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
, u64
*enabled
, u64
*running
)
1779 struct perf_event
*child
;
1785 mutex_lock(&event
->child_mutex
);
1786 total
+= perf_event_read(event
);
1787 *enabled
+= event
->total_time_enabled
+
1788 atomic64_read(&event
->child_total_time_enabled
);
1789 *running
+= event
->total_time_running
+
1790 atomic64_read(&event
->child_total_time_running
);
1792 list_for_each_entry(child
, &event
->child_list
, child_list
) {
1793 total
+= perf_event_read(child
);
1794 *enabled
+= child
->total_time_enabled
;
1795 *running
+= child
->total_time_running
;
1797 mutex_unlock(&event
->child_mutex
);
1801 EXPORT_SYMBOL_GPL(perf_event_read_value
);
1803 static int perf_event_read_group(struct perf_event
*event
,
1804 u64 read_format
, char __user
*buf
)
1806 struct perf_event
*leader
= event
->group_leader
, *sub
;
1807 int n
= 0, size
= 0, ret
= -EFAULT
;
1808 struct perf_event_context
*ctx
= leader
->ctx
;
1810 u64 count
, enabled
, running
;
1812 mutex_lock(&ctx
->mutex
);
1813 count
= perf_event_read_value(leader
, &enabled
, &running
);
1815 values
[n
++] = 1 + leader
->nr_siblings
;
1816 if (read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
)
1817 values
[n
++] = enabled
;
1818 if (read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
)
1819 values
[n
++] = running
;
1820 values
[n
++] = count
;
1821 if (read_format
& PERF_FORMAT_ID
)
1822 values
[n
++] = primary_event_id(leader
);
1824 size
= n
* sizeof(u64
);
1826 if (copy_to_user(buf
, values
, size
))
1831 list_for_each_entry(sub
, &leader
->sibling_list
, group_entry
) {
1834 values
[n
++] = perf_event_read_value(sub
, &enabled
, &running
);
1835 if (read_format
& PERF_FORMAT_ID
)
1836 values
[n
++] = primary_event_id(sub
);
1838 size
= n
* sizeof(u64
);
1840 if (copy_to_user(buf
+ size
, values
, size
)) {
1848 mutex_unlock(&ctx
->mutex
);
1853 static int perf_event_read_one(struct perf_event
*event
,
1854 u64 read_format
, char __user
*buf
)
1856 u64 enabled
, running
;
1860 values
[n
++] = perf_event_read_value(event
, &enabled
, &running
);
1861 if (read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
)
1862 values
[n
++] = enabled
;
1863 if (read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
)
1864 values
[n
++] = running
;
1865 if (read_format
& PERF_FORMAT_ID
)
1866 values
[n
++] = primary_event_id(event
);
1868 if (copy_to_user(buf
, values
, n
* sizeof(u64
)))
1871 return n
* sizeof(u64
);
1875 * Read the performance event - simple non blocking version for now
1878 perf_read_hw(struct perf_event
*event
, char __user
*buf
, size_t count
)
1880 u64 read_format
= event
->attr
.read_format
;
1884 * Return end-of-file for a read on a event that is in
1885 * error state (i.e. because it was pinned but it couldn't be
1886 * scheduled on to the CPU at some point).
1888 if (event
->state
== PERF_EVENT_STATE_ERROR
)
1891 if (count
< perf_event_read_size(event
))
1894 WARN_ON_ONCE(event
->ctx
->parent_ctx
);
1895 if (read_format
& PERF_FORMAT_GROUP
)
1896 ret
= perf_event_read_group(event
, read_format
, buf
);
1898 ret
= perf_event_read_one(event
, read_format
, buf
);
1904 perf_read(struct file
*file
, char __user
*buf
, size_t count
, loff_t
*ppos
)
1906 struct perf_event
*event
= file
->private_data
;
1908 return perf_read_hw(event
, buf
, count
);
1911 static unsigned int perf_poll(struct file
*file
, poll_table
*wait
)
1913 struct perf_event
*event
= file
->private_data
;
1914 struct perf_mmap_data
*data
;
1915 unsigned int events
= POLL_HUP
;
1918 data
= rcu_dereference(event
->data
);
1920 events
= atomic_xchg(&data
->poll
, 0);
1923 poll_wait(file
, &event
->waitq
, wait
);
1928 static void perf_event_reset(struct perf_event
*event
)
1930 (void)perf_event_read(event
);
1931 atomic64_set(&event
->count
, 0);
1932 perf_event_update_userpage(event
);
1936 * Holding the top-level event's child_mutex means that any
1937 * descendant process that has inherited this event will block
1938 * in sync_child_event if it goes to exit, thus satisfying the
1939 * task existence requirements of perf_event_enable/disable.
1941 static void perf_event_for_each_child(struct perf_event
*event
,
1942 void (*func
)(struct perf_event
*))
1944 struct perf_event
*child
;
1946 WARN_ON_ONCE(event
->ctx
->parent_ctx
);
1947 mutex_lock(&event
->child_mutex
);
1949 list_for_each_entry(child
, &event
->child_list
, child_list
)
1951 mutex_unlock(&event
->child_mutex
);
1954 static void perf_event_for_each(struct perf_event
*event
,
1955 void (*func
)(struct perf_event
*))
1957 struct perf_event_context
*ctx
= event
->ctx
;
1958 struct perf_event
*sibling
;
1960 WARN_ON_ONCE(ctx
->parent_ctx
);
1961 mutex_lock(&ctx
->mutex
);
1962 event
= event
->group_leader
;
1964 perf_event_for_each_child(event
, func
);
1966 list_for_each_entry(sibling
, &event
->sibling_list
, group_entry
)
1967 perf_event_for_each_child(event
, func
);
1968 mutex_unlock(&ctx
->mutex
);
1971 static int perf_event_period(struct perf_event
*event
, u64 __user
*arg
)
1973 struct perf_event_context
*ctx
= event
->ctx
;
1978 if (!event
->attr
.sample_period
)
1981 size
= copy_from_user(&value
, arg
, sizeof(value
));
1982 if (size
!= sizeof(value
))
1988 spin_lock_irq(&ctx
->lock
);
1989 if (event
->attr
.freq
) {
1990 if (value
> sysctl_perf_event_sample_rate
) {
1995 event
->attr
.sample_freq
= value
;
1997 event
->attr
.sample_period
= value
;
1998 event
->hw
.sample_period
= value
;
2001 spin_unlock_irq(&ctx
->lock
);
2006 static int perf_event_set_output(struct perf_event
*event
, int output_fd
);
2007 static int perf_event_set_filter(struct perf_event
*event
, void __user
*arg
);
2009 static long perf_ioctl(struct file
*file
, unsigned int cmd
, unsigned long arg
)
2011 struct perf_event
*event
= file
->private_data
;
2012 void (*func
)(struct perf_event
*);
2016 case PERF_EVENT_IOC_ENABLE
:
2017 func
= perf_event_enable
;
2019 case PERF_EVENT_IOC_DISABLE
:
2020 func
= perf_event_disable
;
2022 case PERF_EVENT_IOC_RESET
:
2023 func
= perf_event_reset
;
2026 case PERF_EVENT_IOC_REFRESH
:
2027 return perf_event_refresh(event
, arg
);
2029 case PERF_EVENT_IOC_PERIOD
:
2030 return perf_event_period(event
, (u64 __user
*)arg
);
2032 case PERF_EVENT_IOC_SET_OUTPUT
:
2033 return perf_event_set_output(event
, arg
);
2035 case PERF_EVENT_IOC_SET_FILTER
:
2036 return perf_event_set_filter(event
, (void __user
*)arg
);
2042 if (flags
& PERF_IOC_FLAG_GROUP
)
2043 perf_event_for_each(event
, func
);
2045 perf_event_for_each_child(event
, func
);
2050 int perf_event_task_enable(void)
2052 struct perf_event
*event
;
2054 mutex_lock(¤t
->perf_event_mutex
);
2055 list_for_each_entry(event
, ¤t
->perf_event_list
, owner_entry
)
2056 perf_event_for_each_child(event
, perf_event_enable
);
2057 mutex_unlock(¤t
->perf_event_mutex
);
2062 int perf_event_task_disable(void)
2064 struct perf_event
*event
;
2066 mutex_lock(¤t
->perf_event_mutex
);
2067 list_for_each_entry(event
, ¤t
->perf_event_list
, owner_entry
)
2068 perf_event_for_each_child(event
, perf_event_disable
);
2069 mutex_unlock(¤t
->perf_event_mutex
);
2074 #ifndef PERF_EVENT_INDEX_OFFSET
2075 # define PERF_EVENT_INDEX_OFFSET 0
2078 static int perf_event_index(struct perf_event
*event
)
2080 if (event
->state
!= PERF_EVENT_STATE_ACTIVE
)
2083 return event
->hw
.idx
+ 1 - PERF_EVENT_INDEX_OFFSET
;
2087 * Callers need to ensure there can be no nesting of this function, otherwise
2088 * the seqlock logic goes bad. We can not serialize this because the arch
2089 * code calls this from NMI context.
2091 void perf_event_update_userpage(struct perf_event
*event
)
2093 struct perf_event_mmap_page
*userpg
;
2094 struct perf_mmap_data
*data
;
2097 data
= rcu_dereference(event
->data
);
2101 userpg
= data
->user_page
;
2104 * Disable preemption so as to not let the corresponding user-space
2105 * spin too long if we get preempted.
2110 userpg
->index
= perf_event_index(event
);
2111 userpg
->offset
= atomic64_read(&event
->count
);
2112 if (event
->state
== PERF_EVENT_STATE_ACTIVE
)
2113 userpg
->offset
-= atomic64_read(&event
->hw
.prev_count
);
2115 userpg
->time_enabled
= event
->total_time_enabled
+
2116 atomic64_read(&event
->child_total_time_enabled
);
2118 userpg
->time_running
= event
->total_time_running
+
2119 atomic64_read(&event
->child_total_time_running
);
2128 static unsigned long perf_data_size(struct perf_mmap_data
*data
)
2130 return data
->nr_pages
<< (PAGE_SHIFT
+ data
->data_order
);
2133 #ifndef CONFIG_PERF_USE_VMALLOC
2136 * Back perf_mmap() with regular GFP_KERNEL-0 pages.
2139 static struct page
*
2140 perf_mmap_to_page(struct perf_mmap_data
*data
, unsigned long pgoff
)
2142 if (pgoff
> data
->nr_pages
)
2146 return virt_to_page(data
->user_page
);
2148 return virt_to_page(data
->data_pages
[pgoff
- 1]);
2151 static struct perf_mmap_data
*
2152 perf_mmap_data_alloc(struct perf_event
*event
, int nr_pages
)
2154 struct perf_mmap_data
*data
;
2158 WARN_ON(atomic_read(&event
->mmap_count
));
2160 size
= sizeof(struct perf_mmap_data
);
2161 size
+= nr_pages
* sizeof(void *);
2163 data
= kzalloc(size
, GFP_KERNEL
);
2167 data
->user_page
= (void *)get_zeroed_page(GFP_KERNEL
);
2168 if (!data
->user_page
)
2169 goto fail_user_page
;
2171 for (i
= 0; i
< nr_pages
; i
++) {
2172 data
->data_pages
[i
] = (void *)get_zeroed_page(GFP_KERNEL
);
2173 if (!data
->data_pages
[i
])
2174 goto fail_data_pages
;
2177 data
->data_order
= 0;
2178 data
->nr_pages
= nr_pages
;
2183 for (i
--; i
>= 0; i
--)
2184 free_page((unsigned long)data
->data_pages
[i
]);
2186 free_page((unsigned long)data
->user_page
);
2195 static void perf_mmap_free_page(unsigned long addr
)
2197 struct page
*page
= virt_to_page((void *)addr
);
2199 page
->mapping
= NULL
;
2203 static void perf_mmap_data_free(struct perf_mmap_data
*data
)
2207 perf_mmap_free_page((unsigned long)data
->user_page
);
2208 for (i
= 0; i
< data
->nr_pages
; i
++)
2209 perf_mmap_free_page((unsigned long)data
->data_pages
[i
]);
2215 * Back perf_mmap() with vmalloc memory.
2217 * Required for architectures that have d-cache aliasing issues.
2220 static struct page
*
2221 perf_mmap_to_page(struct perf_mmap_data
*data
, unsigned long pgoff
)
2223 if (pgoff
> (1UL << data
->data_order
))
2226 return vmalloc_to_page((void *)data
->user_page
+ pgoff
* PAGE_SIZE
);
2229 static void perf_mmap_unmark_page(void *addr
)
2231 struct page
*page
= vmalloc_to_page(addr
);
2233 page
->mapping
= NULL
;
2236 static void perf_mmap_data_free_work(struct work_struct
*work
)
2238 struct perf_mmap_data
*data
;
2242 data
= container_of(work
, struct perf_mmap_data
, work
);
2243 nr
= 1 << data
->data_order
;
2245 base
= data
->user_page
;
2246 for (i
= 0; i
< nr
+ 1; i
++)
2247 perf_mmap_unmark_page(base
+ (i
* PAGE_SIZE
));
2252 static void perf_mmap_data_free(struct perf_mmap_data
*data
)
2254 schedule_work(&data
->work
);
2257 static struct perf_mmap_data
*
2258 perf_mmap_data_alloc(struct perf_event
*event
, int nr_pages
)
2260 struct perf_mmap_data
*data
;
2264 WARN_ON(atomic_read(&event
->mmap_count
));
2266 size
= sizeof(struct perf_mmap_data
);
2267 size
+= sizeof(void *);
2269 data
= kzalloc(size
, GFP_KERNEL
);
2273 INIT_WORK(&data
->work
, perf_mmap_data_free_work
);
2275 all_buf
= vmalloc_user((nr_pages
+ 1) * PAGE_SIZE
);
2279 data
->user_page
= all_buf
;
2280 data
->data_pages
[0] = all_buf
+ PAGE_SIZE
;
2281 data
->data_order
= ilog2(nr_pages
);
2295 static int perf_mmap_fault(struct vm_area_struct
*vma
, struct vm_fault
*vmf
)
2297 struct perf_event
*event
= vma
->vm_file
->private_data
;
2298 struct perf_mmap_data
*data
;
2299 int ret
= VM_FAULT_SIGBUS
;
2301 if (vmf
->flags
& FAULT_FLAG_MKWRITE
) {
2302 if (vmf
->pgoff
== 0)
2308 data
= rcu_dereference(event
->data
);
2312 if (vmf
->pgoff
&& (vmf
->flags
& FAULT_FLAG_WRITE
))
2315 vmf
->page
= perf_mmap_to_page(data
, vmf
->pgoff
);
2319 get_page(vmf
->page
);
2320 vmf
->page
->mapping
= vma
->vm_file
->f_mapping
;
2321 vmf
->page
->index
= vmf
->pgoff
;
2331 perf_mmap_data_init(struct perf_event
*event
, struct perf_mmap_data
*data
)
2333 long max_size
= perf_data_size(data
);
2335 atomic_set(&data
->lock
, -1);
2337 if (event
->attr
.watermark
) {
2338 data
->watermark
= min_t(long, max_size
,
2339 event
->attr
.wakeup_watermark
);
2342 if (!data
->watermark
)
2343 data
->watermark
= max_size
/ 2;
2346 rcu_assign_pointer(event
->data
, data
);
2349 static void perf_mmap_data_free_rcu(struct rcu_head
*rcu_head
)
2351 struct perf_mmap_data
*data
;
2353 data
= container_of(rcu_head
, struct perf_mmap_data
, rcu_head
);
2354 perf_mmap_data_free(data
);
2358 static void perf_mmap_data_release(struct perf_event
*event
)
2360 struct perf_mmap_data
*data
= event
->data
;
2362 WARN_ON(atomic_read(&event
->mmap_count
));
2364 rcu_assign_pointer(event
->data
, NULL
);
2365 call_rcu(&data
->rcu_head
, perf_mmap_data_free_rcu
);
2368 static void perf_mmap_open(struct vm_area_struct
*vma
)
2370 struct perf_event
*event
= vma
->vm_file
->private_data
;
2372 atomic_inc(&event
->mmap_count
);
2375 static void perf_mmap_close(struct vm_area_struct
*vma
)
2377 struct perf_event
*event
= vma
->vm_file
->private_data
;
2379 WARN_ON_ONCE(event
->ctx
->parent_ctx
);
2380 if (atomic_dec_and_mutex_lock(&event
->mmap_count
, &event
->mmap_mutex
)) {
2381 unsigned long size
= perf_data_size(event
->data
);
2382 struct user_struct
*user
= current_user();
2384 atomic_long_sub((size
>> PAGE_SHIFT
) + 1, &user
->locked_vm
);
2385 vma
->vm_mm
->locked_vm
-= event
->data
->nr_locked
;
2386 perf_mmap_data_release(event
);
2387 mutex_unlock(&event
->mmap_mutex
);
2391 static const struct vm_operations_struct perf_mmap_vmops
= {
2392 .open
= perf_mmap_open
,
2393 .close
= perf_mmap_close
,
2394 .fault
= perf_mmap_fault
,
2395 .page_mkwrite
= perf_mmap_fault
,
2398 static int perf_mmap(struct file
*file
, struct vm_area_struct
*vma
)
2400 struct perf_event
*event
= file
->private_data
;
2401 unsigned long user_locked
, user_lock_limit
;
2402 struct user_struct
*user
= current_user();
2403 unsigned long locked
, lock_limit
;
2404 struct perf_mmap_data
*data
;
2405 unsigned long vma_size
;
2406 unsigned long nr_pages
;
2407 long user_extra
, extra
;
2410 if (!(vma
->vm_flags
& VM_SHARED
))
2413 vma_size
= vma
->vm_end
- vma
->vm_start
;
2414 nr_pages
= (vma_size
/ PAGE_SIZE
) - 1;
2417 * If we have data pages ensure they're a power-of-two number, so we
2418 * can do bitmasks instead of modulo.
2420 if (nr_pages
!= 0 && !is_power_of_2(nr_pages
))
2423 if (vma_size
!= PAGE_SIZE
* (1 + nr_pages
))
2426 if (vma
->vm_pgoff
!= 0)
2429 WARN_ON_ONCE(event
->ctx
->parent_ctx
);
2430 mutex_lock(&event
->mmap_mutex
);
2431 if (event
->output
) {
2436 if (atomic_inc_not_zero(&event
->mmap_count
)) {
2437 if (nr_pages
!= event
->data
->nr_pages
)
2442 user_extra
= nr_pages
+ 1;
2443 user_lock_limit
= sysctl_perf_event_mlock
>> (PAGE_SHIFT
- 10);
2446 * Increase the limit linearly with more CPUs:
2448 user_lock_limit
*= num_online_cpus();
2450 user_locked
= atomic_long_read(&user
->locked_vm
) + user_extra
;
2453 if (user_locked
> user_lock_limit
)
2454 extra
= user_locked
- user_lock_limit
;
2456 lock_limit
= current
->signal
->rlim
[RLIMIT_MEMLOCK
].rlim_cur
;
2457 lock_limit
>>= PAGE_SHIFT
;
2458 locked
= vma
->vm_mm
->locked_vm
+ extra
;
2460 if ((locked
> lock_limit
) && perf_paranoid_tracepoint_raw() &&
2461 !capable(CAP_IPC_LOCK
)) {
2466 WARN_ON(event
->data
);
2468 data
= perf_mmap_data_alloc(event
, nr_pages
);
2474 perf_mmap_data_init(event
, data
);
2476 atomic_set(&event
->mmap_count
, 1);
2477 atomic_long_add(user_extra
, &user
->locked_vm
);
2478 vma
->vm_mm
->locked_vm
+= extra
;
2479 event
->data
->nr_locked
= extra
;
2480 if (vma
->vm_flags
& VM_WRITE
)
2481 event
->data
->writable
= 1;
2484 mutex_unlock(&event
->mmap_mutex
);
2486 vma
->vm_flags
|= VM_RESERVED
;
2487 vma
->vm_ops
= &perf_mmap_vmops
;
2492 static int perf_fasync(int fd
, struct file
*filp
, int on
)
2494 struct inode
*inode
= filp
->f_path
.dentry
->d_inode
;
2495 struct perf_event
*event
= filp
->private_data
;
2498 mutex_lock(&inode
->i_mutex
);
2499 retval
= fasync_helper(fd
, filp
, on
, &event
->fasync
);
2500 mutex_unlock(&inode
->i_mutex
);
2508 static const struct file_operations perf_fops
= {
2509 .release
= perf_release
,
2512 .unlocked_ioctl
= perf_ioctl
,
2513 .compat_ioctl
= perf_ioctl
,
2515 .fasync
= perf_fasync
,
2521 * If there's data, ensure we set the poll() state and publish everything
2522 * to user-space before waking everybody up.
2525 void perf_event_wakeup(struct perf_event
*event
)
2527 wake_up_all(&event
->waitq
);
2529 if (event
->pending_kill
) {
2530 kill_fasync(&event
->fasync
, SIGIO
, event
->pending_kill
);
2531 event
->pending_kill
= 0;
2538 * Handle the case where we need to wakeup up from NMI (or rq->lock) context.
2540 * The NMI bit means we cannot possibly take locks. Therefore, maintain a
2541 * single linked list and use cmpxchg() to add entries lockless.
2544 static void perf_pending_event(struct perf_pending_entry
*entry
)
2546 struct perf_event
*event
= container_of(entry
,
2547 struct perf_event
, pending
);
2549 if (event
->pending_disable
) {
2550 event
->pending_disable
= 0;
2551 __perf_event_disable(event
);
2554 if (event
->pending_wakeup
) {
2555 event
->pending_wakeup
= 0;
2556 perf_event_wakeup(event
);
2560 #define PENDING_TAIL ((struct perf_pending_entry *)-1UL)
2562 static DEFINE_PER_CPU(struct perf_pending_entry
*, perf_pending_head
) = {
2566 static void perf_pending_queue(struct perf_pending_entry
*entry
,
2567 void (*func
)(struct perf_pending_entry
*))
2569 struct perf_pending_entry
**head
;
2571 if (cmpxchg(&entry
->next
, NULL
, PENDING_TAIL
) != NULL
)
2576 head
= &get_cpu_var(perf_pending_head
);
2579 entry
->next
= *head
;
2580 } while (cmpxchg(head
, entry
->next
, entry
) != entry
->next
);
2582 set_perf_event_pending();
2584 put_cpu_var(perf_pending_head
);
2587 static int __perf_pending_run(void)
2589 struct perf_pending_entry
*list
;
2592 list
= xchg(&__get_cpu_var(perf_pending_head
), PENDING_TAIL
);
2593 while (list
!= PENDING_TAIL
) {
2594 void (*func
)(struct perf_pending_entry
*);
2595 struct perf_pending_entry
*entry
= list
;
2602 * Ensure we observe the unqueue before we issue the wakeup,
2603 * so that we won't be waiting forever.
2604 * -- see perf_not_pending().
2615 static inline int perf_not_pending(struct perf_event
*event
)
2618 * If we flush on whatever cpu we run, there is a chance we don't
2622 __perf_pending_run();
2626 * Ensure we see the proper queue state before going to sleep
2627 * so that we do not miss the wakeup. -- see perf_pending_handle()
2630 return event
->pending
.next
== NULL
;
2633 static void perf_pending_sync(struct perf_event
*event
)
2635 wait_event(event
->waitq
, perf_not_pending(event
));
2638 void perf_event_do_pending(void)
2640 __perf_pending_run();
2644 * Callchain support -- arch specific
2647 __weak
struct perf_callchain_entry
*perf_callchain(struct pt_regs
*regs
)
2655 static bool perf_output_space(struct perf_mmap_data
*data
, unsigned long tail
,
2656 unsigned long offset
, unsigned long head
)
2660 if (!data
->writable
)
2663 mask
= perf_data_size(data
) - 1;
2665 offset
= (offset
- tail
) & mask
;
2666 head
= (head
- tail
) & mask
;
2668 if ((int)(head
- offset
) < 0)
2674 static void perf_output_wakeup(struct perf_output_handle
*handle
)
2676 atomic_set(&handle
->data
->poll
, POLL_IN
);
2679 handle
->event
->pending_wakeup
= 1;
2680 perf_pending_queue(&handle
->event
->pending
,
2681 perf_pending_event
);
2683 perf_event_wakeup(handle
->event
);
2687 * Curious locking construct.
2689 * We need to ensure a later event_id doesn't publish a head when a former
2690 * event_id isn't done writing. However since we need to deal with NMIs we
2691 * cannot fully serialize things.
2693 * What we do is serialize between CPUs so we only have to deal with NMI
2694 * nesting on a single CPU.
2696 * We only publish the head (and generate a wakeup) when the outer-most
2697 * event_id completes.
2699 static void perf_output_lock(struct perf_output_handle
*handle
)
2701 struct perf_mmap_data
*data
= handle
->data
;
2702 int cur
, cpu
= get_cpu();
2707 cur
= atomic_cmpxchg(&data
->lock
, -1, cpu
);
2719 static void perf_output_unlock(struct perf_output_handle
*handle
)
2721 struct perf_mmap_data
*data
= handle
->data
;
2725 data
->done_head
= data
->head
;
2727 if (!handle
->locked
)
2732 * The xchg implies a full barrier that ensures all writes are done
2733 * before we publish the new head, matched by a rmb() in userspace when
2734 * reading this position.
2736 while ((head
= atomic_long_xchg(&data
->done_head
, 0)))
2737 data
->user_page
->data_head
= head
;
2740 * NMI can happen here, which means we can miss a done_head update.
2743 cpu
= atomic_xchg(&data
->lock
, -1);
2744 WARN_ON_ONCE(cpu
!= smp_processor_id());
2747 * Therefore we have to validate we did not indeed do so.
2749 if (unlikely(atomic_long_read(&data
->done_head
))) {
2751 * Since we had it locked, we can lock it again.
2753 while (atomic_cmpxchg(&data
->lock
, -1, cpu
) != -1)
2759 if (atomic_xchg(&data
->wakeup
, 0))
2760 perf_output_wakeup(handle
);
2765 void perf_output_copy(struct perf_output_handle
*handle
,
2766 const void *buf
, unsigned int len
)
2768 unsigned int pages_mask
;
2769 unsigned long offset
;
2773 offset
= handle
->offset
;
2774 pages_mask
= handle
->data
->nr_pages
- 1;
2775 pages
= handle
->data
->data_pages
;
2778 unsigned long page_offset
;
2779 unsigned long page_size
;
2782 nr
= (offset
>> PAGE_SHIFT
) & pages_mask
;
2783 page_size
= 1UL << (handle
->data
->data_order
+ PAGE_SHIFT
);
2784 page_offset
= offset
& (page_size
- 1);
2785 size
= min_t(unsigned int, page_size
- page_offset
, len
);
2787 memcpy(pages
[nr
] + page_offset
, buf
, size
);
2794 handle
->offset
= offset
;
2797 * Check we didn't copy past our reservation window, taking the
2798 * possible unsigned int wrap into account.
2800 WARN_ON_ONCE(((long)(handle
->head
- handle
->offset
)) < 0);
2803 int perf_output_begin(struct perf_output_handle
*handle
,
2804 struct perf_event
*event
, unsigned int size
,
2805 int nmi
, int sample
)
2807 struct perf_event
*output_event
;
2808 struct perf_mmap_data
*data
;
2809 unsigned long tail
, offset
, head
;
2812 struct perf_event_header header
;
2819 * For inherited events we send all the output towards the parent.
2822 event
= event
->parent
;
2824 output_event
= rcu_dereference(event
->output
);
2826 event
= output_event
;
2828 data
= rcu_dereference(event
->data
);
2832 handle
->data
= data
;
2833 handle
->event
= event
;
2835 handle
->sample
= sample
;
2837 if (!data
->nr_pages
)
2840 have_lost
= atomic_read(&data
->lost
);
2842 size
+= sizeof(lost_event
);
2844 perf_output_lock(handle
);
2848 * Userspace could choose to issue a mb() before updating the
2849 * tail pointer. So that all reads will be completed before the
2852 tail
= ACCESS_ONCE(data
->user_page
->data_tail
);
2854 offset
= head
= atomic_long_read(&data
->head
);
2856 if (unlikely(!perf_output_space(data
, tail
, offset
, head
)))
2858 } while (atomic_long_cmpxchg(&data
->head
, offset
, head
) != offset
);
2860 handle
->offset
= offset
;
2861 handle
->head
= head
;
2863 if (head
- tail
> data
->watermark
)
2864 atomic_set(&data
->wakeup
, 1);
2867 lost_event
.header
.type
= PERF_RECORD_LOST
;
2868 lost_event
.header
.misc
= 0;
2869 lost_event
.header
.size
= sizeof(lost_event
);
2870 lost_event
.id
= event
->id
;
2871 lost_event
.lost
= atomic_xchg(&data
->lost
, 0);
2873 perf_output_put(handle
, lost_event
);
2879 atomic_inc(&data
->lost
);
2880 perf_output_unlock(handle
);
2887 void perf_output_end(struct perf_output_handle
*handle
)
2889 struct perf_event
*event
= handle
->event
;
2890 struct perf_mmap_data
*data
= handle
->data
;
2892 int wakeup_events
= event
->attr
.wakeup_events
;
2894 if (handle
->sample
&& wakeup_events
) {
2895 int events
= atomic_inc_return(&data
->events
);
2896 if (events
>= wakeup_events
) {
2897 atomic_sub(wakeup_events
, &data
->events
);
2898 atomic_set(&data
->wakeup
, 1);
2902 perf_output_unlock(handle
);
2906 static u32
perf_event_pid(struct perf_event
*event
, struct task_struct
*p
)
2909 * only top level events have the pid namespace they were created in
2912 event
= event
->parent
;
2914 return task_tgid_nr_ns(p
, event
->ns
);
2917 static u32
perf_event_tid(struct perf_event
*event
, struct task_struct
*p
)
2920 * only top level events have the pid namespace they were created in
2923 event
= event
->parent
;
2925 return task_pid_nr_ns(p
, event
->ns
);
2928 static void perf_output_read_one(struct perf_output_handle
*handle
,
2929 struct perf_event
*event
)
2931 u64 read_format
= event
->attr
.read_format
;
2935 values
[n
++] = atomic64_read(&event
->count
);
2936 if (read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
) {
2937 values
[n
++] = event
->total_time_enabled
+
2938 atomic64_read(&event
->child_total_time_enabled
);
2940 if (read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
) {
2941 values
[n
++] = event
->total_time_running
+
2942 atomic64_read(&event
->child_total_time_running
);
2944 if (read_format
& PERF_FORMAT_ID
)
2945 values
[n
++] = primary_event_id(event
);
2947 perf_output_copy(handle
, values
, n
* sizeof(u64
));
2951 * XXX PERF_FORMAT_GROUP vs inherited events seems difficult.
2953 static void perf_output_read_group(struct perf_output_handle
*handle
,
2954 struct perf_event
*event
)
2956 struct perf_event
*leader
= event
->group_leader
, *sub
;
2957 u64 read_format
= event
->attr
.read_format
;
2961 values
[n
++] = 1 + leader
->nr_siblings
;
2963 if (read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
)
2964 values
[n
++] = leader
->total_time_enabled
;
2966 if (read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
)
2967 values
[n
++] = leader
->total_time_running
;
2969 if (leader
!= event
)
2970 leader
->pmu
->read(leader
);
2972 values
[n
++] = atomic64_read(&leader
->count
);
2973 if (read_format
& PERF_FORMAT_ID
)
2974 values
[n
++] = primary_event_id(leader
);
2976 perf_output_copy(handle
, values
, n
* sizeof(u64
));
2978 list_for_each_entry(sub
, &leader
->sibling_list
, group_entry
) {
2982 sub
->pmu
->read(sub
);
2984 values
[n
++] = atomic64_read(&sub
->count
);
2985 if (read_format
& PERF_FORMAT_ID
)
2986 values
[n
++] = primary_event_id(sub
);
2988 perf_output_copy(handle
, values
, n
* sizeof(u64
));
2992 static void perf_output_read(struct perf_output_handle
*handle
,
2993 struct perf_event
*event
)
2995 if (event
->attr
.read_format
& PERF_FORMAT_GROUP
)
2996 perf_output_read_group(handle
, event
);
2998 perf_output_read_one(handle
, event
);
3001 void perf_output_sample(struct perf_output_handle
*handle
,
3002 struct perf_event_header
*header
,
3003 struct perf_sample_data
*data
,
3004 struct perf_event
*event
)
3006 u64 sample_type
= data
->type
;
3008 perf_output_put(handle
, *header
);
3010 if (sample_type
& PERF_SAMPLE_IP
)
3011 perf_output_put(handle
, data
->ip
);
3013 if (sample_type
& PERF_SAMPLE_TID
)
3014 perf_output_put(handle
, data
->tid_entry
);
3016 if (sample_type
& PERF_SAMPLE_TIME
)
3017 perf_output_put(handle
, data
->time
);
3019 if (sample_type
& PERF_SAMPLE_ADDR
)
3020 perf_output_put(handle
, data
->addr
);
3022 if (sample_type
& PERF_SAMPLE_ID
)
3023 perf_output_put(handle
, data
->id
);
3025 if (sample_type
& PERF_SAMPLE_STREAM_ID
)
3026 perf_output_put(handle
, data
->stream_id
);
3028 if (sample_type
& PERF_SAMPLE_CPU
)
3029 perf_output_put(handle
, data
->cpu_entry
);
3031 if (sample_type
& PERF_SAMPLE_PERIOD
)
3032 perf_output_put(handle
, data
->period
);
3034 if (sample_type
& PERF_SAMPLE_READ
)
3035 perf_output_read(handle
, event
);
3037 if (sample_type
& PERF_SAMPLE_CALLCHAIN
) {
3038 if (data
->callchain
) {
3041 if (data
->callchain
)
3042 size
+= data
->callchain
->nr
;
3044 size
*= sizeof(u64
);
3046 perf_output_copy(handle
, data
->callchain
, size
);
3049 perf_output_put(handle
, nr
);
3053 if (sample_type
& PERF_SAMPLE_RAW
) {
3055 perf_output_put(handle
, data
->raw
->size
);
3056 perf_output_copy(handle
, data
->raw
->data
,
3063 .size
= sizeof(u32
),
3066 perf_output_put(handle
, raw
);
3071 void perf_prepare_sample(struct perf_event_header
*header
,
3072 struct perf_sample_data
*data
,
3073 struct perf_event
*event
,
3074 struct pt_regs
*regs
)
3076 u64 sample_type
= event
->attr
.sample_type
;
3078 data
->type
= sample_type
;
3080 header
->type
= PERF_RECORD_SAMPLE
;
3081 header
->size
= sizeof(*header
);
3084 header
->misc
|= perf_misc_flags(regs
);
3086 if (sample_type
& PERF_SAMPLE_IP
) {
3087 data
->ip
= perf_instruction_pointer(regs
);
3089 header
->size
+= sizeof(data
->ip
);
3092 if (sample_type
& PERF_SAMPLE_TID
) {
3093 /* namespace issues */
3094 data
->tid_entry
.pid
= perf_event_pid(event
, current
);
3095 data
->tid_entry
.tid
= perf_event_tid(event
, current
);
3097 header
->size
+= sizeof(data
->tid_entry
);
3100 if (sample_type
& PERF_SAMPLE_TIME
) {
3101 data
->time
= perf_clock();
3103 header
->size
+= sizeof(data
->time
);
3106 if (sample_type
& PERF_SAMPLE_ADDR
)
3107 header
->size
+= sizeof(data
->addr
);
3109 if (sample_type
& PERF_SAMPLE_ID
) {
3110 data
->id
= primary_event_id(event
);
3112 header
->size
+= sizeof(data
->id
);
3115 if (sample_type
& PERF_SAMPLE_STREAM_ID
) {
3116 data
->stream_id
= event
->id
;
3118 header
->size
+= sizeof(data
->stream_id
);
3121 if (sample_type
& PERF_SAMPLE_CPU
) {
3122 data
->cpu_entry
.cpu
= raw_smp_processor_id();
3123 data
->cpu_entry
.reserved
= 0;
3125 header
->size
+= sizeof(data
->cpu_entry
);
3128 if (sample_type
& PERF_SAMPLE_PERIOD
)
3129 header
->size
+= sizeof(data
->period
);
3131 if (sample_type
& PERF_SAMPLE_READ
)
3132 header
->size
+= perf_event_read_size(event
);
3134 if (sample_type
& PERF_SAMPLE_CALLCHAIN
) {
3137 data
->callchain
= perf_callchain(regs
);
3139 if (data
->callchain
)
3140 size
+= data
->callchain
->nr
;
3142 header
->size
+= size
* sizeof(u64
);
3145 if (sample_type
& PERF_SAMPLE_RAW
) {
3146 int size
= sizeof(u32
);
3149 size
+= data
->raw
->size
;
3151 size
+= sizeof(u32
);
3153 WARN_ON_ONCE(size
& (sizeof(u64
)-1));
3154 header
->size
+= size
;
3158 static void perf_event_output(struct perf_event
*event
, int nmi
,
3159 struct perf_sample_data
*data
,
3160 struct pt_regs
*regs
)
3162 struct perf_output_handle handle
;
3163 struct perf_event_header header
;
3165 perf_prepare_sample(&header
, data
, event
, regs
);
3167 if (perf_output_begin(&handle
, event
, header
.size
, nmi
, 1))
3170 perf_output_sample(&handle
, &header
, data
, event
);
3172 perf_output_end(&handle
);
3179 struct perf_read_event
{
3180 struct perf_event_header header
;
3187 perf_event_read_event(struct perf_event
*event
,
3188 struct task_struct
*task
)
3190 struct perf_output_handle handle
;
3191 struct perf_read_event read_event
= {
3193 .type
= PERF_RECORD_READ
,
3195 .size
= sizeof(read_event
) + perf_event_read_size(event
),
3197 .pid
= perf_event_pid(event
, task
),
3198 .tid
= perf_event_tid(event
, task
),
3202 ret
= perf_output_begin(&handle
, event
, read_event
.header
.size
, 0, 0);
3206 perf_output_put(&handle
, read_event
);
3207 perf_output_read(&handle
, event
);
3209 perf_output_end(&handle
);
3213 * task tracking -- fork/exit
3215 * enabled by: attr.comm | attr.mmap | attr.task
3218 struct perf_task_event
{
3219 struct task_struct
*task
;
3220 struct perf_event_context
*task_ctx
;
3223 struct perf_event_header header
;
3233 static void perf_event_task_output(struct perf_event
*event
,
3234 struct perf_task_event
*task_event
)
3236 struct perf_output_handle handle
;
3238 struct task_struct
*task
= task_event
->task
;
3241 size
= task_event
->event_id
.header
.size
;
3242 ret
= perf_output_begin(&handle
, event
, size
, 0, 0);
3247 task_event
->event_id
.pid
= perf_event_pid(event
, task
);
3248 task_event
->event_id
.ppid
= perf_event_pid(event
, current
);
3250 task_event
->event_id
.tid
= perf_event_tid(event
, task
);
3251 task_event
->event_id
.ptid
= perf_event_tid(event
, current
);
3253 task_event
->event_id
.time
= perf_clock();
3255 perf_output_put(&handle
, task_event
->event_id
);
3257 perf_output_end(&handle
);
3260 static int perf_event_task_match(struct perf_event
*event
)
3262 if (event
->attr
.comm
|| event
->attr
.mmap
|| event
->attr
.task
)
3268 static void perf_event_task_ctx(struct perf_event_context
*ctx
,
3269 struct perf_task_event
*task_event
)
3271 struct perf_event
*event
;
3273 list_for_each_entry_rcu(event
, &ctx
->event_list
, event_entry
) {
3274 if (perf_event_task_match(event
))
3275 perf_event_task_output(event
, task_event
);
3279 static void perf_event_task_event(struct perf_task_event
*task_event
)
3281 struct perf_cpu_context
*cpuctx
;
3282 struct perf_event_context
*ctx
= task_event
->task_ctx
;
3285 cpuctx
= &get_cpu_var(perf_cpu_context
);
3286 perf_event_task_ctx(&cpuctx
->ctx
, task_event
);
3287 put_cpu_var(perf_cpu_context
);
3290 ctx
= rcu_dereference(task_event
->task
->perf_event_ctxp
);
3292 perf_event_task_ctx(ctx
, task_event
);
3296 static void perf_event_task(struct task_struct
*task
,
3297 struct perf_event_context
*task_ctx
,
3300 struct perf_task_event task_event
;
3302 if (!atomic_read(&nr_comm_events
) &&
3303 !atomic_read(&nr_mmap_events
) &&
3304 !atomic_read(&nr_task_events
))
3307 task_event
= (struct perf_task_event
){
3309 .task_ctx
= task_ctx
,
3312 .type
= new ? PERF_RECORD_FORK
: PERF_RECORD_EXIT
,
3314 .size
= sizeof(task_event
.event_id
),
3323 perf_event_task_event(&task_event
);
3326 void perf_event_fork(struct task_struct
*task
)
3328 perf_event_task(task
, NULL
, 1);
3335 struct perf_comm_event
{
3336 struct task_struct
*task
;
3341 struct perf_event_header header
;
3348 static void perf_event_comm_output(struct perf_event
*event
,
3349 struct perf_comm_event
*comm_event
)
3351 struct perf_output_handle handle
;
3352 int size
= comm_event
->event_id
.header
.size
;
3353 int ret
= perf_output_begin(&handle
, event
, size
, 0, 0);
3358 comm_event
->event_id
.pid
= perf_event_pid(event
, comm_event
->task
);
3359 comm_event
->event_id
.tid
= perf_event_tid(event
, comm_event
->task
);
3361 perf_output_put(&handle
, comm_event
->event_id
);
3362 perf_output_copy(&handle
, comm_event
->comm
,
3363 comm_event
->comm_size
);
3364 perf_output_end(&handle
);
3367 static int perf_event_comm_match(struct perf_event
*event
)
3369 if (event
->attr
.comm
)
3375 static void perf_event_comm_ctx(struct perf_event_context
*ctx
,
3376 struct perf_comm_event
*comm_event
)
3378 struct perf_event
*event
;
3380 list_for_each_entry_rcu(event
, &ctx
->event_list
, event_entry
) {
3381 if (perf_event_comm_match(event
))
3382 perf_event_comm_output(event
, comm_event
);
3386 static void perf_event_comm_event(struct perf_comm_event
*comm_event
)
3388 struct perf_cpu_context
*cpuctx
;
3389 struct perf_event_context
*ctx
;
3391 char comm
[TASK_COMM_LEN
];
3393 memset(comm
, 0, sizeof(comm
));
3394 strlcpy(comm
, comm_event
->task
->comm
, sizeof(comm
));
3395 size
= ALIGN(strlen(comm
)+1, sizeof(u64
));
3397 comm_event
->comm
= comm
;
3398 comm_event
->comm_size
= size
;
3400 comm_event
->event_id
.header
.size
= sizeof(comm_event
->event_id
) + size
;
3403 cpuctx
= &get_cpu_var(perf_cpu_context
);
3404 perf_event_comm_ctx(&cpuctx
->ctx
, comm_event
);
3405 put_cpu_var(perf_cpu_context
);
3408 * doesn't really matter which of the child contexts the
3409 * events ends up in.
3411 ctx
= rcu_dereference(current
->perf_event_ctxp
);
3413 perf_event_comm_ctx(ctx
, comm_event
);
3417 void perf_event_comm(struct task_struct
*task
)
3419 struct perf_comm_event comm_event
;
3421 if (task
->perf_event_ctxp
)
3422 perf_event_enable_on_exec(task
);
3424 if (!atomic_read(&nr_comm_events
))
3427 comm_event
= (struct perf_comm_event
){
3433 .type
= PERF_RECORD_COMM
,
3442 perf_event_comm_event(&comm_event
);
3449 struct perf_mmap_event
{
3450 struct vm_area_struct
*vma
;
3452 const char *file_name
;
3456 struct perf_event_header header
;
3466 static void perf_event_mmap_output(struct perf_event
*event
,
3467 struct perf_mmap_event
*mmap_event
)
3469 struct perf_output_handle handle
;
3470 int size
= mmap_event
->event_id
.header
.size
;
3471 int ret
= perf_output_begin(&handle
, event
, size
, 0, 0);
3476 mmap_event
->event_id
.pid
= perf_event_pid(event
, current
);
3477 mmap_event
->event_id
.tid
= perf_event_tid(event
, current
);
3479 perf_output_put(&handle
, mmap_event
->event_id
);
3480 perf_output_copy(&handle
, mmap_event
->file_name
,
3481 mmap_event
->file_size
);
3482 perf_output_end(&handle
);
3485 static int perf_event_mmap_match(struct perf_event
*event
,
3486 struct perf_mmap_event
*mmap_event
)
3488 if (event
->attr
.mmap
)
3494 static void perf_event_mmap_ctx(struct perf_event_context
*ctx
,
3495 struct perf_mmap_event
*mmap_event
)
3497 struct perf_event
*event
;
3499 list_for_each_entry_rcu(event
, &ctx
->event_list
, event_entry
) {
3500 if (perf_event_mmap_match(event
, mmap_event
))
3501 perf_event_mmap_output(event
, mmap_event
);
3505 static void perf_event_mmap_event(struct perf_mmap_event
*mmap_event
)
3507 struct perf_cpu_context
*cpuctx
;
3508 struct perf_event_context
*ctx
;
3509 struct vm_area_struct
*vma
= mmap_event
->vma
;
3510 struct file
*file
= vma
->vm_file
;
3516 memset(tmp
, 0, sizeof(tmp
));
3520 * d_path works from the end of the buffer backwards, so we
3521 * need to add enough zero bytes after the string to handle
3522 * the 64bit alignment we do later.
3524 buf
= kzalloc(PATH_MAX
+ sizeof(u64
), GFP_KERNEL
);
3526 name
= strncpy(tmp
, "//enomem", sizeof(tmp
));
3529 name
= d_path(&file
->f_path
, buf
, PATH_MAX
);
3531 name
= strncpy(tmp
, "//toolong", sizeof(tmp
));
3535 if (arch_vma_name(mmap_event
->vma
)) {
3536 name
= strncpy(tmp
, arch_vma_name(mmap_event
->vma
),
3542 name
= strncpy(tmp
, "[vdso]", sizeof(tmp
));
3546 name
= strncpy(tmp
, "//anon", sizeof(tmp
));
3551 size
= ALIGN(strlen(name
)+1, sizeof(u64
));
3553 mmap_event
->file_name
= name
;
3554 mmap_event
->file_size
= size
;
3556 mmap_event
->event_id
.header
.size
= sizeof(mmap_event
->event_id
) + size
;
3559 cpuctx
= &get_cpu_var(perf_cpu_context
);
3560 perf_event_mmap_ctx(&cpuctx
->ctx
, mmap_event
);
3561 put_cpu_var(perf_cpu_context
);
3564 * doesn't really matter which of the child contexts the
3565 * events ends up in.
3567 ctx
= rcu_dereference(current
->perf_event_ctxp
);
3569 perf_event_mmap_ctx(ctx
, mmap_event
);
3575 void __perf_event_mmap(struct vm_area_struct
*vma
)
3577 struct perf_mmap_event mmap_event
;
3579 if (!atomic_read(&nr_mmap_events
))
3582 mmap_event
= (struct perf_mmap_event
){
3588 .type
= PERF_RECORD_MMAP
,
3594 .start
= vma
->vm_start
,
3595 .len
= vma
->vm_end
- vma
->vm_start
,
3596 .pgoff
= vma
->vm_pgoff
,
3600 perf_event_mmap_event(&mmap_event
);
3604 * IRQ throttle logging
3607 static void perf_log_throttle(struct perf_event
*event
, int enable
)
3609 struct perf_output_handle handle
;
3613 struct perf_event_header header
;
3617 } throttle_event
= {
3619 .type
= PERF_RECORD_THROTTLE
,
3621 .size
= sizeof(throttle_event
),
3623 .time
= perf_clock(),
3624 .id
= primary_event_id(event
),
3625 .stream_id
= event
->id
,
3629 throttle_event
.header
.type
= PERF_RECORD_UNTHROTTLE
;
3631 ret
= perf_output_begin(&handle
, event
, sizeof(throttle_event
), 1, 0);
3635 perf_output_put(&handle
, throttle_event
);
3636 perf_output_end(&handle
);
3640 * Generic event overflow handling, sampling.
3643 static int __perf_event_overflow(struct perf_event
*event
, int nmi
,
3644 int throttle
, struct perf_sample_data
*data
,
3645 struct pt_regs
*regs
)
3647 int events
= atomic_read(&event
->event_limit
);
3648 struct hw_perf_event
*hwc
= &event
->hw
;
3651 throttle
= (throttle
&& event
->pmu
->unthrottle
!= NULL
);
3656 if (hwc
->interrupts
!= MAX_INTERRUPTS
) {
3658 if (HZ
* hwc
->interrupts
>
3659 (u64
)sysctl_perf_event_sample_rate
) {
3660 hwc
->interrupts
= MAX_INTERRUPTS
;
3661 perf_log_throttle(event
, 0);
3666 * Keep re-disabling events even though on the previous
3667 * pass we disabled it - just in case we raced with a
3668 * sched-in and the event got enabled again:
3674 if (event
->attr
.freq
) {
3675 u64 now
= perf_clock();
3676 s64 delta
= now
- hwc
->freq_stamp
;
3678 hwc
->freq_stamp
= now
;
3680 if (delta
> 0 && delta
< TICK_NSEC
)
3681 perf_adjust_period(event
, NSEC_PER_SEC
/ (int)delta
);
3685 * XXX event_limit might not quite work as expected on inherited
3689 event
->pending_kill
= POLL_IN
;
3690 if (events
&& atomic_dec_and_test(&event
->event_limit
)) {
3692 event
->pending_kill
= POLL_HUP
;
3694 event
->pending_disable
= 1;
3695 perf_pending_queue(&event
->pending
,
3696 perf_pending_event
);
3698 perf_event_disable(event
);
3701 if (event
->overflow_handler
)
3702 event
->overflow_handler(event
, nmi
, data
, regs
);
3704 perf_event_output(event
, nmi
, data
, regs
);
3709 int perf_event_overflow(struct perf_event
*event
, int nmi
,
3710 struct perf_sample_data
*data
,
3711 struct pt_regs
*regs
)
3713 return __perf_event_overflow(event
, nmi
, 1, data
, regs
);
3717 * Generic software event infrastructure
3721 * We directly increment event->count and keep a second value in
3722 * event->hw.period_left to count intervals. This period event
3723 * is kept in the range [-sample_period, 0] so that we can use the
3727 static u64
perf_swevent_set_period(struct perf_event
*event
)
3729 struct hw_perf_event
*hwc
= &event
->hw
;
3730 u64 period
= hwc
->last_period
;
3734 hwc
->last_period
= hwc
->sample_period
;
3737 old
= val
= atomic64_read(&hwc
->period_left
);
3741 nr
= div64_u64(period
+ val
, period
);
3742 offset
= nr
* period
;
3744 if (atomic64_cmpxchg(&hwc
->period_left
, old
, val
) != old
)
3750 static void perf_swevent_overflow(struct perf_event
*event
, u64 overflow
,
3751 int nmi
, struct perf_sample_data
*data
,
3752 struct pt_regs
*regs
)
3754 struct hw_perf_event
*hwc
= &event
->hw
;
3757 data
->period
= event
->hw
.last_period
;
3759 overflow
= perf_swevent_set_period(event
);
3761 if (hwc
->interrupts
== MAX_INTERRUPTS
)
3764 for (; overflow
; overflow
--) {
3765 if (__perf_event_overflow(event
, nmi
, throttle
,
3768 * We inhibit the overflow from happening when
3769 * hwc->interrupts == MAX_INTERRUPTS.
3777 static void perf_swevent_unthrottle(struct perf_event
*event
)
3780 * Nothing to do, we already reset hwc->interrupts.
3784 static void perf_swevent_add(struct perf_event
*event
, u64 nr
,
3785 int nmi
, struct perf_sample_data
*data
,
3786 struct pt_regs
*regs
)
3788 struct hw_perf_event
*hwc
= &event
->hw
;
3790 atomic64_add(nr
, &event
->count
);
3795 if (!hwc
->sample_period
)
3798 if (nr
== 1 && hwc
->sample_period
== 1 && !event
->attr
.freq
)
3799 return perf_swevent_overflow(event
, 1, nmi
, data
, regs
);
3801 if (atomic64_add_negative(nr
, &hwc
->period_left
))
3804 perf_swevent_overflow(event
, 0, nmi
, data
, regs
);
3807 static int perf_swevent_is_counting(struct perf_event
*event
)
3810 * The event is active, we're good!
3812 if (event
->state
== PERF_EVENT_STATE_ACTIVE
)
3816 * The event is off/error, not counting.
3818 if (event
->state
!= PERF_EVENT_STATE_INACTIVE
)
3822 * The event is inactive, if the context is active
3823 * we're part of a group that didn't make it on the 'pmu',
3826 if (event
->ctx
->is_active
)
3830 * We're inactive and the context is too, this means the
3831 * task is scheduled out, we're counting events that happen
3832 * to us, like migration events.
3837 static int perf_tp_event_match(struct perf_event
*event
,
3838 struct perf_sample_data
*data
);
3840 static int perf_swevent_match(struct perf_event
*event
,
3841 enum perf_type_id type
,
3843 struct perf_sample_data
*data
,
3844 struct pt_regs
*regs
)
3846 if (!perf_swevent_is_counting(event
))
3849 if (event
->attr
.type
!= type
)
3851 if (event
->attr
.config
!= event_id
)
3855 if (event
->attr
.exclude_user
&& user_mode(regs
))
3858 if (event
->attr
.exclude_kernel
&& !user_mode(regs
))
3862 if (event
->attr
.type
== PERF_TYPE_TRACEPOINT
&&
3863 !perf_tp_event_match(event
, data
))
3869 static void perf_swevent_ctx_event(struct perf_event_context
*ctx
,
3870 enum perf_type_id type
,
3871 u32 event_id
, u64 nr
, int nmi
,
3872 struct perf_sample_data
*data
,
3873 struct pt_regs
*regs
)
3875 struct perf_event
*event
;
3877 list_for_each_entry_rcu(event
, &ctx
->event_list
, event_entry
) {
3878 if (perf_swevent_match(event
, type
, event_id
, data
, regs
))
3879 perf_swevent_add(event
, nr
, nmi
, data
, regs
);
3884 * Must be called with preemption disabled
3886 int perf_swevent_get_recursion_context(int **recursion
)
3888 struct perf_cpu_context
*cpuctx
= &__get_cpu_var(perf_cpu_context
);
3891 *recursion
= &cpuctx
->recursion
[3];
3893 *recursion
= &cpuctx
->recursion
[2];
3894 else if (in_softirq())
3895 *recursion
= &cpuctx
->recursion
[1];
3897 *recursion
= &cpuctx
->recursion
[0];
3907 void perf_swevent_put_recursion_context(int *recursion
)
3912 static void __do_perf_sw_event(enum perf_type_id type
, u32 event_id
,
3914 struct perf_sample_data
*data
,
3915 struct pt_regs
*regs
)
3917 struct perf_event_context
*ctx
;
3918 struct perf_cpu_context
*cpuctx
= &__get_cpu_var(perf_cpu_context
);
3921 perf_swevent_ctx_event(&cpuctx
->ctx
, type
, event_id
,
3922 nr
, nmi
, data
, regs
);
3924 * doesn't really matter which of the child contexts the
3925 * events ends up in.
3927 ctx
= rcu_dereference(current
->perf_event_ctxp
);
3929 perf_swevent_ctx_event(ctx
, type
, event_id
, nr
, nmi
, data
, regs
);
3933 static void do_perf_sw_event(enum perf_type_id type
, u32 event_id
,
3935 struct perf_sample_data
*data
,
3936 struct pt_regs
*regs
)
3942 if (perf_swevent_get_recursion_context(&recursion
))
3945 __do_perf_sw_event(type
, event_id
, nr
, nmi
, data
, regs
);
3947 perf_swevent_put_recursion_context(recursion
);
3952 void __perf_sw_event(u32 event_id
, u64 nr
, int nmi
,
3953 struct pt_regs
*regs
, u64 addr
)
3955 struct perf_sample_data data
= {
3959 do_perf_sw_event(PERF_TYPE_SOFTWARE
, event_id
, nr
, nmi
,
3963 static void perf_swevent_read(struct perf_event
*event
)
3967 static int perf_swevent_enable(struct perf_event
*event
)
3969 struct hw_perf_event
*hwc
= &event
->hw
;
3971 if (hwc
->sample_period
) {
3972 hwc
->last_period
= hwc
->sample_period
;
3973 perf_swevent_set_period(event
);
3978 static void perf_swevent_disable(struct perf_event
*event
)
3982 static const struct pmu perf_ops_generic
= {
3983 .enable
= perf_swevent_enable
,
3984 .disable
= perf_swevent_disable
,
3985 .read
= perf_swevent_read
,
3986 .unthrottle
= perf_swevent_unthrottle
,
3990 * hrtimer based swevent callback
3993 static enum hrtimer_restart
perf_swevent_hrtimer(struct hrtimer
*hrtimer
)
3995 enum hrtimer_restart ret
= HRTIMER_RESTART
;
3996 struct perf_sample_data data
;
3997 struct pt_regs
*regs
;
3998 struct perf_event
*event
;
4001 event
= container_of(hrtimer
, struct perf_event
, hw
.hrtimer
);
4002 event
->pmu
->read(event
);
4005 regs
= get_irq_regs();
4007 * In case we exclude kernel IPs or are somehow not in interrupt
4008 * context, provide the next best thing, the user IP.
4010 if ((event
->attr
.exclude_kernel
|| !regs
) &&
4011 !event
->attr
.exclude_user
)
4012 regs
= task_pt_regs(current
);
4015 if (!(event
->attr
.exclude_idle
&& current
->pid
== 0))
4016 if (perf_event_overflow(event
, 0, &data
, regs
))
4017 ret
= HRTIMER_NORESTART
;
4020 period
= max_t(u64
, 10000, event
->hw
.sample_period
);
4021 hrtimer_forward_now(hrtimer
, ns_to_ktime(period
));
4026 static void perf_swevent_start_hrtimer(struct perf_event
*event
)
4028 struct hw_perf_event
*hwc
= &event
->hw
;
4030 hrtimer_init(&hwc
->hrtimer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
4031 hwc
->hrtimer
.function
= perf_swevent_hrtimer
;
4032 if (hwc
->sample_period
) {
4035 if (hwc
->remaining
) {
4036 if (hwc
->remaining
< 0)
4039 period
= hwc
->remaining
;
4042 period
= max_t(u64
, 10000, hwc
->sample_period
);
4044 __hrtimer_start_range_ns(&hwc
->hrtimer
,
4045 ns_to_ktime(period
), 0,
4046 HRTIMER_MODE_REL
, 0);
4050 static void perf_swevent_cancel_hrtimer(struct perf_event
*event
)
4052 struct hw_perf_event
*hwc
= &event
->hw
;
4054 if (hwc
->sample_period
) {
4055 ktime_t remaining
= hrtimer_get_remaining(&hwc
->hrtimer
);
4056 hwc
->remaining
= ktime_to_ns(remaining
);
4058 hrtimer_cancel(&hwc
->hrtimer
);
4063 * Software event: cpu wall time clock
4066 static void cpu_clock_perf_event_update(struct perf_event
*event
)
4068 int cpu
= raw_smp_processor_id();
4072 now
= cpu_clock(cpu
);
4073 prev
= atomic64_read(&event
->hw
.prev_count
);
4074 atomic64_set(&event
->hw
.prev_count
, now
);
4075 atomic64_add(now
- prev
, &event
->count
);
4078 static int cpu_clock_perf_event_enable(struct perf_event
*event
)
4080 struct hw_perf_event
*hwc
= &event
->hw
;
4081 int cpu
= raw_smp_processor_id();
4083 atomic64_set(&hwc
->prev_count
, cpu_clock(cpu
));
4084 perf_swevent_start_hrtimer(event
);
4089 static void cpu_clock_perf_event_disable(struct perf_event
*event
)
4091 perf_swevent_cancel_hrtimer(event
);
4092 cpu_clock_perf_event_update(event
);
4095 static void cpu_clock_perf_event_read(struct perf_event
*event
)
4097 cpu_clock_perf_event_update(event
);
4100 static const struct pmu perf_ops_cpu_clock
= {
4101 .enable
= cpu_clock_perf_event_enable
,
4102 .disable
= cpu_clock_perf_event_disable
,
4103 .read
= cpu_clock_perf_event_read
,
4107 * Software event: task time clock
4110 static void task_clock_perf_event_update(struct perf_event
*event
, u64 now
)
4115 prev
= atomic64_xchg(&event
->hw
.prev_count
, now
);
4117 atomic64_add(delta
, &event
->count
);
4120 static int task_clock_perf_event_enable(struct perf_event
*event
)
4122 struct hw_perf_event
*hwc
= &event
->hw
;
4125 now
= event
->ctx
->time
;
4127 atomic64_set(&hwc
->prev_count
, now
);
4129 perf_swevent_start_hrtimer(event
);
4134 static void task_clock_perf_event_disable(struct perf_event
*event
)
4136 perf_swevent_cancel_hrtimer(event
);
4137 task_clock_perf_event_update(event
, event
->ctx
->time
);
4141 static void task_clock_perf_event_read(struct perf_event
*event
)
4146 update_context_time(event
->ctx
);
4147 time
= event
->ctx
->time
;
4149 u64 now
= perf_clock();
4150 u64 delta
= now
- event
->ctx
->timestamp
;
4151 time
= event
->ctx
->time
+ delta
;
4154 task_clock_perf_event_update(event
, time
);
4157 static const struct pmu perf_ops_task_clock
= {
4158 .enable
= task_clock_perf_event_enable
,
4159 .disable
= task_clock_perf_event_disable
,
4160 .read
= task_clock_perf_event_read
,
4163 #ifdef CONFIG_EVENT_PROFILE
4165 void perf_tp_event(int event_id
, u64 addr
, u64 count
, void *record
,
4168 struct perf_raw_record raw
= {
4173 struct perf_sample_data data
= {
4178 struct pt_regs
*regs
= get_irq_regs();
4181 regs
= task_pt_regs(current
);
4183 /* Trace events already protected against recursion */
4184 __do_perf_sw_event(PERF_TYPE_TRACEPOINT
, event_id
, count
, 1,
4187 EXPORT_SYMBOL_GPL(perf_tp_event
);
4189 static int perf_tp_event_match(struct perf_event
*event
,
4190 struct perf_sample_data
*data
)
4192 void *record
= data
->raw
->data
;
4194 if (likely(!event
->filter
) || filter_match_preds(event
->filter
, record
))
4199 static void tp_perf_event_destroy(struct perf_event
*event
)
4201 ftrace_profile_disable(event
->attr
.config
);
4204 static const struct pmu
*tp_perf_event_init(struct perf_event
*event
)
4207 * Raw tracepoint data is a severe data leak, only allow root to
4210 if ((event
->attr
.sample_type
& PERF_SAMPLE_RAW
) &&
4211 perf_paranoid_tracepoint_raw() &&
4212 !capable(CAP_SYS_ADMIN
))
4213 return ERR_PTR(-EPERM
);
4215 if (ftrace_profile_enable(event
->attr
.config
))
4218 event
->destroy
= tp_perf_event_destroy
;
4220 return &perf_ops_generic
;
4223 static int perf_event_set_filter(struct perf_event
*event
, void __user
*arg
)
4228 if (event
->attr
.type
!= PERF_TYPE_TRACEPOINT
)
4231 filter_str
= strndup_user(arg
, PAGE_SIZE
);
4232 if (IS_ERR(filter_str
))
4233 return PTR_ERR(filter_str
);
4235 ret
= ftrace_profile_set_filter(event
, event
->attr
.config
, filter_str
);
4241 static void perf_event_free_filter(struct perf_event
*event
)
4243 ftrace_profile_free_filter(event
);
4248 static int perf_tp_event_match(struct perf_event
*event
,
4249 struct perf_sample_data
*data
)
4254 static const struct pmu
*tp_perf_event_init(struct perf_event
*event
)
4259 static int perf_event_set_filter(struct perf_event
*event
, void __user
*arg
)
4264 static void perf_event_free_filter(struct perf_event
*event
)
4268 #endif /* CONFIG_EVENT_PROFILE */
4270 #ifdef CONFIG_HAVE_HW_BREAKPOINT
4271 static void bp_perf_event_destroy(struct perf_event
*event
)
4273 release_bp_slot(event
);
4276 static const struct pmu
*bp_perf_event_init(struct perf_event
*bp
)
4280 * The breakpoint is already filled if we haven't created the counter
4281 * through perf syscall
4282 * FIXME: manage to get trigerred to NULL if it comes from syscalls
4285 err
= register_perf_hw_breakpoint(bp
);
4287 err
= __register_perf_hw_breakpoint(bp
);
4289 return ERR_PTR(err
);
4291 bp
->destroy
= bp_perf_event_destroy
;
4293 return &perf_ops_bp
;
4296 void perf_bp_event(struct perf_event
*bp
, void *regs
)
4301 static void bp_perf_event_destroy(struct perf_event
*event
)
4305 static const struct pmu
*bp_perf_event_init(struct perf_event
*bp
)
4310 void perf_bp_event(struct perf_event
*bp
, void *regs
)
4315 atomic_t perf_swevent_enabled
[PERF_COUNT_SW_MAX
];
4317 static void sw_perf_event_destroy(struct perf_event
*event
)
4319 u64 event_id
= event
->attr
.config
;
4321 WARN_ON(event
->parent
);
4323 atomic_dec(&perf_swevent_enabled
[event_id
]);
4326 static const struct pmu
*sw_perf_event_init(struct perf_event
*event
)
4328 const struct pmu
*pmu
= NULL
;
4329 u64 event_id
= event
->attr
.config
;
4332 * Software events (currently) can't in general distinguish
4333 * between user, kernel and hypervisor events.
4334 * However, context switches and cpu migrations are considered
4335 * to be kernel events, and page faults are never hypervisor
4339 case PERF_COUNT_SW_CPU_CLOCK
:
4340 pmu
= &perf_ops_cpu_clock
;
4343 case PERF_COUNT_SW_TASK_CLOCK
:
4345 * If the user instantiates this as a per-cpu event,
4346 * use the cpu_clock event instead.
4348 if (event
->ctx
->task
)
4349 pmu
= &perf_ops_task_clock
;
4351 pmu
= &perf_ops_cpu_clock
;
4354 case PERF_COUNT_SW_PAGE_FAULTS
:
4355 case PERF_COUNT_SW_PAGE_FAULTS_MIN
:
4356 case PERF_COUNT_SW_PAGE_FAULTS_MAJ
:
4357 case PERF_COUNT_SW_CONTEXT_SWITCHES
:
4358 case PERF_COUNT_SW_CPU_MIGRATIONS
:
4359 case PERF_COUNT_SW_ALIGNMENT_FAULTS
:
4360 case PERF_COUNT_SW_EMULATION_FAULTS
:
4361 if (!event
->parent
) {
4362 atomic_inc(&perf_swevent_enabled
[event_id
]);
4363 event
->destroy
= sw_perf_event_destroy
;
4365 pmu
= &perf_ops_generic
;
4373 * Allocate and initialize a event structure
4375 static struct perf_event
*
4376 perf_event_alloc(struct perf_event_attr
*attr
,
4378 struct perf_event_context
*ctx
,
4379 struct perf_event
*group_leader
,
4380 struct perf_event
*parent_event
,
4381 perf_callback_t callback
,
4384 const struct pmu
*pmu
;
4385 struct perf_event
*event
;
4386 struct hw_perf_event
*hwc
;
4389 event
= kzalloc(sizeof(*event
), gfpflags
);
4391 return ERR_PTR(-ENOMEM
);
4394 * Single events are their own group leaders, with an
4395 * empty sibling list:
4398 group_leader
= event
;
4400 mutex_init(&event
->child_mutex
);
4401 INIT_LIST_HEAD(&event
->child_list
);
4403 INIT_LIST_HEAD(&event
->group_entry
);
4404 INIT_LIST_HEAD(&event
->event_entry
);
4405 INIT_LIST_HEAD(&event
->sibling_list
);
4406 init_waitqueue_head(&event
->waitq
);
4408 mutex_init(&event
->mmap_mutex
);
4411 event
->attr
= *attr
;
4412 event
->group_leader
= group_leader
;
4417 event
->parent
= parent_event
;
4419 event
->ns
= get_pid_ns(current
->nsproxy
->pid_ns
);
4420 event
->id
= atomic64_inc_return(&perf_event_id
);
4422 event
->state
= PERF_EVENT_STATE_INACTIVE
;
4424 if (!callback
&& parent_event
)
4425 callback
= parent_event
->callback
;
4427 event
->callback
= callback
;
4430 event
->state
= PERF_EVENT_STATE_OFF
;
4435 hwc
->sample_period
= attr
->sample_period
;
4436 if (attr
->freq
&& attr
->sample_freq
)
4437 hwc
->sample_period
= 1;
4438 hwc
->last_period
= hwc
->sample_period
;
4440 atomic64_set(&hwc
->period_left
, hwc
->sample_period
);
4443 * we currently do not support PERF_FORMAT_GROUP on inherited events
4445 if (attr
->inherit
&& (attr
->read_format
& PERF_FORMAT_GROUP
))
4448 switch (attr
->type
) {
4450 case PERF_TYPE_HARDWARE
:
4451 case PERF_TYPE_HW_CACHE
:
4452 pmu
= hw_perf_event_init(event
);
4455 case PERF_TYPE_SOFTWARE
:
4456 pmu
= sw_perf_event_init(event
);
4459 case PERF_TYPE_TRACEPOINT
:
4460 pmu
= tp_perf_event_init(event
);
4463 case PERF_TYPE_BREAKPOINT
:
4464 pmu
= bp_perf_event_init(event
);
4475 else if (IS_ERR(pmu
))
4480 put_pid_ns(event
->ns
);
4482 return ERR_PTR(err
);
4487 if (!event
->parent
) {
4488 atomic_inc(&nr_events
);
4489 if (event
->attr
.mmap
)
4490 atomic_inc(&nr_mmap_events
);
4491 if (event
->attr
.comm
)
4492 atomic_inc(&nr_comm_events
);
4493 if (event
->attr
.task
)
4494 atomic_inc(&nr_task_events
);
4500 static int perf_copy_attr(struct perf_event_attr __user
*uattr
,
4501 struct perf_event_attr
*attr
)
4506 if (!access_ok(VERIFY_WRITE
, uattr
, PERF_ATTR_SIZE_VER0
))
4510 * zero the full structure, so that a short copy will be nice.
4512 memset(attr
, 0, sizeof(*attr
));
4514 ret
= get_user(size
, &uattr
->size
);
4518 if (size
> PAGE_SIZE
) /* silly large */
4521 if (!size
) /* abi compat */
4522 size
= PERF_ATTR_SIZE_VER0
;
4524 if (size
< PERF_ATTR_SIZE_VER0
)
4528 * If we're handed a bigger struct than we know of,
4529 * ensure all the unknown bits are 0 - i.e. new
4530 * user-space does not rely on any kernel feature
4531 * extensions we dont know about yet.
4533 if (size
> sizeof(*attr
)) {
4534 unsigned char __user
*addr
;
4535 unsigned char __user
*end
;
4538 addr
= (void __user
*)uattr
+ sizeof(*attr
);
4539 end
= (void __user
*)uattr
+ size
;
4541 for (; addr
< end
; addr
++) {
4542 ret
= get_user(val
, addr
);
4548 size
= sizeof(*attr
);
4551 ret
= copy_from_user(attr
, uattr
, size
);
4556 * If the type exists, the corresponding creation will verify
4559 if (attr
->type
>= PERF_TYPE_MAX
)
4562 if (attr
->__reserved_1
|| attr
->__reserved_2
|| attr
->__reserved_3
)
4565 if (attr
->sample_type
& ~(PERF_SAMPLE_MAX
-1))
4568 if (attr
->read_format
& ~(PERF_FORMAT_MAX
-1))
4575 put_user(sizeof(*attr
), &uattr
->size
);
4580 static int perf_event_set_output(struct perf_event
*event
, int output_fd
)
4582 struct perf_event
*output_event
= NULL
;
4583 struct file
*output_file
= NULL
;
4584 struct perf_event
*old_output
;
4585 int fput_needed
= 0;
4591 output_file
= fget_light(output_fd
, &fput_needed
);
4595 if (output_file
->f_op
!= &perf_fops
)
4598 output_event
= output_file
->private_data
;
4600 /* Don't chain output fds */
4601 if (output_event
->output
)
4604 /* Don't set an output fd when we already have an output channel */
4608 atomic_long_inc(&output_file
->f_count
);
4611 mutex_lock(&event
->mmap_mutex
);
4612 old_output
= event
->output
;
4613 rcu_assign_pointer(event
->output
, output_event
);
4614 mutex_unlock(&event
->mmap_mutex
);
4618 * we need to make sure no existing perf_output_*()
4619 * is still referencing this event.
4622 fput(old_output
->filp
);
4627 fput_light(output_file
, fput_needed
);
4632 * sys_perf_event_open - open a performance event, associate it to a task/cpu
4634 * @attr_uptr: event_id type attributes for monitoring/sampling
4637 * @group_fd: group leader event fd
4639 SYSCALL_DEFINE5(perf_event_open
,
4640 struct perf_event_attr __user
*, attr_uptr
,
4641 pid_t
, pid
, int, cpu
, int, group_fd
, unsigned long, flags
)
4643 struct perf_event
*event
, *group_leader
;
4644 struct perf_event_attr attr
;
4645 struct perf_event_context
*ctx
;
4646 struct file
*event_file
= NULL
;
4647 struct file
*group_file
= NULL
;
4648 int fput_needed
= 0;
4649 int fput_needed2
= 0;
4652 /* for future expandability... */
4653 if (flags
& ~(PERF_FLAG_FD_NO_GROUP
| PERF_FLAG_FD_OUTPUT
))
4656 err
= perf_copy_attr(attr_uptr
, &attr
);
4660 if (!attr
.exclude_kernel
) {
4661 if (perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN
))
4666 if (attr
.sample_freq
> sysctl_perf_event_sample_rate
)
4671 * Get the target context (task or percpu):
4673 ctx
= find_get_context(pid
, cpu
);
4675 return PTR_ERR(ctx
);
4678 * Look up the group leader (we will attach this event to it):
4680 group_leader
= NULL
;
4681 if (group_fd
!= -1 && !(flags
& PERF_FLAG_FD_NO_GROUP
)) {
4683 group_file
= fget_light(group_fd
, &fput_needed
);
4685 goto err_put_context
;
4686 if (group_file
->f_op
!= &perf_fops
)
4687 goto err_put_context
;
4689 group_leader
= group_file
->private_data
;
4691 * Do not allow a recursive hierarchy (this new sibling
4692 * becoming part of another group-sibling):
4694 if (group_leader
->group_leader
!= group_leader
)
4695 goto err_put_context
;
4697 * Do not allow to attach to a group in a different
4698 * task or CPU context:
4700 if (group_leader
->ctx
!= ctx
)
4701 goto err_put_context
;
4703 * Only a group leader can be exclusive or pinned
4705 if (attr
.exclusive
|| attr
.pinned
)
4706 goto err_put_context
;
4709 event
= perf_event_alloc(&attr
, cpu
, ctx
, group_leader
,
4710 NULL
, NULL
, GFP_KERNEL
);
4711 err
= PTR_ERR(event
);
4713 goto err_put_context
;
4715 err
= anon_inode_getfd("[perf_event]", &perf_fops
, event
, 0);
4717 goto err_free_put_context
;
4719 event_file
= fget_light(err
, &fput_needed2
);
4721 goto err_free_put_context
;
4723 if (flags
& PERF_FLAG_FD_OUTPUT
) {
4724 err
= perf_event_set_output(event
, group_fd
);
4726 goto err_fput_free_put_context
;
4729 event
->filp
= event_file
;
4730 WARN_ON_ONCE(ctx
->parent_ctx
);
4731 mutex_lock(&ctx
->mutex
);
4732 perf_install_in_context(ctx
, event
, cpu
);
4734 mutex_unlock(&ctx
->mutex
);
4736 event
->owner
= current
;
4737 get_task_struct(current
);
4738 mutex_lock(¤t
->perf_event_mutex
);
4739 list_add_tail(&event
->owner_entry
, ¤t
->perf_event_list
);
4740 mutex_unlock(¤t
->perf_event_mutex
);
4742 err_fput_free_put_context
:
4743 fput_light(event_file
, fput_needed2
);
4745 err_free_put_context
:
4753 fput_light(group_file
, fput_needed
);
4759 * perf_event_create_kernel_counter
4761 * @attr: attributes of the counter to create
4762 * @cpu: cpu in which the counter is bound
4763 * @pid: task to profile
4766 perf_event_create_kernel_counter(struct perf_event_attr
*attr
, int cpu
,
4767 pid_t pid
, perf_callback_t callback
)
4769 struct perf_event
*event
;
4770 struct perf_event_context
*ctx
;
4774 * Get the target context (task or percpu):
4777 ctx
= find_get_context(pid
, cpu
);
4781 event
= perf_event_alloc(attr
, cpu
, ctx
, NULL
,
4782 NULL
, callback
, GFP_KERNEL
);
4783 err
= PTR_ERR(event
);
4785 goto err_put_context
;
4788 WARN_ON_ONCE(ctx
->parent_ctx
);
4789 mutex_lock(&ctx
->mutex
);
4790 perf_install_in_context(ctx
, event
, cpu
);
4792 mutex_unlock(&ctx
->mutex
);
4794 event
->owner
= current
;
4795 get_task_struct(current
);
4796 mutex_lock(¤t
->perf_event_mutex
);
4797 list_add_tail(&event
->owner_entry
, ¤t
->perf_event_list
);
4798 mutex_unlock(¤t
->perf_event_mutex
);
4808 EXPORT_SYMBOL_GPL(perf_event_create_kernel_counter
);
4811 * inherit a event from parent task to child task:
4813 static struct perf_event
*
4814 inherit_event(struct perf_event
*parent_event
,
4815 struct task_struct
*parent
,
4816 struct perf_event_context
*parent_ctx
,
4817 struct task_struct
*child
,
4818 struct perf_event
*group_leader
,
4819 struct perf_event_context
*child_ctx
)
4821 struct perf_event
*child_event
;
4824 * Instead of creating recursive hierarchies of events,
4825 * we link inherited events back to the original parent,
4826 * which has a filp for sure, which we use as the reference
4829 if (parent_event
->parent
)
4830 parent_event
= parent_event
->parent
;
4832 child_event
= perf_event_alloc(&parent_event
->attr
,
4833 parent_event
->cpu
, child_ctx
,
4834 group_leader
, parent_event
,
4836 if (IS_ERR(child_event
))
4841 * Make the child state follow the state of the parent event,
4842 * not its attr.disabled bit. We hold the parent's mutex,
4843 * so we won't race with perf_event_{en, dis}able_family.
4845 if (parent_event
->state
>= PERF_EVENT_STATE_INACTIVE
)
4846 child_event
->state
= PERF_EVENT_STATE_INACTIVE
;
4848 child_event
->state
= PERF_EVENT_STATE_OFF
;
4850 if (parent_event
->attr
.freq
)
4851 child_event
->hw
.sample_period
= parent_event
->hw
.sample_period
;
4853 child_event
->overflow_handler
= parent_event
->overflow_handler
;
4856 * Link it up in the child's context:
4858 add_event_to_ctx(child_event
, child_ctx
);
4861 * Get a reference to the parent filp - we will fput it
4862 * when the child event exits. This is safe to do because
4863 * we are in the parent and we know that the filp still
4864 * exists and has a nonzero count:
4866 atomic_long_inc(&parent_event
->filp
->f_count
);
4869 * Link this into the parent event's child list
4871 WARN_ON_ONCE(parent_event
->ctx
->parent_ctx
);
4872 mutex_lock(&parent_event
->child_mutex
);
4873 list_add_tail(&child_event
->child_list
, &parent_event
->child_list
);
4874 mutex_unlock(&parent_event
->child_mutex
);
4879 static int inherit_group(struct perf_event
*parent_event
,
4880 struct task_struct
*parent
,
4881 struct perf_event_context
*parent_ctx
,
4882 struct task_struct
*child
,
4883 struct perf_event_context
*child_ctx
)
4885 struct perf_event
*leader
;
4886 struct perf_event
*sub
;
4887 struct perf_event
*child_ctr
;
4889 leader
= inherit_event(parent_event
, parent
, parent_ctx
,
4890 child
, NULL
, child_ctx
);
4892 return PTR_ERR(leader
);
4893 list_for_each_entry(sub
, &parent_event
->sibling_list
, group_entry
) {
4894 child_ctr
= inherit_event(sub
, parent
, parent_ctx
,
4895 child
, leader
, child_ctx
);
4896 if (IS_ERR(child_ctr
))
4897 return PTR_ERR(child_ctr
);
4902 static void sync_child_event(struct perf_event
*child_event
,
4903 struct task_struct
*child
)
4905 struct perf_event
*parent_event
= child_event
->parent
;
4908 if (child_event
->attr
.inherit_stat
)
4909 perf_event_read_event(child_event
, child
);
4911 child_val
= atomic64_read(&child_event
->count
);
4914 * Add back the child's count to the parent's count:
4916 atomic64_add(child_val
, &parent_event
->count
);
4917 atomic64_add(child_event
->total_time_enabled
,
4918 &parent_event
->child_total_time_enabled
);
4919 atomic64_add(child_event
->total_time_running
,
4920 &parent_event
->child_total_time_running
);
4923 * Remove this event from the parent's list
4925 WARN_ON_ONCE(parent_event
->ctx
->parent_ctx
);
4926 mutex_lock(&parent_event
->child_mutex
);
4927 list_del_init(&child_event
->child_list
);
4928 mutex_unlock(&parent_event
->child_mutex
);
4931 * Release the parent event, if this was the last
4934 fput(parent_event
->filp
);
4938 __perf_event_exit_task(struct perf_event
*child_event
,
4939 struct perf_event_context
*child_ctx
,
4940 struct task_struct
*child
)
4942 struct perf_event
*parent_event
;
4944 update_event_times(child_event
);
4945 perf_event_remove_from_context(child_event
);
4947 parent_event
= child_event
->parent
;
4949 * It can happen that parent exits first, and has events
4950 * that are still around due to the child reference. These
4951 * events need to be zapped - but otherwise linger.
4954 sync_child_event(child_event
, child
);
4955 free_event(child_event
);
4960 * When a child task exits, feed back event values to parent events.
4962 void perf_event_exit_task(struct task_struct
*child
)
4964 struct perf_event
*child_event
, *tmp
;
4965 struct perf_event_context
*child_ctx
;
4966 unsigned long flags
;
4968 if (likely(!child
->perf_event_ctxp
)) {
4969 perf_event_task(child
, NULL
, 0);
4973 local_irq_save(flags
);
4975 * We can't reschedule here because interrupts are disabled,
4976 * and either child is current or it is a task that can't be
4977 * scheduled, so we are now safe from rescheduling changing
4980 child_ctx
= child
->perf_event_ctxp
;
4981 __perf_event_task_sched_out(child_ctx
);
4984 * Take the context lock here so that if find_get_context is
4985 * reading child->perf_event_ctxp, we wait until it has
4986 * incremented the context's refcount before we do put_ctx below.
4988 spin_lock(&child_ctx
->lock
);
4989 child
->perf_event_ctxp
= NULL
;
4991 * If this context is a clone; unclone it so it can't get
4992 * swapped to another process while we're removing all
4993 * the events from it.
4995 unclone_ctx(child_ctx
);
4996 spin_unlock_irqrestore(&child_ctx
->lock
, flags
);
4999 * Report the task dead after unscheduling the events so that we
5000 * won't get any samples after PERF_RECORD_EXIT. We can however still
5001 * get a few PERF_RECORD_READ events.
5003 perf_event_task(child
, child_ctx
, 0);
5006 * We can recurse on the same lock type through:
5008 * __perf_event_exit_task()
5009 * sync_child_event()
5010 * fput(parent_event->filp)
5012 * mutex_lock(&ctx->mutex)
5014 * But since its the parent context it won't be the same instance.
5016 mutex_lock_nested(&child_ctx
->mutex
, SINGLE_DEPTH_NESTING
);
5019 list_for_each_entry_safe(child_event
, tmp
, &child_ctx
->group_list
,
5021 __perf_event_exit_task(child_event
, child_ctx
, child
);
5024 * If the last event was a group event, it will have appended all
5025 * its siblings to the list, but we obtained 'tmp' before that which
5026 * will still point to the list head terminating the iteration.
5028 if (!list_empty(&child_ctx
->group_list
))
5031 mutex_unlock(&child_ctx
->mutex
);
5037 * free an unexposed, unused context as created by inheritance by
5038 * init_task below, used by fork() in case of fail.
5040 void perf_event_free_task(struct task_struct
*task
)
5042 struct perf_event_context
*ctx
= task
->perf_event_ctxp
;
5043 struct perf_event
*event
, *tmp
;
5048 mutex_lock(&ctx
->mutex
);
5050 list_for_each_entry_safe(event
, tmp
, &ctx
->group_list
, group_entry
) {
5051 struct perf_event
*parent
= event
->parent
;
5053 if (WARN_ON_ONCE(!parent
))
5056 mutex_lock(&parent
->child_mutex
);
5057 list_del_init(&event
->child_list
);
5058 mutex_unlock(&parent
->child_mutex
);
5062 list_del_event(event
, ctx
);
5066 if (!list_empty(&ctx
->group_list
))
5069 mutex_unlock(&ctx
->mutex
);
5075 * Initialize the perf_event context in task_struct
5077 int perf_event_init_task(struct task_struct
*child
)
5079 struct perf_event_context
*child_ctx
, *parent_ctx
;
5080 struct perf_event_context
*cloned_ctx
;
5081 struct perf_event
*event
;
5082 struct task_struct
*parent
= current
;
5083 int inherited_all
= 1;
5086 child
->perf_event_ctxp
= NULL
;
5088 mutex_init(&child
->perf_event_mutex
);
5089 INIT_LIST_HEAD(&child
->perf_event_list
);
5091 if (likely(!parent
->perf_event_ctxp
))
5095 * This is executed from the parent task context, so inherit
5096 * events that have been marked for cloning.
5097 * First allocate and initialize a context for the child.
5100 child_ctx
= kmalloc(sizeof(struct perf_event_context
), GFP_KERNEL
);
5104 __perf_event_init_context(child_ctx
, child
);
5105 child
->perf_event_ctxp
= child_ctx
;
5106 get_task_struct(child
);
5109 * If the parent's context is a clone, pin it so it won't get
5112 parent_ctx
= perf_pin_task_context(parent
);
5115 * No need to check if parent_ctx != NULL here; since we saw
5116 * it non-NULL earlier, the only reason for it to become NULL
5117 * is if we exit, and since we're currently in the middle of
5118 * a fork we can't be exiting at the same time.
5122 * Lock the parent list. No need to lock the child - not PID
5123 * hashed yet and not running, so nobody can access it.
5125 mutex_lock(&parent_ctx
->mutex
);
5128 * We dont have to disable NMIs - we are only looking at
5129 * the list, not manipulating it:
5131 list_for_each_entry(event
, &parent_ctx
->group_list
, group_entry
) {
5133 if (!event
->attr
.inherit
) {
5138 ret
= inherit_group(event
, parent
, parent_ctx
,
5146 if (inherited_all
) {
5148 * Mark the child context as a clone of the parent
5149 * context, or of whatever the parent is a clone of.
5150 * Note that if the parent is a clone, it could get
5151 * uncloned at any point, but that doesn't matter
5152 * because the list of events and the generation
5153 * count can't have changed since we took the mutex.
5155 cloned_ctx
= rcu_dereference(parent_ctx
->parent_ctx
);
5157 child_ctx
->parent_ctx
= cloned_ctx
;
5158 child_ctx
->parent_gen
= parent_ctx
->parent_gen
;
5160 child_ctx
->parent_ctx
= parent_ctx
;
5161 child_ctx
->parent_gen
= parent_ctx
->generation
;
5163 get_ctx(child_ctx
->parent_ctx
);
5166 mutex_unlock(&parent_ctx
->mutex
);
5168 perf_unpin_context(parent_ctx
);
5173 static void __cpuinit
perf_event_init_cpu(int cpu
)
5175 struct perf_cpu_context
*cpuctx
;
5177 cpuctx
= &per_cpu(perf_cpu_context
, cpu
);
5178 __perf_event_init_context(&cpuctx
->ctx
, NULL
);
5180 spin_lock(&perf_resource_lock
);
5181 cpuctx
->max_pertask
= perf_max_events
- perf_reserved_percpu
;
5182 spin_unlock(&perf_resource_lock
);
5184 hw_perf_event_setup(cpu
);
5187 #ifdef CONFIG_HOTPLUG_CPU
5188 static void __perf_event_exit_cpu(void *info
)
5190 struct perf_cpu_context
*cpuctx
= &__get_cpu_var(perf_cpu_context
);
5191 struct perf_event_context
*ctx
= &cpuctx
->ctx
;
5192 struct perf_event
*event
, *tmp
;
5194 list_for_each_entry_safe(event
, tmp
, &ctx
->group_list
, group_entry
)
5195 __perf_event_remove_from_context(event
);
5197 static void perf_event_exit_cpu(int cpu
)
5199 struct perf_cpu_context
*cpuctx
= &per_cpu(perf_cpu_context
, cpu
);
5200 struct perf_event_context
*ctx
= &cpuctx
->ctx
;
5202 mutex_lock(&ctx
->mutex
);
5203 smp_call_function_single(cpu
, __perf_event_exit_cpu
, NULL
, 1);
5204 mutex_unlock(&ctx
->mutex
);
5207 static inline void perf_event_exit_cpu(int cpu
) { }
5210 static int __cpuinit
5211 perf_cpu_notify(struct notifier_block
*self
, unsigned long action
, void *hcpu
)
5213 unsigned int cpu
= (long)hcpu
;
5217 case CPU_UP_PREPARE
:
5218 case CPU_UP_PREPARE_FROZEN
:
5219 perf_event_init_cpu(cpu
);
5223 case CPU_ONLINE_FROZEN
:
5224 hw_perf_event_setup_online(cpu
);
5227 case CPU_DOWN_PREPARE
:
5228 case CPU_DOWN_PREPARE_FROZEN
:
5229 perf_event_exit_cpu(cpu
);
5240 * This has to have a higher priority than migration_notifier in sched.c.
5242 static struct notifier_block __cpuinitdata perf_cpu_nb
= {
5243 .notifier_call
= perf_cpu_notify
,
5247 void __init
perf_event_init(void)
5249 perf_cpu_notify(&perf_cpu_nb
, (unsigned long)CPU_UP_PREPARE
,
5250 (void *)(long)smp_processor_id());
5251 perf_cpu_notify(&perf_cpu_nb
, (unsigned long)CPU_ONLINE
,
5252 (void *)(long)smp_processor_id());
5253 register_cpu_notifier(&perf_cpu_nb
);
5256 static ssize_t
perf_show_reserve_percpu(struct sysdev_class
*class, char *buf
)
5258 return sprintf(buf
, "%d\n", perf_reserved_percpu
);
5262 perf_set_reserve_percpu(struct sysdev_class
*class,
5266 struct perf_cpu_context
*cpuctx
;
5270 err
= strict_strtoul(buf
, 10, &val
);
5273 if (val
> perf_max_events
)
5276 spin_lock(&perf_resource_lock
);
5277 perf_reserved_percpu
= val
;
5278 for_each_online_cpu(cpu
) {
5279 cpuctx
= &per_cpu(perf_cpu_context
, cpu
);
5280 spin_lock_irq(&cpuctx
->ctx
.lock
);
5281 mpt
= min(perf_max_events
- cpuctx
->ctx
.nr_events
,
5282 perf_max_events
- perf_reserved_percpu
);
5283 cpuctx
->max_pertask
= mpt
;
5284 spin_unlock_irq(&cpuctx
->ctx
.lock
);
5286 spin_unlock(&perf_resource_lock
);
5291 static ssize_t
perf_show_overcommit(struct sysdev_class
*class, char *buf
)
5293 return sprintf(buf
, "%d\n", perf_overcommit
);
5297 perf_set_overcommit(struct sysdev_class
*class, const char *buf
, size_t count
)
5302 err
= strict_strtoul(buf
, 10, &val
);
5308 spin_lock(&perf_resource_lock
);
5309 perf_overcommit
= val
;
5310 spin_unlock(&perf_resource_lock
);
5315 static SYSDEV_CLASS_ATTR(
5318 perf_show_reserve_percpu
,
5319 perf_set_reserve_percpu
5322 static SYSDEV_CLASS_ATTR(
5325 perf_show_overcommit
,
5329 static struct attribute
*perfclass_attrs
[] = {
5330 &attr_reserve_percpu
.attr
,
5331 &attr_overcommit
.attr
,
5335 static struct attribute_group perfclass_attr_group
= {
5336 .attrs
= perfclass_attrs
,
5337 .name
= "perf_events",
5340 static int __init
perf_event_sysfs_init(void)
5342 return sysfs_create_group(&cpu_sysdev_class
.kset
.kobj
,
5343 &perfclass_attr_group
);
5345 device_initcall(perf_event_sysfs_init
);