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/slab.h>
19 #include <linux/hash.h>
20 #include <linux/sysfs.h>
21 #include <linux/dcache.h>
22 #include <linux/percpu.h>
23 #include <linux/ptrace.h>
24 #include <linux/vmstat.h>
25 #include <linux/vmalloc.h>
26 #include <linux/hardirq.h>
27 #include <linux/rculist.h>
28 #include <linux/uaccess.h>
29 #include <linux/syscalls.h>
30 #include <linux/anon_inodes.h>
31 #include <linux/kernel_stat.h>
32 #include <linux/perf_event.h>
33 #include <linux/ftrace_event.h>
35 #include <asm/irq_regs.h>
37 static atomic_t nr_events __read_mostly
;
38 static atomic_t nr_mmap_events __read_mostly
;
39 static atomic_t nr_comm_events __read_mostly
;
40 static atomic_t nr_task_events __read_mostly
;
42 static LIST_HEAD(pmus
);
43 static DEFINE_MUTEX(pmus_lock
);
44 static struct srcu_struct pmus_srcu
;
47 * perf event paranoia level:
48 * -1 - not paranoid at all
49 * 0 - disallow raw tracepoint access for unpriv
50 * 1 - disallow cpu events for unpriv
51 * 2 - disallow kernel profiling for unpriv
53 int sysctl_perf_event_paranoid __read_mostly
= 1;
55 int sysctl_perf_event_mlock __read_mostly
= 512; /* 'free' kb per user */
58 * max perf event sample rate
60 int sysctl_perf_event_sample_rate __read_mostly
= 100000;
62 static atomic64_t perf_event_id
;
64 void __weak
perf_event_print_debug(void) { }
66 extern __weak
const char *perf_pmu_name(void)
71 void perf_pmu_disable(struct pmu
*pmu
)
73 int *count
= this_cpu_ptr(pmu
->pmu_disable_count
);
75 pmu
->pmu_disable(pmu
);
78 void perf_pmu_enable(struct pmu
*pmu
)
80 int *count
= this_cpu_ptr(pmu
->pmu_disable_count
);
85 static DEFINE_PER_CPU(struct list_head
, rotation_list
);
88 * perf_pmu_rotate_start() and perf_rotate_context() are fully serialized
89 * because they're strictly cpu affine and rotate_start is called with IRQs
90 * disabled, while rotate_context is called from IRQ context.
92 static void perf_pmu_rotate_start(struct pmu
*pmu
)
94 struct perf_cpu_context
*cpuctx
= this_cpu_ptr(pmu
->pmu_cpu_context
);
95 struct list_head
*head
= &__get_cpu_var(rotation_list
);
97 WARN_ON(!irqs_disabled());
99 if (list_empty(&cpuctx
->rotation_list
))
100 list_add(&cpuctx
->rotation_list
, head
);
103 static void get_ctx(struct perf_event_context
*ctx
)
105 WARN_ON(!atomic_inc_not_zero(&ctx
->refcount
));
108 static void free_ctx(struct rcu_head
*head
)
110 struct perf_event_context
*ctx
;
112 ctx
= container_of(head
, struct perf_event_context
, rcu_head
);
116 static void put_ctx(struct perf_event_context
*ctx
)
118 if (atomic_dec_and_test(&ctx
->refcount
)) {
120 put_ctx(ctx
->parent_ctx
);
122 put_task_struct(ctx
->task
);
123 call_rcu(&ctx
->rcu_head
, free_ctx
);
127 static void unclone_ctx(struct perf_event_context
*ctx
)
129 if (ctx
->parent_ctx
) {
130 put_ctx(ctx
->parent_ctx
);
131 ctx
->parent_ctx
= NULL
;
136 * If we inherit events we want to return the parent event id
139 static u64
primary_event_id(struct perf_event
*event
)
144 id
= event
->parent
->id
;
150 * Get the perf_event_context for a task and lock it.
151 * This has to cope with with the fact that until it is locked,
152 * the context could get moved to another task.
154 static struct perf_event_context
*
155 perf_lock_task_context(struct task_struct
*task
, int ctxn
, unsigned long *flags
)
157 struct perf_event_context
*ctx
;
161 ctx
= rcu_dereference(task
->perf_event_ctxp
[ctxn
]);
164 * If this context is a clone of another, it might
165 * get swapped for another underneath us by
166 * perf_event_task_sched_out, though the
167 * rcu_read_lock() protects us from any context
168 * getting freed. Lock the context and check if it
169 * got swapped before we could get the lock, and retry
170 * if so. If we locked the right context, then it
171 * can't get swapped on us any more.
173 raw_spin_lock_irqsave(&ctx
->lock
, *flags
);
174 if (ctx
!= rcu_dereference(task
->perf_event_ctxp
[ctxn
])) {
175 raw_spin_unlock_irqrestore(&ctx
->lock
, *flags
);
179 if (!atomic_inc_not_zero(&ctx
->refcount
)) {
180 raw_spin_unlock_irqrestore(&ctx
->lock
, *flags
);
189 * Get the context for a task and increment its pin_count so it
190 * can't get swapped to another task. This also increments its
191 * reference count so that the context can't get freed.
193 static struct perf_event_context
*
194 perf_pin_task_context(struct task_struct
*task
, int ctxn
)
196 struct perf_event_context
*ctx
;
199 ctx
= perf_lock_task_context(task
, ctxn
, &flags
);
202 raw_spin_unlock_irqrestore(&ctx
->lock
, flags
);
207 static void perf_unpin_context(struct perf_event_context
*ctx
)
211 raw_spin_lock_irqsave(&ctx
->lock
, flags
);
213 raw_spin_unlock_irqrestore(&ctx
->lock
, flags
);
217 static inline u64
perf_clock(void)
219 return local_clock();
223 * Update the record of the current time in a context.
225 static void update_context_time(struct perf_event_context
*ctx
)
227 u64 now
= perf_clock();
229 ctx
->time
+= now
- ctx
->timestamp
;
230 ctx
->timestamp
= now
;
234 * Update the total_time_enabled and total_time_running fields for a event.
236 static void update_event_times(struct perf_event
*event
)
238 struct perf_event_context
*ctx
= event
->ctx
;
241 if (event
->state
< PERF_EVENT_STATE_INACTIVE
||
242 event
->group_leader
->state
< PERF_EVENT_STATE_INACTIVE
)
248 run_end
= event
->tstamp_stopped
;
250 event
->total_time_enabled
= run_end
- event
->tstamp_enabled
;
252 if (event
->state
== PERF_EVENT_STATE_INACTIVE
)
253 run_end
= event
->tstamp_stopped
;
257 event
->total_time_running
= run_end
- event
->tstamp_running
;
261 * Update total_time_enabled and total_time_running for all events in a group.
263 static void update_group_times(struct perf_event
*leader
)
265 struct perf_event
*event
;
267 update_event_times(leader
);
268 list_for_each_entry(event
, &leader
->sibling_list
, group_entry
)
269 update_event_times(event
);
272 static struct list_head
*
273 ctx_group_list(struct perf_event
*event
, struct perf_event_context
*ctx
)
275 if (event
->attr
.pinned
)
276 return &ctx
->pinned_groups
;
278 return &ctx
->flexible_groups
;
282 * Add a event from the lists for its context.
283 * Must be called with ctx->mutex and ctx->lock held.
286 list_add_event(struct perf_event
*event
, struct perf_event_context
*ctx
)
288 WARN_ON_ONCE(event
->attach_state
& PERF_ATTACH_CONTEXT
);
289 event
->attach_state
|= PERF_ATTACH_CONTEXT
;
292 * If we're a stand alone event or group leader, we go to the context
293 * list, group events are kept attached to the group so that
294 * perf_group_detach can, at all times, locate all siblings.
296 if (event
->group_leader
== event
) {
297 struct list_head
*list
;
299 if (is_software_event(event
))
300 event
->group_flags
|= PERF_GROUP_SOFTWARE
;
302 list
= ctx_group_list(event
, ctx
);
303 list_add_tail(&event
->group_entry
, list
);
306 list_add_rcu(&event
->event_entry
, &ctx
->event_list
);
308 perf_pmu_rotate_start(ctx
->pmu
);
310 if (event
->attr
.inherit_stat
)
314 static void perf_group_attach(struct perf_event
*event
)
316 struct perf_event
*group_leader
= event
->group_leader
;
318 WARN_ON_ONCE(event
->attach_state
& PERF_ATTACH_GROUP
);
319 event
->attach_state
|= PERF_ATTACH_GROUP
;
321 if (group_leader
== event
)
324 if (group_leader
->group_flags
& PERF_GROUP_SOFTWARE
&&
325 !is_software_event(event
))
326 group_leader
->group_flags
&= ~PERF_GROUP_SOFTWARE
;
328 list_add_tail(&event
->group_entry
, &group_leader
->sibling_list
);
329 group_leader
->nr_siblings
++;
333 * Remove a event from the lists for its context.
334 * Must be called with ctx->mutex and ctx->lock held.
337 list_del_event(struct perf_event
*event
, struct perf_event_context
*ctx
)
340 * We can have double detach due to exit/hot-unplug + close.
342 if (!(event
->attach_state
& PERF_ATTACH_CONTEXT
))
345 event
->attach_state
&= ~PERF_ATTACH_CONTEXT
;
348 if (event
->attr
.inherit_stat
)
351 list_del_rcu(&event
->event_entry
);
353 if (event
->group_leader
== event
)
354 list_del_init(&event
->group_entry
);
356 update_group_times(event
);
359 * If event was in error state, then keep it
360 * that way, otherwise bogus counts will be
361 * returned on read(). The only way to get out
362 * of error state is by explicit re-enabling
365 if (event
->state
> PERF_EVENT_STATE_OFF
)
366 event
->state
= PERF_EVENT_STATE_OFF
;
369 static void perf_group_detach(struct perf_event
*event
)
371 struct perf_event
*sibling
, *tmp
;
372 struct list_head
*list
= NULL
;
375 * We can have double detach due to exit/hot-unplug + close.
377 if (!(event
->attach_state
& PERF_ATTACH_GROUP
))
380 event
->attach_state
&= ~PERF_ATTACH_GROUP
;
383 * If this is a sibling, remove it from its group.
385 if (event
->group_leader
!= event
) {
386 list_del_init(&event
->group_entry
);
387 event
->group_leader
->nr_siblings
--;
391 if (!list_empty(&event
->group_entry
))
392 list
= &event
->group_entry
;
395 * If this was a group event with sibling events then
396 * upgrade the siblings to singleton events by adding them
397 * to whatever list we are on.
399 list_for_each_entry_safe(sibling
, tmp
, &event
->sibling_list
, group_entry
) {
401 list_move_tail(&sibling
->group_entry
, list
);
402 sibling
->group_leader
= sibling
;
404 /* Inherit group flags from the previous leader */
405 sibling
->group_flags
= event
->group_flags
;
410 event_filter_match(struct perf_event
*event
)
412 return event
->cpu
== -1 || event
->cpu
== smp_processor_id();
416 __event_sched_out(struct perf_event
*event
,
417 struct perf_cpu_context
*cpuctx
,
418 struct perf_event_context
*ctx
)
422 * An event which could not be activated because of
423 * filter mismatch still needs to have its timings
424 * maintained, otherwise bogus information is return
425 * via read() for time_enabled, time_running:
427 if (event
->state
== PERF_EVENT_STATE_INACTIVE
428 && !event_filter_match(event
)) {
429 delta
= ctx
->time
- event
->tstamp_stopped
;
430 event
->tstamp_running
+= delta
;
431 event
->tstamp_stopped
= ctx
->time
;
434 if (event
->state
!= PERF_EVENT_STATE_ACTIVE
)
437 event
->state
= PERF_EVENT_STATE_INACTIVE
;
438 if (event
->pending_disable
) {
439 event
->pending_disable
= 0;
440 event
->state
= PERF_EVENT_STATE_OFF
;
442 event
->pmu
->del(event
, 0);
445 if (!is_software_event(event
))
446 cpuctx
->active_oncpu
--;
448 if (event
->attr
.exclusive
|| !cpuctx
->active_oncpu
)
449 cpuctx
->exclusive
= 0;
454 event_sched_out(struct perf_event
*event
,
455 struct perf_cpu_context
*cpuctx
,
456 struct perf_event_context
*ctx
)
460 ret
= __event_sched_out(event
, cpuctx
, ctx
);
462 event
->tstamp_stopped
= ctx
->time
;
466 group_sched_out(struct perf_event
*group_event
,
467 struct perf_cpu_context
*cpuctx
,
468 struct perf_event_context
*ctx
)
470 struct perf_event
*event
;
471 int state
= group_event
->state
;
473 event_sched_out(group_event
, cpuctx
, ctx
);
476 * Schedule out siblings (if any):
478 list_for_each_entry(event
, &group_event
->sibling_list
, group_entry
)
479 event_sched_out(event
, cpuctx
, ctx
);
481 if (state
== PERF_EVENT_STATE_ACTIVE
&& group_event
->attr
.exclusive
)
482 cpuctx
->exclusive
= 0;
485 static inline struct perf_cpu_context
*
486 __get_cpu_context(struct perf_event_context
*ctx
)
488 return this_cpu_ptr(ctx
->pmu
->pmu_cpu_context
);
492 * Cross CPU call to remove a performance event
494 * We disable the event on the hardware level first. After that we
495 * remove it from the context list.
497 static void __perf_event_remove_from_context(void *info
)
499 struct perf_event
*event
= info
;
500 struct perf_event_context
*ctx
= event
->ctx
;
501 struct perf_cpu_context
*cpuctx
= __get_cpu_context(ctx
);
504 * If this is a task context, we need to check whether it is
505 * the current task context of this cpu. If not it has been
506 * scheduled out before the smp call arrived.
508 if (ctx
->task
&& cpuctx
->task_ctx
!= ctx
)
511 raw_spin_lock(&ctx
->lock
);
513 event_sched_out(event
, cpuctx
, ctx
);
515 list_del_event(event
, ctx
);
517 raw_spin_unlock(&ctx
->lock
);
522 * Remove the event from a task's (or a CPU's) list of events.
524 * Must be called with ctx->mutex held.
526 * CPU events are removed with a smp call. For task events we only
527 * call when the task is on a CPU.
529 * If event->ctx is a cloned context, callers must make sure that
530 * every task struct that event->ctx->task could possibly point to
531 * remains valid. This is OK when called from perf_release since
532 * that only calls us on the top-level context, which can't be a clone.
533 * When called from perf_event_exit_task, it's OK because the
534 * context has been detached from its task.
536 static void perf_event_remove_from_context(struct perf_event
*event
)
538 struct perf_event_context
*ctx
= event
->ctx
;
539 struct task_struct
*task
= ctx
->task
;
543 * Per cpu events are removed via an smp call and
544 * the removal is always successful.
546 smp_call_function_single(event
->cpu
,
547 __perf_event_remove_from_context
,
553 task_oncpu_function_call(task
, __perf_event_remove_from_context
,
556 raw_spin_lock_irq(&ctx
->lock
);
558 * If the context is active we need to retry the smp call.
560 if (ctx
->nr_active
&& !list_empty(&event
->group_entry
)) {
561 raw_spin_unlock_irq(&ctx
->lock
);
566 * The lock prevents that this context is scheduled in so we
567 * can remove the event safely, if the call above did not
570 if (!list_empty(&event
->group_entry
))
571 list_del_event(event
, ctx
);
572 raw_spin_unlock_irq(&ctx
->lock
);
576 * Cross CPU call to disable a performance event
578 static void __perf_event_disable(void *info
)
580 struct perf_event
*event
= info
;
581 struct perf_event_context
*ctx
= event
->ctx
;
582 struct perf_cpu_context
*cpuctx
= __get_cpu_context(ctx
);
585 * If this is a per-task event, need to check whether this
586 * event's task is the current task on this cpu.
588 if (ctx
->task
&& cpuctx
->task_ctx
!= ctx
)
591 raw_spin_lock(&ctx
->lock
);
594 * If the event is on, turn it off.
595 * If it is in error state, leave it in error state.
597 if (event
->state
>= PERF_EVENT_STATE_INACTIVE
) {
598 update_context_time(ctx
);
599 update_group_times(event
);
600 if (event
== event
->group_leader
)
601 group_sched_out(event
, cpuctx
, ctx
);
603 event_sched_out(event
, cpuctx
, ctx
);
604 event
->state
= PERF_EVENT_STATE_OFF
;
607 raw_spin_unlock(&ctx
->lock
);
613 * If event->ctx is a cloned context, callers must make sure that
614 * every task struct that event->ctx->task could possibly point to
615 * remains valid. This condition is satisifed when called through
616 * perf_event_for_each_child or perf_event_for_each because they
617 * hold the top-level event's child_mutex, so any descendant that
618 * goes to exit will block in sync_child_event.
619 * When called from perf_pending_event it's OK because event->ctx
620 * is the current context on this CPU and preemption is disabled,
621 * hence we can't get into perf_event_task_sched_out for this context.
623 void perf_event_disable(struct perf_event
*event
)
625 struct perf_event_context
*ctx
= event
->ctx
;
626 struct task_struct
*task
= ctx
->task
;
630 * Disable the event on the cpu that it's on
632 smp_call_function_single(event
->cpu
, __perf_event_disable
,
638 task_oncpu_function_call(task
, __perf_event_disable
, event
);
640 raw_spin_lock_irq(&ctx
->lock
);
642 * If the event is still active, we need to retry the cross-call.
644 if (event
->state
== PERF_EVENT_STATE_ACTIVE
) {
645 raw_spin_unlock_irq(&ctx
->lock
);
650 * Since we have the lock this context can't be scheduled
651 * in, so we can change the state safely.
653 if (event
->state
== PERF_EVENT_STATE_INACTIVE
) {
654 update_group_times(event
);
655 event
->state
= PERF_EVENT_STATE_OFF
;
658 raw_spin_unlock_irq(&ctx
->lock
);
662 __event_sched_in(struct perf_event
*event
,
663 struct perf_cpu_context
*cpuctx
,
664 struct perf_event_context
*ctx
)
666 if (event
->state
<= PERF_EVENT_STATE_OFF
)
669 event
->state
= PERF_EVENT_STATE_ACTIVE
;
670 event
->oncpu
= smp_processor_id();
672 * The new state must be visible before we turn it on in the hardware:
676 if (event
->pmu
->add(event
, PERF_EF_START
)) {
677 event
->state
= PERF_EVENT_STATE_INACTIVE
;
682 if (!is_software_event(event
))
683 cpuctx
->active_oncpu
++;
686 if (event
->attr
.exclusive
)
687 cpuctx
->exclusive
= 1;
693 event_sched_in(struct perf_event
*event
,
694 struct perf_cpu_context
*cpuctx
,
695 struct perf_event_context
*ctx
)
697 int ret
= __event_sched_in(event
, cpuctx
, ctx
);
700 event
->tstamp_running
+= ctx
->time
- event
->tstamp_stopped
;
705 group_commit_event_sched_in(struct perf_event
*group_event
,
706 struct perf_cpu_context
*cpuctx
,
707 struct perf_event_context
*ctx
)
709 struct perf_event
*event
;
712 group_event
->tstamp_running
+= now
- group_event
->tstamp_stopped
;
714 * Schedule in siblings as one group (if any):
716 list_for_each_entry(event
, &group_event
->sibling_list
, group_entry
) {
717 event
->tstamp_running
+= now
- event
->tstamp_stopped
;
722 group_sched_in(struct perf_event
*group_event
,
723 struct perf_cpu_context
*cpuctx
,
724 struct perf_event_context
*ctx
)
726 struct perf_event
*event
, *partial_group
= NULL
;
727 struct pmu
*pmu
= group_event
->pmu
;
729 if (group_event
->state
== PERF_EVENT_STATE_OFF
)
735 * use __event_sched_in() to delay updating tstamp_running
736 * until the transaction is committed. In case of failure
737 * we will keep an unmodified tstamp_running which is a
738 * requirement to get correct timing information
740 if (__event_sched_in(group_event
, cpuctx
, ctx
)) {
741 pmu
->cancel_txn(pmu
);
746 * Schedule in siblings as one group (if any):
748 list_for_each_entry(event
, &group_event
->sibling_list
, group_entry
) {
749 if (__event_sched_in(event
, cpuctx
, ctx
)) {
750 partial_group
= event
;
755 if (!pmu
->commit_txn(pmu
)) {
756 /* commit tstamp_running */
757 group_commit_event_sched_in(group_event
, cpuctx
, ctx
);
762 * Groups can be scheduled in as one unit only, so undo any
763 * partial group before returning:
765 * use __event_sched_out() to avoid updating tstamp_stopped
766 * because the event never actually ran
768 list_for_each_entry(event
, &group_event
->sibling_list
, group_entry
) {
769 if (event
== partial_group
)
771 __event_sched_out(event
, cpuctx
, ctx
);
773 __event_sched_out(group_event
, cpuctx
, ctx
);
775 pmu
->cancel_txn(pmu
);
781 * Work out whether we can put this event group on the CPU now.
783 static int group_can_go_on(struct perf_event
*event
,
784 struct perf_cpu_context
*cpuctx
,
788 * Groups consisting entirely of software events can always go on.
790 if (event
->group_flags
& PERF_GROUP_SOFTWARE
)
793 * If an exclusive group is already on, no other hardware
796 if (cpuctx
->exclusive
)
799 * If this group is exclusive and there are already
800 * events on the CPU, it can't go on.
802 if (event
->attr
.exclusive
&& cpuctx
->active_oncpu
)
805 * Otherwise, try to add it if all previous groups were able
811 static void add_event_to_ctx(struct perf_event
*event
,
812 struct perf_event_context
*ctx
)
814 list_add_event(event
, ctx
);
815 perf_group_attach(event
);
816 event
->tstamp_enabled
= ctx
->time
;
817 event
->tstamp_running
= ctx
->time
;
818 event
->tstamp_stopped
= ctx
->time
;
822 * Cross CPU call to install and enable a performance event
824 * Must be called with ctx->mutex held
826 static void __perf_install_in_context(void *info
)
828 struct perf_event
*event
= info
;
829 struct perf_event_context
*ctx
= event
->ctx
;
830 struct perf_event
*leader
= event
->group_leader
;
831 struct perf_cpu_context
*cpuctx
= __get_cpu_context(ctx
);
835 * If this is a task context, we need to check whether it is
836 * the current task context of this cpu. If not it has been
837 * scheduled out before the smp call arrived.
838 * Or possibly this is the right context but it isn't
839 * on this cpu because it had no events.
841 if (ctx
->task
&& cpuctx
->task_ctx
!= ctx
) {
842 if (cpuctx
->task_ctx
|| ctx
->task
!= current
)
844 cpuctx
->task_ctx
= ctx
;
847 raw_spin_lock(&ctx
->lock
);
849 update_context_time(ctx
);
851 add_event_to_ctx(event
, ctx
);
853 if (event
->cpu
!= -1 && event
->cpu
!= smp_processor_id())
857 * Don't put the event on if it is disabled or if
858 * it is in a group and the group isn't on.
860 if (event
->state
!= PERF_EVENT_STATE_INACTIVE
||
861 (leader
!= event
&& leader
->state
!= PERF_EVENT_STATE_ACTIVE
))
865 * An exclusive event can't go on if there are already active
866 * hardware events, and no hardware event can go on if there
867 * is already an exclusive event on.
869 if (!group_can_go_on(event
, cpuctx
, 1))
872 err
= event_sched_in(event
, cpuctx
, ctx
);
876 * This event couldn't go on. If it is in a group
877 * then we have to pull the whole group off.
878 * If the event group is pinned then put it in error state.
881 group_sched_out(leader
, cpuctx
, ctx
);
882 if (leader
->attr
.pinned
) {
883 update_group_times(leader
);
884 leader
->state
= PERF_EVENT_STATE_ERROR
;
889 raw_spin_unlock(&ctx
->lock
);
893 * Attach a performance event to a context
895 * First we add the event to the list with the hardware enable bit
896 * in event->hw_config cleared.
898 * If the event is attached to a task which is on a CPU we use a smp
899 * call to enable it in the task context. The task might have been
900 * scheduled away, but we check this in the smp call again.
902 * Must be called with ctx->mutex held.
905 perf_install_in_context(struct perf_event_context
*ctx
,
906 struct perf_event
*event
,
909 struct task_struct
*task
= ctx
->task
;
915 * Per cpu events are installed via an smp call and
916 * the install is always successful.
918 smp_call_function_single(cpu
, __perf_install_in_context
,
924 task_oncpu_function_call(task
, __perf_install_in_context
,
927 raw_spin_lock_irq(&ctx
->lock
);
929 * we need to retry the smp call.
931 if (ctx
->is_active
&& list_empty(&event
->group_entry
)) {
932 raw_spin_unlock_irq(&ctx
->lock
);
937 * The lock prevents that this context is scheduled in so we
938 * can add the event safely, if it the call above did not
941 if (list_empty(&event
->group_entry
))
942 add_event_to_ctx(event
, ctx
);
943 raw_spin_unlock_irq(&ctx
->lock
);
947 * Put a event into inactive state and update time fields.
948 * Enabling the leader of a group effectively enables all
949 * the group members that aren't explicitly disabled, so we
950 * have to update their ->tstamp_enabled also.
951 * Note: this works for group members as well as group leaders
952 * since the non-leader members' sibling_lists will be empty.
954 static void __perf_event_mark_enabled(struct perf_event
*event
,
955 struct perf_event_context
*ctx
)
957 struct perf_event
*sub
;
959 event
->state
= PERF_EVENT_STATE_INACTIVE
;
960 event
->tstamp_enabled
= ctx
->time
- event
->total_time_enabled
;
961 list_for_each_entry(sub
, &event
->sibling_list
, group_entry
) {
962 if (sub
->state
>= PERF_EVENT_STATE_INACTIVE
) {
963 sub
->tstamp_enabled
=
964 ctx
->time
- sub
->total_time_enabled
;
970 * Cross CPU call to enable a performance event
972 static void __perf_event_enable(void *info
)
974 struct perf_event
*event
= info
;
975 struct perf_event_context
*ctx
= event
->ctx
;
976 struct perf_event
*leader
= event
->group_leader
;
977 struct perf_cpu_context
*cpuctx
= __get_cpu_context(ctx
);
981 * If this is a per-task event, need to check whether this
982 * event's task is the current task on this cpu.
984 if (ctx
->task
&& cpuctx
->task_ctx
!= ctx
) {
985 if (cpuctx
->task_ctx
|| ctx
->task
!= current
)
987 cpuctx
->task_ctx
= ctx
;
990 raw_spin_lock(&ctx
->lock
);
992 update_context_time(ctx
);
994 if (event
->state
>= PERF_EVENT_STATE_INACTIVE
)
996 __perf_event_mark_enabled(event
, ctx
);
998 if (event
->cpu
!= -1 && event
->cpu
!= smp_processor_id())
1002 * If the event is in a group and isn't the group leader,
1003 * then don't put it on unless the group is on.
1005 if (leader
!= event
&& leader
->state
!= PERF_EVENT_STATE_ACTIVE
)
1008 if (!group_can_go_on(event
, cpuctx
, 1)) {
1011 if (event
== leader
)
1012 err
= group_sched_in(event
, cpuctx
, ctx
);
1014 err
= event_sched_in(event
, cpuctx
, ctx
);
1019 * If this event can't go on and it's part of a
1020 * group, then the whole group has to come off.
1022 if (leader
!= event
)
1023 group_sched_out(leader
, cpuctx
, ctx
);
1024 if (leader
->attr
.pinned
) {
1025 update_group_times(leader
);
1026 leader
->state
= PERF_EVENT_STATE_ERROR
;
1031 raw_spin_unlock(&ctx
->lock
);
1037 * If event->ctx is a cloned context, callers must make sure that
1038 * every task struct that event->ctx->task could possibly point to
1039 * remains valid. This condition is satisfied when called through
1040 * perf_event_for_each_child or perf_event_for_each as described
1041 * for perf_event_disable.
1043 void perf_event_enable(struct perf_event
*event
)
1045 struct perf_event_context
*ctx
= event
->ctx
;
1046 struct task_struct
*task
= ctx
->task
;
1050 * Enable the event on the cpu that it's on
1052 smp_call_function_single(event
->cpu
, __perf_event_enable
,
1057 raw_spin_lock_irq(&ctx
->lock
);
1058 if (event
->state
>= PERF_EVENT_STATE_INACTIVE
)
1062 * If the event is in error state, clear that first.
1063 * That way, if we see the event in error state below, we
1064 * know that it has gone back into error state, as distinct
1065 * from the task having been scheduled away before the
1066 * cross-call arrived.
1068 if (event
->state
== PERF_EVENT_STATE_ERROR
)
1069 event
->state
= PERF_EVENT_STATE_OFF
;
1072 raw_spin_unlock_irq(&ctx
->lock
);
1073 task_oncpu_function_call(task
, __perf_event_enable
, event
);
1075 raw_spin_lock_irq(&ctx
->lock
);
1078 * If the context is active and the event is still off,
1079 * we need to retry the cross-call.
1081 if (ctx
->is_active
&& event
->state
== PERF_EVENT_STATE_OFF
)
1085 * Since we have the lock this context can't be scheduled
1086 * in, so we can change the state safely.
1088 if (event
->state
== PERF_EVENT_STATE_OFF
)
1089 __perf_event_mark_enabled(event
, ctx
);
1092 raw_spin_unlock_irq(&ctx
->lock
);
1095 static int perf_event_refresh(struct perf_event
*event
, int refresh
)
1098 * not supported on inherited events
1100 if (event
->attr
.inherit
)
1103 atomic_add(refresh
, &event
->event_limit
);
1104 perf_event_enable(event
);
1110 EVENT_FLEXIBLE
= 0x1,
1112 EVENT_ALL
= EVENT_FLEXIBLE
| EVENT_PINNED
,
1115 static void ctx_sched_out(struct perf_event_context
*ctx
,
1116 struct perf_cpu_context
*cpuctx
,
1117 enum event_type_t event_type
)
1119 struct perf_event
*event
;
1121 raw_spin_lock(&ctx
->lock
);
1122 perf_pmu_disable(ctx
->pmu
);
1124 if (likely(!ctx
->nr_events
))
1126 update_context_time(ctx
);
1128 if (!ctx
->nr_active
)
1131 if (event_type
& EVENT_PINNED
) {
1132 list_for_each_entry(event
, &ctx
->pinned_groups
, group_entry
)
1133 group_sched_out(event
, cpuctx
, ctx
);
1136 if (event_type
& EVENT_FLEXIBLE
) {
1137 list_for_each_entry(event
, &ctx
->flexible_groups
, group_entry
)
1138 group_sched_out(event
, cpuctx
, ctx
);
1141 perf_pmu_enable(ctx
->pmu
);
1142 raw_spin_unlock(&ctx
->lock
);
1146 * Test whether two contexts are equivalent, i.e. whether they
1147 * have both been cloned from the same version of the same context
1148 * and they both have the same number of enabled events.
1149 * If the number of enabled events is the same, then the set
1150 * of enabled events should be the same, because these are both
1151 * inherited contexts, therefore we can't access individual events
1152 * in them directly with an fd; we can only enable/disable all
1153 * events via prctl, or enable/disable all events in a family
1154 * via ioctl, which will have the same effect on both contexts.
1156 static int context_equiv(struct perf_event_context
*ctx1
,
1157 struct perf_event_context
*ctx2
)
1159 return ctx1
->parent_ctx
&& ctx1
->parent_ctx
== ctx2
->parent_ctx
1160 && ctx1
->parent_gen
== ctx2
->parent_gen
1161 && !ctx1
->pin_count
&& !ctx2
->pin_count
;
1164 static void __perf_event_sync_stat(struct perf_event
*event
,
1165 struct perf_event
*next_event
)
1169 if (!event
->attr
.inherit_stat
)
1173 * Update the event value, we cannot use perf_event_read()
1174 * because we're in the middle of a context switch and have IRQs
1175 * disabled, which upsets smp_call_function_single(), however
1176 * we know the event must be on the current CPU, therefore we
1177 * don't need to use it.
1179 switch (event
->state
) {
1180 case PERF_EVENT_STATE_ACTIVE
:
1181 event
->pmu
->read(event
);
1184 case PERF_EVENT_STATE_INACTIVE
:
1185 update_event_times(event
);
1193 * In order to keep per-task stats reliable we need to flip the event
1194 * values when we flip the contexts.
1196 value
= local64_read(&next_event
->count
);
1197 value
= local64_xchg(&event
->count
, value
);
1198 local64_set(&next_event
->count
, value
);
1200 swap(event
->total_time_enabled
, next_event
->total_time_enabled
);
1201 swap(event
->total_time_running
, next_event
->total_time_running
);
1204 * Since we swizzled the values, update the user visible data too.
1206 perf_event_update_userpage(event
);
1207 perf_event_update_userpage(next_event
);
1210 #define list_next_entry(pos, member) \
1211 list_entry(pos->member.next, typeof(*pos), member)
1213 static void perf_event_sync_stat(struct perf_event_context
*ctx
,
1214 struct perf_event_context
*next_ctx
)
1216 struct perf_event
*event
, *next_event
;
1221 update_context_time(ctx
);
1223 event
= list_first_entry(&ctx
->event_list
,
1224 struct perf_event
, event_entry
);
1226 next_event
= list_first_entry(&next_ctx
->event_list
,
1227 struct perf_event
, event_entry
);
1229 while (&event
->event_entry
!= &ctx
->event_list
&&
1230 &next_event
->event_entry
!= &next_ctx
->event_list
) {
1232 __perf_event_sync_stat(event
, next_event
);
1234 event
= list_next_entry(event
, event_entry
);
1235 next_event
= list_next_entry(next_event
, event_entry
);
1239 void perf_event_context_sched_out(struct task_struct
*task
, int ctxn
,
1240 struct task_struct
*next
)
1242 struct perf_event_context
*ctx
= task
->perf_event_ctxp
[ctxn
];
1243 struct perf_event_context
*next_ctx
;
1244 struct perf_event_context
*parent
;
1245 struct perf_cpu_context
*cpuctx
;
1251 cpuctx
= __get_cpu_context(ctx
);
1252 if (!cpuctx
->task_ctx
)
1256 parent
= rcu_dereference(ctx
->parent_ctx
);
1257 next_ctx
= next
->perf_event_ctxp
[ctxn
];
1258 if (parent
&& next_ctx
&&
1259 rcu_dereference(next_ctx
->parent_ctx
) == parent
) {
1261 * Looks like the two contexts are clones, so we might be
1262 * able to optimize the context switch. We lock both
1263 * contexts and check that they are clones under the
1264 * lock (including re-checking that neither has been
1265 * uncloned in the meantime). It doesn't matter which
1266 * order we take the locks because no other cpu could
1267 * be trying to lock both of these tasks.
1269 raw_spin_lock(&ctx
->lock
);
1270 raw_spin_lock_nested(&next_ctx
->lock
, SINGLE_DEPTH_NESTING
);
1271 if (context_equiv(ctx
, next_ctx
)) {
1273 * XXX do we need a memory barrier of sorts
1274 * wrt to rcu_dereference() of perf_event_ctxp
1276 task
->perf_event_ctxp
[ctxn
] = next_ctx
;
1277 next
->perf_event_ctxp
[ctxn
] = ctx
;
1279 next_ctx
->task
= task
;
1282 perf_event_sync_stat(ctx
, next_ctx
);
1284 raw_spin_unlock(&next_ctx
->lock
);
1285 raw_spin_unlock(&ctx
->lock
);
1290 ctx_sched_out(ctx
, cpuctx
, EVENT_ALL
);
1291 cpuctx
->task_ctx
= NULL
;
1295 #define for_each_task_context_nr(ctxn) \
1296 for ((ctxn) = 0; (ctxn) < perf_nr_task_contexts; (ctxn)++)
1299 * Called from scheduler to remove the events of the current task,
1300 * with interrupts disabled.
1302 * We stop each event and update the event value in event->count.
1304 * This does not protect us against NMI, but disable()
1305 * sets the disabled bit in the control field of event _before_
1306 * accessing the event control register. If a NMI hits, then it will
1307 * not restart the event.
1309 void perf_event_task_sched_out(struct task_struct
*task
,
1310 struct task_struct
*next
)
1314 perf_sw_event(PERF_COUNT_SW_CONTEXT_SWITCHES
, 1, 1, NULL
, 0);
1316 for_each_task_context_nr(ctxn
)
1317 perf_event_context_sched_out(task
, ctxn
, next
);
1320 static void task_ctx_sched_out(struct perf_event_context
*ctx
,
1321 enum event_type_t event_type
)
1323 struct perf_cpu_context
*cpuctx
= __get_cpu_context(ctx
);
1325 if (!cpuctx
->task_ctx
)
1328 if (WARN_ON_ONCE(ctx
!= cpuctx
->task_ctx
))
1331 ctx_sched_out(ctx
, cpuctx
, event_type
);
1332 cpuctx
->task_ctx
= NULL
;
1336 * Called with IRQs disabled
1338 static void __perf_event_task_sched_out(struct perf_event_context
*ctx
)
1340 task_ctx_sched_out(ctx
, EVENT_ALL
);
1344 * Called with IRQs disabled
1346 static void cpu_ctx_sched_out(struct perf_cpu_context
*cpuctx
,
1347 enum event_type_t event_type
)
1349 ctx_sched_out(&cpuctx
->ctx
, cpuctx
, event_type
);
1353 ctx_pinned_sched_in(struct perf_event_context
*ctx
,
1354 struct perf_cpu_context
*cpuctx
)
1356 struct perf_event
*event
;
1358 list_for_each_entry(event
, &ctx
->pinned_groups
, group_entry
) {
1359 if (event
->state
<= PERF_EVENT_STATE_OFF
)
1361 if (event
->cpu
!= -1 && event
->cpu
!= smp_processor_id())
1364 if (group_can_go_on(event
, cpuctx
, 1))
1365 group_sched_in(event
, cpuctx
, ctx
);
1368 * If this pinned group hasn't been scheduled,
1369 * put it in error state.
1371 if (event
->state
== PERF_EVENT_STATE_INACTIVE
) {
1372 update_group_times(event
);
1373 event
->state
= PERF_EVENT_STATE_ERROR
;
1379 ctx_flexible_sched_in(struct perf_event_context
*ctx
,
1380 struct perf_cpu_context
*cpuctx
)
1382 struct perf_event
*event
;
1385 list_for_each_entry(event
, &ctx
->flexible_groups
, group_entry
) {
1386 /* Ignore events in OFF or ERROR state */
1387 if (event
->state
<= PERF_EVENT_STATE_OFF
)
1390 * Listen to the 'cpu' scheduling filter constraint
1393 if (event
->cpu
!= -1 && event
->cpu
!= smp_processor_id())
1396 if (group_can_go_on(event
, cpuctx
, can_add_hw
)) {
1397 if (group_sched_in(event
, cpuctx
, ctx
))
1404 ctx_sched_in(struct perf_event_context
*ctx
,
1405 struct perf_cpu_context
*cpuctx
,
1406 enum event_type_t event_type
)
1408 raw_spin_lock(&ctx
->lock
);
1410 if (likely(!ctx
->nr_events
))
1413 ctx
->timestamp
= perf_clock();
1416 * First go through the list and put on any pinned groups
1417 * in order to give them the best chance of going on.
1419 if (event_type
& EVENT_PINNED
)
1420 ctx_pinned_sched_in(ctx
, cpuctx
);
1422 /* Then walk through the lower prio flexible groups */
1423 if (event_type
& EVENT_FLEXIBLE
)
1424 ctx_flexible_sched_in(ctx
, cpuctx
);
1427 raw_spin_unlock(&ctx
->lock
);
1430 static void cpu_ctx_sched_in(struct perf_cpu_context
*cpuctx
,
1431 enum event_type_t event_type
)
1433 struct perf_event_context
*ctx
= &cpuctx
->ctx
;
1435 ctx_sched_in(ctx
, cpuctx
, event_type
);
1438 static void task_ctx_sched_in(struct perf_event_context
*ctx
,
1439 enum event_type_t event_type
)
1441 struct perf_cpu_context
*cpuctx
;
1443 cpuctx
= __get_cpu_context(ctx
);
1444 if (cpuctx
->task_ctx
== ctx
)
1447 ctx_sched_in(ctx
, cpuctx
, event_type
);
1448 cpuctx
->task_ctx
= ctx
;
1451 void perf_event_context_sched_in(struct perf_event_context
*ctx
)
1453 struct perf_cpu_context
*cpuctx
;
1455 cpuctx
= __get_cpu_context(ctx
);
1456 if (cpuctx
->task_ctx
== ctx
)
1459 perf_pmu_disable(ctx
->pmu
);
1461 * We want to keep the following priority order:
1462 * cpu pinned (that don't need to move), task pinned,
1463 * cpu flexible, task flexible.
1465 cpu_ctx_sched_out(cpuctx
, EVENT_FLEXIBLE
);
1467 ctx_sched_in(ctx
, cpuctx
, EVENT_PINNED
);
1468 cpu_ctx_sched_in(cpuctx
, EVENT_FLEXIBLE
);
1469 ctx_sched_in(ctx
, cpuctx
, EVENT_FLEXIBLE
);
1471 cpuctx
->task_ctx
= ctx
;
1474 * Since these rotations are per-cpu, we need to ensure the
1475 * cpu-context we got scheduled on is actually rotating.
1477 perf_pmu_rotate_start(ctx
->pmu
);
1478 perf_pmu_enable(ctx
->pmu
);
1482 * Called from scheduler to add the events of the current task
1483 * with interrupts disabled.
1485 * We restore the event value and then enable it.
1487 * This does not protect us against NMI, but enable()
1488 * sets the enabled bit in the control field of event _before_
1489 * accessing the event control register. If a NMI hits, then it will
1490 * keep the event running.
1492 void perf_event_task_sched_in(struct task_struct
*task
)
1494 struct perf_event_context
*ctx
;
1497 for_each_task_context_nr(ctxn
) {
1498 ctx
= task
->perf_event_ctxp
[ctxn
];
1502 perf_event_context_sched_in(ctx
);
1506 #define MAX_INTERRUPTS (~0ULL)
1508 static void perf_log_throttle(struct perf_event
*event
, int enable
);
1510 static u64
perf_calculate_period(struct perf_event
*event
, u64 nsec
, u64 count
)
1512 u64 frequency
= event
->attr
.sample_freq
;
1513 u64 sec
= NSEC_PER_SEC
;
1514 u64 divisor
, dividend
;
1516 int count_fls
, nsec_fls
, frequency_fls
, sec_fls
;
1518 count_fls
= fls64(count
);
1519 nsec_fls
= fls64(nsec
);
1520 frequency_fls
= fls64(frequency
);
1524 * We got @count in @nsec, with a target of sample_freq HZ
1525 * the target period becomes:
1528 * period = -------------------
1529 * @nsec * sample_freq
1534 * Reduce accuracy by one bit such that @a and @b converge
1535 * to a similar magnitude.
1537 #define REDUCE_FLS(a, b) \
1539 if (a##_fls > b##_fls) { \
1549 * Reduce accuracy until either term fits in a u64, then proceed with
1550 * the other, so that finally we can do a u64/u64 division.
1552 while (count_fls
+ sec_fls
> 64 && nsec_fls
+ frequency_fls
> 64) {
1553 REDUCE_FLS(nsec
, frequency
);
1554 REDUCE_FLS(sec
, count
);
1557 if (count_fls
+ sec_fls
> 64) {
1558 divisor
= nsec
* frequency
;
1560 while (count_fls
+ sec_fls
> 64) {
1561 REDUCE_FLS(count
, sec
);
1565 dividend
= count
* sec
;
1567 dividend
= count
* sec
;
1569 while (nsec_fls
+ frequency_fls
> 64) {
1570 REDUCE_FLS(nsec
, frequency
);
1574 divisor
= nsec
* frequency
;
1580 return div64_u64(dividend
, divisor
);
1583 static void perf_adjust_period(struct perf_event
*event
, u64 nsec
, u64 count
)
1585 struct hw_perf_event
*hwc
= &event
->hw
;
1586 s64 period
, sample_period
;
1589 period
= perf_calculate_period(event
, nsec
, count
);
1591 delta
= (s64
)(period
- hwc
->sample_period
);
1592 delta
= (delta
+ 7) / 8; /* low pass filter */
1594 sample_period
= hwc
->sample_period
+ delta
;
1599 hwc
->sample_period
= sample_period
;
1601 if (local64_read(&hwc
->period_left
) > 8*sample_period
) {
1602 event
->pmu
->stop(event
, PERF_EF_UPDATE
);
1603 local64_set(&hwc
->period_left
, 0);
1604 event
->pmu
->start(event
, PERF_EF_RELOAD
);
1608 static void perf_ctx_adjust_freq(struct perf_event_context
*ctx
, u64 period
)
1610 struct perf_event
*event
;
1611 struct hw_perf_event
*hwc
;
1612 u64 interrupts
, now
;
1615 raw_spin_lock(&ctx
->lock
);
1616 list_for_each_entry_rcu(event
, &ctx
->event_list
, event_entry
) {
1617 if (event
->state
!= PERF_EVENT_STATE_ACTIVE
)
1620 if (event
->cpu
!= -1 && event
->cpu
!= smp_processor_id())
1625 interrupts
= hwc
->interrupts
;
1626 hwc
->interrupts
= 0;
1629 * unthrottle events on the tick
1631 if (interrupts
== MAX_INTERRUPTS
) {
1632 perf_log_throttle(event
, 1);
1633 event
->pmu
->start(event
, 0);
1636 if (!event
->attr
.freq
|| !event
->attr
.sample_freq
)
1639 event
->pmu
->read(event
);
1640 now
= local64_read(&event
->count
);
1641 delta
= now
- hwc
->freq_count_stamp
;
1642 hwc
->freq_count_stamp
= now
;
1645 perf_adjust_period(event
, period
, delta
);
1647 raw_spin_unlock(&ctx
->lock
);
1651 * Round-robin a context's events:
1653 static void rotate_ctx(struct perf_event_context
*ctx
)
1655 raw_spin_lock(&ctx
->lock
);
1657 /* Rotate the first entry last of non-pinned groups */
1658 list_rotate_left(&ctx
->flexible_groups
);
1660 raw_spin_unlock(&ctx
->lock
);
1664 * perf_pmu_rotate_start() and perf_rotate_context() are fully serialized
1665 * because they're strictly cpu affine and rotate_start is called with IRQs
1666 * disabled, while rotate_context is called from IRQ context.
1668 static void perf_rotate_context(struct perf_cpu_context
*cpuctx
)
1670 u64 interval
= (u64
)cpuctx
->jiffies_interval
* TICK_NSEC
;
1671 struct perf_event_context
*ctx
= NULL
;
1672 int rotate
= 0, remove
= 1;
1674 if (cpuctx
->ctx
.nr_events
) {
1676 if (cpuctx
->ctx
.nr_events
!= cpuctx
->ctx
.nr_active
)
1680 ctx
= cpuctx
->task_ctx
;
1681 if (ctx
&& ctx
->nr_events
) {
1683 if (ctx
->nr_events
!= ctx
->nr_active
)
1687 perf_pmu_disable(cpuctx
->ctx
.pmu
);
1688 perf_ctx_adjust_freq(&cpuctx
->ctx
, interval
);
1690 perf_ctx_adjust_freq(ctx
, interval
);
1695 cpu_ctx_sched_out(cpuctx
, EVENT_FLEXIBLE
);
1697 task_ctx_sched_out(ctx
, EVENT_FLEXIBLE
);
1699 rotate_ctx(&cpuctx
->ctx
);
1703 cpu_ctx_sched_in(cpuctx
, EVENT_FLEXIBLE
);
1705 task_ctx_sched_in(ctx
, EVENT_FLEXIBLE
);
1709 list_del_init(&cpuctx
->rotation_list
);
1711 perf_pmu_enable(cpuctx
->ctx
.pmu
);
1714 void perf_event_task_tick(void)
1716 struct list_head
*head
= &__get_cpu_var(rotation_list
);
1717 struct perf_cpu_context
*cpuctx
, *tmp
;
1719 WARN_ON(!irqs_disabled());
1721 list_for_each_entry_safe(cpuctx
, tmp
, head
, rotation_list
) {
1722 if (cpuctx
->jiffies_interval
== 1 ||
1723 !(jiffies
% cpuctx
->jiffies_interval
))
1724 perf_rotate_context(cpuctx
);
1728 static int event_enable_on_exec(struct perf_event
*event
,
1729 struct perf_event_context
*ctx
)
1731 if (!event
->attr
.enable_on_exec
)
1734 event
->attr
.enable_on_exec
= 0;
1735 if (event
->state
>= PERF_EVENT_STATE_INACTIVE
)
1738 __perf_event_mark_enabled(event
, ctx
);
1744 * Enable all of a task's events that have been marked enable-on-exec.
1745 * This expects task == current.
1747 static void perf_event_enable_on_exec(struct perf_event_context
*ctx
)
1749 struct perf_event
*event
;
1750 unsigned long flags
;
1754 local_irq_save(flags
);
1755 if (!ctx
|| !ctx
->nr_events
)
1758 task_ctx_sched_out(ctx
, EVENT_ALL
);
1760 raw_spin_lock(&ctx
->lock
);
1762 list_for_each_entry(event
, &ctx
->pinned_groups
, group_entry
) {
1763 ret
= event_enable_on_exec(event
, ctx
);
1768 list_for_each_entry(event
, &ctx
->flexible_groups
, group_entry
) {
1769 ret
= event_enable_on_exec(event
, ctx
);
1775 * Unclone this context if we enabled any event.
1780 raw_spin_unlock(&ctx
->lock
);
1782 perf_event_context_sched_in(ctx
);
1784 local_irq_restore(flags
);
1788 * Cross CPU call to read the hardware event
1790 static void __perf_event_read(void *info
)
1792 struct perf_event
*event
= info
;
1793 struct perf_event_context
*ctx
= event
->ctx
;
1794 struct perf_cpu_context
*cpuctx
= __get_cpu_context(ctx
);
1797 * If this is a task context, we need to check whether it is
1798 * the current task context of this cpu. If not it has been
1799 * scheduled out before the smp call arrived. In that case
1800 * event->count would have been updated to a recent sample
1801 * when the event was scheduled out.
1803 if (ctx
->task
&& cpuctx
->task_ctx
!= ctx
)
1806 raw_spin_lock(&ctx
->lock
);
1807 update_context_time(ctx
);
1808 update_event_times(event
);
1809 raw_spin_unlock(&ctx
->lock
);
1811 event
->pmu
->read(event
);
1814 static inline u64
perf_event_count(struct perf_event
*event
)
1816 return local64_read(&event
->count
) + atomic64_read(&event
->child_count
);
1819 static u64
perf_event_read(struct perf_event
*event
)
1822 * If event is enabled and currently active on a CPU, update the
1823 * value in the event structure:
1825 if (event
->state
== PERF_EVENT_STATE_ACTIVE
) {
1826 smp_call_function_single(event
->oncpu
,
1827 __perf_event_read
, event
, 1);
1828 } else if (event
->state
== PERF_EVENT_STATE_INACTIVE
) {
1829 struct perf_event_context
*ctx
= event
->ctx
;
1830 unsigned long flags
;
1832 raw_spin_lock_irqsave(&ctx
->lock
, flags
);
1834 * may read while context is not active
1835 * (e.g., thread is blocked), in that case
1836 * we cannot update context time
1839 update_context_time(ctx
);
1840 update_event_times(event
);
1841 raw_spin_unlock_irqrestore(&ctx
->lock
, flags
);
1844 return perf_event_count(event
);
1851 struct callchain_cpus_entries
{
1852 struct rcu_head rcu_head
;
1853 struct perf_callchain_entry
*cpu_entries
[0];
1856 static DEFINE_PER_CPU(int, callchain_recursion
[PERF_NR_CONTEXTS
]);
1857 static atomic_t nr_callchain_events
;
1858 static DEFINE_MUTEX(callchain_mutex
);
1859 struct callchain_cpus_entries
*callchain_cpus_entries
;
1862 __weak
void perf_callchain_kernel(struct perf_callchain_entry
*entry
,
1863 struct pt_regs
*regs
)
1867 __weak
void perf_callchain_user(struct perf_callchain_entry
*entry
,
1868 struct pt_regs
*regs
)
1872 static void release_callchain_buffers_rcu(struct rcu_head
*head
)
1874 struct callchain_cpus_entries
*entries
;
1877 entries
= container_of(head
, struct callchain_cpus_entries
, rcu_head
);
1879 for_each_possible_cpu(cpu
)
1880 kfree(entries
->cpu_entries
[cpu
]);
1885 static void release_callchain_buffers(void)
1887 struct callchain_cpus_entries
*entries
;
1889 entries
= callchain_cpus_entries
;
1890 rcu_assign_pointer(callchain_cpus_entries
, NULL
);
1891 call_rcu(&entries
->rcu_head
, release_callchain_buffers_rcu
);
1894 static int alloc_callchain_buffers(void)
1898 struct callchain_cpus_entries
*entries
;
1901 * We can't use the percpu allocation API for data that can be
1902 * accessed from NMI. Use a temporary manual per cpu allocation
1903 * until that gets sorted out.
1905 size
= sizeof(*entries
) + sizeof(struct perf_callchain_entry
*) *
1906 num_possible_cpus();
1908 entries
= kzalloc(size
, GFP_KERNEL
);
1912 size
= sizeof(struct perf_callchain_entry
) * PERF_NR_CONTEXTS
;
1914 for_each_possible_cpu(cpu
) {
1915 entries
->cpu_entries
[cpu
] = kmalloc_node(size
, GFP_KERNEL
,
1917 if (!entries
->cpu_entries
[cpu
])
1921 rcu_assign_pointer(callchain_cpus_entries
, entries
);
1926 for_each_possible_cpu(cpu
)
1927 kfree(entries
->cpu_entries
[cpu
]);
1933 static int get_callchain_buffers(void)
1938 mutex_lock(&callchain_mutex
);
1940 count
= atomic_inc_return(&nr_callchain_events
);
1941 if (WARN_ON_ONCE(count
< 1)) {
1947 /* If the allocation failed, give up */
1948 if (!callchain_cpus_entries
)
1953 err
= alloc_callchain_buffers();
1955 release_callchain_buffers();
1957 mutex_unlock(&callchain_mutex
);
1962 static void put_callchain_buffers(void)
1964 if (atomic_dec_and_mutex_lock(&nr_callchain_events
, &callchain_mutex
)) {
1965 release_callchain_buffers();
1966 mutex_unlock(&callchain_mutex
);
1970 static int get_recursion_context(int *recursion
)
1978 else if (in_softirq())
1983 if (recursion
[rctx
])
1992 static inline void put_recursion_context(int *recursion
, int rctx
)
1998 static struct perf_callchain_entry
*get_callchain_entry(int *rctx
)
2001 struct callchain_cpus_entries
*entries
;
2003 *rctx
= get_recursion_context(__get_cpu_var(callchain_recursion
));
2007 entries
= rcu_dereference(callchain_cpus_entries
);
2011 cpu
= smp_processor_id();
2013 return &entries
->cpu_entries
[cpu
][*rctx
];
2017 put_callchain_entry(int rctx
)
2019 put_recursion_context(__get_cpu_var(callchain_recursion
), rctx
);
2022 static struct perf_callchain_entry
*perf_callchain(struct pt_regs
*regs
)
2025 struct perf_callchain_entry
*entry
;
2028 entry
= get_callchain_entry(&rctx
);
2037 if (!user_mode(regs
)) {
2038 perf_callchain_store(entry
, PERF_CONTEXT_KERNEL
);
2039 perf_callchain_kernel(entry
, regs
);
2041 regs
= task_pt_regs(current
);
2047 perf_callchain_store(entry
, PERF_CONTEXT_USER
);
2048 perf_callchain_user(entry
, regs
);
2052 put_callchain_entry(rctx
);
2058 * Initialize the perf_event context in a task_struct:
2060 static void __perf_event_init_context(struct perf_event_context
*ctx
)
2062 raw_spin_lock_init(&ctx
->lock
);
2063 mutex_init(&ctx
->mutex
);
2064 INIT_LIST_HEAD(&ctx
->pinned_groups
);
2065 INIT_LIST_HEAD(&ctx
->flexible_groups
);
2066 INIT_LIST_HEAD(&ctx
->event_list
);
2067 atomic_set(&ctx
->refcount
, 1);
2070 static struct perf_event_context
*
2071 alloc_perf_context(struct pmu
*pmu
, struct task_struct
*task
)
2073 struct perf_event_context
*ctx
;
2075 ctx
= kzalloc(sizeof(struct perf_event_context
), GFP_KERNEL
);
2079 __perf_event_init_context(ctx
);
2082 get_task_struct(task
);
2089 static struct task_struct
*
2090 find_lively_task_by_vpid(pid_t vpid
)
2092 struct task_struct
*task
;
2099 task
= find_task_by_vpid(vpid
);
2101 get_task_struct(task
);
2105 return ERR_PTR(-ESRCH
);
2108 * Can't attach events to a dying task.
2111 if (task
->flags
& PF_EXITING
)
2114 /* Reuse ptrace permission checks for now. */
2116 if (!ptrace_may_access(task
, PTRACE_MODE_READ
))
2121 put_task_struct(task
);
2122 return ERR_PTR(err
);
2126 static struct perf_event_context
*
2127 find_get_context(struct pmu
*pmu
, struct task_struct
*task
, int cpu
)
2129 struct perf_event_context
*ctx
;
2130 struct perf_cpu_context
*cpuctx
;
2131 unsigned long flags
;
2134 if (!task
&& cpu
!= -1) {
2135 /* Must be root to operate on a CPU event: */
2136 if (perf_paranoid_cpu() && !capable(CAP_SYS_ADMIN
))
2137 return ERR_PTR(-EACCES
);
2139 if (cpu
< 0 || cpu
>= nr_cpumask_bits
)
2140 return ERR_PTR(-EINVAL
);
2143 * We could be clever and allow to attach a event to an
2144 * offline CPU and activate it when the CPU comes up, but
2147 if (!cpu_online(cpu
))
2148 return ERR_PTR(-ENODEV
);
2150 cpuctx
= per_cpu_ptr(pmu
->pmu_cpu_context
, cpu
);
2158 ctxn
= pmu
->task_ctx_nr
;
2163 ctx
= perf_lock_task_context(task
, ctxn
, &flags
);
2166 raw_spin_unlock_irqrestore(&ctx
->lock
, flags
);
2170 ctx
= alloc_perf_context(pmu
, task
);
2177 if (cmpxchg(&task
->perf_event_ctxp
[ctxn
], NULL
, ctx
)) {
2179 * We raced with some other task; use
2180 * the context they set.
2182 put_task_struct(task
);
2188 put_task_struct(task
);
2192 put_task_struct(task
);
2193 return ERR_PTR(err
);
2196 static void perf_event_free_filter(struct perf_event
*event
);
2198 static void free_event_rcu(struct rcu_head
*head
)
2200 struct perf_event
*event
;
2202 event
= container_of(head
, struct perf_event
, rcu_head
);
2204 put_pid_ns(event
->ns
);
2205 perf_event_free_filter(event
);
2209 static void perf_pending_sync(struct perf_event
*event
);
2210 static void perf_buffer_put(struct perf_buffer
*buffer
);
2212 static void free_event(struct perf_event
*event
)
2214 perf_pending_sync(event
);
2216 if (!event
->parent
) {
2217 atomic_dec(&nr_events
);
2218 if (event
->attr
.mmap
|| event
->attr
.mmap_data
)
2219 atomic_dec(&nr_mmap_events
);
2220 if (event
->attr
.comm
)
2221 atomic_dec(&nr_comm_events
);
2222 if (event
->attr
.task
)
2223 atomic_dec(&nr_task_events
);
2224 if (event
->attr
.sample_type
& PERF_SAMPLE_CALLCHAIN
)
2225 put_callchain_buffers();
2228 if (event
->buffer
) {
2229 perf_buffer_put(event
->buffer
);
2230 event
->buffer
= NULL
;
2234 event
->destroy(event
);
2237 put_ctx(event
->ctx
);
2239 call_rcu(&event
->rcu_head
, free_event_rcu
);
2242 int perf_event_release_kernel(struct perf_event
*event
)
2244 struct perf_event_context
*ctx
= event
->ctx
;
2247 * Remove from the PMU, can't get re-enabled since we got
2248 * here because the last ref went.
2250 perf_event_disable(event
);
2252 WARN_ON_ONCE(ctx
->parent_ctx
);
2254 * There are two ways this annotation is useful:
2256 * 1) there is a lock recursion from perf_event_exit_task
2257 * see the comment there.
2259 * 2) there is a lock-inversion with mmap_sem through
2260 * perf_event_read_group(), which takes faults while
2261 * holding ctx->mutex, however this is called after
2262 * the last filedesc died, so there is no possibility
2263 * to trigger the AB-BA case.
2265 mutex_lock_nested(&ctx
->mutex
, SINGLE_DEPTH_NESTING
);
2266 raw_spin_lock_irq(&ctx
->lock
);
2267 perf_group_detach(event
);
2268 list_del_event(event
, ctx
);
2269 raw_spin_unlock_irq(&ctx
->lock
);
2270 mutex_unlock(&ctx
->mutex
);
2272 mutex_lock(&event
->owner
->perf_event_mutex
);
2273 list_del_init(&event
->owner_entry
);
2274 mutex_unlock(&event
->owner
->perf_event_mutex
);
2275 put_task_struct(event
->owner
);
2281 EXPORT_SYMBOL_GPL(perf_event_release_kernel
);
2284 * Called when the last reference to the file is gone.
2286 static int perf_release(struct inode
*inode
, struct file
*file
)
2288 struct perf_event
*event
= file
->private_data
;
2290 file
->private_data
= NULL
;
2292 return perf_event_release_kernel(event
);
2295 static int perf_event_read_size(struct perf_event
*event
)
2297 int entry
= sizeof(u64
); /* value */
2301 if (event
->attr
.read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
)
2302 size
+= sizeof(u64
);
2304 if (event
->attr
.read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
)
2305 size
+= sizeof(u64
);
2307 if (event
->attr
.read_format
& PERF_FORMAT_ID
)
2308 entry
+= sizeof(u64
);
2310 if (event
->attr
.read_format
& PERF_FORMAT_GROUP
) {
2311 nr
+= event
->group_leader
->nr_siblings
;
2312 size
+= sizeof(u64
);
2320 u64
perf_event_read_value(struct perf_event
*event
, u64
*enabled
, u64
*running
)
2322 struct perf_event
*child
;
2328 mutex_lock(&event
->child_mutex
);
2329 total
+= perf_event_read(event
);
2330 *enabled
+= event
->total_time_enabled
+
2331 atomic64_read(&event
->child_total_time_enabled
);
2332 *running
+= event
->total_time_running
+
2333 atomic64_read(&event
->child_total_time_running
);
2335 list_for_each_entry(child
, &event
->child_list
, child_list
) {
2336 total
+= perf_event_read(child
);
2337 *enabled
+= child
->total_time_enabled
;
2338 *running
+= child
->total_time_running
;
2340 mutex_unlock(&event
->child_mutex
);
2344 EXPORT_SYMBOL_GPL(perf_event_read_value
);
2346 static int perf_event_read_group(struct perf_event
*event
,
2347 u64 read_format
, char __user
*buf
)
2349 struct perf_event
*leader
= event
->group_leader
, *sub
;
2350 int n
= 0, size
= 0, ret
= -EFAULT
;
2351 struct perf_event_context
*ctx
= leader
->ctx
;
2353 u64 count
, enabled
, running
;
2355 mutex_lock(&ctx
->mutex
);
2356 count
= perf_event_read_value(leader
, &enabled
, &running
);
2358 values
[n
++] = 1 + leader
->nr_siblings
;
2359 if (read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
)
2360 values
[n
++] = enabled
;
2361 if (read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
)
2362 values
[n
++] = running
;
2363 values
[n
++] = count
;
2364 if (read_format
& PERF_FORMAT_ID
)
2365 values
[n
++] = primary_event_id(leader
);
2367 size
= n
* sizeof(u64
);
2369 if (copy_to_user(buf
, values
, size
))
2374 list_for_each_entry(sub
, &leader
->sibling_list
, group_entry
) {
2377 values
[n
++] = perf_event_read_value(sub
, &enabled
, &running
);
2378 if (read_format
& PERF_FORMAT_ID
)
2379 values
[n
++] = primary_event_id(sub
);
2381 size
= n
* sizeof(u64
);
2383 if (copy_to_user(buf
+ ret
, values
, size
)) {
2391 mutex_unlock(&ctx
->mutex
);
2396 static int perf_event_read_one(struct perf_event
*event
,
2397 u64 read_format
, char __user
*buf
)
2399 u64 enabled
, running
;
2403 values
[n
++] = perf_event_read_value(event
, &enabled
, &running
);
2404 if (read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
)
2405 values
[n
++] = enabled
;
2406 if (read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
)
2407 values
[n
++] = running
;
2408 if (read_format
& PERF_FORMAT_ID
)
2409 values
[n
++] = primary_event_id(event
);
2411 if (copy_to_user(buf
, values
, n
* sizeof(u64
)))
2414 return n
* sizeof(u64
);
2418 * Read the performance event - simple non blocking version for now
2421 perf_read_hw(struct perf_event
*event
, char __user
*buf
, size_t count
)
2423 u64 read_format
= event
->attr
.read_format
;
2427 * Return end-of-file for a read on a event that is in
2428 * error state (i.e. because it was pinned but it couldn't be
2429 * scheduled on to the CPU at some point).
2431 if (event
->state
== PERF_EVENT_STATE_ERROR
)
2434 if (count
< perf_event_read_size(event
))
2437 WARN_ON_ONCE(event
->ctx
->parent_ctx
);
2438 if (read_format
& PERF_FORMAT_GROUP
)
2439 ret
= perf_event_read_group(event
, read_format
, buf
);
2441 ret
= perf_event_read_one(event
, read_format
, buf
);
2447 perf_read(struct file
*file
, char __user
*buf
, size_t count
, loff_t
*ppos
)
2449 struct perf_event
*event
= file
->private_data
;
2451 return perf_read_hw(event
, buf
, count
);
2454 static unsigned int perf_poll(struct file
*file
, poll_table
*wait
)
2456 struct perf_event
*event
= file
->private_data
;
2457 struct perf_buffer
*buffer
;
2458 unsigned int events
= POLL_HUP
;
2461 buffer
= rcu_dereference(event
->buffer
);
2463 events
= atomic_xchg(&buffer
->poll
, 0);
2466 poll_wait(file
, &event
->waitq
, wait
);
2471 static void perf_event_reset(struct perf_event
*event
)
2473 (void)perf_event_read(event
);
2474 local64_set(&event
->count
, 0);
2475 perf_event_update_userpage(event
);
2479 * Holding the top-level event's child_mutex means that any
2480 * descendant process that has inherited this event will block
2481 * in sync_child_event if it goes to exit, thus satisfying the
2482 * task existence requirements of perf_event_enable/disable.
2484 static void perf_event_for_each_child(struct perf_event
*event
,
2485 void (*func
)(struct perf_event
*))
2487 struct perf_event
*child
;
2489 WARN_ON_ONCE(event
->ctx
->parent_ctx
);
2490 mutex_lock(&event
->child_mutex
);
2492 list_for_each_entry(child
, &event
->child_list
, child_list
)
2494 mutex_unlock(&event
->child_mutex
);
2497 static void perf_event_for_each(struct perf_event
*event
,
2498 void (*func
)(struct perf_event
*))
2500 struct perf_event_context
*ctx
= event
->ctx
;
2501 struct perf_event
*sibling
;
2503 WARN_ON_ONCE(ctx
->parent_ctx
);
2504 mutex_lock(&ctx
->mutex
);
2505 event
= event
->group_leader
;
2507 perf_event_for_each_child(event
, func
);
2509 list_for_each_entry(sibling
, &event
->sibling_list
, group_entry
)
2510 perf_event_for_each_child(event
, func
);
2511 mutex_unlock(&ctx
->mutex
);
2514 static int perf_event_period(struct perf_event
*event
, u64 __user
*arg
)
2516 struct perf_event_context
*ctx
= event
->ctx
;
2521 if (!event
->attr
.sample_period
)
2524 size
= copy_from_user(&value
, arg
, sizeof(value
));
2525 if (size
!= sizeof(value
))
2531 raw_spin_lock_irq(&ctx
->lock
);
2532 if (event
->attr
.freq
) {
2533 if (value
> sysctl_perf_event_sample_rate
) {
2538 event
->attr
.sample_freq
= value
;
2540 event
->attr
.sample_period
= value
;
2541 event
->hw
.sample_period
= value
;
2544 raw_spin_unlock_irq(&ctx
->lock
);
2549 static const struct file_operations perf_fops
;
2551 static struct perf_event
*perf_fget_light(int fd
, int *fput_needed
)
2555 file
= fget_light(fd
, fput_needed
);
2557 return ERR_PTR(-EBADF
);
2559 if (file
->f_op
!= &perf_fops
) {
2560 fput_light(file
, *fput_needed
);
2562 return ERR_PTR(-EBADF
);
2565 return file
->private_data
;
2568 static int perf_event_set_output(struct perf_event
*event
,
2569 struct perf_event
*output_event
);
2570 static int perf_event_set_filter(struct perf_event
*event
, void __user
*arg
);
2572 static long perf_ioctl(struct file
*file
, unsigned int cmd
, unsigned long arg
)
2574 struct perf_event
*event
= file
->private_data
;
2575 void (*func
)(struct perf_event
*);
2579 case PERF_EVENT_IOC_ENABLE
:
2580 func
= perf_event_enable
;
2582 case PERF_EVENT_IOC_DISABLE
:
2583 func
= perf_event_disable
;
2585 case PERF_EVENT_IOC_RESET
:
2586 func
= perf_event_reset
;
2589 case PERF_EVENT_IOC_REFRESH
:
2590 return perf_event_refresh(event
, arg
);
2592 case PERF_EVENT_IOC_PERIOD
:
2593 return perf_event_period(event
, (u64 __user
*)arg
);
2595 case PERF_EVENT_IOC_SET_OUTPUT
:
2597 struct perf_event
*output_event
= NULL
;
2598 int fput_needed
= 0;
2602 output_event
= perf_fget_light(arg
, &fput_needed
);
2603 if (IS_ERR(output_event
))
2604 return PTR_ERR(output_event
);
2607 ret
= perf_event_set_output(event
, output_event
);
2609 fput_light(output_event
->filp
, fput_needed
);
2614 case PERF_EVENT_IOC_SET_FILTER
:
2615 return perf_event_set_filter(event
, (void __user
*)arg
);
2621 if (flags
& PERF_IOC_FLAG_GROUP
)
2622 perf_event_for_each(event
, func
);
2624 perf_event_for_each_child(event
, func
);
2629 int perf_event_task_enable(void)
2631 struct perf_event
*event
;
2633 mutex_lock(¤t
->perf_event_mutex
);
2634 list_for_each_entry(event
, ¤t
->perf_event_list
, owner_entry
)
2635 perf_event_for_each_child(event
, perf_event_enable
);
2636 mutex_unlock(¤t
->perf_event_mutex
);
2641 int perf_event_task_disable(void)
2643 struct perf_event
*event
;
2645 mutex_lock(¤t
->perf_event_mutex
);
2646 list_for_each_entry(event
, ¤t
->perf_event_list
, owner_entry
)
2647 perf_event_for_each_child(event
, perf_event_disable
);
2648 mutex_unlock(¤t
->perf_event_mutex
);
2653 #ifndef PERF_EVENT_INDEX_OFFSET
2654 # define PERF_EVENT_INDEX_OFFSET 0
2657 static int perf_event_index(struct perf_event
*event
)
2659 if (event
->hw
.state
& PERF_HES_STOPPED
)
2662 if (event
->state
!= PERF_EVENT_STATE_ACTIVE
)
2665 return event
->hw
.idx
+ 1 - PERF_EVENT_INDEX_OFFSET
;
2669 * Callers need to ensure there can be no nesting of this function, otherwise
2670 * the seqlock logic goes bad. We can not serialize this because the arch
2671 * code calls this from NMI context.
2673 void perf_event_update_userpage(struct perf_event
*event
)
2675 struct perf_event_mmap_page
*userpg
;
2676 struct perf_buffer
*buffer
;
2679 buffer
= rcu_dereference(event
->buffer
);
2683 userpg
= buffer
->user_page
;
2686 * Disable preemption so as to not let the corresponding user-space
2687 * spin too long if we get preempted.
2692 userpg
->index
= perf_event_index(event
);
2693 userpg
->offset
= perf_event_count(event
);
2694 if (event
->state
== PERF_EVENT_STATE_ACTIVE
)
2695 userpg
->offset
-= local64_read(&event
->hw
.prev_count
);
2697 userpg
->time_enabled
= event
->total_time_enabled
+
2698 atomic64_read(&event
->child_total_time_enabled
);
2700 userpg
->time_running
= event
->total_time_running
+
2701 atomic64_read(&event
->child_total_time_running
);
2710 static unsigned long perf_data_size(struct perf_buffer
*buffer
);
2713 perf_buffer_init(struct perf_buffer
*buffer
, long watermark
, int flags
)
2715 long max_size
= perf_data_size(buffer
);
2718 buffer
->watermark
= min(max_size
, watermark
);
2720 if (!buffer
->watermark
)
2721 buffer
->watermark
= max_size
/ 2;
2723 if (flags
& PERF_BUFFER_WRITABLE
)
2724 buffer
->writable
= 1;
2726 atomic_set(&buffer
->refcount
, 1);
2729 #ifndef CONFIG_PERF_USE_VMALLOC
2732 * Back perf_mmap() with regular GFP_KERNEL-0 pages.
2735 static struct page
*
2736 perf_mmap_to_page(struct perf_buffer
*buffer
, unsigned long pgoff
)
2738 if (pgoff
> buffer
->nr_pages
)
2742 return virt_to_page(buffer
->user_page
);
2744 return virt_to_page(buffer
->data_pages
[pgoff
- 1]);
2747 static void *perf_mmap_alloc_page(int cpu
)
2752 node
= (cpu
== -1) ? cpu
: cpu_to_node(cpu
);
2753 page
= alloc_pages_node(node
, GFP_KERNEL
| __GFP_ZERO
, 0);
2757 return page_address(page
);
2760 static struct perf_buffer
*
2761 perf_buffer_alloc(int nr_pages
, long watermark
, int cpu
, int flags
)
2763 struct perf_buffer
*buffer
;
2767 size
= sizeof(struct perf_buffer
);
2768 size
+= nr_pages
* sizeof(void *);
2770 buffer
= kzalloc(size
, GFP_KERNEL
);
2774 buffer
->user_page
= perf_mmap_alloc_page(cpu
);
2775 if (!buffer
->user_page
)
2776 goto fail_user_page
;
2778 for (i
= 0; i
< nr_pages
; i
++) {
2779 buffer
->data_pages
[i
] = perf_mmap_alloc_page(cpu
);
2780 if (!buffer
->data_pages
[i
])
2781 goto fail_data_pages
;
2784 buffer
->nr_pages
= nr_pages
;
2786 perf_buffer_init(buffer
, watermark
, flags
);
2791 for (i
--; i
>= 0; i
--)
2792 free_page((unsigned long)buffer
->data_pages
[i
]);
2794 free_page((unsigned long)buffer
->user_page
);
2803 static void perf_mmap_free_page(unsigned long addr
)
2805 struct page
*page
= virt_to_page((void *)addr
);
2807 page
->mapping
= NULL
;
2811 static void perf_buffer_free(struct perf_buffer
*buffer
)
2815 perf_mmap_free_page((unsigned long)buffer
->user_page
);
2816 for (i
= 0; i
< buffer
->nr_pages
; i
++)
2817 perf_mmap_free_page((unsigned long)buffer
->data_pages
[i
]);
2821 static inline int page_order(struct perf_buffer
*buffer
)
2829 * Back perf_mmap() with vmalloc memory.
2831 * Required for architectures that have d-cache aliasing issues.
2834 static inline int page_order(struct perf_buffer
*buffer
)
2836 return buffer
->page_order
;
2839 static struct page
*
2840 perf_mmap_to_page(struct perf_buffer
*buffer
, unsigned long pgoff
)
2842 if (pgoff
> (1UL << page_order(buffer
)))
2845 return vmalloc_to_page((void *)buffer
->user_page
+ pgoff
* PAGE_SIZE
);
2848 static void perf_mmap_unmark_page(void *addr
)
2850 struct page
*page
= vmalloc_to_page(addr
);
2852 page
->mapping
= NULL
;
2855 static void perf_buffer_free_work(struct work_struct
*work
)
2857 struct perf_buffer
*buffer
;
2861 buffer
= container_of(work
, struct perf_buffer
, work
);
2862 nr
= 1 << page_order(buffer
);
2864 base
= buffer
->user_page
;
2865 for (i
= 0; i
< nr
+ 1; i
++)
2866 perf_mmap_unmark_page(base
+ (i
* PAGE_SIZE
));
2872 static void perf_buffer_free(struct perf_buffer
*buffer
)
2874 schedule_work(&buffer
->work
);
2877 static struct perf_buffer
*
2878 perf_buffer_alloc(int nr_pages
, long watermark
, int cpu
, int flags
)
2880 struct perf_buffer
*buffer
;
2884 size
= sizeof(struct perf_buffer
);
2885 size
+= sizeof(void *);
2887 buffer
= kzalloc(size
, GFP_KERNEL
);
2891 INIT_WORK(&buffer
->work
, perf_buffer_free_work
);
2893 all_buf
= vmalloc_user((nr_pages
+ 1) * PAGE_SIZE
);
2897 buffer
->user_page
= all_buf
;
2898 buffer
->data_pages
[0] = all_buf
+ PAGE_SIZE
;
2899 buffer
->page_order
= ilog2(nr_pages
);
2900 buffer
->nr_pages
= 1;
2902 perf_buffer_init(buffer
, watermark
, flags
);
2915 static unsigned long perf_data_size(struct perf_buffer
*buffer
)
2917 return buffer
->nr_pages
<< (PAGE_SHIFT
+ page_order(buffer
));
2920 static int perf_mmap_fault(struct vm_area_struct
*vma
, struct vm_fault
*vmf
)
2922 struct perf_event
*event
= vma
->vm_file
->private_data
;
2923 struct perf_buffer
*buffer
;
2924 int ret
= VM_FAULT_SIGBUS
;
2926 if (vmf
->flags
& FAULT_FLAG_MKWRITE
) {
2927 if (vmf
->pgoff
== 0)
2933 buffer
= rcu_dereference(event
->buffer
);
2937 if (vmf
->pgoff
&& (vmf
->flags
& FAULT_FLAG_WRITE
))
2940 vmf
->page
= perf_mmap_to_page(buffer
, vmf
->pgoff
);
2944 get_page(vmf
->page
);
2945 vmf
->page
->mapping
= vma
->vm_file
->f_mapping
;
2946 vmf
->page
->index
= vmf
->pgoff
;
2955 static void perf_buffer_free_rcu(struct rcu_head
*rcu_head
)
2957 struct perf_buffer
*buffer
;
2959 buffer
= container_of(rcu_head
, struct perf_buffer
, rcu_head
);
2960 perf_buffer_free(buffer
);
2963 static struct perf_buffer
*perf_buffer_get(struct perf_event
*event
)
2965 struct perf_buffer
*buffer
;
2968 buffer
= rcu_dereference(event
->buffer
);
2970 if (!atomic_inc_not_zero(&buffer
->refcount
))
2978 static void perf_buffer_put(struct perf_buffer
*buffer
)
2980 if (!atomic_dec_and_test(&buffer
->refcount
))
2983 call_rcu(&buffer
->rcu_head
, perf_buffer_free_rcu
);
2986 static void perf_mmap_open(struct vm_area_struct
*vma
)
2988 struct perf_event
*event
= vma
->vm_file
->private_data
;
2990 atomic_inc(&event
->mmap_count
);
2993 static void perf_mmap_close(struct vm_area_struct
*vma
)
2995 struct perf_event
*event
= vma
->vm_file
->private_data
;
2997 if (atomic_dec_and_mutex_lock(&event
->mmap_count
, &event
->mmap_mutex
)) {
2998 unsigned long size
= perf_data_size(event
->buffer
);
2999 struct user_struct
*user
= event
->mmap_user
;
3000 struct perf_buffer
*buffer
= event
->buffer
;
3002 atomic_long_sub((size
>> PAGE_SHIFT
) + 1, &user
->locked_vm
);
3003 vma
->vm_mm
->locked_vm
-= event
->mmap_locked
;
3004 rcu_assign_pointer(event
->buffer
, NULL
);
3005 mutex_unlock(&event
->mmap_mutex
);
3007 perf_buffer_put(buffer
);
3012 static const struct vm_operations_struct perf_mmap_vmops
= {
3013 .open
= perf_mmap_open
,
3014 .close
= perf_mmap_close
,
3015 .fault
= perf_mmap_fault
,
3016 .page_mkwrite
= perf_mmap_fault
,
3019 static int perf_mmap(struct file
*file
, struct vm_area_struct
*vma
)
3021 struct perf_event
*event
= file
->private_data
;
3022 unsigned long user_locked
, user_lock_limit
;
3023 struct user_struct
*user
= current_user();
3024 unsigned long locked
, lock_limit
;
3025 struct perf_buffer
*buffer
;
3026 unsigned long vma_size
;
3027 unsigned long nr_pages
;
3028 long user_extra
, extra
;
3029 int ret
= 0, flags
= 0;
3032 * Don't allow mmap() of inherited per-task counters. This would
3033 * create a performance issue due to all children writing to the
3036 if (event
->cpu
== -1 && event
->attr
.inherit
)
3039 if (!(vma
->vm_flags
& VM_SHARED
))
3042 vma_size
= vma
->vm_end
- vma
->vm_start
;
3043 nr_pages
= (vma_size
/ PAGE_SIZE
) - 1;
3046 * If we have buffer pages ensure they're a power-of-two number, so we
3047 * can do bitmasks instead of modulo.
3049 if (nr_pages
!= 0 && !is_power_of_2(nr_pages
))
3052 if (vma_size
!= PAGE_SIZE
* (1 + nr_pages
))
3055 if (vma
->vm_pgoff
!= 0)
3058 WARN_ON_ONCE(event
->ctx
->parent_ctx
);
3059 mutex_lock(&event
->mmap_mutex
);
3060 if (event
->buffer
) {
3061 if (event
->buffer
->nr_pages
== nr_pages
)
3062 atomic_inc(&event
->buffer
->refcount
);
3068 user_extra
= nr_pages
+ 1;
3069 user_lock_limit
= sysctl_perf_event_mlock
>> (PAGE_SHIFT
- 10);
3072 * Increase the limit linearly with more CPUs:
3074 user_lock_limit
*= num_online_cpus();
3076 user_locked
= atomic_long_read(&user
->locked_vm
) + user_extra
;
3079 if (user_locked
> user_lock_limit
)
3080 extra
= user_locked
- user_lock_limit
;
3082 lock_limit
= rlimit(RLIMIT_MEMLOCK
);
3083 lock_limit
>>= PAGE_SHIFT
;
3084 locked
= vma
->vm_mm
->locked_vm
+ extra
;
3086 if ((locked
> lock_limit
) && perf_paranoid_tracepoint_raw() &&
3087 !capable(CAP_IPC_LOCK
)) {
3092 WARN_ON(event
->buffer
);
3094 if (vma
->vm_flags
& VM_WRITE
)
3095 flags
|= PERF_BUFFER_WRITABLE
;
3097 buffer
= perf_buffer_alloc(nr_pages
, event
->attr
.wakeup_watermark
,
3103 rcu_assign_pointer(event
->buffer
, buffer
);
3105 atomic_long_add(user_extra
, &user
->locked_vm
);
3106 event
->mmap_locked
= extra
;
3107 event
->mmap_user
= get_current_user();
3108 vma
->vm_mm
->locked_vm
+= event
->mmap_locked
;
3112 atomic_inc(&event
->mmap_count
);
3113 mutex_unlock(&event
->mmap_mutex
);
3115 vma
->vm_flags
|= VM_RESERVED
;
3116 vma
->vm_ops
= &perf_mmap_vmops
;
3121 static int perf_fasync(int fd
, struct file
*filp
, int on
)
3123 struct inode
*inode
= filp
->f_path
.dentry
->d_inode
;
3124 struct perf_event
*event
= filp
->private_data
;
3127 mutex_lock(&inode
->i_mutex
);
3128 retval
= fasync_helper(fd
, filp
, on
, &event
->fasync
);
3129 mutex_unlock(&inode
->i_mutex
);
3137 static const struct file_operations perf_fops
= {
3138 .llseek
= no_llseek
,
3139 .release
= perf_release
,
3142 .unlocked_ioctl
= perf_ioctl
,
3143 .compat_ioctl
= perf_ioctl
,
3145 .fasync
= perf_fasync
,
3151 * If there's data, ensure we set the poll() state and publish everything
3152 * to user-space before waking everybody up.
3155 void perf_event_wakeup(struct perf_event
*event
)
3157 wake_up_all(&event
->waitq
);
3159 if (event
->pending_kill
) {
3160 kill_fasync(&event
->fasync
, SIGIO
, event
->pending_kill
);
3161 event
->pending_kill
= 0;
3168 * Handle the case where we need to wakeup up from NMI (or rq->lock) context.
3170 * The NMI bit means we cannot possibly take locks. Therefore, maintain a
3171 * single linked list and use cmpxchg() to add entries lockless.
3174 static void perf_pending_event(struct perf_pending_entry
*entry
)
3176 struct perf_event
*event
= container_of(entry
,
3177 struct perf_event
, pending
);
3179 if (event
->pending_disable
) {
3180 event
->pending_disable
= 0;
3181 __perf_event_disable(event
);
3184 if (event
->pending_wakeup
) {
3185 event
->pending_wakeup
= 0;
3186 perf_event_wakeup(event
);
3190 #define PENDING_TAIL ((struct perf_pending_entry *)-1UL)
3192 static DEFINE_PER_CPU(struct perf_pending_entry
*, perf_pending_head
) = {
3196 static void perf_pending_queue(struct perf_pending_entry
*entry
,
3197 void (*func
)(struct perf_pending_entry
*))
3199 struct perf_pending_entry
**head
;
3201 if (cmpxchg(&entry
->next
, NULL
, PENDING_TAIL
) != NULL
)
3206 head
= &get_cpu_var(perf_pending_head
);
3209 entry
->next
= *head
;
3210 } while (cmpxchg(head
, entry
->next
, entry
) != entry
->next
);
3212 set_perf_event_pending();
3214 put_cpu_var(perf_pending_head
);
3217 static int __perf_pending_run(void)
3219 struct perf_pending_entry
*list
;
3222 list
= xchg(&__get_cpu_var(perf_pending_head
), PENDING_TAIL
);
3223 while (list
!= PENDING_TAIL
) {
3224 void (*func
)(struct perf_pending_entry
*);
3225 struct perf_pending_entry
*entry
= list
;
3232 * Ensure we observe the unqueue before we issue the wakeup,
3233 * so that we won't be waiting forever.
3234 * -- see perf_not_pending().
3245 static inline int perf_not_pending(struct perf_event
*event
)
3248 * If we flush on whatever cpu we run, there is a chance we don't
3252 __perf_pending_run();
3256 * Ensure we see the proper queue state before going to sleep
3257 * so that we do not miss the wakeup. -- see perf_pending_handle()
3260 return event
->pending
.next
== NULL
;
3263 static void perf_pending_sync(struct perf_event
*event
)
3265 wait_event(event
->waitq
, perf_not_pending(event
));
3268 void perf_event_do_pending(void)
3270 __perf_pending_run();
3274 * We assume there is only KVM supporting the callbacks.
3275 * Later on, we might change it to a list if there is
3276 * another virtualization implementation supporting the callbacks.
3278 struct perf_guest_info_callbacks
*perf_guest_cbs
;
3280 int perf_register_guest_info_callbacks(struct perf_guest_info_callbacks
*cbs
)
3282 perf_guest_cbs
= cbs
;
3285 EXPORT_SYMBOL_GPL(perf_register_guest_info_callbacks
);
3287 int perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks
*cbs
)
3289 perf_guest_cbs
= NULL
;
3292 EXPORT_SYMBOL_GPL(perf_unregister_guest_info_callbacks
);
3297 static bool perf_output_space(struct perf_buffer
*buffer
, unsigned long tail
,
3298 unsigned long offset
, unsigned long head
)
3302 if (!buffer
->writable
)
3305 mask
= perf_data_size(buffer
) - 1;
3307 offset
= (offset
- tail
) & mask
;
3308 head
= (head
- tail
) & mask
;
3310 if ((int)(head
- offset
) < 0)
3316 static void perf_output_wakeup(struct perf_output_handle
*handle
)
3318 atomic_set(&handle
->buffer
->poll
, POLL_IN
);
3321 handle
->event
->pending_wakeup
= 1;
3322 perf_pending_queue(&handle
->event
->pending
,
3323 perf_pending_event
);
3325 perf_event_wakeup(handle
->event
);
3329 * We need to ensure a later event_id doesn't publish a head when a former
3330 * event isn't done writing. However since we need to deal with NMIs we
3331 * cannot fully serialize things.
3333 * We only publish the head (and generate a wakeup) when the outer-most
3336 static void perf_output_get_handle(struct perf_output_handle
*handle
)
3338 struct perf_buffer
*buffer
= handle
->buffer
;
3341 local_inc(&buffer
->nest
);
3342 handle
->wakeup
= local_read(&buffer
->wakeup
);
3345 static void perf_output_put_handle(struct perf_output_handle
*handle
)
3347 struct perf_buffer
*buffer
= handle
->buffer
;
3351 head
= local_read(&buffer
->head
);
3354 * IRQ/NMI can happen here, which means we can miss a head update.
3357 if (!local_dec_and_test(&buffer
->nest
))
3361 * Publish the known good head. Rely on the full barrier implied
3362 * by atomic_dec_and_test() order the buffer->head read and this
3365 buffer
->user_page
->data_head
= head
;
3368 * Now check if we missed an update, rely on the (compiler)
3369 * barrier in atomic_dec_and_test() to re-read buffer->head.
3371 if (unlikely(head
!= local_read(&buffer
->head
))) {
3372 local_inc(&buffer
->nest
);
3376 if (handle
->wakeup
!= local_read(&buffer
->wakeup
))
3377 perf_output_wakeup(handle
);
3383 __always_inline
void perf_output_copy(struct perf_output_handle
*handle
,
3384 const void *buf
, unsigned int len
)
3387 unsigned long size
= min_t(unsigned long, handle
->size
, len
);
3389 memcpy(handle
->addr
, buf
, size
);
3392 handle
->addr
+= size
;
3394 handle
->size
-= size
;
3395 if (!handle
->size
) {
3396 struct perf_buffer
*buffer
= handle
->buffer
;
3399 handle
->page
&= buffer
->nr_pages
- 1;
3400 handle
->addr
= buffer
->data_pages
[handle
->page
];
3401 handle
->size
= PAGE_SIZE
<< page_order(buffer
);
3406 int perf_output_begin(struct perf_output_handle
*handle
,
3407 struct perf_event
*event
, unsigned int size
,
3408 int nmi
, int sample
)
3410 struct perf_buffer
*buffer
;
3411 unsigned long tail
, offset
, head
;
3414 struct perf_event_header header
;
3421 * For inherited events we send all the output towards the parent.
3424 event
= event
->parent
;
3426 buffer
= rcu_dereference(event
->buffer
);
3430 handle
->buffer
= buffer
;
3431 handle
->event
= event
;
3433 handle
->sample
= sample
;
3435 if (!buffer
->nr_pages
)
3438 have_lost
= local_read(&buffer
->lost
);
3440 size
+= sizeof(lost_event
);
3442 perf_output_get_handle(handle
);
3446 * Userspace could choose to issue a mb() before updating the
3447 * tail pointer. So that all reads will be completed before the
3450 tail
= ACCESS_ONCE(buffer
->user_page
->data_tail
);
3452 offset
= head
= local_read(&buffer
->head
);
3454 if (unlikely(!perf_output_space(buffer
, tail
, offset
, head
)))
3456 } while (local_cmpxchg(&buffer
->head
, offset
, head
) != offset
);
3458 if (head
- local_read(&buffer
->wakeup
) > buffer
->watermark
)
3459 local_add(buffer
->watermark
, &buffer
->wakeup
);
3461 handle
->page
= offset
>> (PAGE_SHIFT
+ page_order(buffer
));
3462 handle
->page
&= buffer
->nr_pages
- 1;
3463 handle
->size
= offset
& ((PAGE_SIZE
<< page_order(buffer
)) - 1);
3464 handle
->addr
= buffer
->data_pages
[handle
->page
];
3465 handle
->addr
+= handle
->size
;
3466 handle
->size
= (PAGE_SIZE
<< page_order(buffer
)) - handle
->size
;
3469 lost_event
.header
.type
= PERF_RECORD_LOST
;
3470 lost_event
.header
.misc
= 0;
3471 lost_event
.header
.size
= sizeof(lost_event
);
3472 lost_event
.id
= event
->id
;
3473 lost_event
.lost
= local_xchg(&buffer
->lost
, 0);
3475 perf_output_put(handle
, lost_event
);
3481 local_inc(&buffer
->lost
);
3482 perf_output_put_handle(handle
);
3489 void perf_output_end(struct perf_output_handle
*handle
)
3491 struct perf_event
*event
= handle
->event
;
3492 struct perf_buffer
*buffer
= handle
->buffer
;
3494 int wakeup_events
= event
->attr
.wakeup_events
;
3496 if (handle
->sample
&& wakeup_events
) {
3497 int events
= local_inc_return(&buffer
->events
);
3498 if (events
>= wakeup_events
) {
3499 local_sub(wakeup_events
, &buffer
->events
);
3500 local_inc(&buffer
->wakeup
);
3504 perf_output_put_handle(handle
);
3508 static u32
perf_event_pid(struct perf_event
*event
, struct task_struct
*p
)
3511 * only top level events have the pid namespace they were created in
3514 event
= event
->parent
;
3516 return task_tgid_nr_ns(p
, event
->ns
);
3519 static u32
perf_event_tid(struct perf_event
*event
, struct task_struct
*p
)
3522 * only top level events have the pid namespace they were created in
3525 event
= event
->parent
;
3527 return task_pid_nr_ns(p
, event
->ns
);
3530 static void perf_output_read_one(struct perf_output_handle
*handle
,
3531 struct perf_event
*event
)
3533 u64 read_format
= event
->attr
.read_format
;
3537 values
[n
++] = perf_event_count(event
);
3538 if (read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
) {
3539 values
[n
++] = event
->total_time_enabled
+
3540 atomic64_read(&event
->child_total_time_enabled
);
3542 if (read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
) {
3543 values
[n
++] = event
->total_time_running
+
3544 atomic64_read(&event
->child_total_time_running
);
3546 if (read_format
& PERF_FORMAT_ID
)
3547 values
[n
++] = primary_event_id(event
);
3549 perf_output_copy(handle
, values
, n
* sizeof(u64
));
3553 * XXX PERF_FORMAT_GROUP vs inherited events seems difficult.
3555 static void perf_output_read_group(struct perf_output_handle
*handle
,
3556 struct perf_event
*event
)
3558 struct perf_event
*leader
= event
->group_leader
, *sub
;
3559 u64 read_format
= event
->attr
.read_format
;
3563 values
[n
++] = 1 + leader
->nr_siblings
;
3565 if (read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
)
3566 values
[n
++] = leader
->total_time_enabled
;
3568 if (read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
)
3569 values
[n
++] = leader
->total_time_running
;
3571 if (leader
!= event
)
3572 leader
->pmu
->read(leader
);
3574 values
[n
++] = perf_event_count(leader
);
3575 if (read_format
& PERF_FORMAT_ID
)
3576 values
[n
++] = primary_event_id(leader
);
3578 perf_output_copy(handle
, values
, n
* sizeof(u64
));
3580 list_for_each_entry(sub
, &leader
->sibling_list
, group_entry
) {
3584 sub
->pmu
->read(sub
);
3586 values
[n
++] = perf_event_count(sub
);
3587 if (read_format
& PERF_FORMAT_ID
)
3588 values
[n
++] = primary_event_id(sub
);
3590 perf_output_copy(handle
, values
, n
* sizeof(u64
));
3594 static void perf_output_read(struct perf_output_handle
*handle
,
3595 struct perf_event
*event
)
3597 if (event
->attr
.read_format
& PERF_FORMAT_GROUP
)
3598 perf_output_read_group(handle
, event
);
3600 perf_output_read_one(handle
, event
);
3603 void perf_output_sample(struct perf_output_handle
*handle
,
3604 struct perf_event_header
*header
,
3605 struct perf_sample_data
*data
,
3606 struct perf_event
*event
)
3608 u64 sample_type
= data
->type
;
3610 perf_output_put(handle
, *header
);
3612 if (sample_type
& PERF_SAMPLE_IP
)
3613 perf_output_put(handle
, data
->ip
);
3615 if (sample_type
& PERF_SAMPLE_TID
)
3616 perf_output_put(handle
, data
->tid_entry
);
3618 if (sample_type
& PERF_SAMPLE_TIME
)
3619 perf_output_put(handle
, data
->time
);
3621 if (sample_type
& PERF_SAMPLE_ADDR
)
3622 perf_output_put(handle
, data
->addr
);
3624 if (sample_type
& PERF_SAMPLE_ID
)
3625 perf_output_put(handle
, data
->id
);
3627 if (sample_type
& PERF_SAMPLE_STREAM_ID
)
3628 perf_output_put(handle
, data
->stream_id
);
3630 if (sample_type
& PERF_SAMPLE_CPU
)
3631 perf_output_put(handle
, data
->cpu_entry
);
3633 if (sample_type
& PERF_SAMPLE_PERIOD
)
3634 perf_output_put(handle
, data
->period
);
3636 if (sample_type
& PERF_SAMPLE_READ
)
3637 perf_output_read(handle
, event
);
3639 if (sample_type
& PERF_SAMPLE_CALLCHAIN
) {
3640 if (data
->callchain
) {
3643 if (data
->callchain
)
3644 size
+= data
->callchain
->nr
;
3646 size
*= sizeof(u64
);
3648 perf_output_copy(handle
, data
->callchain
, size
);
3651 perf_output_put(handle
, nr
);
3655 if (sample_type
& PERF_SAMPLE_RAW
) {
3657 perf_output_put(handle
, data
->raw
->size
);
3658 perf_output_copy(handle
, data
->raw
->data
,
3665 .size
= sizeof(u32
),
3668 perf_output_put(handle
, raw
);
3673 void perf_prepare_sample(struct perf_event_header
*header
,
3674 struct perf_sample_data
*data
,
3675 struct perf_event
*event
,
3676 struct pt_regs
*regs
)
3678 u64 sample_type
= event
->attr
.sample_type
;
3680 data
->type
= sample_type
;
3682 header
->type
= PERF_RECORD_SAMPLE
;
3683 header
->size
= sizeof(*header
);
3686 header
->misc
|= perf_misc_flags(regs
);
3688 if (sample_type
& PERF_SAMPLE_IP
) {
3689 data
->ip
= perf_instruction_pointer(regs
);
3691 header
->size
+= sizeof(data
->ip
);
3694 if (sample_type
& PERF_SAMPLE_TID
) {
3695 /* namespace issues */
3696 data
->tid_entry
.pid
= perf_event_pid(event
, current
);
3697 data
->tid_entry
.tid
= perf_event_tid(event
, current
);
3699 header
->size
+= sizeof(data
->tid_entry
);
3702 if (sample_type
& PERF_SAMPLE_TIME
) {
3703 data
->time
= perf_clock();
3705 header
->size
+= sizeof(data
->time
);
3708 if (sample_type
& PERF_SAMPLE_ADDR
)
3709 header
->size
+= sizeof(data
->addr
);
3711 if (sample_type
& PERF_SAMPLE_ID
) {
3712 data
->id
= primary_event_id(event
);
3714 header
->size
+= sizeof(data
->id
);
3717 if (sample_type
& PERF_SAMPLE_STREAM_ID
) {
3718 data
->stream_id
= event
->id
;
3720 header
->size
+= sizeof(data
->stream_id
);
3723 if (sample_type
& PERF_SAMPLE_CPU
) {
3724 data
->cpu_entry
.cpu
= raw_smp_processor_id();
3725 data
->cpu_entry
.reserved
= 0;
3727 header
->size
+= sizeof(data
->cpu_entry
);
3730 if (sample_type
& PERF_SAMPLE_PERIOD
)
3731 header
->size
+= sizeof(data
->period
);
3733 if (sample_type
& PERF_SAMPLE_READ
)
3734 header
->size
+= perf_event_read_size(event
);
3736 if (sample_type
& PERF_SAMPLE_CALLCHAIN
) {
3739 data
->callchain
= perf_callchain(regs
);
3741 if (data
->callchain
)
3742 size
+= data
->callchain
->nr
;
3744 header
->size
+= size
* sizeof(u64
);
3747 if (sample_type
& PERF_SAMPLE_RAW
) {
3748 int size
= sizeof(u32
);
3751 size
+= data
->raw
->size
;
3753 size
+= sizeof(u32
);
3755 WARN_ON_ONCE(size
& (sizeof(u64
)-1));
3756 header
->size
+= size
;
3760 static void perf_event_output(struct perf_event
*event
, int nmi
,
3761 struct perf_sample_data
*data
,
3762 struct pt_regs
*regs
)
3764 struct perf_output_handle handle
;
3765 struct perf_event_header header
;
3767 /* protect the callchain buffers */
3770 perf_prepare_sample(&header
, data
, event
, regs
);
3772 if (perf_output_begin(&handle
, event
, header
.size
, nmi
, 1))
3775 perf_output_sample(&handle
, &header
, data
, event
);
3777 perf_output_end(&handle
);
3787 struct perf_read_event
{
3788 struct perf_event_header header
;
3795 perf_event_read_event(struct perf_event
*event
,
3796 struct task_struct
*task
)
3798 struct perf_output_handle handle
;
3799 struct perf_read_event read_event
= {
3801 .type
= PERF_RECORD_READ
,
3803 .size
= sizeof(read_event
) + perf_event_read_size(event
),
3805 .pid
= perf_event_pid(event
, task
),
3806 .tid
= perf_event_tid(event
, task
),
3810 ret
= perf_output_begin(&handle
, event
, read_event
.header
.size
, 0, 0);
3814 perf_output_put(&handle
, read_event
);
3815 perf_output_read(&handle
, event
);
3817 perf_output_end(&handle
);
3821 * task tracking -- fork/exit
3823 * enabled by: attr.comm | attr.mmap | attr.mmap_data | attr.task
3826 struct perf_task_event
{
3827 struct task_struct
*task
;
3828 struct perf_event_context
*task_ctx
;
3831 struct perf_event_header header
;
3841 static void perf_event_task_output(struct perf_event
*event
,
3842 struct perf_task_event
*task_event
)
3844 struct perf_output_handle handle
;
3845 struct task_struct
*task
= task_event
->task
;
3848 size
= task_event
->event_id
.header
.size
;
3849 ret
= perf_output_begin(&handle
, event
, size
, 0, 0);
3854 task_event
->event_id
.pid
= perf_event_pid(event
, task
);
3855 task_event
->event_id
.ppid
= perf_event_pid(event
, current
);
3857 task_event
->event_id
.tid
= perf_event_tid(event
, task
);
3858 task_event
->event_id
.ptid
= perf_event_tid(event
, current
);
3860 perf_output_put(&handle
, task_event
->event_id
);
3862 perf_output_end(&handle
);
3865 static int perf_event_task_match(struct perf_event
*event
)
3867 if (event
->state
< PERF_EVENT_STATE_INACTIVE
)
3870 if (event
->cpu
!= -1 && event
->cpu
!= smp_processor_id())
3873 if (event
->attr
.comm
|| event
->attr
.mmap
||
3874 event
->attr
.mmap_data
|| event
->attr
.task
)
3880 static void perf_event_task_ctx(struct perf_event_context
*ctx
,
3881 struct perf_task_event
*task_event
)
3883 struct perf_event
*event
;
3885 list_for_each_entry_rcu(event
, &ctx
->event_list
, event_entry
) {
3886 if (perf_event_task_match(event
))
3887 perf_event_task_output(event
, task_event
);
3891 static void perf_event_task_event(struct perf_task_event
*task_event
)
3893 struct perf_cpu_context
*cpuctx
;
3894 struct perf_event_context
*ctx
;
3899 list_for_each_entry_rcu(pmu
, &pmus
, entry
) {
3900 cpuctx
= get_cpu_ptr(pmu
->pmu_cpu_context
);
3901 perf_event_task_ctx(&cpuctx
->ctx
, task_event
);
3903 ctx
= task_event
->task_ctx
;
3905 ctxn
= pmu
->task_ctx_nr
;
3908 ctx
= rcu_dereference(current
->perf_event_ctxp
[ctxn
]);
3911 perf_event_task_ctx(ctx
, task_event
);
3913 put_cpu_ptr(pmu
->pmu_cpu_context
);
3918 static void perf_event_task(struct task_struct
*task
,
3919 struct perf_event_context
*task_ctx
,
3922 struct perf_task_event task_event
;
3924 if (!atomic_read(&nr_comm_events
) &&
3925 !atomic_read(&nr_mmap_events
) &&
3926 !atomic_read(&nr_task_events
))
3929 task_event
= (struct perf_task_event
){
3931 .task_ctx
= task_ctx
,
3934 .type
= new ? PERF_RECORD_FORK
: PERF_RECORD_EXIT
,
3936 .size
= sizeof(task_event
.event_id
),
3942 .time
= perf_clock(),
3946 perf_event_task_event(&task_event
);
3949 void perf_event_fork(struct task_struct
*task
)
3951 perf_event_task(task
, NULL
, 1);
3958 struct perf_comm_event
{
3959 struct task_struct
*task
;
3964 struct perf_event_header header
;
3971 static void perf_event_comm_output(struct perf_event
*event
,
3972 struct perf_comm_event
*comm_event
)
3974 struct perf_output_handle handle
;
3975 int size
= comm_event
->event_id
.header
.size
;
3976 int ret
= perf_output_begin(&handle
, event
, size
, 0, 0);
3981 comm_event
->event_id
.pid
= perf_event_pid(event
, comm_event
->task
);
3982 comm_event
->event_id
.tid
= perf_event_tid(event
, comm_event
->task
);
3984 perf_output_put(&handle
, comm_event
->event_id
);
3985 perf_output_copy(&handle
, comm_event
->comm
,
3986 comm_event
->comm_size
);
3987 perf_output_end(&handle
);
3990 static int perf_event_comm_match(struct perf_event
*event
)
3992 if (event
->state
< PERF_EVENT_STATE_INACTIVE
)
3995 if (event
->cpu
!= -1 && event
->cpu
!= smp_processor_id())
3998 if (event
->attr
.comm
)
4004 static void perf_event_comm_ctx(struct perf_event_context
*ctx
,
4005 struct perf_comm_event
*comm_event
)
4007 struct perf_event
*event
;
4009 list_for_each_entry_rcu(event
, &ctx
->event_list
, event_entry
) {
4010 if (perf_event_comm_match(event
))
4011 perf_event_comm_output(event
, comm_event
);
4015 static void perf_event_comm_event(struct perf_comm_event
*comm_event
)
4017 struct perf_cpu_context
*cpuctx
;
4018 struct perf_event_context
*ctx
;
4019 char comm
[TASK_COMM_LEN
];
4024 memset(comm
, 0, sizeof(comm
));
4025 strlcpy(comm
, comm_event
->task
->comm
, sizeof(comm
));
4026 size
= ALIGN(strlen(comm
)+1, sizeof(u64
));
4028 comm_event
->comm
= comm
;
4029 comm_event
->comm_size
= size
;
4031 comm_event
->event_id
.header
.size
= sizeof(comm_event
->event_id
) + size
;
4034 list_for_each_entry_rcu(pmu
, &pmus
, entry
) {
4035 cpuctx
= get_cpu_ptr(pmu
->pmu_cpu_context
);
4036 perf_event_comm_ctx(&cpuctx
->ctx
, comm_event
);
4038 ctxn
= pmu
->task_ctx_nr
;
4042 ctx
= rcu_dereference(current
->perf_event_ctxp
[ctxn
]);
4044 perf_event_comm_ctx(ctx
, comm_event
);
4046 put_cpu_ptr(pmu
->pmu_cpu_context
);
4051 void perf_event_comm(struct task_struct
*task
)
4053 struct perf_comm_event comm_event
;
4054 struct perf_event_context
*ctx
;
4057 for_each_task_context_nr(ctxn
) {
4058 ctx
= task
->perf_event_ctxp
[ctxn
];
4062 perf_event_enable_on_exec(ctx
);
4065 if (!atomic_read(&nr_comm_events
))
4068 comm_event
= (struct perf_comm_event
){
4074 .type
= PERF_RECORD_COMM
,
4083 perf_event_comm_event(&comm_event
);
4090 struct perf_mmap_event
{
4091 struct vm_area_struct
*vma
;
4093 const char *file_name
;
4097 struct perf_event_header header
;
4107 static void perf_event_mmap_output(struct perf_event
*event
,
4108 struct perf_mmap_event
*mmap_event
)
4110 struct perf_output_handle handle
;
4111 int size
= mmap_event
->event_id
.header
.size
;
4112 int ret
= perf_output_begin(&handle
, event
, size
, 0, 0);
4117 mmap_event
->event_id
.pid
= perf_event_pid(event
, current
);
4118 mmap_event
->event_id
.tid
= perf_event_tid(event
, current
);
4120 perf_output_put(&handle
, mmap_event
->event_id
);
4121 perf_output_copy(&handle
, mmap_event
->file_name
,
4122 mmap_event
->file_size
);
4123 perf_output_end(&handle
);
4126 static int perf_event_mmap_match(struct perf_event
*event
,
4127 struct perf_mmap_event
*mmap_event
,
4130 if (event
->state
< PERF_EVENT_STATE_INACTIVE
)
4133 if (event
->cpu
!= -1 && event
->cpu
!= smp_processor_id())
4136 if ((!executable
&& event
->attr
.mmap_data
) ||
4137 (executable
&& event
->attr
.mmap
))
4143 static void perf_event_mmap_ctx(struct perf_event_context
*ctx
,
4144 struct perf_mmap_event
*mmap_event
,
4147 struct perf_event
*event
;
4149 list_for_each_entry_rcu(event
, &ctx
->event_list
, event_entry
) {
4150 if (perf_event_mmap_match(event
, mmap_event
, executable
))
4151 perf_event_mmap_output(event
, mmap_event
);
4155 static void perf_event_mmap_event(struct perf_mmap_event
*mmap_event
)
4157 struct perf_cpu_context
*cpuctx
;
4158 struct perf_event_context
*ctx
;
4159 struct vm_area_struct
*vma
= mmap_event
->vma
;
4160 struct file
*file
= vma
->vm_file
;
4168 memset(tmp
, 0, sizeof(tmp
));
4172 * d_path works from the end of the buffer backwards, so we
4173 * need to add enough zero bytes after the string to handle
4174 * the 64bit alignment we do later.
4176 buf
= kzalloc(PATH_MAX
+ sizeof(u64
), GFP_KERNEL
);
4178 name
= strncpy(tmp
, "//enomem", sizeof(tmp
));
4181 name
= d_path(&file
->f_path
, buf
, PATH_MAX
);
4183 name
= strncpy(tmp
, "//toolong", sizeof(tmp
));
4187 if (arch_vma_name(mmap_event
->vma
)) {
4188 name
= strncpy(tmp
, arch_vma_name(mmap_event
->vma
),
4194 name
= strncpy(tmp
, "[vdso]", sizeof(tmp
));
4196 } else if (vma
->vm_start
<= vma
->vm_mm
->start_brk
&&
4197 vma
->vm_end
>= vma
->vm_mm
->brk
) {
4198 name
= strncpy(tmp
, "[heap]", sizeof(tmp
));
4200 } else if (vma
->vm_start
<= vma
->vm_mm
->start_stack
&&
4201 vma
->vm_end
>= vma
->vm_mm
->start_stack
) {
4202 name
= strncpy(tmp
, "[stack]", sizeof(tmp
));
4206 name
= strncpy(tmp
, "//anon", sizeof(tmp
));
4211 size
= ALIGN(strlen(name
)+1, sizeof(u64
));
4213 mmap_event
->file_name
= name
;
4214 mmap_event
->file_size
= size
;
4216 mmap_event
->event_id
.header
.size
= sizeof(mmap_event
->event_id
) + size
;
4219 list_for_each_entry_rcu(pmu
, &pmus
, entry
) {
4220 cpuctx
= get_cpu_ptr(pmu
->pmu_cpu_context
);
4221 perf_event_mmap_ctx(&cpuctx
->ctx
, mmap_event
,
4222 vma
->vm_flags
& VM_EXEC
);
4224 ctxn
= pmu
->task_ctx_nr
;
4228 ctx
= rcu_dereference(current
->perf_event_ctxp
[ctxn
]);
4230 perf_event_mmap_ctx(ctx
, mmap_event
,
4231 vma
->vm_flags
& VM_EXEC
);
4234 put_cpu_ptr(pmu
->pmu_cpu_context
);
4241 void perf_event_mmap(struct vm_area_struct
*vma
)
4243 struct perf_mmap_event mmap_event
;
4245 if (!atomic_read(&nr_mmap_events
))
4248 mmap_event
= (struct perf_mmap_event
){
4254 .type
= PERF_RECORD_MMAP
,
4255 .misc
= PERF_RECORD_MISC_USER
,
4260 .start
= vma
->vm_start
,
4261 .len
= vma
->vm_end
- vma
->vm_start
,
4262 .pgoff
= (u64
)vma
->vm_pgoff
<< PAGE_SHIFT
,
4266 perf_event_mmap_event(&mmap_event
);
4270 * IRQ throttle logging
4273 static void perf_log_throttle(struct perf_event
*event
, int enable
)
4275 struct perf_output_handle handle
;
4279 struct perf_event_header header
;
4283 } throttle_event
= {
4285 .type
= PERF_RECORD_THROTTLE
,
4287 .size
= sizeof(throttle_event
),
4289 .time
= perf_clock(),
4290 .id
= primary_event_id(event
),
4291 .stream_id
= event
->id
,
4295 throttle_event
.header
.type
= PERF_RECORD_UNTHROTTLE
;
4297 ret
= perf_output_begin(&handle
, event
, sizeof(throttle_event
), 1, 0);
4301 perf_output_put(&handle
, throttle_event
);
4302 perf_output_end(&handle
);
4306 * Generic event overflow handling, sampling.
4309 static int __perf_event_overflow(struct perf_event
*event
, int nmi
,
4310 int throttle
, struct perf_sample_data
*data
,
4311 struct pt_regs
*regs
)
4313 int events
= atomic_read(&event
->event_limit
);
4314 struct hw_perf_event
*hwc
= &event
->hw
;
4320 if (hwc
->interrupts
!= MAX_INTERRUPTS
) {
4322 if (HZ
* hwc
->interrupts
>
4323 (u64
)sysctl_perf_event_sample_rate
) {
4324 hwc
->interrupts
= MAX_INTERRUPTS
;
4325 perf_log_throttle(event
, 0);
4330 * Keep re-disabling events even though on the previous
4331 * pass we disabled it - just in case we raced with a
4332 * sched-in and the event got enabled again:
4338 if (event
->attr
.freq
) {
4339 u64 now
= perf_clock();
4340 s64 delta
= now
- hwc
->freq_time_stamp
;
4342 hwc
->freq_time_stamp
= now
;
4344 if (delta
> 0 && delta
< 2*TICK_NSEC
)
4345 perf_adjust_period(event
, delta
, hwc
->last_period
);
4349 * XXX event_limit might not quite work as expected on inherited
4353 event
->pending_kill
= POLL_IN
;
4354 if (events
&& atomic_dec_and_test(&event
->event_limit
)) {
4356 event
->pending_kill
= POLL_HUP
;
4358 event
->pending_disable
= 1;
4359 perf_pending_queue(&event
->pending
,
4360 perf_pending_event
);
4362 perf_event_disable(event
);
4365 if (event
->overflow_handler
)
4366 event
->overflow_handler(event
, nmi
, data
, regs
);
4368 perf_event_output(event
, nmi
, data
, regs
);
4373 int perf_event_overflow(struct perf_event
*event
, int nmi
,
4374 struct perf_sample_data
*data
,
4375 struct pt_regs
*regs
)
4377 return __perf_event_overflow(event
, nmi
, 1, data
, regs
);
4381 * Generic software event infrastructure
4384 struct swevent_htable
{
4385 struct swevent_hlist
*swevent_hlist
;
4386 struct mutex hlist_mutex
;
4389 /* Recursion avoidance in each contexts */
4390 int recursion
[PERF_NR_CONTEXTS
];
4393 static DEFINE_PER_CPU(struct swevent_htable
, swevent_htable
);
4396 * We directly increment event->count and keep a second value in
4397 * event->hw.period_left to count intervals. This period event
4398 * is kept in the range [-sample_period, 0] so that we can use the
4402 static u64
perf_swevent_set_period(struct perf_event
*event
)
4404 struct hw_perf_event
*hwc
= &event
->hw
;
4405 u64 period
= hwc
->last_period
;
4409 hwc
->last_period
= hwc
->sample_period
;
4412 old
= val
= local64_read(&hwc
->period_left
);
4416 nr
= div64_u64(period
+ val
, period
);
4417 offset
= nr
* period
;
4419 if (local64_cmpxchg(&hwc
->period_left
, old
, val
) != old
)
4425 static void perf_swevent_overflow(struct perf_event
*event
, u64 overflow
,
4426 int nmi
, struct perf_sample_data
*data
,
4427 struct pt_regs
*regs
)
4429 struct hw_perf_event
*hwc
= &event
->hw
;
4432 data
->period
= event
->hw
.last_period
;
4434 overflow
= perf_swevent_set_period(event
);
4436 if (hwc
->interrupts
== MAX_INTERRUPTS
)
4439 for (; overflow
; overflow
--) {
4440 if (__perf_event_overflow(event
, nmi
, throttle
,
4443 * We inhibit the overflow from happening when
4444 * hwc->interrupts == MAX_INTERRUPTS.
4452 static void perf_swevent_event(struct perf_event
*event
, u64 nr
,
4453 int nmi
, struct perf_sample_data
*data
,
4454 struct pt_regs
*regs
)
4456 struct hw_perf_event
*hwc
= &event
->hw
;
4458 local64_add(nr
, &event
->count
);
4463 if (!hwc
->sample_period
)
4466 if (nr
== 1 && hwc
->sample_period
== 1 && !event
->attr
.freq
)
4467 return perf_swevent_overflow(event
, 1, nmi
, data
, regs
);
4469 if (local64_add_negative(nr
, &hwc
->period_left
))
4472 perf_swevent_overflow(event
, 0, nmi
, data
, regs
);
4475 static int perf_exclude_event(struct perf_event
*event
,
4476 struct pt_regs
*regs
)
4478 if (event
->hw
.state
& PERF_HES_STOPPED
)
4482 if (event
->attr
.exclude_user
&& user_mode(regs
))
4485 if (event
->attr
.exclude_kernel
&& !user_mode(regs
))
4492 static int perf_swevent_match(struct perf_event
*event
,
4493 enum perf_type_id type
,
4495 struct perf_sample_data
*data
,
4496 struct pt_regs
*regs
)
4498 if (event
->attr
.type
!= type
)
4501 if (event
->attr
.config
!= event_id
)
4504 if (perf_exclude_event(event
, regs
))
4510 static inline u64
swevent_hash(u64 type
, u32 event_id
)
4512 u64 val
= event_id
| (type
<< 32);
4514 return hash_64(val
, SWEVENT_HLIST_BITS
);
4517 static inline struct hlist_head
*
4518 __find_swevent_head(struct swevent_hlist
*hlist
, u64 type
, u32 event_id
)
4520 u64 hash
= swevent_hash(type
, event_id
);
4522 return &hlist
->heads
[hash
];
4525 /* For the read side: events when they trigger */
4526 static inline struct hlist_head
*
4527 find_swevent_head_rcu(struct swevent_htable
*swhash
, u64 type
, u32 event_id
)
4529 struct swevent_hlist
*hlist
;
4531 hlist
= rcu_dereference(swhash
->swevent_hlist
);
4535 return __find_swevent_head(hlist
, type
, event_id
);
4538 /* For the event head insertion and removal in the hlist */
4539 static inline struct hlist_head
*
4540 find_swevent_head(struct swevent_htable
*swhash
, struct perf_event
*event
)
4542 struct swevent_hlist
*hlist
;
4543 u32 event_id
= event
->attr
.config
;
4544 u64 type
= event
->attr
.type
;
4547 * Event scheduling is always serialized against hlist allocation
4548 * and release. Which makes the protected version suitable here.
4549 * The context lock guarantees that.
4551 hlist
= rcu_dereference_protected(swhash
->swevent_hlist
,
4552 lockdep_is_held(&event
->ctx
->lock
));
4556 return __find_swevent_head(hlist
, type
, event_id
);
4559 static void do_perf_sw_event(enum perf_type_id type
, u32 event_id
,
4561 struct perf_sample_data
*data
,
4562 struct pt_regs
*regs
)
4564 struct swevent_htable
*swhash
= &__get_cpu_var(swevent_htable
);
4565 struct perf_event
*event
;
4566 struct hlist_node
*node
;
4567 struct hlist_head
*head
;
4570 head
= find_swevent_head_rcu(swhash
, type
, event_id
);
4574 hlist_for_each_entry_rcu(event
, node
, head
, hlist_entry
) {
4575 if (perf_swevent_match(event
, type
, event_id
, data
, regs
))
4576 perf_swevent_event(event
, nr
, nmi
, data
, regs
);
4582 int perf_swevent_get_recursion_context(void)
4584 struct swevent_htable
*swhash
= &__get_cpu_var(swevent_htable
);
4586 return get_recursion_context(swhash
->recursion
);
4588 EXPORT_SYMBOL_GPL(perf_swevent_get_recursion_context
);
4590 void inline perf_swevent_put_recursion_context(int rctx
)
4592 struct swevent_htable
*swhash
= &__get_cpu_var(swevent_htable
);
4594 put_recursion_context(swhash
->recursion
, rctx
);
4597 void __perf_sw_event(u32 event_id
, u64 nr
, int nmi
,
4598 struct pt_regs
*regs
, u64 addr
)
4600 struct perf_sample_data data
;
4603 preempt_disable_notrace();
4604 rctx
= perf_swevent_get_recursion_context();
4608 perf_sample_data_init(&data
, addr
);
4610 do_perf_sw_event(PERF_TYPE_SOFTWARE
, event_id
, nr
, nmi
, &data
, regs
);
4612 perf_swevent_put_recursion_context(rctx
);
4613 preempt_enable_notrace();
4616 static void perf_swevent_read(struct perf_event
*event
)
4620 static int perf_swevent_add(struct perf_event
*event
, int flags
)
4622 struct swevent_htable
*swhash
= &__get_cpu_var(swevent_htable
);
4623 struct hw_perf_event
*hwc
= &event
->hw
;
4624 struct hlist_head
*head
;
4626 if (hwc
->sample_period
) {
4627 hwc
->last_period
= hwc
->sample_period
;
4628 perf_swevent_set_period(event
);
4631 hwc
->state
= !(flags
& PERF_EF_START
);
4633 head
= find_swevent_head(swhash
, event
);
4634 if (WARN_ON_ONCE(!head
))
4637 hlist_add_head_rcu(&event
->hlist_entry
, head
);
4642 static void perf_swevent_del(struct perf_event
*event
, int flags
)
4644 hlist_del_rcu(&event
->hlist_entry
);
4647 static void perf_swevent_start(struct perf_event
*event
, int flags
)
4649 event
->hw
.state
= 0;
4652 static void perf_swevent_stop(struct perf_event
*event
, int flags
)
4654 event
->hw
.state
= PERF_HES_STOPPED
;
4657 /* Deref the hlist from the update side */
4658 static inline struct swevent_hlist
*
4659 swevent_hlist_deref(struct swevent_htable
*swhash
)
4661 return rcu_dereference_protected(swhash
->swevent_hlist
,
4662 lockdep_is_held(&swhash
->hlist_mutex
));
4665 static void swevent_hlist_release_rcu(struct rcu_head
*rcu_head
)
4667 struct swevent_hlist
*hlist
;
4669 hlist
= container_of(rcu_head
, struct swevent_hlist
, rcu_head
);
4673 static void swevent_hlist_release(struct swevent_htable
*swhash
)
4675 struct swevent_hlist
*hlist
= swevent_hlist_deref(swhash
);
4680 rcu_assign_pointer(swhash
->swevent_hlist
, NULL
);
4681 call_rcu(&hlist
->rcu_head
, swevent_hlist_release_rcu
);
4684 static void swevent_hlist_put_cpu(struct perf_event
*event
, int cpu
)
4686 struct swevent_htable
*swhash
= &per_cpu(swevent_htable
, cpu
);
4688 mutex_lock(&swhash
->hlist_mutex
);
4690 if (!--swhash
->hlist_refcount
)
4691 swevent_hlist_release(swhash
);
4693 mutex_unlock(&swhash
->hlist_mutex
);
4696 static void swevent_hlist_put(struct perf_event
*event
)
4700 if (event
->cpu
!= -1) {
4701 swevent_hlist_put_cpu(event
, event
->cpu
);
4705 for_each_possible_cpu(cpu
)
4706 swevent_hlist_put_cpu(event
, cpu
);
4709 static int swevent_hlist_get_cpu(struct perf_event
*event
, int cpu
)
4711 struct swevent_htable
*swhash
= &per_cpu(swevent_htable
, cpu
);
4714 mutex_lock(&swhash
->hlist_mutex
);
4716 if (!swevent_hlist_deref(swhash
) && cpu_online(cpu
)) {
4717 struct swevent_hlist
*hlist
;
4719 hlist
= kzalloc(sizeof(*hlist
), GFP_KERNEL
);
4724 rcu_assign_pointer(swhash
->swevent_hlist
, hlist
);
4726 swhash
->hlist_refcount
++;
4728 mutex_unlock(&swhash
->hlist_mutex
);
4733 static int swevent_hlist_get(struct perf_event
*event
)
4736 int cpu
, failed_cpu
;
4738 if (event
->cpu
!= -1)
4739 return swevent_hlist_get_cpu(event
, event
->cpu
);
4742 for_each_possible_cpu(cpu
) {
4743 err
= swevent_hlist_get_cpu(event
, cpu
);
4753 for_each_possible_cpu(cpu
) {
4754 if (cpu
== failed_cpu
)
4756 swevent_hlist_put_cpu(event
, cpu
);
4763 atomic_t perf_swevent_enabled
[PERF_COUNT_SW_MAX
];
4765 static void sw_perf_event_destroy(struct perf_event
*event
)
4767 u64 event_id
= event
->attr
.config
;
4769 WARN_ON(event
->parent
);
4771 atomic_dec(&perf_swevent_enabled
[event_id
]);
4772 swevent_hlist_put(event
);
4775 static int perf_swevent_init(struct perf_event
*event
)
4777 int event_id
= event
->attr
.config
;
4779 if (event
->attr
.type
!= PERF_TYPE_SOFTWARE
)
4783 case PERF_COUNT_SW_CPU_CLOCK
:
4784 case PERF_COUNT_SW_TASK_CLOCK
:
4791 if (event_id
> PERF_COUNT_SW_MAX
)
4794 if (!event
->parent
) {
4797 err
= swevent_hlist_get(event
);
4801 atomic_inc(&perf_swevent_enabled
[event_id
]);
4802 event
->destroy
= sw_perf_event_destroy
;
4808 static struct pmu perf_swevent
= {
4809 .task_ctx_nr
= perf_sw_context
,
4811 .event_init
= perf_swevent_init
,
4812 .add
= perf_swevent_add
,
4813 .del
= perf_swevent_del
,
4814 .start
= perf_swevent_start
,
4815 .stop
= perf_swevent_stop
,
4816 .read
= perf_swevent_read
,
4819 #ifdef CONFIG_EVENT_TRACING
4821 static int perf_tp_filter_match(struct perf_event
*event
,
4822 struct perf_sample_data
*data
)
4824 void *record
= data
->raw
->data
;
4826 if (likely(!event
->filter
) || filter_match_preds(event
->filter
, record
))
4831 static int perf_tp_event_match(struct perf_event
*event
,
4832 struct perf_sample_data
*data
,
4833 struct pt_regs
*regs
)
4836 * All tracepoints are from kernel-space.
4838 if (event
->attr
.exclude_kernel
)
4841 if (!perf_tp_filter_match(event
, data
))
4847 void perf_tp_event(u64 addr
, u64 count
, void *record
, int entry_size
,
4848 struct pt_regs
*regs
, struct hlist_head
*head
, int rctx
)
4850 struct perf_sample_data data
;
4851 struct perf_event
*event
;
4852 struct hlist_node
*node
;
4854 struct perf_raw_record raw
= {
4859 perf_sample_data_init(&data
, addr
);
4862 hlist_for_each_entry_rcu(event
, node
, head
, hlist_entry
) {
4863 if (perf_tp_event_match(event
, &data
, regs
))
4864 perf_swevent_event(event
, count
, 1, &data
, regs
);
4867 perf_swevent_put_recursion_context(rctx
);
4869 EXPORT_SYMBOL_GPL(perf_tp_event
);
4871 static void tp_perf_event_destroy(struct perf_event
*event
)
4873 perf_trace_destroy(event
);
4876 static int perf_tp_event_init(struct perf_event
*event
)
4880 if (event
->attr
.type
!= PERF_TYPE_TRACEPOINT
)
4884 * Raw tracepoint data is a severe data leak, only allow root to
4887 if ((event
->attr
.sample_type
& PERF_SAMPLE_RAW
) &&
4888 perf_paranoid_tracepoint_raw() &&
4889 !capable(CAP_SYS_ADMIN
))
4892 err
= perf_trace_init(event
);
4896 event
->destroy
= tp_perf_event_destroy
;
4901 static struct pmu perf_tracepoint
= {
4902 .task_ctx_nr
= perf_sw_context
,
4904 .event_init
= perf_tp_event_init
,
4905 .add
= perf_trace_add
,
4906 .del
= perf_trace_del
,
4907 .start
= perf_swevent_start
,
4908 .stop
= perf_swevent_stop
,
4909 .read
= perf_swevent_read
,
4912 static inline void perf_tp_register(void)
4914 perf_pmu_register(&perf_tracepoint
);
4917 static int perf_event_set_filter(struct perf_event
*event
, void __user
*arg
)
4922 if (event
->attr
.type
!= PERF_TYPE_TRACEPOINT
)
4925 filter_str
= strndup_user(arg
, PAGE_SIZE
);
4926 if (IS_ERR(filter_str
))
4927 return PTR_ERR(filter_str
);
4929 ret
= ftrace_profile_set_filter(event
, event
->attr
.config
, filter_str
);
4935 static void perf_event_free_filter(struct perf_event
*event
)
4937 ftrace_profile_free_filter(event
);
4942 static inline void perf_tp_register(void)
4946 static int perf_event_set_filter(struct perf_event
*event
, void __user
*arg
)
4951 static void perf_event_free_filter(struct perf_event
*event
)
4955 #endif /* CONFIG_EVENT_TRACING */
4957 #ifdef CONFIG_HAVE_HW_BREAKPOINT
4958 void perf_bp_event(struct perf_event
*bp
, void *data
)
4960 struct perf_sample_data sample
;
4961 struct pt_regs
*regs
= data
;
4963 perf_sample_data_init(&sample
, bp
->attr
.bp_addr
);
4965 if (!bp
->hw
.state
&& !perf_exclude_event(bp
, regs
))
4966 perf_swevent_event(bp
, 1, 1, &sample
, regs
);
4971 * hrtimer based swevent callback
4974 static enum hrtimer_restart
perf_swevent_hrtimer(struct hrtimer
*hrtimer
)
4976 enum hrtimer_restart ret
= HRTIMER_RESTART
;
4977 struct perf_sample_data data
;
4978 struct pt_regs
*regs
;
4979 struct perf_event
*event
;
4982 event
= container_of(hrtimer
, struct perf_event
, hw
.hrtimer
);
4983 event
->pmu
->read(event
);
4985 perf_sample_data_init(&data
, 0);
4986 data
.period
= event
->hw
.last_period
;
4987 regs
= get_irq_regs();
4989 if (regs
&& !perf_exclude_event(event
, regs
)) {
4990 if (!(event
->attr
.exclude_idle
&& current
->pid
== 0))
4991 if (perf_event_overflow(event
, 0, &data
, regs
))
4992 ret
= HRTIMER_NORESTART
;
4995 period
= max_t(u64
, 10000, event
->hw
.sample_period
);
4996 hrtimer_forward_now(hrtimer
, ns_to_ktime(period
));
5001 static void perf_swevent_start_hrtimer(struct perf_event
*event
)
5003 struct hw_perf_event
*hwc
= &event
->hw
;
5005 hrtimer_init(&hwc
->hrtimer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
5006 hwc
->hrtimer
.function
= perf_swevent_hrtimer
;
5007 if (hwc
->sample_period
) {
5008 s64 period
= local64_read(&hwc
->period_left
);
5014 local64_set(&hwc
->period_left
, 0);
5016 period
= max_t(u64
, 10000, hwc
->sample_period
);
5018 __hrtimer_start_range_ns(&hwc
->hrtimer
,
5019 ns_to_ktime(period
), 0,
5020 HRTIMER_MODE_REL_PINNED
, 0);
5024 static void perf_swevent_cancel_hrtimer(struct perf_event
*event
)
5026 struct hw_perf_event
*hwc
= &event
->hw
;
5028 if (hwc
->sample_period
) {
5029 ktime_t remaining
= hrtimer_get_remaining(&hwc
->hrtimer
);
5030 local64_set(&hwc
->period_left
, ktime_to_ns(remaining
));
5032 hrtimer_cancel(&hwc
->hrtimer
);
5037 * Software event: cpu wall time clock
5040 static void cpu_clock_event_update(struct perf_event
*event
)
5045 now
= local_clock();
5046 prev
= local64_xchg(&event
->hw
.prev_count
, now
);
5047 local64_add(now
- prev
, &event
->count
);
5050 static void cpu_clock_event_start(struct perf_event
*event
, int flags
)
5052 local64_set(&event
->hw
.prev_count
, local_clock());
5053 perf_swevent_start_hrtimer(event
);
5056 static void cpu_clock_event_stop(struct perf_event
*event
, int flags
)
5058 perf_swevent_cancel_hrtimer(event
);
5059 cpu_clock_event_update(event
);
5062 static int cpu_clock_event_add(struct perf_event
*event
, int flags
)
5064 if (flags
& PERF_EF_START
)
5065 cpu_clock_event_start(event
, flags
);
5070 static void cpu_clock_event_del(struct perf_event
*event
, int flags
)
5072 cpu_clock_event_stop(event
, flags
);
5075 static void cpu_clock_event_read(struct perf_event
*event
)
5077 cpu_clock_event_update(event
);
5080 static int cpu_clock_event_init(struct perf_event
*event
)
5082 if (event
->attr
.type
!= PERF_TYPE_SOFTWARE
)
5085 if (event
->attr
.config
!= PERF_COUNT_SW_CPU_CLOCK
)
5091 static struct pmu perf_cpu_clock
= {
5092 .task_ctx_nr
= perf_sw_context
,
5094 .event_init
= cpu_clock_event_init
,
5095 .add
= cpu_clock_event_add
,
5096 .del
= cpu_clock_event_del
,
5097 .start
= cpu_clock_event_start
,
5098 .stop
= cpu_clock_event_stop
,
5099 .read
= cpu_clock_event_read
,
5103 * Software event: task time clock
5106 static void task_clock_event_update(struct perf_event
*event
, u64 now
)
5111 prev
= local64_xchg(&event
->hw
.prev_count
, now
);
5113 local64_add(delta
, &event
->count
);
5116 static void task_clock_event_start(struct perf_event
*event
, int flags
)
5118 local64_set(&event
->hw
.prev_count
, event
->ctx
->time
);
5119 perf_swevent_start_hrtimer(event
);
5122 static void task_clock_event_stop(struct perf_event
*event
, int flags
)
5124 perf_swevent_cancel_hrtimer(event
);
5125 task_clock_event_update(event
, event
->ctx
->time
);
5128 static int task_clock_event_add(struct perf_event
*event
, int flags
)
5130 if (flags
& PERF_EF_START
)
5131 task_clock_event_start(event
, flags
);
5136 static void task_clock_event_del(struct perf_event
*event
, int flags
)
5138 task_clock_event_stop(event
, PERF_EF_UPDATE
);
5141 static void task_clock_event_read(struct perf_event
*event
)
5146 update_context_time(event
->ctx
);
5147 time
= event
->ctx
->time
;
5149 u64 now
= perf_clock();
5150 u64 delta
= now
- event
->ctx
->timestamp
;
5151 time
= event
->ctx
->time
+ delta
;
5154 task_clock_event_update(event
, time
);
5157 static int task_clock_event_init(struct perf_event
*event
)
5159 if (event
->attr
.type
!= PERF_TYPE_SOFTWARE
)
5162 if (event
->attr
.config
!= PERF_COUNT_SW_TASK_CLOCK
)
5168 static struct pmu perf_task_clock
= {
5169 .task_ctx_nr
= perf_sw_context
,
5171 .event_init
= task_clock_event_init
,
5172 .add
= task_clock_event_add
,
5173 .del
= task_clock_event_del
,
5174 .start
= task_clock_event_start
,
5175 .stop
= task_clock_event_stop
,
5176 .read
= task_clock_event_read
,
5179 static void perf_pmu_nop_void(struct pmu
*pmu
)
5183 static int perf_pmu_nop_int(struct pmu
*pmu
)
5188 static void perf_pmu_start_txn(struct pmu
*pmu
)
5190 perf_pmu_disable(pmu
);
5193 static int perf_pmu_commit_txn(struct pmu
*pmu
)
5195 perf_pmu_enable(pmu
);
5199 static void perf_pmu_cancel_txn(struct pmu
*pmu
)
5201 perf_pmu_enable(pmu
);
5205 * Ensures all contexts with the same task_ctx_nr have the same
5206 * pmu_cpu_context too.
5208 static void *find_pmu_context(int ctxn
)
5215 list_for_each_entry(pmu
, &pmus
, entry
) {
5216 if (pmu
->task_ctx_nr
== ctxn
)
5217 return pmu
->pmu_cpu_context
;
5223 static void free_pmu_context(void * __percpu cpu_context
)
5227 mutex_lock(&pmus_lock
);
5229 * Like a real lame refcount.
5231 list_for_each_entry(pmu
, &pmus
, entry
) {
5232 if (pmu
->pmu_cpu_context
== cpu_context
)
5236 free_percpu(cpu_context
);
5238 mutex_unlock(&pmus_lock
);
5241 int perf_pmu_register(struct pmu
*pmu
)
5245 mutex_lock(&pmus_lock
);
5247 pmu
->pmu_disable_count
= alloc_percpu(int);
5248 if (!pmu
->pmu_disable_count
)
5251 pmu
->pmu_cpu_context
= find_pmu_context(pmu
->task_ctx_nr
);
5252 if (pmu
->pmu_cpu_context
)
5253 goto got_cpu_context
;
5255 pmu
->pmu_cpu_context
= alloc_percpu(struct perf_cpu_context
);
5256 if (!pmu
->pmu_cpu_context
)
5259 for_each_possible_cpu(cpu
) {
5260 struct perf_cpu_context
*cpuctx
;
5262 cpuctx
= per_cpu_ptr(pmu
->pmu_cpu_context
, cpu
);
5263 __perf_event_init_context(&cpuctx
->ctx
);
5264 cpuctx
->ctx
.type
= cpu_context
;
5265 cpuctx
->ctx
.pmu
= pmu
;
5266 cpuctx
->jiffies_interval
= 1;
5267 INIT_LIST_HEAD(&cpuctx
->rotation_list
);
5271 if (!pmu
->start_txn
) {
5272 if (pmu
->pmu_enable
) {
5274 * If we have pmu_enable/pmu_disable calls, install
5275 * transaction stubs that use that to try and batch
5276 * hardware accesses.
5278 pmu
->start_txn
= perf_pmu_start_txn
;
5279 pmu
->commit_txn
= perf_pmu_commit_txn
;
5280 pmu
->cancel_txn
= perf_pmu_cancel_txn
;
5282 pmu
->start_txn
= perf_pmu_nop_void
;
5283 pmu
->commit_txn
= perf_pmu_nop_int
;
5284 pmu
->cancel_txn
= perf_pmu_nop_void
;
5288 if (!pmu
->pmu_enable
) {
5289 pmu
->pmu_enable
= perf_pmu_nop_void
;
5290 pmu
->pmu_disable
= perf_pmu_nop_void
;
5293 list_add_rcu(&pmu
->entry
, &pmus
);
5296 mutex_unlock(&pmus_lock
);
5301 free_percpu(pmu
->pmu_disable_count
);
5305 void perf_pmu_unregister(struct pmu
*pmu
)
5307 mutex_lock(&pmus_lock
);
5308 list_del_rcu(&pmu
->entry
);
5309 mutex_unlock(&pmus_lock
);
5312 * We dereference the pmu list under both SRCU and regular RCU, so
5313 * synchronize against both of those.
5315 synchronize_srcu(&pmus_srcu
);
5318 free_percpu(pmu
->pmu_disable_count
);
5319 free_pmu_context(pmu
->pmu_cpu_context
);
5322 struct pmu
*perf_init_event(struct perf_event
*event
)
5324 struct pmu
*pmu
= NULL
;
5327 idx
= srcu_read_lock(&pmus_srcu
);
5328 list_for_each_entry_rcu(pmu
, &pmus
, entry
) {
5329 int ret
= pmu
->event_init(event
);
5333 if (ret
!= -ENOENT
) {
5338 pmu
= ERR_PTR(-ENOENT
);
5340 srcu_read_unlock(&pmus_srcu
, idx
);
5346 * Allocate and initialize a event structure
5348 static struct perf_event
*
5349 perf_event_alloc(struct perf_event_attr
*attr
, int cpu
,
5350 struct perf_event
*group_leader
,
5351 struct perf_event
*parent_event
,
5352 perf_overflow_handler_t overflow_handler
)
5355 struct perf_event
*event
;
5356 struct hw_perf_event
*hwc
;
5359 event
= kzalloc(sizeof(*event
), GFP_KERNEL
);
5361 return ERR_PTR(-ENOMEM
);
5364 * Single events are their own group leaders, with an
5365 * empty sibling list:
5368 group_leader
= event
;
5370 mutex_init(&event
->child_mutex
);
5371 INIT_LIST_HEAD(&event
->child_list
);
5373 INIT_LIST_HEAD(&event
->group_entry
);
5374 INIT_LIST_HEAD(&event
->event_entry
);
5375 INIT_LIST_HEAD(&event
->sibling_list
);
5376 init_waitqueue_head(&event
->waitq
);
5378 mutex_init(&event
->mmap_mutex
);
5381 event
->attr
= *attr
;
5382 event
->group_leader
= group_leader
;
5386 event
->parent
= parent_event
;
5388 event
->ns
= get_pid_ns(current
->nsproxy
->pid_ns
);
5389 event
->id
= atomic64_inc_return(&perf_event_id
);
5391 event
->state
= PERF_EVENT_STATE_INACTIVE
;
5393 if (!overflow_handler
&& parent_event
)
5394 overflow_handler
= parent_event
->overflow_handler
;
5396 event
->overflow_handler
= overflow_handler
;
5399 event
->state
= PERF_EVENT_STATE_OFF
;
5404 hwc
->sample_period
= attr
->sample_period
;
5405 if (attr
->freq
&& attr
->sample_freq
)
5406 hwc
->sample_period
= 1;
5407 hwc
->last_period
= hwc
->sample_period
;
5409 local64_set(&hwc
->period_left
, hwc
->sample_period
);
5412 * we currently do not support PERF_FORMAT_GROUP on inherited events
5414 if (attr
->inherit
&& (attr
->read_format
& PERF_FORMAT_GROUP
))
5417 pmu
= perf_init_event(event
);
5423 else if (IS_ERR(pmu
))
5428 put_pid_ns(event
->ns
);
5430 return ERR_PTR(err
);
5435 if (!event
->parent
) {
5436 atomic_inc(&nr_events
);
5437 if (event
->attr
.mmap
|| event
->attr
.mmap_data
)
5438 atomic_inc(&nr_mmap_events
);
5439 if (event
->attr
.comm
)
5440 atomic_inc(&nr_comm_events
);
5441 if (event
->attr
.task
)
5442 atomic_inc(&nr_task_events
);
5443 if (event
->attr
.sample_type
& PERF_SAMPLE_CALLCHAIN
) {
5444 err
= get_callchain_buffers();
5447 return ERR_PTR(err
);
5455 static int perf_copy_attr(struct perf_event_attr __user
*uattr
,
5456 struct perf_event_attr
*attr
)
5461 if (!access_ok(VERIFY_WRITE
, uattr
, PERF_ATTR_SIZE_VER0
))
5465 * zero the full structure, so that a short copy will be nice.
5467 memset(attr
, 0, sizeof(*attr
));
5469 ret
= get_user(size
, &uattr
->size
);
5473 if (size
> PAGE_SIZE
) /* silly large */
5476 if (!size
) /* abi compat */
5477 size
= PERF_ATTR_SIZE_VER0
;
5479 if (size
< PERF_ATTR_SIZE_VER0
)
5483 * If we're handed a bigger struct than we know of,
5484 * ensure all the unknown bits are 0 - i.e. new
5485 * user-space does not rely on any kernel feature
5486 * extensions we dont know about yet.
5488 if (size
> sizeof(*attr
)) {
5489 unsigned char __user
*addr
;
5490 unsigned char __user
*end
;
5493 addr
= (void __user
*)uattr
+ sizeof(*attr
);
5494 end
= (void __user
*)uattr
+ size
;
5496 for (; addr
< end
; addr
++) {
5497 ret
= get_user(val
, addr
);
5503 size
= sizeof(*attr
);
5506 ret
= copy_from_user(attr
, uattr
, size
);
5511 * If the type exists, the corresponding creation will verify
5514 if (attr
->type
>= PERF_TYPE_MAX
)
5517 if (attr
->__reserved_1
)
5520 if (attr
->sample_type
& ~(PERF_SAMPLE_MAX
-1))
5523 if (attr
->read_format
& ~(PERF_FORMAT_MAX
-1))
5530 put_user(sizeof(*attr
), &uattr
->size
);
5536 perf_event_set_output(struct perf_event
*event
, struct perf_event
*output_event
)
5538 struct perf_buffer
*buffer
= NULL
, *old_buffer
= NULL
;
5544 /* don't allow circular references */
5545 if (event
== output_event
)
5549 * Don't allow cross-cpu buffers
5551 if (output_event
->cpu
!= event
->cpu
)
5555 * If its not a per-cpu buffer, it must be the same task.
5557 if (output_event
->cpu
== -1 && output_event
->ctx
!= event
->ctx
)
5561 mutex_lock(&event
->mmap_mutex
);
5562 /* Can't redirect output if we've got an active mmap() */
5563 if (atomic_read(&event
->mmap_count
))
5567 /* get the buffer we want to redirect to */
5568 buffer
= perf_buffer_get(output_event
);
5573 old_buffer
= event
->buffer
;
5574 rcu_assign_pointer(event
->buffer
, buffer
);
5577 mutex_unlock(&event
->mmap_mutex
);
5580 perf_buffer_put(old_buffer
);
5586 * sys_perf_event_open - open a performance event, associate it to a task/cpu
5588 * @attr_uptr: event_id type attributes for monitoring/sampling
5591 * @group_fd: group leader event fd
5593 SYSCALL_DEFINE5(perf_event_open
,
5594 struct perf_event_attr __user
*, attr_uptr
,
5595 pid_t
, pid
, int, cpu
, int, group_fd
, unsigned long, flags
)
5597 struct perf_event
*group_leader
= NULL
, *output_event
= NULL
;
5598 struct perf_event
*event
, *sibling
;
5599 struct perf_event_attr attr
;
5600 struct perf_event_context
*ctx
;
5601 struct file
*event_file
= NULL
;
5602 struct file
*group_file
= NULL
;
5603 struct task_struct
*task
= NULL
;
5607 int fput_needed
= 0;
5610 /* for future expandability... */
5611 if (flags
& ~(PERF_FLAG_FD_NO_GROUP
| PERF_FLAG_FD_OUTPUT
))
5614 err
= perf_copy_attr(attr_uptr
, &attr
);
5618 if (!attr
.exclude_kernel
) {
5619 if (perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN
))
5624 if (attr
.sample_freq
> sysctl_perf_event_sample_rate
)
5628 event_fd
= get_unused_fd_flags(O_RDWR
);
5632 if (group_fd
!= -1) {
5633 group_leader
= perf_fget_light(group_fd
, &fput_needed
);
5634 if (IS_ERR(group_leader
)) {
5635 err
= PTR_ERR(group_leader
);
5638 group_file
= group_leader
->filp
;
5639 if (flags
& PERF_FLAG_FD_OUTPUT
)
5640 output_event
= group_leader
;
5641 if (flags
& PERF_FLAG_FD_NO_GROUP
)
5642 group_leader
= NULL
;
5645 event
= perf_event_alloc(&attr
, cpu
, group_leader
, NULL
, NULL
);
5646 if (IS_ERR(event
)) {
5647 err
= PTR_ERR(event
);
5652 * Special case software events and allow them to be part of
5653 * any hardware group.
5658 (is_software_event(event
) != is_software_event(group_leader
))) {
5659 if (is_software_event(event
)) {
5661 * If event and group_leader are not both a software
5662 * event, and event is, then group leader is not.
5664 * Allow the addition of software events to !software
5665 * groups, this is safe because software events never
5668 pmu
= group_leader
->pmu
;
5669 } else if (is_software_event(group_leader
) &&
5670 (group_leader
->group_flags
& PERF_GROUP_SOFTWARE
)) {
5672 * In case the group is a pure software group, and we
5673 * try to add a hardware event, move the whole group to
5674 * the hardware context.
5681 task
= find_lively_task_by_vpid(pid
);
5683 err
= PTR_ERR(task
);
5689 * Get the target context (task or percpu):
5691 ctx
= find_get_context(pmu
, task
, cpu
);
5698 * Look up the group leader (we will attach this event to it):
5704 * Do not allow a recursive hierarchy (this new sibling
5705 * becoming part of another group-sibling):
5707 if (group_leader
->group_leader
!= group_leader
)
5710 * Do not allow to attach to a group in a different
5711 * task or CPU context:
5714 if (group_leader
->ctx
->type
!= ctx
->type
)
5717 if (group_leader
->ctx
!= ctx
)
5722 * Only a group leader can be exclusive or pinned
5724 if (attr
.exclusive
|| attr
.pinned
)
5729 err
= perf_event_set_output(event
, output_event
);
5734 event_file
= anon_inode_getfile("[perf_event]", &perf_fops
, event
, O_RDWR
);
5735 if (IS_ERR(event_file
)) {
5736 err
= PTR_ERR(event_file
);
5741 struct perf_event_context
*gctx
= group_leader
->ctx
;
5743 mutex_lock(&gctx
->mutex
);
5744 perf_event_remove_from_context(group_leader
);
5745 list_for_each_entry(sibling
, &group_leader
->sibling_list
,
5747 perf_event_remove_from_context(sibling
);
5750 mutex_unlock(&gctx
->mutex
);
5754 event
->filp
= event_file
;
5755 WARN_ON_ONCE(ctx
->parent_ctx
);
5756 mutex_lock(&ctx
->mutex
);
5759 perf_install_in_context(ctx
, group_leader
, cpu
);
5761 list_for_each_entry(sibling
, &group_leader
->sibling_list
,
5763 perf_install_in_context(ctx
, sibling
, cpu
);
5768 perf_install_in_context(ctx
, event
, cpu
);
5770 mutex_unlock(&ctx
->mutex
);
5772 event
->owner
= current
;
5773 get_task_struct(current
);
5774 mutex_lock(¤t
->perf_event_mutex
);
5775 list_add_tail(&event
->owner_entry
, ¤t
->perf_event_list
);
5776 mutex_unlock(¤t
->perf_event_mutex
);
5779 * Drop the reference on the group_event after placing the
5780 * new event on the sibling_list. This ensures destruction
5781 * of the group leader will find the pointer to itself in
5782 * perf_group_detach().
5784 fput_light(group_file
, fput_needed
);
5785 fd_install(event_fd
, event_file
);
5791 fput_light(group_file
, fput_needed
);
5794 put_unused_fd(event_fd
);
5799 * perf_event_create_kernel_counter
5801 * @attr: attributes of the counter to create
5802 * @cpu: cpu in which the counter is bound
5803 * @task: task to profile (NULL for percpu)
5806 perf_event_create_kernel_counter(struct perf_event_attr
*attr
, int cpu
,
5807 struct task_struct
*task
,
5808 perf_overflow_handler_t overflow_handler
)
5810 struct perf_event_context
*ctx
;
5811 struct perf_event
*event
;
5815 * Get the target context (task or percpu):
5818 event
= perf_event_alloc(attr
, cpu
, NULL
, NULL
, overflow_handler
);
5819 if (IS_ERR(event
)) {
5820 err
= PTR_ERR(event
);
5824 ctx
= find_get_context(event
->pmu
, task
, cpu
);
5831 WARN_ON_ONCE(ctx
->parent_ctx
);
5832 mutex_lock(&ctx
->mutex
);
5833 perf_install_in_context(ctx
, event
, cpu
);
5835 mutex_unlock(&ctx
->mutex
);
5837 event
->owner
= current
;
5838 get_task_struct(current
);
5839 mutex_lock(¤t
->perf_event_mutex
);
5840 list_add_tail(&event
->owner_entry
, ¤t
->perf_event_list
);
5841 mutex_unlock(¤t
->perf_event_mutex
);
5848 return ERR_PTR(err
);
5850 EXPORT_SYMBOL_GPL(perf_event_create_kernel_counter
);
5852 static void sync_child_event(struct perf_event
*child_event
,
5853 struct task_struct
*child
)
5855 struct perf_event
*parent_event
= child_event
->parent
;
5858 if (child_event
->attr
.inherit_stat
)
5859 perf_event_read_event(child_event
, child
);
5861 child_val
= perf_event_count(child_event
);
5864 * Add back the child's count to the parent's count:
5866 atomic64_add(child_val
, &parent_event
->child_count
);
5867 atomic64_add(child_event
->total_time_enabled
,
5868 &parent_event
->child_total_time_enabled
);
5869 atomic64_add(child_event
->total_time_running
,
5870 &parent_event
->child_total_time_running
);
5873 * Remove this event from the parent's list
5875 WARN_ON_ONCE(parent_event
->ctx
->parent_ctx
);
5876 mutex_lock(&parent_event
->child_mutex
);
5877 list_del_init(&child_event
->child_list
);
5878 mutex_unlock(&parent_event
->child_mutex
);
5881 * Release the parent event, if this was the last
5884 fput(parent_event
->filp
);
5888 __perf_event_exit_task(struct perf_event
*child_event
,
5889 struct perf_event_context
*child_ctx
,
5890 struct task_struct
*child
)
5892 struct perf_event
*parent_event
;
5894 perf_event_remove_from_context(child_event
);
5896 parent_event
= child_event
->parent
;
5898 * It can happen that parent exits first, and has events
5899 * that are still around due to the child reference. These
5900 * events need to be zapped - but otherwise linger.
5903 sync_child_event(child_event
, child
);
5904 free_event(child_event
);
5908 static void perf_event_exit_task_context(struct task_struct
*child
, int ctxn
)
5910 struct perf_event
*child_event
, *tmp
;
5911 struct perf_event_context
*child_ctx
;
5912 unsigned long flags
;
5914 if (likely(!child
->perf_event_ctxp
[ctxn
])) {
5915 perf_event_task(child
, NULL
, 0);
5919 local_irq_save(flags
);
5921 * We can't reschedule here because interrupts are disabled,
5922 * and either child is current or it is a task that can't be
5923 * scheduled, so we are now safe from rescheduling changing
5926 child_ctx
= child
->perf_event_ctxp
[ctxn
];
5927 __perf_event_task_sched_out(child_ctx
);
5930 * Take the context lock here so that if find_get_context is
5931 * reading child->perf_event_ctxp, we wait until it has
5932 * incremented the context's refcount before we do put_ctx below.
5934 raw_spin_lock(&child_ctx
->lock
);
5935 child
->perf_event_ctxp
[ctxn
] = NULL
;
5937 * If this context is a clone; unclone it so it can't get
5938 * swapped to another process while we're removing all
5939 * the events from it.
5941 unclone_ctx(child_ctx
);
5942 update_context_time(child_ctx
);
5943 raw_spin_unlock_irqrestore(&child_ctx
->lock
, flags
);
5946 * Report the task dead after unscheduling the events so that we
5947 * won't get any samples after PERF_RECORD_EXIT. We can however still
5948 * get a few PERF_RECORD_READ events.
5950 perf_event_task(child
, child_ctx
, 0);
5953 * We can recurse on the same lock type through:
5955 * __perf_event_exit_task()
5956 * sync_child_event()
5957 * fput(parent_event->filp)
5959 * mutex_lock(&ctx->mutex)
5961 * But since its the parent context it won't be the same instance.
5963 mutex_lock(&child_ctx
->mutex
);
5966 list_for_each_entry_safe(child_event
, tmp
, &child_ctx
->pinned_groups
,
5968 __perf_event_exit_task(child_event
, child_ctx
, child
);
5970 list_for_each_entry_safe(child_event
, tmp
, &child_ctx
->flexible_groups
,
5972 __perf_event_exit_task(child_event
, child_ctx
, child
);
5975 * If the last event was a group event, it will have appended all
5976 * its siblings to the list, but we obtained 'tmp' before that which
5977 * will still point to the list head terminating the iteration.
5979 if (!list_empty(&child_ctx
->pinned_groups
) ||
5980 !list_empty(&child_ctx
->flexible_groups
))
5983 mutex_unlock(&child_ctx
->mutex
);
5989 * When a child task exits, feed back event values to parent events.
5991 void perf_event_exit_task(struct task_struct
*child
)
5995 for_each_task_context_nr(ctxn
)
5996 perf_event_exit_task_context(child
, ctxn
);
5999 static void perf_free_event(struct perf_event
*event
,
6000 struct perf_event_context
*ctx
)
6002 struct perf_event
*parent
= event
->parent
;
6004 if (WARN_ON_ONCE(!parent
))
6007 mutex_lock(&parent
->child_mutex
);
6008 list_del_init(&event
->child_list
);
6009 mutex_unlock(&parent
->child_mutex
);
6013 perf_group_detach(event
);
6014 list_del_event(event
, ctx
);
6019 * free an unexposed, unused context as created by inheritance by
6020 * perf_event_init_task below, used by fork() in case of fail.
6022 void perf_event_free_task(struct task_struct
*task
)
6024 struct perf_event_context
*ctx
;
6025 struct perf_event
*event
, *tmp
;
6028 for_each_task_context_nr(ctxn
) {
6029 ctx
= task
->perf_event_ctxp
[ctxn
];
6033 mutex_lock(&ctx
->mutex
);
6035 list_for_each_entry_safe(event
, tmp
, &ctx
->pinned_groups
,
6037 perf_free_event(event
, ctx
);
6039 list_for_each_entry_safe(event
, tmp
, &ctx
->flexible_groups
,
6041 perf_free_event(event
, ctx
);
6043 if (!list_empty(&ctx
->pinned_groups
) ||
6044 !list_empty(&ctx
->flexible_groups
))
6047 mutex_unlock(&ctx
->mutex
);
6053 void perf_event_delayed_put(struct task_struct
*task
)
6057 for_each_task_context_nr(ctxn
)
6058 WARN_ON_ONCE(task
->perf_event_ctxp
[ctxn
]);
6062 * inherit a event from parent task to child task:
6064 static struct perf_event
*
6065 inherit_event(struct perf_event
*parent_event
,
6066 struct task_struct
*parent
,
6067 struct perf_event_context
*parent_ctx
,
6068 struct task_struct
*child
,
6069 struct perf_event
*group_leader
,
6070 struct perf_event_context
*child_ctx
)
6072 struct perf_event
*child_event
;
6073 unsigned long flags
;
6076 * Instead of creating recursive hierarchies of events,
6077 * we link inherited events back to the original parent,
6078 * which has a filp for sure, which we use as the reference
6081 if (parent_event
->parent
)
6082 parent_event
= parent_event
->parent
;
6084 child_event
= perf_event_alloc(&parent_event
->attr
,
6086 group_leader
, parent_event
,
6088 if (IS_ERR(child_event
))
6093 * Make the child state follow the state of the parent event,
6094 * not its attr.disabled bit. We hold the parent's mutex,
6095 * so we won't race with perf_event_{en, dis}able_family.
6097 if (parent_event
->state
>= PERF_EVENT_STATE_INACTIVE
)
6098 child_event
->state
= PERF_EVENT_STATE_INACTIVE
;
6100 child_event
->state
= PERF_EVENT_STATE_OFF
;
6102 if (parent_event
->attr
.freq
) {
6103 u64 sample_period
= parent_event
->hw
.sample_period
;
6104 struct hw_perf_event
*hwc
= &child_event
->hw
;
6106 hwc
->sample_period
= sample_period
;
6107 hwc
->last_period
= sample_period
;
6109 local64_set(&hwc
->period_left
, sample_period
);
6112 child_event
->ctx
= child_ctx
;
6113 child_event
->overflow_handler
= parent_event
->overflow_handler
;
6116 * Link it up in the child's context:
6118 raw_spin_lock_irqsave(&child_ctx
->lock
, flags
);
6119 add_event_to_ctx(child_event
, child_ctx
);
6120 raw_spin_unlock_irqrestore(&child_ctx
->lock
, flags
);
6123 * Get a reference to the parent filp - we will fput it
6124 * when the child event exits. This is safe to do because
6125 * we are in the parent and we know that the filp still
6126 * exists and has a nonzero count:
6128 atomic_long_inc(&parent_event
->filp
->f_count
);
6131 * Link this into the parent event's child list
6133 WARN_ON_ONCE(parent_event
->ctx
->parent_ctx
);
6134 mutex_lock(&parent_event
->child_mutex
);
6135 list_add_tail(&child_event
->child_list
, &parent_event
->child_list
);
6136 mutex_unlock(&parent_event
->child_mutex
);
6141 static int inherit_group(struct perf_event
*parent_event
,
6142 struct task_struct
*parent
,
6143 struct perf_event_context
*parent_ctx
,
6144 struct task_struct
*child
,
6145 struct perf_event_context
*child_ctx
)
6147 struct perf_event
*leader
;
6148 struct perf_event
*sub
;
6149 struct perf_event
*child_ctr
;
6151 leader
= inherit_event(parent_event
, parent
, parent_ctx
,
6152 child
, NULL
, child_ctx
);
6154 return PTR_ERR(leader
);
6155 list_for_each_entry(sub
, &parent_event
->sibling_list
, group_entry
) {
6156 child_ctr
= inherit_event(sub
, parent
, parent_ctx
,
6157 child
, leader
, child_ctx
);
6158 if (IS_ERR(child_ctr
))
6159 return PTR_ERR(child_ctr
);
6165 inherit_task_group(struct perf_event
*event
, struct task_struct
*parent
,
6166 struct perf_event_context
*parent_ctx
,
6167 struct task_struct
*child
, int ctxn
,
6171 struct perf_event_context
*child_ctx
;
6173 if (!event
->attr
.inherit
) {
6178 child_ctx
= child
->perf_event_ctxp
[ctxn
];
6181 * This is executed from the parent task context, so
6182 * inherit events that have been marked for cloning.
6183 * First allocate and initialize a context for the
6187 child_ctx
= alloc_perf_context(event
->pmu
, child
);
6191 child
->perf_event_ctxp
[ctxn
] = child_ctx
;
6194 ret
= inherit_group(event
, parent
, parent_ctx
,
6204 * Initialize the perf_event context in task_struct
6206 int perf_event_init_context(struct task_struct
*child
, int ctxn
)
6208 struct perf_event_context
*child_ctx
, *parent_ctx
;
6209 struct perf_event_context
*cloned_ctx
;
6210 struct perf_event
*event
;
6211 struct task_struct
*parent
= current
;
6212 int inherited_all
= 1;
6215 child
->perf_event_ctxp
[ctxn
] = NULL
;
6217 mutex_init(&child
->perf_event_mutex
);
6218 INIT_LIST_HEAD(&child
->perf_event_list
);
6220 if (likely(!parent
->perf_event_ctxp
[ctxn
]))
6224 * If the parent's context is a clone, pin it so it won't get
6227 parent_ctx
= perf_pin_task_context(parent
, ctxn
);
6230 * No need to check if parent_ctx != NULL here; since we saw
6231 * it non-NULL earlier, the only reason for it to become NULL
6232 * is if we exit, and since we're currently in the middle of
6233 * a fork we can't be exiting at the same time.
6237 * Lock the parent list. No need to lock the child - not PID
6238 * hashed yet and not running, so nobody can access it.
6240 mutex_lock(&parent_ctx
->mutex
);
6243 * We dont have to disable NMIs - we are only looking at
6244 * the list, not manipulating it:
6246 list_for_each_entry(event
, &parent_ctx
->pinned_groups
, group_entry
) {
6247 ret
= inherit_task_group(event
, parent
, parent_ctx
,
6248 child
, ctxn
, &inherited_all
);
6253 list_for_each_entry(event
, &parent_ctx
->flexible_groups
, group_entry
) {
6254 ret
= inherit_task_group(event
, parent
, parent_ctx
,
6255 child
, ctxn
, &inherited_all
);
6260 child_ctx
= child
->perf_event_ctxp
[ctxn
];
6262 if (child_ctx
&& inherited_all
) {
6264 * Mark the child context as a clone of the parent
6265 * context, or of whatever the parent is a clone of.
6266 * Note that if the parent is a clone, it could get
6267 * uncloned at any point, but that doesn't matter
6268 * because the list of events and the generation
6269 * count can't have changed since we took the mutex.
6271 cloned_ctx
= rcu_dereference(parent_ctx
->parent_ctx
);
6273 child_ctx
->parent_ctx
= cloned_ctx
;
6274 child_ctx
->parent_gen
= parent_ctx
->parent_gen
;
6276 child_ctx
->parent_ctx
= parent_ctx
;
6277 child_ctx
->parent_gen
= parent_ctx
->generation
;
6279 get_ctx(child_ctx
->parent_ctx
);
6282 mutex_unlock(&parent_ctx
->mutex
);
6284 perf_unpin_context(parent_ctx
);
6290 * Initialize the perf_event context in task_struct
6292 int perf_event_init_task(struct task_struct
*child
)
6296 for_each_task_context_nr(ctxn
) {
6297 ret
= perf_event_init_context(child
, ctxn
);
6305 static void __init
perf_event_init_all_cpus(void)
6307 struct swevent_htable
*swhash
;
6310 for_each_possible_cpu(cpu
) {
6311 swhash
= &per_cpu(swevent_htable
, cpu
);
6312 mutex_init(&swhash
->hlist_mutex
);
6313 INIT_LIST_HEAD(&per_cpu(rotation_list
, cpu
));
6317 static void __cpuinit
perf_event_init_cpu(int cpu
)
6319 struct swevent_htable
*swhash
= &per_cpu(swevent_htable
, cpu
);
6321 mutex_lock(&swhash
->hlist_mutex
);
6322 if (swhash
->hlist_refcount
> 0) {
6323 struct swevent_hlist
*hlist
;
6325 hlist
= kzalloc_node(sizeof(*hlist
), GFP_KERNEL
, cpu_to_node(cpu
));
6327 rcu_assign_pointer(swhash
->swevent_hlist
, hlist
);
6329 mutex_unlock(&swhash
->hlist_mutex
);
6332 #ifdef CONFIG_HOTPLUG_CPU
6333 static void perf_pmu_rotate_stop(struct pmu
*pmu
)
6335 struct perf_cpu_context
*cpuctx
= this_cpu_ptr(pmu
->pmu_cpu_context
);
6337 WARN_ON(!irqs_disabled());
6339 list_del_init(&cpuctx
->rotation_list
);
6342 static void __perf_event_exit_context(void *__info
)
6344 struct perf_event_context
*ctx
= __info
;
6345 struct perf_event
*event
, *tmp
;
6347 perf_pmu_rotate_stop(ctx
->pmu
);
6349 list_for_each_entry_safe(event
, tmp
, &ctx
->pinned_groups
, group_entry
)
6350 __perf_event_remove_from_context(event
);
6351 list_for_each_entry_safe(event
, tmp
, &ctx
->flexible_groups
, group_entry
)
6352 __perf_event_remove_from_context(event
);
6355 static void perf_event_exit_cpu_context(int cpu
)
6357 struct perf_event_context
*ctx
;
6361 idx
= srcu_read_lock(&pmus_srcu
);
6362 list_for_each_entry_rcu(pmu
, &pmus
, entry
) {
6363 ctx
= &per_cpu_ptr(pmu
->pmu_cpu_context
, cpu
)->ctx
;
6365 mutex_lock(&ctx
->mutex
);
6366 smp_call_function_single(cpu
, __perf_event_exit_context
, ctx
, 1);
6367 mutex_unlock(&ctx
->mutex
);
6369 srcu_read_unlock(&pmus_srcu
, idx
);
6372 static void perf_event_exit_cpu(int cpu
)
6374 struct swevent_htable
*swhash
= &per_cpu(swevent_htable
, cpu
);
6376 mutex_lock(&swhash
->hlist_mutex
);
6377 swevent_hlist_release(swhash
);
6378 mutex_unlock(&swhash
->hlist_mutex
);
6380 perf_event_exit_cpu_context(cpu
);
6383 static inline void perf_event_exit_cpu(int cpu
) { }
6386 static int __cpuinit
6387 perf_cpu_notify(struct notifier_block
*self
, unsigned long action
, void *hcpu
)
6389 unsigned int cpu
= (long)hcpu
;
6391 switch (action
& ~CPU_TASKS_FROZEN
) {
6393 case CPU_UP_PREPARE
:
6394 case CPU_DOWN_FAILED
:
6395 perf_event_init_cpu(cpu
);
6398 case CPU_UP_CANCELED
:
6399 case CPU_DOWN_PREPARE
:
6400 perf_event_exit_cpu(cpu
);
6410 void __init
perf_event_init(void)
6412 perf_event_init_all_cpus();
6413 init_srcu_struct(&pmus_srcu
);
6414 perf_pmu_register(&perf_swevent
);
6415 perf_pmu_register(&perf_cpu_clock
);
6416 perf_pmu_register(&perf_task_clock
);
6418 perf_cpu_notifier(perf_cpu_notify
);