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 static 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 raw_spin_lock_irqsave(&ctx
->lock
, *flags
);
207 if (ctx
!= rcu_dereference(task
->perf_event_ctxp
)) {
208 raw_spin_unlock_irqrestore(&ctx
->lock
, *flags
);
212 if (!atomic_inc_not_zero(&ctx
->refcount
)) {
213 raw_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 raw_spin_unlock_irqrestore(&ctx
->lock
, flags
);
239 static void perf_unpin_context(struct perf_event_context
*ctx
)
243 raw_spin_lock_irqsave(&ctx
->lock
, flags
);
245 raw_spin_unlock_irqrestore(&ctx
->lock
, flags
);
249 static inline u64
perf_clock(void)
251 return cpu_clock(raw_smp_processor_id());
255 * Update the record of the current time in a context.
257 static void update_context_time(struct perf_event_context
*ctx
)
259 u64 now
= perf_clock();
261 ctx
->time
+= now
- ctx
->timestamp
;
262 ctx
->timestamp
= now
;
266 * Update the total_time_enabled and total_time_running fields for a event.
268 static void update_event_times(struct perf_event
*event
)
270 struct perf_event_context
*ctx
= event
->ctx
;
273 if (event
->state
< PERF_EVENT_STATE_INACTIVE
||
274 event
->group_leader
->state
< PERF_EVENT_STATE_INACTIVE
)
280 run_end
= event
->tstamp_stopped
;
282 event
->total_time_enabled
= run_end
- event
->tstamp_enabled
;
284 if (event
->state
== PERF_EVENT_STATE_INACTIVE
)
285 run_end
= event
->tstamp_stopped
;
289 event
->total_time_running
= run_end
- event
->tstamp_running
;
293 * Add a event from the lists for its context.
294 * Must be called with ctx->mutex and ctx->lock held.
297 list_add_event(struct perf_event
*event
, struct perf_event_context
*ctx
)
299 struct perf_event
*group_leader
= event
->group_leader
;
302 * Depending on whether it is a standalone or sibling event,
303 * add it straight to the context's event list, or to the group
304 * leader's sibling list:
306 if (group_leader
== event
)
307 list_add_tail(&event
->group_entry
, &ctx
->group_list
);
309 list_add_tail(&event
->group_entry
, &group_leader
->sibling_list
);
310 group_leader
->nr_siblings
++;
313 list_add_rcu(&event
->event_entry
, &ctx
->event_list
);
315 if (event
->attr
.inherit_stat
)
320 * Remove a event from the lists for its context.
321 * Must be called with ctx->mutex and ctx->lock held.
324 list_del_event(struct perf_event
*event
, struct perf_event_context
*ctx
)
326 struct perf_event
*sibling
, *tmp
;
328 if (list_empty(&event
->group_entry
))
331 if (event
->attr
.inherit_stat
)
334 list_del_init(&event
->group_entry
);
335 list_del_rcu(&event
->event_entry
);
337 if (event
->group_leader
!= event
)
338 event
->group_leader
->nr_siblings
--;
340 update_event_times(event
);
343 * If event was in error state, then keep it
344 * that way, otherwise bogus counts will be
345 * returned on read(). The only way to get out
346 * of error state is by explicit re-enabling
349 if (event
->state
> PERF_EVENT_STATE_OFF
)
350 event
->state
= PERF_EVENT_STATE_OFF
;
353 * If this was a group event with sibling events then
354 * upgrade the siblings to singleton events by adding them
355 * to the context list directly:
357 list_for_each_entry_safe(sibling
, tmp
, &event
->sibling_list
, group_entry
) {
359 list_move_tail(&sibling
->group_entry
, &ctx
->group_list
);
360 sibling
->group_leader
= sibling
;
365 event_sched_out(struct perf_event
*event
,
366 struct perf_cpu_context
*cpuctx
,
367 struct perf_event_context
*ctx
)
369 if (event
->state
!= PERF_EVENT_STATE_ACTIVE
)
372 event
->state
= PERF_EVENT_STATE_INACTIVE
;
373 if (event
->pending_disable
) {
374 event
->pending_disable
= 0;
375 event
->state
= PERF_EVENT_STATE_OFF
;
377 event
->tstamp_stopped
= ctx
->time
;
378 event
->pmu
->disable(event
);
381 if (!is_software_event(event
))
382 cpuctx
->active_oncpu
--;
384 if (event
->attr
.exclusive
|| !cpuctx
->active_oncpu
)
385 cpuctx
->exclusive
= 0;
389 group_sched_out(struct perf_event
*group_event
,
390 struct perf_cpu_context
*cpuctx
,
391 struct perf_event_context
*ctx
)
393 struct perf_event
*event
;
395 if (group_event
->state
!= PERF_EVENT_STATE_ACTIVE
)
398 event_sched_out(group_event
, cpuctx
, ctx
);
401 * Schedule out siblings (if any):
403 list_for_each_entry(event
, &group_event
->sibling_list
, group_entry
)
404 event_sched_out(event
, cpuctx
, ctx
);
406 if (group_event
->attr
.exclusive
)
407 cpuctx
->exclusive
= 0;
411 * Cross CPU call to remove a performance event
413 * We disable the event on the hardware level first. After that we
414 * remove it from the context list.
416 static void __perf_event_remove_from_context(void *info
)
418 struct perf_cpu_context
*cpuctx
= &__get_cpu_var(perf_cpu_context
);
419 struct perf_event
*event
= info
;
420 struct perf_event_context
*ctx
= event
->ctx
;
423 * If this is a task context, we need to check whether it is
424 * the current task context of this cpu. If not it has been
425 * scheduled out before the smp call arrived.
427 if (ctx
->task
&& cpuctx
->task_ctx
!= ctx
)
430 raw_spin_lock(&ctx
->lock
);
432 * Protect the list operation against NMI by disabling the
433 * events on a global level.
437 event_sched_out(event
, cpuctx
, ctx
);
439 list_del_event(event
, ctx
);
443 * Allow more per task events with respect to the
446 cpuctx
->max_pertask
=
447 min(perf_max_events
- ctx
->nr_events
,
448 perf_max_events
- perf_reserved_percpu
);
452 raw_spin_unlock(&ctx
->lock
);
457 * Remove the event from a task's (or a CPU's) list of events.
459 * Must be called with ctx->mutex held.
461 * CPU events are removed with a smp call. For task events we only
462 * call when the task is on a CPU.
464 * If event->ctx is a cloned context, callers must make sure that
465 * every task struct that event->ctx->task could possibly point to
466 * remains valid. This is OK when called from perf_release since
467 * that only calls us on the top-level context, which can't be a clone.
468 * When called from perf_event_exit_task, it's OK because the
469 * context has been detached from its task.
471 static void perf_event_remove_from_context(struct perf_event
*event
)
473 struct perf_event_context
*ctx
= event
->ctx
;
474 struct task_struct
*task
= ctx
->task
;
478 * Per cpu events are removed via an smp call and
479 * the removal is always successful.
481 smp_call_function_single(event
->cpu
,
482 __perf_event_remove_from_context
,
488 task_oncpu_function_call(task
, __perf_event_remove_from_context
,
491 raw_spin_lock_irq(&ctx
->lock
);
493 * If the context is active we need to retry the smp call.
495 if (ctx
->nr_active
&& !list_empty(&event
->group_entry
)) {
496 raw_spin_unlock_irq(&ctx
->lock
);
501 * The lock prevents that this context is scheduled in so we
502 * can remove the event safely, if the call above did not
505 if (!list_empty(&event
->group_entry
))
506 list_del_event(event
, ctx
);
507 raw_spin_unlock_irq(&ctx
->lock
);
511 * Update total_time_enabled and total_time_running for all events in a group.
513 static void update_group_times(struct perf_event
*leader
)
515 struct perf_event
*event
;
517 update_event_times(leader
);
518 list_for_each_entry(event
, &leader
->sibling_list
, group_entry
)
519 update_event_times(event
);
523 * Cross CPU call to disable a performance event
525 static void __perf_event_disable(void *info
)
527 struct perf_event
*event
= info
;
528 struct perf_cpu_context
*cpuctx
= &__get_cpu_var(perf_cpu_context
);
529 struct perf_event_context
*ctx
= event
->ctx
;
532 * If this is a per-task event, need to check whether this
533 * event's task is the current task on this cpu.
535 if (ctx
->task
&& cpuctx
->task_ctx
!= ctx
)
538 raw_spin_lock(&ctx
->lock
);
541 * If the event is on, turn it off.
542 * If it is in error state, leave it in error state.
544 if (event
->state
>= PERF_EVENT_STATE_INACTIVE
) {
545 update_context_time(ctx
);
546 update_group_times(event
);
547 if (event
== event
->group_leader
)
548 group_sched_out(event
, cpuctx
, ctx
);
550 event_sched_out(event
, cpuctx
, ctx
);
551 event
->state
= PERF_EVENT_STATE_OFF
;
554 raw_spin_unlock(&ctx
->lock
);
560 * If event->ctx is a cloned context, callers must make sure that
561 * every task struct that event->ctx->task could possibly point to
562 * remains valid. This condition is satisifed when called through
563 * perf_event_for_each_child or perf_event_for_each because they
564 * hold the top-level event's child_mutex, so any descendant that
565 * goes to exit will block in sync_child_event.
566 * When called from perf_pending_event it's OK because event->ctx
567 * is the current context on this CPU and preemption is disabled,
568 * hence we can't get into perf_event_task_sched_out for this context.
570 void perf_event_disable(struct perf_event
*event
)
572 struct perf_event_context
*ctx
= event
->ctx
;
573 struct task_struct
*task
= ctx
->task
;
577 * Disable the event on the cpu that it's on
579 smp_call_function_single(event
->cpu
, __perf_event_disable
,
585 task_oncpu_function_call(task
, __perf_event_disable
, event
);
587 raw_spin_lock_irq(&ctx
->lock
);
589 * If the event is still active, we need to retry the cross-call.
591 if (event
->state
== PERF_EVENT_STATE_ACTIVE
) {
592 raw_spin_unlock_irq(&ctx
->lock
);
597 * Since we have the lock this context can't be scheduled
598 * in, so we can change the state safely.
600 if (event
->state
== PERF_EVENT_STATE_INACTIVE
) {
601 update_group_times(event
);
602 event
->state
= PERF_EVENT_STATE_OFF
;
605 raw_spin_unlock_irq(&ctx
->lock
);
609 event_sched_in(struct perf_event
*event
,
610 struct perf_cpu_context
*cpuctx
,
611 struct perf_event_context
*ctx
,
614 if (event
->state
<= PERF_EVENT_STATE_OFF
)
617 event
->state
= PERF_EVENT_STATE_ACTIVE
;
618 event
->oncpu
= cpu
; /* TODO: put 'cpu' into cpuctx->cpu */
620 * The new state must be visible before we turn it on in the hardware:
624 if (event
->pmu
->enable(event
)) {
625 event
->state
= PERF_EVENT_STATE_INACTIVE
;
630 event
->tstamp_running
+= ctx
->time
- event
->tstamp_stopped
;
632 if (!is_software_event(event
))
633 cpuctx
->active_oncpu
++;
636 if (event
->attr
.exclusive
)
637 cpuctx
->exclusive
= 1;
643 group_sched_in(struct perf_event
*group_event
,
644 struct perf_cpu_context
*cpuctx
,
645 struct perf_event_context
*ctx
,
648 struct perf_event
*event
, *partial_group
;
651 if (group_event
->state
== PERF_EVENT_STATE_OFF
)
654 ret
= hw_perf_group_sched_in(group_event
, cpuctx
, ctx
, cpu
);
656 return ret
< 0 ? ret
: 0;
658 if (event_sched_in(group_event
, cpuctx
, ctx
, cpu
))
662 * Schedule in siblings as one group (if any):
664 list_for_each_entry(event
, &group_event
->sibling_list
, group_entry
) {
665 if (event_sched_in(event
, cpuctx
, ctx
, cpu
)) {
666 partial_group
= event
;
675 * Groups can be scheduled in as one unit only, so undo any
676 * partial group before returning:
678 list_for_each_entry(event
, &group_event
->sibling_list
, group_entry
) {
679 if (event
== partial_group
)
681 event_sched_out(event
, cpuctx
, ctx
);
683 event_sched_out(group_event
, cpuctx
, ctx
);
689 * Return 1 for a group consisting entirely of software events,
690 * 0 if the group contains any hardware events.
692 static int is_software_only_group(struct perf_event
*leader
)
694 struct perf_event
*event
;
696 if (!is_software_event(leader
))
699 list_for_each_entry(event
, &leader
->sibling_list
, group_entry
)
700 if (!is_software_event(event
))
707 * Work out whether we can put this event group on the CPU now.
709 static int group_can_go_on(struct perf_event
*event
,
710 struct perf_cpu_context
*cpuctx
,
714 * Groups consisting entirely of software events can always go on.
716 if (is_software_only_group(event
))
719 * If an exclusive group is already on, no other hardware
722 if (cpuctx
->exclusive
)
725 * If this group is exclusive and there are already
726 * events on the CPU, it can't go on.
728 if (event
->attr
.exclusive
&& cpuctx
->active_oncpu
)
731 * Otherwise, try to add it if all previous groups were able
737 static void add_event_to_ctx(struct perf_event
*event
,
738 struct perf_event_context
*ctx
)
740 list_add_event(event
, ctx
);
741 event
->tstamp_enabled
= ctx
->time
;
742 event
->tstamp_running
= ctx
->time
;
743 event
->tstamp_stopped
= ctx
->time
;
747 * Cross CPU call to install and enable a performance event
749 * Must be called with ctx->mutex held
751 static void __perf_install_in_context(void *info
)
753 struct perf_cpu_context
*cpuctx
= &__get_cpu_var(perf_cpu_context
);
754 struct perf_event
*event
= info
;
755 struct perf_event_context
*ctx
= event
->ctx
;
756 struct perf_event
*leader
= event
->group_leader
;
757 int cpu
= smp_processor_id();
761 * If this is a task context, we need to check whether it is
762 * the current task context of this cpu. If not it has been
763 * scheduled out before the smp call arrived.
764 * Or possibly this is the right context but it isn't
765 * on this cpu because it had no events.
767 if (ctx
->task
&& cpuctx
->task_ctx
!= ctx
) {
768 if (cpuctx
->task_ctx
|| ctx
->task
!= current
)
770 cpuctx
->task_ctx
= ctx
;
773 raw_spin_lock(&ctx
->lock
);
775 update_context_time(ctx
);
778 * Protect the list operation against NMI by disabling the
779 * events on a global level. NOP for non NMI based events.
783 add_event_to_ctx(event
, ctx
);
785 if (event
->cpu
!= -1 && event
->cpu
!= smp_processor_id())
789 * Don't put the event on if it is disabled or if
790 * it is in a group and the group isn't on.
792 if (event
->state
!= PERF_EVENT_STATE_INACTIVE
||
793 (leader
!= event
&& leader
->state
!= PERF_EVENT_STATE_ACTIVE
))
797 * An exclusive event can't go on if there are already active
798 * hardware events, and no hardware event can go on if there
799 * is already an exclusive event on.
801 if (!group_can_go_on(event
, cpuctx
, 1))
804 err
= event_sched_in(event
, cpuctx
, ctx
, cpu
);
808 * This event couldn't go on. If it is in a group
809 * then we have to pull the whole group off.
810 * If the event group is pinned then put it in error state.
813 group_sched_out(leader
, cpuctx
, ctx
);
814 if (leader
->attr
.pinned
) {
815 update_group_times(leader
);
816 leader
->state
= PERF_EVENT_STATE_ERROR
;
820 if (!err
&& !ctx
->task
&& cpuctx
->max_pertask
)
821 cpuctx
->max_pertask
--;
826 raw_spin_unlock(&ctx
->lock
);
830 * Attach a performance event to a context
832 * First we add the event to the list with the hardware enable bit
833 * in event->hw_config cleared.
835 * If the event is attached to a task which is on a CPU we use a smp
836 * call to enable it in the task context. The task might have been
837 * scheduled away, but we check this in the smp call again.
839 * Must be called with ctx->mutex held.
842 perf_install_in_context(struct perf_event_context
*ctx
,
843 struct perf_event
*event
,
846 struct task_struct
*task
= ctx
->task
;
850 * Per cpu events are installed via an smp call and
851 * the install is always successful.
853 smp_call_function_single(cpu
, __perf_install_in_context
,
859 task_oncpu_function_call(task
, __perf_install_in_context
,
862 raw_spin_lock_irq(&ctx
->lock
);
864 * we need to retry the smp call.
866 if (ctx
->is_active
&& list_empty(&event
->group_entry
)) {
867 raw_spin_unlock_irq(&ctx
->lock
);
872 * The lock prevents that this context is scheduled in so we
873 * can add the event safely, if it the call above did not
876 if (list_empty(&event
->group_entry
))
877 add_event_to_ctx(event
, ctx
);
878 raw_spin_unlock_irq(&ctx
->lock
);
882 * Put a event into inactive state and update time fields.
883 * Enabling the leader of a group effectively enables all
884 * the group members that aren't explicitly disabled, so we
885 * have to update their ->tstamp_enabled also.
886 * Note: this works for group members as well as group leaders
887 * since the non-leader members' sibling_lists will be empty.
889 static void __perf_event_mark_enabled(struct perf_event
*event
,
890 struct perf_event_context
*ctx
)
892 struct perf_event
*sub
;
894 event
->state
= PERF_EVENT_STATE_INACTIVE
;
895 event
->tstamp_enabled
= ctx
->time
- event
->total_time_enabled
;
896 list_for_each_entry(sub
, &event
->sibling_list
, group_entry
)
897 if (sub
->state
>= PERF_EVENT_STATE_INACTIVE
)
898 sub
->tstamp_enabled
=
899 ctx
->time
- sub
->total_time_enabled
;
903 * Cross CPU call to enable a performance event
905 static void __perf_event_enable(void *info
)
907 struct perf_event
*event
= info
;
908 struct perf_cpu_context
*cpuctx
= &__get_cpu_var(perf_cpu_context
);
909 struct perf_event_context
*ctx
= event
->ctx
;
910 struct perf_event
*leader
= event
->group_leader
;
914 * If this is a per-task event, need to check whether this
915 * event's task is the current task on this cpu.
917 if (ctx
->task
&& cpuctx
->task_ctx
!= ctx
) {
918 if (cpuctx
->task_ctx
|| ctx
->task
!= current
)
920 cpuctx
->task_ctx
= ctx
;
923 raw_spin_lock(&ctx
->lock
);
925 update_context_time(ctx
);
927 if (event
->state
>= PERF_EVENT_STATE_INACTIVE
)
929 __perf_event_mark_enabled(event
, ctx
);
931 if (event
->cpu
!= -1 && event
->cpu
!= smp_processor_id())
935 * If the event is in a group and isn't the group leader,
936 * then don't put it on unless the group is on.
938 if (leader
!= event
&& leader
->state
!= PERF_EVENT_STATE_ACTIVE
)
941 if (!group_can_go_on(event
, cpuctx
, 1)) {
946 err
= group_sched_in(event
, cpuctx
, ctx
,
949 err
= event_sched_in(event
, cpuctx
, ctx
,
956 * If this event can't go on and it's part of a
957 * group, then the whole group has to come off.
960 group_sched_out(leader
, cpuctx
, ctx
);
961 if (leader
->attr
.pinned
) {
962 update_group_times(leader
);
963 leader
->state
= PERF_EVENT_STATE_ERROR
;
968 raw_spin_unlock(&ctx
->lock
);
974 * If event->ctx is a cloned context, callers must make sure that
975 * every task struct that event->ctx->task could possibly point to
976 * remains valid. This condition is satisfied when called through
977 * perf_event_for_each_child or perf_event_for_each as described
978 * for perf_event_disable.
980 void perf_event_enable(struct perf_event
*event
)
982 struct perf_event_context
*ctx
= event
->ctx
;
983 struct task_struct
*task
= ctx
->task
;
987 * Enable the event on the cpu that it's on
989 smp_call_function_single(event
->cpu
, __perf_event_enable
,
994 raw_spin_lock_irq(&ctx
->lock
);
995 if (event
->state
>= PERF_EVENT_STATE_INACTIVE
)
999 * If the event is in error state, clear that first.
1000 * That way, if we see the event in error state below, we
1001 * know that it has gone back into error state, as distinct
1002 * from the task having been scheduled away before the
1003 * cross-call arrived.
1005 if (event
->state
== PERF_EVENT_STATE_ERROR
)
1006 event
->state
= PERF_EVENT_STATE_OFF
;
1009 raw_spin_unlock_irq(&ctx
->lock
);
1010 task_oncpu_function_call(task
, __perf_event_enable
, event
);
1012 raw_spin_lock_irq(&ctx
->lock
);
1015 * If the context is active and the event is still off,
1016 * we need to retry the cross-call.
1018 if (ctx
->is_active
&& event
->state
== PERF_EVENT_STATE_OFF
)
1022 * Since we have the lock this context can't be scheduled
1023 * in, so we can change the state safely.
1025 if (event
->state
== PERF_EVENT_STATE_OFF
)
1026 __perf_event_mark_enabled(event
, ctx
);
1029 raw_spin_unlock_irq(&ctx
->lock
);
1032 static int perf_event_refresh(struct perf_event
*event
, int refresh
)
1035 * not supported on inherited events
1037 if (event
->attr
.inherit
)
1040 atomic_add(refresh
, &event
->event_limit
);
1041 perf_event_enable(event
);
1046 void __perf_event_sched_out(struct perf_event_context
*ctx
,
1047 struct perf_cpu_context
*cpuctx
)
1049 struct perf_event
*event
;
1051 raw_spin_lock(&ctx
->lock
);
1053 if (likely(!ctx
->nr_events
))
1055 update_context_time(ctx
);
1058 if (ctx
->nr_active
) {
1059 list_for_each_entry(event
, &ctx
->group_list
, group_entry
)
1060 group_sched_out(event
, cpuctx
, ctx
);
1064 raw_spin_unlock(&ctx
->lock
);
1068 * Test whether two contexts are equivalent, i.e. whether they
1069 * have both been cloned from the same version of the same context
1070 * and they both have the same number of enabled events.
1071 * If the number of enabled events is the same, then the set
1072 * of enabled events should be the same, because these are both
1073 * inherited contexts, therefore we can't access individual events
1074 * in them directly with an fd; we can only enable/disable all
1075 * events via prctl, or enable/disable all events in a family
1076 * via ioctl, which will have the same effect on both contexts.
1078 static int context_equiv(struct perf_event_context
*ctx1
,
1079 struct perf_event_context
*ctx2
)
1081 return ctx1
->parent_ctx
&& ctx1
->parent_ctx
== ctx2
->parent_ctx
1082 && ctx1
->parent_gen
== ctx2
->parent_gen
1083 && !ctx1
->pin_count
&& !ctx2
->pin_count
;
1086 static void __perf_event_sync_stat(struct perf_event
*event
,
1087 struct perf_event
*next_event
)
1091 if (!event
->attr
.inherit_stat
)
1095 * Update the event value, we cannot use perf_event_read()
1096 * because we're in the middle of a context switch and have IRQs
1097 * disabled, which upsets smp_call_function_single(), however
1098 * we know the event must be on the current CPU, therefore we
1099 * don't need to use it.
1101 switch (event
->state
) {
1102 case PERF_EVENT_STATE_ACTIVE
:
1103 event
->pmu
->read(event
);
1106 case PERF_EVENT_STATE_INACTIVE
:
1107 update_event_times(event
);
1115 * In order to keep per-task stats reliable we need to flip the event
1116 * values when we flip the contexts.
1118 value
= atomic64_read(&next_event
->count
);
1119 value
= atomic64_xchg(&event
->count
, value
);
1120 atomic64_set(&next_event
->count
, value
);
1122 swap(event
->total_time_enabled
, next_event
->total_time_enabled
);
1123 swap(event
->total_time_running
, next_event
->total_time_running
);
1126 * Since we swizzled the values, update the user visible data too.
1128 perf_event_update_userpage(event
);
1129 perf_event_update_userpage(next_event
);
1132 #define list_next_entry(pos, member) \
1133 list_entry(pos->member.next, typeof(*pos), member)
1135 static void perf_event_sync_stat(struct perf_event_context
*ctx
,
1136 struct perf_event_context
*next_ctx
)
1138 struct perf_event
*event
, *next_event
;
1143 update_context_time(ctx
);
1145 event
= list_first_entry(&ctx
->event_list
,
1146 struct perf_event
, event_entry
);
1148 next_event
= list_first_entry(&next_ctx
->event_list
,
1149 struct perf_event
, event_entry
);
1151 while (&event
->event_entry
!= &ctx
->event_list
&&
1152 &next_event
->event_entry
!= &next_ctx
->event_list
) {
1154 __perf_event_sync_stat(event
, next_event
);
1156 event
= list_next_entry(event
, event_entry
);
1157 next_event
= list_next_entry(next_event
, event_entry
);
1162 * Called from scheduler to remove the events of the current task,
1163 * with interrupts disabled.
1165 * We stop each event and update the event value in event->count.
1167 * This does not protect us against NMI, but disable()
1168 * sets the disabled bit in the control field of event _before_
1169 * accessing the event control register. If a NMI hits, then it will
1170 * not restart the event.
1172 void perf_event_task_sched_out(struct task_struct
*task
,
1173 struct task_struct
*next
, int cpu
)
1175 struct perf_cpu_context
*cpuctx
= &per_cpu(perf_cpu_context
, cpu
);
1176 struct perf_event_context
*ctx
= task
->perf_event_ctxp
;
1177 struct perf_event_context
*next_ctx
;
1178 struct perf_event_context
*parent
;
1179 struct pt_regs
*regs
;
1182 regs
= task_pt_regs(task
);
1183 perf_sw_event(PERF_COUNT_SW_CONTEXT_SWITCHES
, 1, 1, regs
, 0);
1185 if (likely(!ctx
|| !cpuctx
->task_ctx
))
1189 parent
= rcu_dereference(ctx
->parent_ctx
);
1190 next_ctx
= next
->perf_event_ctxp
;
1191 if (parent
&& next_ctx
&&
1192 rcu_dereference(next_ctx
->parent_ctx
) == parent
) {
1194 * Looks like the two contexts are clones, so we might be
1195 * able to optimize the context switch. We lock both
1196 * contexts and check that they are clones under the
1197 * lock (including re-checking that neither has been
1198 * uncloned in the meantime). It doesn't matter which
1199 * order we take the locks because no other cpu could
1200 * be trying to lock both of these tasks.
1202 raw_spin_lock(&ctx
->lock
);
1203 raw_spin_lock_nested(&next_ctx
->lock
, SINGLE_DEPTH_NESTING
);
1204 if (context_equiv(ctx
, next_ctx
)) {
1206 * XXX do we need a memory barrier of sorts
1207 * wrt to rcu_dereference() of perf_event_ctxp
1209 task
->perf_event_ctxp
= next_ctx
;
1210 next
->perf_event_ctxp
= ctx
;
1212 next_ctx
->task
= task
;
1215 perf_event_sync_stat(ctx
, next_ctx
);
1217 raw_spin_unlock(&next_ctx
->lock
);
1218 raw_spin_unlock(&ctx
->lock
);
1223 __perf_event_sched_out(ctx
, cpuctx
);
1224 cpuctx
->task_ctx
= NULL
;
1229 * Called with IRQs disabled
1231 static void __perf_event_task_sched_out(struct perf_event_context
*ctx
)
1233 struct perf_cpu_context
*cpuctx
= &__get_cpu_var(perf_cpu_context
);
1235 if (!cpuctx
->task_ctx
)
1238 if (WARN_ON_ONCE(ctx
!= cpuctx
->task_ctx
))
1241 __perf_event_sched_out(ctx
, cpuctx
);
1242 cpuctx
->task_ctx
= NULL
;
1246 * Called with IRQs disabled
1248 static void perf_event_cpu_sched_out(struct perf_cpu_context
*cpuctx
)
1250 __perf_event_sched_out(&cpuctx
->ctx
, cpuctx
);
1254 __perf_event_sched_in(struct perf_event_context
*ctx
,
1255 struct perf_cpu_context
*cpuctx
, int cpu
)
1257 struct perf_event
*event
;
1260 raw_spin_lock(&ctx
->lock
);
1262 if (likely(!ctx
->nr_events
))
1265 ctx
->timestamp
= perf_clock();
1270 * First go through the list and put on any pinned groups
1271 * in order to give them the best chance of going on.
1273 list_for_each_entry(event
, &ctx
->group_list
, group_entry
) {
1274 if (event
->state
<= PERF_EVENT_STATE_OFF
||
1275 !event
->attr
.pinned
)
1277 if (event
->cpu
!= -1 && event
->cpu
!= cpu
)
1280 if (group_can_go_on(event
, cpuctx
, 1))
1281 group_sched_in(event
, cpuctx
, ctx
, cpu
);
1284 * If this pinned group hasn't been scheduled,
1285 * put it in error state.
1287 if (event
->state
== PERF_EVENT_STATE_INACTIVE
) {
1288 update_group_times(event
);
1289 event
->state
= PERF_EVENT_STATE_ERROR
;
1293 list_for_each_entry(event
, &ctx
->group_list
, group_entry
) {
1295 * Ignore events in OFF or ERROR state, and
1296 * ignore pinned events since we did them already.
1298 if (event
->state
<= PERF_EVENT_STATE_OFF
||
1303 * Listen to the 'cpu' scheduling filter constraint
1306 if (event
->cpu
!= -1 && event
->cpu
!= cpu
)
1309 if (group_can_go_on(event
, cpuctx
, can_add_hw
))
1310 if (group_sched_in(event
, cpuctx
, ctx
, cpu
))
1315 raw_spin_unlock(&ctx
->lock
);
1319 * Called from scheduler to add the events of the current task
1320 * with interrupts disabled.
1322 * We restore the event value and then enable it.
1324 * This does not protect us against NMI, but enable()
1325 * sets the enabled bit in the control field of event _before_
1326 * accessing the event control register. If a NMI hits, then it will
1327 * keep the event running.
1329 void perf_event_task_sched_in(struct task_struct
*task
, int cpu
)
1331 struct perf_cpu_context
*cpuctx
= &per_cpu(perf_cpu_context
, cpu
);
1332 struct perf_event_context
*ctx
= task
->perf_event_ctxp
;
1336 if (cpuctx
->task_ctx
== ctx
)
1338 __perf_event_sched_in(ctx
, cpuctx
, cpu
);
1339 cpuctx
->task_ctx
= ctx
;
1342 static void perf_event_cpu_sched_in(struct perf_cpu_context
*cpuctx
, int cpu
)
1344 struct perf_event_context
*ctx
= &cpuctx
->ctx
;
1346 __perf_event_sched_in(ctx
, cpuctx
, cpu
);
1349 #define MAX_INTERRUPTS (~0ULL)
1351 static void perf_log_throttle(struct perf_event
*event
, int enable
);
1353 static u64
perf_calculate_period(struct perf_event
*event
, u64 nsec
, u64 count
)
1355 u64 frequency
= event
->attr
.sample_freq
;
1356 u64 sec
= NSEC_PER_SEC
;
1357 u64 divisor
, dividend
;
1359 int count_fls
, nsec_fls
, frequency_fls
, sec_fls
;
1361 count_fls
= fls64(count
);
1362 nsec_fls
= fls64(nsec
);
1363 frequency_fls
= fls64(frequency
);
1367 * We got @count in @nsec, with a target of sample_freq HZ
1368 * the target period becomes:
1371 * period = -------------------
1372 * @nsec * sample_freq
1377 * Reduce accuracy by one bit such that @a and @b converge
1378 * to a similar magnitude.
1380 #define REDUCE_FLS(a, b) \
1382 if (a##_fls > b##_fls) { \
1392 * Reduce accuracy until either term fits in a u64, then proceed with
1393 * the other, so that finally we can do a u64/u64 division.
1395 while (count_fls
+ sec_fls
> 64 && nsec_fls
+ frequency_fls
> 64) {
1396 REDUCE_FLS(nsec
, frequency
);
1397 REDUCE_FLS(sec
, count
);
1400 if (count_fls
+ sec_fls
> 64) {
1401 divisor
= nsec
* frequency
;
1403 while (count_fls
+ sec_fls
> 64) {
1404 REDUCE_FLS(count
, sec
);
1408 dividend
= count
* sec
;
1410 dividend
= count
* sec
;
1412 while (nsec_fls
+ frequency_fls
> 64) {
1413 REDUCE_FLS(nsec
, frequency
);
1417 divisor
= nsec
* frequency
;
1423 return div64_u64(dividend
, divisor
);
1426 static void perf_adjust_period(struct perf_event
*event
, u64 nsec
, u64 count
)
1428 struct hw_perf_event
*hwc
= &event
->hw
;
1429 s64 period
, sample_period
;
1432 period
= perf_calculate_period(event
, nsec
, count
);
1434 delta
= (s64
)(period
- hwc
->sample_period
);
1435 delta
= (delta
+ 7) / 8; /* low pass filter */
1437 sample_period
= hwc
->sample_period
+ delta
;
1442 hwc
->sample_period
= sample_period
;
1444 if (atomic64_read(&hwc
->period_left
) > 8*sample_period
) {
1446 event
->pmu
->disable(event
);
1447 atomic64_set(&hwc
->period_left
, 0);
1448 event
->pmu
->enable(event
);
1453 static void perf_ctx_adjust_freq(struct perf_event_context
*ctx
)
1455 struct perf_event
*event
;
1456 struct hw_perf_event
*hwc
;
1457 u64 interrupts
, now
;
1460 raw_spin_lock(&ctx
->lock
);
1461 list_for_each_entry_rcu(event
, &ctx
->event_list
, event_entry
) {
1462 if (event
->state
!= PERF_EVENT_STATE_ACTIVE
)
1465 if (event
->cpu
!= -1 && event
->cpu
!= smp_processor_id())
1470 interrupts
= hwc
->interrupts
;
1471 hwc
->interrupts
= 0;
1474 * unthrottle events on the tick
1476 if (interrupts
== MAX_INTERRUPTS
) {
1477 perf_log_throttle(event
, 1);
1478 event
->pmu
->unthrottle(event
);
1481 if (!event
->attr
.freq
|| !event
->attr
.sample_freq
)
1484 event
->pmu
->read(event
);
1485 now
= atomic64_read(&event
->count
);
1486 delta
= now
- hwc
->freq_count_stamp
;
1487 hwc
->freq_count_stamp
= now
;
1490 perf_adjust_period(event
, TICK_NSEC
, delta
);
1492 raw_spin_unlock(&ctx
->lock
);
1496 * Round-robin a context's events:
1498 static void rotate_ctx(struct perf_event_context
*ctx
)
1500 struct perf_event
*event
;
1502 if (!ctx
->nr_events
)
1505 raw_spin_lock(&ctx
->lock
);
1507 * Rotate the first entry last (works just fine for group events too):
1510 list_for_each_entry(event
, &ctx
->group_list
, group_entry
) {
1511 list_move_tail(&event
->group_entry
, &ctx
->group_list
);
1516 raw_spin_unlock(&ctx
->lock
);
1519 void perf_event_task_tick(struct task_struct
*curr
, int cpu
)
1521 struct perf_cpu_context
*cpuctx
;
1522 struct perf_event_context
*ctx
;
1524 if (!atomic_read(&nr_events
))
1527 cpuctx
= &per_cpu(perf_cpu_context
, cpu
);
1528 ctx
= curr
->perf_event_ctxp
;
1530 perf_ctx_adjust_freq(&cpuctx
->ctx
);
1532 perf_ctx_adjust_freq(ctx
);
1534 perf_event_cpu_sched_out(cpuctx
);
1536 __perf_event_task_sched_out(ctx
);
1538 rotate_ctx(&cpuctx
->ctx
);
1542 perf_event_cpu_sched_in(cpuctx
, cpu
);
1544 perf_event_task_sched_in(curr
, cpu
);
1548 * Enable all of a task's events that have been marked enable-on-exec.
1549 * This expects task == current.
1551 static void perf_event_enable_on_exec(struct task_struct
*task
)
1553 struct perf_event_context
*ctx
;
1554 struct perf_event
*event
;
1555 unsigned long flags
;
1558 local_irq_save(flags
);
1559 ctx
= task
->perf_event_ctxp
;
1560 if (!ctx
|| !ctx
->nr_events
)
1563 __perf_event_task_sched_out(ctx
);
1565 raw_spin_lock(&ctx
->lock
);
1567 list_for_each_entry(event
, &ctx
->group_list
, group_entry
) {
1568 if (!event
->attr
.enable_on_exec
)
1570 event
->attr
.enable_on_exec
= 0;
1571 if (event
->state
>= PERF_EVENT_STATE_INACTIVE
)
1573 __perf_event_mark_enabled(event
, ctx
);
1578 * Unclone this context if we enabled any event.
1583 raw_spin_unlock(&ctx
->lock
);
1585 perf_event_task_sched_in(task
, smp_processor_id());
1587 local_irq_restore(flags
);
1591 * Cross CPU call to read the hardware event
1593 static void __perf_event_read(void *info
)
1595 struct perf_cpu_context
*cpuctx
= &__get_cpu_var(perf_cpu_context
);
1596 struct perf_event
*event
= info
;
1597 struct perf_event_context
*ctx
= event
->ctx
;
1600 * If this is a task context, we need to check whether it is
1601 * the current task context of this cpu. If not it has been
1602 * scheduled out before the smp call arrived. In that case
1603 * event->count would have been updated to a recent sample
1604 * when the event was scheduled out.
1606 if (ctx
->task
&& cpuctx
->task_ctx
!= ctx
)
1609 raw_spin_lock(&ctx
->lock
);
1610 update_context_time(ctx
);
1611 update_event_times(event
);
1612 raw_spin_unlock(&ctx
->lock
);
1614 event
->pmu
->read(event
);
1617 static u64
perf_event_read(struct perf_event
*event
)
1620 * If event is enabled and currently active on a CPU, update the
1621 * value in the event structure:
1623 if (event
->state
== PERF_EVENT_STATE_ACTIVE
) {
1624 smp_call_function_single(event
->oncpu
,
1625 __perf_event_read
, event
, 1);
1626 } else if (event
->state
== PERF_EVENT_STATE_INACTIVE
) {
1627 struct perf_event_context
*ctx
= event
->ctx
;
1628 unsigned long flags
;
1630 raw_spin_lock_irqsave(&ctx
->lock
, flags
);
1631 update_context_time(ctx
);
1632 update_event_times(event
);
1633 raw_spin_unlock_irqrestore(&ctx
->lock
, flags
);
1636 return atomic64_read(&event
->count
);
1640 * Initialize the perf_event context in a task_struct:
1643 __perf_event_init_context(struct perf_event_context
*ctx
,
1644 struct task_struct
*task
)
1646 raw_spin_lock_init(&ctx
->lock
);
1647 mutex_init(&ctx
->mutex
);
1648 INIT_LIST_HEAD(&ctx
->group_list
);
1649 INIT_LIST_HEAD(&ctx
->event_list
);
1650 atomic_set(&ctx
->refcount
, 1);
1654 static struct perf_event_context
*find_get_context(pid_t pid
, int cpu
)
1656 struct perf_event_context
*ctx
;
1657 struct perf_cpu_context
*cpuctx
;
1658 struct task_struct
*task
;
1659 unsigned long flags
;
1662 if (pid
== -1 && cpu
!= -1) {
1663 /* Must be root to operate on a CPU event: */
1664 if (perf_paranoid_cpu() && !capable(CAP_SYS_ADMIN
))
1665 return ERR_PTR(-EACCES
);
1667 if (cpu
< 0 || cpu
>= nr_cpumask_bits
)
1668 return ERR_PTR(-EINVAL
);
1671 * We could be clever and allow to attach a event to an
1672 * offline CPU and activate it when the CPU comes up, but
1675 if (!cpu_online(cpu
))
1676 return ERR_PTR(-ENODEV
);
1678 cpuctx
= &per_cpu(perf_cpu_context
, cpu
);
1689 task
= find_task_by_vpid(pid
);
1691 get_task_struct(task
);
1695 return ERR_PTR(-ESRCH
);
1698 * Can't attach events to a dying task.
1701 if (task
->flags
& PF_EXITING
)
1704 /* Reuse ptrace permission checks for now. */
1706 if (!ptrace_may_access(task
, PTRACE_MODE_READ
))
1710 ctx
= perf_lock_task_context(task
, &flags
);
1713 raw_spin_unlock_irqrestore(&ctx
->lock
, flags
);
1717 ctx
= kzalloc(sizeof(struct perf_event_context
), GFP_KERNEL
);
1721 __perf_event_init_context(ctx
, task
);
1723 if (cmpxchg(&task
->perf_event_ctxp
, NULL
, ctx
)) {
1725 * We raced with some other task; use
1726 * the context they set.
1731 get_task_struct(task
);
1734 put_task_struct(task
);
1738 put_task_struct(task
);
1739 return ERR_PTR(err
);
1742 static void perf_event_free_filter(struct perf_event
*event
);
1744 static void free_event_rcu(struct rcu_head
*head
)
1746 struct perf_event
*event
;
1748 event
= container_of(head
, struct perf_event
, rcu_head
);
1750 put_pid_ns(event
->ns
);
1751 perf_event_free_filter(event
);
1755 static void perf_pending_sync(struct perf_event
*event
);
1757 static void free_event(struct perf_event
*event
)
1759 perf_pending_sync(event
);
1761 if (!event
->parent
) {
1762 atomic_dec(&nr_events
);
1763 if (event
->attr
.mmap
)
1764 atomic_dec(&nr_mmap_events
);
1765 if (event
->attr
.comm
)
1766 atomic_dec(&nr_comm_events
);
1767 if (event
->attr
.task
)
1768 atomic_dec(&nr_task_events
);
1771 if (event
->output
) {
1772 fput(event
->output
->filp
);
1773 event
->output
= NULL
;
1777 event
->destroy(event
);
1779 put_ctx(event
->ctx
);
1780 call_rcu(&event
->rcu_head
, free_event_rcu
);
1783 int perf_event_release_kernel(struct perf_event
*event
)
1785 struct perf_event_context
*ctx
= event
->ctx
;
1787 WARN_ON_ONCE(ctx
->parent_ctx
);
1788 mutex_lock(&ctx
->mutex
);
1789 perf_event_remove_from_context(event
);
1790 mutex_unlock(&ctx
->mutex
);
1792 mutex_lock(&event
->owner
->perf_event_mutex
);
1793 list_del_init(&event
->owner_entry
);
1794 mutex_unlock(&event
->owner
->perf_event_mutex
);
1795 put_task_struct(event
->owner
);
1801 EXPORT_SYMBOL_GPL(perf_event_release_kernel
);
1804 * Called when the last reference to the file is gone.
1806 static int perf_release(struct inode
*inode
, struct file
*file
)
1808 struct perf_event
*event
= file
->private_data
;
1810 file
->private_data
= NULL
;
1812 return perf_event_release_kernel(event
);
1815 static int perf_event_read_size(struct perf_event
*event
)
1817 int entry
= sizeof(u64
); /* value */
1821 if (event
->attr
.read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
)
1822 size
+= sizeof(u64
);
1824 if (event
->attr
.read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
)
1825 size
+= sizeof(u64
);
1827 if (event
->attr
.read_format
& PERF_FORMAT_ID
)
1828 entry
+= sizeof(u64
);
1830 if (event
->attr
.read_format
& PERF_FORMAT_GROUP
) {
1831 nr
+= event
->group_leader
->nr_siblings
;
1832 size
+= sizeof(u64
);
1840 u64
perf_event_read_value(struct perf_event
*event
, u64
*enabled
, u64
*running
)
1842 struct perf_event
*child
;
1848 mutex_lock(&event
->child_mutex
);
1849 total
+= perf_event_read(event
);
1850 *enabled
+= event
->total_time_enabled
+
1851 atomic64_read(&event
->child_total_time_enabled
);
1852 *running
+= event
->total_time_running
+
1853 atomic64_read(&event
->child_total_time_running
);
1855 list_for_each_entry(child
, &event
->child_list
, child_list
) {
1856 total
+= perf_event_read(child
);
1857 *enabled
+= child
->total_time_enabled
;
1858 *running
+= child
->total_time_running
;
1860 mutex_unlock(&event
->child_mutex
);
1864 EXPORT_SYMBOL_GPL(perf_event_read_value
);
1866 static int perf_event_read_group(struct perf_event
*event
,
1867 u64 read_format
, char __user
*buf
)
1869 struct perf_event
*leader
= event
->group_leader
, *sub
;
1870 int n
= 0, size
= 0, ret
= -EFAULT
;
1871 struct perf_event_context
*ctx
= leader
->ctx
;
1873 u64 count
, enabled
, running
;
1875 mutex_lock(&ctx
->mutex
);
1876 count
= perf_event_read_value(leader
, &enabled
, &running
);
1878 values
[n
++] = 1 + leader
->nr_siblings
;
1879 if (read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
)
1880 values
[n
++] = enabled
;
1881 if (read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
)
1882 values
[n
++] = running
;
1883 values
[n
++] = count
;
1884 if (read_format
& PERF_FORMAT_ID
)
1885 values
[n
++] = primary_event_id(leader
);
1887 size
= n
* sizeof(u64
);
1889 if (copy_to_user(buf
, values
, size
))
1894 list_for_each_entry(sub
, &leader
->sibling_list
, group_entry
) {
1897 values
[n
++] = perf_event_read_value(sub
, &enabled
, &running
);
1898 if (read_format
& PERF_FORMAT_ID
)
1899 values
[n
++] = primary_event_id(sub
);
1901 size
= n
* sizeof(u64
);
1903 if (copy_to_user(buf
+ ret
, values
, size
)) {
1911 mutex_unlock(&ctx
->mutex
);
1916 static int perf_event_read_one(struct perf_event
*event
,
1917 u64 read_format
, char __user
*buf
)
1919 u64 enabled
, running
;
1923 values
[n
++] = perf_event_read_value(event
, &enabled
, &running
);
1924 if (read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
)
1925 values
[n
++] = enabled
;
1926 if (read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
)
1927 values
[n
++] = running
;
1928 if (read_format
& PERF_FORMAT_ID
)
1929 values
[n
++] = primary_event_id(event
);
1931 if (copy_to_user(buf
, values
, n
* sizeof(u64
)))
1934 return n
* sizeof(u64
);
1938 * Read the performance event - simple non blocking version for now
1941 perf_read_hw(struct perf_event
*event
, char __user
*buf
, size_t count
)
1943 u64 read_format
= event
->attr
.read_format
;
1947 * Return end-of-file for a read on a event that is in
1948 * error state (i.e. because it was pinned but it couldn't be
1949 * scheduled on to the CPU at some point).
1951 if (event
->state
== PERF_EVENT_STATE_ERROR
)
1954 if (count
< perf_event_read_size(event
))
1957 WARN_ON_ONCE(event
->ctx
->parent_ctx
);
1958 if (read_format
& PERF_FORMAT_GROUP
)
1959 ret
= perf_event_read_group(event
, read_format
, buf
);
1961 ret
= perf_event_read_one(event
, read_format
, buf
);
1967 perf_read(struct file
*file
, char __user
*buf
, size_t count
, loff_t
*ppos
)
1969 struct perf_event
*event
= file
->private_data
;
1971 return perf_read_hw(event
, buf
, count
);
1974 static unsigned int perf_poll(struct file
*file
, poll_table
*wait
)
1976 struct perf_event
*event
= file
->private_data
;
1977 struct perf_mmap_data
*data
;
1978 unsigned int events
= POLL_HUP
;
1981 data
= rcu_dereference(event
->data
);
1983 events
= atomic_xchg(&data
->poll
, 0);
1986 poll_wait(file
, &event
->waitq
, wait
);
1991 static void perf_event_reset(struct perf_event
*event
)
1993 (void)perf_event_read(event
);
1994 atomic64_set(&event
->count
, 0);
1995 perf_event_update_userpage(event
);
1999 * Holding the top-level event's child_mutex means that any
2000 * descendant process that has inherited this event will block
2001 * in sync_child_event if it goes to exit, thus satisfying the
2002 * task existence requirements of perf_event_enable/disable.
2004 static void perf_event_for_each_child(struct perf_event
*event
,
2005 void (*func
)(struct perf_event
*))
2007 struct perf_event
*child
;
2009 WARN_ON_ONCE(event
->ctx
->parent_ctx
);
2010 mutex_lock(&event
->child_mutex
);
2012 list_for_each_entry(child
, &event
->child_list
, child_list
)
2014 mutex_unlock(&event
->child_mutex
);
2017 static void perf_event_for_each(struct perf_event
*event
,
2018 void (*func
)(struct perf_event
*))
2020 struct perf_event_context
*ctx
= event
->ctx
;
2021 struct perf_event
*sibling
;
2023 WARN_ON_ONCE(ctx
->parent_ctx
);
2024 mutex_lock(&ctx
->mutex
);
2025 event
= event
->group_leader
;
2027 perf_event_for_each_child(event
, func
);
2029 list_for_each_entry(sibling
, &event
->sibling_list
, group_entry
)
2030 perf_event_for_each_child(event
, func
);
2031 mutex_unlock(&ctx
->mutex
);
2034 static int perf_event_period(struct perf_event
*event
, u64 __user
*arg
)
2036 struct perf_event_context
*ctx
= event
->ctx
;
2041 if (!event
->attr
.sample_period
)
2044 size
= copy_from_user(&value
, arg
, sizeof(value
));
2045 if (size
!= sizeof(value
))
2051 raw_spin_lock_irq(&ctx
->lock
);
2052 if (event
->attr
.freq
) {
2053 if (value
> sysctl_perf_event_sample_rate
) {
2058 event
->attr
.sample_freq
= value
;
2060 event
->attr
.sample_period
= value
;
2061 event
->hw
.sample_period
= value
;
2064 raw_spin_unlock_irq(&ctx
->lock
);
2069 static int perf_event_set_output(struct perf_event
*event
, int output_fd
);
2070 static int perf_event_set_filter(struct perf_event
*event
, void __user
*arg
);
2072 static long perf_ioctl(struct file
*file
, unsigned int cmd
, unsigned long arg
)
2074 struct perf_event
*event
= file
->private_data
;
2075 void (*func
)(struct perf_event
*);
2079 case PERF_EVENT_IOC_ENABLE
:
2080 func
= perf_event_enable
;
2082 case PERF_EVENT_IOC_DISABLE
:
2083 func
= perf_event_disable
;
2085 case PERF_EVENT_IOC_RESET
:
2086 func
= perf_event_reset
;
2089 case PERF_EVENT_IOC_REFRESH
:
2090 return perf_event_refresh(event
, arg
);
2092 case PERF_EVENT_IOC_PERIOD
:
2093 return perf_event_period(event
, (u64 __user
*)arg
);
2095 case PERF_EVENT_IOC_SET_OUTPUT
:
2096 return perf_event_set_output(event
, arg
);
2098 case PERF_EVENT_IOC_SET_FILTER
:
2099 return perf_event_set_filter(event
, (void __user
*)arg
);
2105 if (flags
& PERF_IOC_FLAG_GROUP
)
2106 perf_event_for_each(event
, func
);
2108 perf_event_for_each_child(event
, func
);
2113 int perf_event_task_enable(void)
2115 struct perf_event
*event
;
2117 mutex_lock(¤t
->perf_event_mutex
);
2118 list_for_each_entry(event
, ¤t
->perf_event_list
, owner_entry
)
2119 perf_event_for_each_child(event
, perf_event_enable
);
2120 mutex_unlock(¤t
->perf_event_mutex
);
2125 int perf_event_task_disable(void)
2127 struct perf_event
*event
;
2129 mutex_lock(¤t
->perf_event_mutex
);
2130 list_for_each_entry(event
, ¤t
->perf_event_list
, owner_entry
)
2131 perf_event_for_each_child(event
, perf_event_disable
);
2132 mutex_unlock(¤t
->perf_event_mutex
);
2137 #ifndef PERF_EVENT_INDEX_OFFSET
2138 # define PERF_EVENT_INDEX_OFFSET 0
2141 static int perf_event_index(struct perf_event
*event
)
2143 if (event
->state
!= PERF_EVENT_STATE_ACTIVE
)
2146 return event
->hw
.idx
+ 1 - PERF_EVENT_INDEX_OFFSET
;
2150 * Callers need to ensure there can be no nesting of this function, otherwise
2151 * the seqlock logic goes bad. We can not serialize this because the arch
2152 * code calls this from NMI context.
2154 void perf_event_update_userpage(struct perf_event
*event
)
2156 struct perf_event_mmap_page
*userpg
;
2157 struct perf_mmap_data
*data
;
2160 data
= rcu_dereference(event
->data
);
2164 userpg
= data
->user_page
;
2167 * Disable preemption so as to not let the corresponding user-space
2168 * spin too long if we get preempted.
2173 userpg
->index
= perf_event_index(event
);
2174 userpg
->offset
= atomic64_read(&event
->count
);
2175 if (event
->state
== PERF_EVENT_STATE_ACTIVE
)
2176 userpg
->offset
-= atomic64_read(&event
->hw
.prev_count
);
2178 userpg
->time_enabled
= event
->total_time_enabled
+
2179 atomic64_read(&event
->child_total_time_enabled
);
2181 userpg
->time_running
= event
->total_time_running
+
2182 atomic64_read(&event
->child_total_time_running
);
2191 static unsigned long perf_data_size(struct perf_mmap_data
*data
)
2193 return data
->nr_pages
<< (PAGE_SHIFT
+ data
->data_order
);
2196 #ifndef CONFIG_PERF_USE_VMALLOC
2199 * Back perf_mmap() with regular GFP_KERNEL-0 pages.
2202 static struct page
*
2203 perf_mmap_to_page(struct perf_mmap_data
*data
, unsigned long pgoff
)
2205 if (pgoff
> data
->nr_pages
)
2209 return virt_to_page(data
->user_page
);
2211 return virt_to_page(data
->data_pages
[pgoff
- 1]);
2214 static struct perf_mmap_data
*
2215 perf_mmap_data_alloc(struct perf_event
*event
, int nr_pages
)
2217 struct perf_mmap_data
*data
;
2221 WARN_ON(atomic_read(&event
->mmap_count
));
2223 size
= sizeof(struct perf_mmap_data
);
2224 size
+= nr_pages
* sizeof(void *);
2226 data
= kzalloc(size
, GFP_KERNEL
);
2230 data
->user_page
= (void *)get_zeroed_page(GFP_KERNEL
);
2231 if (!data
->user_page
)
2232 goto fail_user_page
;
2234 for (i
= 0; i
< nr_pages
; i
++) {
2235 data
->data_pages
[i
] = (void *)get_zeroed_page(GFP_KERNEL
);
2236 if (!data
->data_pages
[i
])
2237 goto fail_data_pages
;
2240 data
->data_order
= 0;
2241 data
->nr_pages
= nr_pages
;
2246 for (i
--; i
>= 0; i
--)
2247 free_page((unsigned long)data
->data_pages
[i
]);
2249 free_page((unsigned long)data
->user_page
);
2258 static void perf_mmap_free_page(unsigned long addr
)
2260 struct page
*page
= virt_to_page((void *)addr
);
2262 page
->mapping
= NULL
;
2266 static void perf_mmap_data_free(struct perf_mmap_data
*data
)
2270 perf_mmap_free_page((unsigned long)data
->user_page
);
2271 for (i
= 0; i
< data
->nr_pages
; i
++)
2272 perf_mmap_free_page((unsigned long)data
->data_pages
[i
]);
2279 * Back perf_mmap() with vmalloc memory.
2281 * Required for architectures that have d-cache aliasing issues.
2284 static struct page
*
2285 perf_mmap_to_page(struct perf_mmap_data
*data
, unsigned long pgoff
)
2287 if (pgoff
> (1UL << data
->data_order
))
2290 return vmalloc_to_page((void *)data
->user_page
+ pgoff
* PAGE_SIZE
);
2293 static void perf_mmap_unmark_page(void *addr
)
2295 struct page
*page
= vmalloc_to_page(addr
);
2297 page
->mapping
= NULL
;
2300 static void perf_mmap_data_free_work(struct work_struct
*work
)
2302 struct perf_mmap_data
*data
;
2306 data
= container_of(work
, struct perf_mmap_data
, work
);
2307 nr
= 1 << data
->data_order
;
2309 base
= data
->user_page
;
2310 for (i
= 0; i
< nr
+ 1; i
++)
2311 perf_mmap_unmark_page(base
+ (i
* PAGE_SIZE
));
2317 static void perf_mmap_data_free(struct perf_mmap_data
*data
)
2319 schedule_work(&data
->work
);
2322 static struct perf_mmap_data
*
2323 perf_mmap_data_alloc(struct perf_event
*event
, int nr_pages
)
2325 struct perf_mmap_data
*data
;
2329 WARN_ON(atomic_read(&event
->mmap_count
));
2331 size
= sizeof(struct perf_mmap_data
);
2332 size
+= sizeof(void *);
2334 data
= kzalloc(size
, GFP_KERNEL
);
2338 INIT_WORK(&data
->work
, perf_mmap_data_free_work
);
2340 all_buf
= vmalloc_user((nr_pages
+ 1) * PAGE_SIZE
);
2344 data
->user_page
= all_buf
;
2345 data
->data_pages
[0] = all_buf
+ PAGE_SIZE
;
2346 data
->data_order
= ilog2(nr_pages
);
2360 static int perf_mmap_fault(struct vm_area_struct
*vma
, struct vm_fault
*vmf
)
2362 struct perf_event
*event
= vma
->vm_file
->private_data
;
2363 struct perf_mmap_data
*data
;
2364 int ret
= VM_FAULT_SIGBUS
;
2366 if (vmf
->flags
& FAULT_FLAG_MKWRITE
) {
2367 if (vmf
->pgoff
== 0)
2373 data
= rcu_dereference(event
->data
);
2377 if (vmf
->pgoff
&& (vmf
->flags
& FAULT_FLAG_WRITE
))
2380 vmf
->page
= perf_mmap_to_page(data
, vmf
->pgoff
);
2384 get_page(vmf
->page
);
2385 vmf
->page
->mapping
= vma
->vm_file
->f_mapping
;
2386 vmf
->page
->index
= vmf
->pgoff
;
2396 perf_mmap_data_init(struct perf_event
*event
, struct perf_mmap_data
*data
)
2398 long max_size
= perf_data_size(data
);
2400 atomic_set(&data
->lock
, -1);
2402 if (event
->attr
.watermark
) {
2403 data
->watermark
= min_t(long, max_size
,
2404 event
->attr
.wakeup_watermark
);
2407 if (!data
->watermark
)
2408 data
->watermark
= max_size
/ 2;
2411 rcu_assign_pointer(event
->data
, data
);
2414 static void perf_mmap_data_free_rcu(struct rcu_head
*rcu_head
)
2416 struct perf_mmap_data
*data
;
2418 data
= container_of(rcu_head
, struct perf_mmap_data
, rcu_head
);
2419 perf_mmap_data_free(data
);
2422 static void perf_mmap_data_release(struct perf_event
*event
)
2424 struct perf_mmap_data
*data
= event
->data
;
2426 WARN_ON(atomic_read(&event
->mmap_count
));
2428 rcu_assign_pointer(event
->data
, NULL
);
2429 call_rcu(&data
->rcu_head
, perf_mmap_data_free_rcu
);
2432 static void perf_mmap_open(struct vm_area_struct
*vma
)
2434 struct perf_event
*event
= vma
->vm_file
->private_data
;
2436 atomic_inc(&event
->mmap_count
);
2439 static void perf_mmap_close(struct vm_area_struct
*vma
)
2441 struct perf_event
*event
= vma
->vm_file
->private_data
;
2443 WARN_ON_ONCE(event
->ctx
->parent_ctx
);
2444 if (atomic_dec_and_mutex_lock(&event
->mmap_count
, &event
->mmap_mutex
)) {
2445 unsigned long size
= perf_data_size(event
->data
);
2446 struct user_struct
*user
= current_user();
2448 atomic_long_sub((size
>> PAGE_SHIFT
) + 1, &user
->locked_vm
);
2449 vma
->vm_mm
->locked_vm
-= event
->data
->nr_locked
;
2450 perf_mmap_data_release(event
);
2451 mutex_unlock(&event
->mmap_mutex
);
2455 static const struct vm_operations_struct perf_mmap_vmops
= {
2456 .open
= perf_mmap_open
,
2457 .close
= perf_mmap_close
,
2458 .fault
= perf_mmap_fault
,
2459 .page_mkwrite
= perf_mmap_fault
,
2462 static int perf_mmap(struct file
*file
, struct vm_area_struct
*vma
)
2464 struct perf_event
*event
= file
->private_data
;
2465 unsigned long user_locked
, user_lock_limit
;
2466 struct user_struct
*user
= current_user();
2467 unsigned long locked
, lock_limit
;
2468 struct perf_mmap_data
*data
;
2469 unsigned long vma_size
;
2470 unsigned long nr_pages
;
2471 long user_extra
, extra
;
2474 if (!(vma
->vm_flags
& VM_SHARED
))
2477 vma_size
= vma
->vm_end
- vma
->vm_start
;
2478 nr_pages
= (vma_size
/ PAGE_SIZE
) - 1;
2481 * If we have data pages ensure they're a power-of-two number, so we
2482 * can do bitmasks instead of modulo.
2484 if (nr_pages
!= 0 && !is_power_of_2(nr_pages
))
2487 if (vma_size
!= PAGE_SIZE
* (1 + nr_pages
))
2490 if (vma
->vm_pgoff
!= 0)
2493 WARN_ON_ONCE(event
->ctx
->parent_ctx
);
2494 mutex_lock(&event
->mmap_mutex
);
2495 if (event
->output
) {
2500 if (atomic_inc_not_zero(&event
->mmap_count
)) {
2501 if (nr_pages
!= event
->data
->nr_pages
)
2506 user_extra
= nr_pages
+ 1;
2507 user_lock_limit
= sysctl_perf_event_mlock
>> (PAGE_SHIFT
- 10);
2510 * Increase the limit linearly with more CPUs:
2512 user_lock_limit
*= num_online_cpus();
2514 user_locked
= atomic_long_read(&user
->locked_vm
) + user_extra
;
2517 if (user_locked
> user_lock_limit
)
2518 extra
= user_locked
- user_lock_limit
;
2520 lock_limit
= current
->signal
->rlim
[RLIMIT_MEMLOCK
].rlim_cur
;
2521 lock_limit
>>= PAGE_SHIFT
;
2522 locked
= vma
->vm_mm
->locked_vm
+ extra
;
2524 if ((locked
> lock_limit
) && perf_paranoid_tracepoint_raw() &&
2525 !capable(CAP_IPC_LOCK
)) {
2530 WARN_ON(event
->data
);
2532 data
= perf_mmap_data_alloc(event
, nr_pages
);
2538 perf_mmap_data_init(event
, data
);
2540 atomic_set(&event
->mmap_count
, 1);
2541 atomic_long_add(user_extra
, &user
->locked_vm
);
2542 vma
->vm_mm
->locked_vm
+= extra
;
2543 event
->data
->nr_locked
= extra
;
2544 if (vma
->vm_flags
& VM_WRITE
)
2545 event
->data
->writable
= 1;
2548 mutex_unlock(&event
->mmap_mutex
);
2550 vma
->vm_flags
|= VM_RESERVED
;
2551 vma
->vm_ops
= &perf_mmap_vmops
;
2556 static int perf_fasync(int fd
, struct file
*filp
, int on
)
2558 struct inode
*inode
= filp
->f_path
.dentry
->d_inode
;
2559 struct perf_event
*event
= filp
->private_data
;
2562 mutex_lock(&inode
->i_mutex
);
2563 retval
= fasync_helper(fd
, filp
, on
, &event
->fasync
);
2564 mutex_unlock(&inode
->i_mutex
);
2572 static const struct file_operations perf_fops
= {
2573 .release
= perf_release
,
2576 .unlocked_ioctl
= perf_ioctl
,
2577 .compat_ioctl
= perf_ioctl
,
2579 .fasync
= perf_fasync
,
2585 * If there's data, ensure we set the poll() state and publish everything
2586 * to user-space before waking everybody up.
2589 void perf_event_wakeup(struct perf_event
*event
)
2591 wake_up_all(&event
->waitq
);
2593 if (event
->pending_kill
) {
2594 kill_fasync(&event
->fasync
, SIGIO
, event
->pending_kill
);
2595 event
->pending_kill
= 0;
2602 * Handle the case where we need to wakeup up from NMI (or rq->lock) context.
2604 * The NMI bit means we cannot possibly take locks. Therefore, maintain a
2605 * single linked list and use cmpxchg() to add entries lockless.
2608 static void perf_pending_event(struct perf_pending_entry
*entry
)
2610 struct perf_event
*event
= container_of(entry
,
2611 struct perf_event
, pending
);
2613 if (event
->pending_disable
) {
2614 event
->pending_disable
= 0;
2615 __perf_event_disable(event
);
2618 if (event
->pending_wakeup
) {
2619 event
->pending_wakeup
= 0;
2620 perf_event_wakeup(event
);
2624 #define PENDING_TAIL ((struct perf_pending_entry *)-1UL)
2626 static DEFINE_PER_CPU(struct perf_pending_entry
*, perf_pending_head
) = {
2630 static void perf_pending_queue(struct perf_pending_entry
*entry
,
2631 void (*func
)(struct perf_pending_entry
*))
2633 struct perf_pending_entry
**head
;
2635 if (cmpxchg(&entry
->next
, NULL
, PENDING_TAIL
) != NULL
)
2640 head
= &get_cpu_var(perf_pending_head
);
2643 entry
->next
= *head
;
2644 } while (cmpxchg(head
, entry
->next
, entry
) != entry
->next
);
2646 set_perf_event_pending();
2648 put_cpu_var(perf_pending_head
);
2651 static int __perf_pending_run(void)
2653 struct perf_pending_entry
*list
;
2656 list
= xchg(&__get_cpu_var(perf_pending_head
), PENDING_TAIL
);
2657 while (list
!= PENDING_TAIL
) {
2658 void (*func
)(struct perf_pending_entry
*);
2659 struct perf_pending_entry
*entry
= list
;
2666 * Ensure we observe the unqueue before we issue the wakeup,
2667 * so that we won't be waiting forever.
2668 * -- see perf_not_pending().
2679 static inline int perf_not_pending(struct perf_event
*event
)
2682 * If we flush on whatever cpu we run, there is a chance we don't
2686 __perf_pending_run();
2690 * Ensure we see the proper queue state before going to sleep
2691 * so that we do not miss the wakeup. -- see perf_pending_handle()
2694 return event
->pending
.next
== NULL
;
2697 static void perf_pending_sync(struct perf_event
*event
)
2699 wait_event(event
->waitq
, perf_not_pending(event
));
2702 void perf_event_do_pending(void)
2704 __perf_pending_run();
2708 * Callchain support -- arch specific
2711 __weak
struct perf_callchain_entry
*perf_callchain(struct pt_regs
*regs
)
2719 static bool perf_output_space(struct perf_mmap_data
*data
, unsigned long tail
,
2720 unsigned long offset
, unsigned long head
)
2724 if (!data
->writable
)
2727 mask
= perf_data_size(data
) - 1;
2729 offset
= (offset
- tail
) & mask
;
2730 head
= (head
- tail
) & mask
;
2732 if ((int)(head
- offset
) < 0)
2738 static void perf_output_wakeup(struct perf_output_handle
*handle
)
2740 atomic_set(&handle
->data
->poll
, POLL_IN
);
2743 handle
->event
->pending_wakeup
= 1;
2744 perf_pending_queue(&handle
->event
->pending
,
2745 perf_pending_event
);
2747 perf_event_wakeup(handle
->event
);
2751 * Curious locking construct.
2753 * We need to ensure a later event_id doesn't publish a head when a former
2754 * event_id isn't done writing. However since we need to deal with NMIs we
2755 * cannot fully serialize things.
2757 * What we do is serialize between CPUs so we only have to deal with NMI
2758 * nesting on a single CPU.
2760 * We only publish the head (and generate a wakeup) when the outer-most
2761 * event_id completes.
2763 static void perf_output_lock(struct perf_output_handle
*handle
)
2765 struct perf_mmap_data
*data
= handle
->data
;
2766 int cur
, cpu
= get_cpu();
2771 cur
= atomic_cmpxchg(&data
->lock
, -1, cpu
);
2783 static void perf_output_unlock(struct perf_output_handle
*handle
)
2785 struct perf_mmap_data
*data
= handle
->data
;
2789 data
->done_head
= data
->head
;
2791 if (!handle
->locked
)
2796 * The xchg implies a full barrier that ensures all writes are done
2797 * before we publish the new head, matched by a rmb() in userspace when
2798 * reading this position.
2800 while ((head
= atomic_long_xchg(&data
->done_head
, 0)))
2801 data
->user_page
->data_head
= head
;
2804 * NMI can happen here, which means we can miss a done_head update.
2807 cpu
= atomic_xchg(&data
->lock
, -1);
2808 WARN_ON_ONCE(cpu
!= smp_processor_id());
2811 * Therefore we have to validate we did not indeed do so.
2813 if (unlikely(atomic_long_read(&data
->done_head
))) {
2815 * Since we had it locked, we can lock it again.
2817 while (atomic_cmpxchg(&data
->lock
, -1, cpu
) != -1)
2823 if (atomic_xchg(&data
->wakeup
, 0))
2824 perf_output_wakeup(handle
);
2829 void perf_output_copy(struct perf_output_handle
*handle
,
2830 const void *buf
, unsigned int len
)
2832 unsigned int pages_mask
;
2833 unsigned long offset
;
2837 offset
= handle
->offset
;
2838 pages_mask
= handle
->data
->nr_pages
- 1;
2839 pages
= handle
->data
->data_pages
;
2842 unsigned long page_offset
;
2843 unsigned long page_size
;
2846 nr
= (offset
>> PAGE_SHIFT
) & pages_mask
;
2847 page_size
= 1UL << (handle
->data
->data_order
+ PAGE_SHIFT
);
2848 page_offset
= offset
& (page_size
- 1);
2849 size
= min_t(unsigned int, page_size
- page_offset
, len
);
2851 memcpy(pages
[nr
] + page_offset
, buf
, size
);
2858 handle
->offset
= offset
;
2861 * Check we didn't copy past our reservation window, taking the
2862 * possible unsigned int wrap into account.
2864 WARN_ON_ONCE(((long)(handle
->head
- handle
->offset
)) < 0);
2867 int perf_output_begin(struct perf_output_handle
*handle
,
2868 struct perf_event
*event
, unsigned int size
,
2869 int nmi
, int sample
)
2871 struct perf_event
*output_event
;
2872 struct perf_mmap_data
*data
;
2873 unsigned long tail
, offset
, head
;
2876 struct perf_event_header header
;
2883 * For inherited events we send all the output towards the parent.
2886 event
= event
->parent
;
2888 output_event
= rcu_dereference(event
->output
);
2890 event
= output_event
;
2892 data
= rcu_dereference(event
->data
);
2896 handle
->data
= data
;
2897 handle
->event
= event
;
2899 handle
->sample
= sample
;
2901 if (!data
->nr_pages
)
2904 have_lost
= atomic_read(&data
->lost
);
2906 size
+= sizeof(lost_event
);
2908 perf_output_lock(handle
);
2912 * Userspace could choose to issue a mb() before updating the
2913 * tail pointer. So that all reads will be completed before the
2916 tail
= ACCESS_ONCE(data
->user_page
->data_tail
);
2918 offset
= head
= atomic_long_read(&data
->head
);
2920 if (unlikely(!perf_output_space(data
, tail
, offset
, head
)))
2922 } while (atomic_long_cmpxchg(&data
->head
, offset
, head
) != offset
);
2924 handle
->offset
= offset
;
2925 handle
->head
= head
;
2927 if (head
- tail
> data
->watermark
)
2928 atomic_set(&data
->wakeup
, 1);
2931 lost_event
.header
.type
= PERF_RECORD_LOST
;
2932 lost_event
.header
.misc
= 0;
2933 lost_event
.header
.size
= sizeof(lost_event
);
2934 lost_event
.id
= event
->id
;
2935 lost_event
.lost
= atomic_xchg(&data
->lost
, 0);
2937 perf_output_put(handle
, lost_event
);
2943 atomic_inc(&data
->lost
);
2944 perf_output_unlock(handle
);
2951 void perf_output_end(struct perf_output_handle
*handle
)
2953 struct perf_event
*event
= handle
->event
;
2954 struct perf_mmap_data
*data
= handle
->data
;
2956 int wakeup_events
= event
->attr
.wakeup_events
;
2958 if (handle
->sample
&& wakeup_events
) {
2959 int events
= atomic_inc_return(&data
->events
);
2960 if (events
>= wakeup_events
) {
2961 atomic_sub(wakeup_events
, &data
->events
);
2962 atomic_set(&data
->wakeup
, 1);
2966 perf_output_unlock(handle
);
2970 static u32
perf_event_pid(struct perf_event
*event
, struct task_struct
*p
)
2973 * only top level events have the pid namespace they were created in
2976 event
= event
->parent
;
2978 return task_tgid_nr_ns(p
, event
->ns
);
2981 static u32
perf_event_tid(struct perf_event
*event
, struct task_struct
*p
)
2984 * only top level events have the pid namespace they were created in
2987 event
= event
->parent
;
2989 return task_pid_nr_ns(p
, event
->ns
);
2992 static void perf_output_read_one(struct perf_output_handle
*handle
,
2993 struct perf_event
*event
)
2995 u64 read_format
= event
->attr
.read_format
;
2999 values
[n
++] = atomic64_read(&event
->count
);
3000 if (read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
) {
3001 values
[n
++] = event
->total_time_enabled
+
3002 atomic64_read(&event
->child_total_time_enabled
);
3004 if (read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
) {
3005 values
[n
++] = event
->total_time_running
+
3006 atomic64_read(&event
->child_total_time_running
);
3008 if (read_format
& PERF_FORMAT_ID
)
3009 values
[n
++] = primary_event_id(event
);
3011 perf_output_copy(handle
, values
, n
* sizeof(u64
));
3015 * XXX PERF_FORMAT_GROUP vs inherited events seems difficult.
3017 static void perf_output_read_group(struct perf_output_handle
*handle
,
3018 struct perf_event
*event
)
3020 struct perf_event
*leader
= event
->group_leader
, *sub
;
3021 u64 read_format
= event
->attr
.read_format
;
3025 values
[n
++] = 1 + leader
->nr_siblings
;
3027 if (read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
)
3028 values
[n
++] = leader
->total_time_enabled
;
3030 if (read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
)
3031 values
[n
++] = leader
->total_time_running
;
3033 if (leader
!= event
)
3034 leader
->pmu
->read(leader
);
3036 values
[n
++] = atomic64_read(&leader
->count
);
3037 if (read_format
& PERF_FORMAT_ID
)
3038 values
[n
++] = primary_event_id(leader
);
3040 perf_output_copy(handle
, values
, n
* sizeof(u64
));
3042 list_for_each_entry(sub
, &leader
->sibling_list
, group_entry
) {
3046 sub
->pmu
->read(sub
);
3048 values
[n
++] = atomic64_read(&sub
->count
);
3049 if (read_format
& PERF_FORMAT_ID
)
3050 values
[n
++] = primary_event_id(sub
);
3052 perf_output_copy(handle
, values
, n
* sizeof(u64
));
3056 static void perf_output_read(struct perf_output_handle
*handle
,
3057 struct perf_event
*event
)
3059 if (event
->attr
.read_format
& PERF_FORMAT_GROUP
)
3060 perf_output_read_group(handle
, event
);
3062 perf_output_read_one(handle
, event
);
3065 void perf_output_sample(struct perf_output_handle
*handle
,
3066 struct perf_event_header
*header
,
3067 struct perf_sample_data
*data
,
3068 struct perf_event
*event
)
3070 u64 sample_type
= data
->type
;
3072 perf_output_put(handle
, *header
);
3074 if (sample_type
& PERF_SAMPLE_IP
)
3075 perf_output_put(handle
, data
->ip
);
3077 if (sample_type
& PERF_SAMPLE_TID
)
3078 perf_output_put(handle
, data
->tid_entry
);
3080 if (sample_type
& PERF_SAMPLE_TIME
)
3081 perf_output_put(handle
, data
->time
);
3083 if (sample_type
& PERF_SAMPLE_ADDR
)
3084 perf_output_put(handle
, data
->addr
);
3086 if (sample_type
& PERF_SAMPLE_ID
)
3087 perf_output_put(handle
, data
->id
);
3089 if (sample_type
& PERF_SAMPLE_STREAM_ID
)
3090 perf_output_put(handle
, data
->stream_id
);
3092 if (sample_type
& PERF_SAMPLE_CPU
)
3093 perf_output_put(handle
, data
->cpu_entry
);
3095 if (sample_type
& PERF_SAMPLE_PERIOD
)
3096 perf_output_put(handle
, data
->period
);
3098 if (sample_type
& PERF_SAMPLE_READ
)
3099 perf_output_read(handle
, event
);
3101 if (sample_type
& PERF_SAMPLE_CALLCHAIN
) {
3102 if (data
->callchain
) {
3105 if (data
->callchain
)
3106 size
+= data
->callchain
->nr
;
3108 size
*= sizeof(u64
);
3110 perf_output_copy(handle
, data
->callchain
, size
);
3113 perf_output_put(handle
, nr
);
3117 if (sample_type
& PERF_SAMPLE_RAW
) {
3119 perf_output_put(handle
, data
->raw
->size
);
3120 perf_output_copy(handle
, data
->raw
->data
,
3127 .size
= sizeof(u32
),
3130 perf_output_put(handle
, raw
);
3135 void perf_prepare_sample(struct perf_event_header
*header
,
3136 struct perf_sample_data
*data
,
3137 struct perf_event
*event
,
3138 struct pt_regs
*regs
)
3140 u64 sample_type
= event
->attr
.sample_type
;
3142 data
->type
= sample_type
;
3144 header
->type
= PERF_RECORD_SAMPLE
;
3145 header
->size
= sizeof(*header
);
3148 header
->misc
|= perf_misc_flags(regs
);
3150 if (sample_type
& PERF_SAMPLE_IP
) {
3151 data
->ip
= perf_instruction_pointer(regs
);
3153 header
->size
+= sizeof(data
->ip
);
3156 if (sample_type
& PERF_SAMPLE_TID
) {
3157 /* namespace issues */
3158 data
->tid_entry
.pid
= perf_event_pid(event
, current
);
3159 data
->tid_entry
.tid
= perf_event_tid(event
, current
);
3161 header
->size
+= sizeof(data
->tid_entry
);
3164 if (sample_type
& PERF_SAMPLE_TIME
) {
3165 data
->time
= perf_clock();
3167 header
->size
+= sizeof(data
->time
);
3170 if (sample_type
& PERF_SAMPLE_ADDR
)
3171 header
->size
+= sizeof(data
->addr
);
3173 if (sample_type
& PERF_SAMPLE_ID
) {
3174 data
->id
= primary_event_id(event
);
3176 header
->size
+= sizeof(data
->id
);
3179 if (sample_type
& PERF_SAMPLE_STREAM_ID
) {
3180 data
->stream_id
= event
->id
;
3182 header
->size
+= sizeof(data
->stream_id
);
3185 if (sample_type
& PERF_SAMPLE_CPU
) {
3186 data
->cpu_entry
.cpu
= raw_smp_processor_id();
3187 data
->cpu_entry
.reserved
= 0;
3189 header
->size
+= sizeof(data
->cpu_entry
);
3192 if (sample_type
& PERF_SAMPLE_PERIOD
)
3193 header
->size
+= sizeof(data
->period
);
3195 if (sample_type
& PERF_SAMPLE_READ
)
3196 header
->size
+= perf_event_read_size(event
);
3198 if (sample_type
& PERF_SAMPLE_CALLCHAIN
) {
3201 data
->callchain
= perf_callchain(regs
);
3203 if (data
->callchain
)
3204 size
+= data
->callchain
->nr
;
3206 header
->size
+= size
* sizeof(u64
);
3209 if (sample_type
& PERF_SAMPLE_RAW
) {
3210 int size
= sizeof(u32
);
3213 size
+= data
->raw
->size
;
3215 size
+= sizeof(u32
);
3217 WARN_ON_ONCE(size
& (sizeof(u64
)-1));
3218 header
->size
+= size
;
3222 static void perf_event_output(struct perf_event
*event
, int nmi
,
3223 struct perf_sample_data
*data
,
3224 struct pt_regs
*regs
)
3226 struct perf_output_handle handle
;
3227 struct perf_event_header header
;
3229 perf_prepare_sample(&header
, data
, event
, regs
);
3231 if (perf_output_begin(&handle
, event
, header
.size
, nmi
, 1))
3234 perf_output_sample(&handle
, &header
, data
, event
);
3236 perf_output_end(&handle
);
3243 struct perf_read_event
{
3244 struct perf_event_header header
;
3251 perf_event_read_event(struct perf_event
*event
,
3252 struct task_struct
*task
)
3254 struct perf_output_handle handle
;
3255 struct perf_read_event read_event
= {
3257 .type
= PERF_RECORD_READ
,
3259 .size
= sizeof(read_event
) + perf_event_read_size(event
),
3261 .pid
= perf_event_pid(event
, task
),
3262 .tid
= perf_event_tid(event
, task
),
3266 ret
= perf_output_begin(&handle
, event
, read_event
.header
.size
, 0, 0);
3270 perf_output_put(&handle
, read_event
);
3271 perf_output_read(&handle
, event
);
3273 perf_output_end(&handle
);
3277 * task tracking -- fork/exit
3279 * enabled by: attr.comm | attr.mmap | attr.task
3282 struct perf_task_event
{
3283 struct task_struct
*task
;
3284 struct perf_event_context
*task_ctx
;
3287 struct perf_event_header header
;
3297 static void perf_event_task_output(struct perf_event
*event
,
3298 struct perf_task_event
*task_event
)
3300 struct perf_output_handle handle
;
3302 struct task_struct
*task
= task_event
->task
;
3305 size
= task_event
->event_id
.header
.size
;
3306 ret
= perf_output_begin(&handle
, event
, size
, 0, 0);
3311 task_event
->event_id
.pid
= perf_event_pid(event
, task
);
3312 task_event
->event_id
.ppid
= perf_event_pid(event
, current
);
3314 task_event
->event_id
.tid
= perf_event_tid(event
, task
);
3315 task_event
->event_id
.ptid
= perf_event_tid(event
, current
);
3317 perf_output_put(&handle
, task_event
->event_id
);
3319 perf_output_end(&handle
);
3322 static int perf_event_task_match(struct perf_event
*event
)
3324 if (event
->state
< PERF_EVENT_STATE_INACTIVE
)
3327 if (event
->cpu
!= -1 && event
->cpu
!= smp_processor_id())
3330 if (event
->attr
.comm
|| event
->attr
.mmap
|| event
->attr
.task
)
3336 static void perf_event_task_ctx(struct perf_event_context
*ctx
,
3337 struct perf_task_event
*task_event
)
3339 struct perf_event
*event
;
3341 list_for_each_entry_rcu(event
, &ctx
->event_list
, event_entry
) {
3342 if (perf_event_task_match(event
))
3343 perf_event_task_output(event
, task_event
);
3347 static void perf_event_task_event(struct perf_task_event
*task_event
)
3349 struct perf_cpu_context
*cpuctx
;
3350 struct perf_event_context
*ctx
= task_event
->task_ctx
;
3353 cpuctx
= &get_cpu_var(perf_cpu_context
);
3354 perf_event_task_ctx(&cpuctx
->ctx
, task_event
);
3356 ctx
= rcu_dereference(current
->perf_event_ctxp
);
3358 perf_event_task_ctx(ctx
, task_event
);
3359 put_cpu_var(perf_cpu_context
);
3363 static void perf_event_task(struct task_struct
*task
,
3364 struct perf_event_context
*task_ctx
,
3367 struct perf_task_event task_event
;
3369 if (!atomic_read(&nr_comm_events
) &&
3370 !atomic_read(&nr_mmap_events
) &&
3371 !atomic_read(&nr_task_events
))
3374 task_event
= (struct perf_task_event
){
3376 .task_ctx
= task_ctx
,
3379 .type
= new ? PERF_RECORD_FORK
: PERF_RECORD_EXIT
,
3381 .size
= sizeof(task_event
.event_id
),
3387 .time
= perf_clock(),
3391 perf_event_task_event(&task_event
);
3394 void perf_event_fork(struct task_struct
*task
)
3396 perf_event_task(task
, NULL
, 1);
3403 struct perf_comm_event
{
3404 struct task_struct
*task
;
3409 struct perf_event_header header
;
3416 static void perf_event_comm_output(struct perf_event
*event
,
3417 struct perf_comm_event
*comm_event
)
3419 struct perf_output_handle handle
;
3420 int size
= comm_event
->event_id
.header
.size
;
3421 int ret
= perf_output_begin(&handle
, event
, size
, 0, 0);
3426 comm_event
->event_id
.pid
= perf_event_pid(event
, comm_event
->task
);
3427 comm_event
->event_id
.tid
= perf_event_tid(event
, comm_event
->task
);
3429 perf_output_put(&handle
, comm_event
->event_id
);
3430 perf_output_copy(&handle
, comm_event
->comm
,
3431 comm_event
->comm_size
);
3432 perf_output_end(&handle
);
3435 static int perf_event_comm_match(struct perf_event
*event
)
3437 if (event
->state
< PERF_EVENT_STATE_INACTIVE
)
3440 if (event
->cpu
!= -1 && event
->cpu
!= smp_processor_id())
3443 if (event
->attr
.comm
)
3449 static void perf_event_comm_ctx(struct perf_event_context
*ctx
,
3450 struct perf_comm_event
*comm_event
)
3452 struct perf_event
*event
;
3454 list_for_each_entry_rcu(event
, &ctx
->event_list
, event_entry
) {
3455 if (perf_event_comm_match(event
))
3456 perf_event_comm_output(event
, comm_event
);
3460 static void perf_event_comm_event(struct perf_comm_event
*comm_event
)
3462 struct perf_cpu_context
*cpuctx
;
3463 struct perf_event_context
*ctx
;
3465 char comm
[TASK_COMM_LEN
];
3467 memset(comm
, 0, sizeof(comm
));
3468 strlcpy(comm
, comm_event
->task
->comm
, sizeof(comm
));
3469 size
= ALIGN(strlen(comm
)+1, sizeof(u64
));
3471 comm_event
->comm
= comm
;
3472 comm_event
->comm_size
= size
;
3474 comm_event
->event_id
.header
.size
= sizeof(comm_event
->event_id
) + size
;
3477 cpuctx
= &get_cpu_var(perf_cpu_context
);
3478 perf_event_comm_ctx(&cpuctx
->ctx
, comm_event
);
3479 ctx
= rcu_dereference(current
->perf_event_ctxp
);
3481 perf_event_comm_ctx(ctx
, comm_event
);
3482 put_cpu_var(perf_cpu_context
);
3486 void perf_event_comm(struct task_struct
*task
)
3488 struct perf_comm_event comm_event
;
3490 if (task
->perf_event_ctxp
)
3491 perf_event_enable_on_exec(task
);
3493 if (!atomic_read(&nr_comm_events
))
3496 comm_event
= (struct perf_comm_event
){
3502 .type
= PERF_RECORD_COMM
,
3511 perf_event_comm_event(&comm_event
);
3518 struct perf_mmap_event
{
3519 struct vm_area_struct
*vma
;
3521 const char *file_name
;
3525 struct perf_event_header header
;
3535 static void perf_event_mmap_output(struct perf_event
*event
,
3536 struct perf_mmap_event
*mmap_event
)
3538 struct perf_output_handle handle
;
3539 int size
= mmap_event
->event_id
.header
.size
;
3540 int ret
= perf_output_begin(&handle
, event
, size
, 0, 0);
3545 mmap_event
->event_id
.pid
= perf_event_pid(event
, current
);
3546 mmap_event
->event_id
.tid
= perf_event_tid(event
, current
);
3548 perf_output_put(&handle
, mmap_event
->event_id
);
3549 perf_output_copy(&handle
, mmap_event
->file_name
,
3550 mmap_event
->file_size
);
3551 perf_output_end(&handle
);
3554 static int perf_event_mmap_match(struct perf_event
*event
,
3555 struct perf_mmap_event
*mmap_event
)
3557 if (event
->state
< PERF_EVENT_STATE_INACTIVE
)
3560 if (event
->cpu
!= -1 && event
->cpu
!= smp_processor_id())
3563 if (event
->attr
.mmap
)
3569 static void perf_event_mmap_ctx(struct perf_event_context
*ctx
,
3570 struct perf_mmap_event
*mmap_event
)
3572 struct perf_event
*event
;
3574 list_for_each_entry_rcu(event
, &ctx
->event_list
, event_entry
) {
3575 if (perf_event_mmap_match(event
, mmap_event
))
3576 perf_event_mmap_output(event
, mmap_event
);
3580 static void perf_event_mmap_event(struct perf_mmap_event
*mmap_event
)
3582 struct perf_cpu_context
*cpuctx
;
3583 struct perf_event_context
*ctx
;
3584 struct vm_area_struct
*vma
= mmap_event
->vma
;
3585 struct file
*file
= vma
->vm_file
;
3591 memset(tmp
, 0, sizeof(tmp
));
3595 * d_path works from the end of the buffer backwards, so we
3596 * need to add enough zero bytes after the string to handle
3597 * the 64bit alignment we do later.
3599 buf
= kzalloc(PATH_MAX
+ sizeof(u64
), GFP_KERNEL
);
3601 name
= strncpy(tmp
, "//enomem", sizeof(tmp
));
3604 name
= d_path(&file
->f_path
, buf
, PATH_MAX
);
3606 name
= strncpy(tmp
, "//toolong", sizeof(tmp
));
3610 if (arch_vma_name(mmap_event
->vma
)) {
3611 name
= strncpy(tmp
, arch_vma_name(mmap_event
->vma
),
3617 name
= strncpy(tmp
, "[vdso]", sizeof(tmp
));
3621 name
= strncpy(tmp
, "//anon", sizeof(tmp
));
3626 size
= ALIGN(strlen(name
)+1, sizeof(u64
));
3628 mmap_event
->file_name
= name
;
3629 mmap_event
->file_size
= size
;
3631 mmap_event
->event_id
.header
.size
= sizeof(mmap_event
->event_id
) + size
;
3634 cpuctx
= &get_cpu_var(perf_cpu_context
);
3635 perf_event_mmap_ctx(&cpuctx
->ctx
, mmap_event
);
3636 ctx
= rcu_dereference(current
->perf_event_ctxp
);
3638 perf_event_mmap_ctx(ctx
, mmap_event
);
3639 put_cpu_var(perf_cpu_context
);
3645 void __perf_event_mmap(struct vm_area_struct
*vma
)
3647 struct perf_mmap_event mmap_event
;
3649 if (!atomic_read(&nr_mmap_events
))
3652 mmap_event
= (struct perf_mmap_event
){
3658 .type
= PERF_RECORD_MMAP
,
3664 .start
= vma
->vm_start
,
3665 .len
= vma
->vm_end
- vma
->vm_start
,
3666 .pgoff
= vma
->vm_pgoff
,
3670 perf_event_mmap_event(&mmap_event
);
3674 * IRQ throttle logging
3677 static void perf_log_throttle(struct perf_event
*event
, int enable
)
3679 struct perf_output_handle handle
;
3683 struct perf_event_header header
;
3687 } throttle_event
= {
3689 .type
= PERF_RECORD_THROTTLE
,
3691 .size
= sizeof(throttle_event
),
3693 .time
= perf_clock(),
3694 .id
= primary_event_id(event
),
3695 .stream_id
= event
->id
,
3699 throttle_event
.header
.type
= PERF_RECORD_UNTHROTTLE
;
3701 ret
= perf_output_begin(&handle
, event
, sizeof(throttle_event
), 1, 0);
3705 perf_output_put(&handle
, throttle_event
);
3706 perf_output_end(&handle
);
3710 * Generic event overflow handling, sampling.
3713 static int __perf_event_overflow(struct perf_event
*event
, int nmi
,
3714 int throttle
, struct perf_sample_data
*data
,
3715 struct pt_regs
*regs
)
3717 int events
= atomic_read(&event
->event_limit
);
3718 struct hw_perf_event
*hwc
= &event
->hw
;
3721 throttle
= (throttle
&& event
->pmu
->unthrottle
!= NULL
);
3726 if (hwc
->interrupts
!= MAX_INTERRUPTS
) {
3728 if (HZ
* hwc
->interrupts
>
3729 (u64
)sysctl_perf_event_sample_rate
) {
3730 hwc
->interrupts
= MAX_INTERRUPTS
;
3731 perf_log_throttle(event
, 0);
3736 * Keep re-disabling events even though on the previous
3737 * pass we disabled it - just in case we raced with a
3738 * sched-in and the event got enabled again:
3744 if (event
->attr
.freq
) {
3745 u64 now
= perf_clock();
3746 s64 delta
= now
- hwc
->freq_time_stamp
;
3748 hwc
->freq_time_stamp
= now
;
3750 if (delta
> 0 && delta
< 2*TICK_NSEC
)
3751 perf_adjust_period(event
, delta
, hwc
->last_period
);
3755 * XXX event_limit might not quite work as expected on inherited
3759 event
->pending_kill
= POLL_IN
;
3760 if (events
&& atomic_dec_and_test(&event
->event_limit
)) {
3762 event
->pending_kill
= POLL_HUP
;
3764 event
->pending_disable
= 1;
3765 perf_pending_queue(&event
->pending
,
3766 perf_pending_event
);
3768 perf_event_disable(event
);
3771 if (event
->overflow_handler
)
3772 event
->overflow_handler(event
, nmi
, data
, regs
);
3774 perf_event_output(event
, nmi
, data
, regs
);
3779 int perf_event_overflow(struct perf_event
*event
, int nmi
,
3780 struct perf_sample_data
*data
,
3781 struct pt_regs
*regs
)
3783 return __perf_event_overflow(event
, nmi
, 1, data
, regs
);
3787 * Generic software event infrastructure
3791 * We directly increment event->count and keep a second value in
3792 * event->hw.period_left to count intervals. This period event
3793 * is kept in the range [-sample_period, 0] so that we can use the
3797 static u64
perf_swevent_set_period(struct perf_event
*event
)
3799 struct hw_perf_event
*hwc
= &event
->hw
;
3800 u64 period
= hwc
->last_period
;
3804 hwc
->last_period
= hwc
->sample_period
;
3807 old
= val
= atomic64_read(&hwc
->period_left
);
3811 nr
= div64_u64(period
+ val
, period
);
3812 offset
= nr
* period
;
3814 if (atomic64_cmpxchg(&hwc
->period_left
, old
, val
) != old
)
3820 static void perf_swevent_overflow(struct perf_event
*event
, u64 overflow
,
3821 int nmi
, struct perf_sample_data
*data
,
3822 struct pt_regs
*regs
)
3824 struct hw_perf_event
*hwc
= &event
->hw
;
3827 data
->period
= event
->hw
.last_period
;
3829 overflow
= perf_swevent_set_period(event
);
3831 if (hwc
->interrupts
== MAX_INTERRUPTS
)
3834 for (; overflow
; overflow
--) {
3835 if (__perf_event_overflow(event
, nmi
, throttle
,
3838 * We inhibit the overflow from happening when
3839 * hwc->interrupts == MAX_INTERRUPTS.
3847 static void perf_swevent_unthrottle(struct perf_event
*event
)
3850 * Nothing to do, we already reset hwc->interrupts.
3854 static void perf_swevent_add(struct perf_event
*event
, u64 nr
,
3855 int nmi
, struct perf_sample_data
*data
,
3856 struct pt_regs
*regs
)
3858 struct hw_perf_event
*hwc
= &event
->hw
;
3860 atomic64_add(nr
, &event
->count
);
3865 if (!hwc
->sample_period
)
3868 if (nr
== 1 && hwc
->sample_period
== 1 && !event
->attr
.freq
)
3869 return perf_swevent_overflow(event
, 1, nmi
, data
, regs
);
3871 if (atomic64_add_negative(nr
, &hwc
->period_left
))
3874 perf_swevent_overflow(event
, 0, nmi
, data
, regs
);
3877 static int perf_swevent_is_counting(struct perf_event
*event
)
3880 * The event is active, we're good!
3882 if (event
->state
== PERF_EVENT_STATE_ACTIVE
)
3886 * The event is off/error, not counting.
3888 if (event
->state
!= PERF_EVENT_STATE_INACTIVE
)
3892 * The event is inactive, if the context is active
3893 * we're part of a group that didn't make it on the 'pmu',
3896 if (event
->ctx
->is_active
)
3900 * We're inactive and the context is too, this means the
3901 * task is scheduled out, we're counting events that happen
3902 * to us, like migration events.
3907 static int perf_tp_event_match(struct perf_event
*event
,
3908 struct perf_sample_data
*data
);
3910 static int perf_exclude_event(struct perf_event
*event
,
3911 struct pt_regs
*regs
)
3914 if (event
->attr
.exclude_user
&& user_mode(regs
))
3917 if (event
->attr
.exclude_kernel
&& !user_mode(regs
))
3924 static int perf_swevent_match(struct perf_event
*event
,
3925 enum perf_type_id type
,
3927 struct perf_sample_data
*data
,
3928 struct pt_regs
*regs
)
3930 if (event
->cpu
!= -1 && event
->cpu
!= smp_processor_id())
3933 if (!perf_swevent_is_counting(event
))
3936 if (event
->attr
.type
!= type
)
3939 if (event
->attr
.config
!= event_id
)
3942 if (perf_exclude_event(event
, regs
))
3945 if (event
->attr
.type
== PERF_TYPE_TRACEPOINT
&&
3946 !perf_tp_event_match(event
, data
))
3952 static void perf_swevent_ctx_event(struct perf_event_context
*ctx
,
3953 enum perf_type_id type
,
3954 u32 event_id
, u64 nr
, int nmi
,
3955 struct perf_sample_data
*data
,
3956 struct pt_regs
*regs
)
3958 struct perf_event
*event
;
3960 list_for_each_entry_rcu(event
, &ctx
->event_list
, event_entry
) {
3961 if (perf_swevent_match(event
, type
, event_id
, data
, regs
))
3962 perf_swevent_add(event
, nr
, nmi
, data
, regs
);
3966 int perf_swevent_get_recursion_context(void)
3968 struct perf_cpu_context
*cpuctx
= &get_cpu_var(perf_cpu_context
);
3975 else if (in_softirq())
3980 if (cpuctx
->recursion
[rctx
]) {
3981 put_cpu_var(perf_cpu_context
);
3985 cpuctx
->recursion
[rctx
]++;
3990 EXPORT_SYMBOL_GPL(perf_swevent_get_recursion_context
);
3992 void perf_swevent_put_recursion_context(int rctx
)
3994 struct perf_cpu_context
*cpuctx
= &__get_cpu_var(perf_cpu_context
);
3996 cpuctx
->recursion
[rctx
]--;
3997 put_cpu_var(perf_cpu_context
);
3999 EXPORT_SYMBOL_GPL(perf_swevent_put_recursion_context
);
4001 static void do_perf_sw_event(enum perf_type_id type
, u32 event_id
,
4003 struct perf_sample_data
*data
,
4004 struct pt_regs
*regs
)
4006 struct perf_cpu_context
*cpuctx
;
4007 struct perf_event_context
*ctx
;
4009 cpuctx
= &__get_cpu_var(perf_cpu_context
);
4011 perf_swevent_ctx_event(&cpuctx
->ctx
, type
, event_id
,
4012 nr
, nmi
, data
, regs
);
4014 * doesn't really matter which of the child contexts the
4015 * events ends up in.
4017 ctx
= rcu_dereference(current
->perf_event_ctxp
);
4019 perf_swevent_ctx_event(ctx
, type
, event_id
, nr
, nmi
, data
, regs
);
4023 void __perf_sw_event(u32 event_id
, u64 nr
, int nmi
,
4024 struct pt_regs
*regs
, u64 addr
)
4026 struct perf_sample_data data
;
4029 rctx
= perf_swevent_get_recursion_context();
4033 perf_sample_data_init(&data
, addr
);
4035 do_perf_sw_event(PERF_TYPE_SOFTWARE
, event_id
, nr
, nmi
, &data
, regs
);
4037 perf_swevent_put_recursion_context(rctx
);
4040 static void perf_swevent_read(struct perf_event
*event
)
4044 static int perf_swevent_enable(struct perf_event
*event
)
4046 struct hw_perf_event
*hwc
= &event
->hw
;
4048 if (hwc
->sample_period
) {
4049 hwc
->last_period
= hwc
->sample_period
;
4050 perf_swevent_set_period(event
);
4055 static void perf_swevent_disable(struct perf_event
*event
)
4059 static const struct pmu perf_ops_generic
= {
4060 .enable
= perf_swevent_enable
,
4061 .disable
= perf_swevent_disable
,
4062 .read
= perf_swevent_read
,
4063 .unthrottle
= perf_swevent_unthrottle
,
4067 * hrtimer based swevent callback
4070 static enum hrtimer_restart
perf_swevent_hrtimer(struct hrtimer
*hrtimer
)
4072 enum hrtimer_restart ret
= HRTIMER_RESTART
;
4073 struct perf_sample_data data
;
4074 struct pt_regs
*regs
;
4075 struct perf_event
*event
;
4078 event
= container_of(hrtimer
, struct perf_event
, hw
.hrtimer
);
4079 event
->pmu
->read(event
);
4081 perf_sample_data_init(&data
, 0);
4082 data
.period
= event
->hw
.last_period
;
4083 regs
= get_irq_regs();
4085 * In case we exclude kernel IPs or are somehow not in interrupt
4086 * context, provide the next best thing, the user IP.
4088 if ((event
->attr
.exclude_kernel
|| !regs
) &&
4089 !event
->attr
.exclude_user
)
4090 regs
= task_pt_regs(current
);
4093 if (!(event
->attr
.exclude_idle
&& current
->pid
== 0))
4094 if (perf_event_overflow(event
, 0, &data
, regs
))
4095 ret
= HRTIMER_NORESTART
;
4098 period
= max_t(u64
, 10000, event
->hw
.sample_period
);
4099 hrtimer_forward_now(hrtimer
, ns_to_ktime(period
));
4104 static void perf_swevent_start_hrtimer(struct perf_event
*event
)
4106 struct hw_perf_event
*hwc
= &event
->hw
;
4108 hrtimer_init(&hwc
->hrtimer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
4109 hwc
->hrtimer
.function
= perf_swevent_hrtimer
;
4110 if (hwc
->sample_period
) {
4113 if (hwc
->remaining
) {
4114 if (hwc
->remaining
< 0)
4117 period
= hwc
->remaining
;
4120 period
= max_t(u64
, 10000, hwc
->sample_period
);
4122 __hrtimer_start_range_ns(&hwc
->hrtimer
,
4123 ns_to_ktime(period
), 0,
4124 HRTIMER_MODE_REL
, 0);
4128 static void perf_swevent_cancel_hrtimer(struct perf_event
*event
)
4130 struct hw_perf_event
*hwc
= &event
->hw
;
4132 if (hwc
->sample_period
) {
4133 ktime_t remaining
= hrtimer_get_remaining(&hwc
->hrtimer
);
4134 hwc
->remaining
= ktime_to_ns(remaining
);
4136 hrtimer_cancel(&hwc
->hrtimer
);
4141 * Software event: cpu wall time clock
4144 static void cpu_clock_perf_event_update(struct perf_event
*event
)
4146 int cpu
= raw_smp_processor_id();
4150 now
= cpu_clock(cpu
);
4151 prev
= atomic64_xchg(&event
->hw
.prev_count
, now
);
4152 atomic64_add(now
- prev
, &event
->count
);
4155 static int cpu_clock_perf_event_enable(struct perf_event
*event
)
4157 struct hw_perf_event
*hwc
= &event
->hw
;
4158 int cpu
= raw_smp_processor_id();
4160 atomic64_set(&hwc
->prev_count
, cpu_clock(cpu
));
4161 perf_swevent_start_hrtimer(event
);
4166 static void cpu_clock_perf_event_disable(struct perf_event
*event
)
4168 perf_swevent_cancel_hrtimer(event
);
4169 cpu_clock_perf_event_update(event
);
4172 static void cpu_clock_perf_event_read(struct perf_event
*event
)
4174 cpu_clock_perf_event_update(event
);
4177 static const struct pmu perf_ops_cpu_clock
= {
4178 .enable
= cpu_clock_perf_event_enable
,
4179 .disable
= cpu_clock_perf_event_disable
,
4180 .read
= cpu_clock_perf_event_read
,
4184 * Software event: task time clock
4187 static void task_clock_perf_event_update(struct perf_event
*event
, u64 now
)
4192 prev
= atomic64_xchg(&event
->hw
.prev_count
, now
);
4194 atomic64_add(delta
, &event
->count
);
4197 static int task_clock_perf_event_enable(struct perf_event
*event
)
4199 struct hw_perf_event
*hwc
= &event
->hw
;
4202 now
= event
->ctx
->time
;
4204 atomic64_set(&hwc
->prev_count
, now
);
4206 perf_swevent_start_hrtimer(event
);
4211 static void task_clock_perf_event_disable(struct perf_event
*event
)
4213 perf_swevent_cancel_hrtimer(event
);
4214 task_clock_perf_event_update(event
, event
->ctx
->time
);
4218 static void task_clock_perf_event_read(struct perf_event
*event
)
4223 update_context_time(event
->ctx
);
4224 time
= event
->ctx
->time
;
4226 u64 now
= perf_clock();
4227 u64 delta
= now
- event
->ctx
->timestamp
;
4228 time
= event
->ctx
->time
+ delta
;
4231 task_clock_perf_event_update(event
, time
);
4234 static const struct pmu perf_ops_task_clock
= {
4235 .enable
= task_clock_perf_event_enable
,
4236 .disable
= task_clock_perf_event_disable
,
4237 .read
= task_clock_perf_event_read
,
4240 #ifdef CONFIG_EVENT_PROFILE
4242 void perf_tp_event(int event_id
, u64 addr
, u64 count
, void *record
,
4245 struct pt_regs
*regs
= get_irq_regs();
4246 struct perf_sample_data data
;
4247 struct perf_raw_record raw
= {
4252 perf_sample_data_init(&data
, addr
);
4256 regs
= task_pt_regs(current
);
4258 /* Trace events already protected against recursion */
4259 do_perf_sw_event(PERF_TYPE_TRACEPOINT
, event_id
, count
, 1,
4262 EXPORT_SYMBOL_GPL(perf_tp_event
);
4264 static int perf_tp_event_match(struct perf_event
*event
,
4265 struct perf_sample_data
*data
)
4267 void *record
= data
->raw
->data
;
4269 if (likely(!event
->filter
) || filter_match_preds(event
->filter
, record
))
4274 static void tp_perf_event_destroy(struct perf_event
*event
)
4276 ftrace_profile_disable(event
->attr
.config
);
4279 static const struct pmu
*tp_perf_event_init(struct perf_event
*event
)
4282 * Raw tracepoint data is a severe data leak, only allow root to
4285 if ((event
->attr
.sample_type
& PERF_SAMPLE_RAW
) &&
4286 perf_paranoid_tracepoint_raw() &&
4287 !capable(CAP_SYS_ADMIN
))
4288 return ERR_PTR(-EPERM
);
4290 if (ftrace_profile_enable(event
->attr
.config
))
4293 event
->destroy
= tp_perf_event_destroy
;
4295 return &perf_ops_generic
;
4298 static int perf_event_set_filter(struct perf_event
*event
, void __user
*arg
)
4303 if (event
->attr
.type
!= PERF_TYPE_TRACEPOINT
)
4306 filter_str
= strndup_user(arg
, PAGE_SIZE
);
4307 if (IS_ERR(filter_str
))
4308 return PTR_ERR(filter_str
);
4310 ret
= ftrace_profile_set_filter(event
, event
->attr
.config
, filter_str
);
4316 static void perf_event_free_filter(struct perf_event
*event
)
4318 ftrace_profile_free_filter(event
);
4323 static int perf_tp_event_match(struct perf_event
*event
,
4324 struct perf_sample_data
*data
)
4329 static const struct pmu
*tp_perf_event_init(struct perf_event
*event
)
4334 static int perf_event_set_filter(struct perf_event
*event
, void __user
*arg
)
4339 static void perf_event_free_filter(struct perf_event
*event
)
4343 #endif /* CONFIG_EVENT_PROFILE */
4345 #ifdef CONFIG_HAVE_HW_BREAKPOINT
4346 static void bp_perf_event_destroy(struct perf_event
*event
)
4348 release_bp_slot(event
);
4351 static const struct pmu
*bp_perf_event_init(struct perf_event
*bp
)
4355 err
= register_perf_hw_breakpoint(bp
);
4357 return ERR_PTR(err
);
4359 bp
->destroy
= bp_perf_event_destroy
;
4361 return &perf_ops_bp
;
4364 void perf_bp_event(struct perf_event
*bp
, void *data
)
4366 struct perf_sample_data sample
;
4367 struct pt_regs
*regs
= data
;
4369 perf_sample_data_init(&sample
, bp
->attr
.bp_addr
);
4371 if (!perf_exclude_event(bp
, regs
))
4372 perf_swevent_add(bp
, 1, 1, &sample
, regs
);
4375 static const struct pmu
*bp_perf_event_init(struct perf_event
*bp
)
4380 void perf_bp_event(struct perf_event
*bp
, void *regs
)
4385 atomic_t perf_swevent_enabled
[PERF_COUNT_SW_MAX
];
4387 static void sw_perf_event_destroy(struct perf_event
*event
)
4389 u64 event_id
= event
->attr
.config
;
4391 WARN_ON(event
->parent
);
4393 atomic_dec(&perf_swevent_enabled
[event_id
]);
4396 static const struct pmu
*sw_perf_event_init(struct perf_event
*event
)
4398 const struct pmu
*pmu
= NULL
;
4399 u64 event_id
= event
->attr
.config
;
4402 * Software events (currently) can't in general distinguish
4403 * between user, kernel and hypervisor events.
4404 * However, context switches and cpu migrations are considered
4405 * to be kernel events, and page faults are never hypervisor
4409 case PERF_COUNT_SW_CPU_CLOCK
:
4410 pmu
= &perf_ops_cpu_clock
;
4413 case PERF_COUNT_SW_TASK_CLOCK
:
4415 * If the user instantiates this as a per-cpu event,
4416 * use the cpu_clock event instead.
4418 if (event
->ctx
->task
)
4419 pmu
= &perf_ops_task_clock
;
4421 pmu
= &perf_ops_cpu_clock
;
4424 case PERF_COUNT_SW_PAGE_FAULTS
:
4425 case PERF_COUNT_SW_PAGE_FAULTS_MIN
:
4426 case PERF_COUNT_SW_PAGE_FAULTS_MAJ
:
4427 case PERF_COUNT_SW_CONTEXT_SWITCHES
:
4428 case PERF_COUNT_SW_CPU_MIGRATIONS
:
4429 case PERF_COUNT_SW_ALIGNMENT_FAULTS
:
4430 case PERF_COUNT_SW_EMULATION_FAULTS
:
4431 if (!event
->parent
) {
4432 atomic_inc(&perf_swevent_enabled
[event_id
]);
4433 event
->destroy
= sw_perf_event_destroy
;
4435 pmu
= &perf_ops_generic
;
4443 * Allocate and initialize a event structure
4445 static struct perf_event
*
4446 perf_event_alloc(struct perf_event_attr
*attr
,
4448 struct perf_event_context
*ctx
,
4449 struct perf_event
*group_leader
,
4450 struct perf_event
*parent_event
,
4451 perf_overflow_handler_t overflow_handler
,
4454 const struct pmu
*pmu
;
4455 struct perf_event
*event
;
4456 struct hw_perf_event
*hwc
;
4459 event
= kzalloc(sizeof(*event
), gfpflags
);
4461 return ERR_PTR(-ENOMEM
);
4464 * Single events are their own group leaders, with an
4465 * empty sibling list:
4468 group_leader
= event
;
4470 mutex_init(&event
->child_mutex
);
4471 INIT_LIST_HEAD(&event
->child_list
);
4473 INIT_LIST_HEAD(&event
->group_entry
);
4474 INIT_LIST_HEAD(&event
->event_entry
);
4475 INIT_LIST_HEAD(&event
->sibling_list
);
4476 init_waitqueue_head(&event
->waitq
);
4478 mutex_init(&event
->mmap_mutex
);
4481 event
->attr
= *attr
;
4482 event
->group_leader
= group_leader
;
4487 event
->parent
= parent_event
;
4489 event
->ns
= get_pid_ns(current
->nsproxy
->pid_ns
);
4490 event
->id
= atomic64_inc_return(&perf_event_id
);
4492 event
->state
= PERF_EVENT_STATE_INACTIVE
;
4494 if (!overflow_handler
&& parent_event
)
4495 overflow_handler
= parent_event
->overflow_handler
;
4497 event
->overflow_handler
= overflow_handler
;
4500 event
->state
= PERF_EVENT_STATE_OFF
;
4505 hwc
->sample_period
= attr
->sample_period
;
4506 if (attr
->freq
&& attr
->sample_freq
)
4507 hwc
->sample_period
= 1;
4508 hwc
->last_period
= hwc
->sample_period
;
4510 atomic64_set(&hwc
->period_left
, hwc
->sample_period
);
4513 * we currently do not support PERF_FORMAT_GROUP on inherited events
4515 if (attr
->inherit
&& (attr
->read_format
& PERF_FORMAT_GROUP
))
4518 switch (attr
->type
) {
4520 case PERF_TYPE_HARDWARE
:
4521 case PERF_TYPE_HW_CACHE
:
4522 pmu
= hw_perf_event_init(event
);
4525 case PERF_TYPE_SOFTWARE
:
4526 pmu
= sw_perf_event_init(event
);
4529 case PERF_TYPE_TRACEPOINT
:
4530 pmu
= tp_perf_event_init(event
);
4533 case PERF_TYPE_BREAKPOINT
:
4534 pmu
= bp_perf_event_init(event
);
4545 else if (IS_ERR(pmu
))
4550 put_pid_ns(event
->ns
);
4552 return ERR_PTR(err
);
4557 if (!event
->parent
) {
4558 atomic_inc(&nr_events
);
4559 if (event
->attr
.mmap
)
4560 atomic_inc(&nr_mmap_events
);
4561 if (event
->attr
.comm
)
4562 atomic_inc(&nr_comm_events
);
4563 if (event
->attr
.task
)
4564 atomic_inc(&nr_task_events
);
4570 static int perf_copy_attr(struct perf_event_attr __user
*uattr
,
4571 struct perf_event_attr
*attr
)
4576 if (!access_ok(VERIFY_WRITE
, uattr
, PERF_ATTR_SIZE_VER0
))
4580 * zero the full structure, so that a short copy will be nice.
4582 memset(attr
, 0, sizeof(*attr
));
4584 ret
= get_user(size
, &uattr
->size
);
4588 if (size
> PAGE_SIZE
) /* silly large */
4591 if (!size
) /* abi compat */
4592 size
= PERF_ATTR_SIZE_VER0
;
4594 if (size
< PERF_ATTR_SIZE_VER0
)
4598 * If we're handed a bigger struct than we know of,
4599 * ensure all the unknown bits are 0 - i.e. new
4600 * user-space does not rely on any kernel feature
4601 * extensions we dont know about yet.
4603 if (size
> sizeof(*attr
)) {
4604 unsigned char __user
*addr
;
4605 unsigned char __user
*end
;
4608 addr
= (void __user
*)uattr
+ sizeof(*attr
);
4609 end
= (void __user
*)uattr
+ size
;
4611 for (; addr
< end
; addr
++) {
4612 ret
= get_user(val
, addr
);
4618 size
= sizeof(*attr
);
4621 ret
= copy_from_user(attr
, uattr
, size
);
4626 * If the type exists, the corresponding creation will verify
4629 if (attr
->type
>= PERF_TYPE_MAX
)
4632 if (attr
->__reserved_1
)
4635 if (attr
->sample_type
& ~(PERF_SAMPLE_MAX
-1))
4638 if (attr
->read_format
& ~(PERF_FORMAT_MAX
-1))
4645 put_user(sizeof(*attr
), &uattr
->size
);
4650 static int perf_event_set_output(struct perf_event
*event
, int output_fd
)
4652 struct perf_event
*output_event
= NULL
;
4653 struct file
*output_file
= NULL
;
4654 struct perf_event
*old_output
;
4655 int fput_needed
= 0;
4661 output_file
= fget_light(output_fd
, &fput_needed
);
4665 if (output_file
->f_op
!= &perf_fops
)
4668 output_event
= output_file
->private_data
;
4670 /* Don't chain output fds */
4671 if (output_event
->output
)
4674 /* Don't set an output fd when we already have an output channel */
4678 atomic_long_inc(&output_file
->f_count
);
4681 mutex_lock(&event
->mmap_mutex
);
4682 old_output
= event
->output
;
4683 rcu_assign_pointer(event
->output
, output_event
);
4684 mutex_unlock(&event
->mmap_mutex
);
4688 * we need to make sure no existing perf_output_*()
4689 * is still referencing this event.
4692 fput(old_output
->filp
);
4697 fput_light(output_file
, fput_needed
);
4702 * sys_perf_event_open - open a performance event, associate it to a task/cpu
4704 * @attr_uptr: event_id type attributes for monitoring/sampling
4707 * @group_fd: group leader event fd
4709 SYSCALL_DEFINE5(perf_event_open
,
4710 struct perf_event_attr __user
*, attr_uptr
,
4711 pid_t
, pid
, int, cpu
, int, group_fd
, unsigned long, flags
)
4713 struct perf_event
*event
, *group_leader
;
4714 struct perf_event_attr attr
;
4715 struct perf_event_context
*ctx
;
4716 struct file
*event_file
= NULL
;
4717 struct file
*group_file
= NULL
;
4719 int fput_needed
= 0;
4722 /* for future expandability... */
4723 if (flags
& ~(PERF_FLAG_FD_NO_GROUP
| PERF_FLAG_FD_OUTPUT
))
4726 err
= perf_copy_attr(attr_uptr
, &attr
);
4730 if (!attr
.exclude_kernel
) {
4731 if (perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN
))
4736 if (attr
.sample_freq
> sysctl_perf_event_sample_rate
)
4740 event_fd
= get_unused_fd_flags(O_RDWR
);
4745 * Get the target context (task or percpu):
4747 ctx
= find_get_context(pid
, cpu
);
4754 * Look up the group leader (we will attach this event to it):
4756 group_leader
= NULL
;
4757 if (group_fd
!= -1 && !(flags
& PERF_FLAG_FD_NO_GROUP
)) {
4759 group_file
= fget_light(group_fd
, &fput_needed
);
4761 goto err_put_context
;
4762 if (group_file
->f_op
!= &perf_fops
)
4763 goto err_put_context
;
4765 group_leader
= group_file
->private_data
;
4767 * Do not allow a recursive hierarchy (this new sibling
4768 * becoming part of another group-sibling):
4770 if (group_leader
->group_leader
!= group_leader
)
4771 goto err_put_context
;
4773 * Do not allow to attach to a group in a different
4774 * task or CPU context:
4776 if (group_leader
->ctx
!= ctx
)
4777 goto err_put_context
;
4779 * Only a group leader can be exclusive or pinned
4781 if (attr
.exclusive
|| attr
.pinned
)
4782 goto err_put_context
;
4785 event
= perf_event_alloc(&attr
, cpu
, ctx
, group_leader
,
4786 NULL
, NULL
, GFP_KERNEL
);
4787 err
= PTR_ERR(event
);
4789 goto err_put_context
;
4791 event_file
= anon_inode_getfile("[perf_event]", &perf_fops
, event
, O_RDWR
);
4792 if (IS_ERR(event_file
)) {
4793 err
= PTR_ERR(event_file
);
4794 goto err_free_put_context
;
4797 if (flags
& PERF_FLAG_FD_OUTPUT
) {
4798 err
= perf_event_set_output(event
, group_fd
);
4800 goto err_fput_free_put_context
;
4803 event
->filp
= event_file
;
4804 WARN_ON_ONCE(ctx
->parent_ctx
);
4805 mutex_lock(&ctx
->mutex
);
4806 perf_install_in_context(ctx
, event
, cpu
);
4808 mutex_unlock(&ctx
->mutex
);
4810 event
->owner
= current
;
4811 get_task_struct(current
);
4812 mutex_lock(¤t
->perf_event_mutex
);
4813 list_add_tail(&event
->owner_entry
, ¤t
->perf_event_list
);
4814 mutex_unlock(¤t
->perf_event_mutex
);
4816 fput_light(group_file
, fput_needed
);
4817 fd_install(event_fd
, event_file
);
4820 err_fput_free_put_context
:
4822 err_free_put_context
:
4825 fput_light(group_file
, fput_needed
);
4828 put_unused_fd(event_fd
);
4833 * perf_event_create_kernel_counter
4835 * @attr: attributes of the counter to create
4836 * @cpu: cpu in which the counter is bound
4837 * @pid: task to profile
4840 perf_event_create_kernel_counter(struct perf_event_attr
*attr
, int cpu
,
4842 perf_overflow_handler_t overflow_handler
)
4844 struct perf_event
*event
;
4845 struct perf_event_context
*ctx
;
4849 * Get the target context (task or percpu):
4852 ctx
= find_get_context(pid
, cpu
);
4858 event
= perf_event_alloc(attr
, cpu
, ctx
, NULL
,
4859 NULL
, overflow_handler
, GFP_KERNEL
);
4860 if (IS_ERR(event
)) {
4861 err
= PTR_ERR(event
);
4862 goto err_put_context
;
4866 WARN_ON_ONCE(ctx
->parent_ctx
);
4867 mutex_lock(&ctx
->mutex
);
4868 perf_install_in_context(ctx
, event
, cpu
);
4870 mutex_unlock(&ctx
->mutex
);
4872 event
->owner
= current
;
4873 get_task_struct(current
);
4874 mutex_lock(¤t
->perf_event_mutex
);
4875 list_add_tail(&event
->owner_entry
, ¤t
->perf_event_list
);
4876 mutex_unlock(¤t
->perf_event_mutex
);
4883 return ERR_PTR(err
);
4885 EXPORT_SYMBOL_GPL(perf_event_create_kernel_counter
);
4888 * inherit a event from parent task to child task:
4890 static struct perf_event
*
4891 inherit_event(struct perf_event
*parent_event
,
4892 struct task_struct
*parent
,
4893 struct perf_event_context
*parent_ctx
,
4894 struct task_struct
*child
,
4895 struct perf_event
*group_leader
,
4896 struct perf_event_context
*child_ctx
)
4898 struct perf_event
*child_event
;
4901 * Instead of creating recursive hierarchies of events,
4902 * we link inherited events back to the original parent,
4903 * which has a filp for sure, which we use as the reference
4906 if (parent_event
->parent
)
4907 parent_event
= parent_event
->parent
;
4909 child_event
= perf_event_alloc(&parent_event
->attr
,
4910 parent_event
->cpu
, child_ctx
,
4911 group_leader
, parent_event
,
4913 if (IS_ERR(child_event
))
4918 * Make the child state follow the state of the parent event,
4919 * not its attr.disabled bit. We hold the parent's mutex,
4920 * so we won't race with perf_event_{en, dis}able_family.
4922 if (parent_event
->state
>= PERF_EVENT_STATE_INACTIVE
)
4923 child_event
->state
= PERF_EVENT_STATE_INACTIVE
;
4925 child_event
->state
= PERF_EVENT_STATE_OFF
;
4927 if (parent_event
->attr
.freq
)
4928 child_event
->hw
.sample_period
= parent_event
->hw
.sample_period
;
4930 child_event
->overflow_handler
= parent_event
->overflow_handler
;
4933 * Link it up in the child's context:
4935 add_event_to_ctx(child_event
, child_ctx
);
4938 * Get a reference to the parent filp - we will fput it
4939 * when the child event exits. This is safe to do because
4940 * we are in the parent and we know that the filp still
4941 * exists and has a nonzero count:
4943 atomic_long_inc(&parent_event
->filp
->f_count
);
4946 * Link this into the parent event's child list
4948 WARN_ON_ONCE(parent_event
->ctx
->parent_ctx
);
4949 mutex_lock(&parent_event
->child_mutex
);
4950 list_add_tail(&child_event
->child_list
, &parent_event
->child_list
);
4951 mutex_unlock(&parent_event
->child_mutex
);
4956 static int inherit_group(struct perf_event
*parent_event
,
4957 struct task_struct
*parent
,
4958 struct perf_event_context
*parent_ctx
,
4959 struct task_struct
*child
,
4960 struct perf_event_context
*child_ctx
)
4962 struct perf_event
*leader
;
4963 struct perf_event
*sub
;
4964 struct perf_event
*child_ctr
;
4966 leader
= inherit_event(parent_event
, parent
, parent_ctx
,
4967 child
, NULL
, child_ctx
);
4969 return PTR_ERR(leader
);
4970 list_for_each_entry(sub
, &parent_event
->sibling_list
, group_entry
) {
4971 child_ctr
= inherit_event(sub
, parent
, parent_ctx
,
4972 child
, leader
, child_ctx
);
4973 if (IS_ERR(child_ctr
))
4974 return PTR_ERR(child_ctr
);
4979 static void sync_child_event(struct perf_event
*child_event
,
4980 struct task_struct
*child
)
4982 struct perf_event
*parent_event
= child_event
->parent
;
4985 if (child_event
->attr
.inherit_stat
)
4986 perf_event_read_event(child_event
, child
);
4988 child_val
= atomic64_read(&child_event
->count
);
4991 * Add back the child's count to the parent's count:
4993 atomic64_add(child_val
, &parent_event
->count
);
4994 atomic64_add(child_event
->total_time_enabled
,
4995 &parent_event
->child_total_time_enabled
);
4996 atomic64_add(child_event
->total_time_running
,
4997 &parent_event
->child_total_time_running
);
5000 * Remove this event from the parent's list
5002 WARN_ON_ONCE(parent_event
->ctx
->parent_ctx
);
5003 mutex_lock(&parent_event
->child_mutex
);
5004 list_del_init(&child_event
->child_list
);
5005 mutex_unlock(&parent_event
->child_mutex
);
5008 * Release the parent event, if this was the last
5011 fput(parent_event
->filp
);
5015 __perf_event_exit_task(struct perf_event
*child_event
,
5016 struct perf_event_context
*child_ctx
,
5017 struct task_struct
*child
)
5019 struct perf_event
*parent_event
;
5021 perf_event_remove_from_context(child_event
);
5023 parent_event
= child_event
->parent
;
5025 * It can happen that parent exits first, and has events
5026 * that are still around due to the child reference. These
5027 * events need to be zapped - but otherwise linger.
5030 sync_child_event(child_event
, child
);
5031 free_event(child_event
);
5036 * When a child task exits, feed back event values to parent events.
5038 void perf_event_exit_task(struct task_struct
*child
)
5040 struct perf_event
*child_event
, *tmp
;
5041 struct perf_event_context
*child_ctx
;
5042 unsigned long flags
;
5044 if (likely(!child
->perf_event_ctxp
)) {
5045 perf_event_task(child
, NULL
, 0);
5049 local_irq_save(flags
);
5051 * We can't reschedule here because interrupts are disabled,
5052 * and either child is current or it is a task that can't be
5053 * scheduled, so we are now safe from rescheduling changing
5056 child_ctx
= child
->perf_event_ctxp
;
5057 __perf_event_task_sched_out(child_ctx
);
5060 * Take the context lock here so that if find_get_context is
5061 * reading child->perf_event_ctxp, we wait until it has
5062 * incremented the context's refcount before we do put_ctx below.
5064 raw_spin_lock(&child_ctx
->lock
);
5065 child
->perf_event_ctxp
= NULL
;
5067 * If this context is a clone; unclone it so it can't get
5068 * swapped to another process while we're removing all
5069 * the events from it.
5071 unclone_ctx(child_ctx
);
5072 update_context_time(child_ctx
);
5073 raw_spin_unlock_irqrestore(&child_ctx
->lock
, flags
);
5076 * Report the task dead after unscheduling the events so that we
5077 * won't get any samples after PERF_RECORD_EXIT. We can however still
5078 * get a few PERF_RECORD_READ events.
5080 perf_event_task(child
, child_ctx
, 0);
5083 * We can recurse on the same lock type through:
5085 * __perf_event_exit_task()
5086 * sync_child_event()
5087 * fput(parent_event->filp)
5089 * mutex_lock(&ctx->mutex)
5091 * But since its the parent context it won't be the same instance.
5093 mutex_lock_nested(&child_ctx
->mutex
, SINGLE_DEPTH_NESTING
);
5096 list_for_each_entry_safe(child_event
, tmp
, &child_ctx
->group_list
,
5098 __perf_event_exit_task(child_event
, child_ctx
, child
);
5101 * If the last event was a group event, it will have appended all
5102 * its siblings to the list, but we obtained 'tmp' before that which
5103 * will still point to the list head terminating the iteration.
5105 if (!list_empty(&child_ctx
->group_list
))
5108 mutex_unlock(&child_ctx
->mutex
);
5114 * free an unexposed, unused context as created by inheritance by
5115 * init_task below, used by fork() in case of fail.
5117 void perf_event_free_task(struct task_struct
*task
)
5119 struct perf_event_context
*ctx
= task
->perf_event_ctxp
;
5120 struct perf_event
*event
, *tmp
;
5125 mutex_lock(&ctx
->mutex
);
5127 list_for_each_entry_safe(event
, tmp
, &ctx
->group_list
, group_entry
) {
5128 struct perf_event
*parent
= event
->parent
;
5130 if (WARN_ON_ONCE(!parent
))
5133 mutex_lock(&parent
->child_mutex
);
5134 list_del_init(&event
->child_list
);
5135 mutex_unlock(&parent
->child_mutex
);
5139 list_del_event(event
, ctx
);
5143 if (!list_empty(&ctx
->group_list
))
5146 mutex_unlock(&ctx
->mutex
);
5152 * Initialize the perf_event context in task_struct
5154 int perf_event_init_task(struct task_struct
*child
)
5156 struct perf_event_context
*child_ctx
= NULL
, *parent_ctx
;
5157 struct perf_event_context
*cloned_ctx
;
5158 struct perf_event
*event
;
5159 struct task_struct
*parent
= current
;
5160 int inherited_all
= 1;
5163 child
->perf_event_ctxp
= NULL
;
5165 mutex_init(&child
->perf_event_mutex
);
5166 INIT_LIST_HEAD(&child
->perf_event_list
);
5168 if (likely(!parent
->perf_event_ctxp
))
5172 * If the parent's context is a clone, pin it so it won't get
5175 parent_ctx
= perf_pin_task_context(parent
);
5178 * No need to check if parent_ctx != NULL here; since we saw
5179 * it non-NULL earlier, the only reason for it to become NULL
5180 * is if we exit, and since we're currently in the middle of
5181 * a fork we can't be exiting at the same time.
5185 * Lock the parent list. No need to lock the child - not PID
5186 * hashed yet and not running, so nobody can access it.
5188 mutex_lock(&parent_ctx
->mutex
);
5191 * We dont have to disable NMIs - we are only looking at
5192 * the list, not manipulating it:
5194 list_for_each_entry(event
, &parent_ctx
->group_list
, group_entry
) {
5196 if (!event
->attr
.inherit
) {
5201 if (!child
->perf_event_ctxp
) {
5203 * This is executed from the parent task context, so
5204 * inherit events that have been marked for cloning.
5205 * First allocate and initialize a context for the
5209 child_ctx
= kzalloc(sizeof(struct perf_event_context
),
5216 __perf_event_init_context(child_ctx
, child
);
5217 child
->perf_event_ctxp
= child_ctx
;
5218 get_task_struct(child
);
5221 ret
= inherit_group(event
, parent
, parent_ctx
,
5229 if (child_ctx
&& inherited_all
) {
5231 * Mark the child context as a clone of the parent
5232 * context, or of whatever the parent is a clone of.
5233 * Note that if the parent is a clone, it could get
5234 * uncloned at any point, but that doesn't matter
5235 * because the list of events and the generation
5236 * count can't have changed since we took the mutex.
5238 cloned_ctx
= rcu_dereference(parent_ctx
->parent_ctx
);
5240 child_ctx
->parent_ctx
= cloned_ctx
;
5241 child_ctx
->parent_gen
= parent_ctx
->parent_gen
;
5243 child_ctx
->parent_ctx
= parent_ctx
;
5244 child_ctx
->parent_gen
= parent_ctx
->generation
;
5246 get_ctx(child_ctx
->parent_ctx
);
5249 mutex_unlock(&parent_ctx
->mutex
);
5251 perf_unpin_context(parent_ctx
);
5256 static void __init
perf_event_init_all_cpus(void)
5259 struct perf_cpu_context
*cpuctx
;
5261 for_each_possible_cpu(cpu
) {
5262 cpuctx
= &per_cpu(perf_cpu_context
, cpu
);
5263 __perf_event_init_context(&cpuctx
->ctx
, NULL
);
5267 static void __cpuinit
perf_event_init_cpu(int cpu
)
5269 struct perf_cpu_context
*cpuctx
;
5271 cpuctx
= &per_cpu(perf_cpu_context
, cpu
);
5273 spin_lock(&perf_resource_lock
);
5274 cpuctx
->max_pertask
= perf_max_events
- perf_reserved_percpu
;
5275 spin_unlock(&perf_resource_lock
);
5277 hw_perf_event_setup(cpu
);
5280 #ifdef CONFIG_HOTPLUG_CPU
5281 static void __perf_event_exit_cpu(void *info
)
5283 struct perf_cpu_context
*cpuctx
= &__get_cpu_var(perf_cpu_context
);
5284 struct perf_event_context
*ctx
= &cpuctx
->ctx
;
5285 struct perf_event
*event
, *tmp
;
5287 list_for_each_entry_safe(event
, tmp
, &ctx
->group_list
, group_entry
)
5288 __perf_event_remove_from_context(event
);
5290 static void perf_event_exit_cpu(int cpu
)
5292 struct perf_cpu_context
*cpuctx
= &per_cpu(perf_cpu_context
, cpu
);
5293 struct perf_event_context
*ctx
= &cpuctx
->ctx
;
5295 mutex_lock(&ctx
->mutex
);
5296 smp_call_function_single(cpu
, __perf_event_exit_cpu
, NULL
, 1);
5297 mutex_unlock(&ctx
->mutex
);
5300 static inline void perf_event_exit_cpu(int cpu
) { }
5303 static int __cpuinit
5304 perf_cpu_notify(struct notifier_block
*self
, unsigned long action
, void *hcpu
)
5306 unsigned int cpu
= (long)hcpu
;
5310 case CPU_UP_PREPARE
:
5311 case CPU_UP_PREPARE_FROZEN
:
5312 perf_event_init_cpu(cpu
);
5316 case CPU_ONLINE_FROZEN
:
5317 hw_perf_event_setup_online(cpu
);
5320 case CPU_DOWN_PREPARE
:
5321 case CPU_DOWN_PREPARE_FROZEN
:
5322 perf_event_exit_cpu(cpu
);
5333 * This has to have a higher priority than migration_notifier in sched.c.
5335 static struct notifier_block __cpuinitdata perf_cpu_nb
= {
5336 .notifier_call
= perf_cpu_notify
,
5340 void __init
perf_event_init(void)
5342 perf_event_init_all_cpus();
5343 perf_cpu_notify(&perf_cpu_nb
, (unsigned long)CPU_UP_PREPARE
,
5344 (void *)(long)smp_processor_id());
5345 perf_cpu_notify(&perf_cpu_nb
, (unsigned long)CPU_ONLINE
,
5346 (void *)(long)smp_processor_id());
5347 register_cpu_notifier(&perf_cpu_nb
);
5350 static ssize_t
perf_show_reserve_percpu(struct sysdev_class
*class, char *buf
)
5352 return sprintf(buf
, "%d\n", perf_reserved_percpu
);
5356 perf_set_reserve_percpu(struct sysdev_class
*class,
5360 struct perf_cpu_context
*cpuctx
;
5364 err
= strict_strtoul(buf
, 10, &val
);
5367 if (val
> perf_max_events
)
5370 spin_lock(&perf_resource_lock
);
5371 perf_reserved_percpu
= val
;
5372 for_each_online_cpu(cpu
) {
5373 cpuctx
= &per_cpu(perf_cpu_context
, cpu
);
5374 raw_spin_lock_irq(&cpuctx
->ctx
.lock
);
5375 mpt
= min(perf_max_events
- cpuctx
->ctx
.nr_events
,
5376 perf_max_events
- perf_reserved_percpu
);
5377 cpuctx
->max_pertask
= mpt
;
5378 raw_spin_unlock_irq(&cpuctx
->ctx
.lock
);
5380 spin_unlock(&perf_resource_lock
);
5385 static ssize_t
perf_show_overcommit(struct sysdev_class
*class, char *buf
)
5387 return sprintf(buf
, "%d\n", perf_overcommit
);
5391 perf_set_overcommit(struct sysdev_class
*class, const char *buf
, size_t count
)
5396 err
= strict_strtoul(buf
, 10, &val
);
5402 spin_lock(&perf_resource_lock
);
5403 perf_overcommit
= val
;
5404 spin_unlock(&perf_resource_lock
);
5409 static SYSDEV_CLASS_ATTR(
5412 perf_show_reserve_percpu
,
5413 perf_set_reserve_percpu
5416 static SYSDEV_CLASS_ATTR(
5419 perf_show_overcommit
,
5423 static struct attribute
*perfclass_attrs
[] = {
5424 &attr_reserve_percpu
.attr
,
5425 &attr_overcommit
.attr
,
5429 static struct attribute_group perfclass_attr_group
= {
5430 .attrs
= perfclass_attrs
,
5431 .name
= "perf_events",
5434 static int __init
perf_event_sysfs_init(void)
5436 return sysfs_create_group(&cpu_sysdev_class
.kset
.kobj
,
5437 &perfclass_attr_group
);
5439 device_initcall(perf_event_sysfs_init
);