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/sysfs.h>
19 #include <linux/dcache.h>
20 #include <linux/percpu.h>
21 #include <linux/ptrace.h>
22 #include <linux/vmstat.h>
23 #include <linux/vmalloc.h>
24 #include <linux/hardirq.h>
25 #include <linux/rculist.h>
26 #include <linux/uaccess.h>
27 #include <linux/syscalls.h>
28 #include <linux/anon_inodes.h>
29 #include <linux/kernel_stat.h>
30 #include <linux/perf_event.h>
31 #include <linux/ftrace_event.h>
32 #include <linux/hw_breakpoint.h>
34 #include <asm/irq_regs.h>
37 * Each CPU has a list of per CPU events:
39 DEFINE_PER_CPU(struct perf_cpu_context
, perf_cpu_context
);
41 int perf_max_events __read_mostly
= 1;
42 static int perf_reserved_percpu __read_mostly
;
43 static int perf_overcommit __read_mostly
= 1;
45 static atomic_t nr_events __read_mostly
;
46 static atomic_t nr_mmap_events __read_mostly
;
47 static atomic_t nr_comm_events __read_mostly
;
48 static atomic_t nr_task_events __read_mostly
;
51 * perf event paranoia level:
52 * -1 - not paranoid at all
53 * 0 - disallow raw tracepoint access for unpriv
54 * 1 - disallow cpu events for unpriv
55 * 2 - disallow kernel profiling for unpriv
57 int sysctl_perf_event_paranoid __read_mostly
= 1;
59 static inline bool perf_paranoid_tracepoint_raw(void)
61 return sysctl_perf_event_paranoid
> -1;
64 static inline bool perf_paranoid_cpu(void)
66 return sysctl_perf_event_paranoid
> 0;
69 static inline bool perf_paranoid_kernel(void)
71 return sysctl_perf_event_paranoid
> 1;
74 int sysctl_perf_event_mlock __read_mostly
= 512; /* 'free' kb per user */
77 * max perf event sample rate
79 int sysctl_perf_event_sample_rate __read_mostly
= 100000;
81 static atomic64_t perf_event_id
;
84 * Lock for (sysadmin-configurable) event reservations:
86 static DEFINE_SPINLOCK(perf_resource_lock
);
89 * Architecture provided APIs - weak aliases:
91 extern __weak
const struct pmu
*hw_perf_event_init(struct perf_event
*event
)
96 void __weak
hw_perf_disable(void) { barrier(); }
97 void __weak
hw_perf_enable(void) { barrier(); }
99 void __weak
hw_perf_event_setup(int cpu
) { barrier(); }
100 void __weak
hw_perf_event_setup_online(int cpu
) { barrier(); }
103 hw_perf_group_sched_in(struct perf_event
*group_leader
,
104 struct perf_cpu_context
*cpuctx
,
105 struct perf_event_context
*ctx
, int cpu
)
110 void __weak
perf_event_print_debug(void) { }
112 static DEFINE_PER_CPU(int, perf_disable_count
);
114 void __perf_disable(void)
116 __get_cpu_var(perf_disable_count
)++;
119 bool __perf_enable(void)
121 return !--__get_cpu_var(perf_disable_count
);
124 void perf_disable(void)
130 void perf_enable(void)
136 static void get_ctx(struct perf_event_context
*ctx
)
138 WARN_ON(!atomic_inc_not_zero(&ctx
->refcount
));
141 static void free_ctx(struct rcu_head
*head
)
143 struct perf_event_context
*ctx
;
145 ctx
= container_of(head
, struct perf_event_context
, rcu_head
);
149 static void put_ctx(struct perf_event_context
*ctx
)
151 if (atomic_dec_and_test(&ctx
->refcount
)) {
153 put_ctx(ctx
->parent_ctx
);
155 put_task_struct(ctx
->task
);
156 call_rcu(&ctx
->rcu_head
, free_ctx
);
160 static void unclone_ctx(struct perf_event_context
*ctx
)
162 if (ctx
->parent_ctx
) {
163 put_ctx(ctx
->parent_ctx
);
164 ctx
->parent_ctx
= NULL
;
169 * If we inherit events we want to return the parent event id
172 static u64
primary_event_id(struct perf_event
*event
)
177 id
= event
->parent
->id
;
183 * Get the perf_event_context for a task and lock it.
184 * This has to cope with with the fact that until it is locked,
185 * the context could get moved to another task.
187 static struct perf_event_context
*
188 perf_lock_task_context(struct task_struct
*task
, unsigned long *flags
)
190 struct perf_event_context
*ctx
;
194 ctx
= rcu_dereference(task
->perf_event_ctxp
);
197 * If this context is a clone of another, it might
198 * get swapped for another underneath us by
199 * perf_event_task_sched_out, though the
200 * rcu_read_lock() protects us from any context
201 * getting freed. Lock the context and check if it
202 * got swapped before we could get the lock, and retry
203 * if so. If we locked the right context, then it
204 * can't get swapped on us any more.
206 spin_lock_irqsave(&ctx
->lock
, *flags
);
207 if (ctx
!= rcu_dereference(task
->perf_event_ctxp
)) {
208 spin_unlock_irqrestore(&ctx
->lock
, *flags
);
212 if (!atomic_inc_not_zero(&ctx
->refcount
)) {
213 spin_unlock_irqrestore(&ctx
->lock
, *flags
);
222 * Get the context for a task and increment its pin_count so it
223 * can't get swapped to another task. This also increments its
224 * reference count so that the context can't get freed.
226 static struct perf_event_context
*perf_pin_task_context(struct task_struct
*task
)
228 struct perf_event_context
*ctx
;
231 ctx
= perf_lock_task_context(task
, &flags
);
234 spin_unlock_irqrestore(&ctx
->lock
, flags
);
239 static void perf_unpin_context(struct perf_event_context
*ctx
)
243 spin_lock_irqsave(&ctx
->lock
, flags
);
245 spin_unlock_irqrestore(&ctx
->lock
, flags
);
250 * Add a event from the lists for its context.
251 * Must be called with ctx->mutex and ctx->lock held.
254 list_add_event(struct perf_event
*event
, struct perf_event_context
*ctx
)
256 struct perf_event
*group_leader
= event
->group_leader
;
259 * Depending on whether it is a standalone or sibling event,
260 * add it straight to the context's event list, or to the group
261 * leader's sibling list:
263 if (group_leader
== event
)
264 list_add_tail(&event
->group_entry
, &ctx
->group_list
);
266 list_add_tail(&event
->group_entry
, &group_leader
->sibling_list
);
267 group_leader
->nr_siblings
++;
270 list_add_rcu(&event
->event_entry
, &ctx
->event_list
);
272 if (event
->attr
.inherit_stat
)
277 * Remove a event from the lists for its context.
278 * Must be called with ctx->mutex and ctx->lock held.
281 list_del_event(struct perf_event
*event
, struct perf_event_context
*ctx
)
283 struct perf_event
*sibling
, *tmp
;
285 if (list_empty(&event
->group_entry
))
288 if (event
->attr
.inherit_stat
)
291 list_del_init(&event
->group_entry
);
292 list_del_rcu(&event
->event_entry
);
294 if (event
->group_leader
!= event
)
295 event
->group_leader
->nr_siblings
--;
298 * If this was a group event with sibling events then
299 * upgrade the siblings to singleton events by adding them
300 * to the context list directly:
302 list_for_each_entry_safe(sibling
, tmp
, &event
->sibling_list
, group_entry
) {
304 list_move_tail(&sibling
->group_entry
, &ctx
->group_list
);
305 sibling
->group_leader
= sibling
;
310 event_sched_out(struct perf_event
*event
,
311 struct perf_cpu_context
*cpuctx
,
312 struct perf_event_context
*ctx
)
314 if (event
->state
!= PERF_EVENT_STATE_ACTIVE
)
317 event
->state
= PERF_EVENT_STATE_INACTIVE
;
318 if (event
->pending_disable
) {
319 event
->pending_disable
= 0;
320 event
->state
= PERF_EVENT_STATE_OFF
;
322 event
->tstamp_stopped
= ctx
->time
;
323 event
->pmu
->disable(event
);
326 if (!is_software_event(event
))
327 cpuctx
->active_oncpu
--;
329 if (event
->attr
.exclusive
|| !cpuctx
->active_oncpu
)
330 cpuctx
->exclusive
= 0;
334 group_sched_out(struct perf_event
*group_event
,
335 struct perf_cpu_context
*cpuctx
,
336 struct perf_event_context
*ctx
)
338 struct perf_event
*event
;
340 if (group_event
->state
!= PERF_EVENT_STATE_ACTIVE
)
343 event_sched_out(group_event
, cpuctx
, ctx
);
346 * Schedule out siblings (if any):
348 list_for_each_entry(event
, &group_event
->sibling_list
, group_entry
)
349 event_sched_out(event
, cpuctx
, ctx
);
351 if (group_event
->attr
.exclusive
)
352 cpuctx
->exclusive
= 0;
356 * Cross CPU call to remove a performance event
358 * We disable the event on the hardware level first. After that we
359 * remove it from the context list.
361 static void __perf_event_remove_from_context(void *info
)
363 struct perf_cpu_context
*cpuctx
= &__get_cpu_var(perf_cpu_context
);
364 struct perf_event
*event
= info
;
365 struct perf_event_context
*ctx
= event
->ctx
;
368 * If this is a task context, we need to check whether it is
369 * the current task context of this cpu. If not it has been
370 * scheduled out before the smp call arrived.
372 if (ctx
->task
&& cpuctx
->task_ctx
!= ctx
)
375 spin_lock(&ctx
->lock
);
377 * Protect the list operation against NMI by disabling the
378 * events on a global level.
382 event_sched_out(event
, cpuctx
, ctx
);
384 list_del_event(event
, ctx
);
388 * Allow more per task events with respect to the
391 cpuctx
->max_pertask
=
392 min(perf_max_events
- ctx
->nr_events
,
393 perf_max_events
- perf_reserved_percpu
);
397 spin_unlock(&ctx
->lock
);
402 * Remove the event from a task's (or a CPU's) list of events.
404 * Must be called with ctx->mutex held.
406 * CPU events are removed with a smp call. For task events we only
407 * call when the task is on a CPU.
409 * If event->ctx is a cloned context, callers must make sure that
410 * every task struct that event->ctx->task could possibly point to
411 * remains valid. This is OK when called from perf_release since
412 * that only calls us on the top-level context, which can't be a clone.
413 * When called from perf_event_exit_task, it's OK because the
414 * context has been detached from its task.
416 static void perf_event_remove_from_context(struct perf_event
*event
)
418 struct perf_event_context
*ctx
= event
->ctx
;
419 struct task_struct
*task
= ctx
->task
;
423 * Per cpu events are removed via an smp call and
424 * the removal is always sucessful.
426 smp_call_function_single(event
->cpu
,
427 __perf_event_remove_from_context
,
433 task_oncpu_function_call(task
, __perf_event_remove_from_context
,
436 spin_lock_irq(&ctx
->lock
);
438 * If the context is active we need to retry the smp call.
440 if (ctx
->nr_active
&& !list_empty(&event
->group_entry
)) {
441 spin_unlock_irq(&ctx
->lock
);
446 * The lock prevents that this context is scheduled in so we
447 * can remove the event safely, if the call above did not
450 if (!list_empty(&event
->group_entry
)) {
451 list_del_event(event
, ctx
);
453 spin_unlock_irq(&ctx
->lock
);
456 static inline u64
perf_clock(void)
458 return cpu_clock(smp_processor_id());
462 * Update the record of the current time in a context.
464 static void update_context_time(struct perf_event_context
*ctx
)
466 u64 now
= perf_clock();
468 ctx
->time
+= now
- ctx
->timestamp
;
469 ctx
->timestamp
= now
;
473 * Update the total_time_enabled and total_time_running fields for a event.
475 static void update_event_times(struct perf_event
*event
)
477 struct perf_event_context
*ctx
= event
->ctx
;
480 if (event
->state
< PERF_EVENT_STATE_INACTIVE
||
481 event
->group_leader
->state
< PERF_EVENT_STATE_INACTIVE
)
484 event
->total_time_enabled
= ctx
->time
- event
->tstamp_enabled
;
486 if (event
->state
== PERF_EVENT_STATE_INACTIVE
)
487 run_end
= event
->tstamp_stopped
;
491 event
->total_time_running
= run_end
- event
->tstamp_running
;
495 * Update total_time_enabled and total_time_running for all events in a group.
497 static void update_group_times(struct perf_event
*leader
)
499 struct perf_event
*event
;
501 update_event_times(leader
);
502 list_for_each_entry(event
, &leader
->sibling_list
, group_entry
)
503 update_event_times(event
);
507 * Cross CPU call to disable a performance event
509 static void __perf_event_disable(void *info
)
511 struct perf_event
*event
= info
;
512 struct perf_cpu_context
*cpuctx
= &__get_cpu_var(perf_cpu_context
);
513 struct perf_event_context
*ctx
= event
->ctx
;
516 * If this is a per-task event, need to check whether this
517 * event's task is the current task on this cpu.
519 if (ctx
->task
&& cpuctx
->task_ctx
!= ctx
)
522 spin_lock(&ctx
->lock
);
525 * If the event is on, turn it off.
526 * If it is in error state, leave it in error state.
528 if (event
->state
>= PERF_EVENT_STATE_INACTIVE
) {
529 update_context_time(ctx
);
530 update_group_times(event
);
531 if (event
== event
->group_leader
)
532 group_sched_out(event
, cpuctx
, ctx
);
534 event_sched_out(event
, cpuctx
, ctx
);
535 event
->state
= PERF_EVENT_STATE_OFF
;
538 spin_unlock(&ctx
->lock
);
544 * If event->ctx is a cloned context, callers must make sure that
545 * every task struct that event->ctx->task could possibly point to
546 * remains valid. This condition is satisifed when called through
547 * perf_event_for_each_child or perf_event_for_each because they
548 * hold the top-level event's child_mutex, so any descendant that
549 * goes to exit will block in sync_child_event.
550 * When called from perf_pending_event it's OK because event->ctx
551 * is the current context on this CPU and preemption is disabled,
552 * hence we can't get into perf_event_task_sched_out for this context.
554 static void perf_event_disable(struct perf_event
*event
)
556 struct perf_event_context
*ctx
= event
->ctx
;
557 struct task_struct
*task
= ctx
->task
;
561 * Disable the event on the cpu that it's on
563 smp_call_function_single(event
->cpu
, __perf_event_disable
,
569 task_oncpu_function_call(task
, __perf_event_disable
, event
);
571 spin_lock_irq(&ctx
->lock
);
573 * If the event is still active, we need to retry the cross-call.
575 if (event
->state
== PERF_EVENT_STATE_ACTIVE
) {
576 spin_unlock_irq(&ctx
->lock
);
581 * Since we have the lock this context can't be scheduled
582 * in, so we can change the state safely.
584 if (event
->state
== PERF_EVENT_STATE_INACTIVE
) {
585 update_group_times(event
);
586 event
->state
= PERF_EVENT_STATE_OFF
;
589 spin_unlock_irq(&ctx
->lock
);
593 event_sched_in(struct perf_event
*event
,
594 struct perf_cpu_context
*cpuctx
,
595 struct perf_event_context
*ctx
,
598 if (event
->state
<= PERF_EVENT_STATE_OFF
)
601 event
->state
= PERF_EVENT_STATE_ACTIVE
;
602 event
->oncpu
= cpu
; /* TODO: put 'cpu' into cpuctx->cpu */
604 * The new state must be visible before we turn it on in the hardware:
608 if (event
->pmu
->enable(event
)) {
609 event
->state
= PERF_EVENT_STATE_INACTIVE
;
614 event
->tstamp_running
+= ctx
->time
- event
->tstamp_stopped
;
616 if (!is_software_event(event
))
617 cpuctx
->active_oncpu
++;
620 if (event
->attr
.exclusive
)
621 cpuctx
->exclusive
= 1;
627 group_sched_in(struct perf_event
*group_event
,
628 struct perf_cpu_context
*cpuctx
,
629 struct perf_event_context
*ctx
,
632 struct perf_event
*event
, *partial_group
;
635 if (group_event
->state
== PERF_EVENT_STATE_OFF
)
638 ret
= hw_perf_group_sched_in(group_event
, cpuctx
, ctx
, cpu
);
640 return ret
< 0 ? ret
: 0;
642 if (event_sched_in(group_event
, cpuctx
, ctx
, cpu
))
646 * Schedule in siblings as one group (if any):
648 list_for_each_entry(event
, &group_event
->sibling_list
, group_entry
) {
649 if (event_sched_in(event
, cpuctx
, ctx
, cpu
)) {
650 partial_group
= event
;
659 * Groups can be scheduled in as one unit only, so undo any
660 * partial group before returning:
662 list_for_each_entry(event
, &group_event
->sibling_list
, group_entry
) {
663 if (event
== partial_group
)
665 event_sched_out(event
, cpuctx
, ctx
);
667 event_sched_out(group_event
, cpuctx
, ctx
);
673 * Return 1 for a group consisting entirely of software events,
674 * 0 if the group contains any hardware events.
676 static int is_software_only_group(struct perf_event
*leader
)
678 struct perf_event
*event
;
680 if (!is_software_event(leader
))
683 list_for_each_entry(event
, &leader
->sibling_list
, group_entry
)
684 if (!is_software_event(event
))
691 * Work out whether we can put this event group on the CPU now.
693 static int group_can_go_on(struct perf_event
*event
,
694 struct perf_cpu_context
*cpuctx
,
698 * Groups consisting entirely of software events can always go on.
700 if (is_software_only_group(event
))
703 * If an exclusive group is already on, no other hardware
706 if (cpuctx
->exclusive
)
709 * If this group is exclusive and there are already
710 * events on the CPU, it can't go on.
712 if (event
->attr
.exclusive
&& cpuctx
->active_oncpu
)
715 * Otherwise, try to add it if all previous groups were able
721 static void add_event_to_ctx(struct perf_event
*event
,
722 struct perf_event_context
*ctx
)
724 list_add_event(event
, ctx
);
725 event
->tstamp_enabled
= ctx
->time
;
726 event
->tstamp_running
= ctx
->time
;
727 event
->tstamp_stopped
= ctx
->time
;
731 * Cross CPU call to install and enable a performance event
733 * Must be called with ctx->mutex held
735 static void __perf_install_in_context(void *info
)
737 struct perf_cpu_context
*cpuctx
= &__get_cpu_var(perf_cpu_context
);
738 struct perf_event
*event
= info
;
739 struct perf_event_context
*ctx
= event
->ctx
;
740 struct perf_event
*leader
= event
->group_leader
;
741 int cpu
= smp_processor_id();
745 * If this is a task context, we need to check whether it is
746 * the current task context of this cpu. If not it has been
747 * scheduled out before the smp call arrived.
748 * Or possibly this is the right context but it isn't
749 * on this cpu because it had no events.
751 if (ctx
->task
&& cpuctx
->task_ctx
!= ctx
) {
752 if (cpuctx
->task_ctx
|| ctx
->task
!= current
)
754 cpuctx
->task_ctx
= ctx
;
757 spin_lock(&ctx
->lock
);
759 update_context_time(ctx
);
762 * Protect the list operation against NMI by disabling the
763 * events on a global level. NOP for non NMI based events.
767 add_event_to_ctx(event
, ctx
);
770 * Don't put the event on if it is disabled or if
771 * it is in a group and the group isn't on.
773 if (event
->state
!= PERF_EVENT_STATE_INACTIVE
||
774 (leader
!= event
&& leader
->state
!= PERF_EVENT_STATE_ACTIVE
))
778 * An exclusive event can't go on if there are already active
779 * hardware events, and no hardware event can go on if there
780 * is already an exclusive event on.
782 if (!group_can_go_on(event
, cpuctx
, 1))
785 err
= event_sched_in(event
, cpuctx
, ctx
, cpu
);
789 * This event couldn't go on. If it is in a group
790 * then we have to pull the whole group off.
791 * If the event group is pinned then put it in error state.
794 group_sched_out(leader
, cpuctx
, ctx
);
795 if (leader
->attr
.pinned
) {
796 update_group_times(leader
);
797 leader
->state
= PERF_EVENT_STATE_ERROR
;
801 if (!err
&& !ctx
->task
&& cpuctx
->max_pertask
)
802 cpuctx
->max_pertask
--;
807 spin_unlock(&ctx
->lock
);
811 * Attach a performance event to a context
813 * First we add the event to the list with the hardware enable bit
814 * in event->hw_config cleared.
816 * If the event is attached to a task which is on a CPU we use a smp
817 * call to enable it in the task context. The task might have been
818 * scheduled away, but we check this in the smp call again.
820 * Must be called with ctx->mutex held.
823 perf_install_in_context(struct perf_event_context
*ctx
,
824 struct perf_event
*event
,
827 struct task_struct
*task
= ctx
->task
;
831 * Per cpu events are installed via an smp call and
832 * the install is always sucessful.
834 smp_call_function_single(cpu
, __perf_install_in_context
,
840 task_oncpu_function_call(task
, __perf_install_in_context
,
843 spin_lock_irq(&ctx
->lock
);
845 * we need to retry the smp call.
847 if (ctx
->is_active
&& list_empty(&event
->group_entry
)) {
848 spin_unlock_irq(&ctx
->lock
);
853 * The lock prevents that this context is scheduled in so we
854 * can add the event safely, if it the call above did not
857 if (list_empty(&event
->group_entry
))
858 add_event_to_ctx(event
, ctx
);
859 spin_unlock_irq(&ctx
->lock
);
863 * Put a event into inactive state and update time fields.
864 * Enabling the leader of a group effectively enables all
865 * the group members that aren't explicitly disabled, so we
866 * have to update their ->tstamp_enabled also.
867 * Note: this works for group members as well as group leaders
868 * since the non-leader members' sibling_lists will be empty.
870 static void __perf_event_mark_enabled(struct perf_event
*event
,
871 struct perf_event_context
*ctx
)
873 struct perf_event
*sub
;
875 event
->state
= PERF_EVENT_STATE_INACTIVE
;
876 event
->tstamp_enabled
= ctx
->time
- event
->total_time_enabled
;
877 list_for_each_entry(sub
, &event
->sibling_list
, group_entry
)
878 if (sub
->state
>= PERF_EVENT_STATE_INACTIVE
)
879 sub
->tstamp_enabled
=
880 ctx
->time
- sub
->total_time_enabled
;
884 * Cross CPU call to enable a performance event
886 static void __perf_event_enable(void *info
)
888 struct perf_event
*event
= info
;
889 struct perf_cpu_context
*cpuctx
= &__get_cpu_var(perf_cpu_context
);
890 struct perf_event_context
*ctx
= event
->ctx
;
891 struct perf_event
*leader
= event
->group_leader
;
895 * If this is a per-task event, need to check whether this
896 * event's task is the current task on this cpu.
898 if (ctx
->task
&& cpuctx
->task_ctx
!= ctx
) {
899 if (cpuctx
->task_ctx
|| ctx
->task
!= current
)
901 cpuctx
->task_ctx
= ctx
;
904 spin_lock(&ctx
->lock
);
906 update_context_time(ctx
);
908 if (event
->state
>= PERF_EVENT_STATE_INACTIVE
)
910 __perf_event_mark_enabled(event
, ctx
);
913 * If the event is in a group and isn't the group leader,
914 * then don't put it on unless the group is on.
916 if (leader
!= event
&& leader
->state
!= PERF_EVENT_STATE_ACTIVE
)
919 if (!group_can_go_on(event
, cpuctx
, 1)) {
924 err
= group_sched_in(event
, cpuctx
, ctx
,
927 err
= event_sched_in(event
, cpuctx
, ctx
,
934 * If this event can't go on and it's part of a
935 * group, then the whole group has to come off.
938 group_sched_out(leader
, cpuctx
, ctx
);
939 if (leader
->attr
.pinned
) {
940 update_group_times(leader
);
941 leader
->state
= PERF_EVENT_STATE_ERROR
;
946 spin_unlock(&ctx
->lock
);
952 * If event->ctx is a cloned context, callers must make sure that
953 * every task struct that event->ctx->task could possibly point to
954 * remains valid. This condition is satisfied when called through
955 * perf_event_for_each_child or perf_event_for_each as described
956 * for perf_event_disable.
958 static void perf_event_enable(struct perf_event
*event
)
960 struct perf_event_context
*ctx
= event
->ctx
;
961 struct task_struct
*task
= ctx
->task
;
965 * Enable the event on the cpu that it's on
967 smp_call_function_single(event
->cpu
, __perf_event_enable
,
972 spin_lock_irq(&ctx
->lock
);
973 if (event
->state
>= PERF_EVENT_STATE_INACTIVE
)
977 * If the event is in error state, clear that first.
978 * That way, if we see the event in error state below, we
979 * know that it has gone back into error state, as distinct
980 * from the task having been scheduled away before the
981 * cross-call arrived.
983 if (event
->state
== PERF_EVENT_STATE_ERROR
)
984 event
->state
= PERF_EVENT_STATE_OFF
;
987 spin_unlock_irq(&ctx
->lock
);
988 task_oncpu_function_call(task
, __perf_event_enable
, event
);
990 spin_lock_irq(&ctx
->lock
);
993 * If the context is active and the event is still off,
994 * we need to retry the cross-call.
996 if (ctx
->is_active
&& event
->state
== PERF_EVENT_STATE_OFF
)
1000 * Since we have the lock this context can't be scheduled
1001 * in, so we can change the state safely.
1003 if (event
->state
== PERF_EVENT_STATE_OFF
)
1004 __perf_event_mark_enabled(event
, ctx
);
1007 spin_unlock_irq(&ctx
->lock
);
1010 static int perf_event_refresh(struct perf_event
*event
, int refresh
)
1013 * not supported on inherited events
1015 if (event
->attr
.inherit
)
1018 atomic_add(refresh
, &event
->event_limit
);
1019 perf_event_enable(event
);
1024 void __perf_event_sched_out(struct perf_event_context
*ctx
,
1025 struct perf_cpu_context
*cpuctx
)
1027 struct perf_event
*event
;
1029 spin_lock(&ctx
->lock
);
1031 if (likely(!ctx
->nr_events
))
1033 update_context_time(ctx
);
1037 list_for_each_entry(event
, &ctx
->group_list
, group_entry
)
1038 group_sched_out(event
, cpuctx
, ctx
);
1042 spin_unlock(&ctx
->lock
);
1046 * Test whether two contexts are equivalent, i.e. whether they
1047 * have both been cloned from the same version of the same context
1048 * and they both have the same number of enabled events.
1049 * If the number of enabled events is the same, then the set
1050 * of enabled events should be the same, because these are both
1051 * inherited contexts, therefore we can't access individual events
1052 * in them directly with an fd; we can only enable/disable all
1053 * events via prctl, or enable/disable all events in a family
1054 * via ioctl, which will have the same effect on both contexts.
1056 static int context_equiv(struct perf_event_context
*ctx1
,
1057 struct perf_event_context
*ctx2
)
1059 return ctx1
->parent_ctx
&& ctx1
->parent_ctx
== ctx2
->parent_ctx
1060 && ctx1
->parent_gen
== ctx2
->parent_gen
1061 && !ctx1
->pin_count
&& !ctx2
->pin_count
;
1064 static void __perf_event_sync_stat(struct perf_event
*event
,
1065 struct perf_event
*next_event
)
1069 if (!event
->attr
.inherit_stat
)
1073 * Update the event value, we cannot use perf_event_read()
1074 * because we're in the middle of a context switch and have IRQs
1075 * disabled, which upsets smp_call_function_single(), however
1076 * we know the event must be on the current CPU, therefore we
1077 * don't need to use it.
1079 switch (event
->state
) {
1080 case PERF_EVENT_STATE_ACTIVE
:
1081 event
->pmu
->read(event
);
1084 case PERF_EVENT_STATE_INACTIVE
:
1085 update_event_times(event
);
1093 * In order to keep per-task stats reliable we need to flip the event
1094 * values when we flip the contexts.
1096 value
= atomic64_read(&next_event
->count
);
1097 value
= atomic64_xchg(&event
->count
, value
);
1098 atomic64_set(&next_event
->count
, value
);
1100 swap(event
->total_time_enabled
, next_event
->total_time_enabled
);
1101 swap(event
->total_time_running
, next_event
->total_time_running
);
1104 * Since we swizzled the values, update the user visible data too.
1106 perf_event_update_userpage(event
);
1107 perf_event_update_userpage(next_event
);
1110 #define list_next_entry(pos, member) \
1111 list_entry(pos->member.next, typeof(*pos), member)
1113 static void perf_event_sync_stat(struct perf_event_context
*ctx
,
1114 struct perf_event_context
*next_ctx
)
1116 struct perf_event
*event
, *next_event
;
1121 update_context_time(ctx
);
1123 event
= list_first_entry(&ctx
->event_list
,
1124 struct perf_event
, event_entry
);
1126 next_event
= list_first_entry(&next_ctx
->event_list
,
1127 struct perf_event
, event_entry
);
1129 while (&event
->event_entry
!= &ctx
->event_list
&&
1130 &next_event
->event_entry
!= &next_ctx
->event_list
) {
1132 __perf_event_sync_stat(event
, next_event
);
1134 event
= list_next_entry(event
, event_entry
);
1135 next_event
= list_next_entry(next_event
, event_entry
);
1140 * Called from scheduler to remove the events of the current task,
1141 * with interrupts disabled.
1143 * We stop each event and update the event value in event->count.
1145 * This does not protect us against NMI, but disable()
1146 * sets the disabled bit in the control field of event _before_
1147 * accessing the event control register. If a NMI hits, then it will
1148 * not restart the event.
1150 void perf_event_task_sched_out(struct task_struct
*task
,
1151 struct task_struct
*next
, int cpu
)
1153 struct perf_cpu_context
*cpuctx
= &per_cpu(perf_cpu_context
, cpu
);
1154 struct perf_event_context
*ctx
= task
->perf_event_ctxp
;
1155 struct perf_event_context
*next_ctx
;
1156 struct perf_event_context
*parent
;
1157 struct pt_regs
*regs
;
1160 regs
= task_pt_regs(task
);
1161 perf_sw_event(PERF_COUNT_SW_CONTEXT_SWITCHES
, 1, 1, regs
, 0);
1163 if (likely(!ctx
|| !cpuctx
->task_ctx
))
1167 parent
= rcu_dereference(ctx
->parent_ctx
);
1168 next_ctx
= next
->perf_event_ctxp
;
1169 if (parent
&& next_ctx
&&
1170 rcu_dereference(next_ctx
->parent_ctx
) == parent
) {
1172 * Looks like the two contexts are clones, so we might be
1173 * able to optimize the context switch. We lock both
1174 * contexts and check that they are clones under the
1175 * lock (including re-checking that neither has been
1176 * uncloned in the meantime). It doesn't matter which
1177 * order we take the locks because no other cpu could
1178 * be trying to lock both of these tasks.
1180 spin_lock(&ctx
->lock
);
1181 spin_lock_nested(&next_ctx
->lock
, SINGLE_DEPTH_NESTING
);
1182 if (context_equiv(ctx
, next_ctx
)) {
1184 * XXX do we need a memory barrier of sorts
1185 * wrt to rcu_dereference() of perf_event_ctxp
1187 task
->perf_event_ctxp
= next_ctx
;
1188 next
->perf_event_ctxp
= ctx
;
1190 next_ctx
->task
= task
;
1193 perf_event_sync_stat(ctx
, next_ctx
);
1195 spin_unlock(&next_ctx
->lock
);
1196 spin_unlock(&ctx
->lock
);
1201 __perf_event_sched_out(ctx
, cpuctx
);
1202 cpuctx
->task_ctx
= NULL
;
1207 * Called with IRQs disabled
1209 static void __perf_event_task_sched_out(struct perf_event_context
*ctx
)
1211 struct perf_cpu_context
*cpuctx
= &__get_cpu_var(perf_cpu_context
);
1213 if (!cpuctx
->task_ctx
)
1216 if (WARN_ON_ONCE(ctx
!= cpuctx
->task_ctx
))
1219 __perf_event_sched_out(ctx
, cpuctx
);
1220 cpuctx
->task_ctx
= NULL
;
1224 * Called with IRQs disabled
1226 static void perf_event_cpu_sched_out(struct perf_cpu_context
*cpuctx
)
1228 __perf_event_sched_out(&cpuctx
->ctx
, cpuctx
);
1232 __perf_event_sched_in(struct perf_event_context
*ctx
,
1233 struct perf_cpu_context
*cpuctx
, int cpu
)
1235 struct perf_event
*event
;
1238 spin_lock(&ctx
->lock
);
1240 if (likely(!ctx
->nr_events
))
1243 ctx
->timestamp
= perf_clock();
1248 * First go through the list and put on any pinned groups
1249 * in order to give them the best chance of going on.
1251 list_for_each_entry(event
, &ctx
->group_list
, group_entry
) {
1252 if (event
->state
<= PERF_EVENT_STATE_OFF
||
1253 !event
->attr
.pinned
)
1255 if (event
->cpu
!= -1 && event
->cpu
!= cpu
)
1258 if (group_can_go_on(event
, cpuctx
, 1))
1259 group_sched_in(event
, cpuctx
, ctx
, cpu
);
1262 * If this pinned group hasn't been scheduled,
1263 * put it in error state.
1265 if (event
->state
== PERF_EVENT_STATE_INACTIVE
) {
1266 update_group_times(event
);
1267 event
->state
= PERF_EVENT_STATE_ERROR
;
1271 list_for_each_entry(event
, &ctx
->group_list
, group_entry
) {
1273 * Ignore events in OFF or ERROR state, and
1274 * ignore pinned events since we did them already.
1276 if (event
->state
<= PERF_EVENT_STATE_OFF
||
1281 * Listen to the 'cpu' scheduling filter constraint
1284 if (event
->cpu
!= -1 && event
->cpu
!= cpu
)
1287 if (group_can_go_on(event
, cpuctx
, can_add_hw
))
1288 if (group_sched_in(event
, cpuctx
, ctx
, cpu
))
1293 spin_unlock(&ctx
->lock
);
1297 * Called from scheduler to add the events of the current task
1298 * with interrupts disabled.
1300 * We restore the event value and then enable it.
1302 * This does not protect us against NMI, but enable()
1303 * sets the enabled bit in the control field of event _before_
1304 * accessing the event control register. If a NMI hits, then it will
1305 * keep the event running.
1307 void perf_event_task_sched_in(struct task_struct
*task
, int cpu
)
1309 struct perf_cpu_context
*cpuctx
= &per_cpu(perf_cpu_context
, cpu
);
1310 struct perf_event_context
*ctx
= task
->perf_event_ctxp
;
1314 if (cpuctx
->task_ctx
== ctx
)
1316 __perf_event_sched_in(ctx
, cpuctx
, cpu
);
1317 cpuctx
->task_ctx
= ctx
;
1320 static void perf_event_cpu_sched_in(struct perf_cpu_context
*cpuctx
, int cpu
)
1322 struct perf_event_context
*ctx
= &cpuctx
->ctx
;
1324 __perf_event_sched_in(ctx
, cpuctx
, cpu
);
1327 #define MAX_INTERRUPTS (~0ULL)
1329 static void perf_log_throttle(struct perf_event
*event
, int enable
);
1331 static void perf_adjust_period(struct perf_event
*event
, u64 events
)
1333 struct hw_perf_event
*hwc
= &event
->hw
;
1334 u64 period
, sample_period
;
1337 events
*= hwc
->sample_period
;
1338 period
= div64_u64(events
, event
->attr
.sample_freq
);
1340 delta
= (s64
)(period
- hwc
->sample_period
);
1341 delta
= (delta
+ 7) / 8; /* low pass filter */
1343 sample_period
= hwc
->sample_period
+ delta
;
1348 hwc
->sample_period
= sample_period
;
1351 static void perf_ctx_adjust_freq(struct perf_event_context
*ctx
)
1353 struct perf_event
*event
;
1354 struct hw_perf_event
*hwc
;
1355 u64 interrupts
, freq
;
1357 spin_lock(&ctx
->lock
);
1358 list_for_each_entry_rcu(event
, &ctx
->event_list
, event_entry
) {
1359 if (event
->state
!= PERF_EVENT_STATE_ACTIVE
)
1364 interrupts
= hwc
->interrupts
;
1365 hwc
->interrupts
= 0;
1368 * unthrottle events on the tick
1370 if (interrupts
== MAX_INTERRUPTS
) {
1371 perf_log_throttle(event
, 1);
1372 event
->pmu
->unthrottle(event
);
1373 interrupts
= 2*sysctl_perf_event_sample_rate
/HZ
;
1376 if (!event
->attr
.freq
|| !event
->attr
.sample_freq
)
1380 * if the specified freq < HZ then we need to skip ticks
1382 if (event
->attr
.sample_freq
< HZ
) {
1383 freq
= event
->attr
.sample_freq
;
1385 hwc
->freq_count
+= freq
;
1386 hwc
->freq_interrupts
+= interrupts
;
1388 if (hwc
->freq_count
< HZ
)
1391 interrupts
= hwc
->freq_interrupts
;
1392 hwc
->freq_interrupts
= 0;
1393 hwc
->freq_count
-= HZ
;
1397 perf_adjust_period(event
, freq
* interrupts
);
1400 * In order to avoid being stalled by an (accidental) huge
1401 * sample period, force reset the sample period if we didn't
1402 * get any events in this freq period.
1406 event
->pmu
->disable(event
);
1407 atomic64_set(&hwc
->period_left
, 0);
1408 event
->pmu
->enable(event
);
1412 spin_unlock(&ctx
->lock
);
1416 * Round-robin a context's events:
1418 static void rotate_ctx(struct perf_event_context
*ctx
)
1420 struct perf_event
*event
;
1422 if (!ctx
->nr_events
)
1425 spin_lock(&ctx
->lock
);
1427 * Rotate the first entry last (works just fine for group events too):
1430 list_for_each_entry(event
, &ctx
->group_list
, group_entry
) {
1431 list_move_tail(&event
->group_entry
, &ctx
->group_list
);
1436 spin_unlock(&ctx
->lock
);
1439 void perf_event_task_tick(struct task_struct
*curr
, int cpu
)
1441 struct perf_cpu_context
*cpuctx
;
1442 struct perf_event_context
*ctx
;
1444 if (!atomic_read(&nr_events
))
1447 cpuctx
= &per_cpu(perf_cpu_context
, cpu
);
1448 ctx
= curr
->perf_event_ctxp
;
1450 perf_ctx_adjust_freq(&cpuctx
->ctx
);
1452 perf_ctx_adjust_freq(ctx
);
1454 perf_event_cpu_sched_out(cpuctx
);
1456 __perf_event_task_sched_out(ctx
);
1458 rotate_ctx(&cpuctx
->ctx
);
1462 perf_event_cpu_sched_in(cpuctx
, cpu
);
1464 perf_event_task_sched_in(curr
, cpu
);
1468 * Enable all of a task's events that have been marked enable-on-exec.
1469 * This expects task == current.
1471 static void perf_event_enable_on_exec(struct task_struct
*task
)
1473 struct perf_event_context
*ctx
;
1474 struct perf_event
*event
;
1475 unsigned long flags
;
1478 local_irq_save(flags
);
1479 ctx
= task
->perf_event_ctxp
;
1480 if (!ctx
|| !ctx
->nr_events
)
1483 __perf_event_task_sched_out(ctx
);
1485 spin_lock(&ctx
->lock
);
1487 list_for_each_entry(event
, &ctx
->group_list
, group_entry
) {
1488 if (!event
->attr
.enable_on_exec
)
1490 event
->attr
.enable_on_exec
= 0;
1491 if (event
->state
>= PERF_EVENT_STATE_INACTIVE
)
1493 __perf_event_mark_enabled(event
, ctx
);
1498 * Unclone this context if we enabled any event.
1503 spin_unlock(&ctx
->lock
);
1505 perf_event_task_sched_in(task
, smp_processor_id());
1507 local_irq_restore(flags
);
1511 * Cross CPU call to read the hardware event
1513 static void __perf_event_read(void *info
)
1515 struct perf_cpu_context
*cpuctx
= &__get_cpu_var(perf_cpu_context
);
1516 struct perf_event
*event
= info
;
1517 struct perf_event_context
*ctx
= event
->ctx
;
1520 * If this is a task context, we need to check whether it is
1521 * the current task context of this cpu. If not it has been
1522 * scheduled out before the smp call arrived. In that case
1523 * event->count would have been updated to a recent sample
1524 * when the event was scheduled out.
1526 if (ctx
->task
&& cpuctx
->task_ctx
!= ctx
)
1529 spin_lock(&ctx
->lock
);
1530 update_context_time(ctx
);
1531 update_event_times(event
);
1532 spin_unlock(&ctx
->lock
);
1534 event
->pmu
->read(event
);
1537 static u64
perf_event_read(struct perf_event
*event
)
1540 * If event is enabled and currently active on a CPU, update the
1541 * value in the event structure:
1543 if (event
->state
== PERF_EVENT_STATE_ACTIVE
) {
1544 smp_call_function_single(event
->oncpu
,
1545 __perf_event_read
, event
, 1);
1546 } else if (event
->state
== PERF_EVENT_STATE_INACTIVE
) {
1547 struct perf_event_context
*ctx
= event
->ctx
;
1548 unsigned long flags
;
1550 spin_lock_irqsave(&ctx
->lock
, flags
);
1551 update_context_time(ctx
);
1552 update_event_times(event
);
1553 spin_unlock_irqrestore(&ctx
->lock
, flags
);
1556 return atomic64_read(&event
->count
);
1560 * Initialize the perf_event context in a task_struct:
1563 __perf_event_init_context(struct perf_event_context
*ctx
,
1564 struct task_struct
*task
)
1566 memset(ctx
, 0, sizeof(*ctx
));
1567 spin_lock_init(&ctx
->lock
);
1568 mutex_init(&ctx
->mutex
);
1569 INIT_LIST_HEAD(&ctx
->group_list
);
1570 INIT_LIST_HEAD(&ctx
->event_list
);
1571 atomic_set(&ctx
->refcount
, 1);
1575 static struct perf_event_context
*find_get_context(pid_t pid
, int cpu
)
1577 struct perf_event_context
*ctx
;
1578 struct perf_cpu_context
*cpuctx
;
1579 struct task_struct
*task
;
1580 unsigned long flags
;
1584 * If cpu is not a wildcard then this is a percpu event:
1587 /* Must be root to operate on a CPU event: */
1588 if (perf_paranoid_cpu() && !capable(CAP_SYS_ADMIN
))
1589 return ERR_PTR(-EACCES
);
1591 if (cpu
< 0 || cpu
> num_possible_cpus())
1592 return ERR_PTR(-EINVAL
);
1595 * We could be clever and allow to attach a event to an
1596 * offline CPU and activate it when the CPU comes up, but
1599 if (!cpu_isset(cpu
, cpu_online_map
))
1600 return ERR_PTR(-ENODEV
);
1602 cpuctx
= &per_cpu(perf_cpu_context
, cpu
);
1613 task
= find_task_by_vpid(pid
);
1615 get_task_struct(task
);
1619 return ERR_PTR(-ESRCH
);
1622 * Can't attach events to a dying task.
1625 if (task
->flags
& PF_EXITING
)
1628 /* Reuse ptrace permission checks for now. */
1630 if (!ptrace_may_access(task
, PTRACE_MODE_READ
))
1634 ctx
= perf_lock_task_context(task
, &flags
);
1637 spin_unlock_irqrestore(&ctx
->lock
, flags
);
1641 ctx
= kmalloc(sizeof(struct perf_event_context
), GFP_KERNEL
);
1645 __perf_event_init_context(ctx
, task
);
1647 if (cmpxchg(&task
->perf_event_ctxp
, NULL
, ctx
)) {
1649 * We raced with some other task; use
1650 * the context they set.
1655 get_task_struct(task
);
1658 put_task_struct(task
);
1662 put_task_struct(task
);
1663 return ERR_PTR(err
);
1666 static void perf_event_free_filter(struct perf_event
*event
);
1668 static void free_event_rcu(struct rcu_head
*head
)
1670 struct perf_event
*event
;
1672 event
= container_of(head
, struct perf_event
, rcu_head
);
1674 put_pid_ns(event
->ns
);
1675 perf_event_free_filter(event
);
1679 static void perf_pending_sync(struct perf_event
*event
);
1681 static void free_event(struct perf_event
*event
)
1683 perf_pending_sync(event
);
1685 if (!event
->parent
) {
1686 atomic_dec(&nr_events
);
1687 if (event
->attr
.mmap
)
1688 atomic_dec(&nr_mmap_events
);
1689 if (event
->attr
.comm
)
1690 atomic_dec(&nr_comm_events
);
1691 if (event
->attr
.task
)
1692 atomic_dec(&nr_task_events
);
1695 if (event
->output
) {
1696 fput(event
->output
->filp
);
1697 event
->output
= NULL
;
1701 event
->destroy(event
);
1703 put_ctx(event
->ctx
);
1704 call_rcu(&event
->rcu_head
, free_event_rcu
);
1708 * Called when the last reference to the file is gone.
1710 static int perf_release(struct inode
*inode
, struct file
*file
)
1712 struct perf_event
*event
= file
->private_data
;
1713 struct perf_event_context
*ctx
= event
->ctx
;
1715 file
->private_data
= NULL
;
1717 WARN_ON_ONCE(ctx
->parent_ctx
);
1718 mutex_lock(&ctx
->mutex
);
1719 perf_event_remove_from_context(event
);
1720 mutex_unlock(&ctx
->mutex
);
1722 mutex_lock(&event
->owner
->perf_event_mutex
);
1723 list_del_init(&event
->owner_entry
);
1724 mutex_unlock(&event
->owner
->perf_event_mutex
);
1725 put_task_struct(event
->owner
);
1732 int perf_event_release_kernel(struct perf_event
*event
)
1734 struct perf_event_context
*ctx
= event
->ctx
;
1736 WARN_ON_ONCE(ctx
->parent_ctx
);
1737 mutex_lock(&ctx
->mutex
);
1738 perf_event_remove_from_context(event
);
1739 mutex_unlock(&ctx
->mutex
);
1741 mutex_lock(&event
->owner
->perf_event_mutex
);
1742 list_del_init(&event
->owner_entry
);
1743 mutex_unlock(&event
->owner
->perf_event_mutex
);
1744 put_task_struct(event
->owner
);
1750 EXPORT_SYMBOL_GPL(perf_event_release_kernel
);
1752 static int perf_event_read_size(struct perf_event
*event
)
1754 int entry
= sizeof(u64
); /* value */
1758 if (event
->attr
.read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
)
1759 size
+= sizeof(u64
);
1761 if (event
->attr
.read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
)
1762 size
+= sizeof(u64
);
1764 if (event
->attr
.read_format
& PERF_FORMAT_ID
)
1765 entry
+= sizeof(u64
);
1767 if (event
->attr
.read_format
& PERF_FORMAT_GROUP
) {
1768 nr
+= event
->group_leader
->nr_siblings
;
1769 size
+= sizeof(u64
);
1777 u64
perf_event_read_value(struct perf_event
*event
)
1779 struct perf_event
*child
;
1782 total
+= perf_event_read(event
);
1783 list_for_each_entry(child
, &event
->child_list
, child_list
)
1784 total
+= perf_event_read(child
);
1788 EXPORT_SYMBOL_GPL(perf_event_read_value
);
1790 static int perf_event_read_group(struct perf_event
*event
,
1791 u64 read_format
, char __user
*buf
)
1793 struct perf_event
*leader
= event
->group_leader
, *sub
;
1794 int n
= 0, size
= 0, ret
= 0;
1798 count
= perf_event_read_value(leader
);
1800 values
[n
++] = 1 + leader
->nr_siblings
;
1801 if (read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
) {
1802 values
[n
++] = leader
->total_time_enabled
+
1803 atomic64_read(&leader
->child_total_time_enabled
);
1805 if (read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
) {
1806 values
[n
++] = leader
->total_time_running
+
1807 atomic64_read(&leader
->child_total_time_running
);
1809 values
[n
++] = count
;
1810 if (read_format
& PERF_FORMAT_ID
)
1811 values
[n
++] = primary_event_id(leader
);
1813 size
= n
* sizeof(u64
);
1815 if (copy_to_user(buf
, values
, size
))
1820 list_for_each_entry(sub
, &leader
->sibling_list
, group_entry
) {
1823 values
[n
++] = perf_event_read_value(sub
);
1824 if (read_format
& PERF_FORMAT_ID
)
1825 values
[n
++] = primary_event_id(sub
);
1827 size
= n
* sizeof(u64
);
1829 if (copy_to_user(buf
+ size
, values
, size
))
1838 static int perf_event_read_one(struct perf_event
*event
,
1839 u64 read_format
, char __user
*buf
)
1844 values
[n
++] = perf_event_read_value(event
);
1845 if (read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
) {
1846 values
[n
++] = event
->total_time_enabled
+
1847 atomic64_read(&event
->child_total_time_enabled
);
1849 if (read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
) {
1850 values
[n
++] = event
->total_time_running
+
1851 atomic64_read(&event
->child_total_time_running
);
1853 if (read_format
& PERF_FORMAT_ID
)
1854 values
[n
++] = primary_event_id(event
);
1856 if (copy_to_user(buf
, values
, n
* sizeof(u64
)))
1859 return n
* sizeof(u64
);
1863 * Read the performance event - simple non blocking version for now
1866 perf_read_hw(struct perf_event
*event
, char __user
*buf
, size_t count
)
1868 u64 read_format
= event
->attr
.read_format
;
1872 * Return end-of-file for a read on a event that is in
1873 * error state (i.e. because it was pinned but it couldn't be
1874 * scheduled on to the CPU at some point).
1876 if (event
->state
== PERF_EVENT_STATE_ERROR
)
1879 if (count
< perf_event_read_size(event
))
1882 WARN_ON_ONCE(event
->ctx
->parent_ctx
);
1883 mutex_lock(&event
->child_mutex
);
1884 if (read_format
& PERF_FORMAT_GROUP
)
1885 ret
= perf_event_read_group(event
, read_format
, buf
);
1887 ret
= perf_event_read_one(event
, read_format
, buf
);
1888 mutex_unlock(&event
->child_mutex
);
1894 perf_read(struct file
*file
, char __user
*buf
, size_t count
, loff_t
*ppos
)
1896 struct perf_event
*event
= file
->private_data
;
1898 return perf_read_hw(event
, buf
, count
);
1901 static unsigned int perf_poll(struct file
*file
, poll_table
*wait
)
1903 struct perf_event
*event
= file
->private_data
;
1904 struct perf_mmap_data
*data
;
1905 unsigned int events
= POLL_HUP
;
1908 data
= rcu_dereference(event
->data
);
1910 events
= atomic_xchg(&data
->poll
, 0);
1913 poll_wait(file
, &event
->waitq
, wait
);
1918 static void perf_event_reset(struct perf_event
*event
)
1920 (void)perf_event_read(event
);
1921 atomic64_set(&event
->count
, 0);
1922 perf_event_update_userpage(event
);
1926 * Holding the top-level event's child_mutex means that any
1927 * descendant process that has inherited this event will block
1928 * in sync_child_event if it goes to exit, thus satisfying the
1929 * task existence requirements of perf_event_enable/disable.
1931 static void perf_event_for_each_child(struct perf_event
*event
,
1932 void (*func
)(struct perf_event
*))
1934 struct perf_event
*child
;
1936 WARN_ON_ONCE(event
->ctx
->parent_ctx
);
1937 mutex_lock(&event
->child_mutex
);
1939 list_for_each_entry(child
, &event
->child_list
, child_list
)
1941 mutex_unlock(&event
->child_mutex
);
1944 static void perf_event_for_each(struct perf_event
*event
,
1945 void (*func
)(struct perf_event
*))
1947 struct perf_event_context
*ctx
= event
->ctx
;
1948 struct perf_event
*sibling
;
1950 WARN_ON_ONCE(ctx
->parent_ctx
);
1951 mutex_lock(&ctx
->mutex
);
1952 event
= event
->group_leader
;
1954 perf_event_for_each_child(event
, func
);
1956 list_for_each_entry(sibling
, &event
->sibling_list
, group_entry
)
1957 perf_event_for_each_child(event
, func
);
1958 mutex_unlock(&ctx
->mutex
);
1961 static int perf_event_period(struct perf_event
*event
, u64 __user
*arg
)
1963 struct perf_event_context
*ctx
= event
->ctx
;
1968 if (!event
->attr
.sample_period
)
1971 size
= copy_from_user(&value
, arg
, sizeof(value
));
1972 if (size
!= sizeof(value
))
1978 spin_lock_irq(&ctx
->lock
);
1979 if (event
->attr
.freq
) {
1980 if (value
> sysctl_perf_event_sample_rate
) {
1985 event
->attr
.sample_freq
= value
;
1987 event
->attr
.sample_period
= value
;
1988 event
->hw
.sample_period
= value
;
1991 spin_unlock_irq(&ctx
->lock
);
1996 static int perf_event_set_output(struct perf_event
*event
, int output_fd
);
1997 static int perf_event_set_filter(struct perf_event
*event
, void __user
*arg
);
1999 static long perf_ioctl(struct file
*file
, unsigned int cmd
, unsigned long arg
)
2001 struct perf_event
*event
= file
->private_data
;
2002 void (*func
)(struct perf_event
*);
2006 case PERF_EVENT_IOC_ENABLE
:
2007 func
= perf_event_enable
;
2009 case PERF_EVENT_IOC_DISABLE
:
2010 func
= perf_event_disable
;
2012 case PERF_EVENT_IOC_RESET
:
2013 func
= perf_event_reset
;
2016 case PERF_EVENT_IOC_REFRESH
:
2017 return perf_event_refresh(event
, arg
);
2019 case PERF_EVENT_IOC_PERIOD
:
2020 return perf_event_period(event
, (u64 __user
*)arg
);
2022 case PERF_EVENT_IOC_SET_OUTPUT
:
2023 return perf_event_set_output(event
, arg
);
2025 case PERF_EVENT_IOC_SET_FILTER
:
2026 return perf_event_set_filter(event
, (void __user
*)arg
);
2032 if (flags
& PERF_IOC_FLAG_GROUP
)
2033 perf_event_for_each(event
, func
);
2035 perf_event_for_each_child(event
, func
);
2040 int perf_event_task_enable(void)
2042 struct perf_event
*event
;
2044 mutex_lock(¤t
->perf_event_mutex
);
2045 list_for_each_entry(event
, ¤t
->perf_event_list
, owner_entry
)
2046 perf_event_for_each_child(event
, perf_event_enable
);
2047 mutex_unlock(¤t
->perf_event_mutex
);
2052 int perf_event_task_disable(void)
2054 struct perf_event
*event
;
2056 mutex_lock(¤t
->perf_event_mutex
);
2057 list_for_each_entry(event
, ¤t
->perf_event_list
, owner_entry
)
2058 perf_event_for_each_child(event
, perf_event_disable
);
2059 mutex_unlock(¤t
->perf_event_mutex
);
2064 #ifndef PERF_EVENT_INDEX_OFFSET
2065 # define PERF_EVENT_INDEX_OFFSET 0
2068 static int perf_event_index(struct perf_event
*event
)
2070 if (event
->state
!= PERF_EVENT_STATE_ACTIVE
)
2073 return event
->hw
.idx
+ 1 - PERF_EVENT_INDEX_OFFSET
;
2077 * Callers need to ensure there can be no nesting of this function, otherwise
2078 * the seqlock logic goes bad. We can not serialize this because the arch
2079 * code calls this from NMI context.
2081 void perf_event_update_userpage(struct perf_event
*event
)
2083 struct perf_event_mmap_page
*userpg
;
2084 struct perf_mmap_data
*data
;
2087 data
= rcu_dereference(event
->data
);
2091 userpg
= data
->user_page
;
2094 * Disable preemption so as to not let the corresponding user-space
2095 * spin too long if we get preempted.
2100 userpg
->index
= perf_event_index(event
);
2101 userpg
->offset
= atomic64_read(&event
->count
);
2102 if (event
->state
== PERF_EVENT_STATE_ACTIVE
)
2103 userpg
->offset
-= atomic64_read(&event
->hw
.prev_count
);
2105 userpg
->time_enabled
= event
->total_time_enabled
+
2106 atomic64_read(&event
->child_total_time_enabled
);
2108 userpg
->time_running
= event
->total_time_running
+
2109 atomic64_read(&event
->child_total_time_running
);
2118 static unsigned long perf_data_size(struct perf_mmap_data
*data
)
2120 return data
->nr_pages
<< (PAGE_SHIFT
+ data
->data_order
);
2123 #ifndef CONFIG_PERF_USE_VMALLOC
2126 * Back perf_mmap() with regular GFP_KERNEL-0 pages.
2129 static struct page
*
2130 perf_mmap_to_page(struct perf_mmap_data
*data
, unsigned long pgoff
)
2132 if (pgoff
> data
->nr_pages
)
2136 return virt_to_page(data
->user_page
);
2138 return virt_to_page(data
->data_pages
[pgoff
- 1]);
2141 static struct perf_mmap_data
*
2142 perf_mmap_data_alloc(struct perf_event
*event
, int nr_pages
)
2144 struct perf_mmap_data
*data
;
2148 WARN_ON(atomic_read(&event
->mmap_count
));
2150 size
= sizeof(struct perf_mmap_data
);
2151 size
+= nr_pages
* sizeof(void *);
2153 data
= kzalloc(size
, GFP_KERNEL
);
2157 data
->user_page
= (void *)get_zeroed_page(GFP_KERNEL
);
2158 if (!data
->user_page
)
2159 goto fail_user_page
;
2161 for (i
= 0; i
< nr_pages
; i
++) {
2162 data
->data_pages
[i
] = (void *)get_zeroed_page(GFP_KERNEL
);
2163 if (!data
->data_pages
[i
])
2164 goto fail_data_pages
;
2167 data
->data_order
= 0;
2168 data
->nr_pages
= nr_pages
;
2173 for (i
--; i
>= 0; i
--)
2174 free_page((unsigned long)data
->data_pages
[i
]);
2176 free_page((unsigned long)data
->user_page
);
2185 static void perf_mmap_free_page(unsigned long addr
)
2187 struct page
*page
= virt_to_page((void *)addr
);
2189 page
->mapping
= NULL
;
2193 static void perf_mmap_data_free(struct perf_mmap_data
*data
)
2197 perf_mmap_free_page((unsigned long)data
->user_page
);
2198 for (i
= 0; i
< data
->nr_pages
; i
++)
2199 perf_mmap_free_page((unsigned long)data
->data_pages
[i
]);
2205 * Back perf_mmap() with vmalloc memory.
2207 * Required for architectures that have d-cache aliasing issues.
2210 static struct page
*
2211 perf_mmap_to_page(struct perf_mmap_data
*data
, unsigned long pgoff
)
2213 if (pgoff
> (1UL << data
->data_order
))
2216 return vmalloc_to_page((void *)data
->user_page
+ pgoff
* PAGE_SIZE
);
2219 static void perf_mmap_unmark_page(void *addr
)
2221 struct page
*page
= vmalloc_to_page(addr
);
2223 page
->mapping
= NULL
;
2226 static void perf_mmap_data_free_work(struct work_struct
*work
)
2228 struct perf_mmap_data
*data
;
2232 data
= container_of(work
, struct perf_mmap_data
, work
);
2233 nr
= 1 << data
->data_order
;
2235 base
= data
->user_page
;
2236 for (i
= 0; i
< nr
+ 1; i
++)
2237 perf_mmap_unmark_page(base
+ (i
* PAGE_SIZE
));
2242 static void perf_mmap_data_free(struct perf_mmap_data
*data
)
2244 schedule_work(&data
->work
);
2247 static struct perf_mmap_data
*
2248 perf_mmap_data_alloc(struct perf_event
*event
, int nr_pages
)
2250 struct perf_mmap_data
*data
;
2254 WARN_ON(atomic_read(&event
->mmap_count
));
2256 size
= sizeof(struct perf_mmap_data
);
2257 size
+= sizeof(void *);
2259 data
= kzalloc(size
, GFP_KERNEL
);
2263 INIT_WORK(&data
->work
, perf_mmap_data_free_work
);
2265 all_buf
= vmalloc_user((nr_pages
+ 1) * PAGE_SIZE
);
2269 data
->user_page
= all_buf
;
2270 data
->data_pages
[0] = all_buf
+ PAGE_SIZE
;
2271 data
->data_order
= ilog2(nr_pages
);
2285 static int perf_mmap_fault(struct vm_area_struct
*vma
, struct vm_fault
*vmf
)
2287 struct perf_event
*event
= vma
->vm_file
->private_data
;
2288 struct perf_mmap_data
*data
;
2289 int ret
= VM_FAULT_SIGBUS
;
2291 if (vmf
->flags
& FAULT_FLAG_MKWRITE
) {
2292 if (vmf
->pgoff
== 0)
2298 data
= rcu_dereference(event
->data
);
2302 if (vmf
->pgoff
&& (vmf
->flags
& FAULT_FLAG_WRITE
))
2305 vmf
->page
= perf_mmap_to_page(data
, vmf
->pgoff
);
2309 get_page(vmf
->page
);
2310 vmf
->page
->mapping
= vma
->vm_file
->f_mapping
;
2311 vmf
->page
->index
= vmf
->pgoff
;
2321 perf_mmap_data_init(struct perf_event
*event
, struct perf_mmap_data
*data
)
2323 long max_size
= perf_data_size(data
);
2325 atomic_set(&data
->lock
, -1);
2327 if (event
->attr
.watermark
) {
2328 data
->watermark
= min_t(long, max_size
,
2329 event
->attr
.wakeup_watermark
);
2332 if (!data
->watermark
)
2333 data
->watermark
= max_t(long, PAGE_SIZE
, max_size
/ 2);
2336 rcu_assign_pointer(event
->data
, data
);
2339 static void perf_mmap_data_free_rcu(struct rcu_head
*rcu_head
)
2341 struct perf_mmap_data
*data
;
2343 data
= container_of(rcu_head
, struct perf_mmap_data
, rcu_head
);
2344 perf_mmap_data_free(data
);
2348 static void perf_mmap_data_release(struct perf_event
*event
)
2350 struct perf_mmap_data
*data
= event
->data
;
2352 WARN_ON(atomic_read(&event
->mmap_count
));
2354 rcu_assign_pointer(event
->data
, NULL
);
2355 call_rcu(&data
->rcu_head
, perf_mmap_data_free_rcu
);
2358 static void perf_mmap_open(struct vm_area_struct
*vma
)
2360 struct perf_event
*event
= vma
->vm_file
->private_data
;
2362 atomic_inc(&event
->mmap_count
);
2365 static void perf_mmap_close(struct vm_area_struct
*vma
)
2367 struct perf_event
*event
= vma
->vm_file
->private_data
;
2369 WARN_ON_ONCE(event
->ctx
->parent_ctx
);
2370 if (atomic_dec_and_mutex_lock(&event
->mmap_count
, &event
->mmap_mutex
)) {
2371 unsigned long size
= perf_data_size(event
->data
);
2372 struct user_struct
*user
= current_user();
2374 atomic_long_sub((size
>> PAGE_SHIFT
) + 1, &user
->locked_vm
);
2375 vma
->vm_mm
->locked_vm
-= event
->data
->nr_locked
;
2376 perf_mmap_data_release(event
);
2377 mutex_unlock(&event
->mmap_mutex
);
2381 static const struct vm_operations_struct perf_mmap_vmops
= {
2382 .open
= perf_mmap_open
,
2383 .close
= perf_mmap_close
,
2384 .fault
= perf_mmap_fault
,
2385 .page_mkwrite
= perf_mmap_fault
,
2388 static int perf_mmap(struct file
*file
, struct vm_area_struct
*vma
)
2390 struct perf_event
*event
= file
->private_data
;
2391 unsigned long user_locked
, user_lock_limit
;
2392 struct user_struct
*user
= current_user();
2393 unsigned long locked
, lock_limit
;
2394 struct perf_mmap_data
*data
;
2395 unsigned long vma_size
;
2396 unsigned long nr_pages
;
2397 long user_extra
, extra
;
2400 if (!(vma
->vm_flags
& VM_SHARED
))
2403 vma_size
= vma
->vm_end
- vma
->vm_start
;
2404 nr_pages
= (vma_size
/ PAGE_SIZE
) - 1;
2407 * If we have data pages ensure they're a power-of-two number, so we
2408 * can do bitmasks instead of modulo.
2410 if (nr_pages
!= 0 && !is_power_of_2(nr_pages
))
2413 if (vma_size
!= PAGE_SIZE
* (1 + nr_pages
))
2416 if (vma
->vm_pgoff
!= 0)
2419 WARN_ON_ONCE(event
->ctx
->parent_ctx
);
2420 mutex_lock(&event
->mmap_mutex
);
2421 if (event
->output
) {
2426 if (atomic_inc_not_zero(&event
->mmap_count
)) {
2427 if (nr_pages
!= event
->data
->nr_pages
)
2432 user_extra
= nr_pages
+ 1;
2433 user_lock_limit
= sysctl_perf_event_mlock
>> (PAGE_SHIFT
- 10);
2436 * Increase the limit linearly with more CPUs:
2438 user_lock_limit
*= num_online_cpus();
2440 user_locked
= atomic_long_read(&user
->locked_vm
) + user_extra
;
2443 if (user_locked
> user_lock_limit
)
2444 extra
= user_locked
- user_lock_limit
;
2446 lock_limit
= current
->signal
->rlim
[RLIMIT_MEMLOCK
].rlim_cur
;
2447 lock_limit
>>= PAGE_SHIFT
;
2448 locked
= vma
->vm_mm
->locked_vm
+ extra
;
2450 if ((locked
> lock_limit
) && perf_paranoid_tracepoint_raw() &&
2451 !capable(CAP_IPC_LOCK
)) {
2456 WARN_ON(event
->data
);
2458 data
= perf_mmap_data_alloc(event
, nr_pages
);
2464 perf_mmap_data_init(event
, data
);
2466 atomic_set(&event
->mmap_count
, 1);
2467 atomic_long_add(user_extra
, &user
->locked_vm
);
2468 vma
->vm_mm
->locked_vm
+= extra
;
2469 event
->data
->nr_locked
= extra
;
2470 if (vma
->vm_flags
& VM_WRITE
)
2471 event
->data
->writable
= 1;
2474 mutex_unlock(&event
->mmap_mutex
);
2476 vma
->vm_flags
|= VM_RESERVED
;
2477 vma
->vm_ops
= &perf_mmap_vmops
;
2482 static int perf_fasync(int fd
, struct file
*filp
, int on
)
2484 struct inode
*inode
= filp
->f_path
.dentry
->d_inode
;
2485 struct perf_event
*event
= filp
->private_data
;
2488 mutex_lock(&inode
->i_mutex
);
2489 retval
= fasync_helper(fd
, filp
, on
, &event
->fasync
);
2490 mutex_unlock(&inode
->i_mutex
);
2498 static const struct file_operations perf_fops
= {
2499 .release
= perf_release
,
2502 .unlocked_ioctl
= perf_ioctl
,
2503 .compat_ioctl
= perf_ioctl
,
2505 .fasync
= perf_fasync
,
2511 * If there's data, ensure we set the poll() state and publish everything
2512 * to user-space before waking everybody up.
2515 void perf_event_wakeup(struct perf_event
*event
)
2517 wake_up_all(&event
->waitq
);
2519 if (event
->pending_kill
) {
2520 kill_fasync(&event
->fasync
, SIGIO
, event
->pending_kill
);
2521 event
->pending_kill
= 0;
2528 * Handle the case where we need to wakeup up from NMI (or rq->lock) context.
2530 * The NMI bit means we cannot possibly take locks. Therefore, maintain a
2531 * single linked list and use cmpxchg() to add entries lockless.
2534 static void perf_pending_event(struct perf_pending_entry
*entry
)
2536 struct perf_event
*event
= container_of(entry
,
2537 struct perf_event
, pending
);
2539 if (event
->pending_disable
) {
2540 event
->pending_disable
= 0;
2541 __perf_event_disable(event
);
2544 if (event
->pending_wakeup
) {
2545 event
->pending_wakeup
= 0;
2546 perf_event_wakeup(event
);
2550 #define PENDING_TAIL ((struct perf_pending_entry *)-1UL)
2552 static DEFINE_PER_CPU(struct perf_pending_entry
*, perf_pending_head
) = {
2556 static void perf_pending_queue(struct perf_pending_entry
*entry
,
2557 void (*func
)(struct perf_pending_entry
*))
2559 struct perf_pending_entry
**head
;
2561 if (cmpxchg(&entry
->next
, NULL
, PENDING_TAIL
) != NULL
)
2566 head
= &get_cpu_var(perf_pending_head
);
2569 entry
->next
= *head
;
2570 } while (cmpxchg(head
, entry
->next
, entry
) != entry
->next
);
2572 set_perf_event_pending();
2574 put_cpu_var(perf_pending_head
);
2577 static int __perf_pending_run(void)
2579 struct perf_pending_entry
*list
;
2582 list
= xchg(&__get_cpu_var(perf_pending_head
), PENDING_TAIL
);
2583 while (list
!= PENDING_TAIL
) {
2584 void (*func
)(struct perf_pending_entry
*);
2585 struct perf_pending_entry
*entry
= list
;
2592 * Ensure we observe the unqueue before we issue the wakeup,
2593 * so that we won't be waiting forever.
2594 * -- see perf_not_pending().
2605 static inline int perf_not_pending(struct perf_event
*event
)
2608 * If we flush on whatever cpu we run, there is a chance we don't
2612 __perf_pending_run();
2616 * Ensure we see the proper queue state before going to sleep
2617 * so that we do not miss the wakeup. -- see perf_pending_handle()
2620 return event
->pending
.next
== NULL
;
2623 static void perf_pending_sync(struct perf_event
*event
)
2625 wait_event(event
->waitq
, perf_not_pending(event
));
2628 void perf_event_do_pending(void)
2630 __perf_pending_run();
2634 * Callchain support -- arch specific
2637 __weak
struct perf_callchain_entry
*perf_callchain(struct pt_regs
*regs
)
2645 static bool perf_output_space(struct perf_mmap_data
*data
, unsigned long tail
,
2646 unsigned long offset
, unsigned long head
)
2650 if (!data
->writable
)
2653 mask
= perf_data_size(data
) - 1;
2655 offset
= (offset
- tail
) & mask
;
2656 head
= (head
- tail
) & mask
;
2658 if ((int)(head
- offset
) < 0)
2664 static void perf_output_wakeup(struct perf_output_handle
*handle
)
2666 atomic_set(&handle
->data
->poll
, POLL_IN
);
2669 handle
->event
->pending_wakeup
= 1;
2670 perf_pending_queue(&handle
->event
->pending
,
2671 perf_pending_event
);
2673 perf_event_wakeup(handle
->event
);
2677 * Curious locking construct.
2679 * We need to ensure a later event_id doesn't publish a head when a former
2680 * event_id isn't done writing. However since we need to deal with NMIs we
2681 * cannot fully serialize things.
2683 * What we do is serialize between CPUs so we only have to deal with NMI
2684 * nesting on a single CPU.
2686 * We only publish the head (and generate a wakeup) when the outer-most
2687 * event_id completes.
2689 static void perf_output_lock(struct perf_output_handle
*handle
)
2691 struct perf_mmap_data
*data
= handle
->data
;
2692 int cur
, cpu
= get_cpu();
2697 cur
= atomic_cmpxchg(&data
->lock
, -1, cpu
);
2709 static void perf_output_unlock(struct perf_output_handle
*handle
)
2711 struct perf_mmap_data
*data
= handle
->data
;
2715 data
->done_head
= data
->head
;
2717 if (!handle
->locked
)
2722 * The xchg implies a full barrier that ensures all writes are done
2723 * before we publish the new head, matched by a rmb() in userspace when
2724 * reading this position.
2726 while ((head
= atomic_long_xchg(&data
->done_head
, 0)))
2727 data
->user_page
->data_head
= head
;
2730 * NMI can happen here, which means we can miss a done_head update.
2733 cpu
= atomic_xchg(&data
->lock
, -1);
2734 WARN_ON_ONCE(cpu
!= smp_processor_id());
2737 * Therefore we have to validate we did not indeed do so.
2739 if (unlikely(atomic_long_read(&data
->done_head
))) {
2741 * Since we had it locked, we can lock it again.
2743 while (atomic_cmpxchg(&data
->lock
, -1, cpu
) != -1)
2749 if (atomic_xchg(&data
->wakeup
, 0))
2750 perf_output_wakeup(handle
);
2755 void perf_output_copy(struct perf_output_handle
*handle
,
2756 const void *buf
, unsigned int len
)
2758 unsigned int pages_mask
;
2759 unsigned long offset
;
2763 offset
= handle
->offset
;
2764 pages_mask
= handle
->data
->nr_pages
- 1;
2765 pages
= handle
->data
->data_pages
;
2768 unsigned long page_offset
;
2769 unsigned long page_size
;
2772 nr
= (offset
>> PAGE_SHIFT
) & pages_mask
;
2773 page_size
= 1UL << (handle
->data
->data_order
+ PAGE_SHIFT
);
2774 page_offset
= offset
& (page_size
- 1);
2775 size
= min_t(unsigned int, page_size
- page_offset
, len
);
2777 memcpy(pages
[nr
] + page_offset
, buf
, size
);
2784 handle
->offset
= offset
;
2787 * Check we didn't copy past our reservation window, taking the
2788 * possible unsigned int wrap into account.
2790 WARN_ON_ONCE(((long)(handle
->head
- handle
->offset
)) < 0);
2793 int perf_output_begin(struct perf_output_handle
*handle
,
2794 struct perf_event
*event
, unsigned int size
,
2795 int nmi
, int sample
)
2797 struct perf_event
*output_event
;
2798 struct perf_mmap_data
*data
;
2799 unsigned long tail
, offset
, head
;
2802 struct perf_event_header header
;
2809 * For inherited events we send all the output towards the parent.
2812 event
= event
->parent
;
2814 output_event
= rcu_dereference(event
->output
);
2816 event
= output_event
;
2818 data
= rcu_dereference(event
->data
);
2822 handle
->data
= data
;
2823 handle
->event
= event
;
2825 handle
->sample
= sample
;
2827 if (!data
->nr_pages
)
2830 have_lost
= atomic_read(&data
->lost
);
2832 size
+= sizeof(lost_event
);
2834 perf_output_lock(handle
);
2838 * Userspace could choose to issue a mb() before updating the
2839 * tail pointer. So that all reads will be completed before the
2842 tail
= ACCESS_ONCE(data
->user_page
->data_tail
);
2844 offset
= head
= atomic_long_read(&data
->head
);
2846 if (unlikely(!perf_output_space(data
, tail
, offset
, head
)))
2848 } while (atomic_long_cmpxchg(&data
->head
, offset
, head
) != offset
);
2850 handle
->offset
= offset
;
2851 handle
->head
= head
;
2853 if (head
- tail
> data
->watermark
)
2854 atomic_set(&data
->wakeup
, 1);
2857 lost_event
.header
.type
= PERF_RECORD_LOST
;
2858 lost_event
.header
.misc
= 0;
2859 lost_event
.header
.size
= sizeof(lost_event
);
2860 lost_event
.id
= event
->id
;
2861 lost_event
.lost
= atomic_xchg(&data
->lost
, 0);
2863 perf_output_put(handle
, lost_event
);
2869 atomic_inc(&data
->lost
);
2870 perf_output_unlock(handle
);
2877 void perf_output_end(struct perf_output_handle
*handle
)
2879 struct perf_event
*event
= handle
->event
;
2880 struct perf_mmap_data
*data
= handle
->data
;
2882 int wakeup_events
= event
->attr
.wakeup_events
;
2884 if (handle
->sample
&& wakeup_events
) {
2885 int events
= atomic_inc_return(&data
->events
);
2886 if (events
>= wakeup_events
) {
2887 atomic_sub(wakeup_events
, &data
->events
);
2888 atomic_set(&data
->wakeup
, 1);
2892 perf_output_unlock(handle
);
2896 static u32
perf_event_pid(struct perf_event
*event
, struct task_struct
*p
)
2899 * only top level events have the pid namespace they were created in
2902 event
= event
->parent
;
2904 return task_tgid_nr_ns(p
, event
->ns
);
2907 static u32
perf_event_tid(struct perf_event
*event
, struct task_struct
*p
)
2910 * only top level events have the pid namespace they were created in
2913 event
= event
->parent
;
2915 return task_pid_nr_ns(p
, event
->ns
);
2918 static void perf_output_read_one(struct perf_output_handle
*handle
,
2919 struct perf_event
*event
)
2921 u64 read_format
= event
->attr
.read_format
;
2925 values
[n
++] = atomic64_read(&event
->count
);
2926 if (read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
) {
2927 values
[n
++] = event
->total_time_enabled
+
2928 atomic64_read(&event
->child_total_time_enabled
);
2930 if (read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
) {
2931 values
[n
++] = event
->total_time_running
+
2932 atomic64_read(&event
->child_total_time_running
);
2934 if (read_format
& PERF_FORMAT_ID
)
2935 values
[n
++] = primary_event_id(event
);
2937 perf_output_copy(handle
, values
, n
* sizeof(u64
));
2941 * XXX PERF_FORMAT_GROUP vs inherited events seems difficult.
2943 static void perf_output_read_group(struct perf_output_handle
*handle
,
2944 struct perf_event
*event
)
2946 struct perf_event
*leader
= event
->group_leader
, *sub
;
2947 u64 read_format
= event
->attr
.read_format
;
2951 values
[n
++] = 1 + leader
->nr_siblings
;
2953 if (read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
)
2954 values
[n
++] = leader
->total_time_enabled
;
2956 if (read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
)
2957 values
[n
++] = leader
->total_time_running
;
2959 if (leader
!= event
)
2960 leader
->pmu
->read(leader
);
2962 values
[n
++] = atomic64_read(&leader
->count
);
2963 if (read_format
& PERF_FORMAT_ID
)
2964 values
[n
++] = primary_event_id(leader
);
2966 perf_output_copy(handle
, values
, n
* sizeof(u64
));
2968 list_for_each_entry(sub
, &leader
->sibling_list
, group_entry
) {
2972 sub
->pmu
->read(sub
);
2974 values
[n
++] = atomic64_read(&sub
->count
);
2975 if (read_format
& PERF_FORMAT_ID
)
2976 values
[n
++] = primary_event_id(sub
);
2978 perf_output_copy(handle
, values
, n
* sizeof(u64
));
2982 static void perf_output_read(struct perf_output_handle
*handle
,
2983 struct perf_event
*event
)
2985 if (event
->attr
.read_format
& PERF_FORMAT_GROUP
)
2986 perf_output_read_group(handle
, event
);
2988 perf_output_read_one(handle
, event
);
2991 void perf_output_sample(struct perf_output_handle
*handle
,
2992 struct perf_event_header
*header
,
2993 struct perf_sample_data
*data
,
2994 struct perf_event
*event
)
2996 u64 sample_type
= data
->type
;
2998 perf_output_put(handle
, *header
);
3000 if (sample_type
& PERF_SAMPLE_IP
)
3001 perf_output_put(handle
, data
->ip
);
3003 if (sample_type
& PERF_SAMPLE_TID
)
3004 perf_output_put(handle
, data
->tid_entry
);
3006 if (sample_type
& PERF_SAMPLE_TIME
)
3007 perf_output_put(handle
, data
->time
);
3009 if (sample_type
& PERF_SAMPLE_ADDR
)
3010 perf_output_put(handle
, data
->addr
);
3012 if (sample_type
& PERF_SAMPLE_ID
)
3013 perf_output_put(handle
, data
->id
);
3015 if (sample_type
& PERF_SAMPLE_STREAM_ID
)
3016 perf_output_put(handle
, data
->stream_id
);
3018 if (sample_type
& PERF_SAMPLE_CPU
)
3019 perf_output_put(handle
, data
->cpu_entry
);
3021 if (sample_type
& PERF_SAMPLE_PERIOD
)
3022 perf_output_put(handle
, data
->period
);
3024 if (sample_type
& PERF_SAMPLE_READ
)
3025 perf_output_read(handle
, event
);
3027 if (sample_type
& PERF_SAMPLE_CALLCHAIN
) {
3028 if (data
->callchain
) {
3031 if (data
->callchain
)
3032 size
+= data
->callchain
->nr
;
3034 size
*= sizeof(u64
);
3036 perf_output_copy(handle
, data
->callchain
, size
);
3039 perf_output_put(handle
, nr
);
3043 if (sample_type
& PERF_SAMPLE_RAW
) {
3045 perf_output_put(handle
, data
->raw
->size
);
3046 perf_output_copy(handle
, data
->raw
->data
,
3053 .size
= sizeof(u32
),
3056 perf_output_put(handle
, raw
);
3061 void perf_prepare_sample(struct perf_event_header
*header
,
3062 struct perf_sample_data
*data
,
3063 struct perf_event
*event
,
3064 struct pt_regs
*regs
)
3066 u64 sample_type
= event
->attr
.sample_type
;
3068 data
->type
= sample_type
;
3070 header
->type
= PERF_RECORD_SAMPLE
;
3071 header
->size
= sizeof(*header
);
3074 header
->misc
|= perf_misc_flags(regs
);
3076 if (sample_type
& PERF_SAMPLE_IP
) {
3077 data
->ip
= perf_instruction_pointer(regs
);
3079 header
->size
+= sizeof(data
->ip
);
3082 if (sample_type
& PERF_SAMPLE_TID
) {
3083 /* namespace issues */
3084 data
->tid_entry
.pid
= perf_event_pid(event
, current
);
3085 data
->tid_entry
.tid
= perf_event_tid(event
, current
);
3087 header
->size
+= sizeof(data
->tid_entry
);
3090 if (sample_type
& PERF_SAMPLE_TIME
) {
3091 data
->time
= perf_clock();
3093 header
->size
+= sizeof(data
->time
);
3096 if (sample_type
& PERF_SAMPLE_ADDR
)
3097 header
->size
+= sizeof(data
->addr
);
3099 if (sample_type
& PERF_SAMPLE_ID
) {
3100 data
->id
= primary_event_id(event
);
3102 header
->size
+= sizeof(data
->id
);
3105 if (sample_type
& PERF_SAMPLE_STREAM_ID
) {
3106 data
->stream_id
= event
->id
;
3108 header
->size
+= sizeof(data
->stream_id
);
3111 if (sample_type
& PERF_SAMPLE_CPU
) {
3112 data
->cpu_entry
.cpu
= raw_smp_processor_id();
3113 data
->cpu_entry
.reserved
= 0;
3115 header
->size
+= sizeof(data
->cpu_entry
);
3118 if (sample_type
& PERF_SAMPLE_PERIOD
)
3119 header
->size
+= sizeof(data
->period
);
3121 if (sample_type
& PERF_SAMPLE_READ
)
3122 header
->size
+= perf_event_read_size(event
);
3124 if (sample_type
& PERF_SAMPLE_CALLCHAIN
) {
3127 data
->callchain
= perf_callchain(regs
);
3129 if (data
->callchain
)
3130 size
+= data
->callchain
->nr
;
3132 header
->size
+= size
* sizeof(u64
);
3135 if (sample_type
& PERF_SAMPLE_RAW
) {
3136 int size
= sizeof(u32
);
3139 size
+= data
->raw
->size
;
3141 size
+= sizeof(u32
);
3143 WARN_ON_ONCE(size
& (sizeof(u64
)-1));
3144 header
->size
+= size
;
3148 static void perf_event_output(struct perf_event
*event
, int nmi
,
3149 struct perf_sample_data
*data
,
3150 struct pt_regs
*regs
)
3152 struct perf_output_handle handle
;
3153 struct perf_event_header header
;
3155 perf_prepare_sample(&header
, data
, event
, regs
);
3157 if (perf_output_begin(&handle
, event
, header
.size
, nmi
, 1))
3160 perf_output_sample(&handle
, &header
, data
, event
);
3162 perf_output_end(&handle
);
3169 struct perf_read_event
{
3170 struct perf_event_header header
;
3177 perf_event_read_event(struct perf_event
*event
,
3178 struct task_struct
*task
)
3180 struct perf_output_handle handle
;
3181 struct perf_read_event read_event
= {
3183 .type
= PERF_RECORD_READ
,
3185 .size
= sizeof(read_event
) + perf_event_read_size(event
),
3187 .pid
= perf_event_pid(event
, task
),
3188 .tid
= perf_event_tid(event
, task
),
3192 ret
= perf_output_begin(&handle
, event
, read_event
.header
.size
, 0, 0);
3196 perf_output_put(&handle
, read_event
);
3197 perf_output_read(&handle
, event
);
3199 perf_output_end(&handle
);
3203 * task tracking -- fork/exit
3205 * enabled by: attr.comm | attr.mmap | attr.task
3208 struct perf_task_event
{
3209 struct task_struct
*task
;
3210 struct perf_event_context
*task_ctx
;
3213 struct perf_event_header header
;
3223 static void perf_event_task_output(struct perf_event
*event
,
3224 struct perf_task_event
*task_event
)
3226 struct perf_output_handle handle
;
3228 struct task_struct
*task
= task_event
->task
;
3231 size
= task_event
->event_id
.header
.size
;
3232 ret
= perf_output_begin(&handle
, event
, size
, 0, 0);
3237 task_event
->event_id
.pid
= perf_event_pid(event
, task
);
3238 task_event
->event_id
.ppid
= perf_event_pid(event
, current
);
3240 task_event
->event_id
.tid
= perf_event_tid(event
, task
);
3241 task_event
->event_id
.ptid
= perf_event_tid(event
, current
);
3243 task_event
->event_id
.time
= perf_clock();
3245 perf_output_put(&handle
, task_event
->event_id
);
3247 perf_output_end(&handle
);
3250 static int perf_event_task_match(struct perf_event
*event
)
3252 if (event
->attr
.comm
|| event
->attr
.mmap
|| event
->attr
.task
)
3258 static void perf_event_task_ctx(struct perf_event_context
*ctx
,
3259 struct perf_task_event
*task_event
)
3261 struct perf_event
*event
;
3263 list_for_each_entry_rcu(event
, &ctx
->event_list
, event_entry
) {
3264 if (perf_event_task_match(event
))
3265 perf_event_task_output(event
, task_event
);
3269 static void perf_event_task_event(struct perf_task_event
*task_event
)
3271 struct perf_cpu_context
*cpuctx
;
3272 struct perf_event_context
*ctx
= task_event
->task_ctx
;
3275 cpuctx
= &get_cpu_var(perf_cpu_context
);
3276 perf_event_task_ctx(&cpuctx
->ctx
, task_event
);
3277 put_cpu_var(perf_cpu_context
);
3280 ctx
= rcu_dereference(task_event
->task
->perf_event_ctxp
);
3282 perf_event_task_ctx(ctx
, task_event
);
3286 static void perf_event_task(struct task_struct
*task
,
3287 struct perf_event_context
*task_ctx
,
3290 struct perf_task_event task_event
;
3292 if (!atomic_read(&nr_comm_events
) &&
3293 !atomic_read(&nr_mmap_events
) &&
3294 !atomic_read(&nr_task_events
))
3297 task_event
= (struct perf_task_event
){
3299 .task_ctx
= task_ctx
,
3302 .type
= new ? PERF_RECORD_FORK
: PERF_RECORD_EXIT
,
3304 .size
= sizeof(task_event
.event_id
),
3313 perf_event_task_event(&task_event
);
3316 void perf_event_fork(struct task_struct
*task
)
3318 perf_event_task(task
, NULL
, 1);
3325 struct perf_comm_event
{
3326 struct task_struct
*task
;
3331 struct perf_event_header header
;
3338 static void perf_event_comm_output(struct perf_event
*event
,
3339 struct perf_comm_event
*comm_event
)
3341 struct perf_output_handle handle
;
3342 int size
= comm_event
->event_id
.header
.size
;
3343 int ret
= perf_output_begin(&handle
, event
, size
, 0, 0);
3348 comm_event
->event_id
.pid
= perf_event_pid(event
, comm_event
->task
);
3349 comm_event
->event_id
.tid
= perf_event_tid(event
, comm_event
->task
);
3351 perf_output_put(&handle
, comm_event
->event_id
);
3352 perf_output_copy(&handle
, comm_event
->comm
,
3353 comm_event
->comm_size
);
3354 perf_output_end(&handle
);
3357 static int perf_event_comm_match(struct perf_event
*event
)
3359 if (event
->attr
.comm
)
3365 static void perf_event_comm_ctx(struct perf_event_context
*ctx
,
3366 struct perf_comm_event
*comm_event
)
3368 struct perf_event
*event
;
3370 list_for_each_entry_rcu(event
, &ctx
->event_list
, event_entry
) {
3371 if (perf_event_comm_match(event
))
3372 perf_event_comm_output(event
, comm_event
);
3376 static void perf_event_comm_event(struct perf_comm_event
*comm_event
)
3378 struct perf_cpu_context
*cpuctx
;
3379 struct perf_event_context
*ctx
;
3381 char comm
[TASK_COMM_LEN
];
3383 memset(comm
, 0, sizeof(comm
));
3384 strncpy(comm
, comm_event
->task
->comm
, sizeof(comm
));
3385 size
= ALIGN(strlen(comm
)+1, sizeof(u64
));
3387 comm_event
->comm
= comm
;
3388 comm_event
->comm_size
= size
;
3390 comm_event
->event_id
.header
.size
= sizeof(comm_event
->event_id
) + size
;
3393 cpuctx
= &get_cpu_var(perf_cpu_context
);
3394 perf_event_comm_ctx(&cpuctx
->ctx
, comm_event
);
3395 put_cpu_var(perf_cpu_context
);
3398 * doesn't really matter which of the child contexts the
3399 * events ends up in.
3401 ctx
= rcu_dereference(current
->perf_event_ctxp
);
3403 perf_event_comm_ctx(ctx
, comm_event
);
3407 void perf_event_comm(struct task_struct
*task
)
3409 struct perf_comm_event comm_event
;
3411 if (task
->perf_event_ctxp
)
3412 perf_event_enable_on_exec(task
);
3414 if (!atomic_read(&nr_comm_events
))
3417 comm_event
= (struct perf_comm_event
){
3423 .type
= PERF_RECORD_COMM
,
3432 perf_event_comm_event(&comm_event
);
3439 struct perf_mmap_event
{
3440 struct vm_area_struct
*vma
;
3442 const char *file_name
;
3446 struct perf_event_header header
;
3456 static void perf_event_mmap_output(struct perf_event
*event
,
3457 struct perf_mmap_event
*mmap_event
)
3459 struct perf_output_handle handle
;
3460 int size
= mmap_event
->event_id
.header
.size
;
3461 int ret
= perf_output_begin(&handle
, event
, size
, 0, 0);
3466 mmap_event
->event_id
.pid
= perf_event_pid(event
, current
);
3467 mmap_event
->event_id
.tid
= perf_event_tid(event
, current
);
3469 perf_output_put(&handle
, mmap_event
->event_id
);
3470 perf_output_copy(&handle
, mmap_event
->file_name
,
3471 mmap_event
->file_size
);
3472 perf_output_end(&handle
);
3475 static int perf_event_mmap_match(struct perf_event
*event
,
3476 struct perf_mmap_event
*mmap_event
)
3478 if (event
->attr
.mmap
)
3484 static void perf_event_mmap_ctx(struct perf_event_context
*ctx
,
3485 struct perf_mmap_event
*mmap_event
)
3487 struct perf_event
*event
;
3489 list_for_each_entry_rcu(event
, &ctx
->event_list
, event_entry
) {
3490 if (perf_event_mmap_match(event
, mmap_event
))
3491 perf_event_mmap_output(event
, mmap_event
);
3495 static void perf_event_mmap_event(struct perf_mmap_event
*mmap_event
)
3497 struct perf_cpu_context
*cpuctx
;
3498 struct perf_event_context
*ctx
;
3499 struct vm_area_struct
*vma
= mmap_event
->vma
;
3500 struct file
*file
= vma
->vm_file
;
3506 memset(tmp
, 0, sizeof(tmp
));
3510 * d_path works from the end of the buffer backwards, so we
3511 * need to add enough zero bytes after the string to handle
3512 * the 64bit alignment we do later.
3514 buf
= kzalloc(PATH_MAX
+ sizeof(u64
), GFP_KERNEL
);
3516 name
= strncpy(tmp
, "//enomem", sizeof(tmp
));
3519 name
= d_path(&file
->f_path
, buf
, PATH_MAX
);
3521 name
= strncpy(tmp
, "//toolong", sizeof(tmp
));
3525 if (arch_vma_name(mmap_event
->vma
)) {
3526 name
= strncpy(tmp
, arch_vma_name(mmap_event
->vma
),
3532 name
= strncpy(tmp
, "[vdso]", sizeof(tmp
));
3536 name
= strncpy(tmp
, "//anon", sizeof(tmp
));
3541 size
= ALIGN(strlen(name
)+1, sizeof(u64
));
3543 mmap_event
->file_name
= name
;
3544 mmap_event
->file_size
= size
;
3546 mmap_event
->event_id
.header
.size
= sizeof(mmap_event
->event_id
) + size
;
3549 cpuctx
= &get_cpu_var(perf_cpu_context
);
3550 perf_event_mmap_ctx(&cpuctx
->ctx
, mmap_event
);
3551 put_cpu_var(perf_cpu_context
);
3554 * doesn't really matter which of the child contexts the
3555 * events ends up in.
3557 ctx
= rcu_dereference(current
->perf_event_ctxp
);
3559 perf_event_mmap_ctx(ctx
, mmap_event
);
3565 void __perf_event_mmap(struct vm_area_struct
*vma
)
3567 struct perf_mmap_event mmap_event
;
3569 if (!atomic_read(&nr_mmap_events
))
3572 mmap_event
= (struct perf_mmap_event
){
3578 .type
= PERF_RECORD_MMAP
,
3584 .start
= vma
->vm_start
,
3585 .len
= vma
->vm_end
- vma
->vm_start
,
3586 .pgoff
= vma
->vm_pgoff
,
3590 perf_event_mmap_event(&mmap_event
);
3594 * IRQ throttle logging
3597 static void perf_log_throttle(struct perf_event
*event
, int enable
)
3599 struct perf_output_handle handle
;
3603 struct perf_event_header header
;
3607 } throttle_event
= {
3609 .type
= PERF_RECORD_THROTTLE
,
3611 .size
= sizeof(throttle_event
),
3613 .time
= perf_clock(),
3614 .id
= primary_event_id(event
),
3615 .stream_id
= event
->id
,
3619 throttle_event
.header
.type
= PERF_RECORD_UNTHROTTLE
;
3621 ret
= perf_output_begin(&handle
, event
, sizeof(throttle_event
), 1, 0);
3625 perf_output_put(&handle
, throttle_event
);
3626 perf_output_end(&handle
);
3630 * Generic event overflow handling, sampling.
3633 static int __perf_event_overflow(struct perf_event
*event
, int nmi
,
3634 int throttle
, struct perf_sample_data
*data
,
3635 struct pt_regs
*regs
)
3637 int events
= atomic_read(&event
->event_limit
);
3638 struct hw_perf_event
*hwc
= &event
->hw
;
3641 throttle
= (throttle
&& event
->pmu
->unthrottle
!= NULL
);
3646 if (hwc
->interrupts
!= MAX_INTERRUPTS
) {
3648 if (HZ
* hwc
->interrupts
>
3649 (u64
)sysctl_perf_event_sample_rate
) {
3650 hwc
->interrupts
= MAX_INTERRUPTS
;
3651 perf_log_throttle(event
, 0);
3656 * Keep re-disabling events even though on the previous
3657 * pass we disabled it - just in case we raced with a
3658 * sched-in and the event got enabled again:
3664 if (event
->attr
.freq
) {
3665 u64 now
= perf_clock();
3666 s64 delta
= now
- hwc
->freq_stamp
;
3668 hwc
->freq_stamp
= now
;
3670 if (delta
> 0 && delta
< TICK_NSEC
)
3671 perf_adjust_period(event
, NSEC_PER_SEC
/ (int)delta
);
3675 * XXX event_limit might not quite work as expected on inherited
3679 event
->pending_kill
= POLL_IN
;
3680 if (events
&& atomic_dec_and_test(&event
->event_limit
)) {
3682 event
->pending_kill
= POLL_HUP
;
3684 event
->pending_disable
= 1;
3685 perf_pending_queue(&event
->pending
,
3686 perf_pending_event
);
3688 perf_event_disable(event
);
3691 if (event
->overflow_handler
)
3692 event
->overflow_handler(event
, nmi
, data
, regs
);
3694 perf_event_output(event
, nmi
, data
, regs
);
3699 int perf_event_overflow(struct perf_event
*event
, int nmi
,
3700 struct perf_sample_data
*data
,
3701 struct pt_regs
*regs
)
3703 return __perf_event_overflow(event
, nmi
, 1, data
, regs
);
3707 * Generic software event infrastructure
3711 * We directly increment event->count and keep a second value in
3712 * event->hw.period_left to count intervals. This period event
3713 * is kept in the range [-sample_period, 0] so that we can use the
3717 static u64
perf_swevent_set_period(struct perf_event
*event
)
3719 struct hw_perf_event
*hwc
= &event
->hw
;
3720 u64 period
= hwc
->last_period
;
3724 hwc
->last_period
= hwc
->sample_period
;
3727 old
= val
= atomic64_read(&hwc
->period_left
);
3731 nr
= div64_u64(period
+ val
, period
);
3732 offset
= nr
* period
;
3734 if (atomic64_cmpxchg(&hwc
->period_left
, old
, val
) != old
)
3740 static void perf_swevent_overflow(struct perf_event
*event
, u64 overflow
,
3741 int nmi
, struct perf_sample_data
*data
,
3742 struct pt_regs
*regs
)
3744 struct hw_perf_event
*hwc
= &event
->hw
;
3747 data
->period
= event
->hw
.last_period
;
3749 overflow
= perf_swevent_set_period(event
);
3751 if (hwc
->interrupts
== MAX_INTERRUPTS
)
3754 for (; overflow
; overflow
--) {
3755 if (__perf_event_overflow(event
, nmi
, throttle
,
3758 * We inhibit the overflow from happening when
3759 * hwc->interrupts == MAX_INTERRUPTS.
3767 static void perf_swevent_unthrottle(struct perf_event
*event
)
3770 * Nothing to do, we already reset hwc->interrupts.
3774 static void perf_swevent_add(struct perf_event
*event
, u64 nr
,
3775 int nmi
, struct perf_sample_data
*data
,
3776 struct pt_regs
*regs
)
3778 struct hw_perf_event
*hwc
= &event
->hw
;
3780 atomic64_add(nr
, &event
->count
);
3785 if (!hwc
->sample_period
)
3788 if (nr
== 1 && hwc
->sample_period
== 1 && !event
->attr
.freq
)
3789 return perf_swevent_overflow(event
, 1, nmi
, data
, regs
);
3791 if (atomic64_add_negative(nr
, &hwc
->period_left
))
3794 perf_swevent_overflow(event
, 0, nmi
, data
, regs
);
3797 static int perf_swevent_is_counting(struct perf_event
*event
)
3800 * The event is active, we're good!
3802 if (event
->state
== PERF_EVENT_STATE_ACTIVE
)
3806 * The event is off/error, not counting.
3808 if (event
->state
!= PERF_EVENT_STATE_INACTIVE
)
3812 * The event is inactive, if the context is active
3813 * we're part of a group that didn't make it on the 'pmu',
3816 if (event
->ctx
->is_active
)
3820 * We're inactive and the context is too, this means the
3821 * task is scheduled out, we're counting events that happen
3822 * to us, like migration events.
3827 static int perf_tp_event_match(struct perf_event
*event
,
3828 struct perf_sample_data
*data
);
3830 static int perf_swevent_match(struct perf_event
*event
,
3831 enum perf_type_id type
,
3833 struct perf_sample_data
*data
,
3834 struct pt_regs
*regs
)
3836 if (!perf_swevent_is_counting(event
))
3839 if (event
->attr
.type
!= type
)
3841 if (event
->attr
.config
!= event_id
)
3845 if (event
->attr
.exclude_user
&& user_mode(regs
))
3848 if (event
->attr
.exclude_kernel
&& !user_mode(regs
))
3852 if (event
->attr
.type
== PERF_TYPE_TRACEPOINT
&&
3853 !perf_tp_event_match(event
, data
))
3859 static void perf_swevent_ctx_event(struct perf_event_context
*ctx
,
3860 enum perf_type_id type
,
3861 u32 event_id
, u64 nr
, int nmi
,
3862 struct perf_sample_data
*data
,
3863 struct pt_regs
*regs
)
3865 struct perf_event
*event
;
3867 list_for_each_entry_rcu(event
, &ctx
->event_list
, event_entry
) {
3868 if (perf_swevent_match(event
, type
, event_id
, data
, regs
))
3869 perf_swevent_add(event
, nr
, nmi
, data
, regs
);
3873 static int *perf_swevent_recursion_context(struct perf_cpu_context
*cpuctx
)
3876 return &cpuctx
->recursion
[3];
3879 return &cpuctx
->recursion
[2];
3882 return &cpuctx
->recursion
[1];
3884 return &cpuctx
->recursion
[0];
3887 static void do_perf_sw_event(enum perf_type_id type
, u32 event_id
,
3889 struct perf_sample_data
*data
,
3890 struct pt_regs
*regs
)
3892 struct perf_cpu_context
*cpuctx
= &get_cpu_var(perf_cpu_context
);
3893 int *recursion
= perf_swevent_recursion_context(cpuctx
);
3894 struct perf_event_context
*ctx
;
3903 perf_swevent_ctx_event(&cpuctx
->ctx
, type
, event_id
,
3904 nr
, nmi
, data
, regs
);
3906 * doesn't really matter which of the child contexts the
3907 * events ends up in.
3909 ctx
= rcu_dereference(current
->perf_event_ctxp
);
3911 perf_swevent_ctx_event(ctx
, type
, event_id
, nr
, nmi
, data
, regs
);
3918 put_cpu_var(perf_cpu_context
);
3921 void __perf_sw_event(u32 event_id
, u64 nr
, int nmi
,
3922 struct pt_regs
*regs
, u64 addr
)
3924 struct perf_sample_data data
= {
3928 do_perf_sw_event(PERF_TYPE_SOFTWARE
, event_id
, nr
, nmi
,
3932 static void perf_swevent_read(struct perf_event
*event
)
3936 static int perf_swevent_enable(struct perf_event
*event
)
3938 struct hw_perf_event
*hwc
= &event
->hw
;
3940 if (hwc
->sample_period
) {
3941 hwc
->last_period
= hwc
->sample_period
;
3942 perf_swevent_set_period(event
);
3947 static void perf_swevent_disable(struct perf_event
*event
)
3951 static const struct pmu perf_ops_generic
= {
3952 .enable
= perf_swevent_enable
,
3953 .disable
= perf_swevent_disable
,
3954 .read
= perf_swevent_read
,
3955 .unthrottle
= perf_swevent_unthrottle
,
3959 * hrtimer based swevent callback
3962 static enum hrtimer_restart
perf_swevent_hrtimer(struct hrtimer
*hrtimer
)
3964 enum hrtimer_restart ret
= HRTIMER_RESTART
;
3965 struct perf_sample_data data
;
3966 struct pt_regs
*regs
;
3967 struct perf_event
*event
;
3970 event
= container_of(hrtimer
, struct perf_event
, hw
.hrtimer
);
3971 event
->pmu
->read(event
);
3974 regs
= get_irq_regs();
3976 * In case we exclude kernel IPs or are somehow not in interrupt
3977 * context, provide the next best thing, the user IP.
3979 if ((event
->attr
.exclude_kernel
|| !regs
) &&
3980 !event
->attr
.exclude_user
)
3981 regs
= task_pt_regs(current
);
3984 if (!(event
->attr
.exclude_idle
&& current
->pid
== 0))
3985 if (perf_event_overflow(event
, 0, &data
, regs
))
3986 ret
= HRTIMER_NORESTART
;
3989 period
= max_t(u64
, 10000, event
->hw
.sample_period
);
3990 hrtimer_forward_now(hrtimer
, ns_to_ktime(period
));
3995 static void perf_swevent_start_hrtimer(struct perf_event
*event
)
3997 struct hw_perf_event
*hwc
= &event
->hw
;
3999 hrtimer_init(&hwc
->hrtimer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
4000 hwc
->hrtimer
.function
= perf_swevent_hrtimer
;
4001 if (hwc
->sample_period
) {
4004 if (hwc
->remaining
) {
4005 if (hwc
->remaining
< 0)
4008 period
= hwc
->remaining
;
4011 period
= max_t(u64
, 10000, hwc
->sample_period
);
4013 __hrtimer_start_range_ns(&hwc
->hrtimer
,
4014 ns_to_ktime(period
), 0,
4015 HRTIMER_MODE_REL
, 0);
4019 static void perf_swevent_cancel_hrtimer(struct perf_event
*event
)
4021 struct hw_perf_event
*hwc
= &event
->hw
;
4023 if (hwc
->sample_period
) {
4024 ktime_t remaining
= hrtimer_get_remaining(&hwc
->hrtimer
);
4025 hwc
->remaining
= ktime_to_ns(remaining
);
4027 hrtimer_cancel(&hwc
->hrtimer
);
4032 * Software event: cpu wall time clock
4035 static void cpu_clock_perf_event_update(struct perf_event
*event
)
4037 int cpu
= raw_smp_processor_id();
4041 now
= cpu_clock(cpu
);
4042 prev
= atomic64_read(&event
->hw
.prev_count
);
4043 atomic64_set(&event
->hw
.prev_count
, now
);
4044 atomic64_add(now
- prev
, &event
->count
);
4047 static int cpu_clock_perf_event_enable(struct perf_event
*event
)
4049 struct hw_perf_event
*hwc
= &event
->hw
;
4050 int cpu
= raw_smp_processor_id();
4052 atomic64_set(&hwc
->prev_count
, cpu_clock(cpu
));
4053 perf_swevent_start_hrtimer(event
);
4058 static void cpu_clock_perf_event_disable(struct perf_event
*event
)
4060 perf_swevent_cancel_hrtimer(event
);
4061 cpu_clock_perf_event_update(event
);
4064 static void cpu_clock_perf_event_read(struct perf_event
*event
)
4066 cpu_clock_perf_event_update(event
);
4069 static const struct pmu perf_ops_cpu_clock
= {
4070 .enable
= cpu_clock_perf_event_enable
,
4071 .disable
= cpu_clock_perf_event_disable
,
4072 .read
= cpu_clock_perf_event_read
,
4076 * Software event: task time clock
4079 static void task_clock_perf_event_update(struct perf_event
*event
, u64 now
)
4084 prev
= atomic64_xchg(&event
->hw
.prev_count
, now
);
4086 atomic64_add(delta
, &event
->count
);
4089 static int task_clock_perf_event_enable(struct perf_event
*event
)
4091 struct hw_perf_event
*hwc
= &event
->hw
;
4094 now
= event
->ctx
->time
;
4096 atomic64_set(&hwc
->prev_count
, now
);
4098 perf_swevent_start_hrtimer(event
);
4103 static void task_clock_perf_event_disable(struct perf_event
*event
)
4105 perf_swevent_cancel_hrtimer(event
);
4106 task_clock_perf_event_update(event
, event
->ctx
->time
);
4110 static void task_clock_perf_event_read(struct perf_event
*event
)
4115 update_context_time(event
->ctx
);
4116 time
= event
->ctx
->time
;
4118 u64 now
= perf_clock();
4119 u64 delta
= now
- event
->ctx
->timestamp
;
4120 time
= event
->ctx
->time
+ delta
;
4123 task_clock_perf_event_update(event
, time
);
4126 static const struct pmu perf_ops_task_clock
= {
4127 .enable
= task_clock_perf_event_enable
,
4128 .disable
= task_clock_perf_event_disable
,
4129 .read
= task_clock_perf_event_read
,
4132 #ifdef CONFIG_EVENT_PROFILE
4134 void perf_tp_event(int event_id
, u64 addr
, u64 count
, void *record
,
4137 struct perf_raw_record raw
= {
4142 struct perf_sample_data data
= {
4147 struct pt_regs
*regs
= get_irq_regs();
4150 regs
= task_pt_regs(current
);
4152 do_perf_sw_event(PERF_TYPE_TRACEPOINT
, event_id
, count
, 1,
4155 EXPORT_SYMBOL_GPL(perf_tp_event
);
4157 static int perf_tp_event_match(struct perf_event
*event
,
4158 struct perf_sample_data
*data
)
4160 void *record
= data
->raw
->data
;
4162 if (likely(!event
->filter
) || filter_match_preds(event
->filter
, record
))
4167 static void tp_perf_event_destroy(struct perf_event
*event
)
4169 ftrace_profile_disable(event
->attr
.config
);
4172 static const struct pmu
*tp_perf_event_init(struct perf_event
*event
)
4175 * Raw tracepoint data is a severe data leak, only allow root to
4178 if ((event
->attr
.sample_type
& PERF_SAMPLE_RAW
) &&
4179 perf_paranoid_tracepoint_raw() &&
4180 !capable(CAP_SYS_ADMIN
))
4181 return ERR_PTR(-EPERM
);
4183 if (ftrace_profile_enable(event
->attr
.config
))
4186 event
->destroy
= tp_perf_event_destroy
;
4188 return &perf_ops_generic
;
4191 static int perf_event_set_filter(struct perf_event
*event
, void __user
*arg
)
4196 if (event
->attr
.type
!= PERF_TYPE_TRACEPOINT
)
4199 filter_str
= strndup_user(arg
, PAGE_SIZE
);
4200 if (IS_ERR(filter_str
))
4201 return PTR_ERR(filter_str
);
4203 ret
= ftrace_profile_set_filter(event
, event
->attr
.config
, filter_str
);
4209 static void perf_event_free_filter(struct perf_event
*event
)
4211 ftrace_profile_free_filter(event
);
4216 static int perf_tp_event_match(struct perf_event
*event
,
4217 struct perf_sample_data
*data
)
4222 static const struct pmu
*tp_perf_event_init(struct perf_event
*event
)
4227 static int perf_event_set_filter(struct perf_event
*event
, void __user
*arg
)
4232 static void perf_event_free_filter(struct perf_event
*event
)
4236 #endif /* CONFIG_EVENT_PROFILE */
4238 #ifdef CONFIG_HAVE_HW_BREAKPOINT
4239 static void bp_perf_event_destroy(struct perf_event
*event
)
4241 release_bp_slot(event
);
4244 static const struct pmu
*bp_perf_event_init(struct perf_event
*bp
)
4248 * The breakpoint is already filled if we haven't created the counter
4249 * through perf syscall
4250 * FIXME: manage to get trigerred to NULL if it comes from syscalls
4253 err
= register_perf_hw_breakpoint(bp
);
4255 err
= __register_perf_hw_breakpoint(bp
);
4257 return ERR_PTR(err
);
4259 bp
->destroy
= bp_perf_event_destroy
;
4261 return &perf_ops_bp
;
4264 void perf_bp_event(struct perf_event
*bp
, void *regs
)
4269 static void bp_perf_event_destroy(struct perf_event
*event
)
4273 static const struct pmu
*bp_perf_event_init(struct perf_event
*bp
)
4278 void perf_bp_event(struct perf_event
*bp
, void *regs
)
4283 atomic_t perf_swevent_enabled
[PERF_COUNT_SW_MAX
];
4285 static void sw_perf_event_destroy(struct perf_event
*event
)
4287 u64 event_id
= event
->attr
.config
;
4289 WARN_ON(event
->parent
);
4291 atomic_dec(&perf_swevent_enabled
[event_id
]);
4294 static const struct pmu
*sw_perf_event_init(struct perf_event
*event
)
4296 const struct pmu
*pmu
= NULL
;
4297 u64 event_id
= event
->attr
.config
;
4300 * Software events (currently) can't in general distinguish
4301 * between user, kernel and hypervisor events.
4302 * However, context switches and cpu migrations are considered
4303 * to be kernel events, and page faults are never hypervisor
4307 case PERF_COUNT_SW_CPU_CLOCK
:
4308 pmu
= &perf_ops_cpu_clock
;
4311 case PERF_COUNT_SW_TASK_CLOCK
:
4313 * If the user instantiates this as a per-cpu event,
4314 * use the cpu_clock event instead.
4316 if (event
->ctx
->task
)
4317 pmu
= &perf_ops_task_clock
;
4319 pmu
= &perf_ops_cpu_clock
;
4322 case PERF_COUNT_SW_PAGE_FAULTS
:
4323 case PERF_COUNT_SW_PAGE_FAULTS_MIN
:
4324 case PERF_COUNT_SW_PAGE_FAULTS_MAJ
:
4325 case PERF_COUNT_SW_CONTEXT_SWITCHES
:
4326 case PERF_COUNT_SW_CPU_MIGRATIONS
:
4327 case PERF_COUNT_SW_ALIGNMENT_FAULTS
:
4328 case PERF_COUNT_SW_EMULATION_FAULTS
:
4329 if (!event
->parent
) {
4330 atomic_inc(&perf_swevent_enabled
[event_id
]);
4331 event
->destroy
= sw_perf_event_destroy
;
4333 pmu
= &perf_ops_generic
;
4341 * Allocate and initialize a event structure
4343 static struct perf_event
*
4344 perf_event_alloc(struct perf_event_attr
*attr
,
4346 struct perf_event_context
*ctx
,
4347 struct perf_event
*group_leader
,
4348 struct perf_event
*parent_event
,
4349 perf_callback_t callback
,
4352 const struct pmu
*pmu
;
4353 struct perf_event
*event
;
4354 struct hw_perf_event
*hwc
;
4357 event
= kzalloc(sizeof(*event
), gfpflags
);
4359 return ERR_PTR(-ENOMEM
);
4362 * Single events are their own group leaders, with an
4363 * empty sibling list:
4366 group_leader
= event
;
4368 mutex_init(&event
->child_mutex
);
4369 INIT_LIST_HEAD(&event
->child_list
);
4371 INIT_LIST_HEAD(&event
->group_entry
);
4372 INIT_LIST_HEAD(&event
->event_entry
);
4373 INIT_LIST_HEAD(&event
->sibling_list
);
4374 init_waitqueue_head(&event
->waitq
);
4376 mutex_init(&event
->mmap_mutex
);
4379 event
->attr
= *attr
;
4380 event
->group_leader
= group_leader
;
4385 event
->parent
= parent_event
;
4387 event
->ns
= get_pid_ns(current
->nsproxy
->pid_ns
);
4388 event
->id
= atomic64_inc_return(&perf_event_id
);
4390 event
->state
= PERF_EVENT_STATE_INACTIVE
;
4392 if (!callback
&& parent_event
)
4393 callback
= parent_event
->callback
;
4395 event
->callback
= callback
;
4398 event
->state
= PERF_EVENT_STATE_OFF
;
4403 hwc
->sample_period
= attr
->sample_period
;
4404 if (attr
->freq
&& attr
->sample_freq
)
4405 hwc
->sample_period
= 1;
4406 hwc
->last_period
= hwc
->sample_period
;
4408 atomic64_set(&hwc
->period_left
, hwc
->sample_period
);
4411 * we currently do not support PERF_FORMAT_GROUP on inherited events
4413 if (attr
->inherit
&& (attr
->read_format
& PERF_FORMAT_GROUP
))
4416 switch (attr
->type
) {
4418 case PERF_TYPE_HARDWARE
:
4419 case PERF_TYPE_HW_CACHE
:
4420 pmu
= hw_perf_event_init(event
);
4423 case PERF_TYPE_SOFTWARE
:
4424 pmu
= sw_perf_event_init(event
);
4427 case PERF_TYPE_TRACEPOINT
:
4428 pmu
= tp_perf_event_init(event
);
4431 case PERF_TYPE_BREAKPOINT
:
4432 pmu
= bp_perf_event_init(event
);
4443 else if (IS_ERR(pmu
))
4448 put_pid_ns(event
->ns
);
4450 return ERR_PTR(err
);
4455 if (!event
->parent
) {
4456 atomic_inc(&nr_events
);
4457 if (event
->attr
.mmap
)
4458 atomic_inc(&nr_mmap_events
);
4459 if (event
->attr
.comm
)
4460 atomic_inc(&nr_comm_events
);
4461 if (event
->attr
.task
)
4462 atomic_inc(&nr_task_events
);
4468 static int perf_copy_attr(struct perf_event_attr __user
*uattr
,
4469 struct perf_event_attr
*attr
)
4474 if (!access_ok(VERIFY_WRITE
, uattr
, PERF_ATTR_SIZE_VER0
))
4478 * zero the full structure, so that a short copy will be nice.
4480 memset(attr
, 0, sizeof(*attr
));
4482 ret
= get_user(size
, &uattr
->size
);
4486 if (size
> PAGE_SIZE
) /* silly large */
4489 if (!size
) /* abi compat */
4490 size
= PERF_ATTR_SIZE_VER0
;
4492 if (size
< PERF_ATTR_SIZE_VER0
)
4496 * If we're handed a bigger struct than we know of,
4497 * ensure all the unknown bits are 0 - i.e. new
4498 * user-space does not rely on any kernel feature
4499 * extensions we dont know about yet.
4501 if (size
> sizeof(*attr
)) {
4502 unsigned char __user
*addr
;
4503 unsigned char __user
*end
;
4506 addr
= (void __user
*)uattr
+ sizeof(*attr
);
4507 end
= (void __user
*)uattr
+ size
;
4509 for (; addr
< end
; addr
++) {
4510 ret
= get_user(val
, addr
);
4516 size
= sizeof(*attr
);
4519 ret
= copy_from_user(attr
, uattr
, size
);
4524 * If the type exists, the corresponding creation will verify
4527 if (attr
->type
>= PERF_TYPE_MAX
)
4530 if (attr
->__reserved_1
|| attr
->__reserved_2
|| attr
->__reserved_3
)
4533 if (attr
->sample_type
& ~(PERF_SAMPLE_MAX
-1))
4536 if (attr
->read_format
& ~(PERF_FORMAT_MAX
-1))
4543 put_user(sizeof(*attr
), &uattr
->size
);
4548 static int perf_event_set_output(struct perf_event
*event
, int output_fd
)
4550 struct perf_event
*output_event
= NULL
;
4551 struct file
*output_file
= NULL
;
4552 struct perf_event
*old_output
;
4553 int fput_needed
= 0;
4559 output_file
= fget_light(output_fd
, &fput_needed
);
4563 if (output_file
->f_op
!= &perf_fops
)
4566 output_event
= output_file
->private_data
;
4568 /* Don't chain output fds */
4569 if (output_event
->output
)
4572 /* Don't set an output fd when we already have an output channel */
4576 atomic_long_inc(&output_file
->f_count
);
4579 mutex_lock(&event
->mmap_mutex
);
4580 old_output
= event
->output
;
4581 rcu_assign_pointer(event
->output
, output_event
);
4582 mutex_unlock(&event
->mmap_mutex
);
4586 * we need to make sure no existing perf_output_*()
4587 * is still referencing this event.
4590 fput(old_output
->filp
);
4595 fput_light(output_file
, fput_needed
);
4600 * sys_perf_event_open - open a performance event, associate it to a task/cpu
4602 * @attr_uptr: event_id type attributes for monitoring/sampling
4605 * @group_fd: group leader event fd
4607 SYSCALL_DEFINE5(perf_event_open
,
4608 struct perf_event_attr __user
*, attr_uptr
,
4609 pid_t
, pid
, int, cpu
, int, group_fd
, unsigned long, flags
)
4611 struct perf_event
*event
, *group_leader
;
4612 struct perf_event_attr attr
;
4613 struct perf_event_context
*ctx
;
4614 struct file
*event_file
= NULL
;
4615 struct file
*group_file
= NULL
;
4616 int fput_needed
= 0;
4617 int fput_needed2
= 0;
4620 /* for future expandability... */
4621 if (flags
& ~(PERF_FLAG_FD_NO_GROUP
| PERF_FLAG_FD_OUTPUT
))
4624 err
= perf_copy_attr(attr_uptr
, &attr
);
4628 if (!attr
.exclude_kernel
) {
4629 if (perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN
))
4634 if (attr
.sample_freq
> sysctl_perf_event_sample_rate
)
4639 * Get the target context (task or percpu):
4641 ctx
= find_get_context(pid
, cpu
);
4643 return PTR_ERR(ctx
);
4646 * Look up the group leader (we will attach this event to it):
4648 group_leader
= NULL
;
4649 if (group_fd
!= -1 && !(flags
& PERF_FLAG_FD_NO_GROUP
)) {
4651 group_file
= fget_light(group_fd
, &fput_needed
);
4653 goto err_put_context
;
4654 if (group_file
->f_op
!= &perf_fops
)
4655 goto err_put_context
;
4657 group_leader
= group_file
->private_data
;
4659 * Do not allow a recursive hierarchy (this new sibling
4660 * becoming part of another group-sibling):
4662 if (group_leader
->group_leader
!= group_leader
)
4663 goto err_put_context
;
4665 * Do not allow to attach to a group in a different
4666 * task or CPU context:
4668 if (group_leader
->ctx
!= ctx
)
4669 goto err_put_context
;
4671 * Only a group leader can be exclusive or pinned
4673 if (attr
.exclusive
|| attr
.pinned
)
4674 goto err_put_context
;
4677 event
= perf_event_alloc(&attr
, cpu
, ctx
, group_leader
,
4678 NULL
, NULL
, GFP_KERNEL
);
4679 err
= PTR_ERR(event
);
4681 goto err_put_context
;
4683 err
= anon_inode_getfd("[perf_event]", &perf_fops
, event
, 0);
4685 goto err_free_put_context
;
4687 event_file
= fget_light(err
, &fput_needed2
);
4689 goto err_free_put_context
;
4691 if (flags
& PERF_FLAG_FD_OUTPUT
) {
4692 err
= perf_event_set_output(event
, group_fd
);
4694 goto err_fput_free_put_context
;
4697 event
->filp
= event_file
;
4698 WARN_ON_ONCE(ctx
->parent_ctx
);
4699 mutex_lock(&ctx
->mutex
);
4700 perf_install_in_context(ctx
, event
, cpu
);
4702 mutex_unlock(&ctx
->mutex
);
4704 event
->owner
= current
;
4705 get_task_struct(current
);
4706 mutex_lock(¤t
->perf_event_mutex
);
4707 list_add_tail(&event
->owner_entry
, ¤t
->perf_event_list
);
4708 mutex_unlock(¤t
->perf_event_mutex
);
4710 err_fput_free_put_context
:
4711 fput_light(event_file
, fput_needed2
);
4713 err_free_put_context
:
4721 fput_light(group_file
, fput_needed
);
4727 * perf_event_create_kernel_counter
4729 * @attr: attributes of the counter to create
4730 * @cpu: cpu in which the counter is bound
4731 * @pid: task to profile
4734 perf_event_create_kernel_counter(struct perf_event_attr
*attr
, int cpu
,
4735 pid_t pid
, perf_callback_t callback
)
4737 struct perf_event
*event
;
4738 struct perf_event_context
*ctx
;
4742 * Get the target context (task or percpu):
4745 ctx
= find_get_context(pid
, cpu
);
4749 event
= perf_event_alloc(attr
, cpu
, ctx
, NULL
,
4750 NULL
, callback
, GFP_KERNEL
);
4751 err
= PTR_ERR(event
);
4753 goto err_put_context
;
4756 WARN_ON_ONCE(ctx
->parent_ctx
);
4757 mutex_lock(&ctx
->mutex
);
4758 perf_install_in_context(ctx
, event
, cpu
);
4760 mutex_unlock(&ctx
->mutex
);
4762 event
->owner
= current
;
4763 get_task_struct(current
);
4764 mutex_lock(¤t
->perf_event_mutex
);
4765 list_add_tail(&event
->owner_entry
, ¤t
->perf_event_list
);
4766 mutex_unlock(¤t
->perf_event_mutex
);
4776 EXPORT_SYMBOL_GPL(perf_event_create_kernel_counter
);
4779 * inherit a event from parent task to child task:
4781 static struct perf_event
*
4782 inherit_event(struct perf_event
*parent_event
,
4783 struct task_struct
*parent
,
4784 struct perf_event_context
*parent_ctx
,
4785 struct task_struct
*child
,
4786 struct perf_event
*group_leader
,
4787 struct perf_event_context
*child_ctx
)
4789 struct perf_event
*child_event
;
4792 * Instead of creating recursive hierarchies of events,
4793 * we link inherited events back to the original parent,
4794 * which has a filp for sure, which we use as the reference
4797 if (parent_event
->parent
)
4798 parent_event
= parent_event
->parent
;
4800 child_event
= perf_event_alloc(&parent_event
->attr
,
4801 parent_event
->cpu
, child_ctx
,
4802 group_leader
, parent_event
,
4804 if (IS_ERR(child_event
))
4809 * Make the child state follow the state of the parent event,
4810 * not its attr.disabled bit. We hold the parent's mutex,
4811 * so we won't race with perf_event_{en, dis}able_family.
4813 if (parent_event
->state
>= PERF_EVENT_STATE_INACTIVE
)
4814 child_event
->state
= PERF_EVENT_STATE_INACTIVE
;
4816 child_event
->state
= PERF_EVENT_STATE_OFF
;
4818 if (parent_event
->attr
.freq
)
4819 child_event
->hw
.sample_period
= parent_event
->hw
.sample_period
;
4821 child_event
->overflow_handler
= parent_event
->overflow_handler
;
4824 * Link it up in the child's context:
4826 add_event_to_ctx(child_event
, child_ctx
);
4829 * Get a reference to the parent filp - we will fput it
4830 * when the child event exits. This is safe to do because
4831 * we are in the parent and we know that the filp still
4832 * exists and has a nonzero count:
4834 atomic_long_inc(&parent_event
->filp
->f_count
);
4837 * Link this into the parent event's child list
4839 WARN_ON_ONCE(parent_event
->ctx
->parent_ctx
);
4840 mutex_lock(&parent_event
->child_mutex
);
4841 list_add_tail(&child_event
->child_list
, &parent_event
->child_list
);
4842 mutex_unlock(&parent_event
->child_mutex
);
4847 static int inherit_group(struct perf_event
*parent_event
,
4848 struct task_struct
*parent
,
4849 struct perf_event_context
*parent_ctx
,
4850 struct task_struct
*child
,
4851 struct perf_event_context
*child_ctx
)
4853 struct perf_event
*leader
;
4854 struct perf_event
*sub
;
4855 struct perf_event
*child_ctr
;
4857 leader
= inherit_event(parent_event
, parent
, parent_ctx
,
4858 child
, NULL
, child_ctx
);
4860 return PTR_ERR(leader
);
4861 list_for_each_entry(sub
, &parent_event
->sibling_list
, group_entry
) {
4862 child_ctr
= inherit_event(sub
, parent
, parent_ctx
,
4863 child
, leader
, child_ctx
);
4864 if (IS_ERR(child_ctr
))
4865 return PTR_ERR(child_ctr
);
4870 static void sync_child_event(struct perf_event
*child_event
,
4871 struct task_struct
*child
)
4873 struct perf_event
*parent_event
= child_event
->parent
;
4876 if (child_event
->attr
.inherit_stat
)
4877 perf_event_read_event(child_event
, child
);
4879 child_val
= atomic64_read(&child_event
->count
);
4882 * Add back the child's count to the parent's count:
4884 atomic64_add(child_val
, &parent_event
->count
);
4885 atomic64_add(child_event
->total_time_enabled
,
4886 &parent_event
->child_total_time_enabled
);
4887 atomic64_add(child_event
->total_time_running
,
4888 &parent_event
->child_total_time_running
);
4891 * Remove this event from the parent's list
4893 WARN_ON_ONCE(parent_event
->ctx
->parent_ctx
);
4894 mutex_lock(&parent_event
->child_mutex
);
4895 list_del_init(&child_event
->child_list
);
4896 mutex_unlock(&parent_event
->child_mutex
);
4899 * Release the parent event, if this was the last
4902 fput(parent_event
->filp
);
4906 __perf_event_exit_task(struct perf_event
*child_event
,
4907 struct perf_event_context
*child_ctx
,
4908 struct task_struct
*child
)
4910 struct perf_event
*parent_event
;
4912 update_event_times(child_event
);
4913 perf_event_remove_from_context(child_event
);
4915 parent_event
= child_event
->parent
;
4917 * It can happen that parent exits first, and has events
4918 * that are still around due to the child reference. These
4919 * events need to be zapped - but otherwise linger.
4922 sync_child_event(child_event
, child
);
4923 free_event(child_event
);
4928 * When a child task exits, feed back event values to parent events.
4930 void perf_event_exit_task(struct task_struct
*child
)
4932 struct perf_event
*child_event
, *tmp
;
4933 struct perf_event_context
*child_ctx
;
4934 unsigned long flags
;
4936 if (likely(!child
->perf_event_ctxp
)) {
4937 perf_event_task(child
, NULL
, 0);
4941 local_irq_save(flags
);
4943 * We can't reschedule here because interrupts are disabled,
4944 * and either child is current or it is a task that can't be
4945 * scheduled, so we are now safe from rescheduling changing
4948 child_ctx
= child
->perf_event_ctxp
;
4949 __perf_event_task_sched_out(child_ctx
);
4952 * Take the context lock here so that if find_get_context is
4953 * reading child->perf_event_ctxp, we wait until it has
4954 * incremented the context's refcount before we do put_ctx below.
4956 spin_lock(&child_ctx
->lock
);
4957 child
->perf_event_ctxp
= NULL
;
4959 * If this context is a clone; unclone it so it can't get
4960 * swapped to another process while we're removing all
4961 * the events from it.
4963 unclone_ctx(child_ctx
);
4964 spin_unlock_irqrestore(&child_ctx
->lock
, flags
);
4967 * Report the task dead after unscheduling the events so that we
4968 * won't get any samples after PERF_RECORD_EXIT. We can however still
4969 * get a few PERF_RECORD_READ events.
4971 perf_event_task(child
, child_ctx
, 0);
4974 * We can recurse on the same lock type through:
4976 * __perf_event_exit_task()
4977 * sync_child_event()
4978 * fput(parent_event->filp)
4980 * mutex_lock(&ctx->mutex)
4982 * But since its the parent context it won't be the same instance.
4984 mutex_lock_nested(&child_ctx
->mutex
, SINGLE_DEPTH_NESTING
);
4987 list_for_each_entry_safe(child_event
, tmp
, &child_ctx
->group_list
,
4989 __perf_event_exit_task(child_event
, child_ctx
, child
);
4992 * If the last event was a group event, it will have appended all
4993 * its siblings to the list, but we obtained 'tmp' before that which
4994 * will still point to the list head terminating the iteration.
4996 if (!list_empty(&child_ctx
->group_list
))
4999 mutex_unlock(&child_ctx
->mutex
);
5005 * free an unexposed, unused context as created by inheritance by
5006 * init_task below, used by fork() in case of fail.
5008 void perf_event_free_task(struct task_struct
*task
)
5010 struct perf_event_context
*ctx
= task
->perf_event_ctxp
;
5011 struct perf_event
*event
, *tmp
;
5016 mutex_lock(&ctx
->mutex
);
5018 list_for_each_entry_safe(event
, tmp
, &ctx
->group_list
, group_entry
) {
5019 struct perf_event
*parent
= event
->parent
;
5021 if (WARN_ON_ONCE(!parent
))
5024 mutex_lock(&parent
->child_mutex
);
5025 list_del_init(&event
->child_list
);
5026 mutex_unlock(&parent
->child_mutex
);
5030 list_del_event(event
, ctx
);
5034 if (!list_empty(&ctx
->group_list
))
5037 mutex_unlock(&ctx
->mutex
);
5043 * Initialize the perf_event context in task_struct
5045 int perf_event_init_task(struct task_struct
*child
)
5047 struct perf_event_context
*child_ctx
, *parent_ctx
;
5048 struct perf_event_context
*cloned_ctx
;
5049 struct perf_event
*event
;
5050 struct task_struct
*parent
= current
;
5051 int inherited_all
= 1;
5054 child
->perf_event_ctxp
= NULL
;
5056 mutex_init(&child
->perf_event_mutex
);
5057 INIT_LIST_HEAD(&child
->perf_event_list
);
5059 if (likely(!parent
->perf_event_ctxp
))
5063 * This is executed from the parent task context, so inherit
5064 * events that have been marked for cloning.
5065 * First allocate and initialize a context for the child.
5068 child_ctx
= kmalloc(sizeof(struct perf_event_context
), GFP_KERNEL
);
5072 __perf_event_init_context(child_ctx
, child
);
5073 child
->perf_event_ctxp
= child_ctx
;
5074 get_task_struct(child
);
5077 * If the parent's context is a clone, pin it so it won't get
5080 parent_ctx
= perf_pin_task_context(parent
);
5083 * No need to check if parent_ctx != NULL here; since we saw
5084 * it non-NULL earlier, the only reason for it to become NULL
5085 * is if we exit, and since we're currently in the middle of
5086 * a fork we can't be exiting at the same time.
5090 * Lock the parent list. No need to lock the child - not PID
5091 * hashed yet and not running, so nobody can access it.
5093 mutex_lock(&parent_ctx
->mutex
);
5096 * We dont have to disable NMIs - we are only looking at
5097 * the list, not manipulating it:
5099 list_for_each_entry(event
, &parent_ctx
->group_list
, group_entry
) {
5101 if (!event
->attr
.inherit
) {
5106 ret
= inherit_group(event
, parent
, parent_ctx
,
5114 if (inherited_all
) {
5116 * Mark the child context as a clone of the parent
5117 * context, or of whatever the parent is a clone of.
5118 * Note that if the parent is a clone, it could get
5119 * uncloned at any point, but that doesn't matter
5120 * because the list of events and the generation
5121 * count can't have changed since we took the mutex.
5123 cloned_ctx
= rcu_dereference(parent_ctx
->parent_ctx
);
5125 child_ctx
->parent_ctx
= cloned_ctx
;
5126 child_ctx
->parent_gen
= parent_ctx
->parent_gen
;
5128 child_ctx
->parent_ctx
= parent_ctx
;
5129 child_ctx
->parent_gen
= parent_ctx
->generation
;
5131 get_ctx(child_ctx
->parent_ctx
);
5134 mutex_unlock(&parent_ctx
->mutex
);
5136 perf_unpin_context(parent_ctx
);
5141 static void __cpuinit
perf_event_init_cpu(int cpu
)
5143 struct perf_cpu_context
*cpuctx
;
5145 cpuctx
= &per_cpu(perf_cpu_context
, cpu
);
5146 __perf_event_init_context(&cpuctx
->ctx
, NULL
);
5148 spin_lock(&perf_resource_lock
);
5149 cpuctx
->max_pertask
= perf_max_events
- perf_reserved_percpu
;
5150 spin_unlock(&perf_resource_lock
);
5152 hw_perf_event_setup(cpu
);
5155 #ifdef CONFIG_HOTPLUG_CPU
5156 static void __perf_event_exit_cpu(void *info
)
5158 struct perf_cpu_context
*cpuctx
= &__get_cpu_var(perf_cpu_context
);
5159 struct perf_event_context
*ctx
= &cpuctx
->ctx
;
5160 struct perf_event
*event
, *tmp
;
5162 list_for_each_entry_safe(event
, tmp
, &ctx
->group_list
, group_entry
)
5163 __perf_event_remove_from_context(event
);
5165 static void perf_event_exit_cpu(int cpu
)
5167 struct perf_cpu_context
*cpuctx
= &per_cpu(perf_cpu_context
, cpu
);
5168 struct perf_event_context
*ctx
= &cpuctx
->ctx
;
5170 mutex_lock(&ctx
->mutex
);
5171 smp_call_function_single(cpu
, __perf_event_exit_cpu
, NULL
, 1);
5172 mutex_unlock(&ctx
->mutex
);
5175 static inline void perf_event_exit_cpu(int cpu
) { }
5178 static int __cpuinit
5179 perf_cpu_notify(struct notifier_block
*self
, unsigned long action
, void *hcpu
)
5181 unsigned int cpu
= (long)hcpu
;
5185 case CPU_UP_PREPARE
:
5186 case CPU_UP_PREPARE_FROZEN
:
5187 perf_event_init_cpu(cpu
);
5191 case CPU_ONLINE_FROZEN
:
5192 hw_perf_event_setup_online(cpu
);
5195 case CPU_DOWN_PREPARE
:
5196 case CPU_DOWN_PREPARE_FROZEN
:
5197 perf_event_exit_cpu(cpu
);
5208 * This has to have a higher priority than migration_notifier in sched.c.
5210 static struct notifier_block __cpuinitdata perf_cpu_nb
= {
5211 .notifier_call
= perf_cpu_notify
,
5215 void __init
perf_event_init(void)
5217 perf_cpu_notify(&perf_cpu_nb
, (unsigned long)CPU_UP_PREPARE
,
5218 (void *)(long)smp_processor_id());
5219 perf_cpu_notify(&perf_cpu_nb
, (unsigned long)CPU_ONLINE
,
5220 (void *)(long)smp_processor_id());
5221 register_cpu_notifier(&perf_cpu_nb
);
5224 static ssize_t
perf_show_reserve_percpu(struct sysdev_class
*class, char *buf
)
5226 return sprintf(buf
, "%d\n", perf_reserved_percpu
);
5230 perf_set_reserve_percpu(struct sysdev_class
*class,
5234 struct perf_cpu_context
*cpuctx
;
5238 err
= strict_strtoul(buf
, 10, &val
);
5241 if (val
> perf_max_events
)
5244 spin_lock(&perf_resource_lock
);
5245 perf_reserved_percpu
= val
;
5246 for_each_online_cpu(cpu
) {
5247 cpuctx
= &per_cpu(perf_cpu_context
, cpu
);
5248 spin_lock_irq(&cpuctx
->ctx
.lock
);
5249 mpt
= min(perf_max_events
- cpuctx
->ctx
.nr_events
,
5250 perf_max_events
- perf_reserved_percpu
);
5251 cpuctx
->max_pertask
= mpt
;
5252 spin_unlock_irq(&cpuctx
->ctx
.lock
);
5254 spin_unlock(&perf_resource_lock
);
5259 static ssize_t
perf_show_overcommit(struct sysdev_class
*class, char *buf
)
5261 return sprintf(buf
, "%d\n", perf_overcommit
);
5265 perf_set_overcommit(struct sysdev_class
*class, const char *buf
, size_t count
)
5270 err
= strict_strtoul(buf
, 10, &val
);
5276 spin_lock(&perf_resource_lock
);
5277 perf_overcommit
= val
;
5278 spin_unlock(&perf_resource_lock
);
5283 static SYSDEV_CLASS_ATTR(
5286 perf_show_reserve_percpu
,
5287 perf_set_reserve_percpu
5290 static SYSDEV_CLASS_ATTR(
5293 perf_show_overcommit
,
5297 static struct attribute
*perfclass_attrs
[] = {
5298 &attr_reserve_percpu
.attr
,
5299 &attr_overcommit
.attr
,
5303 static struct attribute_group perfclass_attr_group
= {
5304 .attrs
= perfclass_attrs
,
5305 .name
= "perf_events",
5308 static int __init
perf_event_sysfs_init(void)
5310 return sysfs_create_group(&cpu_sysdev_class
.kset
.kobj
,
5311 &perfclass_attr_group
);
5313 device_initcall(perf_event_sysfs_init
);