2 * Performance events core code:
4 * Copyright (C) 2008 Thomas Gleixner <tglx@linutronix.de>
5 * Copyright (C) 2008-2009 Red Hat, Inc., Ingo Molnar
6 * Copyright (C) 2008-2009 Red Hat, Inc., Peter Zijlstra <pzijlstr@redhat.com>
7 * Copyright © 2009 Paul Mackerras, IBM Corp. <paulus@au1.ibm.com>
9 * For licensing details see kernel-base/COPYING
14 #include <linux/cpu.h>
15 #include <linux/smp.h>
16 #include <linux/file.h>
17 #include <linux/poll.h>
18 #include <linux/slab.h>
19 #include <linux/hash.h>
20 #include <linux/sysfs.h>
21 #include <linux/dcache.h>
22 #include <linux/percpu.h>
23 #include <linux/ptrace.h>
24 #include <linux/vmstat.h>
25 #include <linux/vmalloc.h>
26 #include <linux/hardirq.h>
27 #include <linux/rculist.h>
28 #include <linux/uaccess.h>
29 #include <linux/syscalls.h>
30 #include <linux/anon_inodes.h>
31 #include <linux/kernel_stat.h>
32 #include <linux/perf_event.h>
33 #include <linux/ftrace_event.h>
34 #include <linux/hw_breakpoint.h>
36 #include <asm/irq_regs.h>
38 atomic_t perf_task_events __read_mostly
;
39 static atomic_t nr_mmap_events __read_mostly
;
40 static atomic_t nr_comm_events __read_mostly
;
41 static atomic_t nr_task_events __read_mostly
;
43 static LIST_HEAD(pmus
);
44 static DEFINE_MUTEX(pmus_lock
);
45 static struct srcu_struct pmus_srcu
;
48 * perf event paranoia level:
49 * -1 - not paranoid at all
50 * 0 - disallow raw tracepoint access for unpriv
51 * 1 - disallow cpu events for unpriv
52 * 2 - disallow kernel profiling for unpriv
54 int sysctl_perf_event_paranoid __read_mostly
= 1;
56 int sysctl_perf_event_mlock __read_mostly
= 512; /* 'free' kb per user */
59 * max perf event sample rate
61 int sysctl_perf_event_sample_rate __read_mostly
= 100000;
63 static atomic64_t perf_event_id
;
65 void __weak
perf_event_print_debug(void) { }
67 extern __weak
const char *perf_pmu_name(void)
72 void perf_pmu_disable(struct pmu
*pmu
)
74 int *count
= this_cpu_ptr(pmu
->pmu_disable_count
);
76 pmu
->pmu_disable(pmu
);
79 void perf_pmu_enable(struct pmu
*pmu
)
81 int *count
= this_cpu_ptr(pmu
->pmu_disable_count
);
86 static DEFINE_PER_CPU(struct list_head
, rotation_list
);
89 * perf_pmu_rotate_start() and perf_rotate_context() are fully serialized
90 * because they're strictly cpu affine and rotate_start is called with IRQs
91 * disabled, while rotate_context is called from IRQ context.
93 static void perf_pmu_rotate_start(struct pmu
*pmu
)
95 struct perf_cpu_context
*cpuctx
= this_cpu_ptr(pmu
->pmu_cpu_context
);
96 struct list_head
*head
= &__get_cpu_var(rotation_list
);
98 WARN_ON(!irqs_disabled());
100 if (list_empty(&cpuctx
->rotation_list
))
101 list_add(&cpuctx
->rotation_list
, head
);
104 static void get_ctx(struct perf_event_context
*ctx
)
106 WARN_ON(!atomic_inc_not_zero(&ctx
->refcount
));
109 static void free_ctx(struct rcu_head
*head
)
111 struct perf_event_context
*ctx
;
113 ctx
= container_of(head
, struct perf_event_context
, rcu_head
);
117 static void put_ctx(struct perf_event_context
*ctx
)
119 if (atomic_dec_and_test(&ctx
->refcount
)) {
121 put_ctx(ctx
->parent_ctx
);
123 put_task_struct(ctx
->task
);
124 call_rcu(&ctx
->rcu_head
, free_ctx
);
128 static void unclone_ctx(struct perf_event_context
*ctx
)
130 if (ctx
->parent_ctx
) {
131 put_ctx(ctx
->parent_ctx
);
132 ctx
->parent_ctx
= NULL
;
137 * If we inherit events we want to return the parent event id
140 static u64
primary_event_id(struct perf_event
*event
)
145 id
= event
->parent
->id
;
151 * Get the perf_event_context for a task and lock it.
152 * This has to cope with with the fact that until it is locked,
153 * the context could get moved to another task.
155 static struct perf_event_context
*
156 perf_lock_task_context(struct task_struct
*task
, int ctxn
, unsigned long *flags
)
158 struct perf_event_context
*ctx
;
162 ctx
= rcu_dereference(task
->perf_event_ctxp
[ctxn
]);
165 * If this context is a clone of another, it might
166 * get swapped for another underneath us by
167 * perf_event_task_sched_out, though the
168 * rcu_read_lock() protects us from any context
169 * getting freed. Lock the context and check if it
170 * got swapped before we could get the lock, and retry
171 * if so. If we locked the right context, then it
172 * can't get swapped on us any more.
174 raw_spin_lock_irqsave(&ctx
->lock
, *flags
);
175 if (ctx
!= rcu_dereference(task
->perf_event_ctxp
[ctxn
])) {
176 raw_spin_unlock_irqrestore(&ctx
->lock
, *flags
);
180 if (!atomic_inc_not_zero(&ctx
->refcount
)) {
181 raw_spin_unlock_irqrestore(&ctx
->lock
, *flags
);
190 * Get the context for a task and increment its pin_count so it
191 * can't get swapped to another task. This also increments its
192 * reference count so that the context can't get freed.
194 static struct perf_event_context
*
195 perf_pin_task_context(struct task_struct
*task
, int ctxn
)
197 struct perf_event_context
*ctx
;
200 ctx
= perf_lock_task_context(task
, ctxn
, &flags
);
203 raw_spin_unlock_irqrestore(&ctx
->lock
, flags
);
208 static void perf_unpin_context(struct perf_event_context
*ctx
)
212 raw_spin_lock_irqsave(&ctx
->lock
, flags
);
214 raw_spin_unlock_irqrestore(&ctx
->lock
, flags
);
218 static inline u64
perf_clock(void)
220 return local_clock();
224 * Update the record of the current time in a context.
226 static void update_context_time(struct perf_event_context
*ctx
)
228 u64 now
= perf_clock();
230 ctx
->time
+= now
- ctx
->timestamp
;
231 ctx
->timestamp
= now
;
235 * Update the total_time_enabled and total_time_running fields for a event.
237 static void update_event_times(struct perf_event
*event
)
239 struct perf_event_context
*ctx
= event
->ctx
;
242 if (event
->state
< PERF_EVENT_STATE_INACTIVE
||
243 event
->group_leader
->state
< PERF_EVENT_STATE_INACTIVE
)
249 run_end
= event
->tstamp_stopped
;
251 event
->total_time_enabled
= run_end
- event
->tstamp_enabled
;
253 if (event
->state
== PERF_EVENT_STATE_INACTIVE
)
254 run_end
= event
->tstamp_stopped
;
258 event
->total_time_running
= run_end
- event
->tstamp_running
;
262 * Update total_time_enabled and total_time_running for all events in a group.
264 static void update_group_times(struct perf_event
*leader
)
266 struct perf_event
*event
;
268 update_event_times(leader
);
269 list_for_each_entry(event
, &leader
->sibling_list
, group_entry
)
270 update_event_times(event
);
273 static struct list_head
*
274 ctx_group_list(struct perf_event
*event
, struct perf_event_context
*ctx
)
276 if (event
->attr
.pinned
)
277 return &ctx
->pinned_groups
;
279 return &ctx
->flexible_groups
;
283 * Add a event from the lists for its context.
284 * Must be called with ctx->mutex and ctx->lock held.
287 list_add_event(struct perf_event
*event
, struct perf_event_context
*ctx
)
289 WARN_ON_ONCE(event
->attach_state
& PERF_ATTACH_CONTEXT
);
290 event
->attach_state
|= PERF_ATTACH_CONTEXT
;
293 * If we're a stand alone event or group leader, we go to the context
294 * list, group events are kept attached to the group so that
295 * perf_group_detach can, at all times, locate all siblings.
297 if (event
->group_leader
== event
) {
298 struct list_head
*list
;
300 if (is_software_event(event
))
301 event
->group_flags
|= PERF_GROUP_SOFTWARE
;
303 list
= ctx_group_list(event
, ctx
);
304 list_add_tail(&event
->group_entry
, list
);
307 list_add_rcu(&event
->event_entry
, &ctx
->event_list
);
309 perf_pmu_rotate_start(ctx
->pmu
);
311 if (event
->attr
.inherit_stat
)
315 static void perf_group_attach(struct perf_event
*event
)
317 struct perf_event
*group_leader
= event
->group_leader
;
320 * We can have double attach due to group movement in perf_event_open.
322 if (event
->attach_state
& PERF_ATTACH_GROUP
)
325 event
->attach_state
|= PERF_ATTACH_GROUP
;
327 if (group_leader
== event
)
330 if (group_leader
->group_flags
& PERF_GROUP_SOFTWARE
&&
331 !is_software_event(event
))
332 group_leader
->group_flags
&= ~PERF_GROUP_SOFTWARE
;
334 list_add_tail(&event
->group_entry
, &group_leader
->sibling_list
);
335 group_leader
->nr_siblings
++;
339 * Remove a event from the lists for its context.
340 * Must be called with ctx->mutex and ctx->lock held.
343 list_del_event(struct perf_event
*event
, struct perf_event_context
*ctx
)
346 * We can have double detach due to exit/hot-unplug + close.
348 if (!(event
->attach_state
& PERF_ATTACH_CONTEXT
))
351 event
->attach_state
&= ~PERF_ATTACH_CONTEXT
;
354 if (event
->attr
.inherit_stat
)
357 list_del_rcu(&event
->event_entry
);
359 if (event
->group_leader
== event
)
360 list_del_init(&event
->group_entry
);
362 update_group_times(event
);
365 * If event was in error state, then keep it
366 * that way, otherwise bogus counts will be
367 * returned on read(). The only way to get out
368 * of error state is by explicit re-enabling
371 if (event
->state
> PERF_EVENT_STATE_OFF
)
372 event
->state
= PERF_EVENT_STATE_OFF
;
375 static void perf_group_detach(struct perf_event
*event
)
377 struct perf_event
*sibling
, *tmp
;
378 struct list_head
*list
= NULL
;
381 * We can have double detach due to exit/hot-unplug + close.
383 if (!(event
->attach_state
& PERF_ATTACH_GROUP
))
386 event
->attach_state
&= ~PERF_ATTACH_GROUP
;
389 * If this is a sibling, remove it from its group.
391 if (event
->group_leader
!= event
) {
392 list_del_init(&event
->group_entry
);
393 event
->group_leader
->nr_siblings
--;
397 if (!list_empty(&event
->group_entry
))
398 list
= &event
->group_entry
;
401 * If this was a group event with sibling events then
402 * upgrade the siblings to singleton events by adding them
403 * to whatever list we are on.
405 list_for_each_entry_safe(sibling
, tmp
, &event
->sibling_list
, group_entry
) {
407 list_move_tail(&sibling
->group_entry
, list
);
408 sibling
->group_leader
= sibling
;
410 /* Inherit group flags from the previous leader */
411 sibling
->group_flags
= event
->group_flags
;
416 event_filter_match(struct perf_event
*event
)
418 return event
->cpu
== -1 || event
->cpu
== smp_processor_id();
422 event_sched_out(struct perf_event
*event
,
423 struct perf_cpu_context
*cpuctx
,
424 struct perf_event_context
*ctx
)
428 * An event which could not be activated because of
429 * filter mismatch still needs to have its timings
430 * maintained, otherwise bogus information is return
431 * via read() for time_enabled, time_running:
433 if (event
->state
== PERF_EVENT_STATE_INACTIVE
434 && !event_filter_match(event
)) {
435 delta
= ctx
->time
- event
->tstamp_stopped
;
436 event
->tstamp_running
+= delta
;
437 event
->tstamp_stopped
= ctx
->time
;
440 if (event
->state
!= PERF_EVENT_STATE_ACTIVE
)
443 event
->state
= PERF_EVENT_STATE_INACTIVE
;
444 if (event
->pending_disable
) {
445 event
->pending_disable
= 0;
446 event
->state
= PERF_EVENT_STATE_OFF
;
448 event
->tstamp_stopped
= ctx
->time
;
449 event
->pmu
->del(event
, 0);
452 if (!is_software_event(event
))
453 cpuctx
->active_oncpu
--;
455 if (event
->attr
.exclusive
|| !cpuctx
->active_oncpu
)
456 cpuctx
->exclusive
= 0;
460 group_sched_out(struct perf_event
*group_event
,
461 struct perf_cpu_context
*cpuctx
,
462 struct perf_event_context
*ctx
)
464 struct perf_event
*event
;
465 int state
= group_event
->state
;
467 event_sched_out(group_event
, cpuctx
, ctx
);
470 * Schedule out siblings (if any):
472 list_for_each_entry(event
, &group_event
->sibling_list
, group_entry
)
473 event_sched_out(event
, cpuctx
, ctx
);
475 if (state
== PERF_EVENT_STATE_ACTIVE
&& group_event
->attr
.exclusive
)
476 cpuctx
->exclusive
= 0;
479 static inline struct perf_cpu_context
*
480 __get_cpu_context(struct perf_event_context
*ctx
)
482 return this_cpu_ptr(ctx
->pmu
->pmu_cpu_context
);
486 * Cross CPU call to remove a performance event
488 * We disable the event on the hardware level first. After that we
489 * remove it from the context list.
491 static void __perf_event_remove_from_context(void *info
)
493 struct perf_event
*event
= info
;
494 struct perf_event_context
*ctx
= event
->ctx
;
495 struct perf_cpu_context
*cpuctx
= __get_cpu_context(ctx
);
498 * If this is a task context, we need to check whether it is
499 * the current task context of this cpu. If not it has been
500 * scheduled out before the smp call arrived.
502 if (ctx
->task
&& cpuctx
->task_ctx
!= ctx
)
505 raw_spin_lock(&ctx
->lock
);
507 event_sched_out(event
, cpuctx
, ctx
);
509 list_del_event(event
, ctx
);
511 raw_spin_unlock(&ctx
->lock
);
516 * Remove the event from a task's (or a CPU's) list of events.
518 * Must be called with ctx->mutex held.
520 * CPU events are removed with a smp call. For task events we only
521 * call when the task is on a CPU.
523 * If event->ctx is a cloned context, callers must make sure that
524 * every task struct that event->ctx->task could possibly point to
525 * remains valid. This is OK when called from perf_release since
526 * that only calls us on the top-level context, which can't be a clone.
527 * When called from perf_event_exit_task, it's OK because the
528 * context has been detached from its task.
530 static void perf_event_remove_from_context(struct perf_event
*event
)
532 struct perf_event_context
*ctx
= event
->ctx
;
533 struct task_struct
*task
= ctx
->task
;
537 * Per cpu events are removed via an smp call and
538 * the removal is always successful.
540 smp_call_function_single(event
->cpu
,
541 __perf_event_remove_from_context
,
547 task_oncpu_function_call(task
, __perf_event_remove_from_context
,
550 raw_spin_lock_irq(&ctx
->lock
);
552 * If the context is active we need to retry the smp call.
554 if (ctx
->nr_active
&& !list_empty(&event
->group_entry
)) {
555 raw_spin_unlock_irq(&ctx
->lock
);
560 * The lock prevents that this context is scheduled in so we
561 * can remove the event safely, if the call above did not
564 if (!list_empty(&event
->group_entry
))
565 list_del_event(event
, ctx
);
566 raw_spin_unlock_irq(&ctx
->lock
);
570 * Cross CPU call to disable a performance event
572 static void __perf_event_disable(void *info
)
574 struct perf_event
*event
= info
;
575 struct perf_event_context
*ctx
= event
->ctx
;
576 struct perf_cpu_context
*cpuctx
= __get_cpu_context(ctx
);
579 * If this is a per-task event, need to check whether this
580 * event's task is the current task on this cpu.
582 if (ctx
->task
&& cpuctx
->task_ctx
!= ctx
)
585 raw_spin_lock(&ctx
->lock
);
588 * If the event is on, turn it off.
589 * If it is in error state, leave it in error state.
591 if (event
->state
>= PERF_EVENT_STATE_INACTIVE
) {
592 update_context_time(ctx
);
593 update_group_times(event
);
594 if (event
== event
->group_leader
)
595 group_sched_out(event
, cpuctx
, ctx
);
597 event_sched_out(event
, cpuctx
, ctx
);
598 event
->state
= PERF_EVENT_STATE_OFF
;
601 raw_spin_unlock(&ctx
->lock
);
607 * If event->ctx is a cloned context, callers must make sure that
608 * every task struct that event->ctx->task could possibly point to
609 * remains valid. This condition is satisifed when called through
610 * perf_event_for_each_child or perf_event_for_each because they
611 * hold the top-level event's child_mutex, so any descendant that
612 * goes to exit will block in sync_child_event.
613 * When called from perf_pending_event it's OK because event->ctx
614 * is the current context on this CPU and preemption is disabled,
615 * hence we can't get into perf_event_task_sched_out for this context.
617 void perf_event_disable(struct perf_event
*event
)
619 struct perf_event_context
*ctx
= event
->ctx
;
620 struct task_struct
*task
= ctx
->task
;
624 * Disable the event on the cpu that it's on
626 smp_call_function_single(event
->cpu
, __perf_event_disable
,
632 task_oncpu_function_call(task
, __perf_event_disable
, event
);
634 raw_spin_lock_irq(&ctx
->lock
);
636 * If the event is still active, we need to retry the cross-call.
638 if (event
->state
== PERF_EVENT_STATE_ACTIVE
) {
639 raw_spin_unlock_irq(&ctx
->lock
);
644 * Since we have the lock this context can't be scheduled
645 * in, so we can change the state safely.
647 if (event
->state
== PERF_EVENT_STATE_INACTIVE
) {
648 update_group_times(event
);
649 event
->state
= PERF_EVENT_STATE_OFF
;
652 raw_spin_unlock_irq(&ctx
->lock
);
656 event_sched_in(struct perf_event
*event
,
657 struct perf_cpu_context
*cpuctx
,
658 struct perf_event_context
*ctx
)
660 if (event
->state
<= PERF_EVENT_STATE_OFF
)
663 event
->state
= PERF_EVENT_STATE_ACTIVE
;
664 event
->oncpu
= smp_processor_id();
666 * The new state must be visible before we turn it on in the hardware:
670 if (event
->pmu
->add(event
, PERF_EF_START
)) {
671 event
->state
= PERF_EVENT_STATE_INACTIVE
;
676 event
->tstamp_running
+= ctx
->time
- event
->tstamp_stopped
;
678 event
->shadow_ctx_time
= ctx
->time
- ctx
->timestamp
;
680 if (!is_software_event(event
))
681 cpuctx
->active_oncpu
++;
684 if (event
->attr
.exclusive
)
685 cpuctx
->exclusive
= 1;
691 group_sched_in(struct perf_event
*group_event
,
692 struct perf_cpu_context
*cpuctx
,
693 struct perf_event_context
*ctx
)
695 struct perf_event
*event
, *partial_group
= NULL
;
696 struct pmu
*pmu
= group_event
->pmu
;
698 bool simulate
= false;
700 if (group_event
->state
== PERF_EVENT_STATE_OFF
)
705 if (event_sched_in(group_event
, cpuctx
, ctx
)) {
706 pmu
->cancel_txn(pmu
);
711 * Schedule in siblings as one group (if any):
713 list_for_each_entry(event
, &group_event
->sibling_list
, group_entry
) {
714 if (event_sched_in(event
, cpuctx
, ctx
)) {
715 partial_group
= event
;
720 if (!pmu
->commit_txn(pmu
))
725 * Groups can be scheduled in as one unit only, so undo any
726 * partial group before returning:
727 * The events up to the failed event are scheduled out normally,
728 * tstamp_stopped will be updated.
730 * The failed events and the remaining siblings need to have
731 * their timings updated as if they had gone thru event_sched_in()
732 * and event_sched_out(). This is required to get consistent timings
733 * across the group. This also takes care of the case where the group
734 * could never be scheduled by ensuring tstamp_stopped is set to mark
735 * the time the event was actually stopped, such that time delta
736 * calculation in update_event_times() is correct.
738 list_for_each_entry(event
, &group_event
->sibling_list
, group_entry
) {
739 if (event
== partial_group
)
743 event
->tstamp_running
+= now
- event
->tstamp_stopped
;
744 event
->tstamp_stopped
= now
;
746 event_sched_out(event
, cpuctx
, ctx
);
749 event_sched_out(group_event
, cpuctx
, ctx
);
751 pmu
->cancel_txn(pmu
);
757 * Work out whether we can put this event group on the CPU now.
759 static int group_can_go_on(struct perf_event
*event
,
760 struct perf_cpu_context
*cpuctx
,
764 * Groups consisting entirely of software events can always go on.
766 if (event
->group_flags
& PERF_GROUP_SOFTWARE
)
769 * If an exclusive group is already on, no other hardware
772 if (cpuctx
->exclusive
)
775 * If this group is exclusive and there are already
776 * events on the CPU, it can't go on.
778 if (event
->attr
.exclusive
&& cpuctx
->active_oncpu
)
781 * Otherwise, try to add it if all previous groups were able
787 static void add_event_to_ctx(struct perf_event
*event
,
788 struct perf_event_context
*ctx
)
790 list_add_event(event
, ctx
);
791 perf_group_attach(event
);
792 event
->tstamp_enabled
= ctx
->time
;
793 event
->tstamp_running
= ctx
->time
;
794 event
->tstamp_stopped
= ctx
->time
;
798 * Cross CPU call to install and enable a performance event
800 * Must be called with ctx->mutex held
802 static void __perf_install_in_context(void *info
)
804 struct perf_event
*event
= info
;
805 struct perf_event_context
*ctx
= event
->ctx
;
806 struct perf_event
*leader
= event
->group_leader
;
807 struct perf_cpu_context
*cpuctx
= __get_cpu_context(ctx
);
811 * If this is a task context, we need to check whether it is
812 * the current task context of this cpu. If not it has been
813 * scheduled out before the smp call arrived.
814 * Or possibly this is the right context but it isn't
815 * on this cpu because it had no events.
817 if (ctx
->task
&& cpuctx
->task_ctx
!= ctx
) {
818 if (cpuctx
->task_ctx
|| ctx
->task
!= current
)
820 cpuctx
->task_ctx
= ctx
;
823 raw_spin_lock(&ctx
->lock
);
825 update_context_time(ctx
);
827 add_event_to_ctx(event
, ctx
);
829 if (event
->cpu
!= -1 && event
->cpu
!= smp_processor_id())
833 * Don't put the event on if it is disabled or if
834 * it is in a group and the group isn't on.
836 if (event
->state
!= PERF_EVENT_STATE_INACTIVE
||
837 (leader
!= event
&& leader
->state
!= PERF_EVENT_STATE_ACTIVE
))
841 * An exclusive event can't go on if there are already active
842 * hardware events, and no hardware event can go on if there
843 * is already an exclusive event on.
845 if (!group_can_go_on(event
, cpuctx
, 1))
848 err
= event_sched_in(event
, cpuctx
, ctx
);
852 * This event couldn't go on. If it is in a group
853 * then we have to pull the whole group off.
854 * If the event group is pinned then put it in error state.
857 group_sched_out(leader
, cpuctx
, ctx
);
858 if (leader
->attr
.pinned
) {
859 update_group_times(leader
);
860 leader
->state
= PERF_EVENT_STATE_ERROR
;
865 raw_spin_unlock(&ctx
->lock
);
869 * Attach a performance event to a context
871 * First we add the event to the list with the hardware enable bit
872 * in event->hw_config cleared.
874 * If the event is attached to a task which is on a CPU we use a smp
875 * call to enable it in the task context. The task might have been
876 * scheduled away, but we check this in the smp call again.
878 * Must be called with ctx->mutex held.
881 perf_install_in_context(struct perf_event_context
*ctx
,
882 struct perf_event
*event
,
885 struct task_struct
*task
= ctx
->task
;
891 * Per cpu events are installed via an smp call and
892 * the install is always successful.
894 smp_call_function_single(cpu
, __perf_install_in_context
,
900 task_oncpu_function_call(task
, __perf_install_in_context
,
903 raw_spin_lock_irq(&ctx
->lock
);
905 * we need to retry the smp call.
907 if (ctx
->is_active
&& list_empty(&event
->group_entry
)) {
908 raw_spin_unlock_irq(&ctx
->lock
);
913 * The lock prevents that this context is scheduled in so we
914 * can add the event safely, if it the call above did not
917 if (list_empty(&event
->group_entry
))
918 add_event_to_ctx(event
, ctx
);
919 raw_spin_unlock_irq(&ctx
->lock
);
923 * Put a event into inactive state and update time fields.
924 * Enabling the leader of a group effectively enables all
925 * the group members that aren't explicitly disabled, so we
926 * have to update their ->tstamp_enabled also.
927 * Note: this works for group members as well as group leaders
928 * since the non-leader members' sibling_lists will be empty.
930 static void __perf_event_mark_enabled(struct perf_event
*event
,
931 struct perf_event_context
*ctx
)
933 struct perf_event
*sub
;
935 event
->state
= PERF_EVENT_STATE_INACTIVE
;
936 event
->tstamp_enabled
= ctx
->time
- event
->total_time_enabled
;
937 list_for_each_entry(sub
, &event
->sibling_list
, group_entry
) {
938 if (sub
->state
>= PERF_EVENT_STATE_INACTIVE
) {
939 sub
->tstamp_enabled
=
940 ctx
->time
- sub
->total_time_enabled
;
946 * Cross CPU call to enable a performance event
948 static void __perf_event_enable(void *info
)
950 struct perf_event
*event
= info
;
951 struct perf_event_context
*ctx
= event
->ctx
;
952 struct perf_event
*leader
= event
->group_leader
;
953 struct perf_cpu_context
*cpuctx
= __get_cpu_context(ctx
);
957 * If this is a per-task event, need to check whether this
958 * event's task is the current task on this cpu.
960 if (ctx
->task
&& cpuctx
->task_ctx
!= ctx
) {
961 if (cpuctx
->task_ctx
|| ctx
->task
!= current
)
963 cpuctx
->task_ctx
= ctx
;
966 raw_spin_lock(&ctx
->lock
);
968 update_context_time(ctx
);
970 if (event
->state
>= PERF_EVENT_STATE_INACTIVE
)
972 __perf_event_mark_enabled(event
, ctx
);
974 if (event
->cpu
!= -1 && event
->cpu
!= smp_processor_id())
978 * If the event is in a group and isn't the group leader,
979 * then don't put it on unless the group is on.
981 if (leader
!= event
&& leader
->state
!= PERF_EVENT_STATE_ACTIVE
)
984 if (!group_can_go_on(event
, cpuctx
, 1)) {
988 err
= group_sched_in(event
, cpuctx
, ctx
);
990 err
= event_sched_in(event
, cpuctx
, ctx
);
995 * If this event can't go on and it's part of a
996 * group, then the whole group has to come off.
999 group_sched_out(leader
, cpuctx
, ctx
);
1000 if (leader
->attr
.pinned
) {
1001 update_group_times(leader
);
1002 leader
->state
= PERF_EVENT_STATE_ERROR
;
1007 raw_spin_unlock(&ctx
->lock
);
1013 * If event->ctx is a cloned context, callers must make sure that
1014 * every task struct that event->ctx->task could possibly point to
1015 * remains valid. This condition is satisfied when called through
1016 * perf_event_for_each_child or perf_event_for_each as described
1017 * for perf_event_disable.
1019 void perf_event_enable(struct perf_event
*event
)
1021 struct perf_event_context
*ctx
= event
->ctx
;
1022 struct task_struct
*task
= ctx
->task
;
1026 * Enable the event on the cpu that it's on
1028 smp_call_function_single(event
->cpu
, __perf_event_enable
,
1033 raw_spin_lock_irq(&ctx
->lock
);
1034 if (event
->state
>= PERF_EVENT_STATE_INACTIVE
)
1038 * If the event is in error state, clear that first.
1039 * That way, if we see the event in error state below, we
1040 * know that it has gone back into error state, as distinct
1041 * from the task having been scheduled away before the
1042 * cross-call arrived.
1044 if (event
->state
== PERF_EVENT_STATE_ERROR
)
1045 event
->state
= PERF_EVENT_STATE_OFF
;
1048 raw_spin_unlock_irq(&ctx
->lock
);
1049 task_oncpu_function_call(task
, __perf_event_enable
, event
);
1051 raw_spin_lock_irq(&ctx
->lock
);
1054 * If the context is active and the event is still off,
1055 * we need to retry the cross-call.
1057 if (ctx
->is_active
&& event
->state
== PERF_EVENT_STATE_OFF
)
1061 * Since we have the lock this context can't be scheduled
1062 * in, so we can change the state safely.
1064 if (event
->state
== PERF_EVENT_STATE_OFF
)
1065 __perf_event_mark_enabled(event
, ctx
);
1068 raw_spin_unlock_irq(&ctx
->lock
);
1071 static int perf_event_refresh(struct perf_event
*event
, int refresh
)
1074 * not supported on inherited events
1076 if (event
->attr
.inherit
)
1079 atomic_add(refresh
, &event
->event_limit
);
1080 perf_event_enable(event
);
1086 EVENT_FLEXIBLE
= 0x1,
1088 EVENT_ALL
= EVENT_FLEXIBLE
| EVENT_PINNED
,
1091 static void ctx_sched_out(struct perf_event_context
*ctx
,
1092 struct perf_cpu_context
*cpuctx
,
1093 enum event_type_t event_type
)
1095 struct perf_event
*event
;
1097 raw_spin_lock(&ctx
->lock
);
1098 perf_pmu_disable(ctx
->pmu
);
1100 if (likely(!ctx
->nr_events
))
1102 update_context_time(ctx
);
1104 if (!ctx
->nr_active
)
1107 if (event_type
& EVENT_PINNED
) {
1108 list_for_each_entry(event
, &ctx
->pinned_groups
, group_entry
)
1109 group_sched_out(event
, cpuctx
, ctx
);
1112 if (event_type
& EVENT_FLEXIBLE
) {
1113 list_for_each_entry(event
, &ctx
->flexible_groups
, group_entry
)
1114 group_sched_out(event
, cpuctx
, ctx
);
1117 perf_pmu_enable(ctx
->pmu
);
1118 raw_spin_unlock(&ctx
->lock
);
1122 * Test whether two contexts are equivalent, i.e. whether they
1123 * have both been cloned from the same version of the same context
1124 * and they both have the same number of enabled events.
1125 * If the number of enabled events is the same, then the set
1126 * of enabled events should be the same, because these are both
1127 * inherited contexts, therefore we can't access individual events
1128 * in them directly with an fd; we can only enable/disable all
1129 * events via prctl, or enable/disable all events in a family
1130 * via ioctl, which will have the same effect on both contexts.
1132 static int context_equiv(struct perf_event_context
*ctx1
,
1133 struct perf_event_context
*ctx2
)
1135 return ctx1
->parent_ctx
&& ctx1
->parent_ctx
== ctx2
->parent_ctx
1136 && ctx1
->parent_gen
== ctx2
->parent_gen
1137 && !ctx1
->pin_count
&& !ctx2
->pin_count
;
1140 static void __perf_event_sync_stat(struct perf_event
*event
,
1141 struct perf_event
*next_event
)
1145 if (!event
->attr
.inherit_stat
)
1149 * Update the event value, we cannot use perf_event_read()
1150 * because we're in the middle of a context switch and have IRQs
1151 * disabled, which upsets smp_call_function_single(), however
1152 * we know the event must be on the current CPU, therefore we
1153 * don't need to use it.
1155 switch (event
->state
) {
1156 case PERF_EVENT_STATE_ACTIVE
:
1157 event
->pmu
->read(event
);
1160 case PERF_EVENT_STATE_INACTIVE
:
1161 update_event_times(event
);
1169 * In order to keep per-task stats reliable we need to flip the event
1170 * values when we flip the contexts.
1172 value
= local64_read(&next_event
->count
);
1173 value
= local64_xchg(&event
->count
, value
);
1174 local64_set(&next_event
->count
, value
);
1176 swap(event
->total_time_enabled
, next_event
->total_time_enabled
);
1177 swap(event
->total_time_running
, next_event
->total_time_running
);
1180 * Since we swizzled the values, update the user visible data too.
1182 perf_event_update_userpage(event
);
1183 perf_event_update_userpage(next_event
);
1186 #define list_next_entry(pos, member) \
1187 list_entry(pos->member.next, typeof(*pos), member)
1189 static void perf_event_sync_stat(struct perf_event_context
*ctx
,
1190 struct perf_event_context
*next_ctx
)
1192 struct perf_event
*event
, *next_event
;
1197 update_context_time(ctx
);
1199 event
= list_first_entry(&ctx
->event_list
,
1200 struct perf_event
, event_entry
);
1202 next_event
= list_first_entry(&next_ctx
->event_list
,
1203 struct perf_event
, event_entry
);
1205 while (&event
->event_entry
!= &ctx
->event_list
&&
1206 &next_event
->event_entry
!= &next_ctx
->event_list
) {
1208 __perf_event_sync_stat(event
, next_event
);
1210 event
= list_next_entry(event
, event_entry
);
1211 next_event
= list_next_entry(next_event
, event_entry
);
1215 void perf_event_context_sched_out(struct task_struct
*task
, int ctxn
,
1216 struct task_struct
*next
)
1218 struct perf_event_context
*ctx
= task
->perf_event_ctxp
[ctxn
];
1219 struct perf_event_context
*next_ctx
;
1220 struct perf_event_context
*parent
;
1221 struct perf_cpu_context
*cpuctx
;
1227 cpuctx
= __get_cpu_context(ctx
);
1228 if (!cpuctx
->task_ctx
)
1232 parent
= rcu_dereference(ctx
->parent_ctx
);
1233 next_ctx
= next
->perf_event_ctxp
[ctxn
];
1234 if (parent
&& next_ctx
&&
1235 rcu_dereference(next_ctx
->parent_ctx
) == parent
) {
1237 * Looks like the two contexts are clones, so we might be
1238 * able to optimize the context switch. We lock both
1239 * contexts and check that they are clones under the
1240 * lock (including re-checking that neither has been
1241 * uncloned in the meantime). It doesn't matter which
1242 * order we take the locks because no other cpu could
1243 * be trying to lock both of these tasks.
1245 raw_spin_lock(&ctx
->lock
);
1246 raw_spin_lock_nested(&next_ctx
->lock
, SINGLE_DEPTH_NESTING
);
1247 if (context_equiv(ctx
, next_ctx
)) {
1249 * XXX do we need a memory barrier of sorts
1250 * wrt to rcu_dereference() of perf_event_ctxp
1252 task
->perf_event_ctxp
[ctxn
] = next_ctx
;
1253 next
->perf_event_ctxp
[ctxn
] = ctx
;
1255 next_ctx
->task
= task
;
1258 perf_event_sync_stat(ctx
, next_ctx
);
1260 raw_spin_unlock(&next_ctx
->lock
);
1261 raw_spin_unlock(&ctx
->lock
);
1266 ctx_sched_out(ctx
, cpuctx
, EVENT_ALL
);
1267 cpuctx
->task_ctx
= NULL
;
1271 #define for_each_task_context_nr(ctxn) \
1272 for ((ctxn) = 0; (ctxn) < perf_nr_task_contexts; (ctxn)++)
1275 * Called from scheduler to remove the events of the current task,
1276 * with interrupts disabled.
1278 * We stop each event and update the event value in event->count.
1280 * This does not protect us against NMI, but disable()
1281 * sets the disabled bit in the control field of event _before_
1282 * accessing the event control register. If a NMI hits, then it will
1283 * not restart the event.
1285 void __perf_event_task_sched_out(struct task_struct
*task
,
1286 struct task_struct
*next
)
1290 perf_sw_event(PERF_COUNT_SW_CONTEXT_SWITCHES
, 1, 1, NULL
, 0);
1292 for_each_task_context_nr(ctxn
)
1293 perf_event_context_sched_out(task
, ctxn
, next
);
1296 static void task_ctx_sched_out(struct perf_event_context
*ctx
,
1297 enum event_type_t event_type
)
1299 struct perf_cpu_context
*cpuctx
= __get_cpu_context(ctx
);
1301 if (!cpuctx
->task_ctx
)
1304 if (WARN_ON_ONCE(ctx
!= cpuctx
->task_ctx
))
1307 ctx_sched_out(ctx
, cpuctx
, event_type
);
1308 cpuctx
->task_ctx
= NULL
;
1312 * Called with IRQs disabled
1314 static void cpu_ctx_sched_out(struct perf_cpu_context
*cpuctx
,
1315 enum event_type_t event_type
)
1317 ctx_sched_out(&cpuctx
->ctx
, cpuctx
, event_type
);
1321 ctx_pinned_sched_in(struct perf_event_context
*ctx
,
1322 struct perf_cpu_context
*cpuctx
)
1324 struct perf_event
*event
;
1326 list_for_each_entry(event
, &ctx
->pinned_groups
, group_entry
) {
1327 if (event
->state
<= PERF_EVENT_STATE_OFF
)
1329 if (event
->cpu
!= -1 && event
->cpu
!= smp_processor_id())
1332 if (group_can_go_on(event
, cpuctx
, 1))
1333 group_sched_in(event
, cpuctx
, ctx
);
1336 * If this pinned group hasn't been scheduled,
1337 * put it in error state.
1339 if (event
->state
== PERF_EVENT_STATE_INACTIVE
) {
1340 update_group_times(event
);
1341 event
->state
= PERF_EVENT_STATE_ERROR
;
1347 ctx_flexible_sched_in(struct perf_event_context
*ctx
,
1348 struct perf_cpu_context
*cpuctx
)
1350 struct perf_event
*event
;
1353 list_for_each_entry(event
, &ctx
->flexible_groups
, group_entry
) {
1354 /* Ignore events in OFF or ERROR state */
1355 if (event
->state
<= PERF_EVENT_STATE_OFF
)
1358 * Listen to the 'cpu' scheduling filter constraint
1361 if (event
->cpu
!= -1 && event
->cpu
!= smp_processor_id())
1364 if (group_can_go_on(event
, cpuctx
, can_add_hw
)) {
1365 if (group_sched_in(event
, cpuctx
, ctx
))
1372 ctx_sched_in(struct perf_event_context
*ctx
,
1373 struct perf_cpu_context
*cpuctx
,
1374 enum event_type_t event_type
)
1376 raw_spin_lock(&ctx
->lock
);
1378 if (likely(!ctx
->nr_events
))
1381 ctx
->timestamp
= perf_clock();
1384 * First go through the list and put on any pinned groups
1385 * in order to give them the best chance of going on.
1387 if (event_type
& EVENT_PINNED
)
1388 ctx_pinned_sched_in(ctx
, cpuctx
);
1390 /* Then walk through the lower prio flexible groups */
1391 if (event_type
& EVENT_FLEXIBLE
)
1392 ctx_flexible_sched_in(ctx
, cpuctx
);
1395 raw_spin_unlock(&ctx
->lock
);
1398 static void cpu_ctx_sched_in(struct perf_cpu_context
*cpuctx
,
1399 enum event_type_t event_type
)
1401 struct perf_event_context
*ctx
= &cpuctx
->ctx
;
1403 ctx_sched_in(ctx
, cpuctx
, event_type
);
1406 static void task_ctx_sched_in(struct perf_event_context
*ctx
,
1407 enum event_type_t event_type
)
1409 struct perf_cpu_context
*cpuctx
;
1411 cpuctx
= __get_cpu_context(ctx
);
1412 if (cpuctx
->task_ctx
== ctx
)
1415 ctx_sched_in(ctx
, cpuctx
, event_type
);
1416 cpuctx
->task_ctx
= ctx
;
1419 void perf_event_context_sched_in(struct perf_event_context
*ctx
)
1421 struct perf_cpu_context
*cpuctx
;
1423 cpuctx
= __get_cpu_context(ctx
);
1424 if (cpuctx
->task_ctx
== ctx
)
1427 perf_pmu_disable(ctx
->pmu
);
1429 * We want to keep the following priority order:
1430 * cpu pinned (that don't need to move), task pinned,
1431 * cpu flexible, task flexible.
1433 cpu_ctx_sched_out(cpuctx
, EVENT_FLEXIBLE
);
1435 ctx_sched_in(ctx
, cpuctx
, EVENT_PINNED
);
1436 cpu_ctx_sched_in(cpuctx
, EVENT_FLEXIBLE
);
1437 ctx_sched_in(ctx
, cpuctx
, EVENT_FLEXIBLE
);
1439 cpuctx
->task_ctx
= ctx
;
1442 * Since these rotations are per-cpu, we need to ensure the
1443 * cpu-context we got scheduled on is actually rotating.
1445 perf_pmu_rotate_start(ctx
->pmu
);
1446 perf_pmu_enable(ctx
->pmu
);
1450 * Called from scheduler to add the events of the current task
1451 * with interrupts disabled.
1453 * We restore the event value and then enable it.
1455 * This does not protect us against NMI, but enable()
1456 * sets the enabled bit in the control field of event _before_
1457 * accessing the event control register. If a NMI hits, then it will
1458 * keep the event running.
1460 void __perf_event_task_sched_in(struct task_struct
*task
)
1462 struct perf_event_context
*ctx
;
1465 for_each_task_context_nr(ctxn
) {
1466 ctx
= task
->perf_event_ctxp
[ctxn
];
1470 perf_event_context_sched_in(ctx
);
1474 #define MAX_INTERRUPTS (~0ULL)
1476 static void perf_log_throttle(struct perf_event
*event
, int enable
);
1478 static u64
perf_calculate_period(struct perf_event
*event
, u64 nsec
, u64 count
)
1480 u64 frequency
= event
->attr
.sample_freq
;
1481 u64 sec
= NSEC_PER_SEC
;
1482 u64 divisor
, dividend
;
1484 int count_fls
, nsec_fls
, frequency_fls
, sec_fls
;
1486 count_fls
= fls64(count
);
1487 nsec_fls
= fls64(nsec
);
1488 frequency_fls
= fls64(frequency
);
1492 * We got @count in @nsec, with a target of sample_freq HZ
1493 * the target period becomes:
1496 * period = -------------------
1497 * @nsec * sample_freq
1502 * Reduce accuracy by one bit such that @a and @b converge
1503 * to a similar magnitude.
1505 #define REDUCE_FLS(a, b) \
1507 if (a##_fls > b##_fls) { \
1517 * Reduce accuracy until either term fits in a u64, then proceed with
1518 * the other, so that finally we can do a u64/u64 division.
1520 while (count_fls
+ sec_fls
> 64 && nsec_fls
+ frequency_fls
> 64) {
1521 REDUCE_FLS(nsec
, frequency
);
1522 REDUCE_FLS(sec
, count
);
1525 if (count_fls
+ sec_fls
> 64) {
1526 divisor
= nsec
* frequency
;
1528 while (count_fls
+ sec_fls
> 64) {
1529 REDUCE_FLS(count
, sec
);
1533 dividend
= count
* sec
;
1535 dividend
= count
* sec
;
1537 while (nsec_fls
+ frequency_fls
> 64) {
1538 REDUCE_FLS(nsec
, frequency
);
1542 divisor
= nsec
* frequency
;
1548 return div64_u64(dividend
, divisor
);
1551 static void perf_adjust_period(struct perf_event
*event
, u64 nsec
, u64 count
)
1553 struct hw_perf_event
*hwc
= &event
->hw
;
1554 s64 period
, sample_period
;
1557 period
= perf_calculate_period(event
, nsec
, count
);
1559 delta
= (s64
)(period
- hwc
->sample_period
);
1560 delta
= (delta
+ 7) / 8; /* low pass filter */
1562 sample_period
= hwc
->sample_period
+ delta
;
1567 hwc
->sample_period
= sample_period
;
1569 if (local64_read(&hwc
->period_left
) > 8*sample_period
) {
1570 event
->pmu
->stop(event
, PERF_EF_UPDATE
);
1571 local64_set(&hwc
->period_left
, 0);
1572 event
->pmu
->start(event
, PERF_EF_RELOAD
);
1576 static void perf_ctx_adjust_freq(struct perf_event_context
*ctx
, u64 period
)
1578 struct perf_event
*event
;
1579 struct hw_perf_event
*hwc
;
1580 u64 interrupts
, now
;
1583 raw_spin_lock(&ctx
->lock
);
1584 list_for_each_entry_rcu(event
, &ctx
->event_list
, event_entry
) {
1585 if (event
->state
!= PERF_EVENT_STATE_ACTIVE
)
1588 if (event
->cpu
!= -1 && event
->cpu
!= smp_processor_id())
1593 interrupts
= hwc
->interrupts
;
1594 hwc
->interrupts
= 0;
1597 * unthrottle events on the tick
1599 if (interrupts
== MAX_INTERRUPTS
) {
1600 perf_log_throttle(event
, 1);
1601 event
->pmu
->start(event
, 0);
1604 if (!event
->attr
.freq
|| !event
->attr
.sample_freq
)
1607 event
->pmu
->read(event
);
1608 now
= local64_read(&event
->count
);
1609 delta
= now
- hwc
->freq_count_stamp
;
1610 hwc
->freq_count_stamp
= now
;
1613 perf_adjust_period(event
, period
, delta
);
1615 raw_spin_unlock(&ctx
->lock
);
1619 * Round-robin a context's events:
1621 static void rotate_ctx(struct perf_event_context
*ctx
)
1623 raw_spin_lock(&ctx
->lock
);
1626 * Rotate the first entry last of non-pinned groups. Rotation might be
1627 * disabled by the inheritance code.
1629 if (!ctx
->rotate_disable
)
1630 list_rotate_left(&ctx
->flexible_groups
);
1632 raw_spin_unlock(&ctx
->lock
);
1636 * perf_pmu_rotate_start() and perf_rotate_context() are fully serialized
1637 * because they're strictly cpu affine and rotate_start is called with IRQs
1638 * disabled, while rotate_context is called from IRQ context.
1640 static void perf_rotate_context(struct perf_cpu_context
*cpuctx
)
1642 u64 interval
= (u64
)cpuctx
->jiffies_interval
* TICK_NSEC
;
1643 struct perf_event_context
*ctx
= NULL
;
1644 int rotate
= 0, remove
= 1;
1646 if (cpuctx
->ctx
.nr_events
) {
1648 if (cpuctx
->ctx
.nr_events
!= cpuctx
->ctx
.nr_active
)
1652 ctx
= cpuctx
->task_ctx
;
1653 if (ctx
&& ctx
->nr_events
) {
1655 if (ctx
->nr_events
!= ctx
->nr_active
)
1659 perf_pmu_disable(cpuctx
->ctx
.pmu
);
1660 perf_ctx_adjust_freq(&cpuctx
->ctx
, interval
);
1662 perf_ctx_adjust_freq(ctx
, interval
);
1667 cpu_ctx_sched_out(cpuctx
, EVENT_FLEXIBLE
);
1669 task_ctx_sched_out(ctx
, EVENT_FLEXIBLE
);
1671 rotate_ctx(&cpuctx
->ctx
);
1675 cpu_ctx_sched_in(cpuctx
, EVENT_FLEXIBLE
);
1677 task_ctx_sched_in(ctx
, EVENT_FLEXIBLE
);
1681 list_del_init(&cpuctx
->rotation_list
);
1683 perf_pmu_enable(cpuctx
->ctx
.pmu
);
1686 void perf_event_task_tick(void)
1688 struct list_head
*head
= &__get_cpu_var(rotation_list
);
1689 struct perf_cpu_context
*cpuctx
, *tmp
;
1691 WARN_ON(!irqs_disabled());
1693 list_for_each_entry_safe(cpuctx
, tmp
, head
, rotation_list
) {
1694 if (cpuctx
->jiffies_interval
== 1 ||
1695 !(jiffies
% cpuctx
->jiffies_interval
))
1696 perf_rotate_context(cpuctx
);
1700 static int event_enable_on_exec(struct perf_event
*event
,
1701 struct perf_event_context
*ctx
)
1703 if (!event
->attr
.enable_on_exec
)
1706 event
->attr
.enable_on_exec
= 0;
1707 if (event
->state
>= PERF_EVENT_STATE_INACTIVE
)
1710 __perf_event_mark_enabled(event
, ctx
);
1716 * Enable all of a task's events that have been marked enable-on-exec.
1717 * This expects task == current.
1719 static void perf_event_enable_on_exec(struct perf_event_context
*ctx
)
1721 struct perf_event
*event
;
1722 unsigned long flags
;
1726 local_irq_save(flags
);
1727 if (!ctx
|| !ctx
->nr_events
)
1730 task_ctx_sched_out(ctx
, EVENT_ALL
);
1732 raw_spin_lock(&ctx
->lock
);
1734 list_for_each_entry(event
, &ctx
->pinned_groups
, group_entry
) {
1735 ret
= event_enable_on_exec(event
, ctx
);
1740 list_for_each_entry(event
, &ctx
->flexible_groups
, group_entry
) {
1741 ret
= event_enable_on_exec(event
, ctx
);
1747 * Unclone this context if we enabled any event.
1752 raw_spin_unlock(&ctx
->lock
);
1754 perf_event_context_sched_in(ctx
);
1756 local_irq_restore(flags
);
1760 * Cross CPU call to read the hardware event
1762 static void __perf_event_read(void *info
)
1764 struct perf_event
*event
= info
;
1765 struct perf_event_context
*ctx
= event
->ctx
;
1766 struct perf_cpu_context
*cpuctx
= __get_cpu_context(ctx
);
1769 * If this is a task context, we need to check whether it is
1770 * the current task context of this cpu. If not it has been
1771 * scheduled out before the smp call arrived. In that case
1772 * event->count would have been updated to a recent sample
1773 * when the event was scheduled out.
1775 if (ctx
->task
&& cpuctx
->task_ctx
!= ctx
)
1778 raw_spin_lock(&ctx
->lock
);
1779 update_context_time(ctx
);
1780 update_event_times(event
);
1781 raw_spin_unlock(&ctx
->lock
);
1783 event
->pmu
->read(event
);
1786 static inline u64
perf_event_count(struct perf_event
*event
)
1788 return local64_read(&event
->count
) + atomic64_read(&event
->child_count
);
1791 static u64
perf_event_read(struct perf_event
*event
)
1794 * If event is enabled and currently active on a CPU, update the
1795 * value in the event structure:
1797 if (event
->state
== PERF_EVENT_STATE_ACTIVE
) {
1798 smp_call_function_single(event
->oncpu
,
1799 __perf_event_read
, event
, 1);
1800 } else if (event
->state
== PERF_EVENT_STATE_INACTIVE
) {
1801 struct perf_event_context
*ctx
= event
->ctx
;
1802 unsigned long flags
;
1804 raw_spin_lock_irqsave(&ctx
->lock
, flags
);
1806 * may read while context is not active
1807 * (e.g., thread is blocked), in that case
1808 * we cannot update context time
1811 update_context_time(ctx
);
1812 update_event_times(event
);
1813 raw_spin_unlock_irqrestore(&ctx
->lock
, flags
);
1816 return perf_event_count(event
);
1823 struct callchain_cpus_entries
{
1824 struct rcu_head rcu_head
;
1825 struct perf_callchain_entry
*cpu_entries
[0];
1828 static DEFINE_PER_CPU(int, callchain_recursion
[PERF_NR_CONTEXTS
]);
1829 static atomic_t nr_callchain_events
;
1830 static DEFINE_MUTEX(callchain_mutex
);
1831 struct callchain_cpus_entries
*callchain_cpus_entries
;
1834 __weak
void perf_callchain_kernel(struct perf_callchain_entry
*entry
,
1835 struct pt_regs
*regs
)
1839 __weak
void perf_callchain_user(struct perf_callchain_entry
*entry
,
1840 struct pt_regs
*regs
)
1844 static void release_callchain_buffers_rcu(struct rcu_head
*head
)
1846 struct callchain_cpus_entries
*entries
;
1849 entries
= container_of(head
, struct callchain_cpus_entries
, rcu_head
);
1851 for_each_possible_cpu(cpu
)
1852 kfree(entries
->cpu_entries
[cpu
]);
1857 static void release_callchain_buffers(void)
1859 struct callchain_cpus_entries
*entries
;
1861 entries
= callchain_cpus_entries
;
1862 rcu_assign_pointer(callchain_cpus_entries
, NULL
);
1863 call_rcu(&entries
->rcu_head
, release_callchain_buffers_rcu
);
1866 static int alloc_callchain_buffers(void)
1870 struct callchain_cpus_entries
*entries
;
1873 * We can't use the percpu allocation API for data that can be
1874 * accessed from NMI. Use a temporary manual per cpu allocation
1875 * until that gets sorted out.
1877 size
= sizeof(*entries
) + sizeof(struct perf_callchain_entry
*) *
1878 num_possible_cpus();
1880 entries
= kzalloc(size
, GFP_KERNEL
);
1884 size
= sizeof(struct perf_callchain_entry
) * PERF_NR_CONTEXTS
;
1886 for_each_possible_cpu(cpu
) {
1887 entries
->cpu_entries
[cpu
] = kmalloc_node(size
, GFP_KERNEL
,
1889 if (!entries
->cpu_entries
[cpu
])
1893 rcu_assign_pointer(callchain_cpus_entries
, entries
);
1898 for_each_possible_cpu(cpu
)
1899 kfree(entries
->cpu_entries
[cpu
]);
1905 static int get_callchain_buffers(void)
1910 mutex_lock(&callchain_mutex
);
1912 count
= atomic_inc_return(&nr_callchain_events
);
1913 if (WARN_ON_ONCE(count
< 1)) {
1919 /* If the allocation failed, give up */
1920 if (!callchain_cpus_entries
)
1925 err
= alloc_callchain_buffers();
1927 release_callchain_buffers();
1929 mutex_unlock(&callchain_mutex
);
1934 static void put_callchain_buffers(void)
1936 if (atomic_dec_and_mutex_lock(&nr_callchain_events
, &callchain_mutex
)) {
1937 release_callchain_buffers();
1938 mutex_unlock(&callchain_mutex
);
1942 static int get_recursion_context(int *recursion
)
1950 else if (in_softirq())
1955 if (recursion
[rctx
])
1964 static inline void put_recursion_context(int *recursion
, int rctx
)
1970 static struct perf_callchain_entry
*get_callchain_entry(int *rctx
)
1973 struct callchain_cpus_entries
*entries
;
1975 *rctx
= get_recursion_context(__get_cpu_var(callchain_recursion
));
1979 entries
= rcu_dereference(callchain_cpus_entries
);
1983 cpu
= smp_processor_id();
1985 return &entries
->cpu_entries
[cpu
][*rctx
];
1989 put_callchain_entry(int rctx
)
1991 put_recursion_context(__get_cpu_var(callchain_recursion
), rctx
);
1994 static struct perf_callchain_entry
*perf_callchain(struct pt_regs
*regs
)
1997 struct perf_callchain_entry
*entry
;
2000 entry
= get_callchain_entry(&rctx
);
2009 if (!user_mode(regs
)) {
2010 perf_callchain_store(entry
, PERF_CONTEXT_KERNEL
);
2011 perf_callchain_kernel(entry
, regs
);
2013 regs
= task_pt_regs(current
);
2019 perf_callchain_store(entry
, PERF_CONTEXT_USER
);
2020 perf_callchain_user(entry
, regs
);
2024 put_callchain_entry(rctx
);
2030 * Initialize the perf_event context in a task_struct:
2032 static void __perf_event_init_context(struct perf_event_context
*ctx
)
2034 raw_spin_lock_init(&ctx
->lock
);
2035 mutex_init(&ctx
->mutex
);
2036 INIT_LIST_HEAD(&ctx
->pinned_groups
);
2037 INIT_LIST_HEAD(&ctx
->flexible_groups
);
2038 INIT_LIST_HEAD(&ctx
->event_list
);
2039 atomic_set(&ctx
->refcount
, 1);
2042 static struct perf_event_context
*
2043 alloc_perf_context(struct pmu
*pmu
, struct task_struct
*task
)
2045 struct perf_event_context
*ctx
;
2047 ctx
= kzalloc(sizeof(struct perf_event_context
), GFP_KERNEL
);
2051 __perf_event_init_context(ctx
);
2054 get_task_struct(task
);
2061 static struct task_struct
*
2062 find_lively_task_by_vpid(pid_t vpid
)
2064 struct task_struct
*task
;
2071 task
= find_task_by_vpid(vpid
);
2073 get_task_struct(task
);
2077 return ERR_PTR(-ESRCH
);
2080 * Can't attach events to a dying task.
2083 if (task
->flags
& PF_EXITING
)
2086 /* Reuse ptrace permission checks for now. */
2088 if (!ptrace_may_access(task
, PTRACE_MODE_READ
))
2093 put_task_struct(task
);
2094 return ERR_PTR(err
);
2098 static struct perf_event_context
*
2099 find_get_context(struct pmu
*pmu
, struct task_struct
*task
, int cpu
)
2101 struct perf_event_context
*ctx
;
2102 struct perf_cpu_context
*cpuctx
;
2103 unsigned long flags
;
2106 if (!task
&& cpu
!= -1) {
2107 /* Must be root to operate on a CPU event: */
2108 if (perf_paranoid_cpu() && !capable(CAP_SYS_ADMIN
))
2109 return ERR_PTR(-EACCES
);
2111 if (cpu
< 0 || cpu
>= nr_cpumask_bits
)
2112 return ERR_PTR(-EINVAL
);
2115 * We could be clever and allow to attach a event to an
2116 * offline CPU and activate it when the CPU comes up, but
2119 if (!cpu_online(cpu
))
2120 return ERR_PTR(-ENODEV
);
2122 cpuctx
= per_cpu_ptr(pmu
->pmu_cpu_context
, cpu
);
2130 ctxn
= pmu
->task_ctx_nr
;
2135 ctx
= perf_lock_task_context(task
, ctxn
, &flags
);
2138 raw_spin_unlock_irqrestore(&ctx
->lock
, flags
);
2142 ctx
= alloc_perf_context(pmu
, task
);
2149 if (cmpxchg(&task
->perf_event_ctxp
[ctxn
], NULL
, ctx
)) {
2151 * We raced with some other task; use
2152 * the context they set.
2154 put_task_struct(task
);
2163 return ERR_PTR(err
);
2166 static void perf_event_free_filter(struct perf_event
*event
);
2168 static void free_event_rcu(struct rcu_head
*head
)
2170 struct perf_event
*event
;
2172 event
= container_of(head
, struct perf_event
, rcu_head
);
2174 put_pid_ns(event
->ns
);
2175 perf_event_free_filter(event
);
2179 static void perf_buffer_put(struct perf_buffer
*buffer
);
2181 static void free_event(struct perf_event
*event
)
2183 irq_work_sync(&event
->pending
);
2185 if (!event
->parent
) {
2186 if (event
->attach_state
& PERF_ATTACH_TASK
)
2187 jump_label_dec(&perf_task_events
);
2188 if (event
->attr
.mmap
|| event
->attr
.mmap_data
)
2189 atomic_dec(&nr_mmap_events
);
2190 if (event
->attr
.comm
)
2191 atomic_dec(&nr_comm_events
);
2192 if (event
->attr
.task
)
2193 atomic_dec(&nr_task_events
);
2194 if (event
->attr
.sample_type
& PERF_SAMPLE_CALLCHAIN
)
2195 put_callchain_buffers();
2198 if (event
->buffer
) {
2199 perf_buffer_put(event
->buffer
);
2200 event
->buffer
= NULL
;
2204 event
->destroy(event
);
2207 put_ctx(event
->ctx
);
2209 call_rcu(&event
->rcu_head
, free_event_rcu
);
2212 int perf_event_release_kernel(struct perf_event
*event
)
2214 struct perf_event_context
*ctx
= event
->ctx
;
2217 * Remove from the PMU, can't get re-enabled since we got
2218 * here because the last ref went.
2220 perf_event_disable(event
);
2222 WARN_ON_ONCE(ctx
->parent_ctx
);
2224 * There are two ways this annotation is useful:
2226 * 1) there is a lock recursion from perf_event_exit_task
2227 * see the comment there.
2229 * 2) there is a lock-inversion with mmap_sem through
2230 * perf_event_read_group(), which takes faults while
2231 * holding ctx->mutex, however this is called after
2232 * the last filedesc died, so there is no possibility
2233 * to trigger the AB-BA case.
2235 mutex_lock_nested(&ctx
->mutex
, SINGLE_DEPTH_NESTING
);
2236 raw_spin_lock_irq(&ctx
->lock
);
2237 perf_group_detach(event
);
2238 list_del_event(event
, ctx
);
2239 raw_spin_unlock_irq(&ctx
->lock
);
2240 mutex_unlock(&ctx
->mutex
);
2246 EXPORT_SYMBOL_GPL(perf_event_release_kernel
);
2249 * Called when the last reference to the file is gone.
2251 static int perf_release(struct inode
*inode
, struct file
*file
)
2253 struct perf_event
*event
= file
->private_data
;
2254 struct task_struct
*owner
;
2256 file
->private_data
= NULL
;
2259 owner
= ACCESS_ONCE(event
->owner
);
2261 * Matches the smp_wmb() in perf_event_exit_task(). If we observe
2262 * !owner it means the list deletion is complete and we can indeed
2263 * free this event, otherwise we need to serialize on
2264 * owner->perf_event_mutex.
2266 smp_read_barrier_depends();
2269 * Since delayed_put_task_struct() also drops the last
2270 * task reference we can safely take a new reference
2271 * while holding the rcu_read_lock().
2273 get_task_struct(owner
);
2278 mutex_lock(&owner
->perf_event_mutex
);
2280 * We have to re-check the event->owner field, if it is cleared
2281 * we raced with perf_event_exit_task(), acquiring the mutex
2282 * ensured they're done, and we can proceed with freeing the
2286 list_del_init(&event
->owner_entry
);
2287 mutex_unlock(&owner
->perf_event_mutex
);
2288 put_task_struct(owner
);
2291 return perf_event_release_kernel(event
);
2294 static int perf_event_read_size(struct perf_event
*event
)
2296 int entry
= sizeof(u64
); /* value */
2300 if (event
->attr
.read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
)
2301 size
+= sizeof(u64
);
2303 if (event
->attr
.read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
)
2304 size
+= sizeof(u64
);
2306 if (event
->attr
.read_format
& PERF_FORMAT_ID
)
2307 entry
+= sizeof(u64
);
2309 if (event
->attr
.read_format
& PERF_FORMAT_GROUP
) {
2310 nr
+= event
->group_leader
->nr_siblings
;
2311 size
+= sizeof(u64
);
2319 u64
perf_event_read_value(struct perf_event
*event
, u64
*enabled
, u64
*running
)
2321 struct perf_event
*child
;
2327 mutex_lock(&event
->child_mutex
);
2328 total
+= perf_event_read(event
);
2329 *enabled
+= event
->total_time_enabled
+
2330 atomic64_read(&event
->child_total_time_enabled
);
2331 *running
+= event
->total_time_running
+
2332 atomic64_read(&event
->child_total_time_running
);
2334 list_for_each_entry(child
, &event
->child_list
, child_list
) {
2335 total
+= perf_event_read(child
);
2336 *enabled
+= child
->total_time_enabled
;
2337 *running
+= child
->total_time_running
;
2339 mutex_unlock(&event
->child_mutex
);
2343 EXPORT_SYMBOL_GPL(perf_event_read_value
);
2345 static int perf_event_read_group(struct perf_event
*event
,
2346 u64 read_format
, char __user
*buf
)
2348 struct perf_event
*leader
= event
->group_leader
, *sub
;
2349 int n
= 0, size
= 0, ret
= -EFAULT
;
2350 struct perf_event_context
*ctx
= leader
->ctx
;
2352 u64 count
, enabled
, running
;
2354 mutex_lock(&ctx
->mutex
);
2355 count
= perf_event_read_value(leader
, &enabled
, &running
);
2357 values
[n
++] = 1 + leader
->nr_siblings
;
2358 if (read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
)
2359 values
[n
++] = enabled
;
2360 if (read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
)
2361 values
[n
++] = running
;
2362 values
[n
++] = count
;
2363 if (read_format
& PERF_FORMAT_ID
)
2364 values
[n
++] = primary_event_id(leader
);
2366 size
= n
* sizeof(u64
);
2368 if (copy_to_user(buf
, values
, size
))
2373 list_for_each_entry(sub
, &leader
->sibling_list
, group_entry
) {
2376 values
[n
++] = perf_event_read_value(sub
, &enabled
, &running
);
2377 if (read_format
& PERF_FORMAT_ID
)
2378 values
[n
++] = primary_event_id(sub
);
2380 size
= n
* sizeof(u64
);
2382 if (copy_to_user(buf
+ ret
, values
, size
)) {
2390 mutex_unlock(&ctx
->mutex
);
2395 static int perf_event_read_one(struct perf_event
*event
,
2396 u64 read_format
, char __user
*buf
)
2398 u64 enabled
, running
;
2402 values
[n
++] = perf_event_read_value(event
, &enabled
, &running
);
2403 if (read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
)
2404 values
[n
++] = enabled
;
2405 if (read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
)
2406 values
[n
++] = running
;
2407 if (read_format
& PERF_FORMAT_ID
)
2408 values
[n
++] = primary_event_id(event
);
2410 if (copy_to_user(buf
, values
, n
* sizeof(u64
)))
2413 return n
* sizeof(u64
);
2417 * Read the performance event - simple non blocking version for now
2420 perf_read_hw(struct perf_event
*event
, char __user
*buf
, size_t count
)
2422 u64 read_format
= event
->attr
.read_format
;
2426 * Return end-of-file for a read on a event that is in
2427 * error state (i.e. because it was pinned but it couldn't be
2428 * scheduled on to the CPU at some point).
2430 if (event
->state
== PERF_EVENT_STATE_ERROR
)
2433 if (count
< perf_event_read_size(event
))
2436 WARN_ON_ONCE(event
->ctx
->parent_ctx
);
2437 if (read_format
& PERF_FORMAT_GROUP
)
2438 ret
= perf_event_read_group(event
, read_format
, buf
);
2440 ret
= perf_event_read_one(event
, read_format
, buf
);
2446 perf_read(struct file
*file
, char __user
*buf
, size_t count
, loff_t
*ppos
)
2448 struct perf_event
*event
= file
->private_data
;
2450 return perf_read_hw(event
, buf
, count
);
2453 static unsigned int perf_poll(struct file
*file
, poll_table
*wait
)
2455 struct perf_event
*event
= file
->private_data
;
2456 struct perf_buffer
*buffer
;
2457 unsigned int events
= POLL_HUP
;
2460 buffer
= rcu_dereference(event
->buffer
);
2462 events
= atomic_xchg(&buffer
->poll
, 0);
2465 poll_wait(file
, &event
->waitq
, wait
);
2470 static void perf_event_reset(struct perf_event
*event
)
2472 (void)perf_event_read(event
);
2473 local64_set(&event
->count
, 0);
2474 perf_event_update_userpage(event
);
2478 * Holding the top-level event's child_mutex means that any
2479 * descendant process that has inherited this event will block
2480 * in sync_child_event if it goes to exit, thus satisfying the
2481 * task existence requirements of perf_event_enable/disable.
2483 static void perf_event_for_each_child(struct perf_event
*event
,
2484 void (*func
)(struct perf_event
*))
2486 struct perf_event
*child
;
2488 WARN_ON_ONCE(event
->ctx
->parent_ctx
);
2489 mutex_lock(&event
->child_mutex
);
2491 list_for_each_entry(child
, &event
->child_list
, child_list
)
2493 mutex_unlock(&event
->child_mutex
);
2496 static void perf_event_for_each(struct perf_event
*event
,
2497 void (*func
)(struct perf_event
*))
2499 struct perf_event_context
*ctx
= event
->ctx
;
2500 struct perf_event
*sibling
;
2502 WARN_ON_ONCE(ctx
->parent_ctx
);
2503 mutex_lock(&ctx
->mutex
);
2504 event
= event
->group_leader
;
2506 perf_event_for_each_child(event
, func
);
2508 list_for_each_entry(sibling
, &event
->sibling_list
, group_entry
)
2509 perf_event_for_each_child(event
, func
);
2510 mutex_unlock(&ctx
->mutex
);
2513 static int perf_event_period(struct perf_event
*event
, u64 __user
*arg
)
2515 struct perf_event_context
*ctx
= event
->ctx
;
2519 if (!event
->attr
.sample_period
)
2522 if (copy_from_user(&value
, arg
, sizeof(value
)))
2528 raw_spin_lock_irq(&ctx
->lock
);
2529 if (event
->attr
.freq
) {
2530 if (value
> sysctl_perf_event_sample_rate
) {
2535 event
->attr
.sample_freq
= value
;
2537 event
->attr
.sample_period
= value
;
2538 event
->hw
.sample_period
= value
;
2541 raw_spin_unlock_irq(&ctx
->lock
);
2546 static const struct file_operations perf_fops
;
2548 static struct perf_event
*perf_fget_light(int fd
, int *fput_needed
)
2552 file
= fget_light(fd
, fput_needed
);
2554 return ERR_PTR(-EBADF
);
2556 if (file
->f_op
!= &perf_fops
) {
2557 fput_light(file
, *fput_needed
);
2559 return ERR_PTR(-EBADF
);
2562 return file
->private_data
;
2565 static int perf_event_set_output(struct perf_event
*event
,
2566 struct perf_event
*output_event
);
2567 static int perf_event_set_filter(struct perf_event
*event
, void __user
*arg
);
2569 static long perf_ioctl(struct file
*file
, unsigned int cmd
, unsigned long arg
)
2571 struct perf_event
*event
= file
->private_data
;
2572 void (*func
)(struct perf_event
*);
2576 case PERF_EVENT_IOC_ENABLE
:
2577 func
= perf_event_enable
;
2579 case PERF_EVENT_IOC_DISABLE
:
2580 func
= perf_event_disable
;
2582 case PERF_EVENT_IOC_RESET
:
2583 func
= perf_event_reset
;
2586 case PERF_EVENT_IOC_REFRESH
:
2587 return perf_event_refresh(event
, arg
);
2589 case PERF_EVENT_IOC_PERIOD
:
2590 return perf_event_period(event
, (u64 __user
*)arg
);
2592 case PERF_EVENT_IOC_SET_OUTPUT
:
2594 struct perf_event
*output_event
= NULL
;
2595 int fput_needed
= 0;
2599 output_event
= perf_fget_light(arg
, &fput_needed
);
2600 if (IS_ERR(output_event
))
2601 return PTR_ERR(output_event
);
2604 ret
= perf_event_set_output(event
, output_event
);
2606 fput_light(output_event
->filp
, fput_needed
);
2611 case PERF_EVENT_IOC_SET_FILTER
:
2612 return perf_event_set_filter(event
, (void __user
*)arg
);
2618 if (flags
& PERF_IOC_FLAG_GROUP
)
2619 perf_event_for_each(event
, func
);
2621 perf_event_for_each_child(event
, func
);
2626 int perf_event_task_enable(void)
2628 struct perf_event
*event
;
2630 mutex_lock(¤t
->perf_event_mutex
);
2631 list_for_each_entry(event
, ¤t
->perf_event_list
, owner_entry
)
2632 perf_event_for_each_child(event
, perf_event_enable
);
2633 mutex_unlock(¤t
->perf_event_mutex
);
2638 int perf_event_task_disable(void)
2640 struct perf_event
*event
;
2642 mutex_lock(¤t
->perf_event_mutex
);
2643 list_for_each_entry(event
, ¤t
->perf_event_list
, owner_entry
)
2644 perf_event_for_each_child(event
, perf_event_disable
);
2645 mutex_unlock(¤t
->perf_event_mutex
);
2650 #ifndef PERF_EVENT_INDEX_OFFSET
2651 # define PERF_EVENT_INDEX_OFFSET 0
2654 static int perf_event_index(struct perf_event
*event
)
2656 if (event
->hw
.state
& PERF_HES_STOPPED
)
2659 if (event
->state
!= PERF_EVENT_STATE_ACTIVE
)
2662 return event
->hw
.idx
+ 1 - PERF_EVENT_INDEX_OFFSET
;
2666 * Callers need to ensure there can be no nesting of this function, otherwise
2667 * the seqlock logic goes bad. We can not serialize this because the arch
2668 * code calls this from NMI context.
2670 void perf_event_update_userpage(struct perf_event
*event
)
2672 struct perf_event_mmap_page
*userpg
;
2673 struct perf_buffer
*buffer
;
2676 buffer
= rcu_dereference(event
->buffer
);
2680 userpg
= buffer
->user_page
;
2683 * Disable preemption so as to not let the corresponding user-space
2684 * spin too long if we get preempted.
2689 userpg
->index
= perf_event_index(event
);
2690 userpg
->offset
= perf_event_count(event
);
2691 if (event
->state
== PERF_EVENT_STATE_ACTIVE
)
2692 userpg
->offset
-= local64_read(&event
->hw
.prev_count
);
2694 userpg
->time_enabled
= event
->total_time_enabled
+
2695 atomic64_read(&event
->child_total_time_enabled
);
2697 userpg
->time_running
= event
->total_time_running
+
2698 atomic64_read(&event
->child_total_time_running
);
2707 static unsigned long perf_data_size(struct perf_buffer
*buffer
);
2710 perf_buffer_init(struct perf_buffer
*buffer
, long watermark
, int flags
)
2712 long max_size
= perf_data_size(buffer
);
2715 buffer
->watermark
= min(max_size
, watermark
);
2717 if (!buffer
->watermark
)
2718 buffer
->watermark
= max_size
/ 2;
2720 if (flags
& PERF_BUFFER_WRITABLE
)
2721 buffer
->writable
= 1;
2723 atomic_set(&buffer
->refcount
, 1);
2726 #ifndef CONFIG_PERF_USE_VMALLOC
2729 * Back perf_mmap() with regular GFP_KERNEL-0 pages.
2732 static struct page
*
2733 perf_mmap_to_page(struct perf_buffer
*buffer
, unsigned long pgoff
)
2735 if (pgoff
> buffer
->nr_pages
)
2739 return virt_to_page(buffer
->user_page
);
2741 return virt_to_page(buffer
->data_pages
[pgoff
- 1]);
2744 static void *perf_mmap_alloc_page(int cpu
)
2749 node
= (cpu
== -1) ? cpu
: cpu_to_node(cpu
);
2750 page
= alloc_pages_node(node
, GFP_KERNEL
| __GFP_ZERO
, 0);
2754 return page_address(page
);
2757 static struct perf_buffer
*
2758 perf_buffer_alloc(int nr_pages
, long watermark
, int cpu
, int flags
)
2760 struct perf_buffer
*buffer
;
2764 size
= sizeof(struct perf_buffer
);
2765 size
+= nr_pages
* sizeof(void *);
2767 buffer
= kzalloc(size
, GFP_KERNEL
);
2771 buffer
->user_page
= perf_mmap_alloc_page(cpu
);
2772 if (!buffer
->user_page
)
2773 goto fail_user_page
;
2775 for (i
= 0; i
< nr_pages
; i
++) {
2776 buffer
->data_pages
[i
] = perf_mmap_alloc_page(cpu
);
2777 if (!buffer
->data_pages
[i
])
2778 goto fail_data_pages
;
2781 buffer
->nr_pages
= nr_pages
;
2783 perf_buffer_init(buffer
, watermark
, flags
);
2788 for (i
--; i
>= 0; i
--)
2789 free_page((unsigned long)buffer
->data_pages
[i
]);
2791 free_page((unsigned long)buffer
->user_page
);
2800 static void perf_mmap_free_page(unsigned long addr
)
2802 struct page
*page
= virt_to_page((void *)addr
);
2804 page
->mapping
= NULL
;
2808 static void perf_buffer_free(struct perf_buffer
*buffer
)
2812 perf_mmap_free_page((unsigned long)buffer
->user_page
);
2813 for (i
= 0; i
< buffer
->nr_pages
; i
++)
2814 perf_mmap_free_page((unsigned long)buffer
->data_pages
[i
]);
2818 static inline int page_order(struct perf_buffer
*buffer
)
2826 * Back perf_mmap() with vmalloc memory.
2828 * Required for architectures that have d-cache aliasing issues.
2831 static inline int page_order(struct perf_buffer
*buffer
)
2833 return buffer
->page_order
;
2836 static struct page
*
2837 perf_mmap_to_page(struct perf_buffer
*buffer
, unsigned long pgoff
)
2839 if (pgoff
> (1UL << page_order(buffer
)))
2842 return vmalloc_to_page((void *)buffer
->user_page
+ pgoff
* PAGE_SIZE
);
2845 static void perf_mmap_unmark_page(void *addr
)
2847 struct page
*page
= vmalloc_to_page(addr
);
2849 page
->mapping
= NULL
;
2852 static void perf_buffer_free_work(struct work_struct
*work
)
2854 struct perf_buffer
*buffer
;
2858 buffer
= container_of(work
, struct perf_buffer
, work
);
2859 nr
= 1 << page_order(buffer
);
2861 base
= buffer
->user_page
;
2862 for (i
= 0; i
< nr
+ 1; i
++)
2863 perf_mmap_unmark_page(base
+ (i
* PAGE_SIZE
));
2869 static void perf_buffer_free(struct perf_buffer
*buffer
)
2871 schedule_work(&buffer
->work
);
2874 static struct perf_buffer
*
2875 perf_buffer_alloc(int nr_pages
, long watermark
, int cpu
, int flags
)
2877 struct perf_buffer
*buffer
;
2881 size
= sizeof(struct perf_buffer
);
2882 size
+= sizeof(void *);
2884 buffer
= kzalloc(size
, GFP_KERNEL
);
2888 INIT_WORK(&buffer
->work
, perf_buffer_free_work
);
2890 all_buf
= vmalloc_user((nr_pages
+ 1) * PAGE_SIZE
);
2894 buffer
->user_page
= all_buf
;
2895 buffer
->data_pages
[0] = all_buf
+ PAGE_SIZE
;
2896 buffer
->page_order
= ilog2(nr_pages
);
2897 buffer
->nr_pages
= 1;
2899 perf_buffer_init(buffer
, watermark
, flags
);
2912 static unsigned long perf_data_size(struct perf_buffer
*buffer
)
2914 return buffer
->nr_pages
<< (PAGE_SHIFT
+ page_order(buffer
));
2917 static int perf_mmap_fault(struct vm_area_struct
*vma
, struct vm_fault
*vmf
)
2919 struct perf_event
*event
= vma
->vm_file
->private_data
;
2920 struct perf_buffer
*buffer
;
2921 int ret
= VM_FAULT_SIGBUS
;
2923 if (vmf
->flags
& FAULT_FLAG_MKWRITE
) {
2924 if (vmf
->pgoff
== 0)
2930 buffer
= rcu_dereference(event
->buffer
);
2934 if (vmf
->pgoff
&& (vmf
->flags
& FAULT_FLAG_WRITE
))
2937 vmf
->page
= perf_mmap_to_page(buffer
, vmf
->pgoff
);
2941 get_page(vmf
->page
);
2942 vmf
->page
->mapping
= vma
->vm_file
->f_mapping
;
2943 vmf
->page
->index
= vmf
->pgoff
;
2952 static void perf_buffer_free_rcu(struct rcu_head
*rcu_head
)
2954 struct perf_buffer
*buffer
;
2956 buffer
= container_of(rcu_head
, struct perf_buffer
, rcu_head
);
2957 perf_buffer_free(buffer
);
2960 static struct perf_buffer
*perf_buffer_get(struct perf_event
*event
)
2962 struct perf_buffer
*buffer
;
2965 buffer
= rcu_dereference(event
->buffer
);
2967 if (!atomic_inc_not_zero(&buffer
->refcount
))
2975 static void perf_buffer_put(struct perf_buffer
*buffer
)
2977 if (!atomic_dec_and_test(&buffer
->refcount
))
2980 call_rcu(&buffer
->rcu_head
, perf_buffer_free_rcu
);
2983 static void perf_mmap_open(struct vm_area_struct
*vma
)
2985 struct perf_event
*event
= vma
->vm_file
->private_data
;
2987 atomic_inc(&event
->mmap_count
);
2990 static void perf_mmap_close(struct vm_area_struct
*vma
)
2992 struct perf_event
*event
= vma
->vm_file
->private_data
;
2994 if (atomic_dec_and_mutex_lock(&event
->mmap_count
, &event
->mmap_mutex
)) {
2995 unsigned long size
= perf_data_size(event
->buffer
);
2996 struct user_struct
*user
= event
->mmap_user
;
2997 struct perf_buffer
*buffer
= event
->buffer
;
2999 atomic_long_sub((size
>> PAGE_SHIFT
) + 1, &user
->locked_vm
);
3000 vma
->vm_mm
->locked_vm
-= event
->mmap_locked
;
3001 rcu_assign_pointer(event
->buffer
, NULL
);
3002 mutex_unlock(&event
->mmap_mutex
);
3004 perf_buffer_put(buffer
);
3009 static const struct vm_operations_struct perf_mmap_vmops
= {
3010 .open
= perf_mmap_open
,
3011 .close
= perf_mmap_close
,
3012 .fault
= perf_mmap_fault
,
3013 .page_mkwrite
= perf_mmap_fault
,
3016 static int perf_mmap(struct file
*file
, struct vm_area_struct
*vma
)
3018 struct perf_event
*event
= file
->private_data
;
3019 unsigned long user_locked
, user_lock_limit
;
3020 struct user_struct
*user
= current_user();
3021 unsigned long locked
, lock_limit
;
3022 struct perf_buffer
*buffer
;
3023 unsigned long vma_size
;
3024 unsigned long nr_pages
;
3025 long user_extra
, extra
;
3026 int ret
= 0, flags
= 0;
3029 * Don't allow mmap() of inherited per-task counters. This would
3030 * create a performance issue due to all children writing to the
3033 if (event
->cpu
== -1 && event
->attr
.inherit
)
3036 if (!(vma
->vm_flags
& VM_SHARED
))
3039 vma_size
= vma
->vm_end
- vma
->vm_start
;
3040 nr_pages
= (vma_size
/ PAGE_SIZE
) - 1;
3043 * If we have buffer pages ensure they're a power-of-two number, so we
3044 * can do bitmasks instead of modulo.
3046 if (nr_pages
!= 0 && !is_power_of_2(nr_pages
))
3049 if (vma_size
!= PAGE_SIZE
* (1 + nr_pages
))
3052 if (vma
->vm_pgoff
!= 0)
3055 WARN_ON_ONCE(event
->ctx
->parent_ctx
);
3056 mutex_lock(&event
->mmap_mutex
);
3057 if (event
->buffer
) {
3058 if (event
->buffer
->nr_pages
== nr_pages
)
3059 atomic_inc(&event
->buffer
->refcount
);
3065 user_extra
= nr_pages
+ 1;
3066 user_lock_limit
= sysctl_perf_event_mlock
>> (PAGE_SHIFT
- 10);
3069 * Increase the limit linearly with more CPUs:
3071 user_lock_limit
*= num_online_cpus();
3073 user_locked
= atomic_long_read(&user
->locked_vm
) + user_extra
;
3076 if (user_locked
> user_lock_limit
)
3077 extra
= user_locked
- user_lock_limit
;
3079 lock_limit
= rlimit(RLIMIT_MEMLOCK
);
3080 lock_limit
>>= PAGE_SHIFT
;
3081 locked
= vma
->vm_mm
->locked_vm
+ extra
;
3083 if ((locked
> lock_limit
) && perf_paranoid_tracepoint_raw() &&
3084 !capable(CAP_IPC_LOCK
)) {
3089 WARN_ON(event
->buffer
);
3091 if (vma
->vm_flags
& VM_WRITE
)
3092 flags
|= PERF_BUFFER_WRITABLE
;
3094 buffer
= perf_buffer_alloc(nr_pages
, event
->attr
.wakeup_watermark
,
3100 rcu_assign_pointer(event
->buffer
, buffer
);
3102 atomic_long_add(user_extra
, &user
->locked_vm
);
3103 event
->mmap_locked
= extra
;
3104 event
->mmap_user
= get_current_user();
3105 vma
->vm_mm
->locked_vm
+= event
->mmap_locked
;
3109 atomic_inc(&event
->mmap_count
);
3110 mutex_unlock(&event
->mmap_mutex
);
3112 vma
->vm_flags
|= VM_RESERVED
;
3113 vma
->vm_ops
= &perf_mmap_vmops
;
3118 static int perf_fasync(int fd
, struct file
*filp
, int on
)
3120 struct inode
*inode
= filp
->f_path
.dentry
->d_inode
;
3121 struct perf_event
*event
= filp
->private_data
;
3124 mutex_lock(&inode
->i_mutex
);
3125 retval
= fasync_helper(fd
, filp
, on
, &event
->fasync
);
3126 mutex_unlock(&inode
->i_mutex
);
3134 static const struct file_operations perf_fops
= {
3135 .llseek
= no_llseek
,
3136 .release
= perf_release
,
3139 .unlocked_ioctl
= perf_ioctl
,
3140 .compat_ioctl
= perf_ioctl
,
3142 .fasync
= perf_fasync
,
3148 * If there's data, ensure we set the poll() state and publish everything
3149 * to user-space before waking everybody up.
3152 void perf_event_wakeup(struct perf_event
*event
)
3154 wake_up_all(&event
->waitq
);
3156 if (event
->pending_kill
) {
3157 kill_fasync(&event
->fasync
, SIGIO
, event
->pending_kill
);
3158 event
->pending_kill
= 0;
3162 static void perf_pending_event(struct irq_work
*entry
)
3164 struct perf_event
*event
= container_of(entry
,
3165 struct perf_event
, pending
);
3167 if (event
->pending_disable
) {
3168 event
->pending_disable
= 0;
3169 __perf_event_disable(event
);
3172 if (event
->pending_wakeup
) {
3173 event
->pending_wakeup
= 0;
3174 perf_event_wakeup(event
);
3179 * We assume there is only KVM supporting the callbacks.
3180 * Later on, we might change it to a list if there is
3181 * another virtualization implementation supporting the callbacks.
3183 struct perf_guest_info_callbacks
*perf_guest_cbs
;
3185 int perf_register_guest_info_callbacks(struct perf_guest_info_callbacks
*cbs
)
3187 perf_guest_cbs
= cbs
;
3190 EXPORT_SYMBOL_GPL(perf_register_guest_info_callbacks
);
3192 int perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks
*cbs
)
3194 perf_guest_cbs
= NULL
;
3197 EXPORT_SYMBOL_GPL(perf_unregister_guest_info_callbacks
);
3202 static bool perf_output_space(struct perf_buffer
*buffer
, unsigned long tail
,
3203 unsigned long offset
, unsigned long head
)
3207 if (!buffer
->writable
)
3210 mask
= perf_data_size(buffer
) - 1;
3212 offset
= (offset
- tail
) & mask
;
3213 head
= (head
- tail
) & mask
;
3215 if ((int)(head
- offset
) < 0)
3221 static void perf_output_wakeup(struct perf_output_handle
*handle
)
3223 atomic_set(&handle
->buffer
->poll
, POLL_IN
);
3226 handle
->event
->pending_wakeup
= 1;
3227 irq_work_queue(&handle
->event
->pending
);
3229 perf_event_wakeup(handle
->event
);
3233 * We need to ensure a later event_id doesn't publish a head when a former
3234 * event isn't done writing. However since we need to deal with NMIs we
3235 * cannot fully serialize things.
3237 * We only publish the head (and generate a wakeup) when the outer-most
3240 static void perf_output_get_handle(struct perf_output_handle
*handle
)
3242 struct perf_buffer
*buffer
= handle
->buffer
;
3245 local_inc(&buffer
->nest
);
3246 handle
->wakeup
= local_read(&buffer
->wakeup
);
3249 static void perf_output_put_handle(struct perf_output_handle
*handle
)
3251 struct perf_buffer
*buffer
= handle
->buffer
;
3255 head
= local_read(&buffer
->head
);
3258 * IRQ/NMI can happen here, which means we can miss a head update.
3261 if (!local_dec_and_test(&buffer
->nest
))
3265 * Publish the known good head. Rely on the full barrier implied
3266 * by atomic_dec_and_test() order the buffer->head read and this
3269 buffer
->user_page
->data_head
= head
;
3272 * Now check if we missed an update, rely on the (compiler)
3273 * barrier in atomic_dec_and_test() to re-read buffer->head.
3275 if (unlikely(head
!= local_read(&buffer
->head
))) {
3276 local_inc(&buffer
->nest
);
3280 if (handle
->wakeup
!= local_read(&buffer
->wakeup
))
3281 perf_output_wakeup(handle
);
3287 __always_inline
void perf_output_copy(struct perf_output_handle
*handle
,
3288 const void *buf
, unsigned int len
)
3291 unsigned long size
= min_t(unsigned long, handle
->size
, len
);
3293 memcpy(handle
->addr
, buf
, size
);
3296 handle
->addr
+= size
;
3298 handle
->size
-= size
;
3299 if (!handle
->size
) {
3300 struct perf_buffer
*buffer
= handle
->buffer
;
3303 handle
->page
&= buffer
->nr_pages
- 1;
3304 handle
->addr
= buffer
->data_pages
[handle
->page
];
3305 handle
->size
= PAGE_SIZE
<< page_order(buffer
);
3310 int perf_output_begin(struct perf_output_handle
*handle
,
3311 struct perf_event
*event
, unsigned int size
,
3312 int nmi
, int sample
)
3314 struct perf_buffer
*buffer
;
3315 unsigned long tail
, offset
, head
;
3318 struct perf_event_header header
;
3325 * For inherited events we send all the output towards the parent.
3328 event
= event
->parent
;
3330 buffer
= rcu_dereference(event
->buffer
);
3334 handle
->buffer
= buffer
;
3335 handle
->event
= event
;
3337 handle
->sample
= sample
;
3339 if (!buffer
->nr_pages
)
3342 have_lost
= local_read(&buffer
->lost
);
3344 size
+= sizeof(lost_event
);
3346 perf_output_get_handle(handle
);
3350 * Userspace could choose to issue a mb() before updating the
3351 * tail pointer. So that all reads will be completed before the
3354 tail
= ACCESS_ONCE(buffer
->user_page
->data_tail
);
3356 offset
= head
= local_read(&buffer
->head
);
3358 if (unlikely(!perf_output_space(buffer
, tail
, offset
, head
)))
3360 } while (local_cmpxchg(&buffer
->head
, offset
, head
) != offset
);
3362 if (head
- local_read(&buffer
->wakeup
) > buffer
->watermark
)
3363 local_add(buffer
->watermark
, &buffer
->wakeup
);
3365 handle
->page
= offset
>> (PAGE_SHIFT
+ page_order(buffer
));
3366 handle
->page
&= buffer
->nr_pages
- 1;
3367 handle
->size
= offset
& ((PAGE_SIZE
<< page_order(buffer
)) - 1);
3368 handle
->addr
= buffer
->data_pages
[handle
->page
];
3369 handle
->addr
+= handle
->size
;
3370 handle
->size
= (PAGE_SIZE
<< page_order(buffer
)) - handle
->size
;
3373 lost_event
.header
.type
= PERF_RECORD_LOST
;
3374 lost_event
.header
.misc
= 0;
3375 lost_event
.header
.size
= sizeof(lost_event
);
3376 lost_event
.id
= event
->id
;
3377 lost_event
.lost
= local_xchg(&buffer
->lost
, 0);
3379 perf_output_put(handle
, lost_event
);
3385 local_inc(&buffer
->lost
);
3386 perf_output_put_handle(handle
);
3393 void perf_output_end(struct perf_output_handle
*handle
)
3395 struct perf_event
*event
= handle
->event
;
3396 struct perf_buffer
*buffer
= handle
->buffer
;
3398 int wakeup_events
= event
->attr
.wakeup_events
;
3400 if (handle
->sample
&& wakeup_events
) {
3401 int events
= local_inc_return(&buffer
->events
);
3402 if (events
>= wakeup_events
) {
3403 local_sub(wakeup_events
, &buffer
->events
);
3404 local_inc(&buffer
->wakeup
);
3408 perf_output_put_handle(handle
);
3412 static u32
perf_event_pid(struct perf_event
*event
, struct task_struct
*p
)
3415 * only top level events have the pid namespace they were created in
3418 event
= event
->parent
;
3420 return task_tgid_nr_ns(p
, event
->ns
);
3423 static u32
perf_event_tid(struct perf_event
*event
, struct task_struct
*p
)
3426 * only top level events have the pid namespace they were created in
3429 event
= event
->parent
;
3431 return task_pid_nr_ns(p
, event
->ns
);
3434 static void perf_output_read_one(struct perf_output_handle
*handle
,
3435 struct perf_event
*event
,
3436 u64 enabled
, u64 running
)
3438 u64 read_format
= event
->attr
.read_format
;
3442 values
[n
++] = perf_event_count(event
);
3443 if (read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
) {
3444 values
[n
++] = enabled
+
3445 atomic64_read(&event
->child_total_time_enabled
);
3447 if (read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
) {
3448 values
[n
++] = running
+
3449 atomic64_read(&event
->child_total_time_running
);
3451 if (read_format
& PERF_FORMAT_ID
)
3452 values
[n
++] = primary_event_id(event
);
3454 perf_output_copy(handle
, values
, n
* sizeof(u64
));
3458 * XXX PERF_FORMAT_GROUP vs inherited events seems difficult.
3460 static void perf_output_read_group(struct perf_output_handle
*handle
,
3461 struct perf_event
*event
,
3462 u64 enabled
, u64 running
)
3464 struct perf_event
*leader
= event
->group_leader
, *sub
;
3465 u64 read_format
= event
->attr
.read_format
;
3469 values
[n
++] = 1 + leader
->nr_siblings
;
3471 if (read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
)
3472 values
[n
++] = enabled
;
3474 if (read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
)
3475 values
[n
++] = running
;
3477 if (leader
!= event
)
3478 leader
->pmu
->read(leader
);
3480 values
[n
++] = perf_event_count(leader
);
3481 if (read_format
& PERF_FORMAT_ID
)
3482 values
[n
++] = primary_event_id(leader
);
3484 perf_output_copy(handle
, values
, n
* sizeof(u64
));
3486 list_for_each_entry(sub
, &leader
->sibling_list
, group_entry
) {
3490 sub
->pmu
->read(sub
);
3492 values
[n
++] = perf_event_count(sub
);
3493 if (read_format
& PERF_FORMAT_ID
)
3494 values
[n
++] = primary_event_id(sub
);
3496 perf_output_copy(handle
, values
, n
* sizeof(u64
));
3500 #define PERF_FORMAT_TOTAL_TIMES (PERF_FORMAT_TOTAL_TIME_ENABLED|\
3501 PERF_FORMAT_TOTAL_TIME_RUNNING)
3503 static void perf_output_read(struct perf_output_handle
*handle
,
3504 struct perf_event
*event
)
3506 u64 enabled
= 0, running
= 0, now
, ctx_time
;
3507 u64 read_format
= event
->attr
.read_format
;
3510 * compute total_time_enabled, total_time_running
3511 * based on snapshot values taken when the event
3512 * was last scheduled in.
3514 * we cannot simply called update_context_time()
3515 * because of locking issue as we are called in
3518 if (read_format
& PERF_FORMAT_TOTAL_TIMES
) {
3520 ctx_time
= event
->shadow_ctx_time
+ now
;
3521 enabled
= ctx_time
- event
->tstamp_enabled
;
3522 running
= ctx_time
- event
->tstamp_running
;
3525 if (event
->attr
.read_format
& PERF_FORMAT_GROUP
)
3526 perf_output_read_group(handle
, event
, enabled
, running
);
3528 perf_output_read_one(handle
, event
, enabled
, running
);
3531 void perf_output_sample(struct perf_output_handle
*handle
,
3532 struct perf_event_header
*header
,
3533 struct perf_sample_data
*data
,
3534 struct perf_event
*event
)
3536 u64 sample_type
= data
->type
;
3538 perf_output_put(handle
, *header
);
3540 if (sample_type
& PERF_SAMPLE_IP
)
3541 perf_output_put(handle
, data
->ip
);
3543 if (sample_type
& PERF_SAMPLE_TID
)
3544 perf_output_put(handle
, data
->tid_entry
);
3546 if (sample_type
& PERF_SAMPLE_TIME
)
3547 perf_output_put(handle
, data
->time
);
3549 if (sample_type
& PERF_SAMPLE_ADDR
)
3550 perf_output_put(handle
, data
->addr
);
3552 if (sample_type
& PERF_SAMPLE_ID
)
3553 perf_output_put(handle
, data
->id
);
3555 if (sample_type
& PERF_SAMPLE_STREAM_ID
)
3556 perf_output_put(handle
, data
->stream_id
);
3558 if (sample_type
& PERF_SAMPLE_CPU
)
3559 perf_output_put(handle
, data
->cpu_entry
);
3561 if (sample_type
& PERF_SAMPLE_PERIOD
)
3562 perf_output_put(handle
, data
->period
);
3564 if (sample_type
& PERF_SAMPLE_READ
)
3565 perf_output_read(handle
, event
);
3567 if (sample_type
& PERF_SAMPLE_CALLCHAIN
) {
3568 if (data
->callchain
) {
3571 if (data
->callchain
)
3572 size
+= data
->callchain
->nr
;
3574 size
*= sizeof(u64
);
3576 perf_output_copy(handle
, data
->callchain
, size
);
3579 perf_output_put(handle
, nr
);
3583 if (sample_type
& PERF_SAMPLE_RAW
) {
3585 perf_output_put(handle
, data
->raw
->size
);
3586 perf_output_copy(handle
, data
->raw
->data
,
3593 .size
= sizeof(u32
),
3596 perf_output_put(handle
, raw
);
3601 void perf_prepare_sample(struct perf_event_header
*header
,
3602 struct perf_sample_data
*data
,
3603 struct perf_event
*event
,
3604 struct pt_regs
*regs
)
3606 u64 sample_type
= event
->attr
.sample_type
;
3608 data
->type
= sample_type
;
3610 header
->type
= PERF_RECORD_SAMPLE
;
3611 header
->size
= sizeof(*header
);
3614 header
->misc
|= perf_misc_flags(regs
);
3616 if (sample_type
& PERF_SAMPLE_IP
) {
3617 data
->ip
= perf_instruction_pointer(regs
);
3619 header
->size
+= sizeof(data
->ip
);
3622 if (sample_type
& PERF_SAMPLE_TID
) {
3623 /* namespace issues */
3624 data
->tid_entry
.pid
= perf_event_pid(event
, current
);
3625 data
->tid_entry
.tid
= perf_event_tid(event
, current
);
3627 header
->size
+= sizeof(data
->tid_entry
);
3630 if (sample_type
& PERF_SAMPLE_TIME
) {
3631 data
->time
= perf_clock();
3633 header
->size
+= sizeof(data
->time
);
3636 if (sample_type
& PERF_SAMPLE_ADDR
)
3637 header
->size
+= sizeof(data
->addr
);
3639 if (sample_type
& PERF_SAMPLE_ID
) {
3640 data
->id
= primary_event_id(event
);
3642 header
->size
+= sizeof(data
->id
);
3645 if (sample_type
& PERF_SAMPLE_STREAM_ID
) {
3646 data
->stream_id
= event
->id
;
3648 header
->size
+= sizeof(data
->stream_id
);
3651 if (sample_type
& PERF_SAMPLE_CPU
) {
3652 data
->cpu_entry
.cpu
= raw_smp_processor_id();
3653 data
->cpu_entry
.reserved
= 0;
3655 header
->size
+= sizeof(data
->cpu_entry
);
3658 if (sample_type
& PERF_SAMPLE_PERIOD
)
3659 header
->size
+= sizeof(data
->period
);
3661 if (sample_type
& PERF_SAMPLE_READ
)
3662 header
->size
+= perf_event_read_size(event
);
3664 if (sample_type
& PERF_SAMPLE_CALLCHAIN
) {
3667 data
->callchain
= perf_callchain(regs
);
3669 if (data
->callchain
)
3670 size
+= data
->callchain
->nr
;
3672 header
->size
+= size
* sizeof(u64
);
3675 if (sample_type
& PERF_SAMPLE_RAW
) {
3676 int size
= sizeof(u32
);
3679 size
+= data
->raw
->size
;
3681 size
+= sizeof(u32
);
3683 WARN_ON_ONCE(size
& (sizeof(u64
)-1));
3684 header
->size
+= size
;
3688 static void perf_event_output(struct perf_event
*event
, int nmi
,
3689 struct perf_sample_data
*data
,
3690 struct pt_regs
*regs
)
3692 struct perf_output_handle handle
;
3693 struct perf_event_header header
;
3695 /* protect the callchain buffers */
3698 perf_prepare_sample(&header
, data
, event
, regs
);
3700 if (perf_output_begin(&handle
, event
, header
.size
, nmi
, 1))
3703 perf_output_sample(&handle
, &header
, data
, event
);
3705 perf_output_end(&handle
);
3715 struct perf_read_event
{
3716 struct perf_event_header header
;
3723 perf_event_read_event(struct perf_event
*event
,
3724 struct task_struct
*task
)
3726 struct perf_output_handle handle
;
3727 struct perf_read_event read_event
= {
3729 .type
= PERF_RECORD_READ
,
3731 .size
= sizeof(read_event
) + perf_event_read_size(event
),
3733 .pid
= perf_event_pid(event
, task
),
3734 .tid
= perf_event_tid(event
, task
),
3738 ret
= perf_output_begin(&handle
, event
, read_event
.header
.size
, 0, 0);
3742 perf_output_put(&handle
, read_event
);
3743 perf_output_read(&handle
, event
);
3745 perf_output_end(&handle
);
3749 * task tracking -- fork/exit
3751 * enabled by: attr.comm | attr.mmap | attr.mmap_data | attr.task
3754 struct perf_task_event
{
3755 struct task_struct
*task
;
3756 struct perf_event_context
*task_ctx
;
3759 struct perf_event_header header
;
3769 static void perf_event_task_output(struct perf_event
*event
,
3770 struct perf_task_event
*task_event
)
3772 struct perf_output_handle handle
;
3773 struct task_struct
*task
= task_event
->task
;
3776 size
= task_event
->event_id
.header
.size
;
3777 ret
= perf_output_begin(&handle
, event
, size
, 0, 0);
3782 task_event
->event_id
.pid
= perf_event_pid(event
, task
);
3783 task_event
->event_id
.ppid
= perf_event_pid(event
, current
);
3785 task_event
->event_id
.tid
= perf_event_tid(event
, task
);
3786 task_event
->event_id
.ptid
= perf_event_tid(event
, current
);
3788 perf_output_put(&handle
, task_event
->event_id
);
3790 perf_output_end(&handle
);
3793 static int perf_event_task_match(struct perf_event
*event
)
3795 if (event
->state
< PERF_EVENT_STATE_INACTIVE
)
3798 if (event
->cpu
!= -1 && event
->cpu
!= smp_processor_id())
3801 if (event
->attr
.comm
|| event
->attr
.mmap
||
3802 event
->attr
.mmap_data
|| event
->attr
.task
)
3808 static void perf_event_task_ctx(struct perf_event_context
*ctx
,
3809 struct perf_task_event
*task_event
)
3811 struct perf_event
*event
;
3813 list_for_each_entry_rcu(event
, &ctx
->event_list
, event_entry
) {
3814 if (perf_event_task_match(event
))
3815 perf_event_task_output(event
, task_event
);
3819 static void perf_event_task_event(struct perf_task_event
*task_event
)
3821 struct perf_cpu_context
*cpuctx
;
3822 struct perf_event_context
*ctx
;
3827 list_for_each_entry_rcu(pmu
, &pmus
, entry
) {
3828 cpuctx
= get_cpu_ptr(pmu
->pmu_cpu_context
);
3829 perf_event_task_ctx(&cpuctx
->ctx
, task_event
);
3831 ctx
= task_event
->task_ctx
;
3833 ctxn
= pmu
->task_ctx_nr
;
3836 ctx
= rcu_dereference(current
->perf_event_ctxp
[ctxn
]);
3839 perf_event_task_ctx(ctx
, task_event
);
3841 put_cpu_ptr(pmu
->pmu_cpu_context
);
3846 static void perf_event_task(struct task_struct
*task
,
3847 struct perf_event_context
*task_ctx
,
3850 struct perf_task_event task_event
;
3852 if (!atomic_read(&nr_comm_events
) &&
3853 !atomic_read(&nr_mmap_events
) &&
3854 !atomic_read(&nr_task_events
))
3857 task_event
= (struct perf_task_event
){
3859 .task_ctx
= task_ctx
,
3862 .type
= new ? PERF_RECORD_FORK
: PERF_RECORD_EXIT
,
3864 .size
= sizeof(task_event
.event_id
),
3870 .time
= perf_clock(),
3874 perf_event_task_event(&task_event
);
3877 void perf_event_fork(struct task_struct
*task
)
3879 perf_event_task(task
, NULL
, 1);
3886 struct perf_comm_event
{
3887 struct task_struct
*task
;
3892 struct perf_event_header header
;
3899 static void perf_event_comm_output(struct perf_event
*event
,
3900 struct perf_comm_event
*comm_event
)
3902 struct perf_output_handle handle
;
3903 int size
= comm_event
->event_id
.header
.size
;
3904 int ret
= perf_output_begin(&handle
, event
, size
, 0, 0);
3909 comm_event
->event_id
.pid
= perf_event_pid(event
, comm_event
->task
);
3910 comm_event
->event_id
.tid
= perf_event_tid(event
, comm_event
->task
);
3912 perf_output_put(&handle
, comm_event
->event_id
);
3913 perf_output_copy(&handle
, comm_event
->comm
,
3914 comm_event
->comm_size
);
3915 perf_output_end(&handle
);
3918 static int perf_event_comm_match(struct perf_event
*event
)
3920 if (event
->state
< PERF_EVENT_STATE_INACTIVE
)
3923 if (event
->cpu
!= -1 && event
->cpu
!= smp_processor_id())
3926 if (event
->attr
.comm
)
3932 static void perf_event_comm_ctx(struct perf_event_context
*ctx
,
3933 struct perf_comm_event
*comm_event
)
3935 struct perf_event
*event
;
3937 list_for_each_entry_rcu(event
, &ctx
->event_list
, event_entry
) {
3938 if (perf_event_comm_match(event
))
3939 perf_event_comm_output(event
, comm_event
);
3943 static void perf_event_comm_event(struct perf_comm_event
*comm_event
)
3945 struct perf_cpu_context
*cpuctx
;
3946 struct perf_event_context
*ctx
;
3947 char comm
[TASK_COMM_LEN
];
3952 memset(comm
, 0, sizeof(comm
));
3953 strlcpy(comm
, comm_event
->task
->comm
, sizeof(comm
));
3954 size
= ALIGN(strlen(comm
)+1, sizeof(u64
));
3956 comm_event
->comm
= comm
;
3957 comm_event
->comm_size
= size
;
3959 comm_event
->event_id
.header
.size
= sizeof(comm_event
->event_id
) + size
;
3962 list_for_each_entry_rcu(pmu
, &pmus
, entry
) {
3963 cpuctx
= get_cpu_ptr(pmu
->pmu_cpu_context
);
3964 perf_event_comm_ctx(&cpuctx
->ctx
, comm_event
);
3966 ctxn
= pmu
->task_ctx_nr
;
3970 ctx
= rcu_dereference(current
->perf_event_ctxp
[ctxn
]);
3972 perf_event_comm_ctx(ctx
, comm_event
);
3974 put_cpu_ptr(pmu
->pmu_cpu_context
);
3979 void perf_event_comm(struct task_struct
*task
)
3981 struct perf_comm_event comm_event
;
3982 struct perf_event_context
*ctx
;
3985 for_each_task_context_nr(ctxn
) {
3986 ctx
= task
->perf_event_ctxp
[ctxn
];
3990 perf_event_enable_on_exec(ctx
);
3993 if (!atomic_read(&nr_comm_events
))
3996 comm_event
= (struct perf_comm_event
){
4002 .type
= PERF_RECORD_COMM
,
4011 perf_event_comm_event(&comm_event
);
4018 struct perf_mmap_event
{
4019 struct vm_area_struct
*vma
;
4021 const char *file_name
;
4025 struct perf_event_header header
;
4035 static void perf_event_mmap_output(struct perf_event
*event
,
4036 struct perf_mmap_event
*mmap_event
)
4038 struct perf_output_handle handle
;
4039 int size
= mmap_event
->event_id
.header
.size
;
4040 int ret
= perf_output_begin(&handle
, event
, size
, 0, 0);
4045 mmap_event
->event_id
.pid
= perf_event_pid(event
, current
);
4046 mmap_event
->event_id
.tid
= perf_event_tid(event
, current
);
4048 perf_output_put(&handle
, mmap_event
->event_id
);
4049 perf_output_copy(&handle
, mmap_event
->file_name
,
4050 mmap_event
->file_size
);
4051 perf_output_end(&handle
);
4054 static int perf_event_mmap_match(struct perf_event
*event
,
4055 struct perf_mmap_event
*mmap_event
,
4058 if (event
->state
< PERF_EVENT_STATE_INACTIVE
)
4061 if (event
->cpu
!= -1 && event
->cpu
!= smp_processor_id())
4064 if ((!executable
&& event
->attr
.mmap_data
) ||
4065 (executable
&& event
->attr
.mmap
))
4071 static void perf_event_mmap_ctx(struct perf_event_context
*ctx
,
4072 struct perf_mmap_event
*mmap_event
,
4075 struct perf_event
*event
;
4077 list_for_each_entry_rcu(event
, &ctx
->event_list
, event_entry
) {
4078 if (perf_event_mmap_match(event
, mmap_event
, executable
))
4079 perf_event_mmap_output(event
, mmap_event
);
4083 static void perf_event_mmap_event(struct perf_mmap_event
*mmap_event
)
4085 struct perf_cpu_context
*cpuctx
;
4086 struct perf_event_context
*ctx
;
4087 struct vm_area_struct
*vma
= mmap_event
->vma
;
4088 struct file
*file
= vma
->vm_file
;
4096 memset(tmp
, 0, sizeof(tmp
));
4100 * d_path works from the end of the buffer backwards, so we
4101 * need to add enough zero bytes after the string to handle
4102 * the 64bit alignment we do later.
4104 buf
= kzalloc(PATH_MAX
+ sizeof(u64
), GFP_KERNEL
);
4106 name
= strncpy(tmp
, "//enomem", sizeof(tmp
));
4109 name
= d_path(&file
->f_path
, buf
, PATH_MAX
);
4111 name
= strncpy(tmp
, "//toolong", sizeof(tmp
));
4115 if (arch_vma_name(mmap_event
->vma
)) {
4116 name
= strncpy(tmp
, arch_vma_name(mmap_event
->vma
),
4122 name
= strncpy(tmp
, "[vdso]", sizeof(tmp
));
4124 } else if (vma
->vm_start
<= vma
->vm_mm
->start_brk
&&
4125 vma
->vm_end
>= vma
->vm_mm
->brk
) {
4126 name
= strncpy(tmp
, "[heap]", sizeof(tmp
));
4128 } else if (vma
->vm_start
<= vma
->vm_mm
->start_stack
&&
4129 vma
->vm_end
>= vma
->vm_mm
->start_stack
) {
4130 name
= strncpy(tmp
, "[stack]", sizeof(tmp
));
4134 name
= strncpy(tmp
, "//anon", sizeof(tmp
));
4139 size
= ALIGN(strlen(name
)+1, sizeof(u64
));
4141 mmap_event
->file_name
= name
;
4142 mmap_event
->file_size
= size
;
4144 mmap_event
->event_id
.header
.size
= sizeof(mmap_event
->event_id
) + size
;
4147 list_for_each_entry_rcu(pmu
, &pmus
, entry
) {
4148 cpuctx
= get_cpu_ptr(pmu
->pmu_cpu_context
);
4149 perf_event_mmap_ctx(&cpuctx
->ctx
, mmap_event
,
4150 vma
->vm_flags
& VM_EXEC
);
4152 ctxn
= pmu
->task_ctx_nr
;
4156 ctx
= rcu_dereference(current
->perf_event_ctxp
[ctxn
]);
4158 perf_event_mmap_ctx(ctx
, mmap_event
,
4159 vma
->vm_flags
& VM_EXEC
);
4162 put_cpu_ptr(pmu
->pmu_cpu_context
);
4169 void perf_event_mmap(struct vm_area_struct
*vma
)
4171 struct perf_mmap_event mmap_event
;
4173 if (!atomic_read(&nr_mmap_events
))
4176 mmap_event
= (struct perf_mmap_event
){
4182 .type
= PERF_RECORD_MMAP
,
4183 .misc
= PERF_RECORD_MISC_USER
,
4188 .start
= vma
->vm_start
,
4189 .len
= vma
->vm_end
- vma
->vm_start
,
4190 .pgoff
= (u64
)vma
->vm_pgoff
<< PAGE_SHIFT
,
4194 perf_event_mmap_event(&mmap_event
);
4198 * IRQ throttle logging
4201 static void perf_log_throttle(struct perf_event
*event
, int enable
)
4203 struct perf_output_handle handle
;
4207 struct perf_event_header header
;
4211 } throttle_event
= {
4213 .type
= PERF_RECORD_THROTTLE
,
4215 .size
= sizeof(throttle_event
),
4217 .time
= perf_clock(),
4218 .id
= primary_event_id(event
),
4219 .stream_id
= event
->id
,
4223 throttle_event
.header
.type
= PERF_RECORD_UNTHROTTLE
;
4225 ret
= perf_output_begin(&handle
, event
, sizeof(throttle_event
), 1, 0);
4229 perf_output_put(&handle
, throttle_event
);
4230 perf_output_end(&handle
);
4234 * Generic event overflow handling, sampling.
4237 static int __perf_event_overflow(struct perf_event
*event
, int nmi
,
4238 int throttle
, struct perf_sample_data
*data
,
4239 struct pt_regs
*regs
)
4241 int events
= atomic_read(&event
->event_limit
);
4242 struct hw_perf_event
*hwc
= &event
->hw
;
4248 if (hwc
->interrupts
!= MAX_INTERRUPTS
) {
4250 if (HZ
* hwc
->interrupts
>
4251 (u64
)sysctl_perf_event_sample_rate
) {
4252 hwc
->interrupts
= MAX_INTERRUPTS
;
4253 perf_log_throttle(event
, 0);
4258 * Keep re-disabling events even though on the previous
4259 * pass we disabled it - just in case we raced with a
4260 * sched-in and the event got enabled again:
4266 if (event
->attr
.freq
) {
4267 u64 now
= perf_clock();
4268 s64 delta
= now
- hwc
->freq_time_stamp
;
4270 hwc
->freq_time_stamp
= now
;
4272 if (delta
> 0 && delta
< 2*TICK_NSEC
)
4273 perf_adjust_period(event
, delta
, hwc
->last_period
);
4277 * XXX event_limit might not quite work as expected on inherited
4281 event
->pending_kill
= POLL_IN
;
4282 if (events
&& atomic_dec_and_test(&event
->event_limit
)) {
4284 event
->pending_kill
= POLL_HUP
;
4286 event
->pending_disable
= 1;
4287 irq_work_queue(&event
->pending
);
4289 perf_event_disable(event
);
4292 if (event
->overflow_handler
)
4293 event
->overflow_handler(event
, nmi
, data
, regs
);
4295 perf_event_output(event
, nmi
, data
, regs
);
4300 int perf_event_overflow(struct perf_event
*event
, int nmi
,
4301 struct perf_sample_data
*data
,
4302 struct pt_regs
*regs
)
4304 return __perf_event_overflow(event
, nmi
, 1, data
, regs
);
4308 * Generic software event infrastructure
4311 struct swevent_htable
{
4312 struct swevent_hlist
*swevent_hlist
;
4313 struct mutex hlist_mutex
;
4316 /* Recursion avoidance in each contexts */
4317 int recursion
[PERF_NR_CONTEXTS
];
4320 static DEFINE_PER_CPU(struct swevent_htable
, swevent_htable
);
4323 * We directly increment event->count and keep a second value in
4324 * event->hw.period_left to count intervals. This period event
4325 * is kept in the range [-sample_period, 0] so that we can use the
4329 static u64
perf_swevent_set_period(struct perf_event
*event
)
4331 struct hw_perf_event
*hwc
= &event
->hw
;
4332 u64 period
= hwc
->last_period
;
4336 hwc
->last_period
= hwc
->sample_period
;
4339 old
= val
= local64_read(&hwc
->period_left
);
4343 nr
= div64_u64(period
+ val
, period
);
4344 offset
= nr
* period
;
4346 if (local64_cmpxchg(&hwc
->period_left
, old
, val
) != old
)
4352 static void perf_swevent_overflow(struct perf_event
*event
, u64 overflow
,
4353 int nmi
, struct perf_sample_data
*data
,
4354 struct pt_regs
*regs
)
4356 struct hw_perf_event
*hwc
= &event
->hw
;
4359 data
->period
= event
->hw
.last_period
;
4361 overflow
= perf_swevent_set_period(event
);
4363 if (hwc
->interrupts
== MAX_INTERRUPTS
)
4366 for (; overflow
; overflow
--) {
4367 if (__perf_event_overflow(event
, nmi
, throttle
,
4370 * We inhibit the overflow from happening when
4371 * hwc->interrupts == MAX_INTERRUPTS.
4379 static void perf_swevent_event(struct perf_event
*event
, u64 nr
,
4380 int nmi
, struct perf_sample_data
*data
,
4381 struct pt_regs
*regs
)
4383 struct hw_perf_event
*hwc
= &event
->hw
;
4385 local64_add(nr
, &event
->count
);
4390 if (!hwc
->sample_period
)
4393 if (nr
== 1 && hwc
->sample_period
== 1 && !event
->attr
.freq
)
4394 return perf_swevent_overflow(event
, 1, nmi
, data
, regs
);
4396 if (local64_add_negative(nr
, &hwc
->period_left
))
4399 perf_swevent_overflow(event
, 0, nmi
, data
, regs
);
4402 static int perf_exclude_event(struct perf_event
*event
,
4403 struct pt_regs
*regs
)
4405 if (event
->hw
.state
& PERF_HES_STOPPED
)
4409 if (event
->attr
.exclude_user
&& user_mode(regs
))
4412 if (event
->attr
.exclude_kernel
&& !user_mode(regs
))
4419 static int perf_swevent_match(struct perf_event
*event
,
4420 enum perf_type_id type
,
4422 struct perf_sample_data
*data
,
4423 struct pt_regs
*regs
)
4425 if (event
->attr
.type
!= type
)
4428 if (event
->attr
.config
!= event_id
)
4431 if (perf_exclude_event(event
, regs
))
4437 static inline u64
swevent_hash(u64 type
, u32 event_id
)
4439 u64 val
= event_id
| (type
<< 32);
4441 return hash_64(val
, SWEVENT_HLIST_BITS
);
4444 static inline struct hlist_head
*
4445 __find_swevent_head(struct swevent_hlist
*hlist
, u64 type
, u32 event_id
)
4447 u64 hash
= swevent_hash(type
, event_id
);
4449 return &hlist
->heads
[hash
];
4452 /* For the read side: events when they trigger */
4453 static inline struct hlist_head
*
4454 find_swevent_head_rcu(struct swevent_htable
*swhash
, u64 type
, u32 event_id
)
4456 struct swevent_hlist
*hlist
;
4458 hlist
= rcu_dereference(swhash
->swevent_hlist
);
4462 return __find_swevent_head(hlist
, type
, event_id
);
4465 /* For the event head insertion and removal in the hlist */
4466 static inline struct hlist_head
*
4467 find_swevent_head(struct swevent_htable
*swhash
, struct perf_event
*event
)
4469 struct swevent_hlist
*hlist
;
4470 u32 event_id
= event
->attr
.config
;
4471 u64 type
= event
->attr
.type
;
4474 * Event scheduling is always serialized against hlist allocation
4475 * and release. Which makes the protected version suitable here.
4476 * The context lock guarantees that.
4478 hlist
= rcu_dereference_protected(swhash
->swevent_hlist
,
4479 lockdep_is_held(&event
->ctx
->lock
));
4483 return __find_swevent_head(hlist
, type
, event_id
);
4486 static void do_perf_sw_event(enum perf_type_id type
, u32 event_id
,
4488 struct perf_sample_data
*data
,
4489 struct pt_regs
*regs
)
4491 struct swevent_htable
*swhash
= &__get_cpu_var(swevent_htable
);
4492 struct perf_event
*event
;
4493 struct hlist_node
*node
;
4494 struct hlist_head
*head
;
4497 head
= find_swevent_head_rcu(swhash
, type
, event_id
);
4501 hlist_for_each_entry_rcu(event
, node
, head
, hlist_entry
) {
4502 if (perf_swevent_match(event
, type
, event_id
, data
, regs
))
4503 perf_swevent_event(event
, nr
, nmi
, data
, regs
);
4509 int perf_swevent_get_recursion_context(void)
4511 struct swevent_htable
*swhash
= &__get_cpu_var(swevent_htable
);
4513 return get_recursion_context(swhash
->recursion
);
4515 EXPORT_SYMBOL_GPL(perf_swevent_get_recursion_context
);
4517 void inline perf_swevent_put_recursion_context(int rctx
)
4519 struct swevent_htable
*swhash
= &__get_cpu_var(swevent_htable
);
4521 put_recursion_context(swhash
->recursion
, rctx
);
4524 void __perf_sw_event(u32 event_id
, u64 nr
, int nmi
,
4525 struct pt_regs
*regs
, u64 addr
)
4527 struct perf_sample_data data
;
4530 preempt_disable_notrace();
4531 rctx
= perf_swevent_get_recursion_context();
4535 perf_sample_data_init(&data
, addr
);
4537 do_perf_sw_event(PERF_TYPE_SOFTWARE
, event_id
, nr
, nmi
, &data
, regs
);
4539 perf_swevent_put_recursion_context(rctx
);
4540 preempt_enable_notrace();
4543 static void perf_swevent_read(struct perf_event
*event
)
4547 static int perf_swevent_add(struct perf_event
*event
, int flags
)
4549 struct swevent_htable
*swhash
= &__get_cpu_var(swevent_htable
);
4550 struct hw_perf_event
*hwc
= &event
->hw
;
4551 struct hlist_head
*head
;
4553 if (hwc
->sample_period
) {
4554 hwc
->last_period
= hwc
->sample_period
;
4555 perf_swevent_set_period(event
);
4558 hwc
->state
= !(flags
& PERF_EF_START
);
4560 head
= find_swevent_head(swhash
, event
);
4561 if (WARN_ON_ONCE(!head
))
4564 hlist_add_head_rcu(&event
->hlist_entry
, head
);
4569 static void perf_swevent_del(struct perf_event
*event
, int flags
)
4571 hlist_del_rcu(&event
->hlist_entry
);
4574 static void perf_swevent_start(struct perf_event
*event
, int flags
)
4576 event
->hw
.state
= 0;
4579 static void perf_swevent_stop(struct perf_event
*event
, int flags
)
4581 event
->hw
.state
= PERF_HES_STOPPED
;
4584 /* Deref the hlist from the update side */
4585 static inline struct swevent_hlist
*
4586 swevent_hlist_deref(struct swevent_htable
*swhash
)
4588 return rcu_dereference_protected(swhash
->swevent_hlist
,
4589 lockdep_is_held(&swhash
->hlist_mutex
));
4592 static void swevent_hlist_release_rcu(struct rcu_head
*rcu_head
)
4594 struct swevent_hlist
*hlist
;
4596 hlist
= container_of(rcu_head
, struct swevent_hlist
, rcu_head
);
4600 static void swevent_hlist_release(struct swevent_htable
*swhash
)
4602 struct swevent_hlist
*hlist
= swevent_hlist_deref(swhash
);
4607 rcu_assign_pointer(swhash
->swevent_hlist
, NULL
);
4608 call_rcu(&hlist
->rcu_head
, swevent_hlist_release_rcu
);
4611 static void swevent_hlist_put_cpu(struct perf_event
*event
, int cpu
)
4613 struct swevent_htable
*swhash
= &per_cpu(swevent_htable
, cpu
);
4615 mutex_lock(&swhash
->hlist_mutex
);
4617 if (!--swhash
->hlist_refcount
)
4618 swevent_hlist_release(swhash
);
4620 mutex_unlock(&swhash
->hlist_mutex
);
4623 static void swevent_hlist_put(struct perf_event
*event
)
4627 if (event
->cpu
!= -1) {
4628 swevent_hlist_put_cpu(event
, event
->cpu
);
4632 for_each_possible_cpu(cpu
)
4633 swevent_hlist_put_cpu(event
, cpu
);
4636 static int swevent_hlist_get_cpu(struct perf_event
*event
, int cpu
)
4638 struct swevent_htable
*swhash
= &per_cpu(swevent_htable
, cpu
);
4641 mutex_lock(&swhash
->hlist_mutex
);
4643 if (!swevent_hlist_deref(swhash
) && cpu_online(cpu
)) {
4644 struct swevent_hlist
*hlist
;
4646 hlist
= kzalloc(sizeof(*hlist
), GFP_KERNEL
);
4651 rcu_assign_pointer(swhash
->swevent_hlist
, hlist
);
4653 swhash
->hlist_refcount
++;
4655 mutex_unlock(&swhash
->hlist_mutex
);
4660 static int swevent_hlist_get(struct perf_event
*event
)
4663 int cpu
, failed_cpu
;
4665 if (event
->cpu
!= -1)
4666 return swevent_hlist_get_cpu(event
, event
->cpu
);
4669 for_each_possible_cpu(cpu
) {
4670 err
= swevent_hlist_get_cpu(event
, cpu
);
4680 for_each_possible_cpu(cpu
) {
4681 if (cpu
== failed_cpu
)
4683 swevent_hlist_put_cpu(event
, cpu
);
4690 atomic_t perf_swevent_enabled
[PERF_COUNT_SW_MAX
];
4692 static void sw_perf_event_destroy(struct perf_event
*event
)
4694 u64 event_id
= event
->attr
.config
;
4696 WARN_ON(event
->parent
);
4698 jump_label_dec(&perf_swevent_enabled
[event_id
]);
4699 swevent_hlist_put(event
);
4702 static int perf_swevent_init(struct perf_event
*event
)
4704 int event_id
= event
->attr
.config
;
4706 if (event
->attr
.type
!= PERF_TYPE_SOFTWARE
)
4710 case PERF_COUNT_SW_CPU_CLOCK
:
4711 case PERF_COUNT_SW_TASK_CLOCK
:
4718 if (event_id
> PERF_COUNT_SW_MAX
)
4721 if (!event
->parent
) {
4724 err
= swevent_hlist_get(event
);
4728 jump_label_inc(&perf_swevent_enabled
[event_id
]);
4729 event
->destroy
= sw_perf_event_destroy
;
4735 static struct pmu perf_swevent
= {
4736 .task_ctx_nr
= perf_sw_context
,
4738 .event_init
= perf_swevent_init
,
4739 .add
= perf_swevent_add
,
4740 .del
= perf_swevent_del
,
4741 .start
= perf_swevent_start
,
4742 .stop
= perf_swevent_stop
,
4743 .read
= perf_swevent_read
,
4746 #ifdef CONFIG_EVENT_TRACING
4748 static int perf_tp_filter_match(struct perf_event
*event
,
4749 struct perf_sample_data
*data
)
4751 void *record
= data
->raw
->data
;
4753 if (likely(!event
->filter
) || filter_match_preds(event
->filter
, record
))
4758 static int perf_tp_event_match(struct perf_event
*event
,
4759 struct perf_sample_data
*data
,
4760 struct pt_regs
*regs
)
4763 * All tracepoints are from kernel-space.
4765 if (event
->attr
.exclude_kernel
)
4768 if (!perf_tp_filter_match(event
, data
))
4774 void perf_tp_event(u64 addr
, u64 count
, void *record
, int entry_size
,
4775 struct pt_regs
*regs
, struct hlist_head
*head
, int rctx
)
4777 struct perf_sample_data data
;
4778 struct perf_event
*event
;
4779 struct hlist_node
*node
;
4781 struct perf_raw_record raw
= {
4786 perf_sample_data_init(&data
, addr
);
4789 hlist_for_each_entry_rcu(event
, node
, head
, hlist_entry
) {
4790 if (perf_tp_event_match(event
, &data
, regs
))
4791 perf_swevent_event(event
, count
, 1, &data
, regs
);
4794 perf_swevent_put_recursion_context(rctx
);
4796 EXPORT_SYMBOL_GPL(perf_tp_event
);
4798 static void tp_perf_event_destroy(struct perf_event
*event
)
4800 perf_trace_destroy(event
);
4803 static int perf_tp_event_init(struct perf_event
*event
)
4807 if (event
->attr
.type
!= PERF_TYPE_TRACEPOINT
)
4811 * Raw tracepoint data is a severe data leak, only allow root to
4814 if ((event
->attr
.sample_type
& PERF_SAMPLE_RAW
) &&
4815 perf_paranoid_tracepoint_raw() &&
4816 !capable(CAP_SYS_ADMIN
))
4819 err
= perf_trace_init(event
);
4823 event
->destroy
= tp_perf_event_destroy
;
4828 static struct pmu perf_tracepoint
= {
4829 .task_ctx_nr
= perf_sw_context
,
4831 .event_init
= perf_tp_event_init
,
4832 .add
= perf_trace_add
,
4833 .del
= perf_trace_del
,
4834 .start
= perf_swevent_start
,
4835 .stop
= perf_swevent_stop
,
4836 .read
= perf_swevent_read
,
4839 static inline void perf_tp_register(void)
4841 perf_pmu_register(&perf_tracepoint
);
4844 static int perf_event_set_filter(struct perf_event
*event
, void __user
*arg
)
4849 if (event
->attr
.type
!= PERF_TYPE_TRACEPOINT
)
4852 filter_str
= strndup_user(arg
, PAGE_SIZE
);
4853 if (IS_ERR(filter_str
))
4854 return PTR_ERR(filter_str
);
4856 ret
= ftrace_profile_set_filter(event
, event
->attr
.config
, filter_str
);
4862 static void perf_event_free_filter(struct perf_event
*event
)
4864 ftrace_profile_free_filter(event
);
4869 static inline void perf_tp_register(void)
4873 static int perf_event_set_filter(struct perf_event
*event
, void __user
*arg
)
4878 static void perf_event_free_filter(struct perf_event
*event
)
4882 #endif /* CONFIG_EVENT_TRACING */
4884 #ifdef CONFIG_HAVE_HW_BREAKPOINT
4885 void perf_bp_event(struct perf_event
*bp
, void *data
)
4887 struct perf_sample_data sample
;
4888 struct pt_regs
*regs
= data
;
4890 perf_sample_data_init(&sample
, bp
->attr
.bp_addr
);
4892 if (!bp
->hw
.state
&& !perf_exclude_event(bp
, regs
))
4893 perf_swevent_event(bp
, 1, 1, &sample
, regs
);
4898 * hrtimer based swevent callback
4901 static enum hrtimer_restart
perf_swevent_hrtimer(struct hrtimer
*hrtimer
)
4903 enum hrtimer_restart ret
= HRTIMER_RESTART
;
4904 struct perf_sample_data data
;
4905 struct pt_regs
*regs
;
4906 struct perf_event
*event
;
4909 event
= container_of(hrtimer
, struct perf_event
, hw
.hrtimer
);
4910 event
->pmu
->read(event
);
4912 perf_sample_data_init(&data
, 0);
4913 data
.period
= event
->hw
.last_period
;
4914 regs
= get_irq_regs();
4916 if (regs
&& !perf_exclude_event(event
, regs
)) {
4917 if (!(event
->attr
.exclude_idle
&& current
->pid
== 0))
4918 if (perf_event_overflow(event
, 0, &data
, regs
))
4919 ret
= HRTIMER_NORESTART
;
4922 period
= max_t(u64
, 10000, event
->hw
.sample_period
);
4923 hrtimer_forward_now(hrtimer
, ns_to_ktime(period
));
4928 static void perf_swevent_start_hrtimer(struct perf_event
*event
)
4930 struct hw_perf_event
*hwc
= &event
->hw
;
4932 hrtimer_init(&hwc
->hrtimer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
4933 hwc
->hrtimer
.function
= perf_swevent_hrtimer
;
4934 if (hwc
->sample_period
) {
4935 s64 period
= local64_read(&hwc
->period_left
);
4941 local64_set(&hwc
->period_left
, 0);
4943 period
= max_t(u64
, 10000, hwc
->sample_period
);
4945 __hrtimer_start_range_ns(&hwc
->hrtimer
,
4946 ns_to_ktime(period
), 0,
4947 HRTIMER_MODE_REL_PINNED
, 0);
4951 static void perf_swevent_cancel_hrtimer(struct perf_event
*event
)
4953 struct hw_perf_event
*hwc
= &event
->hw
;
4955 if (hwc
->sample_period
) {
4956 ktime_t remaining
= hrtimer_get_remaining(&hwc
->hrtimer
);
4957 local64_set(&hwc
->period_left
, ktime_to_ns(remaining
));
4959 hrtimer_cancel(&hwc
->hrtimer
);
4964 * Software event: cpu wall time clock
4967 static void cpu_clock_event_update(struct perf_event
*event
)
4972 now
= local_clock();
4973 prev
= local64_xchg(&event
->hw
.prev_count
, now
);
4974 local64_add(now
- prev
, &event
->count
);
4977 static void cpu_clock_event_start(struct perf_event
*event
, int flags
)
4979 local64_set(&event
->hw
.prev_count
, local_clock());
4980 perf_swevent_start_hrtimer(event
);
4983 static void cpu_clock_event_stop(struct perf_event
*event
, int flags
)
4985 perf_swevent_cancel_hrtimer(event
);
4986 cpu_clock_event_update(event
);
4989 static int cpu_clock_event_add(struct perf_event
*event
, int flags
)
4991 if (flags
& PERF_EF_START
)
4992 cpu_clock_event_start(event
, flags
);
4997 static void cpu_clock_event_del(struct perf_event
*event
, int flags
)
4999 cpu_clock_event_stop(event
, flags
);
5002 static void cpu_clock_event_read(struct perf_event
*event
)
5004 cpu_clock_event_update(event
);
5007 static int cpu_clock_event_init(struct perf_event
*event
)
5009 if (event
->attr
.type
!= PERF_TYPE_SOFTWARE
)
5012 if (event
->attr
.config
!= PERF_COUNT_SW_CPU_CLOCK
)
5018 static struct pmu perf_cpu_clock
= {
5019 .task_ctx_nr
= perf_sw_context
,
5021 .event_init
= cpu_clock_event_init
,
5022 .add
= cpu_clock_event_add
,
5023 .del
= cpu_clock_event_del
,
5024 .start
= cpu_clock_event_start
,
5025 .stop
= cpu_clock_event_stop
,
5026 .read
= cpu_clock_event_read
,
5030 * Software event: task time clock
5033 static void task_clock_event_update(struct perf_event
*event
, u64 now
)
5038 prev
= local64_xchg(&event
->hw
.prev_count
, now
);
5040 local64_add(delta
, &event
->count
);
5043 static void task_clock_event_start(struct perf_event
*event
, int flags
)
5045 local64_set(&event
->hw
.prev_count
, event
->ctx
->time
);
5046 perf_swevent_start_hrtimer(event
);
5049 static void task_clock_event_stop(struct perf_event
*event
, int flags
)
5051 perf_swevent_cancel_hrtimer(event
);
5052 task_clock_event_update(event
, event
->ctx
->time
);
5055 static int task_clock_event_add(struct perf_event
*event
, int flags
)
5057 if (flags
& PERF_EF_START
)
5058 task_clock_event_start(event
, flags
);
5063 static void task_clock_event_del(struct perf_event
*event
, int flags
)
5065 task_clock_event_stop(event
, PERF_EF_UPDATE
);
5068 static void task_clock_event_read(struct perf_event
*event
)
5073 update_context_time(event
->ctx
);
5074 time
= event
->ctx
->time
;
5076 u64 now
= perf_clock();
5077 u64 delta
= now
- event
->ctx
->timestamp
;
5078 time
= event
->ctx
->time
+ delta
;
5081 task_clock_event_update(event
, time
);
5084 static int task_clock_event_init(struct perf_event
*event
)
5086 if (event
->attr
.type
!= PERF_TYPE_SOFTWARE
)
5089 if (event
->attr
.config
!= PERF_COUNT_SW_TASK_CLOCK
)
5095 static struct pmu perf_task_clock
= {
5096 .task_ctx_nr
= perf_sw_context
,
5098 .event_init
= task_clock_event_init
,
5099 .add
= task_clock_event_add
,
5100 .del
= task_clock_event_del
,
5101 .start
= task_clock_event_start
,
5102 .stop
= task_clock_event_stop
,
5103 .read
= task_clock_event_read
,
5106 static void perf_pmu_nop_void(struct pmu
*pmu
)
5110 static int perf_pmu_nop_int(struct pmu
*pmu
)
5115 static void perf_pmu_start_txn(struct pmu
*pmu
)
5117 perf_pmu_disable(pmu
);
5120 static int perf_pmu_commit_txn(struct pmu
*pmu
)
5122 perf_pmu_enable(pmu
);
5126 static void perf_pmu_cancel_txn(struct pmu
*pmu
)
5128 perf_pmu_enable(pmu
);
5132 * Ensures all contexts with the same task_ctx_nr have the same
5133 * pmu_cpu_context too.
5135 static void *find_pmu_context(int ctxn
)
5142 list_for_each_entry(pmu
, &pmus
, entry
) {
5143 if (pmu
->task_ctx_nr
== ctxn
)
5144 return pmu
->pmu_cpu_context
;
5150 static void free_pmu_context(void * __percpu cpu_context
)
5154 mutex_lock(&pmus_lock
);
5156 * Like a real lame refcount.
5158 list_for_each_entry(pmu
, &pmus
, entry
) {
5159 if (pmu
->pmu_cpu_context
== cpu_context
)
5163 free_percpu(cpu_context
);
5165 mutex_unlock(&pmus_lock
);
5168 int perf_pmu_register(struct pmu
*pmu
)
5172 mutex_lock(&pmus_lock
);
5174 pmu
->pmu_disable_count
= alloc_percpu(int);
5175 if (!pmu
->pmu_disable_count
)
5178 pmu
->pmu_cpu_context
= find_pmu_context(pmu
->task_ctx_nr
);
5179 if (pmu
->pmu_cpu_context
)
5180 goto got_cpu_context
;
5182 pmu
->pmu_cpu_context
= alloc_percpu(struct perf_cpu_context
);
5183 if (!pmu
->pmu_cpu_context
)
5186 for_each_possible_cpu(cpu
) {
5187 struct perf_cpu_context
*cpuctx
;
5189 cpuctx
= per_cpu_ptr(pmu
->pmu_cpu_context
, cpu
);
5190 __perf_event_init_context(&cpuctx
->ctx
);
5191 cpuctx
->ctx
.type
= cpu_context
;
5192 cpuctx
->ctx
.pmu
= pmu
;
5193 cpuctx
->jiffies_interval
= 1;
5194 INIT_LIST_HEAD(&cpuctx
->rotation_list
);
5198 if (!pmu
->start_txn
) {
5199 if (pmu
->pmu_enable
) {
5201 * If we have pmu_enable/pmu_disable calls, install
5202 * transaction stubs that use that to try and batch
5203 * hardware accesses.
5205 pmu
->start_txn
= perf_pmu_start_txn
;
5206 pmu
->commit_txn
= perf_pmu_commit_txn
;
5207 pmu
->cancel_txn
= perf_pmu_cancel_txn
;
5209 pmu
->start_txn
= perf_pmu_nop_void
;
5210 pmu
->commit_txn
= perf_pmu_nop_int
;
5211 pmu
->cancel_txn
= perf_pmu_nop_void
;
5215 if (!pmu
->pmu_enable
) {
5216 pmu
->pmu_enable
= perf_pmu_nop_void
;
5217 pmu
->pmu_disable
= perf_pmu_nop_void
;
5220 list_add_rcu(&pmu
->entry
, &pmus
);
5223 mutex_unlock(&pmus_lock
);
5228 free_percpu(pmu
->pmu_disable_count
);
5232 void perf_pmu_unregister(struct pmu
*pmu
)
5234 mutex_lock(&pmus_lock
);
5235 list_del_rcu(&pmu
->entry
);
5236 mutex_unlock(&pmus_lock
);
5239 * We dereference the pmu list under both SRCU and regular RCU, so
5240 * synchronize against both of those.
5242 synchronize_srcu(&pmus_srcu
);
5245 free_percpu(pmu
->pmu_disable_count
);
5246 free_pmu_context(pmu
->pmu_cpu_context
);
5249 struct pmu
*perf_init_event(struct perf_event
*event
)
5251 struct pmu
*pmu
= NULL
;
5254 idx
= srcu_read_lock(&pmus_srcu
);
5255 list_for_each_entry_rcu(pmu
, &pmus
, entry
) {
5256 int ret
= pmu
->event_init(event
);
5260 if (ret
!= -ENOENT
) {
5265 pmu
= ERR_PTR(-ENOENT
);
5267 srcu_read_unlock(&pmus_srcu
, idx
);
5273 * Allocate and initialize a event structure
5275 static struct perf_event
*
5276 perf_event_alloc(struct perf_event_attr
*attr
, int cpu
,
5277 struct task_struct
*task
,
5278 struct perf_event
*group_leader
,
5279 struct perf_event
*parent_event
,
5280 perf_overflow_handler_t overflow_handler
)
5283 struct perf_event
*event
;
5284 struct hw_perf_event
*hwc
;
5287 event
= kzalloc(sizeof(*event
), GFP_KERNEL
);
5289 return ERR_PTR(-ENOMEM
);
5292 * Single events are their own group leaders, with an
5293 * empty sibling list:
5296 group_leader
= event
;
5298 mutex_init(&event
->child_mutex
);
5299 INIT_LIST_HEAD(&event
->child_list
);
5301 INIT_LIST_HEAD(&event
->group_entry
);
5302 INIT_LIST_HEAD(&event
->event_entry
);
5303 INIT_LIST_HEAD(&event
->sibling_list
);
5304 init_waitqueue_head(&event
->waitq
);
5305 init_irq_work(&event
->pending
, perf_pending_event
);
5307 mutex_init(&event
->mmap_mutex
);
5310 event
->attr
= *attr
;
5311 event
->group_leader
= group_leader
;
5315 event
->parent
= parent_event
;
5317 event
->ns
= get_pid_ns(current
->nsproxy
->pid_ns
);
5318 event
->id
= atomic64_inc_return(&perf_event_id
);
5320 event
->state
= PERF_EVENT_STATE_INACTIVE
;
5323 event
->attach_state
= PERF_ATTACH_TASK
;
5324 #ifdef CONFIG_HAVE_HW_BREAKPOINT
5326 * hw_breakpoint is a bit difficult here..
5328 if (attr
->type
== PERF_TYPE_BREAKPOINT
)
5329 event
->hw
.bp_target
= task
;
5333 if (!overflow_handler
&& parent_event
)
5334 overflow_handler
= parent_event
->overflow_handler
;
5336 event
->overflow_handler
= overflow_handler
;
5339 event
->state
= PERF_EVENT_STATE_OFF
;
5344 hwc
->sample_period
= attr
->sample_period
;
5345 if (attr
->freq
&& attr
->sample_freq
)
5346 hwc
->sample_period
= 1;
5347 hwc
->last_period
= hwc
->sample_period
;
5349 local64_set(&hwc
->period_left
, hwc
->sample_period
);
5352 * we currently do not support PERF_FORMAT_GROUP on inherited events
5354 if (attr
->inherit
&& (attr
->read_format
& PERF_FORMAT_GROUP
))
5357 pmu
= perf_init_event(event
);
5363 else if (IS_ERR(pmu
))
5368 put_pid_ns(event
->ns
);
5370 return ERR_PTR(err
);
5375 if (!event
->parent
) {
5376 if (event
->attach_state
& PERF_ATTACH_TASK
)
5377 jump_label_inc(&perf_task_events
);
5378 if (event
->attr
.mmap
|| event
->attr
.mmap_data
)
5379 atomic_inc(&nr_mmap_events
);
5380 if (event
->attr
.comm
)
5381 atomic_inc(&nr_comm_events
);
5382 if (event
->attr
.task
)
5383 atomic_inc(&nr_task_events
);
5384 if (event
->attr
.sample_type
& PERF_SAMPLE_CALLCHAIN
) {
5385 err
= get_callchain_buffers();
5388 return ERR_PTR(err
);
5396 static int perf_copy_attr(struct perf_event_attr __user
*uattr
,
5397 struct perf_event_attr
*attr
)
5402 if (!access_ok(VERIFY_WRITE
, uattr
, PERF_ATTR_SIZE_VER0
))
5406 * zero the full structure, so that a short copy will be nice.
5408 memset(attr
, 0, sizeof(*attr
));
5410 ret
= get_user(size
, &uattr
->size
);
5414 if (size
> PAGE_SIZE
) /* silly large */
5417 if (!size
) /* abi compat */
5418 size
= PERF_ATTR_SIZE_VER0
;
5420 if (size
< PERF_ATTR_SIZE_VER0
)
5424 * If we're handed a bigger struct than we know of,
5425 * ensure all the unknown bits are 0 - i.e. new
5426 * user-space does not rely on any kernel feature
5427 * extensions we dont know about yet.
5429 if (size
> sizeof(*attr
)) {
5430 unsigned char __user
*addr
;
5431 unsigned char __user
*end
;
5434 addr
= (void __user
*)uattr
+ sizeof(*attr
);
5435 end
= (void __user
*)uattr
+ size
;
5437 for (; addr
< end
; addr
++) {
5438 ret
= get_user(val
, addr
);
5444 size
= sizeof(*attr
);
5447 ret
= copy_from_user(attr
, uattr
, size
);
5452 * If the type exists, the corresponding creation will verify
5455 if (attr
->type
>= PERF_TYPE_MAX
)
5458 if (attr
->__reserved_1
)
5461 if (attr
->sample_type
& ~(PERF_SAMPLE_MAX
-1))
5464 if (attr
->read_format
& ~(PERF_FORMAT_MAX
-1))
5471 put_user(sizeof(*attr
), &uattr
->size
);
5477 perf_event_set_output(struct perf_event
*event
, struct perf_event
*output_event
)
5479 struct perf_buffer
*buffer
= NULL
, *old_buffer
= NULL
;
5485 /* don't allow circular references */
5486 if (event
== output_event
)
5490 * Don't allow cross-cpu buffers
5492 if (output_event
->cpu
!= event
->cpu
)
5496 * If its not a per-cpu buffer, it must be the same task.
5498 if (output_event
->cpu
== -1 && output_event
->ctx
!= event
->ctx
)
5502 mutex_lock(&event
->mmap_mutex
);
5503 /* Can't redirect output if we've got an active mmap() */
5504 if (atomic_read(&event
->mmap_count
))
5508 /* get the buffer we want to redirect to */
5509 buffer
= perf_buffer_get(output_event
);
5514 old_buffer
= event
->buffer
;
5515 rcu_assign_pointer(event
->buffer
, buffer
);
5518 mutex_unlock(&event
->mmap_mutex
);
5521 perf_buffer_put(old_buffer
);
5527 * sys_perf_event_open - open a performance event, associate it to a task/cpu
5529 * @attr_uptr: event_id type attributes for monitoring/sampling
5532 * @group_fd: group leader event fd
5534 SYSCALL_DEFINE5(perf_event_open
,
5535 struct perf_event_attr __user
*, attr_uptr
,
5536 pid_t
, pid
, int, cpu
, int, group_fd
, unsigned long, flags
)
5538 struct perf_event
*group_leader
= NULL
, *output_event
= NULL
;
5539 struct perf_event
*event
, *sibling
;
5540 struct perf_event_attr attr
;
5541 struct perf_event_context
*ctx
;
5542 struct file
*event_file
= NULL
;
5543 struct file
*group_file
= NULL
;
5544 struct task_struct
*task
= NULL
;
5548 int fput_needed
= 0;
5551 /* for future expandability... */
5552 if (flags
& ~(PERF_FLAG_FD_NO_GROUP
| PERF_FLAG_FD_OUTPUT
))
5555 err
= perf_copy_attr(attr_uptr
, &attr
);
5559 if (!attr
.exclude_kernel
) {
5560 if (perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN
))
5565 if (attr
.sample_freq
> sysctl_perf_event_sample_rate
)
5569 event_fd
= get_unused_fd_flags(O_RDWR
);
5573 if (group_fd
!= -1) {
5574 group_leader
= perf_fget_light(group_fd
, &fput_needed
);
5575 if (IS_ERR(group_leader
)) {
5576 err
= PTR_ERR(group_leader
);
5579 group_file
= group_leader
->filp
;
5580 if (flags
& PERF_FLAG_FD_OUTPUT
)
5581 output_event
= group_leader
;
5582 if (flags
& PERF_FLAG_FD_NO_GROUP
)
5583 group_leader
= NULL
;
5587 task
= find_lively_task_by_vpid(pid
);
5589 err
= PTR_ERR(task
);
5594 event
= perf_event_alloc(&attr
, cpu
, task
, group_leader
, NULL
, NULL
);
5595 if (IS_ERR(event
)) {
5596 err
= PTR_ERR(event
);
5601 * Special case software events and allow them to be part of
5602 * any hardware group.
5607 (is_software_event(event
) != is_software_event(group_leader
))) {
5608 if (is_software_event(event
)) {
5610 * If event and group_leader are not both a software
5611 * event, and event is, then group leader is not.
5613 * Allow the addition of software events to !software
5614 * groups, this is safe because software events never
5617 pmu
= group_leader
->pmu
;
5618 } else if (is_software_event(group_leader
) &&
5619 (group_leader
->group_flags
& PERF_GROUP_SOFTWARE
)) {
5621 * In case the group is a pure software group, and we
5622 * try to add a hardware event, move the whole group to
5623 * the hardware context.
5630 * Get the target context (task or percpu):
5632 ctx
= find_get_context(pmu
, task
, cpu
);
5639 * Look up the group leader (we will attach this event to it):
5645 * Do not allow a recursive hierarchy (this new sibling
5646 * becoming part of another group-sibling):
5648 if (group_leader
->group_leader
!= group_leader
)
5651 * Do not allow to attach to a group in a different
5652 * task or CPU context:
5655 if (group_leader
->ctx
->type
!= ctx
->type
)
5658 if (group_leader
->ctx
!= ctx
)
5663 * Only a group leader can be exclusive or pinned
5665 if (attr
.exclusive
|| attr
.pinned
)
5670 err
= perf_event_set_output(event
, output_event
);
5675 event_file
= anon_inode_getfile("[perf_event]", &perf_fops
, event
, O_RDWR
);
5676 if (IS_ERR(event_file
)) {
5677 err
= PTR_ERR(event_file
);
5682 struct perf_event_context
*gctx
= group_leader
->ctx
;
5684 mutex_lock(&gctx
->mutex
);
5685 perf_event_remove_from_context(group_leader
);
5686 list_for_each_entry(sibling
, &group_leader
->sibling_list
,
5688 perf_event_remove_from_context(sibling
);
5691 mutex_unlock(&gctx
->mutex
);
5695 event
->filp
= event_file
;
5696 WARN_ON_ONCE(ctx
->parent_ctx
);
5697 mutex_lock(&ctx
->mutex
);
5700 perf_install_in_context(ctx
, group_leader
, cpu
);
5702 list_for_each_entry(sibling
, &group_leader
->sibling_list
,
5704 perf_install_in_context(ctx
, sibling
, cpu
);
5709 perf_install_in_context(ctx
, event
, cpu
);
5711 mutex_unlock(&ctx
->mutex
);
5713 event
->owner
= current
;
5715 mutex_lock(¤t
->perf_event_mutex
);
5716 list_add_tail(&event
->owner_entry
, ¤t
->perf_event_list
);
5717 mutex_unlock(¤t
->perf_event_mutex
);
5720 * Drop the reference on the group_event after placing the
5721 * new event on the sibling_list. This ensures destruction
5722 * of the group leader will find the pointer to itself in
5723 * perf_group_detach().
5725 fput_light(group_file
, fput_needed
);
5726 fd_install(event_fd
, event_file
);
5735 put_task_struct(task
);
5737 fput_light(group_file
, fput_needed
);
5739 put_unused_fd(event_fd
);
5744 * perf_event_create_kernel_counter
5746 * @attr: attributes of the counter to create
5747 * @cpu: cpu in which the counter is bound
5748 * @task: task to profile (NULL for percpu)
5751 perf_event_create_kernel_counter(struct perf_event_attr
*attr
, int cpu
,
5752 struct task_struct
*task
,
5753 perf_overflow_handler_t overflow_handler
)
5755 struct perf_event_context
*ctx
;
5756 struct perf_event
*event
;
5760 * Get the target context (task or percpu):
5763 event
= perf_event_alloc(attr
, cpu
, task
, NULL
, NULL
, overflow_handler
);
5764 if (IS_ERR(event
)) {
5765 err
= PTR_ERR(event
);
5769 ctx
= find_get_context(event
->pmu
, task
, cpu
);
5776 WARN_ON_ONCE(ctx
->parent_ctx
);
5777 mutex_lock(&ctx
->mutex
);
5778 perf_install_in_context(ctx
, event
, cpu
);
5780 mutex_unlock(&ctx
->mutex
);
5787 return ERR_PTR(err
);
5789 EXPORT_SYMBOL_GPL(perf_event_create_kernel_counter
);
5791 static void sync_child_event(struct perf_event
*child_event
,
5792 struct task_struct
*child
)
5794 struct perf_event
*parent_event
= child_event
->parent
;
5797 if (child_event
->attr
.inherit_stat
)
5798 perf_event_read_event(child_event
, child
);
5800 child_val
= perf_event_count(child_event
);
5803 * Add back the child's count to the parent's count:
5805 atomic64_add(child_val
, &parent_event
->child_count
);
5806 atomic64_add(child_event
->total_time_enabled
,
5807 &parent_event
->child_total_time_enabled
);
5808 atomic64_add(child_event
->total_time_running
,
5809 &parent_event
->child_total_time_running
);
5812 * Remove this event from the parent's list
5814 WARN_ON_ONCE(parent_event
->ctx
->parent_ctx
);
5815 mutex_lock(&parent_event
->child_mutex
);
5816 list_del_init(&child_event
->child_list
);
5817 mutex_unlock(&parent_event
->child_mutex
);
5820 * Release the parent event, if this was the last
5823 fput(parent_event
->filp
);
5827 __perf_event_exit_task(struct perf_event
*child_event
,
5828 struct perf_event_context
*child_ctx
,
5829 struct task_struct
*child
)
5831 struct perf_event
*parent_event
;
5833 perf_event_remove_from_context(child_event
);
5835 parent_event
= child_event
->parent
;
5837 * It can happen that parent exits first, and has events
5838 * that are still around due to the child reference. These
5839 * events need to be zapped - but otherwise linger.
5842 sync_child_event(child_event
, child
);
5843 free_event(child_event
);
5847 static void perf_event_exit_task_context(struct task_struct
*child
, int ctxn
)
5849 struct perf_event
*child_event
, *tmp
;
5850 struct perf_event_context
*child_ctx
;
5851 unsigned long flags
;
5853 if (likely(!child
->perf_event_ctxp
[ctxn
])) {
5854 perf_event_task(child
, NULL
, 0);
5858 local_irq_save(flags
);
5860 * We can't reschedule here because interrupts are disabled,
5861 * and either child is current or it is a task that can't be
5862 * scheduled, so we are now safe from rescheduling changing
5865 child_ctx
= child
->perf_event_ctxp
[ctxn
];
5866 task_ctx_sched_out(child_ctx
, EVENT_ALL
);
5869 * Take the context lock here so that if find_get_context is
5870 * reading child->perf_event_ctxp, we wait until it has
5871 * incremented the context's refcount before we do put_ctx below.
5873 raw_spin_lock(&child_ctx
->lock
);
5874 child
->perf_event_ctxp
[ctxn
] = NULL
;
5876 * If this context is a clone; unclone it so it can't get
5877 * swapped to another process while we're removing all
5878 * the events from it.
5880 unclone_ctx(child_ctx
);
5881 update_context_time(child_ctx
);
5882 raw_spin_unlock_irqrestore(&child_ctx
->lock
, flags
);
5885 * Report the task dead after unscheduling the events so that we
5886 * won't get any samples after PERF_RECORD_EXIT. We can however still
5887 * get a few PERF_RECORD_READ events.
5889 perf_event_task(child
, child_ctx
, 0);
5892 * We can recurse on the same lock type through:
5894 * __perf_event_exit_task()
5895 * sync_child_event()
5896 * fput(parent_event->filp)
5898 * mutex_lock(&ctx->mutex)
5900 * But since its the parent context it won't be the same instance.
5902 mutex_lock(&child_ctx
->mutex
);
5905 list_for_each_entry_safe(child_event
, tmp
, &child_ctx
->pinned_groups
,
5907 __perf_event_exit_task(child_event
, child_ctx
, child
);
5909 list_for_each_entry_safe(child_event
, tmp
, &child_ctx
->flexible_groups
,
5911 __perf_event_exit_task(child_event
, child_ctx
, child
);
5914 * If the last event was a group event, it will have appended all
5915 * its siblings to the list, but we obtained 'tmp' before that which
5916 * will still point to the list head terminating the iteration.
5918 if (!list_empty(&child_ctx
->pinned_groups
) ||
5919 !list_empty(&child_ctx
->flexible_groups
))
5922 mutex_unlock(&child_ctx
->mutex
);
5928 * When a child task exits, feed back event values to parent events.
5930 void perf_event_exit_task(struct task_struct
*child
)
5932 struct perf_event
*event
, *tmp
;
5935 mutex_lock(&child
->perf_event_mutex
);
5936 list_for_each_entry_safe(event
, tmp
, &child
->perf_event_list
,
5938 list_del_init(&event
->owner_entry
);
5941 * Ensure the list deletion is visible before we clear
5942 * the owner, closes a race against perf_release() where
5943 * we need to serialize on the owner->perf_event_mutex.
5946 event
->owner
= NULL
;
5948 mutex_unlock(&child
->perf_event_mutex
);
5950 for_each_task_context_nr(ctxn
)
5951 perf_event_exit_task_context(child
, ctxn
);
5954 static void perf_free_event(struct perf_event
*event
,
5955 struct perf_event_context
*ctx
)
5957 struct perf_event
*parent
= event
->parent
;
5959 if (WARN_ON_ONCE(!parent
))
5962 mutex_lock(&parent
->child_mutex
);
5963 list_del_init(&event
->child_list
);
5964 mutex_unlock(&parent
->child_mutex
);
5968 perf_group_detach(event
);
5969 list_del_event(event
, ctx
);
5974 * free an unexposed, unused context as created by inheritance by
5975 * perf_event_init_task below, used by fork() in case of fail.
5977 void perf_event_free_task(struct task_struct
*task
)
5979 struct perf_event_context
*ctx
;
5980 struct perf_event
*event
, *tmp
;
5983 for_each_task_context_nr(ctxn
) {
5984 ctx
= task
->perf_event_ctxp
[ctxn
];
5988 mutex_lock(&ctx
->mutex
);
5990 list_for_each_entry_safe(event
, tmp
, &ctx
->pinned_groups
,
5992 perf_free_event(event
, ctx
);
5994 list_for_each_entry_safe(event
, tmp
, &ctx
->flexible_groups
,
5996 perf_free_event(event
, ctx
);
5998 if (!list_empty(&ctx
->pinned_groups
) ||
5999 !list_empty(&ctx
->flexible_groups
))
6002 mutex_unlock(&ctx
->mutex
);
6008 void perf_event_delayed_put(struct task_struct
*task
)
6012 for_each_task_context_nr(ctxn
)
6013 WARN_ON_ONCE(task
->perf_event_ctxp
[ctxn
]);
6017 * inherit a event from parent task to child task:
6019 static struct perf_event
*
6020 inherit_event(struct perf_event
*parent_event
,
6021 struct task_struct
*parent
,
6022 struct perf_event_context
*parent_ctx
,
6023 struct task_struct
*child
,
6024 struct perf_event
*group_leader
,
6025 struct perf_event_context
*child_ctx
)
6027 struct perf_event
*child_event
;
6028 unsigned long flags
;
6031 * Instead of creating recursive hierarchies of events,
6032 * we link inherited events back to the original parent,
6033 * which has a filp for sure, which we use as the reference
6036 if (parent_event
->parent
)
6037 parent_event
= parent_event
->parent
;
6039 child_event
= perf_event_alloc(&parent_event
->attr
,
6042 group_leader
, parent_event
,
6044 if (IS_ERR(child_event
))
6049 * Make the child state follow the state of the parent event,
6050 * not its attr.disabled bit. We hold the parent's mutex,
6051 * so we won't race with perf_event_{en, dis}able_family.
6053 if (parent_event
->state
>= PERF_EVENT_STATE_INACTIVE
)
6054 child_event
->state
= PERF_EVENT_STATE_INACTIVE
;
6056 child_event
->state
= PERF_EVENT_STATE_OFF
;
6058 if (parent_event
->attr
.freq
) {
6059 u64 sample_period
= parent_event
->hw
.sample_period
;
6060 struct hw_perf_event
*hwc
= &child_event
->hw
;
6062 hwc
->sample_period
= sample_period
;
6063 hwc
->last_period
= sample_period
;
6065 local64_set(&hwc
->period_left
, sample_period
);
6068 child_event
->ctx
= child_ctx
;
6069 child_event
->overflow_handler
= parent_event
->overflow_handler
;
6072 * Link it up in the child's context:
6074 raw_spin_lock_irqsave(&child_ctx
->lock
, flags
);
6075 add_event_to_ctx(child_event
, child_ctx
);
6076 raw_spin_unlock_irqrestore(&child_ctx
->lock
, flags
);
6079 * Get a reference to the parent filp - we will fput it
6080 * when the child event exits. This is safe to do because
6081 * we are in the parent and we know that the filp still
6082 * exists and has a nonzero count:
6084 atomic_long_inc(&parent_event
->filp
->f_count
);
6087 * Link this into the parent event's child list
6089 WARN_ON_ONCE(parent_event
->ctx
->parent_ctx
);
6090 mutex_lock(&parent_event
->child_mutex
);
6091 list_add_tail(&child_event
->child_list
, &parent_event
->child_list
);
6092 mutex_unlock(&parent_event
->child_mutex
);
6097 static int inherit_group(struct perf_event
*parent_event
,
6098 struct task_struct
*parent
,
6099 struct perf_event_context
*parent_ctx
,
6100 struct task_struct
*child
,
6101 struct perf_event_context
*child_ctx
)
6103 struct perf_event
*leader
;
6104 struct perf_event
*sub
;
6105 struct perf_event
*child_ctr
;
6107 leader
= inherit_event(parent_event
, parent
, parent_ctx
,
6108 child
, NULL
, child_ctx
);
6110 return PTR_ERR(leader
);
6111 list_for_each_entry(sub
, &parent_event
->sibling_list
, group_entry
) {
6112 child_ctr
= inherit_event(sub
, parent
, parent_ctx
,
6113 child
, leader
, child_ctx
);
6114 if (IS_ERR(child_ctr
))
6115 return PTR_ERR(child_ctr
);
6121 inherit_task_group(struct perf_event
*event
, struct task_struct
*parent
,
6122 struct perf_event_context
*parent_ctx
,
6123 struct task_struct
*child
, int ctxn
,
6127 struct perf_event_context
*child_ctx
;
6129 if (!event
->attr
.inherit
) {
6134 child_ctx
= child
->perf_event_ctxp
[ctxn
];
6137 * This is executed from the parent task context, so
6138 * inherit events that have been marked for cloning.
6139 * First allocate and initialize a context for the
6143 child_ctx
= alloc_perf_context(event
->pmu
, child
);
6147 child
->perf_event_ctxp
[ctxn
] = child_ctx
;
6150 ret
= inherit_group(event
, parent
, parent_ctx
,
6160 * Initialize the perf_event context in task_struct
6162 int perf_event_init_context(struct task_struct
*child
, int ctxn
)
6164 struct perf_event_context
*child_ctx
, *parent_ctx
;
6165 struct perf_event_context
*cloned_ctx
;
6166 struct perf_event
*event
;
6167 struct task_struct
*parent
= current
;
6168 int inherited_all
= 1;
6169 unsigned long flags
;
6172 child
->perf_event_ctxp
[ctxn
] = NULL
;
6174 mutex_init(&child
->perf_event_mutex
);
6175 INIT_LIST_HEAD(&child
->perf_event_list
);
6177 if (likely(!parent
->perf_event_ctxp
[ctxn
]))
6181 * If the parent's context is a clone, pin it so it won't get
6184 parent_ctx
= perf_pin_task_context(parent
, ctxn
);
6187 * No need to check if parent_ctx != NULL here; since we saw
6188 * it non-NULL earlier, the only reason for it to become NULL
6189 * is if we exit, and since we're currently in the middle of
6190 * a fork we can't be exiting at the same time.
6194 * Lock the parent list. No need to lock the child - not PID
6195 * hashed yet and not running, so nobody can access it.
6197 mutex_lock(&parent_ctx
->mutex
);
6200 * We dont have to disable NMIs - we are only looking at
6201 * the list, not manipulating it:
6203 list_for_each_entry(event
, &parent_ctx
->pinned_groups
, group_entry
) {
6204 ret
= inherit_task_group(event
, parent
, parent_ctx
,
6205 child
, ctxn
, &inherited_all
);
6211 * We can't hold ctx->lock when iterating the ->flexible_group list due
6212 * to allocations, but we need to prevent rotation because
6213 * rotate_ctx() will change the list from interrupt context.
6215 raw_spin_lock_irqsave(&parent_ctx
->lock
, flags
);
6216 parent_ctx
->rotate_disable
= 1;
6217 raw_spin_unlock_irqrestore(&parent_ctx
->lock
, flags
);
6219 list_for_each_entry(event
, &parent_ctx
->flexible_groups
, group_entry
) {
6220 ret
= inherit_task_group(event
, parent
, parent_ctx
,
6221 child
, ctxn
, &inherited_all
);
6226 raw_spin_lock_irqsave(&parent_ctx
->lock
, flags
);
6227 parent_ctx
->rotate_disable
= 0;
6228 raw_spin_unlock_irqrestore(&parent_ctx
->lock
, flags
);
6230 child_ctx
= child
->perf_event_ctxp
[ctxn
];
6232 if (child_ctx
&& inherited_all
) {
6234 * Mark the child context as a clone of the parent
6235 * context, or of whatever the parent is a clone of.
6236 * Note that if the parent is a clone, it could get
6237 * uncloned at any point, but that doesn't matter
6238 * because the list of events and the generation
6239 * count can't have changed since we took the mutex.
6241 cloned_ctx
= rcu_dereference(parent_ctx
->parent_ctx
);
6243 child_ctx
->parent_ctx
= cloned_ctx
;
6244 child_ctx
->parent_gen
= parent_ctx
->parent_gen
;
6246 child_ctx
->parent_ctx
= parent_ctx
;
6247 child_ctx
->parent_gen
= parent_ctx
->generation
;
6249 get_ctx(child_ctx
->parent_ctx
);
6252 mutex_unlock(&parent_ctx
->mutex
);
6254 perf_unpin_context(parent_ctx
);
6260 * Initialize the perf_event context in task_struct
6262 int perf_event_init_task(struct task_struct
*child
)
6266 for_each_task_context_nr(ctxn
) {
6267 ret
= perf_event_init_context(child
, ctxn
);
6275 static void __init
perf_event_init_all_cpus(void)
6277 struct swevent_htable
*swhash
;
6280 for_each_possible_cpu(cpu
) {
6281 swhash
= &per_cpu(swevent_htable
, cpu
);
6282 mutex_init(&swhash
->hlist_mutex
);
6283 INIT_LIST_HEAD(&per_cpu(rotation_list
, cpu
));
6287 static void __cpuinit
perf_event_init_cpu(int cpu
)
6289 struct swevent_htable
*swhash
= &per_cpu(swevent_htable
, cpu
);
6291 mutex_lock(&swhash
->hlist_mutex
);
6292 if (swhash
->hlist_refcount
> 0) {
6293 struct swevent_hlist
*hlist
;
6295 hlist
= kzalloc_node(sizeof(*hlist
), GFP_KERNEL
, cpu_to_node(cpu
));
6297 rcu_assign_pointer(swhash
->swevent_hlist
, hlist
);
6299 mutex_unlock(&swhash
->hlist_mutex
);
6302 #ifdef CONFIG_HOTPLUG_CPU
6303 static void perf_pmu_rotate_stop(struct pmu
*pmu
)
6305 struct perf_cpu_context
*cpuctx
= this_cpu_ptr(pmu
->pmu_cpu_context
);
6307 WARN_ON(!irqs_disabled());
6309 list_del_init(&cpuctx
->rotation_list
);
6312 static void __perf_event_exit_context(void *__info
)
6314 struct perf_event_context
*ctx
= __info
;
6315 struct perf_event
*event
, *tmp
;
6317 perf_pmu_rotate_stop(ctx
->pmu
);
6319 list_for_each_entry_safe(event
, tmp
, &ctx
->pinned_groups
, group_entry
)
6320 __perf_event_remove_from_context(event
);
6321 list_for_each_entry_safe(event
, tmp
, &ctx
->flexible_groups
, group_entry
)
6322 __perf_event_remove_from_context(event
);
6325 static void perf_event_exit_cpu_context(int cpu
)
6327 struct perf_event_context
*ctx
;
6331 idx
= srcu_read_lock(&pmus_srcu
);
6332 list_for_each_entry_rcu(pmu
, &pmus
, entry
) {
6333 ctx
= &per_cpu_ptr(pmu
->pmu_cpu_context
, cpu
)->ctx
;
6335 mutex_lock(&ctx
->mutex
);
6336 smp_call_function_single(cpu
, __perf_event_exit_context
, ctx
, 1);
6337 mutex_unlock(&ctx
->mutex
);
6339 srcu_read_unlock(&pmus_srcu
, idx
);
6342 static void perf_event_exit_cpu(int cpu
)
6344 struct swevent_htable
*swhash
= &per_cpu(swevent_htable
, cpu
);
6346 mutex_lock(&swhash
->hlist_mutex
);
6347 swevent_hlist_release(swhash
);
6348 mutex_unlock(&swhash
->hlist_mutex
);
6350 perf_event_exit_cpu_context(cpu
);
6353 static inline void perf_event_exit_cpu(int cpu
) { }
6356 static int __cpuinit
6357 perf_cpu_notify(struct notifier_block
*self
, unsigned long action
, void *hcpu
)
6359 unsigned int cpu
= (long)hcpu
;
6361 switch (action
& ~CPU_TASKS_FROZEN
) {
6363 case CPU_UP_PREPARE
:
6364 case CPU_DOWN_FAILED
:
6365 perf_event_init_cpu(cpu
);
6368 case CPU_UP_CANCELED
:
6369 case CPU_DOWN_PREPARE
:
6370 perf_event_exit_cpu(cpu
);
6380 void __init
perf_event_init(void)
6384 perf_event_init_all_cpus();
6385 init_srcu_struct(&pmus_srcu
);
6386 perf_pmu_register(&perf_swevent
);
6387 perf_pmu_register(&perf_cpu_clock
);
6388 perf_pmu_register(&perf_task_clock
);
6390 perf_cpu_notifier(perf_cpu_notify
);
6392 ret
= init_hw_breakpoint();
6393 WARN(ret
, "hw_breakpoint initialization failed with: %d", ret
);