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 cpu_clock(raw_smp_processor_id());
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_sched_out(struct perf_event
*event
,
407 struct perf_cpu_context
*cpuctx
,
408 struct perf_event_context
*ctx
)
410 if (event
->state
!= PERF_EVENT_STATE_ACTIVE
)
413 event
->state
= PERF_EVENT_STATE_INACTIVE
;
414 if (event
->pending_disable
) {
415 event
->pending_disable
= 0;
416 event
->state
= PERF_EVENT_STATE_OFF
;
418 event
->tstamp_stopped
= ctx
->time
;
419 event
->pmu
->disable(event
);
422 if (!is_software_event(event
))
423 cpuctx
->active_oncpu
--;
425 if (event
->attr
.exclusive
|| !cpuctx
->active_oncpu
)
426 cpuctx
->exclusive
= 0;
430 group_sched_out(struct perf_event
*group_event
,
431 struct perf_cpu_context
*cpuctx
,
432 struct perf_event_context
*ctx
)
434 struct perf_event
*event
;
436 if (group_event
->state
!= PERF_EVENT_STATE_ACTIVE
)
439 event_sched_out(group_event
, cpuctx
, ctx
);
442 * Schedule out siblings (if any):
444 list_for_each_entry(event
, &group_event
->sibling_list
, group_entry
)
445 event_sched_out(event
, cpuctx
, ctx
);
447 if (group_event
->attr
.exclusive
)
448 cpuctx
->exclusive
= 0;
452 * Cross CPU call to remove a performance event
454 * We disable the event on the hardware level first. After that we
455 * remove it from the context list.
457 static void __perf_event_remove_from_context(void *info
)
459 struct perf_cpu_context
*cpuctx
= &__get_cpu_var(perf_cpu_context
);
460 struct perf_event
*event
= info
;
461 struct perf_event_context
*ctx
= event
->ctx
;
464 * If this is a task context, we need to check whether it is
465 * the current task context of this cpu. If not it has been
466 * scheduled out before the smp call arrived.
468 if (ctx
->task
&& cpuctx
->task_ctx
!= ctx
)
471 raw_spin_lock(&ctx
->lock
);
473 * Protect the list operation against NMI by disabling the
474 * events on a global level.
478 event_sched_out(event
, cpuctx
, ctx
);
480 list_del_event(event
, ctx
);
484 * Allow more per task events with respect to the
487 cpuctx
->max_pertask
=
488 min(perf_max_events
- ctx
->nr_events
,
489 perf_max_events
- perf_reserved_percpu
);
493 raw_spin_unlock(&ctx
->lock
);
498 * Remove the event from a task's (or a CPU's) list of events.
500 * Must be called with ctx->mutex held.
502 * CPU events are removed with a smp call. For task events we only
503 * call when the task is on a CPU.
505 * If event->ctx is a cloned context, callers must make sure that
506 * every task struct that event->ctx->task could possibly point to
507 * remains valid. This is OK when called from perf_release since
508 * that only calls us on the top-level context, which can't be a clone.
509 * When called from perf_event_exit_task, it's OK because the
510 * context has been detached from its task.
512 static void perf_event_remove_from_context(struct perf_event
*event
)
514 struct perf_event_context
*ctx
= event
->ctx
;
515 struct task_struct
*task
= ctx
->task
;
519 * Per cpu events are removed via an smp call and
520 * the removal is always successful.
522 smp_call_function_single(event
->cpu
,
523 __perf_event_remove_from_context
,
529 task_oncpu_function_call(task
, __perf_event_remove_from_context
,
532 raw_spin_lock_irq(&ctx
->lock
);
534 * If the context is active we need to retry the smp call.
536 if (ctx
->nr_active
&& !list_empty(&event
->group_entry
)) {
537 raw_spin_unlock_irq(&ctx
->lock
);
542 * The lock prevents that this context is scheduled in so we
543 * can remove the event safely, if the call above did not
546 if (!list_empty(&event
->group_entry
))
547 list_del_event(event
, ctx
);
548 raw_spin_unlock_irq(&ctx
->lock
);
552 * Cross CPU call to disable a performance event
554 static void __perf_event_disable(void *info
)
556 struct perf_event
*event
= info
;
557 struct perf_cpu_context
*cpuctx
= &__get_cpu_var(perf_cpu_context
);
558 struct perf_event_context
*ctx
= event
->ctx
;
561 * If this is a per-task event, need to check whether this
562 * event's task is the current task on this cpu.
564 if (ctx
->task
&& cpuctx
->task_ctx
!= ctx
)
567 raw_spin_lock(&ctx
->lock
);
570 * If the event is on, turn it off.
571 * If it is in error state, leave it in error state.
573 if (event
->state
>= PERF_EVENT_STATE_INACTIVE
) {
574 update_context_time(ctx
);
575 update_group_times(event
);
576 if (event
== event
->group_leader
)
577 group_sched_out(event
, cpuctx
, ctx
);
579 event_sched_out(event
, cpuctx
, ctx
);
580 event
->state
= PERF_EVENT_STATE_OFF
;
583 raw_spin_unlock(&ctx
->lock
);
589 * If event->ctx is a cloned context, callers must make sure that
590 * every task struct that event->ctx->task could possibly point to
591 * remains valid. This condition is satisifed when called through
592 * perf_event_for_each_child or perf_event_for_each because they
593 * hold the top-level event's child_mutex, so any descendant that
594 * goes to exit will block in sync_child_event.
595 * When called from perf_pending_event it's OK because event->ctx
596 * is the current context on this CPU and preemption is disabled,
597 * hence we can't get into perf_event_task_sched_out for this context.
599 void perf_event_disable(struct perf_event
*event
)
601 struct perf_event_context
*ctx
= event
->ctx
;
602 struct task_struct
*task
= ctx
->task
;
606 * Disable the event on the cpu that it's on
608 smp_call_function_single(event
->cpu
, __perf_event_disable
,
614 task_oncpu_function_call(task
, __perf_event_disable
, event
);
616 raw_spin_lock_irq(&ctx
->lock
);
618 * If the event is still active, we need to retry the cross-call.
620 if (event
->state
== PERF_EVENT_STATE_ACTIVE
) {
621 raw_spin_unlock_irq(&ctx
->lock
);
626 * Since we have the lock this context can't be scheduled
627 * in, so we can change the state safely.
629 if (event
->state
== PERF_EVENT_STATE_INACTIVE
) {
630 update_group_times(event
);
631 event
->state
= PERF_EVENT_STATE_OFF
;
634 raw_spin_unlock_irq(&ctx
->lock
);
638 event_sched_in(struct perf_event
*event
,
639 struct perf_cpu_context
*cpuctx
,
640 struct perf_event_context
*ctx
)
642 if (event
->state
<= PERF_EVENT_STATE_OFF
)
645 event
->state
= PERF_EVENT_STATE_ACTIVE
;
646 event
->oncpu
= smp_processor_id();
648 * The new state must be visible before we turn it on in the hardware:
652 if (event
->pmu
->enable(event
)) {
653 event
->state
= PERF_EVENT_STATE_INACTIVE
;
658 event
->tstamp_running
+= ctx
->time
- event
->tstamp_stopped
;
660 if (!is_software_event(event
))
661 cpuctx
->active_oncpu
++;
664 if (event
->attr
.exclusive
)
665 cpuctx
->exclusive
= 1;
671 group_sched_in(struct perf_event
*group_event
,
672 struct perf_cpu_context
*cpuctx
,
673 struct perf_event_context
*ctx
)
675 struct perf_event
*event
, *partial_group
= NULL
;
676 const struct pmu
*pmu
= group_event
->pmu
;
679 if (group_event
->state
== PERF_EVENT_STATE_OFF
)
682 /* Check if group transaction availabe */
689 if (event_sched_in(group_event
, cpuctx
, ctx
)) {
691 pmu
->cancel_txn(pmu
);
696 * Schedule in siblings as one group (if any):
698 list_for_each_entry(event
, &group_event
->sibling_list
, group_entry
) {
699 if (event_sched_in(event
, cpuctx
, ctx
)) {
700 partial_group
= event
;
705 if (!txn
|| !pmu
->commit_txn(pmu
))
710 * Groups can be scheduled in as one unit only, so undo any
711 * partial group before returning:
713 list_for_each_entry(event
, &group_event
->sibling_list
, group_entry
) {
714 if (event
== partial_group
)
716 event_sched_out(event
, cpuctx
, ctx
);
718 event_sched_out(group_event
, cpuctx
, ctx
);
721 pmu
->cancel_txn(pmu
);
727 * Work out whether we can put this event group on the CPU now.
729 static int group_can_go_on(struct perf_event
*event
,
730 struct perf_cpu_context
*cpuctx
,
734 * Groups consisting entirely of software events can always go on.
736 if (event
->group_flags
& PERF_GROUP_SOFTWARE
)
739 * If an exclusive group is already on, no other hardware
742 if (cpuctx
->exclusive
)
745 * If this group is exclusive and there are already
746 * events on the CPU, it can't go on.
748 if (event
->attr
.exclusive
&& cpuctx
->active_oncpu
)
751 * Otherwise, try to add it if all previous groups were able
757 static void add_event_to_ctx(struct perf_event
*event
,
758 struct perf_event_context
*ctx
)
760 list_add_event(event
, ctx
);
761 perf_group_attach(event
);
762 event
->tstamp_enabled
= ctx
->time
;
763 event
->tstamp_running
= ctx
->time
;
764 event
->tstamp_stopped
= ctx
->time
;
768 * Cross CPU call to install and enable a performance event
770 * Must be called with ctx->mutex held
772 static void __perf_install_in_context(void *info
)
774 struct perf_cpu_context
*cpuctx
= &__get_cpu_var(perf_cpu_context
);
775 struct perf_event
*event
= info
;
776 struct perf_event_context
*ctx
= event
->ctx
;
777 struct perf_event
*leader
= event
->group_leader
;
781 * If this is a task context, we need to check whether it is
782 * the current task context of this cpu. If not it has been
783 * scheduled out before the smp call arrived.
784 * Or possibly this is the right context but it isn't
785 * on this cpu because it had no events.
787 if (ctx
->task
&& cpuctx
->task_ctx
!= ctx
) {
788 if (cpuctx
->task_ctx
|| ctx
->task
!= current
)
790 cpuctx
->task_ctx
= ctx
;
793 raw_spin_lock(&ctx
->lock
);
795 update_context_time(ctx
);
798 * Protect the list operation against NMI by disabling the
799 * events on a global level. NOP for non NMI based events.
803 add_event_to_ctx(event
, ctx
);
805 if (event
->cpu
!= -1 && event
->cpu
!= smp_processor_id())
809 * Don't put the event on if it is disabled or if
810 * it is in a group and the group isn't on.
812 if (event
->state
!= PERF_EVENT_STATE_INACTIVE
||
813 (leader
!= event
&& leader
->state
!= PERF_EVENT_STATE_ACTIVE
))
817 * An exclusive event can't go on if there are already active
818 * hardware events, and no hardware event can go on if there
819 * is already an exclusive event on.
821 if (!group_can_go_on(event
, cpuctx
, 1))
824 err
= event_sched_in(event
, cpuctx
, ctx
);
828 * This event couldn't go on. If it is in a group
829 * then we have to pull the whole group off.
830 * If the event group is pinned then put it in error state.
833 group_sched_out(leader
, cpuctx
, ctx
);
834 if (leader
->attr
.pinned
) {
835 update_group_times(leader
);
836 leader
->state
= PERF_EVENT_STATE_ERROR
;
840 if (!err
&& !ctx
->task
&& cpuctx
->max_pertask
)
841 cpuctx
->max_pertask
--;
846 raw_spin_unlock(&ctx
->lock
);
850 * Attach a performance event to a context
852 * First we add the event to the list with the hardware enable bit
853 * in event->hw_config cleared.
855 * If the event is attached to a task which is on a CPU we use a smp
856 * call to enable it in the task context. The task might have been
857 * scheduled away, but we check this in the smp call again.
859 * Must be called with ctx->mutex held.
862 perf_install_in_context(struct perf_event_context
*ctx
,
863 struct perf_event
*event
,
866 struct task_struct
*task
= ctx
->task
;
870 * Per cpu events are installed via an smp call and
871 * the install is always successful.
873 smp_call_function_single(cpu
, __perf_install_in_context
,
879 task_oncpu_function_call(task
, __perf_install_in_context
,
882 raw_spin_lock_irq(&ctx
->lock
);
884 * we need to retry the smp call.
886 if (ctx
->is_active
&& list_empty(&event
->group_entry
)) {
887 raw_spin_unlock_irq(&ctx
->lock
);
892 * The lock prevents that this context is scheduled in so we
893 * can add the event safely, if it the call above did not
896 if (list_empty(&event
->group_entry
))
897 add_event_to_ctx(event
, ctx
);
898 raw_spin_unlock_irq(&ctx
->lock
);
902 * Put a event into inactive state and update time fields.
903 * Enabling the leader of a group effectively enables all
904 * the group members that aren't explicitly disabled, so we
905 * have to update their ->tstamp_enabled also.
906 * Note: this works for group members as well as group leaders
907 * since the non-leader members' sibling_lists will be empty.
909 static void __perf_event_mark_enabled(struct perf_event
*event
,
910 struct perf_event_context
*ctx
)
912 struct perf_event
*sub
;
914 event
->state
= PERF_EVENT_STATE_INACTIVE
;
915 event
->tstamp_enabled
= ctx
->time
- event
->total_time_enabled
;
916 list_for_each_entry(sub
, &event
->sibling_list
, group_entry
)
917 if (sub
->state
>= PERF_EVENT_STATE_INACTIVE
)
918 sub
->tstamp_enabled
=
919 ctx
->time
- sub
->total_time_enabled
;
923 * Cross CPU call to enable a performance event
925 static void __perf_event_enable(void *info
)
927 struct perf_event
*event
= info
;
928 struct perf_cpu_context
*cpuctx
= &__get_cpu_var(perf_cpu_context
);
929 struct perf_event_context
*ctx
= event
->ctx
;
930 struct perf_event
*leader
= event
->group_leader
;
934 * If this is a per-task event, need to check whether this
935 * event's task is the current task on this cpu.
937 if (ctx
->task
&& cpuctx
->task_ctx
!= ctx
) {
938 if (cpuctx
->task_ctx
|| ctx
->task
!= current
)
940 cpuctx
->task_ctx
= ctx
;
943 raw_spin_lock(&ctx
->lock
);
945 update_context_time(ctx
);
947 if (event
->state
>= PERF_EVENT_STATE_INACTIVE
)
949 __perf_event_mark_enabled(event
, ctx
);
951 if (event
->cpu
!= -1 && event
->cpu
!= smp_processor_id())
955 * If the event is in a group and isn't the group leader,
956 * then don't put it on unless the group is on.
958 if (leader
!= event
&& leader
->state
!= PERF_EVENT_STATE_ACTIVE
)
961 if (!group_can_go_on(event
, cpuctx
, 1)) {
966 err
= group_sched_in(event
, cpuctx
, ctx
);
968 err
= event_sched_in(event
, cpuctx
, ctx
);
974 * If this event can't go on and it's part of a
975 * group, then the whole group has to come off.
978 group_sched_out(leader
, cpuctx
, ctx
);
979 if (leader
->attr
.pinned
) {
980 update_group_times(leader
);
981 leader
->state
= PERF_EVENT_STATE_ERROR
;
986 raw_spin_unlock(&ctx
->lock
);
992 * If event->ctx is a cloned context, callers must make sure that
993 * every task struct that event->ctx->task could possibly point to
994 * remains valid. This condition is satisfied when called through
995 * perf_event_for_each_child or perf_event_for_each as described
996 * for perf_event_disable.
998 void perf_event_enable(struct perf_event
*event
)
1000 struct perf_event_context
*ctx
= event
->ctx
;
1001 struct task_struct
*task
= ctx
->task
;
1005 * Enable the event on the cpu that it's on
1007 smp_call_function_single(event
->cpu
, __perf_event_enable
,
1012 raw_spin_lock_irq(&ctx
->lock
);
1013 if (event
->state
>= PERF_EVENT_STATE_INACTIVE
)
1017 * If the event is in error state, clear that first.
1018 * That way, if we see the event in error state below, we
1019 * know that it has gone back into error state, as distinct
1020 * from the task having been scheduled away before the
1021 * cross-call arrived.
1023 if (event
->state
== PERF_EVENT_STATE_ERROR
)
1024 event
->state
= PERF_EVENT_STATE_OFF
;
1027 raw_spin_unlock_irq(&ctx
->lock
);
1028 task_oncpu_function_call(task
, __perf_event_enable
, event
);
1030 raw_spin_lock_irq(&ctx
->lock
);
1033 * If the context is active and the event is still off,
1034 * we need to retry the cross-call.
1036 if (ctx
->is_active
&& event
->state
== PERF_EVENT_STATE_OFF
)
1040 * Since we have the lock this context can't be scheduled
1041 * in, so we can change the state safely.
1043 if (event
->state
== PERF_EVENT_STATE_OFF
)
1044 __perf_event_mark_enabled(event
, ctx
);
1047 raw_spin_unlock_irq(&ctx
->lock
);
1050 static int perf_event_refresh(struct perf_event
*event
, int refresh
)
1053 * not supported on inherited events
1055 if (event
->attr
.inherit
)
1058 atomic_add(refresh
, &event
->event_limit
);
1059 perf_event_enable(event
);
1065 EVENT_FLEXIBLE
= 0x1,
1067 EVENT_ALL
= EVENT_FLEXIBLE
| EVENT_PINNED
,
1070 static void ctx_sched_out(struct perf_event_context
*ctx
,
1071 struct perf_cpu_context
*cpuctx
,
1072 enum event_type_t event_type
)
1074 struct perf_event
*event
;
1076 raw_spin_lock(&ctx
->lock
);
1078 if (likely(!ctx
->nr_events
))
1080 update_context_time(ctx
);
1083 if (!ctx
->nr_active
)
1086 if (event_type
& EVENT_PINNED
)
1087 list_for_each_entry(event
, &ctx
->pinned_groups
, group_entry
)
1088 group_sched_out(event
, cpuctx
, ctx
);
1090 if (event_type
& EVENT_FLEXIBLE
)
1091 list_for_each_entry(event
, &ctx
->flexible_groups
, group_entry
)
1092 group_sched_out(event
, cpuctx
, ctx
);
1097 raw_spin_unlock(&ctx
->lock
);
1101 * Test whether two contexts are equivalent, i.e. whether they
1102 * have both been cloned from the same version of the same context
1103 * and they both have the same number of enabled events.
1104 * If the number of enabled events is the same, then the set
1105 * of enabled events should be the same, because these are both
1106 * inherited contexts, therefore we can't access individual events
1107 * in them directly with an fd; we can only enable/disable all
1108 * events via prctl, or enable/disable all events in a family
1109 * via ioctl, which will have the same effect on both contexts.
1111 static int context_equiv(struct perf_event_context
*ctx1
,
1112 struct perf_event_context
*ctx2
)
1114 return ctx1
->parent_ctx
&& ctx1
->parent_ctx
== ctx2
->parent_ctx
1115 && ctx1
->parent_gen
== ctx2
->parent_gen
1116 && !ctx1
->pin_count
&& !ctx2
->pin_count
;
1119 static void __perf_event_sync_stat(struct perf_event
*event
,
1120 struct perf_event
*next_event
)
1124 if (!event
->attr
.inherit_stat
)
1128 * Update the event value, we cannot use perf_event_read()
1129 * because we're in the middle of a context switch and have IRQs
1130 * disabled, which upsets smp_call_function_single(), however
1131 * we know the event must be on the current CPU, therefore we
1132 * don't need to use it.
1134 switch (event
->state
) {
1135 case PERF_EVENT_STATE_ACTIVE
:
1136 event
->pmu
->read(event
);
1139 case PERF_EVENT_STATE_INACTIVE
:
1140 update_event_times(event
);
1148 * In order to keep per-task stats reliable we need to flip the event
1149 * values when we flip the contexts.
1151 value
= local64_read(&next_event
->count
);
1152 value
= local64_xchg(&event
->count
, value
);
1153 local64_set(&next_event
->count
, value
);
1155 swap(event
->total_time_enabled
, next_event
->total_time_enabled
);
1156 swap(event
->total_time_running
, next_event
->total_time_running
);
1159 * Since we swizzled the values, update the user visible data too.
1161 perf_event_update_userpage(event
);
1162 perf_event_update_userpage(next_event
);
1165 #define list_next_entry(pos, member) \
1166 list_entry(pos->member.next, typeof(*pos), member)
1168 static void perf_event_sync_stat(struct perf_event_context
*ctx
,
1169 struct perf_event_context
*next_ctx
)
1171 struct perf_event
*event
, *next_event
;
1176 update_context_time(ctx
);
1178 event
= list_first_entry(&ctx
->event_list
,
1179 struct perf_event
, event_entry
);
1181 next_event
= list_first_entry(&next_ctx
->event_list
,
1182 struct perf_event
, event_entry
);
1184 while (&event
->event_entry
!= &ctx
->event_list
&&
1185 &next_event
->event_entry
!= &next_ctx
->event_list
) {
1187 __perf_event_sync_stat(event
, next_event
);
1189 event
= list_next_entry(event
, event_entry
);
1190 next_event
= list_next_entry(next_event
, event_entry
);
1195 * Called from scheduler to remove the events of the current task,
1196 * with interrupts disabled.
1198 * We stop each event and update the event value in event->count.
1200 * This does not protect us against NMI, but disable()
1201 * sets the disabled bit in the control field of event _before_
1202 * accessing the event control register. If a NMI hits, then it will
1203 * not restart the event.
1205 void perf_event_task_sched_out(struct task_struct
*task
,
1206 struct task_struct
*next
)
1208 struct perf_cpu_context
*cpuctx
= &__get_cpu_var(perf_cpu_context
);
1209 struct perf_event_context
*ctx
= task
->perf_event_ctxp
;
1210 struct perf_event_context
*next_ctx
;
1211 struct perf_event_context
*parent
;
1214 perf_sw_event(PERF_COUNT_SW_CONTEXT_SWITCHES
, 1, 1, NULL
, 0);
1216 if (likely(!ctx
|| !cpuctx
->task_ctx
))
1220 parent
= rcu_dereference(ctx
->parent_ctx
);
1221 next_ctx
= next
->perf_event_ctxp
;
1222 if (parent
&& next_ctx
&&
1223 rcu_dereference(next_ctx
->parent_ctx
) == parent
) {
1225 * Looks like the two contexts are clones, so we might be
1226 * able to optimize the context switch. We lock both
1227 * contexts and check that they are clones under the
1228 * lock (including re-checking that neither has been
1229 * uncloned in the meantime). It doesn't matter which
1230 * order we take the locks because no other cpu could
1231 * be trying to lock both of these tasks.
1233 raw_spin_lock(&ctx
->lock
);
1234 raw_spin_lock_nested(&next_ctx
->lock
, SINGLE_DEPTH_NESTING
);
1235 if (context_equiv(ctx
, next_ctx
)) {
1237 * XXX do we need a memory barrier of sorts
1238 * wrt to rcu_dereference() of perf_event_ctxp
1240 task
->perf_event_ctxp
= next_ctx
;
1241 next
->perf_event_ctxp
= ctx
;
1243 next_ctx
->task
= task
;
1246 perf_event_sync_stat(ctx
, next_ctx
);
1248 raw_spin_unlock(&next_ctx
->lock
);
1249 raw_spin_unlock(&ctx
->lock
);
1254 ctx_sched_out(ctx
, cpuctx
, EVENT_ALL
);
1255 cpuctx
->task_ctx
= NULL
;
1259 static void task_ctx_sched_out(struct perf_event_context
*ctx
,
1260 enum event_type_t event_type
)
1262 struct perf_cpu_context
*cpuctx
= &__get_cpu_var(perf_cpu_context
);
1264 if (!cpuctx
->task_ctx
)
1267 if (WARN_ON_ONCE(ctx
!= cpuctx
->task_ctx
))
1270 ctx_sched_out(ctx
, cpuctx
, event_type
);
1271 cpuctx
->task_ctx
= NULL
;
1275 * Called with IRQs disabled
1277 static void __perf_event_task_sched_out(struct perf_event_context
*ctx
)
1279 task_ctx_sched_out(ctx
, EVENT_ALL
);
1283 * Called with IRQs disabled
1285 static void cpu_ctx_sched_out(struct perf_cpu_context
*cpuctx
,
1286 enum event_type_t event_type
)
1288 ctx_sched_out(&cpuctx
->ctx
, cpuctx
, event_type
);
1292 ctx_pinned_sched_in(struct perf_event_context
*ctx
,
1293 struct perf_cpu_context
*cpuctx
)
1295 struct perf_event
*event
;
1297 list_for_each_entry(event
, &ctx
->pinned_groups
, group_entry
) {
1298 if (event
->state
<= PERF_EVENT_STATE_OFF
)
1300 if (event
->cpu
!= -1 && event
->cpu
!= smp_processor_id())
1303 if (group_can_go_on(event
, cpuctx
, 1))
1304 group_sched_in(event
, cpuctx
, ctx
);
1307 * If this pinned group hasn't been scheduled,
1308 * put it in error state.
1310 if (event
->state
== PERF_EVENT_STATE_INACTIVE
) {
1311 update_group_times(event
);
1312 event
->state
= PERF_EVENT_STATE_ERROR
;
1318 ctx_flexible_sched_in(struct perf_event_context
*ctx
,
1319 struct perf_cpu_context
*cpuctx
)
1321 struct perf_event
*event
;
1324 list_for_each_entry(event
, &ctx
->flexible_groups
, group_entry
) {
1325 /* Ignore events in OFF or ERROR state */
1326 if (event
->state
<= PERF_EVENT_STATE_OFF
)
1329 * Listen to the 'cpu' scheduling filter constraint
1332 if (event
->cpu
!= -1 && event
->cpu
!= smp_processor_id())
1335 if (group_can_go_on(event
, cpuctx
, can_add_hw
))
1336 if (group_sched_in(event
, cpuctx
, ctx
))
1342 ctx_sched_in(struct perf_event_context
*ctx
,
1343 struct perf_cpu_context
*cpuctx
,
1344 enum event_type_t event_type
)
1346 raw_spin_lock(&ctx
->lock
);
1348 if (likely(!ctx
->nr_events
))
1351 ctx
->timestamp
= perf_clock();
1356 * First go through the list and put on any pinned groups
1357 * in order to give them the best chance of going on.
1359 if (event_type
& EVENT_PINNED
)
1360 ctx_pinned_sched_in(ctx
, cpuctx
);
1362 /* Then walk through the lower prio flexible groups */
1363 if (event_type
& EVENT_FLEXIBLE
)
1364 ctx_flexible_sched_in(ctx
, cpuctx
);
1368 raw_spin_unlock(&ctx
->lock
);
1371 static void cpu_ctx_sched_in(struct perf_cpu_context
*cpuctx
,
1372 enum event_type_t event_type
)
1374 struct perf_event_context
*ctx
= &cpuctx
->ctx
;
1376 ctx_sched_in(ctx
, cpuctx
, event_type
);
1379 static void task_ctx_sched_in(struct task_struct
*task
,
1380 enum event_type_t event_type
)
1382 struct perf_cpu_context
*cpuctx
= &__get_cpu_var(perf_cpu_context
);
1383 struct perf_event_context
*ctx
= task
->perf_event_ctxp
;
1387 if (cpuctx
->task_ctx
== ctx
)
1389 ctx_sched_in(ctx
, cpuctx
, event_type
);
1390 cpuctx
->task_ctx
= ctx
;
1393 * Called from scheduler to add the events of the current task
1394 * with interrupts disabled.
1396 * We restore the event value and then enable it.
1398 * This does not protect us against NMI, but enable()
1399 * sets the enabled bit in the control field of event _before_
1400 * accessing the event control register. If a NMI hits, then it will
1401 * keep the event running.
1403 void perf_event_task_sched_in(struct task_struct
*task
)
1405 struct perf_cpu_context
*cpuctx
= &__get_cpu_var(perf_cpu_context
);
1406 struct perf_event_context
*ctx
= task
->perf_event_ctxp
;
1411 if (cpuctx
->task_ctx
== ctx
)
1417 * We want to keep the following priority order:
1418 * cpu pinned (that don't need to move), task pinned,
1419 * cpu flexible, task flexible.
1421 cpu_ctx_sched_out(cpuctx
, EVENT_FLEXIBLE
);
1423 ctx_sched_in(ctx
, cpuctx
, EVENT_PINNED
);
1424 cpu_ctx_sched_in(cpuctx
, EVENT_FLEXIBLE
);
1425 ctx_sched_in(ctx
, cpuctx
, EVENT_FLEXIBLE
);
1427 cpuctx
->task_ctx
= ctx
;
1432 #define MAX_INTERRUPTS (~0ULL)
1434 static void perf_log_throttle(struct perf_event
*event
, int enable
);
1436 static u64
perf_calculate_period(struct perf_event
*event
, u64 nsec
, u64 count
)
1438 u64 frequency
= event
->attr
.sample_freq
;
1439 u64 sec
= NSEC_PER_SEC
;
1440 u64 divisor
, dividend
;
1442 int count_fls
, nsec_fls
, frequency_fls
, sec_fls
;
1444 count_fls
= fls64(count
);
1445 nsec_fls
= fls64(nsec
);
1446 frequency_fls
= fls64(frequency
);
1450 * We got @count in @nsec, with a target of sample_freq HZ
1451 * the target period becomes:
1454 * period = -------------------
1455 * @nsec * sample_freq
1460 * Reduce accuracy by one bit such that @a and @b converge
1461 * to a similar magnitude.
1463 #define REDUCE_FLS(a, b) \
1465 if (a##_fls > b##_fls) { \
1475 * Reduce accuracy until either term fits in a u64, then proceed with
1476 * the other, so that finally we can do a u64/u64 division.
1478 while (count_fls
+ sec_fls
> 64 && nsec_fls
+ frequency_fls
> 64) {
1479 REDUCE_FLS(nsec
, frequency
);
1480 REDUCE_FLS(sec
, count
);
1483 if (count_fls
+ sec_fls
> 64) {
1484 divisor
= nsec
* frequency
;
1486 while (count_fls
+ sec_fls
> 64) {
1487 REDUCE_FLS(count
, sec
);
1491 dividend
= count
* sec
;
1493 dividend
= count
* sec
;
1495 while (nsec_fls
+ frequency_fls
> 64) {
1496 REDUCE_FLS(nsec
, frequency
);
1500 divisor
= nsec
* frequency
;
1506 return div64_u64(dividend
, divisor
);
1509 static void perf_event_stop(struct perf_event
*event
)
1511 if (!event
->pmu
->stop
)
1512 return event
->pmu
->disable(event
);
1514 return event
->pmu
->stop(event
);
1517 static int perf_event_start(struct perf_event
*event
)
1519 if (!event
->pmu
->start
)
1520 return event
->pmu
->enable(event
);
1522 return event
->pmu
->start(event
);
1525 static void perf_adjust_period(struct perf_event
*event
, u64 nsec
, u64 count
)
1527 struct hw_perf_event
*hwc
= &event
->hw
;
1528 s64 period
, sample_period
;
1531 period
= perf_calculate_period(event
, nsec
, count
);
1533 delta
= (s64
)(period
- hwc
->sample_period
);
1534 delta
= (delta
+ 7) / 8; /* low pass filter */
1536 sample_period
= hwc
->sample_period
+ delta
;
1541 hwc
->sample_period
= sample_period
;
1543 if (local64_read(&hwc
->period_left
) > 8*sample_period
) {
1545 perf_event_stop(event
);
1546 local64_set(&hwc
->period_left
, 0);
1547 perf_event_start(event
);
1552 static void perf_ctx_adjust_freq(struct perf_event_context
*ctx
)
1554 struct perf_event
*event
;
1555 struct hw_perf_event
*hwc
;
1556 u64 interrupts
, now
;
1559 raw_spin_lock(&ctx
->lock
);
1560 list_for_each_entry_rcu(event
, &ctx
->event_list
, event_entry
) {
1561 if (event
->state
!= PERF_EVENT_STATE_ACTIVE
)
1564 if (event
->cpu
!= -1 && event
->cpu
!= smp_processor_id())
1569 interrupts
= hwc
->interrupts
;
1570 hwc
->interrupts
= 0;
1573 * unthrottle events on the tick
1575 if (interrupts
== MAX_INTERRUPTS
) {
1576 perf_log_throttle(event
, 1);
1578 event
->pmu
->unthrottle(event
);
1582 if (!event
->attr
.freq
|| !event
->attr
.sample_freq
)
1586 event
->pmu
->read(event
);
1587 now
= local64_read(&event
->count
);
1588 delta
= now
- hwc
->freq_count_stamp
;
1589 hwc
->freq_count_stamp
= now
;
1592 perf_adjust_period(event
, TICK_NSEC
, delta
);
1595 raw_spin_unlock(&ctx
->lock
);
1599 * Round-robin a context's events:
1601 static void rotate_ctx(struct perf_event_context
*ctx
)
1603 raw_spin_lock(&ctx
->lock
);
1605 /* Rotate the first entry last of non-pinned groups */
1606 list_rotate_left(&ctx
->flexible_groups
);
1608 raw_spin_unlock(&ctx
->lock
);
1611 void perf_event_task_tick(struct task_struct
*curr
)
1613 struct perf_cpu_context
*cpuctx
;
1614 struct perf_event_context
*ctx
;
1617 if (!atomic_read(&nr_events
))
1620 cpuctx
= &__get_cpu_var(perf_cpu_context
);
1621 if (cpuctx
->ctx
.nr_events
&&
1622 cpuctx
->ctx
.nr_events
!= cpuctx
->ctx
.nr_active
)
1625 ctx
= curr
->perf_event_ctxp
;
1626 if (ctx
&& ctx
->nr_events
&& ctx
->nr_events
!= ctx
->nr_active
)
1629 perf_ctx_adjust_freq(&cpuctx
->ctx
);
1631 perf_ctx_adjust_freq(ctx
);
1637 cpu_ctx_sched_out(cpuctx
, EVENT_FLEXIBLE
);
1639 task_ctx_sched_out(ctx
, EVENT_FLEXIBLE
);
1641 rotate_ctx(&cpuctx
->ctx
);
1645 cpu_ctx_sched_in(cpuctx
, EVENT_FLEXIBLE
);
1647 task_ctx_sched_in(curr
, EVENT_FLEXIBLE
);
1651 static int event_enable_on_exec(struct perf_event
*event
,
1652 struct perf_event_context
*ctx
)
1654 if (!event
->attr
.enable_on_exec
)
1657 event
->attr
.enable_on_exec
= 0;
1658 if (event
->state
>= PERF_EVENT_STATE_INACTIVE
)
1661 __perf_event_mark_enabled(event
, ctx
);
1667 * Enable all of a task's events that have been marked enable-on-exec.
1668 * This expects task == current.
1670 static void perf_event_enable_on_exec(struct task_struct
*task
)
1672 struct perf_event_context
*ctx
;
1673 struct perf_event
*event
;
1674 unsigned long flags
;
1678 local_irq_save(flags
);
1679 ctx
= task
->perf_event_ctxp
;
1680 if (!ctx
|| !ctx
->nr_events
)
1683 __perf_event_task_sched_out(ctx
);
1685 raw_spin_lock(&ctx
->lock
);
1687 list_for_each_entry(event
, &ctx
->pinned_groups
, group_entry
) {
1688 ret
= event_enable_on_exec(event
, ctx
);
1693 list_for_each_entry(event
, &ctx
->flexible_groups
, group_entry
) {
1694 ret
= event_enable_on_exec(event
, ctx
);
1700 * Unclone this context if we enabled any event.
1705 raw_spin_unlock(&ctx
->lock
);
1707 perf_event_task_sched_in(task
);
1709 local_irq_restore(flags
);
1713 * Cross CPU call to read the hardware event
1715 static void __perf_event_read(void *info
)
1717 struct perf_cpu_context
*cpuctx
= &__get_cpu_var(perf_cpu_context
);
1718 struct perf_event
*event
= info
;
1719 struct perf_event_context
*ctx
= event
->ctx
;
1722 * If this is a task context, we need to check whether it is
1723 * the current task context of this cpu. If not it has been
1724 * scheduled out before the smp call arrived. In that case
1725 * event->count would have been updated to a recent sample
1726 * when the event was scheduled out.
1728 if (ctx
->task
&& cpuctx
->task_ctx
!= ctx
)
1731 raw_spin_lock(&ctx
->lock
);
1732 update_context_time(ctx
);
1733 update_event_times(event
);
1734 raw_spin_unlock(&ctx
->lock
);
1736 event
->pmu
->read(event
);
1739 static inline u64
perf_event_count(struct perf_event
*event
)
1741 return local64_read(&event
->count
) + atomic64_read(&event
->child_count
);
1744 static u64
perf_event_read(struct perf_event
*event
)
1747 * If event is enabled and currently active on a CPU, update the
1748 * value in the event structure:
1750 if (event
->state
== PERF_EVENT_STATE_ACTIVE
) {
1751 smp_call_function_single(event
->oncpu
,
1752 __perf_event_read
, event
, 1);
1753 } else if (event
->state
== PERF_EVENT_STATE_INACTIVE
) {
1754 struct perf_event_context
*ctx
= event
->ctx
;
1755 unsigned long flags
;
1757 raw_spin_lock_irqsave(&ctx
->lock
, flags
);
1758 update_context_time(ctx
);
1759 update_event_times(event
);
1760 raw_spin_unlock_irqrestore(&ctx
->lock
, flags
);
1763 return perf_event_count(event
);
1770 struct callchain_cpus_entries
{
1771 struct rcu_head rcu_head
;
1772 struct perf_callchain_entry
*cpu_entries
[0];
1775 static DEFINE_PER_CPU(int, callchain_recursion
[PERF_NR_CONTEXTS
]);
1776 static atomic_t nr_callchain_events
;
1777 static DEFINE_MUTEX(callchain_mutex
);
1778 struct callchain_cpus_entries
*callchain_cpus_entries
;
1781 __weak
void perf_callchain_kernel(struct perf_callchain_entry
*entry
,
1782 struct pt_regs
*regs
)
1786 __weak
void perf_callchain_user(struct perf_callchain_entry
*entry
,
1787 struct pt_regs
*regs
)
1791 static void release_callchain_buffers_rcu(struct rcu_head
*head
)
1793 struct callchain_cpus_entries
*entries
;
1796 entries
= container_of(head
, struct callchain_cpus_entries
, rcu_head
);
1798 for_each_possible_cpu(cpu
)
1799 kfree(entries
->cpu_entries
[cpu
]);
1804 static void release_callchain_buffers(void)
1806 struct callchain_cpus_entries
*entries
;
1808 entries
= callchain_cpus_entries
;
1809 rcu_assign_pointer(callchain_cpus_entries
, NULL
);
1810 call_rcu(&entries
->rcu_head
, release_callchain_buffers_rcu
);
1813 static int alloc_callchain_buffers(void)
1817 struct callchain_cpus_entries
*entries
;
1820 * We can't use the percpu allocation API for data that can be
1821 * accessed from NMI. Use a temporary manual per cpu allocation
1822 * until that gets sorted out.
1824 size
= sizeof(*entries
) + sizeof(struct perf_callchain_entry
*) *
1825 num_possible_cpus();
1827 entries
= kzalloc(size
, GFP_KERNEL
);
1831 size
= sizeof(struct perf_callchain_entry
) * PERF_NR_CONTEXTS
;
1833 for_each_possible_cpu(cpu
) {
1834 entries
->cpu_entries
[cpu
] = kmalloc_node(size
, GFP_KERNEL
,
1836 if (!entries
->cpu_entries
[cpu
])
1840 rcu_assign_pointer(callchain_cpus_entries
, entries
);
1845 for_each_possible_cpu(cpu
)
1846 kfree(entries
->cpu_entries
[cpu
]);
1852 static int get_callchain_buffers(void)
1857 mutex_lock(&callchain_mutex
);
1859 count
= atomic_inc_return(&nr_callchain_events
);
1860 if (WARN_ON_ONCE(count
< 1)) {
1866 /* If the allocation failed, give up */
1867 if (!callchain_cpus_entries
)
1872 err
= alloc_callchain_buffers();
1874 release_callchain_buffers();
1876 mutex_unlock(&callchain_mutex
);
1881 static void put_callchain_buffers(void)
1883 if (atomic_dec_and_mutex_lock(&nr_callchain_events
, &callchain_mutex
)) {
1884 release_callchain_buffers();
1885 mutex_unlock(&callchain_mutex
);
1889 static int get_recursion_context(int *recursion
)
1897 else if (in_softirq())
1902 if (recursion
[rctx
])
1911 static inline void put_recursion_context(int *recursion
, int rctx
)
1917 static struct perf_callchain_entry
*get_callchain_entry(int *rctx
)
1920 struct callchain_cpus_entries
*entries
;
1922 *rctx
= get_recursion_context(__get_cpu_var(callchain_recursion
));
1926 entries
= rcu_dereference(callchain_cpus_entries
);
1930 cpu
= smp_processor_id();
1932 return &entries
->cpu_entries
[cpu
][*rctx
];
1936 put_callchain_entry(int rctx
)
1938 put_recursion_context(__get_cpu_var(callchain_recursion
), rctx
);
1941 static struct perf_callchain_entry
*perf_callchain(struct pt_regs
*regs
)
1944 struct perf_callchain_entry
*entry
;
1947 entry
= get_callchain_entry(&rctx
);
1956 if (!user_mode(regs
)) {
1957 perf_callchain_store(entry
, PERF_CONTEXT_KERNEL
);
1958 perf_callchain_kernel(entry
, regs
);
1960 regs
= task_pt_regs(current
);
1966 perf_callchain_store(entry
, PERF_CONTEXT_USER
);
1967 perf_callchain_user(entry
, regs
);
1971 put_callchain_entry(rctx
);
1977 * Initialize the perf_event context in a task_struct:
1980 __perf_event_init_context(struct perf_event_context
*ctx
,
1981 struct task_struct
*task
)
1983 raw_spin_lock_init(&ctx
->lock
);
1984 mutex_init(&ctx
->mutex
);
1985 INIT_LIST_HEAD(&ctx
->pinned_groups
);
1986 INIT_LIST_HEAD(&ctx
->flexible_groups
);
1987 INIT_LIST_HEAD(&ctx
->event_list
);
1988 atomic_set(&ctx
->refcount
, 1);
1992 static struct perf_event_context
*find_get_context(pid_t pid
, int cpu
)
1994 struct perf_event_context
*ctx
;
1995 struct perf_cpu_context
*cpuctx
;
1996 struct task_struct
*task
;
1997 unsigned long flags
;
2000 if (pid
== -1 && cpu
!= -1) {
2001 /* Must be root to operate on a CPU event: */
2002 if (perf_paranoid_cpu() && !capable(CAP_SYS_ADMIN
))
2003 return ERR_PTR(-EACCES
);
2005 if (cpu
< 0 || cpu
>= nr_cpumask_bits
)
2006 return ERR_PTR(-EINVAL
);
2009 * We could be clever and allow to attach a event to an
2010 * offline CPU and activate it when the CPU comes up, but
2013 if (!cpu_online(cpu
))
2014 return ERR_PTR(-ENODEV
);
2016 cpuctx
= &per_cpu(perf_cpu_context
, cpu
);
2027 task
= find_task_by_vpid(pid
);
2029 get_task_struct(task
);
2033 return ERR_PTR(-ESRCH
);
2036 * Can't attach events to a dying task.
2039 if (task
->flags
& PF_EXITING
)
2042 /* Reuse ptrace permission checks for now. */
2044 if (!ptrace_may_access(task
, PTRACE_MODE_READ
))
2048 ctx
= perf_lock_task_context(task
, &flags
);
2051 raw_spin_unlock_irqrestore(&ctx
->lock
, flags
);
2055 ctx
= kzalloc(sizeof(struct perf_event_context
), GFP_KERNEL
);
2059 __perf_event_init_context(ctx
, task
);
2061 if (cmpxchg(&task
->perf_event_ctxp
, NULL
, ctx
)) {
2063 * We raced with some other task; use
2064 * the context they set.
2069 get_task_struct(task
);
2072 put_task_struct(task
);
2076 put_task_struct(task
);
2077 return ERR_PTR(err
);
2080 static void perf_event_free_filter(struct perf_event
*event
);
2082 static void free_event_rcu(struct rcu_head
*head
)
2084 struct perf_event
*event
;
2086 event
= container_of(head
, struct perf_event
, rcu_head
);
2088 put_pid_ns(event
->ns
);
2089 perf_event_free_filter(event
);
2093 static void perf_pending_sync(struct perf_event
*event
);
2094 static void perf_buffer_put(struct perf_buffer
*buffer
);
2096 static void free_event(struct perf_event
*event
)
2098 perf_pending_sync(event
);
2100 if (!event
->parent
) {
2101 atomic_dec(&nr_events
);
2102 if (event
->attr
.mmap
|| event
->attr
.mmap_data
)
2103 atomic_dec(&nr_mmap_events
);
2104 if (event
->attr
.comm
)
2105 atomic_dec(&nr_comm_events
);
2106 if (event
->attr
.task
)
2107 atomic_dec(&nr_task_events
);
2108 if (event
->attr
.sample_type
& PERF_SAMPLE_CALLCHAIN
)
2109 put_callchain_buffers();
2112 if (event
->buffer
) {
2113 perf_buffer_put(event
->buffer
);
2114 event
->buffer
= NULL
;
2118 event
->destroy(event
);
2120 put_ctx(event
->ctx
);
2121 call_rcu(&event
->rcu_head
, free_event_rcu
);
2124 int perf_event_release_kernel(struct perf_event
*event
)
2126 struct perf_event_context
*ctx
= event
->ctx
;
2129 * Remove from the PMU, can't get re-enabled since we got
2130 * here because the last ref went.
2132 perf_event_disable(event
);
2134 WARN_ON_ONCE(ctx
->parent_ctx
);
2136 * There are two ways this annotation is useful:
2138 * 1) there is a lock recursion from perf_event_exit_task
2139 * see the comment there.
2141 * 2) there is a lock-inversion with mmap_sem through
2142 * perf_event_read_group(), which takes faults while
2143 * holding ctx->mutex, however this is called after
2144 * the last filedesc died, so there is no possibility
2145 * to trigger the AB-BA case.
2147 mutex_lock_nested(&ctx
->mutex
, SINGLE_DEPTH_NESTING
);
2148 raw_spin_lock_irq(&ctx
->lock
);
2149 perf_group_detach(event
);
2150 list_del_event(event
, ctx
);
2151 raw_spin_unlock_irq(&ctx
->lock
);
2152 mutex_unlock(&ctx
->mutex
);
2154 mutex_lock(&event
->owner
->perf_event_mutex
);
2155 list_del_init(&event
->owner_entry
);
2156 mutex_unlock(&event
->owner
->perf_event_mutex
);
2157 put_task_struct(event
->owner
);
2163 EXPORT_SYMBOL_GPL(perf_event_release_kernel
);
2166 * Called when the last reference to the file is gone.
2168 static int perf_release(struct inode
*inode
, struct file
*file
)
2170 struct perf_event
*event
= file
->private_data
;
2172 file
->private_data
= NULL
;
2174 return perf_event_release_kernel(event
);
2177 static int perf_event_read_size(struct perf_event
*event
)
2179 int entry
= sizeof(u64
); /* value */
2183 if (event
->attr
.read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
)
2184 size
+= sizeof(u64
);
2186 if (event
->attr
.read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
)
2187 size
+= sizeof(u64
);
2189 if (event
->attr
.read_format
& PERF_FORMAT_ID
)
2190 entry
+= sizeof(u64
);
2192 if (event
->attr
.read_format
& PERF_FORMAT_GROUP
) {
2193 nr
+= event
->group_leader
->nr_siblings
;
2194 size
+= sizeof(u64
);
2202 u64
perf_event_read_value(struct perf_event
*event
, u64
*enabled
, u64
*running
)
2204 struct perf_event
*child
;
2210 mutex_lock(&event
->child_mutex
);
2211 total
+= perf_event_read(event
);
2212 *enabled
+= event
->total_time_enabled
+
2213 atomic64_read(&event
->child_total_time_enabled
);
2214 *running
+= event
->total_time_running
+
2215 atomic64_read(&event
->child_total_time_running
);
2217 list_for_each_entry(child
, &event
->child_list
, child_list
) {
2218 total
+= perf_event_read(child
);
2219 *enabled
+= child
->total_time_enabled
;
2220 *running
+= child
->total_time_running
;
2222 mutex_unlock(&event
->child_mutex
);
2226 EXPORT_SYMBOL_GPL(perf_event_read_value
);
2228 static int perf_event_read_group(struct perf_event
*event
,
2229 u64 read_format
, char __user
*buf
)
2231 struct perf_event
*leader
= event
->group_leader
, *sub
;
2232 int n
= 0, size
= 0, ret
= -EFAULT
;
2233 struct perf_event_context
*ctx
= leader
->ctx
;
2235 u64 count
, enabled
, running
;
2237 mutex_lock(&ctx
->mutex
);
2238 count
= perf_event_read_value(leader
, &enabled
, &running
);
2240 values
[n
++] = 1 + leader
->nr_siblings
;
2241 if (read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
)
2242 values
[n
++] = enabled
;
2243 if (read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
)
2244 values
[n
++] = running
;
2245 values
[n
++] = count
;
2246 if (read_format
& PERF_FORMAT_ID
)
2247 values
[n
++] = primary_event_id(leader
);
2249 size
= n
* sizeof(u64
);
2251 if (copy_to_user(buf
, values
, size
))
2256 list_for_each_entry(sub
, &leader
->sibling_list
, group_entry
) {
2259 values
[n
++] = perf_event_read_value(sub
, &enabled
, &running
);
2260 if (read_format
& PERF_FORMAT_ID
)
2261 values
[n
++] = primary_event_id(sub
);
2263 size
= n
* sizeof(u64
);
2265 if (copy_to_user(buf
+ ret
, values
, size
)) {
2273 mutex_unlock(&ctx
->mutex
);
2278 static int perf_event_read_one(struct perf_event
*event
,
2279 u64 read_format
, char __user
*buf
)
2281 u64 enabled
, running
;
2285 values
[n
++] = perf_event_read_value(event
, &enabled
, &running
);
2286 if (read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
)
2287 values
[n
++] = enabled
;
2288 if (read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
)
2289 values
[n
++] = running
;
2290 if (read_format
& PERF_FORMAT_ID
)
2291 values
[n
++] = primary_event_id(event
);
2293 if (copy_to_user(buf
, values
, n
* sizeof(u64
)))
2296 return n
* sizeof(u64
);
2300 * Read the performance event - simple non blocking version for now
2303 perf_read_hw(struct perf_event
*event
, char __user
*buf
, size_t count
)
2305 u64 read_format
= event
->attr
.read_format
;
2309 * Return end-of-file for a read on a event that is in
2310 * error state (i.e. because it was pinned but it couldn't be
2311 * scheduled on to the CPU at some point).
2313 if (event
->state
== PERF_EVENT_STATE_ERROR
)
2316 if (count
< perf_event_read_size(event
))
2319 WARN_ON_ONCE(event
->ctx
->parent_ctx
);
2320 if (read_format
& PERF_FORMAT_GROUP
)
2321 ret
= perf_event_read_group(event
, read_format
, buf
);
2323 ret
= perf_event_read_one(event
, read_format
, buf
);
2329 perf_read(struct file
*file
, char __user
*buf
, size_t count
, loff_t
*ppos
)
2331 struct perf_event
*event
= file
->private_data
;
2333 return perf_read_hw(event
, buf
, count
);
2336 static unsigned int perf_poll(struct file
*file
, poll_table
*wait
)
2338 struct perf_event
*event
= file
->private_data
;
2339 struct perf_buffer
*buffer
;
2340 unsigned int events
= POLL_HUP
;
2343 buffer
= rcu_dereference(event
->buffer
);
2345 events
= atomic_xchg(&buffer
->poll
, 0);
2348 poll_wait(file
, &event
->waitq
, wait
);
2353 static void perf_event_reset(struct perf_event
*event
)
2355 (void)perf_event_read(event
);
2356 local64_set(&event
->count
, 0);
2357 perf_event_update_userpage(event
);
2361 * Holding the top-level event's child_mutex means that any
2362 * descendant process that has inherited this event will block
2363 * in sync_child_event if it goes to exit, thus satisfying the
2364 * task existence requirements of perf_event_enable/disable.
2366 static void perf_event_for_each_child(struct perf_event
*event
,
2367 void (*func
)(struct perf_event
*))
2369 struct perf_event
*child
;
2371 WARN_ON_ONCE(event
->ctx
->parent_ctx
);
2372 mutex_lock(&event
->child_mutex
);
2374 list_for_each_entry(child
, &event
->child_list
, child_list
)
2376 mutex_unlock(&event
->child_mutex
);
2379 static void perf_event_for_each(struct perf_event
*event
,
2380 void (*func
)(struct perf_event
*))
2382 struct perf_event_context
*ctx
= event
->ctx
;
2383 struct perf_event
*sibling
;
2385 WARN_ON_ONCE(ctx
->parent_ctx
);
2386 mutex_lock(&ctx
->mutex
);
2387 event
= event
->group_leader
;
2389 perf_event_for_each_child(event
, func
);
2391 list_for_each_entry(sibling
, &event
->sibling_list
, group_entry
)
2392 perf_event_for_each_child(event
, func
);
2393 mutex_unlock(&ctx
->mutex
);
2396 static int perf_event_period(struct perf_event
*event
, u64 __user
*arg
)
2398 struct perf_event_context
*ctx
= event
->ctx
;
2403 if (!event
->attr
.sample_period
)
2406 size
= copy_from_user(&value
, arg
, sizeof(value
));
2407 if (size
!= sizeof(value
))
2413 raw_spin_lock_irq(&ctx
->lock
);
2414 if (event
->attr
.freq
) {
2415 if (value
> sysctl_perf_event_sample_rate
) {
2420 event
->attr
.sample_freq
= value
;
2422 event
->attr
.sample_period
= value
;
2423 event
->hw
.sample_period
= value
;
2426 raw_spin_unlock_irq(&ctx
->lock
);
2431 static const struct file_operations perf_fops
;
2433 static struct perf_event
*perf_fget_light(int fd
, int *fput_needed
)
2437 file
= fget_light(fd
, fput_needed
);
2439 return ERR_PTR(-EBADF
);
2441 if (file
->f_op
!= &perf_fops
) {
2442 fput_light(file
, *fput_needed
);
2444 return ERR_PTR(-EBADF
);
2447 return file
->private_data
;
2450 static int perf_event_set_output(struct perf_event
*event
,
2451 struct perf_event
*output_event
);
2452 static int perf_event_set_filter(struct perf_event
*event
, void __user
*arg
);
2454 static long perf_ioctl(struct file
*file
, unsigned int cmd
, unsigned long arg
)
2456 struct perf_event
*event
= file
->private_data
;
2457 void (*func
)(struct perf_event
*);
2461 case PERF_EVENT_IOC_ENABLE
:
2462 func
= perf_event_enable
;
2464 case PERF_EVENT_IOC_DISABLE
:
2465 func
= perf_event_disable
;
2467 case PERF_EVENT_IOC_RESET
:
2468 func
= perf_event_reset
;
2471 case PERF_EVENT_IOC_REFRESH
:
2472 return perf_event_refresh(event
, arg
);
2474 case PERF_EVENT_IOC_PERIOD
:
2475 return perf_event_period(event
, (u64 __user
*)arg
);
2477 case PERF_EVENT_IOC_SET_OUTPUT
:
2479 struct perf_event
*output_event
= NULL
;
2480 int fput_needed
= 0;
2484 output_event
= perf_fget_light(arg
, &fput_needed
);
2485 if (IS_ERR(output_event
))
2486 return PTR_ERR(output_event
);
2489 ret
= perf_event_set_output(event
, output_event
);
2491 fput_light(output_event
->filp
, fput_needed
);
2496 case PERF_EVENT_IOC_SET_FILTER
:
2497 return perf_event_set_filter(event
, (void __user
*)arg
);
2503 if (flags
& PERF_IOC_FLAG_GROUP
)
2504 perf_event_for_each(event
, func
);
2506 perf_event_for_each_child(event
, func
);
2511 int perf_event_task_enable(void)
2513 struct perf_event
*event
;
2515 mutex_lock(¤t
->perf_event_mutex
);
2516 list_for_each_entry(event
, ¤t
->perf_event_list
, owner_entry
)
2517 perf_event_for_each_child(event
, perf_event_enable
);
2518 mutex_unlock(¤t
->perf_event_mutex
);
2523 int perf_event_task_disable(void)
2525 struct perf_event
*event
;
2527 mutex_lock(¤t
->perf_event_mutex
);
2528 list_for_each_entry(event
, ¤t
->perf_event_list
, owner_entry
)
2529 perf_event_for_each_child(event
, perf_event_disable
);
2530 mutex_unlock(¤t
->perf_event_mutex
);
2535 #ifndef PERF_EVENT_INDEX_OFFSET
2536 # define PERF_EVENT_INDEX_OFFSET 0
2539 static int perf_event_index(struct perf_event
*event
)
2541 if (event
->state
!= PERF_EVENT_STATE_ACTIVE
)
2544 return event
->hw
.idx
+ 1 - PERF_EVENT_INDEX_OFFSET
;
2548 * Callers need to ensure there can be no nesting of this function, otherwise
2549 * the seqlock logic goes bad. We can not serialize this because the arch
2550 * code calls this from NMI context.
2552 void perf_event_update_userpage(struct perf_event
*event
)
2554 struct perf_event_mmap_page
*userpg
;
2555 struct perf_buffer
*buffer
;
2558 buffer
= rcu_dereference(event
->buffer
);
2562 userpg
= buffer
->user_page
;
2565 * Disable preemption so as to not let the corresponding user-space
2566 * spin too long if we get preempted.
2571 userpg
->index
= perf_event_index(event
);
2572 userpg
->offset
= perf_event_count(event
);
2573 if (event
->state
== PERF_EVENT_STATE_ACTIVE
)
2574 userpg
->offset
-= local64_read(&event
->hw
.prev_count
);
2576 userpg
->time_enabled
= event
->total_time_enabled
+
2577 atomic64_read(&event
->child_total_time_enabled
);
2579 userpg
->time_running
= event
->total_time_running
+
2580 atomic64_read(&event
->child_total_time_running
);
2589 static unsigned long perf_data_size(struct perf_buffer
*buffer
);
2592 perf_buffer_init(struct perf_buffer
*buffer
, long watermark
, int flags
)
2594 long max_size
= perf_data_size(buffer
);
2597 buffer
->watermark
= min(max_size
, watermark
);
2599 if (!buffer
->watermark
)
2600 buffer
->watermark
= max_size
/ 2;
2602 if (flags
& PERF_BUFFER_WRITABLE
)
2603 buffer
->writable
= 1;
2605 atomic_set(&buffer
->refcount
, 1);
2608 #ifndef CONFIG_PERF_USE_VMALLOC
2611 * Back perf_mmap() with regular GFP_KERNEL-0 pages.
2614 static struct page
*
2615 perf_mmap_to_page(struct perf_buffer
*buffer
, unsigned long pgoff
)
2617 if (pgoff
> buffer
->nr_pages
)
2621 return virt_to_page(buffer
->user_page
);
2623 return virt_to_page(buffer
->data_pages
[pgoff
- 1]);
2626 static void *perf_mmap_alloc_page(int cpu
)
2631 node
= (cpu
== -1) ? cpu
: cpu_to_node(cpu
);
2632 page
= alloc_pages_node(node
, GFP_KERNEL
| __GFP_ZERO
, 0);
2636 return page_address(page
);
2639 static struct perf_buffer
*
2640 perf_buffer_alloc(int nr_pages
, long watermark
, int cpu
, int flags
)
2642 struct perf_buffer
*buffer
;
2646 size
= sizeof(struct perf_buffer
);
2647 size
+= nr_pages
* sizeof(void *);
2649 buffer
= kzalloc(size
, GFP_KERNEL
);
2653 buffer
->user_page
= perf_mmap_alloc_page(cpu
);
2654 if (!buffer
->user_page
)
2655 goto fail_user_page
;
2657 for (i
= 0; i
< nr_pages
; i
++) {
2658 buffer
->data_pages
[i
] = perf_mmap_alloc_page(cpu
);
2659 if (!buffer
->data_pages
[i
])
2660 goto fail_data_pages
;
2663 buffer
->nr_pages
= nr_pages
;
2665 perf_buffer_init(buffer
, watermark
, flags
);
2670 for (i
--; i
>= 0; i
--)
2671 free_page((unsigned long)buffer
->data_pages
[i
]);
2673 free_page((unsigned long)buffer
->user_page
);
2682 static void perf_mmap_free_page(unsigned long addr
)
2684 struct page
*page
= virt_to_page((void *)addr
);
2686 page
->mapping
= NULL
;
2690 static void perf_buffer_free(struct perf_buffer
*buffer
)
2694 perf_mmap_free_page((unsigned long)buffer
->user_page
);
2695 for (i
= 0; i
< buffer
->nr_pages
; i
++)
2696 perf_mmap_free_page((unsigned long)buffer
->data_pages
[i
]);
2700 static inline int page_order(struct perf_buffer
*buffer
)
2708 * Back perf_mmap() with vmalloc memory.
2710 * Required for architectures that have d-cache aliasing issues.
2713 static inline int page_order(struct perf_buffer
*buffer
)
2715 return buffer
->page_order
;
2718 static struct page
*
2719 perf_mmap_to_page(struct perf_buffer
*buffer
, unsigned long pgoff
)
2721 if (pgoff
> (1UL << page_order(buffer
)))
2724 return vmalloc_to_page((void *)buffer
->user_page
+ pgoff
* PAGE_SIZE
);
2727 static void perf_mmap_unmark_page(void *addr
)
2729 struct page
*page
= vmalloc_to_page(addr
);
2731 page
->mapping
= NULL
;
2734 static void perf_buffer_free_work(struct work_struct
*work
)
2736 struct perf_buffer
*buffer
;
2740 buffer
= container_of(work
, struct perf_buffer
, work
);
2741 nr
= 1 << page_order(buffer
);
2743 base
= buffer
->user_page
;
2744 for (i
= 0; i
< nr
+ 1; i
++)
2745 perf_mmap_unmark_page(base
+ (i
* PAGE_SIZE
));
2751 static void perf_buffer_free(struct perf_buffer
*buffer
)
2753 schedule_work(&buffer
->work
);
2756 static struct perf_buffer
*
2757 perf_buffer_alloc(int nr_pages
, long watermark
, int cpu
, int flags
)
2759 struct perf_buffer
*buffer
;
2763 size
= sizeof(struct perf_buffer
);
2764 size
+= sizeof(void *);
2766 buffer
= kzalloc(size
, GFP_KERNEL
);
2770 INIT_WORK(&buffer
->work
, perf_buffer_free_work
);
2772 all_buf
= vmalloc_user((nr_pages
+ 1) * PAGE_SIZE
);
2776 buffer
->user_page
= all_buf
;
2777 buffer
->data_pages
[0] = all_buf
+ PAGE_SIZE
;
2778 buffer
->page_order
= ilog2(nr_pages
);
2779 buffer
->nr_pages
= 1;
2781 perf_buffer_init(buffer
, watermark
, flags
);
2794 static unsigned long perf_data_size(struct perf_buffer
*buffer
)
2796 return buffer
->nr_pages
<< (PAGE_SHIFT
+ page_order(buffer
));
2799 static int perf_mmap_fault(struct vm_area_struct
*vma
, struct vm_fault
*vmf
)
2801 struct perf_event
*event
= vma
->vm_file
->private_data
;
2802 struct perf_buffer
*buffer
;
2803 int ret
= VM_FAULT_SIGBUS
;
2805 if (vmf
->flags
& FAULT_FLAG_MKWRITE
) {
2806 if (vmf
->pgoff
== 0)
2812 buffer
= rcu_dereference(event
->buffer
);
2816 if (vmf
->pgoff
&& (vmf
->flags
& FAULT_FLAG_WRITE
))
2819 vmf
->page
= perf_mmap_to_page(buffer
, vmf
->pgoff
);
2823 get_page(vmf
->page
);
2824 vmf
->page
->mapping
= vma
->vm_file
->f_mapping
;
2825 vmf
->page
->index
= vmf
->pgoff
;
2834 static void perf_buffer_free_rcu(struct rcu_head
*rcu_head
)
2836 struct perf_buffer
*buffer
;
2838 buffer
= container_of(rcu_head
, struct perf_buffer
, rcu_head
);
2839 perf_buffer_free(buffer
);
2842 static struct perf_buffer
*perf_buffer_get(struct perf_event
*event
)
2844 struct perf_buffer
*buffer
;
2847 buffer
= rcu_dereference(event
->buffer
);
2849 if (!atomic_inc_not_zero(&buffer
->refcount
))
2857 static void perf_buffer_put(struct perf_buffer
*buffer
)
2859 if (!atomic_dec_and_test(&buffer
->refcount
))
2862 call_rcu(&buffer
->rcu_head
, perf_buffer_free_rcu
);
2865 static void perf_mmap_open(struct vm_area_struct
*vma
)
2867 struct perf_event
*event
= vma
->vm_file
->private_data
;
2869 atomic_inc(&event
->mmap_count
);
2872 static void perf_mmap_close(struct vm_area_struct
*vma
)
2874 struct perf_event
*event
= vma
->vm_file
->private_data
;
2876 if (atomic_dec_and_mutex_lock(&event
->mmap_count
, &event
->mmap_mutex
)) {
2877 unsigned long size
= perf_data_size(event
->buffer
);
2878 struct user_struct
*user
= event
->mmap_user
;
2879 struct perf_buffer
*buffer
= event
->buffer
;
2881 atomic_long_sub((size
>> PAGE_SHIFT
) + 1, &user
->locked_vm
);
2882 vma
->vm_mm
->locked_vm
-= event
->mmap_locked
;
2883 rcu_assign_pointer(event
->buffer
, NULL
);
2884 mutex_unlock(&event
->mmap_mutex
);
2886 perf_buffer_put(buffer
);
2891 static const struct vm_operations_struct perf_mmap_vmops
= {
2892 .open
= perf_mmap_open
,
2893 .close
= perf_mmap_close
,
2894 .fault
= perf_mmap_fault
,
2895 .page_mkwrite
= perf_mmap_fault
,
2898 static int perf_mmap(struct file
*file
, struct vm_area_struct
*vma
)
2900 struct perf_event
*event
= file
->private_data
;
2901 unsigned long user_locked
, user_lock_limit
;
2902 struct user_struct
*user
= current_user();
2903 unsigned long locked
, lock_limit
;
2904 struct perf_buffer
*buffer
;
2905 unsigned long vma_size
;
2906 unsigned long nr_pages
;
2907 long user_extra
, extra
;
2908 int ret
= 0, flags
= 0;
2911 * Don't allow mmap() of inherited per-task counters. This would
2912 * create a performance issue due to all children writing to the
2915 if (event
->cpu
== -1 && event
->attr
.inherit
)
2918 if (!(vma
->vm_flags
& VM_SHARED
))
2921 vma_size
= vma
->vm_end
- vma
->vm_start
;
2922 nr_pages
= (vma_size
/ PAGE_SIZE
) - 1;
2925 * If we have buffer pages ensure they're a power-of-two number, so we
2926 * can do bitmasks instead of modulo.
2928 if (nr_pages
!= 0 && !is_power_of_2(nr_pages
))
2931 if (vma_size
!= PAGE_SIZE
* (1 + nr_pages
))
2934 if (vma
->vm_pgoff
!= 0)
2937 WARN_ON_ONCE(event
->ctx
->parent_ctx
);
2938 mutex_lock(&event
->mmap_mutex
);
2939 if (event
->buffer
) {
2940 if (event
->buffer
->nr_pages
== nr_pages
)
2941 atomic_inc(&event
->buffer
->refcount
);
2947 user_extra
= nr_pages
+ 1;
2948 user_lock_limit
= sysctl_perf_event_mlock
>> (PAGE_SHIFT
- 10);
2951 * Increase the limit linearly with more CPUs:
2953 user_lock_limit
*= num_online_cpus();
2955 user_locked
= atomic_long_read(&user
->locked_vm
) + user_extra
;
2958 if (user_locked
> user_lock_limit
)
2959 extra
= user_locked
- user_lock_limit
;
2961 lock_limit
= rlimit(RLIMIT_MEMLOCK
);
2962 lock_limit
>>= PAGE_SHIFT
;
2963 locked
= vma
->vm_mm
->locked_vm
+ extra
;
2965 if ((locked
> lock_limit
) && perf_paranoid_tracepoint_raw() &&
2966 !capable(CAP_IPC_LOCK
)) {
2971 WARN_ON(event
->buffer
);
2973 if (vma
->vm_flags
& VM_WRITE
)
2974 flags
|= PERF_BUFFER_WRITABLE
;
2976 buffer
= perf_buffer_alloc(nr_pages
, event
->attr
.wakeup_watermark
,
2982 rcu_assign_pointer(event
->buffer
, buffer
);
2984 atomic_long_add(user_extra
, &user
->locked_vm
);
2985 event
->mmap_locked
= extra
;
2986 event
->mmap_user
= get_current_user();
2987 vma
->vm_mm
->locked_vm
+= event
->mmap_locked
;
2991 atomic_inc(&event
->mmap_count
);
2992 mutex_unlock(&event
->mmap_mutex
);
2994 vma
->vm_flags
|= VM_RESERVED
;
2995 vma
->vm_ops
= &perf_mmap_vmops
;
3000 static int perf_fasync(int fd
, struct file
*filp
, int on
)
3002 struct inode
*inode
= filp
->f_path
.dentry
->d_inode
;
3003 struct perf_event
*event
= filp
->private_data
;
3006 mutex_lock(&inode
->i_mutex
);
3007 retval
= fasync_helper(fd
, filp
, on
, &event
->fasync
);
3008 mutex_unlock(&inode
->i_mutex
);
3016 static const struct file_operations perf_fops
= {
3017 .llseek
= no_llseek
,
3018 .release
= perf_release
,
3021 .unlocked_ioctl
= perf_ioctl
,
3022 .compat_ioctl
= perf_ioctl
,
3024 .fasync
= perf_fasync
,
3030 * If there's data, ensure we set the poll() state and publish everything
3031 * to user-space before waking everybody up.
3034 void perf_event_wakeup(struct perf_event
*event
)
3036 wake_up_all(&event
->waitq
);
3038 if (event
->pending_kill
) {
3039 kill_fasync(&event
->fasync
, SIGIO
, event
->pending_kill
);
3040 event
->pending_kill
= 0;
3047 * Handle the case where we need to wakeup up from NMI (or rq->lock) context.
3049 * The NMI bit means we cannot possibly take locks. Therefore, maintain a
3050 * single linked list and use cmpxchg() to add entries lockless.
3053 static void perf_pending_event(struct perf_pending_entry
*entry
)
3055 struct perf_event
*event
= container_of(entry
,
3056 struct perf_event
, pending
);
3058 if (event
->pending_disable
) {
3059 event
->pending_disable
= 0;
3060 __perf_event_disable(event
);
3063 if (event
->pending_wakeup
) {
3064 event
->pending_wakeup
= 0;
3065 perf_event_wakeup(event
);
3069 #define PENDING_TAIL ((struct perf_pending_entry *)-1UL)
3071 static DEFINE_PER_CPU(struct perf_pending_entry
*, perf_pending_head
) = {
3075 static void perf_pending_queue(struct perf_pending_entry
*entry
,
3076 void (*func
)(struct perf_pending_entry
*))
3078 struct perf_pending_entry
**head
;
3080 if (cmpxchg(&entry
->next
, NULL
, PENDING_TAIL
) != NULL
)
3085 head
= &get_cpu_var(perf_pending_head
);
3088 entry
->next
= *head
;
3089 } while (cmpxchg(head
, entry
->next
, entry
) != entry
->next
);
3091 set_perf_event_pending();
3093 put_cpu_var(perf_pending_head
);
3096 static int __perf_pending_run(void)
3098 struct perf_pending_entry
*list
;
3101 list
= xchg(&__get_cpu_var(perf_pending_head
), PENDING_TAIL
);
3102 while (list
!= PENDING_TAIL
) {
3103 void (*func
)(struct perf_pending_entry
*);
3104 struct perf_pending_entry
*entry
= list
;
3111 * Ensure we observe the unqueue before we issue the wakeup,
3112 * so that we won't be waiting forever.
3113 * -- see perf_not_pending().
3124 static inline int perf_not_pending(struct perf_event
*event
)
3127 * If we flush on whatever cpu we run, there is a chance we don't
3131 __perf_pending_run();
3135 * Ensure we see the proper queue state before going to sleep
3136 * so that we do not miss the wakeup. -- see perf_pending_handle()
3139 return event
->pending
.next
== NULL
;
3142 static void perf_pending_sync(struct perf_event
*event
)
3144 wait_event(event
->waitq
, perf_not_pending(event
));
3147 void perf_event_do_pending(void)
3149 __perf_pending_run();
3153 * We assume there is only KVM supporting the callbacks.
3154 * Later on, we might change it to a list if there is
3155 * another virtualization implementation supporting the callbacks.
3157 struct perf_guest_info_callbacks
*perf_guest_cbs
;
3159 int perf_register_guest_info_callbacks(struct perf_guest_info_callbacks
*cbs
)
3161 perf_guest_cbs
= cbs
;
3164 EXPORT_SYMBOL_GPL(perf_register_guest_info_callbacks
);
3166 int perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks
*cbs
)
3168 perf_guest_cbs
= NULL
;
3171 EXPORT_SYMBOL_GPL(perf_unregister_guest_info_callbacks
);
3176 static bool perf_output_space(struct perf_buffer
*buffer
, unsigned long tail
,
3177 unsigned long offset
, unsigned long head
)
3181 if (!buffer
->writable
)
3184 mask
= perf_data_size(buffer
) - 1;
3186 offset
= (offset
- tail
) & mask
;
3187 head
= (head
- tail
) & mask
;
3189 if ((int)(head
- offset
) < 0)
3195 static void perf_output_wakeup(struct perf_output_handle
*handle
)
3197 atomic_set(&handle
->buffer
->poll
, POLL_IN
);
3200 handle
->event
->pending_wakeup
= 1;
3201 perf_pending_queue(&handle
->event
->pending
,
3202 perf_pending_event
);
3204 perf_event_wakeup(handle
->event
);
3208 * We need to ensure a later event_id doesn't publish a head when a former
3209 * event isn't done writing. However since we need to deal with NMIs we
3210 * cannot fully serialize things.
3212 * We only publish the head (and generate a wakeup) when the outer-most
3215 static void perf_output_get_handle(struct perf_output_handle
*handle
)
3217 struct perf_buffer
*buffer
= handle
->buffer
;
3220 local_inc(&buffer
->nest
);
3221 handle
->wakeup
= local_read(&buffer
->wakeup
);
3224 static void perf_output_put_handle(struct perf_output_handle
*handle
)
3226 struct perf_buffer
*buffer
= handle
->buffer
;
3230 head
= local_read(&buffer
->head
);
3233 * IRQ/NMI can happen here, which means we can miss a head update.
3236 if (!local_dec_and_test(&buffer
->nest
))
3240 * Publish the known good head. Rely on the full barrier implied
3241 * by atomic_dec_and_test() order the buffer->head read and this
3244 buffer
->user_page
->data_head
= head
;
3247 * Now check if we missed an update, rely on the (compiler)
3248 * barrier in atomic_dec_and_test() to re-read buffer->head.
3250 if (unlikely(head
!= local_read(&buffer
->head
))) {
3251 local_inc(&buffer
->nest
);
3255 if (handle
->wakeup
!= local_read(&buffer
->wakeup
))
3256 perf_output_wakeup(handle
);
3262 __always_inline
void perf_output_copy(struct perf_output_handle
*handle
,
3263 const void *buf
, unsigned int len
)
3266 unsigned long size
= min_t(unsigned long, handle
->size
, len
);
3268 memcpy(handle
->addr
, buf
, size
);
3271 handle
->addr
+= size
;
3273 handle
->size
-= size
;
3274 if (!handle
->size
) {
3275 struct perf_buffer
*buffer
= handle
->buffer
;
3278 handle
->page
&= buffer
->nr_pages
- 1;
3279 handle
->addr
= buffer
->data_pages
[handle
->page
];
3280 handle
->size
= PAGE_SIZE
<< page_order(buffer
);
3285 int perf_output_begin(struct perf_output_handle
*handle
,
3286 struct perf_event
*event
, unsigned int size
,
3287 int nmi
, int sample
)
3289 struct perf_buffer
*buffer
;
3290 unsigned long tail
, offset
, head
;
3293 struct perf_event_header header
;
3300 * For inherited events we send all the output towards the parent.
3303 event
= event
->parent
;
3305 buffer
= rcu_dereference(event
->buffer
);
3309 handle
->buffer
= buffer
;
3310 handle
->event
= event
;
3312 handle
->sample
= sample
;
3314 if (!buffer
->nr_pages
)
3317 have_lost
= local_read(&buffer
->lost
);
3319 size
+= sizeof(lost_event
);
3321 perf_output_get_handle(handle
);
3325 * Userspace could choose to issue a mb() before updating the
3326 * tail pointer. So that all reads will be completed before the
3329 tail
= ACCESS_ONCE(buffer
->user_page
->data_tail
);
3331 offset
= head
= local_read(&buffer
->head
);
3333 if (unlikely(!perf_output_space(buffer
, tail
, offset
, head
)))
3335 } while (local_cmpxchg(&buffer
->head
, offset
, head
) != offset
);
3337 if (head
- local_read(&buffer
->wakeup
) > buffer
->watermark
)
3338 local_add(buffer
->watermark
, &buffer
->wakeup
);
3340 handle
->page
= offset
>> (PAGE_SHIFT
+ page_order(buffer
));
3341 handle
->page
&= buffer
->nr_pages
- 1;
3342 handle
->size
= offset
& ((PAGE_SIZE
<< page_order(buffer
)) - 1);
3343 handle
->addr
= buffer
->data_pages
[handle
->page
];
3344 handle
->addr
+= handle
->size
;
3345 handle
->size
= (PAGE_SIZE
<< page_order(buffer
)) - handle
->size
;
3348 lost_event
.header
.type
= PERF_RECORD_LOST
;
3349 lost_event
.header
.misc
= 0;
3350 lost_event
.header
.size
= sizeof(lost_event
);
3351 lost_event
.id
= event
->id
;
3352 lost_event
.lost
= local_xchg(&buffer
->lost
, 0);
3354 perf_output_put(handle
, lost_event
);
3360 local_inc(&buffer
->lost
);
3361 perf_output_put_handle(handle
);
3368 void perf_output_end(struct perf_output_handle
*handle
)
3370 struct perf_event
*event
= handle
->event
;
3371 struct perf_buffer
*buffer
= handle
->buffer
;
3373 int wakeup_events
= event
->attr
.wakeup_events
;
3375 if (handle
->sample
&& wakeup_events
) {
3376 int events
= local_inc_return(&buffer
->events
);
3377 if (events
>= wakeup_events
) {
3378 local_sub(wakeup_events
, &buffer
->events
);
3379 local_inc(&buffer
->wakeup
);
3383 perf_output_put_handle(handle
);
3387 static u32
perf_event_pid(struct perf_event
*event
, struct task_struct
*p
)
3390 * only top level events have the pid namespace they were created in
3393 event
= event
->parent
;
3395 return task_tgid_nr_ns(p
, event
->ns
);
3398 static u32
perf_event_tid(struct perf_event
*event
, struct task_struct
*p
)
3401 * only top level events have the pid namespace they were created in
3404 event
= event
->parent
;
3406 return task_pid_nr_ns(p
, event
->ns
);
3409 static void perf_output_read_one(struct perf_output_handle
*handle
,
3410 struct perf_event
*event
)
3412 u64 read_format
= event
->attr
.read_format
;
3416 values
[n
++] = perf_event_count(event
);
3417 if (read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
) {
3418 values
[n
++] = event
->total_time_enabled
+
3419 atomic64_read(&event
->child_total_time_enabled
);
3421 if (read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
) {
3422 values
[n
++] = event
->total_time_running
+
3423 atomic64_read(&event
->child_total_time_running
);
3425 if (read_format
& PERF_FORMAT_ID
)
3426 values
[n
++] = primary_event_id(event
);
3428 perf_output_copy(handle
, values
, n
* sizeof(u64
));
3432 * XXX PERF_FORMAT_GROUP vs inherited events seems difficult.
3434 static void perf_output_read_group(struct perf_output_handle
*handle
,
3435 struct perf_event
*event
)
3437 struct perf_event
*leader
= event
->group_leader
, *sub
;
3438 u64 read_format
= event
->attr
.read_format
;
3442 values
[n
++] = 1 + leader
->nr_siblings
;
3444 if (read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
)
3445 values
[n
++] = leader
->total_time_enabled
;
3447 if (read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
)
3448 values
[n
++] = leader
->total_time_running
;
3450 if (leader
!= event
)
3451 leader
->pmu
->read(leader
);
3453 values
[n
++] = perf_event_count(leader
);
3454 if (read_format
& PERF_FORMAT_ID
)
3455 values
[n
++] = primary_event_id(leader
);
3457 perf_output_copy(handle
, values
, n
* sizeof(u64
));
3459 list_for_each_entry(sub
, &leader
->sibling_list
, group_entry
) {
3463 sub
->pmu
->read(sub
);
3465 values
[n
++] = perf_event_count(sub
);
3466 if (read_format
& PERF_FORMAT_ID
)
3467 values
[n
++] = primary_event_id(sub
);
3469 perf_output_copy(handle
, values
, n
* sizeof(u64
));
3473 static void perf_output_read(struct perf_output_handle
*handle
,
3474 struct perf_event
*event
)
3476 if (event
->attr
.read_format
& PERF_FORMAT_GROUP
)
3477 perf_output_read_group(handle
, event
);
3479 perf_output_read_one(handle
, event
);
3482 void perf_output_sample(struct perf_output_handle
*handle
,
3483 struct perf_event_header
*header
,
3484 struct perf_sample_data
*data
,
3485 struct perf_event
*event
)
3487 u64 sample_type
= data
->type
;
3489 perf_output_put(handle
, *header
);
3491 if (sample_type
& PERF_SAMPLE_IP
)
3492 perf_output_put(handle
, data
->ip
);
3494 if (sample_type
& PERF_SAMPLE_TID
)
3495 perf_output_put(handle
, data
->tid_entry
);
3497 if (sample_type
& PERF_SAMPLE_TIME
)
3498 perf_output_put(handle
, data
->time
);
3500 if (sample_type
& PERF_SAMPLE_ADDR
)
3501 perf_output_put(handle
, data
->addr
);
3503 if (sample_type
& PERF_SAMPLE_ID
)
3504 perf_output_put(handle
, data
->id
);
3506 if (sample_type
& PERF_SAMPLE_STREAM_ID
)
3507 perf_output_put(handle
, data
->stream_id
);
3509 if (sample_type
& PERF_SAMPLE_CPU
)
3510 perf_output_put(handle
, data
->cpu_entry
);
3512 if (sample_type
& PERF_SAMPLE_PERIOD
)
3513 perf_output_put(handle
, data
->period
);
3515 if (sample_type
& PERF_SAMPLE_READ
)
3516 perf_output_read(handle
, event
);
3518 if (sample_type
& PERF_SAMPLE_CALLCHAIN
) {
3519 if (data
->callchain
) {
3522 if (data
->callchain
)
3523 size
+= data
->callchain
->nr
;
3525 size
*= sizeof(u64
);
3527 perf_output_copy(handle
, data
->callchain
, size
);
3530 perf_output_put(handle
, nr
);
3534 if (sample_type
& PERF_SAMPLE_RAW
) {
3536 perf_output_put(handle
, data
->raw
->size
);
3537 perf_output_copy(handle
, data
->raw
->data
,
3544 .size
= sizeof(u32
),
3547 perf_output_put(handle
, raw
);
3552 void perf_prepare_sample(struct perf_event_header
*header
,
3553 struct perf_sample_data
*data
,
3554 struct perf_event
*event
,
3555 struct pt_regs
*regs
)
3557 u64 sample_type
= event
->attr
.sample_type
;
3559 data
->type
= sample_type
;
3561 header
->type
= PERF_RECORD_SAMPLE
;
3562 header
->size
= sizeof(*header
);
3565 header
->misc
|= perf_misc_flags(regs
);
3567 if (sample_type
& PERF_SAMPLE_IP
) {
3568 data
->ip
= perf_instruction_pointer(regs
);
3570 header
->size
+= sizeof(data
->ip
);
3573 if (sample_type
& PERF_SAMPLE_TID
) {
3574 /* namespace issues */
3575 data
->tid_entry
.pid
= perf_event_pid(event
, current
);
3576 data
->tid_entry
.tid
= perf_event_tid(event
, current
);
3578 header
->size
+= sizeof(data
->tid_entry
);
3581 if (sample_type
& PERF_SAMPLE_TIME
) {
3582 data
->time
= perf_clock();
3584 header
->size
+= sizeof(data
->time
);
3587 if (sample_type
& PERF_SAMPLE_ADDR
)
3588 header
->size
+= sizeof(data
->addr
);
3590 if (sample_type
& PERF_SAMPLE_ID
) {
3591 data
->id
= primary_event_id(event
);
3593 header
->size
+= sizeof(data
->id
);
3596 if (sample_type
& PERF_SAMPLE_STREAM_ID
) {
3597 data
->stream_id
= event
->id
;
3599 header
->size
+= sizeof(data
->stream_id
);
3602 if (sample_type
& PERF_SAMPLE_CPU
) {
3603 data
->cpu_entry
.cpu
= raw_smp_processor_id();
3604 data
->cpu_entry
.reserved
= 0;
3606 header
->size
+= sizeof(data
->cpu_entry
);
3609 if (sample_type
& PERF_SAMPLE_PERIOD
)
3610 header
->size
+= sizeof(data
->period
);
3612 if (sample_type
& PERF_SAMPLE_READ
)
3613 header
->size
+= perf_event_read_size(event
);
3615 if (sample_type
& PERF_SAMPLE_CALLCHAIN
) {
3618 data
->callchain
= perf_callchain(regs
);
3620 if (data
->callchain
)
3621 size
+= data
->callchain
->nr
;
3623 header
->size
+= size
* sizeof(u64
);
3626 if (sample_type
& PERF_SAMPLE_RAW
) {
3627 int size
= sizeof(u32
);
3630 size
+= data
->raw
->size
;
3632 size
+= sizeof(u32
);
3634 WARN_ON_ONCE(size
& (sizeof(u64
)-1));
3635 header
->size
+= size
;
3639 static void perf_event_output(struct perf_event
*event
, int nmi
,
3640 struct perf_sample_data
*data
,
3641 struct pt_regs
*regs
)
3643 struct perf_output_handle handle
;
3644 struct perf_event_header header
;
3646 /* protect the callchain buffers */
3649 perf_prepare_sample(&header
, data
, event
, regs
);
3651 if (perf_output_begin(&handle
, event
, header
.size
, nmi
, 1))
3654 perf_output_sample(&handle
, &header
, data
, event
);
3656 perf_output_end(&handle
);
3666 struct perf_read_event
{
3667 struct perf_event_header header
;
3674 perf_event_read_event(struct perf_event
*event
,
3675 struct task_struct
*task
)
3677 struct perf_output_handle handle
;
3678 struct perf_read_event read_event
= {
3680 .type
= PERF_RECORD_READ
,
3682 .size
= sizeof(read_event
) + perf_event_read_size(event
),
3684 .pid
= perf_event_pid(event
, task
),
3685 .tid
= perf_event_tid(event
, task
),
3689 ret
= perf_output_begin(&handle
, event
, read_event
.header
.size
, 0, 0);
3693 perf_output_put(&handle
, read_event
);
3694 perf_output_read(&handle
, event
);
3696 perf_output_end(&handle
);
3700 * task tracking -- fork/exit
3702 * enabled by: attr.comm | attr.mmap | attr.mmap_data | attr.task
3705 struct perf_task_event
{
3706 struct task_struct
*task
;
3707 struct perf_event_context
*task_ctx
;
3710 struct perf_event_header header
;
3720 static void perf_event_task_output(struct perf_event
*event
,
3721 struct perf_task_event
*task_event
)
3723 struct perf_output_handle handle
;
3724 struct task_struct
*task
= task_event
->task
;
3727 size
= task_event
->event_id
.header
.size
;
3728 ret
= perf_output_begin(&handle
, event
, size
, 0, 0);
3733 task_event
->event_id
.pid
= perf_event_pid(event
, task
);
3734 task_event
->event_id
.ppid
= perf_event_pid(event
, current
);
3736 task_event
->event_id
.tid
= perf_event_tid(event
, task
);
3737 task_event
->event_id
.ptid
= perf_event_tid(event
, current
);
3739 perf_output_put(&handle
, task_event
->event_id
);
3741 perf_output_end(&handle
);
3744 static int perf_event_task_match(struct perf_event
*event
)
3746 if (event
->state
< PERF_EVENT_STATE_INACTIVE
)
3749 if (event
->cpu
!= -1 && event
->cpu
!= smp_processor_id())
3752 if (event
->attr
.comm
|| event
->attr
.mmap
||
3753 event
->attr
.mmap_data
|| event
->attr
.task
)
3759 static void perf_event_task_ctx(struct perf_event_context
*ctx
,
3760 struct perf_task_event
*task_event
)
3762 struct perf_event
*event
;
3764 list_for_each_entry_rcu(event
, &ctx
->event_list
, event_entry
) {
3765 if (perf_event_task_match(event
))
3766 perf_event_task_output(event
, task_event
);
3770 static void perf_event_task_event(struct perf_task_event
*task_event
)
3772 struct perf_cpu_context
*cpuctx
;
3773 struct perf_event_context
*ctx
= task_event
->task_ctx
;
3776 cpuctx
= &get_cpu_var(perf_cpu_context
);
3777 perf_event_task_ctx(&cpuctx
->ctx
, task_event
);
3779 ctx
= rcu_dereference(current
->perf_event_ctxp
);
3781 perf_event_task_ctx(ctx
, task_event
);
3782 put_cpu_var(perf_cpu_context
);
3786 static void perf_event_task(struct task_struct
*task
,
3787 struct perf_event_context
*task_ctx
,
3790 struct perf_task_event task_event
;
3792 if (!atomic_read(&nr_comm_events
) &&
3793 !atomic_read(&nr_mmap_events
) &&
3794 !atomic_read(&nr_task_events
))
3797 task_event
= (struct perf_task_event
){
3799 .task_ctx
= task_ctx
,
3802 .type
= new ? PERF_RECORD_FORK
: PERF_RECORD_EXIT
,
3804 .size
= sizeof(task_event
.event_id
),
3810 .time
= perf_clock(),
3814 perf_event_task_event(&task_event
);
3817 void perf_event_fork(struct task_struct
*task
)
3819 perf_event_task(task
, NULL
, 1);
3826 struct perf_comm_event
{
3827 struct task_struct
*task
;
3832 struct perf_event_header header
;
3839 static void perf_event_comm_output(struct perf_event
*event
,
3840 struct perf_comm_event
*comm_event
)
3842 struct perf_output_handle handle
;
3843 int size
= comm_event
->event_id
.header
.size
;
3844 int ret
= perf_output_begin(&handle
, event
, size
, 0, 0);
3849 comm_event
->event_id
.pid
= perf_event_pid(event
, comm_event
->task
);
3850 comm_event
->event_id
.tid
= perf_event_tid(event
, comm_event
->task
);
3852 perf_output_put(&handle
, comm_event
->event_id
);
3853 perf_output_copy(&handle
, comm_event
->comm
,
3854 comm_event
->comm_size
);
3855 perf_output_end(&handle
);
3858 static int perf_event_comm_match(struct perf_event
*event
)
3860 if (event
->state
< PERF_EVENT_STATE_INACTIVE
)
3863 if (event
->cpu
!= -1 && event
->cpu
!= smp_processor_id())
3866 if (event
->attr
.comm
)
3872 static void perf_event_comm_ctx(struct perf_event_context
*ctx
,
3873 struct perf_comm_event
*comm_event
)
3875 struct perf_event
*event
;
3877 list_for_each_entry_rcu(event
, &ctx
->event_list
, event_entry
) {
3878 if (perf_event_comm_match(event
))
3879 perf_event_comm_output(event
, comm_event
);
3883 static void perf_event_comm_event(struct perf_comm_event
*comm_event
)
3885 struct perf_cpu_context
*cpuctx
;
3886 struct perf_event_context
*ctx
;
3888 char comm
[TASK_COMM_LEN
];
3890 memset(comm
, 0, sizeof(comm
));
3891 strlcpy(comm
, comm_event
->task
->comm
, sizeof(comm
));
3892 size
= ALIGN(strlen(comm
)+1, sizeof(u64
));
3894 comm_event
->comm
= comm
;
3895 comm_event
->comm_size
= size
;
3897 comm_event
->event_id
.header
.size
= sizeof(comm_event
->event_id
) + size
;
3900 cpuctx
= &get_cpu_var(perf_cpu_context
);
3901 perf_event_comm_ctx(&cpuctx
->ctx
, comm_event
);
3902 ctx
= rcu_dereference(current
->perf_event_ctxp
);
3904 perf_event_comm_ctx(ctx
, comm_event
);
3905 put_cpu_var(perf_cpu_context
);
3909 void perf_event_comm(struct task_struct
*task
)
3911 struct perf_comm_event comm_event
;
3913 if (task
->perf_event_ctxp
)
3914 perf_event_enable_on_exec(task
);
3916 if (!atomic_read(&nr_comm_events
))
3919 comm_event
= (struct perf_comm_event
){
3925 .type
= PERF_RECORD_COMM
,
3934 perf_event_comm_event(&comm_event
);
3941 struct perf_mmap_event
{
3942 struct vm_area_struct
*vma
;
3944 const char *file_name
;
3948 struct perf_event_header header
;
3958 static void perf_event_mmap_output(struct perf_event
*event
,
3959 struct perf_mmap_event
*mmap_event
)
3961 struct perf_output_handle handle
;
3962 int size
= mmap_event
->event_id
.header
.size
;
3963 int ret
= perf_output_begin(&handle
, event
, size
, 0, 0);
3968 mmap_event
->event_id
.pid
= perf_event_pid(event
, current
);
3969 mmap_event
->event_id
.tid
= perf_event_tid(event
, current
);
3971 perf_output_put(&handle
, mmap_event
->event_id
);
3972 perf_output_copy(&handle
, mmap_event
->file_name
,
3973 mmap_event
->file_size
);
3974 perf_output_end(&handle
);
3977 static int perf_event_mmap_match(struct perf_event
*event
,
3978 struct perf_mmap_event
*mmap_event
,
3981 if (event
->state
< PERF_EVENT_STATE_INACTIVE
)
3984 if (event
->cpu
!= -1 && event
->cpu
!= smp_processor_id())
3987 if ((!executable
&& event
->attr
.mmap_data
) ||
3988 (executable
&& event
->attr
.mmap
))
3994 static void perf_event_mmap_ctx(struct perf_event_context
*ctx
,
3995 struct perf_mmap_event
*mmap_event
,
3998 struct perf_event
*event
;
4000 list_for_each_entry_rcu(event
, &ctx
->event_list
, event_entry
) {
4001 if (perf_event_mmap_match(event
, mmap_event
, executable
))
4002 perf_event_mmap_output(event
, mmap_event
);
4006 static void perf_event_mmap_event(struct perf_mmap_event
*mmap_event
)
4008 struct perf_cpu_context
*cpuctx
;
4009 struct perf_event_context
*ctx
;
4010 struct vm_area_struct
*vma
= mmap_event
->vma
;
4011 struct file
*file
= vma
->vm_file
;
4017 memset(tmp
, 0, sizeof(tmp
));
4021 * d_path works from the end of the buffer backwards, so we
4022 * need to add enough zero bytes after the string to handle
4023 * the 64bit alignment we do later.
4025 buf
= kzalloc(PATH_MAX
+ sizeof(u64
), GFP_KERNEL
);
4027 name
= strncpy(tmp
, "//enomem", sizeof(tmp
));
4030 name
= d_path(&file
->f_path
, buf
, PATH_MAX
);
4032 name
= strncpy(tmp
, "//toolong", sizeof(tmp
));
4036 if (arch_vma_name(mmap_event
->vma
)) {
4037 name
= strncpy(tmp
, arch_vma_name(mmap_event
->vma
),
4043 name
= strncpy(tmp
, "[vdso]", sizeof(tmp
));
4045 } else if (vma
->vm_start
<= vma
->vm_mm
->start_brk
&&
4046 vma
->vm_end
>= vma
->vm_mm
->brk
) {
4047 name
= strncpy(tmp
, "[heap]", sizeof(tmp
));
4049 } else if (vma
->vm_start
<= vma
->vm_mm
->start_stack
&&
4050 vma
->vm_end
>= vma
->vm_mm
->start_stack
) {
4051 name
= strncpy(tmp
, "[stack]", sizeof(tmp
));
4055 name
= strncpy(tmp
, "//anon", sizeof(tmp
));
4060 size
= ALIGN(strlen(name
)+1, sizeof(u64
));
4062 mmap_event
->file_name
= name
;
4063 mmap_event
->file_size
= size
;
4065 mmap_event
->event_id
.header
.size
= sizeof(mmap_event
->event_id
) + size
;
4068 cpuctx
= &get_cpu_var(perf_cpu_context
);
4069 perf_event_mmap_ctx(&cpuctx
->ctx
, mmap_event
, vma
->vm_flags
& VM_EXEC
);
4070 ctx
= rcu_dereference(current
->perf_event_ctxp
);
4072 perf_event_mmap_ctx(ctx
, mmap_event
, vma
->vm_flags
& VM_EXEC
);
4073 put_cpu_var(perf_cpu_context
);
4079 void perf_event_mmap(struct vm_area_struct
*vma
)
4081 struct perf_mmap_event mmap_event
;
4083 if (!atomic_read(&nr_mmap_events
))
4086 mmap_event
= (struct perf_mmap_event
){
4092 .type
= PERF_RECORD_MMAP
,
4093 .misc
= PERF_RECORD_MISC_USER
,
4098 .start
= vma
->vm_start
,
4099 .len
= vma
->vm_end
- vma
->vm_start
,
4100 .pgoff
= (u64
)vma
->vm_pgoff
<< PAGE_SHIFT
,
4104 perf_event_mmap_event(&mmap_event
);
4108 * IRQ throttle logging
4111 static void perf_log_throttle(struct perf_event
*event
, int enable
)
4113 struct perf_output_handle handle
;
4117 struct perf_event_header header
;
4121 } throttle_event
= {
4123 .type
= PERF_RECORD_THROTTLE
,
4125 .size
= sizeof(throttle_event
),
4127 .time
= perf_clock(),
4128 .id
= primary_event_id(event
),
4129 .stream_id
= event
->id
,
4133 throttle_event
.header
.type
= PERF_RECORD_UNTHROTTLE
;
4135 ret
= perf_output_begin(&handle
, event
, sizeof(throttle_event
), 1, 0);
4139 perf_output_put(&handle
, throttle_event
);
4140 perf_output_end(&handle
);
4144 * Generic event overflow handling, sampling.
4147 static int __perf_event_overflow(struct perf_event
*event
, int nmi
,
4148 int throttle
, struct perf_sample_data
*data
,
4149 struct pt_regs
*regs
)
4151 int events
= atomic_read(&event
->event_limit
);
4152 struct hw_perf_event
*hwc
= &event
->hw
;
4155 throttle
= (throttle
&& event
->pmu
->unthrottle
!= NULL
);
4160 if (hwc
->interrupts
!= MAX_INTERRUPTS
) {
4162 if (HZ
* hwc
->interrupts
>
4163 (u64
)sysctl_perf_event_sample_rate
) {
4164 hwc
->interrupts
= MAX_INTERRUPTS
;
4165 perf_log_throttle(event
, 0);
4170 * Keep re-disabling events even though on the previous
4171 * pass we disabled it - just in case we raced with a
4172 * sched-in and the event got enabled again:
4178 if (event
->attr
.freq
) {
4179 u64 now
= perf_clock();
4180 s64 delta
= now
- hwc
->freq_time_stamp
;
4182 hwc
->freq_time_stamp
= now
;
4184 if (delta
> 0 && delta
< 2*TICK_NSEC
)
4185 perf_adjust_period(event
, delta
, hwc
->last_period
);
4189 * XXX event_limit might not quite work as expected on inherited
4193 event
->pending_kill
= POLL_IN
;
4194 if (events
&& atomic_dec_and_test(&event
->event_limit
)) {
4196 event
->pending_kill
= POLL_HUP
;
4198 event
->pending_disable
= 1;
4199 perf_pending_queue(&event
->pending
,
4200 perf_pending_event
);
4202 perf_event_disable(event
);
4205 if (event
->overflow_handler
)
4206 event
->overflow_handler(event
, nmi
, data
, regs
);
4208 perf_event_output(event
, nmi
, data
, regs
);
4213 int perf_event_overflow(struct perf_event
*event
, int nmi
,
4214 struct perf_sample_data
*data
,
4215 struct pt_regs
*regs
)
4217 return __perf_event_overflow(event
, nmi
, 1, data
, regs
);
4221 * Generic software event infrastructure
4225 * We directly increment event->count and keep a second value in
4226 * event->hw.period_left to count intervals. This period event
4227 * is kept in the range [-sample_period, 0] so that we can use the
4231 static u64
perf_swevent_set_period(struct perf_event
*event
)
4233 struct hw_perf_event
*hwc
= &event
->hw
;
4234 u64 period
= hwc
->last_period
;
4238 hwc
->last_period
= hwc
->sample_period
;
4241 old
= val
= local64_read(&hwc
->period_left
);
4245 nr
= div64_u64(period
+ val
, period
);
4246 offset
= nr
* period
;
4248 if (local64_cmpxchg(&hwc
->period_left
, old
, val
) != old
)
4254 static void perf_swevent_overflow(struct perf_event
*event
, u64 overflow
,
4255 int nmi
, struct perf_sample_data
*data
,
4256 struct pt_regs
*regs
)
4258 struct hw_perf_event
*hwc
= &event
->hw
;
4261 data
->period
= event
->hw
.last_period
;
4263 overflow
= perf_swevent_set_period(event
);
4265 if (hwc
->interrupts
== MAX_INTERRUPTS
)
4268 for (; overflow
; overflow
--) {
4269 if (__perf_event_overflow(event
, nmi
, throttle
,
4272 * We inhibit the overflow from happening when
4273 * hwc->interrupts == MAX_INTERRUPTS.
4281 static void perf_swevent_add(struct perf_event
*event
, u64 nr
,
4282 int nmi
, struct perf_sample_data
*data
,
4283 struct pt_regs
*regs
)
4285 struct hw_perf_event
*hwc
= &event
->hw
;
4287 local64_add(nr
, &event
->count
);
4292 if (!hwc
->sample_period
)
4295 if (nr
== 1 && hwc
->sample_period
== 1 && !event
->attr
.freq
)
4296 return perf_swevent_overflow(event
, 1, nmi
, data
, regs
);
4298 if (local64_add_negative(nr
, &hwc
->period_left
))
4301 perf_swevent_overflow(event
, 0, nmi
, data
, regs
);
4304 static int perf_exclude_event(struct perf_event
*event
,
4305 struct pt_regs
*regs
)
4308 if (event
->attr
.exclude_user
&& user_mode(regs
))
4311 if (event
->attr
.exclude_kernel
&& !user_mode(regs
))
4318 static int perf_swevent_match(struct perf_event
*event
,
4319 enum perf_type_id type
,
4321 struct perf_sample_data
*data
,
4322 struct pt_regs
*regs
)
4324 if (event
->attr
.type
!= type
)
4327 if (event
->attr
.config
!= event_id
)
4330 if (perf_exclude_event(event
, regs
))
4336 static inline u64
swevent_hash(u64 type
, u32 event_id
)
4338 u64 val
= event_id
| (type
<< 32);
4340 return hash_64(val
, SWEVENT_HLIST_BITS
);
4343 static inline struct hlist_head
*
4344 __find_swevent_head(struct swevent_hlist
*hlist
, u64 type
, u32 event_id
)
4346 u64 hash
= swevent_hash(type
, event_id
);
4348 return &hlist
->heads
[hash
];
4351 /* For the read side: events when they trigger */
4352 static inline struct hlist_head
*
4353 find_swevent_head_rcu(struct perf_cpu_context
*ctx
, u64 type
, u32 event_id
)
4355 struct swevent_hlist
*hlist
;
4357 hlist
= rcu_dereference(ctx
->swevent_hlist
);
4361 return __find_swevent_head(hlist
, type
, event_id
);
4364 /* For the event head insertion and removal in the hlist */
4365 static inline struct hlist_head
*
4366 find_swevent_head(struct perf_cpu_context
*ctx
, struct perf_event
*event
)
4368 struct swevent_hlist
*hlist
;
4369 u32 event_id
= event
->attr
.config
;
4370 u64 type
= event
->attr
.type
;
4373 * Event scheduling is always serialized against hlist allocation
4374 * and release. Which makes the protected version suitable here.
4375 * The context lock guarantees that.
4377 hlist
= rcu_dereference_protected(ctx
->swevent_hlist
,
4378 lockdep_is_held(&event
->ctx
->lock
));
4382 return __find_swevent_head(hlist
, type
, event_id
);
4385 static void do_perf_sw_event(enum perf_type_id type
, u32 event_id
,
4387 struct perf_sample_data
*data
,
4388 struct pt_regs
*regs
)
4390 struct perf_cpu_context
*cpuctx
;
4391 struct perf_event
*event
;
4392 struct hlist_node
*node
;
4393 struct hlist_head
*head
;
4395 cpuctx
= &__get_cpu_var(perf_cpu_context
);
4399 head
= find_swevent_head_rcu(cpuctx
, type
, event_id
);
4404 hlist_for_each_entry_rcu(event
, node
, head
, hlist_entry
) {
4405 if (perf_swevent_match(event
, type
, event_id
, data
, regs
))
4406 perf_swevent_add(event
, nr
, nmi
, data
, regs
);
4412 int perf_swevent_get_recursion_context(void)
4414 struct perf_cpu_context
*cpuctx
= &__get_cpu_var(perf_cpu_context
);
4416 return get_recursion_context(cpuctx
->recursion
);
4418 EXPORT_SYMBOL_GPL(perf_swevent_get_recursion_context
);
4420 void inline perf_swevent_put_recursion_context(int rctx
)
4422 struct perf_cpu_context
*cpuctx
= &__get_cpu_var(perf_cpu_context
);
4424 put_recursion_context(cpuctx
->recursion
, rctx
);
4427 void __perf_sw_event(u32 event_id
, u64 nr
, int nmi
,
4428 struct pt_regs
*regs
, u64 addr
)
4430 struct perf_sample_data data
;
4433 preempt_disable_notrace();
4434 rctx
= perf_swevent_get_recursion_context();
4438 perf_sample_data_init(&data
, addr
);
4440 do_perf_sw_event(PERF_TYPE_SOFTWARE
, event_id
, nr
, nmi
, &data
, regs
);
4442 perf_swevent_put_recursion_context(rctx
);
4443 preempt_enable_notrace();
4446 static void perf_swevent_read(struct perf_event
*event
)
4450 static int perf_swevent_enable(struct perf_event
*event
)
4452 struct hw_perf_event
*hwc
= &event
->hw
;
4453 struct perf_cpu_context
*cpuctx
;
4454 struct hlist_head
*head
;
4456 cpuctx
= &__get_cpu_var(perf_cpu_context
);
4458 if (hwc
->sample_period
) {
4459 hwc
->last_period
= hwc
->sample_period
;
4460 perf_swevent_set_period(event
);
4463 head
= find_swevent_head(cpuctx
, event
);
4464 if (WARN_ON_ONCE(!head
))
4467 hlist_add_head_rcu(&event
->hlist_entry
, head
);
4472 static void perf_swevent_disable(struct perf_event
*event
)
4474 hlist_del_rcu(&event
->hlist_entry
);
4477 static void perf_swevent_void(struct perf_event
*event
)
4481 static int perf_swevent_int(struct perf_event
*event
)
4486 static const struct pmu perf_ops_generic
= {
4487 .enable
= perf_swevent_enable
,
4488 .disable
= perf_swevent_disable
,
4489 .start
= perf_swevent_int
,
4490 .stop
= perf_swevent_void
,
4491 .read
= perf_swevent_read
,
4492 .unthrottle
= perf_swevent_void
, /* hwc->interrupts already reset */
4496 * hrtimer based swevent callback
4499 static enum hrtimer_restart
perf_swevent_hrtimer(struct hrtimer
*hrtimer
)
4501 enum hrtimer_restart ret
= HRTIMER_RESTART
;
4502 struct perf_sample_data data
;
4503 struct pt_regs
*regs
;
4504 struct perf_event
*event
;
4507 event
= container_of(hrtimer
, struct perf_event
, hw
.hrtimer
);
4508 event
->pmu
->read(event
);
4510 perf_sample_data_init(&data
, 0);
4511 data
.period
= event
->hw
.last_period
;
4512 regs
= get_irq_regs();
4514 if (regs
&& !perf_exclude_event(event
, regs
)) {
4515 if (!(event
->attr
.exclude_idle
&& current
->pid
== 0))
4516 if (perf_event_overflow(event
, 0, &data
, regs
))
4517 ret
= HRTIMER_NORESTART
;
4520 period
= max_t(u64
, 10000, event
->hw
.sample_period
);
4521 hrtimer_forward_now(hrtimer
, ns_to_ktime(period
));
4526 static void perf_swevent_start_hrtimer(struct perf_event
*event
)
4528 struct hw_perf_event
*hwc
= &event
->hw
;
4530 hrtimer_init(&hwc
->hrtimer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
4531 hwc
->hrtimer
.function
= perf_swevent_hrtimer
;
4532 if (hwc
->sample_period
) {
4535 if (hwc
->remaining
) {
4536 if (hwc
->remaining
< 0)
4539 period
= hwc
->remaining
;
4542 period
= max_t(u64
, 10000, hwc
->sample_period
);
4544 __hrtimer_start_range_ns(&hwc
->hrtimer
,
4545 ns_to_ktime(period
), 0,
4546 HRTIMER_MODE_REL
, 0);
4550 static void perf_swevent_cancel_hrtimer(struct perf_event
*event
)
4552 struct hw_perf_event
*hwc
= &event
->hw
;
4554 if (hwc
->sample_period
) {
4555 ktime_t remaining
= hrtimer_get_remaining(&hwc
->hrtimer
);
4556 hwc
->remaining
= ktime_to_ns(remaining
);
4558 hrtimer_cancel(&hwc
->hrtimer
);
4563 * Software event: cpu wall time clock
4566 static void cpu_clock_perf_event_update(struct perf_event
*event
)
4568 int cpu
= raw_smp_processor_id();
4572 now
= cpu_clock(cpu
);
4573 prev
= local64_xchg(&event
->hw
.prev_count
, now
);
4574 local64_add(now
- prev
, &event
->count
);
4577 static int cpu_clock_perf_event_enable(struct perf_event
*event
)
4579 struct hw_perf_event
*hwc
= &event
->hw
;
4580 int cpu
= raw_smp_processor_id();
4582 local64_set(&hwc
->prev_count
, cpu_clock(cpu
));
4583 perf_swevent_start_hrtimer(event
);
4588 static void cpu_clock_perf_event_disable(struct perf_event
*event
)
4590 perf_swevent_cancel_hrtimer(event
);
4591 cpu_clock_perf_event_update(event
);
4594 static void cpu_clock_perf_event_read(struct perf_event
*event
)
4596 cpu_clock_perf_event_update(event
);
4599 static const struct pmu perf_ops_cpu_clock
= {
4600 .enable
= cpu_clock_perf_event_enable
,
4601 .disable
= cpu_clock_perf_event_disable
,
4602 .read
= cpu_clock_perf_event_read
,
4606 * Software event: task time clock
4609 static void task_clock_perf_event_update(struct perf_event
*event
, u64 now
)
4614 prev
= local64_xchg(&event
->hw
.prev_count
, now
);
4616 local64_add(delta
, &event
->count
);
4619 static int task_clock_perf_event_enable(struct perf_event
*event
)
4621 struct hw_perf_event
*hwc
= &event
->hw
;
4624 now
= event
->ctx
->time
;
4626 local64_set(&hwc
->prev_count
, now
);
4628 perf_swevent_start_hrtimer(event
);
4633 static void task_clock_perf_event_disable(struct perf_event
*event
)
4635 perf_swevent_cancel_hrtimer(event
);
4636 task_clock_perf_event_update(event
, event
->ctx
->time
);
4640 static void task_clock_perf_event_read(struct perf_event
*event
)
4645 update_context_time(event
->ctx
);
4646 time
= event
->ctx
->time
;
4648 u64 now
= perf_clock();
4649 u64 delta
= now
- event
->ctx
->timestamp
;
4650 time
= event
->ctx
->time
+ delta
;
4653 task_clock_perf_event_update(event
, time
);
4656 static const struct pmu perf_ops_task_clock
= {
4657 .enable
= task_clock_perf_event_enable
,
4658 .disable
= task_clock_perf_event_disable
,
4659 .read
= task_clock_perf_event_read
,
4662 /* Deref the hlist from the update side */
4663 static inline struct swevent_hlist
*
4664 swevent_hlist_deref(struct perf_cpu_context
*cpuctx
)
4666 return rcu_dereference_protected(cpuctx
->swevent_hlist
,
4667 lockdep_is_held(&cpuctx
->hlist_mutex
));
4670 static void swevent_hlist_release_rcu(struct rcu_head
*rcu_head
)
4672 struct swevent_hlist
*hlist
;
4674 hlist
= container_of(rcu_head
, struct swevent_hlist
, rcu_head
);
4678 static void swevent_hlist_release(struct perf_cpu_context
*cpuctx
)
4680 struct swevent_hlist
*hlist
= swevent_hlist_deref(cpuctx
);
4685 rcu_assign_pointer(cpuctx
->swevent_hlist
, NULL
);
4686 call_rcu(&hlist
->rcu_head
, swevent_hlist_release_rcu
);
4689 static void swevent_hlist_put_cpu(struct perf_event
*event
, int cpu
)
4691 struct perf_cpu_context
*cpuctx
= &per_cpu(perf_cpu_context
, cpu
);
4693 mutex_lock(&cpuctx
->hlist_mutex
);
4695 if (!--cpuctx
->hlist_refcount
)
4696 swevent_hlist_release(cpuctx
);
4698 mutex_unlock(&cpuctx
->hlist_mutex
);
4701 static void swevent_hlist_put(struct perf_event
*event
)
4705 if (event
->cpu
!= -1) {
4706 swevent_hlist_put_cpu(event
, event
->cpu
);
4710 for_each_possible_cpu(cpu
)
4711 swevent_hlist_put_cpu(event
, cpu
);
4714 static int swevent_hlist_get_cpu(struct perf_event
*event
, int cpu
)
4716 struct perf_cpu_context
*cpuctx
= &per_cpu(perf_cpu_context
, cpu
);
4719 mutex_lock(&cpuctx
->hlist_mutex
);
4721 if (!swevent_hlist_deref(cpuctx
) && cpu_online(cpu
)) {
4722 struct swevent_hlist
*hlist
;
4724 hlist
= kzalloc(sizeof(*hlist
), GFP_KERNEL
);
4729 rcu_assign_pointer(cpuctx
->swevent_hlist
, hlist
);
4731 cpuctx
->hlist_refcount
++;
4733 mutex_unlock(&cpuctx
->hlist_mutex
);
4738 static int swevent_hlist_get(struct perf_event
*event
)
4741 int cpu
, failed_cpu
;
4743 if (event
->cpu
!= -1)
4744 return swevent_hlist_get_cpu(event
, event
->cpu
);
4747 for_each_possible_cpu(cpu
) {
4748 err
= swevent_hlist_get_cpu(event
, cpu
);
4758 for_each_possible_cpu(cpu
) {
4759 if (cpu
== failed_cpu
)
4761 swevent_hlist_put_cpu(event
, cpu
);
4768 #ifdef CONFIG_EVENT_TRACING
4770 static const struct pmu perf_ops_tracepoint
= {
4771 .enable
= perf_trace_enable
,
4772 .disable
= perf_trace_disable
,
4773 .start
= perf_swevent_int
,
4774 .stop
= perf_swevent_void
,
4775 .read
= perf_swevent_read
,
4776 .unthrottle
= perf_swevent_void
,
4779 static int perf_tp_filter_match(struct perf_event
*event
,
4780 struct perf_sample_data
*data
)
4782 void *record
= data
->raw
->data
;
4784 if (likely(!event
->filter
) || filter_match_preds(event
->filter
, record
))
4789 static int perf_tp_event_match(struct perf_event
*event
,
4790 struct perf_sample_data
*data
,
4791 struct pt_regs
*regs
)
4794 * All tracepoints are from kernel-space.
4796 if (event
->attr
.exclude_kernel
)
4799 if (!perf_tp_filter_match(event
, data
))
4805 void perf_tp_event(u64 addr
, u64 count
, void *record
, int entry_size
,
4806 struct pt_regs
*regs
, struct hlist_head
*head
, int rctx
)
4808 struct perf_sample_data data
;
4809 struct perf_event
*event
;
4810 struct hlist_node
*node
;
4812 struct perf_raw_record raw
= {
4817 perf_sample_data_init(&data
, addr
);
4820 hlist_for_each_entry_rcu(event
, node
, head
, hlist_entry
) {
4821 if (perf_tp_event_match(event
, &data
, regs
))
4822 perf_swevent_add(event
, count
, 1, &data
, regs
);
4825 perf_swevent_put_recursion_context(rctx
);
4827 EXPORT_SYMBOL_GPL(perf_tp_event
);
4829 static void tp_perf_event_destroy(struct perf_event
*event
)
4831 perf_trace_destroy(event
);
4834 static const struct pmu
*tp_perf_event_init(struct perf_event
*event
)
4839 * Raw tracepoint data is a severe data leak, only allow root to
4842 if ((event
->attr
.sample_type
& PERF_SAMPLE_RAW
) &&
4843 perf_paranoid_tracepoint_raw() &&
4844 !capable(CAP_SYS_ADMIN
))
4845 return ERR_PTR(-EPERM
);
4847 err
= perf_trace_init(event
);
4851 event
->destroy
= tp_perf_event_destroy
;
4853 return &perf_ops_tracepoint
;
4856 static int perf_event_set_filter(struct perf_event
*event
, void __user
*arg
)
4861 if (event
->attr
.type
!= PERF_TYPE_TRACEPOINT
)
4864 filter_str
= strndup_user(arg
, PAGE_SIZE
);
4865 if (IS_ERR(filter_str
))
4866 return PTR_ERR(filter_str
);
4868 ret
= ftrace_profile_set_filter(event
, event
->attr
.config
, filter_str
);
4874 static void perf_event_free_filter(struct perf_event
*event
)
4876 ftrace_profile_free_filter(event
);
4881 static const struct pmu
*tp_perf_event_init(struct perf_event
*event
)
4886 static int perf_event_set_filter(struct perf_event
*event
, void __user
*arg
)
4891 static void perf_event_free_filter(struct perf_event
*event
)
4895 #endif /* CONFIG_EVENT_TRACING */
4897 #ifdef CONFIG_HAVE_HW_BREAKPOINT
4898 static void bp_perf_event_destroy(struct perf_event
*event
)
4900 release_bp_slot(event
);
4903 static const struct pmu
*bp_perf_event_init(struct perf_event
*bp
)
4907 err
= register_perf_hw_breakpoint(bp
);
4909 return ERR_PTR(err
);
4911 bp
->destroy
= bp_perf_event_destroy
;
4913 return &perf_ops_bp
;
4916 void perf_bp_event(struct perf_event
*bp
, void *data
)
4918 struct perf_sample_data sample
;
4919 struct pt_regs
*regs
= data
;
4921 perf_sample_data_init(&sample
, bp
->attr
.bp_addr
);
4923 if (!perf_exclude_event(bp
, regs
))
4924 perf_swevent_add(bp
, 1, 1, &sample
, regs
);
4927 static const struct pmu
*bp_perf_event_init(struct perf_event
*bp
)
4932 void perf_bp_event(struct perf_event
*bp
, void *regs
)
4937 atomic_t perf_swevent_enabled
[PERF_COUNT_SW_MAX
];
4939 static void sw_perf_event_destroy(struct perf_event
*event
)
4941 u64 event_id
= event
->attr
.config
;
4943 WARN_ON(event
->parent
);
4945 atomic_dec(&perf_swevent_enabled
[event_id
]);
4946 swevent_hlist_put(event
);
4949 static const struct pmu
*sw_perf_event_init(struct perf_event
*event
)
4951 const struct pmu
*pmu
= NULL
;
4952 u64 event_id
= event
->attr
.config
;
4955 * Software events (currently) can't in general distinguish
4956 * between user, kernel and hypervisor events.
4957 * However, context switches and cpu migrations are considered
4958 * to be kernel events, and page faults are never hypervisor
4962 case PERF_COUNT_SW_CPU_CLOCK
:
4963 pmu
= &perf_ops_cpu_clock
;
4966 case PERF_COUNT_SW_TASK_CLOCK
:
4968 * If the user instantiates this as a per-cpu event,
4969 * use the cpu_clock event instead.
4971 if (event
->ctx
->task
)
4972 pmu
= &perf_ops_task_clock
;
4974 pmu
= &perf_ops_cpu_clock
;
4977 case PERF_COUNT_SW_PAGE_FAULTS
:
4978 case PERF_COUNT_SW_PAGE_FAULTS_MIN
:
4979 case PERF_COUNT_SW_PAGE_FAULTS_MAJ
:
4980 case PERF_COUNT_SW_CONTEXT_SWITCHES
:
4981 case PERF_COUNT_SW_CPU_MIGRATIONS
:
4982 case PERF_COUNT_SW_ALIGNMENT_FAULTS
:
4983 case PERF_COUNT_SW_EMULATION_FAULTS
:
4984 if (!event
->parent
) {
4987 err
= swevent_hlist_get(event
);
4989 return ERR_PTR(err
);
4991 atomic_inc(&perf_swevent_enabled
[event_id
]);
4992 event
->destroy
= sw_perf_event_destroy
;
4994 pmu
= &perf_ops_generic
;
5002 * Allocate and initialize a event structure
5004 static struct perf_event
*
5005 perf_event_alloc(struct perf_event_attr
*attr
,
5007 struct perf_event_context
*ctx
,
5008 struct perf_event
*group_leader
,
5009 struct perf_event
*parent_event
,
5010 perf_overflow_handler_t overflow_handler
,
5013 const struct pmu
*pmu
;
5014 struct perf_event
*event
;
5015 struct hw_perf_event
*hwc
;
5018 event
= kzalloc(sizeof(*event
), gfpflags
);
5020 return ERR_PTR(-ENOMEM
);
5023 * Single events are their own group leaders, with an
5024 * empty sibling list:
5027 group_leader
= event
;
5029 mutex_init(&event
->child_mutex
);
5030 INIT_LIST_HEAD(&event
->child_list
);
5032 INIT_LIST_HEAD(&event
->group_entry
);
5033 INIT_LIST_HEAD(&event
->event_entry
);
5034 INIT_LIST_HEAD(&event
->sibling_list
);
5035 init_waitqueue_head(&event
->waitq
);
5037 mutex_init(&event
->mmap_mutex
);
5040 event
->attr
= *attr
;
5041 event
->group_leader
= group_leader
;
5046 event
->parent
= parent_event
;
5048 event
->ns
= get_pid_ns(current
->nsproxy
->pid_ns
);
5049 event
->id
= atomic64_inc_return(&perf_event_id
);
5051 event
->state
= PERF_EVENT_STATE_INACTIVE
;
5053 if (!overflow_handler
&& parent_event
)
5054 overflow_handler
= parent_event
->overflow_handler
;
5056 event
->overflow_handler
= overflow_handler
;
5059 event
->state
= PERF_EVENT_STATE_OFF
;
5064 hwc
->sample_period
= attr
->sample_period
;
5065 if (attr
->freq
&& attr
->sample_freq
)
5066 hwc
->sample_period
= 1;
5067 hwc
->last_period
= hwc
->sample_period
;
5069 local64_set(&hwc
->period_left
, hwc
->sample_period
);
5072 * we currently do not support PERF_FORMAT_GROUP on inherited events
5074 if (attr
->inherit
&& (attr
->read_format
& PERF_FORMAT_GROUP
))
5077 switch (attr
->type
) {
5079 case PERF_TYPE_HARDWARE
:
5080 case PERF_TYPE_HW_CACHE
:
5081 pmu
= hw_perf_event_init(event
);
5084 case PERF_TYPE_SOFTWARE
:
5085 pmu
= sw_perf_event_init(event
);
5088 case PERF_TYPE_TRACEPOINT
:
5089 pmu
= tp_perf_event_init(event
);
5092 case PERF_TYPE_BREAKPOINT
:
5093 pmu
= bp_perf_event_init(event
);
5104 else if (IS_ERR(pmu
))
5109 put_pid_ns(event
->ns
);
5111 return ERR_PTR(err
);
5116 if (!event
->parent
) {
5117 atomic_inc(&nr_events
);
5118 if (event
->attr
.mmap
|| event
->attr
.mmap_data
)
5119 atomic_inc(&nr_mmap_events
);
5120 if (event
->attr
.comm
)
5121 atomic_inc(&nr_comm_events
);
5122 if (event
->attr
.task
)
5123 atomic_inc(&nr_task_events
);
5124 if (event
->attr
.sample_type
& PERF_SAMPLE_CALLCHAIN
) {
5125 err
= get_callchain_buffers();
5128 return ERR_PTR(err
);
5136 static int perf_copy_attr(struct perf_event_attr __user
*uattr
,
5137 struct perf_event_attr
*attr
)
5142 if (!access_ok(VERIFY_WRITE
, uattr
, PERF_ATTR_SIZE_VER0
))
5146 * zero the full structure, so that a short copy will be nice.
5148 memset(attr
, 0, sizeof(*attr
));
5150 ret
= get_user(size
, &uattr
->size
);
5154 if (size
> PAGE_SIZE
) /* silly large */
5157 if (!size
) /* abi compat */
5158 size
= PERF_ATTR_SIZE_VER0
;
5160 if (size
< PERF_ATTR_SIZE_VER0
)
5164 * If we're handed a bigger struct than we know of,
5165 * ensure all the unknown bits are 0 - i.e. new
5166 * user-space does not rely on any kernel feature
5167 * extensions we dont know about yet.
5169 if (size
> sizeof(*attr
)) {
5170 unsigned char __user
*addr
;
5171 unsigned char __user
*end
;
5174 addr
= (void __user
*)uattr
+ sizeof(*attr
);
5175 end
= (void __user
*)uattr
+ size
;
5177 for (; addr
< end
; addr
++) {
5178 ret
= get_user(val
, addr
);
5184 size
= sizeof(*attr
);
5187 ret
= copy_from_user(attr
, uattr
, size
);
5192 * If the type exists, the corresponding creation will verify
5195 if (attr
->type
>= PERF_TYPE_MAX
)
5198 if (attr
->__reserved_1
)
5201 if (attr
->sample_type
& ~(PERF_SAMPLE_MAX
-1))
5204 if (attr
->read_format
& ~(PERF_FORMAT_MAX
-1))
5211 put_user(sizeof(*attr
), &uattr
->size
);
5217 perf_event_set_output(struct perf_event
*event
, struct perf_event
*output_event
)
5219 struct perf_buffer
*buffer
= NULL
, *old_buffer
= NULL
;
5225 /* don't allow circular references */
5226 if (event
== output_event
)
5230 * Don't allow cross-cpu buffers
5232 if (output_event
->cpu
!= event
->cpu
)
5236 * If its not a per-cpu buffer, it must be the same task.
5238 if (output_event
->cpu
== -1 && output_event
->ctx
!= event
->ctx
)
5242 mutex_lock(&event
->mmap_mutex
);
5243 /* Can't redirect output if we've got an active mmap() */
5244 if (atomic_read(&event
->mmap_count
))
5248 /* get the buffer we want to redirect to */
5249 buffer
= perf_buffer_get(output_event
);
5254 old_buffer
= event
->buffer
;
5255 rcu_assign_pointer(event
->buffer
, buffer
);
5258 mutex_unlock(&event
->mmap_mutex
);
5261 perf_buffer_put(old_buffer
);
5267 * sys_perf_event_open - open a performance event, associate it to a task/cpu
5269 * @attr_uptr: event_id type attributes for monitoring/sampling
5272 * @group_fd: group leader event fd
5274 SYSCALL_DEFINE5(perf_event_open
,
5275 struct perf_event_attr __user
*, attr_uptr
,
5276 pid_t
, pid
, int, cpu
, int, group_fd
, unsigned long, flags
)
5278 struct perf_event
*event
, *group_leader
= NULL
, *output_event
= NULL
;
5279 struct perf_event_attr attr
;
5280 struct perf_event_context
*ctx
;
5281 struct file
*event_file
= NULL
;
5282 struct file
*group_file
= NULL
;
5284 int fput_needed
= 0;
5287 /* for future expandability... */
5288 if (flags
& ~(PERF_FLAG_FD_NO_GROUP
| PERF_FLAG_FD_OUTPUT
))
5291 err
= perf_copy_attr(attr_uptr
, &attr
);
5295 if (!attr
.exclude_kernel
) {
5296 if (perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN
))
5301 if (attr
.sample_freq
> sysctl_perf_event_sample_rate
)
5305 event_fd
= get_unused_fd_flags(O_RDWR
);
5310 * Get the target context (task or percpu):
5312 ctx
= find_get_context(pid
, cpu
);
5318 if (group_fd
!= -1) {
5319 group_leader
= perf_fget_light(group_fd
, &fput_needed
);
5320 if (IS_ERR(group_leader
)) {
5321 err
= PTR_ERR(group_leader
);
5322 goto err_put_context
;
5324 group_file
= group_leader
->filp
;
5325 if (flags
& PERF_FLAG_FD_OUTPUT
)
5326 output_event
= group_leader
;
5327 if (flags
& PERF_FLAG_FD_NO_GROUP
)
5328 group_leader
= NULL
;
5332 * Look up the group leader (we will attach this event to it):
5338 * Do not allow a recursive hierarchy (this new sibling
5339 * becoming part of another group-sibling):
5341 if (group_leader
->group_leader
!= group_leader
)
5342 goto err_put_context
;
5344 * Do not allow to attach to a group in a different
5345 * task or CPU context:
5347 if (group_leader
->ctx
!= ctx
)
5348 goto err_put_context
;
5350 * Only a group leader can be exclusive or pinned
5352 if (attr
.exclusive
|| attr
.pinned
)
5353 goto err_put_context
;
5356 event
= perf_event_alloc(&attr
, cpu
, ctx
, group_leader
,
5357 NULL
, NULL
, GFP_KERNEL
);
5358 if (IS_ERR(event
)) {
5359 err
= PTR_ERR(event
);
5360 goto err_put_context
;
5364 err
= perf_event_set_output(event
, output_event
);
5366 goto err_free_put_context
;
5369 event_file
= anon_inode_getfile("[perf_event]", &perf_fops
, event
, O_RDWR
);
5370 if (IS_ERR(event_file
)) {
5371 err
= PTR_ERR(event_file
);
5372 goto err_free_put_context
;
5375 event
->filp
= event_file
;
5376 WARN_ON_ONCE(ctx
->parent_ctx
);
5377 mutex_lock(&ctx
->mutex
);
5378 perf_install_in_context(ctx
, event
, cpu
);
5380 mutex_unlock(&ctx
->mutex
);
5382 event
->owner
= current
;
5383 get_task_struct(current
);
5384 mutex_lock(¤t
->perf_event_mutex
);
5385 list_add_tail(&event
->owner_entry
, ¤t
->perf_event_list
);
5386 mutex_unlock(¤t
->perf_event_mutex
);
5389 * Drop the reference on the group_event after placing the
5390 * new event on the sibling_list. This ensures destruction
5391 * of the group leader will find the pointer to itself in
5392 * perf_group_detach().
5394 fput_light(group_file
, fput_needed
);
5395 fd_install(event_fd
, event_file
);
5398 err_free_put_context
:
5401 fput_light(group_file
, fput_needed
);
5404 put_unused_fd(event_fd
);
5409 * perf_event_create_kernel_counter
5411 * @attr: attributes of the counter to create
5412 * @cpu: cpu in which the counter is bound
5413 * @pid: task to profile
5416 perf_event_create_kernel_counter(struct perf_event_attr
*attr
, int cpu
,
5418 perf_overflow_handler_t overflow_handler
)
5420 struct perf_event
*event
;
5421 struct perf_event_context
*ctx
;
5425 * Get the target context (task or percpu):
5428 ctx
= find_get_context(pid
, cpu
);
5434 event
= perf_event_alloc(attr
, cpu
, ctx
, NULL
,
5435 NULL
, overflow_handler
, GFP_KERNEL
);
5436 if (IS_ERR(event
)) {
5437 err
= PTR_ERR(event
);
5438 goto err_put_context
;
5442 WARN_ON_ONCE(ctx
->parent_ctx
);
5443 mutex_lock(&ctx
->mutex
);
5444 perf_install_in_context(ctx
, event
, cpu
);
5446 mutex_unlock(&ctx
->mutex
);
5448 event
->owner
= current
;
5449 get_task_struct(current
);
5450 mutex_lock(¤t
->perf_event_mutex
);
5451 list_add_tail(&event
->owner_entry
, ¤t
->perf_event_list
);
5452 mutex_unlock(¤t
->perf_event_mutex
);
5459 return ERR_PTR(err
);
5461 EXPORT_SYMBOL_GPL(perf_event_create_kernel_counter
);
5464 * inherit a event from parent task to child task:
5466 static struct perf_event
*
5467 inherit_event(struct perf_event
*parent_event
,
5468 struct task_struct
*parent
,
5469 struct perf_event_context
*parent_ctx
,
5470 struct task_struct
*child
,
5471 struct perf_event
*group_leader
,
5472 struct perf_event_context
*child_ctx
)
5474 struct perf_event
*child_event
;
5477 * Instead of creating recursive hierarchies of events,
5478 * we link inherited events back to the original parent,
5479 * which has a filp for sure, which we use as the reference
5482 if (parent_event
->parent
)
5483 parent_event
= parent_event
->parent
;
5485 child_event
= perf_event_alloc(&parent_event
->attr
,
5486 parent_event
->cpu
, child_ctx
,
5487 group_leader
, parent_event
,
5489 if (IS_ERR(child_event
))
5494 * Make the child state follow the state of the parent event,
5495 * not its attr.disabled bit. We hold the parent's mutex,
5496 * so we won't race with perf_event_{en, dis}able_family.
5498 if (parent_event
->state
>= PERF_EVENT_STATE_INACTIVE
)
5499 child_event
->state
= PERF_EVENT_STATE_INACTIVE
;
5501 child_event
->state
= PERF_EVENT_STATE_OFF
;
5503 if (parent_event
->attr
.freq
) {
5504 u64 sample_period
= parent_event
->hw
.sample_period
;
5505 struct hw_perf_event
*hwc
= &child_event
->hw
;
5507 hwc
->sample_period
= sample_period
;
5508 hwc
->last_period
= sample_period
;
5510 local64_set(&hwc
->period_left
, sample_period
);
5513 child_event
->overflow_handler
= parent_event
->overflow_handler
;
5516 * Link it up in the child's context:
5518 add_event_to_ctx(child_event
, child_ctx
);
5521 * Get a reference to the parent filp - we will fput it
5522 * when the child event exits. This is safe to do because
5523 * we are in the parent and we know that the filp still
5524 * exists and has a nonzero count:
5526 atomic_long_inc(&parent_event
->filp
->f_count
);
5529 * Link this into the parent event's child list
5531 WARN_ON_ONCE(parent_event
->ctx
->parent_ctx
);
5532 mutex_lock(&parent_event
->child_mutex
);
5533 list_add_tail(&child_event
->child_list
, &parent_event
->child_list
);
5534 mutex_unlock(&parent_event
->child_mutex
);
5539 static int inherit_group(struct perf_event
*parent_event
,
5540 struct task_struct
*parent
,
5541 struct perf_event_context
*parent_ctx
,
5542 struct task_struct
*child
,
5543 struct perf_event_context
*child_ctx
)
5545 struct perf_event
*leader
;
5546 struct perf_event
*sub
;
5547 struct perf_event
*child_ctr
;
5549 leader
= inherit_event(parent_event
, parent
, parent_ctx
,
5550 child
, NULL
, child_ctx
);
5552 return PTR_ERR(leader
);
5553 list_for_each_entry(sub
, &parent_event
->sibling_list
, group_entry
) {
5554 child_ctr
= inherit_event(sub
, parent
, parent_ctx
,
5555 child
, leader
, child_ctx
);
5556 if (IS_ERR(child_ctr
))
5557 return PTR_ERR(child_ctr
);
5562 static void sync_child_event(struct perf_event
*child_event
,
5563 struct task_struct
*child
)
5565 struct perf_event
*parent_event
= child_event
->parent
;
5568 if (child_event
->attr
.inherit_stat
)
5569 perf_event_read_event(child_event
, child
);
5571 child_val
= perf_event_count(child_event
);
5574 * Add back the child's count to the parent's count:
5576 atomic64_add(child_val
, &parent_event
->child_count
);
5577 atomic64_add(child_event
->total_time_enabled
,
5578 &parent_event
->child_total_time_enabled
);
5579 atomic64_add(child_event
->total_time_running
,
5580 &parent_event
->child_total_time_running
);
5583 * Remove this event from the parent's list
5585 WARN_ON_ONCE(parent_event
->ctx
->parent_ctx
);
5586 mutex_lock(&parent_event
->child_mutex
);
5587 list_del_init(&child_event
->child_list
);
5588 mutex_unlock(&parent_event
->child_mutex
);
5591 * Release the parent event, if this was the last
5594 fput(parent_event
->filp
);
5598 __perf_event_exit_task(struct perf_event
*child_event
,
5599 struct perf_event_context
*child_ctx
,
5600 struct task_struct
*child
)
5602 struct perf_event
*parent_event
;
5604 perf_event_remove_from_context(child_event
);
5606 parent_event
= child_event
->parent
;
5608 * It can happen that parent exits first, and has events
5609 * that are still around due to the child reference. These
5610 * events need to be zapped - but otherwise linger.
5613 sync_child_event(child_event
, child
);
5614 free_event(child_event
);
5619 * When a child task exits, feed back event values to parent events.
5621 void perf_event_exit_task(struct task_struct
*child
)
5623 struct perf_event
*child_event
, *tmp
;
5624 struct perf_event_context
*child_ctx
;
5625 unsigned long flags
;
5627 if (likely(!child
->perf_event_ctxp
)) {
5628 perf_event_task(child
, NULL
, 0);
5632 local_irq_save(flags
);
5634 * We can't reschedule here because interrupts are disabled,
5635 * and either child is current or it is a task that can't be
5636 * scheduled, so we are now safe from rescheduling changing
5639 child_ctx
= child
->perf_event_ctxp
;
5640 __perf_event_task_sched_out(child_ctx
);
5643 * Take the context lock here so that if find_get_context is
5644 * reading child->perf_event_ctxp, we wait until it has
5645 * incremented the context's refcount before we do put_ctx below.
5647 raw_spin_lock(&child_ctx
->lock
);
5648 child
->perf_event_ctxp
= NULL
;
5650 * If this context is a clone; unclone it so it can't get
5651 * swapped to another process while we're removing all
5652 * the events from it.
5654 unclone_ctx(child_ctx
);
5655 update_context_time(child_ctx
);
5656 raw_spin_unlock_irqrestore(&child_ctx
->lock
, flags
);
5659 * Report the task dead after unscheduling the events so that we
5660 * won't get any samples after PERF_RECORD_EXIT. We can however still
5661 * get a few PERF_RECORD_READ events.
5663 perf_event_task(child
, child_ctx
, 0);
5666 * We can recurse on the same lock type through:
5668 * __perf_event_exit_task()
5669 * sync_child_event()
5670 * fput(parent_event->filp)
5672 * mutex_lock(&ctx->mutex)
5674 * But since its the parent context it won't be the same instance.
5676 mutex_lock(&child_ctx
->mutex
);
5679 list_for_each_entry_safe(child_event
, tmp
, &child_ctx
->pinned_groups
,
5681 __perf_event_exit_task(child_event
, child_ctx
, child
);
5683 list_for_each_entry_safe(child_event
, tmp
, &child_ctx
->flexible_groups
,
5685 __perf_event_exit_task(child_event
, child_ctx
, child
);
5688 * If the last event was a group event, it will have appended all
5689 * its siblings to the list, but we obtained 'tmp' before that which
5690 * will still point to the list head terminating the iteration.
5692 if (!list_empty(&child_ctx
->pinned_groups
) ||
5693 !list_empty(&child_ctx
->flexible_groups
))
5696 mutex_unlock(&child_ctx
->mutex
);
5701 static void perf_free_event(struct perf_event
*event
,
5702 struct perf_event_context
*ctx
)
5704 struct perf_event
*parent
= event
->parent
;
5706 if (WARN_ON_ONCE(!parent
))
5709 mutex_lock(&parent
->child_mutex
);
5710 list_del_init(&event
->child_list
);
5711 mutex_unlock(&parent
->child_mutex
);
5715 perf_group_detach(event
);
5716 list_del_event(event
, ctx
);
5721 * free an unexposed, unused context as created by inheritance by
5722 * init_task below, used by fork() in case of fail.
5724 void perf_event_free_task(struct task_struct
*task
)
5726 struct perf_event_context
*ctx
= task
->perf_event_ctxp
;
5727 struct perf_event
*event
, *tmp
;
5732 mutex_lock(&ctx
->mutex
);
5734 list_for_each_entry_safe(event
, tmp
, &ctx
->pinned_groups
, group_entry
)
5735 perf_free_event(event
, ctx
);
5737 list_for_each_entry_safe(event
, tmp
, &ctx
->flexible_groups
,
5739 perf_free_event(event
, ctx
);
5741 if (!list_empty(&ctx
->pinned_groups
) ||
5742 !list_empty(&ctx
->flexible_groups
))
5745 mutex_unlock(&ctx
->mutex
);
5751 inherit_task_group(struct perf_event
*event
, struct task_struct
*parent
,
5752 struct perf_event_context
*parent_ctx
,
5753 struct task_struct
*child
,
5757 struct perf_event_context
*child_ctx
= child
->perf_event_ctxp
;
5759 if (!event
->attr
.inherit
) {
5766 * This is executed from the parent task context, so
5767 * inherit events that have been marked for cloning.
5768 * First allocate and initialize a context for the
5772 child_ctx
= kzalloc(sizeof(struct perf_event_context
),
5777 __perf_event_init_context(child_ctx
, child
);
5778 child
->perf_event_ctxp
= child_ctx
;
5779 get_task_struct(child
);
5782 ret
= inherit_group(event
, parent
, parent_ctx
,
5793 * Initialize the perf_event context in task_struct
5795 int perf_event_init_task(struct task_struct
*child
)
5797 struct perf_event_context
*child_ctx
, *parent_ctx
;
5798 struct perf_event_context
*cloned_ctx
;
5799 struct perf_event
*event
;
5800 struct task_struct
*parent
= current
;
5801 int inherited_all
= 1;
5804 child
->perf_event_ctxp
= NULL
;
5806 mutex_init(&child
->perf_event_mutex
);
5807 INIT_LIST_HEAD(&child
->perf_event_list
);
5809 if (likely(!parent
->perf_event_ctxp
))
5813 * If the parent's context is a clone, pin it so it won't get
5816 parent_ctx
= perf_pin_task_context(parent
);
5819 * No need to check if parent_ctx != NULL here; since we saw
5820 * it non-NULL earlier, the only reason for it to become NULL
5821 * is if we exit, and since we're currently in the middle of
5822 * a fork we can't be exiting at the same time.
5826 * Lock the parent list. No need to lock the child - not PID
5827 * hashed yet and not running, so nobody can access it.
5829 mutex_lock(&parent_ctx
->mutex
);
5832 * We dont have to disable NMIs - we are only looking at
5833 * the list, not manipulating it:
5835 list_for_each_entry(event
, &parent_ctx
->pinned_groups
, group_entry
) {
5836 ret
= inherit_task_group(event
, parent
, parent_ctx
, child
,
5842 list_for_each_entry(event
, &parent_ctx
->flexible_groups
, group_entry
) {
5843 ret
= inherit_task_group(event
, parent
, parent_ctx
, child
,
5849 child_ctx
= child
->perf_event_ctxp
;
5851 if (child_ctx
&& inherited_all
) {
5853 * Mark the child context as a clone of the parent
5854 * context, or of whatever the parent is a clone of.
5855 * Note that if the parent is a clone, it could get
5856 * uncloned at any point, but that doesn't matter
5857 * because the list of events and the generation
5858 * count can't have changed since we took the mutex.
5860 cloned_ctx
= rcu_dereference(parent_ctx
->parent_ctx
);
5862 child_ctx
->parent_ctx
= cloned_ctx
;
5863 child_ctx
->parent_gen
= parent_ctx
->parent_gen
;
5865 child_ctx
->parent_ctx
= parent_ctx
;
5866 child_ctx
->parent_gen
= parent_ctx
->generation
;
5868 get_ctx(child_ctx
->parent_ctx
);
5871 mutex_unlock(&parent_ctx
->mutex
);
5873 perf_unpin_context(parent_ctx
);
5878 static void __init
perf_event_init_all_cpus(void)
5881 struct perf_cpu_context
*cpuctx
;
5883 for_each_possible_cpu(cpu
) {
5884 cpuctx
= &per_cpu(perf_cpu_context
, cpu
);
5885 mutex_init(&cpuctx
->hlist_mutex
);
5886 __perf_event_init_context(&cpuctx
->ctx
, NULL
);
5890 static void __cpuinit
perf_event_init_cpu(int cpu
)
5892 struct perf_cpu_context
*cpuctx
;
5894 cpuctx
= &per_cpu(perf_cpu_context
, cpu
);
5896 spin_lock(&perf_resource_lock
);
5897 cpuctx
->max_pertask
= perf_max_events
- perf_reserved_percpu
;
5898 spin_unlock(&perf_resource_lock
);
5900 mutex_lock(&cpuctx
->hlist_mutex
);
5901 if (cpuctx
->hlist_refcount
> 0) {
5902 struct swevent_hlist
*hlist
;
5904 hlist
= kzalloc(sizeof(*hlist
), GFP_KERNEL
);
5905 WARN_ON_ONCE(!hlist
);
5906 rcu_assign_pointer(cpuctx
->swevent_hlist
, hlist
);
5908 mutex_unlock(&cpuctx
->hlist_mutex
);
5911 #ifdef CONFIG_HOTPLUG_CPU
5912 static void __perf_event_exit_cpu(void *info
)
5914 struct perf_cpu_context
*cpuctx
= &__get_cpu_var(perf_cpu_context
);
5915 struct perf_event_context
*ctx
= &cpuctx
->ctx
;
5916 struct perf_event
*event
, *tmp
;
5918 list_for_each_entry_safe(event
, tmp
, &ctx
->pinned_groups
, group_entry
)
5919 __perf_event_remove_from_context(event
);
5920 list_for_each_entry_safe(event
, tmp
, &ctx
->flexible_groups
, group_entry
)
5921 __perf_event_remove_from_context(event
);
5923 static void perf_event_exit_cpu(int cpu
)
5925 struct perf_cpu_context
*cpuctx
= &per_cpu(perf_cpu_context
, cpu
);
5926 struct perf_event_context
*ctx
= &cpuctx
->ctx
;
5928 mutex_lock(&cpuctx
->hlist_mutex
);
5929 swevent_hlist_release(cpuctx
);
5930 mutex_unlock(&cpuctx
->hlist_mutex
);
5932 mutex_lock(&ctx
->mutex
);
5933 smp_call_function_single(cpu
, __perf_event_exit_cpu
, NULL
, 1);
5934 mutex_unlock(&ctx
->mutex
);
5937 static inline void perf_event_exit_cpu(int cpu
) { }
5940 static int __cpuinit
5941 perf_cpu_notify(struct notifier_block
*self
, unsigned long action
, void *hcpu
)
5943 unsigned int cpu
= (long)hcpu
;
5947 case CPU_UP_PREPARE
:
5948 case CPU_UP_PREPARE_FROZEN
:
5949 perf_event_init_cpu(cpu
);
5952 case CPU_DOWN_PREPARE
:
5953 case CPU_DOWN_PREPARE_FROZEN
:
5954 perf_event_exit_cpu(cpu
);
5965 * This has to have a higher priority than migration_notifier in sched.c.
5967 static struct notifier_block __cpuinitdata perf_cpu_nb
= {
5968 .notifier_call
= perf_cpu_notify
,
5972 void __init
perf_event_init(void)
5974 perf_event_init_all_cpus();
5975 perf_cpu_notify(&perf_cpu_nb
, (unsigned long)CPU_UP_PREPARE
,
5976 (void *)(long)smp_processor_id());
5977 perf_cpu_notify(&perf_cpu_nb
, (unsigned long)CPU_ONLINE
,
5978 (void *)(long)smp_processor_id());
5979 register_cpu_notifier(&perf_cpu_nb
);
5982 static ssize_t
perf_show_reserve_percpu(struct sysdev_class
*class,
5983 struct sysdev_class_attribute
*attr
,
5986 return sprintf(buf
, "%d\n", perf_reserved_percpu
);
5990 perf_set_reserve_percpu(struct sysdev_class
*class,
5991 struct sysdev_class_attribute
*attr
,
5995 struct perf_cpu_context
*cpuctx
;
5999 err
= strict_strtoul(buf
, 10, &val
);
6002 if (val
> perf_max_events
)
6005 spin_lock(&perf_resource_lock
);
6006 perf_reserved_percpu
= val
;
6007 for_each_online_cpu(cpu
) {
6008 cpuctx
= &per_cpu(perf_cpu_context
, cpu
);
6009 raw_spin_lock_irq(&cpuctx
->ctx
.lock
);
6010 mpt
= min(perf_max_events
- cpuctx
->ctx
.nr_events
,
6011 perf_max_events
- perf_reserved_percpu
);
6012 cpuctx
->max_pertask
= mpt
;
6013 raw_spin_unlock_irq(&cpuctx
->ctx
.lock
);
6015 spin_unlock(&perf_resource_lock
);
6020 static ssize_t
perf_show_overcommit(struct sysdev_class
*class,
6021 struct sysdev_class_attribute
*attr
,
6024 return sprintf(buf
, "%d\n", perf_overcommit
);
6028 perf_set_overcommit(struct sysdev_class
*class,
6029 struct sysdev_class_attribute
*attr
,
6030 const char *buf
, size_t count
)
6035 err
= strict_strtoul(buf
, 10, &val
);
6041 spin_lock(&perf_resource_lock
);
6042 perf_overcommit
= val
;
6043 spin_unlock(&perf_resource_lock
);
6048 static SYSDEV_CLASS_ATTR(
6051 perf_show_reserve_percpu
,
6052 perf_set_reserve_percpu
6055 static SYSDEV_CLASS_ATTR(
6058 perf_show_overcommit
,
6062 static struct attribute
*perfclass_attrs
[] = {
6063 &attr_reserve_percpu
.attr
,
6064 &attr_overcommit
.attr
,
6068 static struct attribute_group perfclass_attr_group
= {
6069 .attrs
= perfclass_attrs
,
6070 .name
= "perf_events",
6073 static int __init
perf_event_sysfs_init(void)
6075 return sysfs_create_group(&cpu_sysdev_class
.kset
.kobj
,
6076 &perfclass_attr_group
);
6078 device_initcall(perf_event_sysfs_init
);