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
;
1420 return div64_u64(dividend
, divisor
);
1423 static void perf_adjust_period(struct perf_event
*event
, u64 nsec
, u64 count
)
1425 struct hw_perf_event
*hwc
= &event
->hw
;
1426 u64 period
, sample_period
;
1429 period
= perf_calculate_period(event
, nsec
, count
);
1431 delta
= (s64
)(period
- hwc
->sample_period
);
1432 delta
= (delta
+ 7) / 8; /* low pass filter */
1434 sample_period
= hwc
->sample_period
+ delta
;
1439 hwc
->sample_period
= sample_period
;
1441 if (atomic64_read(&hwc
->period_left
) > 8*sample_period
) {
1443 event
->pmu
->disable(event
);
1444 atomic64_set(&hwc
->period_left
, 0);
1445 event
->pmu
->enable(event
);
1450 static void perf_ctx_adjust_freq(struct perf_event_context
*ctx
)
1452 struct perf_event
*event
;
1453 struct hw_perf_event
*hwc
;
1454 u64 interrupts
, now
;
1457 raw_spin_lock(&ctx
->lock
);
1458 list_for_each_entry_rcu(event
, &ctx
->event_list
, event_entry
) {
1459 if (event
->state
!= PERF_EVENT_STATE_ACTIVE
)
1462 if (event
->cpu
!= -1 && event
->cpu
!= smp_processor_id())
1467 interrupts
= hwc
->interrupts
;
1468 hwc
->interrupts
= 0;
1471 * unthrottle events on the tick
1473 if (interrupts
== MAX_INTERRUPTS
) {
1474 perf_log_throttle(event
, 1);
1475 event
->pmu
->unthrottle(event
);
1478 if (!event
->attr
.freq
|| !event
->attr
.sample_freq
)
1481 event
->pmu
->read(event
);
1482 now
= atomic64_read(&event
->count
);
1483 delta
= now
- hwc
->freq_count_stamp
;
1484 hwc
->freq_count_stamp
= now
;
1487 perf_adjust_period(event
, TICK_NSEC
, delta
);
1489 raw_spin_unlock(&ctx
->lock
);
1493 * Round-robin a context's events:
1495 static void rotate_ctx(struct perf_event_context
*ctx
)
1497 struct perf_event
*event
;
1499 if (!ctx
->nr_events
)
1502 raw_spin_lock(&ctx
->lock
);
1504 * Rotate the first entry last (works just fine for group events too):
1507 list_for_each_entry(event
, &ctx
->group_list
, group_entry
) {
1508 list_move_tail(&event
->group_entry
, &ctx
->group_list
);
1513 raw_spin_unlock(&ctx
->lock
);
1516 void perf_event_task_tick(struct task_struct
*curr
, int cpu
)
1518 struct perf_cpu_context
*cpuctx
;
1519 struct perf_event_context
*ctx
;
1521 if (!atomic_read(&nr_events
))
1524 cpuctx
= &per_cpu(perf_cpu_context
, cpu
);
1525 ctx
= curr
->perf_event_ctxp
;
1527 perf_ctx_adjust_freq(&cpuctx
->ctx
);
1529 perf_ctx_adjust_freq(ctx
);
1531 perf_event_cpu_sched_out(cpuctx
);
1533 __perf_event_task_sched_out(ctx
);
1535 rotate_ctx(&cpuctx
->ctx
);
1539 perf_event_cpu_sched_in(cpuctx
, cpu
);
1541 perf_event_task_sched_in(curr
, cpu
);
1545 * Enable all of a task's events that have been marked enable-on-exec.
1546 * This expects task == current.
1548 static void perf_event_enable_on_exec(struct task_struct
*task
)
1550 struct perf_event_context
*ctx
;
1551 struct perf_event
*event
;
1552 unsigned long flags
;
1555 local_irq_save(flags
);
1556 ctx
= task
->perf_event_ctxp
;
1557 if (!ctx
|| !ctx
->nr_events
)
1560 __perf_event_task_sched_out(ctx
);
1562 raw_spin_lock(&ctx
->lock
);
1564 list_for_each_entry(event
, &ctx
->group_list
, group_entry
) {
1565 if (!event
->attr
.enable_on_exec
)
1567 event
->attr
.enable_on_exec
= 0;
1568 if (event
->state
>= PERF_EVENT_STATE_INACTIVE
)
1570 __perf_event_mark_enabled(event
, ctx
);
1575 * Unclone this context if we enabled any event.
1580 raw_spin_unlock(&ctx
->lock
);
1582 perf_event_task_sched_in(task
, smp_processor_id());
1584 local_irq_restore(flags
);
1588 * Cross CPU call to read the hardware event
1590 static void __perf_event_read(void *info
)
1592 struct perf_cpu_context
*cpuctx
= &__get_cpu_var(perf_cpu_context
);
1593 struct perf_event
*event
= info
;
1594 struct perf_event_context
*ctx
= event
->ctx
;
1597 * If this is a task context, we need to check whether it is
1598 * the current task context of this cpu. If not it has been
1599 * scheduled out before the smp call arrived. In that case
1600 * event->count would have been updated to a recent sample
1601 * when the event was scheduled out.
1603 if (ctx
->task
&& cpuctx
->task_ctx
!= ctx
)
1606 raw_spin_lock(&ctx
->lock
);
1607 update_context_time(ctx
);
1608 update_event_times(event
);
1609 raw_spin_unlock(&ctx
->lock
);
1611 event
->pmu
->read(event
);
1614 static u64
perf_event_read(struct perf_event
*event
)
1617 * If event is enabled and currently active on a CPU, update the
1618 * value in the event structure:
1620 if (event
->state
== PERF_EVENT_STATE_ACTIVE
) {
1621 smp_call_function_single(event
->oncpu
,
1622 __perf_event_read
, event
, 1);
1623 } else if (event
->state
== PERF_EVENT_STATE_INACTIVE
) {
1624 struct perf_event_context
*ctx
= event
->ctx
;
1625 unsigned long flags
;
1627 raw_spin_lock_irqsave(&ctx
->lock
, flags
);
1628 update_context_time(ctx
);
1629 update_event_times(event
);
1630 raw_spin_unlock_irqrestore(&ctx
->lock
, flags
);
1633 return atomic64_read(&event
->count
);
1637 * Initialize the perf_event context in a task_struct:
1640 __perf_event_init_context(struct perf_event_context
*ctx
,
1641 struct task_struct
*task
)
1643 raw_spin_lock_init(&ctx
->lock
);
1644 mutex_init(&ctx
->mutex
);
1645 INIT_LIST_HEAD(&ctx
->group_list
);
1646 INIT_LIST_HEAD(&ctx
->event_list
);
1647 atomic_set(&ctx
->refcount
, 1);
1651 static struct perf_event_context
*find_get_context(pid_t pid
, int cpu
)
1653 struct perf_event_context
*ctx
;
1654 struct perf_cpu_context
*cpuctx
;
1655 struct task_struct
*task
;
1656 unsigned long flags
;
1659 if (pid
== -1 && cpu
!= -1) {
1660 /* Must be root to operate on a CPU event: */
1661 if (perf_paranoid_cpu() && !capable(CAP_SYS_ADMIN
))
1662 return ERR_PTR(-EACCES
);
1664 if (cpu
< 0 || cpu
>= nr_cpumask_bits
)
1665 return ERR_PTR(-EINVAL
);
1668 * We could be clever and allow to attach a event to an
1669 * offline CPU and activate it when the CPU comes up, but
1672 if (!cpu_online(cpu
))
1673 return ERR_PTR(-ENODEV
);
1675 cpuctx
= &per_cpu(perf_cpu_context
, cpu
);
1686 task
= find_task_by_vpid(pid
);
1688 get_task_struct(task
);
1692 return ERR_PTR(-ESRCH
);
1695 * Can't attach events to a dying task.
1698 if (task
->flags
& PF_EXITING
)
1701 /* Reuse ptrace permission checks for now. */
1703 if (!ptrace_may_access(task
, PTRACE_MODE_READ
))
1707 ctx
= perf_lock_task_context(task
, &flags
);
1710 raw_spin_unlock_irqrestore(&ctx
->lock
, flags
);
1714 ctx
= kzalloc(sizeof(struct perf_event_context
), GFP_KERNEL
);
1718 __perf_event_init_context(ctx
, task
);
1720 if (cmpxchg(&task
->perf_event_ctxp
, NULL
, ctx
)) {
1722 * We raced with some other task; use
1723 * the context they set.
1728 get_task_struct(task
);
1731 put_task_struct(task
);
1735 put_task_struct(task
);
1736 return ERR_PTR(err
);
1739 static void perf_event_free_filter(struct perf_event
*event
);
1741 static void free_event_rcu(struct rcu_head
*head
)
1743 struct perf_event
*event
;
1745 event
= container_of(head
, struct perf_event
, rcu_head
);
1747 put_pid_ns(event
->ns
);
1748 perf_event_free_filter(event
);
1752 static void perf_pending_sync(struct perf_event
*event
);
1754 static void free_event(struct perf_event
*event
)
1756 perf_pending_sync(event
);
1758 if (!event
->parent
) {
1759 atomic_dec(&nr_events
);
1760 if (event
->attr
.mmap
)
1761 atomic_dec(&nr_mmap_events
);
1762 if (event
->attr
.comm
)
1763 atomic_dec(&nr_comm_events
);
1764 if (event
->attr
.task
)
1765 atomic_dec(&nr_task_events
);
1768 if (event
->output
) {
1769 fput(event
->output
->filp
);
1770 event
->output
= NULL
;
1774 event
->destroy(event
);
1776 put_ctx(event
->ctx
);
1777 call_rcu(&event
->rcu_head
, free_event_rcu
);
1780 int perf_event_release_kernel(struct perf_event
*event
)
1782 struct perf_event_context
*ctx
= event
->ctx
;
1784 WARN_ON_ONCE(ctx
->parent_ctx
);
1785 mutex_lock(&ctx
->mutex
);
1786 perf_event_remove_from_context(event
);
1787 mutex_unlock(&ctx
->mutex
);
1789 mutex_lock(&event
->owner
->perf_event_mutex
);
1790 list_del_init(&event
->owner_entry
);
1791 mutex_unlock(&event
->owner
->perf_event_mutex
);
1792 put_task_struct(event
->owner
);
1798 EXPORT_SYMBOL_GPL(perf_event_release_kernel
);
1801 * Called when the last reference to the file is gone.
1803 static int perf_release(struct inode
*inode
, struct file
*file
)
1805 struct perf_event
*event
= file
->private_data
;
1807 file
->private_data
= NULL
;
1809 return perf_event_release_kernel(event
);
1812 static int perf_event_read_size(struct perf_event
*event
)
1814 int entry
= sizeof(u64
); /* value */
1818 if (event
->attr
.read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
)
1819 size
+= sizeof(u64
);
1821 if (event
->attr
.read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
)
1822 size
+= sizeof(u64
);
1824 if (event
->attr
.read_format
& PERF_FORMAT_ID
)
1825 entry
+= sizeof(u64
);
1827 if (event
->attr
.read_format
& PERF_FORMAT_GROUP
) {
1828 nr
+= event
->group_leader
->nr_siblings
;
1829 size
+= sizeof(u64
);
1837 u64
perf_event_read_value(struct perf_event
*event
, u64
*enabled
, u64
*running
)
1839 struct perf_event
*child
;
1845 mutex_lock(&event
->child_mutex
);
1846 total
+= perf_event_read(event
);
1847 *enabled
+= event
->total_time_enabled
+
1848 atomic64_read(&event
->child_total_time_enabled
);
1849 *running
+= event
->total_time_running
+
1850 atomic64_read(&event
->child_total_time_running
);
1852 list_for_each_entry(child
, &event
->child_list
, child_list
) {
1853 total
+= perf_event_read(child
);
1854 *enabled
+= child
->total_time_enabled
;
1855 *running
+= child
->total_time_running
;
1857 mutex_unlock(&event
->child_mutex
);
1861 EXPORT_SYMBOL_GPL(perf_event_read_value
);
1863 static int perf_event_read_group(struct perf_event
*event
,
1864 u64 read_format
, char __user
*buf
)
1866 struct perf_event
*leader
= event
->group_leader
, *sub
;
1867 int n
= 0, size
= 0, ret
= -EFAULT
;
1868 struct perf_event_context
*ctx
= leader
->ctx
;
1870 u64 count
, enabled
, running
;
1872 mutex_lock(&ctx
->mutex
);
1873 count
= perf_event_read_value(leader
, &enabled
, &running
);
1875 values
[n
++] = 1 + leader
->nr_siblings
;
1876 if (read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
)
1877 values
[n
++] = enabled
;
1878 if (read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
)
1879 values
[n
++] = running
;
1880 values
[n
++] = count
;
1881 if (read_format
& PERF_FORMAT_ID
)
1882 values
[n
++] = primary_event_id(leader
);
1884 size
= n
* sizeof(u64
);
1886 if (copy_to_user(buf
, values
, size
))
1891 list_for_each_entry(sub
, &leader
->sibling_list
, group_entry
) {
1894 values
[n
++] = perf_event_read_value(sub
, &enabled
, &running
);
1895 if (read_format
& PERF_FORMAT_ID
)
1896 values
[n
++] = primary_event_id(sub
);
1898 size
= n
* sizeof(u64
);
1900 if (copy_to_user(buf
+ ret
, values
, size
)) {
1908 mutex_unlock(&ctx
->mutex
);
1913 static int perf_event_read_one(struct perf_event
*event
,
1914 u64 read_format
, char __user
*buf
)
1916 u64 enabled
, running
;
1920 values
[n
++] = perf_event_read_value(event
, &enabled
, &running
);
1921 if (read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
)
1922 values
[n
++] = enabled
;
1923 if (read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
)
1924 values
[n
++] = running
;
1925 if (read_format
& PERF_FORMAT_ID
)
1926 values
[n
++] = primary_event_id(event
);
1928 if (copy_to_user(buf
, values
, n
* sizeof(u64
)))
1931 return n
* sizeof(u64
);
1935 * Read the performance event - simple non blocking version for now
1938 perf_read_hw(struct perf_event
*event
, char __user
*buf
, size_t count
)
1940 u64 read_format
= event
->attr
.read_format
;
1944 * Return end-of-file for a read on a event that is in
1945 * error state (i.e. because it was pinned but it couldn't be
1946 * scheduled on to the CPU at some point).
1948 if (event
->state
== PERF_EVENT_STATE_ERROR
)
1951 if (count
< perf_event_read_size(event
))
1954 WARN_ON_ONCE(event
->ctx
->parent_ctx
);
1955 if (read_format
& PERF_FORMAT_GROUP
)
1956 ret
= perf_event_read_group(event
, read_format
, buf
);
1958 ret
= perf_event_read_one(event
, read_format
, buf
);
1964 perf_read(struct file
*file
, char __user
*buf
, size_t count
, loff_t
*ppos
)
1966 struct perf_event
*event
= file
->private_data
;
1968 return perf_read_hw(event
, buf
, count
);
1971 static unsigned int perf_poll(struct file
*file
, poll_table
*wait
)
1973 struct perf_event
*event
= file
->private_data
;
1974 struct perf_mmap_data
*data
;
1975 unsigned int events
= POLL_HUP
;
1978 data
= rcu_dereference(event
->data
);
1980 events
= atomic_xchg(&data
->poll
, 0);
1983 poll_wait(file
, &event
->waitq
, wait
);
1988 static void perf_event_reset(struct perf_event
*event
)
1990 (void)perf_event_read(event
);
1991 atomic64_set(&event
->count
, 0);
1992 perf_event_update_userpage(event
);
1996 * Holding the top-level event's child_mutex means that any
1997 * descendant process that has inherited this event will block
1998 * in sync_child_event if it goes to exit, thus satisfying the
1999 * task existence requirements of perf_event_enable/disable.
2001 static void perf_event_for_each_child(struct perf_event
*event
,
2002 void (*func
)(struct perf_event
*))
2004 struct perf_event
*child
;
2006 WARN_ON_ONCE(event
->ctx
->parent_ctx
);
2007 mutex_lock(&event
->child_mutex
);
2009 list_for_each_entry(child
, &event
->child_list
, child_list
)
2011 mutex_unlock(&event
->child_mutex
);
2014 static void perf_event_for_each(struct perf_event
*event
,
2015 void (*func
)(struct perf_event
*))
2017 struct perf_event_context
*ctx
= event
->ctx
;
2018 struct perf_event
*sibling
;
2020 WARN_ON_ONCE(ctx
->parent_ctx
);
2021 mutex_lock(&ctx
->mutex
);
2022 event
= event
->group_leader
;
2024 perf_event_for_each_child(event
, func
);
2026 list_for_each_entry(sibling
, &event
->sibling_list
, group_entry
)
2027 perf_event_for_each_child(event
, func
);
2028 mutex_unlock(&ctx
->mutex
);
2031 static int perf_event_period(struct perf_event
*event
, u64 __user
*arg
)
2033 struct perf_event_context
*ctx
= event
->ctx
;
2038 if (!event
->attr
.sample_period
)
2041 size
= copy_from_user(&value
, arg
, sizeof(value
));
2042 if (size
!= sizeof(value
))
2048 raw_spin_lock_irq(&ctx
->lock
);
2049 if (event
->attr
.freq
) {
2050 if (value
> sysctl_perf_event_sample_rate
) {
2055 event
->attr
.sample_freq
= value
;
2057 event
->attr
.sample_period
= value
;
2058 event
->hw
.sample_period
= value
;
2061 raw_spin_unlock_irq(&ctx
->lock
);
2066 static int perf_event_set_output(struct perf_event
*event
, int output_fd
);
2067 static int perf_event_set_filter(struct perf_event
*event
, void __user
*arg
);
2069 static long perf_ioctl(struct file
*file
, unsigned int cmd
, unsigned long arg
)
2071 struct perf_event
*event
= file
->private_data
;
2072 void (*func
)(struct perf_event
*);
2076 case PERF_EVENT_IOC_ENABLE
:
2077 func
= perf_event_enable
;
2079 case PERF_EVENT_IOC_DISABLE
:
2080 func
= perf_event_disable
;
2082 case PERF_EVENT_IOC_RESET
:
2083 func
= perf_event_reset
;
2086 case PERF_EVENT_IOC_REFRESH
:
2087 return perf_event_refresh(event
, arg
);
2089 case PERF_EVENT_IOC_PERIOD
:
2090 return perf_event_period(event
, (u64 __user
*)arg
);
2092 case PERF_EVENT_IOC_SET_OUTPUT
:
2093 return perf_event_set_output(event
, arg
);
2095 case PERF_EVENT_IOC_SET_FILTER
:
2096 return perf_event_set_filter(event
, (void __user
*)arg
);
2102 if (flags
& PERF_IOC_FLAG_GROUP
)
2103 perf_event_for_each(event
, func
);
2105 perf_event_for_each_child(event
, func
);
2110 int perf_event_task_enable(void)
2112 struct perf_event
*event
;
2114 mutex_lock(¤t
->perf_event_mutex
);
2115 list_for_each_entry(event
, ¤t
->perf_event_list
, owner_entry
)
2116 perf_event_for_each_child(event
, perf_event_enable
);
2117 mutex_unlock(¤t
->perf_event_mutex
);
2122 int perf_event_task_disable(void)
2124 struct perf_event
*event
;
2126 mutex_lock(¤t
->perf_event_mutex
);
2127 list_for_each_entry(event
, ¤t
->perf_event_list
, owner_entry
)
2128 perf_event_for_each_child(event
, perf_event_disable
);
2129 mutex_unlock(¤t
->perf_event_mutex
);
2134 #ifndef PERF_EVENT_INDEX_OFFSET
2135 # define PERF_EVENT_INDEX_OFFSET 0
2138 static int perf_event_index(struct perf_event
*event
)
2140 if (event
->state
!= PERF_EVENT_STATE_ACTIVE
)
2143 return event
->hw
.idx
+ 1 - PERF_EVENT_INDEX_OFFSET
;
2147 * Callers need to ensure there can be no nesting of this function, otherwise
2148 * the seqlock logic goes bad. We can not serialize this because the arch
2149 * code calls this from NMI context.
2151 void perf_event_update_userpage(struct perf_event
*event
)
2153 struct perf_event_mmap_page
*userpg
;
2154 struct perf_mmap_data
*data
;
2157 data
= rcu_dereference(event
->data
);
2161 userpg
= data
->user_page
;
2164 * Disable preemption so as to not let the corresponding user-space
2165 * spin too long if we get preempted.
2170 userpg
->index
= perf_event_index(event
);
2171 userpg
->offset
= atomic64_read(&event
->count
);
2172 if (event
->state
== PERF_EVENT_STATE_ACTIVE
)
2173 userpg
->offset
-= atomic64_read(&event
->hw
.prev_count
);
2175 userpg
->time_enabled
= event
->total_time_enabled
+
2176 atomic64_read(&event
->child_total_time_enabled
);
2178 userpg
->time_running
= event
->total_time_running
+
2179 atomic64_read(&event
->child_total_time_running
);
2188 static unsigned long perf_data_size(struct perf_mmap_data
*data
)
2190 return data
->nr_pages
<< (PAGE_SHIFT
+ data
->data_order
);
2193 #ifndef CONFIG_PERF_USE_VMALLOC
2196 * Back perf_mmap() with regular GFP_KERNEL-0 pages.
2199 static struct page
*
2200 perf_mmap_to_page(struct perf_mmap_data
*data
, unsigned long pgoff
)
2202 if (pgoff
> data
->nr_pages
)
2206 return virt_to_page(data
->user_page
);
2208 return virt_to_page(data
->data_pages
[pgoff
- 1]);
2211 static struct perf_mmap_data
*
2212 perf_mmap_data_alloc(struct perf_event
*event
, int nr_pages
)
2214 struct perf_mmap_data
*data
;
2218 WARN_ON(atomic_read(&event
->mmap_count
));
2220 size
= sizeof(struct perf_mmap_data
);
2221 size
+= nr_pages
* sizeof(void *);
2223 data
= kzalloc(size
, GFP_KERNEL
);
2227 data
->user_page
= (void *)get_zeroed_page(GFP_KERNEL
);
2228 if (!data
->user_page
)
2229 goto fail_user_page
;
2231 for (i
= 0; i
< nr_pages
; i
++) {
2232 data
->data_pages
[i
] = (void *)get_zeroed_page(GFP_KERNEL
);
2233 if (!data
->data_pages
[i
])
2234 goto fail_data_pages
;
2237 data
->data_order
= 0;
2238 data
->nr_pages
= nr_pages
;
2243 for (i
--; i
>= 0; i
--)
2244 free_page((unsigned long)data
->data_pages
[i
]);
2246 free_page((unsigned long)data
->user_page
);
2255 static void perf_mmap_free_page(unsigned long addr
)
2257 struct page
*page
= virt_to_page((void *)addr
);
2259 page
->mapping
= NULL
;
2263 static void perf_mmap_data_free(struct perf_mmap_data
*data
)
2267 perf_mmap_free_page((unsigned long)data
->user_page
);
2268 for (i
= 0; i
< data
->nr_pages
; i
++)
2269 perf_mmap_free_page((unsigned long)data
->data_pages
[i
]);
2276 * Back perf_mmap() with vmalloc memory.
2278 * Required for architectures that have d-cache aliasing issues.
2281 static struct page
*
2282 perf_mmap_to_page(struct perf_mmap_data
*data
, unsigned long pgoff
)
2284 if (pgoff
> (1UL << data
->data_order
))
2287 return vmalloc_to_page((void *)data
->user_page
+ pgoff
* PAGE_SIZE
);
2290 static void perf_mmap_unmark_page(void *addr
)
2292 struct page
*page
= vmalloc_to_page(addr
);
2294 page
->mapping
= NULL
;
2297 static void perf_mmap_data_free_work(struct work_struct
*work
)
2299 struct perf_mmap_data
*data
;
2303 data
= container_of(work
, struct perf_mmap_data
, work
);
2304 nr
= 1 << data
->data_order
;
2306 base
= data
->user_page
;
2307 for (i
= 0; i
< nr
+ 1; i
++)
2308 perf_mmap_unmark_page(base
+ (i
* PAGE_SIZE
));
2314 static void perf_mmap_data_free(struct perf_mmap_data
*data
)
2316 schedule_work(&data
->work
);
2319 static struct perf_mmap_data
*
2320 perf_mmap_data_alloc(struct perf_event
*event
, int nr_pages
)
2322 struct perf_mmap_data
*data
;
2326 WARN_ON(atomic_read(&event
->mmap_count
));
2328 size
= sizeof(struct perf_mmap_data
);
2329 size
+= sizeof(void *);
2331 data
= kzalloc(size
, GFP_KERNEL
);
2335 INIT_WORK(&data
->work
, perf_mmap_data_free_work
);
2337 all_buf
= vmalloc_user((nr_pages
+ 1) * PAGE_SIZE
);
2341 data
->user_page
= all_buf
;
2342 data
->data_pages
[0] = all_buf
+ PAGE_SIZE
;
2343 data
->data_order
= ilog2(nr_pages
);
2357 static int perf_mmap_fault(struct vm_area_struct
*vma
, struct vm_fault
*vmf
)
2359 struct perf_event
*event
= vma
->vm_file
->private_data
;
2360 struct perf_mmap_data
*data
;
2361 int ret
= VM_FAULT_SIGBUS
;
2363 if (vmf
->flags
& FAULT_FLAG_MKWRITE
) {
2364 if (vmf
->pgoff
== 0)
2370 data
= rcu_dereference(event
->data
);
2374 if (vmf
->pgoff
&& (vmf
->flags
& FAULT_FLAG_WRITE
))
2377 vmf
->page
= perf_mmap_to_page(data
, vmf
->pgoff
);
2381 get_page(vmf
->page
);
2382 vmf
->page
->mapping
= vma
->vm_file
->f_mapping
;
2383 vmf
->page
->index
= vmf
->pgoff
;
2393 perf_mmap_data_init(struct perf_event
*event
, struct perf_mmap_data
*data
)
2395 long max_size
= perf_data_size(data
);
2397 atomic_set(&data
->lock
, -1);
2399 if (event
->attr
.watermark
) {
2400 data
->watermark
= min_t(long, max_size
,
2401 event
->attr
.wakeup_watermark
);
2404 if (!data
->watermark
)
2405 data
->watermark
= max_size
/ 2;
2408 rcu_assign_pointer(event
->data
, data
);
2411 static void perf_mmap_data_free_rcu(struct rcu_head
*rcu_head
)
2413 struct perf_mmap_data
*data
;
2415 data
= container_of(rcu_head
, struct perf_mmap_data
, rcu_head
);
2416 perf_mmap_data_free(data
);
2419 static void perf_mmap_data_release(struct perf_event
*event
)
2421 struct perf_mmap_data
*data
= event
->data
;
2423 WARN_ON(atomic_read(&event
->mmap_count
));
2425 rcu_assign_pointer(event
->data
, NULL
);
2426 call_rcu(&data
->rcu_head
, perf_mmap_data_free_rcu
);
2429 static void perf_mmap_open(struct vm_area_struct
*vma
)
2431 struct perf_event
*event
= vma
->vm_file
->private_data
;
2433 atomic_inc(&event
->mmap_count
);
2436 static void perf_mmap_close(struct vm_area_struct
*vma
)
2438 struct perf_event
*event
= vma
->vm_file
->private_data
;
2440 WARN_ON_ONCE(event
->ctx
->parent_ctx
);
2441 if (atomic_dec_and_mutex_lock(&event
->mmap_count
, &event
->mmap_mutex
)) {
2442 unsigned long size
= perf_data_size(event
->data
);
2443 struct user_struct
*user
= current_user();
2445 atomic_long_sub((size
>> PAGE_SHIFT
) + 1, &user
->locked_vm
);
2446 vma
->vm_mm
->locked_vm
-= event
->data
->nr_locked
;
2447 perf_mmap_data_release(event
);
2448 mutex_unlock(&event
->mmap_mutex
);
2452 static const struct vm_operations_struct perf_mmap_vmops
= {
2453 .open
= perf_mmap_open
,
2454 .close
= perf_mmap_close
,
2455 .fault
= perf_mmap_fault
,
2456 .page_mkwrite
= perf_mmap_fault
,
2459 static int perf_mmap(struct file
*file
, struct vm_area_struct
*vma
)
2461 struct perf_event
*event
= file
->private_data
;
2462 unsigned long user_locked
, user_lock_limit
;
2463 struct user_struct
*user
= current_user();
2464 unsigned long locked
, lock_limit
;
2465 struct perf_mmap_data
*data
;
2466 unsigned long vma_size
;
2467 unsigned long nr_pages
;
2468 long user_extra
, extra
;
2471 if (!(vma
->vm_flags
& VM_SHARED
))
2474 vma_size
= vma
->vm_end
- vma
->vm_start
;
2475 nr_pages
= (vma_size
/ PAGE_SIZE
) - 1;
2478 * If we have data pages ensure they're a power-of-two number, so we
2479 * can do bitmasks instead of modulo.
2481 if (nr_pages
!= 0 && !is_power_of_2(nr_pages
))
2484 if (vma_size
!= PAGE_SIZE
* (1 + nr_pages
))
2487 if (vma
->vm_pgoff
!= 0)
2490 WARN_ON_ONCE(event
->ctx
->parent_ctx
);
2491 mutex_lock(&event
->mmap_mutex
);
2492 if (event
->output
) {
2497 if (atomic_inc_not_zero(&event
->mmap_count
)) {
2498 if (nr_pages
!= event
->data
->nr_pages
)
2503 user_extra
= nr_pages
+ 1;
2504 user_lock_limit
= sysctl_perf_event_mlock
>> (PAGE_SHIFT
- 10);
2507 * Increase the limit linearly with more CPUs:
2509 user_lock_limit
*= num_online_cpus();
2511 user_locked
= atomic_long_read(&user
->locked_vm
) + user_extra
;
2514 if (user_locked
> user_lock_limit
)
2515 extra
= user_locked
- user_lock_limit
;
2517 lock_limit
= current
->signal
->rlim
[RLIMIT_MEMLOCK
].rlim_cur
;
2518 lock_limit
>>= PAGE_SHIFT
;
2519 locked
= vma
->vm_mm
->locked_vm
+ extra
;
2521 if ((locked
> lock_limit
) && perf_paranoid_tracepoint_raw() &&
2522 !capable(CAP_IPC_LOCK
)) {
2527 WARN_ON(event
->data
);
2529 data
= perf_mmap_data_alloc(event
, nr_pages
);
2535 perf_mmap_data_init(event
, data
);
2537 atomic_set(&event
->mmap_count
, 1);
2538 atomic_long_add(user_extra
, &user
->locked_vm
);
2539 vma
->vm_mm
->locked_vm
+= extra
;
2540 event
->data
->nr_locked
= extra
;
2541 if (vma
->vm_flags
& VM_WRITE
)
2542 event
->data
->writable
= 1;
2545 mutex_unlock(&event
->mmap_mutex
);
2547 vma
->vm_flags
|= VM_RESERVED
;
2548 vma
->vm_ops
= &perf_mmap_vmops
;
2553 static int perf_fasync(int fd
, struct file
*filp
, int on
)
2555 struct inode
*inode
= filp
->f_path
.dentry
->d_inode
;
2556 struct perf_event
*event
= filp
->private_data
;
2559 mutex_lock(&inode
->i_mutex
);
2560 retval
= fasync_helper(fd
, filp
, on
, &event
->fasync
);
2561 mutex_unlock(&inode
->i_mutex
);
2569 static const struct file_operations perf_fops
= {
2570 .release
= perf_release
,
2573 .unlocked_ioctl
= perf_ioctl
,
2574 .compat_ioctl
= perf_ioctl
,
2576 .fasync
= perf_fasync
,
2582 * If there's data, ensure we set the poll() state and publish everything
2583 * to user-space before waking everybody up.
2586 void perf_event_wakeup(struct perf_event
*event
)
2588 wake_up_all(&event
->waitq
);
2590 if (event
->pending_kill
) {
2591 kill_fasync(&event
->fasync
, SIGIO
, event
->pending_kill
);
2592 event
->pending_kill
= 0;
2599 * Handle the case where we need to wakeup up from NMI (or rq->lock) context.
2601 * The NMI bit means we cannot possibly take locks. Therefore, maintain a
2602 * single linked list and use cmpxchg() to add entries lockless.
2605 static void perf_pending_event(struct perf_pending_entry
*entry
)
2607 struct perf_event
*event
= container_of(entry
,
2608 struct perf_event
, pending
);
2610 if (event
->pending_disable
) {
2611 event
->pending_disable
= 0;
2612 __perf_event_disable(event
);
2615 if (event
->pending_wakeup
) {
2616 event
->pending_wakeup
= 0;
2617 perf_event_wakeup(event
);
2621 #define PENDING_TAIL ((struct perf_pending_entry *)-1UL)
2623 static DEFINE_PER_CPU(struct perf_pending_entry
*, perf_pending_head
) = {
2627 static void perf_pending_queue(struct perf_pending_entry
*entry
,
2628 void (*func
)(struct perf_pending_entry
*))
2630 struct perf_pending_entry
**head
;
2632 if (cmpxchg(&entry
->next
, NULL
, PENDING_TAIL
) != NULL
)
2637 head
= &get_cpu_var(perf_pending_head
);
2640 entry
->next
= *head
;
2641 } while (cmpxchg(head
, entry
->next
, entry
) != entry
->next
);
2643 set_perf_event_pending();
2645 put_cpu_var(perf_pending_head
);
2648 static int __perf_pending_run(void)
2650 struct perf_pending_entry
*list
;
2653 list
= xchg(&__get_cpu_var(perf_pending_head
), PENDING_TAIL
);
2654 while (list
!= PENDING_TAIL
) {
2655 void (*func
)(struct perf_pending_entry
*);
2656 struct perf_pending_entry
*entry
= list
;
2663 * Ensure we observe the unqueue before we issue the wakeup,
2664 * so that we won't be waiting forever.
2665 * -- see perf_not_pending().
2676 static inline int perf_not_pending(struct perf_event
*event
)
2679 * If we flush on whatever cpu we run, there is a chance we don't
2683 __perf_pending_run();
2687 * Ensure we see the proper queue state before going to sleep
2688 * so that we do not miss the wakeup. -- see perf_pending_handle()
2691 return event
->pending
.next
== NULL
;
2694 static void perf_pending_sync(struct perf_event
*event
)
2696 wait_event(event
->waitq
, perf_not_pending(event
));
2699 void perf_event_do_pending(void)
2701 __perf_pending_run();
2705 * Callchain support -- arch specific
2708 __weak
struct perf_callchain_entry
*perf_callchain(struct pt_regs
*regs
)
2716 static bool perf_output_space(struct perf_mmap_data
*data
, unsigned long tail
,
2717 unsigned long offset
, unsigned long head
)
2721 if (!data
->writable
)
2724 mask
= perf_data_size(data
) - 1;
2726 offset
= (offset
- tail
) & mask
;
2727 head
= (head
- tail
) & mask
;
2729 if ((int)(head
- offset
) < 0)
2735 static void perf_output_wakeup(struct perf_output_handle
*handle
)
2737 atomic_set(&handle
->data
->poll
, POLL_IN
);
2740 handle
->event
->pending_wakeup
= 1;
2741 perf_pending_queue(&handle
->event
->pending
,
2742 perf_pending_event
);
2744 perf_event_wakeup(handle
->event
);
2748 * Curious locking construct.
2750 * We need to ensure a later event_id doesn't publish a head when a former
2751 * event_id isn't done writing. However since we need to deal with NMIs we
2752 * cannot fully serialize things.
2754 * What we do is serialize between CPUs so we only have to deal with NMI
2755 * nesting on a single CPU.
2757 * We only publish the head (and generate a wakeup) when the outer-most
2758 * event_id completes.
2760 static void perf_output_lock(struct perf_output_handle
*handle
)
2762 struct perf_mmap_data
*data
= handle
->data
;
2763 int cur
, cpu
= get_cpu();
2768 cur
= atomic_cmpxchg(&data
->lock
, -1, cpu
);
2780 static void perf_output_unlock(struct perf_output_handle
*handle
)
2782 struct perf_mmap_data
*data
= handle
->data
;
2786 data
->done_head
= data
->head
;
2788 if (!handle
->locked
)
2793 * The xchg implies a full barrier that ensures all writes are done
2794 * before we publish the new head, matched by a rmb() in userspace when
2795 * reading this position.
2797 while ((head
= atomic_long_xchg(&data
->done_head
, 0)))
2798 data
->user_page
->data_head
= head
;
2801 * NMI can happen here, which means we can miss a done_head update.
2804 cpu
= atomic_xchg(&data
->lock
, -1);
2805 WARN_ON_ONCE(cpu
!= smp_processor_id());
2808 * Therefore we have to validate we did not indeed do so.
2810 if (unlikely(atomic_long_read(&data
->done_head
))) {
2812 * Since we had it locked, we can lock it again.
2814 while (atomic_cmpxchg(&data
->lock
, -1, cpu
) != -1)
2820 if (atomic_xchg(&data
->wakeup
, 0))
2821 perf_output_wakeup(handle
);
2826 void perf_output_copy(struct perf_output_handle
*handle
,
2827 const void *buf
, unsigned int len
)
2829 unsigned int pages_mask
;
2830 unsigned long offset
;
2834 offset
= handle
->offset
;
2835 pages_mask
= handle
->data
->nr_pages
- 1;
2836 pages
= handle
->data
->data_pages
;
2839 unsigned long page_offset
;
2840 unsigned long page_size
;
2843 nr
= (offset
>> PAGE_SHIFT
) & pages_mask
;
2844 page_size
= 1UL << (handle
->data
->data_order
+ PAGE_SHIFT
);
2845 page_offset
= offset
& (page_size
- 1);
2846 size
= min_t(unsigned int, page_size
- page_offset
, len
);
2848 memcpy(pages
[nr
] + page_offset
, buf
, size
);
2855 handle
->offset
= offset
;
2858 * Check we didn't copy past our reservation window, taking the
2859 * possible unsigned int wrap into account.
2861 WARN_ON_ONCE(((long)(handle
->head
- handle
->offset
)) < 0);
2864 int perf_output_begin(struct perf_output_handle
*handle
,
2865 struct perf_event
*event
, unsigned int size
,
2866 int nmi
, int sample
)
2868 struct perf_event
*output_event
;
2869 struct perf_mmap_data
*data
;
2870 unsigned long tail
, offset
, head
;
2873 struct perf_event_header header
;
2880 * For inherited events we send all the output towards the parent.
2883 event
= event
->parent
;
2885 output_event
= rcu_dereference(event
->output
);
2887 event
= output_event
;
2889 data
= rcu_dereference(event
->data
);
2893 handle
->data
= data
;
2894 handle
->event
= event
;
2896 handle
->sample
= sample
;
2898 if (!data
->nr_pages
)
2901 have_lost
= atomic_read(&data
->lost
);
2903 size
+= sizeof(lost_event
);
2905 perf_output_lock(handle
);
2909 * Userspace could choose to issue a mb() before updating the
2910 * tail pointer. So that all reads will be completed before the
2913 tail
= ACCESS_ONCE(data
->user_page
->data_tail
);
2915 offset
= head
= atomic_long_read(&data
->head
);
2917 if (unlikely(!perf_output_space(data
, tail
, offset
, head
)))
2919 } while (atomic_long_cmpxchg(&data
->head
, offset
, head
) != offset
);
2921 handle
->offset
= offset
;
2922 handle
->head
= head
;
2924 if (head
- tail
> data
->watermark
)
2925 atomic_set(&data
->wakeup
, 1);
2928 lost_event
.header
.type
= PERF_RECORD_LOST
;
2929 lost_event
.header
.misc
= 0;
2930 lost_event
.header
.size
= sizeof(lost_event
);
2931 lost_event
.id
= event
->id
;
2932 lost_event
.lost
= atomic_xchg(&data
->lost
, 0);
2934 perf_output_put(handle
, lost_event
);
2940 atomic_inc(&data
->lost
);
2941 perf_output_unlock(handle
);
2948 void perf_output_end(struct perf_output_handle
*handle
)
2950 struct perf_event
*event
= handle
->event
;
2951 struct perf_mmap_data
*data
= handle
->data
;
2953 int wakeup_events
= event
->attr
.wakeup_events
;
2955 if (handle
->sample
&& wakeup_events
) {
2956 int events
= atomic_inc_return(&data
->events
);
2957 if (events
>= wakeup_events
) {
2958 atomic_sub(wakeup_events
, &data
->events
);
2959 atomic_set(&data
->wakeup
, 1);
2963 perf_output_unlock(handle
);
2967 static u32
perf_event_pid(struct perf_event
*event
, struct task_struct
*p
)
2970 * only top level events have the pid namespace they were created in
2973 event
= event
->parent
;
2975 return task_tgid_nr_ns(p
, event
->ns
);
2978 static u32
perf_event_tid(struct perf_event
*event
, struct task_struct
*p
)
2981 * only top level events have the pid namespace they were created in
2984 event
= event
->parent
;
2986 return task_pid_nr_ns(p
, event
->ns
);
2989 static void perf_output_read_one(struct perf_output_handle
*handle
,
2990 struct perf_event
*event
)
2992 u64 read_format
= event
->attr
.read_format
;
2996 values
[n
++] = atomic64_read(&event
->count
);
2997 if (read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
) {
2998 values
[n
++] = event
->total_time_enabled
+
2999 atomic64_read(&event
->child_total_time_enabled
);
3001 if (read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
) {
3002 values
[n
++] = event
->total_time_running
+
3003 atomic64_read(&event
->child_total_time_running
);
3005 if (read_format
& PERF_FORMAT_ID
)
3006 values
[n
++] = primary_event_id(event
);
3008 perf_output_copy(handle
, values
, n
* sizeof(u64
));
3012 * XXX PERF_FORMAT_GROUP vs inherited events seems difficult.
3014 static void perf_output_read_group(struct perf_output_handle
*handle
,
3015 struct perf_event
*event
)
3017 struct perf_event
*leader
= event
->group_leader
, *sub
;
3018 u64 read_format
= event
->attr
.read_format
;
3022 values
[n
++] = 1 + leader
->nr_siblings
;
3024 if (read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
)
3025 values
[n
++] = leader
->total_time_enabled
;
3027 if (read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
)
3028 values
[n
++] = leader
->total_time_running
;
3030 if (leader
!= event
)
3031 leader
->pmu
->read(leader
);
3033 values
[n
++] = atomic64_read(&leader
->count
);
3034 if (read_format
& PERF_FORMAT_ID
)
3035 values
[n
++] = primary_event_id(leader
);
3037 perf_output_copy(handle
, values
, n
* sizeof(u64
));
3039 list_for_each_entry(sub
, &leader
->sibling_list
, group_entry
) {
3043 sub
->pmu
->read(sub
);
3045 values
[n
++] = atomic64_read(&sub
->count
);
3046 if (read_format
& PERF_FORMAT_ID
)
3047 values
[n
++] = primary_event_id(sub
);
3049 perf_output_copy(handle
, values
, n
* sizeof(u64
));
3053 static void perf_output_read(struct perf_output_handle
*handle
,
3054 struct perf_event
*event
)
3056 if (event
->attr
.read_format
& PERF_FORMAT_GROUP
)
3057 perf_output_read_group(handle
, event
);
3059 perf_output_read_one(handle
, event
);
3062 void perf_output_sample(struct perf_output_handle
*handle
,
3063 struct perf_event_header
*header
,
3064 struct perf_sample_data
*data
,
3065 struct perf_event
*event
)
3067 u64 sample_type
= data
->type
;
3069 perf_output_put(handle
, *header
);
3071 if (sample_type
& PERF_SAMPLE_IP
)
3072 perf_output_put(handle
, data
->ip
);
3074 if (sample_type
& PERF_SAMPLE_TID
)
3075 perf_output_put(handle
, data
->tid_entry
);
3077 if (sample_type
& PERF_SAMPLE_TIME
)
3078 perf_output_put(handle
, data
->time
);
3080 if (sample_type
& PERF_SAMPLE_ADDR
)
3081 perf_output_put(handle
, data
->addr
);
3083 if (sample_type
& PERF_SAMPLE_ID
)
3084 perf_output_put(handle
, data
->id
);
3086 if (sample_type
& PERF_SAMPLE_STREAM_ID
)
3087 perf_output_put(handle
, data
->stream_id
);
3089 if (sample_type
& PERF_SAMPLE_CPU
)
3090 perf_output_put(handle
, data
->cpu_entry
);
3092 if (sample_type
& PERF_SAMPLE_PERIOD
)
3093 perf_output_put(handle
, data
->period
);
3095 if (sample_type
& PERF_SAMPLE_READ
)
3096 perf_output_read(handle
, event
);
3098 if (sample_type
& PERF_SAMPLE_CALLCHAIN
) {
3099 if (data
->callchain
) {
3102 if (data
->callchain
)
3103 size
+= data
->callchain
->nr
;
3105 size
*= sizeof(u64
);
3107 perf_output_copy(handle
, data
->callchain
, size
);
3110 perf_output_put(handle
, nr
);
3114 if (sample_type
& PERF_SAMPLE_RAW
) {
3116 perf_output_put(handle
, data
->raw
->size
);
3117 perf_output_copy(handle
, data
->raw
->data
,
3124 .size
= sizeof(u32
),
3127 perf_output_put(handle
, raw
);
3132 void perf_prepare_sample(struct perf_event_header
*header
,
3133 struct perf_sample_data
*data
,
3134 struct perf_event
*event
,
3135 struct pt_regs
*regs
)
3137 u64 sample_type
= event
->attr
.sample_type
;
3139 data
->type
= sample_type
;
3141 header
->type
= PERF_RECORD_SAMPLE
;
3142 header
->size
= sizeof(*header
);
3145 header
->misc
|= perf_misc_flags(regs
);
3147 if (sample_type
& PERF_SAMPLE_IP
) {
3148 data
->ip
= perf_instruction_pointer(regs
);
3150 header
->size
+= sizeof(data
->ip
);
3153 if (sample_type
& PERF_SAMPLE_TID
) {
3154 /* namespace issues */
3155 data
->tid_entry
.pid
= perf_event_pid(event
, current
);
3156 data
->tid_entry
.tid
= perf_event_tid(event
, current
);
3158 header
->size
+= sizeof(data
->tid_entry
);
3161 if (sample_type
& PERF_SAMPLE_TIME
) {
3162 data
->time
= perf_clock();
3164 header
->size
+= sizeof(data
->time
);
3167 if (sample_type
& PERF_SAMPLE_ADDR
)
3168 header
->size
+= sizeof(data
->addr
);
3170 if (sample_type
& PERF_SAMPLE_ID
) {
3171 data
->id
= primary_event_id(event
);
3173 header
->size
+= sizeof(data
->id
);
3176 if (sample_type
& PERF_SAMPLE_STREAM_ID
) {
3177 data
->stream_id
= event
->id
;
3179 header
->size
+= sizeof(data
->stream_id
);
3182 if (sample_type
& PERF_SAMPLE_CPU
) {
3183 data
->cpu_entry
.cpu
= raw_smp_processor_id();
3184 data
->cpu_entry
.reserved
= 0;
3186 header
->size
+= sizeof(data
->cpu_entry
);
3189 if (sample_type
& PERF_SAMPLE_PERIOD
)
3190 header
->size
+= sizeof(data
->period
);
3192 if (sample_type
& PERF_SAMPLE_READ
)
3193 header
->size
+= perf_event_read_size(event
);
3195 if (sample_type
& PERF_SAMPLE_CALLCHAIN
) {
3198 data
->callchain
= perf_callchain(regs
);
3200 if (data
->callchain
)
3201 size
+= data
->callchain
->nr
;
3203 header
->size
+= size
* sizeof(u64
);
3206 if (sample_type
& PERF_SAMPLE_RAW
) {
3207 int size
= sizeof(u32
);
3210 size
+= data
->raw
->size
;
3212 size
+= sizeof(u32
);
3214 WARN_ON_ONCE(size
& (sizeof(u64
)-1));
3215 header
->size
+= size
;
3219 static void perf_event_output(struct perf_event
*event
, int nmi
,
3220 struct perf_sample_data
*data
,
3221 struct pt_regs
*regs
)
3223 struct perf_output_handle handle
;
3224 struct perf_event_header header
;
3226 perf_prepare_sample(&header
, data
, event
, regs
);
3228 if (perf_output_begin(&handle
, event
, header
.size
, nmi
, 1))
3231 perf_output_sample(&handle
, &header
, data
, event
);
3233 perf_output_end(&handle
);
3240 struct perf_read_event
{
3241 struct perf_event_header header
;
3248 perf_event_read_event(struct perf_event
*event
,
3249 struct task_struct
*task
)
3251 struct perf_output_handle handle
;
3252 struct perf_read_event read_event
= {
3254 .type
= PERF_RECORD_READ
,
3256 .size
= sizeof(read_event
) + perf_event_read_size(event
),
3258 .pid
= perf_event_pid(event
, task
),
3259 .tid
= perf_event_tid(event
, task
),
3263 ret
= perf_output_begin(&handle
, event
, read_event
.header
.size
, 0, 0);
3267 perf_output_put(&handle
, read_event
);
3268 perf_output_read(&handle
, event
);
3270 perf_output_end(&handle
);
3274 * task tracking -- fork/exit
3276 * enabled by: attr.comm | attr.mmap | attr.task
3279 struct perf_task_event
{
3280 struct task_struct
*task
;
3281 struct perf_event_context
*task_ctx
;
3284 struct perf_event_header header
;
3294 static void perf_event_task_output(struct perf_event
*event
,
3295 struct perf_task_event
*task_event
)
3297 struct perf_output_handle handle
;
3299 struct task_struct
*task
= task_event
->task
;
3302 size
= task_event
->event_id
.header
.size
;
3303 ret
= perf_output_begin(&handle
, event
, size
, 0, 0);
3308 task_event
->event_id
.pid
= perf_event_pid(event
, task
);
3309 task_event
->event_id
.ppid
= perf_event_pid(event
, current
);
3311 task_event
->event_id
.tid
= perf_event_tid(event
, task
);
3312 task_event
->event_id
.ptid
= perf_event_tid(event
, current
);
3314 perf_output_put(&handle
, task_event
->event_id
);
3316 perf_output_end(&handle
);
3319 static int perf_event_task_match(struct perf_event
*event
)
3321 if (event
->state
< PERF_EVENT_STATE_INACTIVE
)
3324 if (event
->cpu
!= -1 && event
->cpu
!= smp_processor_id())
3327 if (event
->attr
.comm
|| event
->attr
.mmap
|| event
->attr
.task
)
3333 static void perf_event_task_ctx(struct perf_event_context
*ctx
,
3334 struct perf_task_event
*task_event
)
3336 struct perf_event
*event
;
3338 list_for_each_entry_rcu(event
, &ctx
->event_list
, event_entry
) {
3339 if (perf_event_task_match(event
))
3340 perf_event_task_output(event
, task_event
);
3344 static void perf_event_task_event(struct perf_task_event
*task_event
)
3346 struct perf_cpu_context
*cpuctx
;
3347 struct perf_event_context
*ctx
= task_event
->task_ctx
;
3350 cpuctx
= &get_cpu_var(perf_cpu_context
);
3351 perf_event_task_ctx(&cpuctx
->ctx
, task_event
);
3353 ctx
= rcu_dereference(current
->perf_event_ctxp
);
3355 perf_event_task_ctx(ctx
, task_event
);
3356 put_cpu_var(perf_cpu_context
);
3360 static void perf_event_task(struct task_struct
*task
,
3361 struct perf_event_context
*task_ctx
,
3364 struct perf_task_event task_event
;
3366 if (!atomic_read(&nr_comm_events
) &&
3367 !atomic_read(&nr_mmap_events
) &&
3368 !atomic_read(&nr_task_events
))
3371 task_event
= (struct perf_task_event
){
3373 .task_ctx
= task_ctx
,
3376 .type
= new ? PERF_RECORD_FORK
: PERF_RECORD_EXIT
,
3378 .size
= sizeof(task_event
.event_id
),
3384 .time
= perf_clock(),
3388 perf_event_task_event(&task_event
);
3391 void perf_event_fork(struct task_struct
*task
)
3393 perf_event_task(task
, NULL
, 1);
3400 struct perf_comm_event
{
3401 struct task_struct
*task
;
3406 struct perf_event_header header
;
3413 static void perf_event_comm_output(struct perf_event
*event
,
3414 struct perf_comm_event
*comm_event
)
3416 struct perf_output_handle handle
;
3417 int size
= comm_event
->event_id
.header
.size
;
3418 int ret
= perf_output_begin(&handle
, event
, size
, 0, 0);
3423 comm_event
->event_id
.pid
= perf_event_pid(event
, comm_event
->task
);
3424 comm_event
->event_id
.tid
= perf_event_tid(event
, comm_event
->task
);
3426 perf_output_put(&handle
, comm_event
->event_id
);
3427 perf_output_copy(&handle
, comm_event
->comm
,
3428 comm_event
->comm_size
);
3429 perf_output_end(&handle
);
3432 static int perf_event_comm_match(struct perf_event
*event
)
3434 if (event
->state
< PERF_EVENT_STATE_INACTIVE
)
3437 if (event
->cpu
!= -1 && event
->cpu
!= smp_processor_id())
3440 if (event
->attr
.comm
)
3446 static void perf_event_comm_ctx(struct perf_event_context
*ctx
,
3447 struct perf_comm_event
*comm_event
)
3449 struct perf_event
*event
;
3451 list_for_each_entry_rcu(event
, &ctx
->event_list
, event_entry
) {
3452 if (perf_event_comm_match(event
))
3453 perf_event_comm_output(event
, comm_event
);
3457 static void perf_event_comm_event(struct perf_comm_event
*comm_event
)
3459 struct perf_cpu_context
*cpuctx
;
3460 struct perf_event_context
*ctx
;
3462 char comm
[TASK_COMM_LEN
];
3464 memset(comm
, 0, sizeof(comm
));
3465 strlcpy(comm
, comm_event
->task
->comm
, sizeof(comm
));
3466 size
= ALIGN(strlen(comm
)+1, sizeof(u64
));
3468 comm_event
->comm
= comm
;
3469 comm_event
->comm_size
= size
;
3471 comm_event
->event_id
.header
.size
= sizeof(comm_event
->event_id
) + size
;
3474 cpuctx
= &get_cpu_var(perf_cpu_context
);
3475 perf_event_comm_ctx(&cpuctx
->ctx
, comm_event
);
3476 ctx
= rcu_dereference(current
->perf_event_ctxp
);
3478 perf_event_comm_ctx(ctx
, comm_event
);
3479 put_cpu_var(perf_cpu_context
);
3483 void perf_event_comm(struct task_struct
*task
)
3485 struct perf_comm_event comm_event
;
3487 if (task
->perf_event_ctxp
)
3488 perf_event_enable_on_exec(task
);
3490 if (!atomic_read(&nr_comm_events
))
3493 comm_event
= (struct perf_comm_event
){
3499 .type
= PERF_RECORD_COMM
,
3508 perf_event_comm_event(&comm_event
);
3515 struct perf_mmap_event
{
3516 struct vm_area_struct
*vma
;
3518 const char *file_name
;
3522 struct perf_event_header header
;
3532 static void perf_event_mmap_output(struct perf_event
*event
,
3533 struct perf_mmap_event
*mmap_event
)
3535 struct perf_output_handle handle
;
3536 int size
= mmap_event
->event_id
.header
.size
;
3537 int ret
= perf_output_begin(&handle
, event
, size
, 0, 0);
3542 mmap_event
->event_id
.pid
= perf_event_pid(event
, current
);
3543 mmap_event
->event_id
.tid
= perf_event_tid(event
, current
);
3545 perf_output_put(&handle
, mmap_event
->event_id
);
3546 perf_output_copy(&handle
, mmap_event
->file_name
,
3547 mmap_event
->file_size
);
3548 perf_output_end(&handle
);
3551 static int perf_event_mmap_match(struct perf_event
*event
,
3552 struct perf_mmap_event
*mmap_event
)
3554 if (event
->state
< PERF_EVENT_STATE_INACTIVE
)
3557 if (event
->cpu
!= -1 && event
->cpu
!= smp_processor_id())
3560 if (event
->attr
.mmap
)
3566 static void perf_event_mmap_ctx(struct perf_event_context
*ctx
,
3567 struct perf_mmap_event
*mmap_event
)
3569 struct perf_event
*event
;
3571 list_for_each_entry_rcu(event
, &ctx
->event_list
, event_entry
) {
3572 if (perf_event_mmap_match(event
, mmap_event
))
3573 perf_event_mmap_output(event
, mmap_event
);
3577 static void perf_event_mmap_event(struct perf_mmap_event
*mmap_event
)
3579 struct perf_cpu_context
*cpuctx
;
3580 struct perf_event_context
*ctx
;
3581 struct vm_area_struct
*vma
= mmap_event
->vma
;
3582 struct file
*file
= vma
->vm_file
;
3588 memset(tmp
, 0, sizeof(tmp
));
3592 * d_path works from the end of the buffer backwards, so we
3593 * need to add enough zero bytes after the string to handle
3594 * the 64bit alignment we do later.
3596 buf
= kzalloc(PATH_MAX
+ sizeof(u64
), GFP_KERNEL
);
3598 name
= strncpy(tmp
, "//enomem", sizeof(tmp
));
3601 name
= d_path(&file
->f_path
, buf
, PATH_MAX
);
3603 name
= strncpy(tmp
, "//toolong", sizeof(tmp
));
3607 if (arch_vma_name(mmap_event
->vma
)) {
3608 name
= strncpy(tmp
, arch_vma_name(mmap_event
->vma
),
3614 name
= strncpy(tmp
, "[vdso]", sizeof(tmp
));
3618 name
= strncpy(tmp
, "//anon", sizeof(tmp
));
3623 size
= ALIGN(strlen(name
)+1, sizeof(u64
));
3625 mmap_event
->file_name
= name
;
3626 mmap_event
->file_size
= size
;
3628 mmap_event
->event_id
.header
.size
= sizeof(mmap_event
->event_id
) + size
;
3631 cpuctx
= &get_cpu_var(perf_cpu_context
);
3632 perf_event_mmap_ctx(&cpuctx
->ctx
, mmap_event
);
3633 ctx
= rcu_dereference(current
->perf_event_ctxp
);
3635 perf_event_mmap_ctx(ctx
, mmap_event
);
3636 put_cpu_var(perf_cpu_context
);
3642 void __perf_event_mmap(struct vm_area_struct
*vma
)
3644 struct perf_mmap_event mmap_event
;
3646 if (!atomic_read(&nr_mmap_events
))
3649 mmap_event
= (struct perf_mmap_event
){
3655 .type
= PERF_RECORD_MMAP
,
3661 .start
= vma
->vm_start
,
3662 .len
= vma
->vm_end
- vma
->vm_start
,
3663 .pgoff
= vma
->vm_pgoff
,
3667 perf_event_mmap_event(&mmap_event
);
3671 * IRQ throttle logging
3674 static void perf_log_throttle(struct perf_event
*event
, int enable
)
3676 struct perf_output_handle handle
;
3680 struct perf_event_header header
;
3684 } throttle_event
= {
3686 .type
= PERF_RECORD_THROTTLE
,
3688 .size
= sizeof(throttle_event
),
3690 .time
= perf_clock(),
3691 .id
= primary_event_id(event
),
3692 .stream_id
= event
->id
,
3696 throttle_event
.header
.type
= PERF_RECORD_UNTHROTTLE
;
3698 ret
= perf_output_begin(&handle
, event
, sizeof(throttle_event
), 1, 0);
3702 perf_output_put(&handle
, throttle_event
);
3703 perf_output_end(&handle
);
3707 * Generic event overflow handling, sampling.
3710 static int __perf_event_overflow(struct perf_event
*event
, int nmi
,
3711 int throttle
, struct perf_sample_data
*data
,
3712 struct pt_regs
*regs
)
3714 int events
= atomic_read(&event
->event_limit
);
3715 struct hw_perf_event
*hwc
= &event
->hw
;
3718 throttle
= (throttle
&& event
->pmu
->unthrottle
!= NULL
);
3723 if (hwc
->interrupts
!= MAX_INTERRUPTS
) {
3725 if (HZ
* hwc
->interrupts
>
3726 (u64
)sysctl_perf_event_sample_rate
) {
3727 hwc
->interrupts
= MAX_INTERRUPTS
;
3728 perf_log_throttle(event
, 0);
3733 * Keep re-disabling events even though on the previous
3734 * pass we disabled it - just in case we raced with a
3735 * sched-in and the event got enabled again:
3741 if (event
->attr
.freq
) {
3742 u64 now
= perf_clock();
3743 s64 delta
= now
- hwc
->freq_time_stamp
;
3745 hwc
->freq_time_stamp
= now
;
3747 if (delta
> 0 && delta
< 2*TICK_NSEC
)
3748 perf_adjust_period(event
, delta
, hwc
->last_period
);
3752 * XXX event_limit might not quite work as expected on inherited
3756 event
->pending_kill
= POLL_IN
;
3757 if (events
&& atomic_dec_and_test(&event
->event_limit
)) {
3759 event
->pending_kill
= POLL_HUP
;
3761 event
->pending_disable
= 1;
3762 perf_pending_queue(&event
->pending
,
3763 perf_pending_event
);
3765 perf_event_disable(event
);
3768 if (event
->overflow_handler
)
3769 event
->overflow_handler(event
, nmi
, data
, regs
);
3771 perf_event_output(event
, nmi
, data
, regs
);
3776 int perf_event_overflow(struct perf_event
*event
, int nmi
,
3777 struct perf_sample_data
*data
,
3778 struct pt_regs
*regs
)
3780 return __perf_event_overflow(event
, nmi
, 1, data
, regs
);
3784 * Generic software event infrastructure
3788 * We directly increment event->count and keep a second value in
3789 * event->hw.period_left to count intervals. This period event
3790 * is kept in the range [-sample_period, 0] so that we can use the
3794 static u64
perf_swevent_set_period(struct perf_event
*event
)
3796 struct hw_perf_event
*hwc
= &event
->hw
;
3797 u64 period
= hwc
->last_period
;
3801 hwc
->last_period
= hwc
->sample_period
;
3804 old
= val
= atomic64_read(&hwc
->period_left
);
3808 nr
= div64_u64(period
+ val
, period
);
3809 offset
= nr
* period
;
3811 if (atomic64_cmpxchg(&hwc
->period_left
, old
, val
) != old
)
3817 static void perf_swevent_overflow(struct perf_event
*event
, u64 overflow
,
3818 int nmi
, struct perf_sample_data
*data
,
3819 struct pt_regs
*regs
)
3821 struct hw_perf_event
*hwc
= &event
->hw
;
3824 data
->period
= event
->hw
.last_period
;
3826 overflow
= perf_swevent_set_period(event
);
3828 if (hwc
->interrupts
== MAX_INTERRUPTS
)
3831 for (; overflow
; overflow
--) {
3832 if (__perf_event_overflow(event
, nmi
, throttle
,
3835 * We inhibit the overflow from happening when
3836 * hwc->interrupts == MAX_INTERRUPTS.
3844 static void perf_swevent_unthrottle(struct perf_event
*event
)
3847 * Nothing to do, we already reset hwc->interrupts.
3851 static void perf_swevent_add(struct perf_event
*event
, u64 nr
,
3852 int nmi
, struct perf_sample_data
*data
,
3853 struct pt_regs
*regs
)
3855 struct hw_perf_event
*hwc
= &event
->hw
;
3857 atomic64_add(nr
, &event
->count
);
3862 if (!hwc
->sample_period
)
3865 if (nr
== 1 && hwc
->sample_period
== 1 && !event
->attr
.freq
)
3866 return perf_swevent_overflow(event
, 1, nmi
, data
, regs
);
3868 if (atomic64_add_negative(nr
, &hwc
->period_left
))
3871 perf_swevent_overflow(event
, 0, nmi
, data
, regs
);
3874 static int perf_swevent_is_counting(struct perf_event
*event
)
3877 * The event is active, we're good!
3879 if (event
->state
== PERF_EVENT_STATE_ACTIVE
)
3883 * The event is off/error, not counting.
3885 if (event
->state
!= PERF_EVENT_STATE_INACTIVE
)
3889 * The event is inactive, if the context is active
3890 * we're part of a group that didn't make it on the 'pmu',
3893 if (event
->ctx
->is_active
)
3897 * We're inactive and the context is too, this means the
3898 * task is scheduled out, we're counting events that happen
3899 * to us, like migration events.
3904 static int perf_tp_event_match(struct perf_event
*event
,
3905 struct perf_sample_data
*data
);
3907 static int perf_exclude_event(struct perf_event
*event
,
3908 struct pt_regs
*regs
)
3911 if (event
->attr
.exclude_user
&& user_mode(regs
))
3914 if (event
->attr
.exclude_kernel
&& !user_mode(regs
))
3921 static int perf_swevent_match(struct perf_event
*event
,
3922 enum perf_type_id type
,
3924 struct perf_sample_data
*data
,
3925 struct pt_regs
*regs
)
3927 if (event
->cpu
!= -1 && event
->cpu
!= smp_processor_id())
3930 if (!perf_swevent_is_counting(event
))
3933 if (event
->attr
.type
!= type
)
3936 if (event
->attr
.config
!= event_id
)
3939 if (perf_exclude_event(event
, regs
))
3942 if (event
->attr
.type
== PERF_TYPE_TRACEPOINT
&&
3943 !perf_tp_event_match(event
, data
))
3949 static void perf_swevent_ctx_event(struct perf_event_context
*ctx
,
3950 enum perf_type_id type
,
3951 u32 event_id
, u64 nr
, int nmi
,
3952 struct perf_sample_data
*data
,
3953 struct pt_regs
*regs
)
3955 struct perf_event
*event
;
3957 list_for_each_entry_rcu(event
, &ctx
->event_list
, event_entry
) {
3958 if (perf_swevent_match(event
, type
, event_id
, data
, regs
))
3959 perf_swevent_add(event
, nr
, nmi
, data
, regs
);
3963 int perf_swevent_get_recursion_context(void)
3965 struct perf_cpu_context
*cpuctx
= &get_cpu_var(perf_cpu_context
);
3972 else if (in_softirq())
3977 if (cpuctx
->recursion
[rctx
]) {
3978 put_cpu_var(perf_cpu_context
);
3982 cpuctx
->recursion
[rctx
]++;
3987 EXPORT_SYMBOL_GPL(perf_swevent_get_recursion_context
);
3989 void perf_swevent_put_recursion_context(int rctx
)
3991 struct perf_cpu_context
*cpuctx
= &__get_cpu_var(perf_cpu_context
);
3993 cpuctx
->recursion
[rctx
]--;
3994 put_cpu_var(perf_cpu_context
);
3996 EXPORT_SYMBOL_GPL(perf_swevent_put_recursion_context
);
3998 static void do_perf_sw_event(enum perf_type_id type
, u32 event_id
,
4000 struct perf_sample_data
*data
,
4001 struct pt_regs
*regs
)
4003 struct perf_cpu_context
*cpuctx
;
4004 struct perf_event_context
*ctx
;
4006 cpuctx
= &__get_cpu_var(perf_cpu_context
);
4008 perf_swevent_ctx_event(&cpuctx
->ctx
, type
, event_id
,
4009 nr
, nmi
, data
, regs
);
4011 * doesn't really matter which of the child contexts the
4012 * events ends up in.
4014 ctx
= rcu_dereference(current
->perf_event_ctxp
);
4016 perf_swevent_ctx_event(ctx
, type
, event_id
, nr
, nmi
, data
, regs
);
4020 void __perf_sw_event(u32 event_id
, u64 nr
, int nmi
,
4021 struct pt_regs
*regs
, u64 addr
)
4023 struct perf_sample_data data
;
4026 rctx
= perf_swevent_get_recursion_context();
4030 perf_sample_data_init(&data
, addr
);
4032 do_perf_sw_event(PERF_TYPE_SOFTWARE
, event_id
, nr
, nmi
, &data
, regs
);
4034 perf_swevent_put_recursion_context(rctx
);
4037 static void perf_swevent_read(struct perf_event
*event
)
4041 static int perf_swevent_enable(struct perf_event
*event
)
4043 struct hw_perf_event
*hwc
= &event
->hw
;
4045 if (hwc
->sample_period
) {
4046 hwc
->last_period
= hwc
->sample_period
;
4047 perf_swevent_set_period(event
);
4052 static void perf_swevent_disable(struct perf_event
*event
)
4056 static const struct pmu perf_ops_generic
= {
4057 .enable
= perf_swevent_enable
,
4058 .disable
= perf_swevent_disable
,
4059 .read
= perf_swevent_read
,
4060 .unthrottle
= perf_swevent_unthrottle
,
4064 * hrtimer based swevent callback
4067 static enum hrtimer_restart
perf_swevent_hrtimer(struct hrtimer
*hrtimer
)
4069 enum hrtimer_restart ret
= HRTIMER_RESTART
;
4070 struct perf_sample_data data
;
4071 struct pt_regs
*regs
;
4072 struct perf_event
*event
;
4075 event
= container_of(hrtimer
, struct perf_event
, hw
.hrtimer
);
4076 event
->pmu
->read(event
);
4078 perf_sample_data_init(&data
, 0);
4079 data
.period
= event
->hw
.last_period
;
4080 regs
= get_irq_regs();
4082 * In case we exclude kernel IPs or are somehow not in interrupt
4083 * context, provide the next best thing, the user IP.
4085 if ((event
->attr
.exclude_kernel
|| !regs
) &&
4086 !event
->attr
.exclude_user
)
4087 regs
= task_pt_regs(current
);
4090 if (!(event
->attr
.exclude_idle
&& current
->pid
== 0))
4091 if (perf_event_overflow(event
, 0, &data
, regs
))
4092 ret
= HRTIMER_NORESTART
;
4095 period
= max_t(u64
, 10000, event
->hw
.sample_period
);
4096 hrtimer_forward_now(hrtimer
, ns_to_ktime(period
));
4101 static void perf_swevent_start_hrtimer(struct perf_event
*event
)
4103 struct hw_perf_event
*hwc
= &event
->hw
;
4105 hrtimer_init(&hwc
->hrtimer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
4106 hwc
->hrtimer
.function
= perf_swevent_hrtimer
;
4107 if (hwc
->sample_period
) {
4110 if (hwc
->remaining
) {
4111 if (hwc
->remaining
< 0)
4114 period
= hwc
->remaining
;
4117 period
= max_t(u64
, 10000, hwc
->sample_period
);
4119 __hrtimer_start_range_ns(&hwc
->hrtimer
,
4120 ns_to_ktime(period
), 0,
4121 HRTIMER_MODE_REL
, 0);
4125 static void perf_swevent_cancel_hrtimer(struct perf_event
*event
)
4127 struct hw_perf_event
*hwc
= &event
->hw
;
4129 if (hwc
->sample_period
) {
4130 ktime_t remaining
= hrtimer_get_remaining(&hwc
->hrtimer
);
4131 hwc
->remaining
= ktime_to_ns(remaining
);
4133 hrtimer_cancel(&hwc
->hrtimer
);
4138 * Software event: cpu wall time clock
4141 static void cpu_clock_perf_event_update(struct perf_event
*event
)
4143 int cpu
= raw_smp_processor_id();
4147 now
= cpu_clock(cpu
);
4148 prev
= atomic64_xchg(&event
->hw
.prev_count
, now
);
4149 atomic64_add(now
- prev
, &event
->count
);
4152 static int cpu_clock_perf_event_enable(struct perf_event
*event
)
4154 struct hw_perf_event
*hwc
= &event
->hw
;
4155 int cpu
= raw_smp_processor_id();
4157 atomic64_set(&hwc
->prev_count
, cpu_clock(cpu
));
4158 perf_swevent_start_hrtimer(event
);
4163 static void cpu_clock_perf_event_disable(struct perf_event
*event
)
4165 perf_swevent_cancel_hrtimer(event
);
4166 cpu_clock_perf_event_update(event
);
4169 static void cpu_clock_perf_event_read(struct perf_event
*event
)
4171 cpu_clock_perf_event_update(event
);
4174 static const struct pmu perf_ops_cpu_clock
= {
4175 .enable
= cpu_clock_perf_event_enable
,
4176 .disable
= cpu_clock_perf_event_disable
,
4177 .read
= cpu_clock_perf_event_read
,
4181 * Software event: task time clock
4184 static void task_clock_perf_event_update(struct perf_event
*event
, u64 now
)
4189 prev
= atomic64_xchg(&event
->hw
.prev_count
, now
);
4191 atomic64_add(delta
, &event
->count
);
4194 static int task_clock_perf_event_enable(struct perf_event
*event
)
4196 struct hw_perf_event
*hwc
= &event
->hw
;
4199 now
= event
->ctx
->time
;
4201 atomic64_set(&hwc
->prev_count
, now
);
4203 perf_swevent_start_hrtimer(event
);
4208 static void task_clock_perf_event_disable(struct perf_event
*event
)
4210 perf_swevent_cancel_hrtimer(event
);
4211 task_clock_perf_event_update(event
, event
->ctx
->time
);
4215 static void task_clock_perf_event_read(struct perf_event
*event
)
4220 update_context_time(event
->ctx
);
4221 time
= event
->ctx
->time
;
4223 u64 now
= perf_clock();
4224 u64 delta
= now
- event
->ctx
->timestamp
;
4225 time
= event
->ctx
->time
+ delta
;
4228 task_clock_perf_event_update(event
, time
);
4231 static const struct pmu perf_ops_task_clock
= {
4232 .enable
= task_clock_perf_event_enable
,
4233 .disable
= task_clock_perf_event_disable
,
4234 .read
= task_clock_perf_event_read
,
4237 #ifdef CONFIG_EVENT_PROFILE
4239 void perf_tp_event(int event_id
, u64 addr
, u64 count
, void *record
,
4242 struct pt_regs
*regs
= get_irq_regs();
4243 struct perf_sample_data data
;
4244 struct perf_raw_record raw
= {
4249 perf_sample_data_init(&data
, addr
);
4253 regs
= task_pt_regs(current
);
4255 /* Trace events already protected against recursion */
4256 do_perf_sw_event(PERF_TYPE_TRACEPOINT
, event_id
, count
, 1,
4259 EXPORT_SYMBOL_GPL(perf_tp_event
);
4261 static int perf_tp_event_match(struct perf_event
*event
,
4262 struct perf_sample_data
*data
)
4264 void *record
= data
->raw
->data
;
4266 if (likely(!event
->filter
) || filter_match_preds(event
->filter
, record
))
4271 static void tp_perf_event_destroy(struct perf_event
*event
)
4273 ftrace_profile_disable(event
->attr
.config
);
4276 static const struct pmu
*tp_perf_event_init(struct perf_event
*event
)
4279 * Raw tracepoint data is a severe data leak, only allow root to
4282 if ((event
->attr
.sample_type
& PERF_SAMPLE_RAW
) &&
4283 perf_paranoid_tracepoint_raw() &&
4284 !capable(CAP_SYS_ADMIN
))
4285 return ERR_PTR(-EPERM
);
4287 if (ftrace_profile_enable(event
->attr
.config
))
4290 event
->destroy
= tp_perf_event_destroy
;
4292 return &perf_ops_generic
;
4295 static int perf_event_set_filter(struct perf_event
*event
, void __user
*arg
)
4300 if (event
->attr
.type
!= PERF_TYPE_TRACEPOINT
)
4303 filter_str
= strndup_user(arg
, PAGE_SIZE
);
4304 if (IS_ERR(filter_str
))
4305 return PTR_ERR(filter_str
);
4307 ret
= ftrace_profile_set_filter(event
, event
->attr
.config
, filter_str
);
4313 static void perf_event_free_filter(struct perf_event
*event
)
4315 ftrace_profile_free_filter(event
);
4320 static int perf_tp_event_match(struct perf_event
*event
,
4321 struct perf_sample_data
*data
)
4326 static const struct pmu
*tp_perf_event_init(struct perf_event
*event
)
4331 static int perf_event_set_filter(struct perf_event
*event
, void __user
*arg
)
4336 static void perf_event_free_filter(struct perf_event
*event
)
4340 #endif /* CONFIG_EVENT_PROFILE */
4342 #ifdef CONFIG_HAVE_HW_BREAKPOINT
4343 static void bp_perf_event_destroy(struct perf_event
*event
)
4345 release_bp_slot(event
);
4348 static const struct pmu
*bp_perf_event_init(struct perf_event
*bp
)
4352 err
= register_perf_hw_breakpoint(bp
);
4354 return ERR_PTR(err
);
4356 bp
->destroy
= bp_perf_event_destroy
;
4358 return &perf_ops_bp
;
4361 void perf_bp_event(struct perf_event
*bp
, void *data
)
4363 struct perf_sample_data sample
;
4364 struct pt_regs
*regs
= data
;
4366 perf_sample_data_init(&sample
, bp
->attr
.bp_addr
);
4368 if (!perf_exclude_event(bp
, regs
))
4369 perf_swevent_add(bp
, 1, 1, &sample
, regs
);
4372 static const struct pmu
*bp_perf_event_init(struct perf_event
*bp
)
4377 void perf_bp_event(struct perf_event
*bp
, void *regs
)
4382 atomic_t perf_swevent_enabled
[PERF_COUNT_SW_MAX
];
4384 static void sw_perf_event_destroy(struct perf_event
*event
)
4386 u64 event_id
= event
->attr
.config
;
4388 WARN_ON(event
->parent
);
4390 atomic_dec(&perf_swevent_enabled
[event_id
]);
4393 static const struct pmu
*sw_perf_event_init(struct perf_event
*event
)
4395 const struct pmu
*pmu
= NULL
;
4396 u64 event_id
= event
->attr
.config
;
4399 * Software events (currently) can't in general distinguish
4400 * between user, kernel and hypervisor events.
4401 * However, context switches and cpu migrations are considered
4402 * to be kernel events, and page faults are never hypervisor
4406 case PERF_COUNT_SW_CPU_CLOCK
:
4407 pmu
= &perf_ops_cpu_clock
;
4410 case PERF_COUNT_SW_TASK_CLOCK
:
4412 * If the user instantiates this as a per-cpu event,
4413 * use the cpu_clock event instead.
4415 if (event
->ctx
->task
)
4416 pmu
= &perf_ops_task_clock
;
4418 pmu
= &perf_ops_cpu_clock
;
4421 case PERF_COUNT_SW_PAGE_FAULTS
:
4422 case PERF_COUNT_SW_PAGE_FAULTS_MIN
:
4423 case PERF_COUNT_SW_PAGE_FAULTS_MAJ
:
4424 case PERF_COUNT_SW_CONTEXT_SWITCHES
:
4425 case PERF_COUNT_SW_CPU_MIGRATIONS
:
4426 case PERF_COUNT_SW_ALIGNMENT_FAULTS
:
4427 case PERF_COUNT_SW_EMULATION_FAULTS
:
4428 if (!event
->parent
) {
4429 atomic_inc(&perf_swevent_enabled
[event_id
]);
4430 event
->destroy
= sw_perf_event_destroy
;
4432 pmu
= &perf_ops_generic
;
4440 * Allocate and initialize a event structure
4442 static struct perf_event
*
4443 perf_event_alloc(struct perf_event_attr
*attr
,
4445 struct perf_event_context
*ctx
,
4446 struct perf_event
*group_leader
,
4447 struct perf_event
*parent_event
,
4448 perf_overflow_handler_t overflow_handler
,
4451 const struct pmu
*pmu
;
4452 struct perf_event
*event
;
4453 struct hw_perf_event
*hwc
;
4456 event
= kzalloc(sizeof(*event
), gfpflags
);
4458 return ERR_PTR(-ENOMEM
);
4461 * Single events are their own group leaders, with an
4462 * empty sibling list:
4465 group_leader
= event
;
4467 mutex_init(&event
->child_mutex
);
4468 INIT_LIST_HEAD(&event
->child_list
);
4470 INIT_LIST_HEAD(&event
->group_entry
);
4471 INIT_LIST_HEAD(&event
->event_entry
);
4472 INIT_LIST_HEAD(&event
->sibling_list
);
4473 init_waitqueue_head(&event
->waitq
);
4475 mutex_init(&event
->mmap_mutex
);
4478 event
->attr
= *attr
;
4479 event
->group_leader
= group_leader
;
4484 event
->parent
= parent_event
;
4486 event
->ns
= get_pid_ns(current
->nsproxy
->pid_ns
);
4487 event
->id
= atomic64_inc_return(&perf_event_id
);
4489 event
->state
= PERF_EVENT_STATE_INACTIVE
;
4491 if (!overflow_handler
&& parent_event
)
4492 overflow_handler
= parent_event
->overflow_handler
;
4494 event
->overflow_handler
= overflow_handler
;
4497 event
->state
= PERF_EVENT_STATE_OFF
;
4502 hwc
->sample_period
= attr
->sample_period
;
4503 if (attr
->freq
&& attr
->sample_freq
)
4504 hwc
->sample_period
= 1;
4505 hwc
->last_period
= hwc
->sample_period
;
4507 atomic64_set(&hwc
->period_left
, hwc
->sample_period
);
4510 * we currently do not support PERF_FORMAT_GROUP on inherited events
4512 if (attr
->inherit
&& (attr
->read_format
& PERF_FORMAT_GROUP
))
4515 switch (attr
->type
) {
4517 case PERF_TYPE_HARDWARE
:
4518 case PERF_TYPE_HW_CACHE
:
4519 pmu
= hw_perf_event_init(event
);
4522 case PERF_TYPE_SOFTWARE
:
4523 pmu
= sw_perf_event_init(event
);
4526 case PERF_TYPE_TRACEPOINT
:
4527 pmu
= tp_perf_event_init(event
);
4530 case PERF_TYPE_BREAKPOINT
:
4531 pmu
= bp_perf_event_init(event
);
4542 else if (IS_ERR(pmu
))
4547 put_pid_ns(event
->ns
);
4549 return ERR_PTR(err
);
4554 if (!event
->parent
) {
4555 atomic_inc(&nr_events
);
4556 if (event
->attr
.mmap
)
4557 atomic_inc(&nr_mmap_events
);
4558 if (event
->attr
.comm
)
4559 atomic_inc(&nr_comm_events
);
4560 if (event
->attr
.task
)
4561 atomic_inc(&nr_task_events
);
4567 static int perf_copy_attr(struct perf_event_attr __user
*uattr
,
4568 struct perf_event_attr
*attr
)
4573 if (!access_ok(VERIFY_WRITE
, uattr
, PERF_ATTR_SIZE_VER0
))
4577 * zero the full structure, so that a short copy will be nice.
4579 memset(attr
, 0, sizeof(*attr
));
4581 ret
= get_user(size
, &uattr
->size
);
4585 if (size
> PAGE_SIZE
) /* silly large */
4588 if (!size
) /* abi compat */
4589 size
= PERF_ATTR_SIZE_VER0
;
4591 if (size
< PERF_ATTR_SIZE_VER0
)
4595 * If we're handed a bigger struct than we know of,
4596 * ensure all the unknown bits are 0 - i.e. new
4597 * user-space does not rely on any kernel feature
4598 * extensions we dont know about yet.
4600 if (size
> sizeof(*attr
)) {
4601 unsigned char __user
*addr
;
4602 unsigned char __user
*end
;
4605 addr
= (void __user
*)uattr
+ sizeof(*attr
);
4606 end
= (void __user
*)uattr
+ size
;
4608 for (; addr
< end
; addr
++) {
4609 ret
= get_user(val
, addr
);
4615 size
= sizeof(*attr
);
4618 ret
= copy_from_user(attr
, uattr
, size
);
4623 * If the type exists, the corresponding creation will verify
4626 if (attr
->type
>= PERF_TYPE_MAX
)
4629 if (attr
->__reserved_1
)
4632 if (attr
->sample_type
& ~(PERF_SAMPLE_MAX
-1))
4635 if (attr
->read_format
& ~(PERF_FORMAT_MAX
-1))
4642 put_user(sizeof(*attr
), &uattr
->size
);
4647 static int perf_event_set_output(struct perf_event
*event
, int output_fd
)
4649 struct perf_event
*output_event
= NULL
;
4650 struct file
*output_file
= NULL
;
4651 struct perf_event
*old_output
;
4652 int fput_needed
= 0;
4658 output_file
= fget_light(output_fd
, &fput_needed
);
4662 if (output_file
->f_op
!= &perf_fops
)
4665 output_event
= output_file
->private_data
;
4667 /* Don't chain output fds */
4668 if (output_event
->output
)
4671 /* Don't set an output fd when we already have an output channel */
4675 atomic_long_inc(&output_file
->f_count
);
4678 mutex_lock(&event
->mmap_mutex
);
4679 old_output
= event
->output
;
4680 rcu_assign_pointer(event
->output
, output_event
);
4681 mutex_unlock(&event
->mmap_mutex
);
4685 * we need to make sure no existing perf_output_*()
4686 * is still referencing this event.
4689 fput(old_output
->filp
);
4694 fput_light(output_file
, fput_needed
);
4699 * sys_perf_event_open - open a performance event, associate it to a task/cpu
4701 * @attr_uptr: event_id type attributes for monitoring/sampling
4704 * @group_fd: group leader event fd
4706 SYSCALL_DEFINE5(perf_event_open
,
4707 struct perf_event_attr __user
*, attr_uptr
,
4708 pid_t
, pid
, int, cpu
, int, group_fd
, unsigned long, flags
)
4710 struct perf_event
*event
, *group_leader
;
4711 struct perf_event_attr attr
;
4712 struct perf_event_context
*ctx
;
4713 struct file
*event_file
= NULL
;
4714 struct file
*group_file
= NULL
;
4715 int fput_needed
= 0;
4716 int fput_needed2
= 0;
4719 /* for future expandability... */
4720 if (flags
& ~(PERF_FLAG_FD_NO_GROUP
| PERF_FLAG_FD_OUTPUT
))
4723 err
= perf_copy_attr(attr_uptr
, &attr
);
4727 if (!attr
.exclude_kernel
) {
4728 if (perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN
))
4733 if (attr
.sample_freq
> sysctl_perf_event_sample_rate
)
4738 * Get the target context (task or percpu):
4740 ctx
= find_get_context(pid
, cpu
);
4742 return PTR_ERR(ctx
);
4745 * Look up the group leader (we will attach this event to it):
4747 group_leader
= NULL
;
4748 if (group_fd
!= -1 && !(flags
& PERF_FLAG_FD_NO_GROUP
)) {
4750 group_file
= fget_light(group_fd
, &fput_needed
);
4752 goto err_put_context
;
4753 if (group_file
->f_op
!= &perf_fops
)
4754 goto err_put_context
;
4756 group_leader
= group_file
->private_data
;
4758 * Do not allow a recursive hierarchy (this new sibling
4759 * becoming part of another group-sibling):
4761 if (group_leader
->group_leader
!= group_leader
)
4762 goto err_put_context
;
4764 * Do not allow to attach to a group in a different
4765 * task or CPU context:
4767 if (group_leader
->ctx
!= ctx
)
4768 goto err_put_context
;
4770 * Only a group leader can be exclusive or pinned
4772 if (attr
.exclusive
|| attr
.pinned
)
4773 goto err_put_context
;
4776 event
= perf_event_alloc(&attr
, cpu
, ctx
, group_leader
,
4777 NULL
, NULL
, GFP_KERNEL
);
4778 err
= PTR_ERR(event
);
4780 goto err_put_context
;
4782 err
= anon_inode_getfd("[perf_event]", &perf_fops
, event
, O_RDWR
);
4784 goto err_free_put_context
;
4786 event_file
= fget_light(err
, &fput_needed2
);
4788 goto err_free_put_context
;
4790 if (flags
& PERF_FLAG_FD_OUTPUT
) {
4791 err
= perf_event_set_output(event
, group_fd
);
4793 goto err_fput_free_put_context
;
4796 event
->filp
= event_file
;
4797 WARN_ON_ONCE(ctx
->parent_ctx
);
4798 mutex_lock(&ctx
->mutex
);
4799 perf_install_in_context(ctx
, event
, cpu
);
4801 mutex_unlock(&ctx
->mutex
);
4803 event
->owner
= current
;
4804 get_task_struct(current
);
4805 mutex_lock(¤t
->perf_event_mutex
);
4806 list_add_tail(&event
->owner_entry
, ¤t
->perf_event_list
);
4807 mutex_unlock(¤t
->perf_event_mutex
);
4809 err_fput_free_put_context
:
4810 fput_light(event_file
, fput_needed2
);
4812 err_free_put_context
:
4820 fput_light(group_file
, fput_needed
);
4826 * perf_event_create_kernel_counter
4828 * @attr: attributes of the counter to create
4829 * @cpu: cpu in which the counter is bound
4830 * @pid: task to profile
4833 perf_event_create_kernel_counter(struct perf_event_attr
*attr
, int cpu
,
4835 perf_overflow_handler_t overflow_handler
)
4837 struct perf_event
*event
;
4838 struct perf_event_context
*ctx
;
4842 * Get the target context (task or percpu):
4845 ctx
= find_get_context(pid
, cpu
);
4851 event
= perf_event_alloc(attr
, cpu
, ctx
, NULL
,
4852 NULL
, overflow_handler
, GFP_KERNEL
);
4853 if (IS_ERR(event
)) {
4854 err
= PTR_ERR(event
);
4855 goto err_put_context
;
4859 WARN_ON_ONCE(ctx
->parent_ctx
);
4860 mutex_lock(&ctx
->mutex
);
4861 perf_install_in_context(ctx
, event
, cpu
);
4863 mutex_unlock(&ctx
->mutex
);
4865 event
->owner
= current
;
4866 get_task_struct(current
);
4867 mutex_lock(¤t
->perf_event_mutex
);
4868 list_add_tail(&event
->owner_entry
, ¤t
->perf_event_list
);
4869 mutex_unlock(¤t
->perf_event_mutex
);
4876 return ERR_PTR(err
);
4878 EXPORT_SYMBOL_GPL(perf_event_create_kernel_counter
);
4881 * inherit a event from parent task to child task:
4883 static struct perf_event
*
4884 inherit_event(struct perf_event
*parent_event
,
4885 struct task_struct
*parent
,
4886 struct perf_event_context
*parent_ctx
,
4887 struct task_struct
*child
,
4888 struct perf_event
*group_leader
,
4889 struct perf_event_context
*child_ctx
)
4891 struct perf_event
*child_event
;
4894 * Instead of creating recursive hierarchies of events,
4895 * we link inherited events back to the original parent,
4896 * which has a filp for sure, which we use as the reference
4899 if (parent_event
->parent
)
4900 parent_event
= parent_event
->parent
;
4902 child_event
= perf_event_alloc(&parent_event
->attr
,
4903 parent_event
->cpu
, child_ctx
,
4904 group_leader
, parent_event
,
4906 if (IS_ERR(child_event
))
4911 * Make the child state follow the state of the parent event,
4912 * not its attr.disabled bit. We hold the parent's mutex,
4913 * so we won't race with perf_event_{en, dis}able_family.
4915 if (parent_event
->state
>= PERF_EVENT_STATE_INACTIVE
)
4916 child_event
->state
= PERF_EVENT_STATE_INACTIVE
;
4918 child_event
->state
= PERF_EVENT_STATE_OFF
;
4920 if (parent_event
->attr
.freq
)
4921 child_event
->hw
.sample_period
= parent_event
->hw
.sample_period
;
4923 child_event
->overflow_handler
= parent_event
->overflow_handler
;
4926 * Link it up in the child's context:
4928 add_event_to_ctx(child_event
, child_ctx
);
4931 * Get a reference to the parent filp - we will fput it
4932 * when the child event exits. This is safe to do because
4933 * we are in the parent and we know that the filp still
4934 * exists and has a nonzero count:
4936 atomic_long_inc(&parent_event
->filp
->f_count
);
4939 * Link this into the parent event's child list
4941 WARN_ON_ONCE(parent_event
->ctx
->parent_ctx
);
4942 mutex_lock(&parent_event
->child_mutex
);
4943 list_add_tail(&child_event
->child_list
, &parent_event
->child_list
);
4944 mutex_unlock(&parent_event
->child_mutex
);
4949 static int inherit_group(struct perf_event
*parent_event
,
4950 struct task_struct
*parent
,
4951 struct perf_event_context
*parent_ctx
,
4952 struct task_struct
*child
,
4953 struct perf_event_context
*child_ctx
)
4955 struct perf_event
*leader
;
4956 struct perf_event
*sub
;
4957 struct perf_event
*child_ctr
;
4959 leader
= inherit_event(parent_event
, parent
, parent_ctx
,
4960 child
, NULL
, child_ctx
);
4962 return PTR_ERR(leader
);
4963 list_for_each_entry(sub
, &parent_event
->sibling_list
, group_entry
) {
4964 child_ctr
= inherit_event(sub
, parent
, parent_ctx
,
4965 child
, leader
, child_ctx
);
4966 if (IS_ERR(child_ctr
))
4967 return PTR_ERR(child_ctr
);
4972 static void sync_child_event(struct perf_event
*child_event
,
4973 struct task_struct
*child
)
4975 struct perf_event
*parent_event
= child_event
->parent
;
4978 if (child_event
->attr
.inherit_stat
)
4979 perf_event_read_event(child_event
, child
);
4981 child_val
= atomic64_read(&child_event
->count
);
4984 * Add back the child's count to the parent's count:
4986 atomic64_add(child_val
, &parent_event
->count
);
4987 atomic64_add(child_event
->total_time_enabled
,
4988 &parent_event
->child_total_time_enabled
);
4989 atomic64_add(child_event
->total_time_running
,
4990 &parent_event
->child_total_time_running
);
4993 * Remove this event from the parent's list
4995 WARN_ON_ONCE(parent_event
->ctx
->parent_ctx
);
4996 mutex_lock(&parent_event
->child_mutex
);
4997 list_del_init(&child_event
->child_list
);
4998 mutex_unlock(&parent_event
->child_mutex
);
5001 * Release the parent event, if this was the last
5004 fput(parent_event
->filp
);
5008 __perf_event_exit_task(struct perf_event
*child_event
,
5009 struct perf_event_context
*child_ctx
,
5010 struct task_struct
*child
)
5012 struct perf_event
*parent_event
;
5014 perf_event_remove_from_context(child_event
);
5016 parent_event
= child_event
->parent
;
5018 * It can happen that parent exits first, and has events
5019 * that are still around due to the child reference. These
5020 * events need to be zapped - but otherwise linger.
5023 sync_child_event(child_event
, child
);
5024 free_event(child_event
);
5029 * When a child task exits, feed back event values to parent events.
5031 void perf_event_exit_task(struct task_struct
*child
)
5033 struct perf_event
*child_event
, *tmp
;
5034 struct perf_event_context
*child_ctx
;
5035 unsigned long flags
;
5037 if (likely(!child
->perf_event_ctxp
)) {
5038 perf_event_task(child
, NULL
, 0);
5042 local_irq_save(flags
);
5044 * We can't reschedule here because interrupts are disabled,
5045 * and either child is current or it is a task that can't be
5046 * scheduled, so we are now safe from rescheduling changing
5049 child_ctx
= child
->perf_event_ctxp
;
5050 __perf_event_task_sched_out(child_ctx
);
5053 * Take the context lock here so that if find_get_context is
5054 * reading child->perf_event_ctxp, we wait until it has
5055 * incremented the context's refcount before we do put_ctx below.
5057 raw_spin_lock(&child_ctx
->lock
);
5058 child
->perf_event_ctxp
= NULL
;
5060 * If this context is a clone; unclone it so it can't get
5061 * swapped to another process while we're removing all
5062 * the events from it.
5064 unclone_ctx(child_ctx
);
5065 update_context_time(child_ctx
);
5066 raw_spin_unlock_irqrestore(&child_ctx
->lock
, flags
);
5069 * Report the task dead after unscheduling the events so that we
5070 * won't get any samples after PERF_RECORD_EXIT. We can however still
5071 * get a few PERF_RECORD_READ events.
5073 perf_event_task(child
, child_ctx
, 0);
5076 * We can recurse on the same lock type through:
5078 * __perf_event_exit_task()
5079 * sync_child_event()
5080 * fput(parent_event->filp)
5082 * mutex_lock(&ctx->mutex)
5084 * But since its the parent context it won't be the same instance.
5086 mutex_lock_nested(&child_ctx
->mutex
, SINGLE_DEPTH_NESTING
);
5089 list_for_each_entry_safe(child_event
, tmp
, &child_ctx
->group_list
,
5091 __perf_event_exit_task(child_event
, child_ctx
, child
);
5094 * If the last event was a group event, it will have appended all
5095 * its siblings to the list, but we obtained 'tmp' before that which
5096 * will still point to the list head terminating the iteration.
5098 if (!list_empty(&child_ctx
->group_list
))
5101 mutex_unlock(&child_ctx
->mutex
);
5107 * free an unexposed, unused context as created by inheritance by
5108 * init_task below, used by fork() in case of fail.
5110 void perf_event_free_task(struct task_struct
*task
)
5112 struct perf_event_context
*ctx
= task
->perf_event_ctxp
;
5113 struct perf_event
*event
, *tmp
;
5118 mutex_lock(&ctx
->mutex
);
5120 list_for_each_entry_safe(event
, tmp
, &ctx
->group_list
, group_entry
) {
5121 struct perf_event
*parent
= event
->parent
;
5123 if (WARN_ON_ONCE(!parent
))
5126 mutex_lock(&parent
->child_mutex
);
5127 list_del_init(&event
->child_list
);
5128 mutex_unlock(&parent
->child_mutex
);
5132 list_del_event(event
, ctx
);
5136 if (!list_empty(&ctx
->group_list
))
5139 mutex_unlock(&ctx
->mutex
);
5145 * Initialize the perf_event context in task_struct
5147 int perf_event_init_task(struct task_struct
*child
)
5149 struct perf_event_context
*child_ctx
= NULL
, *parent_ctx
;
5150 struct perf_event_context
*cloned_ctx
;
5151 struct perf_event
*event
;
5152 struct task_struct
*parent
= current
;
5153 int inherited_all
= 1;
5156 child
->perf_event_ctxp
= NULL
;
5158 mutex_init(&child
->perf_event_mutex
);
5159 INIT_LIST_HEAD(&child
->perf_event_list
);
5161 if (likely(!parent
->perf_event_ctxp
))
5165 * If the parent's context is a clone, pin it so it won't get
5168 parent_ctx
= perf_pin_task_context(parent
);
5171 * No need to check if parent_ctx != NULL here; since we saw
5172 * it non-NULL earlier, the only reason for it to become NULL
5173 * is if we exit, and since we're currently in the middle of
5174 * a fork we can't be exiting at the same time.
5178 * Lock the parent list. No need to lock the child - not PID
5179 * hashed yet and not running, so nobody can access it.
5181 mutex_lock(&parent_ctx
->mutex
);
5184 * We dont have to disable NMIs - we are only looking at
5185 * the list, not manipulating it:
5187 list_for_each_entry(event
, &parent_ctx
->group_list
, group_entry
) {
5189 if (!event
->attr
.inherit
) {
5194 if (!child
->perf_event_ctxp
) {
5196 * This is executed from the parent task context, so
5197 * inherit events that have been marked for cloning.
5198 * First allocate and initialize a context for the
5202 child_ctx
= kzalloc(sizeof(struct perf_event_context
),
5209 __perf_event_init_context(child_ctx
, child
);
5210 child
->perf_event_ctxp
= child_ctx
;
5211 get_task_struct(child
);
5214 ret
= inherit_group(event
, parent
, parent_ctx
,
5222 if (child_ctx
&& inherited_all
) {
5224 * Mark the child context as a clone of the parent
5225 * context, or of whatever the parent is a clone of.
5226 * Note that if the parent is a clone, it could get
5227 * uncloned at any point, but that doesn't matter
5228 * because the list of events and the generation
5229 * count can't have changed since we took the mutex.
5231 cloned_ctx
= rcu_dereference(parent_ctx
->parent_ctx
);
5233 child_ctx
->parent_ctx
= cloned_ctx
;
5234 child_ctx
->parent_gen
= parent_ctx
->parent_gen
;
5236 child_ctx
->parent_ctx
= parent_ctx
;
5237 child_ctx
->parent_gen
= parent_ctx
->generation
;
5239 get_ctx(child_ctx
->parent_ctx
);
5242 mutex_unlock(&parent_ctx
->mutex
);
5244 perf_unpin_context(parent_ctx
);
5249 static void __init
perf_event_init_all_cpus(void)
5252 struct perf_cpu_context
*cpuctx
;
5254 for_each_possible_cpu(cpu
) {
5255 cpuctx
= &per_cpu(perf_cpu_context
, cpu
);
5256 __perf_event_init_context(&cpuctx
->ctx
, NULL
);
5260 static void __cpuinit
perf_event_init_cpu(int cpu
)
5262 struct perf_cpu_context
*cpuctx
;
5264 cpuctx
= &per_cpu(perf_cpu_context
, cpu
);
5266 spin_lock(&perf_resource_lock
);
5267 cpuctx
->max_pertask
= perf_max_events
- perf_reserved_percpu
;
5268 spin_unlock(&perf_resource_lock
);
5270 hw_perf_event_setup(cpu
);
5273 #ifdef CONFIG_HOTPLUG_CPU
5274 static void __perf_event_exit_cpu(void *info
)
5276 struct perf_cpu_context
*cpuctx
= &__get_cpu_var(perf_cpu_context
);
5277 struct perf_event_context
*ctx
= &cpuctx
->ctx
;
5278 struct perf_event
*event
, *tmp
;
5280 list_for_each_entry_safe(event
, tmp
, &ctx
->group_list
, group_entry
)
5281 __perf_event_remove_from_context(event
);
5283 static void perf_event_exit_cpu(int cpu
)
5285 struct perf_cpu_context
*cpuctx
= &per_cpu(perf_cpu_context
, cpu
);
5286 struct perf_event_context
*ctx
= &cpuctx
->ctx
;
5288 mutex_lock(&ctx
->mutex
);
5289 smp_call_function_single(cpu
, __perf_event_exit_cpu
, NULL
, 1);
5290 mutex_unlock(&ctx
->mutex
);
5293 static inline void perf_event_exit_cpu(int cpu
) { }
5296 static int __cpuinit
5297 perf_cpu_notify(struct notifier_block
*self
, unsigned long action
, void *hcpu
)
5299 unsigned int cpu
= (long)hcpu
;
5303 case CPU_UP_PREPARE
:
5304 case CPU_UP_PREPARE_FROZEN
:
5305 perf_event_init_cpu(cpu
);
5309 case CPU_ONLINE_FROZEN
:
5310 hw_perf_event_setup_online(cpu
);
5313 case CPU_DOWN_PREPARE
:
5314 case CPU_DOWN_PREPARE_FROZEN
:
5315 perf_event_exit_cpu(cpu
);
5326 * This has to have a higher priority than migration_notifier in sched.c.
5328 static struct notifier_block __cpuinitdata perf_cpu_nb
= {
5329 .notifier_call
= perf_cpu_notify
,
5333 void __init
perf_event_init(void)
5335 perf_event_init_all_cpus();
5336 perf_cpu_notify(&perf_cpu_nb
, (unsigned long)CPU_UP_PREPARE
,
5337 (void *)(long)smp_processor_id());
5338 perf_cpu_notify(&perf_cpu_nb
, (unsigned long)CPU_ONLINE
,
5339 (void *)(long)smp_processor_id());
5340 register_cpu_notifier(&perf_cpu_nb
);
5343 static ssize_t
perf_show_reserve_percpu(struct sysdev_class
*class, char *buf
)
5345 return sprintf(buf
, "%d\n", perf_reserved_percpu
);
5349 perf_set_reserve_percpu(struct sysdev_class
*class,
5353 struct perf_cpu_context
*cpuctx
;
5357 err
= strict_strtoul(buf
, 10, &val
);
5360 if (val
> perf_max_events
)
5363 spin_lock(&perf_resource_lock
);
5364 perf_reserved_percpu
= val
;
5365 for_each_online_cpu(cpu
) {
5366 cpuctx
= &per_cpu(perf_cpu_context
, cpu
);
5367 raw_spin_lock_irq(&cpuctx
->ctx
.lock
);
5368 mpt
= min(perf_max_events
- cpuctx
->ctx
.nr_events
,
5369 perf_max_events
- perf_reserved_percpu
);
5370 cpuctx
->max_pertask
= mpt
;
5371 raw_spin_unlock_irq(&cpuctx
->ctx
.lock
);
5373 spin_unlock(&perf_resource_lock
);
5378 static ssize_t
perf_show_overcommit(struct sysdev_class
*class, char *buf
)
5380 return sprintf(buf
, "%d\n", perf_overcommit
);
5384 perf_set_overcommit(struct sysdev_class
*class, const char *buf
, size_t count
)
5389 err
= strict_strtoul(buf
, 10, &val
);
5395 spin_lock(&perf_resource_lock
);
5396 perf_overcommit
= val
;
5397 spin_unlock(&perf_resource_lock
);
5402 static SYSDEV_CLASS_ATTR(
5405 perf_show_reserve_percpu
,
5406 perf_set_reserve_percpu
5409 static SYSDEV_CLASS_ATTR(
5412 perf_show_overcommit
,
5416 static struct attribute
*perfclass_attrs
[] = {
5417 &attr_reserve_percpu
.attr
,
5418 &attr_overcommit
.attr
,
5422 static struct attribute_group perfclass_attr_group
= {
5423 .attrs
= perfclass_attrs
,
5424 .name
= "perf_events",
5427 static int __init
perf_event_sysfs_init(void)
5429 return sysfs_create_group(&cpu_sysdev_class
.kset
.kobj
,
5430 &perfclass_attr_group
);
5432 device_initcall(perf_event_sysfs_init
);