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_buffer_put(struct perf_buffer
*buffer
);
2211 static void free_event(struct perf_event
*event
)
2213 irq_work_sync(&event
->pending
);
2215 if (!event
->parent
) {
2216 atomic_dec(&nr_events
);
2217 if (event
->attr
.mmap
|| event
->attr
.mmap_data
)
2218 atomic_dec(&nr_mmap_events
);
2219 if (event
->attr
.comm
)
2220 atomic_dec(&nr_comm_events
);
2221 if (event
->attr
.task
)
2222 atomic_dec(&nr_task_events
);
2223 if (event
->attr
.sample_type
& PERF_SAMPLE_CALLCHAIN
)
2224 put_callchain_buffers();
2227 if (event
->buffer
) {
2228 perf_buffer_put(event
->buffer
);
2229 event
->buffer
= NULL
;
2233 event
->destroy(event
);
2236 put_ctx(event
->ctx
);
2238 call_rcu(&event
->rcu_head
, free_event_rcu
);
2241 int perf_event_release_kernel(struct perf_event
*event
)
2243 struct perf_event_context
*ctx
= event
->ctx
;
2246 * Remove from the PMU, can't get re-enabled since we got
2247 * here because the last ref went.
2249 perf_event_disable(event
);
2251 WARN_ON_ONCE(ctx
->parent_ctx
);
2253 * There are two ways this annotation is useful:
2255 * 1) there is a lock recursion from perf_event_exit_task
2256 * see the comment there.
2258 * 2) there is a lock-inversion with mmap_sem through
2259 * perf_event_read_group(), which takes faults while
2260 * holding ctx->mutex, however this is called after
2261 * the last filedesc died, so there is no possibility
2262 * to trigger the AB-BA case.
2264 mutex_lock_nested(&ctx
->mutex
, SINGLE_DEPTH_NESTING
);
2265 raw_spin_lock_irq(&ctx
->lock
);
2266 perf_group_detach(event
);
2267 list_del_event(event
, ctx
);
2268 raw_spin_unlock_irq(&ctx
->lock
);
2269 mutex_unlock(&ctx
->mutex
);
2271 mutex_lock(&event
->owner
->perf_event_mutex
);
2272 list_del_init(&event
->owner_entry
);
2273 mutex_unlock(&event
->owner
->perf_event_mutex
);
2274 put_task_struct(event
->owner
);
2280 EXPORT_SYMBOL_GPL(perf_event_release_kernel
);
2283 * Called when the last reference to the file is gone.
2285 static int perf_release(struct inode
*inode
, struct file
*file
)
2287 struct perf_event
*event
= file
->private_data
;
2289 file
->private_data
= NULL
;
2291 return perf_event_release_kernel(event
);
2294 static int perf_event_read_size(struct perf_event
*event
)
2296 int entry
= sizeof(u64
); /* value */
2300 if (event
->attr
.read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
)
2301 size
+= sizeof(u64
);
2303 if (event
->attr
.read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
)
2304 size
+= sizeof(u64
);
2306 if (event
->attr
.read_format
& PERF_FORMAT_ID
)
2307 entry
+= sizeof(u64
);
2309 if (event
->attr
.read_format
& PERF_FORMAT_GROUP
) {
2310 nr
+= event
->group_leader
->nr_siblings
;
2311 size
+= sizeof(u64
);
2319 u64
perf_event_read_value(struct perf_event
*event
, u64
*enabled
, u64
*running
)
2321 struct perf_event
*child
;
2327 mutex_lock(&event
->child_mutex
);
2328 total
+= perf_event_read(event
);
2329 *enabled
+= event
->total_time_enabled
+
2330 atomic64_read(&event
->child_total_time_enabled
);
2331 *running
+= event
->total_time_running
+
2332 atomic64_read(&event
->child_total_time_running
);
2334 list_for_each_entry(child
, &event
->child_list
, child_list
) {
2335 total
+= perf_event_read(child
);
2336 *enabled
+= child
->total_time_enabled
;
2337 *running
+= child
->total_time_running
;
2339 mutex_unlock(&event
->child_mutex
);
2343 EXPORT_SYMBOL_GPL(perf_event_read_value
);
2345 static int perf_event_read_group(struct perf_event
*event
,
2346 u64 read_format
, char __user
*buf
)
2348 struct perf_event
*leader
= event
->group_leader
, *sub
;
2349 int n
= 0, size
= 0, ret
= -EFAULT
;
2350 struct perf_event_context
*ctx
= leader
->ctx
;
2352 u64 count
, enabled
, running
;
2354 mutex_lock(&ctx
->mutex
);
2355 count
= perf_event_read_value(leader
, &enabled
, &running
);
2357 values
[n
++] = 1 + leader
->nr_siblings
;
2358 if (read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
)
2359 values
[n
++] = enabled
;
2360 if (read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
)
2361 values
[n
++] = running
;
2362 values
[n
++] = count
;
2363 if (read_format
& PERF_FORMAT_ID
)
2364 values
[n
++] = primary_event_id(leader
);
2366 size
= n
* sizeof(u64
);
2368 if (copy_to_user(buf
, values
, size
))
2373 list_for_each_entry(sub
, &leader
->sibling_list
, group_entry
) {
2376 values
[n
++] = perf_event_read_value(sub
, &enabled
, &running
);
2377 if (read_format
& PERF_FORMAT_ID
)
2378 values
[n
++] = primary_event_id(sub
);
2380 size
= n
* sizeof(u64
);
2382 if (copy_to_user(buf
+ ret
, values
, size
)) {
2390 mutex_unlock(&ctx
->mutex
);
2395 static int perf_event_read_one(struct perf_event
*event
,
2396 u64 read_format
, char __user
*buf
)
2398 u64 enabled
, running
;
2402 values
[n
++] = perf_event_read_value(event
, &enabled
, &running
);
2403 if (read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
)
2404 values
[n
++] = enabled
;
2405 if (read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
)
2406 values
[n
++] = running
;
2407 if (read_format
& PERF_FORMAT_ID
)
2408 values
[n
++] = primary_event_id(event
);
2410 if (copy_to_user(buf
, values
, n
* sizeof(u64
)))
2413 return n
* sizeof(u64
);
2417 * Read the performance event - simple non blocking version for now
2420 perf_read_hw(struct perf_event
*event
, char __user
*buf
, size_t count
)
2422 u64 read_format
= event
->attr
.read_format
;
2426 * Return end-of-file for a read on a event that is in
2427 * error state (i.e. because it was pinned but it couldn't be
2428 * scheduled on to the CPU at some point).
2430 if (event
->state
== PERF_EVENT_STATE_ERROR
)
2433 if (count
< perf_event_read_size(event
))
2436 WARN_ON_ONCE(event
->ctx
->parent_ctx
);
2437 if (read_format
& PERF_FORMAT_GROUP
)
2438 ret
= perf_event_read_group(event
, read_format
, buf
);
2440 ret
= perf_event_read_one(event
, read_format
, buf
);
2446 perf_read(struct file
*file
, char __user
*buf
, size_t count
, loff_t
*ppos
)
2448 struct perf_event
*event
= file
->private_data
;
2450 return perf_read_hw(event
, buf
, count
);
2453 static unsigned int perf_poll(struct file
*file
, poll_table
*wait
)
2455 struct perf_event
*event
= file
->private_data
;
2456 struct perf_buffer
*buffer
;
2457 unsigned int events
= POLL_HUP
;
2460 buffer
= rcu_dereference(event
->buffer
);
2462 events
= atomic_xchg(&buffer
->poll
, 0);
2465 poll_wait(file
, &event
->waitq
, wait
);
2470 static void perf_event_reset(struct perf_event
*event
)
2472 (void)perf_event_read(event
);
2473 local64_set(&event
->count
, 0);
2474 perf_event_update_userpage(event
);
2478 * Holding the top-level event's child_mutex means that any
2479 * descendant process that has inherited this event will block
2480 * in sync_child_event if it goes to exit, thus satisfying the
2481 * task existence requirements of perf_event_enable/disable.
2483 static void perf_event_for_each_child(struct perf_event
*event
,
2484 void (*func
)(struct perf_event
*))
2486 struct perf_event
*child
;
2488 WARN_ON_ONCE(event
->ctx
->parent_ctx
);
2489 mutex_lock(&event
->child_mutex
);
2491 list_for_each_entry(child
, &event
->child_list
, child_list
)
2493 mutex_unlock(&event
->child_mutex
);
2496 static void perf_event_for_each(struct perf_event
*event
,
2497 void (*func
)(struct perf_event
*))
2499 struct perf_event_context
*ctx
= event
->ctx
;
2500 struct perf_event
*sibling
;
2502 WARN_ON_ONCE(ctx
->parent_ctx
);
2503 mutex_lock(&ctx
->mutex
);
2504 event
= event
->group_leader
;
2506 perf_event_for_each_child(event
, func
);
2508 list_for_each_entry(sibling
, &event
->sibling_list
, group_entry
)
2509 perf_event_for_each_child(event
, func
);
2510 mutex_unlock(&ctx
->mutex
);
2513 static int perf_event_period(struct perf_event
*event
, u64 __user
*arg
)
2515 struct perf_event_context
*ctx
= event
->ctx
;
2520 if (!event
->attr
.sample_period
)
2523 size
= copy_from_user(&value
, arg
, sizeof(value
));
2524 if (size
!= sizeof(value
))
2530 raw_spin_lock_irq(&ctx
->lock
);
2531 if (event
->attr
.freq
) {
2532 if (value
> sysctl_perf_event_sample_rate
) {
2537 event
->attr
.sample_freq
= value
;
2539 event
->attr
.sample_period
= value
;
2540 event
->hw
.sample_period
= value
;
2543 raw_spin_unlock_irq(&ctx
->lock
);
2548 static const struct file_operations perf_fops
;
2550 static struct perf_event
*perf_fget_light(int fd
, int *fput_needed
)
2554 file
= fget_light(fd
, fput_needed
);
2556 return ERR_PTR(-EBADF
);
2558 if (file
->f_op
!= &perf_fops
) {
2559 fput_light(file
, *fput_needed
);
2561 return ERR_PTR(-EBADF
);
2564 return file
->private_data
;
2567 static int perf_event_set_output(struct perf_event
*event
,
2568 struct perf_event
*output_event
);
2569 static int perf_event_set_filter(struct perf_event
*event
, void __user
*arg
);
2571 static long perf_ioctl(struct file
*file
, unsigned int cmd
, unsigned long arg
)
2573 struct perf_event
*event
= file
->private_data
;
2574 void (*func
)(struct perf_event
*);
2578 case PERF_EVENT_IOC_ENABLE
:
2579 func
= perf_event_enable
;
2581 case PERF_EVENT_IOC_DISABLE
:
2582 func
= perf_event_disable
;
2584 case PERF_EVENT_IOC_RESET
:
2585 func
= perf_event_reset
;
2588 case PERF_EVENT_IOC_REFRESH
:
2589 return perf_event_refresh(event
, arg
);
2591 case PERF_EVENT_IOC_PERIOD
:
2592 return perf_event_period(event
, (u64 __user
*)arg
);
2594 case PERF_EVENT_IOC_SET_OUTPUT
:
2596 struct perf_event
*output_event
= NULL
;
2597 int fput_needed
= 0;
2601 output_event
= perf_fget_light(arg
, &fput_needed
);
2602 if (IS_ERR(output_event
))
2603 return PTR_ERR(output_event
);
2606 ret
= perf_event_set_output(event
, output_event
);
2608 fput_light(output_event
->filp
, fput_needed
);
2613 case PERF_EVENT_IOC_SET_FILTER
:
2614 return perf_event_set_filter(event
, (void __user
*)arg
);
2620 if (flags
& PERF_IOC_FLAG_GROUP
)
2621 perf_event_for_each(event
, func
);
2623 perf_event_for_each_child(event
, func
);
2628 int perf_event_task_enable(void)
2630 struct perf_event
*event
;
2632 mutex_lock(¤t
->perf_event_mutex
);
2633 list_for_each_entry(event
, ¤t
->perf_event_list
, owner_entry
)
2634 perf_event_for_each_child(event
, perf_event_enable
);
2635 mutex_unlock(¤t
->perf_event_mutex
);
2640 int perf_event_task_disable(void)
2642 struct perf_event
*event
;
2644 mutex_lock(¤t
->perf_event_mutex
);
2645 list_for_each_entry(event
, ¤t
->perf_event_list
, owner_entry
)
2646 perf_event_for_each_child(event
, perf_event_disable
);
2647 mutex_unlock(¤t
->perf_event_mutex
);
2652 #ifndef PERF_EVENT_INDEX_OFFSET
2653 # define PERF_EVENT_INDEX_OFFSET 0
2656 static int perf_event_index(struct perf_event
*event
)
2658 if (event
->hw
.state
& PERF_HES_STOPPED
)
2661 if (event
->state
!= PERF_EVENT_STATE_ACTIVE
)
2664 return event
->hw
.idx
+ 1 - PERF_EVENT_INDEX_OFFSET
;
2668 * Callers need to ensure there can be no nesting of this function, otherwise
2669 * the seqlock logic goes bad. We can not serialize this because the arch
2670 * code calls this from NMI context.
2672 void perf_event_update_userpage(struct perf_event
*event
)
2674 struct perf_event_mmap_page
*userpg
;
2675 struct perf_buffer
*buffer
;
2678 buffer
= rcu_dereference(event
->buffer
);
2682 userpg
= buffer
->user_page
;
2685 * Disable preemption so as to not let the corresponding user-space
2686 * spin too long if we get preempted.
2691 userpg
->index
= perf_event_index(event
);
2692 userpg
->offset
= perf_event_count(event
);
2693 if (event
->state
== PERF_EVENT_STATE_ACTIVE
)
2694 userpg
->offset
-= local64_read(&event
->hw
.prev_count
);
2696 userpg
->time_enabled
= event
->total_time_enabled
+
2697 atomic64_read(&event
->child_total_time_enabled
);
2699 userpg
->time_running
= event
->total_time_running
+
2700 atomic64_read(&event
->child_total_time_running
);
2709 static unsigned long perf_data_size(struct perf_buffer
*buffer
);
2712 perf_buffer_init(struct perf_buffer
*buffer
, long watermark
, int flags
)
2714 long max_size
= perf_data_size(buffer
);
2717 buffer
->watermark
= min(max_size
, watermark
);
2719 if (!buffer
->watermark
)
2720 buffer
->watermark
= max_size
/ 2;
2722 if (flags
& PERF_BUFFER_WRITABLE
)
2723 buffer
->writable
= 1;
2725 atomic_set(&buffer
->refcount
, 1);
2728 #ifndef CONFIG_PERF_USE_VMALLOC
2731 * Back perf_mmap() with regular GFP_KERNEL-0 pages.
2734 static struct page
*
2735 perf_mmap_to_page(struct perf_buffer
*buffer
, unsigned long pgoff
)
2737 if (pgoff
> buffer
->nr_pages
)
2741 return virt_to_page(buffer
->user_page
);
2743 return virt_to_page(buffer
->data_pages
[pgoff
- 1]);
2746 static void *perf_mmap_alloc_page(int cpu
)
2751 node
= (cpu
== -1) ? cpu
: cpu_to_node(cpu
);
2752 page
= alloc_pages_node(node
, GFP_KERNEL
| __GFP_ZERO
, 0);
2756 return page_address(page
);
2759 static struct perf_buffer
*
2760 perf_buffer_alloc(int nr_pages
, long watermark
, int cpu
, int flags
)
2762 struct perf_buffer
*buffer
;
2766 size
= sizeof(struct perf_buffer
);
2767 size
+= nr_pages
* sizeof(void *);
2769 buffer
= kzalloc(size
, GFP_KERNEL
);
2773 buffer
->user_page
= perf_mmap_alloc_page(cpu
);
2774 if (!buffer
->user_page
)
2775 goto fail_user_page
;
2777 for (i
= 0; i
< nr_pages
; i
++) {
2778 buffer
->data_pages
[i
] = perf_mmap_alloc_page(cpu
);
2779 if (!buffer
->data_pages
[i
])
2780 goto fail_data_pages
;
2783 buffer
->nr_pages
= nr_pages
;
2785 perf_buffer_init(buffer
, watermark
, flags
);
2790 for (i
--; i
>= 0; i
--)
2791 free_page((unsigned long)buffer
->data_pages
[i
]);
2793 free_page((unsigned long)buffer
->user_page
);
2802 static void perf_mmap_free_page(unsigned long addr
)
2804 struct page
*page
= virt_to_page((void *)addr
);
2806 page
->mapping
= NULL
;
2810 static void perf_buffer_free(struct perf_buffer
*buffer
)
2814 perf_mmap_free_page((unsigned long)buffer
->user_page
);
2815 for (i
= 0; i
< buffer
->nr_pages
; i
++)
2816 perf_mmap_free_page((unsigned long)buffer
->data_pages
[i
]);
2820 static inline int page_order(struct perf_buffer
*buffer
)
2828 * Back perf_mmap() with vmalloc memory.
2830 * Required for architectures that have d-cache aliasing issues.
2833 static inline int page_order(struct perf_buffer
*buffer
)
2835 return buffer
->page_order
;
2838 static struct page
*
2839 perf_mmap_to_page(struct perf_buffer
*buffer
, unsigned long pgoff
)
2841 if (pgoff
> (1UL << page_order(buffer
)))
2844 return vmalloc_to_page((void *)buffer
->user_page
+ pgoff
* PAGE_SIZE
);
2847 static void perf_mmap_unmark_page(void *addr
)
2849 struct page
*page
= vmalloc_to_page(addr
);
2851 page
->mapping
= NULL
;
2854 static void perf_buffer_free_work(struct work_struct
*work
)
2856 struct perf_buffer
*buffer
;
2860 buffer
= container_of(work
, struct perf_buffer
, work
);
2861 nr
= 1 << page_order(buffer
);
2863 base
= buffer
->user_page
;
2864 for (i
= 0; i
< nr
+ 1; i
++)
2865 perf_mmap_unmark_page(base
+ (i
* PAGE_SIZE
));
2871 static void perf_buffer_free(struct perf_buffer
*buffer
)
2873 schedule_work(&buffer
->work
);
2876 static struct perf_buffer
*
2877 perf_buffer_alloc(int nr_pages
, long watermark
, int cpu
, int flags
)
2879 struct perf_buffer
*buffer
;
2883 size
= sizeof(struct perf_buffer
);
2884 size
+= sizeof(void *);
2886 buffer
= kzalloc(size
, GFP_KERNEL
);
2890 INIT_WORK(&buffer
->work
, perf_buffer_free_work
);
2892 all_buf
= vmalloc_user((nr_pages
+ 1) * PAGE_SIZE
);
2896 buffer
->user_page
= all_buf
;
2897 buffer
->data_pages
[0] = all_buf
+ PAGE_SIZE
;
2898 buffer
->page_order
= ilog2(nr_pages
);
2899 buffer
->nr_pages
= 1;
2901 perf_buffer_init(buffer
, watermark
, flags
);
2914 static unsigned long perf_data_size(struct perf_buffer
*buffer
)
2916 return buffer
->nr_pages
<< (PAGE_SHIFT
+ page_order(buffer
));
2919 static int perf_mmap_fault(struct vm_area_struct
*vma
, struct vm_fault
*vmf
)
2921 struct perf_event
*event
= vma
->vm_file
->private_data
;
2922 struct perf_buffer
*buffer
;
2923 int ret
= VM_FAULT_SIGBUS
;
2925 if (vmf
->flags
& FAULT_FLAG_MKWRITE
) {
2926 if (vmf
->pgoff
== 0)
2932 buffer
= rcu_dereference(event
->buffer
);
2936 if (vmf
->pgoff
&& (vmf
->flags
& FAULT_FLAG_WRITE
))
2939 vmf
->page
= perf_mmap_to_page(buffer
, vmf
->pgoff
);
2943 get_page(vmf
->page
);
2944 vmf
->page
->mapping
= vma
->vm_file
->f_mapping
;
2945 vmf
->page
->index
= vmf
->pgoff
;
2954 static void perf_buffer_free_rcu(struct rcu_head
*rcu_head
)
2956 struct perf_buffer
*buffer
;
2958 buffer
= container_of(rcu_head
, struct perf_buffer
, rcu_head
);
2959 perf_buffer_free(buffer
);
2962 static struct perf_buffer
*perf_buffer_get(struct perf_event
*event
)
2964 struct perf_buffer
*buffer
;
2967 buffer
= rcu_dereference(event
->buffer
);
2969 if (!atomic_inc_not_zero(&buffer
->refcount
))
2977 static void perf_buffer_put(struct perf_buffer
*buffer
)
2979 if (!atomic_dec_and_test(&buffer
->refcount
))
2982 call_rcu(&buffer
->rcu_head
, perf_buffer_free_rcu
);
2985 static void perf_mmap_open(struct vm_area_struct
*vma
)
2987 struct perf_event
*event
= vma
->vm_file
->private_data
;
2989 atomic_inc(&event
->mmap_count
);
2992 static void perf_mmap_close(struct vm_area_struct
*vma
)
2994 struct perf_event
*event
= vma
->vm_file
->private_data
;
2996 if (atomic_dec_and_mutex_lock(&event
->mmap_count
, &event
->mmap_mutex
)) {
2997 unsigned long size
= perf_data_size(event
->buffer
);
2998 struct user_struct
*user
= event
->mmap_user
;
2999 struct perf_buffer
*buffer
= event
->buffer
;
3001 atomic_long_sub((size
>> PAGE_SHIFT
) + 1, &user
->locked_vm
);
3002 vma
->vm_mm
->locked_vm
-= event
->mmap_locked
;
3003 rcu_assign_pointer(event
->buffer
, NULL
);
3004 mutex_unlock(&event
->mmap_mutex
);
3006 perf_buffer_put(buffer
);
3011 static const struct vm_operations_struct perf_mmap_vmops
= {
3012 .open
= perf_mmap_open
,
3013 .close
= perf_mmap_close
,
3014 .fault
= perf_mmap_fault
,
3015 .page_mkwrite
= perf_mmap_fault
,
3018 static int perf_mmap(struct file
*file
, struct vm_area_struct
*vma
)
3020 struct perf_event
*event
= file
->private_data
;
3021 unsigned long user_locked
, user_lock_limit
;
3022 struct user_struct
*user
= current_user();
3023 unsigned long locked
, lock_limit
;
3024 struct perf_buffer
*buffer
;
3025 unsigned long vma_size
;
3026 unsigned long nr_pages
;
3027 long user_extra
, extra
;
3028 int ret
= 0, flags
= 0;
3031 * Don't allow mmap() of inherited per-task counters. This would
3032 * create a performance issue due to all children writing to the
3035 if (event
->cpu
== -1 && event
->attr
.inherit
)
3038 if (!(vma
->vm_flags
& VM_SHARED
))
3041 vma_size
= vma
->vm_end
- vma
->vm_start
;
3042 nr_pages
= (vma_size
/ PAGE_SIZE
) - 1;
3045 * If we have buffer pages ensure they're a power-of-two number, so we
3046 * can do bitmasks instead of modulo.
3048 if (nr_pages
!= 0 && !is_power_of_2(nr_pages
))
3051 if (vma_size
!= PAGE_SIZE
* (1 + nr_pages
))
3054 if (vma
->vm_pgoff
!= 0)
3057 WARN_ON_ONCE(event
->ctx
->parent_ctx
);
3058 mutex_lock(&event
->mmap_mutex
);
3059 if (event
->buffer
) {
3060 if (event
->buffer
->nr_pages
== nr_pages
)
3061 atomic_inc(&event
->buffer
->refcount
);
3067 user_extra
= nr_pages
+ 1;
3068 user_lock_limit
= sysctl_perf_event_mlock
>> (PAGE_SHIFT
- 10);
3071 * Increase the limit linearly with more CPUs:
3073 user_lock_limit
*= num_online_cpus();
3075 user_locked
= atomic_long_read(&user
->locked_vm
) + user_extra
;
3078 if (user_locked
> user_lock_limit
)
3079 extra
= user_locked
- user_lock_limit
;
3081 lock_limit
= rlimit(RLIMIT_MEMLOCK
);
3082 lock_limit
>>= PAGE_SHIFT
;
3083 locked
= vma
->vm_mm
->locked_vm
+ extra
;
3085 if ((locked
> lock_limit
) && perf_paranoid_tracepoint_raw() &&
3086 !capable(CAP_IPC_LOCK
)) {
3091 WARN_ON(event
->buffer
);
3093 if (vma
->vm_flags
& VM_WRITE
)
3094 flags
|= PERF_BUFFER_WRITABLE
;
3096 buffer
= perf_buffer_alloc(nr_pages
, event
->attr
.wakeup_watermark
,
3102 rcu_assign_pointer(event
->buffer
, buffer
);
3104 atomic_long_add(user_extra
, &user
->locked_vm
);
3105 event
->mmap_locked
= extra
;
3106 event
->mmap_user
= get_current_user();
3107 vma
->vm_mm
->locked_vm
+= event
->mmap_locked
;
3111 atomic_inc(&event
->mmap_count
);
3112 mutex_unlock(&event
->mmap_mutex
);
3114 vma
->vm_flags
|= VM_RESERVED
;
3115 vma
->vm_ops
= &perf_mmap_vmops
;
3120 static int perf_fasync(int fd
, struct file
*filp
, int on
)
3122 struct inode
*inode
= filp
->f_path
.dentry
->d_inode
;
3123 struct perf_event
*event
= filp
->private_data
;
3126 mutex_lock(&inode
->i_mutex
);
3127 retval
= fasync_helper(fd
, filp
, on
, &event
->fasync
);
3128 mutex_unlock(&inode
->i_mutex
);
3136 static const struct file_operations perf_fops
= {
3137 .llseek
= no_llseek
,
3138 .release
= perf_release
,
3141 .unlocked_ioctl
= perf_ioctl
,
3142 .compat_ioctl
= perf_ioctl
,
3144 .fasync
= perf_fasync
,
3150 * If there's data, ensure we set the poll() state and publish everything
3151 * to user-space before waking everybody up.
3154 void perf_event_wakeup(struct perf_event
*event
)
3156 wake_up_all(&event
->waitq
);
3158 if (event
->pending_kill
) {
3159 kill_fasync(&event
->fasync
, SIGIO
, event
->pending_kill
);
3160 event
->pending_kill
= 0;
3164 static void perf_pending_event(struct irq_work
*entry
)
3166 struct perf_event
*event
= container_of(entry
,
3167 struct perf_event
, pending
);
3169 if (event
->pending_disable
) {
3170 event
->pending_disable
= 0;
3171 __perf_event_disable(event
);
3174 if (event
->pending_wakeup
) {
3175 event
->pending_wakeup
= 0;
3176 perf_event_wakeup(event
);
3181 * We assume there is only KVM supporting the callbacks.
3182 * Later on, we might change it to a list if there is
3183 * another virtualization implementation supporting the callbacks.
3185 struct perf_guest_info_callbacks
*perf_guest_cbs
;
3187 int perf_register_guest_info_callbacks(struct perf_guest_info_callbacks
*cbs
)
3189 perf_guest_cbs
= cbs
;
3192 EXPORT_SYMBOL_GPL(perf_register_guest_info_callbacks
);
3194 int perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks
*cbs
)
3196 perf_guest_cbs
= NULL
;
3199 EXPORT_SYMBOL_GPL(perf_unregister_guest_info_callbacks
);
3204 static bool perf_output_space(struct perf_buffer
*buffer
, unsigned long tail
,
3205 unsigned long offset
, unsigned long head
)
3209 if (!buffer
->writable
)
3212 mask
= perf_data_size(buffer
) - 1;
3214 offset
= (offset
- tail
) & mask
;
3215 head
= (head
- tail
) & mask
;
3217 if ((int)(head
- offset
) < 0)
3223 static void perf_output_wakeup(struct perf_output_handle
*handle
)
3225 atomic_set(&handle
->buffer
->poll
, POLL_IN
);
3228 handle
->event
->pending_wakeup
= 1;
3229 irq_work_queue(&handle
->event
->pending
);
3231 perf_event_wakeup(handle
->event
);
3235 * We need to ensure a later event_id doesn't publish a head when a former
3236 * event isn't done writing. However since we need to deal with NMIs we
3237 * cannot fully serialize things.
3239 * We only publish the head (and generate a wakeup) when the outer-most
3242 static void perf_output_get_handle(struct perf_output_handle
*handle
)
3244 struct perf_buffer
*buffer
= handle
->buffer
;
3247 local_inc(&buffer
->nest
);
3248 handle
->wakeup
= local_read(&buffer
->wakeup
);
3251 static void perf_output_put_handle(struct perf_output_handle
*handle
)
3253 struct perf_buffer
*buffer
= handle
->buffer
;
3257 head
= local_read(&buffer
->head
);
3260 * IRQ/NMI can happen here, which means we can miss a head update.
3263 if (!local_dec_and_test(&buffer
->nest
))
3267 * Publish the known good head. Rely on the full barrier implied
3268 * by atomic_dec_and_test() order the buffer->head read and this
3271 buffer
->user_page
->data_head
= head
;
3274 * Now check if we missed an update, rely on the (compiler)
3275 * barrier in atomic_dec_and_test() to re-read buffer->head.
3277 if (unlikely(head
!= local_read(&buffer
->head
))) {
3278 local_inc(&buffer
->nest
);
3282 if (handle
->wakeup
!= local_read(&buffer
->wakeup
))
3283 perf_output_wakeup(handle
);
3289 __always_inline
void perf_output_copy(struct perf_output_handle
*handle
,
3290 const void *buf
, unsigned int len
)
3293 unsigned long size
= min_t(unsigned long, handle
->size
, len
);
3295 memcpy(handle
->addr
, buf
, size
);
3298 handle
->addr
+= size
;
3300 handle
->size
-= size
;
3301 if (!handle
->size
) {
3302 struct perf_buffer
*buffer
= handle
->buffer
;
3305 handle
->page
&= buffer
->nr_pages
- 1;
3306 handle
->addr
= buffer
->data_pages
[handle
->page
];
3307 handle
->size
= PAGE_SIZE
<< page_order(buffer
);
3312 int perf_output_begin(struct perf_output_handle
*handle
,
3313 struct perf_event
*event
, unsigned int size
,
3314 int nmi
, int sample
)
3316 struct perf_buffer
*buffer
;
3317 unsigned long tail
, offset
, head
;
3320 struct perf_event_header header
;
3327 * For inherited events we send all the output towards the parent.
3330 event
= event
->parent
;
3332 buffer
= rcu_dereference(event
->buffer
);
3336 handle
->buffer
= buffer
;
3337 handle
->event
= event
;
3339 handle
->sample
= sample
;
3341 if (!buffer
->nr_pages
)
3344 have_lost
= local_read(&buffer
->lost
);
3346 size
+= sizeof(lost_event
);
3348 perf_output_get_handle(handle
);
3352 * Userspace could choose to issue a mb() before updating the
3353 * tail pointer. So that all reads will be completed before the
3356 tail
= ACCESS_ONCE(buffer
->user_page
->data_tail
);
3358 offset
= head
= local_read(&buffer
->head
);
3360 if (unlikely(!perf_output_space(buffer
, tail
, offset
, head
)))
3362 } while (local_cmpxchg(&buffer
->head
, offset
, head
) != offset
);
3364 if (head
- local_read(&buffer
->wakeup
) > buffer
->watermark
)
3365 local_add(buffer
->watermark
, &buffer
->wakeup
);
3367 handle
->page
= offset
>> (PAGE_SHIFT
+ page_order(buffer
));
3368 handle
->page
&= buffer
->nr_pages
- 1;
3369 handle
->size
= offset
& ((PAGE_SIZE
<< page_order(buffer
)) - 1);
3370 handle
->addr
= buffer
->data_pages
[handle
->page
];
3371 handle
->addr
+= handle
->size
;
3372 handle
->size
= (PAGE_SIZE
<< page_order(buffer
)) - handle
->size
;
3375 lost_event
.header
.type
= PERF_RECORD_LOST
;
3376 lost_event
.header
.misc
= 0;
3377 lost_event
.header
.size
= sizeof(lost_event
);
3378 lost_event
.id
= event
->id
;
3379 lost_event
.lost
= local_xchg(&buffer
->lost
, 0);
3381 perf_output_put(handle
, lost_event
);
3387 local_inc(&buffer
->lost
);
3388 perf_output_put_handle(handle
);
3395 void perf_output_end(struct perf_output_handle
*handle
)
3397 struct perf_event
*event
= handle
->event
;
3398 struct perf_buffer
*buffer
= handle
->buffer
;
3400 int wakeup_events
= event
->attr
.wakeup_events
;
3402 if (handle
->sample
&& wakeup_events
) {
3403 int events
= local_inc_return(&buffer
->events
);
3404 if (events
>= wakeup_events
) {
3405 local_sub(wakeup_events
, &buffer
->events
);
3406 local_inc(&buffer
->wakeup
);
3410 perf_output_put_handle(handle
);
3414 static u32
perf_event_pid(struct perf_event
*event
, struct task_struct
*p
)
3417 * only top level events have the pid namespace they were created in
3420 event
= event
->parent
;
3422 return task_tgid_nr_ns(p
, event
->ns
);
3425 static u32
perf_event_tid(struct perf_event
*event
, struct task_struct
*p
)
3428 * only top level events have the pid namespace they were created in
3431 event
= event
->parent
;
3433 return task_pid_nr_ns(p
, event
->ns
);
3436 static void perf_output_read_one(struct perf_output_handle
*handle
,
3437 struct perf_event
*event
)
3439 u64 read_format
= event
->attr
.read_format
;
3443 values
[n
++] = perf_event_count(event
);
3444 if (read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
) {
3445 values
[n
++] = event
->total_time_enabled
+
3446 atomic64_read(&event
->child_total_time_enabled
);
3448 if (read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
) {
3449 values
[n
++] = event
->total_time_running
+
3450 atomic64_read(&event
->child_total_time_running
);
3452 if (read_format
& PERF_FORMAT_ID
)
3453 values
[n
++] = primary_event_id(event
);
3455 perf_output_copy(handle
, values
, n
* sizeof(u64
));
3459 * XXX PERF_FORMAT_GROUP vs inherited events seems difficult.
3461 static void perf_output_read_group(struct perf_output_handle
*handle
,
3462 struct perf_event
*event
)
3464 struct perf_event
*leader
= event
->group_leader
, *sub
;
3465 u64 read_format
= event
->attr
.read_format
;
3469 values
[n
++] = 1 + leader
->nr_siblings
;
3471 if (read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
)
3472 values
[n
++] = leader
->total_time_enabled
;
3474 if (read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
)
3475 values
[n
++] = leader
->total_time_running
;
3477 if (leader
!= event
)
3478 leader
->pmu
->read(leader
);
3480 values
[n
++] = perf_event_count(leader
);
3481 if (read_format
& PERF_FORMAT_ID
)
3482 values
[n
++] = primary_event_id(leader
);
3484 perf_output_copy(handle
, values
, n
* sizeof(u64
));
3486 list_for_each_entry(sub
, &leader
->sibling_list
, group_entry
) {
3490 sub
->pmu
->read(sub
);
3492 values
[n
++] = perf_event_count(sub
);
3493 if (read_format
& PERF_FORMAT_ID
)
3494 values
[n
++] = primary_event_id(sub
);
3496 perf_output_copy(handle
, values
, n
* sizeof(u64
));
3500 static void perf_output_read(struct perf_output_handle
*handle
,
3501 struct perf_event
*event
)
3503 if (event
->attr
.read_format
& PERF_FORMAT_GROUP
)
3504 perf_output_read_group(handle
, event
);
3506 perf_output_read_one(handle
, event
);
3509 void perf_output_sample(struct perf_output_handle
*handle
,
3510 struct perf_event_header
*header
,
3511 struct perf_sample_data
*data
,
3512 struct perf_event
*event
)
3514 u64 sample_type
= data
->type
;
3516 perf_output_put(handle
, *header
);
3518 if (sample_type
& PERF_SAMPLE_IP
)
3519 perf_output_put(handle
, data
->ip
);
3521 if (sample_type
& PERF_SAMPLE_TID
)
3522 perf_output_put(handle
, data
->tid_entry
);
3524 if (sample_type
& PERF_SAMPLE_TIME
)
3525 perf_output_put(handle
, data
->time
);
3527 if (sample_type
& PERF_SAMPLE_ADDR
)
3528 perf_output_put(handle
, data
->addr
);
3530 if (sample_type
& PERF_SAMPLE_ID
)
3531 perf_output_put(handle
, data
->id
);
3533 if (sample_type
& PERF_SAMPLE_STREAM_ID
)
3534 perf_output_put(handle
, data
->stream_id
);
3536 if (sample_type
& PERF_SAMPLE_CPU
)
3537 perf_output_put(handle
, data
->cpu_entry
);
3539 if (sample_type
& PERF_SAMPLE_PERIOD
)
3540 perf_output_put(handle
, data
->period
);
3542 if (sample_type
& PERF_SAMPLE_READ
)
3543 perf_output_read(handle
, event
);
3545 if (sample_type
& PERF_SAMPLE_CALLCHAIN
) {
3546 if (data
->callchain
) {
3549 if (data
->callchain
)
3550 size
+= data
->callchain
->nr
;
3552 size
*= sizeof(u64
);
3554 perf_output_copy(handle
, data
->callchain
, size
);
3557 perf_output_put(handle
, nr
);
3561 if (sample_type
& PERF_SAMPLE_RAW
) {
3563 perf_output_put(handle
, data
->raw
->size
);
3564 perf_output_copy(handle
, data
->raw
->data
,
3571 .size
= sizeof(u32
),
3574 perf_output_put(handle
, raw
);
3579 void perf_prepare_sample(struct perf_event_header
*header
,
3580 struct perf_sample_data
*data
,
3581 struct perf_event
*event
,
3582 struct pt_regs
*regs
)
3584 u64 sample_type
= event
->attr
.sample_type
;
3586 data
->type
= sample_type
;
3588 header
->type
= PERF_RECORD_SAMPLE
;
3589 header
->size
= sizeof(*header
);
3592 header
->misc
|= perf_misc_flags(regs
);
3594 if (sample_type
& PERF_SAMPLE_IP
) {
3595 data
->ip
= perf_instruction_pointer(regs
);
3597 header
->size
+= sizeof(data
->ip
);
3600 if (sample_type
& PERF_SAMPLE_TID
) {
3601 /* namespace issues */
3602 data
->tid_entry
.pid
= perf_event_pid(event
, current
);
3603 data
->tid_entry
.tid
= perf_event_tid(event
, current
);
3605 header
->size
+= sizeof(data
->tid_entry
);
3608 if (sample_type
& PERF_SAMPLE_TIME
) {
3609 data
->time
= perf_clock();
3611 header
->size
+= sizeof(data
->time
);
3614 if (sample_type
& PERF_SAMPLE_ADDR
)
3615 header
->size
+= sizeof(data
->addr
);
3617 if (sample_type
& PERF_SAMPLE_ID
) {
3618 data
->id
= primary_event_id(event
);
3620 header
->size
+= sizeof(data
->id
);
3623 if (sample_type
& PERF_SAMPLE_STREAM_ID
) {
3624 data
->stream_id
= event
->id
;
3626 header
->size
+= sizeof(data
->stream_id
);
3629 if (sample_type
& PERF_SAMPLE_CPU
) {
3630 data
->cpu_entry
.cpu
= raw_smp_processor_id();
3631 data
->cpu_entry
.reserved
= 0;
3633 header
->size
+= sizeof(data
->cpu_entry
);
3636 if (sample_type
& PERF_SAMPLE_PERIOD
)
3637 header
->size
+= sizeof(data
->period
);
3639 if (sample_type
& PERF_SAMPLE_READ
)
3640 header
->size
+= perf_event_read_size(event
);
3642 if (sample_type
& PERF_SAMPLE_CALLCHAIN
) {
3645 data
->callchain
= perf_callchain(regs
);
3647 if (data
->callchain
)
3648 size
+= data
->callchain
->nr
;
3650 header
->size
+= size
* sizeof(u64
);
3653 if (sample_type
& PERF_SAMPLE_RAW
) {
3654 int size
= sizeof(u32
);
3657 size
+= data
->raw
->size
;
3659 size
+= sizeof(u32
);
3661 WARN_ON_ONCE(size
& (sizeof(u64
)-1));
3662 header
->size
+= size
;
3666 static void perf_event_output(struct perf_event
*event
, int nmi
,
3667 struct perf_sample_data
*data
,
3668 struct pt_regs
*regs
)
3670 struct perf_output_handle handle
;
3671 struct perf_event_header header
;
3673 /* protect the callchain buffers */
3676 perf_prepare_sample(&header
, data
, event
, regs
);
3678 if (perf_output_begin(&handle
, event
, header
.size
, nmi
, 1))
3681 perf_output_sample(&handle
, &header
, data
, event
);
3683 perf_output_end(&handle
);
3693 struct perf_read_event
{
3694 struct perf_event_header header
;
3701 perf_event_read_event(struct perf_event
*event
,
3702 struct task_struct
*task
)
3704 struct perf_output_handle handle
;
3705 struct perf_read_event read_event
= {
3707 .type
= PERF_RECORD_READ
,
3709 .size
= sizeof(read_event
) + perf_event_read_size(event
),
3711 .pid
= perf_event_pid(event
, task
),
3712 .tid
= perf_event_tid(event
, task
),
3716 ret
= perf_output_begin(&handle
, event
, read_event
.header
.size
, 0, 0);
3720 perf_output_put(&handle
, read_event
);
3721 perf_output_read(&handle
, event
);
3723 perf_output_end(&handle
);
3727 * task tracking -- fork/exit
3729 * enabled by: attr.comm | attr.mmap | attr.mmap_data | attr.task
3732 struct perf_task_event
{
3733 struct task_struct
*task
;
3734 struct perf_event_context
*task_ctx
;
3737 struct perf_event_header header
;
3747 static void perf_event_task_output(struct perf_event
*event
,
3748 struct perf_task_event
*task_event
)
3750 struct perf_output_handle handle
;
3751 struct task_struct
*task
= task_event
->task
;
3754 size
= task_event
->event_id
.header
.size
;
3755 ret
= perf_output_begin(&handle
, event
, size
, 0, 0);
3760 task_event
->event_id
.pid
= perf_event_pid(event
, task
);
3761 task_event
->event_id
.ppid
= perf_event_pid(event
, current
);
3763 task_event
->event_id
.tid
= perf_event_tid(event
, task
);
3764 task_event
->event_id
.ptid
= perf_event_tid(event
, current
);
3766 perf_output_put(&handle
, task_event
->event_id
);
3768 perf_output_end(&handle
);
3771 static int perf_event_task_match(struct perf_event
*event
)
3773 if (event
->state
< PERF_EVENT_STATE_INACTIVE
)
3776 if (event
->cpu
!= -1 && event
->cpu
!= smp_processor_id())
3779 if (event
->attr
.comm
|| event
->attr
.mmap
||
3780 event
->attr
.mmap_data
|| event
->attr
.task
)
3786 static void perf_event_task_ctx(struct perf_event_context
*ctx
,
3787 struct perf_task_event
*task_event
)
3789 struct perf_event
*event
;
3791 list_for_each_entry_rcu(event
, &ctx
->event_list
, event_entry
) {
3792 if (perf_event_task_match(event
))
3793 perf_event_task_output(event
, task_event
);
3797 static void perf_event_task_event(struct perf_task_event
*task_event
)
3799 struct perf_cpu_context
*cpuctx
;
3800 struct perf_event_context
*ctx
;
3805 list_for_each_entry_rcu(pmu
, &pmus
, entry
) {
3806 cpuctx
= get_cpu_ptr(pmu
->pmu_cpu_context
);
3807 perf_event_task_ctx(&cpuctx
->ctx
, task_event
);
3809 ctx
= task_event
->task_ctx
;
3811 ctxn
= pmu
->task_ctx_nr
;
3814 ctx
= rcu_dereference(current
->perf_event_ctxp
[ctxn
]);
3817 perf_event_task_ctx(ctx
, task_event
);
3819 put_cpu_ptr(pmu
->pmu_cpu_context
);
3824 static void perf_event_task(struct task_struct
*task
,
3825 struct perf_event_context
*task_ctx
,
3828 struct perf_task_event task_event
;
3830 if (!atomic_read(&nr_comm_events
) &&
3831 !atomic_read(&nr_mmap_events
) &&
3832 !atomic_read(&nr_task_events
))
3835 task_event
= (struct perf_task_event
){
3837 .task_ctx
= task_ctx
,
3840 .type
= new ? PERF_RECORD_FORK
: PERF_RECORD_EXIT
,
3842 .size
= sizeof(task_event
.event_id
),
3848 .time
= perf_clock(),
3852 perf_event_task_event(&task_event
);
3855 void perf_event_fork(struct task_struct
*task
)
3857 perf_event_task(task
, NULL
, 1);
3864 struct perf_comm_event
{
3865 struct task_struct
*task
;
3870 struct perf_event_header header
;
3877 static void perf_event_comm_output(struct perf_event
*event
,
3878 struct perf_comm_event
*comm_event
)
3880 struct perf_output_handle handle
;
3881 int size
= comm_event
->event_id
.header
.size
;
3882 int ret
= perf_output_begin(&handle
, event
, size
, 0, 0);
3887 comm_event
->event_id
.pid
= perf_event_pid(event
, comm_event
->task
);
3888 comm_event
->event_id
.tid
= perf_event_tid(event
, comm_event
->task
);
3890 perf_output_put(&handle
, comm_event
->event_id
);
3891 perf_output_copy(&handle
, comm_event
->comm
,
3892 comm_event
->comm_size
);
3893 perf_output_end(&handle
);
3896 static int perf_event_comm_match(struct perf_event
*event
)
3898 if (event
->state
< PERF_EVENT_STATE_INACTIVE
)
3901 if (event
->cpu
!= -1 && event
->cpu
!= smp_processor_id())
3904 if (event
->attr
.comm
)
3910 static void perf_event_comm_ctx(struct perf_event_context
*ctx
,
3911 struct perf_comm_event
*comm_event
)
3913 struct perf_event
*event
;
3915 list_for_each_entry_rcu(event
, &ctx
->event_list
, event_entry
) {
3916 if (perf_event_comm_match(event
))
3917 perf_event_comm_output(event
, comm_event
);
3921 static void perf_event_comm_event(struct perf_comm_event
*comm_event
)
3923 struct perf_cpu_context
*cpuctx
;
3924 struct perf_event_context
*ctx
;
3925 char comm
[TASK_COMM_LEN
];
3930 memset(comm
, 0, sizeof(comm
));
3931 strlcpy(comm
, comm_event
->task
->comm
, sizeof(comm
));
3932 size
= ALIGN(strlen(comm
)+1, sizeof(u64
));
3934 comm_event
->comm
= comm
;
3935 comm_event
->comm_size
= size
;
3937 comm_event
->event_id
.header
.size
= sizeof(comm_event
->event_id
) + size
;
3940 list_for_each_entry_rcu(pmu
, &pmus
, entry
) {
3941 cpuctx
= get_cpu_ptr(pmu
->pmu_cpu_context
);
3942 perf_event_comm_ctx(&cpuctx
->ctx
, comm_event
);
3944 ctxn
= pmu
->task_ctx_nr
;
3948 ctx
= rcu_dereference(current
->perf_event_ctxp
[ctxn
]);
3950 perf_event_comm_ctx(ctx
, comm_event
);
3952 put_cpu_ptr(pmu
->pmu_cpu_context
);
3957 void perf_event_comm(struct task_struct
*task
)
3959 struct perf_comm_event comm_event
;
3960 struct perf_event_context
*ctx
;
3963 for_each_task_context_nr(ctxn
) {
3964 ctx
= task
->perf_event_ctxp
[ctxn
];
3968 perf_event_enable_on_exec(ctx
);
3971 if (!atomic_read(&nr_comm_events
))
3974 comm_event
= (struct perf_comm_event
){
3980 .type
= PERF_RECORD_COMM
,
3989 perf_event_comm_event(&comm_event
);
3996 struct perf_mmap_event
{
3997 struct vm_area_struct
*vma
;
3999 const char *file_name
;
4003 struct perf_event_header header
;
4013 static void perf_event_mmap_output(struct perf_event
*event
,
4014 struct perf_mmap_event
*mmap_event
)
4016 struct perf_output_handle handle
;
4017 int size
= mmap_event
->event_id
.header
.size
;
4018 int ret
= perf_output_begin(&handle
, event
, size
, 0, 0);
4023 mmap_event
->event_id
.pid
= perf_event_pid(event
, current
);
4024 mmap_event
->event_id
.tid
= perf_event_tid(event
, current
);
4026 perf_output_put(&handle
, mmap_event
->event_id
);
4027 perf_output_copy(&handle
, mmap_event
->file_name
,
4028 mmap_event
->file_size
);
4029 perf_output_end(&handle
);
4032 static int perf_event_mmap_match(struct perf_event
*event
,
4033 struct perf_mmap_event
*mmap_event
,
4036 if (event
->state
< PERF_EVENT_STATE_INACTIVE
)
4039 if (event
->cpu
!= -1 && event
->cpu
!= smp_processor_id())
4042 if ((!executable
&& event
->attr
.mmap_data
) ||
4043 (executable
&& event
->attr
.mmap
))
4049 static void perf_event_mmap_ctx(struct perf_event_context
*ctx
,
4050 struct perf_mmap_event
*mmap_event
,
4053 struct perf_event
*event
;
4055 list_for_each_entry_rcu(event
, &ctx
->event_list
, event_entry
) {
4056 if (perf_event_mmap_match(event
, mmap_event
, executable
))
4057 perf_event_mmap_output(event
, mmap_event
);
4061 static void perf_event_mmap_event(struct perf_mmap_event
*mmap_event
)
4063 struct perf_cpu_context
*cpuctx
;
4064 struct perf_event_context
*ctx
;
4065 struct vm_area_struct
*vma
= mmap_event
->vma
;
4066 struct file
*file
= vma
->vm_file
;
4074 memset(tmp
, 0, sizeof(tmp
));
4078 * d_path works from the end of the buffer backwards, so we
4079 * need to add enough zero bytes after the string to handle
4080 * the 64bit alignment we do later.
4082 buf
= kzalloc(PATH_MAX
+ sizeof(u64
), GFP_KERNEL
);
4084 name
= strncpy(tmp
, "//enomem", sizeof(tmp
));
4087 name
= d_path(&file
->f_path
, buf
, PATH_MAX
);
4089 name
= strncpy(tmp
, "//toolong", sizeof(tmp
));
4093 if (arch_vma_name(mmap_event
->vma
)) {
4094 name
= strncpy(tmp
, arch_vma_name(mmap_event
->vma
),
4100 name
= strncpy(tmp
, "[vdso]", sizeof(tmp
));
4102 } else if (vma
->vm_start
<= vma
->vm_mm
->start_brk
&&
4103 vma
->vm_end
>= vma
->vm_mm
->brk
) {
4104 name
= strncpy(tmp
, "[heap]", sizeof(tmp
));
4106 } else if (vma
->vm_start
<= vma
->vm_mm
->start_stack
&&
4107 vma
->vm_end
>= vma
->vm_mm
->start_stack
) {
4108 name
= strncpy(tmp
, "[stack]", sizeof(tmp
));
4112 name
= strncpy(tmp
, "//anon", sizeof(tmp
));
4117 size
= ALIGN(strlen(name
)+1, sizeof(u64
));
4119 mmap_event
->file_name
= name
;
4120 mmap_event
->file_size
= size
;
4122 mmap_event
->event_id
.header
.size
= sizeof(mmap_event
->event_id
) + size
;
4125 list_for_each_entry_rcu(pmu
, &pmus
, entry
) {
4126 cpuctx
= get_cpu_ptr(pmu
->pmu_cpu_context
);
4127 perf_event_mmap_ctx(&cpuctx
->ctx
, mmap_event
,
4128 vma
->vm_flags
& VM_EXEC
);
4130 ctxn
= pmu
->task_ctx_nr
;
4134 ctx
= rcu_dereference(current
->perf_event_ctxp
[ctxn
]);
4136 perf_event_mmap_ctx(ctx
, mmap_event
,
4137 vma
->vm_flags
& VM_EXEC
);
4140 put_cpu_ptr(pmu
->pmu_cpu_context
);
4147 void perf_event_mmap(struct vm_area_struct
*vma
)
4149 struct perf_mmap_event mmap_event
;
4151 if (!atomic_read(&nr_mmap_events
))
4154 mmap_event
= (struct perf_mmap_event
){
4160 .type
= PERF_RECORD_MMAP
,
4161 .misc
= PERF_RECORD_MISC_USER
,
4166 .start
= vma
->vm_start
,
4167 .len
= vma
->vm_end
- vma
->vm_start
,
4168 .pgoff
= (u64
)vma
->vm_pgoff
<< PAGE_SHIFT
,
4172 perf_event_mmap_event(&mmap_event
);
4176 * IRQ throttle logging
4179 static void perf_log_throttle(struct perf_event
*event
, int enable
)
4181 struct perf_output_handle handle
;
4185 struct perf_event_header header
;
4189 } throttle_event
= {
4191 .type
= PERF_RECORD_THROTTLE
,
4193 .size
= sizeof(throttle_event
),
4195 .time
= perf_clock(),
4196 .id
= primary_event_id(event
),
4197 .stream_id
= event
->id
,
4201 throttle_event
.header
.type
= PERF_RECORD_UNTHROTTLE
;
4203 ret
= perf_output_begin(&handle
, event
, sizeof(throttle_event
), 1, 0);
4207 perf_output_put(&handle
, throttle_event
);
4208 perf_output_end(&handle
);
4212 * Generic event overflow handling, sampling.
4215 static int __perf_event_overflow(struct perf_event
*event
, int nmi
,
4216 int throttle
, struct perf_sample_data
*data
,
4217 struct pt_regs
*regs
)
4219 int events
= atomic_read(&event
->event_limit
);
4220 struct hw_perf_event
*hwc
= &event
->hw
;
4226 if (hwc
->interrupts
!= MAX_INTERRUPTS
) {
4228 if (HZ
* hwc
->interrupts
>
4229 (u64
)sysctl_perf_event_sample_rate
) {
4230 hwc
->interrupts
= MAX_INTERRUPTS
;
4231 perf_log_throttle(event
, 0);
4236 * Keep re-disabling events even though on the previous
4237 * pass we disabled it - just in case we raced with a
4238 * sched-in and the event got enabled again:
4244 if (event
->attr
.freq
) {
4245 u64 now
= perf_clock();
4246 s64 delta
= now
- hwc
->freq_time_stamp
;
4248 hwc
->freq_time_stamp
= now
;
4250 if (delta
> 0 && delta
< 2*TICK_NSEC
)
4251 perf_adjust_period(event
, delta
, hwc
->last_period
);
4255 * XXX event_limit might not quite work as expected on inherited
4259 event
->pending_kill
= POLL_IN
;
4260 if (events
&& atomic_dec_and_test(&event
->event_limit
)) {
4262 event
->pending_kill
= POLL_HUP
;
4264 event
->pending_disable
= 1;
4265 irq_work_queue(&event
->pending
);
4267 perf_event_disable(event
);
4270 if (event
->overflow_handler
)
4271 event
->overflow_handler(event
, nmi
, data
, regs
);
4273 perf_event_output(event
, nmi
, data
, regs
);
4278 int perf_event_overflow(struct perf_event
*event
, int nmi
,
4279 struct perf_sample_data
*data
,
4280 struct pt_regs
*regs
)
4282 return __perf_event_overflow(event
, nmi
, 1, data
, regs
);
4286 * Generic software event infrastructure
4289 struct swevent_htable
{
4290 struct swevent_hlist
*swevent_hlist
;
4291 struct mutex hlist_mutex
;
4294 /* Recursion avoidance in each contexts */
4295 int recursion
[PERF_NR_CONTEXTS
];
4298 static DEFINE_PER_CPU(struct swevent_htable
, swevent_htable
);
4301 * We directly increment event->count and keep a second value in
4302 * event->hw.period_left to count intervals. This period event
4303 * is kept in the range [-sample_period, 0] so that we can use the
4307 static u64
perf_swevent_set_period(struct perf_event
*event
)
4309 struct hw_perf_event
*hwc
= &event
->hw
;
4310 u64 period
= hwc
->last_period
;
4314 hwc
->last_period
= hwc
->sample_period
;
4317 old
= val
= local64_read(&hwc
->period_left
);
4321 nr
= div64_u64(period
+ val
, period
);
4322 offset
= nr
* period
;
4324 if (local64_cmpxchg(&hwc
->period_left
, old
, val
) != old
)
4330 static void perf_swevent_overflow(struct perf_event
*event
, u64 overflow
,
4331 int nmi
, struct perf_sample_data
*data
,
4332 struct pt_regs
*regs
)
4334 struct hw_perf_event
*hwc
= &event
->hw
;
4337 data
->period
= event
->hw
.last_period
;
4339 overflow
= perf_swevent_set_period(event
);
4341 if (hwc
->interrupts
== MAX_INTERRUPTS
)
4344 for (; overflow
; overflow
--) {
4345 if (__perf_event_overflow(event
, nmi
, throttle
,
4348 * We inhibit the overflow from happening when
4349 * hwc->interrupts == MAX_INTERRUPTS.
4357 static void perf_swevent_event(struct perf_event
*event
, u64 nr
,
4358 int nmi
, struct perf_sample_data
*data
,
4359 struct pt_regs
*regs
)
4361 struct hw_perf_event
*hwc
= &event
->hw
;
4363 local64_add(nr
, &event
->count
);
4368 if (!hwc
->sample_period
)
4371 if (nr
== 1 && hwc
->sample_period
== 1 && !event
->attr
.freq
)
4372 return perf_swevent_overflow(event
, 1, nmi
, data
, regs
);
4374 if (local64_add_negative(nr
, &hwc
->period_left
))
4377 perf_swevent_overflow(event
, 0, nmi
, data
, regs
);
4380 static int perf_exclude_event(struct perf_event
*event
,
4381 struct pt_regs
*regs
)
4383 if (event
->hw
.state
& PERF_HES_STOPPED
)
4387 if (event
->attr
.exclude_user
&& user_mode(regs
))
4390 if (event
->attr
.exclude_kernel
&& !user_mode(regs
))
4397 static int perf_swevent_match(struct perf_event
*event
,
4398 enum perf_type_id type
,
4400 struct perf_sample_data
*data
,
4401 struct pt_regs
*regs
)
4403 if (event
->attr
.type
!= type
)
4406 if (event
->attr
.config
!= event_id
)
4409 if (perf_exclude_event(event
, regs
))
4415 static inline u64
swevent_hash(u64 type
, u32 event_id
)
4417 u64 val
= event_id
| (type
<< 32);
4419 return hash_64(val
, SWEVENT_HLIST_BITS
);
4422 static inline struct hlist_head
*
4423 __find_swevent_head(struct swevent_hlist
*hlist
, u64 type
, u32 event_id
)
4425 u64 hash
= swevent_hash(type
, event_id
);
4427 return &hlist
->heads
[hash
];
4430 /* For the read side: events when they trigger */
4431 static inline struct hlist_head
*
4432 find_swevent_head_rcu(struct swevent_htable
*swhash
, u64 type
, u32 event_id
)
4434 struct swevent_hlist
*hlist
;
4436 hlist
= rcu_dereference(swhash
->swevent_hlist
);
4440 return __find_swevent_head(hlist
, type
, event_id
);
4443 /* For the event head insertion and removal in the hlist */
4444 static inline struct hlist_head
*
4445 find_swevent_head(struct swevent_htable
*swhash
, struct perf_event
*event
)
4447 struct swevent_hlist
*hlist
;
4448 u32 event_id
= event
->attr
.config
;
4449 u64 type
= event
->attr
.type
;
4452 * Event scheduling is always serialized against hlist allocation
4453 * and release. Which makes the protected version suitable here.
4454 * The context lock guarantees that.
4456 hlist
= rcu_dereference_protected(swhash
->swevent_hlist
,
4457 lockdep_is_held(&event
->ctx
->lock
));
4461 return __find_swevent_head(hlist
, type
, event_id
);
4464 static void do_perf_sw_event(enum perf_type_id type
, u32 event_id
,
4466 struct perf_sample_data
*data
,
4467 struct pt_regs
*regs
)
4469 struct swevent_htable
*swhash
= &__get_cpu_var(swevent_htable
);
4470 struct perf_event
*event
;
4471 struct hlist_node
*node
;
4472 struct hlist_head
*head
;
4475 head
= find_swevent_head_rcu(swhash
, type
, event_id
);
4479 hlist_for_each_entry_rcu(event
, node
, head
, hlist_entry
) {
4480 if (perf_swevent_match(event
, type
, event_id
, data
, regs
))
4481 perf_swevent_event(event
, nr
, nmi
, data
, regs
);
4487 int perf_swevent_get_recursion_context(void)
4489 struct swevent_htable
*swhash
= &__get_cpu_var(swevent_htable
);
4491 return get_recursion_context(swhash
->recursion
);
4493 EXPORT_SYMBOL_GPL(perf_swevent_get_recursion_context
);
4495 void inline perf_swevent_put_recursion_context(int rctx
)
4497 struct swevent_htable
*swhash
= &__get_cpu_var(swevent_htable
);
4499 put_recursion_context(swhash
->recursion
, rctx
);
4502 void __perf_sw_event(u32 event_id
, u64 nr
, int nmi
,
4503 struct pt_regs
*regs
, u64 addr
)
4505 struct perf_sample_data data
;
4508 preempt_disable_notrace();
4509 rctx
= perf_swevent_get_recursion_context();
4513 perf_sample_data_init(&data
, addr
);
4515 do_perf_sw_event(PERF_TYPE_SOFTWARE
, event_id
, nr
, nmi
, &data
, regs
);
4517 perf_swevent_put_recursion_context(rctx
);
4518 preempt_enable_notrace();
4521 static void perf_swevent_read(struct perf_event
*event
)
4525 static int perf_swevent_add(struct perf_event
*event
, int flags
)
4527 struct swevent_htable
*swhash
= &__get_cpu_var(swevent_htable
);
4528 struct hw_perf_event
*hwc
= &event
->hw
;
4529 struct hlist_head
*head
;
4531 if (hwc
->sample_period
) {
4532 hwc
->last_period
= hwc
->sample_period
;
4533 perf_swevent_set_period(event
);
4536 hwc
->state
= !(flags
& PERF_EF_START
);
4538 head
= find_swevent_head(swhash
, event
);
4539 if (WARN_ON_ONCE(!head
))
4542 hlist_add_head_rcu(&event
->hlist_entry
, head
);
4547 static void perf_swevent_del(struct perf_event
*event
, int flags
)
4549 hlist_del_rcu(&event
->hlist_entry
);
4552 static void perf_swevent_start(struct perf_event
*event
, int flags
)
4554 event
->hw
.state
= 0;
4557 static void perf_swevent_stop(struct perf_event
*event
, int flags
)
4559 event
->hw
.state
= PERF_HES_STOPPED
;
4562 /* Deref the hlist from the update side */
4563 static inline struct swevent_hlist
*
4564 swevent_hlist_deref(struct swevent_htable
*swhash
)
4566 return rcu_dereference_protected(swhash
->swevent_hlist
,
4567 lockdep_is_held(&swhash
->hlist_mutex
));
4570 static void swevent_hlist_release_rcu(struct rcu_head
*rcu_head
)
4572 struct swevent_hlist
*hlist
;
4574 hlist
= container_of(rcu_head
, struct swevent_hlist
, rcu_head
);
4578 static void swevent_hlist_release(struct swevent_htable
*swhash
)
4580 struct swevent_hlist
*hlist
= swevent_hlist_deref(swhash
);
4585 rcu_assign_pointer(swhash
->swevent_hlist
, NULL
);
4586 call_rcu(&hlist
->rcu_head
, swevent_hlist_release_rcu
);
4589 static void swevent_hlist_put_cpu(struct perf_event
*event
, int cpu
)
4591 struct swevent_htable
*swhash
= &per_cpu(swevent_htable
, cpu
);
4593 mutex_lock(&swhash
->hlist_mutex
);
4595 if (!--swhash
->hlist_refcount
)
4596 swevent_hlist_release(swhash
);
4598 mutex_unlock(&swhash
->hlist_mutex
);
4601 static void swevent_hlist_put(struct perf_event
*event
)
4605 if (event
->cpu
!= -1) {
4606 swevent_hlist_put_cpu(event
, event
->cpu
);
4610 for_each_possible_cpu(cpu
)
4611 swevent_hlist_put_cpu(event
, cpu
);
4614 static int swevent_hlist_get_cpu(struct perf_event
*event
, int cpu
)
4616 struct swevent_htable
*swhash
= &per_cpu(swevent_htable
, cpu
);
4619 mutex_lock(&swhash
->hlist_mutex
);
4621 if (!swevent_hlist_deref(swhash
) && cpu_online(cpu
)) {
4622 struct swevent_hlist
*hlist
;
4624 hlist
= kzalloc(sizeof(*hlist
), GFP_KERNEL
);
4629 rcu_assign_pointer(swhash
->swevent_hlist
, hlist
);
4631 swhash
->hlist_refcount
++;
4633 mutex_unlock(&swhash
->hlist_mutex
);
4638 static int swevent_hlist_get(struct perf_event
*event
)
4641 int cpu
, failed_cpu
;
4643 if (event
->cpu
!= -1)
4644 return swevent_hlist_get_cpu(event
, event
->cpu
);
4647 for_each_possible_cpu(cpu
) {
4648 err
= swevent_hlist_get_cpu(event
, cpu
);
4658 for_each_possible_cpu(cpu
) {
4659 if (cpu
== failed_cpu
)
4661 swevent_hlist_put_cpu(event
, cpu
);
4668 atomic_t perf_swevent_enabled
[PERF_COUNT_SW_MAX
];
4670 static void sw_perf_event_destroy(struct perf_event
*event
)
4672 u64 event_id
= event
->attr
.config
;
4674 WARN_ON(event
->parent
);
4676 atomic_dec(&perf_swevent_enabled
[event_id
]);
4677 swevent_hlist_put(event
);
4680 static int perf_swevent_init(struct perf_event
*event
)
4682 int event_id
= event
->attr
.config
;
4684 if (event
->attr
.type
!= PERF_TYPE_SOFTWARE
)
4688 case PERF_COUNT_SW_CPU_CLOCK
:
4689 case PERF_COUNT_SW_TASK_CLOCK
:
4696 if (event_id
> PERF_COUNT_SW_MAX
)
4699 if (!event
->parent
) {
4702 err
= swevent_hlist_get(event
);
4706 atomic_inc(&perf_swevent_enabled
[event_id
]);
4707 event
->destroy
= sw_perf_event_destroy
;
4713 static struct pmu perf_swevent
= {
4714 .task_ctx_nr
= perf_sw_context
,
4716 .event_init
= perf_swevent_init
,
4717 .add
= perf_swevent_add
,
4718 .del
= perf_swevent_del
,
4719 .start
= perf_swevent_start
,
4720 .stop
= perf_swevent_stop
,
4721 .read
= perf_swevent_read
,
4724 #ifdef CONFIG_EVENT_TRACING
4726 static int perf_tp_filter_match(struct perf_event
*event
,
4727 struct perf_sample_data
*data
)
4729 void *record
= data
->raw
->data
;
4731 if (likely(!event
->filter
) || filter_match_preds(event
->filter
, record
))
4736 static int perf_tp_event_match(struct perf_event
*event
,
4737 struct perf_sample_data
*data
,
4738 struct pt_regs
*regs
)
4741 * All tracepoints are from kernel-space.
4743 if (event
->attr
.exclude_kernel
)
4746 if (!perf_tp_filter_match(event
, data
))
4752 void perf_tp_event(u64 addr
, u64 count
, void *record
, int entry_size
,
4753 struct pt_regs
*regs
, struct hlist_head
*head
, int rctx
)
4755 struct perf_sample_data data
;
4756 struct perf_event
*event
;
4757 struct hlist_node
*node
;
4759 struct perf_raw_record raw
= {
4764 perf_sample_data_init(&data
, addr
);
4767 hlist_for_each_entry_rcu(event
, node
, head
, hlist_entry
) {
4768 if (perf_tp_event_match(event
, &data
, regs
))
4769 perf_swevent_event(event
, count
, 1, &data
, regs
);
4772 perf_swevent_put_recursion_context(rctx
);
4774 EXPORT_SYMBOL_GPL(perf_tp_event
);
4776 static void tp_perf_event_destroy(struct perf_event
*event
)
4778 perf_trace_destroy(event
);
4781 static int perf_tp_event_init(struct perf_event
*event
)
4785 if (event
->attr
.type
!= PERF_TYPE_TRACEPOINT
)
4789 * Raw tracepoint data is a severe data leak, only allow root to
4792 if ((event
->attr
.sample_type
& PERF_SAMPLE_RAW
) &&
4793 perf_paranoid_tracepoint_raw() &&
4794 !capable(CAP_SYS_ADMIN
))
4797 err
= perf_trace_init(event
);
4801 event
->destroy
= tp_perf_event_destroy
;
4806 static struct pmu perf_tracepoint
= {
4807 .task_ctx_nr
= perf_sw_context
,
4809 .event_init
= perf_tp_event_init
,
4810 .add
= perf_trace_add
,
4811 .del
= perf_trace_del
,
4812 .start
= perf_swevent_start
,
4813 .stop
= perf_swevent_stop
,
4814 .read
= perf_swevent_read
,
4817 static inline void perf_tp_register(void)
4819 perf_pmu_register(&perf_tracepoint
);
4822 static int perf_event_set_filter(struct perf_event
*event
, void __user
*arg
)
4827 if (event
->attr
.type
!= PERF_TYPE_TRACEPOINT
)
4830 filter_str
= strndup_user(arg
, PAGE_SIZE
);
4831 if (IS_ERR(filter_str
))
4832 return PTR_ERR(filter_str
);
4834 ret
= ftrace_profile_set_filter(event
, event
->attr
.config
, filter_str
);
4840 static void perf_event_free_filter(struct perf_event
*event
)
4842 ftrace_profile_free_filter(event
);
4847 static inline void perf_tp_register(void)
4851 static int perf_event_set_filter(struct perf_event
*event
, void __user
*arg
)
4856 static void perf_event_free_filter(struct perf_event
*event
)
4860 #endif /* CONFIG_EVENT_TRACING */
4862 #ifdef CONFIG_HAVE_HW_BREAKPOINT
4863 void perf_bp_event(struct perf_event
*bp
, void *data
)
4865 struct perf_sample_data sample
;
4866 struct pt_regs
*regs
= data
;
4868 perf_sample_data_init(&sample
, bp
->attr
.bp_addr
);
4870 if (!bp
->hw
.state
&& !perf_exclude_event(bp
, regs
))
4871 perf_swevent_event(bp
, 1, 1, &sample
, regs
);
4876 * hrtimer based swevent callback
4879 static enum hrtimer_restart
perf_swevent_hrtimer(struct hrtimer
*hrtimer
)
4881 enum hrtimer_restart ret
= HRTIMER_RESTART
;
4882 struct perf_sample_data data
;
4883 struct pt_regs
*regs
;
4884 struct perf_event
*event
;
4887 event
= container_of(hrtimer
, struct perf_event
, hw
.hrtimer
);
4888 event
->pmu
->read(event
);
4890 perf_sample_data_init(&data
, 0);
4891 data
.period
= event
->hw
.last_period
;
4892 regs
= get_irq_regs();
4894 if (regs
&& !perf_exclude_event(event
, regs
)) {
4895 if (!(event
->attr
.exclude_idle
&& current
->pid
== 0))
4896 if (perf_event_overflow(event
, 0, &data
, regs
))
4897 ret
= HRTIMER_NORESTART
;
4900 period
= max_t(u64
, 10000, event
->hw
.sample_period
);
4901 hrtimer_forward_now(hrtimer
, ns_to_ktime(period
));
4906 static void perf_swevent_start_hrtimer(struct perf_event
*event
)
4908 struct hw_perf_event
*hwc
= &event
->hw
;
4910 hrtimer_init(&hwc
->hrtimer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
4911 hwc
->hrtimer
.function
= perf_swevent_hrtimer
;
4912 if (hwc
->sample_period
) {
4913 s64 period
= local64_read(&hwc
->period_left
);
4919 local64_set(&hwc
->period_left
, 0);
4921 period
= max_t(u64
, 10000, hwc
->sample_period
);
4923 __hrtimer_start_range_ns(&hwc
->hrtimer
,
4924 ns_to_ktime(period
), 0,
4925 HRTIMER_MODE_REL_PINNED
, 0);
4929 static void perf_swevent_cancel_hrtimer(struct perf_event
*event
)
4931 struct hw_perf_event
*hwc
= &event
->hw
;
4933 if (hwc
->sample_period
) {
4934 ktime_t remaining
= hrtimer_get_remaining(&hwc
->hrtimer
);
4935 local64_set(&hwc
->period_left
, ktime_to_ns(remaining
));
4937 hrtimer_cancel(&hwc
->hrtimer
);
4942 * Software event: cpu wall time clock
4945 static void cpu_clock_event_update(struct perf_event
*event
)
4950 now
= local_clock();
4951 prev
= local64_xchg(&event
->hw
.prev_count
, now
);
4952 local64_add(now
- prev
, &event
->count
);
4955 static void cpu_clock_event_start(struct perf_event
*event
, int flags
)
4957 local64_set(&event
->hw
.prev_count
, local_clock());
4958 perf_swevent_start_hrtimer(event
);
4961 static void cpu_clock_event_stop(struct perf_event
*event
, int flags
)
4963 perf_swevent_cancel_hrtimer(event
);
4964 cpu_clock_event_update(event
);
4967 static int cpu_clock_event_add(struct perf_event
*event
, int flags
)
4969 if (flags
& PERF_EF_START
)
4970 cpu_clock_event_start(event
, flags
);
4975 static void cpu_clock_event_del(struct perf_event
*event
, int flags
)
4977 cpu_clock_event_stop(event
, flags
);
4980 static void cpu_clock_event_read(struct perf_event
*event
)
4982 cpu_clock_event_update(event
);
4985 static int cpu_clock_event_init(struct perf_event
*event
)
4987 if (event
->attr
.type
!= PERF_TYPE_SOFTWARE
)
4990 if (event
->attr
.config
!= PERF_COUNT_SW_CPU_CLOCK
)
4996 static struct pmu perf_cpu_clock
= {
4997 .task_ctx_nr
= perf_sw_context
,
4999 .event_init
= cpu_clock_event_init
,
5000 .add
= cpu_clock_event_add
,
5001 .del
= cpu_clock_event_del
,
5002 .start
= cpu_clock_event_start
,
5003 .stop
= cpu_clock_event_stop
,
5004 .read
= cpu_clock_event_read
,
5008 * Software event: task time clock
5011 static void task_clock_event_update(struct perf_event
*event
, u64 now
)
5016 prev
= local64_xchg(&event
->hw
.prev_count
, now
);
5018 local64_add(delta
, &event
->count
);
5021 static void task_clock_event_start(struct perf_event
*event
, int flags
)
5023 local64_set(&event
->hw
.prev_count
, event
->ctx
->time
);
5024 perf_swevent_start_hrtimer(event
);
5027 static void task_clock_event_stop(struct perf_event
*event
, int flags
)
5029 perf_swevent_cancel_hrtimer(event
);
5030 task_clock_event_update(event
, event
->ctx
->time
);
5033 static int task_clock_event_add(struct perf_event
*event
, int flags
)
5035 if (flags
& PERF_EF_START
)
5036 task_clock_event_start(event
, flags
);
5041 static void task_clock_event_del(struct perf_event
*event
, int flags
)
5043 task_clock_event_stop(event
, PERF_EF_UPDATE
);
5046 static void task_clock_event_read(struct perf_event
*event
)
5051 update_context_time(event
->ctx
);
5052 time
= event
->ctx
->time
;
5054 u64 now
= perf_clock();
5055 u64 delta
= now
- event
->ctx
->timestamp
;
5056 time
= event
->ctx
->time
+ delta
;
5059 task_clock_event_update(event
, time
);
5062 static int task_clock_event_init(struct perf_event
*event
)
5064 if (event
->attr
.type
!= PERF_TYPE_SOFTWARE
)
5067 if (event
->attr
.config
!= PERF_COUNT_SW_TASK_CLOCK
)
5073 static struct pmu perf_task_clock
= {
5074 .task_ctx_nr
= perf_sw_context
,
5076 .event_init
= task_clock_event_init
,
5077 .add
= task_clock_event_add
,
5078 .del
= task_clock_event_del
,
5079 .start
= task_clock_event_start
,
5080 .stop
= task_clock_event_stop
,
5081 .read
= task_clock_event_read
,
5084 static void perf_pmu_nop_void(struct pmu
*pmu
)
5088 static int perf_pmu_nop_int(struct pmu
*pmu
)
5093 static void perf_pmu_start_txn(struct pmu
*pmu
)
5095 perf_pmu_disable(pmu
);
5098 static int perf_pmu_commit_txn(struct pmu
*pmu
)
5100 perf_pmu_enable(pmu
);
5104 static void perf_pmu_cancel_txn(struct pmu
*pmu
)
5106 perf_pmu_enable(pmu
);
5110 * Ensures all contexts with the same task_ctx_nr have the same
5111 * pmu_cpu_context too.
5113 static void *find_pmu_context(int ctxn
)
5120 list_for_each_entry(pmu
, &pmus
, entry
) {
5121 if (pmu
->task_ctx_nr
== ctxn
)
5122 return pmu
->pmu_cpu_context
;
5128 static void free_pmu_context(void * __percpu cpu_context
)
5132 mutex_lock(&pmus_lock
);
5134 * Like a real lame refcount.
5136 list_for_each_entry(pmu
, &pmus
, entry
) {
5137 if (pmu
->pmu_cpu_context
== cpu_context
)
5141 free_percpu(cpu_context
);
5143 mutex_unlock(&pmus_lock
);
5146 int perf_pmu_register(struct pmu
*pmu
)
5150 mutex_lock(&pmus_lock
);
5152 pmu
->pmu_disable_count
= alloc_percpu(int);
5153 if (!pmu
->pmu_disable_count
)
5156 pmu
->pmu_cpu_context
= find_pmu_context(pmu
->task_ctx_nr
);
5157 if (pmu
->pmu_cpu_context
)
5158 goto got_cpu_context
;
5160 pmu
->pmu_cpu_context
= alloc_percpu(struct perf_cpu_context
);
5161 if (!pmu
->pmu_cpu_context
)
5164 for_each_possible_cpu(cpu
) {
5165 struct perf_cpu_context
*cpuctx
;
5167 cpuctx
= per_cpu_ptr(pmu
->pmu_cpu_context
, cpu
);
5168 __perf_event_init_context(&cpuctx
->ctx
);
5169 cpuctx
->ctx
.type
= cpu_context
;
5170 cpuctx
->ctx
.pmu
= pmu
;
5171 cpuctx
->jiffies_interval
= 1;
5172 INIT_LIST_HEAD(&cpuctx
->rotation_list
);
5176 if (!pmu
->start_txn
) {
5177 if (pmu
->pmu_enable
) {
5179 * If we have pmu_enable/pmu_disable calls, install
5180 * transaction stubs that use that to try and batch
5181 * hardware accesses.
5183 pmu
->start_txn
= perf_pmu_start_txn
;
5184 pmu
->commit_txn
= perf_pmu_commit_txn
;
5185 pmu
->cancel_txn
= perf_pmu_cancel_txn
;
5187 pmu
->start_txn
= perf_pmu_nop_void
;
5188 pmu
->commit_txn
= perf_pmu_nop_int
;
5189 pmu
->cancel_txn
= perf_pmu_nop_void
;
5193 if (!pmu
->pmu_enable
) {
5194 pmu
->pmu_enable
= perf_pmu_nop_void
;
5195 pmu
->pmu_disable
= perf_pmu_nop_void
;
5198 list_add_rcu(&pmu
->entry
, &pmus
);
5201 mutex_unlock(&pmus_lock
);
5206 free_percpu(pmu
->pmu_disable_count
);
5210 void perf_pmu_unregister(struct pmu
*pmu
)
5212 mutex_lock(&pmus_lock
);
5213 list_del_rcu(&pmu
->entry
);
5214 mutex_unlock(&pmus_lock
);
5217 * We dereference the pmu list under both SRCU and regular RCU, so
5218 * synchronize against both of those.
5220 synchronize_srcu(&pmus_srcu
);
5223 free_percpu(pmu
->pmu_disable_count
);
5224 free_pmu_context(pmu
->pmu_cpu_context
);
5227 struct pmu
*perf_init_event(struct perf_event
*event
)
5229 struct pmu
*pmu
= NULL
;
5232 idx
= srcu_read_lock(&pmus_srcu
);
5233 list_for_each_entry_rcu(pmu
, &pmus
, entry
) {
5234 int ret
= pmu
->event_init(event
);
5238 if (ret
!= -ENOENT
) {
5243 pmu
= ERR_PTR(-ENOENT
);
5245 srcu_read_unlock(&pmus_srcu
, idx
);
5251 * Allocate and initialize a event structure
5253 static struct perf_event
*
5254 perf_event_alloc(struct perf_event_attr
*attr
, int cpu
,
5255 struct perf_event
*group_leader
,
5256 struct perf_event
*parent_event
,
5257 perf_overflow_handler_t overflow_handler
)
5260 struct perf_event
*event
;
5261 struct hw_perf_event
*hwc
;
5264 event
= kzalloc(sizeof(*event
), GFP_KERNEL
);
5266 return ERR_PTR(-ENOMEM
);
5269 * Single events are their own group leaders, with an
5270 * empty sibling list:
5273 group_leader
= event
;
5275 mutex_init(&event
->child_mutex
);
5276 INIT_LIST_HEAD(&event
->child_list
);
5278 INIT_LIST_HEAD(&event
->group_entry
);
5279 INIT_LIST_HEAD(&event
->event_entry
);
5280 INIT_LIST_HEAD(&event
->sibling_list
);
5281 init_waitqueue_head(&event
->waitq
);
5282 init_irq_work(&event
->pending
, perf_pending_event
);
5284 mutex_init(&event
->mmap_mutex
);
5287 event
->attr
= *attr
;
5288 event
->group_leader
= group_leader
;
5292 event
->parent
= parent_event
;
5294 event
->ns
= get_pid_ns(current
->nsproxy
->pid_ns
);
5295 event
->id
= atomic64_inc_return(&perf_event_id
);
5297 event
->state
= PERF_EVENT_STATE_INACTIVE
;
5299 if (!overflow_handler
&& parent_event
)
5300 overflow_handler
= parent_event
->overflow_handler
;
5302 event
->overflow_handler
= overflow_handler
;
5305 event
->state
= PERF_EVENT_STATE_OFF
;
5310 hwc
->sample_period
= attr
->sample_period
;
5311 if (attr
->freq
&& attr
->sample_freq
)
5312 hwc
->sample_period
= 1;
5313 hwc
->last_period
= hwc
->sample_period
;
5315 local64_set(&hwc
->period_left
, hwc
->sample_period
);
5318 * we currently do not support PERF_FORMAT_GROUP on inherited events
5320 if (attr
->inherit
&& (attr
->read_format
& PERF_FORMAT_GROUP
))
5323 pmu
= perf_init_event(event
);
5329 else if (IS_ERR(pmu
))
5334 put_pid_ns(event
->ns
);
5336 return ERR_PTR(err
);
5341 if (!event
->parent
) {
5342 atomic_inc(&nr_events
);
5343 if (event
->attr
.mmap
|| event
->attr
.mmap_data
)
5344 atomic_inc(&nr_mmap_events
);
5345 if (event
->attr
.comm
)
5346 atomic_inc(&nr_comm_events
);
5347 if (event
->attr
.task
)
5348 atomic_inc(&nr_task_events
);
5349 if (event
->attr
.sample_type
& PERF_SAMPLE_CALLCHAIN
) {
5350 err
= get_callchain_buffers();
5353 return ERR_PTR(err
);
5361 static int perf_copy_attr(struct perf_event_attr __user
*uattr
,
5362 struct perf_event_attr
*attr
)
5367 if (!access_ok(VERIFY_WRITE
, uattr
, PERF_ATTR_SIZE_VER0
))
5371 * zero the full structure, so that a short copy will be nice.
5373 memset(attr
, 0, sizeof(*attr
));
5375 ret
= get_user(size
, &uattr
->size
);
5379 if (size
> PAGE_SIZE
) /* silly large */
5382 if (!size
) /* abi compat */
5383 size
= PERF_ATTR_SIZE_VER0
;
5385 if (size
< PERF_ATTR_SIZE_VER0
)
5389 * If we're handed a bigger struct than we know of,
5390 * ensure all the unknown bits are 0 - i.e. new
5391 * user-space does not rely on any kernel feature
5392 * extensions we dont know about yet.
5394 if (size
> sizeof(*attr
)) {
5395 unsigned char __user
*addr
;
5396 unsigned char __user
*end
;
5399 addr
= (void __user
*)uattr
+ sizeof(*attr
);
5400 end
= (void __user
*)uattr
+ size
;
5402 for (; addr
< end
; addr
++) {
5403 ret
= get_user(val
, addr
);
5409 size
= sizeof(*attr
);
5412 ret
= copy_from_user(attr
, uattr
, size
);
5417 * If the type exists, the corresponding creation will verify
5420 if (attr
->type
>= PERF_TYPE_MAX
)
5423 if (attr
->__reserved_1
)
5426 if (attr
->sample_type
& ~(PERF_SAMPLE_MAX
-1))
5429 if (attr
->read_format
& ~(PERF_FORMAT_MAX
-1))
5436 put_user(sizeof(*attr
), &uattr
->size
);
5442 perf_event_set_output(struct perf_event
*event
, struct perf_event
*output_event
)
5444 struct perf_buffer
*buffer
= NULL
, *old_buffer
= NULL
;
5450 /* don't allow circular references */
5451 if (event
== output_event
)
5455 * Don't allow cross-cpu buffers
5457 if (output_event
->cpu
!= event
->cpu
)
5461 * If its not a per-cpu buffer, it must be the same task.
5463 if (output_event
->cpu
== -1 && output_event
->ctx
!= event
->ctx
)
5467 mutex_lock(&event
->mmap_mutex
);
5468 /* Can't redirect output if we've got an active mmap() */
5469 if (atomic_read(&event
->mmap_count
))
5473 /* get the buffer we want to redirect to */
5474 buffer
= perf_buffer_get(output_event
);
5479 old_buffer
= event
->buffer
;
5480 rcu_assign_pointer(event
->buffer
, buffer
);
5483 mutex_unlock(&event
->mmap_mutex
);
5486 perf_buffer_put(old_buffer
);
5492 * sys_perf_event_open - open a performance event, associate it to a task/cpu
5494 * @attr_uptr: event_id type attributes for monitoring/sampling
5497 * @group_fd: group leader event fd
5499 SYSCALL_DEFINE5(perf_event_open
,
5500 struct perf_event_attr __user
*, attr_uptr
,
5501 pid_t
, pid
, int, cpu
, int, group_fd
, unsigned long, flags
)
5503 struct perf_event
*group_leader
= NULL
, *output_event
= NULL
;
5504 struct perf_event
*event
, *sibling
;
5505 struct perf_event_attr attr
;
5506 struct perf_event_context
*ctx
;
5507 struct file
*event_file
= NULL
;
5508 struct file
*group_file
= NULL
;
5509 struct task_struct
*task
= NULL
;
5513 int fput_needed
= 0;
5516 /* for future expandability... */
5517 if (flags
& ~(PERF_FLAG_FD_NO_GROUP
| PERF_FLAG_FD_OUTPUT
))
5520 err
= perf_copy_attr(attr_uptr
, &attr
);
5524 if (!attr
.exclude_kernel
) {
5525 if (perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN
))
5530 if (attr
.sample_freq
> sysctl_perf_event_sample_rate
)
5534 event_fd
= get_unused_fd_flags(O_RDWR
);
5538 if (group_fd
!= -1) {
5539 group_leader
= perf_fget_light(group_fd
, &fput_needed
);
5540 if (IS_ERR(group_leader
)) {
5541 err
= PTR_ERR(group_leader
);
5544 group_file
= group_leader
->filp
;
5545 if (flags
& PERF_FLAG_FD_OUTPUT
)
5546 output_event
= group_leader
;
5547 if (flags
& PERF_FLAG_FD_NO_GROUP
)
5548 group_leader
= NULL
;
5551 event
= perf_event_alloc(&attr
, cpu
, group_leader
, NULL
, NULL
);
5552 if (IS_ERR(event
)) {
5553 err
= PTR_ERR(event
);
5558 * Special case software events and allow them to be part of
5559 * any hardware group.
5564 (is_software_event(event
) != is_software_event(group_leader
))) {
5565 if (is_software_event(event
)) {
5567 * If event and group_leader are not both a software
5568 * event, and event is, then group leader is not.
5570 * Allow the addition of software events to !software
5571 * groups, this is safe because software events never
5574 pmu
= group_leader
->pmu
;
5575 } else if (is_software_event(group_leader
) &&
5576 (group_leader
->group_flags
& PERF_GROUP_SOFTWARE
)) {
5578 * In case the group is a pure software group, and we
5579 * try to add a hardware event, move the whole group to
5580 * the hardware context.
5587 task
= find_lively_task_by_vpid(pid
);
5589 err
= PTR_ERR(task
);
5595 * Get the target context (task or percpu):
5597 ctx
= find_get_context(pmu
, task
, cpu
);
5604 * Look up the group leader (we will attach this event to it):
5610 * Do not allow a recursive hierarchy (this new sibling
5611 * becoming part of another group-sibling):
5613 if (group_leader
->group_leader
!= group_leader
)
5616 * Do not allow to attach to a group in a different
5617 * task or CPU context:
5620 if (group_leader
->ctx
->type
!= ctx
->type
)
5623 if (group_leader
->ctx
!= ctx
)
5628 * Only a group leader can be exclusive or pinned
5630 if (attr
.exclusive
|| attr
.pinned
)
5635 err
= perf_event_set_output(event
, output_event
);
5640 event_file
= anon_inode_getfile("[perf_event]", &perf_fops
, event
, O_RDWR
);
5641 if (IS_ERR(event_file
)) {
5642 err
= PTR_ERR(event_file
);
5647 struct perf_event_context
*gctx
= group_leader
->ctx
;
5649 mutex_lock(&gctx
->mutex
);
5650 perf_event_remove_from_context(group_leader
);
5651 list_for_each_entry(sibling
, &group_leader
->sibling_list
,
5653 perf_event_remove_from_context(sibling
);
5656 mutex_unlock(&gctx
->mutex
);
5660 event
->filp
= event_file
;
5661 WARN_ON_ONCE(ctx
->parent_ctx
);
5662 mutex_lock(&ctx
->mutex
);
5665 perf_install_in_context(ctx
, group_leader
, cpu
);
5667 list_for_each_entry(sibling
, &group_leader
->sibling_list
,
5669 perf_install_in_context(ctx
, sibling
, cpu
);
5674 perf_install_in_context(ctx
, event
, cpu
);
5676 mutex_unlock(&ctx
->mutex
);
5678 event
->owner
= current
;
5679 get_task_struct(current
);
5680 mutex_lock(¤t
->perf_event_mutex
);
5681 list_add_tail(&event
->owner_entry
, ¤t
->perf_event_list
);
5682 mutex_unlock(¤t
->perf_event_mutex
);
5685 * Drop the reference on the group_event after placing the
5686 * new event on the sibling_list. This ensures destruction
5687 * of the group leader will find the pointer to itself in
5688 * perf_group_detach().
5690 fput_light(group_file
, fput_needed
);
5691 fd_install(event_fd
, event_file
);
5697 fput_light(group_file
, fput_needed
);
5700 put_unused_fd(event_fd
);
5705 * perf_event_create_kernel_counter
5707 * @attr: attributes of the counter to create
5708 * @cpu: cpu in which the counter is bound
5709 * @task: task to profile (NULL for percpu)
5712 perf_event_create_kernel_counter(struct perf_event_attr
*attr
, int cpu
,
5713 struct task_struct
*task
,
5714 perf_overflow_handler_t overflow_handler
)
5716 struct perf_event_context
*ctx
;
5717 struct perf_event
*event
;
5721 * Get the target context (task or percpu):
5724 event
= perf_event_alloc(attr
, cpu
, NULL
, NULL
, overflow_handler
);
5725 if (IS_ERR(event
)) {
5726 err
= PTR_ERR(event
);
5730 ctx
= find_get_context(event
->pmu
, task
, cpu
);
5737 WARN_ON_ONCE(ctx
->parent_ctx
);
5738 mutex_lock(&ctx
->mutex
);
5739 perf_install_in_context(ctx
, event
, cpu
);
5741 mutex_unlock(&ctx
->mutex
);
5743 event
->owner
= current
;
5744 get_task_struct(current
);
5745 mutex_lock(¤t
->perf_event_mutex
);
5746 list_add_tail(&event
->owner_entry
, ¤t
->perf_event_list
);
5747 mutex_unlock(¤t
->perf_event_mutex
);
5754 return ERR_PTR(err
);
5756 EXPORT_SYMBOL_GPL(perf_event_create_kernel_counter
);
5758 static void sync_child_event(struct perf_event
*child_event
,
5759 struct task_struct
*child
)
5761 struct perf_event
*parent_event
= child_event
->parent
;
5764 if (child_event
->attr
.inherit_stat
)
5765 perf_event_read_event(child_event
, child
);
5767 child_val
= perf_event_count(child_event
);
5770 * Add back the child's count to the parent's count:
5772 atomic64_add(child_val
, &parent_event
->child_count
);
5773 atomic64_add(child_event
->total_time_enabled
,
5774 &parent_event
->child_total_time_enabled
);
5775 atomic64_add(child_event
->total_time_running
,
5776 &parent_event
->child_total_time_running
);
5779 * Remove this event from the parent's list
5781 WARN_ON_ONCE(parent_event
->ctx
->parent_ctx
);
5782 mutex_lock(&parent_event
->child_mutex
);
5783 list_del_init(&child_event
->child_list
);
5784 mutex_unlock(&parent_event
->child_mutex
);
5787 * Release the parent event, if this was the last
5790 fput(parent_event
->filp
);
5794 __perf_event_exit_task(struct perf_event
*child_event
,
5795 struct perf_event_context
*child_ctx
,
5796 struct task_struct
*child
)
5798 struct perf_event
*parent_event
;
5800 perf_event_remove_from_context(child_event
);
5802 parent_event
= child_event
->parent
;
5804 * It can happen that parent exits first, and has events
5805 * that are still around due to the child reference. These
5806 * events need to be zapped - but otherwise linger.
5809 sync_child_event(child_event
, child
);
5810 free_event(child_event
);
5814 static void perf_event_exit_task_context(struct task_struct
*child
, int ctxn
)
5816 struct perf_event
*child_event
, *tmp
;
5817 struct perf_event_context
*child_ctx
;
5818 unsigned long flags
;
5820 if (likely(!child
->perf_event_ctxp
[ctxn
])) {
5821 perf_event_task(child
, NULL
, 0);
5825 local_irq_save(flags
);
5827 * We can't reschedule here because interrupts are disabled,
5828 * and either child is current or it is a task that can't be
5829 * scheduled, so we are now safe from rescheduling changing
5832 child_ctx
= child
->perf_event_ctxp
[ctxn
];
5833 __perf_event_task_sched_out(child_ctx
);
5836 * Take the context lock here so that if find_get_context is
5837 * reading child->perf_event_ctxp, we wait until it has
5838 * incremented the context's refcount before we do put_ctx below.
5840 raw_spin_lock(&child_ctx
->lock
);
5841 child
->perf_event_ctxp
[ctxn
] = NULL
;
5843 * If this context is a clone; unclone it so it can't get
5844 * swapped to another process while we're removing all
5845 * the events from it.
5847 unclone_ctx(child_ctx
);
5848 update_context_time(child_ctx
);
5849 raw_spin_unlock_irqrestore(&child_ctx
->lock
, flags
);
5852 * Report the task dead after unscheduling the events so that we
5853 * won't get any samples after PERF_RECORD_EXIT. We can however still
5854 * get a few PERF_RECORD_READ events.
5856 perf_event_task(child
, child_ctx
, 0);
5859 * We can recurse on the same lock type through:
5861 * __perf_event_exit_task()
5862 * sync_child_event()
5863 * fput(parent_event->filp)
5865 * mutex_lock(&ctx->mutex)
5867 * But since its the parent context it won't be the same instance.
5869 mutex_lock(&child_ctx
->mutex
);
5872 list_for_each_entry_safe(child_event
, tmp
, &child_ctx
->pinned_groups
,
5874 __perf_event_exit_task(child_event
, child_ctx
, child
);
5876 list_for_each_entry_safe(child_event
, tmp
, &child_ctx
->flexible_groups
,
5878 __perf_event_exit_task(child_event
, child_ctx
, child
);
5881 * If the last event was a group event, it will have appended all
5882 * its siblings to the list, but we obtained 'tmp' before that which
5883 * will still point to the list head terminating the iteration.
5885 if (!list_empty(&child_ctx
->pinned_groups
) ||
5886 !list_empty(&child_ctx
->flexible_groups
))
5889 mutex_unlock(&child_ctx
->mutex
);
5895 * When a child task exits, feed back event values to parent events.
5897 void perf_event_exit_task(struct task_struct
*child
)
5901 for_each_task_context_nr(ctxn
)
5902 perf_event_exit_task_context(child
, ctxn
);
5905 static void perf_free_event(struct perf_event
*event
,
5906 struct perf_event_context
*ctx
)
5908 struct perf_event
*parent
= event
->parent
;
5910 if (WARN_ON_ONCE(!parent
))
5913 mutex_lock(&parent
->child_mutex
);
5914 list_del_init(&event
->child_list
);
5915 mutex_unlock(&parent
->child_mutex
);
5919 perf_group_detach(event
);
5920 list_del_event(event
, ctx
);
5925 * free an unexposed, unused context as created by inheritance by
5926 * perf_event_init_task below, used by fork() in case of fail.
5928 void perf_event_free_task(struct task_struct
*task
)
5930 struct perf_event_context
*ctx
;
5931 struct perf_event
*event
, *tmp
;
5934 for_each_task_context_nr(ctxn
) {
5935 ctx
= task
->perf_event_ctxp
[ctxn
];
5939 mutex_lock(&ctx
->mutex
);
5941 list_for_each_entry_safe(event
, tmp
, &ctx
->pinned_groups
,
5943 perf_free_event(event
, ctx
);
5945 list_for_each_entry_safe(event
, tmp
, &ctx
->flexible_groups
,
5947 perf_free_event(event
, ctx
);
5949 if (!list_empty(&ctx
->pinned_groups
) ||
5950 !list_empty(&ctx
->flexible_groups
))
5953 mutex_unlock(&ctx
->mutex
);
5959 void perf_event_delayed_put(struct task_struct
*task
)
5963 for_each_task_context_nr(ctxn
)
5964 WARN_ON_ONCE(task
->perf_event_ctxp
[ctxn
]);
5968 * inherit a event from parent task to child task:
5970 static struct perf_event
*
5971 inherit_event(struct perf_event
*parent_event
,
5972 struct task_struct
*parent
,
5973 struct perf_event_context
*parent_ctx
,
5974 struct task_struct
*child
,
5975 struct perf_event
*group_leader
,
5976 struct perf_event_context
*child_ctx
)
5978 struct perf_event
*child_event
;
5979 unsigned long flags
;
5982 * Instead of creating recursive hierarchies of events,
5983 * we link inherited events back to the original parent,
5984 * which has a filp for sure, which we use as the reference
5987 if (parent_event
->parent
)
5988 parent_event
= parent_event
->parent
;
5990 child_event
= perf_event_alloc(&parent_event
->attr
,
5992 group_leader
, parent_event
,
5994 if (IS_ERR(child_event
))
5999 * Make the child state follow the state of the parent event,
6000 * not its attr.disabled bit. We hold the parent's mutex,
6001 * so we won't race with perf_event_{en, dis}able_family.
6003 if (parent_event
->state
>= PERF_EVENT_STATE_INACTIVE
)
6004 child_event
->state
= PERF_EVENT_STATE_INACTIVE
;
6006 child_event
->state
= PERF_EVENT_STATE_OFF
;
6008 if (parent_event
->attr
.freq
) {
6009 u64 sample_period
= parent_event
->hw
.sample_period
;
6010 struct hw_perf_event
*hwc
= &child_event
->hw
;
6012 hwc
->sample_period
= sample_period
;
6013 hwc
->last_period
= sample_period
;
6015 local64_set(&hwc
->period_left
, sample_period
);
6018 child_event
->ctx
= child_ctx
;
6019 child_event
->overflow_handler
= parent_event
->overflow_handler
;
6022 * Link it up in the child's context:
6024 raw_spin_lock_irqsave(&child_ctx
->lock
, flags
);
6025 add_event_to_ctx(child_event
, child_ctx
);
6026 raw_spin_unlock_irqrestore(&child_ctx
->lock
, flags
);
6029 * Get a reference to the parent filp - we will fput it
6030 * when the child event exits. This is safe to do because
6031 * we are in the parent and we know that the filp still
6032 * exists and has a nonzero count:
6034 atomic_long_inc(&parent_event
->filp
->f_count
);
6037 * Link this into the parent event's child list
6039 WARN_ON_ONCE(parent_event
->ctx
->parent_ctx
);
6040 mutex_lock(&parent_event
->child_mutex
);
6041 list_add_tail(&child_event
->child_list
, &parent_event
->child_list
);
6042 mutex_unlock(&parent_event
->child_mutex
);
6047 static int inherit_group(struct perf_event
*parent_event
,
6048 struct task_struct
*parent
,
6049 struct perf_event_context
*parent_ctx
,
6050 struct task_struct
*child
,
6051 struct perf_event_context
*child_ctx
)
6053 struct perf_event
*leader
;
6054 struct perf_event
*sub
;
6055 struct perf_event
*child_ctr
;
6057 leader
= inherit_event(parent_event
, parent
, parent_ctx
,
6058 child
, NULL
, child_ctx
);
6060 return PTR_ERR(leader
);
6061 list_for_each_entry(sub
, &parent_event
->sibling_list
, group_entry
) {
6062 child_ctr
= inherit_event(sub
, parent
, parent_ctx
,
6063 child
, leader
, child_ctx
);
6064 if (IS_ERR(child_ctr
))
6065 return PTR_ERR(child_ctr
);
6071 inherit_task_group(struct perf_event
*event
, struct task_struct
*parent
,
6072 struct perf_event_context
*parent_ctx
,
6073 struct task_struct
*child
, int ctxn
,
6077 struct perf_event_context
*child_ctx
;
6079 if (!event
->attr
.inherit
) {
6084 child_ctx
= child
->perf_event_ctxp
[ctxn
];
6087 * This is executed from the parent task context, so
6088 * inherit events that have been marked for cloning.
6089 * First allocate and initialize a context for the
6093 child_ctx
= alloc_perf_context(event
->pmu
, child
);
6097 child
->perf_event_ctxp
[ctxn
] = child_ctx
;
6100 ret
= inherit_group(event
, parent
, parent_ctx
,
6110 * Initialize the perf_event context in task_struct
6112 int perf_event_init_context(struct task_struct
*child
, int ctxn
)
6114 struct perf_event_context
*child_ctx
, *parent_ctx
;
6115 struct perf_event_context
*cloned_ctx
;
6116 struct perf_event
*event
;
6117 struct task_struct
*parent
= current
;
6118 int inherited_all
= 1;
6121 child
->perf_event_ctxp
[ctxn
] = NULL
;
6123 mutex_init(&child
->perf_event_mutex
);
6124 INIT_LIST_HEAD(&child
->perf_event_list
);
6126 if (likely(!parent
->perf_event_ctxp
[ctxn
]))
6130 * If the parent's context is a clone, pin it so it won't get
6133 parent_ctx
= perf_pin_task_context(parent
, ctxn
);
6136 * No need to check if parent_ctx != NULL here; since we saw
6137 * it non-NULL earlier, the only reason for it to become NULL
6138 * is if we exit, and since we're currently in the middle of
6139 * a fork we can't be exiting at the same time.
6143 * Lock the parent list. No need to lock the child - not PID
6144 * hashed yet and not running, so nobody can access it.
6146 mutex_lock(&parent_ctx
->mutex
);
6149 * We dont have to disable NMIs - we are only looking at
6150 * the list, not manipulating it:
6152 list_for_each_entry(event
, &parent_ctx
->pinned_groups
, group_entry
) {
6153 ret
= inherit_task_group(event
, parent
, parent_ctx
,
6154 child
, ctxn
, &inherited_all
);
6159 list_for_each_entry(event
, &parent_ctx
->flexible_groups
, group_entry
) {
6160 ret
= inherit_task_group(event
, parent
, parent_ctx
,
6161 child
, ctxn
, &inherited_all
);
6166 child_ctx
= child
->perf_event_ctxp
[ctxn
];
6168 if (child_ctx
&& inherited_all
) {
6170 * Mark the child context as a clone of the parent
6171 * context, or of whatever the parent is a clone of.
6172 * Note that if the parent is a clone, it could get
6173 * uncloned at any point, but that doesn't matter
6174 * because the list of events and the generation
6175 * count can't have changed since we took the mutex.
6177 cloned_ctx
= rcu_dereference(parent_ctx
->parent_ctx
);
6179 child_ctx
->parent_ctx
= cloned_ctx
;
6180 child_ctx
->parent_gen
= parent_ctx
->parent_gen
;
6182 child_ctx
->parent_ctx
= parent_ctx
;
6183 child_ctx
->parent_gen
= parent_ctx
->generation
;
6185 get_ctx(child_ctx
->parent_ctx
);
6188 mutex_unlock(&parent_ctx
->mutex
);
6190 perf_unpin_context(parent_ctx
);
6196 * Initialize the perf_event context in task_struct
6198 int perf_event_init_task(struct task_struct
*child
)
6202 for_each_task_context_nr(ctxn
) {
6203 ret
= perf_event_init_context(child
, ctxn
);
6211 static void __init
perf_event_init_all_cpus(void)
6213 struct swevent_htable
*swhash
;
6216 for_each_possible_cpu(cpu
) {
6217 swhash
= &per_cpu(swevent_htable
, cpu
);
6218 mutex_init(&swhash
->hlist_mutex
);
6219 INIT_LIST_HEAD(&per_cpu(rotation_list
, cpu
));
6223 static void __cpuinit
perf_event_init_cpu(int cpu
)
6225 struct swevent_htable
*swhash
= &per_cpu(swevent_htable
, cpu
);
6227 mutex_lock(&swhash
->hlist_mutex
);
6228 if (swhash
->hlist_refcount
> 0) {
6229 struct swevent_hlist
*hlist
;
6231 hlist
= kzalloc_node(sizeof(*hlist
), GFP_KERNEL
, cpu_to_node(cpu
));
6233 rcu_assign_pointer(swhash
->swevent_hlist
, hlist
);
6235 mutex_unlock(&swhash
->hlist_mutex
);
6238 #ifdef CONFIG_HOTPLUG_CPU
6239 static void perf_pmu_rotate_stop(struct pmu
*pmu
)
6241 struct perf_cpu_context
*cpuctx
= this_cpu_ptr(pmu
->pmu_cpu_context
);
6243 WARN_ON(!irqs_disabled());
6245 list_del_init(&cpuctx
->rotation_list
);
6248 static void __perf_event_exit_context(void *__info
)
6250 struct perf_event_context
*ctx
= __info
;
6251 struct perf_event
*event
, *tmp
;
6253 perf_pmu_rotate_stop(ctx
->pmu
);
6255 list_for_each_entry_safe(event
, tmp
, &ctx
->pinned_groups
, group_entry
)
6256 __perf_event_remove_from_context(event
);
6257 list_for_each_entry_safe(event
, tmp
, &ctx
->flexible_groups
, group_entry
)
6258 __perf_event_remove_from_context(event
);
6261 static void perf_event_exit_cpu_context(int cpu
)
6263 struct perf_event_context
*ctx
;
6267 idx
= srcu_read_lock(&pmus_srcu
);
6268 list_for_each_entry_rcu(pmu
, &pmus
, entry
) {
6269 ctx
= &per_cpu_ptr(pmu
->pmu_cpu_context
, cpu
)->ctx
;
6271 mutex_lock(&ctx
->mutex
);
6272 smp_call_function_single(cpu
, __perf_event_exit_context
, ctx
, 1);
6273 mutex_unlock(&ctx
->mutex
);
6275 srcu_read_unlock(&pmus_srcu
, idx
);
6278 static void perf_event_exit_cpu(int cpu
)
6280 struct swevent_htable
*swhash
= &per_cpu(swevent_htable
, cpu
);
6282 mutex_lock(&swhash
->hlist_mutex
);
6283 swevent_hlist_release(swhash
);
6284 mutex_unlock(&swhash
->hlist_mutex
);
6286 perf_event_exit_cpu_context(cpu
);
6289 static inline void perf_event_exit_cpu(int cpu
) { }
6292 static int __cpuinit
6293 perf_cpu_notify(struct notifier_block
*self
, unsigned long action
, void *hcpu
)
6295 unsigned int cpu
= (long)hcpu
;
6297 switch (action
& ~CPU_TASKS_FROZEN
) {
6299 case CPU_UP_PREPARE
:
6300 case CPU_DOWN_FAILED
:
6301 perf_event_init_cpu(cpu
);
6304 case CPU_UP_CANCELED
:
6305 case CPU_DOWN_PREPARE
:
6306 perf_event_exit_cpu(cpu
);
6316 void __init
perf_event_init(void)
6318 perf_event_init_all_cpus();
6319 init_srcu_struct(&pmus_srcu
);
6320 perf_pmu_register(&perf_swevent
);
6321 perf_pmu_register(&perf_cpu_clock
);
6322 perf_pmu_register(&perf_task_clock
);
6324 perf_cpu_notifier(perf_cpu_notify
);