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/idr.h>
17 #include <linux/file.h>
18 #include <linux/poll.h>
19 #include <linux/slab.h>
20 #include <linux/hash.h>
21 #include <linux/sysfs.h>
22 #include <linux/dcache.h>
23 #include <linux/percpu.h>
24 #include <linux/ptrace.h>
25 #include <linux/reboot.h>
26 #include <linux/vmstat.h>
27 #include <linux/device.h>
28 #include <linux/vmalloc.h>
29 #include <linux/hardirq.h>
30 #include <linux/rculist.h>
31 #include <linux/uaccess.h>
32 #include <linux/syscalls.h>
33 #include <linux/anon_inodes.h>
34 #include <linux/kernel_stat.h>
35 #include <linux/perf_event.h>
36 #include <linux/ftrace_event.h>
37 #include <linux/hw_breakpoint.h>
39 #include <asm/irq_regs.h>
41 atomic_t perf_task_events __read_mostly
;
42 static atomic_t nr_mmap_events __read_mostly
;
43 static atomic_t nr_comm_events __read_mostly
;
44 static atomic_t nr_task_events __read_mostly
;
46 static LIST_HEAD(pmus
);
47 static DEFINE_MUTEX(pmus_lock
);
48 static struct srcu_struct pmus_srcu
;
51 * perf event paranoia level:
52 * -1 - not paranoid at all
53 * 0 - disallow raw tracepoint access for unpriv
54 * 1 - disallow cpu events for unpriv
55 * 2 - disallow kernel profiling for unpriv
57 int sysctl_perf_event_paranoid __read_mostly
= 1;
59 int sysctl_perf_event_mlock __read_mostly
= 512; /* 'free' kb per user */
62 * max perf event sample rate
64 int sysctl_perf_event_sample_rate __read_mostly
= 100000;
66 static atomic64_t perf_event_id
;
68 void __weak
perf_event_print_debug(void) { }
70 extern __weak
const char *perf_pmu_name(void)
75 void perf_pmu_disable(struct pmu
*pmu
)
77 int *count
= this_cpu_ptr(pmu
->pmu_disable_count
);
79 pmu
->pmu_disable(pmu
);
82 void perf_pmu_enable(struct pmu
*pmu
)
84 int *count
= this_cpu_ptr(pmu
->pmu_disable_count
);
89 static DEFINE_PER_CPU(struct list_head
, rotation_list
);
92 * perf_pmu_rotate_start() and perf_rotate_context() are fully serialized
93 * because they're strictly cpu affine and rotate_start is called with IRQs
94 * disabled, while rotate_context is called from IRQ context.
96 static void perf_pmu_rotate_start(struct pmu
*pmu
)
98 struct perf_cpu_context
*cpuctx
= this_cpu_ptr(pmu
->pmu_cpu_context
);
99 struct list_head
*head
= &__get_cpu_var(rotation_list
);
101 WARN_ON(!irqs_disabled());
103 if (list_empty(&cpuctx
->rotation_list
))
104 list_add(&cpuctx
->rotation_list
, head
);
107 static void get_ctx(struct perf_event_context
*ctx
)
109 WARN_ON(!atomic_inc_not_zero(&ctx
->refcount
));
112 static void free_ctx(struct rcu_head
*head
)
114 struct perf_event_context
*ctx
;
116 ctx
= container_of(head
, struct perf_event_context
, rcu_head
);
120 static void put_ctx(struct perf_event_context
*ctx
)
122 if (atomic_dec_and_test(&ctx
->refcount
)) {
124 put_ctx(ctx
->parent_ctx
);
126 put_task_struct(ctx
->task
);
127 call_rcu(&ctx
->rcu_head
, free_ctx
);
131 static void unclone_ctx(struct perf_event_context
*ctx
)
133 if (ctx
->parent_ctx
) {
134 put_ctx(ctx
->parent_ctx
);
135 ctx
->parent_ctx
= NULL
;
139 static u32
perf_event_pid(struct perf_event
*event
, struct task_struct
*p
)
142 * only top level events have the pid namespace they were created in
145 event
= event
->parent
;
147 return task_tgid_nr_ns(p
, event
->ns
);
150 static u32
perf_event_tid(struct perf_event
*event
, struct task_struct
*p
)
153 * only top level events have the pid namespace they were created in
156 event
= event
->parent
;
158 return task_pid_nr_ns(p
, event
->ns
);
162 * If we inherit events we want to return the parent event id
165 static u64
primary_event_id(struct perf_event
*event
)
170 id
= event
->parent
->id
;
176 * Get the perf_event_context for a task and lock it.
177 * This has to cope with with the fact that until it is locked,
178 * the context could get moved to another task.
180 static struct perf_event_context
*
181 perf_lock_task_context(struct task_struct
*task
, int ctxn
, unsigned long *flags
)
183 struct perf_event_context
*ctx
;
187 ctx
= rcu_dereference(task
->perf_event_ctxp
[ctxn
]);
190 * If this context is a clone of another, it might
191 * get swapped for another underneath us by
192 * perf_event_task_sched_out, though the
193 * rcu_read_lock() protects us from any context
194 * getting freed. Lock the context and check if it
195 * got swapped before we could get the lock, and retry
196 * if so. If we locked the right context, then it
197 * can't get swapped on us any more.
199 raw_spin_lock_irqsave(&ctx
->lock
, *flags
);
200 if (ctx
!= rcu_dereference(task
->perf_event_ctxp
[ctxn
])) {
201 raw_spin_unlock_irqrestore(&ctx
->lock
, *flags
);
205 if (!atomic_inc_not_zero(&ctx
->refcount
)) {
206 raw_spin_unlock_irqrestore(&ctx
->lock
, *flags
);
215 * Get the context for a task and increment its pin_count so it
216 * can't get swapped to another task. This also increments its
217 * reference count so that the context can't get freed.
219 static struct perf_event_context
*
220 perf_pin_task_context(struct task_struct
*task
, int ctxn
)
222 struct perf_event_context
*ctx
;
225 ctx
= perf_lock_task_context(task
, ctxn
, &flags
);
228 raw_spin_unlock_irqrestore(&ctx
->lock
, flags
);
233 static void perf_unpin_context(struct perf_event_context
*ctx
)
237 raw_spin_lock_irqsave(&ctx
->lock
, flags
);
239 raw_spin_unlock_irqrestore(&ctx
->lock
, flags
);
243 static inline u64
perf_clock(void)
245 return local_clock();
249 * Update the record of the current time in a context.
251 static void update_context_time(struct perf_event_context
*ctx
)
253 u64 now
= perf_clock();
255 ctx
->time
+= now
- ctx
->timestamp
;
256 ctx
->timestamp
= now
;
260 * Update the total_time_enabled and total_time_running fields for a event.
262 static void update_event_times(struct perf_event
*event
)
264 struct perf_event_context
*ctx
= event
->ctx
;
267 if (event
->state
< PERF_EVENT_STATE_INACTIVE
||
268 event
->group_leader
->state
< PERF_EVENT_STATE_INACTIVE
)
274 run_end
= event
->tstamp_stopped
;
276 event
->total_time_enabled
= run_end
- event
->tstamp_enabled
;
278 if (event
->state
== PERF_EVENT_STATE_INACTIVE
)
279 run_end
= event
->tstamp_stopped
;
283 event
->total_time_running
= run_end
- event
->tstamp_running
;
287 * Update total_time_enabled and total_time_running for all events in a group.
289 static void update_group_times(struct perf_event
*leader
)
291 struct perf_event
*event
;
293 update_event_times(leader
);
294 list_for_each_entry(event
, &leader
->sibling_list
, group_entry
)
295 update_event_times(event
);
298 static struct list_head
*
299 ctx_group_list(struct perf_event
*event
, struct perf_event_context
*ctx
)
301 if (event
->attr
.pinned
)
302 return &ctx
->pinned_groups
;
304 return &ctx
->flexible_groups
;
308 * Add a event from the lists for its context.
309 * Must be called with ctx->mutex and ctx->lock held.
312 list_add_event(struct perf_event
*event
, struct perf_event_context
*ctx
)
314 WARN_ON_ONCE(event
->attach_state
& PERF_ATTACH_CONTEXT
);
315 event
->attach_state
|= PERF_ATTACH_CONTEXT
;
318 * If we're a stand alone event or group leader, we go to the context
319 * list, group events are kept attached to the group so that
320 * perf_group_detach can, at all times, locate all siblings.
322 if (event
->group_leader
== event
) {
323 struct list_head
*list
;
325 if (is_software_event(event
))
326 event
->group_flags
|= PERF_GROUP_SOFTWARE
;
328 list
= ctx_group_list(event
, ctx
);
329 list_add_tail(&event
->group_entry
, list
);
332 list_add_rcu(&event
->event_entry
, &ctx
->event_list
);
334 perf_pmu_rotate_start(ctx
->pmu
);
336 if (event
->attr
.inherit_stat
)
341 * Called at perf_event creation and when events are attached/detached from a
344 static void perf_event__read_size(struct perf_event
*event
)
346 int entry
= sizeof(u64
); /* value */
350 if (event
->attr
.read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
)
353 if (event
->attr
.read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
)
356 if (event
->attr
.read_format
& PERF_FORMAT_ID
)
357 entry
+= sizeof(u64
);
359 if (event
->attr
.read_format
& PERF_FORMAT_GROUP
) {
360 nr
+= event
->group_leader
->nr_siblings
;
365 event
->read_size
= size
;
368 static void perf_event__header_size(struct perf_event
*event
)
370 struct perf_sample_data
*data
;
371 u64 sample_type
= event
->attr
.sample_type
;
374 perf_event__read_size(event
);
376 if (sample_type
& PERF_SAMPLE_IP
)
377 size
+= sizeof(data
->ip
);
379 if (sample_type
& PERF_SAMPLE_ADDR
)
380 size
+= sizeof(data
->addr
);
382 if (sample_type
& PERF_SAMPLE_PERIOD
)
383 size
+= sizeof(data
->period
);
385 if (sample_type
& PERF_SAMPLE_READ
)
386 size
+= event
->read_size
;
388 event
->header_size
= size
;
391 static void perf_event__id_header_size(struct perf_event
*event
)
393 struct perf_sample_data
*data
;
394 u64 sample_type
= event
->attr
.sample_type
;
397 if (sample_type
& PERF_SAMPLE_TID
)
398 size
+= sizeof(data
->tid_entry
);
400 if (sample_type
& PERF_SAMPLE_TIME
)
401 size
+= sizeof(data
->time
);
403 if (sample_type
& PERF_SAMPLE_ID
)
404 size
+= sizeof(data
->id
);
406 if (sample_type
& PERF_SAMPLE_STREAM_ID
)
407 size
+= sizeof(data
->stream_id
);
409 if (sample_type
& PERF_SAMPLE_CPU
)
410 size
+= sizeof(data
->cpu_entry
);
412 event
->id_header_size
= size
;
415 static void perf_group_attach(struct perf_event
*event
)
417 struct perf_event
*group_leader
= event
->group_leader
, *pos
;
420 * We can have double attach due to group movement in perf_event_open.
422 if (event
->attach_state
& PERF_ATTACH_GROUP
)
425 event
->attach_state
|= PERF_ATTACH_GROUP
;
427 if (group_leader
== event
)
430 if (group_leader
->group_flags
& PERF_GROUP_SOFTWARE
&&
431 !is_software_event(event
))
432 group_leader
->group_flags
&= ~PERF_GROUP_SOFTWARE
;
434 list_add_tail(&event
->group_entry
, &group_leader
->sibling_list
);
435 group_leader
->nr_siblings
++;
437 perf_event__header_size(group_leader
);
439 list_for_each_entry(pos
, &group_leader
->sibling_list
, group_entry
)
440 perf_event__header_size(pos
);
444 * Remove a event from the lists for its context.
445 * Must be called with ctx->mutex and ctx->lock held.
448 list_del_event(struct perf_event
*event
, struct perf_event_context
*ctx
)
451 * We can have double detach due to exit/hot-unplug + close.
453 if (!(event
->attach_state
& PERF_ATTACH_CONTEXT
))
456 event
->attach_state
&= ~PERF_ATTACH_CONTEXT
;
459 if (event
->attr
.inherit_stat
)
462 list_del_rcu(&event
->event_entry
);
464 if (event
->group_leader
== event
)
465 list_del_init(&event
->group_entry
);
467 update_group_times(event
);
470 * If event was in error state, then keep it
471 * that way, otherwise bogus counts will be
472 * returned on read(). The only way to get out
473 * of error state is by explicit re-enabling
476 if (event
->state
> PERF_EVENT_STATE_OFF
)
477 event
->state
= PERF_EVENT_STATE_OFF
;
480 static void perf_group_detach(struct perf_event
*event
)
482 struct perf_event
*sibling
, *tmp
;
483 struct list_head
*list
= NULL
;
486 * We can have double detach due to exit/hot-unplug + close.
488 if (!(event
->attach_state
& PERF_ATTACH_GROUP
))
491 event
->attach_state
&= ~PERF_ATTACH_GROUP
;
494 * If this is a sibling, remove it from its group.
496 if (event
->group_leader
!= event
) {
497 list_del_init(&event
->group_entry
);
498 event
->group_leader
->nr_siblings
--;
502 if (!list_empty(&event
->group_entry
))
503 list
= &event
->group_entry
;
506 * If this was a group event with sibling events then
507 * upgrade the siblings to singleton events by adding them
508 * to whatever list we are on.
510 list_for_each_entry_safe(sibling
, tmp
, &event
->sibling_list
, group_entry
) {
512 list_move_tail(&sibling
->group_entry
, list
);
513 sibling
->group_leader
= sibling
;
515 /* Inherit group flags from the previous leader */
516 sibling
->group_flags
= event
->group_flags
;
520 perf_event__header_size(event
->group_leader
);
522 list_for_each_entry(tmp
, &event
->group_leader
->sibling_list
, group_entry
)
523 perf_event__header_size(tmp
);
527 event_filter_match(struct perf_event
*event
)
529 return event
->cpu
== -1 || event
->cpu
== smp_processor_id();
533 event_sched_out(struct perf_event
*event
,
534 struct perf_cpu_context
*cpuctx
,
535 struct perf_event_context
*ctx
)
539 * An event which could not be activated because of
540 * filter mismatch still needs to have its timings
541 * maintained, otherwise bogus information is return
542 * via read() for time_enabled, time_running:
544 if (event
->state
== PERF_EVENT_STATE_INACTIVE
545 && !event_filter_match(event
)) {
546 delta
= ctx
->time
- event
->tstamp_stopped
;
547 event
->tstamp_running
+= delta
;
548 event
->tstamp_stopped
= ctx
->time
;
551 if (event
->state
!= PERF_EVENT_STATE_ACTIVE
)
554 event
->state
= PERF_EVENT_STATE_INACTIVE
;
555 if (event
->pending_disable
) {
556 event
->pending_disable
= 0;
557 event
->state
= PERF_EVENT_STATE_OFF
;
559 event
->tstamp_stopped
= ctx
->time
;
560 event
->pmu
->del(event
, 0);
563 if (!is_software_event(event
))
564 cpuctx
->active_oncpu
--;
566 if (event
->attr
.exclusive
|| !cpuctx
->active_oncpu
)
567 cpuctx
->exclusive
= 0;
571 group_sched_out(struct perf_event
*group_event
,
572 struct perf_cpu_context
*cpuctx
,
573 struct perf_event_context
*ctx
)
575 struct perf_event
*event
;
576 int state
= group_event
->state
;
578 event_sched_out(group_event
, cpuctx
, ctx
);
581 * Schedule out siblings (if any):
583 list_for_each_entry(event
, &group_event
->sibling_list
, group_entry
)
584 event_sched_out(event
, cpuctx
, ctx
);
586 if (state
== PERF_EVENT_STATE_ACTIVE
&& group_event
->attr
.exclusive
)
587 cpuctx
->exclusive
= 0;
590 static inline struct perf_cpu_context
*
591 __get_cpu_context(struct perf_event_context
*ctx
)
593 return this_cpu_ptr(ctx
->pmu
->pmu_cpu_context
);
597 * Cross CPU call to remove a performance event
599 * We disable the event on the hardware level first. After that we
600 * remove it from the context list.
602 static void __perf_event_remove_from_context(void *info
)
604 struct perf_event
*event
= info
;
605 struct perf_event_context
*ctx
= event
->ctx
;
606 struct perf_cpu_context
*cpuctx
= __get_cpu_context(ctx
);
609 * If this is a task context, we need to check whether it is
610 * the current task context of this cpu. If not it has been
611 * scheduled out before the smp call arrived.
613 if (ctx
->task
&& cpuctx
->task_ctx
!= ctx
)
616 raw_spin_lock(&ctx
->lock
);
618 event_sched_out(event
, cpuctx
, ctx
);
620 list_del_event(event
, ctx
);
622 raw_spin_unlock(&ctx
->lock
);
627 * Remove the event from a task's (or a CPU's) list of events.
629 * Must be called with ctx->mutex held.
631 * CPU events are removed with a smp call. For task events we only
632 * call when the task is on a CPU.
634 * If event->ctx is a cloned context, callers must make sure that
635 * every task struct that event->ctx->task could possibly point to
636 * remains valid. This is OK when called from perf_release since
637 * that only calls us on the top-level context, which can't be a clone.
638 * When called from perf_event_exit_task, it's OK because the
639 * context has been detached from its task.
641 static void perf_event_remove_from_context(struct perf_event
*event
)
643 struct perf_event_context
*ctx
= event
->ctx
;
644 struct task_struct
*task
= ctx
->task
;
648 * Per cpu events are removed via an smp call and
649 * the removal is always successful.
651 smp_call_function_single(event
->cpu
,
652 __perf_event_remove_from_context
,
658 task_oncpu_function_call(task
, __perf_event_remove_from_context
,
661 raw_spin_lock_irq(&ctx
->lock
);
663 * If the context is active we need to retry the smp call.
665 if (ctx
->nr_active
&& !list_empty(&event
->group_entry
)) {
666 raw_spin_unlock_irq(&ctx
->lock
);
671 * The lock prevents that this context is scheduled in so we
672 * can remove the event safely, if the call above did not
675 if (!list_empty(&event
->group_entry
))
676 list_del_event(event
, ctx
);
677 raw_spin_unlock_irq(&ctx
->lock
);
681 * Cross CPU call to disable a performance event
683 static void __perf_event_disable(void *info
)
685 struct perf_event
*event
= info
;
686 struct perf_event_context
*ctx
= event
->ctx
;
687 struct perf_cpu_context
*cpuctx
= __get_cpu_context(ctx
);
690 * If this is a per-task event, need to check whether this
691 * event's task is the current task on this cpu.
693 if (ctx
->task
&& cpuctx
->task_ctx
!= ctx
)
696 raw_spin_lock(&ctx
->lock
);
699 * If the event is on, turn it off.
700 * If it is in error state, leave it in error state.
702 if (event
->state
>= PERF_EVENT_STATE_INACTIVE
) {
703 update_context_time(ctx
);
704 update_group_times(event
);
705 if (event
== event
->group_leader
)
706 group_sched_out(event
, cpuctx
, ctx
);
708 event_sched_out(event
, cpuctx
, ctx
);
709 event
->state
= PERF_EVENT_STATE_OFF
;
712 raw_spin_unlock(&ctx
->lock
);
718 * If event->ctx is a cloned context, callers must make sure that
719 * every task struct that event->ctx->task could possibly point to
720 * remains valid. This condition is satisifed when called through
721 * perf_event_for_each_child or perf_event_for_each because they
722 * hold the top-level event's child_mutex, so any descendant that
723 * goes to exit will block in sync_child_event.
724 * When called from perf_pending_event it's OK because event->ctx
725 * is the current context on this CPU and preemption is disabled,
726 * hence we can't get into perf_event_task_sched_out for this context.
728 void perf_event_disable(struct perf_event
*event
)
730 struct perf_event_context
*ctx
= event
->ctx
;
731 struct task_struct
*task
= ctx
->task
;
735 * Disable the event on the cpu that it's on
737 smp_call_function_single(event
->cpu
, __perf_event_disable
,
743 task_oncpu_function_call(task
, __perf_event_disable
, event
);
745 raw_spin_lock_irq(&ctx
->lock
);
747 * If the event is still active, we need to retry the cross-call.
749 if (event
->state
== PERF_EVENT_STATE_ACTIVE
) {
750 raw_spin_unlock_irq(&ctx
->lock
);
755 * Since we have the lock this context can't be scheduled
756 * in, so we can change the state safely.
758 if (event
->state
== PERF_EVENT_STATE_INACTIVE
) {
759 update_group_times(event
);
760 event
->state
= PERF_EVENT_STATE_OFF
;
763 raw_spin_unlock_irq(&ctx
->lock
);
767 event_sched_in(struct perf_event
*event
,
768 struct perf_cpu_context
*cpuctx
,
769 struct perf_event_context
*ctx
)
771 if (event
->state
<= PERF_EVENT_STATE_OFF
)
774 event
->state
= PERF_EVENT_STATE_ACTIVE
;
775 event
->oncpu
= smp_processor_id();
777 * The new state must be visible before we turn it on in the hardware:
781 if (event
->pmu
->add(event
, PERF_EF_START
)) {
782 event
->state
= PERF_EVENT_STATE_INACTIVE
;
787 event
->tstamp_running
+= ctx
->time
- event
->tstamp_stopped
;
789 event
->shadow_ctx_time
= ctx
->time
- ctx
->timestamp
;
791 if (!is_software_event(event
))
792 cpuctx
->active_oncpu
++;
795 if (event
->attr
.exclusive
)
796 cpuctx
->exclusive
= 1;
802 group_sched_in(struct perf_event
*group_event
,
803 struct perf_cpu_context
*cpuctx
,
804 struct perf_event_context
*ctx
)
806 struct perf_event
*event
, *partial_group
= NULL
;
807 struct pmu
*pmu
= group_event
->pmu
;
809 bool simulate
= false;
811 if (group_event
->state
== PERF_EVENT_STATE_OFF
)
816 if (event_sched_in(group_event
, cpuctx
, ctx
)) {
817 pmu
->cancel_txn(pmu
);
822 * Schedule in siblings as one group (if any):
824 list_for_each_entry(event
, &group_event
->sibling_list
, group_entry
) {
825 if (event_sched_in(event
, cpuctx
, ctx
)) {
826 partial_group
= event
;
831 if (!pmu
->commit_txn(pmu
))
836 * Groups can be scheduled in as one unit only, so undo any
837 * partial group before returning:
838 * The events up to the failed event are scheduled out normally,
839 * tstamp_stopped will be updated.
841 * The failed events and the remaining siblings need to have
842 * their timings updated as if they had gone thru event_sched_in()
843 * and event_sched_out(). This is required to get consistent timings
844 * across the group. This also takes care of the case where the group
845 * could never be scheduled by ensuring tstamp_stopped is set to mark
846 * the time the event was actually stopped, such that time delta
847 * calculation in update_event_times() is correct.
849 list_for_each_entry(event
, &group_event
->sibling_list
, group_entry
) {
850 if (event
== partial_group
)
854 event
->tstamp_running
+= now
- event
->tstamp_stopped
;
855 event
->tstamp_stopped
= now
;
857 event_sched_out(event
, cpuctx
, ctx
);
860 event_sched_out(group_event
, cpuctx
, ctx
);
862 pmu
->cancel_txn(pmu
);
868 * Work out whether we can put this event group on the CPU now.
870 static int group_can_go_on(struct perf_event
*event
,
871 struct perf_cpu_context
*cpuctx
,
875 * Groups consisting entirely of software events can always go on.
877 if (event
->group_flags
& PERF_GROUP_SOFTWARE
)
880 * If an exclusive group is already on, no other hardware
883 if (cpuctx
->exclusive
)
886 * If this group is exclusive and there are already
887 * events on the CPU, it can't go on.
889 if (event
->attr
.exclusive
&& cpuctx
->active_oncpu
)
892 * Otherwise, try to add it if all previous groups were able
898 static void add_event_to_ctx(struct perf_event
*event
,
899 struct perf_event_context
*ctx
)
901 list_add_event(event
, ctx
);
902 perf_group_attach(event
);
903 event
->tstamp_enabled
= ctx
->time
;
904 event
->tstamp_running
= ctx
->time
;
905 event
->tstamp_stopped
= ctx
->time
;
909 * Cross CPU call to install and enable a performance event
911 * Must be called with ctx->mutex held
913 static void __perf_install_in_context(void *info
)
915 struct perf_event
*event
= info
;
916 struct perf_event_context
*ctx
= event
->ctx
;
917 struct perf_event
*leader
= event
->group_leader
;
918 struct perf_cpu_context
*cpuctx
= __get_cpu_context(ctx
);
922 * If this is a task context, we need to check whether it is
923 * the current task context of this cpu. If not it has been
924 * scheduled out before the smp call arrived.
925 * Or possibly this is the right context but it isn't
926 * on this cpu because it had no events.
928 if (ctx
->task
&& cpuctx
->task_ctx
!= ctx
) {
929 if (cpuctx
->task_ctx
|| ctx
->task
!= current
)
931 cpuctx
->task_ctx
= ctx
;
934 raw_spin_lock(&ctx
->lock
);
936 update_context_time(ctx
);
938 add_event_to_ctx(event
, ctx
);
940 if (event
->cpu
!= -1 && event
->cpu
!= smp_processor_id())
944 * Don't put the event on if it is disabled or if
945 * it is in a group and the group isn't on.
947 if (event
->state
!= PERF_EVENT_STATE_INACTIVE
||
948 (leader
!= event
&& leader
->state
!= PERF_EVENT_STATE_ACTIVE
))
952 * An exclusive event can't go on if there are already active
953 * hardware events, and no hardware event can go on if there
954 * is already an exclusive event on.
956 if (!group_can_go_on(event
, cpuctx
, 1))
959 err
= event_sched_in(event
, cpuctx
, ctx
);
963 * This event couldn't go on. If it is in a group
964 * then we have to pull the whole group off.
965 * If the event group is pinned then put it in error state.
968 group_sched_out(leader
, cpuctx
, ctx
);
969 if (leader
->attr
.pinned
) {
970 update_group_times(leader
);
971 leader
->state
= PERF_EVENT_STATE_ERROR
;
976 raw_spin_unlock(&ctx
->lock
);
980 * Attach a performance event to a context
982 * First we add the event to the list with the hardware enable bit
983 * in event->hw_config cleared.
985 * If the event is attached to a task which is on a CPU we use a smp
986 * call to enable it in the task context. The task might have been
987 * scheduled away, but we check this in the smp call again.
989 * Must be called with ctx->mutex held.
992 perf_install_in_context(struct perf_event_context
*ctx
,
993 struct perf_event
*event
,
996 struct task_struct
*task
= ctx
->task
;
1002 * Per cpu events are installed via an smp call and
1003 * the install is always successful.
1005 smp_call_function_single(cpu
, __perf_install_in_context
,
1011 task_oncpu_function_call(task
, __perf_install_in_context
,
1014 raw_spin_lock_irq(&ctx
->lock
);
1016 * we need to retry the smp call.
1018 if (ctx
->is_active
&& list_empty(&event
->group_entry
)) {
1019 raw_spin_unlock_irq(&ctx
->lock
);
1024 * The lock prevents that this context is scheduled in so we
1025 * can add the event safely, if it the call above did not
1028 if (list_empty(&event
->group_entry
))
1029 add_event_to_ctx(event
, ctx
);
1030 raw_spin_unlock_irq(&ctx
->lock
);
1034 * Put a event into inactive state and update time fields.
1035 * Enabling the leader of a group effectively enables all
1036 * the group members that aren't explicitly disabled, so we
1037 * have to update their ->tstamp_enabled also.
1038 * Note: this works for group members as well as group leaders
1039 * since the non-leader members' sibling_lists will be empty.
1041 static void __perf_event_mark_enabled(struct perf_event
*event
,
1042 struct perf_event_context
*ctx
)
1044 struct perf_event
*sub
;
1046 event
->state
= PERF_EVENT_STATE_INACTIVE
;
1047 event
->tstamp_enabled
= ctx
->time
- event
->total_time_enabled
;
1048 list_for_each_entry(sub
, &event
->sibling_list
, group_entry
) {
1049 if (sub
->state
>= PERF_EVENT_STATE_INACTIVE
) {
1050 sub
->tstamp_enabled
=
1051 ctx
->time
- sub
->total_time_enabled
;
1057 * Cross CPU call to enable a performance event
1059 static void __perf_event_enable(void *info
)
1061 struct perf_event
*event
= info
;
1062 struct perf_event_context
*ctx
= event
->ctx
;
1063 struct perf_event
*leader
= event
->group_leader
;
1064 struct perf_cpu_context
*cpuctx
= __get_cpu_context(ctx
);
1068 * If this is a per-task event, need to check whether this
1069 * event's task is the current task on this cpu.
1071 if (ctx
->task
&& cpuctx
->task_ctx
!= ctx
) {
1072 if (cpuctx
->task_ctx
|| ctx
->task
!= current
)
1074 cpuctx
->task_ctx
= ctx
;
1077 raw_spin_lock(&ctx
->lock
);
1079 update_context_time(ctx
);
1081 if (event
->state
>= PERF_EVENT_STATE_INACTIVE
)
1083 __perf_event_mark_enabled(event
, ctx
);
1085 if (event
->cpu
!= -1 && event
->cpu
!= smp_processor_id())
1089 * If the event is in a group and isn't the group leader,
1090 * then don't put it on unless the group is on.
1092 if (leader
!= event
&& leader
->state
!= PERF_EVENT_STATE_ACTIVE
)
1095 if (!group_can_go_on(event
, cpuctx
, 1)) {
1098 if (event
== leader
)
1099 err
= group_sched_in(event
, cpuctx
, ctx
);
1101 err
= event_sched_in(event
, cpuctx
, ctx
);
1106 * If this event can't go on and it's part of a
1107 * group, then the whole group has to come off.
1109 if (leader
!= event
)
1110 group_sched_out(leader
, cpuctx
, ctx
);
1111 if (leader
->attr
.pinned
) {
1112 update_group_times(leader
);
1113 leader
->state
= PERF_EVENT_STATE_ERROR
;
1118 raw_spin_unlock(&ctx
->lock
);
1124 * If event->ctx is a cloned context, callers must make sure that
1125 * every task struct that event->ctx->task could possibly point to
1126 * remains valid. This condition is satisfied when called through
1127 * perf_event_for_each_child or perf_event_for_each as described
1128 * for perf_event_disable.
1130 void perf_event_enable(struct perf_event
*event
)
1132 struct perf_event_context
*ctx
= event
->ctx
;
1133 struct task_struct
*task
= ctx
->task
;
1137 * Enable the event on the cpu that it's on
1139 smp_call_function_single(event
->cpu
, __perf_event_enable
,
1144 raw_spin_lock_irq(&ctx
->lock
);
1145 if (event
->state
>= PERF_EVENT_STATE_INACTIVE
)
1149 * If the event is in error state, clear that first.
1150 * That way, if we see the event in error state below, we
1151 * know that it has gone back into error state, as distinct
1152 * from the task having been scheduled away before the
1153 * cross-call arrived.
1155 if (event
->state
== PERF_EVENT_STATE_ERROR
)
1156 event
->state
= PERF_EVENT_STATE_OFF
;
1159 raw_spin_unlock_irq(&ctx
->lock
);
1160 task_oncpu_function_call(task
, __perf_event_enable
, event
);
1162 raw_spin_lock_irq(&ctx
->lock
);
1165 * If the context is active and the event is still off,
1166 * we need to retry the cross-call.
1168 if (ctx
->is_active
&& event
->state
== PERF_EVENT_STATE_OFF
)
1172 * Since we have the lock this context can't be scheduled
1173 * in, so we can change the state safely.
1175 if (event
->state
== PERF_EVENT_STATE_OFF
)
1176 __perf_event_mark_enabled(event
, ctx
);
1179 raw_spin_unlock_irq(&ctx
->lock
);
1182 static int perf_event_refresh(struct perf_event
*event
, int refresh
)
1185 * not supported on inherited events
1187 if (event
->attr
.inherit
|| !is_sampling_event(event
))
1190 atomic_add(refresh
, &event
->event_limit
);
1191 perf_event_enable(event
);
1197 EVENT_FLEXIBLE
= 0x1,
1199 EVENT_ALL
= EVENT_FLEXIBLE
| EVENT_PINNED
,
1202 static void ctx_sched_out(struct perf_event_context
*ctx
,
1203 struct perf_cpu_context
*cpuctx
,
1204 enum event_type_t event_type
)
1206 struct perf_event
*event
;
1208 raw_spin_lock(&ctx
->lock
);
1209 perf_pmu_disable(ctx
->pmu
);
1211 if (likely(!ctx
->nr_events
))
1213 update_context_time(ctx
);
1215 if (!ctx
->nr_active
)
1218 if (event_type
& EVENT_PINNED
) {
1219 list_for_each_entry(event
, &ctx
->pinned_groups
, group_entry
)
1220 group_sched_out(event
, cpuctx
, ctx
);
1223 if (event_type
& EVENT_FLEXIBLE
) {
1224 list_for_each_entry(event
, &ctx
->flexible_groups
, group_entry
)
1225 group_sched_out(event
, cpuctx
, ctx
);
1228 perf_pmu_enable(ctx
->pmu
);
1229 raw_spin_unlock(&ctx
->lock
);
1233 * Test whether two contexts are equivalent, i.e. whether they
1234 * have both been cloned from the same version of the same context
1235 * and they both have the same number of enabled events.
1236 * If the number of enabled events is the same, then the set
1237 * of enabled events should be the same, because these are both
1238 * inherited contexts, therefore we can't access individual events
1239 * in them directly with an fd; we can only enable/disable all
1240 * events via prctl, or enable/disable all events in a family
1241 * via ioctl, which will have the same effect on both contexts.
1243 static int context_equiv(struct perf_event_context
*ctx1
,
1244 struct perf_event_context
*ctx2
)
1246 return ctx1
->parent_ctx
&& ctx1
->parent_ctx
== ctx2
->parent_ctx
1247 && ctx1
->parent_gen
== ctx2
->parent_gen
1248 && !ctx1
->pin_count
&& !ctx2
->pin_count
;
1251 static void __perf_event_sync_stat(struct perf_event
*event
,
1252 struct perf_event
*next_event
)
1256 if (!event
->attr
.inherit_stat
)
1260 * Update the event value, we cannot use perf_event_read()
1261 * because we're in the middle of a context switch and have IRQs
1262 * disabled, which upsets smp_call_function_single(), however
1263 * we know the event must be on the current CPU, therefore we
1264 * don't need to use it.
1266 switch (event
->state
) {
1267 case PERF_EVENT_STATE_ACTIVE
:
1268 event
->pmu
->read(event
);
1271 case PERF_EVENT_STATE_INACTIVE
:
1272 update_event_times(event
);
1280 * In order to keep per-task stats reliable we need to flip the event
1281 * values when we flip the contexts.
1283 value
= local64_read(&next_event
->count
);
1284 value
= local64_xchg(&event
->count
, value
);
1285 local64_set(&next_event
->count
, value
);
1287 swap(event
->total_time_enabled
, next_event
->total_time_enabled
);
1288 swap(event
->total_time_running
, next_event
->total_time_running
);
1291 * Since we swizzled the values, update the user visible data too.
1293 perf_event_update_userpage(event
);
1294 perf_event_update_userpage(next_event
);
1297 #define list_next_entry(pos, member) \
1298 list_entry(pos->member.next, typeof(*pos), member)
1300 static void perf_event_sync_stat(struct perf_event_context
*ctx
,
1301 struct perf_event_context
*next_ctx
)
1303 struct perf_event
*event
, *next_event
;
1308 update_context_time(ctx
);
1310 event
= list_first_entry(&ctx
->event_list
,
1311 struct perf_event
, event_entry
);
1313 next_event
= list_first_entry(&next_ctx
->event_list
,
1314 struct perf_event
, event_entry
);
1316 while (&event
->event_entry
!= &ctx
->event_list
&&
1317 &next_event
->event_entry
!= &next_ctx
->event_list
) {
1319 __perf_event_sync_stat(event
, next_event
);
1321 event
= list_next_entry(event
, event_entry
);
1322 next_event
= list_next_entry(next_event
, event_entry
);
1326 void perf_event_context_sched_out(struct task_struct
*task
, int ctxn
,
1327 struct task_struct
*next
)
1329 struct perf_event_context
*ctx
= task
->perf_event_ctxp
[ctxn
];
1330 struct perf_event_context
*next_ctx
;
1331 struct perf_event_context
*parent
;
1332 struct perf_cpu_context
*cpuctx
;
1338 cpuctx
= __get_cpu_context(ctx
);
1339 if (!cpuctx
->task_ctx
)
1343 parent
= rcu_dereference(ctx
->parent_ctx
);
1344 next_ctx
= next
->perf_event_ctxp
[ctxn
];
1345 if (parent
&& next_ctx
&&
1346 rcu_dereference(next_ctx
->parent_ctx
) == parent
) {
1348 * Looks like the two contexts are clones, so we might be
1349 * able to optimize the context switch. We lock both
1350 * contexts and check that they are clones under the
1351 * lock (including re-checking that neither has been
1352 * uncloned in the meantime). It doesn't matter which
1353 * order we take the locks because no other cpu could
1354 * be trying to lock both of these tasks.
1356 raw_spin_lock(&ctx
->lock
);
1357 raw_spin_lock_nested(&next_ctx
->lock
, SINGLE_DEPTH_NESTING
);
1358 if (context_equiv(ctx
, next_ctx
)) {
1360 * XXX do we need a memory barrier of sorts
1361 * wrt to rcu_dereference() of perf_event_ctxp
1363 task
->perf_event_ctxp
[ctxn
] = next_ctx
;
1364 next
->perf_event_ctxp
[ctxn
] = ctx
;
1366 next_ctx
->task
= task
;
1369 perf_event_sync_stat(ctx
, next_ctx
);
1371 raw_spin_unlock(&next_ctx
->lock
);
1372 raw_spin_unlock(&ctx
->lock
);
1377 ctx_sched_out(ctx
, cpuctx
, EVENT_ALL
);
1378 cpuctx
->task_ctx
= NULL
;
1382 #define for_each_task_context_nr(ctxn) \
1383 for ((ctxn) = 0; (ctxn) < perf_nr_task_contexts; (ctxn)++)
1386 * Called from scheduler to remove the events of the current task,
1387 * with interrupts disabled.
1389 * We stop each event and update the event value in event->count.
1391 * This does not protect us against NMI, but disable()
1392 * sets the disabled bit in the control field of event _before_
1393 * accessing the event control register. If a NMI hits, then it will
1394 * not restart the event.
1396 void __perf_event_task_sched_out(struct task_struct
*task
,
1397 struct task_struct
*next
)
1401 for_each_task_context_nr(ctxn
)
1402 perf_event_context_sched_out(task
, ctxn
, next
);
1405 static void task_ctx_sched_out(struct perf_event_context
*ctx
,
1406 enum event_type_t event_type
)
1408 struct perf_cpu_context
*cpuctx
= __get_cpu_context(ctx
);
1410 if (!cpuctx
->task_ctx
)
1413 if (WARN_ON_ONCE(ctx
!= cpuctx
->task_ctx
))
1416 ctx_sched_out(ctx
, cpuctx
, event_type
);
1417 cpuctx
->task_ctx
= NULL
;
1421 * Called with IRQs disabled
1423 static void cpu_ctx_sched_out(struct perf_cpu_context
*cpuctx
,
1424 enum event_type_t event_type
)
1426 ctx_sched_out(&cpuctx
->ctx
, cpuctx
, event_type
);
1430 ctx_pinned_sched_in(struct perf_event_context
*ctx
,
1431 struct perf_cpu_context
*cpuctx
)
1433 struct perf_event
*event
;
1435 list_for_each_entry(event
, &ctx
->pinned_groups
, group_entry
) {
1436 if (event
->state
<= PERF_EVENT_STATE_OFF
)
1438 if (event
->cpu
!= -1 && event
->cpu
!= smp_processor_id())
1441 if (group_can_go_on(event
, cpuctx
, 1))
1442 group_sched_in(event
, cpuctx
, ctx
);
1445 * If this pinned group hasn't been scheduled,
1446 * put it in error state.
1448 if (event
->state
== PERF_EVENT_STATE_INACTIVE
) {
1449 update_group_times(event
);
1450 event
->state
= PERF_EVENT_STATE_ERROR
;
1456 ctx_flexible_sched_in(struct perf_event_context
*ctx
,
1457 struct perf_cpu_context
*cpuctx
)
1459 struct perf_event
*event
;
1462 list_for_each_entry(event
, &ctx
->flexible_groups
, group_entry
) {
1463 /* Ignore events in OFF or ERROR state */
1464 if (event
->state
<= PERF_EVENT_STATE_OFF
)
1467 * Listen to the 'cpu' scheduling filter constraint
1470 if (event
->cpu
!= -1 && event
->cpu
!= smp_processor_id())
1473 if (group_can_go_on(event
, cpuctx
, can_add_hw
)) {
1474 if (group_sched_in(event
, cpuctx
, ctx
))
1481 ctx_sched_in(struct perf_event_context
*ctx
,
1482 struct perf_cpu_context
*cpuctx
,
1483 enum event_type_t event_type
)
1485 raw_spin_lock(&ctx
->lock
);
1487 if (likely(!ctx
->nr_events
))
1490 ctx
->timestamp
= perf_clock();
1493 * First go through the list and put on any pinned groups
1494 * in order to give them the best chance of going on.
1496 if (event_type
& EVENT_PINNED
)
1497 ctx_pinned_sched_in(ctx
, cpuctx
);
1499 /* Then walk through the lower prio flexible groups */
1500 if (event_type
& EVENT_FLEXIBLE
)
1501 ctx_flexible_sched_in(ctx
, cpuctx
);
1504 raw_spin_unlock(&ctx
->lock
);
1507 static void cpu_ctx_sched_in(struct perf_cpu_context
*cpuctx
,
1508 enum event_type_t event_type
)
1510 struct perf_event_context
*ctx
= &cpuctx
->ctx
;
1512 ctx_sched_in(ctx
, cpuctx
, event_type
);
1515 static void task_ctx_sched_in(struct perf_event_context
*ctx
,
1516 enum event_type_t event_type
)
1518 struct perf_cpu_context
*cpuctx
;
1520 cpuctx
= __get_cpu_context(ctx
);
1521 if (cpuctx
->task_ctx
== ctx
)
1524 ctx_sched_in(ctx
, cpuctx
, event_type
);
1525 cpuctx
->task_ctx
= ctx
;
1528 void perf_event_context_sched_in(struct perf_event_context
*ctx
)
1530 struct perf_cpu_context
*cpuctx
;
1532 cpuctx
= __get_cpu_context(ctx
);
1533 if (cpuctx
->task_ctx
== ctx
)
1536 perf_pmu_disable(ctx
->pmu
);
1538 * We want to keep the following priority order:
1539 * cpu pinned (that don't need to move), task pinned,
1540 * cpu flexible, task flexible.
1542 cpu_ctx_sched_out(cpuctx
, EVENT_FLEXIBLE
);
1544 ctx_sched_in(ctx
, cpuctx
, EVENT_PINNED
);
1545 cpu_ctx_sched_in(cpuctx
, EVENT_FLEXIBLE
);
1546 ctx_sched_in(ctx
, cpuctx
, EVENT_FLEXIBLE
);
1548 cpuctx
->task_ctx
= ctx
;
1551 * Since these rotations are per-cpu, we need to ensure the
1552 * cpu-context we got scheduled on is actually rotating.
1554 perf_pmu_rotate_start(ctx
->pmu
);
1555 perf_pmu_enable(ctx
->pmu
);
1559 * Called from scheduler to add the events of the current task
1560 * with interrupts disabled.
1562 * We restore the event value and then enable it.
1564 * This does not protect us against NMI, but enable()
1565 * sets the enabled bit in the control field of event _before_
1566 * accessing the event control register. If a NMI hits, then it will
1567 * keep the event running.
1569 void __perf_event_task_sched_in(struct task_struct
*task
)
1571 struct perf_event_context
*ctx
;
1574 for_each_task_context_nr(ctxn
) {
1575 ctx
= task
->perf_event_ctxp
[ctxn
];
1579 perf_event_context_sched_in(ctx
);
1583 #define MAX_INTERRUPTS (~0ULL)
1585 static void perf_log_throttle(struct perf_event
*event
, int enable
);
1587 static u64
perf_calculate_period(struct perf_event
*event
, u64 nsec
, u64 count
)
1589 u64 frequency
= event
->attr
.sample_freq
;
1590 u64 sec
= NSEC_PER_SEC
;
1591 u64 divisor
, dividend
;
1593 int count_fls
, nsec_fls
, frequency_fls
, sec_fls
;
1595 count_fls
= fls64(count
);
1596 nsec_fls
= fls64(nsec
);
1597 frequency_fls
= fls64(frequency
);
1601 * We got @count in @nsec, with a target of sample_freq HZ
1602 * the target period becomes:
1605 * period = -------------------
1606 * @nsec * sample_freq
1611 * Reduce accuracy by one bit such that @a and @b converge
1612 * to a similar magnitude.
1614 #define REDUCE_FLS(a, b) \
1616 if (a##_fls > b##_fls) { \
1626 * Reduce accuracy until either term fits in a u64, then proceed with
1627 * the other, so that finally we can do a u64/u64 division.
1629 while (count_fls
+ sec_fls
> 64 && nsec_fls
+ frequency_fls
> 64) {
1630 REDUCE_FLS(nsec
, frequency
);
1631 REDUCE_FLS(sec
, count
);
1634 if (count_fls
+ sec_fls
> 64) {
1635 divisor
= nsec
* frequency
;
1637 while (count_fls
+ sec_fls
> 64) {
1638 REDUCE_FLS(count
, sec
);
1642 dividend
= count
* sec
;
1644 dividend
= count
* sec
;
1646 while (nsec_fls
+ frequency_fls
> 64) {
1647 REDUCE_FLS(nsec
, frequency
);
1651 divisor
= nsec
* frequency
;
1657 return div64_u64(dividend
, divisor
);
1660 static void perf_adjust_period(struct perf_event
*event
, u64 nsec
, u64 count
)
1662 struct hw_perf_event
*hwc
= &event
->hw
;
1663 s64 period
, sample_period
;
1666 period
= perf_calculate_period(event
, nsec
, count
);
1668 delta
= (s64
)(period
- hwc
->sample_period
);
1669 delta
= (delta
+ 7) / 8; /* low pass filter */
1671 sample_period
= hwc
->sample_period
+ delta
;
1676 hwc
->sample_period
= sample_period
;
1678 if (local64_read(&hwc
->period_left
) > 8*sample_period
) {
1679 event
->pmu
->stop(event
, PERF_EF_UPDATE
);
1680 local64_set(&hwc
->period_left
, 0);
1681 event
->pmu
->start(event
, PERF_EF_RELOAD
);
1685 static void perf_ctx_adjust_freq(struct perf_event_context
*ctx
, u64 period
)
1687 struct perf_event
*event
;
1688 struct hw_perf_event
*hwc
;
1689 u64 interrupts
, now
;
1692 raw_spin_lock(&ctx
->lock
);
1693 list_for_each_entry_rcu(event
, &ctx
->event_list
, event_entry
) {
1694 if (event
->state
!= PERF_EVENT_STATE_ACTIVE
)
1697 if (event
->cpu
!= -1 && event
->cpu
!= smp_processor_id())
1702 interrupts
= hwc
->interrupts
;
1703 hwc
->interrupts
= 0;
1706 * unthrottle events on the tick
1708 if (interrupts
== MAX_INTERRUPTS
) {
1709 perf_log_throttle(event
, 1);
1710 event
->pmu
->start(event
, 0);
1713 if (!event
->attr
.freq
|| !event
->attr
.sample_freq
)
1716 event
->pmu
->read(event
);
1717 now
= local64_read(&event
->count
);
1718 delta
= now
- hwc
->freq_count_stamp
;
1719 hwc
->freq_count_stamp
= now
;
1722 perf_adjust_period(event
, period
, delta
);
1724 raw_spin_unlock(&ctx
->lock
);
1728 * Round-robin a context's events:
1730 static void rotate_ctx(struct perf_event_context
*ctx
)
1732 raw_spin_lock(&ctx
->lock
);
1735 * Rotate the first entry last of non-pinned groups. Rotation might be
1736 * disabled by the inheritance code.
1738 if (!ctx
->rotate_disable
)
1739 list_rotate_left(&ctx
->flexible_groups
);
1741 raw_spin_unlock(&ctx
->lock
);
1745 * perf_pmu_rotate_start() and perf_rotate_context() are fully serialized
1746 * because they're strictly cpu affine and rotate_start is called with IRQs
1747 * disabled, while rotate_context is called from IRQ context.
1749 static void perf_rotate_context(struct perf_cpu_context
*cpuctx
)
1751 u64 interval
= (u64
)cpuctx
->jiffies_interval
* TICK_NSEC
;
1752 struct perf_event_context
*ctx
= NULL
;
1753 int rotate
= 0, remove
= 1;
1755 if (cpuctx
->ctx
.nr_events
) {
1757 if (cpuctx
->ctx
.nr_events
!= cpuctx
->ctx
.nr_active
)
1761 ctx
= cpuctx
->task_ctx
;
1762 if (ctx
&& ctx
->nr_events
) {
1764 if (ctx
->nr_events
!= ctx
->nr_active
)
1768 perf_pmu_disable(cpuctx
->ctx
.pmu
);
1769 perf_ctx_adjust_freq(&cpuctx
->ctx
, interval
);
1771 perf_ctx_adjust_freq(ctx
, interval
);
1776 cpu_ctx_sched_out(cpuctx
, EVENT_FLEXIBLE
);
1778 task_ctx_sched_out(ctx
, EVENT_FLEXIBLE
);
1780 rotate_ctx(&cpuctx
->ctx
);
1784 cpu_ctx_sched_in(cpuctx
, EVENT_FLEXIBLE
);
1786 task_ctx_sched_in(ctx
, EVENT_FLEXIBLE
);
1790 list_del_init(&cpuctx
->rotation_list
);
1792 perf_pmu_enable(cpuctx
->ctx
.pmu
);
1795 void perf_event_task_tick(void)
1797 struct list_head
*head
= &__get_cpu_var(rotation_list
);
1798 struct perf_cpu_context
*cpuctx
, *tmp
;
1800 WARN_ON(!irqs_disabled());
1802 list_for_each_entry_safe(cpuctx
, tmp
, head
, rotation_list
) {
1803 if (cpuctx
->jiffies_interval
== 1 ||
1804 !(jiffies
% cpuctx
->jiffies_interval
))
1805 perf_rotate_context(cpuctx
);
1809 static int event_enable_on_exec(struct perf_event
*event
,
1810 struct perf_event_context
*ctx
)
1812 if (!event
->attr
.enable_on_exec
)
1815 event
->attr
.enable_on_exec
= 0;
1816 if (event
->state
>= PERF_EVENT_STATE_INACTIVE
)
1819 __perf_event_mark_enabled(event
, ctx
);
1825 * Enable all of a task's events that have been marked enable-on-exec.
1826 * This expects task == current.
1828 static void perf_event_enable_on_exec(struct perf_event_context
*ctx
)
1830 struct perf_event
*event
;
1831 unsigned long flags
;
1835 local_irq_save(flags
);
1836 if (!ctx
|| !ctx
->nr_events
)
1839 task_ctx_sched_out(ctx
, EVENT_ALL
);
1841 raw_spin_lock(&ctx
->lock
);
1843 list_for_each_entry(event
, &ctx
->pinned_groups
, group_entry
) {
1844 ret
= event_enable_on_exec(event
, ctx
);
1849 list_for_each_entry(event
, &ctx
->flexible_groups
, group_entry
) {
1850 ret
= event_enable_on_exec(event
, ctx
);
1856 * Unclone this context if we enabled any event.
1861 raw_spin_unlock(&ctx
->lock
);
1863 perf_event_context_sched_in(ctx
);
1865 local_irq_restore(flags
);
1869 * Cross CPU call to read the hardware event
1871 static void __perf_event_read(void *info
)
1873 struct perf_event
*event
= info
;
1874 struct perf_event_context
*ctx
= event
->ctx
;
1875 struct perf_cpu_context
*cpuctx
= __get_cpu_context(ctx
);
1878 * If this is a task context, we need to check whether it is
1879 * the current task context of this cpu. If not it has been
1880 * scheduled out before the smp call arrived. In that case
1881 * event->count would have been updated to a recent sample
1882 * when the event was scheduled out.
1884 if (ctx
->task
&& cpuctx
->task_ctx
!= ctx
)
1887 raw_spin_lock(&ctx
->lock
);
1888 update_context_time(ctx
);
1889 update_event_times(event
);
1890 raw_spin_unlock(&ctx
->lock
);
1892 event
->pmu
->read(event
);
1895 static inline u64
perf_event_count(struct perf_event
*event
)
1897 return local64_read(&event
->count
) + atomic64_read(&event
->child_count
);
1900 static u64
perf_event_read(struct perf_event
*event
)
1903 * If event is enabled and currently active on a CPU, update the
1904 * value in the event structure:
1906 if (event
->state
== PERF_EVENT_STATE_ACTIVE
) {
1907 smp_call_function_single(event
->oncpu
,
1908 __perf_event_read
, event
, 1);
1909 } else if (event
->state
== PERF_EVENT_STATE_INACTIVE
) {
1910 struct perf_event_context
*ctx
= event
->ctx
;
1911 unsigned long flags
;
1913 raw_spin_lock_irqsave(&ctx
->lock
, flags
);
1915 * may read while context is not active
1916 * (e.g., thread is blocked), in that case
1917 * we cannot update context time
1920 update_context_time(ctx
);
1921 update_event_times(event
);
1922 raw_spin_unlock_irqrestore(&ctx
->lock
, flags
);
1925 return perf_event_count(event
);
1932 struct callchain_cpus_entries
{
1933 struct rcu_head rcu_head
;
1934 struct perf_callchain_entry
*cpu_entries
[0];
1937 static DEFINE_PER_CPU(int, callchain_recursion
[PERF_NR_CONTEXTS
]);
1938 static atomic_t nr_callchain_events
;
1939 static DEFINE_MUTEX(callchain_mutex
);
1940 struct callchain_cpus_entries
*callchain_cpus_entries
;
1943 __weak
void perf_callchain_kernel(struct perf_callchain_entry
*entry
,
1944 struct pt_regs
*regs
)
1948 __weak
void perf_callchain_user(struct perf_callchain_entry
*entry
,
1949 struct pt_regs
*regs
)
1953 static void release_callchain_buffers_rcu(struct rcu_head
*head
)
1955 struct callchain_cpus_entries
*entries
;
1958 entries
= container_of(head
, struct callchain_cpus_entries
, rcu_head
);
1960 for_each_possible_cpu(cpu
)
1961 kfree(entries
->cpu_entries
[cpu
]);
1966 static void release_callchain_buffers(void)
1968 struct callchain_cpus_entries
*entries
;
1970 entries
= callchain_cpus_entries
;
1971 rcu_assign_pointer(callchain_cpus_entries
, NULL
);
1972 call_rcu(&entries
->rcu_head
, release_callchain_buffers_rcu
);
1975 static int alloc_callchain_buffers(void)
1979 struct callchain_cpus_entries
*entries
;
1982 * We can't use the percpu allocation API for data that can be
1983 * accessed from NMI. Use a temporary manual per cpu allocation
1984 * until that gets sorted out.
1986 size
= sizeof(*entries
) + sizeof(struct perf_callchain_entry
*) *
1987 num_possible_cpus();
1989 entries
= kzalloc(size
, GFP_KERNEL
);
1993 size
= sizeof(struct perf_callchain_entry
) * PERF_NR_CONTEXTS
;
1995 for_each_possible_cpu(cpu
) {
1996 entries
->cpu_entries
[cpu
] = kmalloc_node(size
, GFP_KERNEL
,
1998 if (!entries
->cpu_entries
[cpu
])
2002 rcu_assign_pointer(callchain_cpus_entries
, entries
);
2007 for_each_possible_cpu(cpu
)
2008 kfree(entries
->cpu_entries
[cpu
]);
2014 static int get_callchain_buffers(void)
2019 mutex_lock(&callchain_mutex
);
2021 count
= atomic_inc_return(&nr_callchain_events
);
2022 if (WARN_ON_ONCE(count
< 1)) {
2028 /* If the allocation failed, give up */
2029 if (!callchain_cpus_entries
)
2034 err
= alloc_callchain_buffers();
2036 release_callchain_buffers();
2038 mutex_unlock(&callchain_mutex
);
2043 static void put_callchain_buffers(void)
2045 if (atomic_dec_and_mutex_lock(&nr_callchain_events
, &callchain_mutex
)) {
2046 release_callchain_buffers();
2047 mutex_unlock(&callchain_mutex
);
2051 static int get_recursion_context(int *recursion
)
2059 else if (in_softirq())
2064 if (recursion
[rctx
])
2073 static inline void put_recursion_context(int *recursion
, int rctx
)
2079 static struct perf_callchain_entry
*get_callchain_entry(int *rctx
)
2082 struct callchain_cpus_entries
*entries
;
2084 *rctx
= get_recursion_context(__get_cpu_var(callchain_recursion
));
2088 entries
= rcu_dereference(callchain_cpus_entries
);
2092 cpu
= smp_processor_id();
2094 return &entries
->cpu_entries
[cpu
][*rctx
];
2098 put_callchain_entry(int rctx
)
2100 put_recursion_context(__get_cpu_var(callchain_recursion
), rctx
);
2103 static struct perf_callchain_entry
*perf_callchain(struct pt_regs
*regs
)
2106 struct perf_callchain_entry
*entry
;
2109 entry
= get_callchain_entry(&rctx
);
2118 if (!user_mode(regs
)) {
2119 perf_callchain_store(entry
, PERF_CONTEXT_KERNEL
);
2120 perf_callchain_kernel(entry
, regs
);
2122 regs
= task_pt_regs(current
);
2128 perf_callchain_store(entry
, PERF_CONTEXT_USER
);
2129 perf_callchain_user(entry
, regs
);
2133 put_callchain_entry(rctx
);
2139 * Initialize the perf_event context in a task_struct:
2141 static void __perf_event_init_context(struct perf_event_context
*ctx
)
2143 raw_spin_lock_init(&ctx
->lock
);
2144 mutex_init(&ctx
->mutex
);
2145 INIT_LIST_HEAD(&ctx
->pinned_groups
);
2146 INIT_LIST_HEAD(&ctx
->flexible_groups
);
2147 INIT_LIST_HEAD(&ctx
->event_list
);
2148 atomic_set(&ctx
->refcount
, 1);
2151 static struct perf_event_context
*
2152 alloc_perf_context(struct pmu
*pmu
, struct task_struct
*task
)
2154 struct perf_event_context
*ctx
;
2156 ctx
= kzalloc(sizeof(struct perf_event_context
), GFP_KERNEL
);
2160 __perf_event_init_context(ctx
);
2163 get_task_struct(task
);
2170 static struct task_struct
*
2171 find_lively_task_by_vpid(pid_t vpid
)
2173 struct task_struct
*task
;
2180 task
= find_task_by_vpid(vpid
);
2182 get_task_struct(task
);
2186 return ERR_PTR(-ESRCH
);
2189 * Can't attach events to a dying task.
2192 if (task
->flags
& PF_EXITING
)
2195 /* Reuse ptrace permission checks for now. */
2197 if (!ptrace_may_access(task
, PTRACE_MODE_READ
))
2202 put_task_struct(task
);
2203 return ERR_PTR(err
);
2207 static struct perf_event_context
*
2208 find_get_context(struct pmu
*pmu
, struct task_struct
*task
, int cpu
)
2210 struct perf_event_context
*ctx
;
2211 struct perf_cpu_context
*cpuctx
;
2212 unsigned long flags
;
2215 if (!task
&& cpu
!= -1) {
2216 /* Must be root to operate on a CPU event: */
2217 if (perf_paranoid_cpu() && !capable(CAP_SYS_ADMIN
))
2218 return ERR_PTR(-EACCES
);
2220 if (cpu
< 0 || cpu
>= nr_cpumask_bits
)
2221 return ERR_PTR(-EINVAL
);
2224 * We could be clever and allow to attach a event to an
2225 * offline CPU and activate it when the CPU comes up, but
2228 if (!cpu_online(cpu
))
2229 return ERR_PTR(-ENODEV
);
2231 cpuctx
= per_cpu_ptr(pmu
->pmu_cpu_context
, cpu
);
2239 ctxn
= pmu
->task_ctx_nr
;
2244 ctx
= perf_lock_task_context(task
, ctxn
, &flags
);
2247 raw_spin_unlock_irqrestore(&ctx
->lock
, flags
);
2251 ctx
= alloc_perf_context(pmu
, task
);
2258 if (cmpxchg(&task
->perf_event_ctxp
[ctxn
], NULL
, ctx
)) {
2260 * We raced with some other task; use
2261 * the context they set.
2263 put_task_struct(task
);
2272 return ERR_PTR(err
);
2275 static void perf_event_free_filter(struct perf_event
*event
);
2277 static void free_event_rcu(struct rcu_head
*head
)
2279 struct perf_event
*event
;
2281 event
= container_of(head
, struct perf_event
, rcu_head
);
2283 put_pid_ns(event
->ns
);
2284 perf_event_free_filter(event
);
2288 static void perf_buffer_put(struct perf_buffer
*buffer
);
2290 static void free_event(struct perf_event
*event
)
2292 irq_work_sync(&event
->pending
);
2294 if (!event
->parent
) {
2295 if (event
->attach_state
& PERF_ATTACH_TASK
)
2296 jump_label_dec(&perf_task_events
);
2297 if (event
->attr
.mmap
|| event
->attr
.mmap_data
)
2298 atomic_dec(&nr_mmap_events
);
2299 if (event
->attr
.comm
)
2300 atomic_dec(&nr_comm_events
);
2301 if (event
->attr
.task
)
2302 atomic_dec(&nr_task_events
);
2303 if (event
->attr
.sample_type
& PERF_SAMPLE_CALLCHAIN
)
2304 put_callchain_buffers();
2307 if (event
->buffer
) {
2308 perf_buffer_put(event
->buffer
);
2309 event
->buffer
= NULL
;
2313 event
->destroy(event
);
2316 put_ctx(event
->ctx
);
2318 call_rcu(&event
->rcu_head
, free_event_rcu
);
2321 int perf_event_release_kernel(struct perf_event
*event
)
2323 struct perf_event_context
*ctx
= event
->ctx
;
2326 * Remove from the PMU, can't get re-enabled since we got
2327 * here because the last ref went.
2329 perf_event_disable(event
);
2331 WARN_ON_ONCE(ctx
->parent_ctx
);
2333 * There are two ways this annotation is useful:
2335 * 1) there is a lock recursion from perf_event_exit_task
2336 * see the comment there.
2338 * 2) there is a lock-inversion with mmap_sem through
2339 * perf_event_read_group(), which takes faults while
2340 * holding ctx->mutex, however this is called after
2341 * the last filedesc died, so there is no possibility
2342 * to trigger the AB-BA case.
2344 mutex_lock_nested(&ctx
->mutex
, SINGLE_DEPTH_NESTING
);
2345 raw_spin_lock_irq(&ctx
->lock
);
2346 perf_group_detach(event
);
2347 list_del_event(event
, ctx
);
2348 raw_spin_unlock_irq(&ctx
->lock
);
2349 mutex_unlock(&ctx
->mutex
);
2355 EXPORT_SYMBOL_GPL(perf_event_release_kernel
);
2358 * Called when the last reference to the file is gone.
2360 static int perf_release(struct inode
*inode
, struct file
*file
)
2362 struct perf_event
*event
= file
->private_data
;
2363 struct task_struct
*owner
;
2365 file
->private_data
= NULL
;
2368 owner
= ACCESS_ONCE(event
->owner
);
2370 * Matches the smp_wmb() in perf_event_exit_task(). If we observe
2371 * !owner it means the list deletion is complete and we can indeed
2372 * free this event, otherwise we need to serialize on
2373 * owner->perf_event_mutex.
2375 smp_read_barrier_depends();
2378 * Since delayed_put_task_struct() also drops the last
2379 * task reference we can safely take a new reference
2380 * while holding the rcu_read_lock().
2382 get_task_struct(owner
);
2387 mutex_lock(&owner
->perf_event_mutex
);
2389 * We have to re-check the event->owner field, if it is cleared
2390 * we raced with perf_event_exit_task(), acquiring the mutex
2391 * ensured they're done, and we can proceed with freeing the
2395 list_del_init(&event
->owner_entry
);
2396 mutex_unlock(&owner
->perf_event_mutex
);
2397 put_task_struct(owner
);
2400 return perf_event_release_kernel(event
);
2403 u64
perf_event_read_value(struct perf_event
*event
, u64
*enabled
, u64
*running
)
2405 struct perf_event
*child
;
2411 mutex_lock(&event
->child_mutex
);
2412 total
+= perf_event_read(event
);
2413 *enabled
+= event
->total_time_enabled
+
2414 atomic64_read(&event
->child_total_time_enabled
);
2415 *running
+= event
->total_time_running
+
2416 atomic64_read(&event
->child_total_time_running
);
2418 list_for_each_entry(child
, &event
->child_list
, child_list
) {
2419 total
+= perf_event_read(child
);
2420 *enabled
+= child
->total_time_enabled
;
2421 *running
+= child
->total_time_running
;
2423 mutex_unlock(&event
->child_mutex
);
2427 EXPORT_SYMBOL_GPL(perf_event_read_value
);
2429 static int perf_event_read_group(struct perf_event
*event
,
2430 u64 read_format
, char __user
*buf
)
2432 struct perf_event
*leader
= event
->group_leader
, *sub
;
2433 int n
= 0, size
= 0, ret
= -EFAULT
;
2434 struct perf_event_context
*ctx
= leader
->ctx
;
2436 u64 count
, enabled
, running
;
2438 mutex_lock(&ctx
->mutex
);
2439 count
= perf_event_read_value(leader
, &enabled
, &running
);
2441 values
[n
++] = 1 + leader
->nr_siblings
;
2442 if (read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
)
2443 values
[n
++] = enabled
;
2444 if (read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
)
2445 values
[n
++] = running
;
2446 values
[n
++] = count
;
2447 if (read_format
& PERF_FORMAT_ID
)
2448 values
[n
++] = primary_event_id(leader
);
2450 size
= n
* sizeof(u64
);
2452 if (copy_to_user(buf
, values
, size
))
2457 list_for_each_entry(sub
, &leader
->sibling_list
, group_entry
) {
2460 values
[n
++] = perf_event_read_value(sub
, &enabled
, &running
);
2461 if (read_format
& PERF_FORMAT_ID
)
2462 values
[n
++] = primary_event_id(sub
);
2464 size
= n
* sizeof(u64
);
2466 if (copy_to_user(buf
+ ret
, values
, size
)) {
2474 mutex_unlock(&ctx
->mutex
);
2479 static int perf_event_read_one(struct perf_event
*event
,
2480 u64 read_format
, char __user
*buf
)
2482 u64 enabled
, running
;
2486 values
[n
++] = perf_event_read_value(event
, &enabled
, &running
);
2487 if (read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
)
2488 values
[n
++] = enabled
;
2489 if (read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
)
2490 values
[n
++] = running
;
2491 if (read_format
& PERF_FORMAT_ID
)
2492 values
[n
++] = primary_event_id(event
);
2494 if (copy_to_user(buf
, values
, n
* sizeof(u64
)))
2497 return n
* sizeof(u64
);
2501 * Read the performance event - simple non blocking version for now
2504 perf_read_hw(struct perf_event
*event
, char __user
*buf
, size_t count
)
2506 u64 read_format
= event
->attr
.read_format
;
2510 * Return end-of-file for a read on a event that is in
2511 * error state (i.e. because it was pinned but it couldn't be
2512 * scheduled on to the CPU at some point).
2514 if (event
->state
== PERF_EVENT_STATE_ERROR
)
2517 if (count
< event
->read_size
)
2520 WARN_ON_ONCE(event
->ctx
->parent_ctx
);
2521 if (read_format
& PERF_FORMAT_GROUP
)
2522 ret
= perf_event_read_group(event
, read_format
, buf
);
2524 ret
= perf_event_read_one(event
, read_format
, buf
);
2530 perf_read(struct file
*file
, char __user
*buf
, size_t count
, loff_t
*ppos
)
2532 struct perf_event
*event
= file
->private_data
;
2534 return perf_read_hw(event
, buf
, count
);
2537 static unsigned int perf_poll(struct file
*file
, poll_table
*wait
)
2539 struct perf_event
*event
= file
->private_data
;
2540 struct perf_buffer
*buffer
;
2541 unsigned int events
= POLL_HUP
;
2544 buffer
= rcu_dereference(event
->buffer
);
2546 events
= atomic_xchg(&buffer
->poll
, 0);
2549 poll_wait(file
, &event
->waitq
, wait
);
2554 static void perf_event_reset(struct perf_event
*event
)
2556 (void)perf_event_read(event
);
2557 local64_set(&event
->count
, 0);
2558 perf_event_update_userpage(event
);
2562 * Holding the top-level event's child_mutex means that any
2563 * descendant process that has inherited this event will block
2564 * in sync_child_event if it goes to exit, thus satisfying the
2565 * task existence requirements of perf_event_enable/disable.
2567 static void perf_event_for_each_child(struct perf_event
*event
,
2568 void (*func
)(struct perf_event
*))
2570 struct perf_event
*child
;
2572 WARN_ON_ONCE(event
->ctx
->parent_ctx
);
2573 mutex_lock(&event
->child_mutex
);
2575 list_for_each_entry(child
, &event
->child_list
, child_list
)
2577 mutex_unlock(&event
->child_mutex
);
2580 static void perf_event_for_each(struct perf_event
*event
,
2581 void (*func
)(struct perf_event
*))
2583 struct perf_event_context
*ctx
= event
->ctx
;
2584 struct perf_event
*sibling
;
2586 WARN_ON_ONCE(ctx
->parent_ctx
);
2587 mutex_lock(&ctx
->mutex
);
2588 event
= event
->group_leader
;
2590 perf_event_for_each_child(event
, func
);
2592 list_for_each_entry(sibling
, &event
->sibling_list
, group_entry
)
2593 perf_event_for_each_child(event
, func
);
2594 mutex_unlock(&ctx
->mutex
);
2597 static int perf_event_period(struct perf_event
*event
, u64 __user
*arg
)
2599 struct perf_event_context
*ctx
= event
->ctx
;
2603 if (!is_sampling_event(event
))
2606 if (copy_from_user(&value
, arg
, sizeof(value
)))
2612 raw_spin_lock_irq(&ctx
->lock
);
2613 if (event
->attr
.freq
) {
2614 if (value
> sysctl_perf_event_sample_rate
) {
2619 event
->attr
.sample_freq
= value
;
2621 event
->attr
.sample_period
= value
;
2622 event
->hw
.sample_period
= value
;
2625 raw_spin_unlock_irq(&ctx
->lock
);
2630 static const struct file_operations perf_fops
;
2632 static struct perf_event
*perf_fget_light(int fd
, int *fput_needed
)
2636 file
= fget_light(fd
, fput_needed
);
2638 return ERR_PTR(-EBADF
);
2640 if (file
->f_op
!= &perf_fops
) {
2641 fput_light(file
, *fput_needed
);
2643 return ERR_PTR(-EBADF
);
2646 return file
->private_data
;
2649 static int perf_event_set_output(struct perf_event
*event
,
2650 struct perf_event
*output_event
);
2651 static int perf_event_set_filter(struct perf_event
*event
, void __user
*arg
);
2653 static long perf_ioctl(struct file
*file
, unsigned int cmd
, unsigned long arg
)
2655 struct perf_event
*event
= file
->private_data
;
2656 void (*func
)(struct perf_event
*);
2660 case PERF_EVENT_IOC_ENABLE
:
2661 func
= perf_event_enable
;
2663 case PERF_EVENT_IOC_DISABLE
:
2664 func
= perf_event_disable
;
2666 case PERF_EVENT_IOC_RESET
:
2667 func
= perf_event_reset
;
2670 case PERF_EVENT_IOC_REFRESH
:
2671 return perf_event_refresh(event
, arg
);
2673 case PERF_EVENT_IOC_PERIOD
:
2674 return perf_event_period(event
, (u64 __user
*)arg
);
2676 case PERF_EVENT_IOC_SET_OUTPUT
:
2678 struct perf_event
*output_event
= NULL
;
2679 int fput_needed
= 0;
2683 output_event
= perf_fget_light(arg
, &fput_needed
);
2684 if (IS_ERR(output_event
))
2685 return PTR_ERR(output_event
);
2688 ret
= perf_event_set_output(event
, output_event
);
2690 fput_light(output_event
->filp
, fput_needed
);
2695 case PERF_EVENT_IOC_SET_FILTER
:
2696 return perf_event_set_filter(event
, (void __user
*)arg
);
2702 if (flags
& PERF_IOC_FLAG_GROUP
)
2703 perf_event_for_each(event
, func
);
2705 perf_event_for_each_child(event
, func
);
2710 int perf_event_task_enable(void)
2712 struct perf_event
*event
;
2714 mutex_lock(¤t
->perf_event_mutex
);
2715 list_for_each_entry(event
, ¤t
->perf_event_list
, owner_entry
)
2716 perf_event_for_each_child(event
, perf_event_enable
);
2717 mutex_unlock(¤t
->perf_event_mutex
);
2722 int perf_event_task_disable(void)
2724 struct perf_event
*event
;
2726 mutex_lock(¤t
->perf_event_mutex
);
2727 list_for_each_entry(event
, ¤t
->perf_event_list
, owner_entry
)
2728 perf_event_for_each_child(event
, perf_event_disable
);
2729 mutex_unlock(¤t
->perf_event_mutex
);
2734 #ifndef PERF_EVENT_INDEX_OFFSET
2735 # define PERF_EVENT_INDEX_OFFSET 0
2738 static int perf_event_index(struct perf_event
*event
)
2740 if (event
->hw
.state
& PERF_HES_STOPPED
)
2743 if (event
->state
!= PERF_EVENT_STATE_ACTIVE
)
2746 return event
->hw
.idx
+ 1 - PERF_EVENT_INDEX_OFFSET
;
2750 * Callers need to ensure there can be no nesting of this function, otherwise
2751 * the seqlock logic goes bad. We can not serialize this because the arch
2752 * code calls this from NMI context.
2754 void perf_event_update_userpage(struct perf_event
*event
)
2756 struct perf_event_mmap_page
*userpg
;
2757 struct perf_buffer
*buffer
;
2760 buffer
= rcu_dereference(event
->buffer
);
2764 userpg
= buffer
->user_page
;
2767 * Disable preemption so as to not let the corresponding user-space
2768 * spin too long if we get preempted.
2773 userpg
->index
= perf_event_index(event
);
2774 userpg
->offset
= perf_event_count(event
);
2775 if (event
->state
== PERF_EVENT_STATE_ACTIVE
)
2776 userpg
->offset
-= local64_read(&event
->hw
.prev_count
);
2778 userpg
->time_enabled
= event
->total_time_enabled
+
2779 atomic64_read(&event
->child_total_time_enabled
);
2781 userpg
->time_running
= event
->total_time_running
+
2782 atomic64_read(&event
->child_total_time_running
);
2791 static unsigned long perf_data_size(struct perf_buffer
*buffer
);
2794 perf_buffer_init(struct perf_buffer
*buffer
, long watermark
, int flags
)
2796 long max_size
= perf_data_size(buffer
);
2799 buffer
->watermark
= min(max_size
, watermark
);
2801 if (!buffer
->watermark
)
2802 buffer
->watermark
= max_size
/ 2;
2804 if (flags
& PERF_BUFFER_WRITABLE
)
2805 buffer
->writable
= 1;
2807 atomic_set(&buffer
->refcount
, 1);
2810 #ifndef CONFIG_PERF_USE_VMALLOC
2813 * Back perf_mmap() with regular GFP_KERNEL-0 pages.
2816 static struct page
*
2817 perf_mmap_to_page(struct perf_buffer
*buffer
, unsigned long pgoff
)
2819 if (pgoff
> buffer
->nr_pages
)
2823 return virt_to_page(buffer
->user_page
);
2825 return virt_to_page(buffer
->data_pages
[pgoff
- 1]);
2828 static void *perf_mmap_alloc_page(int cpu
)
2833 node
= (cpu
== -1) ? cpu
: cpu_to_node(cpu
);
2834 page
= alloc_pages_node(node
, GFP_KERNEL
| __GFP_ZERO
, 0);
2838 return page_address(page
);
2841 static struct perf_buffer
*
2842 perf_buffer_alloc(int nr_pages
, long watermark
, int cpu
, int flags
)
2844 struct perf_buffer
*buffer
;
2848 size
= sizeof(struct perf_buffer
);
2849 size
+= nr_pages
* sizeof(void *);
2851 buffer
= kzalloc(size
, GFP_KERNEL
);
2855 buffer
->user_page
= perf_mmap_alloc_page(cpu
);
2856 if (!buffer
->user_page
)
2857 goto fail_user_page
;
2859 for (i
= 0; i
< nr_pages
; i
++) {
2860 buffer
->data_pages
[i
] = perf_mmap_alloc_page(cpu
);
2861 if (!buffer
->data_pages
[i
])
2862 goto fail_data_pages
;
2865 buffer
->nr_pages
= nr_pages
;
2867 perf_buffer_init(buffer
, watermark
, flags
);
2872 for (i
--; i
>= 0; i
--)
2873 free_page((unsigned long)buffer
->data_pages
[i
]);
2875 free_page((unsigned long)buffer
->user_page
);
2884 static void perf_mmap_free_page(unsigned long addr
)
2886 struct page
*page
= virt_to_page((void *)addr
);
2888 page
->mapping
= NULL
;
2892 static void perf_buffer_free(struct perf_buffer
*buffer
)
2896 perf_mmap_free_page((unsigned long)buffer
->user_page
);
2897 for (i
= 0; i
< buffer
->nr_pages
; i
++)
2898 perf_mmap_free_page((unsigned long)buffer
->data_pages
[i
]);
2902 static inline int page_order(struct perf_buffer
*buffer
)
2910 * Back perf_mmap() with vmalloc memory.
2912 * Required for architectures that have d-cache aliasing issues.
2915 static inline int page_order(struct perf_buffer
*buffer
)
2917 return buffer
->page_order
;
2920 static struct page
*
2921 perf_mmap_to_page(struct perf_buffer
*buffer
, unsigned long pgoff
)
2923 if (pgoff
> (1UL << page_order(buffer
)))
2926 return vmalloc_to_page((void *)buffer
->user_page
+ pgoff
* PAGE_SIZE
);
2929 static void perf_mmap_unmark_page(void *addr
)
2931 struct page
*page
= vmalloc_to_page(addr
);
2933 page
->mapping
= NULL
;
2936 static void perf_buffer_free_work(struct work_struct
*work
)
2938 struct perf_buffer
*buffer
;
2942 buffer
= container_of(work
, struct perf_buffer
, work
);
2943 nr
= 1 << page_order(buffer
);
2945 base
= buffer
->user_page
;
2946 for (i
= 0; i
< nr
+ 1; i
++)
2947 perf_mmap_unmark_page(base
+ (i
* PAGE_SIZE
));
2953 static void perf_buffer_free(struct perf_buffer
*buffer
)
2955 schedule_work(&buffer
->work
);
2958 static struct perf_buffer
*
2959 perf_buffer_alloc(int nr_pages
, long watermark
, int cpu
, int flags
)
2961 struct perf_buffer
*buffer
;
2965 size
= sizeof(struct perf_buffer
);
2966 size
+= sizeof(void *);
2968 buffer
= kzalloc(size
, GFP_KERNEL
);
2972 INIT_WORK(&buffer
->work
, perf_buffer_free_work
);
2974 all_buf
= vmalloc_user((nr_pages
+ 1) * PAGE_SIZE
);
2978 buffer
->user_page
= all_buf
;
2979 buffer
->data_pages
[0] = all_buf
+ PAGE_SIZE
;
2980 buffer
->page_order
= ilog2(nr_pages
);
2981 buffer
->nr_pages
= 1;
2983 perf_buffer_init(buffer
, watermark
, flags
);
2996 static unsigned long perf_data_size(struct perf_buffer
*buffer
)
2998 return buffer
->nr_pages
<< (PAGE_SHIFT
+ page_order(buffer
));
3001 static int perf_mmap_fault(struct vm_area_struct
*vma
, struct vm_fault
*vmf
)
3003 struct perf_event
*event
= vma
->vm_file
->private_data
;
3004 struct perf_buffer
*buffer
;
3005 int ret
= VM_FAULT_SIGBUS
;
3007 if (vmf
->flags
& FAULT_FLAG_MKWRITE
) {
3008 if (vmf
->pgoff
== 0)
3014 buffer
= rcu_dereference(event
->buffer
);
3018 if (vmf
->pgoff
&& (vmf
->flags
& FAULT_FLAG_WRITE
))
3021 vmf
->page
= perf_mmap_to_page(buffer
, vmf
->pgoff
);
3025 get_page(vmf
->page
);
3026 vmf
->page
->mapping
= vma
->vm_file
->f_mapping
;
3027 vmf
->page
->index
= vmf
->pgoff
;
3036 static void perf_buffer_free_rcu(struct rcu_head
*rcu_head
)
3038 struct perf_buffer
*buffer
;
3040 buffer
= container_of(rcu_head
, struct perf_buffer
, rcu_head
);
3041 perf_buffer_free(buffer
);
3044 static struct perf_buffer
*perf_buffer_get(struct perf_event
*event
)
3046 struct perf_buffer
*buffer
;
3049 buffer
= rcu_dereference(event
->buffer
);
3051 if (!atomic_inc_not_zero(&buffer
->refcount
))
3059 static void perf_buffer_put(struct perf_buffer
*buffer
)
3061 if (!atomic_dec_and_test(&buffer
->refcount
))
3064 call_rcu(&buffer
->rcu_head
, perf_buffer_free_rcu
);
3067 static void perf_mmap_open(struct vm_area_struct
*vma
)
3069 struct perf_event
*event
= vma
->vm_file
->private_data
;
3071 atomic_inc(&event
->mmap_count
);
3074 static void perf_mmap_close(struct vm_area_struct
*vma
)
3076 struct perf_event
*event
= vma
->vm_file
->private_data
;
3078 if (atomic_dec_and_mutex_lock(&event
->mmap_count
, &event
->mmap_mutex
)) {
3079 unsigned long size
= perf_data_size(event
->buffer
);
3080 struct user_struct
*user
= event
->mmap_user
;
3081 struct perf_buffer
*buffer
= event
->buffer
;
3083 atomic_long_sub((size
>> PAGE_SHIFT
) + 1, &user
->locked_vm
);
3084 vma
->vm_mm
->locked_vm
-= event
->mmap_locked
;
3085 rcu_assign_pointer(event
->buffer
, NULL
);
3086 mutex_unlock(&event
->mmap_mutex
);
3088 perf_buffer_put(buffer
);
3093 static const struct vm_operations_struct perf_mmap_vmops
= {
3094 .open
= perf_mmap_open
,
3095 .close
= perf_mmap_close
,
3096 .fault
= perf_mmap_fault
,
3097 .page_mkwrite
= perf_mmap_fault
,
3100 static int perf_mmap(struct file
*file
, struct vm_area_struct
*vma
)
3102 struct perf_event
*event
= file
->private_data
;
3103 unsigned long user_locked
, user_lock_limit
;
3104 struct user_struct
*user
= current_user();
3105 unsigned long locked
, lock_limit
;
3106 struct perf_buffer
*buffer
;
3107 unsigned long vma_size
;
3108 unsigned long nr_pages
;
3109 long user_extra
, extra
;
3110 int ret
= 0, flags
= 0;
3113 * Don't allow mmap() of inherited per-task counters. This would
3114 * create a performance issue due to all children writing to the
3117 if (event
->cpu
== -1 && event
->attr
.inherit
)
3120 if (!(vma
->vm_flags
& VM_SHARED
))
3123 vma_size
= vma
->vm_end
- vma
->vm_start
;
3124 nr_pages
= (vma_size
/ PAGE_SIZE
) - 1;
3127 * If we have buffer pages ensure they're a power-of-two number, so we
3128 * can do bitmasks instead of modulo.
3130 if (nr_pages
!= 0 && !is_power_of_2(nr_pages
))
3133 if (vma_size
!= PAGE_SIZE
* (1 + nr_pages
))
3136 if (vma
->vm_pgoff
!= 0)
3139 WARN_ON_ONCE(event
->ctx
->parent_ctx
);
3140 mutex_lock(&event
->mmap_mutex
);
3141 if (event
->buffer
) {
3142 if (event
->buffer
->nr_pages
== nr_pages
)
3143 atomic_inc(&event
->buffer
->refcount
);
3149 user_extra
= nr_pages
+ 1;
3150 user_lock_limit
= sysctl_perf_event_mlock
>> (PAGE_SHIFT
- 10);
3153 * Increase the limit linearly with more CPUs:
3155 user_lock_limit
*= num_online_cpus();
3157 user_locked
= atomic_long_read(&user
->locked_vm
) + user_extra
;
3160 if (user_locked
> user_lock_limit
)
3161 extra
= user_locked
- user_lock_limit
;
3163 lock_limit
= rlimit(RLIMIT_MEMLOCK
);
3164 lock_limit
>>= PAGE_SHIFT
;
3165 locked
= vma
->vm_mm
->locked_vm
+ extra
;
3167 if ((locked
> lock_limit
) && perf_paranoid_tracepoint_raw() &&
3168 !capable(CAP_IPC_LOCK
)) {
3173 WARN_ON(event
->buffer
);
3175 if (vma
->vm_flags
& VM_WRITE
)
3176 flags
|= PERF_BUFFER_WRITABLE
;
3178 buffer
= perf_buffer_alloc(nr_pages
, event
->attr
.wakeup_watermark
,
3184 rcu_assign_pointer(event
->buffer
, buffer
);
3186 atomic_long_add(user_extra
, &user
->locked_vm
);
3187 event
->mmap_locked
= extra
;
3188 event
->mmap_user
= get_current_user();
3189 vma
->vm_mm
->locked_vm
+= event
->mmap_locked
;
3193 atomic_inc(&event
->mmap_count
);
3194 mutex_unlock(&event
->mmap_mutex
);
3196 vma
->vm_flags
|= VM_RESERVED
;
3197 vma
->vm_ops
= &perf_mmap_vmops
;
3202 static int perf_fasync(int fd
, struct file
*filp
, int on
)
3204 struct inode
*inode
= filp
->f_path
.dentry
->d_inode
;
3205 struct perf_event
*event
= filp
->private_data
;
3208 mutex_lock(&inode
->i_mutex
);
3209 retval
= fasync_helper(fd
, filp
, on
, &event
->fasync
);
3210 mutex_unlock(&inode
->i_mutex
);
3218 static const struct file_operations perf_fops
= {
3219 .llseek
= no_llseek
,
3220 .release
= perf_release
,
3223 .unlocked_ioctl
= perf_ioctl
,
3224 .compat_ioctl
= perf_ioctl
,
3226 .fasync
= perf_fasync
,
3232 * If there's data, ensure we set the poll() state and publish everything
3233 * to user-space before waking everybody up.
3236 void perf_event_wakeup(struct perf_event
*event
)
3238 wake_up_all(&event
->waitq
);
3240 if (event
->pending_kill
) {
3241 kill_fasync(&event
->fasync
, SIGIO
, event
->pending_kill
);
3242 event
->pending_kill
= 0;
3246 static void perf_pending_event(struct irq_work
*entry
)
3248 struct perf_event
*event
= container_of(entry
,
3249 struct perf_event
, pending
);
3251 if (event
->pending_disable
) {
3252 event
->pending_disable
= 0;
3253 __perf_event_disable(event
);
3256 if (event
->pending_wakeup
) {
3257 event
->pending_wakeup
= 0;
3258 perf_event_wakeup(event
);
3263 * We assume there is only KVM supporting the callbacks.
3264 * Later on, we might change it to a list if there is
3265 * another virtualization implementation supporting the callbacks.
3267 struct perf_guest_info_callbacks
*perf_guest_cbs
;
3269 int perf_register_guest_info_callbacks(struct perf_guest_info_callbacks
*cbs
)
3271 perf_guest_cbs
= cbs
;
3274 EXPORT_SYMBOL_GPL(perf_register_guest_info_callbacks
);
3276 int perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks
*cbs
)
3278 perf_guest_cbs
= NULL
;
3281 EXPORT_SYMBOL_GPL(perf_unregister_guest_info_callbacks
);
3286 static bool perf_output_space(struct perf_buffer
*buffer
, unsigned long tail
,
3287 unsigned long offset
, unsigned long head
)
3291 if (!buffer
->writable
)
3294 mask
= perf_data_size(buffer
) - 1;
3296 offset
= (offset
- tail
) & mask
;
3297 head
= (head
- tail
) & mask
;
3299 if ((int)(head
- offset
) < 0)
3305 static void perf_output_wakeup(struct perf_output_handle
*handle
)
3307 atomic_set(&handle
->buffer
->poll
, POLL_IN
);
3310 handle
->event
->pending_wakeup
= 1;
3311 irq_work_queue(&handle
->event
->pending
);
3313 perf_event_wakeup(handle
->event
);
3317 * We need to ensure a later event_id doesn't publish a head when a former
3318 * event isn't done writing. However since we need to deal with NMIs we
3319 * cannot fully serialize things.
3321 * We only publish the head (and generate a wakeup) when the outer-most
3324 static void perf_output_get_handle(struct perf_output_handle
*handle
)
3326 struct perf_buffer
*buffer
= handle
->buffer
;
3329 local_inc(&buffer
->nest
);
3330 handle
->wakeup
= local_read(&buffer
->wakeup
);
3333 static void perf_output_put_handle(struct perf_output_handle
*handle
)
3335 struct perf_buffer
*buffer
= handle
->buffer
;
3339 head
= local_read(&buffer
->head
);
3342 * IRQ/NMI can happen here, which means we can miss a head update.
3345 if (!local_dec_and_test(&buffer
->nest
))
3349 * Publish the known good head. Rely on the full barrier implied
3350 * by atomic_dec_and_test() order the buffer->head read and this
3353 buffer
->user_page
->data_head
= head
;
3356 * Now check if we missed an update, rely on the (compiler)
3357 * barrier in atomic_dec_and_test() to re-read buffer->head.
3359 if (unlikely(head
!= local_read(&buffer
->head
))) {
3360 local_inc(&buffer
->nest
);
3364 if (handle
->wakeup
!= local_read(&buffer
->wakeup
))
3365 perf_output_wakeup(handle
);
3371 __always_inline
void perf_output_copy(struct perf_output_handle
*handle
,
3372 const void *buf
, unsigned int len
)
3375 unsigned long size
= min_t(unsigned long, handle
->size
, len
);
3377 memcpy(handle
->addr
, buf
, size
);
3380 handle
->addr
+= size
;
3382 handle
->size
-= size
;
3383 if (!handle
->size
) {
3384 struct perf_buffer
*buffer
= handle
->buffer
;
3387 handle
->page
&= buffer
->nr_pages
- 1;
3388 handle
->addr
= buffer
->data_pages
[handle
->page
];
3389 handle
->size
= PAGE_SIZE
<< page_order(buffer
);
3394 static void __perf_event_header__init_id(struct perf_event_header
*header
,
3395 struct perf_sample_data
*data
,
3396 struct perf_event
*event
)
3398 u64 sample_type
= event
->attr
.sample_type
;
3400 data
->type
= sample_type
;
3401 header
->size
+= event
->id_header_size
;
3403 if (sample_type
& PERF_SAMPLE_TID
) {
3404 /* namespace issues */
3405 data
->tid_entry
.pid
= perf_event_pid(event
, current
);
3406 data
->tid_entry
.tid
= perf_event_tid(event
, current
);
3409 if (sample_type
& PERF_SAMPLE_TIME
)
3410 data
->time
= perf_clock();
3412 if (sample_type
& PERF_SAMPLE_ID
)
3413 data
->id
= primary_event_id(event
);
3415 if (sample_type
& PERF_SAMPLE_STREAM_ID
)
3416 data
->stream_id
= event
->id
;
3418 if (sample_type
& PERF_SAMPLE_CPU
) {
3419 data
->cpu_entry
.cpu
= raw_smp_processor_id();
3420 data
->cpu_entry
.reserved
= 0;
3424 static void perf_event_header__init_id(struct perf_event_header
*header
,
3425 struct perf_sample_data
*data
,
3426 struct perf_event
*event
)
3428 if (event
->attr
.sample_id_all
)
3429 __perf_event_header__init_id(header
, data
, event
);
3432 static void __perf_event__output_id_sample(struct perf_output_handle
*handle
,
3433 struct perf_sample_data
*data
)
3435 u64 sample_type
= data
->type
;
3437 if (sample_type
& PERF_SAMPLE_TID
)
3438 perf_output_put(handle
, data
->tid_entry
);
3440 if (sample_type
& PERF_SAMPLE_TIME
)
3441 perf_output_put(handle
, data
->time
);
3443 if (sample_type
& PERF_SAMPLE_ID
)
3444 perf_output_put(handle
, data
->id
);
3446 if (sample_type
& PERF_SAMPLE_STREAM_ID
)
3447 perf_output_put(handle
, data
->stream_id
);
3449 if (sample_type
& PERF_SAMPLE_CPU
)
3450 perf_output_put(handle
, data
->cpu_entry
);
3453 static void perf_event__output_id_sample(struct perf_event
*event
,
3454 struct perf_output_handle
*handle
,
3455 struct perf_sample_data
*sample
)
3457 if (event
->attr
.sample_id_all
)
3458 __perf_event__output_id_sample(handle
, sample
);
3461 int perf_output_begin(struct perf_output_handle
*handle
,
3462 struct perf_event
*event
, unsigned int size
,
3463 int nmi
, int sample
)
3465 struct perf_buffer
*buffer
;
3466 unsigned long tail
, offset
, head
;
3468 struct perf_sample_data sample_data
;
3470 struct perf_event_header header
;
3477 * For inherited events we send all the output towards the parent.
3480 event
= event
->parent
;
3482 buffer
= rcu_dereference(event
->buffer
);
3486 handle
->buffer
= buffer
;
3487 handle
->event
= event
;
3489 handle
->sample
= sample
;
3491 if (!buffer
->nr_pages
)
3494 have_lost
= local_read(&buffer
->lost
);
3496 lost_event
.header
.size
= sizeof(lost_event
);
3497 perf_event_header__init_id(&lost_event
.header
, &sample_data
,
3499 size
+= lost_event
.header
.size
;
3502 perf_output_get_handle(handle
);
3506 * Userspace could choose to issue a mb() before updating the
3507 * tail pointer. So that all reads will be completed before the
3510 tail
= ACCESS_ONCE(buffer
->user_page
->data_tail
);
3512 offset
= head
= local_read(&buffer
->head
);
3514 if (unlikely(!perf_output_space(buffer
, tail
, offset
, head
)))
3516 } while (local_cmpxchg(&buffer
->head
, offset
, head
) != offset
);
3518 if (head
- local_read(&buffer
->wakeup
) > buffer
->watermark
)
3519 local_add(buffer
->watermark
, &buffer
->wakeup
);
3521 handle
->page
= offset
>> (PAGE_SHIFT
+ page_order(buffer
));
3522 handle
->page
&= buffer
->nr_pages
- 1;
3523 handle
->size
= offset
& ((PAGE_SIZE
<< page_order(buffer
)) - 1);
3524 handle
->addr
= buffer
->data_pages
[handle
->page
];
3525 handle
->addr
+= handle
->size
;
3526 handle
->size
= (PAGE_SIZE
<< page_order(buffer
)) - handle
->size
;
3529 lost_event
.header
.type
= PERF_RECORD_LOST
;
3530 lost_event
.header
.misc
= 0;
3531 lost_event
.id
= event
->id
;
3532 lost_event
.lost
= local_xchg(&buffer
->lost
, 0);
3534 perf_output_put(handle
, lost_event
);
3535 perf_event__output_id_sample(event
, handle
, &sample_data
);
3541 local_inc(&buffer
->lost
);
3542 perf_output_put_handle(handle
);
3549 void perf_output_end(struct perf_output_handle
*handle
)
3551 struct perf_event
*event
= handle
->event
;
3552 struct perf_buffer
*buffer
= handle
->buffer
;
3554 int wakeup_events
= event
->attr
.wakeup_events
;
3556 if (handle
->sample
&& wakeup_events
) {
3557 int events
= local_inc_return(&buffer
->events
);
3558 if (events
>= wakeup_events
) {
3559 local_sub(wakeup_events
, &buffer
->events
);
3560 local_inc(&buffer
->wakeup
);
3564 perf_output_put_handle(handle
);
3568 static void perf_output_read_one(struct perf_output_handle
*handle
,
3569 struct perf_event
*event
,
3570 u64 enabled
, u64 running
)
3572 u64 read_format
= event
->attr
.read_format
;
3576 values
[n
++] = perf_event_count(event
);
3577 if (read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
) {
3578 values
[n
++] = enabled
+
3579 atomic64_read(&event
->child_total_time_enabled
);
3581 if (read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
) {
3582 values
[n
++] = running
+
3583 atomic64_read(&event
->child_total_time_running
);
3585 if (read_format
& PERF_FORMAT_ID
)
3586 values
[n
++] = primary_event_id(event
);
3588 perf_output_copy(handle
, values
, n
* sizeof(u64
));
3592 * XXX PERF_FORMAT_GROUP vs inherited events seems difficult.
3594 static void perf_output_read_group(struct perf_output_handle
*handle
,
3595 struct perf_event
*event
,
3596 u64 enabled
, u64 running
)
3598 struct perf_event
*leader
= event
->group_leader
, *sub
;
3599 u64 read_format
= event
->attr
.read_format
;
3603 values
[n
++] = 1 + leader
->nr_siblings
;
3605 if (read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
)
3606 values
[n
++] = enabled
;
3608 if (read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
)
3609 values
[n
++] = running
;
3611 if (leader
!= event
)
3612 leader
->pmu
->read(leader
);
3614 values
[n
++] = perf_event_count(leader
);
3615 if (read_format
& PERF_FORMAT_ID
)
3616 values
[n
++] = primary_event_id(leader
);
3618 perf_output_copy(handle
, values
, n
* sizeof(u64
));
3620 list_for_each_entry(sub
, &leader
->sibling_list
, group_entry
) {
3624 sub
->pmu
->read(sub
);
3626 values
[n
++] = perf_event_count(sub
);
3627 if (read_format
& PERF_FORMAT_ID
)
3628 values
[n
++] = primary_event_id(sub
);
3630 perf_output_copy(handle
, values
, n
* sizeof(u64
));
3634 #define PERF_FORMAT_TOTAL_TIMES (PERF_FORMAT_TOTAL_TIME_ENABLED|\
3635 PERF_FORMAT_TOTAL_TIME_RUNNING)
3637 static void perf_output_read(struct perf_output_handle
*handle
,
3638 struct perf_event
*event
)
3640 u64 enabled
= 0, running
= 0, now
, ctx_time
;
3641 u64 read_format
= event
->attr
.read_format
;
3644 * compute total_time_enabled, total_time_running
3645 * based on snapshot values taken when the event
3646 * was last scheduled in.
3648 * we cannot simply called update_context_time()
3649 * because of locking issue as we are called in
3652 if (read_format
& PERF_FORMAT_TOTAL_TIMES
) {
3654 ctx_time
= event
->shadow_ctx_time
+ now
;
3655 enabled
= ctx_time
- event
->tstamp_enabled
;
3656 running
= ctx_time
- event
->tstamp_running
;
3659 if (event
->attr
.read_format
& PERF_FORMAT_GROUP
)
3660 perf_output_read_group(handle
, event
, enabled
, running
);
3662 perf_output_read_one(handle
, event
, enabled
, running
);
3665 void perf_output_sample(struct perf_output_handle
*handle
,
3666 struct perf_event_header
*header
,
3667 struct perf_sample_data
*data
,
3668 struct perf_event
*event
)
3670 u64 sample_type
= data
->type
;
3672 perf_output_put(handle
, *header
);
3674 if (sample_type
& PERF_SAMPLE_IP
)
3675 perf_output_put(handle
, data
->ip
);
3677 if (sample_type
& PERF_SAMPLE_TID
)
3678 perf_output_put(handle
, data
->tid_entry
);
3680 if (sample_type
& PERF_SAMPLE_TIME
)
3681 perf_output_put(handle
, data
->time
);
3683 if (sample_type
& PERF_SAMPLE_ADDR
)
3684 perf_output_put(handle
, data
->addr
);
3686 if (sample_type
& PERF_SAMPLE_ID
)
3687 perf_output_put(handle
, data
->id
);
3689 if (sample_type
& PERF_SAMPLE_STREAM_ID
)
3690 perf_output_put(handle
, data
->stream_id
);
3692 if (sample_type
& PERF_SAMPLE_CPU
)
3693 perf_output_put(handle
, data
->cpu_entry
);
3695 if (sample_type
& PERF_SAMPLE_PERIOD
)
3696 perf_output_put(handle
, data
->period
);
3698 if (sample_type
& PERF_SAMPLE_READ
)
3699 perf_output_read(handle
, event
);
3701 if (sample_type
& PERF_SAMPLE_CALLCHAIN
) {
3702 if (data
->callchain
) {
3705 if (data
->callchain
)
3706 size
+= data
->callchain
->nr
;
3708 size
*= sizeof(u64
);
3710 perf_output_copy(handle
, data
->callchain
, size
);
3713 perf_output_put(handle
, nr
);
3717 if (sample_type
& PERF_SAMPLE_RAW
) {
3719 perf_output_put(handle
, data
->raw
->size
);
3720 perf_output_copy(handle
, data
->raw
->data
,
3727 .size
= sizeof(u32
),
3730 perf_output_put(handle
, raw
);
3735 void perf_prepare_sample(struct perf_event_header
*header
,
3736 struct perf_sample_data
*data
,
3737 struct perf_event
*event
,
3738 struct pt_regs
*regs
)
3740 u64 sample_type
= event
->attr
.sample_type
;
3742 header
->type
= PERF_RECORD_SAMPLE
;
3743 header
->size
= sizeof(*header
) + event
->header_size
;
3746 header
->misc
|= perf_misc_flags(regs
);
3748 __perf_event_header__init_id(header
, data
, event
);
3750 if (sample_type
& PERF_SAMPLE_IP
)
3751 data
->ip
= perf_instruction_pointer(regs
);
3753 if (sample_type
& PERF_SAMPLE_CALLCHAIN
) {
3756 data
->callchain
= perf_callchain(regs
);
3758 if (data
->callchain
)
3759 size
+= data
->callchain
->nr
;
3761 header
->size
+= size
* sizeof(u64
);
3764 if (sample_type
& PERF_SAMPLE_RAW
) {
3765 int size
= sizeof(u32
);
3768 size
+= data
->raw
->size
;
3770 size
+= sizeof(u32
);
3772 WARN_ON_ONCE(size
& (sizeof(u64
)-1));
3773 header
->size
+= size
;
3777 static void perf_event_output(struct perf_event
*event
, int nmi
,
3778 struct perf_sample_data
*data
,
3779 struct pt_regs
*regs
)
3781 struct perf_output_handle handle
;
3782 struct perf_event_header header
;
3784 /* protect the callchain buffers */
3787 perf_prepare_sample(&header
, data
, event
, regs
);
3789 if (perf_output_begin(&handle
, event
, header
.size
, nmi
, 1))
3792 perf_output_sample(&handle
, &header
, data
, event
);
3794 perf_output_end(&handle
);
3804 struct perf_read_event
{
3805 struct perf_event_header header
;
3812 perf_event_read_event(struct perf_event
*event
,
3813 struct task_struct
*task
)
3815 struct perf_output_handle handle
;
3816 struct perf_sample_data sample
;
3817 struct perf_read_event read_event
= {
3819 .type
= PERF_RECORD_READ
,
3821 .size
= sizeof(read_event
) + event
->read_size
,
3823 .pid
= perf_event_pid(event
, task
),
3824 .tid
= perf_event_tid(event
, task
),
3828 perf_event_header__init_id(&read_event
.header
, &sample
, event
);
3829 ret
= perf_output_begin(&handle
, event
, read_event
.header
.size
, 0, 0);
3833 perf_output_put(&handle
, read_event
);
3834 perf_output_read(&handle
, event
);
3835 perf_event__output_id_sample(event
, &handle
, &sample
);
3837 perf_output_end(&handle
);
3841 * task tracking -- fork/exit
3843 * enabled by: attr.comm | attr.mmap | attr.mmap_data | attr.task
3846 struct perf_task_event
{
3847 struct task_struct
*task
;
3848 struct perf_event_context
*task_ctx
;
3851 struct perf_event_header header
;
3861 static void perf_event_task_output(struct perf_event
*event
,
3862 struct perf_task_event
*task_event
)
3864 struct perf_output_handle handle
;
3865 struct perf_sample_data sample
;
3866 struct task_struct
*task
= task_event
->task
;
3867 int ret
, size
= task_event
->event_id
.header
.size
;
3869 perf_event_header__init_id(&task_event
->event_id
.header
, &sample
, event
);
3871 ret
= perf_output_begin(&handle
, event
,
3872 task_event
->event_id
.header
.size
, 0, 0);
3876 task_event
->event_id
.pid
= perf_event_pid(event
, task
);
3877 task_event
->event_id
.ppid
= perf_event_pid(event
, current
);
3879 task_event
->event_id
.tid
= perf_event_tid(event
, task
);
3880 task_event
->event_id
.ptid
= perf_event_tid(event
, current
);
3882 perf_output_put(&handle
, task_event
->event_id
);
3884 perf_event__output_id_sample(event
, &handle
, &sample
);
3886 perf_output_end(&handle
);
3888 task_event
->event_id
.header
.size
= size
;
3891 static int perf_event_task_match(struct perf_event
*event
)
3893 if (event
->state
< PERF_EVENT_STATE_INACTIVE
)
3896 if (event
->cpu
!= -1 && event
->cpu
!= smp_processor_id())
3899 if (event
->attr
.comm
|| event
->attr
.mmap
||
3900 event
->attr
.mmap_data
|| event
->attr
.task
)
3906 static void perf_event_task_ctx(struct perf_event_context
*ctx
,
3907 struct perf_task_event
*task_event
)
3909 struct perf_event
*event
;
3911 list_for_each_entry_rcu(event
, &ctx
->event_list
, event_entry
) {
3912 if (perf_event_task_match(event
))
3913 perf_event_task_output(event
, task_event
);
3917 static void perf_event_task_event(struct perf_task_event
*task_event
)
3919 struct perf_cpu_context
*cpuctx
;
3920 struct perf_event_context
*ctx
;
3925 list_for_each_entry_rcu(pmu
, &pmus
, entry
) {
3926 cpuctx
= get_cpu_ptr(pmu
->pmu_cpu_context
);
3927 if (cpuctx
->active_pmu
!= pmu
)
3929 perf_event_task_ctx(&cpuctx
->ctx
, task_event
);
3931 ctx
= task_event
->task_ctx
;
3933 ctxn
= pmu
->task_ctx_nr
;
3936 ctx
= rcu_dereference(current
->perf_event_ctxp
[ctxn
]);
3939 perf_event_task_ctx(ctx
, task_event
);
3941 put_cpu_ptr(pmu
->pmu_cpu_context
);
3946 static void perf_event_task(struct task_struct
*task
,
3947 struct perf_event_context
*task_ctx
,
3950 struct perf_task_event task_event
;
3952 if (!atomic_read(&nr_comm_events
) &&
3953 !atomic_read(&nr_mmap_events
) &&
3954 !atomic_read(&nr_task_events
))
3957 task_event
= (struct perf_task_event
){
3959 .task_ctx
= task_ctx
,
3962 .type
= new ? PERF_RECORD_FORK
: PERF_RECORD_EXIT
,
3964 .size
= sizeof(task_event
.event_id
),
3970 .time
= perf_clock(),
3974 perf_event_task_event(&task_event
);
3977 void perf_event_fork(struct task_struct
*task
)
3979 perf_event_task(task
, NULL
, 1);
3986 struct perf_comm_event
{
3987 struct task_struct
*task
;
3992 struct perf_event_header header
;
3999 static void perf_event_comm_output(struct perf_event
*event
,
4000 struct perf_comm_event
*comm_event
)
4002 struct perf_output_handle handle
;
4003 struct perf_sample_data sample
;
4004 int size
= comm_event
->event_id
.header
.size
;
4007 perf_event_header__init_id(&comm_event
->event_id
.header
, &sample
, event
);
4008 ret
= perf_output_begin(&handle
, event
,
4009 comm_event
->event_id
.header
.size
, 0, 0);
4014 comm_event
->event_id
.pid
= perf_event_pid(event
, comm_event
->task
);
4015 comm_event
->event_id
.tid
= perf_event_tid(event
, comm_event
->task
);
4017 perf_output_put(&handle
, comm_event
->event_id
);
4018 perf_output_copy(&handle
, comm_event
->comm
,
4019 comm_event
->comm_size
);
4021 perf_event__output_id_sample(event
, &handle
, &sample
);
4023 perf_output_end(&handle
);
4025 comm_event
->event_id
.header
.size
= size
;
4028 static int perf_event_comm_match(struct perf_event
*event
)
4030 if (event
->state
< PERF_EVENT_STATE_INACTIVE
)
4033 if (event
->cpu
!= -1 && event
->cpu
!= smp_processor_id())
4036 if (event
->attr
.comm
)
4042 static void perf_event_comm_ctx(struct perf_event_context
*ctx
,
4043 struct perf_comm_event
*comm_event
)
4045 struct perf_event
*event
;
4047 list_for_each_entry_rcu(event
, &ctx
->event_list
, event_entry
) {
4048 if (perf_event_comm_match(event
))
4049 perf_event_comm_output(event
, comm_event
);
4053 static void perf_event_comm_event(struct perf_comm_event
*comm_event
)
4055 struct perf_cpu_context
*cpuctx
;
4056 struct perf_event_context
*ctx
;
4057 char comm
[TASK_COMM_LEN
];
4062 memset(comm
, 0, sizeof(comm
));
4063 strlcpy(comm
, comm_event
->task
->comm
, sizeof(comm
));
4064 size
= ALIGN(strlen(comm
)+1, sizeof(u64
));
4066 comm_event
->comm
= comm
;
4067 comm_event
->comm_size
= size
;
4069 comm_event
->event_id
.header
.size
= sizeof(comm_event
->event_id
) + size
;
4071 list_for_each_entry_rcu(pmu
, &pmus
, entry
) {
4072 cpuctx
= get_cpu_ptr(pmu
->pmu_cpu_context
);
4073 if (cpuctx
->active_pmu
!= pmu
)
4075 perf_event_comm_ctx(&cpuctx
->ctx
, comm_event
);
4077 ctxn
= pmu
->task_ctx_nr
;
4081 ctx
= rcu_dereference(current
->perf_event_ctxp
[ctxn
]);
4083 perf_event_comm_ctx(ctx
, comm_event
);
4085 put_cpu_ptr(pmu
->pmu_cpu_context
);
4090 void perf_event_comm(struct task_struct
*task
)
4092 struct perf_comm_event comm_event
;
4093 struct perf_event_context
*ctx
;
4096 for_each_task_context_nr(ctxn
) {
4097 ctx
= task
->perf_event_ctxp
[ctxn
];
4101 perf_event_enable_on_exec(ctx
);
4104 if (!atomic_read(&nr_comm_events
))
4107 comm_event
= (struct perf_comm_event
){
4113 .type
= PERF_RECORD_COMM
,
4122 perf_event_comm_event(&comm_event
);
4129 struct perf_mmap_event
{
4130 struct vm_area_struct
*vma
;
4132 const char *file_name
;
4136 struct perf_event_header header
;
4146 static void perf_event_mmap_output(struct perf_event
*event
,
4147 struct perf_mmap_event
*mmap_event
)
4149 struct perf_output_handle handle
;
4150 struct perf_sample_data sample
;
4151 int size
= mmap_event
->event_id
.header
.size
;
4154 perf_event_header__init_id(&mmap_event
->event_id
.header
, &sample
, event
);
4155 ret
= perf_output_begin(&handle
, event
,
4156 mmap_event
->event_id
.header
.size
, 0, 0);
4160 mmap_event
->event_id
.pid
= perf_event_pid(event
, current
);
4161 mmap_event
->event_id
.tid
= perf_event_tid(event
, current
);
4163 perf_output_put(&handle
, mmap_event
->event_id
);
4164 perf_output_copy(&handle
, mmap_event
->file_name
,
4165 mmap_event
->file_size
);
4167 perf_event__output_id_sample(event
, &handle
, &sample
);
4169 perf_output_end(&handle
);
4171 mmap_event
->event_id
.header
.size
= size
;
4174 static int perf_event_mmap_match(struct perf_event
*event
,
4175 struct perf_mmap_event
*mmap_event
,
4178 if (event
->state
< PERF_EVENT_STATE_INACTIVE
)
4181 if (event
->cpu
!= -1 && event
->cpu
!= smp_processor_id())
4184 if ((!executable
&& event
->attr
.mmap_data
) ||
4185 (executable
&& event
->attr
.mmap
))
4191 static void perf_event_mmap_ctx(struct perf_event_context
*ctx
,
4192 struct perf_mmap_event
*mmap_event
,
4195 struct perf_event
*event
;
4197 list_for_each_entry_rcu(event
, &ctx
->event_list
, event_entry
) {
4198 if (perf_event_mmap_match(event
, mmap_event
, executable
))
4199 perf_event_mmap_output(event
, mmap_event
);
4203 static void perf_event_mmap_event(struct perf_mmap_event
*mmap_event
)
4205 struct perf_cpu_context
*cpuctx
;
4206 struct perf_event_context
*ctx
;
4207 struct vm_area_struct
*vma
= mmap_event
->vma
;
4208 struct file
*file
= vma
->vm_file
;
4216 memset(tmp
, 0, sizeof(tmp
));
4220 * d_path works from the end of the buffer backwards, so we
4221 * need to add enough zero bytes after the string to handle
4222 * the 64bit alignment we do later.
4224 buf
= kzalloc(PATH_MAX
+ sizeof(u64
), GFP_KERNEL
);
4226 name
= strncpy(tmp
, "//enomem", sizeof(tmp
));
4229 name
= d_path(&file
->f_path
, buf
, PATH_MAX
);
4231 name
= strncpy(tmp
, "//toolong", sizeof(tmp
));
4235 if (arch_vma_name(mmap_event
->vma
)) {
4236 name
= strncpy(tmp
, arch_vma_name(mmap_event
->vma
),
4242 name
= strncpy(tmp
, "[vdso]", sizeof(tmp
));
4244 } else if (vma
->vm_start
<= vma
->vm_mm
->start_brk
&&
4245 vma
->vm_end
>= vma
->vm_mm
->brk
) {
4246 name
= strncpy(tmp
, "[heap]", sizeof(tmp
));
4248 } else if (vma
->vm_start
<= vma
->vm_mm
->start_stack
&&
4249 vma
->vm_end
>= vma
->vm_mm
->start_stack
) {
4250 name
= strncpy(tmp
, "[stack]", sizeof(tmp
));
4254 name
= strncpy(tmp
, "//anon", sizeof(tmp
));
4259 size
= ALIGN(strlen(name
)+1, sizeof(u64
));
4261 mmap_event
->file_name
= name
;
4262 mmap_event
->file_size
= size
;
4264 mmap_event
->event_id
.header
.size
= sizeof(mmap_event
->event_id
) + size
;
4267 list_for_each_entry_rcu(pmu
, &pmus
, entry
) {
4268 cpuctx
= get_cpu_ptr(pmu
->pmu_cpu_context
);
4269 if (cpuctx
->active_pmu
!= pmu
)
4271 perf_event_mmap_ctx(&cpuctx
->ctx
, mmap_event
,
4272 vma
->vm_flags
& VM_EXEC
);
4274 ctxn
= pmu
->task_ctx_nr
;
4278 ctx
= rcu_dereference(current
->perf_event_ctxp
[ctxn
]);
4280 perf_event_mmap_ctx(ctx
, mmap_event
,
4281 vma
->vm_flags
& VM_EXEC
);
4284 put_cpu_ptr(pmu
->pmu_cpu_context
);
4291 void perf_event_mmap(struct vm_area_struct
*vma
)
4293 struct perf_mmap_event mmap_event
;
4295 if (!atomic_read(&nr_mmap_events
))
4298 mmap_event
= (struct perf_mmap_event
){
4304 .type
= PERF_RECORD_MMAP
,
4305 .misc
= PERF_RECORD_MISC_USER
,
4310 .start
= vma
->vm_start
,
4311 .len
= vma
->vm_end
- vma
->vm_start
,
4312 .pgoff
= (u64
)vma
->vm_pgoff
<< PAGE_SHIFT
,
4316 perf_event_mmap_event(&mmap_event
);
4320 * IRQ throttle logging
4323 static void perf_log_throttle(struct perf_event
*event
, int enable
)
4325 struct perf_output_handle handle
;
4326 struct perf_sample_data sample
;
4330 struct perf_event_header header
;
4334 } throttle_event
= {
4336 .type
= PERF_RECORD_THROTTLE
,
4338 .size
= sizeof(throttle_event
),
4340 .time
= perf_clock(),
4341 .id
= primary_event_id(event
),
4342 .stream_id
= event
->id
,
4346 throttle_event
.header
.type
= PERF_RECORD_UNTHROTTLE
;
4348 perf_event_header__init_id(&throttle_event
.header
, &sample
, event
);
4350 ret
= perf_output_begin(&handle
, event
,
4351 throttle_event
.header
.size
, 1, 0);
4355 perf_output_put(&handle
, throttle_event
);
4356 perf_event__output_id_sample(event
, &handle
, &sample
);
4357 perf_output_end(&handle
);
4361 * Generic event overflow handling, sampling.
4364 static int __perf_event_overflow(struct perf_event
*event
, int nmi
,
4365 int throttle
, struct perf_sample_data
*data
,
4366 struct pt_regs
*regs
)
4368 int events
= atomic_read(&event
->event_limit
);
4369 struct hw_perf_event
*hwc
= &event
->hw
;
4373 * Non-sampling counters might still use the PMI to fold short
4374 * hardware counters, ignore those.
4376 if (unlikely(!is_sampling_event(event
)))
4382 if (hwc
->interrupts
!= MAX_INTERRUPTS
) {
4384 if (HZ
* hwc
->interrupts
>
4385 (u64
)sysctl_perf_event_sample_rate
) {
4386 hwc
->interrupts
= MAX_INTERRUPTS
;
4387 perf_log_throttle(event
, 0);
4392 * Keep re-disabling events even though on the previous
4393 * pass we disabled it - just in case we raced with a
4394 * sched-in and the event got enabled again:
4400 if (event
->attr
.freq
) {
4401 u64 now
= perf_clock();
4402 s64 delta
= now
- hwc
->freq_time_stamp
;
4404 hwc
->freq_time_stamp
= now
;
4406 if (delta
> 0 && delta
< 2*TICK_NSEC
)
4407 perf_adjust_period(event
, delta
, hwc
->last_period
);
4411 * XXX event_limit might not quite work as expected on inherited
4415 event
->pending_kill
= POLL_IN
;
4416 if (events
&& atomic_dec_and_test(&event
->event_limit
)) {
4418 event
->pending_kill
= POLL_HUP
;
4420 event
->pending_disable
= 1;
4421 irq_work_queue(&event
->pending
);
4423 perf_event_disable(event
);
4426 if (event
->overflow_handler
)
4427 event
->overflow_handler(event
, nmi
, data
, regs
);
4429 perf_event_output(event
, nmi
, data
, regs
);
4434 int perf_event_overflow(struct perf_event
*event
, int nmi
,
4435 struct perf_sample_data
*data
,
4436 struct pt_regs
*regs
)
4438 return __perf_event_overflow(event
, nmi
, 1, data
, regs
);
4442 * Generic software event infrastructure
4445 struct swevent_htable
{
4446 struct swevent_hlist
*swevent_hlist
;
4447 struct mutex hlist_mutex
;
4450 /* Recursion avoidance in each contexts */
4451 int recursion
[PERF_NR_CONTEXTS
];
4454 static DEFINE_PER_CPU(struct swevent_htable
, swevent_htable
);
4457 * We directly increment event->count and keep a second value in
4458 * event->hw.period_left to count intervals. This period event
4459 * is kept in the range [-sample_period, 0] so that we can use the
4463 static u64
perf_swevent_set_period(struct perf_event
*event
)
4465 struct hw_perf_event
*hwc
= &event
->hw
;
4466 u64 period
= hwc
->last_period
;
4470 hwc
->last_period
= hwc
->sample_period
;
4473 old
= val
= local64_read(&hwc
->period_left
);
4477 nr
= div64_u64(period
+ val
, period
);
4478 offset
= nr
* period
;
4480 if (local64_cmpxchg(&hwc
->period_left
, old
, val
) != old
)
4486 static void perf_swevent_overflow(struct perf_event
*event
, u64 overflow
,
4487 int nmi
, struct perf_sample_data
*data
,
4488 struct pt_regs
*regs
)
4490 struct hw_perf_event
*hwc
= &event
->hw
;
4493 data
->period
= event
->hw
.last_period
;
4495 overflow
= perf_swevent_set_period(event
);
4497 if (hwc
->interrupts
== MAX_INTERRUPTS
)
4500 for (; overflow
; overflow
--) {
4501 if (__perf_event_overflow(event
, nmi
, throttle
,
4504 * We inhibit the overflow from happening when
4505 * hwc->interrupts == MAX_INTERRUPTS.
4513 static void perf_swevent_event(struct perf_event
*event
, u64 nr
,
4514 int nmi
, struct perf_sample_data
*data
,
4515 struct pt_regs
*regs
)
4517 struct hw_perf_event
*hwc
= &event
->hw
;
4519 local64_add(nr
, &event
->count
);
4524 if (!is_sampling_event(event
))
4527 if (nr
== 1 && hwc
->sample_period
== 1 && !event
->attr
.freq
)
4528 return perf_swevent_overflow(event
, 1, nmi
, data
, regs
);
4530 if (local64_add_negative(nr
, &hwc
->period_left
))
4533 perf_swevent_overflow(event
, 0, nmi
, data
, regs
);
4536 static int perf_exclude_event(struct perf_event
*event
,
4537 struct pt_regs
*regs
)
4539 if (event
->hw
.state
& PERF_HES_STOPPED
)
4543 if (event
->attr
.exclude_user
&& user_mode(regs
))
4546 if (event
->attr
.exclude_kernel
&& !user_mode(regs
))
4553 static int perf_swevent_match(struct perf_event
*event
,
4554 enum perf_type_id type
,
4556 struct perf_sample_data
*data
,
4557 struct pt_regs
*regs
)
4559 if (event
->attr
.type
!= type
)
4562 if (event
->attr
.config
!= event_id
)
4565 if (perf_exclude_event(event
, regs
))
4571 static inline u64
swevent_hash(u64 type
, u32 event_id
)
4573 u64 val
= event_id
| (type
<< 32);
4575 return hash_64(val
, SWEVENT_HLIST_BITS
);
4578 static inline struct hlist_head
*
4579 __find_swevent_head(struct swevent_hlist
*hlist
, u64 type
, u32 event_id
)
4581 u64 hash
= swevent_hash(type
, event_id
);
4583 return &hlist
->heads
[hash
];
4586 /* For the read side: events when they trigger */
4587 static inline struct hlist_head
*
4588 find_swevent_head_rcu(struct swevent_htable
*swhash
, u64 type
, u32 event_id
)
4590 struct swevent_hlist
*hlist
;
4592 hlist
= rcu_dereference(swhash
->swevent_hlist
);
4596 return __find_swevent_head(hlist
, type
, event_id
);
4599 /* For the event head insertion and removal in the hlist */
4600 static inline struct hlist_head
*
4601 find_swevent_head(struct swevent_htable
*swhash
, struct perf_event
*event
)
4603 struct swevent_hlist
*hlist
;
4604 u32 event_id
= event
->attr
.config
;
4605 u64 type
= event
->attr
.type
;
4608 * Event scheduling is always serialized against hlist allocation
4609 * and release. Which makes the protected version suitable here.
4610 * The context lock guarantees that.
4612 hlist
= rcu_dereference_protected(swhash
->swevent_hlist
,
4613 lockdep_is_held(&event
->ctx
->lock
));
4617 return __find_swevent_head(hlist
, type
, event_id
);
4620 static void do_perf_sw_event(enum perf_type_id type
, u32 event_id
,
4622 struct perf_sample_data
*data
,
4623 struct pt_regs
*regs
)
4625 struct swevent_htable
*swhash
= &__get_cpu_var(swevent_htable
);
4626 struct perf_event
*event
;
4627 struct hlist_node
*node
;
4628 struct hlist_head
*head
;
4631 head
= find_swevent_head_rcu(swhash
, type
, event_id
);
4635 hlist_for_each_entry_rcu(event
, node
, head
, hlist_entry
) {
4636 if (perf_swevent_match(event
, type
, event_id
, data
, regs
))
4637 perf_swevent_event(event
, nr
, nmi
, data
, regs
);
4643 int perf_swevent_get_recursion_context(void)
4645 struct swevent_htable
*swhash
= &__get_cpu_var(swevent_htable
);
4647 return get_recursion_context(swhash
->recursion
);
4649 EXPORT_SYMBOL_GPL(perf_swevent_get_recursion_context
);
4651 void inline perf_swevent_put_recursion_context(int rctx
)
4653 struct swevent_htable
*swhash
= &__get_cpu_var(swevent_htable
);
4655 put_recursion_context(swhash
->recursion
, rctx
);
4658 void __perf_sw_event(u32 event_id
, u64 nr
, int nmi
,
4659 struct pt_regs
*regs
, u64 addr
)
4661 struct perf_sample_data data
;
4664 preempt_disable_notrace();
4665 rctx
= perf_swevent_get_recursion_context();
4669 perf_sample_data_init(&data
, addr
);
4671 do_perf_sw_event(PERF_TYPE_SOFTWARE
, event_id
, nr
, nmi
, &data
, regs
);
4673 perf_swevent_put_recursion_context(rctx
);
4674 preempt_enable_notrace();
4677 static void perf_swevent_read(struct perf_event
*event
)
4681 static int perf_swevent_add(struct perf_event
*event
, int flags
)
4683 struct swevent_htable
*swhash
= &__get_cpu_var(swevent_htable
);
4684 struct hw_perf_event
*hwc
= &event
->hw
;
4685 struct hlist_head
*head
;
4687 if (is_sampling_event(event
)) {
4688 hwc
->last_period
= hwc
->sample_period
;
4689 perf_swevent_set_period(event
);
4692 hwc
->state
= !(flags
& PERF_EF_START
);
4694 head
= find_swevent_head(swhash
, event
);
4695 if (WARN_ON_ONCE(!head
))
4698 hlist_add_head_rcu(&event
->hlist_entry
, head
);
4703 static void perf_swevent_del(struct perf_event
*event
, int flags
)
4705 hlist_del_rcu(&event
->hlist_entry
);
4708 static void perf_swevent_start(struct perf_event
*event
, int flags
)
4710 event
->hw
.state
= 0;
4713 static void perf_swevent_stop(struct perf_event
*event
, int flags
)
4715 event
->hw
.state
= PERF_HES_STOPPED
;
4718 /* Deref the hlist from the update side */
4719 static inline struct swevent_hlist
*
4720 swevent_hlist_deref(struct swevent_htable
*swhash
)
4722 return rcu_dereference_protected(swhash
->swevent_hlist
,
4723 lockdep_is_held(&swhash
->hlist_mutex
));
4726 static void swevent_hlist_release_rcu(struct rcu_head
*rcu_head
)
4728 struct swevent_hlist
*hlist
;
4730 hlist
= container_of(rcu_head
, struct swevent_hlist
, rcu_head
);
4734 static void swevent_hlist_release(struct swevent_htable
*swhash
)
4736 struct swevent_hlist
*hlist
= swevent_hlist_deref(swhash
);
4741 rcu_assign_pointer(swhash
->swevent_hlist
, NULL
);
4742 call_rcu(&hlist
->rcu_head
, swevent_hlist_release_rcu
);
4745 static void swevent_hlist_put_cpu(struct perf_event
*event
, int cpu
)
4747 struct swevent_htable
*swhash
= &per_cpu(swevent_htable
, cpu
);
4749 mutex_lock(&swhash
->hlist_mutex
);
4751 if (!--swhash
->hlist_refcount
)
4752 swevent_hlist_release(swhash
);
4754 mutex_unlock(&swhash
->hlist_mutex
);
4757 static void swevent_hlist_put(struct perf_event
*event
)
4761 if (event
->cpu
!= -1) {
4762 swevent_hlist_put_cpu(event
, event
->cpu
);
4766 for_each_possible_cpu(cpu
)
4767 swevent_hlist_put_cpu(event
, cpu
);
4770 static int swevent_hlist_get_cpu(struct perf_event
*event
, int cpu
)
4772 struct swevent_htable
*swhash
= &per_cpu(swevent_htable
, cpu
);
4775 mutex_lock(&swhash
->hlist_mutex
);
4777 if (!swevent_hlist_deref(swhash
) && cpu_online(cpu
)) {
4778 struct swevent_hlist
*hlist
;
4780 hlist
= kzalloc(sizeof(*hlist
), GFP_KERNEL
);
4785 rcu_assign_pointer(swhash
->swevent_hlist
, hlist
);
4787 swhash
->hlist_refcount
++;
4789 mutex_unlock(&swhash
->hlist_mutex
);
4794 static int swevent_hlist_get(struct perf_event
*event
)
4797 int cpu
, failed_cpu
;
4799 if (event
->cpu
!= -1)
4800 return swevent_hlist_get_cpu(event
, event
->cpu
);
4803 for_each_possible_cpu(cpu
) {
4804 err
= swevent_hlist_get_cpu(event
, cpu
);
4814 for_each_possible_cpu(cpu
) {
4815 if (cpu
== failed_cpu
)
4817 swevent_hlist_put_cpu(event
, cpu
);
4824 atomic_t perf_swevent_enabled
[PERF_COUNT_SW_MAX
];
4826 static void sw_perf_event_destroy(struct perf_event
*event
)
4828 u64 event_id
= event
->attr
.config
;
4830 WARN_ON(event
->parent
);
4832 jump_label_dec(&perf_swevent_enabled
[event_id
]);
4833 swevent_hlist_put(event
);
4836 static int perf_swevent_init(struct perf_event
*event
)
4838 int event_id
= event
->attr
.config
;
4840 if (event
->attr
.type
!= PERF_TYPE_SOFTWARE
)
4844 case PERF_COUNT_SW_CPU_CLOCK
:
4845 case PERF_COUNT_SW_TASK_CLOCK
:
4852 if (event_id
>= PERF_COUNT_SW_MAX
)
4855 if (!event
->parent
) {
4858 err
= swevent_hlist_get(event
);
4862 jump_label_inc(&perf_swevent_enabled
[event_id
]);
4863 event
->destroy
= sw_perf_event_destroy
;
4869 static struct pmu perf_swevent
= {
4870 .task_ctx_nr
= perf_sw_context
,
4872 .event_init
= perf_swevent_init
,
4873 .add
= perf_swevent_add
,
4874 .del
= perf_swevent_del
,
4875 .start
= perf_swevent_start
,
4876 .stop
= perf_swevent_stop
,
4877 .read
= perf_swevent_read
,
4880 #ifdef CONFIG_EVENT_TRACING
4882 static int perf_tp_filter_match(struct perf_event
*event
,
4883 struct perf_sample_data
*data
)
4885 void *record
= data
->raw
->data
;
4887 if (likely(!event
->filter
) || filter_match_preds(event
->filter
, record
))
4892 static int perf_tp_event_match(struct perf_event
*event
,
4893 struct perf_sample_data
*data
,
4894 struct pt_regs
*regs
)
4897 * All tracepoints are from kernel-space.
4899 if (event
->attr
.exclude_kernel
)
4902 if (!perf_tp_filter_match(event
, data
))
4908 void perf_tp_event(u64 addr
, u64 count
, void *record
, int entry_size
,
4909 struct pt_regs
*regs
, struct hlist_head
*head
, int rctx
)
4911 struct perf_sample_data data
;
4912 struct perf_event
*event
;
4913 struct hlist_node
*node
;
4915 struct perf_raw_record raw
= {
4920 perf_sample_data_init(&data
, addr
);
4923 hlist_for_each_entry_rcu(event
, node
, head
, hlist_entry
) {
4924 if (perf_tp_event_match(event
, &data
, regs
))
4925 perf_swevent_event(event
, count
, 1, &data
, regs
);
4928 perf_swevent_put_recursion_context(rctx
);
4930 EXPORT_SYMBOL_GPL(perf_tp_event
);
4932 static void tp_perf_event_destroy(struct perf_event
*event
)
4934 perf_trace_destroy(event
);
4937 static int perf_tp_event_init(struct perf_event
*event
)
4941 if (event
->attr
.type
!= PERF_TYPE_TRACEPOINT
)
4944 err
= perf_trace_init(event
);
4948 event
->destroy
= tp_perf_event_destroy
;
4953 static struct pmu perf_tracepoint
= {
4954 .task_ctx_nr
= perf_sw_context
,
4956 .event_init
= perf_tp_event_init
,
4957 .add
= perf_trace_add
,
4958 .del
= perf_trace_del
,
4959 .start
= perf_swevent_start
,
4960 .stop
= perf_swevent_stop
,
4961 .read
= perf_swevent_read
,
4964 static inline void perf_tp_register(void)
4966 perf_pmu_register(&perf_tracepoint
, "tracepoint", PERF_TYPE_TRACEPOINT
);
4969 static int perf_event_set_filter(struct perf_event
*event
, void __user
*arg
)
4974 if (event
->attr
.type
!= PERF_TYPE_TRACEPOINT
)
4977 filter_str
= strndup_user(arg
, PAGE_SIZE
);
4978 if (IS_ERR(filter_str
))
4979 return PTR_ERR(filter_str
);
4981 ret
= ftrace_profile_set_filter(event
, event
->attr
.config
, filter_str
);
4987 static void perf_event_free_filter(struct perf_event
*event
)
4989 ftrace_profile_free_filter(event
);
4994 static inline void perf_tp_register(void)
4998 static int perf_event_set_filter(struct perf_event
*event
, void __user
*arg
)
5003 static void perf_event_free_filter(struct perf_event
*event
)
5007 #endif /* CONFIG_EVENT_TRACING */
5009 #ifdef CONFIG_HAVE_HW_BREAKPOINT
5010 void perf_bp_event(struct perf_event
*bp
, void *data
)
5012 struct perf_sample_data sample
;
5013 struct pt_regs
*regs
= data
;
5015 perf_sample_data_init(&sample
, bp
->attr
.bp_addr
);
5017 if (!bp
->hw
.state
&& !perf_exclude_event(bp
, regs
))
5018 perf_swevent_event(bp
, 1, 1, &sample
, regs
);
5023 * hrtimer based swevent callback
5026 static enum hrtimer_restart
perf_swevent_hrtimer(struct hrtimer
*hrtimer
)
5028 enum hrtimer_restart ret
= HRTIMER_RESTART
;
5029 struct perf_sample_data data
;
5030 struct pt_regs
*regs
;
5031 struct perf_event
*event
;
5034 event
= container_of(hrtimer
, struct perf_event
, hw
.hrtimer
);
5035 event
->pmu
->read(event
);
5037 perf_sample_data_init(&data
, 0);
5038 data
.period
= event
->hw
.last_period
;
5039 regs
= get_irq_regs();
5041 if (regs
&& !perf_exclude_event(event
, regs
)) {
5042 if (!(event
->attr
.exclude_idle
&& current
->pid
== 0))
5043 if (perf_event_overflow(event
, 0, &data
, regs
))
5044 ret
= HRTIMER_NORESTART
;
5047 period
= max_t(u64
, 10000, event
->hw
.sample_period
);
5048 hrtimer_forward_now(hrtimer
, ns_to_ktime(period
));
5053 static void perf_swevent_start_hrtimer(struct perf_event
*event
)
5055 struct hw_perf_event
*hwc
= &event
->hw
;
5058 if (!is_sampling_event(event
))
5061 hrtimer_init(&hwc
->hrtimer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
5062 hwc
->hrtimer
.function
= perf_swevent_hrtimer
;
5064 period
= local64_read(&hwc
->period_left
);
5069 local64_set(&hwc
->period_left
, 0);
5071 period
= max_t(u64
, 10000, hwc
->sample_period
);
5073 __hrtimer_start_range_ns(&hwc
->hrtimer
,
5074 ns_to_ktime(period
), 0,
5075 HRTIMER_MODE_REL_PINNED
, 0);
5078 static void perf_swevent_cancel_hrtimer(struct perf_event
*event
)
5080 struct hw_perf_event
*hwc
= &event
->hw
;
5082 if (is_sampling_event(event
)) {
5083 ktime_t remaining
= hrtimer_get_remaining(&hwc
->hrtimer
);
5084 local64_set(&hwc
->period_left
, ktime_to_ns(remaining
));
5086 hrtimer_cancel(&hwc
->hrtimer
);
5091 * Software event: cpu wall time clock
5094 static void cpu_clock_event_update(struct perf_event
*event
)
5099 now
= local_clock();
5100 prev
= local64_xchg(&event
->hw
.prev_count
, now
);
5101 local64_add(now
- prev
, &event
->count
);
5104 static void cpu_clock_event_start(struct perf_event
*event
, int flags
)
5106 local64_set(&event
->hw
.prev_count
, local_clock());
5107 perf_swevent_start_hrtimer(event
);
5110 static void cpu_clock_event_stop(struct perf_event
*event
, int flags
)
5112 perf_swevent_cancel_hrtimer(event
);
5113 cpu_clock_event_update(event
);
5116 static int cpu_clock_event_add(struct perf_event
*event
, int flags
)
5118 if (flags
& PERF_EF_START
)
5119 cpu_clock_event_start(event
, flags
);
5124 static void cpu_clock_event_del(struct perf_event
*event
, int flags
)
5126 cpu_clock_event_stop(event
, flags
);
5129 static void cpu_clock_event_read(struct perf_event
*event
)
5131 cpu_clock_event_update(event
);
5134 static int cpu_clock_event_init(struct perf_event
*event
)
5136 if (event
->attr
.type
!= PERF_TYPE_SOFTWARE
)
5139 if (event
->attr
.config
!= PERF_COUNT_SW_CPU_CLOCK
)
5145 static struct pmu perf_cpu_clock
= {
5146 .task_ctx_nr
= perf_sw_context
,
5148 .event_init
= cpu_clock_event_init
,
5149 .add
= cpu_clock_event_add
,
5150 .del
= cpu_clock_event_del
,
5151 .start
= cpu_clock_event_start
,
5152 .stop
= cpu_clock_event_stop
,
5153 .read
= cpu_clock_event_read
,
5157 * Software event: task time clock
5160 static void task_clock_event_update(struct perf_event
*event
, u64 now
)
5165 prev
= local64_xchg(&event
->hw
.prev_count
, now
);
5167 local64_add(delta
, &event
->count
);
5170 static void task_clock_event_start(struct perf_event
*event
, int flags
)
5172 local64_set(&event
->hw
.prev_count
, event
->ctx
->time
);
5173 perf_swevent_start_hrtimer(event
);
5176 static void task_clock_event_stop(struct perf_event
*event
, int flags
)
5178 perf_swevent_cancel_hrtimer(event
);
5179 task_clock_event_update(event
, event
->ctx
->time
);
5182 static int task_clock_event_add(struct perf_event
*event
, int flags
)
5184 if (flags
& PERF_EF_START
)
5185 task_clock_event_start(event
, flags
);
5190 static void task_clock_event_del(struct perf_event
*event
, int flags
)
5192 task_clock_event_stop(event
, PERF_EF_UPDATE
);
5195 static void task_clock_event_read(struct perf_event
*event
)
5200 update_context_time(event
->ctx
);
5201 time
= event
->ctx
->time
;
5203 u64 now
= perf_clock();
5204 u64 delta
= now
- event
->ctx
->timestamp
;
5205 time
= event
->ctx
->time
+ delta
;
5208 task_clock_event_update(event
, time
);
5211 static int task_clock_event_init(struct perf_event
*event
)
5213 if (event
->attr
.type
!= PERF_TYPE_SOFTWARE
)
5216 if (event
->attr
.config
!= PERF_COUNT_SW_TASK_CLOCK
)
5222 static struct pmu perf_task_clock
= {
5223 .task_ctx_nr
= perf_sw_context
,
5225 .event_init
= task_clock_event_init
,
5226 .add
= task_clock_event_add
,
5227 .del
= task_clock_event_del
,
5228 .start
= task_clock_event_start
,
5229 .stop
= task_clock_event_stop
,
5230 .read
= task_clock_event_read
,
5233 static void perf_pmu_nop_void(struct pmu
*pmu
)
5237 static int perf_pmu_nop_int(struct pmu
*pmu
)
5242 static void perf_pmu_start_txn(struct pmu
*pmu
)
5244 perf_pmu_disable(pmu
);
5247 static int perf_pmu_commit_txn(struct pmu
*pmu
)
5249 perf_pmu_enable(pmu
);
5253 static void perf_pmu_cancel_txn(struct pmu
*pmu
)
5255 perf_pmu_enable(pmu
);
5259 * Ensures all contexts with the same task_ctx_nr have the same
5260 * pmu_cpu_context too.
5262 static void *find_pmu_context(int ctxn
)
5269 list_for_each_entry(pmu
, &pmus
, entry
) {
5270 if (pmu
->task_ctx_nr
== ctxn
)
5271 return pmu
->pmu_cpu_context
;
5277 static void update_pmu_context(struct pmu
*pmu
, struct pmu
*old_pmu
)
5281 for_each_possible_cpu(cpu
) {
5282 struct perf_cpu_context
*cpuctx
;
5284 cpuctx
= per_cpu_ptr(pmu
->pmu_cpu_context
, cpu
);
5286 if (cpuctx
->active_pmu
== old_pmu
)
5287 cpuctx
->active_pmu
= pmu
;
5291 static void free_pmu_context(struct pmu
*pmu
)
5295 mutex_lock(&pmus_lock
);
5297 * Like a real lame refcount.
5299 list_for_each_entry(i
, &pmus
, entry
) {
5300 if (i
->pmu_cpu_context
== pmu
->pmu_cpu_context
) {
5301 update_pmu_context(i
, pmu
);
5306 free_percpu(pmu
->pmu_cpu_context
);
5308 mutex_unlock(&pmus_lock
);
5310 static struct idr pmu_idr
;
5313 type_show(struct device
*dev
, struct device_attribute
*attr
, char *page
)
5315 struct pmu
*pmu
= dev_get_drvdata(dev
);
5317 return snprintf(page
, PAGE_SIZE
-1, "%d\n", pmu
->type
);
5320 static struct device_attribute pmu_dev_attrs
[] = {
5325 static int pmu_bus_running
;
5326 static struct bus_type pmu_bus
= {
5327 .name
= "event_source",
5328 .dev_attrs
= pmu_dev_attrs
,
5331 static void pmu_dev_release(struct device
*dev
)
5336 static int pmu_dev_alloc(struct pmu
*pmu
)
5340 pmu
->dev
= kzalloc(sizeof(struct device
), GFP_KERNEL
);
5344 device_initialize(pmu
->dev
);
5345 ret
= dev_set_name(pmu
->dev
, "%s", pmu
->name
);
5349 dev_set_drvdata(pmu
->dev
, pmu
);
5350 pmu
->dev
->bus
= &pmu_bus
;
5351 pmu
->dev
->release
= pmu_dev_release
;
5352 ret
= device_add(pmu
->dev
);
5360 put_device(pmu
->dev
);
5364 int perf_pmu_register(struct pmu
*pmu
, char *name
, int type
)
5368 mutex_lock(&pmus_lock
);
5370 pmu
->pmu_disable_count
= alloc_percpu(int);
5371 if (!pmu
->pmu_disable_count
)
5380 int err
= idr_pre_get(&pmu_idr
, GFP_KERNEL
);
5384 err
= idr_get_new_above(&pmu_idr
, pmu
, PERF_TYPE_MAX
, &type
);
5392 if (pmu_bus_running
) {
5393 ret
= pmu_dev_alloc(pmu
);
5399 pmu
->pmu_cpu_context
= find_pmu_context(pmu
->task_ctx_nr
);
5400 if (pmu
->pmu_cpu_context
)
5401 goto got_cpu_context
;
5403 pmu
->pmu_cpu_context
= alloc_percpu(struct perf_cpu_context
);
5404 if (!pmu
->pmu_cpu_context
)
5407 for_each_possible_cpu(cpu
) {
5408 struct perf_cpu_context
*cpuctx
;
5410 cpuctx
= per_cpu_ptr(pmu
->pmu_cpu_context
, cpu
);
5411 __perf_event_init_context(&cpuctx
->ctx
);
5412 cpuctx
->ctx
.type
= cpu_context
;
5413 cpuctx
->ctx
.pmu
= pmu
;
5414 cpuctx
->jiffies_interval
= 1;
5415 INIT_LIST_HEAD(&cpuctx
->rotation_list
);
5416 cpuctx
->active_pmu
= pmu
;
5420 if (!pmu
->start_txn
) {
5421 if (pmu
->pmu_enable
) {
5423 * If we have pmu_enable/pmu_disable calls, install
5424 * transaction stubs that use that to try and batch
5425 * hardware accesses.
5427 pmu
->start_txn
= perf_pmu_start_txn
;
5428 pmu
->commit_txn
= perf_pmu_commit_txn
;
5429 pmu
->cancel_txn
= perf_pmu_cancel_txn
;
5431 pmu
->start_txn
= perf_pmu_nop_void
;
5432 pmu
->commit_txn
= perf_pmu_nop_int
;
5433 pmu
->cancel_txn
= perf_pmu_nop_void
;
5437 if (!pmu
->pmu_enable
) {
5438 pmu
->pmu_enable
= perf_pmu_nop_void
;
5439 pmu
->pmu_disable
= perf_pmu_nop_void
;
5442 list_add_rcu(&pmu
->entry
, &pmus
);
5445 mutex_unlock(&pmus_lock
);
5450 device_del(pmu
->dev
);
5451 put_device(pmu
->dev
);
5454 if (pmu
->type
>= PERF_TYPE_MAX
)
5455 idr_remove(&pmu_idr
, pmu
->type
);
5458 free_percpu(pmu
->pmu_disable_count
);
5462 void perf_pmu_unregister(struct pmu
*pmu
)
5464 mutex_lock(&pmus_lock
);
5465 list_del_rcu(&pmu
->entry
);
5466 mutex_unlock(&pmus_lock
);
5469 * We dereference the pmu list under both SRCU and regular RCU, so
5470 * synchronize against both of those.
5472 synchronize_srcu(&pmus_srcu
);
5475 free_percpu(pmu
->pmu_disable_count
);
5476 if (pmu
->type
>= PERF_TYPE_MAX
)
5477 idr_remove(&pmu_idr
, pmu
->type
);
5478 device_del(pmu
->dev
);
5479 put_device(pmu
->dev
);
5480 free_pmu_context(pmu
);
5483 struct pmu
*perf_init_event(struct perf_event
*event
)
5485 struct pmu
*pmu
= NULL
;
5488 idx
= srcu_read_lock(&pmus_srcu
);
5491 pmu
= idr_find(&pmu_idr
, event
->attr
.type
);
5496 list_for_each_entry_rcu(pmu
, &pmus
, entry
) {
5497 int ret
= pmu
->event_init(event
);
5501 if (ret
!= -ENOENT
) {
5506 pmu
= ERR_PTR(-ENOENT
);
5508 srcu_read_unlock(&pmus_srcu
, idx
);
5514 * Allocate and initialize a event structure
5516 static struct perf_event
*
5517 perf_event_alloc(struct perf_event_attr
*attr
, int cpu
,
5518 struct task_struct
*task
,
5519 struct perf_event
*group_leader
,
5520 struct perf_event
*parent_event
,
5521 perf_overflow_handler_t overflow_handler
)
5524 struct perf_event
*event
;
5525 struct hw_perf_event
*hwc
;
5528 event
= kzalloc(sizeof(*event
), GFP_KERNEL
);
5530 return ERR_PTR(-ENOMEM
);
5533 * Single events are their own group leaders, with an
5534 * empty sibling list:
5537 group_leader
= event
;
5539 mutex_init(&event
->child_mutex
);
5540 INIT_LIST_HEAD(&event
->child_list
);
5542 INIT_LIST_HEAD(&event
->group_entry
);
5543 INIT_LIST_HEAD(&event
->event_entry
);
5544 INIT_LIST_HEAD(&event
->sibling_list
);
5545 init_waitqueue_head(&event
->waitq
);
5546 init_irq_work(&event
->pending
, perf_pending_event
);
5548 mutex_init(&event
->mmap_mutex
);
5551 event
->attr
= *attr
;
5552 event
->group_leader
= group_leader
;
5556 event
->parent
= parent_event
;
5558 event
->ns
= get_pid_ns(current
->nsproxy
->pid_ns
);
5559 event
->id
= atomic64_inc_return(&perf_event_id
);
5561 event
->state
= PERF_EVENT_STATE_INACTIVE
;
5564 event
->attach_state
= PERF_ATTACH_TASK
;
5565 #ifdef CONFIG_HAVE_HW_BREAKPOINT
5567 * hw_breakpoint is a bit difficult here..
5569 if (attr
->type
== PERF_TYPE_BREAKPOINT
)
5570 event
->hw
.bp_target
= task
;
5574 if (!overflow_handler
&& parent_event
)
5575 overflow_handler
= parent_event
->overflow_handler
;
5577 event
->overflow_handler
= overflow_handler
;
5580 event
->state
= PERF_EVENT_STATE_OFF
;
5585 hwc
->sample_period
= attr
->sample_period
;
5586 if (attr
->freq
&& attr
->sample_freq
)
5587 hwc
->sample_period
= 1;
5588 hwc
->last_period
= hwc
->sample_period
;
5590 local64_set(&hwc
->period_left
, hwc
->sample_period
);
5593 * we currently do not support PERF_FORMAT_GROUP on inherited events
5595 if (attr
->inherit
&& (attr
->read_format
& PERF_FORMAT_GROUP
))
5598 pmu
= perf_init_event(event
);
5604 else if (IS_ERR(pmu
))
5609 put_pid_ns(event
->ns
);
5611 return ERR_PTR(err
);
5616 if (!event
->parent
) {
5617 if (event
->attach_state
& PERF_ATTACH_TASK
)
5618 jump_label_inc(&perf_task_events
);
5619 if (event
->attr
.mmap
|| event
->attr
.mmap_data
)
5620 atomic_inc(&nr_mmap_events
);
5621 if (event
->attr
.comm
)
5622 atomic_inc(&nr_comm_events
);
5623 if (event
->attr
.task
)
5624 atomic_inc(&nr_task_events
);
5625 if (event
->attr
.sample_type
& PERF_SAMPLE_CALLCHAIN
) {
5626 err
= get_callchain_buffers();
5629 return ERR_PTR(err
);
5637 static int perf_copy_attr(struct perf_event_attr __user
*uattr
,
5638 struct perf_event_attr
*attr
)
5643 if (!access_ok(VERIFY_WRITE
, uattr
, PERF_ATTR_SIZE_VER0
))
5647 * zero the full structure, so that a short copy will be nice.
5649 memset(attr
, 0, sizeof(*attr
));
5651 ret
= get_user(size
, &uattr
->size
);
5655 if (size
> PAGE_SIZE
) /* silly large */
5658 if (!size
) /* abi compat */
5659 size
= PERF_ATTR_SIZE_VER0
;
5661 if (size
< PERF_ATTR_SIZE_VER0
)
5665 * If we're handed a bigger struct than we know of,
5666 * ensure all the unknown bits are 0 - i.e. new
5667 * user-space does not rely on any kernel feature
5668 * extensions we dont know about yet.
5670 if (size
> sizeof(*attr
)) {
5671 unsigned char __user
*addr
;
5672 unsigned char __user
*end
;
5675 addr
= (void __user
*)uattr
+ sizeof(*attr
);
5676 end
= (void __user
*)uattr
+ size
;
5678 for (; addr
< end
; addr
++) {
5679 ret
= get_user(val
, addr
);
5685 size
= sizeof(*attr
);
5688 ret
= copy_from_user(attr
, uattr
, size
);
5693 * If the type exists, the corresponding creation will verify
5696 if (attr
->type
>= PERF_TYPE_MAX
)
5699 if (attr
->__reserved_1
)
5702 if (attr
->sample_type
& ~(PERF_SAMPLE_MAX
-1))
5705 if (attr
->read_format
& ~(PERF_FORMAT_MAX
-1))
5712 put_user(sizeof(*attr
), &uattr
->size
);
5718 perf_event_set_output(struct perf_event
*event
, struct perf_event
*output_event
)
5720 struct perf_buffer
*buffer
= NULL
, *old_buffer
= NULL
;
5726 /* don't allow circular references */
5727 if (event
== output_event
)
5731 * Don't allow cross-cpu buffers
5733 if (output_event
->cpu
!= event
->cpu
)
5737 * If its not a per-cpu buffer, it must be the same task.
5739 if (output_event
->cpu
== -1 && output_event
->ctx
!= event
->ctx
)
5743 mutex_lock(&event
->mmap_mutex
);
5744 /* Can't redirect output if we've got an active mmap() */
5745 if (atomic_read(&event
->mmap_count
))
5749 /* get the buffer we want to redirect to */
5750 buffer
= perf_buffer_get(output_event
);
5755 old_buffer
= event
->buffer
;
5756 rcu_assign_pointer(event
->buffer
, buffer
);
5759 mutex_unlock(&event
->mmap_mutex
);
5762 perf_buffer_put(old_buffer
);
5768 * sys_perf_event_open - open a performance event, associate it to a task/cpu
5770 * @attr_uptr: event_id type attributes for monitoring/sampling
5773 * @group_fd: group leader event fd
5775 SYSCALL_DEFINE5(perf_event_open
,
5776 struct perf_event_attr __user
*, attr_uptr
,
5777 pid_t
, pid
, int, cpu
, int, group_fd
, unsigned long, flags
)
5779 struct perf_event
*group_leader
= NULL
, *output_event
= NULL
;
5780 struct perf_event
*event
, *sibling
;
5781 struct perf_event_attr attr
;
5782 struct perf_event_context
*ctx
;
5783 struct file
*event_file
= NULL
;
5784 struct file
*group_file
= NULL
;
5785 struct task_struct
*task
= NULL
;
5789 int fput_needed
= 0;
5792 /* for future expandability... */
5793 if (flags
& ~(PERF_FLAG_FD_NO_GROUP
| PERF_FLAG_FD_OUTPUT
))
5796 err
= perf_copy_attr(attr_uptr
, &attr
);
5800 if (!attr
.exclude_kernel
) {
5801 if (perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN
))
5806 if (attr
.sample_freq
> sysctl_perf_event_sample_rate
)
5810 event_fd
= get_unused_fd_flags(O_RDWR
);
5814 if (group_fd
!= -1) {
5815 group_leader
= perf_fget_light(group_fd
, &fput_needed
);
5816 if (IS_ERR(group_leader
)) {
5817 err
= PTR_ERR(group_leader
);
5820 group_file
= group_leader
->filp
;
5821 if (flags
& PERF_FLAG_FD_OUTPUT
)
5822 output_event
= group_leader
;
5823 if (flags
& PERF_FLAG_FD_NO_GROUP
)
5824 group_leader
= NULL
;
5828 task
= find_lively_task_by_vpid(pid
);
5830 err
= PTR_ERR(task
);
5835 event
= perf_event_alloc(&attr
, cpu
, task
, group_leader
, NULL
, NULL
);
5836 if (IS_ERR(event
)) {
5837 err
= PTR_ERR(event
);
5842 * Special case software events and allow them to be part of
5843 * any hardware group.
5848 (is_software_event(event
) != is_software_event(group_leader
))) {
5849 if (is_software_event(event
)) {
5851 * If event and group_leader are not both a software
5852 * event, and event is, then group leader is not.
5854 * Allow the addition of software events to !software
5855 * groups, this is safe because software events never
5858 pmu
= group_leader
->pmu
;
5859 } else if (is_software_event(group_leader
) &&
5860 (group_leader
->group_flags
& PERF_GROUP_SOFTWARE
)) {
5862 * In case the group is a pure software group, and we
5863 * try to add a hardware event, move the whole group to
5864 * the hardware context.
5871 * Get the target context (task or percpu):
5873 ctx
= find_get_context(pmu
, task
, cpu
);
5880 * Look up the group leader (we will attach this event to it):
5886 * Do not allow a recursive hierarchy (this new sibling
5887 * becoming part of another group-sibling):
5889 if (group_leader
->group_leader
!= group_leader
)
5892 * Do not allow to attach to a group in a different
5893 * task or CPU context:
5896 if (group_leader
->ctx
->type
!= ctx
->type
)
5899 if (group_leader
->ctx
!= ctx
)
5904 * Only a group leader can be exclusive or pinned
5906 if (attr
.exclusive
|| attr
.pinned
)
5911 err
= perf_event_set_output(event
, output_event
);
5916 event_file
= anon_inode_getfile("[perf_event]", &perf_fops
, event
, O_RDWR
);
5917 if (IS_ERR(event_file
)) {
5918 err
= PTR_ERR(event_file
);
5923 struct perf_event_context
*gctx
= group_leader
->ctx
;
5925 mutex_lock(&gctx
->mutex
);
5926 perf_event_remove_from_context(group_leader
);
5927 list_for_each_entry(sibling
, &group_leader
->sibling_list
,
5929 perf_event_remove_from_context(sibling
);
5932 mutex_unlock(&gctx
->mutex
);
5936 event
->filp
= event_file
;
5937 WARN_ON_ONCE(ctx
->parent_ctx
);
5938 mutex_lock(&ctx
->mutex
);
5941 perf_install_in_context(ctx
, group_leader
, cpu
);
5943 list_for_each_entry(sibling
, &group_leader
->sibling_list
,
5945 perf_install_in_context(ctx
, sibling
, cpu
);
5950 perf_install_in_context(ctx
, event
, cpu
);
5952 mutex_unlock(&ctx
->mutex
);
5954 event
->owner
= current
;
5956 mutex_lock(¤t
->perf_event_mutex
);
5957 list_add_tail(&event
->owner_entry
, ¤t
->perf_event_list
);
5958 mutex_unlock(¤t
->perf_event_mutex
);
5961 * Precalculate sample_data sizes
5963 perf_event__header_size(event
);
5964 perf_event__id_header_size(event
);
5967 * Drop the reference on the group_event after placing the
5968 * new event on the sibling_list. This ensures destruction
5969 * of the group leader will find the pointer to itself in
5970 * perf_group_detach().
5972 fput_light(group_file
, fput_needed
);
5973 fd_install(event_fd
, event_file
);
5982 put_task_struct(task
);
5984 fput_light(group_file
, fput_needed
);
5986 put_unused_fd(event_fd
);
5991 * perf_event_create_kernel_counter
5993 * @attr: attributes of the counter to create
5994 * @cpu: cpu in which the counter is bound
5995 * @task: task to profile (NULL for percpu)
5998 perf_event_create_kernel_counter(struct perf_event_attr
*attr
, int cpu
,
5999 struct task_struct
*task
,
6000 perf_overflow_handler_t overflow_handler
)
6002 struct perf_event_context
*ctx
;
6003 struct perf_event
*event
;
6007 * Get the target context (task or percpu):
6010 event
= perf_event_alloc(attr
, cpu
, task
, NULL
, NULL
, overflow_handler
);
6011 if (IS_ERR(event
)) {
6012 err
= PTR_ERR(event
);
6016 ctx
= find_get_context(event
->pmu
, task
, cpu
);
6023 WARN_ON_ONCE(ctx
->parent_ctx
);
6024 mutex_lock(&ctx
->mutex
);
6025 perf_install_in_context(ctx
, event
, cpu
);
6027 mutex_unlock(&ctx
->mutex
);
6034 return ERR_PTR(err
);
6036 EXPORT_SYMBOL_GPL(perf_event_create_kernel_counter
);
6038 static void sync_child_event(struct perf_event
*child_event
,
6039 struct task_struct
*child
)
6041 struct perf_event
*parent_event
= child_event
->parent
;
6044 if (child_event
->attr
.inherit_stat
)
6045 perf_event_read_event(child_event
, child
);
6047 child_val
= perf_event_count(child_event
);
6050 * Add back the child's count to the parent's count:
6052 atomic64_add(child_val
, &parent_event
->child_count
);
6053 atomic64_add(child_event
->total_time_enabled
,
6054 &parent_event
->child_total_time_enabled
);
6055 atomic64_add(child_event
->total_time_running
,
6056 &parent_event
->child_total_time_running
);
6059 * Remove this event from the parent's list
6061 WARN_ON_ONCE(parent_event
->ctx
->parent_ctx
);
6062 mutex_lock(&parent_event
->child_mutex
);
6063 list_del_init(&child_event
->child_list
);
6064 mutex_unlock(&parent_event
->child_mutex
);
6067 * Release the parent event, if this was the last
6070 fput(parent_event
->filp
);
6074 __perf_event_exit_task(struct perf_event
*child_event
,
6075 struct perf_event_context
*child_ctx
,
6076 struct task_struct
*child
)
6078 struct perf_event
*parent_event
;
6080 perf_event_remove_from_context(child_event
);
6082 parent_event
= child_event
->parent
;
6084 * It can happen that parent exits first, and has events
6085 * that are still around due to the child reference. These
6086 * events need to be zapped - but otherwise linger.
6089 sync_child_event(child_event
, child
);
6090 free_event(child_event
);
6094 static void perf_event_exit_task_context(struct task_struct
*child
, int ctxn
)
6096 struct perf_event
*child_event
, *tmp
;
6097 struct perf_event_context
*child_ctx
;
6098 unsigned long flags
;
6100 if (likely(!child
->perf_event_ctxp
[ctxn
])) {
6101 perf_event_task(child
, NULL
, 0);
6105 local_irq_save(flags
);
6107 * We can't reschedule here because interrupts are disabled,
6108 * and either child is current or it is a task that can't be
6109 * scheduled, so we are now safe from rescheduling changing
6112 child_ctx
= child
->perf_event_ctxp
[ctxn
];
6113 task_ctx_sched_out(child_ctx
, EVENT_ALL
);
6116 * Take the context lock here so that if find_get_context is
6117 * reading child->perf_event_ctxp, we wait until it has
6118 * incremented the context's refcount before we do put_ctx below.
6120 raw_spin_lock(&child_ctx
->lock
);
6121 child
->perf_event_ctxp
[ctxn
] = NULL
;
6123 * If this context is a clone; unclone it so it can't get
6124 * swapped to another process while we're removing all
6125 * the events from it.
6127 unclone_ctx(child_ctx
);
6128 update_context_time(child_ctx
);
6129 raw_spin_unlock_irqrestore(&child_ctx
->lock
, flags
);
6132 * Report the task dead after unscheduling the events so that we
6133 * won't get any samples after PERF_RECORD_EXIT. We can however still
6134 * get a few PERF_RECORD_READ events.
6136 perf_event_task(child
, child_ctx
, 0);
6139 * We can recurse on the same lock type through:
6141 * __perf_event_exit_task()
6142 * sync_child_event()
6143 * fput(parent_event->filp)
6145 * mutex_lock(&ctx->mutex)
6147 * But since its the parent context it won't be the same instance.
6149 mutex_lock(&child_ctx
->mutex
);
6152 list_for_each_entry_safe(child_event
, tmp
, &child_ctx
->pinned_groups
,
6154 __perf_event_exit_task(child_event
, child_ctx
, child
);
6156 list_for_each_entry_safe(child_event
, tmp
, &child_ctx
->flexible_groups
,
6158 __perf_event_exit_task(child_event
, child_ctx
, child
);
6161 * If the last event was a group event, it will have appended all
6162 * its siblings to the list, but we obtained 'tmp' before that which
6163 * will still point to the list head terminating the iteration.
6165 if (!list_empty(&child_ctx
->pinned_groups
) ||
6166 !list_empty(&child_ctx
->flexible_groups
))
6169 mutex_unlock(&child_ctx
->mutex
);
6175 * When a child task exits, feed back event values to parent events.
6177 void perf_event_exit_task(struct task_struct
*child
)
6179 struct perf_event
*event
, *tmp
;
6182 mutex_lock(&child
->perf_event_mutex
);
6183 list_for_each_entry_safe(event
, tmp
, &child
->perf_event_list
,
6185 list_del_init(&event
->owner_entry
);
6188 * Ensure the list deletion is visible before we clear
6189 * the owner, closes a race against perf_release() where
6190 * we need to serialize on the owner->perf_event_mutex.
6193 event
->owner
= NULL
;
6195 mutex_unlock(&child
->perf_event_mutex
);
6197 for_each_task_context_nr(ctxn
)
6198 perf_event_exit_task_context(child
, ctxn
);
6201 static void perf_free_event(struct perf_event
*event
,
6202 struct perf_event_context
*ctx
)
6204 struct perf_event
*parent
= event
->parent
;
6206 if (WARN_ON_ONCE(!parent
))
6209 mutex_lock(&parent
->child_mutex
);
6210 list_del_init(&event
->child_list
);
6211 mutex_unlock(&parent
->child_mutex
);
6215 perf_group_detach(event
);
6216 list_del_event(event
, ctx
);
6221 * free an unexposed, unused context as created by inheritance by
6222 * perf_event_init_task below, used by fork() in case of fail.
6224 void perf_event_free_task(struct task_struct
*task
)
6226 struct perf_event_context
*ctx
;
6227 struct perf_event
*event
, *tmp
;
6230 for_each_task_context_nr(ctxn
) {
6231 ctx
= task
->perf_event_ctxp
[ctxn
];
6235 mutex_lock(&ctx
->mutex
);
6237 list_for_each_entry_safe(event
, tmp
, &ctx
->pinned_groups
,
6239 perf_free_event(event
, ctx
);
6241 list_for_each_entry_safe(event
, tmp
, &ctx
->flexible_groups
,
6243 perf_free_event(event
, ctx
);
6245 if (!list_empty(&ctx
->pinned_groups
) ||
6246 !list_empty(&ctx
->flexible_groups
))
6249 mutex_unlock(&ctx
->mutex
);
6255 void perf_event_delayed_put(struct task_struct
*task
)
6259 for_each_task_context_nr(ctxn
)
6260 WARN_ON_ONCE(task
->perf_event_ctxp
[ctxn
]);
6264 * inherit a event from parent task to child task:
6266 static struct perf_event
*
6267 inherit_event(struct perf_event
*parent_event
,
6268 struct task_struct
*parent
,
6269 struct perf_event_context
*parent_ctx
,
6270 struct task_struct
*child
,
6271 struct perf_event
*group_leader
,
6272 struct perf_event_context
*child_ctx
)
6274 struct perf_event
*child_event
;
6275 unsigned long flags
;
6278 * Instead of creating recursive hierarchies of events,
6279 * we link inherited events back to the original parent,
6280 * which has a filp for sure, which we use as the reference
6283 if (parent_event
->parent
)
6284 parent_event
= parent_event
->parent
;
6286 child_event
= perf_event_alloc(&parent_event
->attr
,
6289 group_leader
, parent_event
,
6291 if (IS_ERR(child_event
))
6296 * Make the child state follow the state of the parent event,
6297 * not its attr.disabled bit. We hold the parent's mutex,
6298 * so we won't race with perf_event_{en, dis}able_family.
6300 if (parent_event
->state
>= PERF_EVENT_STATE_INACTIVE
)
6301 child_event
->state
= PERF_EVENT_STATE_INACTIVE
;
6303 child_event
->state
= PERF_EVENT_STATE_OFF
;
6305 if (parent_event
->attr
.freq
) {
6306 u64 sample_period
= parent_event
->hw
.sample_period
;
6307 struct hw_perf_event
*hwc
= &child_event
->hw
;
6309 hwc
->sample_period
= sample_period
;
6310 hwc
->last_period
= sample_period
;
6312 local64_set(&hwc
->period_left
, sample_period
);
6315 child_event
->ctx
= child_ctx
;
6316 child_event
->overflow_handler
= parent_event
->overflow_handler
;
6319 * Precalculate sample_data sizes
6321 perf_event__header_size(child_event
);
6322 perf_event__id_header_size(child_event
);
6325 * Link it up in the child's context:
6327 raw_spin_lock_irqsave(&child_ctx
->lock
, flags
);
6328 add_event_to_ctx(child_event
, child_ctx
);
6329 raw_spin_unlock_irqrestore(&child_ctx
->lock
, flags
);
6332 * Get a reference to the parent filp - we will fput it
6333 * when the child event exits. This is safe to do because
6334 * we are in the parent and we know that the filp still
6335 * exists and has a nonzero count:
6337 atomic_long_inc(&parent_event
->filp
->f_count
);
6340 * Link this into the parent event's child list
6342 WARN_ON_ONCE(parent_event
->ctx
->parent_ctx
);
6343 mutex_lock(&parent_event
->child_mutex
);
6344 list_add_tail(&child_event
->child_list
, &parent_event
->child_list
);
6345 mutex_unlock(&parent_event
->child_mutex
);
6350 static int inherit_group(struct perf_event
*parent_event
,
6351 struct task_struct
*parent
,
6352 struct perf_event_context
*parent_ctx
,
6353 struct task_struct
*child
,
6354 struct perf_event_context
*child_ctx
)
6356 struct perf_event
*leader
;
6357 struct perf_event
*sub
;
6358 struct perf_event
*child_ctr
;
6360 leader
= inherit_event(parent_event
, parent
, parent_ctx
,
6361 child
, NULL
, child_ctx
);
6363 return PTR_ERR(leader
);
6364 list_for_each_entry(sub
, &parent_event
->sibling_list
, group_entry
) {
6365 child_ctr
= inherit_event(sub
, parent
, parent_ctx
,
6366 child
, leader
, child_ctx
);
6367 if (IS_ERR(child_ctr
))
6368 return PTR_ERR(child_ctr
);
6374 inherit_task_group(struct perf_event
*event
, struct task_struct
*parent
,
6375 struct perf_event_context
*parent_ctx
,
6376 struct task_struct
*child
, int ctxn
,
6380 struct perf_event_context
*child_ctx
;
6382 if (!event
->attr
.inherit
) {
6387 child_ctx
= child
->perf_event_ctxp
[ctxn
];
6390 * This is executed from the parent task context, so
6391 * inherit events that have been marked for cloning.
6392 * First allocate and initialize a context for the
6396 child_ctx
= alloc_perf_context(event
->pmu
, child
);
6400 child
->perf_event_ctxp
[ctxn
] = child_ctx
;
6403 ret
= inherit_group(event
, parent
, parent_ctx
,
6413 * Initialize the perf_event context in task_struct
6415 int perf_event_init_context(struct task_struct
*child
, int ctxn
)
6417 struct perf_event_context
*child_ctx
, *parent_ctx
;
6418 struct perf_event_context
*cloned_ctx
;
6419 struct perf_event
*event
;
6420 struct task_struct
*parent
= current
;
6421 int inherited_all
= 1;
6422 unsigned long flags
;
6425 child
->perf_event_ctxp
[ctxn
] = NULL
;
6427 mutex_init(&child
->perf_event_mutex
);
6428 INIT_LIST_HEAD(&child
->perf_event_list
);
6430 if (likely(!parent
->perf_event_ctxp
[ctxn
]))
6434 * If the parent's context is a clone, pin it so it won't get
6437 parent_ctx
= perf_pin_task_context(parent
, ctxn
);
6440 * No need to check if parent_ctx != NULL here; since we saw
6441 * it non-NULL earlier, the only reason for it to become NULL
6442 * is if we exit, and since we're currently in the middle of
6443 * a fork we can't be exiting at the same time.
6447 * Lock the parent list. No need to lock the child - not PID
6448 * hashed yet and not running, so nobody can access it.
6450 mutex_lock(&parent_ctx
->mutex
);
6453 * We dont have to disable NMIs - we are only looking at
6454 * the list, not manipulating it:
6456 list_for_each_entry(event
, &parent_ctx
->pinned_groups
, group_entry
) {
6457 ret
= inherit_task_group(event
, parent
, parent_ctx
,
6458 child
, ctxn
, &inherited_all
);
6464 * We can't hold ctx->lock when iterating the ->flexible_group list due
6465 * to allocations, but we need to prevent rotation because
6466 * rotate_ctx() will change the list from interrupt context.
6468 raw_spin_lock_irqsave(&parent_ctx
->lock
, flags
);
6469 parent_ctx
->rotate_disable
= 1;
6470 raw_spin_unlock_irqrestore(&parent_ctx
->lock
, flags
);
6472 list_for_each_entry(event
, &parent_ctx
->flexible_groups
, group_entry
) {
6473 ret
= inherit_task_group(event
, parent
, parent_ctx
,
6474 child
, ctxn
, &inherited_all
);
6479 raw_spin_lock_irqsave(&parent_ctx
->lock
, flags
);
6480 parent_ctx
->rotate_disable
= 0;
6481 raw_spin_unlock_irqrestore(&parent_ctx
->lock
, flags
);
6483 child_ctx
= child
->perf_event_ctxp
[ctxn
];
6485 if (child_ctx
&& inherited_all
) {
6487 * Mark the child context as a clone of the parent
6488 * context, or of whatever the parent is a clone of.
6489 * Note that if the parent is a clone, it could get
6490 * uncloned at any point, but that doesn't matter
6491 * because the list of events and the generation
6492 * count can't have changed since we took the mutex.
6494 cloned_ctx
= rcu_dereference(parent_ctx
->parent_ctx
);
6496 child_ctx
->parent_ctx
= cloned_ctx
;
6497 child_ctx
->parent_gen
= parent_ctx
->parent_gen
;
6499 child_ctx
->parent_ctx
= parent_ctx
;
6500 child_ctx
->parent_gen
= parent_ctx
->generation
;
6502 get_ctx(child_ctx
->parent_ctx
);
6505 mutex_unlock(&parent_ctx
->mutex
);
6507 perf_unpin_context(parent_ctx
);
6513 * Initialize the perf_event context in task_struct
6515 int perf_event_init_task(struct task_struct
*child
)
6519 for_each_task_context_nr(ctxn
) {
6520 ret
= perf_event_init_context(child
, ctxn
);
6528 static void __init
perf_event_init_all_cpus(void)
6530 struct swevent_htable
*swhash
;
6533 for_each_possible_cpu(cpu
) {
6534 swhash
= &per_cpu(swevent_htable
, cpu
);
6535 mutex_init(&swhash
->hlist_mutex
);
6536 INIT_LIST_HEAD(&per_cpu(rotation_list
, cpu
));
6540 static void __cpuinit
perf_event_init_cpu(int cpu
)
6542 struct swevent_htable
*swhash
= &per_cpu(swevent_htable
, cpu
);
6544 mutex_lock(&swhash
->hlist_mutex
);
6545 if (swhash
->hlist_refcount
> 0) {
6546 struct swevent_hlist
*hlist
;
6548 hlist
= kzalloc_node(sizeof(*hlist
), GFP_KERNEL
, cpu_to_node(cpu
));
6550 rcu_assign_pointer(swhash
->swevent_hlist
, hlist
);
6552 mutex_unlock(&swhash
->hlist_mutex
);
6555 #if defined CONFIG_HOTPLUG_CPU || defined CONFIG_KEXEC
6556 static void perf_pmu_rotate_stop(struct pmu
*pmu
)
6558 struct perf_cpu_context
*cpuctx
= this_cpu_ptr(pmu
->pmu_cpu_context
);
6560 WARN_ON(!irqs_disabled());
6562 list_del_init(&cpuctx
->rotation_list
);
6565 static void __perf_event_exit_context(void *__info
)
6567 struct perf_event_context
*ctx
= __info
;
6568 struct perf_event
*event
, *tmp
;
6570 perf_pmu_rotate_stop(ctx
->pmu
);
6572 list_for_each_entry_safe(event
, tmp
, &ctx
->pinned_groups
, group_entry
)
6573 __perf_event_remove_from_context(event
);
6574 list_for_each_entry_safe(event
, tmp
, &ctx
->flexible_groups
, group_entry
)
6575 __perf_event_remove_from_context(event
);
6578 static void perf_event_exit_cpu_context(int cpu
)
6580 struct perf_event_context
*ctx
;
6584 idx
= srcu_read_lock(&pmus_srcu
);
6585 list_for_each_entry_rcu(pmu
, &pmus
, entry
) {
6586 ctx
= &per_cpu_ptr(pmu
->pmu_cpu_context
, cpu
)->ctx
;
6588 mutex_lock(&ctx
->mutex
);
6589 smp_call_function_single(cpu
, __perf_event_exit_context
, ctx
, 1);
6590 mutex_unlock(&ctx
->mutex
);
6592 srcu_read_unlock(&pmus_srcu
, idx
);
6595 static void perf_event_exit_cpu(int cpu
)
6597 struct swevent_htable
*swhash
= &per_cpu(swevent_htable
, cpu
);
6599 mutex_lock(&swhash
->hlist_mutex
);
6600 swevent_hlist_release(swhash
);
6601 mutex_unlock(&swhash
->hlist_mutex
);
6603 perf_event_exit_cpu_context(cpu
);
6606 static inline void perf_event_exit_cpu(int cpu
) { }
6610 perf_reboot(struct notifier_block
*notifier
, unsigned long val
, void *v
)
6614 for_each_online_cpu(cpu
)
6615 perf_event_exit_cpu(cpu
);
6621 * Run the perf reboot notifier at the very last possible moment so that
6622 * the generic watchdog code runs as long as possible.
6624 static struct notifier_block perf_reboot_notifier
= {
6625 .notifier_call
= perf_reboot
,
6626 .priority
= INT_MIN
,
6629 static int __cpuinit
6630 perf_cpu_notify(struct notifier_block
*self
, unsigned long action
, void *hcpu
)
6632 unsigned int cpu
= (long)hcpu
;
6634 switch (action
& ~CPU_TASKS_FROZEN
) {
6636 case CPU_UP_PREPARE
:
6637 case CPU_DOWN_FAILED
:
6638 perf_event_init_cpu(cpu
);
6641 case CPU_UP_CANCELED
:
6642 case CPU_DOWN_PREPARE
:
6643 perf_event_exit_cpu(cpu
);
6653 void __init
perf_event_init(void)
6659 perf_event_init_all_cpus();
6660 init_srcu_struct(&pmus_srcu
);
6661 perf_pmu_register(&perf_swevent
, "software", PERF_TYPE_SOFTWARE
);
6662 perf_pmu_register(&perf_cpu_clock
, NULL
, -1);
6663 perf_pmu_register(&perf_task_clock
, NULL
, -1);
6665 perf_cpu_notifier(perf_cpu_notify
);
6666 register_reboot_notifier(&perf_reboot_notifier
);
6668 ret
= init_hw_breakpoint();
6669 WARN(ret
, "hw_breakpoint initialization failed with: %d", ret
);
6672 static int __init
perf_event_sysfs_init(void)
6677 mutex_lock(&pmus_lock
);
6679 ret
= bus_register(&pmu_bus
);
6683 list_for_each_entry(pmu
, &pmus
, entry
) {
6684 if (!pmu
->name
|| pmu
->type
< 0)
6687 ret
= pmu_dev_alloc(pmu
);
6688 WARN(ret
, "Failed to register pmu: %s, reason %d\n", pmu
->name
, ret
);
6690 pmu_bus_running
= 1;
6694 mutex_unlock(&pmus_lock
);
6698 device_initcall(perf_event_sysfs_init
);