nsec_fls
+ frequency_fls
> 64) {
1510 REDUCE_FLS(nsec
, frequency
);
1514 divisor
= nsec
* frequency
;
1520 return div64_u64(dividend
, divisor
);
1523 static void perf_event_stop(struct perf_event
*event
)
1525 if (!event
->pmu
->stop
)
1526 return event
->pmu
->disable(event
);
1528 return event
->pmu
->stop(event
);
1531 static int perf_event_start(struct perf_event
*event
)
1533 if (!event
->pmu
->start
)
1534 return event
->pmu
->enable(event
);
1536 return event
->pmu
->start(event
);
1539 static void perf_adjust_period(struct perf_event
*event
, u64 nsec
, u64 count
)
1541 struct hw_perf_event
*hwc
= &event
->hw
;
1542 s64 period
, sample_period
;
1545 period
= perf_calculate_period(event
, nsec
, count
);
1547 delta
= (s64
)(period
- hwc
->sample_period
);
1548 delta
= (delta
+ 7) / 8; /* low pass filter */
1550 sample_period
= hwc
->sample_period
+ delta
;
1555 hwc
->sample_period
= sample_period
;
1557 if (local64_read(&hwc
->period_left
) > 8*sample_period
) {
1559 perf_event_stop(event
);
1560 local64_set(&hwc
->period_left
, 0);
1561 perf_event_start(event
);
1566 static void perf_ctx_adjust_freq(struct perf_event_context
*ctx
)
1568 struct perf_event
*event
;
1569 struct hw_perf_event
*hwc
;
1570 u64 interrupts
, now
;
1573 raw_spin_lock(&ctx
->lock
);
1574 list_for_each_entry_rcu(event
, &ctx
->event_list
, event_entry
) {
1575 if (event
->state
!= PERF_EVENT_STATE_ACTIVE
)
1578 if (event
->cpu
!= -1 && event
->cpu
!= smp_processor_id())
1583 interrupts
= hwc
->interrupts
;
1584 hwc
->interrupts
= 0;
1587 * unthrottle events on the tick
1589 if (interrupts
== MAX_INTERRUPTS
) {
1590 perf_log_throttle(event
, 1);
1592 event
->pmu
->unthrottle(event
);
1596 if (!event
->attr
.freq
|| !event
->attr
.sample_freq
)
1600 event
->pmu
->read(event
);
1601 now
= local64_read(&event
->count
);
1602 delta
= now
- hwc
->freq_count_stamp
;
1603 hwc
->freq_count_stamp
= now
;
1606 perf_adjust_period(event
, TICK_NSEC
, delta
);
1609 raw_spin_unlock(&ctx
->lock
);
1613 * Round-robin a context's events:
1615 static void rotate_ctx(struct perf_event_context
*ctx
)
1617 raw_spin_lock(&ctx
->lock
);
1620 * Rotate the first entry last of non-pinned groups. Rotation might be
1621 * disabled by the inheritance code.
1623 if (!ctx
->rotate_disable
)
1624 list_rotate_left(&ctx
->flexible_groups
);
1626 raw_spin_unlock(&ctx
->lock
);
1629 void perf_event_task_tick(struct task_struct
*curr
)
1631 struct perf_cpu_context
*cpuctx
;
1632 struct perf_event_context
*ctx
;
1635 if (!atomic_read(&nr_events
))
1638 cpuctx
= &__get_cpu_var(perf_cpu_context
);
1639 if (cpuctx
->ctx
.nr_events
&&
1640 cpuctx
->ctx
.nr_events
!= cpuctx
->ctx
.nr_active
)
1643 ctx
= curr
->perf_event_ctxp
;
1644 if (ctx
&& ctx
->nr_events
&& ctx
->nr_events
!= ctx
->nr_active
)
1647 perf_ctx_adjust_freq(&cpuctx
->ctx
);
1649 perf_ctx_adjust_freq(ctx
);
1655 cpu_ctx_sched_out(cpuctx
, EVENT_FLEXIBLE
);
1657 task_ctx_sched_out(ctx
, EVENT_FLEXIBLE
);
1659 rotate_ctx(&cpuctx
->ctx
);
1663 cpu_ctx_sched_in(cpuctx
, EVENT_FLEXIBLE
);
1665 task_ctx_sched_in(curr
, EVENT_FLEXIBLE
);
1669 static int event_enable_on_exec(struct perf_event
*event
,
1670 struct perf_event_context
*ctx
)
1672 if (!event
->attr
.enable_on_exec
)
1675 event
->attr
.enable_on_exec
= 0;
1676 if (event
->state
>= PERF_EVENT_STATE_INACTIVE
)
1679 __perf_event_mark_enabled(event
, ctx
);
1685 * Enable all of a task's events that have been marked enable-on-exec.
1686 * This expects task == current.
1688 static void perf_event_enable_on_exec(struct task_struct
*task
)
1690 struct perf_event_context
*ctx
;
1691 struct perf_event
*event
;
1692 unsigned long flags
;
1696 local_irq_save(flags
);
1697 ctx
= task
->perf_event_ctxp
;
1698 if (!ctx
|| !ctx
->nr_events
)
1701 __perf_event_task_sched_out(ctx
);
1703 raw_spin_lock(&ctx
->lock
);
1705 list_for_each_entry(event
, &ctx
->pinned_groups
, group_entry
) {
1706 ret
= event_enable_on_exec(event
, ctx
);
1711 list_for_each_entry(event
, &ctx
->flexible_groups
, group_entry
) {
1712 ret
= event_enable_on_exec(event
, ctx
);
1718 * Unclone this context if we enabled any event.
1723 raw_spin_unlock(&ctx
->lock
);
1725 perf_event_task_sched_in(task
);
1727 local_irq_restore(flags
);
1731 * Cross CPU call to read the hardware event
1733 static void __perf_event_read(void *info
)
1735 struct perf_cpu_context
*cpuctx
= &__get_cpu_var(perf_cpu_context
);
1736 struct perf_event
*event
= info
;
1737 struct perf_event_context
*ctx
= event
->ctx
;
1740 * If this is a task context, we need to check whether it is
1741 * the current task context of this cpu. If not it has been
1742 * scheduled out before the smp call arrived. In that case
1743 * event->count would have been updated to a recent sample
1744 * when the event was scheduled out.
1746 if (ctx
->task
&& cpuctx
->task_ctx
!= ctx
)
1749 raw_spin_lock(&ctx
->lock
);
1750 update_context_time(ctx
);
1751 update_event_times(event
);
1752 raw_spin_unlock(&ctx
->lock
);
1754 event
->pmu
->read(event
);
1757 static inline u64
perf_event_count(struct perf_event
*event
)
1759 return local64_read(&event
->count
) + atomic64_read(&event
->child_count
);
1762 static u64
perf_event_read(struct perf_event
*event
)
1765 * If event is enabled and currently active on a CPU, update the
1766 * value in the event structure:
1768 if (event
->state
== PERF_EVENT_STATE_ACTIVE
) {
1769 smp_call_function_single(event
->oncpu
,
1770 __perf_event_read
, event
, 1);
1771 } else if (event
->state
== PERF_EVENT_STATE_INACTIVE
) {
1772 struct perf_event_context
*ctx
= event
->ctx
;
1773 unsigned long flags
;
1775 raw_spin_lock_irqsave(&ctx
->lock
, flags
);
1777 * may read while context is not active
1778 * (e.g., thread is blocked), in that case
1779 * we cannot update context time
1782 update_context_time(ctx
);
1783 update_event_times(event
);
1784 raw_spin_unlock_irqrestore(&ctx
->lock
, flags
);
1787 return perf_event_count(event
);
1791 * Initialize the perf_event context in a task_struct:
1794 __perf_event_init_context(struct perf_event_context
*ctx
,
1795 struct task_struct
*task
)
1797 raw_spin_lock_init(&ctx
->lock
);
1798 mutex_init(&ctx
->mutex
);
1799 INIT_LIST_HEAD(&ctx
->pinned_groups
);
1800 INIT_LIST_HEAD(&ctx
->flexible_groups
);
1801 INIT_LIST_HEAD(&ctx
->event_list
);
1802 atomic_set(&ctx
->refcount
, 1);
1806 static struct perf_event_context
*find_get_context(pid_t pid
, int cpu
)
1808 struct perf_event_context
*ctx
;
1809 struct perf_cpu_context
*cpuctx
;
1810 struct task_struct
*task
;
1811 unsigned long flags
;
1814 if (pid
== -1 && cpu
!= -1) {
1815 /* Must be root to operate on a CPU event: */
1816 if (perf_paranoid_cpu() && !capable(CAP_SYS_ADMIN
))
1817 return ERR_PTR(-EACCES
);
1819 if (cpu
< 0 || cpu
>= nr_cpumask_bits
)
1820 return ERR_PTR(-EINVAL
);
1823 * We could be clever and allow to attach a event to an
1824 * offline CPU and activate it when the CPU comes up, but
1827 if (!cpu_online(cpu
))
1828 return ERR_PTR(-ENODEV
);
1830 cpuctx
= &per_cpu(perf_cpu_context
, cpu
);
1841 task
= find_task_by_vpid(pid
);
1843 get_task_struct(task
);
1847 return ERR_PTR(-ESRCH
);
1850 * Can't attach events to a dying task.
1853 if (task
->flags
& PF_EXITING
)
1856 /* Reuse ptrace permission checks for now. */
1858 if (!ptrace_may_access(task
, PTRACE_MODE_READ
))
1862 ctx
= perf_lock_task_context(task
, &flags
);
1865 raw_spin_unlock_irqrestore(&ctx
->lock
, flags
);
1869 ctx
= kzalloc(sizeof(struct perf_event_context
), GFP_KERNEL
);
1873 __perf_event_init_context(ctx
, task
);
1875 if (cmpxchg(&task
->perf_event_ctxp
, NULL
, ctx
)) {
1877 * We raced with some other task; use
1878 * the context they set.
1883 get_task_struct(task
);
1886 put_task_struct(task
);
1890 put_task_struct(task
);
1891 return ERR_PTR(err
);
1894 static void perf_event_free_filter(struct perf_event
*event
);
1896 static void free_event_rcu(struct rcu_head
*head
)
1898 struct perf_event
*event
;
1900 event
= container_of(head
, struct perf_event
, rcu_head
);
1902 put_pid_ns(event
->ns
);
1903 perf_event_free_filter(event
);
1907 static void perf_pending_sync(struct perf_event
*event
);
1908 static void perf_buffer_put(struct perf_buffer
*buffer
);
1910 static void free_event(struct perf_event
*event
)
1912 perf_pending_sync(event
);
1914 if (!event
->parent
) {
1915 atomic_dec(&nr_events
);
1916 if (event
->attr
.mmap
|| event
->attr
.mmap_data
)
1917 atomic_dec(&nr_mmap_events
);
1918 if (event
->attr
.comm
)
1919 atomic_dec(&nr_comm_events
);
1920 if (event
->attr
.task
)
1921 atomic_dec(&nr_task_events
);
1924 if (event
->buffer
) {
1925 perf_buffer_put(event
->buffer
);
1926 event
->buffer
= NULL
;
1930 event
->destroy(event
);
1932 put_ctx(event
->ctx
);
1933 call_rcu(&event
->rcu_head
, free_event_rcu
);
1936 int perf_event_release_kernel(struct perf_event
*event
)
1938 struct perf_event_context
*ctx
= event
->ctx
;
1941 * Remove from the PMU, can't get re-enabled since we got
1942 * here because the last ref went.
1944 perf_event_disable(event
);
1946 WARN_ON_ONCE(ctx
->parent_ctx
);
1948 * There are two ways this annotation is useful:
1950 * 1) there is a lock recursion from perf_event_exit_task
1951 * see the comment there.
1953 * 2) there is a lock-inversion with mmap_sem through
1954 * perf_event_read_group(), which takes faults while
1955 * holding ctx->mutex, however this is called after
1956 * the last filedesc died, so there is no possibility
1957 * to trigger the AB-BA case.
1959 mutex_lock_nested(&ctx
->mutex
, SINGLE_DEPTH_NESTING
);
1960 raw_spin_lock_irq(&ctx
->lock
);
1961 perf_group_detach(event
);
1962 list_del_event(event
, ctx
);
1963 raw_spin_unlock_irq(&ctx
->lock
);
1964 mutex_unlock(&ctx
->mutex
);
1966 mutex_lock(&event
->owner
->perf_event_mutex
);
1967 list_del_init(&event
->owner_entry
);
1968 mutex_unlock(&event
->owner
->perf_event_mutex
);
1969 put_task_struct(event
->owner
);
1975 EXPORT_SYMBOL_GPL(perf_event_release_kernel
);
1978 * Called when the last reference to the file is gone.
1980 static int perf_release(struct inode
*inode
, struct file
*file
)
1982 struct perf_event
*event
= file
->private_data
;
1984 file
->private_data
= NULL
;
1986 return perf_event_release_kernel(event
);
1989 static int perf_event_read_size(struct perf_event
*event
)
1991 int entry
= sizeof(u64
); /* value */
1995 if (event
->attr
.read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
)
1996 size
+= sizeof(u64
);
1998 if (event
->attr
.read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
)
1999 size
+= sizeof(u64
);
2001 if (event
->attr
.read_format
& PERF_FORMAT_ID
)
2002 entry
+= sizeof(u64
);
2004 if (event
->attr
.read_format
& PERF_FORMAT_GROUP
) {
2005 nr
+= event
->group_leader
->nr_siblings
;
2006 size
+= sizeof(u64
);
2014 u64
perf_event_read_value(struct perf_event
*event
, u64
*enabled
, u64
*running
)
2016 struct perf_event
*child
;
2022 mutex_lock(&event
->child_mutex
);
2023 total
+= perf_event_read(event
);
2024 *enabled
+= event
->total_time_enabled
+
2025 atomic64_read(&event
->child_total_time_enabled
);
2026 *running
+= event
->total_time_running
+
2027 atomic64_read(&event
->child_total_time_running
);
2029 list_for_each_entry(child
, &event
->child_list
, child_list
) {
2030 total
+= perf_event_read(child
);
2031 *enabled
+= child
->total_time_enabled
;
2032 *running
+= child
->total_time_running
;
2034 mutex_unlock(&event
->child_mutex
);
2038 EXPORT_SYMBOL_GPL(perf_event_read_value
);
2040 static int perf_event_read_group(struct perf_event
*event
,
2041 u64 read_format
, char __user
*buf
)
2043 struct perf_event
*leader
= event
->group_leader
, *sub
;
2044 int n
= 0, size
= 0, ret
= -EFAULT
;
2045 struct perf_event_context
*ctx
= leader
->ctx
;
2047 u64 count
, enabled
, running
;
2049 mutex_lock(&ctx
->mutex
);
2050 count
= perf_event_read_value(leader
, &enabled
, &running
);
2052 values
[n
++] = 1 + leader
->nr_siblings
;
2053 if (read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
)
2054 values
[n
++] = enabled
;
2055 if (read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
)
2056 values
[n
++] = running
;
2057 values
[n
++] = count
;
2058 if (read_format
& PERF_FORMAT_ID
)
2059 values
[n
++] = primary_event_id(leader
);
2061 size
= n
* sizeof(u64
);
2063 if (copy_to_user(buf
, values
, size
))
2068 list_for_each_entry(sub
, &leader
->sibling_list
, group_entry
) {
2071 values
[n
++] = perf_event_read_value(sub
, &enabled
, &running
);
2072 if (read_format
& PERF_FORMAT_ID
)
2073 values
[n
++] = primary_event_id(sub
);
2075 size
= n
* sizeof(u64
);
2077 if (copy_to_user(buf
+ ret
, values
, size
)) {
2085 mutex_unlock(&ctx
->mutex
);
2090 static int perf_event_read_one(struct perf_event
*event
,
2091 u64 read_format
, char __user
*buf
)
2093 u64 enabled
, running
;
2097 values
[n
++] = perf_event_read_value(event
, &enabled
, &running
);
2098 if (read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
)
2099 values
[n
++] = enabled
;
2100 if (read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
)
2101 values
[n
++] = running
;
2102 if (read_format
& PERF_FORMAT_ID
)
2103 values
[n
++] = primary_event_id(event
);
2105 if (copy_to_user(buf
, values
, n
* sizeof(u64
)))
2108 return n
* sizeof(u64
);
2112 * Read the performance event - simple non blocking version for now
2115 perf_read_hw(struct perf_event
*event
, char __user
*buf
, size_t count
)
2117 u64 read_format
= event
->attr
.read_format
;
2121 * Return end-of-file for a read on a event that is in
2122 * error state (i.e. because it was pinned but it couldn't be
2123 * scheduled on to the CPU at some point).
2125 if (event
->state
== PERF_EVENT_STATE_ERROR
)
2128 if (count
< perf_event_read_size(event
))
2131 WARN_ON_ONCE(event
->ctx
->parent_ctx
);
2132 if (read_format
& PERF_FORMAT_GROUP
)
2133 ret
= perf_event_read_group(event
, read_format
, buf
);
2135 ret
= perf_event_read_one(event
, read_format
, buf
);
2141 perf_read(struct file
*file
, char __user
*buf
, size_t count
, loff_t
*ppos
)
2143 struct perf_event
*event
= file
->private_data
;
2145 return perf_read_hw(event
, buf
, count
);
2148 static unsigned int perf_poll(struct file
*file
, poll_table
*wait
)
2150 struct perf_event
*event
= file
->private_data
;
2151 struct perf_buffer
*buffer
;
2152 unsigned int events
= POLL_HUP
;
2155 buffer
= rcu_dereference(event
->buffer
);
2157 events
= atomic_xchg(&buffer
->poll
, 0);
2160 poll_wait(file
, &event
->waitq
, wait
);
2165 static void perf_event_reset(struct perf_event
*event
)
2167 (void)perf_event_read(event
);
2168 local64_set(&event
->count
, 0);
2169 perf_event_update_userpage(event
);
2173 * Holding the top-level event's child_mutex means that any
2174 * descendant process that has inherited this event will block
2175 * in sync_child_event if it goes to exit, thus satisfying the
2176 * task existence requirements of perf_event_enable/disable.
2178 static void perf_event_for_each_child(struct perf_event
*event
,
2179 void (*func
)(struct perf_event
*))
2181 struct perf_event
*child
;
2183 WARN_ON_ONCE(event
->ctx
->parent_ctx
);
2184 mutex_lock(&event
->child_mutex
);
2186 list_for_each_entry(child
, &event
->child_list
, child_list
)
2188 mutex_unlock(&event
->child_mutex
);
2191 static void perf_event_for_each(struct perf_event
*event
,
2192 void (*func
)(struct perf_event
*))
2194 struct perf_event_context
*ctx
= event
->ctx
;
2195 struct perf_event
*sibling
;
2197 WARN_ON_ONCE(ctx
->parent_ctx
);
2198 mutex_lock(&ctx
->mutex
);
2199 event
= event
->group_leader
;
2201 perf_event_for_each_child(event
, func
);
2203 list_for_each_entry(sibling
, &event
->sibling_list
, group_entry
)
2204 perf_event_for_each_child(event
, func
);
2205 mutex_unlock(&ctx
->mutex
);
2208 static int perf_event_period(struct perf_event
*event
, u64 __user
*arg
)
2210 struct perf_event_context
*ctx
= event
->ctx
;
2214 if (!event
->attr
.sample_period
)
2217 if (copy_from_user(&value
, arg
, sizeof(value
)))
2223 raw_spin_lock_irq(&ctx
->lock
);
2224 if (event
->attr
.freq
) {
2225 if (value
> sysctl_perf_event_sample_rate
) {
2230 event
->attr
.sample_freq
= value
;
2232 event
->attr
.sample_period
= value
;
2233 event
->hw
.sample_period
= value
;
2236 raw_spin_unlock_irq(&ctx
->lock
);
2241 static const struct file_operations perf_fops
;
2243 static struct perf_event
*perf_fget_light(int fd
, int *fput_needed
)
2247 file
= fget_light(fd
, fput_needed
);
2249 return ERR_PTR(-EBADF
);
2251 if (file
->f_op
!= &perf_fops
) {
2252 fput_light(file
, *fput_needed
);
2254 return ERR_PTR(-EBADF
);
2257 return file
->private_data
;
2260 static int perf_event_set_output(struct perf_event
*event
,
2261 struct perf_event
*output_event
);
2262 static int perf_event_set_filter(struct perf_event
*event
, void __user
*arg
);
2264 static long perf_ioctl(struct file
*file
, unsigned int cmd
, unsigned long arg
)
2266 struct perf_event
*event
= file
->private_data
;
2267 void (*func
)(struct perf_event
*);
2271 case PERF_EVENT_IOC_ENABLE
:
2272 func
= perf_event_enable
;
2274 case PERF_EVENT_IOC_DISABLE
:
2275 func
= perf_event_disable
;
2277 case PERF_EVENT_IOC_RESET
:
2278 func
= perf_event_reset
;
2281 case PERF_EVENT_IOC_REFRESH
:
2282 return perf_event_refresh(event
, arg
);
2284 case PERF_EVENT_IOC_PERIOD
:
2285 return perf_event_period(event
, (u64 __user
*)arg
);
2287 case PERF_EVENT_IOC_SET_OUTPUT
:
2289 struct perf_event
*output_event
= NULL
;
2290 int fput_needed
= 0;
2294 output_event
= perf_fget_light(arg
, &fput_needed
);
2295 if (IS_ERR(output_event
))
2296 return PTR_ERR(output_event
);
2299 ret
= perf_event_set_output(event
, output_event
);
2301 fput_light(output_event
->filp
, fput_needed
);
2306 case PERF_EVENT_IOC_SET_FILTER
:
2307 return perf_event_set_filter(event
, (void __user
*)arg
);
2313 if (flags
& PERF_IOC_FLAG_GROUP
)
2314 perf_event_for_each(event
, func
);
2316 perf_event_for_each_child(event
, func
);
2321 int perf_event_task_enable(void)
2323 struct perf_event
*event
;
2325 mutex_lock(¤t
->perf_event_mutex
);
2326 list_for_each_entry(event
, ¤t
->perf_event_list
, owner_entry
)
2327 perf_event_for_each_child(event
, perf_event_enable
);
2328 mutex_unlock(¤t
->perf_event_mutex
);
2333 int perf_event_task_disable(void)
2335 struct perf_event
*event
;
2337 mutex_lock(¤t
->perf_event_mutex
);
2338 list_for_each_entry(event
, ¤t
->perf_event_list
, owner_entry
)
2339 perf_event_for_each_child(event
, perf_event_disable
);
2340 mutex_unlock(¤t
->perf_event_mutex
);
2345 #ifndef PERF_EVENT_INDEX_OFFSET
2346 # define PERF_EVENT_INDEX_OFFSET 0
2349 static int perf_event_index(struct perf_event
*event
)
2351 if (event
->state
!= PERF_EVENT_STATE_ACTIVE
)
2354 return event
->hw
.idx
+ 1 - PERF_EVENT_INDEX_OFFSET
;
2358 * Callers need to ensure there can be no nesting of this function, otherwise
2359 * the seqlock logic goes bad. We can not serialize this because the arch
2360 * code calls this from NMI context.
2362 void perf_event_update_userpage(struct perf_event
*event
)
2364 struct perf_event_mmap_page
*userpg
;
2365 struct perf_buffer
*buffer
;
2368 buffer
= rcu_dereference(event
->buffer
);
2372 userpg
= buffer
->user_page
;
2375 * Disable preemption so as to not let the corresponding user-space
2376 * spin too long if we get preempted.
2381 userpg
->index
= perf_event_index(event
);
2382 userpg
->offset
= perf_event_count(event
);
2383 if (event
->state
== PERF_EVENT_STATE_ACTIVE
)
2384 userpg
->offset
-= local64_read(&event
->hw
.prev_count
);
2386 userpg
->time_enabled
= event
->total_time_enabled
+
2387 atomic64_read(&event
->child_total_time_enabled
);
2389 userpg
->time_running
= event
->total_time_running
+
2390 atomic64_read(&event
->child_total_time_running
);
2399 static unsigned long perf_data_size(struct perf_buffer
*buffer
);
2402 perf_buffer_init(struct perf_buffer
*buffer
, long watermark
, int flags
)
2404 long max_size
= perf_data_size(buffer
);
2407 buffer
->watermark
= min(max_size
, watermark
);
2409 if (!buffer
->watermark
)
2410 buffer
->watermark
= max_size
/ 2;
2412 if (flags
& PERF_BUFFER_WRITABLE
)
2413 buffer
->writable
= 1;
2415 atomic_set(&buffer
->refcount
, 1);
2418 #ifndef CONFIG_PERF_USE_VMALLOC
2421 * Back perf_mmap() with regular GFP_KERNEL-0 pages.
2424 static struct page
*
2425 perf_mmap_to_page(struct perf_buffer
*buffer
, unsigned long pgoff
)
2427 if (pgoff
> buffer
->nr_pages
)
2431 return virt_to_page(buffer
->user_page
);
2433 return virt_to_page(buffer
->data_pages
[pgoff
- 1]);
2436 static void *perf_mmap_alloc_page(int cpu
)
2441 node
= (cpu
== -1) ? cpu
: cpu_to_node(cpu
);
2442 page
= alloc_pages_node(node
, GFP_KERNEL
| __GFP_ZERO
, 0);
2446 return page_address(page
);
2449 static struct perf_buffer
*
2450 perf_buffer_alloc(int nr_pages
, long watermark
, int cpu
, int flags
)
2452 struct perf_buffer
*buffer
;
2456 size
= sizeof(struct perf_buffer
);
2457 size
+= nr_pages
* sizeof(void *);
2459 buffer
= kzalloc(size
, GFP_KERNEL
);
2463 buffer
->user_page
= perf_mmap_alloc_page(cpu
);
2464 if (!buffer
->user_page
)
2465 goto fail_user_page
;
2467 for (i
= 0; i
< nr_pages
; i
++) {
2468 buffer
->data_pages
[i
] = perf_mmap_alloc_page(cpu
);
2469 if (!buffer
->data_pages
[i
])
2470 goto fail_data_pages
;
2473 buffer
->nr_pages
= nr_pages
;
2475 perf_buffer_init(buffer
, watermark
, flags
);
2480 for (i
--; i
>= 0; i
--)
2481 free_page((unsigned long)buffer
->data_pages
[i
]);
2483 free_page((unsigned long)buffer
->user_page
);
2492 static void perf_mmap_free_page(unsigned long addr
)
2494 struct page
*page
= virt_to_page((void *)addr
);
2496 page
->mapping
= NULL
;
2500 static void perf_buffer_free(struct perf_buffer
*buffer
)
2504 perf_mmap_free_page((unsigned long)buffer
->user_page
);
2505 for (i
= 0; i
< buffer
->nr_pages
; i
++)
2506 perf_mmap_free_page((unsigned long)buffer
->data_pages
[i
]);
2510 static inline int page_order(struct perf_buffer
*buffer
)
2518 * Back perf_mmap() with vmalloc memory.
2520 * Required for architectures that have d-cache aliasing issues.
2523 static inline int page_order(struct perf_buffer
*buffer
)
2525 return buffer
->page_order
;
2528 static struct page
*
2529 perf_mmap_to_page(struct perf_buffer
*buffer
, unsigned long pgoff
)
2531 if (pgoff
> (1UL << page_order(buffer
)))
2534 return vmalloc_to_page((void *)buffer
->user_page
+ pgoff
* PAGE_SIZE
);
2537 static void perf_mmap_unmark_page(void *addr
)
2539 struct page
*page
= vmalloc_to_page(addr
);
2541 page
->mapping
= NULL
;
2544 static void perf_buffer_free_work(struct work_struct
*work
)
2546 struct perf_buffer
*buffer
;
2550 buffer
= container_of(work
, struct perf_buffer
, work
);
2551 nr
= 1 << page_order(buffer
);
2553 base
= buffer
->user_page
;
2554 for (i
= 0; i
< nr
+ 1; i
++)
2555 perf_mmap_unmark_page(base
+ (i
* PAGE_SIZE
));
2561 static void perf_buffer_free(struct perf_buffer
*buffer
)
2563 schedule_work(&buffer
->work
);
2566 static struct perf_buffer
*
2567 perf_buffer_alloc(int nr_pages
, long watermark
, int cpu
, int flags
)
2569 struct perf_buffer
*buffer
;
2573 size
= sizeof(struct perf_buffer
);
2574 size
+= sizeof(void *);
2576 buffer
= kzalloc(size
, GFP_KERNEL
);
2580 INIT_WORK(&buffer
->work
, perf_buffer_free_work
);
2582 all_buf
= vmalloc_user((nr_pages
+ 1) * PAGE_SIZE
);
2586 buffer
->user_page
= all_buf
;
2587 buffer
->data_pages
[0] = all_buf
+ PAGE_SIZE
;
2588 buffer
->page_order
= ilog2(nr_pages
);
2589 buffer
->nr_pages
= 1;
2591 perf_buffer_init(buffer
, watermark
, flags
);
2604 static unsigned long perf_data_size(struct perf_buffer
*buffer
)
2606 return buffer
->nr_pages
<< (PAGE_SHIFT
+ page_order(buffer
));
2609 static int perf_mmap_fault(struct vm_area_struct
*vma
, struct vm_fault
*vmf
)
2611 struct perf_event
*event
= vma
->vm_file
->private_data
;
2612 struct perf_buffer
*buffer
;
2613 int ret
= VM_FAULT_SIGBUS
;
2615 if (vmf
->flags
& FAULT_FLAG_MKWRITE
) {
2616 if (vmf
->pgoff
== 0)
2622 buffer
= rcu_dereference(event
->buffer
);
2626 if (vmf
->pgoff
&& (vmf
->flags
& FAULT_FLAG_WRITE
))
2629 vmf
->page
= perf_mmap_to_page(buffer
, vmf
->pgoff
);
2633 get_page(vmf
->page
);
2634 vmf
->page
->mapping
= vma
->vm_file
->f_mapping
;
2635 vmf
->page
->index
= vmf
->pgoff
;
2644 static void perf_buffer_free_rcu(struct rcu_head
*rcu_head
)
2646 struct perf_buffer
*buffer
;
2648 buffer
= container_of(rcu_head
, struct perf_buffer
, rcu_head
);
2649 perf_buffer_free(buffer
);
2652 static struct perf_buffer
*perf_buffer_get(struct perf_event
*event
)
2654 struct perf_buffer
*buffer
;
2657 buffer
= rcu_dereference(event
->buffer
);
2659 if (!atomic_inc_not_zero(&buffer
->refcount
))
2667 static void perf_buffer_put(struct perf_buffer
*buffer
)
2669 if (!atomic_dec_and_test(&buffer
->refcount
))
2672 call_rcu(&buffer
->rcu_head
, perf_buffer_free_rcu
);
2675 static void perf_mmap_open(struct vm_area_struct
*vma
)
2677 struct perf_event
*event
= vma
->vm_file
->private_data
;
2679 atomic_inc(&event
->mmap_count
);
2682 static void perf_mmap_close(struct vm_area_struct
*vma
)
2684 struct perf_event
*event
= vma
->vm_file
->private_data
;
2686 if (atomic_dec_and_mutex_lock(&event
->mmap_count
, &event
->mmap_mutex
)) {
2687 unsigned long size
= perf_data_size(event
->buffer
);
2688 struct user_struct
*user
= event
->mmap_user
;
2689 struct perf_buffer
*buffer
= event
->buffer
;
2691 atomic_long_sub((size
>> PAGE_SHIFT
) + 1, &user
->locked_vm
);
2692 vma
->vm_mm
->locked_vm
-= event
->mmap_locked
;
2693 rcu_assign_pointer(event
->buffer
, NULL
);
2694 mutex_unlock(&event
->mmap_mutex
);
2696 perf_buffer_put(buffer
);
2701 static const struct vm_operations_struct perf_mmap_vmops
= {
2702 .open
= perf_mmap_open
,
2703 .close
= perf_mmap_close
,
2704 .fault
= perf_mmap_fault
,
2705 .page_mkwrite
= perf_mmap_fault
,
2708 static int perf_mmap(struct file
*file
, struct vm_area_struct
*vma
)
2710 struct perf_event
*event
= file
->private_data
;
2711 unsigned long user_locked
, user_lock_limit
;
2712 struct user_struct
*user
= current_user();
2713 unsigned long locked
, lock_limit
;
2714 struct perf_buffer
*buffer
;
2715 unsigned long vma_size
;
2716 unsigned long nr_pages
;
2717 long user_extra
, extra
;
2718 int ret
= 0, flags
= 0;
2721 * Don't allow mmap() of inherited per-task counters. This would
2722 * create a performance issue due to all children writing to the
2725 if (event
->cpu
== -1 && event
->attr
.inherit
)
2728 if (!(vma
->vm_flags
& VM_SHARED
))
2731 vma_size
= vma
->vm_end
- vma
->vm_start
;
2732 nr_pages
= (vma_size
/ PAGE_SIZE
) - 1;
2735 * If we have buffer pages ensure they're a power-of-two number, so we
2736 * can do bitmasks instead of modulo.
2738 if (nr_pages
!= 0 && !is_power_of_2(nr_pages
))
2741 if (vma_size
!= PAGE_SIZE
* (1 + nr_pages
))
2744 if (vma
->vm_pgoff
!= 0)
2747 WARN_ON_ONCE(event
->ctx
->parent_ctx
);
2748 mutex_lock(&event
->mmap_mutex
);
2749 if (event
->buffer
) {
2750 if (event
->buffer
->nr_pages
== nr_pages
)
2751 atomic_inc(&event
->buffer
->refcount
);
2757 user_extra
= nr_pages
+ 1;
2758 user_lock_limit
= sysctl_perf_event_mlock
>> (PAGE_SHIFT
- 10);
2761 * Increase the limit linearly with more CPUs:
2763 user_lock_limit
*= num_online_cpus();
2765 user_locked
= atomic_long_read(&user
->locked_vm
) + user_extra
;
2768 if (user_locked
> user_lock_limit
)
2769 extra
= user_locked
- user_lock_limit
;
2771 lock_limit
= rlimit(RLIMIT_MEMLOCK
);
2772 lock_limit
>>= PAGE_SHIFT
;
2773 locked
= vma
->vm_mm
->locked_vm
+ extra
;
2775 if ((locked
> lock_limit
) && perf_paranoid_tracepoint_raw() &&
2776 !capable(CAP_IPC_LOCK
)) {
2781 WARN_ON(event
->buffer
);
2783 if (vma
->vm_flags
& VM_WRITE
)
2784 flags
|= PERF_BUFFER_WRITABLE
;
2786 buffer
= perf_buffer_alloc(nr_pages
, event
->attr
.wakeup_watermark
,
2792 rcu_assign_pointer(event
->buffer
, buffer
);
2794 atomic_long_add(user_extra
, &user
->locked_vm
);
2795 event
->mmap_locked
= extra
;
2796 event
->mmap_user
= get_current_user();
2797 vma
->vm_mm
->locked_vm
+= event
->mmap_locked
;
2801 atomic_inc(&event
->mmap_count
);
2802 mutex_unlock(&event
->mmap_mutex
);
2804 vma
->vm_flags
|= VM_RESERVED
;
2805 vma
->vm_ops
= &perf_mmap_vmops
;
2810 static int perf_fasync(int fd
, struct file
*filp
, int on
)
2812 struct inode
*inode
= filp
->f_path
.dentry
->d_inode
;
2813 struct perf_event
*event
= filp
->private_data
;
2816 mutex_lock(&inode
->i_mutex
);
2817 retval
= fasync_helper(fd
, filp
, on
, &event
->fasync
);
2818 mutex_unlock(&inode
->i_mutex
);
2826 static const struct file_operations perf_fops
= {
2827 .llseek
= no_llseek
,
2828 .release
= perf_release
,
2831 .unlocked_ioctl
= perf_ioctl
,
2832 .compat_ioctl
= perf_ioctl
,
2834 .fasync
= perf_fasync
,
2840 * If there's data, ensure we set the poll() state and publish everything
2841 * to user-space before waking everybody up.
2844 void perf_event_wakeup(struct perf_event
*event
)
2846 wake_up_all(&event
->waitq
);
2848 if (event
->pending_kill
) {
2849 kill_fasync(&event
->fasync
, SIGIO
, event
->pending_kill
);
2850 event
->pending_kill
= 0;
2857 * Handle the case where we need to wakeup up from NMI (or rq->lock) context.
2859 * The NMI bit means we cannot possibly take locks. Therefore, maintain a
2860 * single linked list and use cmpxchg() to add entries lockless.
2863 static void perf_pending_event(struct perf_pending_entry
*entry
)
2865 struct perf_event
*event
= container_of(entry
,
2866 struct perf_event
, pending
);
2868 if (event
->pending_disable
) {
2869 event
->pending_disable
= 0;
2870 __perf_event_disable(event
);
2873 if (event
->pending_wakeup
) {
2874 event
->pending_wakeup
= 0;
2875 perf_event_wakeup(event
);
2879 #define PENDING_TAIL ((struct perf_pending_entry *)-1UL)
2881 static DEFINE_PER_CPU(struct perf_pending_entry
*, perf_pending_head
) = {
2885 static void perf_pending_queue(struct perf_pending_entry
*entry
,
2886 void (*func
)(struct perf_pending_entry
*))
2888 struct perf_pending_entry
**head
;
2890 if (cmpxchg(&entry
->next
, NULL
, PENDING_TAIL
) != NULL
)
2895 head
= &get_cpu_var(perf_pending_head
);
2898 entry
->next
= *head
;
2899 } while (cmpxchg(head
, entry
->next
, entry
) != entry
->next
);
2901 set_perf_event_pending();
2903 put_cpu_var(perf_pending_head
);
2906 static int __perf_pending_run(void)
2908 struct perf_pending_entry
*list
;
2911 list
= xchg(&__get_cpu_var(perf_pending_head
), PENDING_TAIL
);
2912 while (list
!= PENDING_TAIL
) {
2913 void (*func
)(struct perf_pending_entry
*);
2914 struct perf_pending_entry
*entry
= list
;
2921 * Ensure we observe the unqueue before we issue the wakeup,
2922 * so that we won't be waiting forever.
2923 * -- see perf_not_pending().
2934 static inline int perf_not_pending(struct perf_event
*event
)
2937 * If we flush on whatever cpu we run, there is a chance we don't
2941 __perf_pending_run();
2945 * Ensure we see the proper queue state before going to sleep
2946 * so that we do not miss the wakeup. -- see perf_pending_handle()
2949 return event
->pending
.next
== NULL
;
2952 static void perf_pending_sync(struct perf_event
*event
)
2954 wait_event(event
->waitq
, perf_not_pending(event
));
2957 void perf_event_do_pending(void)
2959 __perf_pending_run();
2963 * Callchain support -- arch specific
2966 __weak
struct perf_callchain_entry
*perf_callchain(struct pt_regs
*regs
)
2973 * We assume there is only KVM supporting the callbacks.
2974 * Later on, we might change it to a list if there is
2975 * another virtualization implementation supporting the callbacks.
2977 struct perf_guest_info_callbacks
*perf_guest_cbs
;
2979 int perf_register_guest_info_callbacks(struct perf_guest_info_callbacks
*cbs
)
2981 perf_guest_cbs
= cbs
;
2984 EXPORT_SYMBOL_GPL(perf_register_guest_info_callbacks
);
2986 int perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks
*cbs
)
2988 perf_guest_cbs
= NULL
;
2991 EXPORT_SYMBOL_GPL(perf_unregister_guest_info_callbacks
);
2996 static bool perf_output_space(struct perf_buffer
*buffer
, unsigned long tail
,
2997 unsigned long offset
, unsigned long head
)
3001 if (!buffer
->writable
)
3004 mask
= perf_data_size(buffer
) - 1;
3006 offset
= (offset
- tail
) & mask
;
3007 head
= (head
- tail
) & mask
;
3009 if ((int)(head
- offset
) < 0)
3015 static void perf_output_wakeup(struct perf_output_handle
*handle
)
3017 atomic_set(&handle
->buffer
->poll
, POLL_IN
);
3020 handle
->event
->pending_wakeup
= 1;
3021 perf_pending_queue(&handle
->event
->pending
,
3022 perf_pending_event
);
3024 perf_event_wakeup(handle
->event
);
3028 * We need to ensure a later event_id doesn't publish a head when a former
3029 * event isn't done writing. However since we need to deal with NMIs we
3030 * cannot fully serialize things.
3032 * We only publish the head (and generate a wakeup) when the outer-most
3035 static void perf_output_get_handle(struct perf_output_handle
*handle
)
3037 struct perf_buffer
*buffer
= handle
->buffer
;
3040 local_inc(&buffer
->nest
);
3041 handle
->wakeup
= local_read(&buffer
->wakeup
);
3044 static void perf_output_put_handle(struct perf_output_handle
*handle
)
3046 struct perf_buffer
*buffer
= handle
->buffer
;
3050 head
= local_read(&buffer
->head
);
3053 * IRQ/NMI can happen here, which means we can miss a head update.
3056 if (!local_dec_and_test(&buffer
->nest
))
3060 * Publish the known good head. Rely on the full barrier implied
3061 * by atomic_dec_and_test() order the buffer->head read and this
3064 buffer
->user_page
->data_head
= head
;
3067 * Now check if we missed an update, rely on the (compiler)
3068 * barrier in atomic_dec_and_test() to re-read buffer->head.
3070 if (unlikely(head
!= local_read(&buffer
->head
))) {
3071 local_inc(&buffer
->nest
);
3075 if (handle
->wakeup
!= local_read(&buffer
->wakeup
))
3076 perf_output_wakeup(handle
);
3082 __always_inline
void perf_output_copy(struct perf_output_handle
*handle
,
3083 const void *buf
, unsigned int len
)
3086 unsigned long size
= min_t(unsigned long, handle
->size
, len
);
3088 memcpy(handle
->addr
, buf
, size
);
3091 handle
->addr
+= size
;
3093 handle
->size
-= size
;
3094 if (!handle
->size
) {
3095 struct perf_buffer
*buffer
= handle
->buffer
;
3098 handle
->page
&= buffer
->nr_pages
- 1;
3099 handle
->addr
= buffer
->data_pages
[handle
->page
];
3100 handle
->size
= PAGE_SIZE
<< page_order(buffer
);
3105 int perf_output_begin(struct perf_output_handle
*handle
,
3106 struct perf_event
*event
, unsigned int size
,
3107 int nmi
, int sample
)
3109 struct perf_buffer
*buffer
;
3110 unsigned long tail
, offset
, head
;
3113 struct perf_event_header header
;
3120 * For inherited events we send all the output towards the parent.
3123 event
= event
->parent
;
3125 buffer
= rcu_dereference(event
->buffer
);
3129 handle
->buffer
= buffer
;
3130 handle
->event
= event
;
3132 handle
->sample
= sample
;
3134 if (!buffer
->nr_pages
)
3137 have_lost
= local_read(&buffer
->lost
);
3139 size
+= sizeof(lost_event
);
3141 perf_output_get_handle(handle
);
3145 * Userspace could choose to issue a mb() before updating the
3146 * tail pointer. So that all reads will be completed before the
3149 tail
= ACCESS_ONCE(buffer
->user_page
->data_tail
);
3151 offset
= head
= local_read(&buffer
->head
);
3153 if (unlikely(!perf_output_space(buffer
, tail
, offset
, head
)))
3155 } while (local_cmpxchg(&buffer
->head
, offset
, head
) != offset
);
3157 if (head
- local_read(&buffer
->wakeup
) > buffer
->watermark
)
3158 local_add(buffer
->watermark
, &buffer
->wakeup
);
3160 handle
->page
= offset
>> (PAGE_SHIFT
+ page_order(buffer
));
3161 handle
->page
&= buffer
->nr_pages
- 1;
3162 handle
->size
= offset
& ((PAGE_SIZE
<< page_order(buffer
)) - 1);
3163 handle
->addr
= buffer
->data_pages
[handle
->page
];
3164 handle
->addr
+= handle
->size
;
3165 handle
->size
= (PAGE_SIZE
<< page_order(buffer
)) - handle
->size
;
3168 lost_event
.header
.type
= PERF_RECORD_LOST
;
3169 lost_event
.header
.misc
= 0;
3170 lost_event
.header
.size
= sizeof(lost_event
);
3171 lost_event
.id
= event
->id
;
3172 lost_event
.lost
= local_xchg(&buffer
->lost
, 0);
3174 perf_output_put(handle
, lost_event
);
3180 local_inc(&buffer
->lost
);
3181 perf_output_put_handle(handle
);
3188 void perf_output_end(struct perf_output_handle
*handle
)
3190 struct perf_event
*event
= handle
->event
;
3191 struct perf_buffer
*buffer
= handle
->buffer
;
3193 int wakeup_events
= event
->attr
.wakeup_events
;
3195 if (handle
->sample
&& wakeup_events
) {
3196 int events
= local_inc_return(&buffer
->events
);
3197 if (events
>= wakeup_events
) {
3198 local_sub(wakeup_events
, &buffer
->events
);
3199 local_inc(&buffer
->wakeup
);
3203 perf_output_put_handle(handle
);
3207 static u32
perf_event_pid(struct perf_event
*event
, struct task_struct
*p
)
3210 * only top level events have the pid namespace they were created in
3213 event
= event
->parent
;
3215 return task_tgid_nr_ns(p
, event
->ns
);
3218 static u32
perf_event_tid(struct perf_event
*event
, struct task_struct
*p
)
3221 * only top level events have the pid namespace they were created in
3224 event
= event
->parent
;
3226 return task_pid_nr_ns(p
, event
->ns
);
3229 static void perf_output_read_one(struct perf_output_handle
*handle
,
3230 struct perf_event
*event
)
3232 u64 read_format
= event
->attr
.read_format
;
3236 values
[n
++] = perf_event_count(event
);
3237 if (read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
) {
3238 values
[n
++] = event
->total_time_enabled
+
3239 atomic64_read(&event
->child_total_time_enabled
);
3241 if (read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
) {
3242 values
[n
++] = event
->total_time_running
+
3243 atomic64_read(&event
->child_total_time_running
);
3245 if (read_format
& PERF_FORMAT_ID
)
3246 values
[n
++] = primary_event_id(event
);
3248 perf_output_copy(handle
, values
, n
* sizeof(u64
));
3251 static void perf_output_read_group(struct perf_output_handle
*handle
,
3252 struct perf_event
*event
)
3254 struct perf_event
*leader
= event
->group_leader
, *sub
;
3255 u64 read_format
= event
->attr
.read_format
;
3259 values
[n
++] = 1 + leader
->nr_siblings
;
3261 if (read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
)
3262 values
[n
++] = leader
->total_time_enabled
;
3264 if (read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
)
3265 values
[n
++] = leader
->total_time_running
;
3267 if (leader
!= event
)
3268 leader
->pmu
->read(leader
);
3270 values
[n
++] = perf_event_count(leader
);
3271 if (read_format
& PERF_FORMAT_ID
)
3272 values
[n
++] = primary_event_id(leader
);
3274 perf_output_copy(handle
, values
, n
* sizeof(u64
));
3276 list_for_each_entry(sub
, &leader
->sibling_list
, group_entry
) {
3280 sub
->pmu
->read(sub
);
3282 values
[n
++] = perf_event_count(sub
);
3283 if (read_format
& PERF_FORMAT_ID
)
3284 values
[n
++] = primary_event_id(sub
);
3286 perf_output_copy(handle
, values
, n
* sizeof(u64
));
3290 static void perf_output_read(struct perf_output_handle
*handle
,
3291 struct perf_event
*event
)
3293 if (event
->attr
.read_format
& PERF_FORMAT_GROUP
)
3294 perf_output_read_group(handle
, event
);
3296 perf_output_read_one(handle
, event
);
3299 void perf_output_sample(struct perf_output_handle
*handle
,
3300 struct perf_event_header
*header
,
3301 struct perf_sample_data
*data
,
3302 struct perf_event
*event
)
3304 u64 sample_type
= data
->type
;
3306 perf_output_put(handle
, *header
);
3308 if (sample_type
& PERF_SAMPLE_IP
)
3309 perf_output_put(handle
, data
->ip
);
3311 if (sample_type
& PERF_SAMPLE_TID
)
3312 perf_output_put(handle
, data
->tid_entry
);
3314 if (sample_type
& PERF_SAMPLE_TIME
)
3315 perf_output_put(handle
, data
->time
);
3317 if (sample_type
& PERF_SAMPLE_ADDR
)
3318 perf_output_put(handle
, data
->addr
);
3320 if (sample_type
& PERF_SAMPLE_ID
)
3321 perf_output_put(handle
, data
->id
);
3323 if (sample_type
& PERF_SAMPLE_STREAM_ID
)
3324 perf_output_put(handle
, data
->stream_id
);
3326 if (sample_type
& PERF_SAMPLE_CPU
)
3327 perf_output_put(handle
, data
->cpu_entry
);
3329 if (sample_type
& PERF_SAMPLE_PERIOD
)
3330 perf_output_put(handle
, data
->period
);
3332 if (sample_type
& PERF_SAMPLE_READ
)
3333 perf_output_read(handle
, event
);
3335 if (sample_type
& PERF_SAMPLE_CALLCHAIN
) {
3336 if (data
->callchain
) {
3339 if (data
->callchain
)
3340 size
+= data
->callchain
->nr
;
3342 size
*= sizeof(u64
);
3344 perf_output_copy(handle
, data
->callchain
, size
);
3347 perf_output_put(handle
, nr
);
3351 if (sample_type
& PERF_SAMPLE_RAW
) {
3353 perf_output_put(handle
, data
->raw
->size
);
3354 perf_output_copy(handle
, data
->raw
->data
,
3361 .size
= sizeof(u32
),
3364 perf_output_put(handle
, raw
);
3369 void perf_prepare_sample(struct perf_event_header
*header
,
3370 struct perf_sample_data
*data
,
3371 struct perf_event
*event
,
3372 struct pt_regs
*regs
)
3374 u64 sample_type
= event
->attr
.sample_type
;
3376 data
->type
= sample_type
;
3378 header
->type
= PERF_RECORD_SAMPLE
;
3379 header
->size
= sizeof(*header
);
3382 header
->misc
|= perf_misc_flags(regs
);
3384 if (sample_type
& PERF_SAMPLE_IP
) {
3385 data
->ip
= perf_instruction_pointer(regs
);
3387 header
->size
+= sizeof(data
->ip
);
3390 if (sample_type
& PERF_SAMPLE_TID
) {
3391 /* namespace issues */
3392 data
->tid_entry
.pid
= perf_event_pid(event
, current
);
3393 data
->tid_entry
.tid
= perf_event_tid(event
, current
);
3395 header
->size
+= sizeof(data
->tid_entry
);
3398 if (sample_type
& PERF_SAMPLE_TIME
) {
3399 data
->time
= perf_clock();
3401 header
->size
+= sizeof(data
->time
);
3404 if (sample_type
& PERF_SAMPLE_ADDR
)
3405 header
->size
+= sizeof(data
->addr
);
3407 if (sample_type
& PERF_SAMPLE_ID
) {
3408 data
->id
= primary_event_id(event
);
3410 header
->size
+= sizeof(data
->id
);
3413 if (sample_type
& PERF_SAMPLE_STREAM_ID
) {
3414 data
->stream_id
= event
->id
;
3416 header
->size
+= sizeof(data
->stream_id
);
3419 if (sample_type
& PERF_SAMPLE_CPU
) {
3420 data
->cpu_entry
.cpu
= raw_smp_processor_id();
3421 data
->cpu_entry
.reserved
= 0;
3423 header
->size
+= sizeof(data
->cpu_entry
);
3426 if (sample_type
& PERF_SAMPLE_PERIOD
)
3427 header
->size
+= sizeof(data
->period
);
3429 if (sample_type
& PERF_SAMPLE_READ
)
3430 header
->size
+= perf_event_read_size(event
);
3432 if (sample_type
& PERF_SAMPLE_CALLCHAIN
) {
3435 data
->callchain
= perf_callchain(regs
);
3437 if (data
->callchain
)
3438 size
+= data
->callchain
->nr
;
3440 header
->size
+= size
* sizeof(u64
);
3443 if (sample_type
& PERF_SAMPLE_RAW
) {
3444 int size
= sizeof(u32
);
3447 size
+= data
->raw
->size
;
3449 size
+= sizeof(u32
);
3451 WARN_ON_ONCE(size
& (sizeof(u64
)-1));
3452 header
->size
+= size
;
3456 static void perf_event_output(struct perf_event
*event
, int nmi
,
3457 struct perf_sample_data
*data
,
3458 struct pt_regs
*regs
)
3460 struct perf_output_handle handle
;
3461 struct perf_event_header header
;
3463 perf_prepare_sample(&header
, data
, event
, regs
);
3465 if (perf_output_begin(&handle
, event
, header
.size
, nmi
, 1))
3468 perf_output_sample(&handle
, &header
, data
, event
);
3470 perf_output_end(&handle
);
3477 struct perf_read_event
{
3478 struct perf_event_header header
;
3485 perf_event_read_event(struct perf_event
*event
,
3486 struct task_struct
*task
)
3488 struct perf_output_handle handle
;
3489 struct perf_read_event read_event
= {
3491 .type
= PERF_RECORD_READ
,
3493 .size
= sizeof(read_event
) + perf_event_read_size(event
),
3495 .pid
= perf_event_pid(event
, task
),
3496 .tid
= perf_event_tid(event
, task
),
3500 ret
= perf_output_begin(&handle
, event
, read_event
.header
.size
, 0, 0);
3504 perf_output_put(&handle
, read_event
);
3505 perf_output_read(&handle
, event
);
3507 perf_output_end(&handle
);
3511 * task tracking -- fork/exit
3513 * enabled by: attr.comm | attr.mmap | attr.mmap_data | attr.task
3516 struct perf_task_event
{
3517 struct task_struct
*task
;
3518 struct perf_event_context
*task_ctx
;
3521 struct perf_event_header header
;
3531 static void perf_event_task_output(struct perf_event
*event
,
3532 struct perf_task_event
*task_event
)
3534 struct perf_output_handle handle
;
3535 struct task_struct
*task
= task_event
->task
;
3538 size
= task_event
->event_id
.header
.size
;
3539 ret
= perf_output_begin(&handle
, event
, size
, 0, 0);
3544 task_event
->event_id
.pid
= perf_event_pid(event
, task
);
3545 task_event
->event_id
.ppid
= perf_event_pid(event
, current
);
3547 task_event
->event_id
.tid
= perf_event_tid(event
, task
);
3548 task_event
->event_id
.ptid
= perf_event_tid(event
, current
);
3550 perf_output_put(&handle
, task_event
->event_id
);
3552 perf_output_end(&handle
);
3555 static int perf_event_task_match(struct perf_event
*event
)
3557 if (event
->state
< PERF_EVENT_STATE_INACTIVE
)
3560 if (event
->cpu
!= -1 && event
->cpu
!= smp_processor_id())
3563 if (event
->attr
.comm
|| event
->attr
.mmap
||
3564 event
->attr
.mmap_data
|| event
->attr
.task
)
3570 static void perf_event_task_ctx(struct perf_event_context
*ctx
,
3571 struct perf_task_event
*task_event
)
3573 struct perf_event
*event
;
3575 list_for_each_entry_rcu(event
, &ctx
->event_list
, event_entry
) {
3576 if (perf_event_task_match(event
))
3577 perf_event_task_output(event
, task_event
);
3581 static void perf_event_task_event(struct perf_task_event
*task_event
)
3583 struct perf_cpu_context
*cpuctx
;
3584 struct perf_event_context
*ctx
= task_event
->task_ctx
;
3587 cpuctx
= &get_cpu_var(perf_cpu_context
);
3588 perf_event_task_ctx(&cpuctx
->ctx
, task_event
);
3590 ctx
= rcu_dereference(current
->perf_event_ctxp
);
3592 perf_event_task_ctx(ctx
, task_event
);
3593 put_cpu_var(perf_cpu_context
);
3597 static void perf_event_task(struct task_struct
*task
,
3598 struct perf_event_context
*task_ctx
,
3601 struct perf_task_event task_event
;
3603 if (!atomic_read(&nr_comm_events
) &&
3604 !atomic_read(&nr_mmap_events
) &&
3605 !atomic_read(&nr_task_events
))
3608 task_event
= (struct perf_task_event
){
3610 .task_ctx
= task_ctx
,
3613 .type
= new ? PERF_RECORD_FORK
: PERF_RECORD_EXIT
,
3615 .size
= sizeof(task_event
.event_id
),
3621 .time
= perf_clock(),
3625 perf_event_task_event(&task_event
);
3628 void perf_event_fork(struct task_struct
*task
)
3630 perf_event_task(task
, NULL
, 1);
3637 struct perf_comm_event
{
3638 struct task_struct
*task
;
3643 struct perf_event_header header
;
3650 static void perf_event_comm_output(struct perf_event
*event
,
3651 struct perf_comm_event
*comm_event
)
3653 struct perf_output_handle handle
;
3654 int size
= comm_event
->event_id
.header
.size
;
3655 int ret
= perf_output_begin(&handle
, event
, size
, 0, 0);
3660 comm_event
->event_id
.pid
= perf_event_pid(event
, comm_event
->task
);
3661 comm_event
->event_id
.tid
= perf_event_tid(event
, comm_event
->task
);
3663 perf_output_put(&handle
, comm_event
->event_id
);
3664 perf_output_copy(&handle
, comm_event
->comm
,
3665 comm_event
->comm_size
);
3666 perf_output_end(&handle
);
3669 static int perf_event_comm_match(struct perf_event
*event
)
3671 if (event
->state
< PERF_EVENT_STATE_INACTIVE
)
3674 if (event
->cpu
!= -1 && event
->cpu
!= smp_processor_id())
3677 if (event
->attr
.comm
)
3683 static void perf_event_comm_ctx(struct perf_event_context
*ctx
,
3684 struct perf_comm_event
*comm_event
)
3686 struct perf_event
*event
;
3688 list_for_each_entry_rcu(event
, &ctx
->event_list
, event_entry
) {
3689 if (perf_event_comm_match(event
))
3690 perf_event_comm_output(event
, comm_event
);
3694 static void perf_event_comm_event(struct perf_comm_event
*comm_event
)
3696 struct perf_cpu_context
*cpuctx
;
3697 struct perf_event_context
*ctx
;
3699 char comm
[TASK_COMM_LEN
];
3701 memset(comm
, 0, sizeof(comm
));
3702 strlcpy(comm
, comm_event
->task
->comm
, sizeof(comm
));
3703 size
= ALIGN(strlen(comm
)+1, sizeof(u64
));
3705 comm_event
->comm
= comm
;
3706 comm_event
->comm_size
= size
;
3708 comm_event
->event_id
.header
.size
= sizeof(comm_event
->event_id
) + size
;
3711 cpuctx
= &get_cpu_var(perf_cpu_context
);
3712 perf_event_comm_ctx(&cpuctx
->ctx
, comm_event
);
3713 ctx
= rcu_dereference(current
->perf_event_ctxp
);
3715 perf_event_comm_ctx(ctx
, comm_event
);
3716 put_cpu_var(perf_cpu_context
);
3720 void perf_event_comm(struct task_struct
*task
)
3722 struct perf_comm_event comm_event
;
3724 if (task
->perf_event_ctxp
)
3725 perf_event_enable_on_exec(task
);
3727 if (!atomic_read(&nr_comm_events
))
3730 comm_event
= (struct perf_comm_event
){
3736 .type
= PERF_RECORD_COMM
,
3745 perf_event_comm_event(&comm_event
);
3752 struct perf_mmap_event
{
3753 struct vm_area_struct
*vma
;
3755 const char *file_name
;
3759 struct perf_event_header header
;
3769 static void perf_event_mmap_output(struct perf_event
*event
,
3770 struct perf_mmap_event
*mmap_event
)
3772 struct perf_output_handle handle
;
3773 int size
= mmap_event
->event_id
.header
.size
;
3774 int ret
= perf_output_begin(&handle
, event
, size
, 0, 0);
3779 mmap_event
->event_id
.pid
= perf_event_pid(event
, current
);
3780 mmap_event
->event_id
.tid
= perf_event_tid(event
, current
);
3782 perf_output_put(&handle
, mmap_event
->event_id
);
3783 perf_output_copy(&handle
, mmap_event
->file_name
,
3784 mmap_event
->file_size
);
3785 perf_output_end(&handle
);
3788 static int perf_event_mmap_match(struct perf_event
*event
,
3789 struct perf_mmap_event
*mmap_event
,
3792 if (event
->state
< PERF_EVENT_STATE_INACTIVE
)
3795 if (event
->cpu
!= -1 && event
->cpu
!= smp_processor_id())
3798 if ((!executable
&& event
->attr
.mmap_data
) ||
3799 (executable
&& event
->attr
.mmap
))
3805 static void perf_event_mmap_ctx(struct perf_event_context
*ctx
,
3806 struct perf_mmap_event
*mmap_event
,
3809 struct perf_event
*event
;
3811 list_for_each_entry_rcu(event
, &ctx
->event_list
, event_entry
) {
3812 if (perf_event_mmap_match(event
, mmap_event
, executable
))
3813 perf_event_mmap_output(event
, mmap_event
);
3817 static void perf_event_mmap_event(struct perf_mmap_event
*mmap_event
)
3819 struct perf_cpu_context
*cpuctx
;
3820 struct perf_event_context
*ctx
;
3821 struct vm_area_struct
*vma
= mmap_event
->vma
;
3822 struct file
*file
= vma
->vm_file
;
3828 memset(tmp
, 0, sizeof(tmp
));
3832 * d_path works from the end of the buffer backwards, so we
3833 * need to add enough zero bytes after the string to handle
3834 * the 64bit alignment we do later.
3836 buf
= kzalloc(PATH_MAX
+ sizeof(u64
), GFP_KERNEL
);
3838 name
= strncpy(tmp
, "//enomem", sizeof(tmp
));
3841 name
= d_path(&file
->f_path
, buf
, PATH_MAX
);
3843 name
= strncpy(tmp
, "//toolong", sizeof(tmp
));
3847 if (arch_vma_name(mmap_event
->vma
)) {
3848 name
= strncpy(tmp
, arch_vma_name(mmap_event
->vma
),
3854 name
= strncpy(tmp
, "[vdso]", sizeof(tmp
));
3856 } else if (vma
->vm_start
<= vma
->vm_mm
->start_brk
&&
3857 vma
->vm_end
>= vma
->vm_mm
->brk
) {
3858 name
= strncpy(tmp
, "[heap]", sizeof(tmp
));
3860 } else if (vma
->vm_start
<= vma
->vm_mm
->start_stack
&&
3861 vma
->vm_end
>= vma
->vm_mm
->start_stack
) {
3862 name
= strncpy(tmp
, "[stack]", sizeof(tmp
));
3866 name
= strncpy(tmp
, "//anon", sizeof(tmp
));
3871 size
= ALIGN(strlen(name
)+1, sizeof(u64
));
3873 mmap_event
->file_name
= name
;
3874 mmap_event
->file_size
= size
;
3876 mmap_event
->event_id
.header
.size
= sizeof(mmap_event
->event_id
) + size
;
3879 cpuctx
= &get_cpu_var(perf_cpu_context
);
3880 perf_event_mmap_ctx(&cpuctx
->ctx
, mmap_event
, vma
->vm_flags
& VM_EXEC
);
3881 ctx
= rcu_dereference(current
->perf_event_ctxp
);
3883 perf_event_mmap_ctx(ctx
, mmap_event
, vma
->vm_flags
& VM_EXEC
);
3884 put_cpu_var(perf_cpu_context
);
3890 void perf_event_mmap(struct vm_area_struct
*vma
)
3892 struct perf_mmap_event mmap_event
;
3894 if (!atomic_read(&nr_mmap_events
))
3897 mmap_event
= (struct perf_mmap_event
){
3903 .type
= PERF_RECORD_MMAP
,
3904 .misc
= PERF_RECORD_MISC_USER
,
3909 .start
= vma
->vm_start
,
3910 .len
= vma
->vm_end
- vma
->vm_start
,
3911 .pgoff
= (u64
)vma
->vm_pgoff
<< PAGE_SHIFT
,
3915 perf_event_mmap_event(&mmap_event
);
3919 * IRQ throttle logging
3922 static void perf_log_throttle(struct perf_event
*event
, int enable
)
3924 struct perf_output_handle handle
;
3928 struct perf_event_header header
;
3932 } throttle_event
= {
3934 .type
= PERF_RECORD_THROTTLE
,
3936 .size
= sizeof(throttle_event
),
3938 .time
= perf_clock(),
3939 .id
= primary_event_id(event
),
3940 .stream_id
= event
->id
,
3944 throttle_event
.header
.type
= PERF_RECORD_UNTHROTTLE
;
3946 ret
= perf_output_begin(&handle
, event
, sizeof(throttle_event
), 1, 0);
3950 perf_output_put(&handle
, throttle_event
);
3951 perf_output_end(&handle
);
3955 * Generic event overflow handling, sampling.
3958 static int __perf_event_overflow(struct perf_event
*event
, int nmi
,
3959 int throttle
, struct perf_sample_data
*data
,
3960 struct pt_regs
*regs
)
3962 int events
= atomic_read(&event
->event_limit
);
3963 struct hw_perf_event
*hwc
= &event
->hw
;
3966 throttle
= (throttle
&& event
->pmu
->unthrottle
!= NULL
);
3971 if (hwc
->interrupts
!= MAX_INTERRUPTS
) {
3973 if (HZ
* hwc
->interrupts
>
3974 (u64
)sysctl_perf_event_sample_rate
) {
3975 hwc
->interrupts
= MAX_INTERRUPTS
;
3976 perf_log_throttle(event
, 0);
3981 * Keep re-disabling events even though on the previous
3982 * pass we disabled it - just in case we raced with a
3983 * sched-in and the event got enabled again:
3989 if (event
->attr
.freq
) {
3990 u64 now
= perf_clock();
3991 s64 delta
= now
- hwc
->freq_time_stamp
;
3993 hwc
->freq_time_stamp
= now
;
3995 if (delta
> 0 && delta
< 2*TICK_NSEC
)
3996 perf_adjust_period(event
, delta
, hwc
->last_period
);
4000 event
->pending_kill
= POLL_IN
;
4001 if (events
&& atomic_dec_and_test(&event
->event_limit
)) {
4003 event
->pending_kill
= POLL_HUP
;
4005 event
->pending_disable
= 1;
4006 perf_pending_queue(&event
->pending
,
4007 perf_pending_event
);
4009 perf_event_disable(event
);
4012 if (event
->overflow_handler
)
4013 event
->overflow_handler(event
, nmi
, data
, regs
);
4015 perf_event_output(event
, nmi
, data
, regs
);
4020 int perf_event_overflow(struct perf_event
*event
, int nmi
,
4021 struct perf_sample_data
*data
,
4022 struct pt_regs
*regs
)
4024 return __perf_event_overflow(event
, nmi
, 1, data
, regs
);
4028 * Generic software event infrastructure
4032 * We directly increment event->count and keep a second value in
4033 * event->hw.period_left to count intervals. This period event
4034 * is kept in the range [-sample_period, 0] so that we can use the
4038 static u64
perf_swevent_set_period(struct perf_event
*event
)
4040 struct hw_perf_event
*hwc
= &event
->hw
;
4041 u64 period
= hwc
->last_period
;
4045 hwc
->last_period
= hwc
->sample_period
;
4048 old
= val
= local64_read(&hwc
->period_left
);
4052 nr
= div64_u64(period
+ val
, period
);
4053 offset
= nr
* period
;
4055 if (local64_cmpxchg(&hwc
->period_left
, old
, val
) != old
)
4061 static void perf_swevent_overflow(struct perf_event
*event
, u64 overflow
,
4062 int nmi
, struct perf_sample_data
*data
,
4063 struct pt_regs
*regs
)
4065 struct hw_perf_event
*hwc
= &event
->hw
;
4068 data
->period
= event
->hw
.last_period
;
4070 overflow
= perf_swevent_set_period(event
);
4072 if (hwc
->interrupts
== MAX_INTERRUPTS
)
4075 for (; overflow
; overflow
--) {
4076 if (__perf_event_overflow(event
, nmi
, throttle
,
4079 * We inhibit the overflow from happening when
4080 * hwc->interrupts == MAX_INTERRUPTS.
4088 static void perf_swevent_add(struct perf_event
*event
, u64 nr
,
4089 int nmi
, struct perf_sample_data
*data
,
4090 struct pt_regs
*regs
)
4092 struct hw_perf_event
*hwc
= &event
->hw
;
4094 local64_add(nr
, &event
->count
);
4099 if (!hwc
->sample_period
)
4102 if (nr
== 1 && hwc
->sample_period
== 1 && !event
->attr
.freq
)
4103 return perf_swevent_overflow(event
, 1, nmi
, data
, regs
);
4105 if (local64_add_negative(nr
, &hwc
->period_left
))
4108 perf_swevent_overflow(event
, 0, nmi
, data
, regs
);
4111 static int perf_exclude_event(struct perf_event
*event
,
4112 struct pt_regs
*regs
)
4115 if (event
->attr
.exclude_user
&& user_mode(regs
))
4118 if (event
->attr
.exclude_kernel
&& !user_mode(regs
))
4125 static int perf_swevent_match(struct perf_event
*event
,
4126 enum perf_type_id type
,
4128 struct perf_sample_data
*data
,
4129 struct pt_regs
*regs
)
4131 if (event
->attr
.type
!= type
)
4134 if (event
->attr
.config
!= event_id
)
4137 if (perf_exclude_event(event
, regs
))
4143 static inline u64
swevent_hash(u64 type
, u32 event_id
)
4145 u64 val
= event_id
| (type
<< 32);
4147 return hash_64(val
, SWEVENT_HLIST_BITS
);
4150 static inline struct hlist_head
*
4151 __find_swevent_head(struct swevent_hlist
*hlist
, u64 type
, u32 event_id
)
4153 u64 hash
= swevent_hash(type
, event_id
);
4155 return &hlist
->heads
[hash
];
4158 /* For the read side: events when they trigger */
4159 static inline struct hlist_head
*
4160 find_swevent_head_rcu(struct perf_cpu_context
*ctx
, u64 type
, u32 event_id
)
4162 struct swevent_hlist
*hlist
;
4164 hlist
= rcu_dereference(ctx
->swevent_hlist
);
4168 return __find_swevent_head(hlist
, type
, event_id
);
4171 /* For the event head insertion and removal in the hlist */
4172 static inline struct hlist_head
*
4173 find_swevent_head(struct perf_cpu_context
*ctx
, struct perf_event
*event
)
4175 struct swevent_hlist
*hlist
;
4176 u32 event_id
= event
->attr
.config
;
4177 u64 type
= event
->attr
.type
;
4180 * Event scheduling is always serialized against hlist allocation
4181 * and release. Which makes the protected version suitable here.
4182 * The context lock guarantees that.
4184 hlist
= rcu_dereference_protected(ctx
->swevent_hlist
,
4185 lockdep_is_held(&event
->ctx
->lock
));
4189 return __find_swevent_head(hlist
, type
, event_id
);
4192 static void do_perf_sw_event(enum perf_type_id type
, u32 event_id
,
4194 struct perf_sample_data
*data
,
4195 struct pt_regs
*regs
)
4197 struct perf_cpu_context
*cpuctx
;
4198 struct perf_event
*event
;
4199 struct hlist_node
*node
;
4200 struct hlist_head
*head
;
4202 cpuctx
= &__get_cpu_var(perf_cpu_context
);
4206 head
= find_swevent_head_rcu(cpuctx
, type
, event_id
);
4211 hlist_for_each_entry_rcu(event
, node
, head
, hlist_entry
) {
4212 if (perf_swevent_match(event
, type
, event_id
, data
, regs
))
4213 perf_swevent_add(event
, nr
, nmi
, data
, regs
);
4219 int perf_swevent_get_recursion_context(void)
4221 struct perf_cpu_context
*cpuctx
= &__get_cpu_var(perf_cpu_context
);
4228 else if (in_softirq())
4233 if (cpuctx
->recursion
[rctx
])
4236 cpuctx
->recursion
[rctx
]++;
4241 EXPORT_SYMBOL_GPL(perf_swevent_get_recursion_context
);
4243 void inline perf_swevent_put_recursion_context(int rctx
)
4245 struct perf_cpu_context
*cpuctx
= &__get_cpu_var(perf_cpu_context
);
4247 cpuctx
->recursion
[rctx
]--;
4250 void __perf_sw_event(u32 event_id
, u64 nr
, int nmi
,
4251 struct pt_regs
*regs
, u64 addr
)
4253 struct perf_sample_data data
;
4256 preempt_disable_notrace();
4257 rctx
= perf_swevent_get_recursion_context();
4261 perf_sample_data_init(&data
, addr
);
4263 do_perf_sw_event(PERF_TYPE_SOFTWARE
, event_id
, nr
, nmi
, &data
, regs
);
4265 perf_swevent_put_recursion_context(rctx
);
4266 preempt_enable_notrace();
4269 static void perf_swevent_read(struct perf_event
*event
)
4273 static int perf_swevent_enable(struct perf_event
*event
)
4275 struct hw_perf_event
*hwc
= &event
->hw
;
4276 struct perf_cpu_context
*cpuctx
;
4277 struct hlist_head
*head
;
4279 cpuctx
= &__get_cpu_var(perf_cpu_context
);
4281 if (hwc
->sample_period
) {
4282 hwc
->last_period
= hwc
->sample_period
;
4283 perf_swevent_set_period(event
);
4286 head
= find_swevent_head(cpuctx
, event
);
4287 if (WARN_ON_ONCE(!head
))
4290 hlist_add_head_rcu(&event
->hlist_entry
, head
);
4295 static void perf_swevent_disable(struct perf_event
*event
)
4297 hlist_del_rcu(&event
->hlist_entry
);
4300 static void perf_swevent_void(struct perf_event
*event
)
4304 static int perf_swevent_int(struct perf_event
*event
)
4309 static const struct pmu perf_ops_generic
= {
4310 .enable
= perf_swevent_enable
,
4311 .disable
= perf_swevent_disable
,
4312 .start
= perf_swevent_int
,
4313 .stop
= perf_swevent_void
,
4314 .read
= perf_swevent_read
,
4315 .unthrottle
= perf_swevent_void
, /* hwc->interrupts already reset */
4319 * hrtimer based swevent callback
4322 static enum hrtimer_restart
perf_swevent_hrtimer(struct hrtimer
*hrtimer
)
4324 enum hrtimer_restart ret
= HRTIMER_RESTART
;
4325 struct perf_sample_data data
;
4326 struct pt_regs
*regs
;
4327 struct perf_event
*event
;
4330 event
= container_of(hrtimer
, struct perf_event
, hw
.hrtimer
);
4331 event
->pmu
->read(event
);
4333 perf_sample_data_init(&data
, 0);
4334 data
.period
= event
->hw
.last_period
;
4335 regs
= get_irq_regs();
4337 if (regs
&& !perf_exclude_event(event
, regs
)) {
4338 if (!(event
->attr
.exclude_idle
&& current
->pid
== 0))
4339 if (perf_event_overflow(event
, 0, &data
, regs
))
4340 ret
= HRTIMER_NORESTART
;
4343 period
= max_t(u64
, 10000, event
->hw
.sample_period
);
4344 hrtimer_forward_now(hrtimer
, ns_to_ktime(period
));
4349 static void perf_swevent_start_hrtimer(struct perf_event
*event
)
4351 struct hw_perf_event
*hwc
= &event
->hw
;
4353 hrtimer_init(&hwc
->hrtimer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
4354 hwc
->hrtimer
.function
= perf_swevent_hrtimer
;
4355 if (hwc
->sample_period
) {
4358 if (hwc
->remaining
) {
4359 if (hwc
->remaining
< 0)
4362 period
= hwc
->remaining
;
4365 period
= max_t(u64
, 10000, hwc
->sample_period
);
4367 __hrtimer_start_range_ns(&hwc
->hrtimer
,
4368 ns_to_ktime(period
), 0,
4369 HRTIMER_MODE_REL
, 0);
4373 static void perf_swevent_cancel_hrtimer(struct perf_event
*event
)
4375 struct hw_perf_event
*hwc
= &event
->hw
;
4377 if (hwc
->sample_period
) {
4378 ktime_t remaining
= hrtimer_get_remaining(&hwc
->hrtimer
);
4379 hwc
->remaining
= ktime_to_ns(remaining
);
4381 hrtimer_cancel(&hwc
->hrtimer
);
4386 * Software event: cpu wall time clock
4389 static void cpu_clock_perf_event_update(struct perf_event
*event
)
4391 int cpu
= raw_smp_processor_id();
4395 now
= cpu_clock(cpu
);
4396 prev
= local64_xchg(&event
->hw
.prev_count
, now
);
4397 local64_add(now
- prev
, &event
->count
);
4400 static int cpu_clock_perf_event_enable(struct perf_event
*event
)
4402 struct hw_perf_event
*hwc
= &event
->hw
;
4403 int cpu
= raw_smp_processor_id();
4405 local64_set(&hwc
->prev_count
, cpu_clock(cpu
));
4406 perf_swevent_start_hrtimer(event
);
4411 static void cpu_clock_perf_event_disable(struct perf_event
*event
)
4413 perf_swevent_cancel_hrtimer(event
);
4414 cpu_clock_perf_event_update(event
);
4417 static void cpu_clock_perf_event_read(struct perf_event
*event
)
4419 cpu_clock_perf_event_update(event
);
4422 static const struct pmu perf_ops_cpu_clock
= {
4423 .enable
= cpu_clock_perf_event_enable
,
4424 .disable
= cpu_clock_perf_event_disable
,
4425 .read
= cpu_clock_perf_event_read
,
4429 * Software event: task time clock
4432 static void task_clock_perf_event_update(struct perf_event
*event
, u64 now
)
4437 prev
= local64_xchg(&event
->hw
.prev_count
, now
);
4439 local64_add(delta
, &event
->count
);
4442 static int task_clock_perf_event_enable(struct perf_event
*event
)
4444 struct hw_perf_event
*hwc
= &event
->hw
;
4447 now
= event
->ctx
->time
;
4449 local64_set(&hwc
->prev_count
, now
);
4451 perf_swevent_start_hrtimer(event
);
4456 static void task_clock_perf_event_disable(struct perf_event
*event
)
4458 perf_swevent_cancel_hrtimer(event
);
4459 task_clock_perf_event_update(event
, event
->ctx
->time
);
4463 static void task_clock_perf_event_read(struct perf_event
*event
)
4468 update_context_time(event
->ctx
);
4469 time
= event
->ctx
->time
;
4471 u64 now
= perf_clock();
4472 u64 delta
= now
- event
->ctx
->timestamp
;
4473 time
= event
->ctx
->time
+ delta
;
4476 task_clock_perf_event_update(event
, time
);
4479 static const struct pmu perf_ops_task_clock
= {
4480 .enable
= task_clock_perf_event_enable
,
4481 .disable
= task_clock_perf_event_disable
,
4482 .read
= task_clock_perf_event_read
,
4485 /* Deref the hlist from the update side */
4486 static inline struct swevent_hlist
*
4487 swevent_hlist_deref(struct perf_cpu_context
*cpuctx
)
4489 return rcu_dereference_protected(cpuctx
->swevent_hlist
,
4490 lockdep_is_held(&cpuctx
->hlist_mutex
));
4493 static void swevent_hlist_release_rcu(struct rcu_head
*rcu_head
)
4495 struct swevent_hlist
*hlist
;
4497 hlist
= container_of(rcu_head
, struct swevent_hlist
, rcu_head
);
4501 static void swevent_hlist_release(struct perf_cpu_context
*cpuctx
)
4503 struct swevent_hlist
*hlist
= swevent_hlist_deref(cpuctx
);
4508 rcu_assign_pointer(cpuctx
->swevent_hlist
, NULL
);
4509 call_rcu(&hlist
->rcu_head
, swevent_hlist_release_rcu
);
4512 static void swevent_hlist_put_cpu(struct perf_event
*event
, int cpu
)
4514 struct perf_cpu_context
*cpuctx
= &per_cpu(perf_cpu_context
, cpu
);
4516 mutex_lock(&cpuctx
->hlist_mutex
);
4518 if (!--cpuctx
->hlist_refcount
)
4519 swevent_hlist_release(cpuctx
);
4521 mutex_unlock(&cpuctx
->hlist_mutex
);
4524 static void swevent_hlist_put(struct perf_event
*event
)
4528 if (event
->cpu
!= -1) {
4529 swevent_hlist_put_cpu(event
, event
->cpu
);
4533 for_each_possible_cpu(cpu
)
4534 swevent_hlist_put_cpu(event
, cpu
);
4537 static int swevent_hlist_get_cpu(struct perf_event
*event
, int cpu
)
4539 struct perf_cpu_context
*cpuctx
= &per_cpu(perf_cpu_context
, cpu
);
4542 mutex_lock(&cpuctx
->hlist_mutex
);
4544 if (!swevent_hlist_deref(cpuctx
) && cpu_online(cpu
)) {
4545 struct swevent_hlist
*hlist
;
4547 hlist
= kzalloc(sizeof(*hlist
), GFP_KERNEL
);
4552 rcu_assign_pointer(cpuctx
->swevent_hlist
, hlist
);
4554 cpuctx
->hlist_refcount
++;
4556 mutex_unlock(&cpuctx
->hlist_mutex
);
4561 static int swevent_hlist_get(struct perf_event
*event
)
4564 int cpu
, failed_cpu
;
4566 if (event
->cpu
!= -1)
4567 return swevent_hlist_get_cpu(event
, event
->cpu
);
4570 for_each_possible_cpu(cpu
) {
4571 err
= swevent_hlist_get_cpu(event
, cpu
);
4581 for_each_possible_cpu(cpu
) {
4582 if (cpu
== failed_cpu
)
4584 swevent_hlist_put_cpu(event
, cpu
);
4591 #ifdef CONFIG_EVENT_TRACING
4593 static const struct pmu perf_ops_tracepoint
= {
4594 .enable
= perf_trace_enable
,
4595 .disable
= perf_trace_disable
,
4596 .start
= perf_swevent_int
,
4597 .stop
= perf_swevent_void
,
4598 .read
= perf_swevent_read
,
4599 .unthrottle
= perf_swevent_void
,
4602 static int perf_tp_filter_match(struct perf_event
*event
,
4603 struct perf_sample_data
*data
)
4605 void *record
= data
->raw
->data
;
4607 if (likely(!event
->filter
) || filter_match_preds(event
->filter
, record
))
4612 static int perf_tp_event_match(struct perf_event
*event
,
4613 struct perf_sample_data
*data
,
4614 struct pt_regs
*regs
)
4617 * All tracepoints are from kernel-space.
4619 if (event
->attr
.exclude_kernel
)
4622 if (!perf_tp_filter_match(event
, data
))
4628 void perf_tp_event(u64 addr
, u64 count
, void *record
, int entry_size
,
4629 struct pt_regs
*regs
, struct hlist_head
*head
, int rctx
)
4631 struct perf_sample_data data
;
4632 struct perf_event
*event
;
4633 struct hlist_node
*node
;
4635 struct perf_raw_record raw
= {
4640 perf_sample_data_init(&data
, addr
);
4643 hlist_for_each_entry_rcu(event
, node
, head
, hlist_entry
) {
4644 if (perf_tp_event_match(event
, &data
, regs
))
4645 perf_swevent_add(event
, count
, 1, &data
, regs
);
4648 perf_swevent_put_recursion_context(rctx
);
4650 EXPORT_SYMBOL_GPL(perf_tp_event
);
4652 static void tp_perf_event_destroy(struct perf_event
*event
)
4654 perf_trace_destroy(event
);
4657 static const struct pmu
*tp_perf_event_init(struct perf_event
*event
)
4662 * Raw tracepoint data is a severe data leak, only allow root to
4665 if ((event
->attr
.sample_type
& PERF_SAMPLE_RAW
) &&
4666 perf_paranoid_tracepoint_raw() &&
4667 !capable(CAP_SYS_ADMIN
))
4668 return ERR_PTR(-EPERM
);
4670 err
= perf_trace_init(event
);
4674 event
->destroy
= tp_perf_event_destroy
;
4676 return &perf_ops_tracepoint
;
4679 static int perf_event_set_filter(struct perf_event
*event
, void __user
*arg
)
4684 if (event
->attr
.type
!= PERF_TYPE_TRACEPOINT
)
4687 filter_str
= strndup_user(arg
, PAGE_SIZE
);
4688 if (IS_ERR(filter_str
))
4689 return PTR_ERR(filter_str
);
4691 ret
= ftrace_profile_set_filter(event
, event
->attr
.config
, filter_str
);
4697 static void perf_event_free_filter(struct perf_event
*event
)
4699 ftrace_profile_free_filter(event
);
4704 static const struct pmu
*tp_perf_event_init(struct perf_event
*event
)
4709 static int perf_event_set_filter(struct perf_event
*event
, void __user
*arg
)
4714 static void perf_event_free_filter(struct perf_event
*event
)
4718 #endif /* CONFIG_EVENT_TRACING */
4720 #ifdef CONFIG_HAVE_HW_BREAKPOINT
4721 static void bp_perf_event_destroy(struct perf_event
*event
)
4723 release_bp_slot(event
);
4726 static const struct pmu
*bp_perf_event_init(struct perf_event
*bp
)
4730 err
= register_perf_hw_breakpoint(bp
);
4732 return ERR_PTR(err
);
4734 bp
->destroy
= bp_perf_event_destroy
;
4736 return &perf_ops_bp
;
4739 void perf_bp_event(struct perf_event
*bp
, void *data
)
4741 struct perf_sample_data sample
;
4742 struct pt_regs
*regs
= data
;
4744 perf_sample_data_init(&sample
, bp
->attr
.bp_addr
);
4746 if (!perf_exclude_event(bp
, regs
))
4747 perf_swevent_add(bp
, 1, 1, &sample
, regs
);
4750 static const struct pmu
*bp_perf_event_init(struct perf_event
*bp
)
4755 void perf_bp_event(struct perf_event
*bp
, void *regs
)
4760 atomic_t perf_swevent_enabled
[PERF_COUNT_SW_MAX
];
4762 static void sw_perf_event_destroy(struct perf_event
*event
)
4764 u64 event_id
= event
->attr
.config
;
4766 WARN_ON(event
->parent
);
4768 atomic_dec(&perf_swevent_enabled
[event_id
]);
4769 swevent_hlist_put(event
);
4772 static const struct pmu
*sw_perf_event_init(struct perf_event
*event
)
4774 const struct pmu
*pmu
= NULL
;
4775 u64 event_id
= event
->attr
.config
;
4778 * Software events (currently) can't in general distinguish
4779 * between user, kernel and hypervisor events.
4780 * However, context switches and cpu migrations are considered
4781 * to be kernel events, and page faults are never hypervisor
4785 case PERF_COUNT_SW_CPU_CLOCK
:
4786 pmu
= &perf_ops_cpu_clock
;
4789 case PERF_COUNT_SW_TASK_CLOCK
:
4791 * If the user instantiates this as a per-cpu event,
4792 * use the cpu_clock event instead.
4794 if (event
->ctx
->task
)
4795 pmu
= &perf_ops_task_clock
;
4797 pmu
= &perf_ops_cpu_clock
;
4800 case PERF_COUNT_SW_PAGE_FAULTS
:
4801 case PERF_COUNT_SW_PAGE_FAULTS_MIN
:
4802 case PERF_COUNT_SW_PAGE_FAULTS_MAJ
:
4803 case PERF_COUNT_SW_CONTEXT_SWITCHES
:
4804 case PERF_COUNT_SW_CPU_MIGRATIONS
:
4805 case PERF_COUNT_SW_ALIGNMENT_FAULTS
:
4806 case PERF_COUNT_SW_EMULATION_FAULTS
:
4807 if (!event
->parent
) {
4810 err
= swevent_hlist_get(event
);
4812 return ERR_PTR(err
);
4814 atomic_inc(&perf_swevent_enabled
[event_id
]);
4815 event
->destroy
= sw_perf_event_destroy
;
4817 pmu
= &perf_ops_generic
;
4825 * Allocate and initialize a event structure
4827 static struct perf_event
*
4828 perf_event_alloc(struct perf_event_attr
*attr
,
4830 struct perf_event_context
*ctx
,
4831 struct perf_event
*group_leader
,
4832 struct perf_event
*parent_event
,
4833 perf_overflow_handler_t overflow_handler
,
4836 const struct pmu
*pmu
;
4837 struct perf_event
*event
;
4838 struct hw_perf_event
*hwc
;
4841 event
= kzalloc(sizeof(*event
), gfpflags
);
4843 return ERR_PTR(-ENOMEM
);
4846 * Single events are their own group leaders, with an
4847 * empty sibling list:
4850 group_leader
= event
;
4852 mutex_init(&event
->child_mutex
);
4853 INIT_LIST_HEAD(&event
->child_list
);
4855 INIT_LIST_HEAD(&event
->group_entry
);
4856 INIT_LIST_HEAD(&event
->event_entry
);
4857 INIT_LIST_HEAD(&event
->sibling_list
);
4858 init_waitqueue_head(&event
->waitq
);
4860 mutex_init(&event
->mmap_mutex
);
4863 event
->attr
= *attr
;
4864 event
->group_leader
= group_leader
;
4869 event
->parent
= parent_event
;
4871 event
->ns
= get_pid_ns(current
->nsproxy
->pid_ns
);
4872 event
->id
= atomic64_inc_return(&perf_event_id
);
4874 event
->state
= PERF_EVENT_STATE_INACTIVE
;
4876 if (!overflow_handler
&& parent_event
)
4877 overflow_handler
= parent_event
->overflow_handler
;
4879 event
->overflow_handler
= overflow_handler
;
4882 event
->state
= PERF_EVENT_STATE_OFF
;
4887 hwc
->sample_period
= attr
->sample_period
;
4888 if (attr
->freq
&& attr
->sample_freq
)
4889 hwc
->sample_period
= 1;
4890 hwc
->last_period
= hwc
->sample_period
;
4892 local64_set(&hwc
->period_left
, hwc
->sample_period
);
4895 * we currently do not support PERF_FORMAT_GROUP on inherited events
4897 if (attr
->inherit
&& (attr
->read_format
& PERF_FORMAT_GROUP
))
4900 switch (attr
->type
) {
4902 case PERF_TYPE_HARDWARE
:
4903 case PERF_TYPE_HW_CACHE
:
4904 pmu
= hw_perf_event_init(event
);
4907 case PERF_TYPE_SOFTWARE
:
4908 pmu
= sw_perf_event_init(event
);
4911 case PERF_TYPE_TRACEPOINT
:
4912 pmu
= tp_perf_event_init(event
);
4915 case PERF_TYPE_BREAKPOINT
:
4916 pmu
= bp_perf_event_init(event
);
4927 else if (IS_ERR(pmu
))
4932 put_pid_ns(event
->ns
);
4934 return ERR_PTR(err
);
4939 if (!event
->parent
) {
4940 atomic_inc(&nr_events
);
4941 if (event
->attr
.mmap
|| event
->attr
.mmap_data
)
4942 atomic_inc(&nr_mmap_events
);
4943 if (event
->attr
.comm
)
4944 atomic_inc(&nr_comm_events
);
4945 if (event
->attr
.task
)
4946 atomic_inc(&nr_task_events
);
4952 static int perf_copy_attr(struct perf_event_attr __user
*uattr
,
4953 struct perf_event_attr
*attr
)
4958 if (!access_ok(VERIFY_WRITE
, uattr
, PERF_ATTR_SIZE_VER0
))
4962 * zero the full structure, so that a short copy will be nice.
4964 memset(attr
, 0, sizeof(*attr
));
4966 ret
= get_user(size
, &uattr
->size
);
4970 if (size
> PAGE_SIZE
) /* silly large */
4973 if (!size
) /* abi compat */
4974 size
= PERF_ATTR_SIZE_VER0
;
4976 if (size
< PERF_ATTR_SIZE_VER0
)
4980 * If we're handed a bigger struct than we know of,
4981 * ensure all the unknown bits are 0 - i.e. new
4982 * user-space does not rely on any kernel feature
4983 * extensions we dont know about yet.
4985 if (size
> sizeof(*attr
)) {
4986 unsigned char __user
*addr
;
4987 unsigned char __user
*end
;
4990 addr
= (void __user
*)uattr
+ sizeof(*attr
);
4991 end
= (void __user
*)uattr
+ size
;
4993 for (; addr
< end
; addr
++) {
4994 ret
= get_user(val
, addr
);
5000 size
= sizeof(*attr
);
5003 ret
= copy_from_user(attr
, uattr
, size
);
5008 * If the type exists, the corresponding creation will verify
5011 if (attr
->type
>= PERF_TYPE_MAX
)
5014 if (attr
->__reserved_1
)
5017 if (attr
->sample_type
& ~(PERF_SAMPLE_MAX
-1))
5020 if (attr
->read_format
& ~(PERF_FORMAT_MAX
-1))
5027 put_user(sizeof(*attr
), &uattr
->size
);
5033 perf_event_set_output(struct perf_event
*event
, struct perf_event
*output_event
)
5035 struct perf_buffer
*buffer
= NULL
, *old_buffer
= NULL
;
5041 /* don't allow circular references */
5042 if (event
== output_event
)
5046 * Don't allow cross-cpu buffers
5048 if (output_event
->cpu
!= event
->cpu
)
5052 * If its not a per-cpu buffer, it must be the same task.
5054 if (output_event
->cpu
== -1 && output_event
->ctx
!= event
->ctx
)
5058 mutex_lock(&event
->mmap_mutex
);
5059 /* Can't redirect output if we've got an active mmap() */
5060 if (atomic_read(&event
->mmap_count
))
5064 /* get the buffer we want to redirect to */
5065 buffer
= perf_buffer_get(output_event
);
5070 old_buffer
= event
->buffer
;
5071 rcu_assign_pointer(event
->buffer
, buffer
);
5074 mutex_unlock(&event
->mmap_mutex
);
5077 perf_buffer_put(old_buffer
);
5083 * sys_perf_event_open - open a performance event, associate it to a task/cpu
5085 * @attr_uptr: event_id type attributes for monitoring/sampling
5088 * @group_fd: group leader event fd
5090 SYSCALL_DEFINE5(perf_event_open
,
5091 struct perf_event_attr __user
*, attr_uptr
,
5092 pid_t
, pid
, int, cpu
, int, group_fd
, unsigned long, flags
)
5094 struct perf_event
*event
, *group_leader
= NULL
, *output_event
= NULL
;
5095 struct perf_event_attr attr
;
5096 struct perf_event_context
*ctx
;
5097 struct file
*event_file
= NULL
;
5098 struct file
*group_file
= NULL
;
5100 int fput_needed
= 0;
5103 /* for future expandability... */
5104 if (flags
& ~(PERF_FLAG_FD_NO_GROUP
| PERF_FLAG_FD_OUTPUT
))
5107 err
= perf_copy_attr(attr_uptr
, &attr
);
5111 if (!attr
.exclude_kernel
) {
5112 if (perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN
))
5117 if (attr
.sample_freq
> sysctl_perf_event_sample_rate
)
5121 event_fd
= get_unused_fd_flags(O_RDWR
);
5126 * Get the target context (task or percpu):
5128 ctx
= find_get_context(pid
, cpu
);
5134 if (group_fd
!= -1) {
5135 group_leader
= perf_fget_light(group_fd
, &fput_needed
);
5136 if (IS_ERR(group_leader
)) {
5137 err
= PTR_ERR(group_leader
);
5138 goto err_put_context
;
5140 group_file
= group_leader
->filp
;
5141 if (flags
& PERF_FLAG_FD_OUTPUT
)
5142 output_event
= group_leader
;
5143 if (flags
& PERF_FLAG_FD_NO_GROUP
)
5144 group_leader
= NULL
;
5148 * Look up the group leader (we will attach this event to it):
5154 * Do not allow a recursive hierarchy (this new sibling
5155 * becoming part of another group-sibling):
5157 if (group_leader
->group_leader
!= group_leader
)
5158 goto err_put_context
;
5160 * Do not allow to attach to a group in a different
5161 * task or CPU context:
5163 if (group_leader
->ctx
!= ctx
)
5164 goto err_put_context
;
5166 * Only a group leader can be exclusive or pinned
5168 if (attr
.exclusive
|| attr
.pinned
)
5169 goto err_put_context
;
5172 event
= perf_event_alloc(&attr
, cpu
, ctx
, group_leader
,
5173 NULL
, NULL
, GFP_KERNEL
);
5174 if (IS_ERR(event
)) {
5175 err
= PTR_ERR(event
);
5176 goto err_put_context
;
5180 err
= perf_event_set_output(event
, output_event
);
5182 goto err_free_put_context
;
5185 event_file
= anon_inode_getfile("[perf_event]", &perf_fops
, event
, O_RDWR
);
5186 if (IS_ERR(event_file
)) {
5187 err
= PTR_ERR(event_file
);
5188 goto err_free_put_context
;
5191 event
->filp
= event_file
;
5192 WARN_ON_ONCE(ctx
->parent_ctx
);
5193 mutex_lock(&ctx
->mutex
);
5194 perf_install_in_context(ctx
, event
, cpu
);
5196 mutex_unlock(&ctx
->mutex
);
5198 event
->owner
= current
;
5199 get_task_struct(current
);
5200 mutex_lock(¤t
->perf_event_mutex
);
5201 list_add_tail(&event
->owner_entry
, ¤t
->perf_event_list
);
5202 mutex_unlock(¤t
->perf_event_mutex
);
5205 * Drop the reference on the group_event after placing the
5206 * new event on the sibling_list. This ensures destruction
5207 * of the group leader will find the pointer to itself in
5208 * perf_group_detach().
5210 fput_light(group_file
, fput_needed
);
5211 fd_install(event_fd
, event_file
);
5214 err_free_put_context
:
5217 fput_light(group_file
, fput_needed
);
5220 put_unused_fd(event_fd
);
5225 * perf_event_create_kernel_counter
5227 * @attr: attributes of the counter to create
5228 * @cpu: cpu in which the counter is bound
5229 * @pid: task to profile
5232 perf_event_create_kernel_counter(struct perf_event_attr
*attr
, int cpu
,
5234 perf_overflow_handler_t overflow_handler
)
5236 struct perf_event
*event
;
5237 struct perf_event_context
*ctx
;
5241 * Get the target context (task or percpu):
5244 ctx
= find_get_context(pid
, cpu
);
5250 event
= perf_event_alloc(attr
, cpu
, ctx
, NULL
,
5251 NULL
, overflow_handler
, GFP_KERNEL
);
5252 if (IS_ERR(event
)) {
5253 err
= PTR_ERR(event
);
5254 goto err_put_context
;
5258 WARN_ON_ONCE(ctx
->parent_ctx
);
5259 mutex_lock(&ctx
->mutex
);
5260 perf_install_in_context(ctx
, event
, cpu
);
5262 mutex_unlock(&ctx
->mutex
);
5264 event
->owner
= current
;
5265 get_task_struct(current
);
5266 mutex_lock(¤t
->perf_event_mutex
);
5267 list_add_tail(&event
->owner_entry
, ¤t
->perf_event_list
);
5268 mutex_unlock(¤t
->perf_event_mutex
);
5275 return ERR_PTR(err
);
5277 EXPORT_SYMBOL_GPL(perf_event_create_kernel_counter
);
5280 * inherit a event from parent task to child task:
5282 static struct perf_event
*
5283 inherit_event(struct perf_event
*parent_event
,
5284 struct task_struct
*parent
,
5285 struct perf_event_context
*parent_ctx
,
5286 struct task_struct
*child
,
5287 struct perf_event
*group_leader
,
5288 struct perf_event_context
*child_ctx
)
5290 struct perf_event
*child_event
;
5293 * Instead of creating recursive hierarchies of events,
5294 * we link inherited events back to the original parent,
5295 * which has a filp for sure, which we use as the reference
5298 if (parent_event
->parent
)
5299 parent_event
= parent_event
->parent
;
5301 child_event
= perf_event_alloc(&parent_event
->attr
,
5302 parent_event
->cpu
, child_ctx
,
5303 group_leader
, parent_event
,
5305 if (IS_ERR(child_event
))
5310 * Make the child state follow the state of the parent event,
5311 * not its attr.disabled bit. We hold the parent's mutex,
5312 * so we won't race with perf_event_{en, dis}able_family.
5314 if (parent_event
->state
>= PERF_EVENT_STATE_INACTIVE
)
5315 child_event
->state
= PERF_EVENT_STATE_INACTIVE
;
5317 child_event
->state
= PERF_EVENT_STATE_OFF
;
5319 if (parent_event
->attr
.freq
) {
5320 u64 sample_period
= parent_event
->hw
.sample_period
;
5321 struct hw_perf_event
*hwc
= &child_event
->hw
;
5323 hwc
->sample_period
= sample_period
;
5324 hwc
->last_period
= sample_period
;
5326 local64_set(&hwc
->period_left
, sample_period
);
5329 child_event
->overflow_handler
= parent_event
->overflow_handler
;
5332 * Link it up in the child's context:
5334 add_event_to_ctx(child_event
, child_ctx
);
5337 * Get a reference to the parent filp - we will fput it
5338 * when the child event exits. This is safe to do because
5339 * we are in the parent and we know that the filp still
5340 * exists and has a nonzero count:
5342 atomic_long_inc(&parent_event
->filp
->f_count
);
5345 * Link this into the parent event's child list
5347 WARN_ON_ONCE(parent_event
->ctx
->parent_ctx
);
5348 mutex_lock(&parent_event
->child_mutex
);
5349 list_add_tail(&child_event
->child_list
, &parent_event
->child_list
);
5350 mutex_unlock(&parent_event
->child_mutex
);
5355 static int inherit_group(struct perf_event
*parent_event
,
5356 struct task_struct
*parent
,
5357 struct perf_event_context
*parent_ctx
,
5358 struct task_struct
*child
,
5359 struct perf_event_context
*child_ctx
)
5361 struct perf_event
*leader
;
5362 struct perf_event
*sub
;
5363 struct perf_event
*child_ctr
;
5365 leader
= inherit_event(parent_event
, parent
, parent_ctx
,
5366 child
, NULL
, child_ctx
);
5368 return PTR_ERR(leader
);
5369 list_for_each_entry(sub
, &parent_event
->sibling_list
, group_entry
) {
5370 child_ctr
= inherit_event(sub
, parent
, parent_ctx
,
5371 child
, leader
, child_ctx
);
5372 if (IS_ERR(child_ctr
))
5373 return PTR_ERR(child_ctr
);
5378 static void sync_child_event(struct perf_event
*child_event
,
5379 struct task_struct
*child
)
5381 struct perf_event
*parent_event
= child_event
->parent
;
5384 if (child_event
->attr
.inherit_stat
)
5385 perf_event_read_event(child_event
, child
);
5387 child_val
= perf_event_count(child_event
);
5390 * Add back the child's count to the parent's count:
5392 atomic64_add(child_val
, &parent_event
->child_count
);
5393 atomic64_add(child_event
->total_time_enabled
,
5394 &parent_event
->child_total_time_enabled
);
5395 atomic64_add(child_event
->total_time_running
,
5396 &parent_event
->child_total_time_running
);
5399 * Remove this event from the parent's list
5401 WARN_ON_ONCE(parent_event
->ctx
->parent_ctx
);
5402 mutex_lock(&parent_event
->child_mutex
);
5403 list_del_init(&child_event
->child_list
);
5404 mutex_unlock(&parent_event
->child_mutex
);
5407 * Release the parent event, if this was the last
5410 fput(parent_event
->filp
);
5414 __perf_event_exit_task(struct perf_event
*child_event
,
5415 struct perf_event_context
*child_ctx
,
5416 struct task_struct
*child
)
5418 struct perf_event
*parent_event
;
5420 perf_event_remove_from_context(child_event
);
5422 parent_event
= child_event
->parent
;
5424 * It can happen that parent exits first, and has events
5425 * that are still around due to the child reference. These
5426 * events need to be zapped - but otherwise linger.
5429 sync_child_event(child_event
, child
);
5430 free_event(child_event
);
5435 * When a child task exits, feed back event values to parent events.
5437 void perf_event_exit_task(struct task_struct
*child
)
5439 struct perf_event
*child_event
, *tmp
;
5440 struct perf_event_context
*child_ctx
;
5441 unsigned long flags
;
5443 if (likely(!child
->perf_event_ctxp
)) {
5444 perf_event_task(child
, NULL
, 0);
5448 local_irq_save(flags
);
5450 * We can't reschedule here because interrupts are disabled,
5451 * and either child is current or it is a task that can't be
5452 * scheduled, so we are now safe from rescheduling changing
5455 child_ctx
= child
->perf_event_ctxp
;
5456 __perf_event_task_sched_out(child_ctx
);
5459 * Take the context lock here so that if find_get_context is
5460 * reading child->perf_event_ctxp, we wait until it has
5461 * incremented the context's refcount before we do put_ctx below.
5463 raw_spin_lock(&child_ctx
->lock
);
5464 child
->perf_event_ctxp
= NULL
;
5466 * If this context is a clone; unclone it so it can't get
5467 * swapped to another process while we're removing all
5468 * the events from it.
5470 unclone_ctx(child_ctx
);
5471 update_context_time(child_ctx
);
5472 raw_spin_unlock_irqrestore(&child_ctx
->lock
, flags
);
5475 * Report the task dead after unscheduling the events so that we
5476 * won't get any samples after PERF_RECORD_EXIT. We can however still
5477 * get a few PERF_RECORD_READ events.
5479 perf_event_task(child
, child_ctx
, 0);
5482 * We can recurse on the same lock type through:
5484 * __perf_event_exit_task()
5485 * sync_child_event()
5486 * fput(parent_event->filp)
5488 * mutex_lock(&ctx->mutex)
5490 * But since its the parent context it won't be the same instance.
5492 mutex_lock(&child_ctx
->mutex
);
5495 list_for_each_entry_safe(child_event
, tmp
, &child_ctx
->pinned_groups
,
5497 __perf_event_exit_task(child_event
, child_ctx
, child
);
5499 list_for_each_entry_safe(child_event
, tmp
, &child_ctx
->flexible_groups
,
5501 __perf_event_exit_task(child_event
, child_ctx
, child
);
5504 * If the last event was a group event, it will have appended all
5505 * its siblings to the list, but we obtained 'tmp' before that which
5506 * will still point to the list head terminating the iteration.
5508 if (!list_empty(&child_ctx
->pinned_groups
) ||
5509 !list_empty(&child_ctx
->flexible_groups
))
5512 mutex_unlock(&child_ctx
->mutex
);
5517 static void perf_free_event(struct perf_event
*event
,
5518 struct perf_event_context
*ctx
)
5520 struct perf_event
*parent
= event
->parent
;
5522 if (WARN_ON_ONCE(!parent
))
5525 mutex_lock(&parent
->child_mutex
);
5526 list_del_init(&event
->child_list
);
5527 mutex_unlock(&parent
->child_mutex
);
5531 perf_group_detach(event
);
5532 list_del_event(event
, ctx
);
5537 * free an unexposed, unused context as created by inheritance by
5538 * init_task below, used by fork() in case of fail.
5540 void perf_event_free_task(struct task_struct
*task
)
5542 struct perf_event_context
*ctx
= task
->perf_event_ctxp
;
5543 struct perf_event
*event
, *tmp
;
5548 mutex_lock(&ctx
->mutex
);
5550 list_for_each_entry_safe(event
, tmp
, &ctx
->pinned_groups
, group_entry
)
5551 perf_free_event(event
, ctx
);
5553 list_for_each_entry_safe(event
, tmp
, &ctx
->flexible_groups
,
5555 perf_free_event(event
, ctx
);
5557 if (!list_empty(&ctx
->pinned_groups
) ||
5558 !list_empty(&ctx
->flexible_groups
))
5561 mutex_unlock(&ctx
->mutex
);
5567 inherit_task_group(struct perf_event
*event
, struct task_struct
*parent
,
5568 struct perf_event_context
*parent_ctx
,
5569 struct task_struct
*child
,
5573 struct perf_event_context
*child_ctx
= child
->perf_event_ctxp
;
5575 if (!event
->attr
.inherit
) {
5582 * This is executed from the parent task context, so
5583 * inherit events that have been marked for cloning.
5584 * First allocate and initialize a context for the
5588 child_ctx
= kzalloc(sizeof(struct perf_event_context
),
5593 __perf_event_init_context(child_ctx
, child
);
5594 child
->perf_event_ctxp
= child_ctx
;
5595 get_task_struct(child
);
5598 ret
= inherit_group(event
, parent
, parent_ctx
,
5609 * Initialize the perf_event context in task_struct
5611 int perf_event_init_task(struct task_struct
*child
)
5613 struct perf_event_context
*child_ctx
, *parent_ctx
;
5614 struct perf_event_context
*cloned_ctx
;
5615 struct perf_event
*event
;
5616 struct task_struct
*parent
= current
;
5617 int inherited_all
= 1;
5618 unsigned long flags
;
5621 child
->perf_event_ctxp
= NULL
;
5623 mutex_init(&child
->perf_event_mutex
);
5624 INIT_LIST_HEAD(&child
->perf_event_list
);
5626 if (likely(!parent
->perf_event_ctxp
))
5630 * If the parent's context is a clone, pin it so it won't get
5633 parent_ctx
= perf_pin_task_context(parent
);
5636 * No need to check if parent_ctx != NULL here; since we saw
5637 * it non-NULL earlier, the only reason for it to become NULL
5638 * is if we exit, and since we're currently in the middle of
5639 * a fork we can't be exiting at the same time.
5643 * Lock the parent list. No need to lock the child - not PID
5644 * hashed yet and not running, so nobody can access it.
5646 mutex_lock(&parent_ctx
->mutex
);
5649 * We dont have to disable NMIs - we are only looking at
5650 * the list, not manipulating it:
5652 list_for_each_entry(event
, &parent_ctx
->pinned_groups
, group_entry
) {
5653 ret
= inherit_task_group(event
, parent
, parent_ctx
, child
,
5660 * We can't hold ctx->lock when iterating the ->flexible_group list due
5661 * to allocations, but we need to prevent rotation because
5662 * rotate_ctx() will change the list from interrupt context.
5664 raw_spin_lock_irqsave(&parent_ctx
->lock
, flags
);
5665 parent_ctx
->rotate_disable
= 1;
5666 raw_spin_unlock_irqrestore(&parent_ctx
->lock
, flags
);
5668 list_for_each_entry(event
, &parent_ctx
->flexible_groups
, group_entry
) {
5669 ret
= inherit_task_group(event
, parent
, parent_ctx
, child
,
5675 raw_spin_lock_irqsave(&parent_ctx
->lock
, flags
);
5676 parent_ctx
->rotate_disable
= 0;
5677 raw_spin_unlock_irqrestore(&parent_ctx
->lock
, flags
);
5679 child_ctx
= child
->perf_event_ctxp
;
5681 if (child_ctx
&& inherited_all
) {
5683 * Mark the child context as a clone of the parent
5684 * context, or of whatever the parent is a clone of.
5685 * Note that if the parent is a clone, it could get
5686 * uncloned at any point, but that doesn't matter
5687 * because the list of events and the generation
5688 * count can't have changed since we took the mutex.
5690 cloned_ctx
= rcu_dereference(parent_ctx
->parent_ctx
);
5692 child_ctx
->parent_ctx
= cloned_ctx
;
5693 child_ctx
->parent_gen
= parent_ctx
->parent_gen
;
5695 child_ctx
->parent_ctx
= parent_ctx
;
5696 child_ctx
->parent_gen
= parent_ctx
->generation
;
5698 get_ctx(child_ctx
->parent_ctx
);
5701 mutex_unlock(&parent_ctx
->mutex
);
5703 perf_unpin_context(parent_ctx
);
5708 static void __init
perf_event_init_all_cpus(void)
5711 struct perf_cpu_context
*cpuctx
;
5713 for_each_possible_cpu(cpu
) {
5714 cpuctx
= &per_cpu(perf_cpu_context
, cpu
);
5715 mutex_init(&cpuctx
->hlist_mutex
);
5716 __perf_event_init_context(&cpuctx
->ctx
, NULL
);
5720 static void __cpuinit
perf_event_init_cpu(int cpu
)
5722 struct perf_cpu_context
*cpuctx
;
5724 cpuctx
= &per_cpu(perf_cpu_context
, cpu
);
5726 spin_lock(&perf_resource_lock
);
5727 cpuctx
->max_pertask
= perf_max_events
- perf_reserved_percpu
;
5728 spin_unlock(&perf_resource_lock
);
5730 mutex_lock(&cpuctx
->hlist_mutex
);
5731 if (cpuctx
->hlist_refcount
> 0) {
5732 struct swevent_hlist
*hlist
;
5734 hlist
= kzalloc(sizeof(*hlist
), GFP_KERNEL
);
5735 WARN_ON_ONCE(!hlist
);
5736 rcu_assign_pointer(cpuctx
->swevent_hlist
, hlist
);
5738 mutex_unlock(&cpuctx
->hlist_mutex
);
5741 #ifdef CONFIG_HOTPLUG_CPU
5742 static void __perf_event_exit_cpu(void *info
)
5744 struct perf_cpu_context
*cpuctx
= &__get_cpu_var(perf_cpu_context
);
5745 struct perf_event_context
*ctx
= &cpuctx
->ctx
;
5746 struct perf_event
*event
, *tmp
;
5748 list_for_each_entry_safe(event
, tmp
, &ctx
->pinned_groups
, group_entry
)
5749 __perf_event_remove_from_context(event
);
5750 list_for_each_entry_safe(event
, tmp
, &ctx
->flexible_groups
, group_entry
)
5751 __perf_event_remove_from_context(event
);
5753 static void perf_event_exit_cpu(int cpu
)
5755 struct perf_cpu_context
*cpuctx
= &per_cpu(perf_cpu_context
, cpu
);
5756 struct perf_event_context
*ctx
= &cpuctx
->ctx
;
5758 mutex_lock(&cpuctx
->hlist_mutex
);
5759 swevent_hlist_release(cpuctx
);
5760 mutex_unlock(&cpuctx
->hlist_mutex
);
5762 mutex_lock(&ctx
->mutex
);
5763 smp_call_function_single(cpu
, __perf_event_exit_cpu
, NULL
, 1);
5764 mutex_unlock(&ctx
->mutex
);
5767 static inline void perf_event_exit_cpu(int cpu
) { }
5770 static int __cpuinit
5771 perf_cpu_notify(struct notifier_block
*self
, unsigned long action
, void *hcpu
)
5773 unsigned int cpu
= (long)hcpu
;
5775 switch (action
& ~CPU_TASKS_FROZEN
) {
5777 case CPU_UP_PREPARE
:
5778 case CPU_DOWN_FAILED
:
5779 perf_event_init_cpu(cpu
);
5782 case CPU_UP_CANCELED
:
5783 case CPU_DOWN_PREPARE
:
5784 perf_event_exit_cpu(cpu
);
5795 * This has to have a higher priority than migration_notifier in sched.c.
5797 static struct notifier_block __cpuinitdata perf_cpu_nb
= {
5798 .notifier_call
= perf_cpu_notify
,
5802 void __init
perf_event_init(void)
5804 perf_event_init_all_cpus();
5805 perf_cpu_notify(&perf_cpu_nb
, (unsigned long)CPU_UP_PREPARE
,
5806 (void *)(long)smp_processor_id());
5807 perf_cpu_notify(&perf_cpu_nb
, (unsigned long)CPU_ONLINE
,
5808 (void *)(long)smp_processor_id());
5809 register_cpu_notifier(&perf_cpu_nb
);
5812 static ssize_t
perf_show_reserve_percpu(struct sysdev_class
*class,
5813 struct sysdev_class_attribute
*attr
,
5816 return sprintf(buf
, "%d\n", perf_reserved_percpu
);
5820 perf_set_reserve_percpu(struct sysdev_class
*class,
5821 struct sysdev_class_attribute
*attr
,
5825 struct perf_cpu_context
*cpuctx
;
5829 err
= strict_strtoul(buf
, 10, &val
);
5832 if (val
> perf_max_events
)
5835 spin_lock(&perf_resource_lock
);
5836 perf_reserved_percpu
= val
;
5837 for_each_online_cpu(cpu
) {
5838 cpuctx
= &per_cpu(perf_cpu_context
, cpu
);
5839 raw_spin_lock_irq(&cpuctx
->ctx
.lock
);
5840 mpt
= min(perf_max_events
- cpuctx
->ctx
.nr_events
,
5841 perf_max_events
- perf_reserved_percpu
);
5842 cpuctx
->max_pertask
= mpt
;
5843 raw_spin_unlock_irq(&cpuctx
->ctx
.lock
);
5845 spin_unlock(&perf_resource_lock
);
5850 static ssize_t
perf_show_overcommit(struct sysdev_class
*class,
5851 struct sysdev_class_attribute
*attr
,
5854 return sprintf(buf
, "%d\n", perf_overcommit
);
5858 perf_set_overcommit(struct sysdev_class
*class,
5859 struct sysdev_class_attribute
*attr
,
5860 const char *buf
, size_t count
)
5865 err
= strict_strtoul(buf
, 10, &val
);
5871 spin_lock(&perf_resource_lock
);
5872 perf_overcommit
= val
;
5873 spin_unlock(&perf_resource_lock
);
5878 static SYSDEV_CLASS_ATTR(
5881 perf_show_reserve_percpu
,
5882 perf_set_reserve_percpu
5885 static SYSDEV_CLASS_ATTR(
5888 perf_show_overcommit
,
5892 static struct attribute
*perfclass_attrs
[] = {
5893 &attr_reserve_percpu
.attr
,
5894 &attr_overcommit
.attr
,
5898 static struct attribute_group perfclass_attr_group
= {
5899 .attrs
= perfclass_attrs
,
5900 .name
= "perf_events",
5903 static int __init
perf_event_sysfs_init(void)
5905 return sysfs_create_group(&cpu_sysdev_class
.kset
.kobj
,
5906 &perfclass_attr_group
);
5908 device_initcall(perf_event_sysfs_init
);