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(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
)
1175 struct perf_cpu_context
*cpuctx
= &__get_cpu_var(perf_cpu_context
);
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
)
1257 int cpu
= smp_processor_id();
1258 struct perf_event
*event
;
1261 raw_spin_lock(&ctx
->lock
);
1263 if (likely(!ctx
->nr_events
))
1266 ctx
->timestamp
= perf_clock();
1271 * First go through the list and put on any pinned groups
1272 * in order to give them the best chance of going on.
1274 list_for_each_entry(event
, &ctx
->group_list
, group_entry
) {
1275 if (event
->state
<= PERF_EVENT_STATE_OFF
||
1276 !event
->attr
.pinned
)
1278 if (event
->cpu
!= -1 && event
->cpu
!= cpu
)
1281 if (group_can_go_on(event
, cpuctx
, 1))
1282 group_sched_in(event
, cpuctx
, ctx
, cpu
);
1285 * If this pinned group hasn't been scheduled,
1286 * put it in error state.
1288 if (event
->state
== PERF_EVENT_STATE_INACTIVE
) {
1289 update_group_times(event
);
1290 event
->state
= PERF_EVENT_STATE_ERROR
;
1294 list_for_each_entry(event
, &ctx
->group_list
, group_entry
) {
1296 * Ignore events in OFF or ERROR state, and
1297 * ignore pinned events since we did them already.
1299 if (event
->state
<= PERF_EVENT_STATE_OFF
||
1304 * Listen to the 'cpu' scheduling filter constraint
1307 if (event
->cpu
!= -1 && event
->cpu
!= cpu
)
1310 if (group_can_go_on(event
, cpuctx
, can_add_hw
))
1311 if (group_sched_in(event
, cpuctx
, ctx
, cpu
))
1316 raw_spin_unlock(&ctx
->lock
);
1320 * Called from scheduler to add the events of the current task
1321 * with interrupts disabled.
1323 * We restore the event value and then enable it.
1325 * This does not protect us against NMI, but enable()
1326 * sets the enabled bit in the control field of event _before_
1327 * accessing the event control register. If a NMI hits, then it will
1328 * keep the event running.
1330 void perf_event_task_sched_in(struct task_struct
*task
)
1332 struct perf_cpu_context
*cpuctx
= &__get_cpu_var(perf_cpu_context
);
1333 struct perf_event_context
*ctx
= task
->perf_event_ctxp
;
1337 if (cpuctx
->task_ctx
== ctx
)
1339 __perf_event_sched_in(ctx
, cpuctx
);
1340 cpuctx
->task_ctx
= ctx
;
1343 static void perf_event_cpu_sched_in(struct perf_cpu_context
*cpuctx
)
1345 struct perf_event_context
*ctx
= &cpuctx
->ctx
;
1347 __perf_event_sched_in(ctx
, cpuctx
);
1350 #define MAX_INTERRUPTS (~0ULL)
1352 static void perf_log_throttle(struct perf_event
*event
, int enable
);
1354 static void perf_adjust_period(struct perf_event
*event
, u64 events
)
1356 struct hw_perf_event
*hwc
= &event
->hw
;
1357 u64 period
, sample_period
;
1360 events
*= hwc
->sample_period
;
1361 period
= div64_u64(events
, event
->attr
.sample_freq
);
1363 delta
= (s64
)(period
- hwc
->sample_period
);
1364 delta
= (delta
+ 7) / 8; /* low pass filter */
1366 sample_period
= hwc
->sample_period
+ delta
;
1371 hwc
->sample_period
= sample_period
;
1374 static void perf_ctx_adjust_freq(struct perf_event_context
*ctx
)
1376 struct perf_event
*event
;
1377 struct hw_perf_event
*hwc
;
1378 u64 interrupts
, freq
;
1380 raw_spin_lock(&ctx
->lock
);
1381 list_for_each_entry_rcu(event
, &ctx
->event_list
, event_entry
) {
1382 if (event
->state
!= PERF_EVENT_STATE_ACTIVE
)
1385 if (event
->cpu
!= -1 && event
->cpu
!= smp_processor_id())
1390 interrupts
= hwc
->interrupts
;
1391 hwc
->interrupts
= 0;
1394 * unthrottle events on the tick
1396 if (interrupts
== MAX_INTERRUPTS
) {
1397 perf_log_throttle(event
, 1);
1398 event
->pmu
->unthrottle(event
);
1399 interrupts
= 2*sysctl_perf_event_sample_rate
/HZ
;
1402 if (!event
->attr
.freq
|| !event
->attr
.sample_freq
)
1406 * if the specified freq < HZ then we need to skip ticks
1408 if (event
->attr
.sample_freq
< HZ
) {
1409 freq
= event
->attr
.sample_freq
;
1411 hwc
->freq_count
+= freq
;
1412 hwc
->freq_interrupts
+= interrupts
;
1414 if (hwc
->freq_count
< HZ
)
1417 interrupts
= hwc
->freq_interrupts
;
1418 hwc
->freq_interrupts
= 0;
1419 hwc
->freq_count
-= HZ
;
1423 perf_adjust_period(event
, freq
* interrupts
);
1426 * In order to avoid being stalled by an (accidental) huge
1427 * sample period, force reset the sample period if we didn't
1428 * get any events in this freq period.
1432 event
->pmu
->disable(event
);
1433 atomic64_set(&hwc
->period_left
, 0);
1434 event
->pmu
->enable(event
);
1438 raw_spin_unlock(&ctx
->lock
);
1442 * Round-robin a context's events:
1444 static void rotate_ctx(struct perf_event_context
*ctx
)
1446 struct perf_event
*event
;
1448 if (!ctx
->nr_events
)
1451 raw_spin_lock(&ctx
->lock
);
1453 * Rotate the first entry last (works just fine for group events too):
1456 list_for_each_entry(event
, &ctx
->group_list
, group_entry
) {
1457 list_move_tail(&event
->group_entry
, &ctx
->group_list
);
1462 raw_spin_unlock(&ctx
->lock
);
1465 void perf_event_task_tick(struct task_struct
*curr
)
1467 struct perf_cpu_context
*cpuctx
;
1468 struct perf_event_context
*ctx
;
1470 if (!atomic_read(&nr_events
))
1473 cpuctx
= &__get_cpu_var(perf_cpu_context
);
1474 ctx
= curr
->perf_event_ctxp
;
1476 perf_ctx_adjust_freq(&cpuctx
->ctx
);
1478 perf_ctx_adjust_freq(ctx
);
1480 perf_event_cpu_sched_out(cpuctx
);
1482 __perf_event_task_sched_out(ctx
);
1484 rotate_ctx(&cpuctx
->ctx
);
1488 perf_event_cpu_sched_in(cpuctx
);
1490 perf_event_task_sched_in(curr
);
1494 * Enable all of a task's events that have been marked enable-on-exec.
1495 * This expects task == current.
1497 static void perf_event_enable_on_exec(struct task_struct
*task
)
1499 struct perf_event_context
*ctx
;
1500 struct perf_event
*event
;
1501 unsigned long flags
;
1504 local_irq_save(flags
);
1505 ctx
= task
->perf_event_ctxp
;
1506 if (!ctx
|| !ctx
->nr_events
)
1509 __perf_event_task_sched_out(ctx
);
1511 raw_spin_lock(&ctx
->lock
);
1513 list_for_each_entry(event
, &ctx
->group_list
, group_entry
) {
1514 if (!event
->attr
.enable_on_exec
)
1516 event
->attr
.enable_on_exec
= 0;
1517 if (event
->state
>= PERF_EVENT_STATE_INACTIVE
)
1519 __perf_event_mark_enabled(event
, ctx
);
1524 * Unclone this context if we enabled any event.
1529 raw_spin_unlock(&ctx
->lock
);
1531 perf_event_task_sched_in(task
);
1533 local_irq_restore(flags
);
1537 * Cross CPU call to read the hardware event
1539 static void __perf_event_read(void *info
)
1541 struct perf_cpu_context
*cpuctx
= &__get_cpu_var(perf_cpu_context
);
1542 struct perf_event
*event
= info
;
1543 struct perf_event_context
*ctx
= event
->ctx
;
1546 * If this is a task context, we need to check whether it is
1547 * the current task context of this cpu. If not it has been
1548 * scheduled out before the smp call arrived. In that case
1549 * event->count would have been updated to a recent sample
1550 * when the event was scheduled out.
1552 if (ctx
->task
&& cpuctx
->task_ctx
!= ctx
)
1555 raw_spin_lock(&ctx
->lock
);
1556 update_context_time(ctx
);
1557 update_event_times(event
);
1558 raw_spin_unlock(&ctx
->lock
);
1560 event
->pmu
->read(event
);
1563 static u64
perf_event_read(struct perf_event
*event
)
1566 * If event is enabled and currently active on a CPU, update the
1567 * value in the event structure:
1569 if (event
->state
== PERF_EVENT_STATE_ACTIVE
) {
1570 smp_call_function_single(event
->oncpu
,
1571 __perf_event_read
, event
, 1);
1572 } else if (event
->state
== PERF_EVENT_STATE_INACTIVE
) {
1573 struct perf_event_context
*ctx
= event
->ctx
;
1574 unsigned long flags
;
1576 raw_spin_lock_irqsave(&ctx
->lock
, flags
);
1577 update_context_time(ctx
);
1578 update_event_times(event
);
1579 raw_spin_unlock_irqrestore(&ctx
->lock
, flags
);
1582 return atomic64_read(&event
->count
);
1586 * Initialize the perf_event context in a task_struct:
1589 __perf_event_init_context(struct perf_event_context
*ctx
,
1590 struct task_struct
*task
)
1592 raw_spin_lock_init(&ctx
->lock
);
1593 mutex_init(&ctx
->mutex
);
1594 INIT_LIST_HEAD(&ctx
->group_list
);
1595 INIT_LIST_HEAD(&ctx
->event_list
);
1596 atomic_set(&ctx
->refcount
, 1);
1600 static struct perf_event_context
*find_get_context(pid_t pid
, int cpu
)
1602 struct perf_event_context
*ctx
;
1603 struct perf_cpu_context
*cpuctx
;
1604 struct task_struct
*task
;
1605 unsigned long flags
;
1608 if (pid
== -1 && cpu
!= -1) {
1609 /* Must be root to operate on a CPU event: */
1610 if (perf_paranoid_cpu() && !capable(CAP_SYS_ADMIN
))
1611 return ERR_PTR(-EACCES
);
1613 if (cpu
< 0 || cpu
>= nr_cpumask_bits
)
1614 return ERR_PTR(-EINVAL
);
1617 * We could be clever and allow to attach a event to an
1618 * offline CPU and activate it when the CPU comes up, but
1621 if (!cpu_online(cpu
))
1622 return ERR_PTR(-ENODEV
);
1624 cpuctx
= &per_cpu(perf_cpu_context
, cpu
);
1635 task
= find_task_by_vpid(pid
);
1637 get_task_struct(task
);
1641 return ERR_PTR(-ESRCH
);
1644 * Can't attach events to a dying task.
1647 if (task
->flags
& PF_EXITING
)
1650 /* Reuse ptrace permission checks for now. */
1652 if (!ptrace_may_access(task
, PTRACE_MODE_READ
))
1656 ctx
= perf_lock_task_context(task
, &flags
);
1659 raw_spin_unlock_irqrestore(&ctx
->lock
, flags
);
1663 ctx
= kzalloc(sizeof(struct perf_event_context
), GFP_KERNEL
);
1667 __perf_event_init_context(ctx
, task
);
1669 if (cmpxchg(&task
->perf_event_ctxp
, NULL
, ctx
)) {
1671 * We raced with some other task; use
1672 * the context they set.
1677 get_task_struct(task
);
1680 put_task_struct(task
);
1684 put_task_struct(task
);
1685 return ERR_PTR(err
);
1688 static void perf_event_free_filter(struct perf_event
*event
);
1690 static void free_event_rcu(struct rcu_head
*head
)
1692 struct perf_event
*event
;
1694 event
= container_of(head
, struct perf_event
, rcu_head
);
1696 put_pid_ns(event
->ns
);
1697 perf_event_free_filter(event
);
1701 static void perf_pending_sync(struct perf_event
*event
);
1703 static void free_event(struct perf_event
*event
)
1705 perf_pending_sync(event
);
1707 if (!event
->parent
) {
1708 atomic_dec(&nr_events
);
1709 if (event
->attr
.mmap
)
1710 atomic_dec(&nr_mmap_events
);
1711 if (event
->attr
.comm
)
1712 atomic_dec(&nr_comm_events
);
1713 if (event
->attr
.task
)
1714 atomic_dec(&nr_task_events
);
1717 if (event
->output
) {
1718 fput(event
->output
->filp
);
1719 event
->output
= NULL
;
1723 event
->destroy(event
);
1725 put_ctx(event
->ctx
);
1726 call_rcu(&event
->rcu_head
, free_event_rcu
);
1729 int perf_event_release_kernel(struct perf_event
*event
)
1731 struct perf_event_context
*ctx
= event
->ctx
;
1733 WARN_ON_ONCE(ctx
->parent_ctx
);
1734 mutex_lock(&ctx
->mutex
);
1735 perf_event_remove_from_context(event
);
1736 mutex_unlock(&ctx
->mutex
);
1738 mutex_lock(&event
->owner
->perf_event_mutex
);
1739 list_del_init(&event
->owner_entry
);
1740 mutex_unlock(&event
->owner
->perf_event_mutex
);
1741 put_task_struct(event
->owner
);
1747 EXPORT_SYMBOL_GPL(perf_event_release_kernel
);
1750 * Called when the last reference to the file is gone.
1752 static int perf_release(struct inode
*inode
, struct file
*file
)
1754 struct perf_event
*event
= file
->private_data
;
1756 file
->private_data
= NULL
;
1758 return perf_event_release_kernel(event
);
1761 static int perf_event_read_size(struct perf_event
*event
)
1763 int entry
= sizeof(u64
); /* value */
1767 if (event
->attr
.read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
)
1768 size
+= sizeof(u64
);
1770 if (event
->attr
.read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
)
1771 size
+= sizeof(u64
);
1773 if (event
->attr
.read_format
& PERF_FORMAT_ID
)
1774 entry
+= sizeof(u64
);
1776 if (event
->attr
.read_format
& PERF_FORMAT_GROUP
) {
1777 nr
+= event
->group_leader
->nr_siblings
;
1778 size
+= sizeof(u64
);
1786 u64
perf_event_read_value(struct perf_event
*event
, u64
*enabled
, u64
*running
)
1788 struct perf_event
*child
;
1794 mutex_lock(&event
->child_mutex
);
1795 total
+= perf_event_read(event
);
1796 *enabled
+= event
->total_time_enabled
+
1797 atomic64_read(&event
->child_total_time_enabled
);
1798 *running
+= event
->total_time_running
+
1799 atomic64_read(&event
->child_total_time_running
);
1801 list_for_each_entry(child
, &event
->child_list
, child_list
) {
1802 total
+= perf_event_read(child
);
1803 *enabled
+= child
->total_time_enabled
;
1804 *running
+= child
->total_time_running
;
1806 mutex_unlock(&event
->child_mutex
);
1810 EXPORT_SYMBOL_GPL(perf_event_read_value
);
1812 static int perf_event_read_group(struct perf_event
*event
,
1813 u64 read_format
, char __user
*buf
)
1815 struct perf_event
*leader
= event
->group_leader
, *sub
;
1816 int n
= 0, size
= 0, ret
= -EFAULT
;
1817 struct perf_event_context
*ctx
= leader
->ctx
;
1819 u64 count
, enabled
, running
;
1821 mutex_lock(&ctx
->mutex
);
1822 count
= perf_event_read_value(leader
, &enabled
, &running
);
1824 values
[n
++] = 1 + leader
->nr_siblings
;
1825 if (read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
)
1826 values
[n
++] = enabled
;
1827 if (read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
)
1828 values
[n
++] = running
;
1829 values
[n
++] = count
;
1830 if (read_format
& PERF_FORMAT_ID
)
1831 values
[n
++] = primary_event_id(leader
);
1833 size
= n
* sizeof(u64
);
1835 if (copy_to_user(buf
, values
, size
))
1840 list_for_each_entry(sub
, &leader
->sibling_list
, group_entry
) {
1843 values
[n
++] = perf_event_read_value(sub
, &enabled
, &running
);
1844 if (read_format
& PERF_FORMAT_ID
)
1845 values
[n
++] = primary_event_id(sub
);
1847 size
= n
* sizeof(u64
);
1849 if (copy_to_user(buf
+ ret
, values
, size
)) {
1857 mutex_unlock(&ctx
->mutex
);
1862 static int perf_event_read_one(struct perf_event
*event
,
1863 u64 read_format
, char __user
*buf
)
1865 u64 enabled
, running
;
1869 values
[n
++] = perf_event_read_value(event
, &enabled
, &running
);
1870 if (read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
)
1871 values
[n
++] = enabled
;
1872 if (read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
)
1873 values
[n
++] = running
;
1874 if (read_format
& PERF_FORMAT_ID
)
1875 values
[n
++] = primary_event_id(event
);
1877 if (copy_to_user(buf
, values
, n
* sizeof(u64
)))
1880 return n
* sizeof(u64
);
1884 * Read the performance event - simple non blocking version for now
1887 perf_read_hw(struct perf_event
*event
, char __user
*buf
, size_t count
)
1889 u64 read_format
= event
->attr
.read_format
;
1893 * Return end-of-file for a read on a event that is in
1894 * error state (i.e. because it was pinned but it couldn't be
1895 * scheduled on to the CPU at some point).
1897 if (event
->state
== PERF_EVENT_STATE_ERROR
)
1900 if (count
< perf_event_read_size(event
))
1903 WARN_ON_ONCE(event
->ctx
->parent_ctx
);
1904 if (read_format
& PERF_FORMAT_GROUP
)
1905 ret
= perf_event_read_group(event
, read_format
, buf
);
1907 ret
= perf_event_read_one(event
, read_format
, buf
);
1913 perf_read(struct file
*file
, char __user
*buf
, size_t count
, loff_t
*ppos
)
1915 struct perf_event
*event
= file
->private_data
;
1917 return perf_read_hw(event
, buf
, count
);
1920 static unsigned int perf_poll(struct file
*file
, poll_table
*wait
)
1922 struct perf_event
*event
= file
->private_data
;
1923 struct perf_mmap_data
*data
;
1924 unsigned int events
= POLL_HUP
;
1927 data
= rcu_dereference(event
->data
);
1929 events
= atomic_xchg(&data
->poll
, 0);
1932 poll_wait(file
, &event
->waitq
, wait
);
1937 static void perf_event_reset(struct perf_event
*event
)
1939 (void)perf_event_read(event
);
1940 atomic64_set(&event
->count
, 0);
1941 perf_event_update_userpage(event
);
1945 * Holding the top-level event's child_mutex means that any
1946 * descendant process that has inherited this event will block
1947 * in sync_child_event if it goes to exit, thus satisfying the
1948 * task existence requirements of perf_event_enable/disable.
1950 static void perf_event_for_each_child(struct perf_event
*event
,
1951 void (*func
)(struct perf_event
*))
1953 struct perf_event
*child
;
1955 WARN_ON_ONCE(event
->ctx
->parent_ctx
);
1956 mutex_lock(&event
->child_mutex
);
1958 list_for_each_entry(child
, &event
->child_list
, child_list
)
1960 mutex_unlock(&event
->child_mutex
);
1963 static void perf_event_for_each(struct perf_event
*event
,
1964 void (*func
)(struct perf_event
*))
1966 struct perf_event_context
*ctx
= event
->ctx
;
1967 struct perf_event
*sibling
;
1969 WARN_ON_ONCE(ctx
->parent_ctx
);
1970 mutex_lock(&ctx
->mutex
);
1971 event
= event
->group_leader
;
1973 perf_event_for_each_child(event
, func
);
1975 list_for_each_entry(sibling
, &event
->sibling_list
, group_entry
)
1976 perf_event_for_each_child(event
, func
);
1977 mutex_unlock(&ctx
->mutex
);
1980 static int perf_event_period(struct perf_event
*event
, u64 __user
*arg
)
1982 struct perf_event_context
*ctx
= event
->ctx
;
1987 if (!event
->attr
.sample_period
)
1990 size
= copy_from_user(&value
, arg
, sizeof(value
));
1991 if (size
!= sizeof(value
))
1997 raw_spin_lock_irq(&ctx
->lock
);
1998 if (event
->attr
.freq
) {
1999 if (value
> sysctl_perf_event_sample_rate
) {
2004 event
->attr
.sample_freq
= value
;
2006 event
->attr
.sample_period
= value
;
2007 event
->hw
.sample_period
= value
;
2010 raw_spin_unlock_irq(&ctx
->lock
);
2015 static int perf_event_set_output(struct perf_event
*event
, int output_fd
);
2016 static int perf_event_set_filter(struct perf_event
*event
, void __user
*arg
);
2018 static long perf_ioctl(struct file
*file
, unsigned int cmd
, unsigned long arg
)
2020 struct perf_event
*event
= file
->private_data
;
2021 void (*func
)(struct perf_event
*);
2025 case PERF_EVENT_IOC_ENABLE
:
2026 func
= perf_event_enable
;
2028 case PERF_EVENT_IOC_DISABLE
:
2029 func
= perf_event_disable
;
2031 case PERF_EVENT_IOC_RESET
:
2032 func
= perf_event_reset
;
2035 case PERF_EVENT_IOC_REFRESH
:
2036 return perf_event_refresh(event
, arg
);
2038 case PERF_EVENT_IOC_PERIOD
:
2039 return perf_event_period(event
, (u64 __user
*)arg
);
2041 case PERF_EVENT_IOC_SET_OUTPUT
:
2042 return perf_event_set_output(event
, arg
);
2044 case PERF_EVENT_IOC_SET_FILTER
:
2045 return perf_event_set_filter(event
, (void __user
*)arg
);
2051 if (flags
& PERF_IOC_FLAG_GROUP
)
2052 perf_event_for_each(event
, func
);
2054 perf_event_for_each_child(event
, func
);
2059 int perf_event_task_enable(void)
2061 struct perf_event
*event
;
2063 mutex_lock(¤t
->perf_event_mutex
);
2064 list_for_each_entry(event
, ¤t
->perf_event_list
, owner_entry
)
2065 perf_event_for_each_child(event
, perf_event_enable
);
2066 mutex_unlock(¤t
->perf_event_mutex
);
2071 int perf_event_task_disable(void)
2073 struct perf_event
*event
;
2075 mutex_lock(¤t
->perf_event_mutex
);
2076 list_for_each_entry(event
, ¤t
->perf_event_list
, owner_entry
)
2077 perf_event_for_each_child(event
, perf_event_disable
);
2078 mutex_unlock(¤t
->perf_event_mutex
);
2083 #ifndef PERF_EVENT_INDEX_OFFSET
2084 # define PERF_EVENT_INDEX_OFFSET 0
2087 static int perf_event_index(struct perf_event
*event
)
2089 if (event
->state
!= PERF_EVENT_STATE_ACTIVE
)
2092 return event
->hw
.idx
+ 1 - PERF_EVENT_INDEX_OFFSET
;
2096 * Callers need to ensure there can be no nesting of this function, otherwise
2097 * the seqlock logic goes bad. We can not serialize this because the arch
2098 * code calls this from NMI context.
2100 void perf_event_update_userpage(struct perf_event
*event
)
2102 struct perf_event_mmap_page
*userpg
;
2103 struct perf_mmap_data
*data
;
2106 data
= rcu_dereference(event
->data
);
2110 userpg
= data
->user_page
;
2113 * Disable preemption so as to not let the corresponding user-space
2114 * spin too long if we get preempted.
2119 userpg
->index
= perf_event_index(event
);
2120 userpg
->offset
= atomic64_read(&event
->count
);
2121 if (event
->state
== PERF_EVENT_STATE_ACTIVE
)
2122 userpg
->offset
-= atomic64_read(&event
->hw
.prev_count
);
2124 userpg
->time_enabled
= event
->total_time_enabled
+
2125 atomic64_read(&event
->child_total_time_enabled
);
2127 userpg
->time_running
= event
->total_time_running
+
2128 atomic64_read(&event
->child_total_time_running
);
2137 static unsigned long perf_data_size(struct perf_mmap_data
*data
)
2139 return data
->nr_pages
<< (PAGE_SHIFT
+ data
->data_order
);
2142 #ifndef CONFIG_PERF_USE_VMALLOC
2145 * Back perf_mmap() with regular GFP_KERNEL-0 pages.
2148 static struct page
*
2149 perf_mmap_to_page(struct perf_mmap_data
*data
, unsigned long pgoff
)
2151 if (pgoff
> data
->nr_pages
)
2155 return virt_to_page(data
->user_page
);
2157 return virt_to_page(data
->data_pages
[pgoff
- 1]);
2160 static struct perf_mmap_data
*
2161 perf_mmap_data_alloc(struct perf_event
*event
, int nr_pages
)
2163 struct perf_mmap_data
*data
;
2167 WARN_ON(atomic_read(&event
->mmap_count
));
2169 size
= sizeof(struct perf_mmap_data
);
2170 size
+= nr_pages
* sizeof(void *);
2172 data
= kzalloc(size
, GFP_KERNEL
);
2176 data
->user_page
= (void *)get_zeroed_page(GFP_KERNEL
);
2177 if (!data
->user_page
)
2178 goto fail_user_page
;
2180 for (i
= 0; i
< nr_pages
; i
++) {
2181 data
->data_pages
[i
] = (void *)get_zeroed_page(GFP_KERNEL
);
2182 if (!data
->data_pages
[i
])
2183 goto fail_data_pages
;
2186 data
->data_order
= 0;
2187 data
->nr_pages
= nr_pages
;
2192 for (i
--; i
>= 0; i
--)
2193 free_page((unsigned long)data
->data_pages
[i
]);
2195 free_page((unsigned long)data
->user_page
);
2204 static void perf_mmap_free_page(unsigned long addr
)
2206 struct page
*page
= virt_to_page((void *)addr
);
2208 page
->mapping
= NULL
;
2212 static void perf_mmap_data_free(struct perf_mmap_data
*data
)
2216 perf_mmap_free_page((unsigned long)data
->user_page
);
2217 for (i
= 0; i
< data
->nr_pages
; i
++)
2218 perf_mmap_free_page((unsigned long)data
->data_pages
[i
]);
2225 * Back perf_mmap() with vmalloc memory.
2227 * Required for architectures that have d-cache aliasing issues.
2230 static struct page
*
2231 perf_mmap_to_page(struct perf_mmap_data
*data
, unsigned long pgoff
)
2233 if (pgoff
> (1UL << data
->data_order
))
2236 return vmalloc_to_page((void *)data
->user_page
+ pgoff
* PAGE_SIZE
);
2239 static void perf_mmap_unmark_page(void *addr
)
2241 struct page
*page
= vmalloc_to_page(addr
);
2243 page
->mapping
= NULL
;
2246 static void perf_mmap_data_free_work(struct work_struct
*work
)
2248 struct perf_mmap_data
*data
;
2252 data
= container_of(work
, struct perf_mmap_data
, work
);
2253 nr
= 1 << data
->data_order
;
2255 base
= data
->user_page
;
2256 for (i
= 0; i
< nr
+ 1; i
++)
2257 perf_mmap_unmark_page(base
+ (i
* PAGE_SIZE
));
2263 static void perf_mmap_data_free(struct perf_mmap_data
*data
)
2265 schedule_work(&data
->work
);
2268 static struct perf_mmap_data
*
2269 perf_mmap_data_alloc(struct perf_event
*event
, int nr_pages
)
2271 struct perf_mmap_data
*data
;
2275 WARN_ON(atomic_read(&event
->mmap_count
));
2277 size
= sizeof(struct perf_mmap_data
);
2278 size
+= sizeof(void *);
2280 data
= kzalloc(size
, GFP_KERNEL
);
2284 INIT_WORK(&data
->work
, perf_mmap_data_free_work
);
2286 all_buf
= vmalloc_user((nr_pages
+ 1) * PAGE_SIZE
);
2290 data
->user_page
= all_buf
;
2291 data
->data_pages
[0] = all_buf
+ PAGE_SIZE
;
2292 data
->data_order
= ilog2(nr_pages
);
2306 static int perf_mmap_fault(struct vm_area_struct
*vma
, struct vm_fault
*vmf
)
2308 struct perf_event
*event
= vma
->vm_file
->private_data
;
2309 struct perf_mmap_data
*data
;
2310 int ret
= VM_FAULT_SIGBUS
;
2312 if (vmf
->flags
& FAULT_FLAG_MKWRITE
) {
2313 if (vmf
->pgoff
== 0)
2319 data
= rcu_dereference(event
->data
);
2323 if (vmf
->pgoff
&& (vmf
->flags
& FAULT_FLAG_WRITE
))
2326 vmf
->page
= perf_mmap_to_page(data
, vmf
->pgoff
);
2330 get_page(vmf
->page
);
2331 vmf
->page
->mapping
= vma
->vm_file
->f_mapping
;
2332 vmf
->page
->index
= vmf
->pgoff
;
2342 perf_mmap_data_init(struct perf_event
*event
, struct perf_mmap_data
*data
)
2344 long max_size
= perf_data_size(data
);
2346 atomic_set(&data
->lock
, -1);
2348 if (event
->attr
.watermark
) {
2349 data
->watermark
= min_t(long, max_size
,
2350 event
->attr
.wakeup_watermark
);
2353 if (!data
->watermark
)
2354 data
->watermark
= max_size
/ 2;
2357 rcu_assign_pointer(event
->data
, data
);
2360 static void perf_mmap_data_free_rcu(struct rcu_head
*rcu_head
)
2362 struct perf_mmap_data
*data
;
2364 data
= container_of(rcu_head
, struct perf_mmap_data
, rcu_head
);
2365 perf_mmap_data_free(data
);
2368 static void perf_mmap_data_release(struct perf_event
*event
)
2370 struct perf_mmap_data
*data
= event
->data
;
2372 WARN_ON(atomic_read(&event
->mmap_count
));
2374 rcu_assign_pointer(event
->data
, NULL
);
2375 call_rcu(&data
->rcu_head
, perf_mmap_data_free_rcu
);
2378 static void perf_mmap_open(struct vm_area_struct
*vma
)
2380 struct perf_event
*event
= vma
->vm_file
->private_data
;
2382 atomic_inc(&event
->mmap_count
);
2385 static void perf_mmap_close(struct vm_area_struct
*vma
)
2387 struct perf_event
*event
= vma
->vm_file
->private_data
;
2389 WARN_ON_ONCE(event
->ctx
->parent_ctx
);
2390 if (atomic_dec_and_mutex_lock(&event
->mmap_count
, &event
->mmap_mutex
)) {
2391 unsigned long size
= perf_data_size(event
->data
);
2392 struct user_struct
*user
= current_user();
2394 atomic_long_sub((size
>> PAGE_SHIFT
) + 1, &user
->locked_vm
);
2395 vma
->vm_mm
->locked_vm
-= event
->data
->nr_locked
;
2396 perf_mmap_data_release(event
);
2397 mutex_unlock(&event
->mmap_mutex
);
2401 static const struct vm_operations_struct perf_mmap_vmops
= {
2402 .open
= perf_mmap_open
,
2403 .close
= perf_mmap_close
,
2404 .fault
= perf_mmap_fault
,
2405 .page_mkwrite
= perf_mmap_fault
,
2408 static int perf_mmap(struct file
*file
, struct vm_area_struct
*vma
)
2410 struct perf_event
*event
= file
->private_data
;
2411 unsigned long user_locked
, user_lock_limit
;
2412 struct user_struct
*user
= current_user();
2413 unsigned long locked
, lock_limit
;
2414 struct perf_mmap_data
*data
;
2415 unsigned long vma_size
;
2416 unsigned long nr_pages
;
2417 long user_extra
, extra
;
2420 if (!(vma
->vm_flags
& VM_SHARED
))
2423 vma_size
= vma
->vm_end
- vma
->vm_start
;
2424 nr_pages
= (vma_size
/ PAGE_SIZE
) - 1;
2427 * If we have data pages ensure they're a power-of-two number, so we
2428 * can do bitmasks instead of modulo.
2430 if (nr_pages
!= 0 && !is_power_of_2(nr_pages
))
2433 if (vma_size
!= PAGE_SIZE
* (1 + nr_pages
))
2436 if (vma
->vm_pgoff
!= 0)
2439 WARN_ON_ONCE(event
->ctx
->parent_ctx
);
2440 mutex_lock(&event
->mmap_mutex
);
2441 if (event
->output
) {
2446 if (atomic_inc_not_zero(&event
->mmap_count
)) {
2447 if (nr_pages
!= event
->data
->nr_pages
)
2452 user_extra
= nr_pages
+ 1;
2453 user_lock_limit
= sysctl_perf_event_mlock
>> (PAGE_SHIFT
- 10);
2456 * Increase the limit linearly with more CPUs:
2458 user_lock_limit
*= num_online_cpus();
2460 user_locked
= atomic_long_read(&user
->locked_vm
) + user_extra
;
2463 if (user_locked
> user_lock_limit
)
2464 extra
= user_locked
- user_lock_limit
;
2466 lock_limit
= current
->signal
->rlim
[RLIMIT_MEMLOCK
].rlim_cur
;
2467 lock_limit
>>= PAGE_SHIFT
;
2468 locked
= vma
->vm_mm
->locked_vm
+ extra
;
2470 if ((locked
> lock_limit
) && perf_paranoid_tracepoint_raw() &&
2471 !capable(CAP_IPC_LOCK
)) {
2476 WARN_ON(event
->data
);
2478 data
= perf_mmap_data_alloc(event
, nr_pages
);
2484 perf_mmap_data_init(event
, data
);
2486 atomic_set(&event
->mmap_count
, 1);
2487 atomic_long_add(user_extra
, &user
->locked_vm
);
2488 vma
->vm_mm
->locked_vm
+= extra
;
2489 event
->data
->nr_locked
= extra
;
2490 if (vma
->vm_flags
& VM_WRITE
)
2491 event
->data
->writable
= 1;
2494 mutex_unlock(&event
->mmap_mutex
);
2496 vma
->vm_flags
|= VM_RESERVED
;
2497 vma
->vm_ops
= &perf_mmap_vmops
;
2502 static int perf_fasync(int fd
, struct file
*filp
, int on
)
2504 struct inode
*inode
= filp
->f_path
.dentry
->d_inode
;
2505 struct perf_event
*event
= filp
->private_data
;
2508 mutex_lock(&inode
->i_mutex
);
2509 retval
= fasync_helper(fd
, filp
, on
, &event
->fasync
);
2510 mutex_unlock(&inode
->i_mutex
);
2518 static const struct file_operations perf_fops
= {
2519 .release
= perf_release
,
2522 .unlocked_ioctl
= perf_ioctl
,
2523 .compat_ioctl
= perf_ioctl
,
2525 .fasync
= perf_fasync
,
2531 * If there's data, ensure we set the poll() state and publish everything
2532 * to user-space before waking everybody up.
2535 void perf_event_wakeup(struct perf_event
*event
)
2537 wake_up_all(&event
->waitq
);
2539 if (event
->pending_kill
) {
2540 kill_fasync(&event
->fasync
, SIGIO
, event
->pending_kill
);
2541 event
->pending_kill
= 0;
2548 * Handle the case where we need to wakeup up from NMI (or rq->lock) context.
2550 * The NMI bit means we cannot possibly take locks. Therefore, maintain a
2551 * single linked list and use cmpxchg() to add entries lockless.
2554 static void perf_pending_event(struct perf_pending_entry
*entry
)
2556 struct perf_event
*event
= container_of(entry
,
2557 struct perf_event
, pending
);
2559 if (event
->pending_disable
) {
2560 event
->pending_disable
= 0;
2561 __perf_event_disable(event
);
2564 if (event
->pending_wakeup
) {
2565 event
->pending_wakeup
= 0;
2566 perf_event_wakeup(event
);
2570 #define PENDING_TAIL ((struct perf_pending_entry *)-1UL)
2572 static DEFINE_PER_CPU(struct perf_pending_entry
*, perf_pending_head
) = {
2576 static void perf_pending_queue(struct perf_pending_entry
*entry
,
2577 void (*func
)(struct perf_pending_entry
*))
2579 struct perf_pending_entry
**head
;
2581 if (cmpxchg(&entry
->next
, NULL
, PENDING_TAIL
) != NULL
)
2586 head
= &get_cpu_var(perf_pending_head
);
2589 entry
->next
= *head
;
2590 } while (cmpxchg(head
, entry
->next
, entry
) != entry
->next
);
2592 set_perf_event_pending();
2594 put_cpu_var(perf_pending_head
);
2597 static int __perf_pending_run(void)
2599 struct perf_pending_entry
*list
;
2602 list
= xchg(&__get_cpu_var(perf_pending_head
), PENDING_TAIL
);
2603 while (list
!= PENDING_TAIL
) {
2604 void (*func
)(struct perf_pending_entry
*);
2605 struct perf_pending_entry
*entry
= list
;
2612 * Ensure we observe the unqueue before we issue the wakeup,
2613 * so that we won't be waiting forever.
2614 * -- see perf_not_pending().
2625 static inline int perf_not_pending(struct perf_event
*event
)
2628 * If we flush on whatever cpu we run, there is a chance we don't
2632 __perf_pending_run();
2636 * Ensure we see the proper queue state before going to sleep
2637 * so that we do not miss the wakeup. -- see perf_pending_handle()
2640 return event
->pending
.next
== NULL
;
2643 static void perf_pending_sync(struct perf_event
*event
)
2645 wait_event(event
->waitq
, perf_not_pending(event
));
2648 void perf_event_do_pending(void)
2650 __perf_pending_run();
2654 * Callchain support -- arch specific
2657 __weak
struct perf_callchain_entry
*perf_callchain(struct pt_regs
*regs
)
2665 static bool perf_output_space(struct perf_mmap_data
*data
, unsigned long tail
,
2666 unsigned long offset
, unsigned long head
)
2670 if (!data
->writable
)
2673 mask
= perf_data_size(data
) - 1;
2675 offset
= (offset
- tail
) & mask
;
2676 head
= (head
- tail
) & mask
;
2678 if ((int)(head
- offset
) < 0)
2684 static void perf_output_wakeup(struct perf_output_handle
*handle
)
2686 atomic_set(&handle
->data
->poll
, POLL_IN
);
2689 handle
->event
->pending_wakeup
= 1;
2690 perf_pending_queue(&handle
->event
->pending
,
2691 perf_pending_event
);
2693 perf_event_wakeup(handle
->event
);
2697 * Curious locking construct.
2699 * We need to ensure a later event_id doesn't publish a head when a former
2700 * event_id isn't done writing. However since we need to deal with NMIs we
2701 * cannot fully serialize things.
2703 * What we do is serialize between CPUs so we only have to deal with NMI
2704 * nesting on a single CPU.
2706 * We only publish the head (and generate a wakeup) when the outer-most
2707 * event_id completes.
2709 static void perf_output_lock(struct perf_output_handle
*handle
)
2711 struct perf_mmap_data
*data
= handle
->data
;
2712 int cur
, cpu
= get_cpu();
2717 cur
= atomic_cmpxchg(&data
->lock
, -1, cpu
);
2729 static void perf_output_unlock(struct perf_output_handle
*handle
)
2731 struct perf_mmap_data
*data
= handle
->data
;
2735 data
->done_head
= data
->head
;
2737 if (!handle
->locked
)
2742 * The xchg implies a full barrier that ensures all writes are done
2743 * before we publish the new head, matched by a rmb() in userspace when
2744 * reading this position.
2746 while ((head
= atomic_long_xchg(&data
->done_head
, 0)))
2747 data
->user_page
->data_head
= head
;
2750 * NMI can happen here, which means we can miss a done_head update.
2753 cpu
= atomic_xchg(&data
->lock
, -1);
2754 WARN_ON_ONCE(cpu
!= smp_processor_id());
2757 * Therefore we have to validate we did not indeed do so.
2759 if (unlikely(atomic_long_read(&data
->done_head
))) {
2761 * Since we had it locked, we can lock it again.
2763 while (atomic_cmpxchg(&data
->lock
, -1, cpu
) != -1)
2769 if (atomic_xchg(&data
->wakeup
, 0))
2770 perf_output_wakeup(handle
);
2775 void perf_output_copy(struct perf_output_handle
*handle
,
2776 const void *buf
, unsigned int len
)
2778 unsigned int pages_mask
;
2779 unsigned long offset
;
2783 offset
= handle
->offset
;
2784 pages_mask
= handle
->data
->nr_pages
- 1;
2785 pages
= handle
->data
->data_pages
;
2788 unsigned long page_offset
;
2789 unsigned long page_size
;
2792 nr
= (offset
>> PAGE_SHIFT
) & pages_mask
;
2793 page_size
= 1UL << (handle
->data
->data_order
+ PAGE_SHIFT
);
2794 page_offset
= offset
& (page_size
- 1);
2795 size
= min_t(unsigned int, page_size
- page_offset
, len
);
2797 memcpy(pages
[nr
] + page_offset
, buf
, size
);
2804 handle
->offset
= offset
;
2807 * Check we didn't copy past our reservation window, taking the
2808 * possible unsigned int wrap into account.
2810 WARN_ON_ONCE(((long)(handle
->head
- handle
->offset
)) < 0);
2813 int perf_output_begin(struct perf_output_handle
*handle
,
2814 struct perf_event
*event
, unsigned int size
,
2815 int nmi
, int sample
)
2817 struct perf_event
*output_event
;
2818 struct perf_mmap_data
*data
;
2819 unsigned long tail
, offset
, head
;
2822 struct perf_event_header header
;
2829 * For inherited events we send all the output towards the parent.
2832 event
= event
->parent
;
2834 output_event
= rcu_dereference(event
->output
);
2836 event
= output_event
;
2838 data
= rcu_dereference(event
->data
);
2842 handle
->data
= data
;
2843 handle
->event
= event
;
2845 handle
->sample
= sample
;
2847 if (!data
->nr_pages
)
2850 have_lost
= atomic_read(&data
->lost
);
2852 size
+= sizeof(lost_event
);
2854 perf_output_lock(handle
);
2858 * Userspace could choose to issue a mb() before updating the
2859 * tail pointer. So that all reads will be completed before the
2862 tail
= ACCESS_ONCE(data
->user_page
->data_tail
);
2864 offset
= head
= atomic_long_read(&data
->head
);
2866 if (unlikely(!perf_output_space(data
, tail
, offset
, head
)))
2868 } while (atomic_long_cmpxchg(&data
->head
, offset
, head
) != offset
);
2870 handle
->offset
= offset
;
2871 handle
->head
= head
;
2873 if (head
- tail
> data
->watermark
)
2874 atomic_set(&data
->wakeup
, 1);
2877 lost_event
.header
.type
= PERF_RECORD_LOST
;
2878 lost_event
.header
.misc
= 0;
2879 lost_event
.header
.size
= sizeof(lost_event
);
2880 lost_event
.id
= event
->id
;
2881 lost_event
.lost
= atomic_xchg(&data
->lost
, 0);
2883 perf_output_put(handle
, lost_event
);
2889 atomic_inc(&data
->lost
);
2890 perf_output_unlock(handle
);
2897 void perf_output_end(struct perf_output_handle
*handle
)
2899 struct perf_event
*event
= handle
->event
;
2900 struct perf_mmap_data
*data
= handle
->data
;
2902 int wakeup_events
= event
->attr
.wakeup_events
;
2904 if (handle
->sample
&& wakeup_events
) {
2905 int events
= atomic_inc_return(&data
->events
);
2906 if (events
>= wakeup_events
) {
2907 atomic_sub(wakeup_events
, &data
->events
);
2908 atomic_set(&data
->wakeup
, 1);
2912 perf_output_unlock(handle
);
2916 static u32
perf_event_pid(struct perf_event
*event
, struct task_struct
*p
)
2919 * only top level events have the pid namespace they were created in
2922 event
= event
->parent
;
2924 return task_tgid_nr_ns(p
, event
->ns
);
2927 static u32
perf_event_tid(struct perf_event
*event
, struct task_struct
*p
)
2930 * only top level events have the pid namespace they were created in
2933 event
= event
->parent
;
2935 return task_pid_nr_ns(p
, event
->ns
);
2938 static void perf_output_read_one(struct perf_output_handle
*handle
,
2939 struct perf_event
*event
)
2941 u64 read_format
= event
->attr
.read_format
;
2945 values
[n
++] = atomic64_read(&event
->count
);
2946 if (read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
) {
2947 values
[n
++] = event
->total_time_enabled
+
2948 atomic64_read(&event
->child_total_time_enabled
);
2950 if (read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
) {
2951 values
[n
++] = event
->total_time_running
+
2952 atomic64_read(&event
->child_total_time_running
);
2954 if (read_format
& PERF_FORMAT_ID
)
2955 values
[n
++] = primary_event_id(event
);
2957 perf_output_copy(handle
, values
, n
* sizeof(u64
));
2961 * XXX PERF_FORMAT_GROUP vs inherited events seems difficult.
2963 static void perf_output_read_group(struct perf_output_handle
*handle
,
2964 struct perf_event
*event
)
2966 struct perf_event
*leader
= event
->group_leader
, *sub
;
2967 u64 read_format
= event
->attr
.read_format
;
2971 values
[n
++] = 1 + leader
->nr_siblings
;
2973 if (read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
)
2974 values
[n
++] = leader
->total_time_enabled
;
2976 if (read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
)
2977 values
[n
++] = leader
->total_time_running
;
2979 if (leader
!= event
)
2980 leader
->pmu
->read(leader
);
2982 values
[n
++] = atomic64_read(&leader
->count
);
2983 if (read_format
& PERF_FORMAT_ID
)
2984 values
[n
++] = primary_event_id(leader
);
2986 perf_output_copy(handle
, values
, n
* sizeof(u64
));
2988 list_for_each_entry(sub
, &leader
->sibling_list
, group_entry
) {
2992 sub
->pmu
->read(sub
);
2994 values
[n
++] = atomic64_read(&sub
->count
);
2995 if (read_format
& PERF_FORMAT_ID
)
2996 values
[n
++] = primary_event_id(sub
);
2998 perf_output_copy(handle
, values
, n
* sizeof(u64
));
3002 static void perf_output_read(struct perf_output_handle
*handle
,
3003 struct perf_event
*event
)
3005 if (event
->attr
.read_format
& PERF_FORMAT_GROUP
)
3006 perf_output_read_group(handle
, event
);
3008 perf_output_read_one(handle
, event
);
3011 void perf_output_sample(struct perf_output_handle
*handle
,
3012 struct perf_event_header
*header
,
3013 struct perf_sample_data
*data
,
3014 struct perf_event
*event
)
3016 u64 sample_type
= data
->type
;
3018 perf_output_put(handle
, *header
);
3020 if (sample_type
& PERF_SAMPLE_IP
)
3021 perf_output_put(handle
, data
->ip
);
3023 if (sample_type
& PERF_SAMPLE_TID
)
3024 perf_output_put(handle
, data
->tid_entry
);
3026 if (sample_type
& PERF_SAMPLE_TIME
)
3027 perf_output_put(handle
, data
->time
);
3029 if (sample_type
& PERF_SAMPLE_ADDR
)
3030 perf_output_put(handle
, data
->addr
);
3032 if (sample_type
& PERF_SAMPLE_ID
)
3033 perf_output_put(handle
, data
->id
);
3035 if (sample_type
& PERF_SAMPLE_STREAM_ID
)
3036 perf_output_put(handle
, data
->stream_id
);
3038 if (sample_type
& PERF_SAMPLE_CPU
)
3039 perf_output_put(handle
, data
->cpu_entry
);
3041 if (sample_type
& PERF_SAMPLE_PERIOD
)
3042 perf_output_put(handle
, data
->period
);
3044 if (sample_type
& PERF_SAMPLE_READ
)
3045 perf_output_read(handle
, event
);
3047 if (sample_type
& PERF_SAMPLE_CALLCHAIN
) {
3048 if (data
->callchain
) {
3051 if (data
->callchain
)
3052 size
+= data
->callchain
->nr
;
3054 size
*= sizeof(u64
);
3056 perf_output_copy(handle
, data
->callchain
, size
);
3059 perf_output_put(handle
, nr
);
3063 if (sample_type
& PERF_SAMPLE_RAW
) {
3065 perf_output_put(handle
, data
->raw
->size
);
3066 perf_output_copy(handle
, data
->raw
->data
,
3073 .size
= sizeof(u32
),
3076 perf_output_put(handle
, raw
);
3081 void perf_prepare_sample(struct perf_event_header
*header
,
3082 struct perf_sample_data
*data
,
3083 struct perf_event
*event
,
3084 struct pt_regs
*regs
)
3086 u64 sample_type
= event
->attr
.sample_type
;
3088 data
->type
= sample_type
;
3090 header
->type
= PERF_RECORD_SAMPLE
;
3091 header
->size
= sizeof(*header
);
3094 header
->misc
|= perf_misc_flags(regs
);
3096 if (sample_type
& PERF_SAMPLE_IP
) {
3097 data
->ip
= perf_instruction_pointer(regs
);
3099 header
->size
+= sizeof(data
->ip
);
3102 if (sample_type
& PERF_SAMPLE_TID
) {
3103 /* namespace issues */
3104 data
->tid_entry
.pid
= perf_event_pid(event
, current
);
3105 data
->tid_entry
.tid
= perf_event_tid(event
, current
);
3107 header
->size
+= sizeof(data
->tid_entry
);
3110 if (sample_type
& PERF_SAMPLE_TIME
) {
3111 data
->time
= perf_clock();
3113 header
->size
+= sizeof(data
->time
);
3116 if (sample_type
& PERF_SAMPLE_ADDR
)
3117 header
->size
+= sizeof(data
->addr
);
3119 if (sample_type
& PERF_SAMPLE_ID
) {
3120 data
->id
= primary_event_id(event
);
3122 header
->size
+= sizeof(data
->id
);
3125 if (sample_type
& PERF_SAMPLE_STREAM_ID
) {
3126 data
->stream_id
= event
->id
;
3128 header
->size
+= sizeof(data
->stream_id
);
3131 if (sample_type
& PERF_SAMPLE_CPU
) {
3132 data
->cpu_entry
.cpu
= raw_smp_processor_id();
3133 data
->cpu_entry
.reserved
= 0;
3135 header
->size
+= sizeof(data
->cpu_entry
);
3138 if (sample_type
& PERF_SAMPLE_PERIOD
)
3139 header
->size
+= sizeof(data
->period
);
3141 if (sample_type
& PERF_SAMPLE_READ
)
3142 header
->size
+= perf_event_read_size(event
);
3144 if (sample_type
& PERF_SAMPLE_CALLCHAIN
) {
3147 data
->callchain
= perf_callchain(regs
);
3149 if (data
->callchain
)
3150 size
+= data
->callchain
->nr
;
3152 header
->size
+= size
* sizeof(u64
);
3155 if (sample_type
& PERF_SAMPLE_RAW
) {
3156 int size
= sizeof(u32
);
3159 size
+= data
->raw
->size
;
3161 size
+= sizeof(u32
);
3163 WARN_ON_ONCE(size
& (sizeof(u64
)-1));
3164 header
->size
+= size
;
3168 static void perf_event_output(struct perf_event
*event
, int nmi
,
3169 struct perf_sample_data
*data
,
3170 struct pt_regs
*regs
)
3172 struct perf_output_handle handle
;
3173 struct perf_event_header header
;
3175 perf_prepare_sample(&header
, data
, event
, regs
);
3177 if (perf_output_begin(&handle
, event
, header
.size
, nmi
, 1))
3180 perf_output_sample(&handle
, &header
, data
, event
);
3182 perf_output_end(&handle
);
3189 struct perf_read_event
{
3190 struct perf_event_header header
;
3197 perf_event_read_event(struct perf_event
*event
,
3198 struct task_struct
*task
)
3200 struct perf_output_handle handle
;
3201 struct perf_read_event read_event
= {
3203 .type
= PERF_RECORD_READ
,
3205 .size
= sizeof(read_event
) + perf_event_read_size(event
),
3207 .pid
= perf_event_pid(event
, task
),
3208 .tid
= perf_event_tid(event
, task
),
3212 ret
= perf_output_begin(&handle
, event
, read_event
.header
.size
, 0, 0);
3216 perf_output_put(&handle
, read_event
);
3217 perf_output_read(&handle
, event
);
3219 perf_output_end(&handle
);
3223 * task tracking -- fork/exit
3225 * enabled by: attr.comm | attr.mmap | attr.task
3228 struct perf_task_event
{
3229 struct task_struct
*task
;
3230 struct perf_event_context
*task_ctx
;
3233 struct perf_event_header header
;
3243 static void perf_event_task_output(struct perf_event
*event
,
3244 struct perf_task_event
*task_event
)
3246 struct perf_output_handle handle
;
3248 struct task_struct
*task
= task_event
->task
;
3251 size
= task_event
->event_id
.header
.size
;
3252 ret
= perf_output_begin(&handle
, event
, size
, 0, 0);
3257 task_event
->event_id
.pid
= perf_event_pid(event
, task
);
3258 task_event
->event_id
.ppid
= perf_event_pid(event
, current
);
3260 task_event
->event_id
.tid
= perf_event_tid(event
, task
);
3261 task_event
->event_id
.ptid
= perf_event_tid(event
, current
);
3263 task_event
->event_id
.time
= perf_clock();
3265 perf_output_put(&handle
, task_event
->event_id
);
3267 perf_output_end(&handle
);
3270 static int perf_event_task_match(struct perf_event
*event
)
3272 if (event
->cpu
!= -1 && event
->cpu
!= smp_processor_id())
3275 if (event
->attr
.comm
|| event
->attr
.mmap
|| event
->attr
.task
)
3281 static void perf_event_task_ctx(struct perf_event_context
*ctx
,
3282 struct perf_task_event
*task_event
)
3284 struct perf_event
*event
;
3286 list_for_each_entry_rcu(event
, &ctx
->event_list
, event_entry
) {
3287 if (perf_event_task_match(event
))
3288 perf_event_task_output(event
, task_event
);
3292 static void perf_event_task_event(struct perf_task_event
*task_event
)
3294 struct perf_cpu_context
*cpuctx
;
3295 struct perf_event_context
*ctx
= task_event
->task_ctx
;
3298 cpuctx
= &get_cpu_var(perf_cpu_context
);
3299 perf_event_task_ctx(&cpuctx
->ctx
, task_event
);
3301 ctx
= rcu_dereference(task_event
->task
->perf_event_ctxp
);
3303 perf_event_task_ctx(ctx
, task_event
);
3304 put_cpu_var(perf_cpu_context
);
3308 static void perf_event_task(struct task_struct
*task
,
3309 struct perf_event_context
*task_ctx
,
3312 struct perf_task_event task_event
;
3314 if (!atomic_read(&nr_comm_events
) &&
3315 !atomic_read(&nr_mmap_events
) &&
3316 !atomic_read(&nr_task_events
))
3319 task_event
= (struct perf_task_event
){
3321 .task_ctx
= task_ctx
,
3324 .type
= new ? PERF_RECORD_FORK
: PERF_RECORD_EXIT
,
3326 .size
= sizeof(task_event
.event_id
),
3335 perf_event_task_event(&task_event
);
3338 void perf_event_fork(struct task_struct
*task
)
3340 perf_event_task(task
, NULL
, 1);
3347 struct perf_comm_event
{
3348 struct task_struct
*task
;
3353 struct perf_event_header header
;
3360 static void perf_event_comm_output(struct perf_event
*event
,
3361 struct perf_comm_event
*comm_event
)
3363 struct perf_output_handle handle
;
3364 int size
= comm_event
->event_id
.header
.size
;
3365 int ret
= perf_output_begin(&handle
, event
, size
, 0, 0);
3370 comm_event
->event_id
.pid
= perf_event_pid(event
, comm_event
->task
);
3371 comm_event
->event_id
.tid
= perf_event_tid(event
, comm_event
->task
);
3373 perf_output_put(&handle
, comm_event
->event_id
);
3374 perf_output_copy(&handle
, comm_event
->comm
,
3375 comm_event
->comm_size
);
3376 perf_output_end(&handle
);
3379 static int perf_event_comm_match(struct perf_event
*event
)
3381 if (event
->cpu
!= -1 && event
->cpu
!= smp_processor_id())
3384 if (event
->attr
.comm
)
3390 static void perf_event_comm_ctx(struct perf_event_context
*ctx
,
3391 struct perf_comm_event
*comm_event
)
3393 struct perf_event
*event
;
3395 list_for_each_entry_rcu(event
, &ctx
->event_list
, event_entry
) {
3396 if (perf_event_comm_match(event
))
3397 perf_event_comm_output(event
, comm_event
);
3401 static void perf_event_comm_event(struct perf_comm_event
*comm_event
)
3403 struct perf_cpu_context
*cpuctx
;
3404 struct perf_event_context
*ctx
;
3406 char comm
[TASK_COMM_LEN
];
3408 memset(comm
, 0, sizeof(comm
));
3409 strlcpy(comm
, comm_event
->task
->comm
, sizeof(comm
));
3410 size
= ALIGN(strlen(comm
)+1, sizeof(u64
));
3412 comm_event
->comm
= comm
;
3413 comm_event
->comm_size
= size
;
3415 comm_event
->event_id
.header
.size
= sizeof(comm_event
->event_id
) + size
;
3418 cpuctx
= &get_cpu_var(perf_cpu_context
);
3419 perf_event_comm_ctx(&cpuctx
->ctx
, comm_event
);
3420 ctx
= rcu_dereference(current
->perf_event_ctxp
);
3422 perf_event_comm_ctx(ctx
, comm_event
);
3423 put_cpu_var(perf_cpu_context
);
3427 void perf_event_comm(struct task_struct
*task
)
3429 struct perf_comm_event comm_event
;
3431 if (task
->perf_event_ctxp
)
3432 perf_event_enable_on_exec(task
);
3434 if (!atomic_read(&nr_comm_events
))
3437 comm_event
= (struct perf_comm_event
){
3443 .type
= PERF_RECORD_COMM
,
3452 perf_event_comm_event(&comm_event
);
3459 struct perf_mmap_event
{
3460 struct vm_area_struct
*vma
;
3462 const char *file_name
;
3466 struct perf_event_header header
;
3476 static void perf_event_mmap_output(struct perf_event
*event
,
3477 struct perf_mmap_event
*mmap_event
)
3479 struct perf_output_handle handle
;
3480 int size
= mmap_event
->event_id
.header
.size
;
3481 int ret
= perf_output_begin(&handle
, event
, size
, 0, 0);
3486 mmap_event
->event_id
.pid
= perf_event_pid(event
, current
);
3487 mmap_event
->event_id
.tid
= perf_event_tid(event
, current
);
3489 perf_output_put(&handle
, mmap_event
->event_id
);
3490 perf_output_copy(&handle
, mmap_event
->file_name
,
3491 mmap_event
->file_size
);
3492 perf_output_end(&handle
);
3495 static int perf_event_mmap_match(struct perf_event
*event
,
3496 struct perf_mmap_event
*mmap_event
)
3498 if (event
->cpu
!= -1 && event
->cpu
!= smp_processor_id())
3501 if (event
->attr
.mmap
)
3507 static void perf_event_mmap_ctx(struct perf_event_context
*ctx
,
3508 struct perf_mmap_event
*mmap_event
)
3510 struct perf_event
*event
;
3512 list_for_each_entry_rcu(event
, &ctx
->event_list
, event_entry
) {
3513 if (perf_event_mmap_match(event
, mmap_event
))
3514 perf_event_mmap_output(event
, mmap_event
);
3518 static void perf_event_mmap_event(struct perf_mmap_event
*mmap_event
)
3520 struct perf_cpu_context
*cpuctx
;
3521 struct perf_event_context
*ctx
;
3522 struct vm_area_struct
*vma
= mmap_event
->vma
;
3523 struct file
*file
= vma
->vm_file
;
3529 memset(tmp
, 0, sizeof(tmp
));
3533 * d_path works from the end of the buffer backwards, so we
3534 * need to add enough zero bytes after the string to handle
3535 * the 64bit alignment we do later.
3537 buf
= kzalloc(PATH_MAX
+ sizeof(u64
), GFP_KERNEL
);
3539 name
= strncpy(tmp
, "//enomem", sizeof(tmp
));
3542 name
= d_path(&file
->f_path
, buf
, PATH_MAX
);
3544 name
= strncpy(tmp
, "//toolong", sizeof(tmp
));
3548 if (arch_vma_name(mmap_event
->vma
)) {
3549 name
= strncpy(tmp
, arch_vma_name(mmap_event
->vma
),
3555 name
= strncpy(tmp
, "[vdso]", sizeof(tmp
));
3559 name
= strncpy(tmp
, "//anon", sizeof(tmp
));
3564 size
= ALIGN(strlen(name
)+1, sizeof(u64
));
3566 mmap_event
->file_name
= name
;
3567 mmap_event
->file_size
= size
;
3569 mmap_event
->event_id
.header
.size
= sizeof(mmap_event
->event_id
) + size
;
3572 cpuctx
= &get_cpu_var(perf_cpu_context
);
3573 perf_event_mmap_ctx(&cpuctx
->ctx
, mmap_event
);
3574 ctx
= rcu_dereference(current
->perf_event_ctxp
);
3576 perf_event_mmap_ctx(ctx
, mmap_event
);
3577 put_cpu_var(perf_cpu_context
);
3583 void __perf_event_mmap(struct vm_area_struct
*vma
)
3585 struct perf_mmap_event mmap_event
;
3587 if (!atomic_read(&nr_mmap_events
))
3590 mmap_event
= (struct perf_mmap_event
){
3596 .type
= PERF_RECORD_MMAP
,
3602 .start
= vma
->vm_start
,
3603 .len
= vma
->vm_end
- vma
->vm_start
,
3604 .pgoff
= vma
->vm_pgoff
,
3608 perf_event_mmap_event(&mmap_event
);
3612 * IRQ throttle logging
3615 static void perf_log_throttle(struct perf_event
*event
, int enable
)
3617 struct perf_output_handle handle
;
3621 struct perf_event_header header
;
3625 } throttle_event
= {
3627 .type
= PERF_RECORD_THROTTLE
,
3629 .size
= sizeof(throttle_event
),
3631 .time
= perf_clock(),
3632 .id
= primary_event_id(event
),
3633 .stream_id
= event
->id
,
3637 throttle_event
.header
.type
= PERF_RECORD_UNTHROTTLE
;
3639 ret
= perf_output_begin(&handle
, event
, sizeof(throttle_event
), 1, 0);
3643 perf_output_put(&handle
, throttle_event
);
3644 perf_output_end(&handle
);
3648 * Generic event overflow handling, sampling.
3651 static int __perf_event_overflow(struct perf_event
*event
, int nmi
,
3652 int throttle
, struct perf_sample_data
*data
,
3653 struct pt_regs
*regs
)
3655 int events
= atomic_read(&event
->event_limit
);
3656 struct hw_perf_event
*hwc
= &event
->hw
;
3659 throttle
= (throttle
&& event
->pmu
->unthrottle
!= NULL
);
3664 if (hwc
->interrupts
!= MAX_INTERRUPTS
) {
3666 if (HZ
* hwc
->interrupts
>
3667 (u64
)sysctl_perf_event_sample_rate
) {
3668 hwc
->interrupts
= MAX_INTERRUPTS
;
3669 perf_log_throttle(event
, 0);
3674 * Keep re-disabling events even though on the previous
3675 * pass we disabled it - just in case we raced with a
3676 * sched-in and the event got enabled again:
3682 if (event
->attr
.freq
) {
3683 u64 now
= perf_clock();
3684 s64 delta
= now
- hwc
->freq_stamp
;
3686 hwc
->freq_stamp
= now
;
3688 if (delta
> 0 && delta
< TICK_NSEC
)
3689 perf_adjust_period(event
, NSEC_PER_SEC
/ (int)delta
);
3693 * XXX event_limit might not quite work as expected on inherited
3697 event
->pending_kill
= POLL_IN
;
3698 if (events
&& atomic_dec_and_test(&event
->event_limit
)) {
3700 event
->pending_kill
= POLL_HUP
;
3702 event
->pending_disable
= 1;
3703 perf_pending_queue(&event
->pending
,
3704 perf_pending_event
);
3706 perf_event_disable(event
);
3709 if (event
->overflow_handler
)
3710 event
->overflow_handler(event
, nmi
, data
, regs
);
3712 perf_event_output(event
, nmi
, data
, regs
);
3717 int perf_event_overflow(struct perf_event
*event
, int nmi
,
3718 struct perf_sample_data
*data
,
3719 struct pt_regs
*regs
)
3721 return __perf_event_overflow(event
, nmi
, 1, data
, regs
);
3725 * Generic software event infrastructure
3729 * We directly increment event->count and keep a second value in
3730 * event->hw.period_left to count intervals. This period event
3731 * is kept in the range [-sample_period, 0] so that we can use the
3735 static u64
perf_swevent_set_period(struct perf_event
*event
)
3737 struct hw_perf_event
*hwc
= &event
->hw
;
3738 u64 period
= hwc
->last_period
;
3742 hwc
->last_period
= hwc
->sample_period
;
3745 old
= val
= atomic64_read(&hwc
->period_left
);
3749 nr
= div64_u64(period
+ val
, period
);
3750 offset
= nr
* period
;
3752 if (atomic64_cmpxchg(&hwc
->period_left
, old
, val
) != old
)
3758 static void perf_swevent_overflow(struct perf_event
*event
, u64 overflow
,
3759 int nmi
, struct perf_sample_data
*data
,
3760 struct pt_regs
*regs
)
3762 struct hw_perf_event
*hwc
= &event
->hw
;
3765 data
->period
= event
->hw
.last_period
;
3767 overflow
= perf_swevent_set_period(event
);
3769 if (hwc
->interrupts
== MAX_INTERRUPTS
)
3772 for (; overflow
; overflow
--) {
3773 if (__perf_event_overflow(event
, nmi
, throttle
,
3776 * We inhibit the overflow from happening when
3777 * hwc->interrupts == MAX_INTERRUPTS.
3785 static void perf_swevent_unthrottle(struct perf_event
*event
)
3788 * Nothing to do, we already reset hwc->interrupts.
3792 static void perf_swevent_add(struct perf_event
*event
, u64 nr
,
3793 int nmi
, struct perf_sample_data
*data
,
3794 struct pt_regs
*regs
)
3796 struct hw_perf_event
*hwc
= &event
->hw
;
3798 atomic64_add(nr
, &event
->count
);
3803 if (!hwc
->sample_period
)
3806 if (nr
== 1 && hwc
->sample_period
== 1 && !event
->attr
.freq
)
3807 return perf_swevent_overflow(event
, 1, nmi
, data
, regs
);
3809 if (atomic64_add_negative(nr
, &hwc
->period_left
))
3812 perf_swevent_overflow(event
, 0, nmi
, data
, regs
);
3815 static int perf_swevent_is_counting(struct perf_event
*event
)
3818 * The event is active, we're good!
3820 if (event
->state
== PERF_EVENT_STATE_ACTIVE
)
3824 * The event is off/error, not counting.
3826 if (event
->state
!= PERF_EVENT_STATE_INACTIVE
)
3830 * The event is inactive, if the context is active
3831 * we're part of a group that didn't make it on the 'pmu',
3834 if (event
->ctx
->is_active
)
3838 * We're inactive and the context is too, this means the
3839 * task is scheduled out, we're counting events that happen
3840 * to us, like migration events.
3845 static int perf_tp_event_match(struct perf_event
*event
,
3846 struct perf_sample_data
*data
);
3848 static int perf_exclude_event(struct perf_event
*event
,
3849 struct pt_regs
*regs
)
3852 if (event
->attr
.exclude_user
&& user_mode(regs
))
3855 if (event
->attr
.exclude_kernel
&& !user_mode(regs
))
3862 static int perf_swevent_match(struct perf_event
*event
,
3863 enum perf_type_id type
,
3865 struct perf_sample_data
*data
,
3866 struct pt_regs
*regs
)
3868 if (event
->cpu
!= -1 && event
->cpu
!= smp_processor_id())
3871 if (!perf_swevent_is_counting(event
))
3874 if (event
->attr
.type
!= type
)
3877 if (event
->attr
.config
!= event_id
)
3880 if (perf_exclude_event(event
, regs
))
3883 if (event
->attr
.type
== PERF_TYPE_TRACEPOINT
&&
3884 !perf_tp_event_match(event
, data
))
3890 static void perf_swevent_ctx_event(struct perf_event_context
*ctx
,
3891 enum perf_type_id type
,
3892 u32 event_id
, u64 nr
, int nmi
,
3893 struct perf_sample_data
*data
,
3894 struct pt_regs
*regs
)
3896 struct perf_event
*event
;
3898 list_for_each_entry_rcu(event
, &ctx
->event_list
, event_entry
) {
3899 if (perf_swevent_match(event
, type
, event_id
, data
, regs
))
3900 perf_swevent_add(event
, nr
, nmi
, data
, regs
);
3904 int perf_swevent_get_recursion_context(void)
3906 struct perf_cpu_context
*cpuctx
= &get_cpu_var(perf_cpu_context
);
3913 else if (in_softirq())
3918 if (cpuctx
->recursion
[rctx
]) {
3919 put_cpu_var(perf_cpu_context
);
3923 cpuctx
->recursion
[rctx
]++;
3928 EXPORT_SYMBOL_GPL(perf_swevent_get_recursion_context
);
3930 void perf_swevent_put_recursion_context(int rctx
)
3932 struct perf_cpu_context
*cpuctx
= &__get_cpu_var(perf_cpu_context
);
3934 cpuctx
->recursion
[rctx
]--;
3935 put_cpu_var(perf_cpu_context
);
3937 EXPORT_SYMBOL_GPL(perf_swevent_put_recursion_context
);
3939 static void do_perf_sw_event(enum perf_type_id type
, u32 event_id
,
3941 struct perf_sample_data
*data
,
3942 struct pt_regs
*regs
)
3944 struct perf_cpu_context
*cpuctx
;
3945 struct perf_event_context
*ctx
;
3947 cpuctx
= &__get_cpu_var(perf_cpu_context
);
3949 perf_swevent_ctx_event(&cpuctx
->ctx
, type
, event_id
,
3950 nr
, nmi
, data
, regs
);
3952 * doesn't really matter which of the child contexts the
3953 * events ends up in.
3955 ctx
= rcu_dereference(current
->perf_event_ctxp
);
3957 perf_swevent_ctx_event(ctx
, type
, event_id
, nr
, nmi
, data
, regs
);
3961 void __perf_sw_event(u32 event_id
, u64 nr
, int nmi
,
3962 struct pt_regs
*regs
, u64 addr
)
3964 struct perf_sample_data data
;
3967 rctx
= perf_swevent_get_recursion_context();
3974 do_perf_sw_event(PERF_TYPE_SOFTWARE
, event_id
, nr
, nmi
, &data
, regs
);
3976 perf_swevent_put_recursion_context(rctx
);
3979 static void perf_swevent_read(struct perf_event
*event
)
3983 static int perf_swevent_enable(struct perf_event
*event
)
3985 struct hw_perf_event
*hwc
= &event
->hw
;
3987 if (hwc
->sample_period
) {
3988 hwc
->last_period
= hwc
->sample_period
;
3989 perf_swevent_set_period(event
);
3994 static void perf_swevent_disable(struct perf_event
*event
)
3998 static const struct pmu perf_ops_generic
= {
3999 .enable
= perf_swevent_enable
,
4000 .disable
= perf_swevent_disable
,
4001 .read
= perf_swevent_read
,
4002 .unthrottle
= perf_swevent_unthrottle
,
4006 * hrtimer based swevent callback
4009 static enum hrtimer_restart
perf_swevent_hrtimer(struct hrtimer
*hrtimer
)
4011 enum hrtimer_restart ret
= HRTIMER_RESTART
;
4012 struct perf_sample_data data
;
4013 struct pt_regs
*regs
;
4014 struct perf_event
*event
;
4017 event
= container_of(hrtimer
, struct perf_event
, hw
.hrtimer
);
4018 event
->pmu
->read(event
);
4022 data
.period
= event
->hw
.last_period
;
4023 regs
= get_irq_regs();
4025 * In case we exclude kernel IPs or are somehow not in interrupt
4026 * context, provide the next best thing, the user IP.
4028 if ((event
->attr
.exclude_kernel
|| !regs
) &&
4029 !event
->attr
.exclude_user
)
4030 regs
= task_pt_regs(current
);
4033 if (!(event
->attr
.exclude_idle
&& current
->pid
== 0))
4034 if (perf_event_overflow(event
, 0, &data
, regs
))
4035 ret
= HRTIMER_NORESTART
;
4038 period
= max_t(u64
, 10000, event
->hw
.sample_period
);
4039 hrtimer_forward_now(hrtimer
, ns_to_ktime(period
));
4044 static void perf_swevent_start_hrtimer(struct perf_event
*event
)
4046 struct hw_perf_event
*hwc
= &event
->hw
;
4048 hrtimer_init(&hwc
->hrtimer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
4049 hwc
->hrtimer
.function
= perf_swevent_hrtimer
;
4050 if (hwc
->sample_period
) {
4053 if (hwc
->remaining
) {
4054 if (hwc
->remaining
< 0)
4057 period
= hwc
->remaining
;
4060 period
= max_t(u64
, 10000, hwc
->sample_period
);
4062 __hrtimer_start_range_ns(&hwc
->hrtimer
,
4063 ns_to_ktime(period
), 0,
4064 HRTIMER_MODE_REL
, 0);
4068 static void perf_swevent_cancel_hrtimer(struct perf_event
*event
)
4070 struct hw_perf_event
*hwc
= &event
->hw
;
4072 if (hwc
->sample_period
) {
4073 ktime_t remaining
= hrtimer_get_remaining(&hwc
->hrtimer
);
4074 hwc
->remaining
= ktime_to_ns(remaining
);
4076 hrtimer_cancel(&hwc
->hrtimer
);
4081 * Software event: cpu wall time clock
4084 static void cpu_clock_perf_event_update(struct perf_event
*event
)
4086 int cpu
= raw_smp_processor_id();
4090 now
= cpu_clock(cpu
);
4091 prev
= atomic64_xchg(&event
->hw
.prev_count
, now
);
4092 atomic64_add(now
- prev
, &event
->count
);
4095 static int cpu_clock_perf_event_enable(struct perf_event
*event
)
4097 struct hw_perf_event
*hwc
= &event
->hw
;
4098 int cpu
= raw_smp_processor_id();
4100 atomic64_set(&hwc
->prev_count
, cpu_clock(cpu
));
4101 perf_swevent_start_hrtimer(event
);
4106 static void cpu_clock_perf_event_disable(struct perf_event
*event
)
4108 perf_swevent_cancel_hrtimer(event
);
4109 cpu_clock_perf_event_update(event
);
4112 static void cpu_clock_perf_event_read(struct perf_event
*event
)
4114 cpu_clock_perf_event_update(event
);
4117 static const struct pmu perf_ops_cpu_clock
= {
4118 .enable
= cpu_clock_perf_event_enable
,
4119 .disable
= cpu_clock_perf_event_disable
,
4120 .read
= cpu_clock_perf_event_read
,
4124 * Software event: task time clock
4127 static void task_clock_perf_event_update(struct perf_event
*event
, u64 now
)
4132 prev
= atomic64_xchg(&event
->hw
.prev_count
, now
);
4134 atomic64_add(delta
, &event
->count
);
4137 static int task_clock_perf_event_enable(struct perf_event
*event
)
4139 struct hw_perf_event
*hwc
= &event
->hw
;
4142 now
= event
->ctx
->time
;
4144 atomic64_set(&hwc
->prev_count
, now
);
4146 perf_swevent_start_hrtimer(event
);
4151 static void task_clock_perf_event_disable(struct perf_event
*event
)
4153 perf_swevent_cancel_hrtimer(event
);
4154 task_clock_perf_event_update(event
, event
->ctx
->time
);
4158 static void task_clock_perf_event_read(struct perf_event
*event
)
4163 update_context_time(event
->ctx
);
4164 time
= event
->ctx
->time
;
4166 u64 now
= perf_clock();
4167 u64 delta
= now
- event
->ctx
->timestamp
;
4168 time
= event
->ctx
->time
+ delta
;
4171 task_clock_perf_event_update(event
, time
);
4174 static const struct pmu perf_ops_task_clock
= {
4175 .enable
= task_clock_perf_event_enable
,
4176 .disable
= task_clock_perf_event_disable
,
4177 .read
= task_clock_perf_event_read
,
4180 #ifdef CONFIG_EVENT_TRACING
4182 void perf_tp_event(int event_id
, u64 addr
, u64 count
, void *record
,
4185 struct perf_raw_record raw
= {
4190 struct perf_sample_data data
= {
4195 struct pt_regs
*regs
= get_irq_regs();
4198 regs
= task_pt_regs(current
);
4200 /* Trace events already protected against recursion */
4201 do_perf_sw_event(PERF_TYPE_TRACEPOINT
, event_id
, count
, 1,
4204 EXPORT_SYMBOL_GPL(perf_tp_event
);
4206 static int perf_tp_event_match(struct perf_event
*event
,
4207 struct perf_sample_data
*data
)
4209 void *record
= data
->raw
->data
;
4211 if (likely(!event
->filter
) || filter_match_preds(event
->filter
, record
))
4216 static void tp_perf_event_destroy(struct perf_event
*event
)
4218 ftrace_profile_disable(event
->attr
.config
);
4221 static const struct pmu
*tp_perf_event_init(struct perf_event
*event
)
4224 * Raw tracepoint data is a severe data leak, only allow root to
4227 if ((event
->attr
.sample_type
& PERF_SAMPLE_RAW
) &&
4228 perf_paranoid_tracepoint_raw() &&
4229 !capable(CAP_SYS_ADMIN
))
4230 return ERR_PTR(-EPERM
);
4232 if (ftrace_profile_enable(event
->attr
.config
))
4235 event
->destroy
= tp_perf_event_destroy
;
4237 return &perf_ops_generic
;
4240 static int perf_event_set_filter(struct perf_event
*event
, void __user
*arg
)
4245 if (event
->attr
.type
!= PERF_TYPE_TRACEPOINT
)
4248 filter_str
= strndup_user(arg
, PAGE_SIZE
);
4249 if (IS_ERR(filter_str
))
4250 return PTR_ERR(filter_str
);
4252 ret
= ftrace_profile_set_filter(event
, event
->attr
.config
, filter_str
);
4258 static void perf_event_free_filter(struct perf_event
*event
)
4260 ftrace_profile_free_filter(event
);
4265 static int perf_tp_event_match(struct perf_event
*event
,
4266 struct perf_sample_data
*data
)
4271 static const struct pmu
*tp_perf_event_init(struct perf_event
*event
)
4276 static int perf_event_set_filter(struct perf_event
*event
, void __user
*arg
)
4281 static void perf_event_free_filter(struct perf_event
*event
)
4285 #endif /* CONFIG_EVENT_TRACING */
4287 #ifdef CONFIG_HAVE_HW_BREAKPOINT
4288 static void bp_perf_event_destroy(struct perf_event
*event
)
4290 release_bp_slot(event
);
4293 static const struct pmu
*bp_perf_event_init(struct perf_event
*bp
)
4297 err
= register_perf_hw_breakpoint(bp
);
4299 return ERR_PTR(err
);
4301 bp
->destroy
= bp_perf_event_destroy
;
4303 return &perf_ops_bp
;
4306 void perf_bp_event(struct perf_event
*bp
, void *data
)
4308 struct perf_sample_data sample
;
4309 struct pt_regs
*regs
= data
;
4312 sample
.addr
= bp
->attr
.bp_addr
;
4314 if (!perf_exclude_event(bp
, regs
))
4315 perf_swevent_add(bp
, 1, 1, &sample
, regs
);
4318 static const struct pmu
*bp_perf_event_init(struct perf_event
*bp
)
4323 void perf_bp_event(struct perf_event
*bp
, void *regs
)
4328 atomic_t perf_swevent_enabled
[PERF_COUNT_SW_MAX
];
4330 static void sw_perf_event_destroy(struct perf_event
*event
)
4332 u64 event_id
= event
->attr
.config
;
4334 WARN_ON(event
->parent
);
4336 atomic_dec(&perf_swevent_enabled
[event_id
]);
4339 static const struct pmu
*sw_perf_event_init(struct perf_event
*event
)
4341 const struct pmu
*pmu
= NULL
;
4342 u64 event_id
= event
->attr
.config
;
4345 * Software events (currently) can't in general distinguish
4346 * between user, kernel and hypervisor events.
4347 * However, context switches and cpu migrations are considered
4348 * to be kernel events, and page faults are never hypervisor
4352 case PERF_COUNT_SW_CPU_CLOCK
:
4353 pmu
= &perf_ops_cpu_clock
;
4356 case PERF_COUNT_SW_TASK_CLOCK
:
4358 * If the user instantiates this as a per-cpu event,
4359 * use the cpu_clock event instead.
4361 if (event
->ctx
->task
)
4362 pmu
= &perf_ops_task_clock
;
4364 pmu
= &perf_ops_cpu_clock
;
4367 case PERF_COUNT_SW_PAGE_FAULTS
:
4368 case PERF_COUNT_SW_PAGE_FAULTS_MIN
:
4369 case PERF_COUNT_SW_PAGE_FAULTS_MAJ
:
4370 case PERF_COUNT_SW_CONTEXT_SWITCHES
:
4371 case PERF_COUNT_SW_CPU_MIGRATIONS
:
4372 case PERF_COUNT_SW_ALIGNMENT_FAULTS
:
4373 case PERF_COUNT_SW_EMULATION_FAULTS
:
4374 if (!event
->parent
) {
4375 atomic_inc(&perf_swevent_enabled
[event_id
]);
4376 event
->destroy
= sw_perf_event_destroy
;
4378 pmu
= &perf_ops_generic
;
4386 * Allocate and initialize a event structure
4388 static struct perf_event
*
4389 perf_event_alloc(struct perf_event_attr
*attr
,
4391 struct perf_event_context
*ctx
,
4392 struct perf_event
*group_leader
,
4393 struct perf_event
*parent_event
,
4394 perf_overflow_handler_t overflow_handler
,
4397 const struct pmu
*pmu
;
4398 struct perf_event
*event
;
4399 struct hw_perf_event
*hwc
;
4402 event
= kzalloc(sizeof(*event
), gfpflags
);
4404 return ERR_PTR(-ENOMEM
);
4407 * Single events are their own group leaders, with an
4408 * empty sibling list:
4411 group_leader
= event
;
4413 mutex_init(&event
->child_mutex
);
4414 INIT_LIST_HEAD(&event
->child_list
);
4416 INIT_LIST_HEAD(&event
->group_entry
);
4417 INIT_LIST_HEAD(&event
->event_entry
);
4418 INIT_LIST_HEAD(&event
->sibling_list
);
4419 init_waitqueue_head(&event
->waitq
);
4421 mutex_init(&event
->mmap_mutex
);
4424 event
->attr
= *attr
;
4425 event
->group_leader
= group_leader
;
4430 event
->parent
= parent_event
;
4432 event
->ns
= get_pid_ns(current
->nsproxy
->pid_ns
);
4433 event
->id
= atomic64_inc_return(&perf_event_id
);
4435 event
->state
= PERF_EVENT_STATE_INACTIVE
;
4437 if (!overflow_handler
&& parent_event
)
4438 overflow_handler
= parent_event
->overflow_handler
;
4440 event
->overflow_handler
= overflow_handler
;
4443 event
->state
= PERF_EVENT_STATE_OFF
;
4448 hwc
->sample_period
= attr
->sample_period
;
4449 if (attr
->freq
&& attr
->sample_freq
)
4450 hwc
->sample_period
= 1;
4451 hwc
->last_period
= hwc
->sample_period
;
4453 atomic64_set(&hwc
->period_left
, hwc
->sample_period
);
4456 * we currently do not support PERF_FORMAT_GROUP on inherited events
4458 if (attr
->inherit
&& (attr
->read_format
& PERF_FORMAT_GROUP
))
4461 switch (attr
->type
) {
4463 case PERF_TYPE_HARDWARE
:
4464 case PERF_TYPE_HW_CACHE
:
4465 pmu
= hw_perf_event_init(event
);
4468 case PERF_TYPE_SOFTWARE
:
4469 pmu
= sw_perf_event_init(event
);
4472 case PERF_TYPE_TRACEPOINT
:
4473 pmu
= tp_perf_event_init(event
);
4476 case PERF_TYPE_BREAKPOINT
:
4477 pmu
= bp_perf_event_init(event
);
4488 else if (IS_ERR(pmu
))
4493 put_pid_ns(event
->ns
);
4495 return ERR_PTR(err
);
4500 if (!event
->parent
) {
4501 atomic_inc(&nr_events
);
4502 if (event
->attr
.mmap
)
4503 atomic_inc(&nr_mmap_events
);
4504 if (event
->attr
.comm
)
4505 atomic_inc(&nr_comm_events
);
4506 if (event
->attr
.task
)
4507 atomic_inc(&nr_task_events
);
4513 static int perf_copy_attr(struct perf_event_attr __user
*uattr
,
4514 struct perf_event_attr
*attr
)
4519 if (!access_ok(VERIFY_WRITE
, uattr
, PERF_ATTR_SIZE_VER0
))
4523 * zero the full structure, so that a short copy will be nice.
4525 memset(attr
, 0, sizeof(*attr
));
4527 ret
= get_user(size
, &uattr
->size
);
4531 if (size
> PAGE_SIZE
) /* silly large */
4534 if (!size
) /* abi compat */
4535 size
= PERF_ATTR_SIZE_VER0
;
4537 if (size
< PERF_ATTR_SIZE_VER0
)
4541 * If we're handed a bigger struct than we know of,
4542 * ensure all the unknown bits are 0 - i.e. new
4543 * user-space does not rely on any kernel feature
4544 * extensions we dont know about yet.
4546 if (size
> sizeof(*attr
)) {
4547 unsigned char __user
*addr
;
4548 unsigned char __user
*end
;
4551 addr
= (void __user
*)uattr
+ sizeof(*attr
);
4552 end
= (void __user
*)uattr
+ size
;
4554 for (; addr
< end
; addr
++) {
4555 ret
= get_user(val
, addr
);
4561 size
= sizeof(*attr
);
4564 ret
= copy_from_user(attr
, uattr
, size
);
4569 * If the type exists, the corresponding creation will verify
4572 if (attr
->type
>= PERF_TYPE_MAX
)
4575 if (attr
->__reserved_1
|| attr
->__reserved_2
)
4578 if (attr
->sample_type
& ~(PERF_SAMPLE_MAX
-1))
4581 if (attr
->read_format
& ~(PERF_FORMAT_MAX
-1))
4588 put_user(sizeof(*attr
), &uattr
->size
);
4593 static int perf_event_set_output(struct perf_event
*event
, int output_fd
)
4595 struct perf_event
*output_event
= NULL
;
4596 struct file
*output_file
= NULL
;
4597 struct perf_event
*old_output
;
4598 int fput_needed
= 0;
4604 output_file
= fget_light(output_fd
, &fput_needed
);
4608 if (output_file
->f_op
!= &perf_fops
)
4611 output_event
= output_file
->private_data
;
4613 /* Don't chain output fds */
4614 if (output_event
->output
)
4617 /* Don't set an output fd when we already have an output channel */
4621 atomic_long_inc(&output_file
->f_count
);
4624 mutex_lock(&event
->mmap_mutex
);
4625 old_output
= event
->output
;
4626 rcu_assign_pointer(event
->output
, output_event
);
4627 mutex_unlock(&event
->mmap_mutex
);
4631 * we need to make sure no existing perf_output_*()
4632 * is still referencing this event.
4635 fput(old_output
->filp
);
4640 fput_light(output_file
, fput_needed
);
4645 * sys_perf_event_open - open a performance event, associate it to a task/cpu
4647 * @attr_uptr: event_id type attributes for monitoring/sampling
4650 * @group_fd: group leader event fd
4652 SYSCALL_DEFINE5(perf_event_open
,
4653 struct perf_event_attr __user
*, attr_uptr
,
4654 pid_t
, pid
, int, cpu
, int, group_fd
, unsigned long, flags
)
4656 struct perf_event
*event
, *group_leader
;
4657 struct perf_event_attr attr
;
4658 struct perf_event_context
*ctx
;
4659 struct file
*event_file
= NULL
;
4660 struct file
*group_file
= NULL
;
4661 int fput_needed
= 0;
4662 int fput_needed2
= 0;
4665 /* for future expandability... */
4666 if (flags
& ~(PERF_FLAG_FD_NO_GROUP
| PERF_FLAG_FD_OUTPUT
))
4669 err
= perf_copy_attr(attr_uptr
, &attr
);
4673 if (!attr
.exclude_kernel
) {
4674 if (perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN
))
4679 if (attr
.sample_freq
> sysctl_perf_event_sample_rate
)
4684 * Get the target context (task or percpu):
4686 ctx
= find_get_context(pid
, cpu
);
4688 return PTR_ERR(ctx
);
4691 * Look up the group leader (we will attach this event to it):
4693 group_leader
= NULL
;
4694 if (group_fd
!= -1 && !(flags
& PERF_FLAG_FD_NO_GROUP
)) {
4696 group_file
= fget_light(group_fd
, &fput_needed
);
4698 goto err_put_context
;
4699 if (group_file
->f_op
!= &perf_fops
)
4700 goto err_put_context
;
4702 group_leader
= group_file
->private_data
;
4704 * Do not allow a recursive hierarchy (this new sibling
4705 * becoming part of another group-sibling):
4707 if (group_leader
->group_leader
!= group_leader
)
4708 goto err_put_context
;
4710 * Do not allow to attach to a group in a different
4711 * task or CPU context:
4713 if (group_leader
->ctx
!= ctx
)
4714 goto err_put_context
;
4716 * Only a group leader can be exclusive or pinned
4718 if (attr
.exclusive
|| attr
.pinned
)
4719 goto err_put_context
;
4722 event
= perf_event_alloc(&attr
, cpu
, ctx
, group_leader
,
4723 NULL
, NULL
, GFP_KERNEL
);
4724 err
= PTR_ERR(event
);
4726 goto err_put_context
;
4728 err
= anon_inode_getfd("[perf_event]", &perf_fops
, event
, O_RDWR
);
4730 goto err_free_put_context
;
4732 event_file
= fget_light(err
, &fput_needed2
);
4734 goto err_free_put_context
;
4736 if (flags
& PERF_FLAG_FD_OUTPUT
) {
4737 err
= perf_event_set_output(event
, group_fd
);
4739 goto err_fput_free_put_context
;
4742 event
->filp
= event_file
;
4743 WARN_ON_ONCE(ctx
->parent_ctx
);
4744 mutex_lock(&ctx
->mutex
);
4745 perf_install_in_context(ctx
, event
, cpu
);
4747 mutex_unlock(&ctx
->mutex
);
4749 event
->owner
= current
;
4750 get_task_struct(current
);
4751 mutex_lock(¤t
->perf_event_mutex
);
4752 list_add_tail(&event
->owner_entry
, ¤t
->perf_event_list
);
4753 mutex_unlock(¤t
->perf_event_mutex
);
4755 err_fput_free_put_context
:
4756 fput_light(event_file
, fput_needed2
);
4758 err_free_put_context
:
4766 fput_light(group_file
, fput_needed
);
4772 * perf_event_create_kernel_counter
4774 * @attr: attributes of the counter to create
4775 * @cpu: cpu in which the counter is bound
4776 * @pid: task to profile
4779 perf_event_create_kernel_counter(struct perf_event_attr
*attr
, int cpu
,
4781 perf_overflow_handler_t overflow_handler
)
4783 struct perf_event
*event
;
4784 struct perf_event_context
*ctx
;
4788 * Get the target context (task or percpu):
4791 ctx
= find_get_context(pid
, cpu
);
4797 event
= perf_event_alloc(attr
, cpu
, ctx
, NULL
,
4798 NULL
, overflow_handler
, GFP_KERNEL
);
4799 if (IS_ERR(event
)) {
4800 err
= PTR_ERR(event
);
4801 goto err_put_context
;
4805 WARN_ON_ONCE(ctx
->parent_ctx
);
4806 mutex_lock(&ctx
->mutex
);
4807 perf_install_in_context(ctx
, event
, cpu
);
4809 mutex_unlock(&ctx
->mutex
);
4811 event
->owner
= current
;
4812 get_task_struct(current
);
4813 mutex_lock(¤t
->perf_event_mutex
);
4814 list_add_tail(&event
->owner_entry
, ¤t
->perf_event_list
);
4815 mutex_unlock(¤t
->perf_event_mutex
);
4822 return ERR_PTR(err
);
4824 EXPORT_SYMBOL_GPL(perf_event_create_kernel_counter
);
4827 * inherit a event from parent task to child task:
4829 static struct perf_event
*
4830 inherit_event(struct perf_event
*parent_event
,
4831 struct task_struct
*parent
,
4832 struct perf_event_context
*parent_ctx
,
4833 struct task_struct
*child
,
4834 struct perf_event
*group_leader
,
4835 struct perf_event_context
*child_ctx
)
4837 struct perf_event
*child_event
;
4840 * Instead of creating recursive hierarchies of events,
4841 * we link inherited events back to the original parent,
4842 * which has a filp for sure, which we use as the reference
4845 if (parent_event
->parent
)
4846 parent_event
= parent_event
->parent
;
4848 child_event
= perf_event_alloc(&parent_event
->attr
,
4849 parent_event
->cpu
, child_ctx
,
4850 group_leader
, parent_event
,
4852 if (IS_ERR(child_event
))
4857 * Make the child state follow the state of the parent event,
4858 * not its attr.disabled bit. We hold the parent's mutex,
4859 * so we won't race with perf_event_{en, dis}able_family.
4861 if (parent_event
->state
>= PERF_EVENT_STATE_INACTIVE
)
4862 child_event
->state
= PERF_EVENT_STATE_INACTIVE
;
4864 child_event
->state
= PERF_EVENT_STATE_OFF
;
4866 if (parent_event
->attr
.freq
)
4867 child_event
->hw
.sample_period
= parent_event
->hw
.sample_period
;
4869 child_event
->overflow_handler
= parent_event
->overflow_handler
;
4872 * Link it up in the child's context:
4874 add_event_to_ctx(child_event
, child_ctx
);
4877 * Get a reference to the parent filp - we will fput it
4878 * when the child event exits. This is safe to do because
4879 * we are in the parent and we know that the filp still
4880 * exists and has a nonzero count:
4882 atomic_long_inc(&parent_event
->filp
->f_count
);
4885 * Link this into the parent event's child list
4887 WARN_ON_ONCE(parent_event
->ctx
->parent_ctx
);
4888 mutex_lock(&parent_event
->child_mutex
);
4889 list_add_tail(&child_event
->child_list
, &parent_event
->child_list
);
4890 mutex_unlock(&parent_event
->child_mutex
);
4895 static int inherit_group(struct perf_event
*parent_event
,
4896 struct task_struct
*parent
,
4897 struct perf_event_context
*parent_ctx
,
4898 struct task_struct
*child
,
4899 struct perf_event_context
*child_ctx
)
4901 struct perf_event
*leader
;
4902 struct perf_event
*sub
;
4903 struct perf_event
*child_ctr
;
4905 leader
= inherit_event(parent_event
, parent
, parent_ctx
,
4906 child
, NULL
, child_ctx
);
4908 return PTR_ERR(leader
);
4909 list_for_each_entry(sub
, &parent_event
->sibling_list
, group_entry
) {
4910 child_ctr
= inherit_event(sub
, parent
, parent_ctx
,
4911 child
, leader
, child_ctx
);
4912 if (IS_ERR(child_ctr
))
4913 return PTR_ERR(child_ctr
);
4918 static void sync_child_event(struct perf_event
*child_event
,
4919 struct task_struct
*child
)
4921 struct perf_event
*parent_event
= child_event
->parent
;
4924 if (child_event
->attr
.inherit_stat
)
4925 perf_event_read_event(child_event
, child
);
4927 child_val
= atomic64_read(&child_event
->count
);
4930 * Add back the child's count to the parent's count:
4932 atomic64_add(child_val
, &parent_event
->count
);
4933 atomic64_add(child_event
->total_time_enabled
,
4934 &parent_event
->child_total_time_enabled
);
4935 atomic64_add(child_event
->total_time_running
,
4936 &parent_event
->child_total_time_running
);
4939 * Remove this event from the parent's list
4941 WARN_ON_ONCE(parent_event
->ctx
->parent_ctx
);
4942 mutex_lock(&parent_event
->child_mutex
);
4943 list_del_init(&child_event
->child_list
);
4944 mutex_unlock(&parent_event
->child_mutex
);
4947 * Release the parent event, if this was the last
4950 fput(parent_event
->filp
);
4954 __perf_event_exit_task(struct perf_event
*child_event
,
4955 struct perf_event_context
*child_ctx
,
4956 struct task_struct
*child
)
4958 struct perf_event
*parent_event
;
4960 perf_event_remove_from_context(child_event
);
4962 parent_event
= child_event
->parent
;
4964 * It can happen that parent exits first, and has events
4965 * that are still around due to the child reference. These
4966 * events need to be zapped - but otherwise linger.
4969 sync_child_event(child_event
, child
);
4970 free_event(child_event
);
4975 * When a child task exits, feed back event values to parent events.
4977 void perf_event_exit_task(struct task_struct
*child
)
4979 struct perf_event
*child_event
, *tmp
;
4980 struct perf_event_context
*child_ctx
;
4981 unsigned long flags
;
4983 if (likely(!child
->perf_event_ctxp
)) {
4984 perf_event_task(child
, NULL
, 0);
4988 local_irq_save(flags
);
4990 * We can't reschedule here because interrupts are disabled,
4991 * and either child is current or it is a task that can't be
4992 * scheduled, so we are now safe from rescheduling changing
4995 child_ctx
= child
->perf_event_ctxp
;
4996 __perf_event_task_sched_out(child_ctx
);
4999 * Take the context lock here so that if find_get_context is
5000 * reading child->perf_event_ctxp, we wait until it has
5001 * incremented the context's refcount before we do put_ctx below.
5003 raw_spin_lock(&child_ctx
->lock
);
5004 child
->perf_event_ctxp
= NULL
;
5006 * If this context is a clone; unclone it so it can't get
5007 * swapped to another process while we're removing all
5008 * the events from it.
5010 unclone_ctx(child_ctx
);
5011 update_context_time(child_ctx
);
5012 raw_spin_unlock_irqrestore(&child_ctx
->lock
, flags
);
5015 * Report the task dead after unscheduling the events so that we
5016 * won't get any samples after PERF_RECORD_EXIT. We can however still
5017 * get a few PERF_RECORD_READ events.
5019 perf_event_task(child
, child_ctx
, 0);
5022 * We can recurse on the same lock type through:
5024 * __perf_event_exit_task()
5025 * sync_child_event()
5026 * fput(parent_event->filp)
5028 * mutex_lock(&ctx->mutex)
5030 * But since its the parent context it won't be the same instance.
5032 mutex_lock_nested(&child_ctx
->mutex
, SINGLE_DEPTH_NESTING
);
5035 list_for_each_entry_safe(child_event
, tmp
, &child_ctx
->group_list
,
5037 __perf_event_exit_task(child_event
, child_ctx
, child
);
5040 * If the last event was a group event, it will have appended all
5041 * its siblings to the list, but we obtained 'tmp' before that which
5042 * will still point to the list head terminating the iteration.
5044 if (!list_empty(&child_ctx
->group_list
))
5047 mutex_unlock(&child_ctx
->mutex
);
5053 * free an unexposed, unused context as created by inheritance by
5054 * init_task below, used by fork() in case of fail.
5056 void perf_event_free_task(struct task_struct
*task
)
5058 struct perf_event_context
*ctx
= task
->perf_event_ctxp
;
5059 struct perf_event
*event
, *tmp
;
5064 mutex_lock(&ctx
->mutex
);
5066 list_for_each_entry_safe(event
, tmp
, &ctx
->group_list
, group_entry
) {
5067 struct perf_event
*parent
= event
->parent
;
5069 if (WARN_ON_ONCE(!parent
))
5072 mutex_lock(&parent
->child_mutex
);
5073 list_del_init(&event
->child_list
);
5074 mutex_unlock(&parent
->child_mutex
);
5078 list_del_event(event
, ctx
);
5082 if (!list_empty(&ctx
->group_list
))
5085 mutex_unlock(&ctx
->mutex
);
5091 * Initialize the perf_event context in task_struct
5093 int perf_event_init_task(struct task_struct
*child
)
5095 struct perf_event_context
*child_ctx
= NULL
, *parent_ctx
;
5096 struct perf_event_context
*cloned_ctx
;
5097 struct perf_event
*event
;
5098 struct task_struct
*parent
= current
;
5099 int inherited_all
= 1;
5102 child
->perf_event_ctxp
= NULL
;
5104 mutex_init(&child
->perf_event_mutex
);
5105 INIT_LIST_HEAD(&child
->perf_event_list
);
5107 if (likely(!parent
->perf_event_ctxp
))
5111 * If the parent's context is a clone, pin it so it won't get
5114 parent_ctx
= perf_pin_task_context(parent
);
5117 * No need to check if parent_ctx != NULL here; since we saw
5118 * it non-NULL earlier, the only reason for it to become NULL
5119 * is if we exit, and since we're currently in the middle of
5120 * a fork we can't be exiting at the same time.
5124 * Lock the parent list. No need to lock the child - not PID
5125 * hashed yet and not running, so nobody can access it.
5127 mutex_lock(&parent_ctx
->mutex
);
5130 * We dont have to disable NMIs - we are only looking at
5131 * the list, not manipulating it:
5133 list_for_each_entry(event
, &parent_ctx
->group_list
, group_entry
) {
5135 if (!event
->attr
.inherit
) {
5140 if (!child
->perf_event_ctxp
) {
5142 * This is executed from the parent task context, so
5143 * inherit events that have been marked for cloning.
5144 * First allocate and initialize a context for the
5148 child_ctx
= kzalloc(sizeof(struct perf_event_context
),
5155 __perf_event_init_context(child_ctx
, child
);
5156 child
->perf_event_ctxp
= child_ctx
;
5157 get_task_struct(child
);
5160 ret
= inherit_group(event
, parent
, parent_ctx
,
5168 if (child_ctx
&& inherited_all
) {
5170 * Mark the child context as a clone of the parent
5171 * context, or of whatever the parent is a clone of.
5172 * Note that if the parent is a clone, it could get
5173 * uncloned at any point, but that doesn't matter
5174 * because the list of events and the generation
5175 * count can't have changed since we took the mutex.
5177 cloned_ctx
= rcu_dereference(parent_ctx
->parent_ctx
);
5179 child_ctx
->parent_ctx
= cloned_ctx
;
5180 child_ctx
->parent_gen
= parent_ctx
->parent_gen
;
5182 child_ctx
->parent_ctx
= parent_ctx
;
5183 child_ctx
->parent_gen
= parent_ctx
->generation
;
5185 get_ctx(child_ctx
->parent_ctx
);
5188 mutex_unlock(&parent_ctx
->mutex
);
5190 perf_unpin_context(parent_ctx
);
5195 static void __cpuinit
perf_event_init_cpu(int cpu
)
5197 struct perf_cpu_context
*cpuctx
;
5199 cpuctx
= &per_cpu(perf_cpu_context
, cpu
);
5200 __perf_event_init_context(&cpuctx
->ctx
, NULL
);
5202 spin_lock(&perf_resource_lock
);
5203 cpuctx
->max_pertask
= perf_max_events
- perf_reserved_percpu
;
5204 spin_unlock(&perf_resource_lock
);
5206 hw_perf_event_setup(cpu
);
5209 #ifdef CONFIG_HOTPLUG_CPU
5210 static void __perf_event_exit_cpu(void *info
)
5212 struct perf_cpu_context
*cpuctx
= &__get_cpu_var(perf_cpu_context
);
5213 struct perf_event_context
*ctx
= &cpuctx
->ctx
;
5214 struct perf_event
*event
, *tmp
;
5216 list_for_each_entry_safe(event
, tmp
, &ctx
->group_list
, group_entry
)
5217 __perf_event_remove_from_context(event
);
5219 static void perf_event_exit_cpu(int cpu
)
5221 struct perf_cpu_context
*cpuctx
= &per_cpu(perf_cpu_context
, cpu
);
5222 struct perf_event_context
*ctx
= &cpuctx
->ctx
;
5224 mutex_lock(&ctx
->mutex
);
5225 smp_call_function_single(cpu
, __perf_event_exit_cpu
, NULL
, 1);
5226 mutex_unlock(&ctx
->mutex
);
5229 static inline void perf_event_exit_cpu(int cpu
) { }
5232 static int __cpuinit
5233 perf_cpu_notify(struct notifier_block
*self
, unsigned long action
, void *hcpu
)
5235 unsigned int cpu
= (long)hcpu
;
5239 case CPU_UP_PREPARE
:
5240 case CPU_UP_PREPARE_FROZEN
:
5241 perf_event_init_cpu(cpu
);
5245 case CPU_ONLINE_FROZEN
:
5246 hw_perf_event_setup_online(cpu
);
5249 case CPU_DOWN_PREPARE
:
5250 case CPU_DOWN_PREPARE_FROZEN
:
5251 perf_event_exit_cpu(cpu
);
5262 * This has to have a higher priority than migration_notifier in sched.c.
5264 static struct notifier_block __cpuinitdata perf_cpu_nb
= {
5265 .notifier_call
= perf_cpu_notify
,
5269 void __init
perf_event_init(void)
5271 perf_cpu_notify(&perf_cpu_nb
, (unsigned long)CPU_UP_PREPARE
,
5272 (void *)(long)smp_processor_id());
5273 perf_cpu_notify(&perf_cpu_nb
, (unsigned long)CPU_ONLINE
,
5274 (void *)(long)smp_processor_id());
5275 register_cpu_notifier(&perf_cpu_nb
);
5278 static ssize_t
perf_show_reserve_percpu(struct sysdev_class
*class, char *buf
)
5280 return sprintf(buf
, "%d\n", perf_reserved_percpu
);
5284 perf_set_reserve_percpu(struct sysdev_class
*class,
5288 struct perf_cpu_context
*cpuctx
;
5292 err
= strict_strtoul(buf
, 10, &val
);
5295 if (val
> perf_max_events
)
5298 spin_lock(&perf_resource_lock
);
5299 perf_reserved_percpu
= val
;
5300 for_each_online_cpu(cpu
) {
5301 cpuctx
= &per_cpu(perf_cpu_context
, cpu
);
5302 raw_spin_lock_irq(&cpuctx
->ctx
.lock
);
5303 mpt
= min(perf_max_events
- cpuctx
->ctx
.nr_events
,
5304 perf_max_events
- perf_reserved_percpu
);
5305 cpuctx
->max_pertask
= mpt
;
5306 raw_spin_unlock_irq(&cpuctx
->ctx
.lock
);
5308 spin_unlock(&perf_resource_lock
);
5313 static ssize_t
perf_show_overcommit(struct sysdev_class
*class, char *buf
)
5315 return sprintf(buf
, "%d\n", perf_overcommit
);
5319 perf_set_overcommit(struct sysdev_class
*class, const char *buf
, size_t count
)
5324 err
= strict_strtoul(buf
, 10, &val
);
5330 spin_lock(&perf_resource_lock
);
5331 perf_overcommit
= val
;
5332 spin_unlock(&perf_resource_lock
);
5337 static SYSDEV_CLASS_ATTR(
5340 perf_show_reserve_percpu
,
5341 perf_set_reserve_percpu
5344 static SYSDEV_CLASS_ATTR(
5347 perf_show_overcommit
,
5351 static struct attribute
*perfclass_attrs
[] = {
5352 &attr_reserve_percpu
.attr
,
5353 &attr_overcommit
.attr
,
5357 static struct attribute_group perfclass_attr_group
= {
5358 .attrs
= perfclass_attrs
,
5359 .name
= "perf_events",
5362 static int __init
perf_event_sysfs_init(void)
5364 return sysfs_create_group(&cpu_sysdev_class
.kset
.kobj
,
5365 &perfclass_attr_group
);
5367 device_initcall(perf_event_sysfs_init
);