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 task_struct
*
2019 find_lively_task_by_vpid(pid_t vpid
)
2021 struct task_struct
*task
;
2028 task
= find_task_by_vpid(vpid
);
2030 get_task_struct(task
);
2034 return ERR_PTR(-ESRCH
);
2037 * Can't attach events to a dying task.
2040 if (task
->flags
& PF_EXITING
)
2043 /* Reuse ptrace permission checks for now. */
2045 if (!ptrace_may_access(task
, PTRACE_MODE_READ
))
2050 put_task_struct(task
);
2051 return ERR_PTR(err
);
2055 static struct perf_event_context
*
2056 find_get_context(struct pmu
*pmu
, struct task_struct
*task
, int cpu
)
2058 struct perf_event_context
*ctx
;
2059 struct perf_cpu_context
*cpuctx
;
2060 unsigned long flags
;
2063 if (!task
&& cpu
!= -1) {
2064 /* Must be root to operate on a CPU event: */
2065 if (perf_paranoid_cpu() && !capable(CAP_SYS_ADMIN
))
2066 return ERR_PTR(-EACCES
);
2068 if (cpu
< 0 || cpu
>= nr_cpumask_bits
)
2069 return ERR_PTR(-EINVAL
);
2072 * We could be clever and allow to attach a event to an
2073 * offline CPU and activate it when the CPU comes up, but
2076 if (!cpu_online(cpu
))
2077 return ERR_PTR(-ENODEV
);
2079 cpuctx
= per_cpu_ptr(pmu
->pmu_cpu_context
, cpu
);
2087 ctxn
= pmu
->task_ctx_nr
;
2092 ctx
= perf_lock_task_context(task
, ctxn
, &flags
);
2095 raw_spin_unlock_irqrestore(&ctx
->lock
, flags
);
2099 ctx
= alloc_perf_context(pmu
, task
);
2106 if (cmpxchg(&task
->perf_event_ctxp
[ctxn
], NULL
, ctx
)) {
2108 * We raced with some other task; use
2109 * the context they set.
2111 put_task_struct(task
);
2117 put_task_struct(task
);
2121 put_task_struct(task
);
2122 return ERR_PTR(err
);
2125 static void perf_event_free_filter(struct perf_event
*event
);
2127 static void free_event_rcu(struct rcu_head
*head
)
2129 struct perf_event
*event
;
2131 event
= container_of(head
, struct perf_event
, rcu_head
);
2133 put_pid_ns(event
->ns
);
2134 perf_event_free_filter(event
);
2138 static void perf_pending_sync(struct perf_event
*event
);
2139 static void perf_buffer_put(struct perf_buffer
*buffer
);
2141 static void free_event(struct perf_event
*event
)
2143 perf_pending_sync(event
);
2145 if (!event
->parent
) {
2146 atomic_dec(&nr_events
);
2147 if (event
->attr
.mmap
|| event
->attr
.mmap_data
)
2148 atomic_dec(&nr_mmap_events
);
2149 if (event
->attr
.comm
)
2150 atomic_dec(&nr_comm_events
);
2151 if (event
->attr
.task
)
2152 atomic_dec(&nr_task_events
);
2153 if (event
->attr
.sample_type
& PERF_SAMPLE_CALLCHAIN
)
2154 put_callchain_buffers();
2157 if (event
->buffer
) {
2158 perf_buffer_put(event
->buffer
);
2159 event
->buffer
= NULL
;
2163 event
->destroy(event
);
2166 put_ctx(event
->ctx
);
2168 call_rcu(&event
->rcu_head
, free_event_rcu
);
2171 int perf_event_release_kernel(struct perf_event
*event
)
2173 struct perf_event_context
*ctx
= event
->ctx
;
2176 * Remove from the PMU, can't get re-enabled since we got
2177 * here because the last ref went.
2179 perf_event_disable(event
);
2181 WARN_ON_ONCE(ctx
->parent_ctx
);
2183 * There are two ways this annotation is useful:
2185 * 1) there is a lock recursion from perf_event_exit_task
2186 * see the comment there.
2188 * 2) there is a lock-inversion with mmap_sem through
2189 * perf_event_read_group(), which takes faults while
2190 * holding ctx->mutex, however this is called after
2191 * the last filedesc died, so there is no possibility
2192 * to trigger the AB-BA case.
2194 mutex_lock_nested(&ctx
->mutex
, SINGLE_DEPTH_NESTING
);
2195 raw_spin_lock_irq(&ctx
->lock
);
2196 perf_group_detach(event
);
2197 list_del_event(event
, ctx
);
2198 raw_spin_unlock_irq(&ctx
->lock
);
2199 mutex_unlock(&ctx
->mutex
);
2201 mutex_lock(&event
->owner
->perf_event_mutex
);
2202 list_del_init(&event
->owner_entry
);
2203 mutex_unlock(&event
->owner
->perf_event_mutex
);
2204 put_task_struct(event
->owner
);
2210 EXPORT_SYMBOL_GPL(perf_event_release_kernel
);
2213 * Called when the last reference to the file is gone.
2215 static int perf_release(struct inode
*inode
, struct file
*file
)
2217 struct perf_event
*event
= file
->private_data
;
2219 file
->private_data
= NULL
;
2221 return perf_event_release_kernel(event
);
2224 static int perf_event_read_size(struct perf_event
*event
)
2226 int entry
= sizeof(u64
); /* value */
2230 if (event
->attr
.read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
)
2231 size
+= sizeof(u64
);
2233 if (event
->attr
.read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
)
2234 size
+= sizeof(u64
);
2236 if (event
->attr
.read_format
& PERF_FORMAT_ID
)
2237 entry
+= sizeof(u64
);
2239 if (event
->attr
.read_format
& PERF_FORMAT_GROUP
) {
2240 nr
+= event
->group_leader
->nr_siblings
;
2241 size
+= sizeof(u64
);
2249 u64
perf_event_read_value(struct perf_event
*event
, u64
*enabled
, u64
*running
)
2251 struct perf_event
*child
;
2257 mutex_lock(&event
->child_mutex
);
2258 total
+= perf_event_read(event
);
2259 *enabled
+= event
->total_time_enabled
+
2260 atomic64_read(&event
->child_total_time_enabled
);
2261 *running
+= event
->total_time_running
+
2262 atomic64_read(&event
->child_total_time_running
);
2264 list_for_each_entry(child
, &event
->child_list
, child_list
) {
2265 total
+= perf_event_read(child
);
2266 *enabled
+= child
->total_time_enabled
;
2267 *running
+= child
->total_time_running
;
2269 mutex_unlock(&event
->child_mutex
);
2273 EXPORT_SYMBOL_GPL(perf_event_read_value
);
2275 static int perf_event_read_group(struct perf_event
*event
,
2276 u64 read_format
, char __user
*buf
)
2278 struct perf_event
*leader
= event
->group_leader
, *sub
;
2279 int n
= 0, size
= 0, ret
= -EFAULT
;
2280 struct perf_event_context
*ctx
= leader
->ctx
;
2282 u64 count
, enabled
, running
;
2284 mutex_lock(&ctx
->mutex
);
2285 count
= perf_event_read_value(leader
, &enabled
, &running
);
2287 values
[n
++] = 1 + leader
->nr_siblings
;
2288 if (read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
)
2289 values
[n
++] = enabled
;
2290 if (read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
)
2291 values
[n
++] = running
;
2292 values
[n
++] = count
;
2293 if (read_format
& PERF_FORMAT_ID
)
2294 values
[n
++] = primary_event_id(leader
);
2296 size
= n
* sizeof(u64
);
2298 if (copy_to_user(buf
, values
, size
))
2303 list_for_each_entry(sub
, &leader
->sibling_list
, group_entry
) {
2306 values
[n
++] = perf_event_read_value(sub
, &enabled
, &running
);
2307 if (read_format
& PERF_FORMAT_ID
)
2308 values
[n
++] = primary_event_id(sub
);
2310 size
= n
* sizeof(u64
);
2312 if (copy_to_user(buf
+ ret
, values
, size
)) {
2320 mutex_unlock(&ctx
->mutex
);
2325 static int perf_event_read_one(struct perf_event
*event
,
2326 u64 read_format
, char __user
*buf
)
2328 u64 enabled
, running
;
2332 values
[n
++] = perf_event_read_value(event
, &enabled
, &running
);
2333 if (read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
)
2334 values
[n
++] = enabled
;
2335 if (read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
)
2336 values
[n
++] = running
;
2337 if (read_format
& PERF_FORMAT_ID
)
2338 values
[n
++] = primary_event_id(event
);
2340 if (copy_to_user(buf
, values
, n
* sizeof(u64
)))
2343 return n
* sizeof(u64
);
2347 * Read the performance event - simple non blocking version for now
2350 perf_read_hw(struct perf_event
*event
, char __user
*buf
, size_t count
)
2352 u64 read_format
= event
->attr
.read_format
;
2356 * Return end-of-file for a read on a event that is in
2357 * error state (i.e. because it was pinned but it couldn't be
2358 * scheduled on to the CPU at some point).
2360 if (event
->state
== PERF_EVENT_STATE_ERROR
)
2363 if (count
< perf_event_read_size(event
))
2366 WARN_ON_ONCE(event
->ctx
->parent_ctx
);
2367 if (read_format
& PERF_FORMAT_GROUP
)
2368 ret
= perf_event_read_group(event
, read_format
, buf
);
2370 ret
= perf_event_read_one(event
, read_format
, buf
);
2376 perf_read(struct file
*file
, char __user
*buf
, size_t count
, loff_t
*ppos
)
2378 struct perf_event
*event
= file
->private_data
;
2380 return perf_read_hw(event
, buf
, count
);
2383 static unsigned int perf_poll(struct file
*file
, poll_table
*wait
)
2385 struct perf_event
*event
= file
->private_data
;
2386 struct perf_buffer
*buffer
;
2387 unsigned int events
= POLL_HUP
;
2390 buffer
= rcu_dereference(event
->buffer
);
2392 events
= atomic_xchg(&buffer
->poll
, 0);
2395 poll_wait(file
, &event
->waitq
, wait
);
2400 static void perf_event_reset(struct perf_event
*event
)
2402 (void)perf_event_read(event
);
2403 local64_set(&event
->count
, 0);
2404 perf_event_update_userpage(event
);
2408 * Holding the top-level event's child_mutex means that any
2409 * descendant process that has inherited this event will block
2410 * in sync_child_event if it goes to exit, thus satisfying the
2411 * task existence requirements of perf_event_enable/disable.
2413 static void perf_event_for_each_child(struct perf_event
*event
,
2414 void (*func
)(struct perf_event
*))
2416 struct perf_event
*child
;
2418 WARN_ON_ONCE(event
->ctx
->parent_ctx
);
2419 mutex_lock(&event
->child_mutex
);
2421 list_for_each_entry(child
, &event
->child_list
, child_list
)
2423 mutex_unlock(&event
->child_mutex
);
2426 static void perf_event_for_each(struct perf_event
*event
,
2427 void (*func
)(struct perf_event
*))
2429 struct perf_event_context
*ctx
= event
->ctx
;
2430 struct perf_event
*sibling
;
2432 WARN_ON_ONCE(ctx
->parent_ctx
);
2433 mutex_lock(&ctx
->mutex
);
2434 event
= event
->group_leader
;
2436 perf_event_for_each_child(event
, func
);
2438 list_for_each_entry(sibling
, &event
->sibling_list
, group_entry
)
2439 perf_event_for_each_child(event
, func
);
2440 mutex_unlock(&ctx
->mutex
);
2443 static int perf_event_period(struct perf_event
*event
, u64 __user
*arg
)
2445 struct perf_event_context
*ctx
= event
->ctx
;
2450 if (!event
->attr
.sample_period
)
2453 size
= copy_from_user(&value
, arg
, sizeof(value
));
2454 if (size
!= sizeof(value
))
2460 raw_spin_lock_irq(&ctx
->lock
);
2461 if (event
->attr
.freq
) {
2462 if (value
> sysctl_perf_event_sample_rate
) {
2467 event
->attr
.sample_freq
= value
;
2469 event
->attr
.sample_period
= value
;
2470 event
->hw
.sample_period
= value
;
2473 raw_spin_unlock_irq(&ctx
->lock
);
2478 static const struct file_operations perf_fops
;
2480 static struct perf_event
*perf_fget_light(int fd
, int *fput_needed
)
2484 file
= fget_light(fd
, fput_needed
);
2486 return ERR_PTR(-EBADF
);
2488 if (file
->f_op
!= &perf_fops
) {
2489 fput_light(file
, *fput_needed
);
2491 return ERR_PTR(-EBADF
);
2494 return file
->private_data
;
2497 static int perf_event_set_output(struct perf_event
*event
,
2498 struct perf_event
*output_event
);
2499 static int perf_event_set_filter(struct perf_event
*event
, void __user
*arg
);
2501 static long perf_ioctl(struct file
*file
, unsigned int cmd
, unsigned long arg
)
2503 struct perf_event
*event
= file
->private_data
;
2504 void (*func
)(struct perf_event
*);
2508 case PERF_EVENT_IOC_ENABLE
:
2509 func
= perf_event_enable
;
2511 case PERF_EVENT_IOC_DISABLE
:
2512 func
= perf_event_disable
;
2514 case PERF_EVENT_IOC_RESET
:
2515 func
= perf_event_reset
;
2518 case PERF_EVENT_IOC_REFRESH
:
2519 return perf_event_refresh(event
, arg
);
2521 case PERF_EVENT_IOC_PERIOD
:
2522 return perf_event_period(event
, (u64 __user
*)arg
);
2524 case PERF_EVENT_IOC_SET_OUTPUT
:
2526 struct perf_event
*output_event
= NULL
;
2527 int fput_needed
= 0;
2531 output_event
= perf_fget_light(arg
, &fput_needed
);
2532 if (IS_ERR(output_event
))
2533 return PTR_ERR(output_event
);
2536 ret
= perf_event_set_output(event
, output_event
);
2538 fput_light(output_event
->filp
, fput_needed
);
2543 case PERF_EVENT_IOC_SET_FILTER
:
2544 return perf_event_set_filter(event
, (void __user
*)arg
);
2550 if (flags
& PERF_IOC_FLAG_GROUP
)
2551 perf_event_for_each(event
, func
);
2553 perf_event_for_each_child(event
, func
);
2558 int perf_event_task_enable(void)
2560 struct perf_event
*event
;
2562 mutex_lock(¤t
->perf_event_mutex
);
2563 list_for_each_entry(event
, ¤t
->perf_event_list
, owner_entry
)
2564 perf_event_for_each_child(event
, perf_event_enable
);
2565 mutex_unlock(¤t
->perf_event_mutex
);
2570 int perf_event_task_disable(void)
2572 struct perf_event
*event
;
2574 mutex_lock(¤t
->perf_event_mutex
);
2575 list_for_each_entry(event
, ¤t
->perf_event_list
, owner_entry
)
2576 perf_event_for_each_child(event
, perf_event_disable
);
2577 mutex_unlock(¤t
->perf_event_mutex
);
2582 #ifndef PERF_EVENT_INDEX_OFFSET
2583 # define PERF_EVENT_INDEX_OFFSET 0
2586 static int perf_event_index(struct perf_event
*event
)
2588 if (event
->hw
.state
& PERF_HES_STOPPED
)
2591 if (event
->state
!= PERF_EVENT_STATE_ACTIVE
)
2594 return event
->hw
.idx
+ 1 - PERF_EVENT_INDEX_OFFSET
;
2598 * Callers need to ensure there can be no nesting of this function, otherwise
2599 * the seqlock logic goes bad. We can not serialize this because the arch
2600 * code calls this from NMI context.
2602 void perf_event_update_userpage(struct perf_event
*event
)
2604 struct perf_event_mmap_page
*userpg
;
2605 struct perf_buffer
*buffer
;
2608 buffer
= rcu_dereference(event
->buffer
);
2612 userpg
= buffer
->user_page
;
2615 * Disable preemption so as to not let the corresponding user-space
2616 * spin too long if we get preempted.
2621 userpg
->index
= perf_event_index(event
);
2622 userpg
->offset
= perf_event_count(event
);
2623 if (event
->state
== PERF_EVENT_STATE_ACTIVE
)
2624 userpg
->offset
-= local64_read(&event
->hw
.prev_count
);
2626 userpg
->time_enabled
= event
->total_time_enabled
+
2627 atomic64_read(&event
->child_total_time_enabled
);
2629 userpg
->time_running
= event
->total_time_running
+
2630 atomic64_read(&event
->child_total_time_running
);
2639 static unsigned long perf_data_size(struct perf_buffer
*buffer
);
2642 perf_buffer_init(struct perf_buffer
*buffer
, long watermark
, int flags
)
2644 long max_size
= perf_data_size(buffer
);
2647 buffer
->watermark
= min(max_size
, watermark
);
2649 if (!buffer
->watermark
)
2650 buffer
->watermark
= max_size
/ 2;
2652 if (flags
& PERF_BUFFER_WRITABLE
)
2653 buffer
->writable
= 1;
2655 atomic_set(&buffer
->refcount
, 1);
2658 #ifndef CONFIG_PERF_USE_VMALLOC
2661 * Back perf_mmap() with regular GFP_KERNEL-0 pages.
2664 static struct page
*
2665 perf_mmap_to_page(struct perf_buffer
*buffer
, unsigned long pgoff
)
2667 if (pgoff
> buffer
->nr_pages
)
2671 return virt_to_page(buffer
->user_page
);
2673 return virt_to_page(buffer
->data_pages
[pgoff
- 1]);
2676 static void *perf_mmap_alloc_page(int cpu
)
2681 node
= (cpu
== -1) ? cpu
: cpu_to_node(cpu
);
2682 page
= alloc_pages_node(node
, GFP_KERNEL
| __GFP_ZERO
, 0);
2686 return page_address(page
);
2689 static struct perf_buffer
*
2690 perf_buffer_alloc(int nr_pages
, long watermark
, int cpu
, int flags
)
2692 struct perf_buffer
*buffer
;
2696 size
= sizeof(struct perf_buffer
);
2697 size
+= nr_pages
* sizeof(void *);
2699 buffer
= kzalloc(size
, GFP_KERNEL
);
2703 buffer
->user_page
= perf_mmap_alloc_page(cpu
);
2704 if (!buffer
->user_page
)
2705 goto fail_user_page
;
2707 for (i
= 0; i
< nr_pages
; i
++) {
2708 buffer
->data_pages
[i
] = perf_mmap_alloc_page(cpu
);
2709 if (!buffer
->data_pages
[i
])
2710 goto fail_data_pages
;
2713 buffer
->nr_pages
= nr_pages
;
2715 perf_buffer_init(buffer
, watermark
, flags
);
2720 for (i
--; i
>= 0; i
--)
2721 free_page((unsigned long)buffer
->data_pages
[i
]);
2723 free_page((unsigned long)buffer
->user_page
);
2732 static void perf_mmap_free_page(unsigned long addr
)
2734 struct page
*page
= virt_to_page((void *)addr
);
2736 page
->mapping
= NULL
;
2740 static void perf_buffer_free(struct perf_buffer
*buffer
)
2744 perf_mmap_free_page((unsigned long)buffer
->user_page
);
2745 for (i
= 0; i
< buffer
->nr_pages
; i
++)
2746 perf_mmap_free_page((unsigned long)buffer
->data_pages
[i
]);
2750 static inline int page_order(struct perf_buffer
*buffer
)
2758 * Back perf_mmap() with vmalloc memory.
2760 * Required for architectures that have d-cache aliasing issues.
2763 static inline int page_order(struct perf_buffer
*buffer
)
2765 return buffer
->page_order
;
2768 static struct page
*
2769 perf_mmap_to_page(struct perf_buffer
*buffer
, unsigned long pgoff
)
2771 if (pgoff
> (1UL << page_order(buffer
)))
2774 return vmalloc_to_page((void *)buffer
->user_page
+ pgoff
* PAGE_SIZE
);
2777 static void perf_mmap_unmark_page(void *addr
)
2779 struct page
*page
= vmalloc_to_page(addr
);
2781 page
->mapping
= NULL
;
2784 static void perf_buffer_free_work(struct work_struct
*work
)
2786 struct perf_buffer
*buffer
;
2790 buffer
= container_of(work
, struct perf_buffer
, work
);
2791 nr
= 1 << page_order(buffer
);
2793 base
= buffer
->user_page
;
2794 for (i
= 0; i
< nr
+ 1; i
++)
2795 perf_mmap_unmark_page(base
+ (i
* PAGE_SIZE
));
2801 static void perf_buffer_free(struct perf_buffer
*buffer
)
2803 schedule_work(&buffer
->work
);
2806 static struct perf_buffer
*
2807 perf_buffer_alloc(int nr_pages
, long watermark
, int cpu
, int flags
)
2809 struct perf_buffer
*buffer
;
2813 size
= sizeof(struct perf_buffer
);
2814 size
+= sizeof(void *);
2816 buffer
= kzalloc(size
, GFP_KERNEL
);
2820 INIT_WORK(&buffer
->work
, perf_buffer_free_work
);
2822 all_buf
= vmalloc_user((nr_pages
+ 1) * PAGE_SIZE
);
2826 buffer
->user_page
= all_buf
;
2827 buffer
->data_pages
[0] = all_buf
+ PAGE_SIZE
;
2828 buffer
->page_order
= ilog2(nr_pages
);
2829 buffer
->nr_pages
= 1;
2831 perf_buffer_init(buffer
, watermark
, flags
);
2844 static unsigned long perf_data_size(struct perf_buffer
*buffer
)
2846 return buffer
->nr_pages
<< (PAGE_SHIFT
+ page_order(buffer
));
2849 static int perf_mmap_fault(struct vm_area_struct
*vma
, struct vm_fault
*vmf
)
2851 struct perf_event
*event
= vma
->vm_file
->private_data
;
2852 struct perf_buffer
*buffer
;
2853 int ret
= VM_FAULT_SIGBUS
;
2855 if (vmf
->flags
& FAULT_FLAG_MKWRITE
) {
2856 if (vmf
->pgoff
== 0)
2862 buffer
= rcu_dereference(event
->buffer
);
2866 if (vmf
->pgoff
&& (vmf
->flags
& FAULT_FLAG_WRITE
))
2869 vmf
->page
= perf_mmap_to_page(buffer
, vmf
->pgoff
);
2873 get_page(vmf
->page
);
2874 vmf
->page
->mapping
= vma
->vm_file
->f_mapping
;
2875 vmf
->page
->index
= vmf
->pgoff
;
2884 static void perf_buffer_free_rcu(struct rcu_head
*rcu_head
)
2886 struct perf_buffer
*buffer
;
2888 buffer
= container_of(rcu_head
, struct perf_buffer
, rcu_head
);
2889 perf_buffer_free(buffer
);
2892 static struct perf_buffer
*perf_buffer_get(struct perf_event
*event
)
2894 struct perf_buffer
*buffer
;
2897 buffer
= rcu_dereference(event
->buffer
);
2899 if (!atomic_inc_not_zero(&buffer
->refcount
))
2907 static void perf_buffer_put(struct perf_buffer
*buffer
)
2909 if (!atomic_dec_and_test(&buffer
->refcount
))
2912 call_rcu(&buffer
->rcu_head
, perf_buffer_free_rcu
);
2915 static void perf_mmap_open(struct vm_area_struct
*vma
)
2917 struct perf_event
*event
= vma
->vm_file
->private_data
;
2919 atomic_inc(&event
->mmap_count
);
2922 static void perf_mmap_close(struct vm_area_struct
*vma
)
2924 struct perf_event
*event
= vma
->vm_file
->private_data
;
2926 if (atomic_dec_and_mutex_lock(&event
->mmap_count
, &event
->mmap_mutex
)) {
2927 unsigned long size
= perf_data_size(event
->buffer
);
2928 struct user_struct
*user
= event
->mmap_user
;
2929 struct perf_buffer
*buffer
= event
->buffer
;
2931 atomic_long_sub((size
>> PAGE_SHIFT
) + 1, &user
->locked_vm
);
2932 vma
->vm_mm
->locked_vm
-= event
->mmap_locked
;
2933 rcu_assign_pointer(event
->buffer
, NULL
);
2934 mutex_unlock(&event
->mmap_mutex
);
2936 perf_buffer_put(buffer
);
2941 static const struct vm_operations_struct perf_mmap_vmops
= {
2942 .open
= perf_mmap_open
,
2943 .close
= perf_mmap_close
,
2944 .fault
= perf_mmap_fault
,
2945 .page_mkwrite
= perf_mmap_fault
,
2948 static int perf_mmap(struct file
*file
, struct vm_area_struct
*vma
)
2950 struct perf_event
*event
= file
->private_data
;
2951 unsigned long user_locked
, user_lock_limit
;
2952 struct user_struct
*user
= current_user();
2953 unsigned long locked
, lock_limit
;
2954 struct perf_buffer
*buffer
;
2955 unsigned long vma_size
;
2956 unsigned long nr_pages
;
2957 long user_extra
, extra
;
2958 int ret
= 0, flags
= 0;
2961 * Don't allow mmap() of inherited per-task counters. This would
2962 * create a performance issue due to all children writing to the
2965 if (event
->cpu
== -1 && event
->attr
.inherit
)
2968 if (!(vma
->vm_flags
& VM_SHARED
))
2971 vma_size
= vma
->vm_end
- vma
->vm_start
;
2972 nr_pages
= (vma_size
/ PAGE_SIZE
) - 1;
2975 * If we have buffer pages ensure they're a power-of-two number, so we
2976 * can do bitmasks instead of modulo.
2978 if (nr_pages
!= 0 && !is_power_of_2(nr_pages
))
2981 if (vma_size
!= PAGE_SIZE
* (1 + nr_pages
))
2984 if (vma
->vm_pgoff
!= 0)
2987 WARN_ON_ONCE(event
->ctx
->parent_ctx
);
2988 mutex_lock(&event
->mmap_mutex
);
2989 if (event
->buffer
) {
2990 if (event
->buffer
->nr_pages
== nr_pages
)
2991 atomic_inc(&event
->buffer
->refcount
);
2997 user_extra
= nr_pages
+ 1;
2998 user_lock_limit
= sysctl_perf_event_mlock
>> (PAGE_SHIFT
- 10);
3001 * Increase the limit linearly with more CPUs:
3003 user_lock_limit
*= num_online_cpus();
3005 user_locked
= atomic_long_read(&user
->locked_vm
) + user_extra
;
3008 if (user_locked
> user_lock_limit
)
3009 extra
= user_locked
- user_lock_limit
;
3011 lock_limit
= rlimit(RLIMIT_MEMLOCK
);
3012 lock_limit
>>= PAGE_SHIFT
;
3013 locked
= vma
->vm_mm
->locked_vm
+ extra
;
3015 if ((locked
> lock_limit
) && perf_paranoid_tracepoint_raw() &&
3016 !capable(CAP_IPC_LOCK
)) {
3021 WARN_ON(event
->buffer
);
3023 if (vma
->vm_flags
& VM_WRITE
)
3024 flags
|= PERF_BUFFER_WRITABLE
;
3026 buffer
= perf_buffer_alloc(nr_pages
, event
->attr
.wakeup_watermark
,
3032 rcu_assign_pointer(event
->buffer
, buffer
);
3034 atomic_long_add(user_extra
, &user
->locked_vm
);
3035 event
->mmap_locked
= extra
;
3036 event
->mmap_user
= get_current_user();
3037 vma
->vm_mm
->locked_vm
+= event
->mmap_locked
;
3041 atomic_inc(&event
->mmap_count
);
3042 mutex_unlock(&event
->mmap_mutex
);
3044 vma
->vm_flags
|= VM_RESERVED
;
3045 vma
->vm_ops
= &perf_mmap_vmops
;
3050 static int perf_fasync(int fd
, struct file
*filp
, int on
)
3052 struct inode
*inode
= filp
->f_path
.dentry
->d_inode
;
3053 struct perf_event
*event
= filp
->private_data
;
3056 mutex_lock(&inode
->i_mutex
);
3057 retval
= fasync_helper(fd
, filp
, on
, &event
->fasync
);
3058 mutex_unlock(&inode
->i_mutex
);
3066 static const struct file_operations perf_fops
= {
3067 .llseek
= no_llseek
,
3068 .release
= perf_release
,
3071 .unlocked_ioctl
= perf_ioctl
,
3072 .compat_ioctl
= perf_ioctl
,
3074 .fasync
= perf_fasync
,
3080 * If there's data, ensure we set the poll() state and publish everything
3081 * to user-space before waking everybody up.
3084 void perf_event_wakeup(struct perf_event
*event
)
3086 wake_up_all(&event
->waitq
);
3088 if (event
->pending_kill
) {
3089 kill_fasync(&event
->fasync
, SIGIO
, event
->pending_kill
);
3090 event
->pending_kill
= 0;
3097 * Handle the case where we need to wakeup up from NMI (or rq->lock) context.
3099 * The NMI bit means we cannot possibly take locks. Therefore, maintain a
3100 * single linked list and use cmpxchg() to add entries lockless.
3103 static void perf_pending_event(struct perf_pending_entry
*entry
)
3105 struct perf_event
*event
= container_of(entry
,
3106 struct perf_event
, pending
);
3108 if (event
->pending_disable
) {
3109 event
->pending_disable
= 0;
3110 __perf_event_disable(event
);
3113 if (event
->pending_wakeup
) {
3114 event
->pending_wakeup
= 0;
3115 perf_event_wakeup(event
);
3119 #define PENDING_TAIL ((struct perf_pending_entry *)-1UL)
3121 static DEFINE_PER_CPU(struct perf_pending_entry
*, perf_pending_head
) = {
3125 static void perf_pending_queue(struct perf_pending_entry
*entry
,
3126 void (*func
)(struct perf_pending_entry
*))
3128 struct perf_pending_entry
**head
;
3130 if (cmpxchg(&entry
->next
, NULL
, PENDING_TAIL
) != NULL
)
3135 head
= &get_cpu_var(perf_pending_head
);
3138 entry
->next
= *head
;
3139 } while (cmpxchg(head
, entry
->next
, entry
) != entry
->next
);
3141 set_perf_event_pending();
3143 put_cpu_var(perf_pending_head
);
3146 static int __perf_pending_run(void)
3148 struct perf_pending_entry
*list
;
3151 list
= xchg(&__get_cpu_var(perf_pending_head
), PENDING_TAIL
);
3152 while (list
!= PENDING_TAIL
) {
3153 void (*func
)(struct perf_pending_entry
*);
3154 struct perf_pending_entry
*entry
= list
;
3161 * Ensure we observe the unqueue before we issue the wakeup,
3162 * so that we won't be waiting forever.
3163 * -- see perf_not_pending().
3174 static inline int perf_not_pending(struct perf_event
*event
)
3177 * If we flush on whatever cpu we run, there is a chance we don't
3181 __perf_pending_run();
3185 * Ensure we see the proper queue state before going to sleep
3186 * so that we do not miss the wakeup. -- see perf_pending_handle()
3189 return event
->pending
.next
== NULL
;
3192 static void perf_pending_sync(struct perf_event
*event
)
3194 wait_event(event
->waitq
, perf_not_pending(event
));
3197 void perf_event_do_pending(void)
3199 __perf_pending_run();
3203 * We assume there is only KVM supporting the callbacks.
3204 * Later on, we might change it to a list if there is
3205 * another virtualization implementation supporting the callbacks.
3207 struct perf_guest_info_callbacks
*perf_guest_cbs
;
3209 int perf_register_guest_info_callbacks(struct perf_guest_info_callbacks
*cbs
)
3211 perf_guest_cbs
= cbs
;
3214 EXPORT_SYMBOL_GPL(perf_register_guest_info_callbacks
);
3216 int perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks
*cbs
)
3218 perf_guest_cbs
= NULL
;
3221 EXPORT_SYMBOL_GPL(perf_unregister_guest_info_callbacks
);
3226 static bool perf_output_space(struct perf_buffer
*buffer
, unsigned long tail
,
3227 unsigned long offset
, unsigned long head
)
3231 if (!buffer
->writable
)
3234 mask
= perf_data_size(buffer
) - 1;
3236 offset
= (offset
- tail
) & mask
;
3237 head
= (head
- tail
) & mask
;
3239 if ((int)(head
- offset
) < 0)
3245 static void perf_output_wakeup(struct perf_output_handle
*handle
)
3247 atomic_set(&handle
->buffer
->poll
, POLL_IN
);
3250 handle
->event
->pending_wakeup
= 1;
3251 perf_pending_queue(&handle
->event
->pending
,
3252 perf_pending_event
);
3254 perf_event_wakeup(handle
->event
);
3258 * We need to ensure a later event_id doesn't publish a head when a former
3259 * event isn't done writing. However since we need to deal with NMIs we
3260 * cannot fully serialize things.
3262 * We only publish the head (and generate a wakeup) when the outer-most
3265 static void perf_output_get_handle(struct perf_output_handle
*handle
)
3267 struct perf_buffer
*buffer
= handle
->buffer
;
3270 local_inc(&buffer
->nest
);
3271 handle
->wakeup
= local_read(&buffer
->wakeup
);
3274 static void perf_output_put_handle(struct perf_output_handle
*handle
)
3276 struct perf_buffer
*buffer
= handle
->buffer
;
3280 head
= local_read(&buffer
->head
);
3283 * IRQ/NMI can happen here, which means we can miss a head update.
3286 if (!local_dec_and_test(&buffer
->nest
))
3290 * Publish the known good head. Rely on the full barrier implied
3291 * by atomic_dec_and_test() order the buffer->head read and this
3294 buffer
->user_page
->data_head
= head
;
3297 * Now check if we missed an update, rely on the (compiler)
3298 * barrier in atomic_dec_and_test() to re-read buffer->head.
3300 if (unlikely(head
!= local_read(&buffer
->head
))) {
3301 local_inc(&buffer
->nest
);
3305 if (handle
->wakeup
!= local_read(&buffer
->wakeup
))
3306 perf_output_wakeup(handle
);
3312 __always_inline
void perf_output_copy(struct perf_output_handle
*handle
,
3313 const void *buf
, unsigned int len
)
3316 unsigned long size
= min_t(unsigned long, handle
->size
, len
);
3318 memcpy(handle
->addr
, buf
, size
);
3321 handle
->addr
+= size
;
3323 handle
->size
-= size
;
3324 if (!handle
->size
) {
3325 struct perf_buffer
*buffer
= handle
->buffer
;
3328 handle
->page
&= buffer
->nr_pages
- 1;
3329 handle
->addr
= buffer
->data_pages
[handle
->page
];
3330 handle
->size
= PAGE_SIZE
<< page_order(buffer
);
3335 int perf_output_begin(struct perf_output_handle
*handle
,
3336 struct perf_event
*event
, unsigned int size
,
3337 int nmi
, int sample
)
3339 struct perf_buffer
*buffer
;
3340 unsigned long tail
, offset
, head
;
3343 struct perf_event_header header
;
3350 * For inherited events we send all the output towards the parent.
3353 event
= event
->parent
;
3355 buffer
= rcu_dereference(event
->buffer
);
3359 handle
->buffer
= buffer
;
3360 handle
->event
= event
;
3362 handle
->sample
= sample
;
3364 if (!buffer
->nr_pages
)
3367 have_lost
= local_read(&buffer
->lost
);
3369 size
+= sizeof(lost_event
);
3371 perf_output_get_handle(handle
);
3375 * Userspace could choose to issue a mb() before updating the
3376 * tail pointer. So that all reads will be completed before the
3379 tail
= ACCESS_ONCE(buffer
->user_page
->data_tail
);
3381 offset
= head
= local_read(&buffer
->head
);
3383 if (unlikely(!perf_output_space(buffer
, tail
, offset
, head
)))
3385 } while (local_cmpxchg(&buffer
->head
, offset
, head
) != offset
);
3387 if (head
- local_read(&buffer
->wakeup
) > buffer
->watermark
)
3388 local_add(buffer
->watermark
, &buffer
->wakeup
);
3390 handle
->page
= offset
>> (PAGE_SHIFT
+ page_order(buffer
));
3391 handle
->page
&= buffer
->nr_pages
- 1;
3392 handle
->size
= offset
& ((PAGE_SIZE
<< page_order(buffer
)) - 1);
3393 handle
->addr
= buffer
->data_pages
[handle
->page
];
3394 handle
->addr
+= handle
->size
;
3395 handle
->size
= (PAGE_SIZE
<< page_order(buffer
)) - handle
->size
;
3398 lost_event
.header
.type
= PERF_RECORD_LOST
;
3399 lost_event
.header
.misc
= 0;
3400 lost_event
.header
.size
= sizeof(lost_event
);
3401 lost_event
.id
= event
->id
;
3402 lost_event
.lost
= local_xchg(&buffer
->lost
, 0);
3404 perf_output_put(handle
, lost_event
);
3410 local_inc(&buffer
->lost
);
3411 perf_output_put_handle(handle
);
3418 void perf_output_end(struct perf_output_handle
*handle
)
3420 struct perf_event
*event
= handle
->event
;
3421 struct perf_buffer
*buffer
= handle
->buffer
;
3423 int wakeup_events
= event
->attr
.wakeup_events
;
3425 if (handle
->sample
&& wakeup_events
) {
3426 int events
= local_inc_return(&buffer
->events
);
3427 if (events
>= wakeup_events
) {
3428 local_sub(wakeup_events
, &buffer
->events
);
3429 local_inc(&buffer
->wakeup
);
3433 perf_output_put_handle(handle
);
3437 static u32
perf_event_pid(struct perf_event
*event
, struct task_struct
*p
)
3440 * only top level events have the pid namespace they were created in
3443 event
= event
->parent
;
3445 return task_tgid_nr_ns(p
, event
->ns
);
3448 static u32
perf_event_tid(struct perf_event
*event
, struct task_struct
*p
)
3451 * only top level events have the pid namespace they were created in
3454 event
= event
->parent
;
3456 return task_pid_nr_ns(p
, event
->ns
);
3459 static void perf_output_read_one(struct perf_output_handle
*handle
,
3460 struct perf_event
*event
)
3462 u64 read_format
= event
->attr
.read_format
;
3466 values
[n
++] = perf_event_count(event
);
3467 if (read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
) {
3468 values
[n
++] = event
->total_time_enabled
+
3469 atomic64_read(&event
->child_total_time_enabled
);
3471 if (read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
) {
3472 values
[n
++] = event
->total_time_running
+
3473 atomic64_read(&event
->child_total_time_running
);
3475 if (read_format
& PERF_FORMAT_ID
)
3476 values
[n
++] = primary_event_id(event
);
3478 perf_output_copy(handle
, values
, n
* sizeof(u64
));
3482 * XXX PERF_FORMAT_GROUP vs inherited events seems difficult.
3484 static void perf_output_read_group(struct perf_output_handle
*handle
,
3485 struct perf_event
*event
)
3487 struct perf_event
*leader
= event
->group_leader
, *sub
;
3488 u64 read_format
= event
->attr
.read_format
;
3492 values
[n
++] = 1 + leader
->nr_siblings
;
3494 if (read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
)
3495 values
[n
++] = leader
->total_time_enabled
;
3497 if (read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
)
3498 values
[n
++] = leader
->total_time_running
;
3500 if (leader
!= event
)
3501 leader
->pmu
->read(leader
);
3503 values
[n
++] = perf_event_count(leader
);
3504 if (read_format
& PERF_FORMAT_ID
)
3505 values
[n
++] = primary_event_id(leader
);
3507 perf_output_copy(handle
, values
, n
* sizeof(u64
));
3509 list_for_each_entry(sub
, &leader
->sibling_list
, group_entry
) {
3513 sub
->pmu
->read(sub
);
3515 values
[n
++] = perf_event_count(sub
);
3516 if (read_format
& PERF_FORMAT_ID
)
3517 values
[n
++] = primary_event_id(sub
);
3519 perf_output_copy(handle
, values
, n
* sizeof(u64
));
3523 static void perf_output_read(struct perf_output_handle
*handle
,
3524 struct perf_event
*event
)
3526 if (event
->attr
.read_format
& PERF_FORMAT_GROUP
)
3527 perf_output_read_group(handle
, event
);
3529 perf_output_read_one(handle
, event
);
3532 void perf_output_sample(struct perf_output_handle
*handle
,
3533 struct perf_event_header
*header
,
3534 struct perf_sample_data
*data
,
3535 struct perf_event
*event
)
3537 u64 sample_type
= data
->type
;
3539 perf_output_put(handle
, *header
);
3541 if (sample_type
& PERF_SAMPLE_IP
)
3542 perf_output_put(handle
, data
->ip
);
3544 if (sample_type
& PERF_SAMPLE_TID
)
3545 perf_output_put(handle
, data
->tid_entry
);
3547 if (sample_type
& PERF_SAMPLE_TIME
)
3548 perf_output_put(handle
, data
->time
);
3550 if (sample_type
& PERF_SAMPLE_ADDR
)
3551 perf_output_put(handle
, data
->addr
);
3553 if (sample_type
& PERF_SAMPLE_ID
)
3554 perf_output_put(handle
, data
->id
);
3556 if (sample_type
& PERF_SAMPLE_STREAM_ID
)
3557 perf_output_put(handle
, data
->stream_id
);
3559 if (sample_type
& PERF_SAMPLE_CPU
)
3560 perf_output_put(handle
, data
->cpu_entry
);
3562 if (sample_type
& PERF_SAMPLE_PERIOD
)
3563 perf_output_put(handle
, data
->period
);
3565 if (sample_type
& PERF_SAMPLE_READ
)
3566 perf_output_read(handle
, event
);
3568 if (sample_type
& PERF_SAMPLE_CALLCHAIN
) {
3569 if (data
->callchain
) {
3572 if (data
->callchain
)
3573 size
+= data
->callchain
->nr
;
3575 size
*= sizeof(u64
);
3577 perf_output_copy(handle
, data
->callchain
, size
);
3580 perf_output_put(handle
, nr
);
3584 if (sample_type
& PERF_SAMPLE_RAW
) {
3586 perf_output_put(handle
, data
->raw
->size
);
3587 perf_output_copy(handle
, data
->raw
->data
,
3594 .size
= sizeof(u32
),
3597 perf_output_put(handle
, raw
);
3602 void perf_prepare_sample(struct perf_event_header
*header
,
3603 struct perf_sample_data
*data
,
3604 struct perf_event
*event
,
3605 struct pt_regs
*regs
)
3607 u64 sample_type
= event
->attr
.sample_type
;
3609 data
->type
= sample_type
;
3611 header
->type
= PERF_RECORD_SAMPLE
;
3612 header
->size
= sizeof(*header
);
3615 header
->misc
|= perf_misc_flags(regs
);
3617 if (sample_type
& PERF_SAMPLE_IP
) {
3618 data
->ip
= perf_instruction_pointer(regs
);
3620 header
->size
+= sizeof(data
->ip
);
3623 if (sample_type
& PERF_SAMPLE_TID
) {
3624 /* namespace issues */
3625 data
->tid_entry
.pid
= perf_event_pid(event
, current
);
3626 data
->tid_entry
.tid
= perf_event_tid(event
, current
);
3628 header
->size
+= sizeof(data
->tid_entry
);
3631 if (sample_type
& PERF_SAMPLE_TIME
) {
3632 data
->time
= perf_clock();
3634 header
->size
+= sizeof(data
->time
);
3637 if (sample_type
& PERF_SAMPLE_ADDR
)
3638 header
->size
+= sizeof(data
->addr
);
3640 if (sample_type
& PERF_SAMPLE_ID
) {
3641 data
->id
= primary_event_id(event
);
3643 header
->size
+= sizeof(data
->id
);
3646 if (sample_type
& PERF_SAMPLE_STREAM_ID
) {
3647 data
->stream_id
= event
->id
;
3649 header
->size
+= sizeof(data
->stream_id
);
3652 if (sample_type
& PERF_SAMPLE_CPU
) {
3653 data
->cpu_entry
.cpu
= raw_smp_processor_id();
3654 data
->cpu_entry
.reserved
= 0;
3656 header
->size
+= sizeof(data
->cpu_entry
);
3659 if (sample_type
& PERF_SAMPLE_PERIOD
)
3660 header
->size
+= sizeof(data
->period
);
3662 if (sample_type
& PERF_SAMPLE_READ
)
3663 header
->size
+= perf_event_read_size(event
);
3665 if (sample_type
& PERF_SAMPLE_CALLCHAIN
) {
3668 data
->callchain
= perf_callchain(regs
);
3670 if (data
->callchain
)
3671 size
+= data
->callchain
->nr
;
3673 header
->size
+= size
* sizeof(u64
);
3676 if (sample_type
& PERF_SAMPLE_RAW
) {
3677 int size
= sizeof(u32
);
3680 size
+= data
->raw
->size
;
3682 size
+= sizeof(u32
);
3684 WARN_ON_ONCE(size
& (sizeof(u64
)-1));
3685 header
->size
+= size
;
3689 static void perf_event_output(struct perf_event
*event
, int nmi
,
3690 struct perf_sample_data
*data
,
3691 struct pt_regs
*regs
)
3693 struct perf_output_handle handle
;
3694 struct perf_event_header header
;
3696 /* protect the callchain buffers */
3699 perf_prepare_sample(&header
, data
, event
, regs
);
3701 if (perf_output_begin(&handle
, event
, header
.size
, nmi
, 1))
3704 perf_output_sample(&handle
, &header
, data
, event
);
3706 perf_output_end(&handle
);
3716 struct perf_read_event
{
3717 struct perf_event_header header
;
3724 perf_event_read_event(struct perf_event
*event
,
3725 struct task_struct
*task
)
3727 struct perf_output_handle handle
;
3728 struct perf_read_event read_event
= {
3730 .type
= PERF_RECORD_READ
,
3732 .size
= sizeof(read_event
) + perf_event_read_size(event
),
3734 .pid
= perf_event_pid(event
, task
),
3735 .tid
= perf_event_tid(event
, task
),
3739 ret
= perf_output_begin(&handle
, event
, read_event
.header
.size
, 0, 0);
3743 perf_output_put(&handle
, read_event
);
3744 perf_output_read(&handle
, event
);
3746 perf_output_end(&handle
);
3750 * task tracking -- fork/exit
3752 * enabled by: attr.comm | attr.mmap | attr.mmap_data | attr.task
3755 struct perf_task_event
{
3756 struct task_struct
*task
;
3757 struct perf_event_context
*task_ctx
;
3760 struct perf_event_header header
;
3770 static void perf_event_task_output(struct perf_event
*event
,
3771 struct perf_task_event
*task_event
)
3773 struct perf_output_handle handle
;
3774 struct task_struct
*task
= task_event
->task
;
3777 size
= task_event
->event_id
.header
.size
;
3778 ret
= perf_output_begin(&handle
, event
, size
, 0, 0);
3783 task_event
->event_id
.pid
= perf_event_pid(event
, task
);
3784 task_event
->event_id
.ppid
= perf_event_pid(event
, current
);
3786 task_event
->event_id
.tid
= perf_event_tid(event
, task
);
3787 task_event
->event_id
.ptid
= perf_event_tid(event
, current
);
3789 perf_output_put(&handle
, task_event
->event_id
);
3791 perf_output_end(&handle
);
3794 static int perf_event_task_match(struct perf_event
*event
)
3796 if (event
->state
< PERF_EVENT_STATE_INACTIVE
)
3799 if (event
->cpu
!= -1 && event
->cpu
!= smp_processor_id())
3802 if (event
->attr
.comm
|| event
->attr
.mmap
||
3803 event
->attr
.mmap_data
|| event
->attr
.task
)
3809 static void perf_event_task_ctx(struct perf_event_context
*ctx
,
3810 struct perf_task_event
*task_event
)
3812 struct perf_event
*event
;
3814 list_for_each_entry_rcu(event
, &ctx
->event_list
, event_entry
) {
3815 if (perf_event_task_match(event
))
3816 perf_event_task_output(event
, task_event
);
3820 static void perf_event_task_event(struct perf_task_event
*task_event
)
3822 struct perf_cpu_context
*cpuctx
;
3823 struct perf_event_context
*ctx
;
3828 list_for_each_entry_rcu(pmu
, &pmus
, entry
) {
3829 cpuctx
= this_cpu_ptr(pmu
->pmu_cpu_context
);
3830 perf_event_task_ctx(&cpuctx
->ctx
, task_event
);
3832 ctx
= task_event
->task_ctx
;
3834 ctxn
= pmu
->task_ctx_nr
;
3837 ctx
= rcu_dereference(current
->perf_event_ctxp
[ctxn
]);
3840 perf_event_task_ctx(ctx
, task_event
);
3845 static void perf_event_task(struct task_struct
*task
,
3846 struct perf_event_context
*task_ctx
,
3849 struct perf_task_event task_event
;
3851 if (!atomic_read(&nr_comm_events
) &&
3852 !atomic_read(&nr_mmap_events
) &&
3853 !atomic_read(&nr_task_events
))
3856 task_event
= (struct perf_task_event
){
3858 .task_ctx
= task_ctx
,
3861 .type
= new ? PERF_RECORD_FORK
: PERF_RECORD_EXIT
,
3863 .size
= sizeof(task_event
.event_id
),
3869 .time
= perf_clock(),
3873 perf_event_task_event(&task_event
);
3876 void perf_event_fork(struct task_struct
*task
)
3878 perf_event_task(task
, NULL
, 1);
3885 struct perf_comm_event
{
3886 struct task_struct
*task
;
3891 struct perf_event_header header
;
3898 static void perf_event_comm_output(struct perf_event
*event
,
3899 struct perf_comm_event
*comm_event
)
3901 struct perf_output_handle handle
;
3902 int size
= comm_event
->event_id
.header
.size
;
3903 int ret
= perf_output_begin(&handle
, event
, size
, 0, 0);
3908 comm_event
->event_id
.pid
= perf_event_pid(event
, comm_event
->task
);
3909 comm_event
->event_id
.tid
= perf_event_tid(event
, comm_event
->task
);
3911 perf_output_put(&handle
, comm_event
->event_id
);
3912 perf_output_copy(&handle
, comm_event
->comm
,
3913 comm_event
->comm_size
);
3914 perf_output_end(&handle
);
3917 static int perf_event_comm_match(struct perf_event
*event
)
3919 if (event
->state
< PERF_EVENT_STATE_INACTIVE
)
3922 if (event
->cpu
!= -1 && event
->cpu
!= smp_processor_id())
3925 if (event
->attr
.comm
)
3931 static void perf_event_comm_ctx(struct perf_event_context
*ctx
,
3932 struct perf_comm_event
*comm_event
)
3934 struct perf_event
*event
;
3936 list_for_each_entry_rcu(event
, &ctx
->event_list
, event_entry
) {
3937 if (perf_event_comm_match(event
))
3938 perf_event_comm_output(event
, comm_event
);
3942 static void perf_event_comm_event(struct perf_comm_event
*comm_event
)
3944 struct perf_cpu_context
*cpuctx
;
3945 struct perf_event_context
*ctx
;
3946 char comm
[TASK_COMM_LEN
];
3951 memset(comm
, 0, sizeof(comm
));
3952 strlcpy(comm
, comm_event
->task
->comm
, sizeof(comm
));
3953 size
= ALIGN(strlen(comm
)+1, sizeof(u64
));
3955 comm_event
->comm
= comm
;
3956 comm_event
->comm_size
= size
;
3958 comm_event
->event_id
.header
.size
= sizeof(comm_event
->event_id
) + size
;
3961 list_for_each_entry_rcu(pmu
, &pmus
, entry
) {
3962 cpuctx
= this_cpu_ptr(pmu
->pmu_cpu_context
);
3963 perf_event_comm_ctx(&cpuctx
->ctx
, comm_event
);
3965 ctxn
= pmu
->task_ctx_nr
;
3969 ctx
= rcu_dereference(current
->perf_event_ctxp
[ctxn
]);
3971 perf_event_comm_ctx(ctx
, comm_event
);
3976 void perf_event_comm(struct task_struct
*task
)
3978 struct perf_comm_event comm_event
;
3979 struct perf_event_context
*ctx
;
3982 for_each_task_context_nr(ctxn
) {
3983 ctx
= task
->perf_event_ctxp
[ctxn
];
3987 perf_event_enable_on_exec(ctx
);
3990 if (!atomic_read(&nr_comm_events
))
3993 comm_event
= (struct perf_comm_event
){
3999 .type
= PERF_RECORD_COMM
,
4008 perf_event_comm_event(&comm_event
);
4015 struct perf_mmap_event
{
4016 struct vm_area_struct
*vma
;
4018 const char *file_name
;
4022 struct perf_event_header header
;
4032 static void perf_event_mmap_output(struct perf_event
*event
,
4033 struct perf_mmap_event
*mmap_event
)
4035 struct perf_output_handle handle
;
4036 int size
= mmap_event
->event_id
.header
.size
;
4037 int ret
= perf_output_begin(&handle
, event
, size
, 0, 0);
4042 mmap_event
->event_id
.pid
= perf_event_pid(event
, current
);
4043 mmap_event
->event_id
.tid
= perf_event_tid(event
, current
);
4045 perf_output_put(&handle
, mmap_event
->event_id
);
4046 perf_output_copy(&handle
, mmap_event
->file_name
,
4047 mmap_event
->file_size
);
4048 perf_output_end(&handle
);
4051 static int perf_event_mmap_match(struct perf_event
*event
,
4052 struct perf_mmap_event
*mmap_event
,
4055 if (event
->state
< PERF_EVENT_STATE_INACTIVE
)
4058 if (event
->cpu
!= -1 && event
->cpu
!= smp_processor_id())
4061 if ((!executable
&& event
->attr
.mmap_data
) ||
4062 (executable
&& event
->attr
.mmap
))
4068 static void perf_event_mmap_ctx(struct perf_event_context
*ctx
,
4069 struct perf_mmap_event
*mmap_event
,
4072 struct perf_event
*event
;
4074 list_for_each_entry_rcu(event
, &ctx
->event_list
, event_entry
) {
4075 if (perf_event_mmap_match(event
, mmap_event
, executable
))
4076 perf_event_mmap_output(event
, mmap_event
);
4080 static void perf_event_mmap_event(struct perf_mmap_event
*mmap_event
)
4082 struct perf_cpu_context
*cpuctx
;
4083 struct perf_event_context
*ctx
;
4084 struct vm_area_struct
*vma
= mmap_event
->vma
;
4085 struct file
*file
= vma
->vm_file
;
4093 memset(tmp
, 0, sizeof(tmp
));
4097 * d_path works from the end of the buffer backwards, so we
4098 * need to add enough zero bytes after the string to handle
4099 * the 64bit alignment we do later.
4101 buf
= kzalloc(PATH_MAX
+ sizeof(u64
), GFP_KERNEL
);
4103 name
= strncpy(tmp
, "//enomem", sizeof(tmp
));
4106 name
= d_path(&file
->f_path
, buf
, PATH_MAX
);
4108 name
= strncpy(tmp
, "//toolong", sizeof(tmp
));
4112 if (arch_vma_name(mmap_event
->vma
)) {
4113 name
= strncpy(tmp
, arch_vma_name(mmap_event
->vma
),
4119 name
= strncpy(tmp
, "[vdso]", sizeof(tmp
));
4121 } else if (vma
->vm_start
<= vma
->vm_mm
->start_brk
&&
4122 vma
->vm_end
>= vma
->vm_mm
->brk
) {
4123 name
= strncpy(tmp
, "[heap]", sizeof(tmp
));
4125 } else if (vma
->vm_start
<= vma
->vm_mm
->start_stack
&&
4126 vma
->vm_end
>= vma
->vm_mm
->start_stack
) {
4127 name
= strncpy(tmp
, "[stack]", sizeof(tmp
));
4131 name
= strncpy(tmp
, "//anon", sizeof(tmp
));
4136 size
= ALIGN(strlen(name
)+1, sizeof(u64
));
4138 mmap_event
->file_name
= name
;
4139 mmap_event
->file_size
= size
;
4141 mmap_event
->event_id
.header
.size
= sizeof(mmap_event
->event_id
) + size
;
4144 list_for_each_entry_rcu(pmu
, &pmus
, entry
) {
4145 cpuctx
= this_cpu_ptr(pmu
->pmu_cpu_context
);
4146 perf_event_mmap_ctx(&cpuctx
->ctx
, mmap_event
,
4147 vma
->vm_flags
& VM_EXEC
);
4149 ctxn
= pmu
->task_ctx_nr
;
4153 ctx
= rcu_dereference(current
->perf_event_ctxp
[ctxn
]);
4155 perf_event_mmap_ctx(ctx
, mmap_event
,
4156 vma
->vm_flags
& VM_EXEC
);
4164 void perf_event_mmap(struct vm_area_struct
*vma
)
4166 struct perf_mmap_event mmap_event
;
4168 if (!atomic_read(&nr_mmap_events
))
4171 mmap_event
= (struct perf_mmap_event
){
4177 .type
= PERF_RECORD_MMAP
,
4178 .misc
= PERF_RECORD_MISC_USER
,
4183 .start
= vma
->vm_start
,
4184 .len
= vma
->vm_end
- vma
->vm_start
,
4185 .pgoff
= (u64
)vma
->vm_pgoff
<< PAGE_SHIFT
,
4189 perf_event_mmap_event(&mmap_event
);
4193 * IRQ throttle logging
4196 static void perf_log_throttle(struct perf_event
*event
, int enable
)
4198 struct perf_output_handle handle
;
4202 struct perf_event_header header
;
4206 } throttle_event
= {
4208 .type
= PERF_RECORD_THROTTLE
,
4210 .size
= sizeof(throttle_event
),
4212 .time
= perf_clock(),
4213 .id
= primary_event_id(event
),
4214 .stream_id
= event
->id
,
4218 throttle_event
.header
.type
= PERF_RECORD_UNTHROTTLE
;
4220 ret
= perf_output_begin(&handle
, event
, sizeof(throttle_event
), 1, 0);
4224 perf_output_put(&handle
, throttle_event
);
4225 perf_output_end(&handle
);
4229 * Generic event overflow handling, sampling.
4232 static int __perf_event_overflow(struct perf_event
*event
, int nmi
,
4233 int throttle
, struct perf_sample_data
*data
,
4234 struct pt_regs
*regs
)
4236 int events
= atomic_read(&event
->event_limit
);
4237 struct hw_perf_event
*hwc
= &event
->hw
;
4243 if (hwc
->interrupts
!= MAX_INTERRUPTS
) {
4245 if (HZ
* hwc
->interrupts
>
4246 (u64
)sysctl_perf_event_sample_rate
) {
4247 hwc
->interrupts
= MAX_INTERRUPTS
;
4248 perf_log_throttle(event
, 0);
4253 * Keep re-disabling events even though on the previous
4254 * pass we disabled it - just in case we raced with a
4255 * sched-in and the event got enabled again:
4261 if (event
->attr
.freq
) {
4262 u64 now
= perf_clock();
4263 s64 delta
= now
- hwc
->freq_time_stamp
;
4265 hwc
->freq_time_stamp
= now
;
4267 if (delta
> 0 && delta
< 2*TICK_NSEC
)
4268 perf_adjust_period(event
, delta
, hwc
->last_period
);
4272 * XXX event_limit might not quite work as expected on inherited
4276 event
->pending_kill
= POLL_IN
;
4277 if (events
&& atomic_dec_and_test(&event
->event_limit
)) {
4279 event
->pending_kill
= POLL_HUP
;
4281 event
->pending_disable
= 1;
4282 perf_pending_queue(&event
->pending
,
4283 perf_pending_event
);
4285 perf_event_disable(event
);
4288 if (event
->overflow_handler
)
4289 event
->overflow_handler(event
, nmi
, data
, regs
);
4291 perf_event_output(event
, nmi
, data
, regs
);
4296 int perf_event_overflow(struct perf_event
*event
, int nmi
,
4297 struct perf_sample_data
*data
,
4298 struct pt_regs
*regs
)
4300 return __perf_event_overflow(event
, nmi
, 1, data
, regs
);
4304 * Generic software event infrastructure
4307 struct swevent_htable
{
4308 struct swevent_hlist
*swevent_hlist
;
4309 struct mutex hlist_mutex
;
4312 /* Recursion avoidance in each contexts */
4313 int recursion
[PERF_NR_CONTEXTS
];
4316 static DEFINE_PER_CPU(struct swevent_htable
, swevent_htable
);
4319 * We directly increment event->count and keep a second value in
4320 * event->hw.period_left to count intervals. This period event
4321 * is kept in the range [-sample_period, 0] so that we can use the
4325 static u64
perf_swevent_set_period(struct perf_event
*event
)
4327 struct hw_perf_event
*hwc
= &event
->hw
;
4328 u64 period
= hwc
->last_period
;
4332 hwc
->last_period
= hwc
->sample_period
;
4335 old
= val
= local64_read(&hwc
->period_left
);
4339 nr
= div64_u64(period
+ val
, period
);
4340 offset
= nr
* period
;
4342 if (local64_cmpxchg(&hwc
->period_left
, old
, val
) != old
)
4348 static void perf_swevent_overflow(struct perf_event
*event
, u64 overflow
,
4349 int nmi
, struct perf_sample_data
*data
,
4350 struct pt_regs
*regs
)
4352 struct hw_perf_event
*hwc
= &event
->hw
;
4355 data
->period
= event
->hw
.last_period
;
4357 overflow
= perf_swevent_set_period(event
);
4359 if (hwc
->interrupts
== MAX_INTERRUPTS
)
4362 for (; overflow
; overflow
--) {
4363 if (__perf_event_overflow(event
, nmi
, throttle
,
4366 * We inhibit the overflow from happening when
4367 * hwc->interrupts == MAX_INTERRUPTS.
4375 static void perf_swevent_event(struct perf_event
*event
, u64 nr
,
4376 int nmi
, struct perf_sample_data
*data
,
4377 struct pt_regs
*regs
)
4379 struct hw_perf_event
*hwc
= &event
->hw
;
4381 local64_add(nr
, &event
->count
);
4386 if (!hwc
->sample_period
)
4389 if (nr
== 1 && hwc
->sample_period
== 1 && !event
->attr
.freq
)
4390 return perf_swevent_overflow(event
, 1, nmi
, data
, regs
);
4392 if (local64_add_negative(nr
, &hwc
->period_left
))
4395 perf_swevent_overflow(event
, 0, nmi
, data
, regs
);
4398 static int perf_exclude_event(struct perf_event
*event
,
4399 struct pt_regs
*regs
)
4401 if (event
->hw
.state
& PERF_HES_STOPPED
)
4405 if (event
->attr
.exclude_user
&& user_mode(regs
))
4408 if (event
->attr
.exclude_kernel
&& !user_mode(regs
))
4415 static int perf_swevent_match(struct perf_event
*event
,
4416 enum perf_type_id type
,
4418 struct perf_sample_data
*data
,
4419 struct pt_regs
*regs
)
4421 if (event
->attr
.type
!= type
)
4424 if (event
->attr
.config
!= event_id
)
4427 if (perf_exclude_event(event
, regs
))
4433 static inline u64
swevent_hash(u64 type
, u32 event_id
)
4435 u64 val
= event_id
| (type
<< 32);
4437 return hash_64(val
, SWEVENT_HLIST_BITS
);
4440 static inline struct hlist_head
*
4441 __find_swevent_head(struct swevent_hlist
*hlist
, u64 type
, u32 event_id
)
4443 u64 hash
= swevent_hash(type
, event_id
);
4445 return &hlist
->heads
[hash
];
4448 /* For the read side: events when they trigger */
4449 static inline struct hlist_head
*
4450 find_swevent_head_rcu(struct swevent_htable
*swhash
, u64 type
, u32 event_id
)
4452 struct swevent_hlist
*hlist
;
4454 hlist
= rcu_dereference(swhash
->swevent_hlist
);
4458 return __find_swevent_head(hlist
, type
, event_id
);
4461 /* For the event head insertion and removal in the hlist */
4462 static inline struct hlist_head
*
4463 find_swevent_head(struct swevent_htable
*swhash
, struct perf_event
*event
)
4465 struct swevent_hlist
*hlist
;
4466 u32 event_id
= event
->attr
.config
;
4467 u64 type
= event
->attr
.type
;
4470 * Event scheduling is always serialized against hlist allocation
4471 * and release. Which makes the protected version suitable here.
4472 * The context lock guarantees that.
4474 hlist
= rcu_dereference_protected(swhash
->swevent_hlist
,
4475 lockdep_is_held(&event
->ctx
->lock
));
4479 return __find_swevent_head(hlist
, type
, event_id
);
4482 static void do_perf_sw_event(enum perf_type_id type
, u32 event_id
,
4484 struct perf_sample_data
*data
,
4485 struct pt_regs
*regs
)
4487 struct swevent_htable
*swhash
= &__get_cpu_var(swevent_htable
);
4488 struct perf_event
*event
;
4489 struct hlist_node
*node
;
4490 struct hlist_head
*head
;
4493 head
= find_swevent_head_rcu(swhash
, type
, event_id
);
4497 hlist_for_each_entry_rcu(event
, node
, head
, hlist_entry
) {
4498 if (perf_swevent_match(event
, type
, event_id
, data
, regs
))
4499 perf_swevent_event(event
, nr
, nmi
, data
, regs
);
4505 int perf_swevent_get_recursion_context(void)
4507 struct swevent_htable
*swhash
= &__get_cpu_var(swevent_htable
);
4509 return get_recursion_context(swhash
->recursion
);
4511 EXPORT_SYMBOL_GPL(perf_swevent_get_recursion_context
);
4513 void inline perf_swevent_put_recursion_context(int rctx
)
4515 struct swevent_htable
*swhash
= &__get_cpu_var(swevent_htable
);
4517 put_recursion_context(swhash
->recursion
, rctx
);
4520 void __perf_sw_event(u32 event_id
, u64 nr
, int nmi
,
4521 struct pt_regs
*regs
, u64 addr
)
4523 struct perf_sample_data data
;
4526 preempt_disable_notrace();
4527 rctx
= perf_swevent_get_recursion_context();
4531 perf_sample_data_init(&data
, addr
);
4533 do_perf_sw_event(PERF_TYPE_SOFTWARE
, event_id
, nr
, nmi
, &data
, regs
);
4535 perf_swevent_put_recursion_context(rctx
);
4536 preempt_enable_notrace();
4539 static void perf_swevent_read(struct perf_event
*event
)
4543 static int perf_swevent_add(struct perf_event
*event
, int flags
)
4545 struct swevent_htable
*swhash
= &__get_cpu_var(swevent_htable
);
4546 struct hw_perf_event
*hwc
= &event
->hw
;
4547 struct hlist_head
*head
;
4549 if (hwc
->sample_period
) {
4550 hwc
->last_period
= hwc
->sample_period
;
4551 perf_swevent_set_period(event
);
4554 hwc
->state
= !(flags
& PERF_EF_START
);
4556 head
= find_swevent_head(swhash
, event
);
4557 if (WARN_ON_ONCE(!head
))
4560 hlist_add_head_rcu(&event
->hlist_entry
, head
);
4565 static void perf_swevent_del(struct perf_event
*event
, int flags
)
4567 hlist_del_rcu(&event
->hlist_entry
);
4570 static void perf_swevent_start(struct perf_event
*event
, int flags
)
4572 event
->hw
.state
= 0;
4575 static void perf_swevent_stop(struct perf_event
*event
, int flags
)
4577 event
->hw
.state
= PERF_HES_STOPPED
;
4580 /* Deref the hlist from the update side */
4581 static inline struct swevent_hlist
*
4582 swevent_hlist_deref(struct swevent_htable
*swhash
)
4584 return rcu_dereference_protected(swhash
->swevent_hlist
,
4585 lockdep_is_held(&swhash
->hlist_mutex
));
4588 static void swevent_hlist_release_rcu(struct rcu_head
*rcu_head
)
4590 struct swevent_hlist
*hlist
;
4592 hlist
= container_of(rcu_head
, struct swevent_hlist
, rcu_head
);
4596 static void swevent_hlist_release(struct swevent_htable
*swhash
)
4598 struct swevent_hlist
*hlist
= swevent_hlist_deref(swhash
);
4603 rcu_assign_pointer(swhash
->swevent_hlist
, NULL
);
4604 call_rcu(&hlist
->rcu_head
, swevent_hlist_release_rcu
);
4607 static void swevent_hlist_put_cpu(struct perf_event
*event
, int cpu
)
4609 struct swevent_htable
*swhash
= &per_cpu(swevent_htable
, cpu
);
4611 mutex_lock(&swhash
->hlist_mutex
);
4613 if (!--swhash
->hlist_refcount
)
4614 swevent_hlist_release(swhash
);
4616 mutex_unlock(&swhash
->hlist_mutex
);
4619 static void swevent_hlist_put(struct perf_event
*event
)
4623 if (event
->cpu
!= -1) {
4624 swevent_hlist_put_cpu(event
, event
->cpu
);
4628 for_each_possible_cpu(cpu
)
4629 swevent_hlist_put_cpu(event
, cpu
);
4632 static int swevent_hlist_get_cpu(struct perf_event
*event
, int cpu
)
4634 struct swevent_htable
*swhash
= &per_cpu(swevent_htable
, cpu
);
4637 mutex_lock(&swhash
->hlist_mutex
);
4639 if (!swevent_hlist_deref(swhash
) && cpu_online(cpu
)) {
4640 struct swevent_hlist
*hlist
;
4642 hlist
= kzalloc(sizeof(*hlist
), GFP_KERNEL
);
4647 rcu_assign_pointer(swhash
->swevent_hlist
, hlist
);
4649 swhash
->hlist_refcount
++;
4651 mutex_unlock(&swhash
->hlist_mutex
);
4656 static int swevent_hlist_get(struct perf_event
*event
)
4659 int cpu
, failed_cpu
;
4661 if (event
->cpu
!= -1)
4662 return swevent_hlist_get_cpu(event
, event
->cpu
);
4665 for_each_possible_cpu(cpu
) {
4666 err
= swevent_hlist_get_cpu(event
, cpu
);
4676 for_each_possible_cpu(cpu
) {
4677 if (cpu
== failed_cpu
)
4679 swevent_hlist_put_cpu(event
, cpu
);
4686 atomic_t perf_swevent_enabled
[PERF_COUNT_SW_MAX
];
4688 static void sw_perf_event_destroy(struct perf_event
*event
)
4690 u64 event_id
= event
->attr
.config
;
4692 WARN_ON(event
->parent
);
4694 atomic_dec(&perf_swevent_enabled
[event_id
]);
4695 swevent_hlist_put(event
);
4698 static int perf_swevent_init(struct perf_event
*event
)
4700 int event_id
= event
->attr
.config
;
4702 if (event
->attr
.type
!= PERF_TYPE_SOFTWARE
)
4706 case PERF_COUNT_SW_CPU_CLOCK
:
4707 case PERF_COUNT_SW_TASK_CLOCK
:
4714 if (event_id
> PERF_COUNT_SW_MAX
)
4717 if (!event
->parent
) {
4720 err
= swevent_hlist_get(event
);
4724 atomic_inc(&perf_swevent_enabled
[event_id
]);
4725 event
->destroy
= sw_perf_event_destroy
;
4731 static struct pmu perf_swevent
= {
4732 .task_ctx_nr
= perf_sw_context
,
4734 .event_init
= perf_swevent_init
,
4735 .add
= perf_swevent_add
,
4736 .del
= perf_swevent_del
,
4737 .start
= perf_swevent_start
,
4738 .stop
= perf_swevent_stop
,
4739 .read
= perf_swevent_read
,
4742 #ifdef CONFIG_EVENT_TRACING
4744 static int perf_tp_filter_match(struct perf_event
*event
,
4745 struct perf_sample_data
*data
)
4747 void *record
= data
->raw
->data
;
4749 if (likely(!event
->filter
) || filter_match_preds(event
->filter
, record
))
4754 static int perf_tp_event_match(struct perf_event
*event
,
4755 struct perf_sample_data
*data
,
4756 struct pt_regs
*regs
)
4759 * All tracepoints are from kernel-space.
4761 if (event
->attr
.exclude_kernel
)
4764 if (!perf_tp_filter_match(event
, data
))
4770 void perf_tp_event(u64 addr
, u64 count
, void *record
, int entry_size
,
4771 struct pt_regs
*regs
, struct hlist_head
*head
, int rctx
)
4773 struct perf_sample_data data
;
4774 struct perf_event
*event
;
4775 struct hlist_node
*node
;
4777 struct perf_raw_record raw
= {
4782 perf_sample_data_init(&data
, addr
);
4785 hlist_for_each_entry_rcu(event
, node
, head
, hlist_entry
) {
4786 if (perf_tp_event_match(event
, &data
, regs
))
4787 perf_swevent_event(event
, count
, 1, &data
, regs
);
4790 perf_swevent_put_recursion_context(rctx
);
4792 EXPORT_SYMBOL_GPL(perf_tp_event
);
4794 static void tp_perf_event_destroy(struct perf_event
*event
)
4796 perf_trace_destroy(event
);
4799 static int perf_tp_event_init(struct perf_event
*event
)
4803 if (event
->attr
.type
!= PERF_TYPE_TRACEPOINT
)
4807 * Raw tracepoint data is a severe data leak, only allow root to
4810 if ((event
->attr
.sample_type
& PERF_SAMPLE_RAW
) &&
4811 perf_paranoid_tracepoint_raw() &&
4812 !capable(CAP_SYS_ADMIN
))
4815 err
= perf_trace_init(event
);
4819 event
->destroy
= tp_perf_event_destroy
;
4824 static struct pmu perf_tracepoint
= {
4825 .task_ctx_nr
= perf_sw_context
,
4827 .event_init
= perf_tp_event_init
,
4828 .add
= perf_trace_add
,
4829 .del
= perf_trace_del
,
4830 .start
= perf_swevent_start
,
4831 .stop
= perf_swevent_stop
,
4832 .read
= perf_swevent_read
,
4835 static inline void perf_tp_register(void)
4837 perf_pmu_register(&perf_tracepoint
);
4840 static int perf_event_set_filter(struct perf_event
*event
, void __user
*arg
)
4845 if (event
->attr
.type
!= PERF_TYPE_TRACEPOINT
)
4848 filter_str
= strndup_user(arg
, PAGE_SIZE
);
4849 if (IS_ERR(filter_str
))
4850 return PTR_ERR(filter_str
);
4852 ret
= ftrace_profile_set_filter(event
, event
->attr
.config
, filter_str
);
4858 static void perf_event_free_filter(struct perf_event
*event
)
4860 ftrace_profile_free_filter(event
);
4865 static inline void perf_tp_register(void)
4869 static int perf_event_set_filter(struct perf_event
*event
, void __user
*arg
)
4874 static void perf_event_free_filter(struct perf_event
*event
)
4878 #endif /* CONFIG_EVENT_TRACING */
4880 #ifdef CONFIG_HAVE_HW_BREAKPOINT
4881 void perf_bp_event(struct perf_event
*bp
, void *data
)
4883 struct perf_sample_data sample
;
4884 struct pt_regs
*regs
= data
;
4886 perf_sample_data_init(&sample
, bp
->attr
.bp_addr
);
4888 if (!bp
->hw
.state
&& !perf_exclude_event(bp
, regs
))
4889 perf_swevent_event(bp
, 1, 1, &sample
, regs
);
4894 * hrtimer based swevent callback
4897 static enum hrtimer_restart
perf_swevent_hrtimer(struct hrtimer
*hrtimer
)
4899 enum hrtimer_restart ret
= HRTIMER_RESTART
;
4900 struct perf_sample_data data
;
4901 struct pt_regs
*regs
;
4902 struct perf_event
*event
;
4905 event
= container_of(hrtimer
, struct perf_event
, hw
.hrtimer
);
4906 event
->pmu
->read(event
);
4908 perf_sample_data_init(&data
, 0);
4909 data
.period
= event
->hw
.last_period
;
4910 regs
= get_irq_regs();
4912 if (regs
&& !perf_exclude_event(event
, regs
)) {
4913 if (!(event
->attr
.exclude_idle
&& current
->pid
== 0))
4914 if (perf_event_overflow(event
, 0, &data
, regs
))
4915 ret
= HRTIMER_NORESTART
;
4918 period
= max_t(u64
, 10000, event
->hw
.sample_period
);
4919 hrtimer_forward_now(hrtimer
, ns_to_ktime(period
));
4924 static void perf_swevent_start_hrtimer(struct perf_event
*event
)
4926 struct hw_perf_event
*hwc
= &event
->hw
;
4928 hrtimer_init(&hwc
->hrtimer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
4929 hwc
->hrtimer
.function
= perf_swevent_hrtimer
;
4930 if (hwc
->sample_period
) {
4931 s64 period
= local64_read(&hwc
->period_left
);
4937 local64_set(&hwc
->period_left
, 0);
4939 period
= max_t(u64
, 10000, hwc
->sample_period
);
4941 __hrtimer_start_range_ns(&hwc
->hrtimer
,
4942 ns_to_ktime(period
), 0,
4943 HRTIMER_MODE_REL_PINNED
, 0);
4947 static void perf_swevent_cancel_hrtimer(struct perf_event
*event
)
4949 struct hw_perf_event
*hwc
= &event
->hw
;
4951 if (hwc
->sample_period
) {
4952 ktime_t remaining
= hrtimer_get_remaining(&hwc
->hrtimer
);
4953 local64_set(&hwc
->period_left
, ktime_to_ns(remaining
));
4955 hrtimer_cancel(&hwc
->hrtimer
);
4960 * Software event: cpu wall time clock
4963 static void cpu_clock_event_update(struct perf_event
*event
)
4968 now
= local_clock();
4969 prev
= local64_xchg(&event
->hw
.prev_count
, now
);
4970 local64_add(now
- prev
, &event
->count
);
4973 static void cpu_clock_event_start(struct perf_event
*event
, int flags
)
4975 local64_set(&event
->hw
.prev_count
, local_clock());
4976 perf_swevent_start_hrtimer(event
);
4979 static void cpu_clock_event_stop(struct perf_event
*event
, int flags
)
4981 perf_swevent_cancel_hrtimer(event
);
4982 cpu_clock_event_update(event
);
4985 static int cpu_clock_event_add(struct perf_event
*event
, int flags
)
4987 if (flags
& PERF_EF_START
)
4988 cpu_clock_event_start(event
, flags
);
4993 static void cpu_clock_event_del(struct perf_event
*event
, int flags
)
4995 cpu_clock_event_stop(event
, flags
);
4998 static void cpu_clock_event_read(struct perf_event
*event
)
5000 cpu_clock_event_update(event
);
5003 static int cpu_clock_event_init(struct perf_event
*event
)
5005 if (event
->attr
.type
!= PERF_TYPE_SOFTWARE
)
5008 if (event
->attr
.config
!= PERF_COUNT_SW_CPU_CLOCK
)
5014 static struct pmu perf_cpu_clock
= {
5015 .task_ctx_nr
= perf_sw_context
,
5017 .event_init
= cpu_clock_event_init
,
5018 .add
= cpu_clock_event_add
,
5019 .del
= cpu_clock_event_del
,
5020 .start
= cpu_clock_event_start
,
5021 .stop
= cpu_clock_event_stop
,
5022 .read
= cpu_clock_event_read
,
5026 * Software event: task time clock
5029 static void task_clock_event_update(struct perf_event
*event
, u64 now
)
5034 prev
= local64_xchg(&event
->hw
.prev_count
, now
);
5036 local64_add(delta
, &event
->count
);
5039 static void task_clock_event_start(struct perf_event
*event
, int flags
)
5041 local64_set(&event
->hw
.prev_count
, event
->ctx
->time
);
5042 perf_swevent_start_hrtimer(event
);
5045 static void task_clock_event_stop(struct perf_event
*event
, int flags
)
5047 perf_swevent_cancel_hrtimer(event
);
5048 task_clock_event_update(event
, event
->ctx
->time
);
5051 static int task_clock_event_add(struct perf_event
*event
, int flags
)
5053 if (flags
& PERF_EF_START
)
5054 task_clock_event_start(event
, flags
);
5059 static void task_clock_event_del(struct perf_event
*event
, int flags
)
5061 task_clock_event_stop(event
, PERF_EF_UPDATE
);
5064 static void task_clock_event_read(struct perf_event
*event
)
5069 update_context_time(event
->ctx
);
5070 time
= event
->ctx
->time
;
5072 u64 now
= perf_clock();
5073 u64 delta
= now
- event
->ctx
->timestamp
;
5074 time
= event
->ctx
->time
+ delta
;
5077 task_clock_event_update(event
, time
);
5080 static int task_clock_event_init(struct perf_event
*event
)
5082 if (event
->attr
.type
!= PERF_TYPE_SOFTWARE
)
5085 if (event
->attr
.config
!= PERF_COUNT_SW_TASK_CLOCK
)
5091 static struct pmu perf_task_clock
= {
5092 .task_ctx_nr
= perf_sw_context
,
5094 .event_init
= task_clock_event_init
,
5095 .add
= task_clock_event_add
,
5096 .del
= task_clock_event_del
,
5097 .start
= task_clock_event_start
,
5098 .stop
= task_clock_event_stop
,
5099 .read
= task_clock_event_read
,
5102 static void perf_pmu_nop_void(struct pmu
*pmu
)
5106 static int perf_pmu_nop_int(struct pmu
*pmu
)
5111 static void perf_pmu_start_txn(struct pmu
*pmu
)
5113 perf_pmu_disable(pmu
);
5116 static int perf_pmu_commit_txn(struct pmu
*pmu
)
5118 perf_pmu_enable(pmu
);
5122 static void perf_pmu_cancel_txn(struct pmu
*pmu
)
5124 perf_pmu_enable(pmu
);
5128 * Ensures all contexts with the same task_ctx_nr have the same
5129 * pmu_cpu_context too.
5131 static void *find_pmu_context(int ctxn
)
5138 list_for_each_entry(pmu
, &pmus
, entry
) {
5139 if (pmu
->task_ctx_nr
== ctxn
)
5140 return pmu
->pmu_cpu_context
;
5146 static void free_pmu_context(void * __percpu cpu_context
)
5150 mutex_lock(&pmus_lock
);
5152 * Like a real lame refcount.
5154 list_for_each_entry(pmu
, &pmus
, entry
) {
5155 if (pmu
->pmu_cpu_context
== cpu_context
)
5159 free_percpu(cpu_context
);
5161 mutex_unlock(&pmus_lock
);
5164 int perf_pmu_register(struct pmu
*pmu
)
5168 mutex_lock(&pmus_lock
);
5170 pmu
->pmu_disable_count
= alloc_percpu(int);
5171 if (!pmu
->pmu_disable_count
)
5174 pmu
->pmu_cpu_context
= find_pmu_context(pmu
->task_ctx_nr
);
5175 if (pmu
->pmu_cpu_context
)
5176 goto got_cpu_context
;
5178 pmu
->pmu_cpu_context
= alloc_percpu(struct perf_cpu_context
);
5179 if (!pmu
->pmu_cpu_context
)
5182 for_each_possible_cpu(cpu
) {
5183 struct perf_cpu_context
*cpuctx
;
5185 cpuctx
= per_cpu_ptr(pmu
->pmu_cpu_context
, cpu
);
5186 __perf_event_init_context(&cpuctx
->ctx
);
5187 cpuctx
->ctx
.type
= cpu_context
;
5188 cpuctx
->ctx
.pmu
= pmu
;
5189 cpuctx
->timer_interval
= TICK_NSEC
;
5190 hrtimer_init(&cpuctx
->timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
5191 cpuctx
->timer
.function
= perf_event_context_tick
;
5195 if (!pmu
->start_txn
) {
5196 if (pmu
->pmu_enable
) {
5198 * If we have pmu_enable/pmu_disable calls, install
5199 * transaction stubs that use that to try and batch
5200 * hardware accesses.
5202 pmu
->start_txn
= perf_pmu_start_txn
;
5203 pmu
->commit_txn
= perf_pmu_commit_txn
;
5204 pmu
->cancel_txn
= perf_pmu_cancel_txn
;
5206 pmu
->start_txn
= perf_pmu_nop_void
;
5207 pmu
->commit_txn
= perf_pmu_nop_int
;
5208 pmu
->cancel_txn
= perf_pmu_nop_void
;
5212 if (!pmu
->pmu_enable
) {
5213 pmu
->pmu_enable
= perf_pmu_nop_void
;
5214 pmu
->pmu_disable
= perf_pmu_nop_void
;
5217 list_add_rcu(&pmu
->entry
, &pmus
);
5220 mutex_unlock(&pmus_lock
);
5225 free_percpu(pmu
->pmu_disable_count
);
5229 void perf_pmu_unregister(struct pmu
*pmu
)
5231 mutex_lock(&pmus_lock
);
5232 list_del_rcu(&pmu
->entry
);
5233 mutex_unlock(&pmus_lock
);
5236 * We dereference the pmu list under both SRCU and regular RCU, so
5237 * synchronize against both of those.
5239 synchronize_srcu(&pmus_srcu
);
5242 free_percpu(pmu
->pmu_disable_count
);
5243 free_pmu_context(pmu
->pmu_cpu_context
);
5246 struct pmu
*perf_init_event(struct perf_event
*event
)
5248 struct pmu
*pmu
= NULL
;
5251 idx
= srcu_read_lock(&pmus_srcu
);
5252 list_for_each_entry_rcu(pmu
, &pmus
, entry
) {
5253 int ret
= pmu
->event_init(event
);
5257 if (ret
!= -ENOENT
) {
5262 pmu
= ERR_PTR(-ENOENT
);
5264 srcu_read_unlock(&pmus_srcu
, idx
);
5270 * Allocate and initialize a event structure
5272 static struct perf_event
*
5273 perf_event_alloc(struct perf_event_attr
*attr
, int cpu
,
5274 struct perf_event
*group_leader
,
5275 struct perf_event
*parent_event
,
5276 perf_overflow_handler_t overflow_handler
)
5279 struct perf_event
*event
;
5280 struct hw_perf_event
*hwc
;
5283 event
= kzalloc(sizeof(*event
), GFP_KERNEL
);
5285 return ERR_PTR(-ENOMEM
);
5288 * Single events are their own group leaders, with an
5289 * empty sibling list:
5292 group_leader
= event
;
5294 mutex_init(&event
->child_mutex
);
5295 INIT_LIST_HEAD(&event
->child_list
);
5297 INIT_LIST_HEAD(&event
->group_entry
);
5298 INIT_LIST_HEAD(&event
->event_entry
);
5299 INIT_LIST_HEAD(&event
->sibling_list
);
5300 init_waitqueue_head(&event
->waitq
);
5302 mutex_init(&event
->mmap_mutex
);
5305 event
->attr
= *attr
;
5306 event
->group_leader
= group_leader
;
5310 event
->parent
= parent_event
;
5312 event
->ns
= get_pid_ns(current
->nsproxy
->pid_ns
);
5313 event
->id
= atomic64_inc_return(&perf_event_id
);
5315 event
->state
= PERF_EVENT_STATE_INACTIVE
;
5317 if (!overflow_handler
&& parent_event
)
5318 overflow_handler
= parent_event
->overflow_handler
;
5320 event
->overflow_handler
= overflow_handler
;
5323 event
->state
= PERF_EVENT_STATE_OFF
;
5328 hwc
->sample_period
= attr
->sample_period
;
5329 if (attr
->freq
&& attr
->sample_freq
)
5330 hwc
->sample_period
= 1;
5331 hwc
->last_period
= hwc
->sample_period
;
5333 local64_set(&hwc
->period_left
, hwc
->sample_period
);
5336 * we currently do not support PERF_FORMAT_GROUP on inherited events
5338 if (attr
->inherit
&& (attr
->read_format
& PERF_FORMAT_GROUP
))
5341 pmu
= perf_init_event(event
);
5347 else if (IS_ERR(pmu
))
5352 put_pid_ns(event
->ns
);
5354 return ERR_PTR(err
);
5359 if (!event
->parent
) {
5360 atomic_inc(&nr_events
);
5361 if (event
->attr
.mmap
|| event
->attr
.mmap_data
)
5362 atomic_inc(&nr_mmap_events
);
5363 if (event
->attr
.comm
)
5364 atomic_inc(&nr_comm_events
);
5365 if (event
->attr
.task
)
5366 atomic_inc(&nr_task_events
);
5367 if (event
->attr
.sample_type
& PERF_SAMPLE_CALLCHAIN
) {
5368 err
= get_callchain_buffers();
5371 return ERR_PTR(err
);
5379 static int perf_copy_attr(struct perf_event_attr __user
*uattr
,
5380 struct perf_event_attr
*attr
)
5385 if (!access_ok(VERIFY_WRITE
, uattr
, PERF_ATTR_SIZE_VER0
))
5389 * zero the full structure, so that a short copy will be nice.
5391 memset(attr
, 0, sizeof(*attr
));
5393 ret
= get_user(size
, &uattr
->size
);
5397 if (size
> PAGE_SIZE
) /* silly large */
5400 if (!size
) /* abi compat */
5401 size
= PERF_ATTR_SIZE_VER0
;
5403 if (size
< PERF_ATTR_SIZE_VER0
)
5407 * If we're handed a bigger struct than we know of,
5408 * ensure all the unknown bits are 0 - i.e. new
5409 * user-space does not rely on any kernel feature
5410 * extensions we dont know about yet.
5412 if (size
> sizeof(*attr
)) {
5413 unsigned char __user
*addr
;
5414 unsigned char __user
*end
;
5417 addr
= (void __user
*)uattr
+ sizeof(*attr
);
5418 end
= (void __user
*)uattr
+ size
;
5420 for (; addr
< end
; addr
++) {
5421 ret
= get_user(val
, addr
);
5427 size
= sizeof(*attr
);
5430 ret
= copy_from_user(attr
, uattr
, size
);
5435 * If the type exists, the corresponding creation will verify
5438 if (attr
->type
>= PERF_TYPE_MAX
)
5441 if (attr
->__reserved_1
)
5444 if (attr
->sample_type
& ~(PERF_SAMPLE_MAX
-1))
5447 if (attr
->read_format
& ~(PERF_FORMAT_MAX
-1))
5454 put_user(sizeof(*attr
), &uattr
->size
);
5460 perf_event_set_output(struct perf_event
*event
, struct perf_event
*output_event
)
5462 struct perf_buffer
*buffer
= NULL
, *old_buffer
= NULL
;
5468 /* don't allow circular references */
5469 if (event
== output_event
)
5473 * Don't allow cross-cpu buffers
5475 if (output_event
->cpu
!= event
->cpu
)
5479 * If its not a per-cpu buffer, it must be the same task.
5481 if (output_event
->cpu
== -1 && output_event
->ctx
!= event
->ctx
)
5485 mutex_lock(&event
->mmap_mutex
);
5486 /* Can't redirect output if we've got an active mmap() */
5487 if (atomic_read(&event
->mmap_count
))
5491 /* get the buffer we want to redirect to */
5492 buffer
= perf_buffer_get(output_event
);
5497 old_buffer
= event
->buffer
;
5498 rcu_assign_pointer(event
->buffer
, buffer
);
5501 mutex_unlock(&event
->mmap_mutex
);
5504 perf_buffer_put(old_buffer
);
5510 * sys_perf_event_open - open a performance event, associate it to a task/cpu
5512 * @attr_uptr: event_id type attributes for monitoring/sampling
5515 * @group_fd: group leader event fd
5517 SYSCALL_DEFINE5(perf_event_open
,
5518 struct perf_event_attr __user
*, attr_uptr
,
5519 pid_t
, pid
, int, cpu
, int, group_fd
, unsigned long, flags
)
5521 struct perf_event
*group_leader
= NULL
, *output_event
= NULL
;
5522 struct perf_event
*event
, *sibling
;
5523 struct perf_event_attr attr
;
5524 struct perf_event_context
*ctx
;
5525 struct file
*event_file
= NULL
;
5526 struct file
*group_file
= NULL
;
5527 struct task_struct
*task
= NULL
;
5531 int fput_needed
= 0;
5534 /* for future expandability... */
5535 if (flags
& ~(PERF_FLAG_FD_NO_GROUP
| PERF_FLAG_FD_OUTPUT
))
5538 err
= perf_copy_attr(attr_uptr
, &attr
);
5542 if (!attr
.exclude_kernel
) {
5543 if (perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN
))
5548 if (attr
.sample_freq
> sysctl_perf_event_sample_rate
)
5552 event_fd
= get_unused_fd_flags(O_RDWR
);
5556 if (group_fd
!= -1) {
5557 group_leader
= perf_fget_light(group_fd
, &fput_needed
);
5558 if (IS_ERR(group_leader
)) {
5559 err
= PTR_ERR(group_leader
);
5562 group_file
= group_leader
->filp
;
5563 if (flags
& PERF_FLAG_FD_OUTPUT
)
5564 output_event
= group_leader
;
5565 if (flags
& PERF_FLAG_FD_NO_GROUP
)
5566 group_leader
= NULL
;
5569 event
= perf_event_alloc(&attr
, cpu
, group_leader
, NULL
, NULL
);
5570 if (IS_ERR(event
)) {
5571 err
= PTR_ERR(event
);
5576 * Special case software events and allow them to be part of
5577 * any hardware group.
5582 (is_software_event(event
) != is_software_event(group_leader
))) {
5583 if (is_software_event(event
)) {
5585 * If event and group_leader are not both a software
5586 * event, and event is, then group leader is not.
5588 * Allow the addition of software events to !software
5589 * groups, this is safe because software events never
5592 pmu
= group_leader
->pmu
;
5593 } else if (is_software_event(group_leader
) &&
5594 (group_leader
->group_flags
& PERF_GROUP_SOFTWARE
)) {
5596 * In case the group is a pure software group, and we
5597 * try to add a hardware event, move the whole group to
5598 * the hardware context.
5605 task
= find_lively_task_by_vpid(pid
);
5608 * Get the target context (task or percpu):
5610 ctx
= find_get_context(pmu
, task
, cpu
);
5617 * Look up the group leader (we will attach this event to it):
5623 * Do not allow a recursive hierarchy (this new sibling
5624 * becoming part of another group-sibling):
5626 if (group_leader
->group_leader
!= group_leader
)
5629 * Do not allow to attach to a group in a different
5630 * task or CPU context:
5633 if (group_leader
->ctx
->type
!= ctx
->type
)
5636 if (group_leader
->ctx
!= ctx
)
5641 * Only a group leader can be exclusive or pinned
5643 if (attr
.exclusive
|| attr
.pinned
)
5648 err
= perf_event_set_output(event
, output_event
);
5653 event_file
= anon_inode_getfile("[perf_event]", &perf_fops
, event
, O_RDWR
);
5654 if (IS_ERR(event_file
)) {
5655 err
= PTR_ERR(event_file
);
5660 struct perf_event_context
*gctx
= group_leader
->ctx
;
5662 mutex_lock(&gctx
->mutex
);
5663 perf_event_remove_from_context(group_leader
);
5664 list_for_each_entry(sibling
, &group_leader
->sibling_list
,
5666 perf_event_remove_from_context(sibling
);
5669 mutex_unlock(&gctx
->mutex
);
5673 event
->filp
= event_file
;
5674 WARN_ON_ONCE(ctx
->parent_ctx
);
5675 mutex_lock(&ctx
->mutex
);
5678 perf_install_in_context(ctx
, group_leader
, cpu
);
5680 list_for_each_entry(sibling
, &group_leader
->sibling_list
,
5682 perf_install_in_context(ctx
, sibling
, cpu
);
5687 perf_install_in_context(ctx
, event
, cpu
);
5689 mutex_unlock(&ctx
->mutex
);
5691 event
->owner
= current
;
5692 get_task_struct(current
);
5693 mutex_lock(¤t
->perf_event_mutex
);
5694 list_add_tail(&event
->owner_entry
, ¤t
->perf_event_list
);
5695 mutex_unlock(¤t
->perf_event_mutex
);
5698 * Drop the reference on the group_event after placing the
5699 * new event on the sibling_list. This ensures destruction
5700 * of the group leader will find the pointer to itself in
5701 * perf_group_detach().
5703 fput_light(group_file
, fput_needed
);
5704 fd_install(event_fd
, event_file
);
5710 fput_light(group_file
, fput_needed
);
5713 put_unused_fd(event_fd
);
5718 * perf_event_create_kernel_counter
5720 * @attr: attributes of the counter to create
5721 * @cpu: cpu in which the counter is bound
5722 * @task: task to profile (NULL for percpu)
5725 perf_event_create_kernel_counter(struct perf_event_attr
*attr
, int cpu
,
5726 struct task_struct
*task
,
5727 perf_overflow_handler_t overflow_handler
)
5729 struct perf_event_context
*ctx
;
5730 struct perf_event
*event
;
5734 * Get the target context (task or percpu):
5737 event
= perf_event_alloc(attr
, cpu
, NULL
, NULL
, overflow_handler
);
5738 if (IS_ERR(event
)) {
5739 err
= PTR_ERR(event
);
5743 ctx
= find_get_context(event
->pmu
, task
, cpu
);
5750 WARN_ON_ONCE(ctx
->parent_ctx
);
5751 mutex_lock(&ctx
->mutex
);
5752 perf_install_in_context(ctx
, event
, cpu
);
5754 mutex_unlock(&ctx
->mutex
);
5756 event
->owner
= current
;
5757 get_task_struct(current
);
5758 mutex_lock(¤t
->perf_event_mutex
);
5759 list_add_tail(&event
->owner_entry
, ¤t
->perf_event_list
);
5760 mutex_unlock(¤t
->perf_event_mutex
);
5767 return ERR_PTR(err
);
5769 EXPORT_SYMBOL_GPL(perf_event_create_kernel_counter
);
5771 static void sync_child_event(struct perf_event
*child_event
,
5772 struct task_struct
*child
)
5774 struct perf_event
*parent_event
= child_event
->parent
;
5777 if (child_event
->attr
.inherit_stat
)
5778 perf_event_read_event(child_event
, child
);
5780 child_val
= perf_event_count(child_event
);
5783 * Add back the child's count to the parent's count:
5785 atomic64_add(child_val
, &parent_event
->child_count
);
5786 atomic64_add(child_event
->total_time_enabled
,
5787 &parent_event
->child_total_time_enabled
);
5788 atomic64_add(child_event
->total_time_running
,
5789 &parent_event
->child_total_time_running
);
5792 * Remove this event from the parent's list
5794 WARN_ON_ONCE(parent_event
->ctx
->parent_ctx
);
5795 mutex_lock(&parent_event
->child_mutex
);
5796 list_del_init(&child_event
->child_list
);
5797 mutex_unlock(&parent_event
->child_mutex
);
5800 * Release the parent event, if this was the last
5803 fput(parent_event
->filp
);
5807 __perf_event_exit_task(struct perf_event
*child_event
,
5808 struct perf_event_context
*child_ctx
,
5809 struct task_struct
*child
)
5811 struct perf_event
*parent_event
;
5813 perf_event_remove_from_context(child_event
);
5815 parent_event
= child_event
->parent
;
5817 * It can happen that parent exits first, and has events
5818 * that are still around due to the child reference. These
5819 * events need to be zapped - but otherwise linger.
5822 sync_child_event(child_event
, child
);
5823 free_event(child_event
);
5827 static void perf_event_exit_task_context(struct task_struct
*child
, int ctxn
)
5829 struct perf_event
*child_event
, *tmp
;
5830 struct perf_event_context
*child_ctx
;
5831 unsigned long flags
;
5833 if (likely(!child
->perf_event_ctxp
[ctxn
])) {
5834 perf_event_task(child
, NULL
, 0);
5838 local_irq_save(flags
);
5840 * We can't reschedule here because interrupts are disabled,
5841 * and either child is current or it is a task that can't be
5842 * scheduled, so we are now safe from rescheduling changing
5845 child_ctx
= child
->perf_event_ctxp
[ctxn
];
5846 __perf_event_task_sched_out(child_ctx
);
5849 * Take the context lock here so that if find_get_context is
5850 * reading child->perf_event_ctxp, we wait until it has
5851 * incremented the context's refcount before we do put_ctx below.
5853 raw_spin_lock(&child_ctx
->lock
);
5854 child
->perf_event_ctxp
[ctxn
] = NULL
;
5856 * If this context is a clone; unclone it so it can't get
5857 * swapped to another process while we're removing all
5858 * the events from it.
5860 unclone_ctx(child_ctx
);
5861 update_context_time(child_ctx
);
5862 raw_spin_unlock_irqrestore(&child_ctx
->lock
, flags
);
5865 * Report the task dead after unscheduling the events so that we
5866 * won't get any samples after PERF_RECORD_EXIT. We can however still
5867 * get a few PERF_RECORD_READ events.
5869 perf_event_task(child
, child_ctx
, 0);
5872 * We can recurse on the same lock type through:
5874 * __perf_event_exit_task()
5875 * sync_child_event()
5876 * fput(parent_event->filp)
5878 * mutex_lock(&ctx->mutex)
5880 * But since its the parent context it won't be the same instance.
5882 mutex_lock(&child_ctx
->mutex
);
5885 list_for_each_entry_safe(child_event
, tmp
, &child_ctx
->pinned_groups
,
5887 __perf_event_exit_task(child_event
, child_ctx
, child
);
5889 list_for_each_entry_safe(child_event
, tmp
, &child_ctx
->flexible_groups
,
5891 __perf_event_exit_task(child_event
, child_ctx
, child
);
5894 * If the last event was a group event, it will have appended all
5895 * its siblings to the list, but we obtained 'tmp' before that which
5896 * will still point to the list head terminating the iteration.
5898 if (!list_empty(&child_ctx
->pinned_groups
) ||
5899 !list_empty(&child_ctx
->flexible_groups
))
5902 mutex_unlock(&child_ctx
->mutex
);
5908 * When a child task exits, feed back event values to parent events.
5910 void perf_event_exit_task(struct task_struct
*child
)
5914 for_each_task_context_nr(ctxn
)
5915 perf_event_exit_task_context(child
, ctxn
);
5918 static void perf_free_event(struct perf_event
*event
,
5919 struct perf_event_context
*ctx
)
5921 struct perf_event
*parent
= event
->parent
;
5923 if (WARN_ON_ONCE(!parent
))
5926 mutex_lock(&parent
->child_mutex
);
5927 list_del_init(&event
->child_list
);
5928 mutex_unlock(&parent
->child_mutex
);
5932 perf_group_detach(event
);
5933 list_del_event(event
, ctx
);
5938 * free an unexposed, unused context as created by inheritance by
5939 * perf_event_init_task below, used by fork() in case of fail.
5941 void perf_event_free_task(struct task_struct
*task
)
5943 struct perf_event_context
*ctx
;
5944 struct perf_event
*event
, *tmp
;
5947 for_each_task_context_nr(ctxn
) {
5948 ctx
= task
->perf_event_ctxp
[ctxn
];
5952 mutex_lock(&ctx
->mutex
);
5954 list_for_each_entry_safe(event
, tmp
, &ctx
->pinned_groups
,
5956 perf_free_event(event
, ctx
);
5958 list_for_each_entry_safe(event
, tmp
, &ctx
->flexible_groups
,
5960 perf_free_event(event
, ctx
);
5962 if (!list_empty(&ctx
->pinned_groups
) ||
5963 !list_empty(&ctx
->flexible_groups
))
5966 mutex_unlock(&ctx
->mutex
);
5972 void perf_event_delayed_put(struct task_struct
*task
)
5976 for_each_task_context_nr(ctxn
)
5977 WARN_ON_ONCE(task
->perf_event_ctxp
[ctxn
]);
5981 * inherit a event from parent task to child task:
5983 static struct perf_event
*
5984 inherit_event(struct perf_event
*parent_event
,
5985 struct task_struct
*parent
,
5986 struct perf_event_context
*parent_ctx
,
5987 struct task_struct
*child
,
5988 struct perf_event
*group_leader
,
5989 struct perf_event_context
*child_ctx
)
5991 struct perf_event
*child_event
;
5992 unsigned long flags
;
5995 * Instead of creating recursive hierarchies of events,
5996 * we link inherited events back to the original parent,
5997 * which has a filp for sure, which we use as the reference
6000 if (parent_event
->parent
)
6001 parent_event
= parent_event
->parent
;
6003 child_event
= perf_event_alloc(&parent_event
->attr
,
6005 group_leader
, parent_event
,
6007 if (IS_ERR(child_event
))
6012 * Make the child state follow the state of the parent event,
6013 * not its attr.disabled bit. We hold the parent's mutex,
6014 * so we won't race with perf_event_{en, dis}able_family.
6016 if (parent_event
->state
>= PERF_EVENT_STATE_INACTIVE
)
6017 child_event
->state
= PERF_EVENT_STATE_INACTIVE
;
6019 child_event
->state
= PERF_EVENT_STATE_OFF
;
6021 if (parent_event
->attr
.freq
) {
6022 u64 sample_period
= parent_event
->hw
.sample_period
;
6023 struct hw_perf_event
*hwc
= &child_event
->hw
;
6025 hwc
->sample_period
= sample_period
;
6026 hwc
->last_period
= sample_period
;
6028 local64_set(&hwc
->period_left
, sample_period
);
6031 child_event
->ctx
= child_ctx
;
6032 child_event
->overflow_handler
= parent_event
->overflow_handler
;
6035 * Link it up in the child's context:
6037 raw_spin_lock_irqsave(&child_ctx
->lock
, flags
);
6038 add_event_to_ctx(child_event
, child_ctx
);
6039 raw_spin_unlock_irqrestore(&child_ctx
->lock
, flags
);
6042 * Get a reference to the parent filp - we will fput it
6043 * when the child event exits. This is safe to do because
6044 * we are in the parent and we know that the filp still
6045 * exists and has a nonzero count:
6047 atomic_long_inc(&parent_event
->filp
->f_count
);
6050 * Link this into the parent event's child list
6052 WARN_ON_ONCE(parent_event
->ctx
->parent_ctx
);
6053 mutex_lock(&parent_event
->child_mutex
);
6054 list_add_tail(&child_event
->child_list
, &parent_event
->child_list
);
6055 mutex_unlock(&parent_event
->child_mutex
);
6060 static int inherit_group(struct perf_event
*parent_event
,
6061 struct task_struct
*parent
,
6062 struct perf_event_context
*parent_ctx
,
6063 struct task_struct
*child
,
6064 struct perf_event_context
*child_ctx
)
6066 struct perf_event
*leader
;
6067 struct perf_event
*sub
;
6068 struct perf_event
*child_ctr
;
6070 leader
= inherit_event(parent_event
, parent
, parent_ctx
,
6071 child
, NULL
, child_ctx
);
6073 return PTR_ERR(leader
);
6074 list_for_each_entry(sub
, &parent_event
->sibling_list
, group_entry
) {
6075 child_ctr
= inherit_event(sub
, parent
, parent_ctx
,
6076 child
, leader
, child_ctx
);
6077 if (IS_ERR(child_ctr
))
6078 return PTR_ERR(child_ctr
);
6084 inherit_task_group(struct perf_event
*event
, struct task_struct
*parent
,
6085 struct perf_event_context
*parent_ctx
,
6086 struct task_struct
*child
, int ctxn
,
6090 struct perf_event_context
*child_ctx
;
6092 if (!event
->attr
.inherit
) {
6097 child_ctx
= child
->perf_event_ctxp
[ctxn
];
6100 * This is executed from the parent task context, so
6101 * inherit events that have been marked for cloning.
6102 * First allocate and initialize a context for the
6106 child_ctx
= alloc_perf_context(event
->pmu
, child
);
6110 child
->perf_event_ctxp
[ctxn
] = child_ctx
;
6113 ret
= inherit_group(event
, parent
, parent_ctx
,
6123 * Initialize the perf_event context in task_struct
6125 int perf_event_init_context(struct task_struct
*child
, int ctxn
)
6127 struct perf_event_context
*child_ctx
, *parent_ctx
;
6128 struct perf_event_context
*cloned_ctx
;
6129 struct perf_event
*event
;
6130 struct task_struct
*parent
= current
;
6131 int inherited_all
= 1;
6134 child
->perf_event_ctxp
[ctxn
] = NULL
;
6136 mutex_init(&child
->perf_event_mutex
);
6137 INIT_LIST_HEAD(&child
->perf_event_list
);
6139 if (likely(!parent
->perf_event_ctxp
[ctxn
]))
6143 * If the parent's context is a clone, pin it so it won't get
6146 parent_ctx
= perf_pin_task_context(parent
, ctxn
);
6149 * No need to check if parent_ctx != NULL here; since we saw
6150 * it non-NULL earlier, the only reason for it to become NULL
6151 * is if we exit, and since we're currently in the middle of
6152 * a fork we can't be exiting at the same time.
6156 * Lock the parent list. No need to lock the child - not PID
6157 * hashed yet and not running, so nobody can access it.
6159 mutex_lock(&parent_ctx
->mutex
);
6162 * We dont have to disable NMIs - we are only looking at
6163 * the list, not manipulating it:
6165 list_for_each_entry(event
, &parent_ctx
->pinned_groups
, group_entry
) {
6166 ret
= inherit_task_group(event
, parent
, parent_ctx
,
6167 child
, ctxn
, &inherited_all
);
6172 list_for_each_entry(event
, &parent_ctx
->flexible_groups
, group_entry
) {
6173 ret
= inherit_task_group(event
, parent
, parent_ctx
,
6174 child
, ctxn
, &inherited_all
);
6179 child_ctx
= child
->perf_event_ctxp
[ctxn
];
6181 if (child_ctx
&& inherited_all
) {
6183 * Mark the child context as a clone of the parent
6184 * context, or of whatever the parent is a clone of.
6185 * Note that if the parent is a clone, it could get
6186 * uncloned at any point, but that doesn't matter
6187 * because the list of events and the generation
6188 * count can't have changed since we took the mutex.
6190 cloned_ctx
= rcu_dereference(parent_ctx
->parent_ctx
);
6192 child_ctx
->parent_ctx
= cloned_ctx
;
6193 child_ctx
->parent_gen
= parent_ctx
->parent_gen
;
6195 child_ctx
->parent_ctx
= parent_ctx
;
6196 child_ctx
->parent_gen
= parent_ctx
->generation
;
6198 get_ctx(child_ctx
->parent_ctx
);
6201 mutex_unlock(&parent_ctx
->mutex
);
6203 perf_unpin_context(parent_ctx
);
6209 * Initialize the perf_event context in task_struct
6211 int perf_event_init_task(struct task_struct
*child
)
6215 for_each_task_context_nr(ctxn
) {
6216 ret
= perf_event_init_context(child
, ctxn
);
6224 static void __init
perf_event_init_all_cpus(void)
6226 struct swevent_htable
*swhash
;
6229 for_each_possible_cpu(cpu
) {
6230 swhash
= &per_cpu(swevent_htable
, cpu
);
6231 mutex_init(&swhash
->hlist_mutex
);
6235 static void __cpuinit
perf_event_init_cpu(int cpu
)
6237 struct swevent_htable
*swhash
= &per_cpu(swevent_htable
, cpu
);
6239 mutex_lock(&swhash
->hlist_mutex
);
6240 if (swhash
->hlist_refcount
> 0) {
6241 struct swevent_hlist
*hlist
;
6243 hlist
= kzalloc_node(sizeof(*hlist
), GFP_KERNEL
, cpu_to_node(cpu
));
6245 rcu_assign_pointer(swhash
->swevent_hlist
, hlist
);
6247 mutex_unlock(&swhash
->hlist_mutex
);
6250 #ifdef CONFIG_HOTPLUG_CPU
6251 static void __perf_event_exit_context(void *__info
)
6253 struct perf_event_context
*ctx
= __info
;
6254 struct perf_event
*event
, *tmp
;
6256 perf_pmu_rotate_stop(ctx
->pmu
);
6258 list_for_each_entry_safe(event
, tmp
, &ctx
->pinned_groups
, group_entry
)
6259 __perf_event_remove_from_context(event
);
6260 list_for_each_entry_safe(event
, tmp
, &ctx
->flexible_groups
, group_entry
)
6261 __perf_event_remove_from_context(event
);
6264 static void perf_event_exit_cpu_context(int cpu
)
6266 struct perf_event_context
*ctx
;
6270 idx
= srcu_read_lock(&pmus_srcu
);
6271 list_for_each_entry_rcu(pmu
, &pmus
, entry
) {
6272 ctx
= &per_cpu_ptr(pmu
->pmu_cpu_context
, cpu
)->ctx
;
6274 mutex_lock(&ctx
->mutex
);
6275 smp_call_function_single(cpu
, __perf_event_exit_context
, ctx
, 1);
6276 mutex_unlock(&ctx
->mutex
);
6278 srcu_read_unlock(&pmus_srcu
, idx
);
6281 static void perf_event_exit_cpu(int cpu
)
6283 struct swevent_htable
*swhash
= &per_cpu(swevent_htable
, cpu
);
6285 mutex_lock(&swhash
->hlist_mutex
);
6286 swevent_hlist_release(swhash
);
6287 mutex_unlock(&swhash
->hlist_mutex
);
6289 perf_event_exit_cpu_context(cpu
);
6292 static inline void perf_event_exit_cpu(int cpu
) { }
6295 static int __cpuinit
6296 perf_cpu_notify(struct notifier_block
*self
, unsigned long action
, void *hcpu
)
6298 unsigned int cpu
= (long)hcpu
;
6300 switch (action
& ~CPU_TASKS_FROZEN
) {
6302 case CPU_UP_PREPARE
:
6303 case CPU_DOWN_FAILED
:
6304 perf_event_init_cpu(cpu
);
6307 case CPU_UP_CANCELED
:
6308 case CPU_DOWN_PREPARE
:
6309 perf_event_exit_cpu(cpu
);
6319 void __init
perf_event_init(void)
6321 perf_event_init_all_cpus();
6322 init_srcu_struct(&pmus_srcu
);
6323 perf_pmu_register(&perf_swevent
);
6324 perf_pmu_register(&perf_cpu_clock
);
6325 perf_pmu_register(&perf_task_clock
);
6327 perf_cpu_notifier(perf_cpu_notify
);