2 * Performance events core code:
4 * Copyright (C) 2008 Thomas Gleixner <tglx@linutronix.de>
5 * Copyright (C) 2008-2009 Red Hat, Inc., Ingo Molnar
6 * Copyright (C) 2008-2009 Red Hat, Inc., Peter Zijlstra <pzijlstr@redhat.com>
7 * Copyright © 2009 Paul Mackerras, IBM Corp. <paulus@au1.ibm.com>
9 * For licensing details see kernel-base/COPYING
14 #include <linux/cpu.h>
15 #include <linux/smp.h>
16 #include <linux/file.h>
17 #include <linux/poll.h>
18 #include <linux/slab.h>
19 #include <linux/hash.h>
20 #include <linux/sysfs.h>
21 #include <linux/dcache.h>
22 #include <linux/percpu.h>
23 #include <linux/ptrace.h>
24 #include <linux/reboot.h>
25 #include <linux/vmstat.h>
26 #include <linux/vmalloc.h>
27 #include <linux/hardirq.h>
28 #include <linux/rculist.h>
29 #include <linux/uaccess.h>
30 #include <linux/syscalls.h>
31 #include <linux/anon_inodes.h>
32 #include <linux/kernel_stat.h>
33 #include <linux/perf_event.h>
34 #include <linux/ftrace_event.h>
35 #include <linux/hw_breakpoint.h>
37 #include <asm/irq_regs.h>
39 atomic_t perf_task_events __read_mostly
;
40 static atomic_t nr_mmap_events __read_mostly
;
41 static atomic_t nr_comm_events __read_mostly
;
42 static atomic_t nr_task_events __read_mostly
;
44 static LIST_HEAD(pmus
);
45 static DEFINE_MUTEX(pmus_lock
);
46 static struct srcu_struct pmus_srcu
;
49 * perf event paranoia level:
50 * -1 - not paranoid at all
51 * 0 - disallow raw tracepoint access for unpriv
52 * 1 - disallow cpu events for unpriv
53 * 2 - disallow kernel profiling for unpriv
55 int sysctl_perf_event_paranoid __read_mostly
= 1;
57 int sysctl_perf_event_mlock __read_mostly
= 512; /* 'free' kb per user */
60 * max perf event sample rate
62 int sysctl_perf_event_sample_rate __read_mostly
= 100000;
64 static atomic64_t perf_event_id
;
66 void __weak
perf_event_print_debug(void) { }
68 extern __weak
const char *perf_pmu_name(void)
73 void perf_pmu_disable(struct pmu
*pmu
)
75 int *count
= this_cpu_ptr(pmu
->pmu_disable_count
);
77 pmu
->pmu_disable(pmu
);
80 void perf_pmu_enable(struct pmu
*pmu
)
82 int *count
= this_cpu_ptr(pmu
->pmu_disable_count
);
87 static DEFINE_PER_CPU(struct list_head
, rotation_list
);
90 * perf_pmu_rotate_start() and perf_rotate_context() are fully serialized
91 * because they're strictly cpu affine and rotate_start is called with IRQs
92 * disabled, while rotate_context is called from IRQ context.
94 static void perf_pmu_rotate_start(struct pmu
*pmu
)
96 struct perf_cpu_context
*cpuctx
= this_cpu_ptr(pmu
->pmu_cpu_context
);
97 struct list_head
*head
= &__get_cpu_var(rotation_list
);
99 WARN_ON(!irqs_disabled());
101 if (list_empty(&cpuctx
->rotation_list
))
102 list_add(&cpuctx
->rotation_list
, head
);
105 static void get_ctx(struct perf_event_context
*ctx
)
107 WARN_ON(!atomic_inc_not_zero(&ctx
->refcount
));
110 static void free_ctx(struct rcu_head
*head
)
112 struct perf_event_context
*ctx
;
114 ctx
= container_of(head
, struct perf_event_context
, rcu_head
);
118 static void put_ctx(struct perf_event_context
*ctx
)
120 if (atomic_dec_and_test(&ctx
->refcount
)) {
122 put_ctx(ctx
->parent_ctx
);
124 put_task_struct(ctx
->task
);
125 call_rcu(&ctx
->rcu_head
, free_ctx
);
129 static void unclone_ctx(struct perf_event_context
*ctx
)
131 if (ctx
->parent_ctx
) {
132 put_ctx(ctx
->parent_ctx
);
133 ctx
->parent_ctx
= NULL
;
137 static u32
perf_event_pid(struct perf_event
*event
, struct task_struct
*p
)
140 * only top level events have the pid namespace they were created in
143 event
= event
->parent
;
145 return task_tgid_nr_ns(p
, event
->ns
);
148 static u32
perf_event_tid(struct perf_event
*event
, struct task_struct
*p
)
151 * only top level events have the pid namespace they were created in
154 event
= event
->parent
;
156 return task_pid_nr_ns(p
, event
->ns
);
160 * If we inherit events we want to return the parent event id
163 static u64
primary_event_id(struct perf_event
*event
)
168 id
= event
->parent
->id
;
174 * Get the perf_event_context for a task and lock it.
175 * This has to cope with with the fact that until it is locked,
176 * the context could get moved to another task.
178 static struct perf_event_context
*
179 perf_lock_task_context(struct task_struct
*task
, int ctxn
, unsigned long *flags
)
181 struct perf_event_context
*ctx
;
185 ctx
= rcu_dereference(task
->perf_event_ctxp
[ctxn
]);
188 * If this context is a clone of another, it might
189 * get swapped for another underneath us by
190 * perf_event_task_sched_out, though the
191 * rcu_read_lock() protects us from any context
192 * getting freed. Lock the context and check if it
193 * got swapped before we could get the lock, and retry
194 * if so. If we locked the right context, then it
195 * can't get swapped on us any more.
197 raw_spin_lock_irqsave(&ctx
->lock
, *flags
);
198 if (ctx
!= rcu_dereference(task
->perf_event_ctxp
[ctxn
])) {
199 raw_spin_unlock_irqrestore(&ctx
->lock
, *flags
);
203 if (!atomic_inc_not_zero(&ctx
->refcount
)) {
204 raw_spin_unlock_irqrestore(&ctx
->lock
, *flags
);
213 * Get the context for a task and increment its pin_count so it
214 * can't get swapped to another task. This also increments its
215 * reference count so that the context can't get freed.
217 static struct perf_event_context
*
218 perf_pin_task_context(struct task_struct
*task
, int ctxn
)
220 struct perf_event_context
*ctx
;
223 ctx
= perf_lock_task_context(task
, ctxn
, &flags
);
226 raw_spin_unlock_irqrestore(&ctx
->lock
, flags
);
231 static void perf_unpin_context(struct perf_event_context
*ctx
)
235 raw_spin_lock_irqsave(&ctx
->lock
, flags
);
237 raw_spin_unlock_irqrestore(&ctx
->lock
, flags
);
241 static inline u64
perf_clock(void)
243 return local_clock();
247 * Update the record of the current time in a context.
249 static void update_context_time(struct perf_event_context
*ctx
)
251 u64 now
= perf_clock();
253 ctx
->time
+= now
- ctx
->timestamp
;
254 ctx
->timestamp
= now
;
258 * Update the total_time_enabled and total_time_running fields for a event.
260 static void update_event_times(struct perf_event
*event
)
262 struct perf_event_context
*ctx
= event
->ctx
;
265 if (event
->state
< PERF_EVENT_STATE_INACTIVE
||
266 event
->group_leader
->state
< PERF_EVENT_STATE_INACTIVE
)
272 run_end
= event
->tstamp_stopped
;
274 event
->total_time_enabled
= run_end
- event
->tstamp_enabled
;
276 if (event
->state
== PERF_EVENT_STATE_INACTIVE
)
277 run_end
= event
->tstamp_stopped
;
281 event
->total_time_running
= run_end
- event
->tstamp_running
;
285 * Update total_time_enabled and total_time_running for all events in a group.
287 static void update_group_times(struct perf_event
*leader
)
289 struct perf_event
*event
;
291 update_event_times(leader
);
292 list_for_each_entry(event
, &leader
->sibling_list
, group_entry
)
293 update_event_times(event
);
296 static struct list_head
*
297 ctx_group_list(struct perf_event
*event
, struct perf_event_context
*ctx
)
299 if (event
->attr
.pinned
)
300 return &ctx
->pinned_groups
;
302 return &ctx
->flexible_groups
;
306 * Add a event from the lists for its context.
307 * Must be called with ctx->mutex and ctx->lock held.
310 list_add_event(struct perf_event
*event
, struct perf_event_context
*ctx
)
312 WARN_ON_ONCE(event
->attach_state
& PERF_ATTACH_CONTEXT
);
313 event
->attach_state
|= PERF_ATTACH_CONTEXT
;
316 * If we're a stand alone event or group leader, we go to the context
317 * list, group events are kept attached to the group so that
318 * perf_group_detach can, at all times, locate all siblings.
320 if (event
->group_leader
== event
) {
321 struct list_head
*list
;
323 if (is_software_event(event
))
324 event
->group_flags
|= PERF_GROUP_SOFTWARE
;
326 list
= ctx_group_list(event
, ctx
);
327 list_add_tail(&event
->group_entry
, list
);
330 list_add_rcu(&event
->event_entry
, &ctx
->event_list
);
332 perf_pmu_rotate_start(ctx
->pmu
);
334 if (event
->attr
.inherit_stat
)
339 * Called at perf_event creation and when events are attached/detached from a
342 static void perf_event__read_size(struct perf_event
*event
)
344 int entry
= sizeof(u64
); /* value */
348 if (event
->attr
.read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
)
351 if (event
->attr
.read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
)
354 if (event
->attr
.read_format
& PERF_FORMAT_ID
)
355 entry
+= sizeof(u64
);
357 if (event
->attr
.read_format
& PERF_FORMAT_GROUP
) {
358 nr
+= event
->group_leader
->nr_siblings
;
363 event
->read_size
= size
;
366 static void perf_event__header_size(struct perf_event
*event
)
368 struct perf_sample_data
*data
;
369 u64 sample_type
= event
->attr
.sample_type
;
372 perf_event__read_size(event
);
374 if (sample_type
& PERF_SAMPLE_IP
)
375 size
+= sizeof(data
->ip
);
377 if (sample_type
& PERF_SAMPLE_ADDR
)
378 size
+= sizeof(data
->addr
);
380 if (sample_type
& PERF_SAMPLE_PERIOD
)
381 size
+= sizeof(data
->period
);
383 if (sample_type
& PERF_SAMPLE_READ
)
384 size
+= event
->read_size
;
386 event
->header_size
= size
;
389 static void perf_event__id_header_size(struct perf_event
*event
)
391 struct perf_sample_data
*data
;
392 u64 sample_type
= event
->attr
.sample_type
;
395 if (sample_type
& PERF_SAMPLE_TID
)
396 size
+= sizeof(data
->tid_entry
);
398 if (sample_type
& PERF_SAMPLE_TIME
)
399 size
+= sizeof(data
->time
);
401 if (sample_type
& PERF_SAMPLE_ID
)
402 size
+= sizeof(data
->id
);
404 if (sample_type
& PERF_SAMPLE_STREAM_ID
)
405 size
+= sizeof(data
->stream_id
);
407 if (sample_type
& PERF_SAMPLE_CPU
)
408 size
+= sizeof(data
->cpu_entry
);
410 event
->id_header_size
= size
;
413 static void perf_group_attach(struct perf_event
*event
)
415 struct perf_event
*group_leader
= event
->group_leader
, *pos
;
418 * We can have double attach due to group movement in perf_event_open.
420 if (event
->attach_state
& PERF_ATTACH_GROUP
)
423 event
->attach_state
|= PERF_ATTACH_GROUP
;
425 if (group_leader
== event
)
428 if (group_leader
->group_flags
& PERF_GROUP_SOFTWARE
&&
429 !is_software_event(event
))
430 group_leader
->group_flags
&= ~PERF_GROUP_SOFTWARE
;
432 list_add_tail(&event
->group_entry
, &group_leader
->sibling_list
);
433 group_leader
->nr_siblings
++;
435 perf_event__header_size(group_leader
);
437 list_for_each_entry(pos
, &group_leader
->sibling_list
, group_entry
)
438 perf_event__header_size(pos
);
442 * Remove a event from the lists for its context.
443 * Must be called with ctx->mutex and ctx->lock held.
446 list_del_event(struct perf_event
*event
, struct perf_event_context
*ctx
)
449 * We can have double detach due to exit/hot-unplug + close.
451 if (!(event
->attach_state
& PERF_ATTACH_CONTEXT
))
454 event
->attach_state
&= ~PERF_ATTACH_CONTEXT
;
457 if (event
->attr
.inherit_stat
)
460 list_del_rcu(&event
->event_entry
);
462 if (event
->group_leader
== event
)
463 list_del_init(&event
->group_entry
);
465 update_group_times(event
);
468 * If event was in error state, then keep it
469 * that way, otherwise bogus counts will be
470 * returned on read(). The only way to get out
471 * of error state is by explicit re-enabling
474 if (event
->state
> PERF_EVENT_STATE_OFF
)
475 event
->state
= PERF_EVENT_STATE_OFF
;
478 static void perf_group_detach(struct perf_event
*event
)
480 struct perf_event
*sibling
, *tmp
;
481 struct list_head
*list
= NULL
;
484 * We can have double detach due to exit/hot-unplug + close.
486 if (!(event
->attach_state
& PERF_ATTACH_GROUP
))
489 event
->attach_state
&= ~PERF_ATTACH_GROUP
;
492 * If this is a sibling, remove it from its group.
494 if (event
->group_leader
!= event
) {
495 list_del_init(&event
->group_entry
);
496 event
->group_leader
->nr_siblings
--;
500 if (!list_empty(&event
->group_entry
))
501 list
= &event
->group_entry
;
504 * If this was a group event with sibling events then
505 * upgrade the siblings to singleton events by adding them
506 * to whatever list we are on.
508 list_for_each_entry_safe(sibling
, tmp
, &event
->sibling_list
, group_entry
) {
510 list_move_tail(&sibling
->group_entry
, list
);
511 sibling
->group_leader
= sibling
;
513 /* Inherit group flags from the previous leader */
514 sibling
->group_flags
= event
->group_flags
;
518 perf_event__header_size(event
->group_leader
);
520 list_for_each_entry(tmp
, &event
->group_leader
->sibling_list
, group_entry
)
521 perf_event__header_size(tmp
);
525 event_filter_match(struct perf_event
*event
)
527 return event
->cpu
== -1 || event
->cpu
== smp_processor_id();
531 event_sched_out(struct perf_event
*event
,
532 struct perf_cpu_context
*cpuctx
,
533 struct perf_event_context
*ctx
)
537 * An event which could not be activated because of
538 * filter mismatch still needs to have its timings
539 * maintained, otherwise bogus information is return
540 * via read() for time_enabled, time_running:
542 if (event
->state
== PERF_EVENT_STATE_INACTIVE
543 && !event_filter_match(event
)) {
544 delta
= ctx
->time
- event
->tstamp_stopped
;
545 event
->tstamp_running
+= delta
;
546 event
->tstamp_stopped
= ctx
->time
;
549 if (event
->state
!= PERF_EVENT_STATE_ACTIVE
)
552 event
->state
= PERF_EVENT_STATE_INACTIVE
;
553 if (event
->pending_disable
) {
554 event
->pending_disable
= 0;
555 event
->state
= PERF_EVENT_STATE_OFF
;
557 event
->tstamp_stopped
= ctx
->time
;
558 event
->pmu
->del(event
, 0);
561 if (!is_software_event(event
))
562 cpuctx
->active_oncpu
--;
564 if (event
->attr
.exclusive
|| !cpuctx
->active_oncpu
)
565 cpuctx
->exclusive
= 0;
569 group_sched_out(struct perf_event
*group_event
,
570 struct perf_cpu_context
*cpuctx
,
571 struct perf_event_context
*ctx
)
573 struct perf_event
*event
;
574 int state
= group_event
->state
;
576 event_sched_out(group_event
, cpuctx
, ctx
);
579 * Schedule out siblings (if any):
581 list_for_each_entry(event
, &group_event
->sibling_list
, group_entry
)
582 event_sched_out(event
, cpuctx
, ctx
);
584 if (state
== PERF_EVENT_STATE_ACTIVE
&& group_event
->attr
.exclusive
)
585 cpuctx
->exclusive
= 0;
588 static inline struct perf_cpu_context
*
589 __get_cpu_context(struct perf_event_context
*ctx
)
591 return this_cpu_ptr(ctx
->pmu
->pmu_cpu_context
);
595 * Cross CPU call to remove a performance event
597 * We disable the event on the hardware level first. After that we
598 * remove it from the context list.
600 static void __perf_event_remove_from_context(void *info
)
602 struct perf_event
*event
= info
;
603 struct perf_event_context
*ctx
= event
->ctx
;
604 struct perf_cpu_context
*cpuctx
= __get_cpu_context(ctx
);
607 * If this is a task context, we need to check whether it is
608 * the current task context of this cpu. If not it has been
609 * scheduled out before the smp call arrived.
611 if (ctx
->task
&& cpuctx
->task_ctx
!= ctx
)
614 raw_spin_lock(&ctx
->lock
);
616 event_sched_out(event
, cpuctx
, ctx
);
618 list_del_event(event
, ctx
);
620 raw_spin_unlock(&ctx
->lock
);
625 * Remove the event from a task's (or a CPU's) list of events.
627 * Must be called with ctx->mutex held.
629 * CPU events are removed with a smp call. For task events we only
630 * call when the task is on a CPU.
632 * If event->ctx is a cloned context, callers must make sure that
633 * every task struct that event->ctx->task could possibly point to
634 * remains valid. This is OK when called from perf_release since
635 * that only calls us on the top-level context, which can't be a clone.
636 * When called from perf_event_exit_task, it's OK because the
637 * context has been detached from its task.
639 static void perf_event_remove_from_context(struct perf_event
*event
)
641 struct perf_event_context
*ctx
= event
->ctx
;
642 struct task_struct
*task
= ctx
->task
;
646 * Per cpu events are removed via an smp call and
647 * the removal is always successful.
649 smp_call_function_single(event
->cpu
,
650 __perf_event_remove_from_context
,
656 task_oncpu_function_call(task
, __perf_event_remove_from_context
,
659 raw_spin_lock_irq(&ctx
->lock
);
661 * If the context is active we need to retry the smp call.
663 if (ctx
->nr_active
&& !list_empty(&event
->group_entry
)) {
664 raw_spin_unlock_irq(&ctx
->lock
);
669 * The lock prevents that this context is scheduled in so we
670 * can remove the event safely, if the call above did not
673 if (!list_empty(&event
->group_entry
))
674 list_del_event(event
, ctx
);
675 raw_spin_unlock_irq(&ctx
->lock
);
679 * Cross CPU call to disable a performance event
681 static void __perf_event_disable(void *info
)
683 struct perf_event
*event
= info
;
684 struct perf_event_context
*ctx
= event
->ctx
;
685 struct perf_cpu_context
*cpuctx
= __get_cpu_context(ctx
);
688 * If this is a per-task event, need to check whether this
689 * event's task is the current task on this cpu.
691 if (ctx
->task
&& cpuctx
->task_ctx
!= ctx
)
694 raw_spin_lock(&ctx
->lock
);
697 * If the event is on, turn it off.
698 * If it is in error state, leave it in error state.
700 if (event
->state
>= PERF_EVENT_STATE_INACTIVE
) {
701 update_context_time(ctx
);
702 update_group_times(event
);
703 if (event
== event
->group_leader
)
704 group_sched_out(event
, cpuctx
, ctx
);
706 event_sched_out(event
, cpuctx
, ctx
);
707 event
->state
= PERF_EVENT_STATE_OFF
;
710 raw_spin_unlock(&ctx
->lock
);
716 * If event->ctx is a cloned context, callers must make sure that
717 * every task struct that event->ctx->task could possibly point to
718 * remains valid. This condition is satisifed when called through
719 * perf_event_for_each_child or perf_event_for_each because they
720 * hold the top-level event's child_mutex, so any descendant that
721 * goes to exit will block in sync_child_event.
722 * When called from perf_pending_event it's OK because event->ctx
723 * is the current context on this CPU and preemption is disabled,
724 * hence we can't get into perf_event_task_sched_out for this context.
726 void perf_event_disable(struct perf_event
*event
)
728 struct perf_event_context
*ctx
= event
->ctx
;
729 struct task_struct
*task
= ctx
->task
;
733 * Disable the event on the cpu that it's on
735 smp_call_function_single(event
->cpu
, __perf_event_disable
,
741 task_oncpu_function_call(task
, __perf_event_disable
, event
);
743 raw_spin_lock_irq(&ctx
->lock
);
745 * If the event is still active, we need to retry the cross-call.
747 if (event
->state
== PERF_EVENT_STATE_ACTIVE
) {
748 raw_spin_unlock_irq(&ctx
->lock
);
753 * Since we have the lock this context can't be scheduled
754 * in, so we can change the state safely.
756 if (event
->state
== PERF_EVENT_STATE_INACTIVE
) {
757 update_group_times(event
);
758 event
->state
= PERF_EVENT_STATE_OFF
;
761 raw_spin_unlock_irq(&ctx
->lock
);
765 event_sched_in(struct perf_event
*event
,
766 struct perf_cpu_context
*cpuctx
,
767 struct perf_event_context
*ctx
)
769 if (event
->state
<= PERF_EVENT_STATE_OFF
)
772 event
->state
= PERF_EVENT_STATE_ACTIVE
;
773 event
->oncpu
= smp_processor_id();
775 * The new state must be visible before we turn it on in the hardware:
779 if (event
->pmu
->add(event
, PERF_EF_START
)) {
780 event
->state
= PERF_EVENT_STATE_INACTIVE
;
785 event
->tstamp_running
+= ctx
->time
- event
->tstamp_stopped
;
787 event
->shadow_ctx_time
= ctx
->time
- ctx
->timestamp
;
789 if (!is_software_event(event
))
790 cpuctx
->active_oncpu
++;
793 if (event
->attr
.exclusive
)
794 cpuctx
->exclusive
= 1;
800 group_sched_in(struct perf_event
*group_event
,
801 struct perf_cpu_context
*cpuctx
,
802 struct perf_event_context
*ctx
)
804 struct perf_event
*event
, *partial_group
= NULL
;
805 struct pmu
*pmu
= group_event
->pmu
;
807 bool simulate
= false;
809 if (group_event
->state
== PERF_EVENT_STATE_OFF
)
814 if (event_sched_in(group_event
, cpuctx
, ctx
)) {
815 pmu
->cancel_txn(pmu
);
820 * Schedule in siblings as one group (if any):
822 list_for_each_entry(event
, &group_event
->sibling_list
, group_entry
) {
823 if (event_sched_in(event
, cpuctx
, ctx
)) {
824 partial_group
= event
;
829 if (!pmu
->commit_txn(pmu
))
834 * Groups can be scheduled in as one unit only, so undo any
835 * partial group before returning:
836 * The events up to the failed event are scheduled out normally,
837 * tstamp_stopped will be updated.
839 * The failed events and the remaining siblings need to have
840 * their timings updated as if they had gone thru event_sched_in()
841 * and event_sched_out(). This is required to get consistent timings
842 * across the group. This also takes care of the case where the group
843 * could never be scheduled by ensuring tstamp_stopped is set to mark
844 * the time the event was actually stopped, such that time delta
845 * calculation in update_event_times() is correct.
847 list_for_each_entry(event
, &group_event
->sibling_list
, group_entry
) {
848 if (event
== partial_group
)
852 event
->tstamp_running
+= now
- event
->tstamp_stopped
;
853 event
->tstamp_stopped
= now
;
855 event_sched_out(event
, cpuctx
, ctx
);
858 event_sched_out(group_event
, cpuctx
, ctx
);
860 pmu
->cancel_txn(pmu
);
866 * Work out whether we can put this event group on the CPU now.
868 static int group_can_go_on(struct perf_event
*event
,
869 struct perf_cpu_context
*cpuctx
,
873 * Groups consisting entirely of software events can always go on.
875 if (event
->group_flags
& PERF_GROUP_SOFTWARE
)
878 * If an exclusive group is already on, no other hardware
881 if (cpuctx
->exclusive
)
884 * If this group is exclusive and there are already
885 * events on the CPU, it can't go on.
887 if (event
->attr
.exclusive
&& cpuctx
->active_oncpu
)
890 * Otherwise, try to add it if all previous groups were able
896 static void add_event_to_ctx(struct perf_event
*event
,
897 struct perf_event_context
*ctx
)
899 list_add_event(event
, ctx
);
900 perf_group_attach(event
);
901 event
->tstamp_enabled
= ctx
->time
;
902 event
->tstamp_running
= ctx
->time
;
903 event
->tstamp_stopped
= ctx
->time
;
907 * Cross CPU call to install and enable a performance event
909 * Must be called with ctx->mutex held
911 static void __perf_install_in_context(void *info
)
913 struct perf_event
*event
= info
;
914 struct perf_event_context
*ctx
= event
->ctx
;
915 struct perf_event
*leader
= event
->group_leader
;
916 struct perf_cpu_context
*cpuctx
= __get_cpu_context(ctx
);
920 * If this is a task context, we need to check whether it is
921 * the current task context of this cpu. If not it has been
922 * scheduled out before the smp call arrived.
923 * Or possibly this is the right context but it isn't
924 * on this cpu because it had no events.
926 if (ctx
->task
&& cpuctx
->task_ctx
!= ctx
) {
927 if (cpuctx
->task_ctx
|| ctx
->task
!= current
)
929 cpuctx
->task_ctx
= ctx
;
932 raw_spin_lock(&ctx
->lock
);
934 update_context_time(ctx
);
936 add_event_to_ctx(event
, ctx
);
938 if (event
->cpu
!= -1 && event
->cpu
!= smp_processor_id())
942 * Don't put the event on if it is disabled or if
943 * it is in a group and the group isn't on.
945 if (event
->state
!= PERF_EVENT_STATE_INACTIVE
||
946 (leader
!= event
&& leader
->state
!= PERF_EVENT_STATE_ACTIVE
))
950 * An exclusive event can't go on if there are already active
951 * hardware events, and no hardware event can go on if there
952 * is already an exclusive event on.
954 if (!group_can_go_on(event
, cpuctx
, 1))
957 err
= event_sched_in(event
, cpuctx
, ctx
);
961 * This event couldn't go on. If it is in a group
962 * then we have to pull the whole group off.
963 * If the event group is pinned then put it in error state.
966 group_sched_out(leader
, cpuctx
, ctx
);
967 if (leader
->attr
.pinned
) {
968 update_group_times(leader
);
969 leader
->state
= PERF_EVENT_STATE_ERROR
;
974 raw_spin_unlock(&ctx
->lock
);
978 * Attach a performance event to a context
980 * First we add the event to the list with the hardware enable bit
981 * in event->hw_config cleared.
983 * If the event is attached to a task which is on a CPU we use a smp
984 * call to enable it in the task context. The task might have been
985 * scheduled away, but we check this in the smp call again.
987 * Must be called with ctx->mutex held.
990 perf_install_in_context(struct perf_event_context
*ctx
,
991 struct perf_event
*event
,
994 struct task_struct
*task
= ctx
->task
;
1000 * Per cpu events are installed via an smp call and
1001 * the install is always successful.
1003 smp_call_function_single(cpu
, __perf_install_in_context
,
1009 task_oncpu_function_call(task
, __perf_install_in_context
,
1012 raw_spin_lock_irq(&ctx
->lock
);
1014 * we need to retry the smp call.
1016 if (ctx
->is_active
&& list_empty(&event
->group_entry
)) {
1017 raw_spin_unlock_irq(&ctx
->lock
);
1022 * The lock prevents that this context is scheduled in so we
1023 * can add the event safely, if it the call above did not
1026 if (list_empty(&event
->group_entry
))
1027 add_event_to_ctx(event
, ctx
);
1028 raw_spin_unlock_irq(&ctx
->lock
);
1032 * Put a event into inactive state and update time fields.
1033 * Enabling the leader of a group effectively enables all
1034 * the group members that aren't explicitly disabled, so we
1035 * have to update their ->tstamp_enabled also.
1036 * Note: this works for group members as well as group leaders
1037 * since the non-leader members' sibling_lists will be empty.
1039 static void __perf_event_mark_enabled(struct perf_event
*event
,
1040 struct perf_event_context
*ctx
)
1042 struct perf_event
*sub
;
1044 event
->state
= PERF_EVENT_STATE_INACTIVE
;
1045 event
->tstamp_enabled
= ctx
->time
- event
->total_time_enabled
;
1046 list_for_each_entry(sub
, &event
->sibling_list
, group_entry
) {
1047 if (sub
->state
>= PERF_EVENT_STATE_INACTIVE
) {
1048 sub
->tstamp_enabled
=
1049 ctx
->time
- sub
->total_time_enabled
;
1055 * Cross CPU call to enable a performance event
1057 static void __perf_event_enable(void *info
)
1059 struct perf_event
*event
= info
;
1060 struct perf_event_context
*ctx
= event
->ctx
;
1061 struct perf_event
*leader
= event
->group_leader
;
1062 struct perf_cpu_context
*cpuctx
= __get_cpu_context(ctx
);
1066 * If this is a per-task event, need to check whether this
1067 * event's task is the current task on this cpu.
1069 if (ctx
->task
&& cpuctx
->task_ctx
!= ctx
) {
1070 if (cpuctx
->task_ctx
|| ctx
->task
!= current
)
1072 cpuctx
->task_ctx
= ctx
;
1075 raw_spin_lock(&ctx
->lock
);
1077 update_context_time(ctx
);
1079 if (event
->state
>= PERF_EVENT_STATE_INACTIVE
)
1081 __perf_event_mark_enabled(event
, ctx
);
1083 if (event
->cpu
!= -1 && event
->cpu
!= smp_processor_id())
1087 * If the event is in a group and isn't the group leader,
1088 * then don't put it on unless the group is on.
1090 if (leader
!= event
&& leader
->state
!= PERF_EVENT_STATE_ACTIVE
)
1093 if (!group_can_go_on(event
, cpuctx
, 1)) {
1096 if (event
== leader
)
1097 err
= group_sched_in(event
, cpuctx
, ctx
);
1099 err
= event_sched_in(event
, cpuctx
, ctx
);
1104 * If this event can't go on and it's part of a
1105 * group, then the whole group has to come off.
1107 if (leader
!= event
)
1108 group_sched_out(leader
, cpuctx
, ctx
);
1109 if (leader
->attr
.pinned
) {
1110 update_group_times(leader
);
1111 leader
->state
= PERF_EVENT_STATE_ERROR
;
1116 raw_spin_unlock(&ctx
->lock
);
1122 * If event->ctx is a cloned context, callers must make sure that
1123 * every task struct that event->ctx->task could possibly point to
1124 * remains valid. This condition is satisfied when called through
1125 * perf_event_for_each_child or perf_event_for_each as described
1126 * for perf_event_disable.
1128 void perf_event_enable(struct perf_event
*event
)
1130 struct perf_event_context
*ctx
= event
->ctx
;
1131 struct task_struct
*task
= ctx
->task
;
1135 * Enable the event on the cpu that it's on
1137 smp_call_function_single(event
->cpu
, __perf_event_enable
,
1142 raw_spin_lock_irq(&ctx
->lock
);
1143 if (event
->state
>= PERF_EVENT_STATE_INACTIVE
)
1147 * If the event is in error state, clear that first.
1148 * That way, if we see the event in error state below, we
1149 * know that it has gone back into error state, as distinct
1150 * from the task having been scheduled away before the
1151 * cross-call arrived.
1153 if (event
->state
== PERF_EVENT_STATE_ERROR
)
1154 event
->state
= PERF_EVENT_STATE_OFF
;
1157 raw_spin_unlock_irq(&ctx
->lock
);
1158 task_oncpu_function_call(task
, __perf_event_enable
, event
);
1160 raw_spin_lock_irq(&ctx
->lock
);
1163 * If the context is active and the event is still off,
1164 * we need to retry the cross-call.
1166 if (ctx
->is_active
&& event
->state
== PERF_EVENT_STATE_OFF
)
1170 * Since we have the lock this context can't be scheduled
1171 * in, so we can change the state safely.
1173 if (event
->state
== PERF_EVENT_STATE_OFF
)
1174 __perf_event_mark_enabled(event
, ctx
);
1177 raw_spin_unlock_irq(&ctx
->lock
);
1180 static int perf_event_refresh(struct perf_event
*event
, int refresh
)
1183 * not supported on inherited events
1185 if (event
->attr
.inherit
|| !is_sampling_event(event
))
1188 atomic_add(refresh
, &event
->event_limit
);
1189 perf_event_enable(event
);
1195 EVENT_FLEXIBLE
= 0x1,
1197 EVENT_ALL
= EVENT_FLEXIBLE
| EVENT_PINNED
,
1200 static void ctx_sched_out(struct perf_event_context
*ctx
,
1201 struct perf_cpu_context
*cpuctx
,
1202 enum event_type_t event_type
)
1204 struct perf_event
*event
;
1206 raw_spin_lock(&ctx
->lock
);
1207 perf_pmu_disable(ctx
->pmu
);
1209 if (likely(!ctx
->nr_events
))
1211 update_context_time(ctx
);
1213 if (!ctx
->nr_active
)
1216 if (event_type
& EVENT_PINNED
) {
1217 list_for_each_entry(event
, &ctx
->pinned_groups
, group_entry
)
1218 group_sched_out(event
, cpuctx
, ctx
);
1221 if (event_type
& EVENT_FLEXIBLE
) {
1222 list_for_each_entry(event
, &ctx
->flexible_groups
, group_entry
)
1223 group_sched_out(event
, cpuctx
, ctx
);
1226 perf_pmu_enable(ctx
->pmu
);
1227 raw_spin_unlock(&ctx
->lock
);
1231 * Test whether two contexts are equivalent, i.e. whether they
1232 * have both been cloned from the same version of the same context
1233 * and they both have the same number of enabled events.
1234 * If the number of enabled events is the same, then the set
1235 * of enabled events should be the same, because these are both
1236 * inherited contexts, therefore we can't access individual events
1237 * in them directly with an fd; we can only enable/disable all
1238 * events via prctl, or enable/disable all events in a family
1239 * via ioctl, which will have the same effect on both contexts.
1241 static int context_equiv(struct perf_event_context
*ctx1
,
1242 struct perf_event_context
*ctx2
)
1244 return ctx1
->parent_ctx
&& ctx1
->parent_ctx
== ctx2
->parent_ctx
1245 && ctx1
->parent_gen
== ctx2
->parent_gen
1246 && !ctx1
->pin_count
&& !ctx2
->pin_count
;
1249 static void __perf_event_sync_stat(struct perf_event
*event
,
1250 struct perf_event
*next_event
)
1254 if (!event
->attr
.inherit_stat
)
1258 * Update the event value, we cannot use perf_event_read()
1259 * because we're in the middle of a context switch and have IRQs
1260 * disabled, which upsets smp_call_function_single(), however
1261 * we know the event must be on the current CPU, therefore we
1262 * don't need to use it.
1264 switch (event
->state
) {
1265 case PERF_EVENT_STATE_ACTIVE
:
1266 event
->pmu
->read(event
);
1269 case PERF_EVENT_STATE_INACTIVE
:
1270 update_event_times(event
);
1278 * In order to keep per-task stats reliable we need to flip the event
1279 * values when we flip the contexts.
1281 value
= local64_read(&next_event
->count
);
1282 value
= local64_xchg(&event
->count
, value
);
1283 local64_set(&next_event
->count
, value
);
1285 swap(event
->total_time_enabled
, next_event
->total_time_enabled
);
1286 swap(event
->total_time_running
, next_event
->total_time_running
);
1289 * Since we swizzled the values, update the user visible data too.
1291 perf_event_update_userpage(event
);
1292 perf_event_update_userpage(next_event
);
1295 #define list_next_entry(pos, member) \
1296 list_entry(pos->member.next, typeof(*pos), member)
1298 static void perf_event_sync_stat(struct perf_event_context
*ctx
,
1299 struct perf_event_context
*next_ctx
)
1301 struct perf_event
*event
, *next_event
;
1306 update_context_time(ctx
);
1308 event
= list_first_entry(&ctx
->event_list
,
1309 struct perf_event
, event_entry
);
1311 next_event
= list_first_entry(&next_ctx
->event_list
,
1312 struct perf_event
, event_entry
);
1314 while (&event
->event_entry
!= &ctx
->event_list
&&
1315 &next_event
->event_entry
!= &next_ctx
->event_list
) {
1317 __perf_event_sync_stat(event
, next_event
);
1319 event
= list_next_entry(event
, event_entry
);
1320 next_event
= list_next_entry(next_event
, event_entry
);
1324 void perf_event_context_sched_out(struct task_struct
*task
, int ctxn
,
1325 struct task_struct
*next
)
1327 struct perf_event_context
*ctx
= task
->perf_event_ctxp
[ctxn
];
1328 struct perf_event_context
*next_ctx
;
1329 struct perf_event_context
*parent
;
1330 struct perf_cpu_context
*cpuctx
;
1336 cpuctx
= __get_cpu_context(ctx
);
1337 if (!cpuctx
->task_ctx
)
1341 parent
= rcu_dereference(ctx
->parent_ctx
);
1342 next_ctx
= next
->perf_event_ctxp
[ctxn
];
1343 if (parent
&& next_ctx
&&
1344 rcu_dereference(next_ctx
->parent_ctx
) == parent
) {
1346 * Looks like the two contexts are clones, so we might be
1347 * able to optimize the context switch. We lock both
1348 * contexts and check that they are clones under the
1349 * lock (including re-checking that neither has been
1350 * uncloned in the meantime). It doesn't matter which
1351 * order we take the locks because no other cpu could
1352 * be trying to lock both of these tasks.
1354 raw_spin_lock(&ctx
->lock
);
1355 raw_spin_lock_nested(&next_ctx
->lock
, SINGLE_DEPTH_NESTING
);
1356 if (context_equiv(ctx
, next_ctx
)) {
1358 * XXX do we need a memory barrier of sorts
1359 * wrt to rcu_dereference() of perf_event_ctxp
1361 task
->perf_event_ctxp
[ctxn
] = next_ctx
;
1362 next
->perf_event_ctxp
[ctxn
] = ctx
;
1364 next_ctx
->task
= task
;
1367 perf_event_sync_stat(ctx
, next_ctx
);
1369 raw_spin_unlock(&next_ctx
->lock
);
1370 raw_spin_unlock(&ctx
->lock
);
1375 ctx_sched_out(ctx
, cpuctx
, EVENT_ALL
);
1376 cpuctx
->task_ctx
= NULL
;
1380 #define for_each_task_context_nr(ctxn) \
1381 for ((ctxn) = 0; (ctxn) < perf_nr_task_contexts; (ctxn)++)
1384 * Called from scheduler to remove the events of the current task,
1385 * with interrupts disabled.
1387 * We stop each event and update the event value in event->count.
1389 * This does not protect us against NMI, but disable()
1390 * sets the disabled bit in the control field of event _before_
1391 * accessing the event control register. If a NMI hits, then it will
1392 * not restart the event.
1394 void __perf_event_task_sched_out(struct task_struct
*task
,
1395 struct task_struct
*next
)
1399 for_each_task_context_nr(ctxn
)
1400 perf_event_context_sched_out(task
, ctxn
, next
);
1403 static void task_ctx_sched_out(struct perf_event_context
*ctx
,
1404 enum event_type_t event_type
)
1406 struct perf_cpu_context
*cpuctx
= __get_cpu_context(ctx
);
1408 if (!cpuctx
->task_ctx
)
1411 if (WARN_ON_ONCE(ctx
!= cpuctx
->task_ctx
))
1414 ctx_sched_out(ctx
, cpuctx
, event_type
);
1415 cpuctx
->task_ctx
= NULL
;
1419 * Called with IRQs disabled
1421 static void cpu_ctx_sched_out(struct perf_cpu_context
*cpuctx
,
1422 enum event_type_t event_type
)
1424 ctx_sched_out(&cpuctx
->ctx
, cpuctx
, event_type
);
1428 ctx_pinned_sched_in(struct perf_event_context
*ctx
,
1429 struct perf_cpu_context
*cpuctx
)
1431 struct perf_event
*event
;
1433 list_for_each_entry(event
, &ctx
->pinned_groups
, group_entry
) {
1434 if (event
->state
<= PERF_EVENT_STATE_OFF
)
1436 if (event
->cpu
!= -1 && event
->cpu
!= smp_processor_id())
1439 if (group_can_go_on(event
, cpuctx
, 1))
1440 group_sched_in(event
, cpuctx
, ctx
);
1443 * If this pinned group hasn't been scheduled,
1444 * put it in error state.
1446 if (event
->state
== PERF_EVENT_STATE_INACTIVE
) {
1447 update_group_times(event
);
1448 event
->state
= PERF_EVENT_STATE_ERROR
;
1454 ctx_flexible_sched_in(struct perf_event_context
*ctx
,
1455 struct perf_cpu_context
*cpuctx
)
1457 struct perf_event
*event
;
1460 list_for_each_entry(event
, &ctx
->flexible_groups
, group_entry
) {
1461 /* Ignore events in OFF or ERROR state */
1462 if (event
->state
<= PERF_EVENT_STATE_OFF
)
1465 * Listen to the 'cpu' scheduling filter constraint
1468 if (event
->cpu
!= -1 && event
->cpu
!= smp_processor_id())
1471 if (group_can_go_on(event
, cpuctx
, can_add_hw
)) {
1472 if (group_sched_in(event
, cpuctx
, ctx
))
1479 ctx_sched_in(struct perf_event_context
*ctx
,
1480 struct perf_cpu_context
*cpuctx
,
1481 enum event_type_t event_type
)
1483 raw_spin_lock(&ctx
->lock
);
1485 if (likely(!ctx
->nr_events
))
1488 ctx
->timestamp
= perf_clock();
1491 * First go through the list and put on any pinned groups
1492 * in order to give them the best chance of going on.
1494 if (event_type
& EVENT_PINNED
)
1495 ctx_pinned_sched_in(ctx
, cpuctx
);
1497 /* Then walk through the lower prio flexible groups */
1498 if (event_type
& EVENT_FLEXIBLE
)
1499 ctx_flexible_sched_in(ctx
, cpuctx
);
1502 raw_spin_unlock(&ctx
->lock
);
1505 static void cpu_ctx_sched_in(struct perf_cpu_context
*cpuctx
,
1506 enum event_type_t event_type
)
1508 struct perf_event_context
*ctx
= &cpuctx
->ctx
;
1510 ctx_sched_in(ctx
, cpuctx
, event_type
);
1513 static void task_ctx_sched_in(struct perf_event_context
*ctx
,
1514 enum event_type_t event_type
)
1516 struct perf_cpu_context
*cpuctx
;
1518 cpuctx
= __get_cpu_context(ctx
);
1519 if (cpuctx
->task_ctx
== ctx
)
1522 ctx_sched_in(ctx
, cpuctx
, event_type
);
1523 cpuctx
->task_ctx
= ctx
;
1526 void perf_event_context_sched_in(struct perf_event_context
*ctx
)
1528 struct perf_cpu_context
*cpuctx
;
1530 cpuctx
= __get_cpu_context(ctx
);
1531 if (cpuctx
->task_ctx
== ctx
)
1534 perf_pmu_disable(ctx
->pmu
);
1536 * We want to keep the following priority order:
1537 * cpu pinned (that don't need to move), task pinned,
1538 * cpu flexible, task flexible.
1540 cpu_ctx_sched_out(cpuctx
, EVENT_FLEXIBLE
);
1542 ctx_sched_in(ctx
, cpuctx
, EVENT_PINNED
);
1543 cpu_ctx_sched_in(cpuctx
, EVENT_FLEXIBLE
);
1544 ctx_sched_in(ctx
, cpuctx
, EVENT_FLEXIBLE
);
1546 cpuctx
->task_ctx
= ctx
;
1549 * Since these rotations are per-cpu, we need to ensure the
1550 * cpu-context we got scheduled on is actually rotating.
1552 perf_pmu_rotate_start(ctx
->pmu
);
1553 perf_pmu_enable(ctx
->pmu
);
1557 * Called from scheduler to add the events of the current task
1558 * with interrupts disabled.
1560 * We restore the event value and then enable it.
1562 * This does not protect us against NMI, but enable()
1563 * sets the enabled bit in the control field of event _before_
1564 * accessing the event control register. If a NMI hits, then it will
1565 * keep the event running.
1567 void __perf_event_task_sched_in(struct task_struct
*task
)
1569 struct perf_event_context
*ctx
;
1572 for_each_task_context_nr(ctxn
) {
1573 ctx
= task
->perf_event_ctxp
[ctxn
];
1577 perf_event_context_sched_in(ctx
);
1581 #define MAX_INTERRUPTS (~0ULL)
1583 static void perf_log_throttle(struct perf_event
*event
, int enable
);
1585 static u64
perf_calculate_period(struct perf_event
*event
, u64 nsec
, u64 count
)
1587 u64 frequency
= event
->attr
.sample_freq
;
1588 u64 sec
= NSEC_PER_SEC
;
1589 u64 divisor
, dividend
;
1591 int count_fls
, nsec_fls
, frequency_fls
, sec_fls
;
1593 count_fls
= fls64(count
);
1594 nsec_fls
= fls64(nsec
);
1595 frequency_fls
= fls64(frequency
);
1599 * We got @count in @nsec, with a target of sample_freq HZ
1600 * the target period becomes:
1603 * period = -------------------
1604 * @nsec * sample_freq
1609 * Reduce accuracy by one bit such that @a and @b converge
1610 * to a similar magnitude.
1612 #define REDUCE_FLS(a, b) \
1614 if (a##_fls > b##_fls) { \
1624 * Reduce accuracy until either term fits in a u64, then proceed with
1625 * the other, so that finally we can do a u64/u64 division.
1627 while (count_fls
+ sec_fls
> 64 && nsec_fls
+ frequency_fls
> 64) {
1628 REDUCE_FLS(nsec
, frequency
);
1629 REDUCE_FLS(sec
, count
);
1632 if (count_fls
+ sec_fls
> 64) {
1633 divisor
= nsec
* frequency
;
1635 while (count_fls
+ sec_fls
> 64) {
1636 REDUCE_FLS(count
, sec
);
1640 dividend
= count
* sec
;
1642 dividend
= count
* sec
;
1644 while (nsec_fls
+ frequency_fls
> 64) {
1645 REDUCE_FLS(nsec
, frequency
);
1649 divisor
= nsec
* frequency
;
1655 return div64_u64(dividend
, divisor
);
1658 static void perf_adjust_period(struct perf_event
*event
, u64 nsec
, u64 count
)
1660 struct hw_perf_event
*hwc
= &event
->hw
;
1661 s64 period
, sample_period
;
1664 period
= perf_calculate_period(event
, nsec
, count
);
1666 delta
= (s64
)(period
- hwc
->sample_period
);
1667 delta
= (delta
+ 7) / 8; /* low pass filter */
1669 sample_period
= hwc
->sample_period
+ delta
;
1674 hwc
->sample_period
= sample_period
;
1676 if (local64_read(&hwc
->period_left
) > 8*sample_period
) {
1677 event
->pmu
->stop(event
, PERF_EF_UPDATE
);
1678 local64_set(&hwc
->period_left
, 0);
1679 event
->pmu
->start(event
, PERF_EF_RELOAD
);
1683 static void perf_ctx_adjust_freq(struct perf_event_context
*ctx
, u64 period
)
1685 struct perf_event
*event
;
1686 struct hw_perf_event
*hwc
;
1687 u64 interrupts
, now
;
1690 raw_spin_lock(&ctx
->lock
);
1691 list_for_each_entry_rcu(event
, &ctx
->event_list
, event_entry
) {
1692 if (event
->state
!= PERF_EVENT_STATE_ACTIVE
)
1695 if (event
->cpu
!= -1 && event
->cpu
!= smp_processor_id())
1700 interrupts
= hwc
->interrupts
;
1701 hwc
->interrupts
= 0;
1704 * unthrottle events on the tick
1706 if (interrupts
== MAX_INTERRUPTS
) {
1707 perf_log_throttle(event
, 1);
1708 event
->pmu
->start(event
, 0);
1711 if (!event
->attr
.freq
|| !event
->attr
.sample_freq
)
1714 event
->pmu
->read(event
);
1715 now
= local64_read(&event
->count
);
1716 delta
= now
- hwc
->freq_count_stamp
;
1717 hwc
->freq_count_stamp
= now
;
1720 perf_adjust_period(event
, period
, delta
);
1722 raw_spin_unlock(&ctx
->lock
);
1726 * Round-robin a context's events:
1728 static void rotate_ctx(struct perf_event_context
*ctx
)
1730 raw_spin_lock(&ctx
->lock
);
1733 * Rotate the first entry last of non-pinned groups. Rotation might be
1734 * disabled by the inheritance code.
1736 if (!ctx
->rotate_disable
)
1737 list_rotate_left(&ctx
->flexible_groups
);
1739 raw_spin_unlock(&ctx
->lock
);
1743 * perf_pmu_rotate_start() and perf_rotate_context() are fully serialized
1744 * because they're strictly cpu affine and rotate_start is called with IRQs
1745 * disabled, while rotate_context is called from IRQ context.
1747 static void perf_rotate_context(struct perf_cpu_context
*cpuctx
)
1749 u64 interval
= (u64
)cpuctx
->jiffies_interval
* TICK_NSEC
;
1750 struct perf_event_context
*ctx
= NULL
;
1751 int rotate
= 0, remove
= 1;
1753 if (cpuctx
->ctx
.nr_events
) {
1755 if (cpuctx
->ctx
.nr_events
!= cpuctx
->ctx
.nr_active
)
1759 ctx
= cpuctx
->task_ctx
;
1760 if (ctx
&& ctx
->nr_events
) {
1762 if (ctx
->nr_events
!= ctx
->nr_active
)
1766 perf_pmu_disable(cpuctx
->ctx
.pmu
);
1767 perf_ctx_adjust_freq(&cpuctx
->ctx
, interval
);
1769 perf_ctx_adjust_freq(ctx
, interval
);
1774 cpu_ctx_sched_out(cpuctx
, EVENT_FLEXIBLE
);
1776 task_ctx_sched_out(ctx
, EVENT_FLEXIBLE
);
1778 rotate_ctx(&cpuctx
->ctx
);
1782 cpu_ctx_sched_in(cpuctx
, EVENT_FLEXIBLE
);
1784 task_ctx_sched_in(ctx
, EVENT_FLEXIBLE
);
1788 list_del_init(&cpuctx
->rotation_list
);
1790 perf_pmu_enable(cpuctx
->ctx
.pmu
);
1793 void perf_event_task_tick(void)
1795 struct list_head
*head
= &__get_cpu_var(rotation_list
);
1796 struct perf_cpu_context
*cpuctx
, *tmp
;
1798 WARN_ON(!irqs_disabled());
1800 list_for_each_entry_safe(cpuctx
, tmp
, head
, rotation_list
) {
1801 if (cpuctx
->jiffies_interval
== 1 ||
1802 !(jiffies
% cpuctx
->jiffies_interval
))
1803 perf_rotate_context(cpuctx
);
1807 static int event_enable_on_exec(struct perf_event
*event
,
1808 struct perf_event_context
*ctx
)
1810 if (!event
->attr
.enable_on_exec
)
1813 event
->attr
.enable_on_exec
= 0;
1814 if (event
->state
>= PERF_EVENT_STATE_INACTIVE
)
1817 __perf_event_mark_enabled(event
, ctx
);
1823 * Enable all of a task's events that have been marked enable-on-exec.
1824 * This expects task == current.
1826 static void perf_event_enable_on_exec(struct perf_event_context
*ctx
)
1828 struct perf_event
*event
;
1829 unsigned long flags
;
1833 local_irq_save(flags
);
1834 if (!ctx
|| !ctx
->nr_events
)
1837 task_ctx_sched_out(ctx
, EVENT_ALL
);
1839 raw_spin_lock(&ctx
->lock
);
1841 list_for_each_entry(event
, &ctx
->pinned_groups
, group_entry
) {
1842 ret
= event_enable_on_exec(event
, ctx
);
1847 list_for_each_entry(event
, &ctx
->flexible_groups
, group_entry
) {
1848 ret
= event_enable_on_exec(event
, ctx
);
1854 * Unclone this context if we enabled any event.
1859 raw_spin_unlock(&ctx
->lock
);
1861 perf_event_context_sched_in(ctx
);
1863 local_irq_restore(flags
);
1867 * Cross CPU call to read the hardware event
1869 static void __perf_event_read(void *info
)
1871 struct perf_event
*event
= info
;
1872 struct perf_event_context
*ctx
= event
->ctx
;
1873 struct perf_cpu_context
*cpuctx
= __get_cpu_context(ctx
);
1876 * If this is a task context, we need to check whether it is
1877 * the current task context of this cpu. If not it has been
1878 * scheduled out before the smp call arrived. In that case
1879 * event->count would have been updated to a recent sample
1880 * when the event was scheduled out.
1882 if (ctx
->task
&& cpuctx
->task_ctx
!= ctx
)
1885 raw_spin_lock(&ctx
->lock
);
1886 update_context_time(ctx
);
1887 update_event_times(event
);
1888 raw_spin_unlock(&ctx
->lock
);
1890 event
->pmu
->read(event
);
1893 static inline u64
perf_event_count(struct perf_event
*event
)
1895 return local64_read(&event
->count
) + atomic64_read(&event
->child_count
);
1898 static u64
perf_event_read(struct perf_event
*event
)
1901 * If event is enabled and currently active on a CPU, update the
1902 * value in the event structure:
1904 if (event
->state
== PERF_EVENT_STATE_ACTIVE
) {
1905 smp_call_function_single(event
->oncpu
,
1906 __perf_event_read
, event
, 1);
1907 } else if (event
->state
== PERF_EVENT_STATE_INACTIVE
) {
1908 struct perf_event_context
*ctx
= event
->ctx
;
1909 unsigned long flags
;
1911 raw_spin_lock_irqsave(&ctx
->lock
, flags
);
1913 * may read while context is not active
1914 * (e.g., thread is blocked), in that case
1915 * we cannot update context time
1918 update_context_time(ctx
);
1919 update_event_times(event
);
1920 raw_spin_unlock_irqrestore(&ctx
->lock
, flags
);
1923 return perf_event_count(event
);
1930 struct callchain_cpus_entries
{
1931 struct rcu_head rcu_head
;
1932 struct perf_callchain_entry
*cpu_entries
[0];
1935 static DEFINE_PER_CPU(int, callchain_recursion
[PERF_NR_CONTEXTS
]);
1936 static atomic_t nr_callchain_events
;
1937 static DEFINE_MUTEX(callchain_mutex
);
1938 struct callchain_cpus_entries
*callchain_cpus_entries
;
1941 __weak
void perf_callchain_kernel(struct perf_callchain_entry
*entry
,
1942 struct pt_regs
*regs
)
1946 __weak
void perf_callchain_user(struct perf_callchain_entry
*entry
,
1947 struct pt_regs
*regs
)
1951 static void release_callchain_buffers_rcu(struct rcu_head
*head
)
1953 struct callchain_cpus_entries
*entries
;
1956 entries
= container_of(head
, struct callchain_cpus_entries
, rcu_head
);
1958 for_each_possible_cpu(cpu
)
1959 kfree(entries
->cpu_entries
[cpu
]);
1964 static void release_callchain_buffers(void)
1966 struct callchain_cpus_entries
*entries
;
1968 entries
= callchain_cpus_entries
;
1969 rcu_assign_pointer(callchain_cpus_entries
, NULL
);
1970 call_rcu(&entries
->rcu_head
, release_callchain_buffers_rcu
);
1973 static int alloc_callchain_buffers(void)
1977 struct callchain_cpus_entries
*entries
;
1980 * We can't use the percpu allocation API for data that can be
1981 * accessed from NMI. Use a temporary manual per cpu allocation
1982 * until that gets sorted out.
1984 size
= sizeof(*entries
) + sizeof(struct perf_callchain_entry
*) *
1985 num_possible_cpus();
1987 entries
= kzalloc(size
, GFP_KERNEL
);
1991 size
= sizeof(struct perf_callchain_entry
) * PERF_NR_CONTEXTS
;
1993 for_each_possible_cpu(cpu
) {
1994 entries
->cpu_entries
[cpu
] = kmalloc_node(size
, GFP_KERNEL
,
1996 if (!entries
->cpu_entries
[cpu
])
2000 rcu_assign_pointer(callchain_cpus_entries
, entries
);
2005 for_each_possible_cpu(cpu
)
2006 kfree(entries
->cpu_entries
[cpu
]);
2012 static int get_callchain_buffers(void)
2017 mutex_lock(&callchain_mutex
);
2019 count
= atomic_inc_return(&nr_callchain_events
);
2020 if (WARN_ON_ONCE(count
< 1)) {
2026 /* If the allocation failed, give up */
2027 if (!callchain_cpus_entries
)
2032 err
= alloc_callchain_buffers();
2034 release_callchain_buffers();
2036 mutex_unlock(&callchain_mutex
);
2041 static void put_callchain_buffers(void)
2043 if (atomic_dec_and_mutex_lock(&nr_callchain_events
, &callchain_mutex
)) {
2044 release_callchain_buffers();
2045 mutex_unlock(&callchain_mutex
);
2049 static int get_recursion_context(int *recursion
)
2057 else if (in_softirq())
2062 if (recursion
[rctx
])
2071 static inline void put_recursion_context(int *recursion
, int rctx
)
2077 static struct perf_callchain_entry
*get_callchain_entry(int *rctx
)
2080 struct callchain_cpus_entries
*entries
;
2082 *rctx
= get_recursion_context(__get_cpu_var(callchain_recursion
));
2086 entries
= rcu_dereference(callchain_cpus_entries
);
2090 cpu
= smp_processor_id();
2092 return &entries
->cpu_entries
[cpu
][*rctx
];
2096 put_callchain_entry(int rctx
)
2098 put_recursion_context(__get_cpu_var(callchain_recursion
), rctx
);
2101 static struct perf_callchain_entry
*perf_callchain(struct pt_regs
*regs
)
2104 struct perf_callchain_entry
*entry
;
2107 entry
= get_callchain_entry(&rctx
);
2116 if (!user_mode(regs
)) {
2117 perf_callchain_store(entry
, PERF_CONTEXT_KERNEL
);
2118 perf_callchain_kernel(entry
, regs
);
2120 regs
= task_pt_regs(current
);
2126 perf_callchain_store(entry
, PERF_CONTEXT_USER
);
2127 perf_callchain_user(entry
, regs
);
2131 put_callchain_entry(rctx
);
2137 * Initialize the perf_event context in a task_struct:
2139 static void __perf_event_init_context(struct perf_event_context
*ctx
)
2141 raw_spin_lock_init(&ctx
->lock
);
2142 mutex_init(&ctx
->mutex
);
2143 INIT_LIST_HEAD(&ctx
->pinned_groups
);
2144 INIT_LIST_HEAD(&ctx
->flexible_groups
);
2145 INIT_LIST_HEAD(&ctx
->event_list
);
2146 atomic_set(&ctx
->refcount
, 1);
2149 static struct perf_event_context
*
2150 alloc_perf_context(struct pmu
*pmu
, struct task_struct
*task
)
2152 struct perf_event_context
*ctx
;
2154 ctx
= kzalloc(sizeof(struct perf_event_context
), GFP_KERNEL
);
2158 __perf_event_init_context(ctx
);
2161 get_task_struct(task
);
2168 static struct task_struct
*
2169 find_lively_task_by_vpid(pid_t vpid
)
2171 struct task_struct
*task
;
2178 task
= find_task_by_vpid(vpid
);
2180 get_task_struct(task
);
2184 return ERR_PTR(-ESRCH
);
2187 * Can't attach events to a dying task.
2190 if (task
->flags
& PF_EXITING
)
2193 /* Reuse ptrace permission checks for now. */
2195 if (!ptrace_may_access(task
, PTRACE_MODE_READ
))
2200 put_task_struct(task
);
2201 return ERR_PTR(err
);
2205 static struct perf_event_context
*
2206 find_get_context(struct pmu
*pmu
, struct task_struct
*task
, int cpu
)
2208 struct perf_event_context
*ctx
;
2209 struct perf_cpu_context
*cpuctx
;
2210 unsigned long flags
;
2213 if (!task
&& cpu
!= -1) {
2214 /* Must be root to operate on a CPU event: */
2215 if (perf_paranoid_cpu() && !capable(CAP_SYS_ADMIN
))
2216 return ERR_PTR(-EACCES
);
2218 if (cpu
< 0 || cpu
>= nr_cpumask_bits
)
2219 return ERR_PTR(-EINVAL
);
2222 * We could be clever and allow to attach a event to an
2223 * offline CPU and activate it when the CPU comes up, but
2226 if (!cpu_online(cpu
))
2227 return ERR_PTR(-ENODEV
);
2229 cpuctx
= per_cpu_ptr(pmu
->pmu_cpu_context
, cpu
);
2237 ctxn
= pmu
->task_ctx_nr
;
2242 ctx
= perf_lock_task_context(task
, ctxn
, &flags
);
2245 raw_spin_unlock_irqrestore(&ctx
->lock
, flags
);
2249 ctx
= alloc_perf_context(pmu
, task
);
2256 if (cmpxchg(&task
->perf_event_ctxp
[ctxn
], NULL
, ctx
)) {
2258 * We raced with some other task; use
2259 * the context they set.
2261 put_task_struct(task
);
2270 return ERR_PTR(err
);
2273 static void perf_event_free_filter(struct perf_event
*event
);
2275 static void free_event_rcu(struct rcu_head
*head
)
2277 struct perf_event
*event
;
2279 event
= container_of(head
, struct perf_event
, rcu_head
);
2281 put_pid_ns(event
->ns
);
2282 perf_event_free_filter(event
);
2286 static void perf_buffer_put(struct perf_buffer
*buffer
);
2288 static void free_event(struct perf_event
*event
)
2290 irq_work_sync(&event
->pending
);
2292 if (!event
->parent
) {
2293 if (event
->attach_state
& PERF_ATTACH_TASK
)
2294 jump_label_dec(&perf_task_events
);
2295 if (event
->attr
.mmap
|| event
->attr
.mmap_data
)
2296 atomic_dec(&nr_mmap_events
);
2297 if (event
->attr
.comm
)
2298 atomic_dec(&nr_comm_events
);
2299 if (event
->attr
.task
)
2300 atomic_dec(&nr_task_events
);
2301 if (event
->attr
.sample_type
& PERF_SAMPLE_CALLCHAIN
)
2302 put_callchain_buffers();
2305 if (event
->buffer
) {
2306 perf_buffer_put(event
->buffer
);
2307 event
->buffer
= NULL
;
2311 event
->destroy(event
);
2314 put_ctx(event
->ctx
);
2316 call_rcu(&event
->rcu_head
, free_event_rcu
);
2319 int perf_event_release_kernel(struct perf_event
*event
)
2321 struct perf_event_context
*ctx
= event
->ctx
;
2324 * Remove from the PMU, can't get re-enabled since we got
2325 * here because the last ref went.
2327 perf_event_disable(event
);
2329 WARN_ON_ONCE(ctx
->parent_ctx
);
2331 * There are two ways this annotation is useful:
2333 * 1) there is a lock recursion from perf_event_exit_task
2334 * see the comment there.
2336 * 2) there is a lock-inversion with mmap_sem through
2337 * perf_event_read_group(), which takes faults while
2338 * holding ctx->mutex, however this is called after
2339 * the last filedesc died, so there is no possibility
2340 * to trigger the AB-BA case.
2342 mutex_lock_nested(&ctx
->mutex
, SINGLE_DEPTH_NESTING
);
2343 raw_spin_lock_irq(&ctx
->lock
);
2344 perf_group_detach(event
);
2345 list_del_event(event
, ctx
);
2346 raw_spin_unlock_irq(&ctx
->lock
);
2347 mutex_unlock(&ctx
->mutex
);
2353 EXPORT_SYMBOL_GPL(perf_event_release_kernel
);
2356 * Called when the last reference to the file is gone.
2358 static int perf_release(struct inode
*inode
, struct file
*file
)
2360 struct perf_event
*event
= file
->private_data
;
2361 struct task_struct
*owner
;
2363 file
->private_data
= NULL
;
2366 owner
= ACCESS_ONCE(event
->owner
);
2368 * Matches the smp_wmb() in perf_event_exit_task(). If we observe
2369 * !owner it means the list deletion is complete and we can indeed
2370 * free this event, otherwise we need to serialize on
2371 * owner->perf_event_mutex.
2373 smp_read_barrier_depends();
2376 * Since delayed_put_task_struct() also drops the last
2377 * task reference we can safely take a new reference
2378 * while holding the rcu_read_lock().
2380 get_task_struct(owner
);
2385 mutex_lock(&owner
->perf_event_mutex
);
2387 * We have to re-check the event->owner field, if it is cleared
2388 * we raced with perf_event_exit_task(), acquiring the mutex
2389 * ensured they're done, and we can proceed with freeing the
2393 list_del_init(&event
->owner_entry
);
2394 mutex_unlock(&owner
->perf_event_mutex
);
2395 put_task_struct(owner
);
2398 return perf_event_release_kernel(event
);
2401 u64
perf_event_read_value(struct perf_event
*event
, u64
*enabled
, u64
*running
)
2403 struct perf_event
*child
;
2409 mutex_lock(&event
->child_mutex
);
2410 total
+= perf_event_read(event
);
2411 *enabled
+= event
->total_time_enabled
+
2412 atomic64_read(&event
->child_total_time_enabled
);
2413 *running
+= event
->total_time_running
+
2414 atomic64_read(&event
->child_total_time_running
);
2416 list_for_each_entry(child
, &event
->child_list
, child_list
) {
2417 total
+= perf_event_read(child
);
2418 *enabled
+= child
->total_time_enabled
;
2419 *running
+= child
->total_time_running
;
2421 mutex_unlock(&event
->child_mutex
);
2425 EXPORT_SYMBOL_GPL(perf_event_read_value
);
2427 static int perf_event_read_group(struct perf_event
*event
,
2428 u64 read_format
, char __user
*buf
)
2430 struct perf_event
*leader
= event
->group_leader
, *sub
;
2431 int n
= 0, size
= 0, ret
= -EFAULT
;
2432 struct perf_event_context
*ctx
= leader
->ctx
;
2434 u64 count
, enabled
, running
;
2436 mutex_lock(&ctx
->mutex
);
2437 count
= perf_event_read_value(leader
, &enabled
, &running
);
2439 values
[n
++] = 1 + leader
->nr_siblings
;
2440 if (read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
)
2441 values
[n
++] = enabled
;
2442 if (read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
)
2443 values
[n
++] = running
;
2444 values
[n
++] = count
;
2445 if (read_format
& PERF_FORMAT_ID
)
2446 values
[n
++] = primary_event_id(leader
);
2448 size
= n
* sizeof(u64
);
2450 if (copy_to_user(buf
, values
, size
))
2455 list_for_each_entry(sub
, &leader
->sibling_list
, group_entry
) {
2458 values
[n
++] = perf_event_read_value(sub
, &enabled
, &running
);
2459 if (read_format
& PERF_FORMAT_ID
)
2460 values
[n
++] = primary_event_id(sub
);
2462 size
= n
* sizeof(u64
);
2464 if (copy_to_user(buf
+ ret
, values
, size
)) {
2472 mutex_unlock(&ctx
->mutex
);
2477 static int perf_event_read_one(struct perf_event
*event
,
2478 u64 read_format
, char __user
*buf
)
2480 u64 enabled
, running
;
2484 values
[n
++] = perf_event_read_value(event
, &enabled
, &running
);
2485 if (read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
)
2486 values
[n
++] = enabled
;
2487 if (read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
)
2488 values
[n
++] = running
;
2489 if (read_format
& PERF_FORMAT_ID
)
2490 values
[n
++] = primary_event_id(event
);
2492 if (copy_to_user(buf
, values
, n
* sizeof(u64
)))
2495 return n
* sizeof(u64
);
2499 * Read the performance event - simple non blocking version for now
2502 perf_read_hw(struct perf_event
*event
, char __user
*buf
, size_t count
)
2504 u64 read_format
= event
->attr
.read_format
;
2508 * Return end-of-file for a read on a event that is in
2509 * error state (i.e. because it was pinned but it couldn't be
2510 * scheduled on to the CPU at some point).
2512 if (event
->state
== PERF_EVENT_STATE_ERROR
)
2515 if (count
< event
->read_size
)
2518 WARN_ON_ONCE(event
->ctx
->parent_ctx
);
2519 if (read_format
& PERF_FORMAT_GROUP
)
2520 ret
= perf_event_read_group(event
, read_format
, buf
);
2522 ret
= perf_event_read_one(event
, read_format
, buf
);
2528 perf_read(struct file
*file
, char __user
*buf
, size_t count
, loff_t
*ppos
)
2530 struct perf_event
*event
= file
->private_data
;
2532 return perf_read_hw(event
, buf
, count
);
2535 static unsigned int perf_poll(struct file
*file
, poll_table
*wait
)
2537 struct perf_event
*event
= file
->private_data
;
2538 struct perf_buffer
*buffer
;
2539 unsigned int events
= POLL_HUP
;
2542 buffer
= rcu_dereference(event
->buffer
);
2544 events
= atomic_xchg(&buffer
->poll
, 0);
2547 poll_wait(file
, &event
->waitq
, wait
);
2552 static void perf_event_reset(struct perf_event
*event
)
2554 (void)perf_event_read(event
);
2555 local64_set(&event
->count
, 0);
2556 perf_event_update_userpage(event
);
2560 * Holding the top-level event's child_mutex means that any
2561 * descendant process that has inherited this event will block
2562 * in sync_child_event if it goes to exit, thus satisfying the
2563 * task existence requirements of perf_event_enable/disable.
2565 static void perf_event_for_each_child(struct perf_event
*event
,
2566 void (*func
)(struct perf_event
*))
2568 struct perf_event
*child
;
2570 WARN_ON_ONCE(event
->ctx
->parent_ctx
);
2571 mutex_lock(&event
->child_mutex
);
2573 list_for_each_entry(child
, &event
->child_list
, child_list
)
2575 mutex_unlock(&event
->child_mutex
);
2578 static void perf_event_for_each(struct perf_event
*event
,
2579 void (*func
)(struct perf_event
*))
2581 struct perf_event_context
*ctx
= event
->ctx
;
2582 struct perf_event
*sibling
;
2584 WARN_ON_ONCE(ctx
->parent_ctx
);
2585 mutex_lock(&ctx
->mutex
);
2586 event
= event
->group_leader
;
2588 perf_event_for_each_child(event
, func
);
2590 list_for_each_entry(sibling
, &event
->sibling_list
, group_entry
)
2591 perf_event_for_each_child(event
, func
);
2592 mutex_unlock(&ctx
->mutex
);
2595 static int perf_event_period(struct perf_event
*event
, u64 __user
*arg
)
2597 struct perf_event_context
*ctx
= event
->ctx
;
2601 if (!is_sampling_event(event
))
2604 if (copy_from_user(&value
, arg
, sizeof(value
)))
2610 raw_spin_lock_irq(&ctx
->lock
);
2611 if (event
->attr
.freq
) {
2612 if (value
> sysctl_perf_event_sample_rate
) {
2617 event
->attr
.sample_freq
= value
;
2619 event
->attr
.sample_period
= value
;
2620 event
->hw
.sample_period
= value
;
2623 raw_spin_unlock_irq(&ctx
->lock
);
2628 static const struct file_operations perf_fops
;
2630 static struct perf_event
*perf_fget_light(int fd
, int *fput_needed
)
2634 file
= fget_light(fd
, fput_needed
);
2636 return ERR_PTR(-EBADF
);
2638 if (file
->f_op
!= &perf_fops
) {
2639 fput_light(file
, *fput_needed
);
2641 return ERR_PTR(-EBADF
);
2644 return file
->private_data
;
2647 static int perf_event_set_output(struct perf_event
*event
,
2648 struct perf_event
*output_event
);
2649 static int perf_event_set_filter(struct perf_event
*event
, void __user
*arg
);
2651 static long perf_ioctl(struct file
*file
, unsigned int cmd
, unsigned long arg
)
2653 struct perf_event
*event
= file
->private_data
;
2654 void (*func
)(struct perf_event
*);
2658 case PERF_EVENT_IOC_ENABLE
:
2659 func
= perf_event_enable
;
2661 case PERF_EVENT_IOC_DISABLE
:
2662 func
= perf_event_disable
;
2664 case PERF_EVENT_IOC_RESET
:
2665 func
= perf_event_reset
;
2668 case PERF_EVENT_IOC_REFRESH
:
2669 return perf_event_refresh(event
, arg
);
2671 case PERF_EVENT_IOC_PERIOD
:
2672 return perf_event_period(event
, (u64 __user
*)arg
);
2674 case PERF_EVENT_IOC_SET_OUTPUT
:
2676 struct perf_event
*output_event
= NULL
;
2677 int fput_needed
= 0;
2681 output_event
= perf_fget_light(arg
, &fput_needed
);
2682 if (IS_ERR(output_event
))
2683 return PTR_ERR(output_event
);
2686 ret
= perf_event_set_output(event
, output_event
);
2688 fput_light(output_event
->filp
, fput_needed
);
2693 case PERF_EVENT_IOC_SET_FILTER
:
2694 return perf_event_set_filter(event
, (void __user
*)arg
);
2700 if (flags
& PERF_IOC_FLAG_GROUP
)
2701 perf_event_for_each(event
, func
);
2703 perf_event_for_each_child(event
, func
);
2708 int perf_event_task_enable(void)
2710 struct perf_event
*event
;
2712 mutex_lock(¤t
->perf_event_mutex
);
2713 list_for_each_entry(event
, ¤t
->perf_event_list
, owner_entry
)
2714 perf_event_for_each_child(event
, perf_event_enable
);
2715 mutex_unlock(¤t
->perf_event_mutex
);
2720 int perf_event_task_disable(void)
2722 struct perf_event
*event
;
2724 mutex_lock(¤t
->perf_event_mutex
);
2725 list_for_each_entry(event
, ¤t
->perf_event_list
, owner_entry
)
2726 perf_event_for_each_child(event
, perf_event_disable
);
2727 mutex_unlock(¤t
->perf_event_mutex
);
2732 #ifndef PERF_EVENT_INDEX_OFFSET
2733 # define PERF_EVENT_INDEX_OFFSET 0
2736 static int perf_event_index(struct perf_event
*event
)
2738 if (event
->hw
.state
& PERF_HES_STOPPED
)
2741 if (event
->state
!= PERF_EVENT_STATE_ACTIVE
)
2744 return event
->hw
.idx
+ 1 - PERF_EVENT_INDEX_OFFSET
;
2748 * Callers need to ensure there can be no nesting of this function, otherwise
2749 * the seqlock logic goes bad. We can not serialize this because the arch
2750 * code calls this from NMI context.
2752 void perf_event_update_userpage(struct perf_event
*event
)
2754 struct perf_event_mmap_page
*userpg
;
2755 struct perf_buffer
*buffer
;
2758 buffer
= rcu_dereference(event
->buffer
);
2762 userpg
= buffer
->user_page
;
2765 * Disable preemption so as to not let the corresponding user-space
2766 * spin too long if we get preempted.
2771 userpg
->index
= perf_event_index(event
);
2772 userpg
->offset
= perf_event_count(event
);
2773 if (event
->state
== PERF_EVENT_STATE_ACTIVE
)
2774 userpg
->offset
-= local64_read(&event
->hw
.prev_count
);
2776 userpg
->time_enabled
= event
->total_time_enabled
+
2777 atomic64_read(&event
->child_total_time_enabled
);
2779 userpg
->time_running
= event
->total_time_running
+
2780 atomic64_read(&event
->child_total_time_running
);
2789 static unsigned long perf_data_size(struct perf_buffer
*buffer
);
2792 perf_buffer_init(struct perf_buffer
*buffer
, long watermark
, int flags
)
2794 long max_size
= perf_data_size(buffer
);
2797 buffer
->watermark
= min(max_size
, watermark
);
2799 if (!buffer
->watermark
)
2800 buffer
->watermark
= max_size
/ 2;
2802 if (flags
& PERF_BUFFER_WRITABLE
)
2803 buffer
->writable
= 1;
2805 atomic_set(&buffer
->refcount
, 1);
2808 #ifndef CONFIG_PERF_USE_VMALLOC
2811 * Back perf_mmap() with regular GFP_KERNEL-0 pages.
2814 static struct page
*
2815 perf_mmap_to_page(struct perf_buffer
*buffer
, unsigned long pgoff
)
2817 if (pgoff
> buffer
->nr_pages
)
2821 return virt_to_page(buffer
->user_page
);
2823 return virt_to_page(buffer
->data_pages
[pgoff
- 1]);
2826 static void *perf_mmap_alloc_page(int cpu
)
2831 node
= (cpu
== -1) ? cpu
: cpu_to_node(cpu
);
2832 page
= alloc_pages_node(node
, GFP_KERNEL
| __GFP_ZERO
, 0);
2836 return page_address(page
);
2839 static struct perf_buffer
*
2840 perf_buffer_alloc(int nr_pages
, long watermark
, int cpu
, int flags
)
2842 struct perf_buffer
*buffer
;
2846 size
= sizeof(struct perf_buffer
);
2847 size
+= nr_pages
* sizeof(void *);
2849 buffer
= kzalloc(size
, GFP_KERNEL
);
2853 buffer
->user_page
= perf_mmap_alloc_page(cpu
);
2854 if (!buffer
->user_page
)
2855 goto fail_user_page
;
2857 for (i
= 0; i
< nr_pages
; i
++) {
2858 buffer
->data_pages
[i
] = perf_mmap_alloc_page(cpu
);
2859 if (!buffer
->data_pages
[i
])
2860 goto fail_data_pages
;
2863 buffer
->nr_pages
= nr_pages
;
2865 perf_buffer_init(buffer
, watermark
, flags
);
2870 for (i
--; i
>= 0; i
--)
2871 free_page((unsigned long)buffer
->data_pages
[i
]);
2873 free_page((unsigned long)buffer
->user_page
);
2882 static void perf_mmap_free_page(unsigned long addr
)
2884 struct page
*page
= virt_to_page((void *)addr
);
2886 page
->mapping
= NULL
;
2890 static void perf_buffer_free(struct perf_buffer
*buffer
)
2894 perf_mmap_free_page((unsigned long)buffer
->user_page
);
2895 for (i
= 0; i
< buffer
->nr_pages
; i
++)
2896 perf_mmap_free_page((unsigned long)buffer
->data_pages
[i
]);
2900 static inline int page_order(struct perf_buffer
*buffer
)
2908 * Back perf_mmap() with vmalloc memory.
2910 * Required for architectures that have d-cache aliasing issues.
2913 static inline int page_order(struct perf_buffer
*buffer
)
2915 return buffer
->page_order
;
2918 static struct page
*
2919 perf_mmap_to_page(struct perf_buffer
*buffer
, unsigned long pgoff
)
2921 if (pgoff
> (1UL << page_order(buffer
)))
2924 return vmalloc_to_page((void *)buffer
->user_page
+ pgoff
* PAGE_SIZE
);
2927 static void perf_mmap_unmark_page(void *addr
)
2929 struct page
*page
= vmalloc_to_page(addr
);
2931 page
->mapping
= NULL
;
2934 static void perf_buffer_free_work(struct work_struct
*work
)
2936 struct perf_buffer
*buffer
;
2940 buffer
= container_of(work
, struct perf_buffer
, work
);
2941 nr
= 1 << page_order(buffer
);
2943 base
= buffer
->user_page
;
2944 for (i
= 0; i
< nr
+ 1; i
++)
2945 perf_mmap_unmark_page(base
+ (i
* PAGE_SIZE
));
2951 static void perf_buffer_free(struct perf_buffer
*buffer
)
2953 schedule_work(&buffer
->work
);
2956 static struct perf_buffer
*
2957 perf_buffer_alloc(int nr_pages
, long watermark
, int cpu
, int flags
)
2959 struct perf_buffer
*buffer
;
2963 size
= sizeof(struct perf_buffer
);
2964 size
+= sizeof(void *);
2966 buffer
= kzalloc(size
, GFP_KERNEL
);
2970 INIT_WORK(&buffer
->work
, perf_buffer_free_work
);
2972 all_buf
= vmalloc_user((nr_pages
+ 1) * PAGE_SIZE
);
2976 buffer
->user_page
= all_buf
;
2977 buffer
->data_pages
[0] = all_buf
+ PAGE_SIZE
;
2978 buffer
->page_order
= ilog2(nr_pages
);
2979 buffer
->nr_pages
= 1;
2981 perf_buffer_init(buffer
, watermark
, flags
);
2994 static unsigned long perf_data_size(struct perf_buffer
*buffer
)
2996 return buffer
->nr_pages
<< (PAGE_SHIFT
+ page_order(buffer
));
2999 static int perf_mmap_fault(struct vm_area_struct
*vma
, struct vm_fault
*vmf
)
3001 struct perf_event
*event
= vma
->vm_file
->private_data
;
3002 struct perf_buffer
*buffer
;
3003 int ret
= VM_FAULT_SIGBUS
;
3005 if (vmf
->flags
& FAULT_FLAG_MKWRITE
) {
3006 if (vmf
->pgoff
== 0)
3012 buffer
= rcu_dereference(event
->buffer
);
3016 if (vmf
->pgoff
&& (vmf
->flags
& FAULT_FLAG_WRITE
))
3019 vmf
->page
= perf_mmap_to_page(buffer
, vmf
->pgoff
);
3023 get_page(vmf
->page
);
3024 vmf
->page
->mapping
= vma
->vm_file
->f_mapping
;
3025 vmf
->page
->index
= vmf
->pgoff
;
3034 static void perf_buffer_free_rcu(struct rcu_head
*rcu_head
)
3036 struct perf_buffer
*buffer
;
3038 buffer
= container_of(rcu_head
, struct perf_buffer
, rcu_head
);
3039 perf_buffer_free(buffer
);
3042 static struct perf_buffer
*perf_buffer_get(struct perf_event
*event
)
3044 struct perf_buffer
*buffer
;
3047 buffer
= rcu_dereference(event
->buffer
);
3049 if (!atomic_inc_not_zero(&buffer
->refcount
))
3057 static void perf_buffer_put(struct perf_buffer
*buffer
)
3059 if (!atomic_dec_and_test(&buffer
->refcount
))
3062 call_rcu(&buffer
->rcu_head
, perf_buffer_free_rcu
);
3065 static void perf_mmap_open(struct vm_area_struct
*vma
)
3067 struct perf_event
*event
= vma
->vm_file
->private_data
;
3069 atomic_inc(&event
->mmap_count
);
3072 static void perf_mmap_close(struct vm_area_struct
*vma
)
3074 struct perf_event
*event
= vma
->vm_file
->private_data
;
3076 if (atomic_dec_and_mutex_lock(&event
->mmap_count
, &event
->mmap_mutex
)) {
3077 unsigned long size
= perf_data_size(event
->buffer
);
3078 struct user_struct
*user
= event
->mmap_user
;
3079 struct perf_buffer
*buffer
= event
->buffer
;
3081 atomic_long_sub((size
>> PAGE_SHIFT
) + 1, &user
->locked_vm
);
3082 vma
->vm_mm
->locked_vm
-= event
->mmap_locked
;
3083 rcu_assign_pointer(event
->buffer
, NULL
);
3084 mutex_unlock(&event
->mmap_mutex
);
3086 perf_buffer_put(buffer
);
3091 static const struct vm_operations_struct perf_mmap_vmops
= {
3092 .open
= perf_mmap_open
,
3093 .close
= perf_mmap_close
,
3094 .fault
= perf_mmap_fault
,
3095 .page_mkwrite
= perf_mmap_fault
,
3098 static int perf_mmap(struct file
*file
, struct vm_area_struct
*vma
)
3100 struct perf_event
*event
= file
->private_data
;
3101 unsigned long user_locked
, user_lock_limit
;
3102 struct user_struct
*user
= current_user();
3103 unsigned long locked
, lock_limit
;
3104 struct perf_buffer
*buffer
;
3105 unsigned long vma_size
;
3106 unsigned long nr_pages
;
3107 long user_extra
, extra
;
3108 int ret
= 0, flags
= 0;
3111 * Don't allow mmap() of inherited per-task counters. This would
3112 * create a performance issue due to all children writing to the
3115 if (event
->cpu
== -1 && event
->attr
.inherit
)
3118 if (!(vma
->vm_flags
& VM_SHARED
))
3121 vma_size
= vma
->vm_end
- vma
->vm_start
;
3122 nr_pages
= (vma_size
/ PAGE_SIZE
) - 1;
3125 * If we have buffer pages ensure they're a power-of-two number, so we
3126 * can do bitmasks instead of modulo.
3128 if (nr_pages
!= 0 && !is_power_of_2(nr_pages
))
3131 if (vma_size
!= PAGE_SIZE
* (1 + nr_pages
))
3134 if (vma
->vm_pgoff
!= 0)
3137 WARN_ON_ONCE(event
->ctx
->parent_ctx
);
3138 mutex_lock(&event
->mmap_mutex
);
3139 if (event
->buffer
) {
3140 if (event
->buffer
->nr_pages
== nr_pages
)
3141 atomic_inc(&event
->buffer
->refcount
);
3147 user_extra
= nr_pages
+ 1;
3148 user_lock_limit
= sysctl_perf_event_mlock
>> (PAGE_SHIFT
- 10);
3151 * Increase the limit linearly with more CPUs:
3153 user_lock_limit
*= num_online_cpus();
3155 user_locked
= atomic_long_read(&user
->locked_vm
) + user_extra
;
3158 if (user_locked
> user_lock_limit
)
3159 extra
= user_locked
- user_lock_limit
;
3161 lock_limit
= rlimit(RLIMIT_MEMLOCK
);
3162 lock_limit
>>= PAGE_SHIFT
;
3163 locked
= vma
->vm_mm
->locked_vm
+ extra
;
3165 if ((locked
> lock_limit
) && perf_paranoid_tracepoint_raw() &&
3166 !capable(CAP_IPC_LOCK
)) {
3171 WARN_ON(event
->buffer
);
3173 if (vma
->vm_flags
& VM_WRITE
)
3174 flags
|= PERF_BUFFER_WRITABLE
;
3176 buffer
= perf_buffer_alloc(nr_pages
, event
->attr
.wakeup_watermark
,
3182 rcu_assign_pointer(event
->buffer
, buffer
);
3184 atomic_long_add(user_extra
, &user
->locked_vm
);
3185 event
->mmap_locked
= extra
;
3186 event
->mmap_user
= get_current_user();
3187 vma
->vm_mm
->locked_vm
+= event
->mmap_locked
;
3191 atomic_inc(&event
->mmap_count
);
3192 mutex_unlock(&event
->mmap_mutex
);
3194 vma
->vm_flags
|= VM_RESERVED
;
3195 vma
->vm_ops
= &perf_mmap_vmops
;
3200 static int perf_fasync(int fd
, struct file
*filp
, int on
)
3202 struct inode
*inode
= filp
->f_path
.dentry
->d_inode
;
3203 struct perf_event
*event
= filp
->private_data
;
3206 mutex_lock(&inode
->i_mutex
);
3207 retval
= fasync_helper(fd
, filp
, on
, &event
->fasync
);
3208 mutex_unlock(&inode
->i_mutex
);
3216 static const struct file_operations perf_fops
= {
3217 .llseek
= no_llseek
,
3218 .release
= perf_release
,
3221 .unlocked_ioctl
= perf_ioctl
,
3222 .compat_ioctl
= perf_ioctl
,
3224 .fasync
= perf_fasync
,
3230 * If there's data, ensure we set the poll() state and publish everything
3231 * to user-space before waking everybody up.
3234 void perf_event_wakeup(struct perf_event
*event
)
3236 wake_up_all(&event
->waitq
);
3238 if (event
->pending_kill
) {
3239 kill_fasync(&event
->fasync
, SIGIO
, event
->pending_kill
);
3240 event
->pending_kill
= 0;
3244 static void perf_pending_event(struct irq_work
*entry
)
3246 struct perf_event
*event
= container_of(entry
,
3247 struct perf_event
, pending
);
3249 if (event
->pending_disable
) {
3250 event
->pending_disable
= 0;
3251 __perf_event_disable(event
);
3254 if (event
->pending_wakeup
) {
3255 event
->pending_wakeup
= 0;
3256 perf_event_wakeup(event
);
3261 * We assume there is only KVM supporting the callbacks.
3262 * Later on, we might change it to a list if there is
3263 * another virtualization implementation supporting the callbacks.
3265 struct perf_guest_info_callbacks
*perf_guest_cbs
;
3267 int perf_register_guest_info_callbacks(struct perf_guest_info_callbacks
*cbs
)
3269 perf_guest_cbs
= cbs
;
3272 EXPORT_SYMBOL_GPL(perf_register_guest_info_callbacks
);
3274 int perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks
*cbs
)
3276 perf_guest_cbs
= NULL
;
3279 EXPORT_SYMBOL_GPL(perf_unregister_guest_info_callbacks
);
3284 static bool perf_output_space(struct perf_buffer
*buffer
, unsigned long tail
,
3285 unsigned long offset
, unsigned long head
)
3289 if (!buffer
->writable
)
3292 mask
= perf_data_size(buffer
) - 1;
3294 offset
= (offset
- tail
) & mask
;
3295 head
= (head
- tail
) & mask
;
3297 if ((int)(head
- offset
) < 0)
3303 static void perf_output_wakeup(struct perf_output_handle
*handle
)
3305 atomic_set(&handle
->buffer
->poll
, POLL_IN
);
3308 handle
->event
->pending_wakeup
= 1;
3309 irq_work_queue(&handle
->event
->pending
);
3311 perf_event_wakeup(handle
->event
);
3315 * We need to ensure a later event_id doesn't publish a head when a former
3316 * event isn't done writing. However since we need to deal with NMIs we
3317 * cannot fully serialize things.
3319 * We only publish the head (and generate a wakeup) when the outer-most
3322 static void perf_output_get_handle(struct perf_output_handle
*handle
)
3324 struct perf_buffer
*buffer
= handle
->buffer
;
3327 local_inc(&buffer
->nest
);
3328 handle
->wakeup
= local_read(&buffer
->wakeup
);
3331 static void perf_output_put_handle(struct perf_output_handle
*handle
)
3333 struct perf_buffer
*buffer
= handle
->buffer
;
3337 head
= local_read(&buffer
->head
);
3340 * IRQ/NMI can happen here, which means we can miss a head update.
3343 if (!local_dec_and_test(&buffer
->nest
))
3347 * Publish the known good head. Rely on the full barrier implied
3348 * by atomic_dec_and_test() order the buffer->head read and this
3351 buffer
->user_page
->data_head
= head
;
3354 * Now check if we missed an update, rely on the (compiler)
3355 * barrier in atomic_dec_and_test() to re-read buffer->head.
3357 if (unlikely(head
!= local_read(&buffer
->head
))) {
3358 local_inc(&buffer
->nest
);
3362 if (handle
->wakeup
!= local_read(&buffer
->wakeup
))
3363 perf_output_wakeup(handle
);
3369 __always_inline
void perf_output_copy(struct perf_output_handle
*handle
,
3370 const void *buf
, unsigned int len
)
3373 unsigned long size
= min_t(unsigned long, handle
->size
, len
);
3375 memcpy(handle
->addr
, buf
, size
);
3378 handle
->addr
+= size
;
3380 handle
->size
-= size
;
3381 if (!handle
->size
) {
3382 struct perf_buffer
*buffer
= handle
->buffer
;
3385 handle
->page
&= buffer
->nr_pages
- 1;
3386 handle
->addr
= buffer
->data_pages
[handle
->page
];
3387 handle
->size
= PAGE_SIZE
<< page_order(buffer
);
3392 static void __perf_event_header__init_id(struct perf_event_header
*header
,
3393 struct perf_sample_data
*data
,
3394 struct perf_event
*event
)
3396 u64 sample_type
= event
->attr
.sample_type
;
3398 data
->type
= sample_type
;
3399 header
->size
+= event
->id_header_size
;
3401 if (sample_type
& PERF_SAMPLE_TID
) {
3402 /* namespace issues */
3403 data
->tid_entry
.pid
= perf_event_pid(event
, current
);
3404 data
->tid_entry
.tid
= perf_event_tid(event
, current
);
3407 if (sample_type
& PERF_SAMPLE_TIME
)
3408 data
->time
= perf_clock();
3410 if (sample_type
& PERF_SAMPLE_ID
)
3411 data
->id
= primary_event_id(event
);
3413 if (sample_type
& PERF_SAMPLE_STREAM_ID
)
3414 data
->stream_id
= event
->id
;
3416 if (sample_type
& PERF_SAMPLE_CPU
) {
3417 data
->cpu_entry
.cpu
= raw_smp_processor_id();
3418 data
->cpu_entry
.reserved
= 0;
3422 static void perf_event_header__init_id(struct perf_event_header
*header
,
3423 struct perf_sample_data
*data
,
3424 struct perf_event
*event
)
3426 if (event
->attr
.sample_id_all
)
3427 __perf_event_header__init_id(header
, data
, event
);
3430 static void __perf_event__output_id_sample(struct perf_output_handle
*handle
,
3431 struct perf_sample_data
*data
)
3433 u64 sample_type
= data
->type
;
3435 if (sample_type
& PERF_SAMPLE_TID
)
3436 perf_output_put(handle
, data
->tid_entry
);
3438 if (sample_type
& PERF_SAMPLE_TIME
)
3439 perf_output_put(handle
, data
->time
);
3441 if (sample_type
& PERF_SAMPLE_ID
)
3442 perf_output_put(handle
, data
->id
);
3444 if (sample_type
& PERF_SAMPLE_STREAM_ID
)
3445 perf_output_put(handle
, data
->stream_id
);
3447 if (sample_type
& PERF_SAMPLE_CPU
)
3448 perf_output_put(handle
, data
->cpu_entry
);
3451 static void perf_event__output_id_sample(struct perf_event
*event
,
3452 struct perf_output_handle
*handle
,
3453 struct perf_sample_data
*sample
)
3455 if (event
->attr
.sample_id_all
)
3456 __perf_event__output_id_sample(handle
, sample
);
3459 int perf_output_begin(struct perf_output_handle
*handle
,
3460 struct perf_event
*event
, unsigned int size
,
3461 int nmi
, int sample
)
3463 struct perf_buffer
*buffer
;
3464 unsigned long tail
, offset
, head
;
3466 struct perf_sample_data sample_data
;
3468 struct perf_event_header header
;
3475 * For inherited events we send all the output towards the parent.
3478 event
= event
->parent
;
3480 buffer
= rcu_dereference(event
->buffer
);
3484 handle
->buffer
= buffer
;
3485 handle
->event
= event
;
3487 handle
->sample
= sample
;
3489 if (!buffer
->nr_pages
)
3492 have_lost
= local_read(&buffer
->lost
);
3494 lost_event
.header
.size
= sizeof(lost_event
);
3495 perf_event_header__init_id(&lost_event
.header
, &sample_data
,
3497 size
+= lost_event
.header
.size
;
3500 perf_output_get_handle(handle
);
3504 * Userspace could choose to issue a mb() before updating the
3505 * tail pointer. So that all reads will be completed before the
3508 tail
= ACCESS_ONCE(buffer
->user_page
->data_tail
);
3510 offset
= head
= local_read(&buffer
->head
);
3512 if (unlikely(!perf_output_space(buffer
, tail
, offset
, head
)))
3514 } while (local_cmpxchg(&buffer
->head
, offset
, head
) != offset
);
3516 if (head
- local_read(&buffer
->wakeup
) > buffer
->watermark
)
3517 local_add(buffer
->watermark
, &buffer
->wakeup
);
3519 handle
->page
= offset
>> (PAGE_SHIFT
+ page_order(buffer
));
3520 handle
->page
&= buffer
->nr_pages
- 1;
3521 handle
->size
= offset
& ((PAGE_SIZE
<< page_order(buffer
)) - 1);
3522 handle
->addr
= buffer
->data_pages
[handle
->page
];
3523 handle
->addr
+= handle
->size
;
3524 handle
->size
= (PAGE_SIZE
<< page_order(buffer
)) - handle
->size
;
3527 lost_event
.header
.type
= PERF_RECORD_LOST
;
3528 lost_event
.header
.misc
= 0;
3529 lost_event
.id
= event
->id
;
3530 lost_event
.lost
= local_xchg(&buffer
->lost
, 0);
3532 perf_output_put(handle
, lost_event
);
3533 perf_event__output_id_sample(event
, handle
, &sample_data
);
3539 local_inc(&buffer
->lost
);
3540 perf_output_put_handle(handle
);
3547 void perf_output_end(struct perf_output_handle
*handle
)
3549 struct perf_event
*event
= handle
->event
;
3550 struct perf_buffer
*buffer
= handle
->buffer
;
3552 int wakeup_events
= event
->attr
.wakeup_events
;
3554 if (handle
->sample
&& wakeup_events
) {
3555 int events
= local_inc_return(&buffer
->events
);
3556 if (events
>= wakeup_events
) {
3557 local_sub(wakeup_events
, &buffer
->events
);
3558 local_inc(&buffer
->wakeup
);
3562 perf_output_put_handle(handle
);
3566 static void perf_output_read_one(struct perf_output_handle
*handle
,
3567 struct perf_event
*event
,
3568 u64 enabled
, u64 running
)
3570 u64 read_format
= event
->attr
.read_format
;
3574 values
[n
++] = perf_event_count(event
);
3575 if (read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
) {
3576 values
[n
++] = enabled
+
3577 atomic64_read(&event
->child_total_time_enabled
);
3579 if (read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
) {
3580 values
[n
++] = running
+
3581 atomic64_read(&event
->child_total_time_running
);
3583 if (read_format
& PERF_FORMAT_ID
)
3584 values
[n
++] = primary_event_id(event
);
3586 perf_output_copy(handle
, values
, n
* sizeof(u64
));
3590 * XXX PERF_FORMAT_GROUP vs inherited events seems difficult.
3592 static void perf_output_read_group(struct perf_output_handle
*handle
,
3593 struct perf_event
*event
,
3594 u64 enabled
, u64 running
)
3596 struct perf_event
*leader
= event
->group_leader
, *sub
;
3597 u64 read_format
= event
->attr
.read_format
;
3601 values
[n
++] = 1 + leader
->nr_siblings
;
3603 if (read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
)
3604 values
[n
++] = enabled
;
3606 if (read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
)
3607 values
[n
++] = running
;
3609 if (leader
!= event
)
3610 leader
->pmu
->read(leader
);
3612 values
[n
++] = perf_event_count(leader
);
3613 if (read_format
& PERF_FORMAT_ID
)
3614 values
[n
++] = primary_event_id(leader
);
3616 perf_output_copy(handle
, values
, n
* sizeof(u64
));
3618 list_for_each_entry(sub
, &leader
->sibling_list
, group_entry
) {
3622 sub
->pmu
->read(sub
);
3624 values
[n
++] = perf_event_count(sub
);
3625 if (read_format
& PERF_FORMAT_ID
)
3626 values
[n
++] = primary_event_id(sub
);
3628 perf_output_copy(handle
, values
, n
* sizeof(u64
));
3632 #define PERF_FORMAT_TOTAL_TIMES (PERF_FORMAT_TOTAL_TIME_ENABLED|\
3633 PERF_FORMAT_TOTAL_TIME_RUNNING)
3635 static void perf_output_read(struct perf_output_handle
*handle
,
3636 struct perf_event
*event
)
3638 u64 enabled
= 0, running
= 0, now
, ctx_time
;
3639 u64 read_format
= event
->attr
.read_format
;
3642 * compute total_time_enabled, total_time_running
3643 * based on snapshot values taken when the event
3644 * was last scheduled in.
3646 * we cannot simply called update_context_time()
3647 * because of locking issue as we are called in
3650 if (read_format
& PERF_FORMAT_TOTAL_TIMES
) {
3652 ctx_time
= event
->shadow_ctx_time
+ now
;
3653 enabled
= ctx_time
- event
->tstamp_enabled
;
3654 running
= ctx_time
- event
->tstamp_running
;
3657 if (event
->attr
.read_format
& PERF_FORMAT_GROUP
)
3658 perf_output_read_group(handle
, event
, enabled
, running
);
3660 perf_output_read_one(handle
, event
, enabled
, running
);
3663 void perf_output_sample(struct perf_output_handle
*handle
,
3664 struct perf_event_header
*header
,
3665 struct perf_sample_data
*data
,
3666 struct perf_event
*event
)
3668 u64 sample_type
= data
->type
;
3670 perf_output_put(handle
, *header
);
3672 if (sample_type
& PERF_SAMPLE_IP
)
3673 perf_output_put(handle
, data
->ip
);
3675 if (sample_type
& PERF_SAMPLE_TID
)
3676 perf_output_put(handle
, data
->tid_entry
);
3678 if (sample_type
& PERF_SAMPLE_TIME
)
3679 perf_output_put(handle
, data
->time
);
3681 if (sample_type
& PERF_SAMPLE_ADDR
)
3682 perf_output_put(handle
, data
->addr
);
3684 if (sample_type
& PERF_SAMPLE_ID
)
3685 perf_output_put(handle
, data
->id
);
3687 if (sample_type
& PERF_SAMPLE_STREAM_ID
)
3688 perf_output_put(handle
, data
->stream_id
);
3690 if (sample_type
& PERF_SAMPLE_CPU
)
3691 perf_output_put(handle
, data
->cpu_entry
);
3693 if (sample_type
& PERF_SAMPLE_PERIOD
)
3694 perf_output_put(handle
, data
->period
);
3696 if (sample_type
& PERF_SAMPLE_READ
)
3697 perf_output_read(handle
, event
);
3699 if (sample_type
& PERF_SAMPLE_CALLCHAIN
) {
3700 if (data
->callchain
) {
3703 if (data
->callchain
)
3704 size
+= data
->callchain
->nr
;
3706 size
*= sizeof(u64
);
3708 perf_output_copy(handle
, data
->callchain
, size
);
3711 perf_output_put(handle
, nr
);
3715 if (sample_type
& PERF_SAMPLE_RAW
) {
3717 perf_output_put(handle
, data
->raw
->size
);
3718 perf_output_copy(handle
, data
->raw
->data
,
3725 .size
= sizeof(u32
),
3728 perf_output_put(handle
, raw
);
3733 void perf_prepare_sample(struct perf_event_header
*header
,
3734 struct perf_sample_data
*data
,
3735 struct perf_event
*event
,
3736 struct pt_regs
*regs
)
3738 u64 sample_type
= event
->attr
.sample_type
;
3740 header
->type
= PERF_RECORD_SAMPLE
;
3741 header
->size
= sizeof(*header
) + event
->header_size
;
3744 header
->misc
|= perf_misc_flags(regs
);
3746 __perf_event_header__init_id(header
, data
, event
);
3748 if (sample_type
& PERF_SAMPLE_IP
)
3749 data
->ip
= perf_instruction_pointer(regs
);
3751 if (sample_type
& PERF_SAMPLE_CALLCHAIN
) {
3754 data
->callchain
= perf_callchain(regs
);
3756 if (data
->callchain
)
3757 size
+= data
->callchain
->nr
;
3759 header
->size
+= size
* sizeof(u64
);
3762 if (sample_type
& PERF_SAMPLE_RAW
) {
3763 int size
= sizeof(u32
);
3766 size
+= data
->raw
->size
;
3768 size
+= sizeof(u32
);
3770 WARN_ON_ONCE(size
& (sizeof(u64
)-1));
3771 header
->size
+= size
;
3775 static void perf_event_output(struct perf_event
*event
, int nmi
,
3776 struct perf_sample_data
*data
,
3777 struct pt_regs
*regs
)
3779 struct perf_output_handle handle
;
3780 struct perf_event_header header
;
3782 /* protect the callchain buffers */
3785 perf_prepare_sample(&header
, data
, event
, regs
);
3787 if (perf_output_begin(&handle
, event
, header
.size
, nmi
, 1))
3790 perf_output_sample(&handle
, &header
, data
, event
);
3792 perf_output_end(&handle
);
3802 struct perf_read_event
{
3803 struct perf_event_header header
;
3810 perf_event_read_event(struct perf_event
*event
,
3811 struct task_struct
*task
)
3813 struct perf_output_handle handle
;
3814 struct perf_sample_data sample
;
3815 struct perf_read_event read_event
= {
3817 .type
= PERF_RECORD_READ
,
3819 .size
= sizeof(read_event
) + event
->read_size
,
3821 .pid
= perf_event_pid(event
, task
),
3822 .tid
= perf_event_tid(event
, task
),
3826 perf_event_header__init_id(&read_event
.header
, &sample
, event
);
3827 ret
= perf_output_begin(&handle
, event
, read_event
.header
.size
, 0, 0);
3831 perf_output_put(&handle
, read_event
);
3832 perf_output_read(&handle
, event
);
3833 perf_event__output_id_sample(event
, &handle
, &sample
);
3835 perf_output_end(&handle
);
3839 * task tracking -- fork/exit
3841 * enabled by: attr.comm | attr.mmap | attr.mmap_data | attr.task
3844 struct perf_task_event
{
3845 struct task_struct
*task
;
3846 struct perf_event_context
*task_ctx
;
3849 struct perf_event_header header
;
3859 static void perf_event_task_output(struct perf_event
*event
,
3860 struct perf_task_event
*task_event
)
3862 struct perf_output_handle handle
;
3863 struct perf_sample_data sample
;
3864 struct task_struct
*task
= task_event
->task
;
3865 int ret
, size
= task_event
->event_id
.header
.size
;
3867 perf_event_header__init_id(&task_event
->event_id
.header
, &sample
, event
);
3869 ret
= perf_output_begin(&handle
, event
,
3870 task_event
->event_id
.header
.size
, 0, 0);
3874 task_event
->event_id
.pid
= perf_event_pid(event
, task
);
3875 task_event
->event_id
.ppid
= perf_event_pid(event
, current
);
3877 task_event
->event_id
.tid
= perf_event_tid(event
, task
);
3878 task_event
->event_id
.ptid
= perf_event_tid(event
, current
);
3880 perf_output_put(&handle
, task_event
->event_id
);
3882 perf_event__output_id_sample(event
, &handle
, &sample
);
3884 perf_output_end(&handle
);
3886 task_event
->event_id
.header
.size
= size
;
3889 static int perf_event_task_match(struct perf_event
*event
)
3891 if (event
->state
< PERF_EVENT_STATE_INACTIVE
)
3894 if (event
->cpu
!= -1 && event
->cpu
!= smp_processor_id())
3897 if (event
->attr
.comm
|| event
->attr
.mmap
||
3898 event
->attr
.mmap_data
|| event
->attr
.task
)
3904 static void perf_event_task_ctx(struct perf_event_context
*ctx
,
3905 struct perf_task_event
*task_event
)
3907 struct perf_event
*event
;
3909 list_for_each_entry_rcu(event
, &ctx
->event_list
, event_entry
) {
3910 if (perf_event_task_match(event
))
3911 perf_event_task_output(event
, task_event
);
3915 static void perf_event_task_event(struct perf_task_event
*task_event
)
3917 struct perf_cpu_context
*cpuctx
;
3918 struct perf_event_context
*ctx
;
3923 list_for_each_entry_rcu(pmu
, &pmus
, entry
) {
3924 cpuctx
= get_cpu_ptr(pmu
->pmu_cpu_context
);
3925 perf_event_task_ctx(&cpuctx
->ctx
, task_event
);
3927 ctx
= task_event
->task_ctx
;
3929 ctxn
= pmu
->task_ctx_nr
;
3932 ctx
= rcu_dereference(current
->perf_event_ctxp
[ctxn
]);
3935 perf_event_task_ctx(ctx
, task_event
);
3937 put_cpu_ptr(pmu
->pmu_cpu_context
);
3942 static void perf_event_task(struct task_struct
*task
,
3943 struct perf_event_context
*task_ctx
,
3946 struct perf_task_event task_event
;
3948 if (!atomic_read(&nr_comm_events
) &&
3949 !atomic_read(&nr_mmap_events
) &&
3950 !atomic_read(&nr_task_events
))
3953 task_event
= (struct perf_task_event
){
3955 .task_ctx
= task_ctx
,
3958 .type
= new ? PERF_RECORD_FORK
: PERF_RECORD_EXIT
,
3960 .size
= sizeof(task_event
.event_id
),
3966 .time
= perf_clock(),
3970 perf_event_task_event(&task_event
);
3973 void perf_event_fork(struct task_struct
*task
)
3975 perf_event_task(task
, NULL
, 1);
3982 struct perf_comm_event
{
3983 struct task_struct
*task
;
3988 struct perf_event_header header
;
3995 static void perf_event_comm_output(struct perf_event
*event
,
3996 struct perf_comm_event
*comm_event
)
3998 struct perf_output_handle handle
;
3999 struct perf_sample_data sample
;
4000 int size
= comm_event
->event_id
.header
.size
;
4003 perf_event_header__init_id(&comm_event
->event_id
.header
, &sample
, event
);
4004 ret
= perf_output_begin(&handle
, event
,
4005 comm_event
->event_id
.header
.size
, 0, 0);
4010 comm_event
->event_id
.pid
= perf_event_pid(event
, comm_event
->task
);
4011 comm_event
->event_id
.tid
= perf_event_tid(event
, comm_event
->task
);
4013 perf_output_put(&handle
, comm_event
->event_id
);
4014 perf_output_copy(&handle
, comm_event
->comm
,
4015 comm_event
->comm_size
);
4017 perf_event__output_id_sample(event
, &handle
, &sample
);
4019 perf_output_end(&handle
);
4021 comm_event
->event_id
.header
.size
= size
;
4024 static int perf_event_comm_match(struct perf_event
*event
)
4026 if (event
->state
< PERF_EVENT_STATE_INACTIVE
)
4029 if (event
->cpu
!= -1 && event
->cpu
!= smp_processor_id())
4032 if (event
->attr
.comm
)
4038 static void perf_event_comm_ctx(struct perf_event_context
*ctx
,
4039 struct perf_comm_event
*comm_event
)
4041 struct perf_event
*event
;
4043 list_for_each_entry_rcu(event
, &ctx
->event_list
, event_entry
) {
4044 if (perf_event_comm_match(event
))
4045 perf_event_comm_output(event
, comm_event
);
4049 static void perf_event_comm_event(struct perf_comm_event
*comm_event
)
4051 struct perf_cpu_context
*cpuctx
;
4052 struct perf_event_context
*ctx
;
4053 char comm
[TASK_COMM_LEN
];
4058 memset(comm
, 0, sizeof(comm
));
4059 strlcpy(comm
, comm_event
->task
->comm
, sizeof(comm
));
4060 size
= ALIGN(strlen(comm
)+1, sizeof(u64
));
4062 comm_event
->comm
= comm
;
4063 comm_event
->comm_size
= size
;
4065 comm_event
->event_id
.header
.size
= sizeof(comm_event
->event_id
) + size
;
4067 list_for_each_entry_rcu(pmu
, &pmus
, entry
) {
4068 cpuctx
= get_cpu_ptr(pmu
->pmu_cpu_context
);
4069 perf_event_comm_ctx(&cpuctx
->ctx
, comm_event
);
4071 ctxn
= pmu
->task_ctx_nr
;
4075 ctx
= rcu_dereference(current
->perf_event_ctxp
[ctxn
]);
4077 perf_event_comm_ctx(ctx
, comm_event
);
4079 put_cpu_ptr(pmu
->pmu_cpu_context
);
4084 void perf_event_comm(struct task_struct
*task
)
4086 struct perf_comm_event comm_event
;
4087 struct perf_event_context
*ctx
;
4090 for_each_task_context_nr(ctxn
) {
4091 ctx
= task
->perf_event_ctxp
[ctxn
];
4095 perf_event_enable_on_exec(ctx
);
4098 if (!atomic_read(&nr_comm_events
))
4101 comm_event
= (struct perf_comm_event
){
4107 .type
= PERF_RECORD_COMM
,
4116 perf_event_comm_event(&comm_event
);
4123 struct perf_mmap_event
{
4124 struct vm_area_struct
*vma
;
4126 const char *file_name
;
4130 struct perf_event_header header
;
4140 static void perf_event_mmap_output(struct perf_event
*event
,
4141 struct perf_mmap_event
*mmap_event
)
4143 struct perf_output_handle handle
;
4144 struct perf_sample_data sample
;
4145 int size
= mmap_event
->event_id
.header
.size
;
4148 perf_event_header__init_id(&mmap_event
->event_id
.header
, &sample
, event
);
4149 ret
= perf_output_begin(&handle
, event
,
4150 mmap_event
->event_id
.header
.size
, 0, 0);
4154 mmap_event
->event_id
.pid
= perf_event_pid(event
, current
);
4155 mmap_event
->event_id
.tid
= perf_event_tid(event
, current
);
4157 perf_output_put(&handle
, mmap_event
->event_id
);
4158 perf_output_copy(&handle
, mmap_event
->file_name
,
4159 mmap_event
->file_size
);
4161 perf_event__output_id_sample(event
, &handle
, &sample
);
4163 perf_output_end(&handle
);
4165 mmap_event
->event_id
.header
.size
= size
;
4168 static int perf_event_mmap_match(struct perf_event
*event
,
4169 struct perf_mmap_event
*mmap_event
,
4172 if (event
->state
< PERF_EVENT_STATE_INACTIVE
)
4175 if (event
->cpu
!= -1 && event
->cpu
!= smp_processor_id())
4178 if ((!executable
&& event
->attr
.mmap_data
) ||
4179 (executable
&& event
->attr
.mmap
))
4185 static void perf_event_mmap_ctx(struct perf_event_context
*ctx
,
4186 struct perf_mmap_event
*mmap_event
,
4189 struct perf_event
*event
;
4191 list_for_each_entry_rcu(event
, &ctx
->event_list
, event_entry
) {
4192 if (perf_event_mmap_match(event
, mmap_event
, executable
))
4193 perf_event_mmap_output(event
, mmap_event
);
4197 static void perf_event_mmap_event(struct perf_mmap_event
*mmap_event
)
4199 struct perf_cpu_context
*cpuctx
;
4200 struct perf_event_context
*ctx
;
4201 struct vm_area_struct
*vma
= mmap_event
->vma
;
4202 struct file
*file
= vma
->vm_file
;
4210 memset(tmp
, 0, sizeof(tmp
));
4214 * d_path works from the end of the buffer backwards, so we
4215 * need to add enough zero bytes after the string to handle
4216 * the 64bit alignment we do later.
4218 buf
= kzalloc(PATH_MAX
+ sizeof(u64
), GFP_KERNEL
);
4220 name
= strncpy(tmp
, "//enomem", sizeof(tmp
));
4223 name
= d_path(&file
->f_path
, buf
, PATH_MAX
);
4225 name
= strncpy(tmp
, "//toolong", sizeof(tmp
));
4229 if (arch_vma_name(mmap_event
->vma
)) {
4230 name
= strncpy(tmp
, arch_vma_name(mmap_event
->vma
),
4236 name
= strncpy(tmp
, "[vdso]", sizeof(tmp
));
4238 } else if (vma
->vm_start
<= vma
->vm_mm
->start_brk
&&
4239 vma
->vm_end
>= vma
->vm_mm
->brk
) {
4240 name
= strncpy(tmp
, "[heap]", sizeof(tmp
));
4242 } else if (vma
->vm_start
<= vma
->vm_mm
->start_stack
&&
4243 vma
->vm_end
>= vma
->vm_mm
->start_stack
) {
4244 name
= strncpy(tmp
, "[stack]", sizeof(tmp
));
4248 name
= strncpy(tmp
, "//anon", sizeof(tmp
));
4253 size
= ALIGN(strlen(name
)+1, sizeof(u64
));
4255 mmap_event
->file_name
= name
;
4256 mmap_event
->file_size
= size
;
4258 mmap_event
->event_id
.header
.size
= sizeof(mmap_event
->event_id
) + size
;
4261 list_for_each_entry_rcu(pmu
, &pmus
, entry
) {
4262 cpuctx
= get_cpu_ptr(pmu
->pmu_cpu_context
);
4263 perf_event_mmap_ctx(&cpuctx
->ctx
, mmap_event
,
4264 vma
->vm_flags
& VM_EXEC
);
4266 ctxn
= pmu
->task_ctx_nr
;
4270 ctx
= rcu_dereference(current
->perf_event_ctxp
[ctxn
]);
4272 perf_event_mmap_ctx(ctx
, mmap_event
,
4273 vma
->vm_flags
& VM_EXEC
);
4276 put_cpu_ptr(pmu
->pmu_cpu_context
);
4283 void perf_event_mmap(struct vm_area_struct
*vma
)
4285 struct perf_mmap_event mmap_event
;
4287 if (!atomic_read(&nr_mmap_events
))
4290 mmap_event
= (struct perf_mmap_event
){
4296 .type
= PERF_RECORD_MMAP
,
4297 .misc
= PERF_RECORD_MISC_USER
,
4302 .start
= vma
->vm_start
,
4303 .len
= vma
->vm_end
- vma
->vm_start
,
4304 .pgoff
= (u64
)vma
->vm_pgoff
<< PAGE_SHIFT
,
4308 perf_event_mmap_event(&mmap_event
);
4312 * IRQ throttle logging
4315 static void perf_log_throttle(struct perf_event
*event
, int enable
)
4317 struct perf_output_handle handle
;
4318 struct perf_sample_data sample
;
4322 struct perf_event_header header
;
4326 } throttle_event
= {
4328 .type
= PERF_RECORD_THROTTLE
,
4330 .size
= sizeof(throttle_event
),
4332 .time
= perf_clock(),
4333 .id
= primary_event_id(event
),
4334 .stream_id
= event
->id
,
4338 throttle_event
.header
.type
= PERF_RECORD_UNTHROTTLE
;
4340 perf_event_header__init_id(&throttle_event
.header
, &sample
, event
);
4342 ret
= perf_output_begin(&handle
, event
,
4343 throttle_event
.header
.size
, 1, 0);
4347 perf_output_put(&handle
, throttle_event
);
4348 perf_event__output_id_sample(event
, &handle
, &sample
);
4349 perf_output_end(&handle
);
4353 * Generic event overflow handling, sampling.
4356 static int __perf_event_overflow(struct perf_event
*event
, int nmi
,
4357 int throttle
, struct perf_sample_data
*data
,
4358 struct pt_regs
*regs
)
4360 int events
= atomic_read(&event
->event_limit
);
4361 struct hw_perf_event
*hwc
= &event
->hw
;
4365 * Non-sampling counters might still use the PMI to fold short
4366 * hardware counters, ignore those.
4368 if (unlikely(!is_sampling_event(event
)))
4374 if (hwc
->interrupts
!= MAX_INTERRUPTS
) {
4376 if (HZ
* hwc
->interrupts
>
4377 (u64
)sysctl_perf_event_sample_rate
) {
4378 hwc
->interrupts
= MAX_INTERRUPTS
;
4379 perf_log_throttle(event
, 0);
4384 * Keep re-disabling events even though on the previous
4385 * pass we disabled it - just in case we raced with a
4386 * sched-in and the event got enabled again:
4392 if (event
->attr
.freq
) {
4393 u64 now
= perf_clock();
4394 s64 delta
= now
- hwc
->freq_time_stamp
;
4396 hwc
->freq_time_stamp
= now
;
4398 if (delta
> 0 && delta
< 2*TICK_NSEC
)
4399 perf_adjust_period(event
, delta
, hwc
->last_period
);
4403 * XXX event_limit might not quite work as expected on inherited
4407 event
->pending_kill
= POLL_IN
;
4408 if (events
&& atomic_dec_and_test(&event
->event_limit
)) {
4410 event
->pending_kill
= POLL_HUP
;
4412 event
->pending_disable
= 1;
4413 irq_work_queue(&event
->pending
);
4415 perf_event_disable(event
);
4418 if (event
->overflow_handler
)
4419 event
->overflow_handler(event
, nmi
, data
, regs
);
4421 perf_event_output(event
, nmi
, data
, regs
);
4426 int perf_event_overflow(struct perf_event
*event
, int nmi
,
4427 struct perf_sample_data
*data
,
4428 struct pt_regs
*regs
)
4430 return __perf_event_overflow(event
, nmi
, 1, data
, regs
);
4434 * Generic software event infrastructure
4437 struct swevent_htable
{
4438 struct swevent_hlist
*swevent_hlist
;
4439 struct mutex hlist_mutex
;
4442 /* Recursion avoidance in each contexts */
4443 int recursion
[PERF_NR_CONTEXTS
];
4446 static DEFINE_PER_CPU(struct swevent_htable
, swevent_htable
);
4449 * We directly increment event->count and keep a second value in
4450 * event->hw.period_left to count intervals. This period event
4451 * is kept in the range [-sample_period, 0] so that we can use the
4455 static u64
perf_swevent_set_period(struct perf_event
*event
)
4457 struct hw_perf_event
*hwc
= &event
->hw
;
4458 u64 period
= hwc
->last_period
;
4462 hwc
->last_period
= hwc
->sample_period
;
4465 old
= val
= local64_read(&hwc
->period_left
);
4469 nr
= div64_u64(period
+ val
, period
);
4470 offset
= nr
* period
;
4472 if (local64_cmpxchg(&hwc
->period_left
, old
, val
) != old
)
4478 static void perf_swevent_overflow(struct perf_event
*event
, u64 overflow
,
4479 int nmi
, struct perf_sample_data
*data
,
4480 struct pt_regs
*regs
)
4482 struct hw_perf_event
*hwc
= &event
->hw
;
4485 data
->period
= event
->hw
.last_period
;
4487 overflow
= perf_swevent_set_period(event
);
4489 if (hwc
->interrupts
== MAX_INTERRUPTS
)
4492 for (; overflow
; overflow
--) {
4493 if (__perf_event_overflow(event
, nmi
, throttle
,
4496 * We inhibit the overflow from happening when
4497 * hwc->interrupts == MAX_INTERRUPTS.
4505 static void perf_swevent_event(struct perf_event
*event
, u64 nr
,
4506 int nmi
, struct perf_sample_data
*data
,
4507 struct pt_regs
*regs
)
4509 struct hw_perf_event
*hwc
= &event
->hw
;
4511 local64_add(nr
, &event
->count
);
4516 if (!is_sampling_event(event
))
4519 if (nr
== 1 && hwc
->sample_period
== 1 && !event
->attr
.freq
)
4520 return perf_swevent_overflow(event
, 1, nmi
, data
, regs
);
4522 if (local64_add_negative(nr
, &hwc
->period_left
))
4525 perf_swevent_overflow(event
, 0, nmi
, data
, regs
);
4528 static int perf_exclude_event(struct perf_event
*event
,
4529 struct pt_regs
*regs
)
4531 if (event
->hw
.state
& PERF_HES_STOPPED
)
4535 if (event
->attr
.exclude_user
&& user_mode(regs
))
4538 if (event
->attr
.exclude_kernel
&& !user_mode(regs
))
4545 static int perf_swevent_match(struct perf_event
*event
,
4546 enum perf_type_id type
,
4548 struct perf_sample_data
*data
,
4549 struct pt_regs
*regs
)
4551 if (event
->attr
.type
!= type
)
4554 if (event
->attr
.config
!= event_id
)
4557 if (perf_exclude_event(event
, regs
))
4563 static inline u64
swevent_hash(u64 type
, u32 event_id
)
4565 u64 val
= event_id
| (type
<< 32);
4567 return hash_64(val
, SWEVENT_HLIST_BITS
);
4570 static inline struct hlist_head
*
4571 __find_swevent_head(struct swevent_hlist
*hlist
, u64 type
, u32 event_id
)
4573 u64 hash
= swevent_hash(type
, event_id
);
4575 return &hlist
->heads
[hash
];
4578 /* For the read side: events when they trigger */
4579 static inline struct hlist_head
*
4580 find_swevent_head_rcu(struct swevent_htable
*swhash
, u64 type
, u32 event_id
)
4582 struct swevent_hlist
*hlist
;
4584 hlist
= rcu_dereference(swhash
->swevent_hlist
);
4588 return __find_swevent_head(hlist
, type
, event_id
);
4591 /* For the event head insertion and removal in the hlist */
4592 static inline struct hlist_head
*
4593 find_swevent_head(struct swevent_htable
*swhash
, struct perf_event
*event
)
4595 struct swevent_hlist
*hlist
;
4596 u32 event_id
= event
->attr
.config
;
4597 u64 type
= event
->attr
.type
;
4600 * Event scheduling is always serialized against hlist allocation
4601 * and release. Which makes the protected version suitable here.
4602 * The context lock guarantees that.
4604 hlist
= rcu_dereference_protected(swhash
->swevent_hlist
,
4605 lockdep_is_held(&event
->ctx
->lock
));
4609 return __find_swevent_head(hlist
, type
, event_id
);
4612 static void do_perf_sw_event(enum perf_type_id type
, u32 event_id
,
4614 struct perf_sample_data
*data
,
4615 struct pt_regs
*regs
)
4617 struct swevent_htable
*swhash
= &__get_cpu_var(swevent_htable
);
4618 struct perf_event
*event
;
4619 struct hlist_node
*node
;
4620 struct hlist_head
*head
;
4623 head
= find_swevent_head_rcu(swhash
, type
, event_id
);
4627 hlist_for_each_entry_rcu(event
, node
, head
, hlist_entry
) {
4628 if (perf_swevent_match(event
, type
, event_id
, data
, regs
))
4629 perf_swevent_event(event
, nr
, nmi
, data
, regs
);
4635 int perf_swevent_get_recursion_context(void)
4637 struct swevent_htable
*swhash
= &__get_cpu_var(swevent_htable
);
4639 return get_recursion_context(swhash
->recursion
);
4641 EXPORT_SYMBOL_GPL(perf_swevent_get_recursion_context
);
4643 void inline perf_swevent_put_recursion_context(int rctx
)
4645 struct swevent_htable
*swhash
= &__get_cpu_var(swevent_htable
);
4647 put_recursion_context(swhash
->recursion
, rctx
);
4650 void __perf_sw_event(u32 event_id
, u64 nr
, int nmi
,
4651 struct pt_regs
*regs
, u64 addr
)
4653 struct perf_sample_data data
;
4656 preempt_disable_notrace();
4657 rctx
= perf_swevent_get_recursion_context();
4661 perf_sample_data_init(&data
, addr
);
4663 do_perf_sw_event(PERF_TYPE_SOFTWARE
, event_id
, nr
, nmi
, &data
, regs
);
4665 perf_swevent_put_recursion_context(rctx
);
4666 preempt_enable_notrace();
4669 static void perf_swevent_read(struct perf_event
*event
)
4673 static int perf_swevent_add(struct perf_event
*event
, int flags
)
4675 struct swevent_htable
*swhash
= &__get_cpu_var(swevent_htable
);
4676 struct hw_perf_event
*hwc
= &event
->hw
;
4677 struct hlist_head
*head
;
4679 if (is_sampling_event(event
)) {
4680 hwc
->last_period
= hwc
->sample_period
;
4681 perf_swevent_set_period(event
);
4684 hwc
->state
= !(flags
& PERF_EF_START
);
4686 head
= find_swevent_head(swhash
, event
);
4687 if (WARN_ON_ONCE(!head
))
4690 hlist_add_head_rcu(&event
->hlist_entry
, head
);
4695 static void perf_swevent_del(struct perf_event
*event
, int flags
)
4697 hlist_del_rcu(&event
->hlist_entry
);
4700 static void perf_swevent_start(struct perf_event
*event
, int flags
)
4702 event
->hw
.state
= 0;
4705 static void perf_swevent_stop(struct perf_event
*event
, int flags
)
4707 event
->hw
.state
= PERF_HES_STOPPED
;
4710 /* Deref the hlist from the update side */
4711 static inline struct swevent_hlist
*
4712 swevent_hlist_deref(struct swevent_htable
*swhash
)
4714 return rcu_dereference_protected(swhash
->swevent_hlist
,
4715 lockdep_is_held(&swhash
->hlist_mutex
));
4718 static void swevent_hlist_release_rcu(struct rcu_head
*rcu_head
)
4720 struct swevent_hlist
*hlist
;
4722 hlist
= container_of(rcu_head
, struct swevent_hlist
, rcu_head
);
4726 static void swevent_hlist_release(struct swevent_htable
*swhash
)
4728 struct swevent_hlist
*hlist
= swevent_hlist_deref(swhash
);
4733 rcu_assign_pointer(swhash
->swevent_hlist
, NULL
);
4734 call_rcu(&hlist
->rcu_head
, swevent_hlist_release_rcu
);
4737 static void swevent_hlist_put_cpu(struct perf_event
*event
, int cpu
)
4739 struct swevent_htable
*swhash
= &per_cpu(swevent_htable
, cpu
);
4741 mutex_lock(&swhash
->hlist_mutex
);
4743 if (!--swhash
->hlist_refcount
)
4744 swevent_hlist_release(swhash
);
4746 mutex_unlock(&swhash
->hlist_mutex
);
4749 static void swevent_hlist_put(struct perf_event
*event
)
4753 if (event
->cpu
!= -1) {
4754 swevent_hlist_put_cpu(event
, event
->cpu
);
4758 for_each_possible_cpu(cpu
)
4759 swevent_hlist_put_cpu(event
, cpu
);
4762 static int swevent_hlist_get_cpu(struct perf_event
*event
, int cpu
)
4764 struct swevent_htable
*swhash
= &per_cpu(swevent_htable
, cpu
);
4767 mutex_lock(&swhash
->hlist_mutex
);
4769 if (!swevent_hlist_deref(swhash
) && cpu_online(cpu
)) {
4770 struct swevent_hlist
*hlist
;
4772 hlist
= kzalloc(sizeof(*hlist
), GFP_KERNEL
);
4777 rcu_assign_pointer(swhash
->swevent_hlist
, hlist
);
4779 swhash
->hlist_refcount
++;
4781 mutex_unlock(&swhash
->hlist_mutex
);
4786 static int swevent_hlist_get(struct perf_event
*event
)
4789 int cpu
, failed_cpu
;
4791 if (event
->cpu
!= -1)
4792 return swevent_hlist_get_cpu(event
, event
->cpu
);
4795 for_each_possible_cpu(cpu
) {
4796 err
= swevent_hlist_get_cpu(event
, cpu
);
4806 for_each_possible_cpu(cpu
) {
4807 if (cpu
== failed_cpu
)
4809 swevent_hlist_put_cpu(event
, cpu
);
4816 atomic_t perf_swevent_enabled
[PERF_COUNT_SW_MAX
];
4818 static void sw_perf_event_destroy(struct perf_event
*event
)
4820 u64 event_id
= event
->attr
.config
;
4822 WARN_ON(event
->parent
);
4824 jump_label_dec(&perf_swevent_enabled
[event_id
]);
4825 swevent_hlist_put(event
);
4828 static int perf_swevent_init(struct perf_event
*event
)
4830 int event_id
= event
->attr
.config
;
4832 if (event
->attr
.type
!= PERF_TYPE_SOFTWARE
)
4836 case PERF_COUNT_SW_CPU_CLOCK
:
4837 case PERF_COUNT_SW_TASK_CLOCK
:
4844 if (event_id
> PERF_COUNT_SW_MAX
)
4847 if (!event
->parent
) {
4850 err
= swevent_hlist_get(event
);
4854 jump_label_inc(&perf_swevent_enabled
[event_id
]);
4855 event
->destroy
= sw_perf_event_destroy
;
4861 static struct pmu perf_swevent
= {
4862 .task_ctx_nr
= perf_sw_context
,
4864 .event_init
= perf_swevent_init
,
4865 .add
= perf_swevent_add
,
4866 .del
= perf_swevent_del
,
4867 .start
= perf_swevent_start
,
4868 .stop
= perf_swevent_stop
,
4869 .read
= perf_swevent_read
,
4872 #ifdef CONFIG_EVENT_TRACING
4874 static int perf_tp_filter_match(struct perf_event
*event
,
4875 struct perf_sample_data
*data
)
4877 void *record
= data
->raw
->data
;
4879 if (likely(!event
->filter
) || filter_match_preds(event
->filter
, record
))
4884 static int perf_tp_event_match(struct perf_event
*event
,
4885 struct perf_sample_data
*data
,
4886 struct pt_regs
*regs
)
4889 * All tracepoints are from kernel-space.
4891 if (event
->attr
.exclude_kernel
)
4894 if (!perf_tp_filter_match(event
, data
))
4900 void perf_tp_event(u64 addr
, u64 count
, void *record
, int entry_size
,
4901 struct pt_regs
*regs
, struct hlist_head
*head
, int rctx
)
4903 struct perf_sample_data data
;
4904 struct perf_event
*event
;
4905 struct hlist_node
*node
;
4907 struct perf_raw_record raw
= {
4912 perf_sample_data_init(&data
, addr
);
4915 hlist_for_each_entry_rcu(event
, node
, head
, hlist_entry
) {
4916 if (perf_tp_event_match(event
, &data
, regs
))
4917 perf_swevent_event(event
, count
, 1, &data
, regs
);
4920 perf_swevent_put_recursion_context(rctx
);
4922 EXPORT_SYMBOL_GPL(perf_tp_event
);
4924 static void tp_perf_event_destroy(struct perf_event
*event
)
4926 perf_trace_destroy(event
);
4929 static int perf_tp_event_init(struct perf_event
*event
)
4933 if (event
->attr
.type
!= PERF_TYPE_TRACEPOINT
)
4936 err
= perf_trace_init(event
);
4940 event
->destroy
= tp_perf_event_destroy
;
4945 static struct pmu perf_tracepoint
= {
4946 .task_ctx_nr
= perf_sw_context
,
4948 .event_init
= perf_tp_event_init
,
4949 .add
= perf_trace_add
,
4950 .del
= perf_trace_del
,
4951 .start
= perf_swevent_start
,
4952 .stop
= perf_swevent_stop
,
4953 .read
= perf_swevent_read
,
4956 static inline void perf_tp_register(void)
4958 perf_pmu_register(&perf_tracepoint
);
4961 static int perf_event_set_filter(struct perf_event
*event
, void __user
*arg
)
4966 if (event
->attr
.type
!= PERF_TYPE_TRACEPOINT
)
4969 filter_str
= strndup_user(arg
, PAGE_SIZE
);
4970 if (IS_ERR(filter_str
))
4971 return PTR_ERR(filter_str
);
4973 ret
= ftrace_profile_set_filter(event
, event
->attr
.config
, filter_str
);
4979 static void perf_event_free_filter(struct perf_event
*event
)
4981 ftrace_profile_free_filter(event
);
4986 static inline void perf_tp_register(void)
4990 static int perf_event_set_filter(struct perf_event
*event
, void __user
*arg
)
4995 static void perf_event_free_filter(struct perf_event
*event
)
4999 #endif /* CONFIG_EVENT_TRACING */
5001 #ifdef CONFIG_HAVE_HW_BREAKPOINT
5002 void perf_bp_event(struct perf_event
*bp
, void *data
)
5004 struct perf_sample_data sample
;
5005 struct pt_regs
*regs
= data
;
5007 perf_sample_data_init(&sample
, bp
->attr
.bp_addr
);
5009 if (!bp
->hw
.state
&& !perf_exclude_event(bp
, regs
))
5010 perf_swevent_event(bp
, 1, 1, &sample
, regs
);
5015 * hrtimer based swevent callback
5018 static enum hrtimer_restart
perf_swevent_hrtimer(struct hrtimer
*hrtimer
)
5020 enum hrtimer_restart ret
= HRTIMER_RESTART
;
5021 struct perf_sample_data data
;
5022 struct pt_regs
*regs
;
5023 struct perf_event
*event
;
5026 event
= container_of(hrtimer
, struct perf_event
, hw
.hrtimer
);
5027 event
->pmu
->read(event
);
5029 perf_sample_data_init(&data
, 0);
5030 data
.period
= event
->hw
.last_period
;
5031 regs
= get_irq_regs();
5033 if (regs
&& !perf_exclude_event(event
, regs
)) {
5034 if (!(event
->attr
.exclude_idle
&& current
->pid
== 0))
5035 if (perf_event_overflow(event
, 0, &data
, regs
))
5036 ret
= HRTIMER_NORESTART
;
5039 period
= max_t(u64
, 10000, event
->hw
.sample_period
);
5040 hrtimer_forward_now(hrtimer
, ns_to_ktime(period
));
5045 static void perf_swevent_start_hrtimer(struct perf_event
*event
)
5047 struct hw_perf_event
*hwc
= &event
->hw
;
5050 if (!is_sampling_event(event
))
5053 hrtimer_init(&hwc
->hrtimer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
5054 hwc
->hrtimer
.function
= perf_swevent_hrtimer
;
5056 period
= local64_read(&hwc
->period_left
);
5061 local64_set(&hwc
->period_left
, 0);
5063 period
= max_t(u64
, 10000, hwc
->sample_period
);
5065 __hrtimer_start_range_ns(&hwc
->hrtimer
,
5066 ns_to_ktime(period
), 0,
5067 HRTIMER_MODE_REL_PINNED
, 0);
5070 static void perf_swevent_cancel_hrtimer(struct perf_event
*event
)
5072 struct hw_perf_event
*hwc
= &event
->hw
;
5074 if (is_sampling_event(event
)) {
5075 ktime_t remaining
= hrtimer_get_remaining(&hwc
->hrtimer
);
5076 local64_set(&hwc
->period_left
, ktime_to_ns(remaining
));
5078 hrtimer_cancel(&hwc
->hrtimer
);
5083 * Software event: cpu wall time clock
5086 static void cpu_clock_event_update(struct perf_event
*event
)
5091 now
= local_clock();
5092 prev
= local64_xchg(&event
->hw
.prev_count
, now
);
5093 local64_add(now
- prev
, &event
->count
);
5096 static void cpu_clock_event_start(struct perf_event
*event
, int flags
)
5098 local64_set(&event
->hw
.prev_count
, local_clock());
5099 perf_swevent_start_hrtimer(event
);
5102 static void cpu_clock_event_stop(struct perf_event
*event
, int flags
)
5104 perf_swevent_cancel_hrtimer(event
);
5105 cpu_clock_event_update(event
);
5108 static int cpu_clock_event_add(struct perf_event
*event
, int flags
)
5110 if (flags
& PERF_EF_START
)
5111 cpu_clock_event_start(event
, flags
);
5116 static void cpu_clock_event_del(struct perf_event
*event
, int flags
)
5118 cpu_clock_event_stop(event
, flags
);
5121 static void cpu_clock_event_read(struct perf_event
*event
)
5123 cpu_clock_event_update(event
);
5126 static int cpu_clock_event_init(struct perf_event
*event
)
5128 if (event
->attr
.type
!= PERF_TYPE_SOFTWARE
)
5131 if (event
->attr
.config
!= PERF_COUNT_SW_CPU_CLOCK
)
5137 static struct pmu perf_cpu_clock
= {
5138 .task_ctx_nr
= perf_sw_context
,
5140 .event_init
= cpu_clock_event_init
,
5141 .add
= cpu_clock_event_add
,
5142 .del
= cpu_clock_event_del
,
5143 .start
= cpu_clock_event_start
,
5144 .stop
= cpu_clock_event_stop
,
5145 .read
= cpu_clock_event_read
,
5149 * Software event: task time clock
5152 static void task_clock_event_update(struct perf_event
*event
, u64 now
)
5157 prev
= local64_xchg(&event
->hw
.prev_count
, now
);
5159 local64_add(delta
, &event
->count
);
5162 static void task_clock_event_start(struct perf_event
*event
, int flags
)
5164 local64_set(&event
->hw
.prev_count
, event
->ctx
->time
);
5165 perf_swevent_start_hrtimer(event
);
5168 static void task_clock_event_stop(struct perf_event
*event
, int flags
)
5170 perf_swevent_cancel_hrtimer(event
);
5171 task_clock_event_update(event
, event
->ctx
->time
);
5174 static int task_clock_event_add(struct perf_event
*event
, int flags
)
5176 if (flags
& PERF_EF_START
)
5177 task_clock_event_start(event
, flags
);
5182 static void task_clock_event_del(struct perf_event
*event
, int flags
)
5184 task_clock_event_stop(event
, PERF_EF_UPDATE
);
5187 static void task_clock_event_read(struct perf_event
*event
)
5192 update_context_time(event
->ctx
);
5193 time
= event
->ctx
->time
;
5195 u64 now
= perf_clock();
5196 u64 delta
= now
- event
->ctx
->timestamp
;
5197 time
= event
->ctx
->time
+ delta
;
5200 task_clock_event_update(event
, time
);
5203 static int task_clock_event_init(struct perf_event
*event
)
5205 if (event
->attr
.type
!= PERF_TYPE_SOFTWARE
)
5208 if (event
->attr
.config
!= PERF_COUNT_SW_TASK_CLOCK
)
5214 static struct pmu perf_task_clock
= {
5215 .task_ctx_nr
= perf_sw_context
,
5217 .event_init
= task_clock_event_init
,
5218 .add
= task_clock_event_add
,
5219 .del
= task_clock_event_del
,
5220 .start
= task_clock_event_start
,
5221 .stop
= task_clock_event_stop
,
5222 .read
= task_clock_event_read
,
5225 static void perf_pmu_nop_void(struct pmu
*pmu
)
5229 static int perf_pmu_nop_int(struct pmu
*pmu
)
5234 static void perf_pmu_start_txn(struct pmu
*pmu
)
5236 perf_pmu_disable(pmu
);
5239 static int perf_pmu_commit_txn(struct pmu
*pmu
)
5241 perf_pmu_enable(pmu
);
5245 static void perf_pmu_cancel_txn(struct pmu
*pmu
)
5247 perf_pmu_enable(pmu
);
5251 * Ensures all contexts with the same task_ctx_nr have the same
5252 * pmu_cpu_context too.
5254 static void *find_pmu_context(int ctxn
)
5261 list_for_each_entry(pmu
, &pmus
, entry
) {
5262 if (pmu
->task_ctx_nr
== ctxn
)
5263 return pmu
->pmu_cpu_context
;
5269 static void free_pmu_context(void * __percpu cpu_context
)
5273 mutex_lock(&pmus_lock
);
5275 * Like a real lame refcount.
5277 list_for_each_entry(pmu
, &pmus
, entry
) {
5278 if (pmu
->pmu_cpu_context
== cpu_context
)
5282 free_percpu(cpu_context
);
5284 mutex_unlock(&pmus_lock
);
5287 int perf_pmu_register(struct pmu
*pmu
)
5291 mutex_lock(&pmus_lock
);
5293 pmu
->pmu_disable_count
= alloc_percpu(int);
5294 if (!pmu
->pmu_disable_count
)
5297 pmu
->pmu_cpu_context
= find_pmu_context(pmu
->task_ctx_nr
);
5298 if (pmu
->pmu_cpu_context
)
5299 goto got_cpu_context
;
5301 pmu
->pmu_cpu_context
= alloc_percpu(struct perf_cpu_context
);
5302 if (!pmu
->pmu_cpu_context
)
5305 for_each_possible_cpu(cpu
) {
5306 struct perf_cpu_context
*cpuctx
;
5308 cpuctx
= per_cpu_ptr(pmu
->pmu_cpu_context
, cpu
);
5309 __perf_event_init_context(&cpuctx
->ctx
);
5310 cpuctx
->ctx
.type
= cpu_context
;
5311 cpuctx
->ctx
.pmu
= pmu
;
5312 cpuctx
->jiffies_interval
= 1;
5313 INIT_LIST_HEAD(&cpuctx
->rotation_list
);
5317 if (!pmu
->start_txn
) {
5318 if (pmu
->pmu_enable
) {
5320 * If we have pmu_enable/pmu_disable calls, install
5321 * transaction stubs that use that to try and batch
5322 * hardware accesses.
5324 pmu
->start_txn
= perf_pmu_start_txn
;
5325 pmu
->commit_txn
= perf_pmu_commit_txn
;
5326 pmu
->cancel_txn
= perf_pmu_cancel_txn
;
5328 pmu
->start_txn
= perf_pmu_nop_void
;
5329 pmu
->commit_txn
= perf_pmu_nop_int
;
5330 pmu
->cancel_txn
= perf_pmu_nop_void
;
5334 if (!pmu
->pmu_enable
) {
5335 pmu
->pmu_enable
= perf_pmu_nop_void
;
5336 pmu
->pmu_disable
= perf_pmu_nop_void
;
5339 list_add_rcu(&pmu
->entry
, &pmus
);
5342 mutex_unlock(&pmus_lock
);
5347 free_percpu(pmu
->pmu_disable_count
);
5351 void perf_pmu_unregister(struct pmu
*pmu
)
5353 mutex_lock(&pmus_lock
);
5354 list_del_rcu(&pmu
->entry
);
5355 mutex_unlock(&pmus_lock
);
5358 * We dereference the pmu list under both SRCU and regular RCU, so
5359 * synchronize against both of those.
5361 synchronize_srcu(&pmus_srcu
);
5364 free_percpu(pmu
->pmu_disable_count
);
5365 free_pmu_context(pmu
->pmu_cpu_context
);
5368 struct pmu
*perf_init_event(struct perf_event
*event
)
5370 struct pmu
*pmu
= NULL
;
5373 idx
= srcu_read_lock(&pmus_srcu
);
5374 list_for_each_entry_rcu(pmu
, &pmus
, entry
) {
5375 int ret
= pmu
->event_init(event
);
5379 if (ret
!= -ENOENT
) {
5384 pmu
= ERR_PTR(-ENOENT
);
5386 srcu_read_unlock(&pmus_srcu
, idx
);
5392 * Allocate and initialize a event structure
5394 static struct perf_event
*
5395 perf_event_alloc(struct perf_event_attr
*attr
, int cpu
,
5396 struct task_struct
*task
,
5397 struct perf_event
*group_leader
,
5398 struct perf_event
*parent_event
,
5399 perf_overflow_handler_t overflow_handler
)
5402 struct perf_event
*event
;
5403 struct hw_perf_event
*hwc
;
5406 event
= kzalloc(sizeof(*event
), GFP_KERNEL
);
5408 return ERR_PTR(-ENOMEM
);
5411 * Single events are their own group leaders, with an
5412 * empty sibling list:
5415 group_leader
= event
;
5417 mutex_init(&event
->child_mutex
);
5418 INIT_LIST_HEAD(&event
->child_list
);
5420 INIT_LIST_HEAD(&event
->group_entry
);
5421 INIT_LIST_HEAD(&event
->event_entry
);
5422 INIT_LIST_HEAD(&event
->sibling_list
);
5423 init_waitqueue_head(&event
->waitq
);
5424 init_irq_work(&event
->pending
, perf_pending_event
);
5426 mutex_init(&event
->mmap_mutex
);
5429 event
->attr
= *attr
;
5430 event
->group_leader
= group_leader
;
5434 event
->parent
= parent_event
;
5436 event
->ns
= get_pid_ns(current
->nsproxy
->pid_ns
);
5437 event
->id
= atomic64_inc_return(&perf_event_id
);
5439 event
->state
= PERF_EVENT_STATE_INACTIVE
;
5442 event
->attach_state
= PERF_ATTACH_TASK
;
5443 #ifdef CONFIG_HAVE_HW_BREAKPOINT
5445 * hw_breakpoint is a bit difficult here..
5447 if (attr
->type
== PERF_TYPE_BREAKPOINT
)
5448 event
->hw
.bp_target
= task
;
5452 if (!overflow_handler
&& parent_event
)
5453 overflow_handler
= parent_event
->overflow_handler
;
5455 event
->overflow_handler
= overflow_handler
;
5458 event
->state
= PERF_EVENT_STATE_OFF
;
5463 hwc
->sample_period
= attr
->sample_period
;
5464 if (attr
->freq
&& attr
->sample_freq
)
5465 hwc
->sample_period
= 1;
5466 hwc
->last_period
= hwc
->sample_period
;
5468 local64_set(&hwc
->period_left
, hwc
->sample_period
);
5471 * we currently do not support PERF_FORMAT_GROUP on inherited events
5473 if (attr
->inherit
&& (attr
->read_format
& PERF_FORMAT_GROUP
))
5476 pmu
= perf_init_event(event
);
5482 else if (IS_ERR(pmu
))
5487 put_pid_ns(event
->ns
);
5489 return ERR_PTR(err
);
5494 if (!event
->parent
) {
5495 if (event
->attach_state
& PERF_ATTACH_TASK
)
5496 jump_label_inc(&perf_task_events
);
5497 if (event
->attr
.mmap
|| event
->attr
.mmap_data
)
5498 atomic_inc(&nr_mmap_events
);
5499 if (event
->attr
.comm
)
5500 atomic_inc(&nr_comm_events
);
5501 if (event
->attr
.task
)
5502 atomic_inc(&nr_task_events
);
5503 if (event
->attr
.sample_type
& PERF_SAMPLE_CALLCHAIN
) {
5504 err
= get_callchain_buffers();
5507 return ERR_PTR(err
);
5515 static int perf_copy_attr(struct perf_event_attr __user
*uattr
,
5516 struct perf_event_attr
*attr
)
5521 if (!access_ok(VERIFY_WRITE
, uattr
, PERF_ATTR_SIZE_VER0
))
5525 * zero the full structure, so that a short copy will be nice.
5527 memset(attr
, 0, sizeof(*attr
));
5529 ret
= get_user(size
, &uattr
->size
);
5533 if (size
> PAGE_SIZE
) /* silly large */
5536 if (!size
) /* abi compat */
5537 size
= PERF_ATTR_SIZE_VER0
;
5539 if (size
< PERF_ATTR_SIZE_VER0
)
5543 * If we're handed a bigger struct than we know of,
5544 * ensure all the unknown bits are 0 - i.e. new
5545 * user-space does not rely on any kernel feature
5546 * extensions we dont know about yet.
5548 if (size
> sizeof(*attr
)) {
5549 unsigned char __user
*addr
;
5550 unsigned char __user
*end
;
5553 addr
= (void __user
*)uattr
+ sizeof(*attr
);
5554 end
= (void __user
*)uattr
+ size
;
5556 for (; addr
< end
; addr
++) {
5557 ret
= get_user(val
, addr
);
5563 size
= sizeof(*attr
);
5566 ret
= copy_from_user(attr
, uattr
, size
);
5571 * If the type exists, the corresponding creation will verify
5574 if (attr
->type
>= PERF_TYPE_MAX
)
5577 if (attr
->__reserved_1
)
5580 if (attr
->sample_type
& ~(PERF_SAMPLE_MAX
-1))
5583 if (attr
->read_format
& ~(PERF_FORMAT_MAX
-1))
5590 put_user(sizeof(*attr
), &uattr
->size
);
5596 perf_event_set_output(struct perf_event
*event
, struct perf_event
*output_event
)
5598 struct perf_buffer
*buffer
= NULL
, *old_buffer
= NULL
;
5604 /* don't allow circular references */
5605 if (event
== output_event
)
5609 * Don't allow cross-cpu buffers
5611 if (output_event
->cpu
!= event
->cpu
)
5615 * If its not a per-cpu buffer, it must be the same task.
5617 if (output_event
->cpu
== -1 && output_event
->ctx
!= event
->ctx
)
5621 mutex_lock(&event
->mmap_mutex
);
5622 /* Can't redirect output if we've got an active mmap() */
5623 if (atomic_read(&event
->mmap_count
))
5627 /* get the buffer we want to redirect to */
5628 buffer
= perf_buffer_get(output_event
);
5633 old_buffer
= event
->buffer
;
5634 rcu_assign_pointer(event
->buffer
, buffer
);
5637 mutex_unlock(&event
->mmap_mutex
);
5640 perf_buffer_put(old_buffer
);
5646 * sys_perf_event_open - open a performance event, associate it to a task/cpu
5648 * @attr_uptr: event_id type attributes for monitoring/sampling
5651 * @group_fd: group leader event fd
5653 SYSCALL_DEFINE5(perf_event_open
,
5654 struct perf_event_attr __user
*, attr_uptr
,
5655 pid_t
, pid
, int, cpu
, int, group_fd
, unsigned long, flags
)
5657 struct perf_event
*group_leader
= NULL
, *output_event
= NULL
;
5658 struct perf_event
*event
, *sibling
;
5659 struct perf_event_attr attr
;
5660 struct perf_event_context
*ctx
;
5661 struct file
*event_file
= NULL
;
5662 struct file
*group_file
= NULL
;
5663 struct task_struct
*task
= NULL
;
5667 int fput_needed
= 0;
5670 /* for future expandability... */
5671 if (flags
& ~(PERF_FLAG_FD_NO_GROUP
| PERF_FLAG_FD_OUTPUT
))
5674 err
= perf_copy_attr(attr_uptr
, &attr
);
5678 if (!attr
.exclude_kernel
) {
5679 if (perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN
))
5684 if (attr
.sample_freq
> sysctl_perf_event_sample_rate
)
5688 event_fd
= get_unused_fd_flags(O_RDWR
);
5692 if (group_fd
!= -1) {
5693 group_leader
= perf_fget_light(group_fd
, &fput_needed
);
5694 if (IS_ERR(group_leader
)) {
5695 err
= PTR_ERR(group_leader
);
5698 group_file
= group_leader
->filp
;
5699 if (flags
& PERF_FLAG_FD_OUTPUT
)
5700 output_event
= group_leader
;
5701 if (flags
& PERF_FLAG_FD_NO_GROUP
)
5702 group_leader
= NULL
;
5706 task
= find_lively_task_by_vpid(pid
);
5708 err
= PTR_ERR(task
);
5713 event
= perf_event_alloc(&attr
, cpu
, task
, group_leader
, NULL
, NULL
);
5714 if (IS_ERR(event
)) {
5715 err
= PTR_ERR(event
);
5720 * Special case software events and allow them to be part of
5721 * any hardware group.
5726 (is_software_event(event
) != is_software_event(group_leader
))) {
5727 if (is_software_event(event
)) {
5729 * If event and group_leader are not both a software
5730 * event, and event is, then group leader is not.
5732 * Allow the addition of software events to !software
5733 * groups, this is safe because software events never
5736 pmu
= group_leader
->pmu
;
5737 } else if (is_software_event(group_leader
) &&
5738 (group_leader
->group_flags
& PERF_GROUP_SOFTWARE
)) {
5740 * In case the group is a pure software group, and we
5741 * try to add a hardware event, move the whole group to
5742 * the hardware context.
5749 * Get the target context (task or percpu):
5751 ctx
= find_get_context(pmu
, task
, cpu
);
5758 * Look up the group leader (we will attach this event to it):
5764 * Do not allow a recursive hierarchy (this new sibling
5765 * becoming part of another group-sibling):
5767 if (group_leader
->group_leader
!= group_leader
)
5770 * Do not allow to attach to a group in a different
5771 * task or CPU context:
5774 if (group_leader
->ctx
->type
!= ctx
->type
)
5777 if (group_leader
->ctx
!= ctx
)
5782 * Only a group leader can be exclusive or pinned
5784 if (attr
.exclusive
|| attr
.pinned
)
5789 err
= perf_event_set_output(event
, output_event
);
5794 event_file
= anon_inode_getfile("[perf_event]", &perf_fops
, event
, O_RDWR
);
5795 if (IS_ERR(event_file
)) {
5796 err
= PTR_ERR(event_file
);
5801 struct perf_event_context
*gctx
= group_leader
->ctx
;
5803 mutex_lock(&gctx
->mutex
);
5804 perf_event_remove_from_context(group_leader
);
5805 list_for_each_entry(sibling
, &group_leader
->sibling_list
,
5807 perf_event_remove_from_context(sibling
);
5810 mutex_unlock(&gctx
->mutex
);
5814 event
->filp
= event_file
;
5815 WARN_ON_ONCE(ctx
->parent_ctx
);
5816 mutex_lock(&ctx
->mutex
);
5819 perf_install_in_context(ctx
, group_leader
, cpu
);
5821 list_for_each_entry(sibling
, &group_leader
->sibling_list
,
5823 perf_install_in_context(ctx
, sibling
, cpu
);
5828 perf_install_in_context(ctx
, event
, cpu
);
5830 mutex_unlock(&ctx
->mutex
);
5832 event
->owner
= current
;
5834 mutex_lock(¤t
->perf_event_mutex
);
5835 list_add_tail(&event
->owner_entry
, ¤t
->perf_event_list
);
5836 mutex_unlock(¤t
->perf_event_mutex
);
5839 * Precalculate sample_data sizes
5841 perf_event__header_size(event
);
5842 perf_event__id_header_size(event
);
5845 * Drop the reference on the group_event after placing the
5846 * new event on the sibling_list. This ensures destruction
5847 * of the group leader will find the pointer to itself in
5848 * perf_group_detach().
5850 fput_light(group_file
, fput_needed
);
5851 fd_install(event_fd
, event_file
);
5860 put_task_struct(task
);
5862 fput_light(group_file
, fput_needed
);
5864 put_unused_fd(event_fd
);
5869 * perf_event_create_kernel_counter
5871 * @attr: attributes of the counter to create
5872 * @cpu: cpu in which the counter is bound
5873 * @task: task to profile (NULL for percpu)
5876 perf_event_create_kernel_counter(struct perf_event_attr
*attr
, int cpu
,
5877 struct task_struct
*task
,
5878 perf_overflow_handler_t overflow_handler
)
5880 struct perf_event_context
*ctx
;
5881 struct perf_event
*event
;
5885 * Get the target context (task or percpu):
5888 event
= perf_event_alloc(attr
, cpu
, task
, NULL
, NULL
, overflow_handler
);
5889 if (IS_ERR(event
)) {
5890 err
= PTR_ERR(event
);
5894 ctx
= find_get_context(event
->pmu
, task
, cpu
);
5901 WARN_ON_ONCE(ctx
->parent_ctx
);
5902 mutex_lock(&ctx
->mutex
);
5903 perf_install_in_context(ctx
, event
, cpu
);
5905 mutex_unlock(&ctx
->mutex
);
5912 return ERR_PTR(err
);
5914 EXPORT_SYMBOL_GPL(perf_event_create_kernel_counter
);
5916 static void sync_child_event(struct perf_event
*child_event
,
5917 struct task_struct
*child
)
5919 struct perf_event
*parent_event
= child_event
->parent
;
5922 if (child_event
->attr
.inherit_stat
)
5923 perf_event_read_event(child_event
, child
);
5925 child_val
= perf_event_count(child_event
);
5928 * Add back the child's count to the parent's count:
5930 atomic64_add(child_val
, &parent_event
->child_count
);
5931 atomic64_add(child_event
->total_time_enabled
,
5932 &parent_event
->child_total_time_enabled
);
5933 atomic64_add(child_event
->total_time_running
,
5934 &parent_event
->child_total_time_running
);
5937 * Remove this event from the parent's list
5939 WARN_ON_ONCE(parent_event
->ctx
->parent_ctx
);
5940 mutex_lock(&parent_event
->child_mutex
);
5941 list_del_init(&child_event
->child_list
);
5942 mutex_unlock(&parent_event
->child_mutex
);
5945 * Release the parent event, if this was the last
5948 fput(parent_event
->filp
);
5952 __perf_event_exit_task(struct perf_event
*child_event
,
5953 struct perf_event_context
*child_ctx
,
5954 struct task_struct
*child
)
5956 struct perf_event
*parent_event
;
5958 perf_event_remove_from_context(child_event
);
5960 parent_event
= child_event
->parent
;
5962 * It can happen that parent exits first, and has events
5963 * that are still around due to the child reference. These
5964 * events need to be zapped - but otherwise linger.
5967 sync_child_event(child_event
, child
);
5968 free_event(child_event
);
5972 static void perf_event_exit_task_context(struct task_struct
*child
, int ctxn
)
5974 struct perf_event
*child_event
, *tmp
;
5975 struct perf_event_context
*child_ctx
;
5976 unsigned long flags
;
5978 if (likely(!child
->perf_event_ctxp
[ctxn
])) {
5979 perf_event_task(child
, NULL
, 0);
5983 local_irq_save(flags
);
5985 * We can't reschedule here because interrupts are disabled,
5986 * and either child is current or it is a task that can't be
5987 * scheduled, so we are now safe from rescheduling changing
5990 child_ctx
= child
->perf_event_ctxp
[ctxn
];
5991 task_ctx_sched_out(child_ctx
, EVENT_ALL
);
5994 * Take the context lock here so that if find_get_context is
5995 * reading child->perf_event_ctxp, we wait until it has
5996 * incremented the context's refcount before we do put_ctx below.
5998 raw_spin_lock(&child_ctx
->lock
);
5999 child
->perf_event_ctxp
[ctxn
] = NULL
;
6001 * If this context is a clone; unclone it so it can't get
6002 * swapped to another process while we're removing all
6003 * the events from it.
6005 unclone_ctx(child_ctx
);
6006 update_context_time(child_ctx
);
6007 raw_spin_unlock_irqrestore(&child_ctx
->lock
, flags
);
6010 * Report the task dead after unscheduling the events so that we
6011 * won't get any samples after PERF_RECORD_EXIT. We can however still
6012 * get a few PERF_RECORD_READ events.
6014 perf_event_task(child
, child_ctx
, 0);
6017 * We can recurse on the same lock type through:
6019 * __perf_event_exit_task()
6020 * sync_child_event()
6021 * fput(parent_event->filp)
6023 * mutex_lock(&ctx->mutex)
6025 * But since its the parent context it won't be the same instance.
6027 mutex_lock(&child_ctx
->mutex
);
6030 list_for_each_entry_safe(child_event
, tmp
, &child_ctx
->pinned_groups
,
6032 __perf_event_exit_task(child_event
, child_ctx
, child
);
6034 list_for_each_entry_safe(child_event
, tmp
, &child_ctx
->flexible_groups
,
6036 __perf_event_exit_task(child_event
, child_ctx
, child
);
6039 * If the last event was a group event, it will have appended all
6040 * its siblings to the list, but we obtained 'tmp' before that which
6041 * will still point to the list head terminating the iteration.
6043 if (!list_empty(&child_ctx
->pinned_groups
) ||
6044 !list_empty(&child_ctx
->flexible_groups
))
6047 mutex_unlock(&child_ctx
->mutex
);
6053 * When a child task exits, feed back event values to parent events.
6055 void perf_event_exit_task(struct task_struct
*child
)
6057 struct perf_event
*event
, *tmp
;
6060 mutex_lock(&child
->perf_event_mutex
);
6061 list_for_each_entry_safe(event
, tmp
, &child
->perf_event_list
,
6063 list_del_init(&event
->owner_entry
);
6066 * Ensure the list deletion is visible before we clear
6067 * the owner, closes a race against perf_release() where
6068 * we need to serialize on the owner->perf_event_mutex.
6071 event
->owner
= NULL
;
6073 mutex_unlock(&child
->perf_event_mutex
);
6075 for_each_task_context_nr(ctxn
)
6076 perf_event_exit_task_context(child
, ctxn
);
6079 static void perf_free_event(struct perf_event
*event
,
6080 struct perf_event_context
*ctx
)
6082 struct perf_event
*parent
= event
->parent
;
6084 if (WARN_ON_ONCE(!parent
))
6087 mutex_lock(&parent
->child_mutex
);
6088 list_del_init(&event
->child_list
);
6089 mutex_unlock(&parent
->child_mutex
);
6093 perf_group_detach(event
);
6094 list_del_event(event
, ctx
);
6099 * free an unexposed, unused context as created by inheritance by
6100 * perf_event_init_task below, used by fork() in case of fail.
6102 void perf_event_free_task(struct task_struct
*task
)
6104 struct perf_event_context
*ctx
;
6105 struct perf_event
*event
, *tmp
;
6108 for_each_task_context_nr(ctxn
) {
6109 ctx
= task
->perf_event_ctxp
[ctxn
];
6113 mutex_lock(&ctx
->mutex
);
6115 list_for_each_entry_safe(event
, tmp
, &ctx
->pinned_groups
,
6117 perf_free_event(event
, ctx
);
6119 list_for_each_entry_safe(event
, tmp
, &ctx
->flexible_groups
,
6121 perf_free_event(event
, ctx
);
6123 if (!list_empty(&ctx
->pinned_groups
) ||
6124 !list_empty(&ctx
->flexible_groups
))
6127 mutex_unlock(&ctx
->mutex
);
6133 void perf_event_delayed_put(struct task_struct
*task
)
6137 for_each_task_context_nr(ctxn
)
6138 WARN_ON_ONCE(task
->perf_event_ctxp
[ctxn
]);
6142 * inherit a event from parent task to child task:
6144 static struct perf_event
*
6145 inherit_event(struct perf_event
*parent_event
,
6146 struct task_struct
*parent
,
6147 struct perf_event_context
*parent_ctx
,
6148 struct task_struct
*child
,
6149 struct perf_event
*group_leader
,
6150 struct perf_event_context
*child_ctx
)
6152 struct perf_event
*child_event
;
6153 unsigned long flags
;
6156 * Instead of creating recursive hierarchies of events,
6157 * we link inherited events back to the original parent,
6158 * which has a filp for sure, which we use as the reference
6161 if (parent_event
->parent
)
6162 parent_event
= parent_event
->parent
;
6164 child_event
= perf_event_alloc(&parent_event
->attr
,
6167 group_leader
, parent_event
,
6169 if (IS_ERR(child_event
))
6174 * Make the child state follow the state of the parent event,
6175 * not its attr.disabled bit. We hold the parent's mutex,
6176 * so we won't race with perf_event_{en, dis}able_family.
6178 if (parent_event
->state
>= PERF_EVENT_STATE_INACTIVE
)
6179 child_event
->state
= PERF_EVENT_STATE_INACTIVE
;
6181 child_event
->state
= PERF_EVENT_STATE_OFF
;
6183 if (parent_event
->attr
.freq
) {
6184 u64 sample_period
= parent_event
->hw
.sample_period
;
6185 struct hw_perf_event
*hwc
= &child_event
->hw
;
6187 hwc
->sample_period
= sample_period
;
6188 hwc
->last_period
= sample_period
;
6190 local64_set(&hwc
->period_left
, sample_period
);
6193 child_event
->ctx
= child_ctx
;
6194 child_event
->overflow_handler
= parent_event
->overflow_handler
;
6197 * Precalculate sample_data sizes
6199 perf_event__header_size(child_event
);
6200 perf_event__id_header_size(child_event
);
6203 * Link it up in the child's context:
6205 raw_spin_lock_irqsave(&child_ctx
->lock
, flags
);
6206 add_event_to_ctx(child_event
, child_ctx
);
6207 raw_spin_unlock_irqrestore(&child_ctx
->lock
, flags
);
6210 * Get a reference to the parent filp - we will fput it
6211 * when the child event exits. This is safe to do because
6212 * we are in the parent and we know that the filp still
6213 * exists and has a nonzero count:
6215 atomic_long_inc(&parent_event
->filp
->f_count
);
6218 * Link this into the parent event's child list
6220 WARN_ON_ONCE(parent_event
->ctx
->parent_ctx
);
6221 mutex_lock(&parent_event
->child_mutex
);
6222 list_add_tail(&child_event
->child_list
, &parent_event
->child_list
);
6223 mutex_unlock(&parent_event
->child_mutex
);
6228 static int inherit_group(struct perf_event
*parent_event
,
6229 struct task_struct
*parent
,
6230 struct perf_event_context
*parent_ctx
,
6231 struct task_struct
*child
,
6232 struct perf_event_context
*child_ctx
)
6234 struct perf_event
*leader
;
6235 struct perf_event
*sub
;
6236 struct perf_event
*child_ctr
;
6238 leader
= inherit_event(parent_event
, parent
, parent_ctx
,
6239 child
, NULL
, child_ctx
);
6241 return PTR_ERR(leader
);
6242 list_for_each_entry(sub
, &parent_event
->sibling_list
, group_entry
) {
6243 child_ctr
= inherit_event(sub
, parent
, parent_ctx
,
6244 child
, leader
, child_ctx
);
6245 if (IS_ERR(child_ctr
))
6246 return PTR_ERR(child_ctr
);
6252 inherit_task_group(struct perf_event
*event
, struct task_struct
*parent
,
6253 struct perf_event_context
*parent_ctx
,
6254 struct task_struct
*child
, int ctxn
,
6258 struct perf_event_context
*child_ctx
;
6260 if (!event
->attr
.inherit
) {
6265 child_ctx
= child
->perf_event_ctxp
[ctxn
];
6268 * This is executed from the parent task context, so
6269 * inherit events that have been marked for cloning.
6270 * First allocate and initialize a context for the
6274 child_ctx
= alloc_perf_context(event
->pmu
, child
);
6278 child
->perf_event_ctxp
[ctxn
] = child_ctx
;
6281 ret
= inherit_group(event
, parent
, parent_ctx
,
6291 * Initialize the perf_event context in task_struct
6293 int perf_event_init_context(struct task_struct
*child
, int ctxn
)
6295 struct perf_event_context
*child_ctx
, *parent_ctx
;
6296 struct perf_event_context
*cloned_ctx
;
6297 struct perf_event
*event
;
6298 struct task_struct
*parent
= current
;
6299 int inherited_all
= 1;
6300 unsigned long flags
;
6303 child
->perf_event_ctxp
[ctxn
] = NULL
;
6305 mutex_init(&child
->perf_event_mutex
);
6306 INIT_LIST_HEAD(&child
->perf_event_list
);
6308 if (likely(!parent
->perf_event_ctxp
[ctxn
]))
6312 * If the parent's context is a clone, pin it so it won't get
6315 parent_ctx
= perf_pin_task_context(parent
, ctxn
);
6318 * No need to check if parent_ctx != NULL here; since we saw
6319 * it non-NULL earlier, the only reason for it to become NULL
6320 * is if we exit, and since we're currently in the middle of
6321 * a fork we can't be exiting at the same time.
6325 * Lock the parent list. No need to lock the child - not PID
6326 * hashed yet and not running, so nobody can access it.
6328 mutex_lock(&parent_ctx
->mutex
);
6331 * We dont have to disable NMIs - we are only looking at
6332 * the list, not manipulating it:
6334 list_for_each_entry(event
, &parent_ctx
->pinned_groups
, group_entry
) {
6335 ret
= inherit_task_group(event
, parent
, parent_ctx
,
6336 child
, ctxn
, &inherited_all
);
6342 * We can't hold ctx->lock when iterating the ->flexible_group list due
6343 * to allocations, but we need to prevent rotation because
6344 * rotate_ctx() will change the list from interrupt context.
6346 raw_spin_lock_irqsave(&parent_ctx
->lock
, flags
);
6347 parent_ctx
->rotate_disable
= 1;
6348 raw_spin_unlock_irqrestore(&parent_ctx
->lock
, flags
);
6350 list_for_each_entry(event
, &parent_ctx
->flexible_groups
, group_entry
) {
6351 ret
= inherit_task_group(event
, parent
, parent_ctx
,
6352 child
, ctxn
, &inherited_all
);
6357 raw_spin_lock_irqsave(&parent_ctx
->lock
, flags
);
6358 parent_ctx
->rotate_disable
= 0;
6359 raw_spin_unlock_irqrestore(&parent_ctx
->lock
, flags
);
6361 child_ctx
= child
->perf_event_ctxp
[ctxn
];
6363 if (child_ctx
&& inherited_all
) {
6365 * Mark the child context as a clone of the parent
6366 * context, or of whatever the parent is a clone of.
6367 * Note that if the parent is a clone, it could get
6368 * uncloned at any point, but that doesn't matter
6369 * because the list of events and the generation
6370 * count can't have changed since we took the mutex.
6372 cloned_ctx
= rcu_dereference(parent_ctx
->parent_ctx
);
6374 child_ctx
->parent_ctx
= cloned_ctx
;
6375 child_ctx
->parent_gen
= parent_ctx
->parent_gen
;
6377 child_ctx
->parent_ctx
= parent_ctx
;
6378 child_ctx
->parent_gen
= parent_ctx
->generation
;
6380 get_ctx(child_ctx
->parent_ctx
);
6383 mutex_unlock(&parent_ctx
->mutex
);
6385 perf_unpin_context(parent_ctx
);
6391 * Initialize the perf_event context in task_struct
6393 int perf_event_init_task(struct task_struct
*child
)
6397 for_each_task_context_nr(ctxn
) {
6398 ret
= perf_event_init_context(child
, ctxn
);
6406 static void __init
perf_event_init_all_cpus(void)
6408 struct swevent_htable
*swhash
;
6411 for_each_possible_cpu(cpu
) {
6412 swhash
= &per_cpu(swevent_htable
, cpu
);
6413 mutex_init(&swhash
->hlist_mutex
);
6414 INIT_LIST_HEAD(&per_cpu(rotation_list
, cpu
));
6418 static void __cpuinit
perf_event_init_cpu(int cpu
)
6420 struct swevent_htable
*swhash
= &per_cpu(swevent_htable
, cpu
);
6422 mutex_lock(&swhash
->hlist_mutex
);
6423 if (swhash
->hlist_refcount
> 0) {
6424 struct swevent_hlist
*hlist
;
6426 hlist
= kzalloc_node(sizeof(*hlist
), GFP_KERNEL
, cpu_to_node(cpu
));
6428 rcu_assign_pointer(swhash
->swevent_hlist
, hlist
);
6430 mutex_unlock(&swhash
->hlist_mutex
);
6433 #if defined CONFIG_HOTPLUG_CPU || defined CONFIG_KEXEC
6434 static void perf_pmu_rotate_stop(struct pmu
*pmu
)
6436 struct perf_cpu_context
*cpuctx
= this_cpu_ptr(pmu
->pmu_cpu_context
);
6438 WARN_ON(!irqs_disabled());
6440 list_del_init(&cpuctx
->rotation_list
);
6443 static void __perf_event_exit_context(void *__info
)
6445 struct perf_event_context
*ctx
= __info
;
6446 struct perf_event
*event
, *tmp
;
6448 perf_pmu_rotate_stop(ctx
->pmu
);
6450 list_for_each_entry_safe(event
, tmp
, &ctx
->pinned_groups
, group_entry
)
6451 __perf_event_remove_from_context(event
);
6452 list_for_each_entry_safe(event
, tmp
, &ctx
->flexible_groups
, group_entry
)
6453 __perf_event_remove_from_context(event
);
6456 static void perf_event_exit_cpu_context(int cpu
)
6458 struct perf_event_context
*ctx
;
6462 idx
= srcu_read_lock(&pmus_srcu
);
6463 list_for_each_entry_rcu(pmu
, &pmus
, entry
) {
6464 ctx
= &per_cpu_ptr(pmu
->pmu_cpu_context
, cpu
)->ctx
;
6466 mutex_lock(&ctx
->mutex
);
6467 smp_call_function_single(cpu
, __perf_event_exit_context
, ctx
, 1);
6468 mutex_unlock(&ctx
->mutex
);
6470 srcu_read_unlock(&pmus_srcu
, idx
);
6473 static void perf_event_exit_cpu(int cpu
)
6475 struct swevent_htable
*swhash
= &per_cpu(swevent_htable
, cpu
);
6477 mutex_lock(&swhash
->hlist_mutex
);
6478 swevent_hlist_release(swhash
);
6479 mutex_unlock(&swhash
->hlist_mutex
);
6481 perf_event_exit_cpu_context(cpu
);
6484 static inline void perf_event_exit_cpu(int cpu
) { }
6488 perf_reboot(struct notifier_block
*notifier
, unsigned long val
, void *v
)
6492 for_each_online_cpu(cpu
)
6493 perf_event_exit_cpu(cpu
);
6499 * Run the perf reboot notifier at the very last possible moment so that
6500 * the generic watchdog code runs as long as possible.
6502 static struct notifier_block perf_reboot_notifier
= {
6503 .notifier_call
= perf_reboot
,
6504 .priority
= INT_MIN
,
6507 static int __cpuinit
6508 perf_cpu_notify(struct notifier_block
*self
, unsigned long action
, void *hcpu
)
6510 unsigned int cpu
= (long)hcpu
;
6512 switch (action
& ~CPU_TASKS_FROZEN
) {
6514 case CPU_UP_PREPARE
:
6515 case CPU_DOWN_FAILED
:
6516 perf_event_init_cpu(cpu
);
6519 case CPU_UP_CANCELED
:
6520 case CPU_DOWN_PREPARE
:
6521 perf_event_exit_cpu(cpu
);
6531 void __init
perf_event_init(void)
6535 perf_event_init_all_cpus();
6536 init_srcu_struct(&pmus_srcu
);
6537 perf_pmu_register(&perf_swevent
);
6538 perf_pmu_register(&perf_cpu_clock
);
6539 perf_pmu_register(&perf_task_clock
);
6541 perf_cpu_notifier(perf_cpu_notify
);
6542 register_reboot_notifier(&perf_reboot_notifier
);
6544 ret
= init_hw_breakpoint();
6545 WARN(ret
, "hw_breakpoint initialization failed with: %d", ret
);