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>
35 #include <asm/irq_regs.h>
37 static atomic_t nr_events __read_mostly
;
38 static atomic_t nr_mmap_events __read_mostly
;
39 static atomic_t nr_comm_events __read_mostly
;
40 static atomic_t nr_task_events __read_mostly
;
42 static LIST_HEAD(pmus
);
43 static DEFINE_MUTEX(pmus_lock
);
44 static struct srcu_struct pmus_srcu
;
47 * perf event paranoia level:
48 * -1 - not paranoid at all
49 * 0 - disallow raw tracepoint access for unpriv
50 * 1 - disallow cpu events for unpriv
51 * 2 - disallow kernel profiling for unpriv
53 int sysctl_perf_event_paranoid __read_mostly
= 1;
55 int sysctl_perf_event_mlock __read_mostly
= 512; /* 'free' kb per user */
58 * max perf event sample rate
60 int sysctl_perf_event_sample_rate __read_mostly
= 100000;
62 static atomic64_t perf_event_id
;
64 void __weak
perf_event_print_debug(void) { }
66 void perf_pmu_disable(struct pmu
*pmu
)
68 int *count
= this_cpu_ptr(pmu
->pmu_disable_count
);
70 pmu
->pmu_disable(pmu
);
73 void perf_pmu_enable(struct pmu
*pmu
)
75 int *count
= this_cpu_ptr(pmu
->pmu_disable_count
);
80 static void perf_pmu_rotate_start(struct pmu
*pmu
)
82 struct perf_cpu_context
*cpuctx
= this_cpu_ptr(pmu
->pmu_cpu_context
);
84 if (hrtimer_active(&cpuctx
->timer
))
87 __hrtimer_start_range_ns(&cpuctx
->timer
,
88 ns_to_ktime(cpuctx
->timer_interval
), 0,
89 HRTIMER_MODE_REL_PINNED
, 0);
92 static void perf_pmu_rotate_stop(struct pmu
*pmu
)
94 struct perf_cpu_context
*cpuctx
= this_cpu_ptr(pmu
->pmu_cpu_context
);
96 hrtimer_cancel(&cpuctx
->timer
);
99 static void get_ctx(struct perf_event_context
*ctx
)
101 WARN_ON(!atomic_inc_not_zero(&ctx
->refcount
));
104 static void free_ctx(struct rcu_head
*head
)
106 struct perf_event_context
*ctx
;
108 ctx
= container_of(head
, struct perf_event_context
, rcu_head
);
112 static void put_ctx(struct perf_event_context
*ctx
)
114 if (atomic_dec_and_test(&ctx
->refcount
)) {
116 put_ctx(ctx
->parent_ctx
);
118 put_task_struct(ctx
->task
);
119 call_rcu(&ctx
->rcu_head
, free_ctx
);
123 static void unclone_ctx(struct perf_event_context
*ctx
)
125 if (ctx
->parent_ctx
) {
126 put_ctx(ctx
->parent_ctx
);
127 ctx
->parent_ctx
= NULL
;
132 * If we inherit events we want to return the parent event id
135 static u64
primary_event_id(struct perf_event
*event
)
140 id
= event
->parent
->id
;
146 * Get the perf_event_context for a task and lock it.
147 * This has to cope with with the fact that until it is locked,
148 * the context could get moved to another task.
150 static struct perf_event_context
*
151 perf_lock_task_context(struct task_struct
*task
, int ctxn
, unsigned long *flags
)
153 struct perf_event_context
*ctx
;
157 ctx
= rcu_dereference(task
->perf_event_ctxp
[ctxn
]);
160 * If this context is a clone of another, it might
161 * get swapped for another underneath us by
162 * perf_event_task_sched_out, though the
163 * rcu_read_lock() protects us from any context
164 * getting freed. Lock the context and check if it
165 * got swapped before we could get the lock, and retry
166 * if so. If we locked the right context, then it
167 * can't get swapped on us any more.
169 raw_spin_lock_irqsave(&ctx
->lock
, *flags
);
170 if (ctx
!= rcu_dereference(task
->perf_event_ctxp
[ctxn
])) {
171 raw_spin_unlock_irqrestore(&ctx
->lock
, *flags
);
175 if (!atomic_inc_not_zero(&ctx
->refcount
)) {
176 raw_spin_unlock_irqrestore(&ctx
->lock
, *flags
);
185 * Get the context for a task and increment its pin_count so it
186 * can't get swapped to another task. This also increments its
187 * reference count so that the context can't get freed.
189 static struct perf_event_context
*
190 perf_pin_task_context(struct task_struct
*task
, int ctxn
)
192 struct perf_event_context
*ctx
;
195 ctx
= perf_lock_task_context(task
, ctxn
, &flags
);
198 raw_spin_unlock_irqrestore(&ctx
->lock
, flags
);
203 static void perf_unpin_context(struct perf_event_context
*ctx
)
207 raw_spin_lock_irqsave(&ctx
->lock
, flags
);
209 raw_spin_unlock_irqrestore(&ctx
->lock
, flags
);
213 static inline u64
perf_clock(void)
215 return local_clock();
219 * Update the record of the current time in a context.
221 static void update_context_time(struct perf_event_context
*ctx
)
223 u64 now
= perf_clock();
225 ctx
->time
+= now
- ctx
->timestamp
;
226 ctx
->timestamp
= now
;
230 * Update the total_time_enabled and total_time_running fields for a event.
232 static void update_event_times(struct perf_event
*event
)
234 struct perf_event_context
*ctx
= event
->ctx
;
237 if (event
->state
< PERF_EVENT_STATE_INACTIVE
||
238 event
->group_leader
->state
< PERF_EVENT_STATE_INACTIVE
)
244 run_end
= event
->tstamp_stopped
;
246 event
->total_time_enabled
= run_end
- event
->tstamp_enabled
;
248 if (event
->state
== PERF_EVENT_STATE_INACTIVE
)
249 run_end
= event
->tstamp_stopped
;
253 event
->total_time_running
= run_end
- event
->tstamp_running
;
257 * Update total_time_enabled and total_time_running for all events in a group.
259 static void update_group_times(struct perf_event
*leader
)
261 struct perf_event
*event
;
263 update_event_times(leader
);
264 list_for_each_entry(event
, &leader
->sibling_list
, group_entry
)
265 update_event_times(event
);
268 static struct list_head
*
269 ctx_group_list(struct perf_event
*event
, struct perf_event_context
*ctx
)
271 if (event
->attr
.pinned
)
272 return &ctx
->pinned_groups
;
274 return &ctx
->flexible_groups
;
278 * Add a event from the lists for its context.
279 * Must be called with ctx->mutex and ctx->lock held.
282 list_add_event(struct perf_event
*event
, struct perf_event_context
*ctx
)
284 WARN_ON_ONCE(event
->attach_state
& PERF_ATTACH_CONTEXT
);
285 event
->attach_state
|= PERF_ATTACH_CONTEXT
;
288 * If we're a stand alone event or group leader, we go to the context
289 * list, group events are kept attached to the group so that
290 * perf_group_detach can, at all times, locate all siblings.
292 if (event
->group_leader
== event
) {
293 struct list_head
*list
;
295 if (is_software_event(event
))
296 event
->group_flags
|= PERF_GROUP_SOFTWARE
;
298 list
= ctx_group_list(event
, ctx
);
299 list_add_tail(&event
->group_entry
, list
);
302 list_add_rcu(&event
->event_entry
, &ctx
->event_list
);
304 perf_pmu_rotate_start(ctx
->pmu
);
306 if (event
->attr
.inherit_stat
)
310 static void perf_group_attach(struct perf_event
*event
)
312 struct perf_event
*group_leader
= event
->group_leader
;
314 WARN_ON_ONCE(event
->attach_state
& PERF_ATTACH_GROUP
);
315 event
->attach_state
|= PERF_ATTACH_GROUP
;
317 if (group_leader
== event
)
320 if (group_leader
->group_flags
& PERF_GROUP_SOFTWARE
&&
321 !is_software_event(event
))
322 group_leader
->group_flags
&= ~PERF_GROUP_SOFTWARE
;
324 list_add_tail(&event
->group_entry
, &group_leader
->sibling_list
);
325 group_leader
->nr_siblings
++;
329 * Remove a event from the lists for its context.
330 * Must be called with ctx->mutex and ctx->lock held.
333 list_del_event(struct perf_event
*event
, struct perf_event_context
*ctx
)
336 * We can have double detach due to exit/hot-unplug + close.
338 if (!(event
->attach_state
& PERF_ATTACH_CONTEXT
))
341 event
->attach_state
&= ~PERF_ATTACH_CONTEXT
;
344 if (event
->attr
.inherit_stat
)
347 list_del_rcu(&event
->event_entry
);
349 if (event
->group_leader
== event
)
350 list_del_init(&event
->group_entry
);
352 update_group_times(event
);
355 * If event was in error state, then keep it
356 * that way, otherwise bogus counts will be
357 * returned on read(). The only way to get out
358 * of error state is by explicit re-enabling
361 if (event
->state
> PERF_EVENT_STATE_OFF
)
362 event
->state
= PERF_EVENT_STATE_OFF
;
365 static void perf_group_detach(struct perf_event
*event
)
367 struct perf_event
*sibling
, *tmp
;
368 struct list_head
*list
= NULL
;
371 * We can have double detach due to exit/hot-unplug + close.
373 if (!(event
->attach_state
& PERF_ATTACH_GROUP
))
376 event
->attach_state
&= ~PERF_ATTACH_GROUP
;
379 * If this is a sibling, remove it from its group.
381 if (event
->group_leader
!= event
) {
382 list_del_init(&event
->group_entry
);
383 event
->group_leader
->nr_siblings
--;
387 if (!list_empty(&event
->group_entry
))
388 list
= &event
->group_entry
;
391 * If this was a group event with sibling events then
392 * upgrade the siblings to singleton events by adding them
393 * to whatever list we are on.
395 list_for_each_entry_safe(sibling
, tmp
, &event
->sibling_list
, group_entry
) {
397 list_move_tail(&sibling
->group_entry
, list
);
398 sibling
->group_leader
= sibling
;
400 /* Inherit group flags from the previous leader */
401 sibling
->group_flags
= event
->group_flags
;
406 event_filter_match(struct perf_event
*event
)
408 return event
->cpu
== -1 || event
->cpu
== smp_processor_id();
412 event_sched_out(struct perf_event
*event
,
413 struct perf_cpu_context
*cpuctx
,
414 struct perf_event_context
*ctx
)
418 * An event which could not be activated because of
419 * filter mismatch still needs to have its timings
420 * maintained, otherwise bogus information is return
421 * via read() for time_enabled, time_running:
423 if (event
->state
== PERF_EVENT_STATE_INACTIVE
424 && !event_filter_match(event
)) {
425 delta
= ctx
->time
- event
->tstamp_stopped
;
426 event
->tstamp_running
+= delta
;
427 event
->tstamp_stopped
= ctx
->time
;
430 if (event
->state
!= PERF_EVENT_STATE_ACTIVE
)
433 event
->state
= PERF_EVENT_STATE_INACTIVE
;
434 if (event
->pending_disable
) {
435 event
->pending_disable
= 0;
436 event
->state
= PERF_EVENT_STATE_OFF
;
438 event
->tstamp_stopped
= ctx
->time
;
439 event
->pmu
->del(event
, 0);
442 if (!is_software_event(event
))
443 cpuctx
->active_oncpu
--;
445 if (event
->attr
.exclusive
|| !cpuctx
->active_oncpu
)
446 cpuctx
->exclusive
= 0;
450 group_sched_out(struct perf_event
*group_event
,
451 struct perf_cpu_context
*cpuctx
,
452 struct perf_event_context
*ctx
)
454 struct perf_event
*event
;
455 int state
= group_event
->state
;
457 event_sched_out(group_event
, cpuctx
, ctx
);
460 * Schedule out siblings (if any):
462 list_for_each_entry(event
, &group_event
->sibling_list
, group_entry
)
463 event_sched_out(event
, cpuctx
, ctx
);
465 if (state
== PERF_EVENT_STATE_ACTIVE
&& group_event
->attr
.exclusive
)
466 cpuctx
->exclusive
= 0;
469 static inline struct perf_cpu_context
*
470 __get_cpu_context(struct perf_event_context
*ctx
)
472 return this_cpu_ptr(ctx
->pmu
->pmu_cpu_context
);
476 * Cross CPU call to remove a performance event
478 * We disable the event on the hardware level first. After that we
479 * remove it from the context list.
481 static void __perf_event_remove_from_context(void *info
)
483 struct perf_event
*event
= info
;
484 struct perf_event_context
*ctx
= event
->ctx
;
485 struct perf_cpu_context
*cpuctx
= __get_cpu_context(ctx
);
488 * If this is a task context, we need to check whether it is
489 * the current task context of this cpu. If not it has been
490 * scheduled out before the smp call arrived.
492 if (ctx
->task
&& cpuctx
->task_ctx
!= ctx
)
495 raw_spin_lock(&ctx
->lock
);
497 event_sched_out(event
, cpuctx
, ctx
);
499 list_del_event(event
, ctx
);
501 raw_spin_unlock(&ctx
->lock
);
506 * Remove the event from a task's (or a CPU's) list of events.
508 * Must be called with ctx->mutex held.
510 * CPU events are removed with a smp call. For task events we only
511 * call when the task is on a CPU.
513 * If event->ctx is a cloned context, callers must make sure that
514 * every task struct that event->ctx->task could possibly point to
515 * remains valid. This is OK when called from perf_release since
516 * that only calls us on the top-level context, which can't be a clone.
517 * When called from perf_event_exit_task, it's OK because the
518 * context has been detached from its task.
520 static void perf_event_remove_from_context(struct perf_event
*event
)
522 struct perf_event_context
*ctx
= event
->ctx
;
523 struct task_struct
*task
= ctx
->task
;
527 * Per cpu events are removed via an smp call and
528 * the removal is always successful.
530 smp_call_function_single(event
->cpu
,
531 __perf_event_remove_from_context
,
537 task_oncpu_function_call(task
, __perf_event_remove_from_context
,
540 raw_spin_lock_irq(&ctx
->lock
);
542 * If the context is active we need to retry the smp call.
544 if (ctx
->nr_active
&& !list_empty(&event
->group_entry
)) {
545 raw_spin_unlock_irq(&ctx
->lock
);
550 * The lock prevents that this context is scheduled in so we
551 * can remove the event safely, if the call above did not
554 if (!list_empty(&event
->group_entry
))
555 list_del_event(event
, ctx
);
556 raw_spin_unlock_irq(&ctx
->lock
);
560 * Cross CPU call to disable a performance event
562 static void __perf_event_disable(void *info
)
564 struct perf_event
*event
= info
;
565 struct perf_event_context
*ctx
= event
->ctx
;
566 struct perf_cpu_context
*cpuctx
= __get_cpu_context(ctx
);
569 * If this is a per-task event, need to check whether this
570 * event's task is the current task on this cpu.
572 if (ctx
->task
&& cpuctx
->task_ctx
!= ctx
)
575 raw_spin_lock(&ctx
->lock
);
578 * If the event is on, turn it off.
579 * If it is in error state, leave it in error state.
581 if (event
->state
>= PERF_EVENT_STATE_INACTIVE
) {
582 update_context_time(ctx
);
583 update_group_times(event
);
584 if (event
== event
->group_leader
)
585 group_sched_out(event
, cpuctx
, ctx
);
587 event_sched_out(event
, cpuctx
, ctx
);
588 event
->state
= PERF_EVENT_STATE_OFF
;
591 raw_spin_unlock(&ctx
->lock
);
597 * If event->ctx is a cloned context, callers must make sure that
598 * every task struct that event->ctx->task could possibly point to
599 * remains valid. This condition is satisifed when called through
600 * perf_event_for_each_child or perf_event_for_each because they
601 * hold the top-level event's child_mutex, so any descendant that
602 * goes to exit will block in sync_child_event.
603 * When called from perf_pending_event it's OK because event->ctx
604 * is the current context on this CPU and preemption is disabled,
605 * hence we can't get into perf_event_task_sched_out for this context.
607 void perf_event_disable(struct perf_event
*event
)
609 struct perf_event_context
*ctx
= event
->ctx
;
610 struct task_struct
*task
= ctx
->task
;
614 * Disable the event on the cpu that it's on
616 smp_call_function_single(event
->cpu
, __perf_event_disable
,
622 task_oncpu_function_call(task
, __perf_event_disable
, event
);
624 raw_spin_lock_irq(&ctx
->lock
);
626 * If the event is still active, we need to retry the cross-call.
628 if (event
->state
== PERF_EVENT_STATE_ACTIVE
) {
629 raw_spin_unlock_irq(&ctx
->lock
);
634 * Since we have the lock this context can't be scheduled
635 * in, so we can change the state safely.
637 if (event
->state
== PERF_EVENT_STATE_INACTIVE
) {
638 update_group_times(event
);
639 event
->state
= PERF_EVENT_STATE_OFF
;
642 raw_spin_unlock_irq(&ctx
->lock
);
646 event_sched_in(struct perf_event
*event
,
647 struct perf_cpu_context
*cpuctx
,
648 struct perf_event_context
*ctx
)
650 if (event
->state
<= PERF_EVENT_STATE_OFF
)
653 event
->state
= PERF_EVENT_STATE_ACTIVE
;
654 event
->oncpu
= smp_processor_id();
656 * The new state must be visible before we turn it on in the hardware:
660 if (event
->pmu
->add(event
, PERF_EF_START
)) {
661 event
->state
= PERF_EVENT_STATE_INACTIVE
;
666 event
->tstamp_running
+= ctx
->time
- event
->tstamp_stopped
;
668 if (!is_software_event(event
))
669 cpuctx
->active_oncpu
++;
672 if (event
->attr
.exclusive
)
673 cpuctx
->exclusive
= 1;
679 group_sched_in(struct perf_event
*group_event
,
680 struct perf_cpu_context
*cpuctx
,
681 struct perf_event_context
*ctx
)
683 struct perf_event
*event
, *partial_group
= NULL
;
684 struct pmu
*pmu
= group_event
->pmu
;
686 if (group_event
->state
== PERF_EVENT_STATE_OFF
)
691 if (event_sched_in(group_event
, cpuctx
, ctx
)) {
692 pmu
->cancel_txn(pmu
);
697 * Schedule in siblings as one group (if any):
699 list_for_each_entry(event
, &group_event
->sibling_list
, group_entry
) {
700 if (event_sched_in(event
, cpuctx
, ctx
)) {
701 partial_group
= event
;
706 if (!pmu
->commit_txn(pmu
))
711 * Groups can be scheduled in as one unit only, so undo any
712 * partial group before returning:
714 list_for_each_entry(event
, &group_event
->sibling_list
, group_entry
) {
715 if (event
== partial_group
)
717 event_sched_out(event
, cpuctx
, ctx
);
719 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_event
*event
= info
;
775 struct perf_event_context
*ctx
= event
->ctx
;
776 struct perf_event
*leader
= event
->group_leader
;
777 struct perf_cpu_context
*cpuctx
= __get_cpu_context(ctx
);
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
);
797 add_event_to_ctx(event
, ctx
);
799 if (event
->cpu
!= -1 && event
->cpu
!= smp_processor_id())
803 * Don't put the event on if it is disabled or if
804 * it is in a group and the group isn't on.
806 if (event
->state
!= PERF_EVENT_STATE_INACTIVE
||
807 (leader
!= event
&& leader
->state
!= PERF_EVENT_STATE_ACTIVE
))
811 * An exclusive event can't go on if there are already active
812 * hardware events, and no hardware event can go on if there
813 * is already an exclusive event on.
815 if (!group_can_go_on(event
, cpuctx
, 1))
818 err
= event_sched_in(event
, cpuctx
, ctx
);
822 * This event couldn't go on. If it is in a group
823 * then we have to pull the whole group off.
824 * If the event group is pinned then put it in error state.
827 group_sched_out(leader
, cpuctx
, ctx
);
828 if (leader
->attr
.pinned
) {
829 update_group_times(leader
);
830 leader
->state
= PERF_EVENT_STATE_ERROR
;
835 raw_spin_unlock(&ctx
->lock
);
839 * Attach a performance event to a context
841 * First we add the event to the list with the hardware enable bit
842 * in event->hw_config cleared.
844 * If the event is attached to a task which is on a CPU we use a smp
845 * call to enable it in the task context. The task might have been
846 * scheduled away, but we check this in the smp call again.
848 * Must be called with ctx->mutex held.
851 perf_install_in_context(struct perf_event_context
*ctx
,
852 struct perf_event
*event
,
855 struct task_struct
*task
= ctx
->task
;
861 * Per cpu events are installed via an smp call and
862 * the install is always successful.
864 smp_call_function_single(cpu
, __perf_install_in_context
,
870 task_oncpu_function_call(task
, __perf_install_in_context
,
873 raw_spin_lock_irq(&ctx
->lock
);
875 * we need to retry the smp call.
877 if (ctx
->is_active
&& list_empty(&event
->group_entry
)) {
878 raw_spin_unlock_irq(&ctx
->lock
);
883 * The lock prevents that this context is scheduled in so we
884 * can add the event safely, if it the call above did not
887 if (list_empty(&event
->group_entry
))
888 add_event_to_ctx(event
, ctx
);
889 raw_spin_unlock_irq(&ctx
->lock
);
893 * Put a event into inactive state and update time fields.
894 * Enabling the leader of a group effectively enables all
895 * the group members that aren't explicitly disabled, so we
896 * have to update their ->tstamp_enabled also.
897 * Note: this works for group members as well as group leaders
898 * since the non-leader members' sibling_lists will be empty.
900 static void __perf_event_mark_enabled(struct perf_event
*event
,
901 struct perf_event_context
*ctx
)
903 struct perf_event
*sub
;
905 event
->state
= PERF_EVENT_STATE_INACTIVE
;
906 event
->tstamp_enabled
= ctx
->time
- event
->total_time_enabled
;
907 list_for_each_entry(sub
, &event
->sibling_list
, group_entry
) {
908 if (sub
->state
>= PERF_EVENT_STATE_INACTIVE
) {
909 sub
->tstamp_enabled
=
910 ctx
->time
- sub
->total_time_enabled
;
916 * Cross CPU call to enable a performance event
918 static void __perf_event_enable(void *info
)
920 struct perf_event
*event
= info
;
921 struct perf_event_context
*ctx
= event
->ctx
;
922 struct perf_event
*leader
= event
->group_leader
;
923 struct perf_cpu_context
*cpuctx
= __get_cpu_context(ctx
);
927 * If this is a per-task event, need to check whether this
928 * event's task is the current task on this cpu.
930 if (ctx
->task
&& cpuctx
->task_ctx
!= ctx
) {
931 if (cpuctx
->task_ctx
|| ctx
->task
!= current
)
933 cpuctx
->task_ctx
= ctx
;
936 raw_spin_lock(&ctx
->lock
);
938 update_context_time(ctx
);
940 if (event
->state
>= PERF_EVENT_STATE_INACTIVE
)
942 __perf_event_mark_enabled(event
, ctx
);
944 if (event
->cpu
!= -1 && event
->cpu
!= smp_processor_id())
948 * If the event is in a group and isn't the group leader,
949 * then don't put it on unless the group is on.
951 if (leader
!= event
&& leader
->state
!= PERF_EVENT_STATE_ACTIVE
)
954 if (!group_can_go_on(event
, cpuctx
, 1)) {
958 err
= group_sched_in(event
, cpuctx
, ctx
);
960 err
= event_sched_in(event
, cpuctx
, ctx
);
965 * If this event can't go on and it's part of a
966 * group, then the whole group has to come off.
969 group_sched_out(leader
, cpuctx
, ctx
);
970 if (leader
->attr
.pinned
) {
971 update_group_times(leader
);
972 leader
->state
= PERF_EVENT_STATE_ERROR
;
977 raw_spin_unlock(&ctx
->lock
);
983 * If event->ctx is a cloned context, callers must make sure that
984 * every task struct that event->ctx->task could possibly point to
985 * remains valid. This condition is satisfied when called through
986 * perf_event_for_each_child or perf_event_for_each as described
987 * for perf_event_disable.
989 void perf_event_enable(struct perf_event
*event
)
991 struct perf_event_context
*ctx
= event
->ctx
;
992 struct task_struct
*task
= ctx
->task
;
996 * Enable the event on the cpu that it's on
998 smp_call_function_single(event
->cpu
, __perf_event_enable
,
1003 raw_spin_lock_irq(&ctx
->lock
);
1004 if (event
->state
>= PERF_EVENT_STATE_INACTIVE
)
1008 * If the event is in error state, clear that first.
1009 * That way, if we see the event in error state below, we
1010 * know that it has gone back into error state, as distinct
1011 * from the task having been scheduled away before the
1012 * cross-call arrived.
1014 if (event
->state
== PERF_EVENT_STATE_ERROR
)
1015 event
->state
= PERF_EVENT_STATE_OFF
;
1018 raw_spin_unlock_irq(&ctx
->lock
);
1019 task_oncpu_function_call(task
, __perf_event_enable
, event
);
1021 raw_spin_lock_irq(&ctx
->lock
);
1024 * If the context is active and the event is still off,
1025 * we need to retry the cross-call.
1027 if (ctx
->is_active
&& event
->state
== PERF_EVENT_STATE_OFF
)
1031 * Since we have the lock this context can't be scheduled
1032 * in, so we can change the state safely.
1034 if (event
->state
== PERF_EVENT_STATE_OFF
)
1035 __perf_event_mark_enabled(event
, ctx
);
1038 raw_spin_unlock_irq(&ctx
->lock
);
1041 static int perf_event_refresh(struct perf_event
*event
, int refresh
)
1044 * not supported on inherited events
1046 if (event
->attr
.inherit
)
1049 atomic_add(refresh
, &event
->event_limit
);
1050 perf_event_enable(event
);
1056 EVENT_FLEXIBLE
= 0x1,
1058 EVENT_ALL
= EVENT_FLEXIBLE
| EVENT_PINNED
,
1061 static void ctx_sched_out(struct perf_event_context
*ctx
,
1062 struct perf_cpu_context
*cpuctx
,
1063 enum event_type_t event_type
)
1065 struct perf_event
*event
;
1067 raw_spin_lock(&ctx
->lock
);
1068 perf_pmu_disable(ctx
->pmu
);
1070 if (likely(!ctx
->nr_events
))
1072 update_context_time(ctx
);
1074 if (!ctx
->nr_active
)
1077 if (event_type
& EVENT_PINNED
) {
1078 list_for_each_entry(event
, &ctx
->pinned_groups
, group_entry
)
1079 group_sched_out(event
, cpuctx
, ctx
);
1082 if (event_type
& EVENT_FLEXIBLE
) {
1083 list_for_each_entry(event
, &ctx
->flexible_groups
, group_entry
)
1084 group_sched_out(event
, cpuctx
, ctx
);
1087 perf_pmu_enable(ctx
->pmu
);
1088 raw_spin_unlock(&ctx
->lock
);
1092 * Test whether two contexts are equivalent, i.e. whether they
1093 * have both been cloned from the same version of the same context
1094 * and they both have the same number of enabled events.
1095 * If the number of enabled events is the same, then the set
1096 * of enabled events should be the same, because these are both
1097 * inherited contexts, therefore we can't access individual events
1098 * in them directly with an fd; we can only enable/disable all
1099 * events via prctl, or enable/disable all events in a family
1100 * via ioctl, which will have the same effect on both contexts.
1102 static int context_equiv(struct perf_event_context
*ctx1
,
1103 struct perf_event_context
*ctx2
)
1105 return ctx1
->parent_ctx
&& ctx1
->parent_ctx
== ctx2
->parent_ctx
1106 && ctx1
->parent_gen
== ctx2
->parent_gen
1107 && !ctx1
->pin_count
&& !ctx2
->pin_count
;
1110 static void __perf_event_sync_stat(struct perf_event
*event
,
1111 struct perf_event
*next_event
)
1115 if (!event
->attr
.inherit_stat
)
1119 * Update the event value, we cannot use perf_event_read()
1120 * because we're in the middle of a context switch and have IRQs
1121 * disabled, which upsets smp_call_function_single(), however
1122 * we know the event must be on the current CPU, therefore we
1123 * don't need to use it.
1125 switch (event
->state
) {
1126 case PERF_EVENT_STATE_ACTIVE
:
1127 event
->pmu
->read(event
);
1130 case PERF_EVENT_STATE_INACTIVE
:
1131 update_event_times(event
);
1139 * In order to keep per-task stats reliable we need to flip the event
1140 * values when we flip the contexts.
1142 value
= local64_read(&next_event
->count
);
1143 value
= local64_xchg(&event
->count
, value
);
1144 local64_set(&next_event
->count
, value
);
1146 swap(event
->total_time_enabled
, next_event
->total_time_enabled
);
1147 swap(event
->total_time_running
, next_event
->total_time_running
);
1150 * Since we swizzled the values, update the user visible data too.
1152 perf_event_update_userpage(event
);
1153 perf_event_update_userpage(next_event
);
1156 #define list_next_entry(pos, member) \
1157 list_entry(pos->member.next, typeof(*pos), member)
1159 static void perf_event_sync_stat(struct perf_event_context
*ctx
,
1160 struct perf_event_context
*next_ctx
)
1162 struct perf_event
*event
, *next_event
;
1167 update_context_time(ctx
);
1169 event
= list_first_entry(&ctx
->event_list
,
1170 struct perf_event
, event_entry
);
1172 next_event
= list_first_entry(&next_ctx
->event_list
,
1173 struct perf_event
, event_entry
);
1175 while (&event
->event_entry
!= &ctx
->event_list
&&
1176 &next_event
->event_entry
!= &next_ctx
->event_list
) {
1178 __perf_event_sync_stat(event
, next_event
);
1180 event
= list_next_entry(event
, event_entry
);
1181 next_event
= list_next_entry(next_event
, event_entry
);
1185 void perf_event_context_sched_out(struct task_struct
*task
, int ctxn
,
1186 struct task_struct
*next
)
1188 struct perf_event_context
*ctx
= task
->perf_event_ctxp
[ctxn
];
1189 struct perf_event_context
*next_ctx
;
1190 struct perf_event_context
*parent
;
1191 struct perf_cpu_context
*cpuctx
;
1197 cpuctx
= __get_cpu_context(ctx
);
1198 if (!cpuctx
->task_ctx
)
1202 parent
= rcu_dereference(ctx
->parent_ctx
);
1203 next_ctx
= next
->perf_event_ctxp
[ctxn
];
1204 if (parent
&& next_ctx
&&
1205 rcu_dereference(next_ctx
->parent_ctx
) == parent
) {
1207 * Looks like the two contexts are clones, so we might be
1208 * able to optimize the context switch. We lock both
1209 * contexts and check that they are clones under the
1210 * lock (including re-checking that neither has been
1211 * uncloned in the meantime). It doesn't matter which
1212 * order we take the locks because no other cpu could
1213 * be trying to lock both of these tasks.
1215 raw_spin_lock(&ctx
->lock
);
1216 raw_spin_lock_nested(&next_ctx
->lock
, SINGLE_DEPTH_NESTING
);
1217 if (context_equiv(ctx
, next_ctx
)) {
1219 * XXX do we need a memory barrier of sorts
1220 * wrt to rcu_dereference() of perf_event_ctxp
1222 task
->perf_event_ctxp
[ctxn
] = next_ctx
;
1223 next
->perf_event_ctxp
[ctxn
] = ctx
;
1225 next_ctx
->task
= task
;
1228 perf_event_sync_stat(ctx
, next_ctx
);
1230 raw_spin_unlock(&next_ctx
->lock
);
1231 raw_spin_unlock(&ctx
->lock
);
1236 ctx_sched_out(ctx
, cpuctx
, EVENT_ALL
);
1237 cpuctx
->task_ctx
= NULL
;
1241 #define for_each_task_context_nr(ctxn) \
1242 for ((ctxn) = 0; (ctxn) < perf_nr_task_contexts; (ctxn)++)
1245 * Called from scheduler to remove the events of the current task,
1246 * with interrupts disabled.
1248 * We stop each event and update the event value in event->count.
1250 * This does not protect us against NMI, but disable()
1251 * sets the disabled bit in the control field of event _before_
1252 * accessing the event control register. If a NMI hits, then it will
1253 * not restart the event.
1255 void perf_event_task_sched_out(struct task_struct
*task
,
1256 struct task_struct
*next
)
1260 perf_sw_event(PERF_COUNT_SW_CONTEXT_SWITCHES
, 1, 1, NULL
, 0);
1262 for_each_task_context_nr(ctxn
)
1263 perf_event_context_sched_out(task
, ctxn
, next
);
1266 static void task_ctx_sched_out(struct perf_event_context
*ctx
,
1267 enum event_type_t event_type
)
1269 struct perf_cpu_context
*cpuctx
= __get_cpu_context(ctx
);
1271 if (!cpuctx
->task_ctx
)
1274 if (WARN_ON_ONCE(ctx
!= cpuctx
->task_ctx
))
1277 ctx_sched_out(ctx
, cpuctx
, event_type
);
1278 cpuctx
->task_ctx
= NULL
;
1282 * Called with IRQs disabled
1284 static void __perf_event_task_sched_out(struct perf_event_context
*ctx
)
1286 task_ctx_sched_out(ctx
, EVENT_ALL
);
1290 * Called with IRQs disabled
1292 static void cpu_ctx_sched_out(struct perf_cpu_context
*cpuctx
,
1293 enum event_type_t event_type
)
1295 ctx_sched_out(&cpuctx
->ctx
, cpuctx
, event_type
);
1299 ctx_pinned_sched_in(struct perf_event_context
*ctx
,
1300 struct perf_cpu_context
*cpuctx
)
1302 struct perf_event
*event
;
1304 list_for_each_entry(event
, &ctx
->pinned_groups
, group_entry
) {
1305 if (event
->state
<= PERF_EVENT_STATE_OFF
)
1307 if (event
->cpu
!= -1 && event
->cpu
!= smp_processor_id())
1310 if (group_can_go_on(event
, cpuctx
, 1))
1311 group_sched_in(event
, cpuctx
, ctx
);
1314 * If this pinned group hasn't been scheduled,
1315 * put it in error state.
1317 if (event
->state
== PERF_EVENT_STATE_INACTIVE
) {
1318 update_group_times(event
);
1319 event
->state
= PERF_EVENT_STATE_ERROR
;
1325 ctx_flexible_sched_in(struct perf_event_context
*ctx
,
1326 struct perf_cpu_context
*cpuctx
)
1328 struct perf_event
*event
;
1331 list_for_each_entry(event
, &ctx
->flexible_groups
, group_entry
) {
1332 /* Ignore events in OFF or ERROR state */
1333 if (event
->state
<= PERF_EVENT_STATE_OFF
)
1336 * Listen to the 'cpu' scheduling filter constraint
1339 if (event
->cpu
!= -1 && event
->cpu
!= smp_processor_id())
1342 if (group_can_go_on(event
, cpuctx
, can_add_hw
)) {
1343 if (group_sched_in(event
, cpuctx
, ctx
))
1350 ctx_sched_in(struct perf_event_context
*ctx
,
1351 struct perf_cpu_context
*cpuctx
,
1352 enum event_type_t event_type
)
1354 raw_spin_lock(&ctx
->lock
);
1356 if (likely(!ctx
->nr_events
))
1359 ctx
->timestamp
= perf_clock();
1362 * First go through the list and put on any pinned groups
1363 * in order to give them the best chance of going on.
1365 if (event_type
& EVENT_PINNED
)
1366 ctx_pinned_sched_in(ctx
, cpuctx
);
1368 /* Then walk through the lower prio flexible groups */
1369 if (event_type
& EVENT_FLEXIBLE
)
1370 ctx_flexible_sched_in(ctx
, cpuctx
);
1373 raw_spin_unlock(&ctx
->lock
);
1376 static void cpu_ctx_sched_in(struct perf_cpu_context
*cpuctx
,
1377 enum event_type_t event_type
)
1379 struct perf_event_context
*ctx
= &cpuctx
->ctx
;
1381 ctx_sched_in(ctx
, cpuctx
, event_type
);
1384 static void task_ctx_sched_in(struct perf_event_context
*ctx
,
1385 enum event_type_t event_type
)
1387 struct perf_cpu_context
*cpuctx
;
1389 cpuctx
= __get_cpu_context(ctx
);
1390 if (cpuctx
->task_ctx
== ctx
)
1393 ctx_sched_in(ctx
, cpuctx
, event_type
);
1394 cpuctx
->task_ctx
= ctx
;
1397 void perf_event_context_sched_in(struct perf_event_context
*ctx
)
1399 struct perf_cpu_context
*cpuctx
;
1401 cpuctx
= __get_cpu_context(ctx
);
1402 if (cpuctx
->task_ctx
== ctx
)
1405 perf_pmu_disable(ctx
->pmu
);
1407 * We want to keep the following priority order:
1408 * cpu pinned (that don't need to move), task pinned,
1409 * cpu flexible, task flexible.
1411 cpu_ctx_sched_out(cpuctx
, EVENT_FLEXIBLE
);
1413 ctx_sched_in(ctx
, cpuctx
, EVENT_PINNED
);
1414 cpu_ctx_sched_in(cpuctx
, EVENT_FLEXIBLE
);
1415 ctx_sched_in(ctx
, cpuctx
, EVENT_FLEXIBLE
);
1417 cpuctx
->task_ctx
= ctx
;
1420 * Since these rotations are per-cpu, we need to ensure the
1421 * cpu-context we got scheduled on is actually rotating.
1423 perf_pmu_rotate_start(ctx
->pmu
);
1424 perf_pmu_enable(ctx
->pmu
);
1428 * Called from scheduler to add the events of the current task
1429 * with interrupts disabled.
1431 * We restore the event value and then enable it.
1433 * This does not protect us against NMI, but enable()
1434 * sets the enabled bit in the control field of event _before_
1435 * accessing the event control register. If a NMI hits, then it will
1436 * keep the event running.
1438 void perf_event_task_sched_in(struct task_struct
*task
)
1440 struct perf_event_context
*ctx
;
1443 for_each_task_context_nr(ctxn
) {
1444 ctx
= task
->perf_event_ctxp
[ctxn
];
1448 perf_event_context_sched_in(ctx
);
1452 #define MAX_INTERRUPTS (~0ULL)
1454 static void perf_log_throttle(struct perf_event
*event
, int enable
);
1456 static u64
perf_calculate_period(struct perf_event
*event
, u64 nsec
, u64 count
)
1458 u64 frequency
= event
->attr
.sample_freq
;
1459 u64 sec
= NSEC_PER_SEC
;
1460 u64 divisor
, dividend
;
1462 int count_fls
, nsec_fls
, frequency_fls
, sec_fls
;
1464 count_fls
= fls64(count
);
1465 nsec_fls
= fls64(nsec
);
1466 frequency_fls
= fls64(frequency
);
1470 * We got @count in @nsec, with a target of sample_freq HZ
1471 * the target period becomes:
1474 * period = -------------------
1475 * @nsec * sample_freq
1480 * Reduce accuracy by one bit such that @a and @b converge
1481 * to a similar magnitude.
1483 #define REDUCE_FLS(a, b) \
1485 if (a##_fls > b##_fls) { \
1495 * Reduce accuracy until either term fits in a u64, then proceed with
1496 * the other, so that finally we can do a u64/u64 division.
1498 while (count_fls
+ sec_fls
> 64 && nsec_fls
+ frequency_fls
> 64) {
1499 REDUCE_FLS(nsec
, frequency
);
1500 REDUCE_FLS(sec
, count
);
1503 if (count_fls
+ sec_fls
> 64) {
1504 divisor
= nsec
* frequency
;
1506 while (count_fls
+ sec_fls
> 64) {
1507 REDUCE_FLS(count
, sec
);
1511 dividend
= count
* sec
;
1513 dividend
= count
* sec
;
1515 while (nsec_fls
+ frequency_fls
> 64) {
1516 REDUCE_FLS(nsec
, frequency
);
1520 divisor
= nsec
* frequency
;
1526 return div64_u64(dividend
, divisor
);
1529 static void perf_adjust_period(struct perf_event
*event
, u64 nsec
, u64 count
)
1531 struct hw_perf_event
*hwc
= &event
->hw
;
1532 s64 period
, sample_period
;
1535 period
= perf_calculate_period(event
, nsec
, count
);
1537 delta
= (s64
)(period
- hwc
->sample_period
);
1538 delta
= (delta
+ 7) / 8; /* low pass filter */
1540 sample_period
= hwc
->sample_period
+ delta
;
1545 hwc
->sample_period
= sample_period
;
1547 if (local64_read(&hwc
->period_left
) > 8*sample_period
) {
1548 event
->pmu
->stop(event
, PERF_EF_UPDATE
);
1549 local64_set(&hwc
->period_left
, 0);
1550 event
->pmu
->start(event
, PERF_EF_RELOAD
);
1554 static void perf_ctx_adjust_freq(struct perf_event_context
*ctx
, u64 period
)
1556 struct perf_event
*event
;
1557 struct hw_perf_event
*hwc
;
1558 u64 interrupts
, now
;
1561 raw_spin_lock(&ctx
->lock
);
1562 list_for_each_entry_rcu(event
, &ctx
->event_list
, event_entry
) {
1563 if (event
->state
!= PERF_EVENT_STATE_ACTIVE
)
1566 if (event
->cpu
!= -1 && event
->cpu
!= smp_processor_id())
1571 interrupts
= hwc
->interrupts
;
1572 hwc
->interrupts
= 0;
1575 * unthrottle events on the tick
1577 if (interrupts
== MAX_INTERRUPTS
) {
1578 perf_log_throttle(event
, 1);
1579 event
->pmu
->start(event
, 0);
1582 if (!event
->attr
.freq
|| !event
->attr
.sample_freq
)
1585 event
->pmu
->read(event
);
1586 now
= local64_read(&event
->count
);
1587 delta
= now
- hwc
->freq_count_stamp
;
1588 hwc
->freq_count_stamp
= now
;
1591 perf_adjust_period(event
, period
, delta
);
1593 raw_spin_unlock(&ctx
->lock
);
1597 * Round-robin a context's events:
1599 static void rotate_ctx(struct perf_event_context
*ctx
)
1601 raw_spin_lock(&ctx
->lock
);
1603 /* Rotate the first entry last of non-pinned groups */
1604 list_rotate_left(&ctx
->flexible_groups
);
1606 raw_spin_unlock(&ctx
->lock
);
1610 * Cannot race with ->pmu_rotate_start() because this is ran from hardirq
1611 * context, and ->pmu_rotate_start() is called with irqs disabled (both are
1612 * cpu affine, so there are no SMP races).
1614 static enum hrtimer_restart
perf_event_context_tick(struct hrtimer
*timer
)
1616 enum hrtimer_restart restart
= HRTIMER_NORESTART
;
1617 struct perf_cpu_context
*cpuctx
;
1618 struct perf_event_context
*ctx
= NULL
;
1621 cpuctx
= container_of(timer
, struct perf_cpu_context
, timer
);
1623 if (cpuctx
->ctx
.nr_events
) {
1624 restart
= HRTIMER_RESTART
;
1625 if (cpuctx
->ctx
.nr_events
!= cpuctx
->ctx
.nr_active
)
1629 ctx
= cpuctx
->task_ctx
;
1630 if (ctx
&& ctx
->nr_events
) {
1631 restart
= HRTIMER_RESTART
;
1632 if (ctx
->nr_events
!= ctx
->nr_active
)
1636 perf_pmu_disable(cpuctx
->ctx
.pmu
);
1637 perf_ctx_adjust_freq(&cpuctx
->ctx
, cpuctx
->timer_interval
);
1639 perf_ctx_adjust_freq(ctx
, cpuctx
->timer_interval
);
1644 cpu_ctx_sched_out(cpuctx
, EVENT_FLEXIBLE
);
1646 task_ctx_sched_out(ctx
, EVENT_FLEXIBLE
);
1648 rotate_ctx(&cpuctx
->ctx
);
1652 cpu_ctx_sched_in(cpuctx
, EVENT_FLEXIBLE
);
1654 task_ctx_sched_in(ctx
, EVENT_FLEXIBLE
);
1657 perf_pmu_enable(cpuctx
->ctx
.pmu
);
1658 hrtimer_forward_now(timer
, ns_to_ktime(cpuctx
->timer_interval
));
1663 static int event_enable_on_exec(struct perf_event
*event
,
1664 struct perf_event_context
*ctx
)
1666 if (!event
->attr
.enable_on_exec
)
1669 event
->attr
.enable_on_exec
= 0;
1670 if (event
->state
>= PERF_EVENT_STATE_INACTIVE
)
1673 __perf_event_mark_enabled(event
, ctx
);
1679 * Enable all of a task's events that have been marked enable-on-exec.
1680 * This expects task == current.
1682 static void perf_event_enable_on_exec(struct perf_event_context
*ctx
)
1684 struct perf_event
*event
;
1685 unsigned long flags
;
1689 local_irq_save(flags
);
1690 if (!ctx
|| !ctx
->nr_events
)
1693 task_ctx_sched_out(ctx
, EVENT_ALL
);
1695 raw_spin_lock(&ctx
->lock
);
1697 list_for_each_entry(event
, &ctx
->pinned_groups
, group_entry
) {
1698 ret
= event_enable_on_exec(event
, ctx
);
1703 list_for_each_entry(event
, &ctx
->flexible_groups
, group_entry
) {
1704 ret
= event_enable_on_exec(event
, ctx
);
1710 * Unclone this context if we enabled any event.
1715 raw_spin_unlock(&ctx
->lock
);
1717 perf_event_context_sched_in(ctx
);
1719 local_irq_restore(flags
);
1723 * Cross CPU call to read the hardware event
1725 static void __perf_event_read(void *info
)
1727 struct perf_event
*event
= info
;
1728 struct perf_event_context
*ctx
= event
->ctx
;
1729 struct perf_cpu_context
*cpuctx
= __get_cpu_context(ctx
);
1732 * If this is a task context, we need to check whether it is
1733 * the current task context of this cpu. If not it has been
1734 * scheduled out before the smp call arrived. In that case
1735 * event->count would have been updated to a recent sample
1736 * when the event was scheduled out.
1738 if (ctx
->task
&& cpuctx
->task_ctx
!= ctx
)
1741 raw_spin_lock(&ctx
->lock
);
1742 update_context_time(ctx
);
1743 update_event_times(event
);
1744 raw_spin_unlock(&ctx
->lock
);
1746 event
->pmu
->read(event
);
1749 static inline u64
perf_event_count(struct perf_event
*event
)
1751 return local64_read(&event
->count
) + atomic64_read(&event
->child_count
);
1754 static u64
perf_event_read(struct perf_event
*event
)
1757 * If event is enabled and currently active on a CPU, update the
1758 * value in the event structure:
1760 if (event
->state
== PERF_EVENT_STATE_ACTIVE
) {
1761 smp_call_function_single(event
->oncpu
,
1762 __perf_event_read
, event
, 1);
1763 } else if (event
->state
== PERF_EVENT_STATE_INACTIVE
) {
1764 struct perf_event_context
*ctx
= event
->ctx
;
1765 unsigned long flags
;
1767 raw_spin_lock_irqsave(&ctx
->lock
, flags
);
1768 update_context_time(ctx
);
1769 update_event_times(event
);
1770 raw_spin_unlock_irqrestore(&ctx
->lock
, flags
);
1773 return perf_event_count(event
);
1780 struct callchain_cpus_entries
{
1781 struct rcu_head rcu_head
;
1782 struct perf_callchain_entry
*cpu_entries
[0];
1785 static DEFINE_PER_CPU(int, callchain_recursion
[PERF_NR_CONTEXTS
]);
1786 static atomic_t nr_callchain_events
;
1787 static DEFINE_MUTEX(callchain_mutex
);
1788 struct callchain_cpus_entries
*callchain_cpus_entries
;
1791 __weak
void perf_callchain_kernel(struct perf_callchain_entry
*entry
,
1792 struct pt_regs
*regs
)
1796 __weak
void perf_callchain_user(struct perf_callchain_entry
*entry
,
1797 struct pt_regs
*regs
)
1801 static void release_callchain_buffers_rcu(struct rcu_head
*head
)
1803 struct callchain_cpus_entries
*entries
;
1806 entries
= container_of(head
, struct callchain_cpus_entries
, rcu_head
);
1808 for_each_possible_cpu(cpu
)
1809 kfree(entries
->cpu_entries
[cpu
]);
1814 static void release_callchain_buffers(void)
1816 struct callchain_cpus_entries
*entries
;
1818 entries
= callchain_cpus_entries
;
1819 rcu_assign_pointer(callchain_cpus_entries
, NULL
);
1820 call_rcu(&entries
->rcu_head
, release_callchain_buffers_rcu
);
1823 static int alloc_callchain_buffers(void)
1827 struct callchain_cpus_entries
*entries
;
1830 * We can't use the percpu allocation API for data that can be
1831 * accessed from NMI. Use a temporary manual per cpu allocation
1832 * until that gets sorted out.
1834 size
= sizeof(*entries
) + sizeof(struct perf_callchain_entry
*) *
1835 num_possible_cpus();
1837 entries
= kzalloc(size
, GFP_KERNEL
);
1841 size
= sizeof(struct perf_callchain_entry
) * PERF_NR_CONTEXTS
;
1843 for_each_possible_cpu(cpu
) {
1844 entries
->cpu_entries
[cpu
] = kmalloc_node(size
, GFP_KERNEL
,
1846 if (!entries
->cpu_entries
[cpu
])
1850 rcu_assign_pointer(callchain_cpus_entries
, entries
);
1855 for_each_possible_cpu(cpu
)
1856 kfree(entries
->cpu_entries
[cpu
]);
1862 static int get_callchain_buffers(void)
1867 mutex_lock(&callchain_mutex
);
1869 count
= atomic_inc_return(&nr_callchain_events
);
1870 if (WARN_ON_ONCE(count
< 1)) {
1876 /* If the allocation failed, give up */
1877 if (!callchain_cpus_entries
)
1882 err
= alloc_callchain_buffers();
1884 release_callchain_buffers();
1886 mutex_unlock(&callchain_mutex
);
1891 static void put_callchain_buffers(void)
1893 if (atomic_dec_and_mutex_lock(&nr_callchain_events
, &callchain_mutex
)) {
1894 release_callchain_buffers();
1895 mutex_unlock(&callchain_mutex
);
1899 static int get_recursion_context(int *recursion
)
1907 else if (in_softirq())
1912 if (recursion
[rctx
])
1921 static inline void put_recursion_context(int *recursion
, int rctx
)
1927 static struct perf_callchain_entry
*get_callchain_entry(int *rctx
)
1930 struct callchain_cpus_entries
*entries
;
1932 *rctx
= get_recursion_context(__get_cpu_var(callchain_recursion
));
1936 entries
= rcu_dereference(callchain_cpus_entries
);
1940 cpu
= smp_processor_id();
1942 return &entries
->cpu_entries
[cpu
][*rctx
];
1946 put_callchain_entry(int rctx
)
1948 put_recursion_context(__get_cpu_var(callchain_recursion
), rctx
);
1951 static struct perf_callchain_entry
*perf_callchain(struct pt_regs
*regs
)
1954 struct perf_callchain_entry
*entry
;
1957 entry
= get_callchain_entry(&rctx
);
1966 if (!user_mode(regs
)) {
1967 perf_callchain_store(entry
, PERF_CONTEXT_KERNEL
);
1968 perf_callchain_kernel(entry
, regs
);
1970 regs
= task_pt_regs(current
);
1976 perf_callchain_store(entry
, PERF_CONTEXT_USER
);
1977 perf_callchain_user(entry
, regs
);
1981 put_callchain_entry(rctx
);
1987 * Initialize the perf_event context in a task_struct:
1989 static void __perf_event_init_context(struct perf_event_context
*ctx
)
1991 raw_spin_lock_init(&ctx
->lock
);
1992 mutex_init(&ctx
->mutex
);
1993 INIT_LIST_HEAD(&ctx
->pinned_groups
);
1994 INIT_LIST_HEAD(&ctx
->flexible_groups
);
1995 INIT_LIST_HEAD(&ctx
->event_list
);
1996 atomic_set(&ctx
->refcount
, 1);
1999 static struct perf_event_context
*
2000 alloc_perf_context(struct pmu
*pmu
, struct task_struct
*task
)
2002 struct perf_event_context
*ctx
;
2004 ctx
= kzalloc(sizeof(struct perf_event_context
), GFP_KERNEL
);
2008 __perf_event_init_context(ctx
);
2011 get_task_struct(task
);
2018 static struct perf_event_context
*
2019 find_get_context(struct pmu
*pmu
, pid_t pid
, int cpu
)
2021 struct perf_event_context
*ctx
;
2022 struct perf_cpu_context
*cpuctx
;
2023 struct task_struct
*task
;
2024 unsigned long flags
;
2027 if (pid
== -1 && cpu
!= -1) {
2028 /* Must be root to operate on a CPU event: */
2029 if (perf_paranoid_cpu() && !capable(CAP_SYS_ADMIN
))
2030 return ERR_PTR(-EACCES
);
2032 if (cpu
< 0 || cpu
>= nr_cpumask_bits
)
2033 return ERR_PTR(-EINVAL
);
2036 * We could be clever and allow to attach a event to an
2037 * offline CPU and activate it when the CPU comes up, but
2040 if (!cpu_online(cpu
))
2041 return ERR_PTR(-ENODEV
);
2043 cpuctx
= per_cpu_ptr(pmu
->pmu_cpu_context
, cpu
);
2054 task
= find_task_by_vpid(pid
);
2056 get_task_struct(task
);
2060 return ERR_PTR(-ESRCH
);
2063 * Can't attach events to a dying task.
2066 if (task
->flags
& PF_EXITING
)
2069 /* Reuse ptrace permission checks for now. */
2071 if (!ptrace_may_access(task
, PTRACE_MODE_READ
))
2075 ctxn
= pmu
->task_ctx_nr
;
2080 ctx
= perf_lock_task_context(task
, ctxn
, &flags
);
2083 raw_spin_unlock_irqrestore(&ctx
->lock
, flags
);
2087 ctx
= alloc_perf_context(pmu
, task
);
2094 if (cmpxchg(&task
->perf_event_ctxp
[ctxn
], NULL
, ctx
)) {
2096 * We raced with some other task; use
2097 * the context they set.
2099 put_task_struct(task
);
2105 put_task_struct(task
);
2109 put_task_struct(task
);
2110 return ERR_PTR(err
);
2113 static void perf_event_free_filter(struct perf_event
*event
);
2115 static void free_event_rcu(struct rcu_head
*head
)
2117 struct perf_event
*event
;
2119 event
= container_of(head
, struct perf_event
, rcu_head
);
2121 put_pid_ns(event
->ns
);
2122 perf_event_free_filter(event
);
2126 static void perf_pending_sync(struct perf_event
*event
);
2127 static void perf_buffer_put(struct perf_buffer
*buffer
);
2129 static void free_event(struct perf_event
*event
)
2131 perf_pending_sync(event
);
2133 if (!event
->parent
) {
2134 atomic_dec(&nr_events
);
2135 if (event
->attr
.mmap
|| event
->attr
.mmap_data
)
2136 atomic_dec(&nr_mmap_events
);
2137 if (event
->attr
.comm
)
2138 atomic_dec(&nr_comm_events
);
2139 if (event
->attr
.task
)
2140 atomic_dec(&nr_task_events
);
2141 if (event
->attr
.sample_type
& PERF_SAMPLE_CALLCHAIN
)
2142 put_callchain_buffers();
2145 if (event
->buffer
) {
2146 perf_buffer_put(event
->buffer
);
2147 event
->buffer
= NULL
;
2151 event
->destroy(event
);
2153 put_ctx(event
->ctx
);
2154 call_rcu(&event
->rcu_head
, free_event_rcu
);
2157 int perf_event_release_kernel(struct perf_event
*event
)
2159 struct perf_event_context
*ctx
= event
->ctx
;
2162 * Remove from the PMU, can't get re-enabled since we got
2163 * here because the last ref went.
2165 perf_event_disable(event
);
2167 WARN_ON_ONCE(ctx
->parent_ctx
);
2169 * There are two ways this annotation is useful:
2171 * 1) there is a lock recursion from perf_event_exit_task
2172 * see the comment there.
2174 * 2) there is a lock-inversion with mmap_sem through
2175 * perf_event_read_group(), which takes faults while
2176 * holding ctx->mutex, however this is called after
2177 * the last filedesc died, so there is no possibility
2178 * to trigger the AB-BA case.
2180 mutex_lock_nested(&ctx
->mutex
, SINGLE_DEPTH_NESTING
);
2181 raw_spin_lock_irq(&ctx
->lock
);
2182 perf_group_detach(event
);
2183 list_del_event(event
, ctx
);
2184 raw_spin_unlock_irq(&ctx
->lock
);
2185 mutex_unlock(&ctx
->mutex
);
2187 mutex_lock(&event
->owner
->perf_event_mutex
);
2188 list_del_init(&event
->owner_entry
);
2189 mutex_unlock(&event
->owner
->perf_event_mutex
);
2190 put_task_struct(event
->owner
);
2196 EXPORT_SYMBOL_GPL(perf_event_release_kernel
);
2199 * Called when the last reference to the file is gone.
2201 static int perf_release(struct inode
*inode
, struct file
*file
)
2203 struct perf_event
*event
= file
->private_data
;
2205 file
->private_data
= NULL
;
2207 return perf_event_release_kernel(event
);
2210 static int perf_event_read_size(struct perf_event
*event
)
2212 int entry
= sizeof(u64
); /* value */
2216 if (event
->attr
.read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
)
2217 size
+= sizeof(u64
);
2219 if (event
->attr
.read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
)
2220 size
+= sizeof(u64
);
2222 if (event
->attr
.read_format
& PERF_FORMAT_ID
)
2223 entry
+= sizeof(u64
);
2225 if (event
->attr
.read_format
& PERF_FORMAT_GROUP
) {
2226 nr
+= event
->group_leader
->nr_siblings
;
2227 size
+= sizeof(u64
);
2235 u64
perf_event_read_value(struct perf_event
*event
, u64
*enabled
, u64
*running
)
2237 struct perf_event
*child
;
2243 mutex_lock(&event
->child_mutex
);
2244 total
+= perf_event_read(event
);
2245 *enabled
+= event
->total_time_enabled
+
2246 atomic64_read(&event
->child_total_time_enabled
);
2247 *running
+= event
->total_time_running
+
2248 atomic64_read(&event
->child_total_time_running
);
2250 list_for_each_entry(child
, &event
->child_list
, child_list
) {
2251 total
+= perf_event_read(child
);
2252 *enabled
+= child
->total_time_enabled
;
2253 *running
+= child
->total_time_running
;
2255 mutex_unlock(&event
->child_mutex
);
2259 EXPORT_SYMBOL_GPL(perf_event_read_value
);
2261 static int perf_event_read_group(struct perf_event
*event
,
2262 u64 read_format
, char __user
*buf
)
2264 struct perf_event
*leader
= event
->group_leader
, *sub
;
2265 int n
= 0, size
= 0, ret
= -EFAULT
;
2266 struct perf_event_context
*ctx
= leader
->ctx
;
2268 u64 count
, enabled
, running
;
2270 mutex_lock(&ctx
->mutex
);
2271 count
= perf_event_read_value(leader
, &enabled
, &running
);
2273 values
[n
++] = 1 + leader
->nr_siblings
;
2274 if (read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
)
2275 values
[n
++] = enabled
;
2276 if (read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
)
2277 values
[n
++] = running
;
2278 values
[n
++] = count
;
2279 if (read_format
& PERF_FORMAT_ID
)
2280 values
[n
++] = primary_event_id(leader
);
2282 size
= n
* sizeof(u64
);
2284 if (copy_to_user(buf
, values
, size
))
2289 list_for_each_entry(sub
, &leader
->sibling_list
, group_entry
) {
2292 values
[n
++] = perf_event_read_value(sub
, &enabled
, &running
);
2293 if (read_format
& PERF_FORMAT_ID
)
2294 values
[n
++] = primary_event_id(sub
);
2296 size
= n
* sizeof(u64
);
2298 if (copy_to_user(buf
+ ret
, values
, size
)) {
2306 mutex_unlock(&ctx
->mutex
);
2311 static int perf_event_read_one(struct perf_event
*event
,
2312 u64 read_format
, char __user
*buf
)
2314 u64 enabled
, running
;
2318 values
[n
++] = perf_event_read_value(event
, &enabled
, &running
);
2319 if (read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
)
2320 values
[n
++] = enabled
;
2321 if (read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
)
2322 values
[n
++] = running
;
2323 if (read_format
& PERF_FORMAT_ID
)
2324 values
[n
++] = primary_event_id(event
);
2326 if (copy_to_user(buf
, values
, n
* sizeof(u64
)))
2329 return n
* sizeof(u64
);
2333 * Read the performance event - simple non blocking version for now
2336 perf_read_hw(struct perf_event
*event
, char __user
*buf
, size_t count
)
2338 u64 read_format
= event
->attr
.read_format
;
2342 * Return end-of-file for a read on a event that is in
2343 * error state (i.e. because it was pinned but it couldn't be
2344 * scheduled on to the CPU at some point).
2346 if (event
->state
== PERF_EVENT_STATE_ERROR
)
2349 if (count
< perf_event_read_size(event
))
2352 WARN_ON_ONCE(event
->ctx
->parent_ctx
);
2353 if (read_format
& PERF_FORMAT_GROUP
)
2354 ret
= perf_event_read_group(event
, read_format
, buf
);
2356 ret
= perf_event_read_one(event
, read_format
, buf
);
2362 perf_read(struct file
*file
, char __user
*buf
, size_t count
, loff_t
*ppos
)
2364 struct perf_event
*event
= file
->private_data
;
2366 return perf_read_hw(event
, buf
, count
);
2369 static unsigned int perf_poll(struct file
*file
, poll_table
*wait
)
2371 struct perf_event
*event
= file
->private_data
;
2372 struct perf_buffer
*buffer
;
2373 unsigned int events
= POLL_HUP
;
2376 buffer
= rcu_dereference(event
->buffer
);
2378 events
= atomic_xchg(&buffer
->poll
, 0);
2381 poll_wait(file
, &event
->waitq
, wait
);
2386 static void perf_event_reset(struct perf_event
*event
)
2388 (void)perf_event_read(event
);
2389 local64_set(&event
->count
, 0);
2390 perf_event_update_userpage(event
);
2394 * Holding the top-level event's child_mutex means that any
2395 * descendant process that has inherited this event will block
2396 * in sync_child_event if it goes to exit, thus satisfying the
2397 * task existence requirements of perf_event_enable/disable.
2399 static void perf_event_for_each_child(struct perf_event
*event
,
2400 void (*func
)(struct perf_event
*))
2402 struct perf_event
*child
;
2404 WARN_ON_ONCE(event
->ctx
->parent_ctx
);
2405 mutex_lock(&event
->child_mutex
);
2407 list_for_each_entry(child
, &event
->child_list
, child_list
)
2409 mutex_unlock(&event
->child_mutex
);
2412 static void perf_event_for_each(struct perf_event
*event
,
2413 void (*func
)(struct perf_event
*))
2415 struct perf_event_context
*ctx
= event
->ctx
;
2416 struct perf_event
*sibling
;
2418 WARN_ON_ONCE(ctx
->parent_ctx
);
2419 mutex_lock(&ctx
->mutex
);
2420 event
= event
->group_leader
;
2422 perf_event_for_each_child(event
, func
);
2424 list_for_each_entry(sibling
, &event
->sibling_list
, group_entry
)
2425 perf_event_for_each_child(event
, func
);
2426 mutex_unlock(&ctx
->mutex
);
2429 static int perf_event_period(struct perf_event
*event
, u64 __user
*arg
)
2431 struct perf_event_context
*ctx
= event
->ctx
;
2436 if (!event
->attr
.sample_period
)
2439 size
= copy_from_user(&value
, arg
, sizeof(value
));
2440 if (size
!= sizeof(value
))
2446 raw_spin_lock_irq(&ctx
->lock
);
2447 if (event
->attr
.freq
) {
2448 if (value
> sysctl_perf_event_sample_rate
) {
2453 event
->attr
.sample_freq
= value
;
2455 event
->attr
.sample_period
= value
;
2456 event
->hw
.sample_period
= value
;
2459 raw_spin_unlock_irq(&ctx
->lock
);
2464 static const struct file_operations perf_fops
;
2466 static struct perf_event
*perf_fget_light(int fd
, int *fput_needed
)
2470 file
= fget_light(fd
, fput_needed
);
2472 return ERR_PTR(-EBADF
);
2474 if (file
->f_op
!= &perf_fops
) {
2475 fput_light(file
, *fput_needed
);
2477 return ERR_PTR(-EBADF
);
2480 return file
->private_data
;
2483 static int perf_event_set_output(struct perf_event
*event
,
2484 struct perf_event
*output_event
);
2485 static int perf_event_set_filter(struct perf_event
*event
, void __user
*arg
);
2487 static long perf_ioctl(struct file
*file
, unsigned int cmd
, unsigned long arg
)
2489 struct perf_event
*event
= file
->private_data
;
2490 void (*func
)(struct perf_event
*);
2494 case PERF_EVENT_IOC_ENABLE
:
2495 func
= perf_event_enable
;
2497 case PERF_EVENT_IOC_DISABLE
:
2498 func
= perf_event_disable
;
2500 case PERF_EVENT_IOC_RESET
:
2501 func
= perf_event_reset
;
2504 case PERF_EVENT_IOC_REFRESH
:
2505 return perf_event_refresh(event
, arg
);
2507 case PERF_EVENT_IOC_PERIOD
:
2508 return perf_event_period(event
, (u64 __user
*)arg
);
2510 case PERF_EVENT_IOC_SET_OUTPUT
:
2512 struct perf_event
*output_event
= NULL
;
2513 int fput_needed
= 0;
2517 output_event
= perf_fget_light(arg
, &fput_needed
);
2518 if (IS_ERR(output_event
))
2519 return PTR_ERR(output_event
);
2522 ret
= perf_event_set_output(event
, output_event
);
2524 fput_light(output_event
->filp
, fput_needed
);
2529 case PERF_EVENT_IOC_SET_FILTER
:
2530 return perf_event_set_filter(event
, (void __user
*)arg
);
2536 if (flags
& PERF_IOC_FLAG_GROUP
)
2537 perf_event_for_each(event
, func
);
2539 perf_event_for_each_child(event
, func
);
2544 int perf_event_task_enable(void)
2546 struct perf_event
*event
;
2548 mutex_lock(¤t
->perf_event_mutex
);
2549 list_for_each_entry(event
, ¤t
->perf_event_list
, owner_entry
)
2550 perf_event_for_each_child(event
, perf_event_enable
);
2551 mutex_unlock(¤t
->perf_event_mutex
);
2556 int perf_event_task_disable(void)
2558 struct perf_event
*event
;
2560 mutex_lock(¤t
->perf_event_mutex
);
2561 list_for_each_entry(event
, ¤t
->perf_event_list
, owner_entry
)
2562 perf_event_for_each_child(event
, perf_event_disable
);
2563 mutex_unlock(¤t
->perf_event_mutex
);
2568 #ifndef PERF_EVENT_INDEX_OFFSET
2569 # define PERF_EVENT_INDEX_OFFSET 0
2572 static int perf_event_index(struct perf_event
*event
)
2574 if (event
->hw
.state
& PERF_HES_STOPPED
)
2577 if (event
->state
!= PERF_EVENT_STATE_ACTIVE
)
2580 return event
->hw
.idx
+ 1 - PERF_EVENT_INDEX_OFFSET
;
2584 * Callers need to ensure there can be no nesting of this function, otherwise
2585 * the seqlock logic goes bad. We can not serialize this because the arch
2586 * code calls this from NMI context.
2588 void perf_event_update_userpage(struct perf_event
*event
)
2590 struct perf_event_mmap_page
*userpg
;
2591 struct perf_buffer
*buffer
;
2594 buffer
= rcu_dereference(event
->buffer
);
2598 userpg
= buffer
->user_page
;
2601 * Disable preemption so as to not let the corresponding user-space
2602 * spin too long if we get preempted.
2607 userpg
->index
= perf_event_index(event
);
2608 userpg
->offset
= perf_event_count(event
);
2609 if (event
->state
== PERF_EVENT_STATE_ACTIVE
)
2610 userpg
->offset
-= local64_read(&event
->hw
.prev_count
);
2612 userpg
->time_enabled
= event
->total_time_enabled
+
2613 atomic64_read(&event
->child_total_time_enabled
);
2615 userpg
->time_running
= event
->total_time_running
+
2616 atomic64_read(&event
->child_total_time_running
);
2625 static unsigned long perf_data_size(struct perf_buffer
*buffer
);
2628 perf_buffer_init(struct perf_buffer
*buffer
, long watermark
, int flags
)
2630 long max_size
= perf_data_size(buffer
);
2633 buffer
->watermark
= min(max_size
, watermark
);
2635 if (!buffer
->watermark
)
2636 buffer
->watermark
= max_size
/ 2;
2638 if (flags
& PERF_BUFFER_WRITABLE
)
2639 buffer
->writable
= 1;
2641 atomic_set(&buffer
->refcount
, 1);
2644 #ifndef CONFIG_PERF_USE_VMALLOC
2647 * Back perf_mmap() with regular GFP_KERNEL-0 pages.
2650 static struct page
*
2651 perf_mmap_to_page(struct perf_buffer
*buffer
, unsigned long pgoff
)
2653 if (pgoff
> buffer
->nr_pages
)
2657 return virt_to_page(buffer
->user_page
);
2659 return virt_to_page(buffer
->data_pages
[pgoff
- 1]);
2662 static void *perf_mmap_alloc_page(int cpu
)
2667 node
= (cpu
== -1) ? cpu
: cpu_to_node(cpu
);
2668 page
= alloc_pages_node(node
, GFP_KERNEL
| __GFP_ZERO
, 0);
2672 return page_address(page
);
2675 static struct perf_buffer
*
2676 perf_buffer_alloc(int nr_pages
, long watermark
, int cpu
, int flags
)
2678 struct perf_buffer
*buffer
;
2682 size
= sizeof(struct perf_buffer
);
2683 size
+= nr_pages
* sizeof(void *);
2685 buffer
= kzalloc(size
, GFP_KERNEL
);
2689 buffer
->user_page
= perf_mmap_alloc_page(cpu
);
2690 if (!buffer
->user_page
)
2691 goto fail_user_page
;
2693 for (i
= 0; i
< nr_pages
; i
++) {
2694 buffer
->data_pages
[i
] = perf_mmap_alloc_page(cpu
);
2695 if (!buffer
->data_pages
[i
])
2696 goto fail_data_pages
;
2699 buffer
->nr_pages
= nr_pages
;
2701 perf_buffer_init(buffer
, watermark
, flags
);
2706 for (i
--; i
>= 0; i
--)
2707 free_page((unsigned long)buffer
->data_pages
[i
]);
2709 free_page((unsigned long)buffer
->user_page
);
2718 static void perf_mmap_free_page(unsigned long addr
)
2720 struct page
*page
= virt_to_page((void *)addr
);
2722 page
->mapping
= NULL
;
2726 static void perf_buffer_free(struct perf_buffer
*buffer
)
2730 perf_mmap_free_page((unsigned long)buffer
->user_page
);
2731 for (i
= 0; i
< buffer
->nr_pages
; i
++)
2732 perf_mmap_free_page((unsigned long)buffer
->data_pages
[i
]);
2736 static inline int page_order(struct perf_buffer
*buffer
)
2744 * Back perf_mmap() with vmalloc memory.
2746 * Required for architectures that have d-cache aliasing issues.
2749 static inline int page_order(struct perf_buffer
*buffer
)
2751 return buffer
->page_order
;
2754 static struct page
*
2755 perf_mmap_to_page(struct perf_buffer
*buffer
, unsigned long pgoff
)
2757 if (pgoff
> (1UL << page_order(buffer
)))
2760 return vmalloc_to_page((void *)buffer
->user_page
+ pgoff
* PAGE_SIZE
);
2763 static void perf_mmap_unmark_page(void *addr
)
2765 struct page
*page
= vmalloc_to_page(addr
);
2767 page
->mapping
= NULL
;
2770 static void perf_buffer_free_work(struct work_struct
*work
)
2772 struct perf_buffer
*buffer
;
2776 buffer
= container_of(work
, struct perf_buffer
, work
);
2777 nr
= 1 << page_order(buffer
);
2779 base
= buffer
->user_page
;
2780 for (i
= 0; i
< nr
+ 1; i
++)
2781 perf_mmap_unmark_page(base
+ (i
* PAGE_SIZE
));
2787 static void perf_buffer_free(struct perf_buffer
*buffer
)
2789 schedule_work(&buffer
->work
);
2792 static struct perf_buffer
*
2793 perf_buffer_alloc(int nr_pages
, long watermark
, int cpu
, int flags
)
2795 struct perf_buffer
*buffer
;
2799 size
= sizeof(struct perf_buffer
);
2800 size
+= sizeof(void *);
2802 buffer
= kzalloc(size
, GFP_KERNEL
);
2806 INIT_WORK(&buffer
->work
, perf_buffer_free_work
);
2808 all_buf
= vmalloc_user((nr_pages
+ 1) * PAGE_SIZE
);
2812 buffer
->user_page
= all_buf
;
2813 buffer
->data_pages
[0] = all_buf
+ PAGE_SIZE
;
2814 buffer
->page_order
= ilog2(nr_pages
);
2815 buffer
->nr_pages
= 1;
2817 perf_buffer_init(buffer
, watermark
, flags
);
2830 static unsigned long perf_data_size(struct perf_buffer
*buffer
)
2832 return buffer
->nr_pages
<< (PAGE_SHIFT
+ page_order(buffer
));
2835 static int perf_mmap_fault(struct vm_area_struct
*vma
, struct vm_fault
*vmf
)
2837 struct perf_event
*event
= vma
->vm_file
->private_data
;
2838 struct perf_buffer
*buffer
;
2839 int ret
= VM_FAULT_SIGBUS
;
2841 if (vmf
->flags
& FAULT_FLAG_MKWRITE
) {
2842 if (vmf
->pgoff
== 0)
2848 buffer
= rcu_dereference(event
->buffer
);
2852 if (vmf
->pgoff
&& (vmf
->flags
& FAULT_FLAG_WRITE
))
2855 vmf
->page
= perf_mmap_to_page(buffer
, vmf
->pgoff
);
2859 get_page(vmf
->page
);
2860 vmf
->page
->mapping
= vma
->vm_file
->f_mapping
;
2861 vmf
->page
->index
= vmf
->pgoff
;
2870 static void perf_buffer_free_rcu(struct rcu_head
*rcu_head
)
2872 struct perf_buffer
*buffer
;
2874 buffer
= container_of(rcu_head
, struct perf_buffer
, rcu_head
);
2875 perf_buffer_free(buffer
);
2878 static struct perf_buffer
*perf_buffer_get(struct perf_event
*event
)
2880 struct perf_buffer
*buffer
;
2883 buffer
= rcu_dereference(event
->buffer
);
2885 if (!atomic_inc_not_zero(&buffer
->refcount
))
2893 static void perf_buffer_put(struct perf_buffer
*buffer
)
2895 if (!atomic_dec_and_test(&buffer
->refcount
))
2898 call_rcu(&buffer
->rcu_head
, perf_buffer_free_rcu
);
2901 static void perf_mmap_open(struct vm_area_struct
*vma
)
2903 struct perf_event
*event
= vma
->vm_file
->private_data
;
2905 atomic_inc(&event
->mmap_count
);
2908 static void perf_mmap_close(struct vm_area_struct
*vma
)
2910 struct perf_event
*event
= vma
->vm_file
->private_data
;
2912 if (atomic_dec_and_mutex_lock(&event
->mmap_count
, &event
->mmap_mutex
)) {
2913 unsigned long size
= perf_data_size(event
->buffer
);
2914 struct user_struct
*user
= event
->mmap_user
;
2915 struct perf_buffer
*buffer
= event
->buffer
;
2917 atomic_long_sub((size
>> PAGE_SHIFT
) + 1, &user
->locked_vm
);
2918 vma
->vm_mm
->locked_vm
-= event
->mmap_locked
;
2919 rcu_assign_pointer(event
->buffer
, NULL
);
2920 mutex_unlock(&event
->mmap_mutex
);
2922 perf_buffer_put(buffer
);
2927 static const struct vm_operations_struct perf_mmap_vmops
= {
2928 .open
= perf_mmap_open
,
2929 .close
= perf_mmap_close
,
2930 .fault
= perf_mmap_fault
,
2931 .page_mkwrite
= perf_mmap_fault
,
2934 static int perf_mmap(struct file
*file
, struct vm_area_struct
*vma
)
2936 struct perf_event
*event
= file
->private_data
;
2937 unsigned long user_locked
, user_lock_limit
;
2938 struct user_struct
*user
= current_user();
2939 unsigned long locked
, lock_limit
;
2940 struct perf_buffer
*buffer
;
2941 unsigned long vma_size
;
2942 unsigned long nr_pages
;
2943 long user_extra
, extra
;
2944 int ret
= 0, flags
= 0;
2947 * Don't allow mmap() of inherited per-task counters. This would
2948 * create a performance issue due to all children writing to the
2951 if (event
->cpu
== -1 && event
->attr
.inherit
)
2954 if (!(vma
->vm_flags
& VM_SHARED
))
2957 vma_size
= vma
->vm_end
- vma
->vm_start
;
2958 nr_pages
= (vma_size
/ PAGE_SIZE
) - 1;
2961 * If we have buffer pages ensure they're a power-of-two number, so we
2962 * can do bitmasks instead of modulo.
2964 if (nr_pages
!= 0 && !is_power_of_2(nr_pages
))
2967 if (vma_size
!= PAGE_SIZE
* (1 + nr_pages
))
2970 if (vma
->vm_pgoff
!= 0)
2973 WARN_ON_ONCE(event
->ctx
->parent_ctx
);
2974 mutex_lock(&event
->mmap_mutex
);
2975 if (event
->buffer
) {
2976 if (event
->buffer
->nr_pages
== nr_pages
)
2977 atomic_inc(&event
->buffer
->refcount
);
2983 user_extra
= nr_pages
+ 1;
2984 user_lock_limit
= sysctl_perf_event_mlock
>> (PAGE_SHIFT
- 10);
2987 * Increase the limit linearly with more CPUs:
2989 user_lock_limit
*= num_online_cpus();
2991 user_locked
= atomic_long_read(&user
->locked_vm
) + user_extra
;
2994 if (user_locked
> user_lock_limit
)
2995 extra
= user_locked
- user_lock_limit
;
2997 lock_limit
= rlimit(RLIMIT_MEMLOCK
);
2998 lock_limit
>>= PAGE_SHIFT
;
2999 locked
= vma
->vm_mm
->locked_vm
+ extra
;
3001 if ((locked
> lock_limit
) && perf_paranoid_tracepoint_raw() &&
3002 !capable(CAP_IPC_LOCK
)) {
3007 WARN_ON(event
->buffer
);
3009 if (vma
->vm_flags
& VM_WRITE
)
3010 flags
|= PERF_BUFFER_WRITABLE
;
3012 buffer
= perf_buffer_alloc(nr_pages
, event
->attr
.wakeup_watermark
,
3018 rcu_assign_pointer(event
->buffer
, buffer
);
3020 atomic_long_add(user_extra
, &user
->locked_vm
);
3021 event
->mmap_locked
= extra
;
3022 event
->mmap_user
= get_current_user();
3023 vma
->vm_mm
->locked_vm
+= event
->mmap_locked
;
3027 atomic_inc(&event
->mmap_count
);
3028 mutex_unlock(&event
->mmap_mutex
);
3030 vma
->vm_flags
|= VM_RESERVED
;
3031 vma
->vm_ops
= &perf_mmap_vmops
;
3036 static int perf_fasync(int fd
, struct file
*filp
, int on
)
3038 struct inode
*inode
= filp
->f_path
.dentry
->d_inode
;
3039 struct perf_event
*event
= filp
->private_data
;
3042 mutex_lock(&inode
->i_mutex
);
3043 retval
= fasync_helper(fd
, filp
, on
, &event
->fasync
);
3044 mutex_unlock(&inode
->i_mutex
);
3052 static const struct file_operations perf_fops
= {
3053 .llseek
= no_llseek
,
3054 .release
= perf_release
,
3057 .unlocked_ioctl
= perf_ioctl
,
3058 .compat_ioctl
= perf_ioctl
,
3060 .fasync
= perf_fasync
,
3066 * If there's data, ensure we set the poll() state and publish everything
3067 * to user-space before waking everybody up.
3070 void perf_event_wakeup(struct perf_event
*event
)
3072 wake_up_all(&event
->waitq
);
3074 if (event
->pending_kill
) {
3075 kill_fasync(&event
->fasync
, SIGIO
, event
->pending_kill
);
3076 event
->pending_kill
= 0;
3083 * Handle the case where we need to wakeup up from NMI (or rq->lock) context.
3085 * The NMI bit means we cannot possibly take locks. Therefore, maintain a
3086 * single linked list and use cmpxchg() to add entries lockless.
3089 static void perf_pending_event(struct perf_pending_entry
*entry
)
3091 struct perf_event
*event
= container_of(entry
,
3092 struct perf_event
, pending
);
3094 if (event
->pending_disable
) {
3095 event
->pending_disable
= 0;
3096 __perf_event_disable(event
);
3099 if (event
->pending_wakeup
) {
3100 event
->pending_wakeup
= 0;
3101 perf_event_wakeup(event
);
3105 #define PENDING_TAIL ((struct perf_pending_entry *)-1UL)
3107 static DEFINE_PER_CPU(struct perf_pending_entry
*, perf_pending_head
) = {
3111 static void perf_pending_queue(struct perf_pending_entry
*entry
,
3112 void (*func
)(struct perf_pending_entry
*))
3114 struct perf_pending_entry
**head
;
3116 if (cmpxchg(&entry
->next
, NULL
, PENDING_TAIL
) != NULL
)
3121 head
= &get_cpu_var(perf_pending_head
);
3124 entry
->next
= *head
;
3125 } while (cmpxchg(head
, entry
->next
, entry
) != entry
->next
);
3127 set_perf_event_pending();
3129 put_cpu_var(perf_pending_head
);
3132 static int __perf_pending_run(void)
3134 struct perf_pending_entry
*list
;
3137 list
= xchg(&__get_cpu_var(perf_pending_head
), PENDING_TAIL
);
3138 while (list
!= PENDING_TAIL
) {
3139 void (*func
)(struct perf_pending_entry
*);
3140 struct perf_pending_entry
*entry
= list
;
3147 * Ensure we observe the unqueue before we issue the wakeup,
3148 * so that we won't be waiting forever.
3149 * -- see perf_not_pending().
3160 static inline int perf_not_pending(struct perf_event
*event
)
3163 * If we flush on whatever cpu we run, there is a chance we don't
3167 __perf_pending_run();
3171 * Ensure we see the proper queue state before going to sleep
3172 * so that we do not miss the wakeup. -- see perf_pending_handle()
3175 return event
->pending
.next
== NULL
;
3178 static void perf_pending_sync(struct perf_event
*event
)
3180 wait_event(event
->waitq
, perf_not_pending(event
));
3183 void perf_event_do_pending(void)
3185 __perf_pending_run();
3189 * We assume there is only KVM supporting the callbacks.
3190 * Later on, we might change it to a list if there is
3191 * another virtualization implementation supporting the callbacks.
3193 struct perf_guest_info_callbacks
*perf_guest_cbs
;
3195 int perf_register_guest_info_callbacks(struct perf_guest_info_callbacks
*cbs
)
3197 perf_guest_cbs
= cbs
;
3200 EXPORT_SYMBOL_GPL(perf_register_guest_info_callbacks
);
3202 int perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks
*cbs
)
3204 perf_guest_cbs
= NULL
;
3207 EXPORT_SYMBOL_GPL(perf_unregister_guest_info_callbacks
);
3212 static bool perf_output_space(struct perf_buffer
*buffer
, unsigned long tail
,
3213 unsigned long offset
, unsigned long head
)
3217 if (!buffer
->writable
)
3220 mask
= perf_data_size(buffer
) - 1;
3222 offset
= (offset
- tail
) & mask
;
3223 head
= (head
- tail
) & mask
;
3225 if ((int)(head
- offset
) < 0)
3231 static void perf_output_wakeup(struct perf_output_handle
*handle
)
3233 atomic_set(&handle
->buffer
->poll
, POLL_IN
);
3236 handle
->event
->pending_wakeup
= 1;
3237 perf_pending_queue(&handle
->event
->pending
,
3238 perf_pending_event
);
3240 perf_event_wakeup(handle
->event
);
3244 * We need to ensure a later event_id doesn't publish a head when a former
3245 * event isn't done writing. However since we need to deal with NMIs we
3246 * cannot fully serialize things.
3248 * We only publish the head (and generate a wakeup) when the outer-most
3251 static void perf_output_get_handle(struct perf_output_handle
*handle
)
3253 struct perf_buffer
*buffer
= handle
->buffer
;
3256 local_inc(&buffer
->nest
);
3257 handle
->wakeup
= local_read(&buffer
->wakeup
);
3260 static void perf_output_put_handle(struct perf_output_handle
*handle
)
3262 struct perf_buffer
*buffer
= handle
->buffer
;
3266 head
= local_read(&buffer
->head
);
3269 * IRQ/NMI can happen here, which means we can miss a head update.
3272 if (!local_dec_and_test(&buffer
->nest
))
3276 * Publish the known good head. Rely on the full barrier implied
3277 * by atomic_dec_and_test() order the buffer->head read and this
3280 buffer
->user_page
->data_head
= head
;
3283 * Now check if we missed an update, rely on the (compiler)
3284 * barrier in atomic_dec_and_test() to re-read buffer->head.
3286 if (unlikely(head
!= local_read(&buffer
->head
))) {
3287 local_inc(&buffer
->nest
);
3291 if (handle
->wakeup
!= local_read(&buffer
->wakeup
))
3292 perf_output_wakeup(handle
);
3298 __always_inline
void perf_output_copy(struct perf_output_handle
*handle
,
3299 const void *buf
, unsigned int len
)
3302 unsigned long size
= min_t(unsigned long, handle
->size
, len
);
3304 memcpy(handle
->addr
, buf
, size
);
3307 handle
->addr
+= size
;
3309 handle
->size
-= size
;
3310 if (!handle
->size
) {
3311 struct perf_buffer
*buffer
= handle
->buffer
;
3314 handle
->page
&= buffer
->nr_pages
- 1;
3315 handle
->addr
= buffer
->data_pages
[handle
->page
];
3316 handle
->size
= PAGE_SIZE
<< page_order(buffer
);
3321 int perf_output_begin(struct perf_output_handle
*handle
,
3322 struct perf_event
*event
, unsigned int size
,
3323 int nmi
, int sample
)
3325 struct perf_buffer
*buffer
;
3326 unsigned long tail
, offset
, head
;
3329 struct perf_event_header header
;
3336 * For inherited events we send all the output towards the parent.
3339 event
= event
->parent
;
3341 buffer
= rcu_dereference(event
->buffer
);
3345 handle
->buffer
= buffer
;
3346 handle
->event
= event
;
3348 handle
->sample
= sample
;
3350 if (!buffer
->nr_pages
)
3353 have_lost
= local_read(&buffer
->lost
);
3355 size
+= sizeof(lost_event
);
3357 perf_output_get_handle(handle
);
3361 * Userspace could choose to issue a mb() before updating the
3362 * tail pointer. So that all reads will be completed before the
3365 tail
= ACCESS_ONCE(buffer
->user_page
->data_tail
);
3367 offset
= head
= local_read(&buffer
->head
);
3369 if (unlikely(!perf_output_space(buffer
, tail
, offset
, head
)))
3371 } while (local_cmpxchg(&buffer
->head
, offset
, head
) != offset
);
3373 if (head
- local_read(&buffer
->wakeup
) > buffer
->watermark
)
3374 local_add(buffer
->watermark
, &buffer
->wakeup
);
3376 handle
->page
= offset
>> (PAGE_SHIFT
+ page_order(buffer
));
3377 handle
->page
&= buffer
->nr_pages
- 1;
3378 handle
->size
= offset
& ((PAGE_SIZE
<< page_order(buffer
)) - 1);
3379 handle
->addr
= buffer
->data_pages
[handle
->page
];
3380 handle
->addr
+= handle
->size
;
3381 handle
->size
= (PAGE_SIZE
<< page_order(buffer
)) - handle
->size
;
3384 lost_event
.header
.type
= PERF_RECORD_LOST
;
3385 lost_event
.header
.misc
= 0;
3386 lost_event
.header
.size
= sizeof(lost_event
);
3387 lost_event
.id
= event
->id
;
3388 lost_event
.lost
= local_xchg(&buffer
->lost
, 0);
3390 perf_output_put(handle
, lost_event
);
3396 local_inc(&buffer
->lost
);
3397 perf_output_put_handle(handle
);
3404 void perf_output_end(struct perf_output_handle
*handle
)
3406 struct perf_event
*event
= handle
->event
;
3407 struct perf_buffer
*buffer
= handle
->buffer
;
3409 int wakeup_events
= event
->attr
.wakeup_events
;
3411 if (handle
->sample
&& wakeup_events
) {
3412 int events
= local_inc_return(&buffer
->events
);
3413 if (events
>= wakeup_events
) {
3414 local_sub(wakeup_events
, &buffer
->events
);
3415 local_inc(&buffer
->wakeup
);
3419 perf_output_put_handle(handle
);
3423 static u32
perf_event_pid(struct perf_event
*event
, struct task_struct
*p
)
3426 * only top level events have the pid namespace they were created in
3429 event
= event
->parent
;
3431 return task_tgid_nr_ns(p
, event
->ns
);
3434 static u32
perf_event_tid(struct perf_event
*event
, struct task_struct
*p
)
3437 * only top level events have the pid namespace they were created in
3440 event
= event
->parent
;
3442 return task_pid_nr_ns(p
, event
->ns
);
3445 static void perf_output_read_one(struct perf_output_handle
*handle
,
3446 struct perf_event
*event
)
3448 u64 read_format
= event
->attr
.read_format
;
3452 values
[n
++] = perf_event_count(event
);
3453 if (read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
) {
3454 values
[n
++] = event
->total_time_enabled
+
3455 atomic64_read(&event
->child_total_time_enabled
);
3457 if (read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
) {
3458 values
[n
++] = event
->total_time_running
+
3459 atomic64_read(&event
->child_total_time_running
);
3461 if (read_format
& PERF_FORMAT_ID
)
3462 values
[n
++] = primary_event_id(event
);
3464 perf_output_copy(handle
, values
, n
* sizeof(u64
));
3468 * XXX PERF_FORMAT_GROUP vs inherited events seems difficult.
3470 static void perf_output_read_group(struct perf_output_handle
*handle
,
3471 struct perf_event
*event
)
3473 struct perf_event
*leader
= event
->group_leader
, *sub
;
3474 u64 read_format
= event
->attr
.read_format
;
3478 values
[n
++] = 1 + leader
->nr_siblings
;
3480 if (read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
)
3481 values
[n
++] = leader
->total_time_enabled
;
3483 if (read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
)
3484 values
[n
++] = leader
->total_time_running
;
3486 if (leader
!= event
)
3487 leader
->pmu
->read(leader
);
3489 values
[n
++] = perf_event_count(leader
);
3490 if (read_format
& PERF_FORMAT_ID
)
3491 values
[n
++] = primary_event_id(leader
);
3493 perf_output_copy(handle
, values
, n
* sizeof(u64
));
3495 list_for_each_entry(sub
, &leader
->sibling_list
, group_entry
) {
3499 sub
->pmu
->read(sub
);
3501 values
[n
++] = perf_event_count(sub
);
3502 if (read_format
& PERF_FORMAT_ID
)
3503 values
[n
++] = primary_event_id(sub
);
3505 perf_output_copy(handle
, values
, n
* sizeof(u64
));
3509 static void perf_output_read(struct perf_output_handle
*handle
,
3510 struct perf_event
*event
)
3512 if (event
->attr
.read_format
& PERF_FORMAT_GROUP
)
3513 perf_output_read_group(handle
, event
);
3515 perf_output_read_one(handle
, event
);
3518 void perf_output_sample(struct perf_output_handle
*handle
,
3519 struct perf_event_header
*header
,
3520 struct perf_sample_data
*data
,
3521 struct perf_event
*event
)
3523 u64 sample_type
= data
->type
;
3525 perf_output_put(handle
, *header
);
3527 if (sample_type
& PERF_SAMPLE_IP
)
3528 perf_output_put(handle
, data
->ip
);
3530 if (sample_type
& PERF_SAMPLE_TID
)
3531 perf_output_put(handle
, data
->tid_entry
);
3533 if (sample_type
& PERF_SAMPLE_TIME
)
3534 perf_output_put(handle
, data
->time
);
3536 if (sample_type
& PERF_SAMPLE_ADDR
)
3537 perf_output_put(handle
, data
->addr
);
3539 if (sample_type
& PERF_SAMPLE_ID
)
3540 perf_output_put(handle
, data
->id
);
3542 if (sample_type
& PERF_SAMPLE_STREAM_ID
)
3543 perf_output_put(handle
, data
->stream_id
);
3545 if (sample_type
& PERF_SAMPLE_CPU
)
3546 perf_output_put(handle
, data
->cpu_entry
);
3548 if (sample_type
& PERF_SAMPLE_PERIOD
)
3549 perf_output_put(handle
, data
->period
);
3551 if (sample_type
& PERF_SAMPLE_READ
)
3552 perf_output_read(handle
, event
);
3554 if (sample_type
& PERF_SAMPLE_CALLCHAIN
) {
3555 if (data
->callchain
) {
3558 if (data
->callchain
)
3559 size
+= data
->callchain
->nr
;
3561 size
*= sizeof(u64
);
3563 perf_output_copy(handle
, data
->callchain
, size
);
3566 perf_output_put(handle
, nr
);
3570 if (sample_type
& PERF_SAMPLE_RAW
) {
3572 perf_output_put(handle
, data
->raw
->size
);
3573 perf_output_copy(handle
, data
->raw
->data
,
3580 .size
= sizeof(u32
),
3583 perf_output_put(handle
, raw
);
3588 void perf_prepare_sample(struct perf_event_header
*header
,
3589 struct perf_sample_data
*data
,
3590 struct perf_event
*event
,
3591 struct pt_regs
*regs
)
3593 u64 sample_type
= event
->attr
.sample_type
;
3595 data
->type
= sample_type
;
3597 header
->type
= PERF_RECORD_SAMPLE
;
3598 header
->size
= sizeof(*header
);
3601 header
->misc
|= perf_misc_flags(regs
);
3603 if (sample_type
& PERF_SAMPLE_IP
) {
3604 data
->ip
= perf_instruction_pointer(regs
);
3606 header
->size
+= sizeof(data
->ip
);
3609 if (sample_type
& PERF_SAMPLE_TID
) {
3610 /* namespace issues */
3611 data
->tid_entry
.pid
= perf_event_pid(event
, current
);
3612 data
->tid_entry
.tid
= perf_event_tid(event
, current
);
3614 header
->size
+= sizeof(data
->tid_entry
);
3617 if (sample_type
& PERF_SAMPLE_TIME
) {
3618 data
->time
= perf_clock();
3620 header
->size
+= sizeof(data
->time
);
3623 if (sample_type
& PERF_SAMPLE_ADDR
)
3624 header
->size
+= sizeof(data
->addr
);
3626 if (sample_type
& PERF_SAMPLE_ID
) {
3627 data
->id
= primary_event_id(event
);
3629 header
->size
+= sizeof(data
->id
);
3632 if (sample_type
& PERF_SAMPLE_STREAM_ID
) {
3633 data
->stream_id
= event
->id
;
3635 header
->size
+= sizeof(data
->stream_id
);
3638 if (sample_type
& PERF_SAMPLE_CPU
) {
3639 data
->cpu_entry
.cpu
= raw_smp_processor_id();
3640 data
->cpu_entry
.reserved
= 0;
3642 header
->size
+= sizeof(data
->cpu_entry
);
3645 if (sample_type
& PERF_SAMPLE_PERIOD
)
3646 header
->size
+= sizeof(data
->period
);
3648 if (sample_type
& PERF_SAMPLE_READ
)
3649 header
->size
+= perf_event_read_size(event
);
3651 if (sample_type
& PERF_SAMPLE_CALLCHAIN
) {
3654 data
->callchain
= perf_callchain(regs
);
3656 if (data
->callchain
)
3657 size
+= data
->callchain
->nr
;
3659 header
->size
+= size
* sizeof(u64
);
3662 if (sample_type
& PERF_SAMPLE_RAW
) {
3663 int size
= sizeof(u32
);
3666 size
+= data
->raw
->size
;
3668 size
+= sizeof(u32
);
3670 WARN_ON_ONCE(size
& (sizeof(u64
)-1));
3671 header
->size
+= size
;
3675 static void perf_event_output(struct perf_event
*event
, int nmi
,
3676 struct perf_sample_data
*data
,
3677 struct pt_regs
*regs
)
3679 struct perf_output_handle handle
;
3680 struct perf_event_header header
;
3682 /* protect the callchain buffers */
3685 perf_prepare_sample(&header
, data
, event
, regs
);
3687 if (perf_output_begin(&handle
, event
, header
.size
, nmi
, 1))
3690 perf_output_sample(&handle
, &header
, data
, event
);
3692 perf_output_end(&handle
);
3702 struct perf_read_event
{
3703 struct perf_event_header header
;
3710 perf_event_read_event(struct perf_event
*event
,
3711 struct task_struct
*task
)
3713 struct perf_output_handle handle
;
3714 struct perf_read_event read_event
= {
3716 .type
= PERF_RECORD_READ
,
3718 .size
= sizeof(read_event
) + perf_event_read_size(event
),
3720 .pid
= perf_event_pid(event
, task
),
3721 .tid
= perf_event_tid(event
, task
),
3725 ret
= perf_output_begin(&handle
, event
, read_event
.header
.size
, 0, 0);
3729 perf_output_put(&handle
, read_event
);
3730 perf_output_read(&handle
, event
);
3732 perf_output_end(&handle
);
3736 * task tracking -- fork/exit
3738 * enabled by: attr.comm | attr.mmap | attr.mmap_data | attr.task
3741 struct perf_task_event
{
3742 struct task_struct
*task
;
3743 struct perf_event_context
*task_ctx
;
3746 struct perf_event_header header
;
3756 static void perf_event_task_output(struct perf_event
*event
,
3757 struct perf_task_event
*task_event
)
3759 struct perf_output_handle handle
;
3760 struct task_struct
*task
= task_event
->task
;
3763 size
= task_event
->event_id
.header
.size
;
3764 ret
= perf_output_begin(&handle
, event
, size
, 0, 0);
3769 task_event
->event_id
.pid
= perf_event_pid(event
, task
);
3770 task_event
->event_id
.ppid
= perf_event_pid(event
, current
);
3772 task_event
->event_id
.tid
= perf_event_tid(event
, task
);
3773 task_event
->event_id
.ptid
= perf_event_tid(event
, current
);
3775 perf_output_put(&handle
, task_event
->event_id
);
3777 perf_output_end(&handle
);
3780 static int perf_event_task_match(struct perf_event
*event
)
3782 if (event
->state
< PERF_EVENT_STATE_INACTIVE
)
3785 if (event
->cpu
!= -1 && event
->cpu
!= smp_processor_id())
3788 if (event
->attr
.comm
|| event
->attr
.mmap
||
3789 event
->attr
.mmap_data
|| event
->attr
.task
)
3795 static void perf_event_task_ctx(struct perf_event_context
*ctx
,
3796 struct perf_task_event
*task_event
)
3798 struct perf_event
*event
;
3800 list_for_each_entry_rcu(event
, &ctx
->event_list
, event_entry
) {
3801 if (perf_event_task_match(event
))
3802 perf_event_task_output(event
, task_event
);
3806 static void perf_event_task_event(struct perf_task_event
*task_event
)
3808 struct perf_cpu_context
*cpuctx
;
3809 struct perf_event_context
*ctx
;
3813 rcu_read_lock_sched();
3814 list_for_each_entry_rcu(pmu
, &pmus
, entry
) {
3815 cpuctx
= this_cpu_ptr(pmu
->pmu_cpu_context
);
3816 perf_event_task_ctx(&cpuctx
->ctx
, task_event
);
3818 ctx
= task_event
->task_ctx
;
3820 ctxn
= pmu
->task_ctx_nr
;
3823 ctx
= rcu_dereference(current
->perf_event_ctxp
[ctxn
]);
3826 perf_event_task_ctx(ctx
, task_event
);
3828 rcu_read_unlock_sched();
3831 static void perf_event_task(struct task_struct
*task
,
3832 struct perf_event_context
*task_ctx
,
3835 struct perf_task_event task_event
;
3837 if (!atomic_read(&nr_comm_events
) &&
3838 !atomic_read(&nr_mmap_events
) &&
3839 !atomic_read(&nr_task_events
))
3842 task_event
= (struct perf_task_event
){
3844 .task_ctx
= task_ctx
,
3847 .type
= new ? PERF_RECORD_FORK
: PERF_RECORD_EXIT
,
3849 .size
= sizeof(task_event
.event_id
),
3855 .time
= perf_clock(),
3859 perf_event_task_event(&task_event
);
3862 void perf_event_fork(struct task_struct
*task
)
3864 perf_event_task(task
, NULL
, 1);
3871 struct perf_comm_event
{
3872 struct task_struct
*task
;
3877 struct perf_event_header header
;
3884 static void perf_event_comm_output(struct perf_event
*event
,
3885 struct perf_comm_event
*comm_event
)
3887 struct perf_output_handle handle
;
3888 int size
= comm_event
->event_id
.header
.size
;
3889 int ret
= perf_output_begin(&handle
, event
, size
, 0, 0);
3894 comm_event
->event_id
.pid
= perf_event_pid(event
, comm_event
->task
);
3895 comm_event
->event_id
.tid
= perf_event_tid(event
, comm_event
->task
);
3897 perf_output_put(&handle
, comm_event
->event_id
);
3898 perf_output_copy(&handle
, comm_event
->comm
,
3899 comm_event
->comm_size
);
3900 perf_output_end(&handle
);
3903 static int perf_event_comm_match(struct perf_event
*event
)
3905 if (event
->state
< PERF_EVENT_STATE_INACTIVE
)
3908 if (event
->cpu
!= -1 && event
->cpu
!= smp_processor_id())
3911 if (event
->attr
.comm
)
3917 static void perf_event_comm_ctx(struct perf_event_context
*ctx
,
3918 struct perf_comm_event
*comm_event
)
3920 struct perf_event
*event
;
3922 list_for_each_entry_rcu(event
, &ctx
->event_list
, event_entry
) {
3923 if (perf_event_comm_match(event
))
3924 perf_event_comm_output(event
, comm_event
);
3928 static void perf_event_comm_event(struct perf_comm_event
*comm_event
)
3930 struct perf_cpu_context
*cpuctx
;
3931 struct perf_event_context
*ctx
;
3932 char comm
[TASK_COMM_LEN
];
3937 memset(comm
, 0, sizeof(comm
));
3938 strlcpy(comm
, comm_event
->task
->comm
, sizeof(comm
));
3939 size
= ALIGN(strlen(comm
)+1, sizeof(u64
));
3941 comm_event
->comm
= comm
;
3942 comm_event
->comm_size
= size
;
3944 comm_event
->event_id
.header
.size
= sizeof(comm_event
->event_id
) + size
;
3946 rcu_read_lock_sched();
3947 list_for_each_entry_rcu(pmu
, &pmus
, entry
) {
3948 cpuctx
= this_cpu_ptr(pmu
->pmu_cpu_context
);
3949 perf_event_comm_ctx(&cpuctx
->ctx
, comm_event
);
3951 ctxn
= pmu
->task_ctx_nr
;
3955 ctx
= rcu_dereference(current
->perf_event_ctxp
[ctxn
]);
3957 perf_event_comm_ctx(ctx
, comm_event
);
3959 rcu_read_unlock_sched();
3962 void perf_event_comm(struct task_struct
*task
)
3964 struct perf_comm_event comm_event
;
3965 struct perf_event_context
*ctx
;
3968 for_each_task_context_nr(ctxn
) {
3969 ctx
= task
->perf_event_ctxp
[ctxn
];
3973 perf_event_enable_on_exec(ctx
);
3976 if (!atomic_read(&nr_comm_events
))
3979 comm_event
= (struct perf_comm_event
){
3985 .type
= PERF_RECORD_COMM
,
3994 perf_event_comm_event(&comm_event
);
4001 struct perf_mmap_event
{
4002 struct vm_area_struct
*vma
;
4004 const char *file_name
;
4008 struct perf_event_header header
;
4018 static void perf_event_mmap_output(struct perf_event
*event
,
4019 struct perf_mmap_event
*mmap_event
)
4021 struct perf_output_handle handle
;
4022 int size
= mmap_event
->event_id
.header
.size
;
4023 int ret
= perf_output_begin(&handle
, event
, size
, 0, 0);
4028 mmap_event
->event_id
.pid
= perf_event_pid(event
, current
);
4029 mmap_event
->event_id
.tid
= perf_event_tid(event
, current
);
4031 perf_output_put(&handle
, mmap_event
->event_id
);
4032 perf_output_copy(&handle
, mmap_event
->file_name
,
4033 mmap_event
->file_size
);
4034 perf_output_end(&handle
);
4037 static int perf_event_mmap_match(struct perf_event
*event
,
4038 struct perf_mmap_event
*mmap_event
,
4041 if (event
->state
< PERF_EVENT_STATE_INACTIVE
)
4044 if (event
->cpu
!= -1 && event
->cpu
!= smp_processor_id())
4047 if ((!executable
&& event
->attr
.mmap_data
) ||
4048 (executable
&& event
->attr
.mmap
))
4054 static void perf_event_mmap_ctx(struct perf_event_context
*ctx
,
4055 struct perf_mmap_event
*mmap_event
,
4058 struct perf_event
*event
;
4060 list_for_each_entry_rcu(event
, &ctx
->event_list
, event_entry
) {
4061 if (perf_event_mmap_match(event
, mmap_event
, executable
))
4062 perf_event_mmap_output(event
, mmap_event
);
4066 static void perf_event_mmap_event(struct perf_mmap_event
*mmap_event
)
4068 struct perf_cpu_context
*cpuctx
;
4069 struct perf_event_context
*ctx
;
4070 struct vm_area_struct
*vma
= mmap_event
->vma
;
4071 struct file
*file
= vma
->vm_file
;
4079 memset(tmp
, 0, sizeof(tmp
));
4083 * d_path works from the end of the buffer backwards, so we
4084 * need to add enough zero bytes after the string to handle
4085 * the 64bit alignment we do later.
4087 buf
= kzalloc(PATH_MAX
+ sizeof(u64
), GFP_KERNEL
);
4089 name
= strncpy(tmp
, "//enomem", sizeof(tmp
));
4092 name
= d_path(&file
->f_path
, buf
, PATH_MAX
);
4094 name
= strncpy(tmp
, "//toolong", sizeof(tmp
));
4098 if (arch_vma_name(mmap_event
->vma
)) {
4099 name
= strncpy(tmp
, arch_vma_name(mmap_event
->vma
),
4105 name
= strncpy(tmp
, "[vdso]", sizeof(tmp
));
4107 } else if (vma
->vm_start
<= vma
->vm_mm
->start_brk
&&
4108 vma
->vm_end
>= vma
->vm_mm
->brk
) {
4109 name
= strncpy(tmp
, "[heap]", sizeof(tmp
));
4111 } else if (vma
->vm_start
<= vma
->vm_mm
->start_stack
&&
4112 vma
->vm_end
>= vma
->vm_mm
->start_stack
) {
4113 name
= strncpy(tmp
, "[stack]", sizeof(tmp
));
4117 name
= strncpy(tmp
, "//anon", sizeof(tmp
));
4122 size
= ALIGN(strlen(name
)+1, sizeof(u64
));
4124 mmap_event
->file_name
= name
;
4125 mmap_event
->file_size
= size
;
4127 mmap_event
->event_id
.header
.size
= sizeof(mmap_event
->event_id
) + size
;
4129 rcu_read_lock_sched();
4130 list_for_each_entry_rcu(pmu
, &pmus
, entry
) {
4131 cpuctx
= this_cpu_ptr(pmu
->pmu_cpu_context
);
4132 perf_event_mmap_ctx(&cpuctx
->ctx
, mmap_event
,
4133 vma
->vm_flags
& VM_EXEC
);
4135 ctxn
= pmu
->task_ctx_nr
;
4139 ctx
= rcu_dereference(current
->perf_event_ctxp
[ctxn
]);
4141 perf_event_mmap_ctx(ctx
, mmap_event
,
4142 vma
->vm_flags
& VM_EXEC
);
4145 rcu_read_unlock_sched();
4150 void perf_event_mmap(struct vm_area_struct
*vma
)
4152 struct perf_mmap_event mmap_event
;
4154 if (!atomic_read(&nr_mmap_events
))
4157 mmap_event
= (struct perf_mmap_event
){
4163 .type
= PERF_RECORD_MMAP
,
4164 .misc
= PERF_RECORD_MISC_USER
,
4169 .start
= vma
->vm_start
,
4170 .len
= vma
->vm_end
- vma
->vm_start
,
4171 .pgoff
= (u64
)vma
->vm_pgoff
<< PAGE_SHIFT
,
4175 perf_event_mmap_event(&mmap_event
);
4179 * IRQ throttle logging
4182 static void perf_log_throttle(struct perf_event
*event
, int enable
)
4184 struct perf_output_handle handle
;
4188 struct perf_event_header header
;
4192 } throttle_event
= {
4194 .type
= PERF_RECORD_THROTTLE
,
4196 .size
= sizeof(throttle_event
),
4198 .time
= perf_clock(),
4199 .id
= primary_event_id(event
),
4200 .stream_id
= event
->id
,
4204 throttle_event
.header
.type
= PERF_RECORD_UNTHROTTLE
;
4206 ret
= perf_output_begin(&handle
, event
, sizeof(throttle_event
), 1, 0);
4210 perf_output_put(&handle
, throttle_event
);
4211 perf_output_end(&handle
);
4215 * Generic event overflow handling, sampling.
4218 static int __perf_event_overflow(struct perf_event
*event
, int nmi
,
4219 int throttle
, struct perf_sample_data
*data
,
4220 struct pt_regs
*regs
)
4222 int events
= atomic_read(&event
->event_limit
);
4223 struct hw_perf_event
*hwc
= &event
->hw
;
4229 if (hwc
->interrupts
!= MAX_INTERRUPTS
) {
4231 if (HZ
* hwc
->interrupts
>
4232 (u64
)sysctl_perf_event_sample_rate
) {
4233 hwc
->interrupts
= MAX_INTERRUPTS
;
4234 perf_log_throttle(event
, 0);
4239 * Keep re-disabling events even though on the previous
4240 * pass we disabled it - just in case we raced with a
4241 * sched-in and the event got enabled again:
4247 if (event
->attr
.freq
) {
4248 u64 now
= perf_clock();
4249 s64 delta
= now
- hwc
->freq_time_stamp
;
4251 hwc
->freq_time_stamp
= now
;
4253 if (delta
> 0 && delta
< 2*TICK_NSEC
)
4254 perf_adjust_period(event
, delta
, hwc
->last_period
);
4258 * XXX event_limit might not quite work as expected on inherited
4262 event
->pending_kill
= POLL_IN
;
4263 if (events
&& atomic_dec_and_test(&event
->event_limit
)) {
4265 event
->pending_kill
= POLL_HUP
;
4267 event
->pending_disable
= 1;
4268 perf_pending_queue(&event
->pending
,
4269 perf_pending_event
);
4271 perf_event_disable(event
);
4274 if (event
->overflow_handler
)
4275 event
->overflow_handler(event
, nmi
, data
, regs
);
4277 perf_event_output(event
, nmi
, data
, regs
);
4282 int perf_event_overflow(struct perf_event
*event
, int nmi
,
4283 struct perf_sample_data
*data
,
4284 struct pt_regs
*regs
)
4286 return __perf_event_overflow(event
, nmi
, 1, data
, regs
);
4290 * Generic software event infrastructure
4293 struct swevent_htable
{
4294 struct swevent_hlist
*swevent_hlist
;
4295 struct mutex hlist_mutex
;
4298 /* Recursion avoidance in each contexts */
4299 int recursion
[PERF_NR_CONTEXTS
];
4302 static DEFINE_PER_CPU(struct swevent_htable
, swevent_htable
);
4305 * We directly increment event->count and keep a second value in
4306 * event->hw.period_left to count intervals. This period event
4307 * is kept in the range [-sample_period, 0] so that we can use the
4311 static u64
perf_swevent_set_period(struct perf_event
*event
)
4313 struct hw_perf_event
*hwc
= &event
->hw
;
4314 u64 period
= hwc
->last_period
;
4318 hwc
->last_period
= hwc
->sample_period
;
4321 old
= val
= local64_read(&hwc
->period_left
);
4325 nr
= div64_u64(period
+ val
, period
);
4326 offset
= nr
* period
;
4328 if (local64_cmpxchg(&hwc
->period_left
, old
, val
) != old
)
4334 static void perf_swevent_overflow(struct perf_event
*event
, u64 overflow
,
4335 int nmi
, struct perf_sample_data
*data
,
4336 struct pt_regs
*regs
)
4338 struct hw_perf_event
*hwc
= &event
->hw
;
4341 data
->period
= event
->hw
.last_period
;
4343 overflow
= perf_swevent_set_period(event
);
4345 if (hwc
->interrupts
== MAX_INTERRUPTS
)
4348 for (; overflow
; overflow
--) {
4349 if (__perf_event_overflow(event
, nmi
, throttle
,
4352 * We inhibit the overflow from happening when
4353 * hwc->interrupts == MAX_INTERRUPTS.
4361 static void perf_swevent_event(struct perf_event
*event
, u64 nr
,
4362 int nmi
, struct perf_sample_data
*data
,
4363 struct pt_regs
*regs
)
4365 struct hw_perf_event
*hwc
= &event
->hw
;
4367 local64_add(nr
, &event
->count
);
4372 if (!hwc
->sample_period
)
4375 if (nr
== 1 && hwc
->sample_period
== 1 && !event
->attr
.freq
)
4376 return perf_swevent_overflow(event
, 1, nmi
, data
, regs
);
4378 if (local64_add_negative(nr
, &hwc
->period_left
))
4381 perf_swevent_overflow(event
, 0, nmi
, data
, regs
);
4384 static int perf_exclude_event(struct perf_event
*event
,
4385 struct pt_regs
*regs
)
4387 if (event
->hw
.state
& PERF_HES_STOPPED
)
4391 if (event
->attr
.exclude_user
&& user_mode(regs
))
4394 if (event
->attr
.exclude_kernel
&& !user_mode(regs
))
4401 static int perf_swevent_match(struct perf_event
*event
,
4402 enum perf_type_id type
,
4404 struct perf_sample_data
*data
,
4405 struct pt_regs
*regs
)
4407 if (event
->attr
.type
!= type
)
4410 if (event
->attr
.config
!= event_id
)
4413 if (perf_exclude_event(event
, regs
))
4419 static inline u64
swevent_hash(u64 type
, u32 event_id
)
4421 u64 val
= event_id
| (type
<< 32);
4423 return hash_64(val
, SWEVENT_HLIST_BITS
);
4426 static inline struct hlist_head
*
4427 __find_swevent_head(struct swevent_hlist
*hlist
, u64 type
, u32 event_id
)
4429 u64 hash
= swevent_hash(type
, event_id
);
4431 return &hlist
->heads
[hash
];
4434 /* For the read side: events when they trigger */
4435 static inline struct hlist_head
*
4436 find_swevent_head_rcu(struct swevent_htable
*swhash
, u64 type
, u32 event_id
)
4438 struct swevent_hlist
*hlist
;
4440 hlist
= rcu_dereference(swhash
->swevent_hlist
);
4444 return __find_swevent_head(hlist
, type
, event_id
);
4447 /* For the event head insertion and removal in the hlist */
4448 static inline struct hlist_head
*
4449 find_swevent_head(struct swevent_htable
*swhash
, struct perf_event
*event
)
4451 struct swevent_hlist
*hlist
;
4452 u32 event_id
= event
->attr
.config
;
4453 u64 type
= event
->attr
.type
;
4456 * Event scheduling is always serialized against hlist allocation
4457 * and release. Which makes the protected version suitable here.
4458 * The context lock guarantees that.
4460 hlist
= rcu_dereference_protected(swhash
->swevent_hlist
,
4461 lockdep_is_held(&event
->ctx
->lock
));
4465 return __find_swevent_head(hlist
, type
, event_id
);
4468 static void do_perf_sw_event(enum perf_type_id type
, u32 event_id
,
4470 struct perf_sample_data
*data
,
4471 struct pt_regs
*regs
)
4473 struct swevent_htable
*swhash
= &__get_cpu_var(swevent_htable
);
4474 struct perf_event
*event
;
4475 struct hlist_node
*node
;
4476 struct hlist_head
*head
;
4479 head
= find_swevent_head_rcu(swhash
, type
, event_id
);
4483 hlist_for_each_entry_rcu(event
, node
, head
, hlist_entry
) {
4484 if (perf_swevent_match(event
, type
, event_id
, data
, regs
))
4485 perf_swevent_event(event
, nr
, nmi
, data
, regs
);
4491 int perf_swevent_get_recursion_context(void)
4493 struct swevent_htable
*swhash
= &__get_cpu_var(swevent_htable
);
4495 return get_recursion_context(swhash
->recursion
);
4497 EXPORT_SYMBOL_GPL(perf_swevent_get_recursion_context
);
4499 void inline perf_swevent_put_recursion_context(int rctx
)
4501 struct swevent_htable
*swhash
= &__get_cpu_var(swevent_htable
);
4503 put_recursion_context(swhash
->recursion
, rctx
);
4506 void __perf_sw_event(u32 event_id
, u64 nr
, int nmi
,
4507 struct pt_regs
*regs
, u64 addr
)
4509 struct perf_sample_data data
;
4512 preempt_disable_notrace();
4513 rctx
= perf_swevent_get_recursion_context();
4517 perf_sample_data_init(&data
, addr
);
4519 do_perf_sw_event(PERF_TYPE_SOFTWARE
, event_id
, nr
, nmi
, &data
, regs
);
4521 perf_swevent_put_recursion_context(rctx
);
4522 preempt_enable_notrace();
4525 static void perf_swevent_read(struct perf_event
*event
)
4529 static int perf_swevent_add(struct perf_event
*event
, int flags
)
4531 struct swevent_htable
*swhash
= &__get_cpu_var(swevent_htable
);
4532 struct hw_perf_event
*hwc
= &event
->hw
;
4533 struct hlist_head
*head
;
4535 if (hwc
->sample_period
) {
4536 hwc
->last_period
= hwc
->sample_period
;
4537 perf_swevent_set_period(event
);
4540 hwc
->state
= !(flags
& PERF_EF_START
);
4542 head
= find_swevent_head(swhash
, event
);
4543 if (WARN_ON_ONCE(!head
))
4546 hlist_add_head_rcu(&event
->hlist_entry
, head
);
4551 static void perf_swevent_del(struct perf_event
*event
, int flags
)
4553 hlist_del_rcu(&event
->hlist_entry
);
4556 static void perf_swevent_start(struct perf_event
*event
, int flags
)
4558 event
->hw
.state
= 0;
4561 static void perf_swevent_stop(struct perf_event
*event
, int flags
)
4563 event
->hw
.state
= PERF_HES_STOPPED
;
4566 /* Deref the hlist from the update side */
4567 static inline struct swevent_hlist
*
4568 swevent_hlist_deref(struct swevent_htable
*swhash
)
4570 return rcu_dereference_protected(swhash
->swevent_hlist
,
4571 lockdep_is_held(&swhash
->hlist_mutex
));
4574 static void swevent_hlist_release_rcu(struct rcu_head
*rcu_head
)
4576 struct swevent_hlist
*hlist
;
4578 hlist
= container_of(rcu_head
, struct swevent_hlist
, rcu_head
);
4582 static void swevent_hlist_release(struct swevent_htable
*swhash
)
4584 struct swevent_hlist
*hlist
= swevent_hlist_deref(swhash
);
4589 rcu_assign_pointer(swhash
->swevent_hlist
, NULL
);
4590 call_rcu(&hlist
->rcu_head
, swevent_hlist_release_rcu
);
4593 static void swevent_hlist_put_cpu(struct perf_event
*event
, int cpu
)
4595 struct swevent_htable
*swhash
= &per_cpu(swevent_htable
, cpu
);
4597 mutex_lock(&swhash
->hlist_mutex
);
4599 if (!--swhash
->hlist_refcount
)
4600 swevent_hlist_release(swhash
);
4602 mutex_unlock(&swhash
->hlist_mutex
);
4605 static void swevent_hlist_put(struct perf_event
*event
)
4609 if (event
->cpu
!= -1) {
4610 swevent_hlist_put_cpu(event
, event
->cpu
);
4614 for_each_possible_cpu(cpu
)
4615 swevent_hlist_put_cpu(event
, cpu
);
4618 static int swevent_hlist_get_cpu(struct perf_event
*event
, int cpu
)
4620 struct swevent_htable
*swhash
= &per_cpu(swevent_htable
, cpu
);
4623 mutex_lock(&swhash
->hlist_mutex
);
4625 if (!swevent_hlist_deref(swhash
) && cpu_online(cpu
)) {
4626 struct swevent_hlist
*hlist
;
4628 hlist
= kzalloc(sizeof(*hlist
), GFP_KERNEL
);
4633 rcu_assign_pointer(swhash
->swevent_hlist
, hlist
);
4635 swhash
->hlist_refcount
++;
4637 mutex_unlock(&swhash
->hlist_mutex
);
4642 static int swevent_hlist_get(struct perf_event
*event
)
4645 int cpu
, failed_cpu
;
4647 if (event
->cpu
!= -1)
4648 return swevent_hlist_get_cpu(event
, event
->cpu
);
4651 for_each_possible_cpu(cpu
) {
4652 err
= swevent_hlist_get_cpu(event
, cpu
);
4662 for_each_possible_cpu(cpu
) {
4663 if (cpu
== failed_cpu
)
4665 swevent_hlist_put_cpu(event
, cpu
);
4672 atomic_t perf_swevent_enabled
[PERF_COUNT_SW_MAX
];
4674 static void sw_perf_event_destroy(struct perf_event
*event
)
4676 u64 event_id
= event
->attr
.config
;
4678 WARN_ON(event
->parent
);
4680 atomic_dec(&perf_swevent_enabled
[event_id
]);
4681 swevent_hlist_put(event
);
4684 static int perf_swevent_init(struct perf_event
*event
)
4686 int event_id
= event
->attr
.config
;
4688 if (event
->attr
.type
!= PERF_TYPE_SOFTWARE
)
4692 case PERF_COUNT_SW_CPU_CLOCK
:
4693 case PERF_COUNT_SW_TASK_CLOCK
:
4700 if (event_id
> PERF_COUNT_SW_MAX
)
4703 if (!event
->parent
) {
4706 err
= swevent_hlist_get(event
);
4710 atomic_inc(&perf_swevent_enabled
[event_id
]);
4711 event
->destroy
= sw_perf_event_destroy
;
4717 static struct pmu perf_swevent
= {
4718 .task_ctx_nr
= perf_sw_context
,
4720 .event_init
= perf_swevent_init
,
4721 .add
= perf_swevent_add
,
4722 .del
= perf_swevent_del
,
4723 .start
= perf_swevent_start
,
4724 .stop
= perf_swevent_stop
,
4725 .read
= perf_swevent_read
,
4728 #ifdef CONFIG_EVENT_TRACING
4730 static int perf_tp_filter_match(struct perf_event
*event
,
4731 struct perf_sample_data
*data
)
4733 void *record
= data
->raw
->data
;
4735 if (likely(!event
->filter
) || filter_match_preds(event
->filter
, record
))
4740 static int perf_tp_event_match(struct perf_event
*event
,
4741 struct perf_sample_data
*data
,
4742 struct pt_regs
*regs
)
4745 * All tracepoints are from kernel-space.
4747 if (event
->attr
.exclude_kernel
)
4750 if (!perf_tp_filter_match(event
, data
))
4756 void perf_tp_event(u64 addr
, u64 count
, void *record
, int entry_size
,
4757 struct pt_regs
*regs
, struct hlist_head
*head
, int rctx
)
4759 struct perf_sample_data data
;
4760 struct perf_event
*event
;
4761 struct hlist_node
*node
;
4763 struct perf_raw_record raw
= {
4768 perf_sample_data_init(&data
, addr
);
4771 hlist_for_each_entry_rcu(event
, node
, head
, hlist_entry
) {
4772 if (perf_tp_event_match(event
, &data
, regs
))
4773 perf_swevent_event(event
, count
, 1, &data
, regs
);
4776 perf_swevent_put_recursion_context(rctx
);
4778 EXPORT_SYMBOL_GPL(perf_tp_event
);
4780 static void tp_perf_event_destroy(struct perf_event
*event
)
4782 perf_trace_destroy(event
);
4785 static int perf_tp_event_init(struct perf_event
*event
)
4789 if (event
->attr
.type
!= PERF_TYPE_TRACEPOINT
)
4793 * Raw tracepoint data is a severe data leak, only allow root to
4796 if ((event
->attr
.sample_type
& PERF_SAMPLE_RAW
) &&
4797 perf_paranoid_tracepoint_raw() &&
4798 !capable(CAP_SYS_ADMIN
))
4801 err
= perf_trace_init(event
);
4805 event
->destroy
= tp_perf_event_destroy
;
4810 static struct pmu perf_tracepoint
= {
4811 .task_ctx_nr
= perf_sw_context
,
4813 .event_init
= perf_tp_event_init
,
4814 .add
= perf_trace_add
,
4815 .del
= perf_trace_del
,
4816 .start
= perf_swevent_start
,
4817 .stop
= perf_swevent_stop
,
4818 .read
= perf_swevent_read
,
4821 static inline void perf_tp_register(void)
4823 perf_pmu_register(&perf_tracepoint
);
4826 static int perf_event_set_filter(struct perf_event
*event
, void __user
*arg
)
4831 if (event
->attr
.type
!= PERF_TYPE_TRACEPOINT
)
4834 filter_str
= strndup_user(arg
, PAGE_SIZE
);
4835 if (IS_ERR(filter_str
))
4836 return PTR_ERR(filter_str
);
4838 ret
= ftrace_profile_set_filter(event
, event
->attr
.config
, filter_str
);
4844 static void perf_event_free_filter(struct perf_event
*event
)
4846 ftrace_profile_free_filter(event
);
4851 static inline void perf_tp_register(void)
4855 static int perf_event_set_filter(struct perf_event
*event
, void __user
*arg
)
4860 static void perf_event_free_filter(struct perf_event
*event
)
4864 #endif /* CONFIG_EVENT_TRACING */
4866 #ifdef CONFIG_HAVE_HW_BREAKPOINT
4867 void perf_bp_event(struct perf_event
*bp
, void *data
)
4869 struct perf_sample_data sample
;
4870 struct pt_regs
*regs
= data
;
4872 perf_sample_data_init(&sample
, bp
->attr
.bp_addr
);
4874 if (!bp
->hw
.state
&& !perf_exclude_event(bp
, regs
))
4875 perf_swevent_event(bp
, 1, 1, &sample
, regs
);
4880 * hrtimer based swevent callback
4883 static enum hrtimer_restart
perf_swevent_hrtimer(struct hrtimer
*hrtimer
)
4885 enum hrtimer_restart ret
= HRTIMER_RESTART
;
4886 struct perf_sample_data data
;
4887 struct pt_regs
*regs
;
4888 struct perf_event
*event
;
4891 event
= container_of(hrtimer
, struct perf_event
, hw
.hrtimer
);
4892 event
->pmu
->read(event
);
4894 perf_sample_data_init(&data
, 0);
4895 data
.period
= event
->hw
.last_period
;
4896 regs
= get_irq_regs();
4898 if (regs
&& !perf_exclude_event(event
, regs
)) {
4899 if (!(event
->attr
.exclude_idle
&& current
->pid
== 0))
4900 if (perf_event_overflow(event
, 0, &data
, regs
))
4901 ret
= HRTIMER_NORESTART
;
4904 period
= max_t(u64
, 10000, event
->hw
.sample_period
);
4905 hrtimer_forward_now(hrtimer
, ns_to_ktime(period
));
4910 static void perf_swevent_start_hrtimer(struct perf_event
*event
)
4912 struct hw_perf_event
*hwc
= &event
->hw
;
4914 hrtimer_init(&hwc
->hrtimer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
4915 hwc
->hrtimer
.function
= perf_swevent_hrtimer
;
4916 if (hwc
->sample_period
) {
4917 s64 period
= local64_read(&hwc
->period_left
);
4923 local64_set(&hwc
->period_left
, 0);
4925 period
= max_t(u64
, 10000, hwc
->sample_period
);
4927 __hrtimer_start_range_ns(&hwc
->hrtimer
,
4928 ns_to_ktime(period
), 0,
4929 HRTIMER_MODE_REL_PINNED
, 0);
4933 static void perf_swevent_cancel_hrtimer(struct perf_event
*event
)
4935 struct hw_perf_event
*hwc
= &event
->hw
;
4937 if (hwc
->sample_period
) {
4938 ktime_t remaining
= hrtimer_get_remaining(&hwc
->hrtimer
);
4939 local64_set(&hwc
->period_left
, ktime_to_ns(remaining
));
4941 hrtimer_cancel(&hwc
->hrtimer
);
4946 * Software event: cpu wall time clock
4949 static void cpu_clock_event_update(struct perf_event
*event
)
4954 now
= local_clock();
4955 prev
= local64_xchg(&event
->hw
.prev_count
, now
);
4956 local64_add(now
- prev
, &event
->count
);
4959 static void cpu_clock_event_start(struct perf_event
*event
, int flags
)
4961 local64_set(&event
->hw
.prev_count
, local_clock());
4962 perf_swevent_start_hrtimer(event
);
4965 static void cpu_clock_event_stop(struct perf_event
*event
, int flags
)
4967 perf_swevent_cancel_hrtimer(event
);
4968 cpu_clock_event_update(event
);
4971 static int cpu_clock_event_add(struct perf_event
*event
, int flags
)
4973 if (flags
& PERF_EF_START
)
4974 cpu_clock_event_start(event
, flags
);
4979 static void cpu_clock_event_del(struct perf_event
*event
, int flags
)
4981 cpu_clock_event_stop(event
, flags
);
4984 static void cpu_clock_event_read(struct perf_event
*event
)
4986 cpu_clock_event_update(event
);
4989 static int cpu_clock_event_init(struct perf_event
*event
)
4991 if (event
->attr
.type
!= PERF_TYPE_SOFTWARE
)
4994 if (event
->attr
.config
!= PERF_COUNT_SW_CPU_CLOCK
)
5000 static struct pmu perf_cpu_clock
= {
5001 .task_ctx_nr
= perf_sw_context
,
5003 .event_init
= cpu_clock_event_init
,
5004 .add
= cpu_clock_event_add
,
5005 .del
= cpu_clock_event_del
,
5006 .start
= cpu_clock_event_start
,
5007 .stop
= cpu_clock_event_stop
,
5008 .read
= cpu_clock_event_read
,
5012 * Software event: task time clock
5015 static void task_clock_event_update(struct perf_event
*event
, u64 now
)
5020 prev
= local64_xchg(&event
->hw
.prev_count
, now
);
5022 local64_add(delta
, &event
->count
);
5025 static void task_clock_event_start(struct perf_event
*event
, int flags
)
5027 local64_set(&event
->hw
.prev_count
, event
->ctx
->time
);
5028 perf_swevent_start_hrtimer(event
);
5031 static void task_clock_event_stop(struct perf_event
*event
, int flags
)
5033 perf_swevent_cancel_hrtimer(event
);
5034 task_clock_event_update(event
, event
->ctx
->time
);
5037 static int task_clock_event_add(struct perf_event
*event
, int flags
)
5039 if (flags
& PERF_EF_START
)
5040 task_clock_event_start(event
, flags
);
5045 static void task_clock_event_del(struct perf_event
*event
, int flags
)
5047 task_clock_event_stop(event
, PERF_EF_UPDATE
);
5050 static void task_clock_event_read(struct perf_event
*event
)
5055 update_context_time(event
->ctx
);
5056 time
= event
->ctx
->time
;
5058 u64 now
= perf_clock();
5059 u64 delta
= now
- event
->ctx
->timestamp
;
5060 time
= event
->ctx
->time
+ delta
;
5063 task_clock_event_update(event
, time
);
5066 static int task_clock_event_init(struct perf_event
*event
)
5068 if (event
->attr
.type
!= PERF_TYPE_SOFTWARE
)
5071 if (event
->attr
.config
!= PERF_COUNT_SW_TASK_CLOCK
)
5077 static struct pmu perf_task_clock
= {
5078 .task_ctx_nr
= perf_sw_context
,
5080 .event_init
= task_clock_event_init
,
5081 .add
= task_clock_event_add
,
5082 .del
= task_clock_event_del
,
5083 .start
= task_clock_event_start
,
5084 .stop
= task_clock_event_stop
,
5085 .read
= task_clock_event_read
,
5088 static void perf_pmu_nop_void(struct pmu
*pmu
)
5092 static int perf_pmu_nop_int(struct pmu
*pmu
)
5097 static void perf_pmu_start_txn(struct pmu
*pmu
)
5099 perf_pmu_disable(pmu
);
5102 static int perf_pmu_commit_txn(struct pmu
*pmu
)
5104 perf_pmu_enable(pmu
);
5108 static void perf_pmu_cancel_txn(struct pmu
*pmu
)
5110 perf_pmu_enable(pmu
);
5114 * Ensures all contexts with the same task_ctx_nr have the same
5115 * pmu_cpu_context too.
5117 static void *find_pmu_context(int ctxn
)
5124 list_for_each_entry(pmu
, &pmus
, entry
) {
5125 if (pmu
->task_ctx_nr
== ctxn
)
5126 return pmu
->pmu_cpu_context
;
5132 static void free_pmu_context(void * __percpu cpu_context
)
5136 mutex_lock(&pmus_lock
);
5138 * Like a real lame refcount.
5140 list_for_each_entry(pmu
, &pmus
, entry
) {
5141 if (pmu
->pmu_cpu_context
== cpu_context
)
5145 free_percpu(cpu_context
);
5147 mutex_unlock(&pmus_lock
);
5150 int perf_pmu_register(struct pmu
*pmu
)
5154 mutex_lock(&pmus_lock
);
5156 pmu
->pmu_disable_count
= alloc_percpu(int);
5157 if (!pmu
->pmu_disable_count
)
5160 pmu
->pmu_cpu_context
= find_pmu_context(pmu
->task_ctx_nr
);
5161 if (pmu
->pmu_cpu_context
)
5162 goto got_cpu_context
;
5164 pmu
->pmu_cpu_context
= alloc_percpu(struct perf_cpu_context
);
5165 if (!pmu
->pmu_cpu_context
)
5168 for_each_possible_cpu(cpu
) {
5169 struct perf_cpu_context
*cpuctx
;
5171 cpuctx
= per_cpu_ptr(pmu
->pmu_cpu_context
, cpu
);
5172 __perf_event_init_context(&cpuctx
->ctx
);
5173 cpuctx
->ctx
.pmu
= pmu
;
5174 cpuctx
->timer_interval
= TICK_NSEC
;
5175 hrtimer_init(&cpuctx
->timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
5176 cpuctx
->timer
.function
= perf_event_context_tick
;
5180 if (!pmu
->start_txn
) {
5181 if (pmu
->pmu_enable
) {
5183 * If we have pmu_enable/pmu_disable calls, install
5184 * transaction stubs that use that to try and batch
5185 * hardware accesses.
5187 pmu
->start_txn
= perf_pmu_start_txn
;
5188 pmu
->commit_txn
= perf_pmu_commit_txn
;
5189 pmu
->cancel_txn
= perf_pmu_cancel_txn
;
5191 pmu
->start_txn
= perf_pmu_nop_void
;
5192 pmu
->commit_txn
= perf_pmu_nop_int
;
5193 pmu
->cancel_txn
= perf_pmu_nop_void
;
5197 if (!pmu
->pmu_enable
) {
5198 pmu
->pmu_enable
= perf_pmu_nop_void
;
5199 pmu
->pmu_disable
= perf_pmu_nop_void
;
5202 list_add_rcu(&pmu
->entry
, &pmus
);
5205 mutex_unlock(&pmus_lock
);
5210 free_percpu(pmu
->pmu_disable_count
);
5214 void perf_pmu_unregister(struct pmu
*pmu
)
5216 mutex_lock(&pmus_lock
);
5217 list_del_rcu(&pmu
->entry
);
5218 mutex_unlock(&pmus_lock
);
5221 * We use the pmu list either under SRCU or preempt_disable,
5222 * synchronize_srcu() implies synchronize_sched() so we're good.
5224 synchronize_srcu(&pmus_srcu
);
5226 free_percpu(pmu
->pmu_disable_count
);
5227 free_pmu_context(pmu
->pmu_cpu_context
);
5230 struct pmu
*perf_init_event(struct perf_event
*event
)
5232 struct pmu
*pmu
= NULL
;
5235 idx
= srcu_read_lock(&pmus_srcu
);
5236 list_for_each_entry_rcu(pmu
, &pmus
, entry
) {
5237 int ret
= pmu
->event_init(event
);
5240 if (ret
!= -ENOENT
) {
5245 srcu_read_unlock(&pmus_srcu
, idx
);
5251 * Allocate and initialize a event structure
5253 static struct perf_event
*
5254 perf_event_alloc(struct perf_event_attr
*attr
, int cpu
,
5255 struct perf_event
*group_leader
,
5256 struct perf_event
*parent_event
,
5257 perf_overflow_handler_t overflow_handler
)
5260 struct perf_event
*event
;
5261 struct hw_perf_event
*hwc
;
5264 event
= kzalloc(sizeof(*event
), GFP_KERNEL
);
5266 return ERR_PTR(-ENOMEM
);
5269 * Single events are their own group leaders, with an
5270 * empty sibling list:
5273 group_leader
= event
;
5275 mutex_init(&event
->child_mutex
);
5276 INIT_LIST_HEAD(&event
->child_list
);
5278 INIT_LIST_HEAD(&event
->group_entry
);
5279 INIT_LIST_HEAD(&event
->event_entry
);
5280 INIT_LIST_HEAD(&event
->sibling_list
);
5281 init_waitqueue_head(&event
->waitq
);
5283 mutex_init(&event
->mmap_mutex
);
5286 event
->attr
= *attr
;
5287 event
->group_leader
= group_leader
;
5291 event
->parent
= parent_event
;
5293 event
->ns
= get_pid_ns(current
->nsproxy
->pid_ns
);
5294 event
->id
= atomic64_inc_return(&perf_event_id
);
5296 event
->state
= PERF_EVENT_STATE_INACTIVE
;
5298 if (!overflow_handler
&& parent_event
)
5299 overflow_handler
= parent_event
->overflow_handler
;
5301 event
->overflow_handler
= overflow_handler
;
5304 event
->state
= PERF_EVENT_STATE_OFF
;
5309 hwc
->sample_period
= attr
->sample_period
;
5310 if (attr
->freq
&& attr
->sample_freq
)
5311 hwc
->sample_period
= 1;
5312 hwc
->last_period
= hwc
->sample_period
;
5314 local64_set(&hwc
->period_left
, hwc
->sample_period
);
5317 * we currently do not support PERF_FORMAT_GROUP on inherited events
5319 if (attr
->inherit
&& (attr
->read_format
& PERF_FORMAT_GROUP
))
5322 pmu
= perf_init_event(event
);
5328 else if (IS_ERR(pmu
))
5333 put_pid_ns(event
->ns
);
5335 return ERR_PTR(err
);
5340 if (!event
->parent
) {
5341 atomic_inc(&nr_events
);
5342 if (event
->attr
.mmap
|| event
->attr
.mmap_data
)
5343 atomic_inc(&nr_mmap_events
);
5344 if (event
->attr
.comm
)
5345 atomic_inc(&nr_comm_events
);
5346 if (event
->attr
.task
)
5347 atomic_inc(&nr_task_events
);
5348 if (event
->attr
.sample_type
& PERF_SAMPLE_CALLCHAIN
) {
5349 err
= get_callchain_buffers();
5352 return ERR_PTR(err
);
5360 static int perf_copy_attr(struct perf_event_attr __user
*uattr
,
5361 struct perf_event_attr
*attr
)
5366 if (!access_ok(VERIFY_WRITE
, uattr
, PERF_ATTR_SIZE_VER0
))
5370 * zero the full structure, so that a short copy will be nice.
5372 memset(attr
, 0, sizeof(*attr
));
5374 ret
= get_user(size
, &uattr
->size
);
5378 if (size
> PAGE_SIZE
) /* silly large */
5381 if (!size
) /* abi compat */
5382 size
= PERF_ATTR_SIZE_VER0
;
5384 if (size
< PERF_ATTR_SIZE_VER0
)
5388 * If we're handed a bigger struct than we know of,
5389 * ensure all the unknown bits are 0 - i.e. new
5390 * user-space does not rely on any kernel feature
5391 * extensions we dont know about yet.
5393 if (size
> sizeof(*attr
)) {
5394 unsigned char __user
*addr
;
5395 unsigned char __user
*end
;
5398 addr
= (void __user
*)uattr
+ sizeof(*attr
);
5399 end
= (void __user
*)uattr
+ size
;
5401 for (; addr
< end
; addr
++) {
5402 ret
= get_user(val
, addr
);
5408 size
= sizeof(*attr
);
5411 ret
= copy_from_user(attr
, uattr
, size
);
5416 * If the type exists, the corresponding creation will verify
5419 if (attr
->type
>= PERF_TYPE_MAX
)
5422 if (attr
->__reserved_1
)
5425 if (attr
->sample_type
& ~(PERF_SAMPLE_MAX
-1))
5428 if (attr
->read_format
& ~(PERF_FORMAT_MAX
-1))
5435 put_user(sizeof(*attr
), &uattr
->size
);
5441 perf_event_set_output(struct perf_event
*event
, struct perf_event
*output_event
)
5443 struct perf_buffer
*buffer
= NULL
, *old_buffer
= NULL
;
5449 /* don't allow circular references */
5450 if (event
== output_event
)
5454 * Don't allow cross-cpu buffers
5456 if (output_event
->cpu
!= event
->cpu
)
5460 * If its not a per-cpu buffer, it must be the same task.
5462 if (output_event
->cpu
== -1 && output_event
->ctx
!= event
->ctx
)
5466 mutex_lock(&event
->mmap_mutex
);
5467 /* Can't redirect output if we've got an active mmap() */
5468 if (atomic_read(&event
->mmap_count
))
5472 /* get the buffer we want to redirect to */
5473 buffer
= perf_buffer_get(output_event
);
5478 old_buffer
= event
->buffer
;
5479 rcu_assign_pointer(event
->buffer
, buffer
);
5482 mutex_unlock(&event
->mmap_mutex
);
5485 perf_buffer_put(old_buffer
);
5491 * sys_perf_event_open - open a performance event, associate it to a task/cpu
5493 * @attr_uptr: event_id type attributes for monitoring/sampling
5496 * @group_fd: group leader event fd
5498 SYSCALL_DEFINE5(perf_event_open
,
5499 struct perf_event_attr __user
*, attr_uptr
,
5500 pid_t
, pid
, int, cpu
, int, group_fd
, unsigned long, flags
)
5502 struct perf_event
*event
, *group_leader
= NULL
, *output_event
= NULL
;
5503 struct perf_event_attr attr
;
5504 struct perf_event_context
*ctx
;
5505 struct file
*event_file
= NULL
;
5506 struct file
*group_file
= NULL
;
5509 int fput_needed
= 0;
5512 /* for future expandability... */
5513 if (flags
& ~(PERF_FLAG_FD_NO_GROUP
| PERF_FLAG_FD_OUTPUT
))
5516 err
= perf_copy_attr(attr_uptr
, &attr
);
5520 if (!attr
.exclude_kernel
) {
5521 if (perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN
))
5526 if (attr
.sample_freq
> sysctl_perf_event_sample_rate
)
5530 event_fd
= get_unused_fd_flags(O_RDWR
);
5534 event
= perf_event_alloc(&attr
, cpu
, group_leader
, NULL
, NULL
);
5535 if (IS_ERR(event
)) {
5536 err
= PTR_ERR(event
);
5540 if (group_fd
!= -1) {
5541 group_leader
= perf_fget_light(group_fd
, &fput_needed
);
5542 if (IS_ERR(group_leader
)) {
5543 err
= PTR_ERR(group_leader
);
5546 group_file
= group_leader
->filp
;
5547 if (flags
& PERF_FLAG_FD_OUTPUT
)
5548 output_event
= group_leader
;
5549 if (flags
& PERF_FLAG_FD_NO_GROUP
)
5550 group_leader
= NULL
;
5554 * Special case software events and allow them to be part of
5555 * any hardware group.
5558 if ((pmu
->task_ctx_nr
== perf_sw_context
) && group_leader
)
5559 pmu
= group_leader
->pmu
;
5562 * Get the target context (task or percpu):
5564 ctx
= find_get_context(pmu
, pid
, cpu
);
5571 * Look up the group leader (we will attach this event to it):
5577 * Do not allow a recursive hierarchy (this new sibling
5578 * becoming part of another group-sibling):
5580 if (group_leader
->group_leader
!= group_leader
)
5583 * Do not allow to attach to a group in a different
5584 * task or CPU context:
5586 if (group_leader
->ctx
!= ctx
)
5589 * Only a group leader can be exclusive or pinned
5591 if (attr
.exclusive
|| attr
.pinned
)
5596 err
= perf_event_set_output(event
, output_event
);
5601 event_file
= anon_inode_getfile("[perf_event]", &perf_fops
, event
, O_RDWR
);
5602 if (IS_ERR(event_file
)) {
5603 err
= PTR_ERR(event_file
);
5607 event
->filp
= event_file
;
5608 WARN_ON_ONCE(ctx
->parent_ctx
);
5609 mutex_lock(&ctx
->mutex
);
5610 perf_install_in_context(ctx
, event
, cpu
);
5612 mutex_unlock(&ctx
->mutex
);
5614 event
->owner
= current
;
5615 get_task_struct(current
);
5616 mutex_lock(¤t
->perf_event_mutex
);
5617 list_add_tail(&event
->owner_entry
, ¤t
->perf_event_list
);
5618 mutex_unlock(¤t
->perf_event_mutex
);
5621 * Drop the reference on the group_event after placing the
5622 * new event on the sibling_list. This ensures destruction
5623 * of the group leader will find the pointer to itself in
5624 * perf_group_detach().
5626 fput_light(group_file
, fput_needed
);
5627 fd_install(event_fd
, event_file
);
5633 fput_light(group_file
, fput_needed
);
5637 put_unused_fd(event_fd
);
5642 * perf_event_create_kernel_counter
5644 * @attr: attributes of the counter to create
5645 * @cpu: cpu in which the counter is bound
5646 * @pid: task to profile
5649 perf_event_create_kernel_counter(struct perf_event_attr
*attr
, int cpu
,
5651 perf_overflow_handler_t overflow_handler
)
5653 struct perf_event_context
*ctx
;
5654 struct perf_event
*event
;
5658 * Get the target context (task or percpu):
5661 event
= perf_event_alloc(attr
, cpu
, NULL
, NULL
, overflow_handler
);
5662 if (IS_ERR(event
)) {
5663 err
= PTR_ERR(event
);
5667 ctx
= find_get_context(event
->pmu
, pid
, cpu
);
5674 WARN_ON_ONCE(ctx
->parent_ctx
);
5675 mutex_lock(&ctx
->mutex
);
5676 perf_install_in_context(ctx
, event
, cpu
);
5678 mutex_unlock(&ctx
->mutex
);
5680 event
->owner
= current
;
5681 get_task_struct(current
);
5682 mutex_lock(¤t
->perf_event_mutex
);
5683 list_add_tail(&event
->owner_entry
, ¤t
->perf_event_list
);
5684 mutex_unlock(¤t
->perf_event_mutex
);
5691 return ERR_PTR(err
);
5693 EXPORT_SYMBOL_GPL(perf_event_create_kernel_counter
);
5695 static void sync_child_event(struct perf_event
*child_event
,
5696 struct task_struct
*child
)
5698 struct perf_event
*parent_event
= child_event
->parent
;
5701 if (child_event
->attr
.inherit_stat
)
5702 perf_event_read_event(child_event
, child
);
5704 child_val
= perf_event_count(child_event
);
5707 * Add back the child's count to the parent's count:
5709 atomic64_add(child_val
, &parent_event
->child_count
);
5710 atomic64_add(child_event
->total_time_enabled
,
5711 &parent_event
->child_total_time_enabled
);
5712 atomic64_add(child_event
->total_time_running
,
5713 &parent_event
->child_total_time_running
);
5716 * Remove this event from the parent's list
5718 WARN_ON_ONCE(parent_event
->ctx
->parent_ctx
);
5719 mutex_lock(&parent_event
->child_mutex
);
5720 list_del_init(&child_event
->child_list
);
5721 mutex_unlock(&parent_event
->child_mutex
);
5724 * Release the parent event, if this was the last
5727 fput(parent_event
->filp
);
5731 __perf_event_exit_task(struct perf_event
*child_event
,
5732 struct perf_event_context
*child_ctx
,
5733 struct task_struct
*child
)
5735 struct perf_event
*parent_event
;
5737 perf_event_remove_from_context(child_event
);
5739 parent_event
= child_event
->parent
;
5741 * It can happen that parent exits first, and has events
5742 * that are still around due to the child reference. These
5743 * events need to be zapped - but otherwise linger.
5746 sync_child_event(child_event
, child
);
5747 free_event(child_event
);
5751 static void perf_event_exit_task_context(struct task_struct
*child
, int ctxn
)
5753 struct perf_event
*child_event
, *tmp
;
5754 struct perf_event_context
*child_ctx
;
5755 unsigned long flags
;
5757 if (likely(!child
->perf_event_ctxp
[ctxn
])) {
5758 perf_event_task(child
, NULL
, 0);
5762 local_irq_save(flags
);
5764 * We can't reschedule here because interrupts are disabled,
5765 * and either child is current or it is a task that can't be
5766 * scheduled, so we are now safe from rescheduling changing
5769 child_ctx
= child
->perf_event_ctxp
[ctxn
];
5770 __perf_event_task_sched_out(child_ctx
);
5773 * Take the context lock here so that if find_get_context is
5774 * reading child->perf_event_ctxp, we wait until it has
5775 * incremented the context's refcount before we do put_ctx below.
5777 raw_spin_lock(&child_ctx
->lock
);
5778 child
->perf_event_ctxp
[ctxn
] = NULL
;
5780 * If this context is a clone; unclone it so it can't get
5781 * swapped to another process while we're removing all
5782 * the events from it.
5784 unclone_ctx(child_ctx
);
5785 update_context_time(child_ctx
);
5786 raw_spin_unlock_irqrestore(&child_ctx
->lock
, flags
);
5789 * Report the task dead after unscheduling the events so that we
5790 * won't get any samples after PERF_RECORD_EXIT. We can however still
5791 * get a few PERF_RECORD_READ events.
5793 perf_event_task(child
, child_ctx
, 0);
5796 * We can recurse on the same lock type through:
5798 * __perf_event_exit_task()
5799 * sync_child_event()
5800 * fput(parent_event->filp)
5802 * mutex_lock(&ctx->mutex)
5804 * But since its the parent context it won't be the same instance.
5806 mutex_lock(&child_ctx
->mutex
);
5809 list_for_each_entry_safe(child_event
, tmp
, &child_ctx
->pinned_groups
,
5811 __perf_event_exit_task(child_event
, child_ctx
, child
);
5813 list_for_each_entry_safe(child_event
, tmp
, &child_ctx
->flexible_groups
,
5815 __perf_event_exit_task(child_event
, child_ctx
, child
);
5818 * If the last event was a group event, it will have appended all
5819 * its siblings to the list, but we obtained 'tmp' before that which
5820 * will still point to the list head terminating the iteration.
5822 if (!list_empty(&child_ctx
->pinned_groups
) ||
5823 !list_empty(&child_ctx
->flexible_groups
))
5826 mutex_unlock(&child_ctx
->mutex
);
5832 * When a child task exits, feed back event values to parent events.
5834 void perf_event_exit_task(struct task_struct
*child
)
5838 for_each_task_context_nr(ctxn
)
5839 perf_event_exit_task_context(child
, ctxn
);
5842 static void perf_free_event(struct perf_event
*event
,
5843 struct perf_event_context
*ctx
)
5845 struct perf_event
*parent
= event
->parent
;
5847 if (WARN_ON_ONCE(!parent
))
5850 mutex_lock(&parent
->child_mutex
);
5851 list_del_init(&event
->child_list
);
5852 mutex_unlock(&parent
->child_mutex
);
5856 perf_group_detach(event
);
5857 list_del_event(event
, ctx
);
5862 * free an unexposed, unused context as created by inheritance by
5863 * perf_event_init_task below, used by fork() in case of fail.
5865 void perf_event_free_task(struct task_struct
*task
)
5867 struct perf_event_context
*ctx
;
5868 struct perf_event
*event
, *tmp
;
5871 for_each_task_context_nr(ctxn
) {
5872 ctx
= task
->perf_event_ctxp
[ctxn
];
5876 mutex_lock(&ctx
->mutex
);
5878 list_for_each_entry_safe(event
, tmp
, &ctx
->pinned_groups
,
5880 perf_free_event(event
, ctx
);
5882 list_for_each_entry_safe(event
, tmp
, &ctx
->flexible_groups
,
5884 perf_free_event(event
, ctx
);
5886 if (!list_empty(&ctx
->pinned_groups
) ||
5887 !list_empty(&ctx
->flexible_groups
))
5890 mutex_unlock(&ctx
->mutex
);
5897 * inherit a event from parent task to child task:
5899 static struct perf_event
*
5900 inherit_event(struct perf_event
*parent_event
,
5901 struct task_struct
*parent
,
5902 struct perf_event_context
*parent_ctx
,
5903 struct task_struct
*child
,
5904 struct perf_event
*group_leader
,
5905 struct perf_event_context
*child_ctx
)
5907 struct perf_event
*child_event
;
5910 * Instead of creating recursive hierarchies of events,
5911 * we link inherited events back to the original parent,
5912 * which has a filp for sure, which we use as the reference
5915 if (parent_event
->parent
)
5916 parent_event
= parent_event
->parent
;
5918 child_event
= perf_event_alloc(&parent_event
->attr
,
5920 group_leader
, parent_event
,
5922 if (IS_ERR(child_event
))
5927 * Make the child state follow the state of the parent event,
5928 * not its attr.disabled bit. We hold the parent's mutex,
5929 * so we won't race with perf_event_{en, dis}able_family.
5931 if (parent_event
->state
>= PERF_EVENT_STATE_INACTIVE
)
5932 child_event
->state
= PERF_EVENT_STATE_INACTIVE
;
5934 child_event
->state
= PERF_EVENT_STATE_OFF
;
5936 if (parent_event
->attr
.freq
) {
5937 u64 sample_period
= parent_event
->hw
.sample_period
;
5938 struct hw_perf_event
*hwc
= &child_event
->hw
;
5940 hwc
->sample_period
= sample_period
;
5941 hwc
->last_period
= sample_period
;
5943 local64_set(&hwc
->period_left
, sample_period
);
5946 child_event
->ctx
= child_ctx
;
5947 child_event
->overflow_handler
= parent_event
->overflow_handler
;
5950 * Link it up in the child's context:
5952 add_event_to_ctx(child_event
, child_ctx
);
5955 * Get a reference to the parent filp - we will fput it
5956 * when the child event exits. This is safe to do because
5957 * we are in the parent and we know that the filp still
5958 * exists and has a nonzero count:
5960 atomic_long_inc(&parent_event
->filp
->f_count
);
5963 * Link this into the parent event's child list
5965 WARN_ON_ONCE(parent_event
->ctx
->parent_ctx
);
5966 mutex_lock(&parent_event
->child_mutex
);
5967 list_add_tail(&child_event
->child_list
, &parent_event
->child_list
);
5968 mutex_unlock(&parent_event
->child_mutex
);
5973 static int inherit_group(struct perf_event
*parent_event
,
5974 struct task_struct
*parent
,
5975 struct perf_event_context
*parent_ctx
,
5976 struct task_struct
*child
,
5977 struct perf_event_context
*child_ctx
)
5979 struct perf_event
*leader
;
5980 struct perf_event
*sub
;
5981 struct perf_event
*child_ctr
;
5983 leader
= inherit_event(parent_event
, parent
, parent_ctx
,
5984 child
, NULL
, child_ctx
);
5986 return PTR_ERR(leader
);
5987 list_for_each_entry(sub
, &parent_event
->sibling_list
, group_entry
) {
5988 child_ctr
= inherit_event(sub
, parent
, parent_ctx
,
5989 child
, leader
, child_ctx
);
5990 if (IS_ERR(child_ctr
))
5991 return PTR_ERR(child_ctr
);
5997 inherit_task_group(struct perf_event
*event
, struct task_struct
*parent
,
5998 struct perf_event_context
*parent_ctx
,
5999 struct task_struct
*child
, int ctxn
,
6003 struct perf_event_context
*child_ctx
;
6005 if (!event
->attr
.inherit
) {
6010 child_ctx
= child
->perf_event_ctxp
[ctxn
];
6013 * This is executed from the parent task context, so
6014 * inherit events that have been marked for cloning.
6015 * First allocate and initialize a context for the
6019 child_ctx
= alloc_perf_context(event
->pmu
, child
);
6023 child
->perf_event_ctxp
[ctxn
] = child_ctx
;
6026 ret
= inherit_group(event
, parent
, parent_ctx
,
6036 * Initialize the perf_event context in task_struct
6038 int perf_event_init_context(struct task_struct
*child
, int ctxn
)
6040 struct perf_event_context
*child_ctx
, *parent_ctx
;
6041 struct perf_event_context
*cloned_ctx
;
6042 struct perf_event
*event
;
6043 struct task_struct
*parent
= current
;
6044 int inherited_all
= 1;
6047 child
->perf_event_ctxp
[ctxn
] = NULL
;
6049 mutex_init(&child
->perf_event_mutex
);
6050 INIT_LIST_HEAD(&child
->perf_event_list
);
6052 if (likely(!parent
->perf_event_ctxp
[ctxn
]))
6056 * If the parent's context is a clone, pin it so it won't get
6059 parent_ctx
= perf_pin_task_context(parent
, ctxn
);
6062 * No need to check if parent_ctx != NULL here; since we saw
6063 * it non-NULL earlier, the only reason for it to become NULL
6064 * is if we exit, and since we're currently in the middle of
6065 * a fork we can't be exiting at the same time.
6069 * Lock the parent list. No need to lock the child - not PID
6070 * hashed yet and not running, so nobody can access it.
6072 mutex_lock(&parent_ctx
->mutex
);
6075 * We dont have to disable NMIs - we are only looking at
6076 * the list, not manipulating it:
6078 list_for_each_entry(event
, &parent_ctx
->pinned_groups
, group_entry
) {
6079 ret
= inherit_task_group(event
, parent
, parent_ctx
,
6080 child
, ctxn
, &inherited_all
);
6085 list_for_each_entry(event
, &parent_ctx
->flexible_groups
, group_entry
) {
6086 ret
= inherit_task_group(event
, parent
, parent_ctx
,
6087 child
, ctxn
, &inherited_all
);
6092 child_ctx
= child
->perf_event_ctxp
[ctxn
];
6094 if (child_ctx
&& inherited_all
) {
6096 * Mark the child context as a clone of the parent
6097 * context, or of whatever the parent is a clone of.
6098 * Note that if the parent is a clone, it could get
6099 * uncloned at any point, but that doesn't matter
6100 * because the list of events and the generation
6101 * count can't have changed since we took the mutex.
6103 cloned_ctx
= rcu_dereference(parent_ctx
->parent_ctx
);
6105 child_ctx
->parent_ctx
= cloned_ctx
;
6106 child_ctx
->parent_gen
= parent_ctx
->parent_gen
;
6108 child_ctx
->parent_ctx
= parent_ctx
;
6109 child_ctx
->parent_gen
= parent_ctx
->generation
;
6111 get_ctx(child_ctx
->parent_ctx
);
6114 mutex_unlock(&parent_ctx
->mutex
);
6116 perf_unpin_context(parent_ctx
);
6122 * Initialize the perf_event context in task_struct
6124 int perf_event_init_task(struct task_struct
*child
)
6128 for_each_task_context_nr(ctxn
) {
6129 ret
= perf_event_init_context(child
, ctxn
);
6137 static void __init
perf_event_init_all_cpus(void)
6139 struct swevent_htable
*swhash
;
6142 for_each_possible_cpu(cpu
) {
6143 swhash
= &per_cpu(swevent_htable
, cpu
);
6144 mutex_init(&swhash
->hlist_mutex
);
6148 static void __cpuinit
perf_event_init_cpu(int cpu
)
6150 struct swevent_htable
*swhash
= &per_cpu(swevent_htable
, cpu
);
6152 mutex_lock(&swhash
->hlist_mutex
);
6153 if (swhash
->hlist_refcount
> 0) {
6154 struct swevent_hlist
*hlist
;
6156 hlist
= kzalloc_node(sizeof(*hlist
), GFP_KERNEL
, cpu_to_node(cpu
));
6158 rcu_assign_pointer(swhash
->swevent_hlist
, hlist
);
6160 mutex_unlock(&swhash
->hlist_mutex
);
6163 #ifdef CONFIG_HOTPLUG_CPU
6164 static void __perf_event_exit_context(void *__info
)
6166 struct perf_event_context
*ctx
= __info
;
6167 struct perf_event
*event
, *tmp
;
6169 perf_pmu_rotate_stop(ctx
->pmu
);
6171 list_for_each_entry_safe(event
, tmp
, &ctx
->pinned_groups
, group_entry
)
6172 __perf_event_remove_from_context(event
);
6173 list_for_each_entry_safe(event
, tmp
, &ctx
->flexible_groups
, group_entry
)
6174 __perf_event_remove_from_context(event
);
6177 static void perf_event_exit_cpu_context(int cpu
)
6179 struct perf_event_context
*ctx
;
6183 idx
= srcu_read_lock(&pmus_srcu
);
6184 list_for_each_entry_rcu(pmu
, &pmus
, entry
) {
6185 ctx
= &this_cpu_ptr(pmu
->pmu_cpu_context
)->ctx
;
6187 mutex_lock(&ctx
->mutex
);
6188 smp_call_function_single(cpu
, __perf_event_exit_context
, ctx
, 1);
6189 mutex_unlock(&ctx
->mutex
);
6191 srcu_read_unlock(&pmus_srcu
, idx
);
6195 static void perf_event_exit_cpu(int cpu
)
6197 struct swevent_htable
*swhash
= &per_cpu(swevent_htable
, cpu
);
6199 mutex_lock(&swhash
->hlist_mutex
);
6200 swevent_hlist_release(swhash
);
6201 mutex_unlock(&swhash
->hlist_mutex
);
6203 perf_event_exit_cpu_context(cpu
);
6206 static inline void perf_event_exit_cpu(int cpu
) { }
6209 static int __cpuinit
6210 perf_cpu_notify(struct notifier_block
*self
, unsigned long action
, void *hcpu
)
6212 unsigned int cpu
= (long)hcpu
;
6214 switch (action
& ~CPU_TASKS_FROZEN
) {
6216 case CPU_UP_PREPARE
:
6217 case CPU_DOWN_FAILED
:
6218 perf_event_init_cpu(cpu
);
6221 case CPU_UP_CANCELED
:
6222 case CPU_DOWN_PREPARE
:
6223 perf_event_exit_cpu(cpu
);
6233 void __init
perf_event_init(void)
6235 perf_event_init_all_cpus();
6236 init_srcu_struct(&pmus_srcu
);
6237 perf_pmu_register(&perf_swevent
);
6238 perf_pmu_register(&perf_cpu_clock
);
6239 perf_pmu_register(&perf_task_clock
);
6241 perf_cpu_notifier(perf_cpu_notify
);