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
;
319 * We can have double attach due to group movement in perf_event_open.
321 if (event
->attach_state
& PERF_ATTACH_GROUP
)
324 event
->attach_state
|= PERF_ATTACH_GROUP
;
326 if (group_leader
== event
)
329 if (group_leader
->group_flags
& PERF_GROUP_SOFTWARE
&&
330 !is_software_event(event
))
331 group_leader
->group_flags
&= ~PERF_GROUP_SOFTWARE
;
333 list_add_tail(&event
->group_entry
, &group_leader
->sibling_list
);
334 group_leader
->nr_siblings
++;
338 * Remove a event from the lists for its context.
339 * Must be called with ctx->mutex and ctx->lock held.
342 list_del_event(struct perf_event
*event
, struct perf_event_context
*ctx
)
345 * We can have double detach due to exit/hot-unplug + close.
347 if (!(event
->attach_state
& PERF_ATTACH_CONTEXT
))
350 event
->attach_state
&= ~PERF_ATTACH_CONTEXT
;
353 if (event
->attr
.inherit_stat
)
356 list_del_rcu(&event
->event_entry
);
358 if (event
->group_leader
== event
)
359 list_del_init(&event
->group_entry
);
361 update_group_times(event
);
364 * If event was in error state, then keep it
365 * that way, otherwise bogus counts will be
366 * returned on read(). The only way to get out
367 * of error state is by explicit re-enabling
370 if (event
->state
> PERF_EVENT_STATE_OFF
)
371 event
->state
= PERF_EVENT_STATE_OFF
;
374 static void perf_group_detach(struct perf_event
*event
)
376 struct perf_event
*sibling
, *tmp
;
377 struct list_head
*list
= NULL
;
380 * We can have double detach due to exit/hot-unplug + close.
382 if (!(event
->attach_state
& PERF_ATTACH_GROUP
))
385 event
->attach_state
&= ~PERF_ATTACH_GROUP
;
388 * If this is a sibling, remove it from its group.
390 if (event
->group_leader
!= event
) {
391 list_del_init(&event
->group_entry
);
392 event
->group_leader
->nr_siblings
--;
396 if (!list_empty(&event
->group_entry
))
397 list
= &event
->group_entry
;
400 * If this was a group event with sibling events then
401 * upgrade the siblings to singleton events by adding them
402 * to whatever list we are on.
404 list_for_each_entry_safe(sibling
, tmp
, &event
->sibling_list
, group_entry
) {
406 list_move_tail(&sibling
->group_entry
, list
);
407 sibling
->group_leader
= sibling
;
409 /* Inherit group flags from the previous leader */
410 sibling
->group_flags
= event
->group_flags
;
415 event_filter_match(struct perf_event
*event
)
417 return event
->cpu
== -1 || event
->cpu
== smp_processor_id();
421 __event_sched_out(struct perf_event
*event
,
422 struct perf_cpu_context
*cpuctx
,
423 struct perf_event_context
*ctx
)
427 * An event which could not be activated because of
428 * filter mismatch still needs to have its timings
429 * maintained, otherwise bogus information is return
430 * via read() for time_enabled, time_running:
432 if (event
->state
== PERF_EVENT_STATE_INACTIVE
433 && !event_filter_match(event
)) {
434 delta
= ctx
->time
- event
->tstamp_stopped
;
435 event
->tstamp_running
+= delta
;
436 event
->tstamp_stopped
= ctx
->time
;
439 if (event
->state
!= PERF_EVENT_STATE_ACTIVE
)
442 event
->state
= PERF_EVENT_STATE_INACTIVE
;
443 if (event
->pending_disable
) {
444 event
->pending_disable
= 0;
445 event
->state
= PERF_EVENT_STATE_OFF
;
447 event
->pmu
->del(event
, 0);
450 if (!is_software_event(event
))
451 cpuctx
->active_oncpu
--;
453 if (event
->attr
.exclusive
|| !cpuctx
->active_oncpu
)
454 cpuctx
->exclusive
= 0;
459 event_sched_out(struct perf_event
*event
,
460 struct perf_cpu_context
*cpuctx
,
461 struct perf_event_context
*ctx
)
465 ret
= __event_sched_out(event
, cpuctx
, ctx
);
467 event
->tstamp_stopped
= ctx
->time
;
471 group_sched_out(struct perf_event
*group_event
,
472 struct perf_cpu_context
*cpuctx
,
473 struct perf_event_context
*ctx
)
475 struct perf_event
*event
;
476 int state
= group_event
->state
;
478 event_sched_out(group_event
, cpuctx
, ctx
);
481 * Schedule out siblings (if any):
483 list_for_each_entry(event
, &group_event
->sibling_list
, group_entry
)
484 event_sched_out(event
, cpuctx
, ctx
);
486 if (state
== PERF_EVENT_STATE_ACTIVE
&& group_event
->attr
.exclusive
)
487 cpuctx
->exclusive
= 0;
490 static inline struct perf_cpu_context
*
491 __get_cpu_context(struct perf_event_context
*ctx
)
493 return this_cpu_ptr(ctx
->pmu
->pmu_cpu_context
);
497 * Cross CPU call to remove a performance event
499 * We disable the event on the hardware level first. After that we
500 * remove it from the context list.
502 static void __perf_event_remove_from_context(void *info
)
504 struct perf_event
*event
= info
;
505 struct perf_event_context
*ctx
= event
->ctx
;
506 struct perf_cpu_context
*cpuctx
= __get_cpu_context(ctx
);
509 * If this is a task context, we need to check whether it is
510 * the current task context of this cpu. If not it has been
511 * scheduled out before the smp call arrived.
513 if (ctx
->task
&& cpuctx
->task_ctx
!= ctx
)
516 raw_spin_lock(&ctx
->lock
);
518 event_sched_out(event
, cpuctx
, ctx
);
520 list_del_event(event
, ctx
);
522 raw_spin_unlock(&ctx
->lock
);
527 * Remove the event from a task's (or a CPU's) list of events.
529 * Must be called with ctx->mutex held.
531 * CPU events are removed with a smp call. For task events we only
532 * call when the task is on a CPU.
534 * If event->ctx is a cloned context, callers must make sure that
535 * every task struct that event->ctx->task could possibly point to
536 * remains valid. This is OK when called from perf_release since
537 * that only calls us on the top-level context, which can't be a clone.
538 * When called from perf_event_exit_task, it's OK because the
539 * context has been detached from its task.
541 static void perf_event_remove_from_context(struct perf_event
*event
)
543 struct perf_event_context
*ctx
= event
->ctx
;
544 struct task_struct
*task
= ctx
->task
;
548 * Per cpu events are removed via an smp call and
549 * the removal is always successful.
551 smp_call_function_single(event
->cpu
,
552 __perf_event_remove_from_context
,
558 task_oncpu_function_call(task
, __perf_event_remove_from_context
,
561 raw_spin_lock_irq(&ctx
->lock
);
563 * If the context is active we need to retry the smp call.
565 if (ctx
->nr_active
&& !list_empty(&event
->group_entry
)) {
566 raw_spin_unlock_irq(&ctx
->lock
);
571 * The lock prevents that this context is scheduled in so we
572 * can remove the event safely, if the call above did not
575 if (!list_empty(&event
->group_entry
))
576 list_del_event(event
, ctx
);
577 raw_spin_unlock_irq(&ctx
->lock
);
581 * Cross CPU call to disable a performance event
583 static void __perf_event_disable(void *info
)
585 struct perf_event
*event
= info
;
586 struct perf_event_context
*ctx
= event
->ctx
;
587 struct perf_cpu_context
*cpuctx
= __get_cpu_context(ctx
);
590 * If this is a per-task event, need to check whether this
591 * event's task is the current task on this cpu.
593 if (ctx
->task
&& cpuctx
->task_ctx
!= ctx
)
596 raw_spin_lock(&ctx
->lock
);
599 * If the event is on, turn it off.
600 * If it is in error state, leave it in error state.
602 if (event
->state
>= PERF_EVENT_STATE_INACTIVE
) {
603 update_context_time(ctx
);
604 update_group_times(event
);
605 if (event
== event
->group_leader
)
606 group_sched_out(event
, cpuctx
, ctx
);
608 event_sched_out(event
, cpuctx
, ctx
);
609 event
->state
= PERF_EVENT_STATE_OFF
;
612 raw_spin_unlock(&ctx
->lock
);
618 * If event->ctx is a cloned context, callers must make sure that
619 * every task struct that event->ctx->task could possibly point to
620 * remains valid. This condition is satisifed when called through
621 * perf_event_for_each_child or perf_event_for_each because they
622 * hold the top-level event's child_mutex, so any descendant that
623 * goes to exit will block in sync_child_event.
624 * When called from perf_pending_event it's OK because event->ctx
625 * is the current context on this CPU and preemption is disabled,
626 * hence we can't get into perf_event_task_sched_out for this context.
628 void perf_event_disable(struct perf_event
*event
)
630 struct perf_event_context
*ctx
= event
->ctx
;
631 struct task_struct
*task
= ctx
->task
;
635 * Disable the event on the cpu that it's on
637 smp_call_function_single(event
->cpu
, __perf_event_disable
,
643 task_oncpu_function_call(task
, __perf_event_disable
, event
);
645 raw_spin_lock_irq(&ctx
->lock
);
647 * If the event is still active, we need to retry the cross-call.
649 if (event
->state
== PERF_EVENT_STATE_ACTIVE
) {
650 raw_spin_unlock_irq(&ctx
->lock
);
655 * Since we have the lock this context can't be scheduled
656 * in, so we can change the state safely.
658 if (event
->state
== PERF_EVENT_STATE_INACTIVE
) {
659 update_group_times(event
);
660 event
->state
= PERF_EVENT_STATE_OFF
;
663 raw_spin_unlock_irq(&ctx
->lock
);
667 __event_sched_in(struct perf_event
*event
,
668 struct perf_cpu_context
*cpuctx
,
669 struct perf_event_context
*ctx
)
671 if (event
->state
<= PERF_EVENT_STATE_OFF
)
674 event
->state
= PERF_EVENT_STATE_ACTIVE
;
675 event
->oncpu
= smp_processor_id();
677 * The new state must be visible before we turn it on in the hardware:
681 if (event
->pmu
->add(event
, PERF_EF_START
)) {
682 event
->state
= PERF_EVENT_STATE_INACTIVE
;
687 if (!is_software_event(event
))
688 cpuctx
->active_oncpu
++;
691 if (event
->attr
.exclusive
)
692 cpuctx
->exclusive
= 1;
698 event_sched_in(struct perf_event
*event
,
699 struct perf_cpu_context
*cpuctx
,
700 struct perf_event_context
*ctx
)
702 int ret
= __event_sched_in(event
, cpuctx
, ctx
);
705 event
->tstamp_running
+= ctx
->time
- event
->tstamp_stopped
;
710 group_commit_event_sched_in(struct perf_event
*group_event
,
711 struct perf_cpu_context
*cpuctx
,
712 struct perf_event_context
*ctx
)
714 struct perf_event
*event
;
717 group_event
->tstamp_running
+= now
- group_event
->tstamp_stopped
;
719 * Schedule in siblings as one group (if any):
721 list_for_each_entry(event
, &group_event
->sibling_list
, group_entry
) {
722 event
->tstamp_running
+= now
- event
->tstamp_stopped
;
727 group_sched_in(struct perf_event
*group_event
,
728 struct perf_cpu_context
*cpuctx
,
729 struct perf_event_context
*ctx
)
731 struct perf_event
*event
, *partial_group
= NULL
;
732 struct pmu
*pmu
= group_event
->pmu
;
734 if (group_event
->state
== PERF_EVENT_STATE_OFF
)
740 * use __event_sched_in() to delay updating tstamp_running
741 * until the transaction is committed. In case of failure
742 * we will keep an unmodified tstamp_running which is a
743 * requirement to get correct timing information
745 if (__event_sched_in(group_event
, cpuctx
, ctx
)) {
746 pmu
->cancel_txn(pmu
);
751 * Schedule in siblings as one group (if any):
753 list_for_each_entry(event
, &group_event
->sibling_list
, group_entry
) {
754 if (__event_sched_in(event
, cpuctx
, ctx
)) {
755 partial_group
= event
;
760 if (!pmu
->commit_txn(pmu
)) {
761 /* commit tstamp_running */
762 group_commit_event_sched_in(group_event
, cpuctx
, ctx
);
767 * Groups can be scheduled in as one unit only, so undo any
768 * partial group before returning:
770 * use __event_sched_out() to avoid updating tstamp_stopped
771 * because the event never actually ran
773 list_for_each_entry(event
, &group_event
->sibling_list
, group_entry
) {
774 if (event
== partial_group
)
776 __event_sched_out(event
, cpuctx
, ctx
);
778 __event_sched_out(group_event
, cpuctx
, ctx
);
780 pmu
->cancel_txn(pmu
);
786 * Work out whether we can put this event group on the CPU now.
788 static int group_can_go_on(struct perf_event
*event
,
789 struct perf_cpu_context
*cpuctx
,
793 * Groups consisting entirely of software events can always go on.
795 if (event
->group_flags
& PERF_GROUP_SOFTWARE
)
798 * If an exclusive group is already on, no other hardware
801 if (cpuctx
->exclusive
)
804 * If this group is exclusive and there are already
805 * events on the CPU, it can't go on.
807 if (event
->attr
.exclusive
&& cpuctx
->active_oncpu
)
810 * Otherwise, try to add it if all previous groups were able
816 static void add_event_to_ctx(struct perf_event
*event
,
817 struct perf_event_context
*ctx
)
819 list_add_event(event
, ctx
);
820 perf_group_attach(event
);
821 event
->tstamp_enabled
= ctx
->time
;
822 event
->tstamp_running
= ctx
->time
;
823 event
->tstamp_stopped
= ctx
->time
;
827 * Cross CPU call to install and enable a performance event
829 * Must be called with ctx->mutex held
831 static void __perf_install_in_context(void *info
)
833 struct perf_event
*event
= info
;
834 struct perf_event_context
*ctx
= event
->ctx
;
835 struct perf_event
*leader
= event
->group_leader
;
836 struct perf_cpu_context
*cpuctx
= __get_cpu_context(ctx
);
840 * If this is a task context, we need to check whether it is
841 * the current task context of this cpu. If not it has been
842 * scheduled out before the smp call arrived.
843 * Or possibly this is the right context but it isn't
844 * on this cpu because it had no events.
846 if (ctx
->task
&& cpuctx
->task_ctx
!= ctx
) {
847 if (cpuctx
->task_ctx
|| ctx
->task
!= current
)
849 cpuctx
->task_ctx
= ctx
;
852 raw_spin_lock(&ctx
->lock
);
854 update_context_time(ctx
);
856 add_event_to_ctx(event
, ctx
);
858 if (event
->cpu
!= -1 && event
->cpu
!= smp_processor_id())
862 * Don't put the event on if it is disabled or if
863 * it is in a group and the group isn't on.
865 if (event
->state
!= PERF_EVENT_STATE_INACTIVE
||
866 (leader
!= event
&& leader
->state
!= PERF_EVENT_STATE_ACTIVE
))
870 * An exclusive event can't go on if there are already active
871 * hardware events, and no hardware event can go on if there
872 * is already an exclusive event on.
874 if (!group_can_go_on(event
, cpuctx
, 1))
877 err
= event_sched_in(event
, cpuctx
, ctx
);
881 * This event couldn't go on. If it is in a group
882 * then we have to pull the whole group off.
883 * If the event group is pinned then put it in error state.
886 group_sched_out(leader
, cpuctx
, ctx
);
887 if (leader
->attr
.pinned
) {
888 update_group_times(leader
);
889 leader
->state
= PERF_EVENT_STATE_ERROR
;
894 raw_spin_unlock(&ctx
->lock
);
898 * Attach a performance event to a context
900 * First we add the event to the list with the hardware enable bit
901 * in event->hw_config cleared.
903 * If the event is attached to a task which is on a CPU we use a smp
904 * call to enable it in the task context. The task might have been
905 * scheduled away, but we check this in the smp call again.
907 * Must be called with ctx->mutex held.
910 perf_install_in_context(struct perf_event_context
*ctx
,
911 struct perf_event
*event
,
914 struct task_struct
*task
= ctx
->task
;
920 * Per cpu events are installed via an smp call and
921 * the install is always successful.
923 smp_call_function_single(cpu
, __perf_install_in_context
,
929 task_oncpu_function_call(task
, __perf_install_in_context
,
932 raw_spin_lock_irq(&ctx
->lock
);
934 * we need to retry the smp call.
936 if (ctx
->is_active
&& list_empty(&event
->group_entry
)) {
937 raw_spin_unlock_irq(&ctx
->lock
);
942 * The lock prevents that this context is scheduled in so we
943 * can add the event safely, if it the call above did not
946 if (list_empty(&event
->group_entry
))
947 add_event_to_ctx(event
, ctx
);
948 raw_spin_unlock_irq(&ctx
->lock
);
952 * Put a event into inactive state and update time fields.
953 * Enabling the leader of a group effectively enables all
954 * the group members that aren't explicitly disabled, so we
955 * have to update their ->tstamp_enabled also.
956 * Note: this works for group members as well as group leaders
957 * since the non-leader members' sibling_lists will be empty.
959 static void __perf_event_mark_enabled(struct perf_event
*event
,
960 struct perf_event_context
*ctx
)
962 struct perf_event
*sub
;
964 event
->state
= PERF_EVENT_STATE_INACTIVE
;
965 event
->tstamp_enabled
= ctx
->time
- event
->total_time_enabled
;
966 list_for_each_entry(sub
, &event
->sibling_list
, group_entry
) {
967 if (sub
->state
>= PERF_EVENT_STATE_INACTIVE
) {
968 sub
->tstamp_enabled
=
969 ctx
->time
- sub
->total_time_enabled
;
975 * Cross CPU call to enable a performance event
977 static void __perf_event_enable(void *info
)
979 struct perf_event
*event
= info
;
980 struct perf_event_context
*ctx
= event
->ctx
;
981 struct perf_event
*leader
= event
->group_leader
;
982 struct perf_cpu_context
*cpuctx
= __get_cpu_context(ctx
);
986 * If this is a per-task event, need to check whether this
987 * event's task is the current task on this cpu.
989 if (ctx
->task
&& cpuctx
->task_ctx
!= ctx
) {
990 if (cpuctx
->task_ctx
|| ctx
->task
!= current
)
992 cpuctx
->task_ctx
= ctx
;
995 raw_spin_lock(&ctx
->lock
);
997 update_context_time(ctx
);
999 if (event
->state
>= PERF_EVENT_STATE_INACTIVE
)
1001 __perf_event_mark_enabled(event
, ctx
);
1003 if (event
->cpu
!= -1 && event
->cpu
!= smp_processor_id())
1007 * If the event is in a group and isn't the group leader,
1008 * then don't put it on unless the group is on.
1010 if (leader
!= event
&& leader
->state
!= PERF_EVENT_STATE_ACTIVE
)
1013 if (!group_can_go_on(event
, cpuctx
, 1)) {
1016 if (event
== leader
)
1017 err
= group_sched_in(event
, cpuctx
, ctx
);
1019 err
= event_sched_in(event
, cpuctx
, ctx
);
1024 * If this event can't go on and it's part of a
1025 * group, then the whole group has to come off.
1027 if (leader
!= event
)
1028 group_sched_out(leader
, cpuctx
, ctx
);
1029 if (leader
->attr
.pinned
) {
1030 update_group_times(leader
);
1031 leader
->state
= PERF_EVENT_STATE_ERROR
;
1036 raw_spin_unlock(&ctx
->lock
);
1042 * If event->ctx is a cloned context, callers must make sure that
1043 * every task struct that event->ctx->task could possibly point to
1044 * remains valid. This condition is satisfied when called through
1045 * perf_event_for_each_child or perf_event_for_each as described
1046 * for perf_event_disable.
1048 void perf_event_enable(struct perf_event
*event
)
1050 struct perf_event_context
*ctx
= event
->ctx
;
1051 struct task_struct
*task
= ctx
->task
;
1055 * Enable the event on the cpu that it's on
1057 smp_call_function_single(event
->cpu
, __perf_event_enable
,
1062 raw_spin_lock_irq(&ctx
->lock
);
1063 if (event
->state
>= PERF_EVENT_STATE_INACTIVE
)
1067 * If the event is in error state, clear that first.
1068 * That way, if we see the event in error state below, we
1069 * know that it has gone back into error state, as distinct
1070 * from the task having been scheduled away before the
1071 * cross-call arrived.
1073 if (event
->state
== PERF_EVENT_STATE_ERROR
)
1074 event
->state
= PERF_EVENT_STATE_OFF
;
1077 raw_spin_unlock_irq(&ctx
->lock
);
1078 task_oncpu_function_call(task
, __perf_event_enable
, event
);
1080 raw_spin_lock_irq(&ctx
->lock
);
1083 * If the context is active and the event is still off,
1084 * we need to retry the cross-call.
1086 if (ctx
->is_active
&& event
->state
== PERF_EVENT_STATE_OFF
)
1090 * Since we have the lock this context can't be scheduled
1091 * in, so we can change the state safely.
1093 if (event
->state
== PERF_EVENT_STATE_OFF
)
1094 __perf_event_mark_enabled(event
, ctx
);
1097 raw_spin_unlock_irq(&ctx
->lock
);
1100 static int perf_event_refresh(struct perf_event
*event
, int refresh
)
1103 * not supported on inherited events
1105 if (event
->attr
.inherit
)
1108 atomic_add(refresh
, &event
->event_limit
);
1109 perf_event_enable(event
);
1115 EVENT_FLEXIBLE
= 0x1,
1117 EVENT_ALL
= EVENT_FLEXIBLE
| EVENT_PINNED
,
1120 static void ctx_sched_out(struct perf_event_context
*ctx
,
1121 struct perf_cpu_context
*cpuctx
,
1122 enum event_type_t event_type
)
1124 struct perf_event
*event
;
1126 raw_spin_lock(&ctx
->lock
);
1127 perf_pmu_disable(ctx
->pmu
);
1129 if (likely(!ctx
->nr_events
))
1131 update_context_time(ctx
);
1133 if (!ctx
->nr_active
)
1136 if (event_type
& EVENT_PINNED
) {
1137 list_for_each_entry(event
, &ctx
->pinned_groups
, group_entry
)
1138 group_sched_out(event
, cpuctx
, ctx
);
1141 if (event_type
& EVENT_FLEXIBLE
) {
1142 list_for_each_entry(event
, &ctx
->flexible_groups
, group_entry
)
1143 group_sched_out(event
, cpuctx
, ctx
);
1146 perf_pmu_enable(ctx
->pmu
);
1147 raw_spin_unlock(&ctx
->lock
);
1151 * Test whether two contexts are equivalent, i.e. whether they
1152 * have both been cloned from the same version of the same context
1153 * and they both have the same number of enabled events.
1154 * If the number of enabled events is the same, then the set
1155 * of enabled events should be the same, because these are both
1156 * inherited contexts, therefore we can't access individual events
1157 * in them directly with an fd; we can only enable/disable all
1158 * events via prctl, or enable/disable all events in a family
1159 * via ioctl, which will have the same effect on both contexts.
1161 static int context_equiv(struct perf_event_context
*ctx1
,
1162 struct perf_event_context
*ctx2
)
1164 return ctx1
->parent_ctx
&& ctx1
->parent_ctx
== ctx2
->parent_ctx
1165 && ctx1
->parent_gen
== ctx2
->parent_gen
1166 && !ctx1
->pin_count
&& !ctx2
->pin_count
;
1169 static void __perf_event_sync_stat(struct perf_event
*event
,
1170 struct perf_event
*next_event
)
1174 if (!event
->attr
.inherit_stat
)
1178 * Update the event value, we cannot use perf_event_read()
1179 * because we're in the middle of a context switch and have IRQs
1180 * disabled, which upsets smp_call_function_single(), however
1181 * we know the event must be on the current CPU, therefore we
1182 * don't need to use it.
1184 switch (event
->state
) {
1185 case PERF_EVENT_STATE_ACTIVE
:
1186 event
->pmu
->read(event
);
1189 case PERF_EVENT_STATE_INACTIVE
:
1190 update_event_times(event
);
1198 * In order to keep per-task stats reliable we need to flip the event
1199 * values when we flip the contexts.
1201 value
= local64_read(&next_event
->count
);
1202 value
= local64_xchg(&event
->count
, value
);
1203 local64_set(&next_event
->count
, value
);
1205 swap(event
->total_time_enabled
, next_event
->total_time_enabled
);
1206 swap(event
->total_time_running
, next_event
->total_time_running
);
1209 * Since we swizzled the values, update the user visible data too.
1211 perf_event_update_userpage(event
);
1212 perf_event_update_userpage(next_event
);
1215 #define list_next_entry(pos, member) \
1216 list_entry(pos->member.next, typeof(*pos), member)
1218 static void perf_event_sync_stat(struct perf_event_context
*ctx
,
1219 struct perf_event_context
*next_ctx
)
1221 struct perf_event
*event
, *next_event
;
1226 update_context_time(ctx
);
1228 event
= list_first_entry(&ctx
->event_list
,
1229 struct perf_event
, event_entry
);
1231 next_event
= list_first_entry(&next_ctx
->event_list
,
1232 struct perf_event
, event_entry
);
1234 while (&event
->event_entry
!= &ctx
->event_list
&&
1235 &next_event
->event_entry
!= &next_ctx
->event_list
) {
1237 __perf_event_sync_stat(event
, next_event
);
1239 event
= list_next_entry(event
, event_entry
);
1240 next_event
= list_next_entry(next_event
, event_entry
);
1244 void perf_event_context_sched_out(struct task_struct
*task
, int ctxn
,
1245 struct task_struct
*next
)
1247 struct perf_event_context
*ctx
= task
->perf_event_ctxp
[ctxn
];
1248 struct perf_event_context
*next_ctx
;
1249 struct perf_event_context
*parent
;
1250 struct perf_cpu_context
*cpuctx
;
1256 cpuctx
= __get_cpu_context(ctx
);
1257 if (!cpuctx
->task_ctx
)
1261 parent
= rcu_dereference(ctx
->parent_ctx
);
1262 next_ctx
= next
->perf_event_ctxp
[ctxn
];
1263 if (parent
&& next_ctx
&&
1264 rcu_dereference(next_ctx
->parent_ctx
) == parent
) {
1266 * Looks like the two contexts are clones, so we might be
1267 * able to optimize the context switch. We lock both
1268 * contexts and check that they are clones under the
1269 * lock (including re-checking that neither has been
1270 * uncloned in the meantime). It doesn't matter which
1271 * order we take the locks because no other cpu could
1272 * be trying to lock both of these tasks.
1274 raw_spin_lock(&ctx
->lock
);
1275 raw_spin_lock_nested(&next_ctx
->lock
, SINGLE_DEPTH_NESTING
);
1276 if (context_equiv(ctx
, next_ctx
)) {
1278 * XXX do we need a memory barrier of sorts
1279 * wrt to rcu_dereference() of perf_event_ctxp
1281 task
->perf_event_ctxp
[ctxn
] = next_ctx
;
1282 next
->perf_event_ctxp
[ctxn
] = ctx
;
1284 next_ctx
->task
= task
;
1287 perf_event_sync_stat(ctx
, next_ctx
);
1289 raw_spin_unlock(&next_ctx
->lock
);
1290 raw_spin_unlock(&ctx
->lock
);
1295 ctx_sched_out(ctx
, cpuctx
, EVENT_ALL
);
1296 cpuctx
->task_ctx
= NULL
;
1300 #define for_each_task_context_nr(ctxn) \
1301 for ((ctxn) = 0; (ctxn) < perf_nr_task_contexts; (ctxn)++)
1304 * Called from scheduler to remove the events of the current task,
1305 * with interrupts disabled.
1307 * We stop each event and update the event value in event->count.
1309 * This does not protect us against NMI, but disable()
1310 * sets the disabled bit in the control field of event _before_
1311 * accessing the event control register. If a NMI hits, then it will
1312 * not restart the event.
1314 void perf_event_task_sched_out(struct task_struct
*task
,
1315 struct task_struct
*next
)
1319 perf_sw_event(PERF_COUNT_SW_CONTEXT_SWITCHES
, 1, 1, NULL
, 0);
1321 for_each_task_context_nr(ctxn
)
1322 perf_event_context_sched_out(task
, ctxn
, next
);
1325 static void task_ctx_sched_out(struct perf_event_context
*ctx
,
1326 enum event_type_t event_type
)
1328 struct perf_cpu_context
*cpuctx
= __get_cpu_context(ctx
);
1330 if (!cpuctx
->task_ctx
)
1333 if (WARN_ON_ONCE(ctx
!= cpuctx
->task_ctx
))
1336 ctx_sched_out(ctx
, cpuctx
, event_type
);
1337 cpuctx
->task_ctx
= NULL
;
1341 * Called with IRQs disabled
1343 static void __perf_event_task_sched_out(struct perf_event_context
*ctx
)
1345 task_ctx_sched_out(ctx
, EVENT_ALL
);
1349 * Called with IRQs disabled
1351 static void cpu_ctx_sched_out(struct perf_cpu_context
*cpuctx
,
1352 enum event_type_t event_type
)
1354 ctx_sched_out(&cpuctx
->ctx
, cpuctx
, event_type
);
1358 ctx_pinned_sched_in(struct perf_event_context
*ctx
,
1359 struct perf_cpu_context
*cpuctx
)
1361 struct perf_event
*event
;
1363 list_for_each_entry(event
, &ctx
->pinned_groups
, group_entry
) {
1364 if (event
->state
<= PERF_EVENT_STATE_OFF
)
1366 if (event
->cpu
!= -1 && event
->cpu
!= smp_processor_id())
1369 if (group_can_go_on(event
, cpuctx
, 1))
1370 group_sched_in(event
, cpuctx
, ctx
);
1373 * If this pinned group hasn't been scheduled,
1374 * put it in error state.
1376 if (event
->state
== PERF_EVENT_STATE_INACTIVE
) {
1377 update_group_times(event
);
1378 event
->state
= PERF_EVENT_STATE_ERROR
;
1384 ctx_flexible_sched_in(struct perf_event_context
*ctx
,
1385 struct perf_cpu_context
*cpuctx
)
1387 struct perf_event
*event
;
1390 list_for_each_entry(event
, &ctx
->flexible_groups
, group_entry
) {
1391 /* Ignore events in OFF or ERROR state */
1392 if (event
->state
<= PERF_EVENT_STATE_OFF
)
1395 * Listen to the 'cpu' scheduling filter constraint
1398 if (event
->cpu
!= -1 && event
->cpu
!= smp_processor_id())
1401 if (group_can_go_on(event
, cpuctx
, can_add_hw
)) {
1402 if (group_sched_in(event
, cpuctx
, ctx
))
1409 ctx_sched_in(struct perf_event_context
*ctx
,
1410 struct perf_cpu_context
*cpuctx
,
1411 enum event_type_t event_type
)
1413 raw_spin_lock(&ctx
->lock
);
1415 if (likely(!ctx
->nr_events
))
1418 ctx
->timestamp
= perf_clock();
1421 * First go through the list and put on any pinned groups
1422 * in order to give them the best chance of going on.
1424 if (event_type
& EVENT_PINNED
)
1425 ctx_pinned_sched_in(ctx
, cpuctx
);
1427 /* Then walk through the lower prio flexible groups */
1428 if (event_type
& EVENT_FLEXIBLE
)
1429 ctx_flexible_sched_in(ctx
, cpuctx
);
1432 raw_spin_unlock(&ctx
->lock
);
1435 static void cpu_ctx_sched_in(struct perf_cpu_context
*cpuctx
,
1436 enum event_type_t event_type
)
1438 struct perf_event_context
*ctx
= &cpuctx
->ctx
;
1440 ctx_sched_in(ctx
, cpuctx
, event_type
);
1443 static void task_ctx_sched_in(struct perf_event_context
*ctx
,
1444 enum event_type_t event_type
)
1446 struct perf_cpu_context
*cpuctx
;
1448 cpuctx
= __get_cpu_context(ctx
);
1449 if (cpuctx
->task_ctx
== ctx
)
1452 ctx_sched_in(ctx
, cpuctx
, event_type
);
1453 cpuctx
->task_ctx
= ctx
;
1456 void perf_event_context_sched_in(struct perf_event_context
*ctx
)
1458 struct perf_cpu_context
*cpuctx
;
1460 cpuctx
= __get_cpu_context(ctx
);
1461 if (cpuctx
->task_ctx
== ctx
)
1464 perf_pmu_disable(ctx
->pmu
);
1466 * We want to keep the following priority order:
1467 * cpu pinned (that don't need to move), task pinned,
1468 * cpu flexible, task flexible.
1470 cpu_ctx_sched_out(cpuctx
, EVENT_FLEXIBLE
);
1472 ctx_sched_in(ctx
, cpuctx
, EVENT_PINNED
);
1473 cpu_ctx_sched_in(cpuctx
, EVENT_FLEXIBLE
);
1474 ctx_sched_in(ctx
, cpuctx
, EVENT_FLEXIBLE
);
1476 cpuctx
->task_ctx
= ctx
;
1479 * Since these rotations are per-cpu, we need to ensure the
1480 * cpu-context we got scheduled on is actually rotating.
1482 perf_pmu_rotate_start(ctx
->pmu
);
1483 perf_pmu_enable(ctx
->pmu
);
1487 * Called from scheduler to add the events of the current task
1488 * with interrupts disabled.
1490 * We restore the event value and then enable it.
1492 * This does not protect us against NMI, but enable()
1493 * sets the enabled bit in the control field of event _before_
1494 * accessing the event control register. If a NMI hits, then it will
1495 * keep the event running.
1497 void perf_event_task_sched_in(struct task_struct
*task
)
1499 struct perf_event_context
*ctx
;
1502 for_each_task_context_nr(ctxn
) {
1503 ctx
= task
->perf_event_ctxp
[ctxn
];
1507 perf_event_context_sched_in(ctx
);
1511 #define MAX_INTERRUPTS (~0ULL)
1513 static void perf_log_throttle(struct perf_event
*event
, int enable
);
1515 static u64
perf_calculate_period(struct perf_event
*event
, u64 nsec
, u64 count
)
1517 u64 frequency
= event
->attr
.sample_freq
;
1518 u64 sec
= NSEC_PER_SEC
;
1519 u64 divisor
, dividend
;
1521 int count_fls
, nsec_fls
, frequency_fls
, sec_fls
;
1523 count_fls
= fls64(count
);
1524 nsec_fls
= fls64(nsec
);
1525 frequency_fls
= fls64(frequency
);
1529 * We got @count in @nsec, with a target of sample_freq HZ
1530 * the target period becomes:
1533 * period = -------------------
1534 * @nsec * sample_freq
1539 * Reduce accuracy by one bit such that @a and @b converge
1540 * to a similar magnitude.
1542 #define REDUCE_FLS(a, b) \
1544 if (a##_fls > b##_fls) { \
1554 * Reduce accuracy until either term fits in a u64, then proceed with
1555 * the other, so that finally we can do a u64/u64 division.
1557 while (count_fls
+ sec_fls
> 64 && nsec_fls
+ frequency_fls
> 64) {
1558 REDUCE_FLS(nsec
, frequency
);
1559 REDUCE_FLS(sec
, count
);
1562 if (count_fls
+ sec_fls
> 64) {
1563 divisor
= nsec
* frequency
;
1565 while (count_fls
+ sec_fls
> 64) {
1566 REDUCE_FLS(count
, sec
);
1570 dividend
= count
* sec
;
1572 dividend
= count
* sec
;
1574 while (nsec_fls
+ frequency_fls
> 64) {
1575 REDUCE_FLS(nsec
, frequency
);
1579 divisor
= nsec
* frequency
;
1585 return div64_u64(dividend
, divisor
);
1588 static void perf_adjust_period(struct perf_event
*event
, u64 nsec
, u64 count
)
1590 struct hw_perf_event
*hwc
= &event
->hw
;
1591 s64 period
, sample_period
;
1594 period
= perf_calculate_period(event
, nsec
, count
);
1596 delta
= (s64
)(period
- hwc
->sample_period
);
1597 delta
= (delta
+ 7) / 8; /* low pass filter */
1599 sample_period
= hwc
->sample_period
+ delta
;
1604 hwc
->sample_period
= sample_period
;
1606 if (local64_read(&hwc
->period_left
) > 8*sample_period
) {
1607 event
->pmu
->stop(event
, PERF_EF_UPDATE
);
1608 local64_set(&hwc
->period_left
, 0);
1609 event
->pmu
->start(event
, PERF_EF_RELOAD
);
1613 static void perf_ctx_adjust_freq(struct perf_event_context
*ctx
, u64 period
)
1615 struct perf_event
*event
;
1616 struct hw_perf_event
*hwc
;
1617 u64 interrupts
, now
;
1620 raw_spin_lock(&ctx
->lock
);
1621 list_for_each_entry_rcu(event
, &ctx
->event_list
, event_entry
) {
1622 if (event
->state
!= PERF_EVENT_STATE_ACTIVE
)
1625 if (event
->cpu
!= -1 && event
->cpu
!= smp_processor_id())
1630 interrupts
= hwc
->interrupts
;
1631 hwc
->interrupts
= 0;
1634 * unthrottle events on the tick
1636 if (interrupts
== MAX_INTERRUPTS
) {
1637 perf_log_throttle(event
, 1);
1638 event
->pmu
->start(event
, 0);
1641 if (!event
->attr
.freq
|| !event
->attr
.sample_freq
)
1644 event
->pmu
->read(event
);
1645 now
= local64_read(&event
->count
);
1646 delta
= now
- hwc
->freq_count_stamp
;
1647 hwc
->freq_count_stamp
= now
;
1650 perf_adjust_period(event
, period
, delta
);
1652 raw_spin_unlock(&ctx
->lock
);
1656 * Round-robin a context's events:
1658 static void rotate_ctx(struct perf_event_context
*ctx
)
1660 raw_spin_lock(&ctx
->lock
);
1662 /* Rotate the first entry last of non-pinned groups */
1663 list_rotate_left(&ctx
->flexible_groups
);
1665 raw_spin_unlock(&ctx
->lock
);
1669 * perf_pmu_rotate_start() and perf_rotate_context() are fully serialized
1670 * because they're strictly cpu affine and rotate_start is called with IRQs
1671 * disabled, while rotate_context is called from IRQ context.
1673 static void perf_rotate_context(struct perf_cpu_context
*cpuctx
)
1675 u64 interval
= (u64
)cpuctx
->jiffies_interval
* TICK_NSEC
;
1676 struct perf_event_context
*ctx
= NULL
;
1677 int rotate
= 0, remove
= 1;
1679 if (cpuctx
->ctx
.nr_events
) {
1681 if (cpuctx
->ctx
.nr_events
!= cpuctx
->ctx
.nr_active
)
1685 ctx
= cpuctx
->task_ctx
;
1686 if (ctx
&& ctx
->nr_events
) {
1688 if (ctx
->nr_events
!= ctx
->nr_active
)
1692 perf_pmu_disable(cpuctx
->ctx
.pmu
);
1693 perf_ctx_adjust_freq(&cpuctx
->ctx
, interval
);
1695 perf_ctx_adjust_freq(ctx
, interval
);
1700 cpu_ctx_sched_out(cpuctx
, EVENT_FLEXIBLE
);
1702 task_ctx_sched_out(ctx
, EVENT_FLEXIBLE
);
1704 rotate_ctx(&cpuctx
->ctx
);
1708 cpu_ctx_sched_in(cpuctx
, EVENT_FLEXIBLE
);
1710 task_ctx_sched_in(ctx
, EVENT_FLEXIBLE
);
1714 list_del_init(&cpuctx
->rotation_list
);
1716 perf_pmu_enable(cpuctx
->ctx
.pmu
);
1719 void perf_event_task_tick(void)
1721 struct list_head
*head
= &__get_cpu_var(rotation_list
);
1722 struct perf_cpu_context
*cpuctx
, *tmp
;
1724 WARN_ON(!irqs_disabled());
1726 list_for_each_entry_safe(cpuctx
, tmp
, head
, rotation_list
) {
1727 if (cpuctx
->jiffies_interval
== 1 ||
1728 !(jiffies
% cpuctx
->jiffies_interval
))
1729 perf_rotate_context(cpuctx
);
1733 static int event_enable_on_exec(struct perf_event
*event
,
1734 struct perf_event_context
*ctx
)
1736 if (!event
->attr
.enable_on_exec
)
1739 event
->attr
.enable_on_exec
= 0;
1740 if (event
->state
>= PERF_EVENT_STATE_INACTIVE
)
1743 __perf_event_mark_enabled(event
, ctx
);
1749 * Enable all of a task's events that have been marked enable-on-exec.
1750 * This expects task == current.
1752 static void perf_event_enable_on_exec(struct perf_event_context
*ctx
)
1754 struct perf_event
*event
;
1755 unsigned long flags
;
1759 local_irq_save(flags
);
1760 if (!ctx
|| !ctx
->nr_events
)
1763 task_ctx_sched_out(ctx
, EVENT_ALL
);
1765 raw_spin_lock(&ctx
->lock
);
1767 list_for_each_entry(event
, &ctx
->pinned_groups
, group_entry
) {
1768 ret
= event_enable_on_exec(event
, ctx
);
1773 list_for_each_entry(event
, &ctx
->flexible_groups
, group_entry
) {
1774 ret
= event_enable_on_exec(event
, ctx
);
1780 * Unclone this context if we enabled any event.
1785 raw_spin_unlock(&ctx
->lock
);
1787 perf_event_context_sched_in(ctx
);
1789 local_irq_restore(flags
);
1793 * Cross CPU call to read the hardware event
1795 static void __perf_event_read(void *info
)
1797 struct perf_event
*event
= info
;
1798 struct perf_event_context
*ctx
= event
->ctx
;
1799 struct perf_cpu_context
*cpuctx
= __get_cpu_context(ctx
);
1802 * If this is a task context, we need to check whether it is
1803 * the current task context of this cpu. If not it has been
1804 * scheduled out before the smp call arrived. In that case
1805 * event->count would have been updated to a recent sample
1806 * when the event was scheduled out.
1808 if (ctx
->task
&& cpuctx
->task_ctx
!= ctx
)
1811 raw_spin_lock(&ctx
->lock
);
1812 update_context_time(ctx
);
1813 update_event_times(event
);
1814 raw_spin_unlock(&ctx
->lock
);
1816 event
->pmu
->read(event
);
1819 static inline u64
perf_event_count(struct perf_event
*event
)
1821 return local64_read(&event
->count
) + atomic64_read(&event
->child_count
);
1824 static u64
perf_event_read(struct perf_event
*event
)
1827 * If event is enabled and currently active on a CPU, update the
1828 * value in the event structure:
1830 if (event
->state
== PERF_EVENT_STATE_ACTIVE
) {
1831 smp_call_function_single(event
->oncpu
,
1832 __perf_event_read
, event
, 1);
1833 } else if (event
->state
== PERF_EVENT_STATE_INACTIVE
) {
1834 struct perf_event_context
*ctx
= event
->ctx
;
1835 unsigned long flags
;
1837 raw_spin_lock_irqsave(&ctx
->lock
, flags
);
1839 * may read while context is not active
1840 * (e.g., thread is blocked), in that case
1841 * we cannot update context time
1844 update_context_time(ctx
);
1845 update_event_times(event
);
1846 raw_spin_unlock_irqrestore(&ctx
->lock
, flags
);
1849 return perf_event_count(event
);
1856 struct callchain_cpus_entries
{
1857 struct rcu_head rcu_head
;
1858 struct perf_callchain_entry
*cpu_entries
[0];
1861 static DEFINE_PER_CPU(int, callchain_recursion
[PERF_NR_CONTEXTS
]);
1862 static atomic_t nr_callchain_events
;
1863 static DEFINE_MUTEX(callchain_mutex
);
1864 struct callchain_cpus_entries
*callchain_cpus_entries
;
1867 __weak
void perf_callchain_kernel(struct perf_callchain_entry
*entry
,
1868 struct pt_regs
*regs
)
1872 __weak
void perf_callchain_user(struct perf_callchain_entry
*entry
,
1873 struct pt_regs
*regs
)
1877 static void release_callchain_buffers_rcu(struct rcu_head
*head
)
1879 struct callchain_cpus_entries
*entries
;
1882 entries
= container_of(head
, struct callchain_cpus_entries
, rcu_head
);
1884 for_each_possible_cpu(cpu
)
1885 kfree(entries
->cpu_entries
[cpu
]);
1890 static void release_callchain_buffers(void)
1892 struct callchain_cpus_entries
*entries
;
1894 entries
= callchain_cpus_entries
;
1895 rcu_assign_pointer(callchain_cpus_entries
, NULL
);
1896 call_rcu(&entries
->rcu_head
, release_callchain_buffers_rcu
);
1899 static int alloc_callchain_buffers(void)
1903 struct callchain_cpus_entries
*entries
;
1906 * We can't use the percpu allocation API for data that can be
1907 * accessed from NMI. Use a temporary manual per cpu allocation
1908 * until that gets sorted out.
1910 size
= sizeof(*entries
) + sizeof(struct perf_callchain_entry
*) *
1911 num_possible_cpus();
1913 entries
= kzalloc(size
, GFP_KERNEL
);
1917 size
= sizeof(struct perf_callchain_entry
) * PERF_NR_CONTEXTS
;
1919 for_each_possible_cpu(cpu
) {
1920 entries
->cpu_entries
[cpu
] = kmalloc_node(size
, GFP_KERNEL
,
1922 if (!entries
->cpu_entries
[cpu
])
1926 rcu_assign_pointer(callchain_cpus_entries
, entries
);
1931 for_each_possible_cpu(cpu
)
1932 kfree(entries
->cpu_entries
[cpu
]);
1938 static int get_callchain_buffers(void)
1943 mutex_lock(&callchain_mutex
);
1945 count
= atomic_inc_return(&nr_callchain_events
);
1946 if (WARN_ON_ONCE(count
< 1)) {
1952 /* If the allocation failed, give up */
1953 if (!callchain_cpus_entries
)
1958 err
= alloc_callchain_buffers();
1960 release_callchain_buffers();
1962 mutex_unlock(&callchain_mutex
);
1967 static void put_callchain_buffers(void)
1969 if (atomic_dec_and_mutex_lock(&nr_callchain_events
, &callchain_mutex
)) {
1970 release_callchain_buffers();
1971 mutex_unlock(&callchain_mutex
);
1975 static int get_recursion_context(int *recursion
)
1983 else if (in_softirq())
1988 if (recursion
[rctx
])
1997 static inline void put_recursion_context(int *recursion
, int rctx
)
2003 static struct perf_callchain_entry
*get_callchain_entry(int *rctx
)
2006 struct callchain_cpus_entries
*entries
;
2008 *rctx
= get_recursion_context(__get_cpu_var(callchain_recursion
));
2012 entries
= rcu_dereference(callchain_cpus_entries
);
2016 cpu
= smp_processor_id();
2018 return &entries
->cpu_entries
[cpu
][*rctx
];
2022 put_callchain_entry(int rctx
)
2024 put_recursion_context(__get_cpu_var(callchain_recursion
), rctx
);
2027 static struct perf_callchain_entry
*perf_callchain(struct pt_regs
*regs
)
2030 struct perf_callchain_entry
*entry
;
2033 entry
= get_callchain_entry(&rctx
);
2042 if (!user_mode(regs
)) {
2043 perf_callchain_store(entry
, PERF_CONTEXT_KERNEL
);
2044 perf_callchain_kernel(entry
, regs
);
2046 regs
= task_pt_regs(current
);
2052 perf_callchain_store(entry
, PERF_CONTEXT_USER
);
2053 perf_callchain_user(entry
, regs
);
2057 put_callchain_entry(rctx
);
2063 * Initialize the perf_event context in a task_struct:
2065 static void __perf_event_init_context(struct perf_event_context
*ctx
)
2067 raw_spin_lock_init(&ctx
->lock
);
2068 mutex_init(&ctx
->mutex
);
2069 INIT_LIST_HEAD(&ctx
->pinned_groups
);
2070 INIT_LIST_HEAD(&ctx
->flexible_groups
);
2071 INIT_LIST_HEAD(&ctx
->event_list
);
2072 atomic_set(&ctx
->refcount
, 1);
2075 static struct perf_event_context
*
2076 alloc_perf_context(struct pmu
*pmu
, struct task_struct
*task
)
2078 struct perf_event_context
*ctx
;
2080 ctx
= kzalloc(sizeof(struct perf_event_context
), GFP_KERNEL
);
2084 __perf_event_init_context(ctx
);
2087 get_task_struct(task
);
2094 static struct task_struct
*
2095 find_lively_task_by_vpid(pid_t vpid
)
2097 struct task_struct
*task
;
2104 task
= find_task_by_vpid(vpid
);
2106 get_task_struct(task
);
2110 return ERR_PTR(-ESRCH
);
2113 * Can't attach events to a dying task.
2116 if (task
->flags
& PF_EXITING
)
2119 /* Reuse ptrace permission checks for now. */
2121 if (!ptrace_may_access(task
, PTRACE_MODE_READ
))
2126 put_task_struct(task
);
2127 return ERR_PTR(err
);
2131 static struct perf_event_context
*
2132 find_get_context(struct pmu
*pmu
, struct task_struct
*task
, int cpu
)
2134 struct perf_event_context
*ctx
;
2135 struct perf_cpu_context
*cpuctx
;
2136 unsigned long flags
;
2139 if (!task
&& cpu
!= -1) {
2140 /* Must be root to operate on a CPU event: */
2141 if (perf_paranoid_cpu() && !capable(CAP_SYS_ADMIN
))
2142 return ERR_PTR(-EACCES
);
2144 if (cpu
< 0 || cpu
>= nr_cpumask_bits
)
2145 return ERR_PTR(-EINVAL
);
2148 * We could be clever and allow to attach a event to an
2149 * offline CPU and activate it when the CPU comes up, but
2152 if (!cpu_online(cpu
))
2153 return ERR_PTR(-ENODEV
);
2155 cpuctx
= per_cpu_ptr(pmu
->pmu_cpu_context
, cpu
);
2163 ctxn
= pmu
->task_ctx_nr
;
2168 ctx
= perf_lock_task_context(task
, ctxn
, &flags
);
2171 raw_spin_unlock_irqrestore(&ctx
->lock
, flags
);
2175 ctx
= alloc_perf_context(pmu
, task
);
2182 if (cmpxchg(&task
->perf_event_ctxp
[ctxn
], NULL
, ctx
)) {
2184 * We raced with some other task; use
2185 * the context they set.
2187 put_task_struct(task
);
2196 return ERR_PTR(err
);
2199 static void perf_event_free_filter(struct perf_event
*event
);
2201 static void free_event_rcu(struct rcu_head
*head
)
2203 struct perf_event
*event
;
2205 event
= container_of(head
, struct perf_event
, rcu_head
);
2207 put_pid_ns(event
->ns
);
2208 perf_event_free_filter(event
);
2212 static void perf_buffer_put(struct perf_buffer
*buffer
);
2214 static void free_event(struct perf_event
*event
)
2216 irq_work_sync(&event
->pending
);
2218 if (!event
->parent
) {
2219 atomic_dec(&nr_events
);
2220 if (event
->attr
.mmap
|| event
->attr
.mmap_data
)
2221 atomic_dec(&nr_mmap_events
);
2222 if (event
->attr
.comm
)
2223 atomic_dec(&nr_comm_events
);
2224 if (event
->attr
.task
)
2225 atomic_dec(&nr_task_events
);
2226 if (event
->attr
.sample_type
& PERF_SAMPLE_CALLCHAIN
)
2227 put_callchain_buffers();
2230 if (event
->buffer
) {
2231 perf_buffer_put(event
->buffer
);
2232 event
->buffer
= NULL
;
2236 event
->destroy(event
);
2239 put_ctx(event
->ctx
);
2241 call_rcu(&event
->rcu_head
, free_event_rcu
);
2244 int perf_event_release_kernel(struct perf_event
*event
)
2246 struct perf_event_context
*ctx
= event
->ctx
;
2249 * Remove from the PMU, can't get re-enabled since we got
2250 * here because the last ref went.
2252 perf_event_disable(event
);
2254 WARN_ON_ONCE(ctx
->parent_ctx
);
2256 * There are two ways this annotation is useful:
2258 * 1) there is a lock recursion from perf_event_exit_task
2259 * see the comment there.
2261 * 2) there is a lock-inversion with mmap_sem through
2262 * perf_event_read_group(), which takes faults while
2263 * holding ctx->mutex, however this is called after
2264 * the last filedesc died, so there is no possibility
2265 * to trigger the AB-BA case.
2267 mutex_lock_nested(&ctx
->mutex
, SINGLE_DEPTH_NESTING
);
2268 raw_spin_lock_irq(&ctx
->lock
);
2269 perf_group_detach(event
);
2270 list_del_event(event
, ctx
);
2271 raw_spin_unlock_irq(&ctx
->lock
);
2272 mutex_unlock(&ctx
->mutex
);
2274 mutex_lock(&event
->owner
->perf_event_mutex
);
2275 list_del_init(&event
->owner_entry
);
2276 mutex_unlock(&event
->owner
->perf_event_mutex
);
2277 put_task_struct(event
->owner
);
2283 EXPORT_SYMBOL_GPL(perf_event_release_kernel
);
2286 * Called when the last reference to the file is gone.
2288 static int perf_release(struct inode
*inode
, struct file
*file
)
2290 struct perf_event
*event
= file
->private_data
;
2292 file
->private_data
= NULL
;
2294 return perf_event_release_kernel(event
);
2297 static int perf_event_read_size(struct perf_event
*event
)
2299 int entry
= sizeof(u64
); /* value */
2303 if (event
->attr
.read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
)
2304 size
+= sizeof(u64
);
2306 if (event
->attr
.read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
)
2307 size
+= sizeof(u64
);
2309 if (event
->attr
.read_format
& PERF_FORMAT_ID
)
2310 entry
+= sizeof(u64
);
2312 if (event
->attr
.read_format
& PERF_FORMAT_GROUP
) {
2313 nr
+= event
->group_leader
->nr_siblings
;
2314 size
+= sizeof(u64
);
2322 u64
perf_event_read_value(struct perf_event
*event
, u64
*enabled
, u64
*running
)
2324 struct perf_event
*child
;
2330 mutex_lock(&event
->child_mutex
);
2331 total
+= perf_event_read(event
);
2332 *enabled
+= event
->total_time_enabled
+
2333 atomic64_read(&event
->child_total_time_enabled
);
2334 *running
+= event
->total_time_running
+
2335 atomic64_read(&event
->child_total_time_running
);
2337 list_for_each_entry(child
, &event
->child_list
, child_list
) {
2338 total
+= perf_event_read(child
);
2339 *enabled
+= child
->total_time_enabled
;
2340 *running
+= child
->total_time_running
;
2342 mutex_unlock(&event
->child_mutex
);
2346 EXPORT_SYMBOL_GPL(perf_event_read_value
);
2348 static int perf_event_read_group(struct perf_event
*event
,
2349 u64 read_format
, char __user
*buf
)
2351 struct perf_event
*leader
= event
->group_leader
, *sub
;
2352 int n
= 0, size
= 0, ret
= -EFAULT
;
2353 struct perf_event_context
*ctx
= leader
->ctx
;
2355 u64 count
, enabled
, running
;
2357 mutex_lock(&ctx
->mutex
);
2358 count
= perf_event_read_value(leader
, &enabled
, &running
);
2360 values
[n
++] = 1 + leader
->nr_siblings
;
2361 if (read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
)
2362 values
[n
++] = enabled
;
2363 if (read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
)
2364 values
[n
++] = running
;
2365 values
[n
++] = count
;
2366 if (read_format
& PERF_FORMAT_ID
)
2367 values
[n
++] = primary_event_id(leader
);
2369 size
= n
* sizeof(u64
);
2371 if (copy_to_user(buf
, values
, size
))
2376 list_for_each_entry(sub
, &leader
->sibling_list
, group_entry
) {
2379 values
[n
++] = perf_event_read_value(sub
, &enabled
, &running
);
2380 if (read_format
& PERF_FORMAT_ID
)
2381 values
[n
++] = primary_event_id(sub
);
2383 size
= n
* sizeof(u64
);
2385 if (copy_to_user(buf
+ ret
, values
, size
)) {
2393 mutex_unlock(&ctx
->mutex
);
2398 static int perf_event_read_one(struct perf_event
*event
,
2399 u64 read_format
, char __user
*buf
)
2401 u64 enabled
, running
;
2405 values
[n
++] = perf_event_read_value(event
, &enabled
, &running
);
2406 if (read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
)
2407 values
[n
++] = enabled
;
2408 if (read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
)
2409 values
[n
++] = running
;
2410 if (read_format
& PERF_FORMAT_ID
)
2411 values
[n
++] = primary_event_id(event
);
2413 if (copy_to_user(buf
, values
, n
* sizeof(u64
)))
2416 return n
* sizeof(u64
);
2420 * Read the performance event - simple non blocking version for now
2423 perf_read_hw(struct perf_event
*event
, char __user
*buf
, size_t count
)
2425 u64 read_format
= event
->attr
.read_format
;
2429 * Return end-of-file for a read on a event that is in
2430 * error state (i.e. because it was pinned but it couldn't be
2431 * scheduled on to the CPU at some point).
2433 if (event
->state
== PERF_EVENT_STATE_ERROR
)
2436 if (count
< perf_event_read_size(event
))
2439 WARN_ON_ONCE(event
->ctx
->parent_ctx
);
2440 if (read_format
& PERF_FORMAT_GROUP
)
2441 ret
= perf_event_read_group(event
, read_format
, buf
);
2443 ret
= perf_event_read_one(event
, read_format
, buf
);
2449 perf_read(struct file
*file
, char __user
*buf
, size_t count
, loff_t
*ppos
)
2451 struct perf_event
*event
= file
->private_data
;
2453 return perf_read_hw(event
, buf
, count
);
2456 static unsigned int perf_poll(struct file
*file
, poll_table
*wait
)
2458 struct perf_event
*event
= file
->private_data
;
2459 struct perf_buffer
*buffer
;
2460 unsigned int events
= POLL_HUP
;
2463 buffer
= rcu_dereference(event
->buffer
);
2465 events
= atomic_xchg(&buffer
->poll
, 0);
2468 poll_wait(file
, &event
->waitq
, wait
);
2473 static void perf_event_reset(struct perf_event
*event
)
2475 (void)perf_event_read(event
);
2476 local64_set(&event
->count
, 0);
2477 perf_event_update_userpage(event
);
2481 * Holding the top-level event's child_mutex means that any
2482 * descendant process that has inherited this event will block
2483 * in sync_child_event if it goes to exit, thus satisfying the
2484 * task existence requirements of perf_event_enable/disable.
2486 static void perf_event_for_each_child(struct perf_event
*event
,
2487 void (*func
)(struct perf_event
*))
2489 struct perf_event
*child
;
2491 WARN_ON_ONCE(event
->ctx
->parent_ctx
);
2492 mutex_lock(&event
->child_mutex
);
2494 list_for_each_entry(child
, &event
->child_list
, child_list
)
2496 mutex_unlock(&event
->child_mutex
);
2499 static void perf_event_for_each(struct perf_event
*event
,
2500 void (*func
)(struct perf_event
*))
2502 struct perf_event_context
*ctx
= event
->ctx
;
2503 struct perf_event
*sibling
;
2505 WARN_ON_ONCE(ctx
->parent_ctx
);
2506 mutex_lock(&ctx
->mutex
);
2507 event
= event
->group_leader
;
2509 perf_event_for_each_child(event
, func
);
2511 list_for_each_entry(sibling
, &event
->sibling_list
, group_entry
)
2512 perf_event_for_each_child(event
, func
);
2513 mutex_unlock(&ctx
->mutex
);
2516 static int perf_event_period(struct perf_event
*event
, u64 __user
*arg
)
2518 struct perf_event_context
*ctx
= event
->ctx
;
2523 if (!event
->attr
.sample_period
)
2526 size
= copy_from_user(&value
, arg
, sizeof(value
));
2527 if (size
!= sizeof(value
))
2533 raw_spin_lock_irq(&ctx
->lock
);
2534 if (event
->attr
.freq
) {
2535 if (value
> sysctl_perf_event_sample_rate
) {
2540 event
->attr
.sample_freq
= value
;
2542 event
->attr
.sample_period
= value
;
2543 event
->hw
.sample_period
= value
;
2546 raw_spin_unlock_irq(&ctx
->lock
);
2551 static const struct file_operations perf_fops
;
2553 static struct perf_event
*perf_fget_light(int fd
, int *fput_needed
)
2557 file
= fget_light(fd
, fput_needed
);
2559 return ERR_PTR(-EBADF
);
2561 if (file
->f_op
!= &perf_fops
) {
2562 fput_light(file
, *fput_needed
);
2564 return ERR_PTR(-EBADF
);
2567 return file
->private_data
;
2570 static int perf_event_set_output(struct perf_event
*event
,
2571 struct perf_event
*output_event
);
2572 static int perf_event_set_filter(struct perf_event
*event
, void __user
*arg
);
2574 static long perf_ioctl(struct file
*file
, unsigned int cmd
, unsigned long arg
)
2576 struct perf_event
*event
= file
->private_data
;
2577 void (*func
)(struct perf_event
*);
2581 case PERF_EVENT_IOC_ENABLE
:
2582 func
= perf_event_enable
;
2584 case PERF_EVENT_IOC_DISABLE
:
2585 func
= perf_event_disable
;
2587 case PERF_EVENT_IOC_RESET
:
2588 func
= perf_event_reset
;
2591 case PERF_EVENT_IOC_REFRESH
:
2592 return perf_event_refresh(event
, arg
);
2594 case PERF_EVENT_IOC_PERIOD
:
2595 return perf_event_period(event
, (u64 __user
*)arg
);
2597 case PERF_EVENT_IOC_SET_OUTPUT
:
2599 struct perf_event
*output_event
= NULL
;
2600 int fput_needed
= 0;
2604 output_event
= perf_fget_light(arg
, &fput_needed
);
2605 if (IS_ERR(output_event
))
2606 return PTR_ERR(output_event
);
2609 ret
= perf_event_set_output(event
, output_event
);
2611 fput_light(output_event
->filp
, fput_needed
);
2616 case PERF_EVENT_IOC_SET_FILTER
:
2617 return perf_event_set_filter(event
, (void __user
*)arg
);
2623 if (flags
& PERF_IOC_FLAG_GROUP
)
2624 perf_event_for_each(event
, func
);
2626 perf_event_for_each_child(event
, func
);
2631 int perf_event_task_enable(void)
2633 struct perf_event
*event
;
2635 mutex_lock(¤t
->perf_event_mutex
);
2636 list_for_each_entry(event
, ¤t
->perf_event_list
, owner_entry
)
2637 perf_event_for_each_child(event
, perf_event_enable
);
2638 mutex_unlock(¤t
->perf_event_mutex
);
2643 int perf_event_task_disable(void)
2645 struct perf_event
*event
;
2647 mutex_lock(¤t
->perf_event_mutex
);
2648 list_for_each_entry(event
, ¤t
->perf_event_list
, owner_entry
)
2649 perf_event_for_each_child(event
, perf_event_disable
);
2650 mutex_unlock(¤t
->perf_event_mutex
);
2655 #ifndef PERF_EVENT_INDEX_OFFSET
2656 # define PERF_EVENT_INDEX_OFFSET 0
2659 static int perf_event_index(struct perf_event
*event
)
2661 if (event
->hw
.state
& PERF_HES_STOPPED
)
2664 if (event
->state
!= PERF_EVENT_STATE_ACTIVE
)
2667 return event
->hw
.idx
+ 1 - PERF_EVENT_INDEX_OFFSET
;
2671 * Callers need to ensure there can be no nesting of this function, otherwise
2672 * the seqlock logic goes bad. We can not serialize this because the arch
2673 * code calls this from NMI context.
2675 void perf_event_update_userpage(struct perf_event
*event
)
2677 struct perf_event_mmap_page
*userpg
;
2678 struct perf_buffer
*buffer
;
2681 buffer
= rcu_dereference(event
->buffer
);
2685 userpg
= buffer
->user_page
;
2688 * Disable preemption so as to not let the corresponding user-space
2689 * spin too long if we get preempted.
2694 userpg
->index
= perf_event_index(event
);
2695 userpg
->offset
= perf_event_count(event
);
2696 if (event
->state
== PERF_EVENT_STATE_ACTIVE
)
2697 userpg
->offset
-= local64_read(&event
->hw
.prev_count
);
2699 userpg
->time_enabled
= event
->total_time_enabled
+
2700 atomic64_read(&event
->child_total_time_enabled
);
2702 userpg
->time_running
= event
->total_time_running
+
2703 atomic64_read(&event
->child_total_time_running
);
2712 static unsigned long perf_data_size(struct perf_buffer
*buffer
);
2715 perf_buffer_init(struct perf_buffer
*buffer
, long watermark
, int flags
)
2717 long max_size
= perf_data_size(buffer
);
2720 buffer
->watermark
= min(max_size
, watermark
);
2722 if (!buffer
->watermark
)
2723 buffer
->watermark
= max_size
/ 2;
2725 if (flags
& PERF_BUFFER_WRITABLE
)
2726 buffer
->writable
= 1;
2728 atomic_set(&buffer
->refcount
, 1);
2731 #ifndef CONFIG_PERF_USE_VMALLOC
2734 * Back perf_mmap() with regular GFP_KERNEL-0 pages.
2737 static struct page
*
2738 perf_mmap_to_page(struct perf_buffer
*buffer
, unsigned long pgoff
)
2740 if (pgoff
> buffer
->nr_pages
)
2744 return virt_to_page(buffer
->user_page
);
2746 return virt_to_page(buffer
->data_pages
[pgoff
- 1]);
2749 static void *perf_mmap_alloc_page(int cpu
)
2754 node
= (cpu
== -1) ? cpu
: cpu_to_node(cpu
);
2755 page
= alloc_pages_node(node
, GFP_KERNEL
| __GFP_ZERO
, 0);
2759 return page_address(page
);
2762 static struct perf_buffer
*
2763 perf_buffer_alloc(int nr_pages
, long watermark
, int cpu
, int flags
)
2765 struct perf_buffer
*buffer
;
2769 size
= sizeof(struct perf_buffer
);
2770 size
+= nr_pages
* sizeof(void *);
2772 buffer
= kzalloc(size
, GFP_KERNEL
);
2776 buffer
->user_page
= perf_mmap_alloc_page(cpu
);
2777 if (!buffer
->user_page
)
2778 goto fail_user_page
;
2780 for (i
= 0; i
< nr_pages
; i
++) {
2781 buffer
->data_pages
[i
] = perf_mmap_alloc_page(cpu
);
2782 if (!buffer
->data_pages
[i
])
2783 goto fail_data_pages
;
2786 buffer
->nr_pages
= nr_pages
;
2788 perf_buffer_init(buffer
, watermark
, flags
);
2793 for (i
--; i
>= 0; i
--)
2794 free_page((unsigned long)buffer
->data_pages
[i
]);
2796 free_page((unsigned long)buffer
->user_page
);
2805 static void perf_mmap_free_page(unsigned long addr
)
2807 struct page
*page
= virt_to_page((void *)addr
);
2809 page
->mapping
= NULL
;
2813 static void perf_buffer_free(struct perf_buffer
*buffer
)
2817 perf_mmap_free_page((unsigned long)buffer
->user_page
);
2818 for (i
= 0; i
< buffer
->nr_pages
; i
++)
2819 perf_mmap_free_page((unsigned long)buffer
->data_pages
[i
]);
2823 static inline int page_order(struct perf_buffer
*buffer
)
2831 * Back perf_mmap() with vmalloc memory.
2833 * Required for architectures that have d-cache aliasing issues.
2836 static inline int page_order(struct perf_buffer
*buffer
)
2838 return buffer
->page_order
;
2841 static struct page
*
2842 perf_mmap_to_page(struct perf_buffer
*buffer
, unsigned long pgoff
)
2844 if (pgoff
> (1UL << page_order(buffer
)))
2847 return vmalloc_to_page((void *)buffer
->user_page
+ pgoff
* PAGE_SIZE
);
2850 static void perf_mmap_unmark_page(void *addr
)
2852 struct page
*page
= vmalloc_to_page(addr
);
2854 page
->mapping
= NULL
;
2857 static void perf_buffer_free_work(struct work_struct
*work
)
2859 struct perf_buffer
*buffer
;
2863 buffer
= container_of(work
, struct perf_buffer
, work
);
2864 nr
= 1 << page_order(buffer
);
2866 base
= buffer
->user_page
;
2867 for (i
= 0; i
< nr
+ 1; i
++)
2868 perf_mmap_unmark_page(base
+ (i
* PAGE_SIZE
));
2874 static void perf_buffer_free(struct perf_buffer
*buffer
)
2876 schedule_work(&buffer
->work
);
2879 static struct perf_buffer
*
2880 perf_buffer_alloc(int nr_pages
, long watermark
, int cpu
, int flags
)
2882 struct perf_buffer
*buffer
;
2886 size
= sizeof(struct perf_buffer
);
2887 size
+= sizeof(void *);
2889 buffer
= kzalloc(size
, GFP_KERNEL
);
2893 INIT_WORK(&buffer
->work
, perf_buffer_free_work
);
2895 all_buf
= vmalloc_user((nr_pages
+ 1) * PAGE_SIZE
);
2899 buffer
->user_page
= all_buf
;
2900 buffer
->data_pages
[0] = all_buf
+ PAGE_SIZE
;
2901 buffer
->page_order
= ilog2(nr_pages
);
2902 buffer
->nr_pages
= 1;
2904 perf_buffer_init(buffer
, watermark
, flags
);
2917 static unsigned long perf_data_size(struct perf_buffer
*buffer
)
2919 return buffer
->nr_pages
<< (PAGE_SHIFT
+ page_order(buffer
));
2922 static int perf_mmap_fault(struct vm_area_struct
*vma
, struct vm_fault
*vmf
)
2924 struct perf_event
*event
= vma
->vm_file
->private_data
;
2925 struct perf_buffer
*buffer
;
2926 int ret
= VM_FAULT_SIGBUS
;
2928 if (vmf
->flags
& FAULT_FLAG_MKWRITE
) {
2929 if (vmf
->pgoff
== 0)
2935 buffer
= rcu_dereference(event
->buffer
);
2939 if (vmf
->pgoff
&& (vmf
->flags
& FAULT_FLAG_WRITE
))
2942 vmf
->page
= perf_mmap_to_page(buffer
, vmf
->pgoff
);
2946 get_page(vmf
->page
);
2947 vmf
->page
->mapping
= vma
->vm_file
->f_mapping
;
2948 vmf
->page
->index
= vmf
->pgoff
;
2957 static void perf_buffer_free_rcu(struct rcu_head
*rcu_head
)
2959 struct perf_buffer
*buffer
;
2961 buffer
= container_of(rcu_head
, struct perf_buffer
, rcu_head
);
2962 perf_buffer_free(buffer
);
2965 static struct perf_buffer
*perf_buffer_get(struct perf_event
*event
)
2967 struct perf_buffer
*buffer
;
2970 buffer
= rcu_dereference(event
->buffer
);
2972 if (!atomic_inc_not_zero(&buffer
->refcount
))
2980 static void perf_buffer_put(struct perf_buffer
*buffer
)
2982 if (!atomic_dec_and_test(&buffer
->refcount
))
2985 call_rcu(&buffer
->rcu_head
, perf_buffer_free_rcu
);
2988 static void perf_mmap_open(struct vm_area_struct
*vma
)
2990 struct perf_event
*event
= vma
->vm_file
->private_data
;
2992 atomic_inc(&event
->mmap_count
);
2995 static void perf_mmap_close(struct vm_area_struct
*vma
)
2997 struct perf_event
*event
= vma
->vm_file
->private_data
;
2999 if (atomic_dec_and_mutex_lock(&event
->mmap_count
, &event
->mmap_mutex
)) {
3000 unsigned long size
= perf_data_size(event
->buffer
);
3001 struct user_struct
*user
= event
->mmap_user
;
3002 struct perf_buffer
*buffer
= event
->buffer
;
3004 atomic_long_sub((size
>> PAGE_SHIFT
) + 1, &user
->locked_vm
);
3005 vma
->vm_mm
->locked_vm
-= event
->mmap_locked
;
3006 rcu_assign_pointer(event
->buffer
, NULL
);
3007 mutex_unlock(&event
->mmap_mutex
);
3009 perf_buffer_put(buffer
);
3014 static const struct vm_operations_struct perf_mmap_vmops
= {
3015 .open
= perf_mmap_open
,
3016 .close
= perf_mmap_close
,
3017 .fault
= perf_mmap_fault
,
3018 .page_mkwrite
= perf_mmap_fault
,
3021 static int perf_mmap(struct file
*file
, struct vm_area_struct
*vma
)
3023 struct perf_event
*event
= file
->private_data
;
3024 unsigned long user_locked
, user_lock_limit
;
3025 struct user_struct
*user
= current_user();
3026 unsigned long locked
, lock_limit
;
3027 struct perf_buffer
*buffer
;
3028 unsigned long vma_size
;
3029 unsigned long nr_pages
;
3030 long user_extra
, extra
;
3031 int ret
= 0, flags
= 0;
3034 * Don't allow mmap() of inherited per-task counters. This would
3035 * create a performance issue due to all children writing to the
3038 if (event
->cpu
== -1 && event
->attr
.inherit
)
3041 if (!(vma
->vm_flags
& VM_SHARED
))
3044 vma_size
= vma
->vm_end
- vma
->vm_start
;
3045 nr_pages
= (vma_size
/ PAGE_SIZE
) - 1;
3048 * If we have buffer pages ensure they're a power-of-two number, so we
3049 * can do bitmasks instead of modulo.
3051 if (nr_pages
!= 0 && !is_power_of_2(nr_pages
))
3054 if (vma_size
!= PAGE_SIZE
* (1 + nr_pages
))
3057 if (vma
->vm_pgoff
!= 0)
3060 WARN_ON_ONCE(event
->ctx
->parent_ctx
);
3061 mutex_lock(&event
->mmap_mutex
);
3062 if (event
->buffer
) {
3063 if (event
->buffer
->nr_pages
== nr_pages
)
3064 atomic_inc(&event
->buffer
->refcount
);
3070 user_extra
= nr_pages
+ 1;
3071 user_lock_limit
= sysctl_perf_event_mlock
>> (PAGE_SHIFT
- 10);
3074 * Increase the limit linearly with more CPUs:
3076 user_lock_limit
*= num_online_cpus();
3078 user_locked
= atomic_long_read(&user
->locked_vm
) + user_extra
;
3081 if (user_locked
> user_lock_limit
)
3082 extra
= user_locked
- user_lock_limit
;
3084 lock_limit
= rlimit(RLIMIT_MEMLOCK
);
3085 lock_limit
>>= PAGE_SHIFT
;
3086 locked
= vma
->vm_mm
->locked_vm
+ extra
;
3088 if ((locked
> lock_limit
) && perf_paranoid_tracepoint_raw() &&
3089 !capable(CAP_IPC_LOCK
)) {
3094 WARN_ON(event
->buffer
);
3096 if (vma
->vm_flags
& VM_WRITE
)
3097 flags
|= PERF_BUFFER_WRITABLE
;
3099 buffer
= perf_buffer_alloc(nr_pages
, event
->attr
.wakeup_watermark
,
3105 rcu_assign_pointer(event
->buffer
, buffer
);
3107 atomic_long_add(user_extra
, &user
->locked_vm
);
3108 event
->mmap_locked
= extra
;
3109 event
->mmap_user
= get_current_user();
3110 vma
->vm_mm
->locked_vm
+= event
->mmap_locked
;
3114 atomic_inc(&event
->mmap_count
);
3115 mutex_unlock(&event
->mmap_mutex
);
3117 vma
->vm_flags
|= VM_RESERVED
;
3118 vma
->vm_ops
= &perf_mmap_vmops
;
3123 static int perf_fasync(int fd
, struct file
*filp
, int on
)
3125 struct inode
*inode
= filp
->f_path
.dentry
->d_inode
;
3126 struct perf_event
*event
= filp
->private_data
;
3129 mutex_lock(&inode
->i_mutex
);
3130 retval
= fasync_helper(fd
, filp
, on
, &event
->fasync
);
3131 mutex_unlock(&inode
->i_mutex
);
3139 static const struct file_operations perf_fops
= {
3140 .llseek
= no_llseek
,
3141 .release
= perf_release
,
3144 .unlocked_ioctl
= perf_ioctl
,
3145 .compat_ioctl
= perf_ioctl
,
3147 .fasync
= perf_fasync
,
3153 * If there's data, ensure we set the poll() state and publish everything
3154 * to user-space before waking everybody up.
3157 void perf_event_wakeup(struct perf_event
*event
)
3159 wake_up_all(&event
->waitq
);
3161 if (event
->pending_kill
) {
3162 kill_fasync(&event
->fasync
, SIGIO
, event
->pending_kill
);
3163 event
->pending_kill
= 0;
3167 static void perf_pending_event(struct irq_work
*entry
)
3169 struct perf_event
*event
= container_of(entry
,
3170 struct perf_event
, pending
);
3172 if (event
->pending_disable
) {
3173 event
->pending_disable
= 0;
3174 __perf_event_disable(event
);
3177 if (event
->pending_wakeup
) {
3178 event
->pending_wakeup
= 0;
3179 perf_event_wakeup(event
);
3184 * We assume there is only KVM supporting the callbacks.
3185 * Later on, we might change it to a list if there is
3186 * another virtualization implementation supporting the callbacks.
3188 struct perf_guest_info_callbacks
*perf_guest_cbs
;
3190 int perf_register_guest_info_callbacks(struct perf_guest_info_callbacks
*cbs
)
3192 perf_guest_cbs
= cbs
;
3195 EXPORT_SYMBOL_GPL(perf_register_guest_info_callbacks
);
3197 int perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks
*cbs
)
3199 perf_guest_cbs
= NULL
;
3202 EXPORT_SYMBOL_GPL(perf_unregister_guest_info_callbacks
);
3207 static bool perf_output_space(struct perf_buffer
*buffer
, unsigned long tail
,
3208 unsigned long offset
, unsigned long head
)
3212 if (!buffer
->writable
)
3215 mask
= perf_data_size(buffer
) - 1;
3217 offset
= (offset
- tail
) & mask
;
3218 head
= (head
- tail
) & mask
;
3220 if ((int)(head
- offset
) < 0)
3226 static void perf_output_wakeup(struct perf_output_handle
*handle
)
3228 atomic_set(&handle
->buffer
->poll
, POLL_IN
);
3231 handle
->event
->pending_wakeup
= 1;
3232 irq_work_queue(&handle
->event
->pending
);
3234 perf_event_wakeup(handle
->event
);
3238 * We need to ensure a later event_id doesn't publish a head when a former
3239 * event isn't done writing. However since we need to deal with NMIs we
3240 * cannot fully serialize things.
3242 * We only publish the head (and generate a wakeup) when the outer-most
3245 static void perf_output_get_handle(struct perf_output_handle
*handle
)
3247 struct perf_buffer
*buffer
= handle
->buffer
;
3250 local_inc(&buffer
->nest
);
3251 handle
->wakeup
= local_read(&buffer
->wakeup
);
3254 static void perf_output_put_handle(struct perf_output_handle
*handle
)
3256 struct perf_buffer
*buffer
= handle
->buffer
;
3260 head
= local_read(&buffer
->head
);
3263 * IRQ/NMI can happen here, which means we can miss a head update.
3266 if (!local_dec_and_test(&buffer
->nest
))
3270 * Publish the known good head. Rely on the full barrier implied
3271 * by atomic_dec_and_test() order the buffer->head read and this
3274 buffer
->user_page
->data_head
= head
;
3277 * Now check if we missed an update, rely on the (compiler)
3278 * barrier in atomic_dec_and_test() to re-read buffer->head.
3280 if (unlikely(head
!= local_read(&buffer
->head
))) {
3281 local_inc(&buffer
->nest
);
3285 if (handle
->wakeup
!= local_read(&buffer
->wakeup
))
3286 perf_output_wakeup(handle
);
3292 __always_inline
void perf_output_copy(struct perf_output_handle
*handle
,
3293 const void *buf
, unsigned int len
)
3296 unsigned long size
= min_t(unsigned long, handle
->size
, len
);
3298 memcpy(handle
->addr
, buf
, size
);
3301 handle
->addr
+= size
;
3303 handle
->size
-= size
;
3304 if (!handle
->size
) {
3305 struct perf_buffer
*buffer
= handle
->buffer
;
3308 handle
->page
&= buffer
->nr_pages
- 1;
3309 handle
->addr
= buffer
->data_pages
[handle
->page
];
3310 handle
->size
= PAGE_SIZE
<< page_order(buffer
);
3315 int perf_output_begin(struct perf_output_handle
*handle
,
3316 struct perf_event
*event
, unsigned int size
,
3317 int nmi
, int sample
)
3319 struct perf_buffer
*buffer
;
3320 unsigned long tail
, offset
, head
;
3323 struct perf_event_header header
;
3330 * For inherited events we send all the output towards the parent.
3333 event
= event
->parent
;
3335 buffer
= rcu_dereference(event
->buffer
);
3339 handle
->buffer
= buffer
;
3340 handle
->event
= event
;
3342 handle
->sample
= sample
;
3344 if (!buffer
->nr_pages
)
3347 have_lost
= local_read(&buffer
->lost
);
3349 size
+= sizeof(lost_event
);
3351 perf_output_get_handle(handle
);
3355 * Userspace could choose to issue a mb() before updating the
3356 * tail pointer. So that all reads will be completed before the
3359 tail
= ACCESS_ONCE(buffer
->user_page
->data_tail
);
3361 offset
= head
= local_read(&buffer
->head
);
3363 if (unlikely(!perf_output_space(buffer
, tail
, offset
, head
)))
3365 } while (local_cmpxchg(&buffer
->head
, offset
, head
) != offset
);
3367 if (head
- local_read(&buffer
->wakeup
) > buffer
->watermark
)
3368 local_add(buffer
->watermark
, &buffer
->wakeup
);
3370 handle
->page
= offset
>> (PAGE_SHIFT
+ page_order(buffer
));
3371 handle
->page
&= buffer
->nr_pages
- 1;
3372 handle
->size
= offset
& ((PAGE_SIZE
<< page_order(buffer
)) - 1);
3373 handle
->addr
= buffer
->data_pages
[handle
->page
];
3374 handle
->addr
+= handle
->size
;
3375 handle
->size
= (PAGE_SIZE
<< page_order(buffer
)) - handle
->size
;
3378 lost_event
.header
.type
= PERF_RECORD_LOST
;
3379 lost_event
.header
.misc
= 0;
3380 lost_event
.header
.size
= sizeof(lost_event
);
3381 lost_event
.id
= event
->id
;
3382 lost_event
.lost
= local_xchg(&buffer
->lost
, 0);
3384 perf_output_put(handle
, lost_event
);
3390 local_inc(&buffer
->lost
);
3391 perf_output_put_handle(handle
);
3398 void perf_output_end(struct perf_output_handle
*handle
)
3400 struct perf_event
*event
= handle
->event
;
3401 struct perf_buffer
*buffer
= handle
->buffer
;
3403 int wakeup_events
= event
->attr
.wakeup_events
;
3405 if (handle
->sample
&& wakeup_events
) {
3406 int events
= local_inc_return(&buffer
->events
);
3407 if (events
>= wakeup_events
) {
3408 local_sub(wakeup_events
, &buffer
->events
);
3409 local_inc(&buffer
->wakeup
);
3413 perf_output_put_handle(handle
);
3417 static u32
perf_event_pid(struct perf_event
*event
, struct task_struct
*p
)
3420 * only top level events have the pid namespace they were created in
3423 event
= event
->parent
;
3425 return task_tgid_nr_ns(p
, event
->ns
);
3428 static u32
perf_event_tid(struct perf_event
*event
, struct task_struct
*p
)
3431 * only top level events have the pid namespace they were created in
3434 event
= event
->parent
;
3436 return task_pid_nr_ns(p
, event
->ns
);
3439 static void perf_output_read_one(struct perf_output_handle
*handle
,
3440 struct perf_event
*event
)
3442 u64 read_format
= event
->attr
.read_format
;
3446 values
[n
++] = perf_event_count(event
);
3447 if (read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
) {
3448 values
[n
++] = event
->total_time_enabled
+
3449 atomic64_read(&event
->child_total_time_enabled
);
3451 if (read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
) {
3452 values
[n
++] = event
->total_time_running
+
3453 atomic64_read(&event
->child_total_time_running
);
3455 if (read_format
& PERF_FORMAT_ID
)
3456 values
[n
++] = primary_event_id(event
);
3458 perf_output_copy(handle
, values
, n
* sizeof(u64
));
3462 * XXX PERF_FORMAT_GROUP vs inherited events seems difficult.
3464 static void perf_output_read_group(struct perf_output_handle
*handle
,
3465 struct perf_event
*event
)
3467 struct perf_event
*leader
= event
->group_leader
, *sub
;
3468 u64 read_format
= event
->attr
.read_format
;
3472 values
[n
++] = 1 + leader
->nr_siblings
;
3474 if (read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
)
3475 values
[n
++] = leader
->total_time_enabled
;
3477 if (read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
)
3478 values
[n
++] = leader
->total_time_running
;
3480 if (leader
!= event
)
3481 leader
->pmu
->read(leader
);
3483 values
[n
++] = perf_event_count(leader
);
3484 if (read_format
& PERF_FORMAT_ID
)
3485 values
[n
++] = primary_event_id(leader
);
3487 perf_output_copy(handle
, values
, n
* sizeof(u64
));
3489 list_for_each_entry(sub
, &leader
->sibling_list
, group_entry
) {
3493 sub
->pmu
->read(sub
);
3495 values
[n
++] = perf_event_count(sub
);
3496 if (read_format
& PERF_FORMAT_ID
)
3497 values
[n
++] = primary_event_id(sub
);
3499 perf_output_copy(handle
, values
, n
* sizeof(u64
));
3503 static void perf_output_read(struct perf_output_handle
*handle
,
3504 struct perf_event
*event
)
3506 if (event
->attr
.read_format
& PERF_FORMAT_GROUP
)
3507 perf_output_read_group(handle
, event
);
3509 perf_output_read_one(handle
, event
);
3512 void perf_output_sample(struct perf_output_handle
*handle
,
3513 struct perf_event_header
*header
,
3514 struct perf_sample_data
*data
,
3515 struct perf_event
*event
)
3517 u64 sample_type
= data
->type
;
3519 perf_output_put(handle
, *header
);
3521 if (sample_type
& PERF_SAMPLE_IP
)
3522 perf_output_put(handle
, data
->ip
);
3524 if (sample_type
& PERF_SAMPLE_TID
)
3525 perf_output_put(handle
, data
->tid_entry
);
3527 if (sample_type
& PERF_SAMPLE_TIME
)
3528 perf_output_put(handle
, data
->time
);
3530 if (sample_type
& PERF_SAMPLE_ADDR
)
3531 perf_output_put(handle
, data
->addr
);
3533 if (sample_type
& PERF_SAMPLE_ID
)
3534 perf_output_put(handle
, data
->id
);
3536 if (sample_type
& PERF_SAMPLE_STREAM_ID
)
3537 perf_output_put(handle
, data
->stream_id
);
3539 if (sample_type
& PERF_SAMPLE_CPU
)
3540 perf_output_put(handle
, data
->cpu_entry
);
3542 if (sample_type
& PERF_SAMPLE_PERIOD
)
3543 perf_output_put(handle
, data
->period
);
3545 if (sample_type
& PERF_SAMPLE_READ
)
3546 perf_output_read(handle
, event
);
3548 if (sample_type
& PERF_SAMPLE_CALLCHAIN
) {
3549 if (data
->callchain
) {
3552 if (data
->callchain
)
3553 size
+= data
->callchain
->nr
;
3555 size
*= sizeof(u64
);
3557 perf_output_copy(handle
, data
->callchain
, size
);
3560 perf_output_put(handle
, nr
);
3564 if (sample_type
& PERF_SAMPLE_RAW
) {
3566 perf_output_put(handle
, data
->raw
->size
);
3567 perf_output_copy(handle
, data
->raw
->data
,
3574 .size
= sizeof(u32
),
3577 perf_output_put(handle
, raw
);
3582 void perf_prepare_sample(struct perf_event_header
*header
,
3583 struct perf_sample_data
*data
,
3584 struct perf_event
*event
,
3585 struct pt_regs
*regs
)
3587 u64 sample_type
= event
->attr
.sample_type
;
3589 data
->type
= sample_type
;
3591 header
->type
= PERF_RECORD_SAMPLE
;
3592 header
->size
= sizeof(*header
);
3595 header
->misc
|= perf_misc_flags(regs
);
3597 if (sample_type
& PERF_SAMPLE_IP
) {
3598 data
->ip
= perf_instruction_pointer(regs
);
3600 header
->size
+= sizeof(data
->ip
);
3603 if (sample_type
& PERF_SAMPLE_TID
) {
3604 /* namespace issues */
3605 data
->tid_entry
.pid
= perf_event_pid(event
, current
);
3606 data
->tid_entry
.tid
= perf_event_tid(event
, current
);
3608 header
->size
+= sizeof(data
->tid_entry
);
3611 if (sample_type
& PERF_SAMPLE_TIME
) {
3612 data
->time
= perf_clock();
3614 header
->size
+= sizeof(data
->time
);
3617 if (sample_type
& PERF_SAMPLE_ADDR
)
3618 header
->size
+= sizeof(data
->addr
);
3620 if (sample_type
& PERF_SAMPLE_ID
) {
3621 data
->id
= primary_event_id(event
);
3623 header
->size
+= sizeof(data
->id
);
3626 if (sample_type
& PERF_SAMPLE_STREAM_ID
) {
3627 data
->stream_id
= event
->id
;
3629 header
->size
+= sizeof(data
->stream_id
);
3632 if (sample_type
& PERF_SAMPLE_CPU
) {
3633 data
->cpu_entry
.cpu
= raw_smp_processor_id();
3634 data
->cpu_entry
.reserved
= 0;
3636 header
->size
+= sizeof(data
->cpu_entry
);
3639 if (sample_type
& PERF_SAMPLE_PERIOD
)
3640 header
->size
+= sizeof(data
->period
);
3642 if (sample_type
& PERF_SAMPLE_READ
)
3643 header
->size
+= perf_event_read_size(event
);
3645 if (sample_type
& PERF_SAMPLE_CALLCHAIN
) {
3648 data
->callchain
= perf_callchain(regs
);
3650 if (data
->callchain
)
3651 size
+= data
->callchain
->nr
;
3653 header
->size
+= size
* sizeof(u64
);
3656 if (sample_type
& PERF_SAMPLE_RAW
) {
3657 int size
= sizeof(u32
);
3660 size
+= data
->raw
->size
;
3662 size
+= sizeof(u32
);
3664 WARN_ON_ONCE(size
& (sizeof(u64
)-1));
3665 header
->size
+= size
;
3669 static void perf_event_output(struct perf_event
*event
, int nmi
,
3670 struct perf_sample_data
*data
,
3671 struct pt_regs
*regs
)
3673 struct perf_output_handle handle
;
3674 struct perf_event_header header
;
3676 /* protect the callchain buffers */
3679 perf_prepare_sample(&header
, data
, event
, regs
);
3681 if (perf_output_begin(&handle
, event
, header
.size
, nmi
, 1))
3684 perf_output_sample(&handle
, &header
, data
, event
);
3686 perf_output_end(&handle
);
3696 struct perf_read_event
{
3697 struct perf_event_header header
;
3704 perf_event_read_event(struct perf_event
*event
,
3705 struct task_struct
*task
)
3707 struct perf_output_handle handle
;
3708 struct perf_read_event read_event
= {
3710 .type
= PERF_RECORD_READ
,
3712 .size
= sizeof(read_event
) + perf_event_read_size(event
),
3714 .pid
= perf_event_pid(event
, task
),
3715 .tid
= perf_event_tid(event
, task
),
3719 ret
= perf_output_begin(&handle
, event
, read_event
.header
.size
, 0, 0);
3723 perf_output_put(&handle
, read_event
);
3724 perf_output_read(&handle
, event
);
3726 perf_output_end(&handle
);
3730 * task tracking -- fork/exit
3732 * enabled by: attr.comm | attr.mmap | attr.mmap_data | attr.task
3735 struct perf_task_event
{
3736 struct task_struct
*task
;
3737 struct perf_event_context
*task_ctx
;
3740 struct perf_event_header header
;
3750 static void perf_event_task_output(struct perf_event
*event
,
3751 struct perf_task_event
*task_event
)
3753 struct perf_output_handle handle
;
3754 struct task_struct
*task
= task_event
->task
;
3757 size
= task_event
->event_id
.header
.size
;
3758 ret
= perf_output_begin(&handle
, event
, size
, 0, 0);
3763 task_event
->event_id
.pid
= perf_event_pid(event
, task
);
3764 task_event
->event_id
.ppid
= perf_event_pid(event
, current
);
3766 task_event
->event_id
.tid
= perf_event_tid(event
, task
);
3767 task_event
->event_id
.ptid
= perf_event_tid(event
, current
);
3769 perf_output_put(&handle
, task_event
->event_id
);
3771 perf_output_end(&handle
);
3774 static int perf_event_task_match(struct perf_event
*event
)
3776 if (event
->state
< PERF_EVENT_STATE_INACTIVE
)
3779 if (event
->cpu
!= -1 && event
->cpu
!= smp_processor_id())
3782 if (event
->attr
.comm
|| event
->attr
.mmap
||
3783 event
->attr
.mmap_data
|| event
->attr
.task
)
3789 static void perf_event_task_ctx(struct perf_event_context
*ctx
,
3790 struct perf_task_event
*task_event
)
3792 struct perf_event
*event
;
3794 list_for_each_entry_rcu(event
, &ctx
->event_list
, event_entry
) {
3795 if (perf_event_task_match(event
))
3796 perf_event_task_output(event
, task_event
);
3800 static void perf_event_task_event(struct perf_task_event
*task_event
)
3802 struct perf_cpu_context
*cpuctx
;
3803 struct perf_event_context
*ctx
;
3808 list_for_each_entry_rcu(pmu
, &pmus
, entry
) {
3809 cpuctx
= get_cpu_ptr(pmu
->pmu_cpu_context
);
3810 perf_event_task_ctx(&cpuctx
->ctx
, task_event
);
3812 ctx
= task_event
->task_ctx
;
3814 ctxn
= pmu
->task_ctx_nr
;
3817 ctx
= rcu_dereference(current
->perf_event_ctxp
[ctxn
]);
3820 perf_event_task_ctx(ctx
, task_event
);
3822 put_cpu_ptr(pmu
->pmu_cpu_context
);
3827 static void perf_event_task(struct task_struct
*task
,
3828 struct perf_event_context
*task_ctx
,
3831 struct perf_task_event task_event
;
3833 if (!atomic_read(&nr_comm_events
) &&
3834 !atomic_read(&nr_mmap_events
) &&
3835 !atomic_read(&nr_task_events
))
3838 task_event
= (struct perf_task_event
){
3840 .task_ctx
= task_ctx
,
3843 .type
= new ? PERF_RECORD_FORK
: PERF_RECORD_EXIT
,
3845 .size
= sizeof(task_event
.event_id
),
3851 .time
= perf_clock(),
3855 perf_event_task_event(&task_event
);
3858 void perf_event_fork(struct task_struct
*task
)
3860 perf_event_task(task
, NULL
, 1);
3867 struct perf_comm_event
{
3868 struct task_struct
*task
;
3873 struct perf_event_header header
;
3880 static void perf_event_comm_output(struct perf_event
*event
,
3881 struct perf_comm_event
*comm_event
)
3883 struct perf_output_handle handle
;
3884 int size
= comm_event
->event_id
.header
.size
;
3885 int ret
= perf_output_begin(&handle
, event
, size
, 0, 0);
3890 comm_event
->event_id
.pid
= perf_event_pid(event
, comm_event
->task
);
3891 comm_event
->event_id
.tid
= perf_event_tid(event
, comm_event
->task
);
3893 perf_output_put(&handle
, comm_event
->event_id
);
3894 perf_output_copy(&handle
, comm_event
->comm
,
3895 comm_event
->comm_size
);
3896 perf_output_end(&handle
);
3899 static int perf_event_comm_match(struct perf_event
*event
)
3901 if (event
->state
< PERF_EVENT_STATE_INACTIVE
)
3904 if (event
->cpu
!= -1 && event
->cpu
!= smp_processor_id())
3907 if (event
->attr
.comm
)
3913 static void perf_event_comm_ctx(struct perf_event_context
*ctx
,
3914 struct perf_comm_event
*comm_event
)
3916 struct perf_event
*event
;
3918 list_for_each_entry_rcu(event
, &ctx
->event_list
, event_entry
) {
3919 if (perf_event_comm_match(event
))
3920 perf_event_comm_output(event
, comm_event
);
3924 static void perf_event_comm_event(struct perf_comm_event
*comm_event
)
3926 struct perf_cpu_context
*cpuctx
;
3927 struct perf_event_context
*ctx
;
3928 char comm
[TASK_COMM_LEN
];
3933 memset(comm
, 0, sizeof(comm
));
3934 strlcpy(comm
, comm_event
->task
->comm
, sizeof(comm
));
3935 size
= ALIGN(strlen(comm
)+1, sizeof(u64
));
3937 comm_event
->comm
= comm
;
3938 comm_event
->comm_size
= size
;
3940 comm_event
->event_id
.header
.size
= sizeof(comm_event
->event_id
) + size
;
3943 list_for_each_entry_rcu(pmu
, &pmus
, entry
) {
3944 cpuctx
= get_cpu_ptr(pmu
->pmu_cpu_context
);
3945 perf_event_comm_ctx(&cpuctx
->ctx
, comm_event
);
3947 ctxn
= pmu
->task_ctx_nr
;
3951 ctx
= rcu_dereference(current
->perf_event_ctxp
[ctxn
]);
3953 perf_event_comm_ctx(ctx
, comm_event
);
3955 put_cpu_ptr(pmu
->pmu_cpu_context
);
3960 void perf_event_comm(struct task_struct
*task
)
3962 struct perf_comm_event comm_event
;
3963 struct perf_event_context
*ctx
;
3966 for_each_task_context_nr(ctxn
) {
3967 ctx
= task
->perf_event_ctxp
[ctxn
];
3971 perf_event_enable_on_exec(ctx
);
3974 if (!atomic_read(&nr_comm_events
))
3977 comm_event
= (struct perf_comm_event
){
3983 .type
= PERF_RECORD_COMM
,
3992 perf_event_comm_event(&comm_event
);
3999 struct perf_mmap_event
{
4000 struct vm_area_struct
*vma
;
4002 const char *file_name
;
4006 struct perf_event_header header
;
4016 static void perf_event_mmap_output(struct perf_event
*event
,
4017 struct perf_mmap_event
*mmap_event
)
4019 struct perf_output_handle handle
;
4020 int size
= mmap_event
->event_id
.header
.size
;
4021 int ret
= perf_output_begin(&handle
, event
, size
, 0, 0);
4026 mmap_event
->event_id
.pid
= perf_event_pid(event
, current
);
4027 mmap_event
->event_id
.tid
= perf_event_tid(event
, current
);
4029 perf_output_put(&handle
, mmap_event
->event_id
);
4030 perf_output_copy(&handle
, mmap_event
->file_name
,
4031 mmap_event
->file_size
);
4032 perf_output_end(&handle
);
4035 static int perf_event_mmap_match(struct perf_event
*event
,
4036 struct perf_mmap_event
*mmap_event
,
4039 if (event
->state
< PERF_EVENT_STATE_INACTIVE
)
4042 if (event
->cpu
!= -1 && event
->cpu
!= smp_processor_id())
4045 if ((!executable
&& event
->attr
.mmap_data
) ||
4046 (executable
&& event
->attr
.mmap
))
4052 static void perf_event_mmap_ctx(struct perf_event_context
*ctx
,
4053 struct perf_mmap_event
*mmap_event
,
4056 struct perf_event
*event
;
4058 list_for_each_entry_rcu(event
, &ctx
->event_list
, event_entry
) {
4059 if (perf_event_mmap_match(event
, mmap_event
, executable
))
4060 perf_event_mmap_output(event
, mmap_event
);
4064 static void perf_event_mmap_event(struct perf_mmap_event
*mmap_event
)
4066 struct perf_cpu_context
*cpuctx
;
4067 struct perf_event_context
*ctx
;
4068 struct vm_area_struct
*vma
= mmap_event
->vma
;
4069 struct file
*file
= vma
->vm_file
;
4077 memset(tmp
, 0, sizeof(tmp
));
4081 * d_path works from the end of the buffer backwards, so we
4082 * need to add enough zero bytes after the string to handle
4083 * the 64bit alignment we do later.
4085 buf
= kzalloc(PATH_MAX
+ sizeof(u64
), GFP_KERNEL
);
4087 name
= strncpy(tmp
, "//enomem", sizeof(tmp
));
4090 name
= d_path(&file
->f_path
, buf
, PATH_MAX
);
4092 name
= strncpy(tmp
, "//toolong", sizeof(tmp
));
4096 if (arch_vma_name(mmap_event
->vma
)) {
4097 name
= strncpy(tmp
, arch_vma_name(mmap_event
->vma
),
4103 name
= strncpy(tmp
, "[vdso]", sizeof(tmp
));
4105 } else if (vma
->vm_start
<= vma
->vm_mm
->start_brk
&&
4106 vma
->vm_end
>= vma
->vm_mm
->brk
) {
4107 name
= strncpy(tmp
, "[heap]", sizeof(tmp
));
4109 } else if (vma
->vm_start
<= vma
->vm_mm
->start_stack
&&
4110 vma
->vm_end
>= vma
->vm_mm
->start_stack
) {
4111 name
= strncpy(tmp
, "[stack]", sizeof(tmp
));
4115 name
= strncpy(tmp
, "//anon", sizeof(tmp
));
4120 size
= ALIGN(strlen(name
)+1, sizeof(u64
));
4122 mmap_event
->file_name
= name
;
4123 mmap_event
->file_size
= size
;
4125 mmap_event
->event_id
.header
.size
= sizeof(mmap_event
->event_id
) + size
;
4128 list_for_each_entry_rcu(pmu
, &pmus
, entry
) {
4129 cpuctx
= get_cpu_ptr(pmu
->pmu_cpu_context
);
4130 perf_event_mmap_ctx(&cpuctx
->ctx
, mmap_event
,
4131 vma
->vm_flags
& VM_EXEC
);
4133 ctxn
= pmu
->task_ctx_nr
;
4137 ctx
= rcu_dereference(current
->perf_event_ctxp
[ctxn
]);
4139 perf_event_mmap_ctx(ctx
, mmap_event
,
4140 vma
->vm_flags
& VM_EXEC
);
4143 put_cpu_ptr(pmu
->pmu_cpu_context
);
4150 void perf_event_mmap(struct vm_area_struct
*vma
)
4152 struct perf_mmap_event mmap_event
;
4154 if (!atomic_read(&nr_mmap_events
))
4157 mmap_event
= (struct perf_mmap_event
){
4163 .type
= PERF_RECORD_MMAP
,
4164 .misc
= PERF_RECORD_MISC_USER
,
4169 .start
= vma
->vm_start
,
4170 .len
= vma
->vm_end
- vma
->vm_start
,
4171 .pgoff
= (u64
)vma
->vm_pgoff
<< PAGE_SHIFT
,
4175 perf_event_mmap_event(&mmap_event
);
4179 * IRQ throttle logging
4182 static void perf_log_throttle(struct perf_event
*event
, int enable
)
4184 struct perf_output_handle handle
;
4188 struct perf_event_header header
;
4192 } throttle_event
= {
4194 .type
= PERF_RECORD_THROTTLE
,
4196 .size
= sizeof(throttle_event
),
4198 .time
= perf_clock(),
4199 .id
= primary_event_id(event
),
4200 .stream_id
= event
->id
,
4204 throttle_event
.header
.type
= PERF_RECORD_UNTHROTTLE
;
4206 ret
= perf_output_begin(&handle
, event
, sizeof(throttle_event
), 1, 0);
4210 perf_output_put(&handle
, throttle_event
);
4211 perf_output_end(&handle
);
4215 * Generic event overflow handling, sampling.
4218 static int __perf_event_overflow(struct perf_event
*event
, int nmi
,
4219 int throttle
, struct perf_sample_data
*data
,
4220 struct pt_regs
*regs
)
4222 int events
= atomic_read(&event
->event_limit
);
4223 struct hw_perf_event
*hwc
= &event
->hw
;
4229 if (hwc
->interrupts
!= MAX_INTERRUPTS
) {
4231 if (HZ
* hwc
->interrupts
>
4232 (u64
)sysctl_perf_event_sample_rate
) {
4233 hwc
->interrupts
= MAX_INTERRUPTS
;
4234 perf_log_throttle(event
, 0);
4239 * Keep re-disabling events even though on the previous
4240 * pass we disabled it - just in case we raced with a
4241 * sched-in and the event got enabled again:
4247 if (event
->attr
.freq
) {
4248 u64 now
= perf_clock();
4249 s64 delta
= now
- hwc
->freq_time_stamp
;
4251 hwc
->freq_time_stamp
= now
;
4253 if (delta
> 0 && delta
< 2*TICK_NSEC
)
4254 perf_adjust_period(event
, delta
, hwc
->last_period
);
4258 * XXX event_limit might not quite work as expected on inherited
4262 event
->pending_kill
= POLL_IN
;
4263 if (events
&& atomic_dec_and_test(&event
->event_limit
)) {
4265 event
->pending_kill
= POLL_HUP
;
4267 event
->pending_disable
= 1;
4268 irq_work_queue(&event
->pending
);
4270 perf_event_disable(event
);
4273 if (event
->overflow_handler
)
4274 event
->overflow_handler(event
, nmi
, data
, regs
);
4276 perf_event_output(event
, nmi
, data
, regs
);
4281 int perf_event_overflow(struct perf_event
*event
, int nmi
,
4282 struct perf_sample_data
*data
,
4283 struct pt_regs
*regs
)
4285 return __perf_event_overflow(event
, nmi
, 1, data
, regs
);
4289 * Generic software event infrastructure
4292 struct swevent_htable
{
4293 struct swevent_hlist
*swevent_hlist
;
4294 struct mutex hlist_mutex
;
4297 /* Recursion avoidance in each contexts */
4298 int recursion
[PERF_NR_CONTEXTS
];
4301 static DEFINE_PER_CPU(struct swevent_htable
, swevent_htable
);
4304 * We directly increment event->count and keep a second value in
4305 * event->hw.period_left to count intervals. This period event
4306 * is kept in the range [-sample_period, 0] so that we can use the
4310 static u64
perf_swevent_set_period(struct perf_event
*event
)
4312 struct hw_perf_event
*hwc
= &event
->hw
;
4313 u64 period
= hwc
->last_period
;
4317 hwc
->last_period
= hwc
->sample_period
;
4320 old
= val
= local64_read(&hwc
->period_left
);
4324 nr
= div64_u64(period
+ val
, period
);
4325 offset
= nr
* period
;
4327 if (local64_cmpxchg(&hwc
->period_left
, old
, val
) != old
)
4333 static void perf_swevent_overflow(struct perf_event
*event
, u64 overflow
,
4334 int nmi
, struct perf_sample_data
*data
,
4335 struct pt_regs
*regs
)
4337 struct hw_perf_event
*hwc
= &event
->hw
;
4340 data
->period
= event
->hw
.last_period
;
4342 overflow
= perf_swevent_set_period(event
);
4344 if (hwc
->interrupts
== MAX_INTERRUPTS
)
4347 for (; overflow
; overflow
--) {
4348 if (__perf_event_overflow(event
, nmi
, throttle
,
4351 * We inhibit the overflow from happening when
4352 * hwc->interrupts == MAX_INTERRUPTS.
4360 static void perf_swevent_event(struct perf_event
*event
, u64 nr
,
4361 int nmi
, struct perf_sample_data
*data
,
4362 struct pt_regs
*regs
)
4364 struct hw_perf_event
*hwc
= &event
->hw
;
4366 local64_add(nr
, &event
->count
);
4371 if (!hwc
->sample_period
)
4374 if (nr
== 1 && hwc
->sample_period
== 1 && !event
->attr
.freq
)
4375 return perf_swevent_overflow(event
, 1, nmi
, data
, regs
);
4377 if (local64_add_negative(nr
, &hwc
->period_left
))
4380 perf_swevent_overflow(event
, 0, nmi
, data
, regs
);
4383 static int perf_exclude_event(struct perf_event
*event
,
4384 struct pt_regs
*regs
)
4386 if (event
->hw
.state
& PERF_HES_STOPPED
)
4390 if (event
->attr
.exclude_user
&& user_mode(regs
))
4393 if (event
->attr
.exclude_kernel
&& !user_mode(regs
))
4400 static int perf_swevent_match(struct perf_event
*event
,
4401 enum perf_type_id type
,
4403 struct perf_sample_data
*data
,
4404 struct pt_regs
*regs
)
4406 if (event
->attr
.type
!= type
)
4409 if (event
->attr
.config
!= event_id
)
4412 if (perf_exclude_event(event
, regs
))
4418 static inline u64
swevent_hash(u64 type
, u32 event_id
)
4420 u64 val
= event_id
| (type
<< 32);
4422 return hash_64(val
, SWEVENT_HLIST_BITS
);
4425 static inline struct hlist_head
*
4426 __find_swevent_head(struct swevent_hlist
*hlist
, u64 type
, u32 event_id
)
4428 u64 hash
= swevent_hash(type
, event_id
);
4430 return &hlist
->heads
[hash
];
4433 /* For the read side: events when they trigger */
4434 static inline struct hlist_head
*
4435 find_swevent_head_rcu(struct swevent_htable
*swhash
, u64 type
, u32 event_id
)
4437 struct swevent_hlist
*hlist
;
4439 hlist
= rcu_dereference(swhash
->swevent_hlist
);
4443 return __find_swevent_head(hlist
, type
, event_id
);
4446 /* For the event head insertion and removal in the hlist */
4447 static inline struct hlist_head
*
4448 find_swevent_head(struct swevent_htable
*swhash
, struct perf_event
*event
)
4450 struct swevent_hlist
*hlist
;
4451 u32 event_id
= event
->attr
.config
;
4452 u64 type
= event
->attr
.type
;
4455 * Event scheduling is always serialized against hlist allocation
4456 * and release. Which makes the protected version suitable here.
4457 * The context lock guarantees that.
4459 hlist
= rcu_dereference_protected(swhash
->swevent_hlist
,
4460 lockdep_is_held(&event
->ctx
->lock
));
4464 return __find_swevent_head(hlist
, type
, event_id
);
4467 static void do_perf_sw_event(enum perf_type_id type
, u32 event_id
,
4469 struct perf_sample_data
*data
,
4470 struct pt_regs
*regs
)
4472 struct swevent_htable
*swhash
= &__get_cpu_var(swevent_htable
);
4473 struct perf_event
*event
;
4474 struct hlist_node
*node
;
4475 struct hlist_head
*head
;
4478 head
= find_swevent_head_rcu(swhash
, type
, event_id
);
4482 hlist_for_each_entry_rcu(event
, node
, head
, hlist_entry
) {
4483 if (perf_swevent_match(event
, type
, event_id
, data
, regs
))
4484 perf_swevent_event(event
, nr
, nmi
, data
, regs
);
4490 int perf_swevent_get_recursion_context(void)
4492 struct swevent_htable
*swhash
= &__get_cpu_var(swevent_htable
);
4494 return get_recursion_context(swhash
->recursion
);
4496 EXPORT_SYMBOL_GPL(perf_swevent_get_recursion_context
);
4498 void inline perf_swevent_put_recursion_context(int rctx
)
4500 struct swevent_htable
*swhash
= &__get_cpu_var(swevent_htable
);
4502 put_recursion_context(swhash
->recursion
, rctx
);
4505 void __perf_sw_event(u32 event_id
, u64 nr
, int nmi
,
4506 struct pt_regs
*regs
, u64 addr
)
4508 struct perf_sample_data data
;
4511 preempt_disable_notrace();
4512 rctx
= perf_swevent_get_recursion_context();
4516 perf_sample_data_init(&data
, addr
);
4518 do_perf_sw_event(PERF_TYPE_SOFTWARE
, event_id
, nr
, nmi
, &data
, regs
);
4520 perf_swevent_put_recursion_context(rctx
);
4521 preempt_enable_notrace();
4524 static void perf_swevent_read(struct perf_event
*event
)
4528 static int perf_swevent_add(struct perf_event
*event
, int flags
)
4530 struct swevent_htable
*swhash
= &__get_cpu_var(swevent_htable
);
4531 struct hw_perf_event
*hwc
= &event
->hw
;
4532 struct hlist_head
*head
;
4534 if (hwc
->sample_period
) {
4535 hwc
->last_period
= hwc
->sample_period
;
4536 perf_swevent_set_period(event
);
4539 hwc
->state
= !(flags
& PERF_EF_START
);
4541 head
= find_swevent_head(swhash
, event
);
4542 if (WARN_ON_ONCE(!head
))
4545 hlist_add_head_rcu(&event
->hlist_entry
, head
);
4550 static void perf_swevent_del(struct perf_event
*event
, int flags
)
4552 hlist_del_rcu(&event
->hlist_entry
);
4555 static void perf_swevent_start(struct perf_event
*event
, int flags
)
4557 event
->hw
.state
= 0;
4560 static void perf_swevent_stop(struct perf_event
*event
, int flags
)
4562 event
->hw
.state
= PERF_HES_STOPPED
;
4565 /* Deref the hlist from the update side */
4566 static inline struct swevent_hlist
*
4567 swevent_hlist_deref(struct swevent_htable
*swhash
)
4569 return rcu_dereference_protected(swhash
->swevent_hlist
,
4570 lockdep_is_held(&swhash
->hlist_mutex
));
4573 static void swevent_hlist_release_rcu(struct rcu_head
*rcu_head
)
4575 struct swevent_hlist
*hlist
;
4577 hlist
= container_of(rcu_head
, struct swevent_hlist
, rcu_head
);
4581 static void swevent_hlist_release(struct swevent_htable
*swhash
)
4583 struct swevent_hlist
*hlist
= swevent_hlist_deref(swhash
);
4588 rcu_assign_pointer(swhash
->swevent_hlist
, NULL
);
4589 call_rcu(&hlist
->rcu_head
, swevent_hlist_release_rcu
);
4592 static void swevent_hlist_put_cpu(struct perf_event
*event
, int cpu
)
4594 struct swevent_htable
*swhash
= &per_cpu(swevent_htable
, cpu
);
4596 mutex_lock(&swhash
->hlist_mutex
);
4598 if (!--swhash
->hlist_refcount
)
4599 swevent_hlist_release(swhash
);
4601 mutex_unlock(&swhash
->hlist_mutex
);
4604 static void swevent_hlist_put(struct perf_event
*event
)
4608 if (event
->cpu
!= -1) {
4609 swevent_hlist_put_cpu(event
, event
->cpu
);
4613 for_each_possible_cpu(cpu
)
4614 swevent_hlist_put_cpu(event
, cpu
);
4617 static int swevent_hlist_get_cpu(struct perf_event
*event
, int cpu
)
4619 struct swevent_htable
*swhash
= &per_cpu(swevent_htable
, cpu
);
4622 mutex_lock(&swhash
->hlist_mutex
);
4624 if (!swevent_hlist_deref(swhash
) && cpu_online(cpu
)) {
4625 struct swevent_hlist
*hlist
;
4627 hlist
= kzalloc(sizeof(*hlist
), GFP_KERNEL
);
4632 rcu_assign_pointer(swhash
->swevent_hlist
, hlist
);
4634 swhash
->hlist_refcount
++;
4636 mutex_unlock(&swhash
->hlist_mutex
);
4641 static int swevent_hlist_get(struct perf_event
*event
)
4644 int cpu
, failed_cpu
;
4646 if (event
->cpu
!= -1)
4647 return swevent_hlist_get_cpu(event
, event
->cpu
);
4650 for_each_possible_cpu(cpu
) {
4651 err
= swevent_hlist_get_cpu(event
, cpu
);
4661 for_each_possible_cpu(cpu
) {
4662 if (cpu
== failed_cpu
)
4664 swevent_hlist_put_cpu(event
, cpu
);
4671 atomic_t perf_swevent_enabled
[PERF_COUNT_SW_MAX
];
4673 static void sw_perf_event_destroy(struct perf_event
*event
)
4675 u64 event_id
= event
->attr
.config
;
4677 WARN_ON(event
->parent
);
4679 atomic_dec(&perf_swevent_enabled
[event_id
]);
4680 swevent_hlist_put(event
);
4683 static int perf_swevent_init(struct perf_event
*event
)
4685 int event_id
= event
->attr
.config
;
4687 if (event
->attr
.type
!= PERF_TYPE_SOFTWARE
)
4691 case PERF_COUNT_SW_CPU_CLOCK
:
4692 case PERF_COUNT_SW_TASK_CLOCK
:
4699 if (event_id
> PERF_COUNT_SW_MAX
)
4702 if (!event
->parent
) {
4705 err
= swevent_hlist_get(event
);
4709 atomic_inc(&perf_swevent_enabled
[event_id
]);
4710 event
->destroy
= sw_perf_event_destroy
;
4716 static struct pmu perf_swevent
= {
4717 .task_ctx_nr
= perf_sw_context
,
4719 .event_init
= perf_swevent_init
,
4720 .add
= perf_swevent_add
,
4721 .del
= perf_swevent_del
,
4722 .start
= perf_swevent_start
,
4723 .stop
= perf_swevent_stop
,
4724 .read
= perf_swevent_read
,
4727 #ifdef CONFIG_EVENT_TRACING
4729 static int perf_tp_filter_match(struct perf_event
*event
,
4730 struct perf_sample_data
*data
)
4732 void *record
= data
->raw
->data
;
4734 if (likely(!event
->filter
) || filter_match_preds(event
->filter
, record
))
4739 static int perf_tp_event_match(struct perf_event
*event
,
4740 struct perf_sample_data
*data
,
4741 struct pt_regs
*regs
)
4744 * All tracepoints are from kernel-space.
4746 if (event
->attr
.exclude_kernel
)
4749 if (!perf_tp_filter_match(event
, data
))
4755 void perf_tp_event(u64 addr
, u64 count
, void *record
, int entry_size
,
4756 struct pt_regs
*regs
, struct hlist_head
*head
, int rctx
)
4758 struct perf_sample_data data
;
4759 struct perf_event
*event
;
4760 struct hlist_node
*node
;
4762 struct perf_raw_record raw
= {
4767 perf_sample_data_init(&data
, addr
);
4770 hlist_for_each_entry_rcu(event
, node
, head
, hlist_entry
) {
4771 if (perf_tp_event_match(event
, &data
, regs
))
4772 perf_swevent_event(event
, count
, 1, &data
, regs
);
4775 perf_swevent_put_recursion_context(rctx
);
4777 EXPORT_SYMBOL_GPL(perf_tp_event
);
4779 static void tp_perf_event_destroy(struct perf_event
*event
)
4781 perf_trace_destroy(event
);
4784 static int perf_tp_event_init(struct perf_event
*event
)
4788 if (event
->attr
.type
!= PERF_TYPE_TRACEPOINT
)
4792 * Raw tracepoint data is a severe data leak, only allow root to
4795 if ((event
->attr
.sample_type
& PERF_SAMPLE_RAW
) &&
4796 perf_paranoid_tracepoint_raw() &&
4797 !capable(CAP_SYS_ADMIN
))
4800 err
= perf_trace_init(event
);
4804 event
->destroy
= tp_perf_event_destroy
;
4809 static struct pmu perf_tracepoint
= {
4810 .task_ctx_nr
= perf_sw_context
,
4812 .event_init
= perf_tp_event_init
,
4813 .add
= perf_trace_add
,
4814 .del
= perf_trace_del
,
4815 .start
= perf_swevent_start
,
4816 .stop
= perf_swevent_stop
,
4817 .read
= perf_swevent_read
,
4820 static inline void perf_tp_register(void)
4822 perf_pmu_register(&perf_tracepoint
);
4825 static int perf_event_set_filter(struct perf_event
*event
, void __user
*arg
)
4830 if (event
->attr
.type
!= PERF_TYPE_TRACEPOINT
)
4833 filter_str
= strndup_user(arg
, PAGE_SIZE
);
4834 if (IS_ERR(filter_str
))
4835 return PTR_ERR(filter_str
);
4837 ret
= ftrace_profile_set_filter(event
, event
->attr
.config
, filter_str
);
4843 static void perf_event_free_filter(struct perf_event
*event
)
4845 ftrace_profile_free_filter(event
);
4850 static inline void perf_tp_register(void)
4854 static int perf_event_set_filter(struct perf_event
*event
, void __user
*arg
)
4859 static void perf_event_free_filter(struct perf_event
*event
)
4863 #endif /* CONFIG_EVENT_TRACING */
4865 #ifdef CONFIG_HAVE_HW_BREAKPOINT
4866 void perf_bp_event(struct perf_event
*bp
, void *data
)
4868 struct perf_sample_data sample
;
4869 struct pt_regs
*regs
= data
;
4871 perf_sample_data_init(&sample
, bp
->attr
.bp_addr
);
4873 if (!bp
->hw
.state
&& !perf_exclude_event(bp
, regs
))
4874 perf_swevent_event(bp
, 1, 1, &sample
, regs
);
4879 * hrtimer based swevent callback
4882 static enum hrtimer_restart
perf_swevent_hrtimer(struct hrtimer
*hrtimer
)
4884 enum hrtimer_restart ret
= HRTIMER_RESTART
;
4885 struct perf_sample_data data
;
4886 struct pt_regs
*regs
;
4887 struct perf_event
*event
;
4890 event
= container_of(hrtimer
, struct perf_event
, hw
.hrtimer
);
4891 event
->pmu
->read(event
);
4893 perf_sample_data_init(&data
, 0);
4894 data
.period
= event
->hw
.last_period
;
4895 regs
= get_irq_regs();
4897 if (regs
&& !perf_exclude_event(event
, regs
)) {
4898 if (!(event
->attr
.exclude_idle
&& current
->pid
== 0))
4899 if (perf_event_overflow(event
, 0, &data
, regs
))
4900 ret
= HRTIMER_NORESTART
;
4903 period
= max_t(u64
, 10000, event
->hw
.sample_period
);
4904 hrtimer_forward_now(hrtimer
, ns_to_ktime(period
));
4909 static void perf_swevent_start_hrtimer(struct perf_event
*event
)
4911 struct hw_perf_event
*hwc
= &event
->hw
;
4913 hrtimer_init(&hwc
->hrtimer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
4914 hwc
->hrtimer
.function
= perf_swevent_hrtimer
;
4915 if (hwc
->sample_period
) {
4916 s64 period
= local64_read(&hwc
->period_left
);
4922 local64_set(&hwc
->period_left
, 0);
4924 period
= max_t(u64
, 10000, hwc
->sample_period
);
4926 __hrtimer_start_range_ns(&hwc
->hrtimer
,
4927 ns_to_ktime(period
), 0,
4928 HRTIMER_MODE_REL_PINNED
, 0);
4932 static void perf_swevent_cancel_hrtimer(struct perf_event
*event
)
4934 struct hw_perf_event
*hwc
= &event
->hw
;
4936 if (hwc
->sample_period
) {
4937 ktime_t remaining
= hrtimer_get_remaining(&hwc
->hrtimer
);
4938 local64_set(&hwc
->period_left
, ktime_to_ns(remaining
));
4940 hrtimer_cancel(&hwc
->hrtimer
);
4945 * Software event: cpu wall time clock
4948 static void cpu_clock_event_update(struct perf_event
*event
)
4953 now
= local_clock();
4954 prev
= local64_xchg(&event
->hw
.prev_count
, now
);
4955 local64_add(now
- prev
, &event
->count
);
4958 static void cpu_clock_event_start(struct perf_event
*event
, int flags
)
4960 local64_set(&event
->hw
.prev_count
, local_clock());
4961 perf_swevent_start_hrtimer(event
);
4964 static void cpu_clock_event_stop(struct perf_event
*event
, int flags
)
4966 perf_swevent_cancel_hrtimer(event
);
4967 cpu_clock_event_update(event
);
4970 static int cpu_clock_event_add(struct perf_event
*event
, int flags
)
4972 if (flags
& PERF_EF_START
)
4973 cpu_clock_event_start(event
, flags
);
4978 static void cpu_clock_event_del(struct perf_event
*event
, int flags
)
4980 cpu_clock_event_stop(event
, flags
);
4983 static void cpu_clock_event_read(struct perf_event
*event
)
4985 cpu_clock_event_update(event
);
4988 static int cpu_clock_event_init(struct perf_event
*event
)
4990 if (event
->attr
.type
!= PERF_TYPE_SOFTWARE
)
4993 if (event
->attr
.config
!= PERF_COUNT_SW_CPU_CLOCK
)
4999 static struct pmu perf_cpu_clock
= {
5000 .task_ctx_nr
= perf_sw_context
,
5002 .event_init
= cpu_clock_event_init
,
5003 .add
= cpu_clock_event_add
,
5004 .del
= cpu_clock_event_del
,
5005 .start
= cpu_clock_event_start
,
5006 .stop
= cpu_clock_event_stop
,
5007 .read
= cpu_clock_event_read
,
5011 * Software event: task time clock
5014 static void task_clock_event_update(struct perf_event
*event
, u64 now
)
5019 prev
= local64_xchg(&event
->hw
.prev_count
, now
);
5021 local64_add(delta
, &event
->count
);
5024 static void task_clock_event_start(struct perf_event
*event
, int flags
)
5026 local64_set(&event
->hw
.prev_count
, event
->ctx
->time
);
5027 perf_swevent_start_hrtimer(event
);
5030 static void task_clock_event_stop(struct perf_event
*event
, int flags
)
5032 perf_swevent_cancel_hrtimer(event
);
5033 task_clock_event_update(event
, event
->ctx
->time
);
5036 static int task_clock_event_add(struct perf_event
*event
, int flags
)
5038 if (flags
& PERF_EF_START
)
5039 task_clock_event_start(event
, flags
);
5044 static void task_clock_event_del(struct perf_event
*event
, int flags
)
5046 task_clock_event_stop(event
, PERF_EF_UPDATE
);
5049 static void task_clock_event_read(struct perf_event
*event
)
5054 update_context_time(event
->ctx
);
5055 time
= event
->ctx
->time
;
5057 u64 now
= perf_clock();
5058 u64 delta
= now
- event
->ctx
->timestamp
;
5059 time
= event
->ctx
->time
+ delta
;
5062 task_clock_event_update(event
, time
);
5065 static int task_clock_event_init(struct perf_event
*event
)
5067 if (event
->attr
.type
!= PERF_TYPE_SOFTWARE
)
5070 if (event
->attr
.config
!= PERF_COUNT_SW_TASK_CLOCK
)
5076 static struct pmu perf_task_clock
= {
5077 .task_ctx_nr
= perf_sw_context
,
5079 .event_init
= task_clock_event_init
,
5080 .add
= task_clock_event_add
,
5081 .del
= task_clock_event_del
,
5082 .start
= task_clock_event_start
,
5083 .stop
= task_clock_event_stop
,
5084 .read
= task_clock_event_read
,
5087 static void perf_pmu_nop_void(struct pmu
*pmu
)
5091 static int perf_pmu_nop_int(struct pmu
*pmu
)
5096 static void perf_pmu_start_txn(struct pmu
*pmu
)
5098 perf_pmu_disable(pmu
);
5101 static int perf_pmu_commit_txn(struct pmu
*pmu
)
5103 perf_pmu_enable(pmu
);
5107 static void perf_pmu_cancel_txn(struct pmu
*pmu
)
5109 perf_pmu_enable(pmu
);
5113 * Ensures all contexts with the same task_ctx_nr have the same
5114 * pmu_cpu_context too.
5116 static void *find_pmu_context(int ctxn
)
5123 list_for_each_entry(pmu
, &pmus
, entry
) {
5124 if (pmu
->task_ctx_nr
== ctxn
)
5125 return pmu
->pmu_cpu_context
;
5131 static void free_pmu_context(void * __percpu cpu_context
)
5135 mutex_lock(&pmus_lock
);
5137 * Like a real lame refcount.
5139 list_for_each_entry(pmu
, &pmus
, entry
) {
5140 if (pmu
->pmu_cpu_context
== cpu_context
)
5144 free_percpu(cpu_context
);
5146 mutex_unlock(&pmus_lock
);
5149 int perf_pmu_register(struct pmu
*pmu
)
5153 mutex_lock(&pmus_lock
);
5155 pmu
->pmu_disable_count
= alloc_percpu(int);
5156 if (!pmu
->pmu_disable_count
)
5159 pmu
->pmu_cpu_context
= find_pmu_context(pmu
->task_ctx_nr
);
5160 if (pmu
->pmu_cpu_context
)
5161 goto got_cpu_context
;
5163 pmu
->pmu_cpu_context
= alloc_percpu(struct perf_cpu_context
);
5164 if (!pmu
->pmu_cpu_context
)
5167 for_each_possible_cpu(cpu
) {
5168 struct perf_cpu_context
*cpuctx
;
5170 cpuctx
= per_cpu_ptr(pmu
->pmu_cpu_context
, cpu
);
5171 __perf_event_init_context(&cpuctx
->ctx
);
5172 cpuctx
->ctx
.type
= cpu_context
;
5173 cpuctx
->ctx
.pmu
= pmu
;
5174 cpuctx
->jiffies_interval
= 1;
5175 INIT_LIST_HEAD(&cpuctx
->rotation_list
);
5179 if (!pmu
->start_txn
) {
5180 if (pmu
->pmu_enable
) {
5182 * If we have pmu_enable/pmu_disable calls, install
5183 * transaction stubs that use that to try and batch
5184 * hardware accesses.
5186 pmu
->start_txn
= perf_pmu_start_txn
;
5187 pmu
->commit_txn
= perf_pmu_commit_txn
;
5188 pmu
->cancel_txn
= perf_pmu_cancel_txn
;
5190 pmu
->start_txn
= perf_pmu_nop_void
;
5191 pmu
->commit_txn
= perf_pmu_nop_int
;
5192 pmu
->cancel_txn
= perf_pmu_nop_void
;
5196 if (!pmu
->pmu_enable
) {
5197 pmu
->pmu_enable
= perf_pmu_nop_void
;
5198 pmu
->pmu_disable
= perf_pmu_nop_void
;
5201 list_add_rcu(&pmu
->entry
, &pmus
);
5204 mutex_unlock(&pmus_lock
);
5209 free_percpu(pmu
->pmu_disable_count
);
5213 void perf_pmu_unregister(struct pmu
*pmu
)
5215 mutex_lock(&pmus_lock
);
5216 list_del_rcu(&pmu
->entry
);
5217 mutex_unlock(&pmus_lock
);
5220 * We dereference the pmu list under both SRCU and regular RCU, so
5221 * synchronize against both of those.
5223 synchronize_srcu(&pmus_srcu
);
5226 free_percpu(pmu
->pmu_disable_count
);
5227 free_pmu_context(pmu
->pmu_cpu_context
);
5230 struct pmu
*perf_init_event(struct perf_event
*event
)
5232 struct pmu
*pmu
= NULL
;
5235 idx
= srcu_read_lock(&pmus_srcu
);
5236 list_for_each_entry_rcu(pmu
, &pmus
, entry
) {
5237 int ret
= pmu
->event_init(event
);
5241 if (ret
!= -ENOENT
) {
5246 pmu
= ERR_PTR(-ENOENT
);
5248 srcu_read_unlock(&pmus_srcu
, idx
);
5254 * Allocate and initialize a event structure
5256 static struct perf_event
*
5257 perf_event_alloc(struct perf_event_attr
*attr
, int cpu
,
5258 struct perf_event
*group_leader
,
5259 struct perf_event
*parent_event
,
5260 perf_overflow_handler_t overflow_handler
)
5263 struct perf_event
*event
;
5264 struct hw_perf_event
*hwc
;
5267 event
= kzalloc(sizeof(*event
), GFP_KERNEL
);
5269 return ERR_PTR(-ENOMEM
);
5272 * Single events are their own group leaders, with an
5273 * empty sibling list:
5276 group_leader
= event
;
5278 mutex_init(&event
->child_mutex
);
5279 INIT_LIST_HEAD(&event
->child_list
);
5281 INIT_LIST_HEAD(&event
->group_entry
);
5282 INIT_LIST_HEAD(&event
->event_entry
);
5283 INIT_LIST_HEAD(&event
->sibling_list
);
5284 init_waitqueue_head(&event
->waitq
);
5285 init_irq_work(&event
->pending
, perf_pending_event
);
5287 mutex_init(&event
->mmap_mutex
);
5290 event
->attr
= *attr
;
5291 event
->group_leader
= group_leader
;
5295 event
->parent
= parent_event
;
5297 event
->ns
= get_pid_ns(current
->nsproxy
->pid_ns
);
5298 event
->id
= atomic64_inc_return(&perf_event_id
);
5300 event
->state
= PERF_EVENT_STATE_INACTIVE
;
5302 if (!overflow_handler
&& parent_event
)
5303 overflow_handler
= parent_event
->overflow_handler
;
5305 event
->overflow_handler
= overflow_handler
;
5308 event
->state
= PERF_EVENT_STATE_OFF
;
5313 hwc
->sample_period
= attr
->sample_period
;
5314 if (attr
->freq
&& attr
->sample_freq
)
5315 hwc
->sample_period
= 1;
5316 hwc
->last_period
= hwc
->sample_period
;
5318 local64_set(&hwc
->period_left
, hwc
->sample_period
);
5321 * we currently do not support PERF_FORMAT_GROUP on inherited events
5323 if (attr
->inherit
&& (attr
->read_format
& PERF_FORMAT_GROUP
))
5326 pmu
= perf_init_event(event
);
5332 else if (IS_ERR(pmu
))
5337 put_pid_ns(event
->ns
);
5339 return ERR_PTR(err
);
5344 if (!event
->parent
) {
5345 atomic_inc(&nr_events
);
5346 if (event
->attr
.mmap
|| event
->attr
.mmap_data
)
5347 atomic_inc(&nr_mmap_events
);
5348 if (event
->attr
.comm
)
5349 atomic_inc(&nr_comm_events
);
5350 if (event
->attr
.task
)
5351 atomic_inc(&nr_task_events
);
5352 if (event
->attr
.sample_type
& PERF_SAMPLE_CALLCHAIN
) {
5353 err
= get_callchain_buffers();
5356 return ERR_PTR(err
);
5364 static int perf_copy_attr(struct perf_event_attr __user
*uattr
,
5365 struct perf_event_attr
*attr
)
5370 if (!access_ok(VERIFY_WRITE
, uattr
, PERF_ATTR_SIZE_VER0
))
5374 * zero the full structure, so that a short copy will be nice.
5376 memset(attr
, 0, sizeof(*attr
));
5378 ret
= get_user(size
, &uattr
->size
);
5382 if (size
> PAGE_SIZE
) /* silly large */
5385 if (!size
) /* abi compat */
5386 size
= PERF_ATTR_SIZE_VER0
;
5388 if (size
< PERF_ATTR_SIZE_VER0
)
5392 * If we're handed a bigger struct than we know of,
5393 * ensure all the unknown bits are 0 - i.e. new
5394 * user-space does not rely on any kernel feature
5395 * extensions we dont know about yet.
5397 if (size
> sizeof(*attr
)) {
5398 unsigned char __user
*addr
;
5399 unsigned char __user
*end
;
5402 addr
= (void __user
*)uattr
+ sizeof(*attr
);
5403 end
= (void __user
*)uattr
+ size
;
5405 for (; addr
< end
; addr
++) {
5406 ret
= get_user(val
, addr
);
5412 size
= sizeof(*attr
);
5415 ret
= copy_from_user(attr
, uattr
, size
);
5420 * If the type exists, the corresponding creation will verify
5423 if (attr
->type
>= PERF_TYPE_MAX
)
5426 if (attr
->__reserved_1
)
5429 if (attr
->sample_type
& ~(PERF_SAMPLE_MAX
-1))
5432 if (attr
->read_format
& ~(PERF_FORMAT_MAX
-1))
5439 put_user(sizeof(*attr
), &uattr
->size
);
5445 perf_event_set_output(struct perf_event
*event
, struct perf_event
*output_event
)
5447 struct perf_buffer
*buffer
= NULL
, *old_buffer
= NULL
;
5453 /* don't allow circular references */
5454 if (event
== output_event
)
5458 * Don't allow cross-cpu buffers
5460 if (output_event
->cpu
!= event
->cpu
)
5464 * If its not a per-cpu buffer, it must be the same task.
5466 if (output_event
->cpu
== -1 && output_event
->ctx
!= event
->ctx
)
5470 mutex_lock(&event
->mmap_mutex
);
5471 /* Can't redirect output if we've got an active mmap() */
5472 if (atomic_read(&event
->mmap_count
))
5476 /* get the buffer we want to redirect to */
5477 buffer
= perf_buffer_get(output_event
);
5482 old_buffer
= event
->buffer
;
5483 rcu_assign_pointer(event
->buffer
, buffer
);
5486 mutex_unlock(&event
->mmap_mutex
);
5489 perf_buffer_put(old_buffer
);
5495 * sys_perf_event_open - open a performance event, associate it to a task/cpu
5497 * @attr_uptr: event_id type attributes for monitoring/sampling
5500 * @group_fd: group leader event fd
5502 SYSCALL_DEFINE5(perf_event_open
,
5503 struct perf_event_attr __user
*, attr_uptr
,
5504 pid_t
, pid
, int, cpu
, int, group_fd
, unsigned long, flags
)
5506 struct perf_event
*group_leader
= NULL
, *output_event
= NULL
;
5507 struct perf_event
*event
, *sibling
;
5508 struct perf_event_attr attr
;
5509 struct perf_event_context
*ctx
;
5510 struct file
*event_file
= NULL
;
5511 struct file
*group_file
= NULL
;
5512 struct task_struct
*task
= NULL
;
5516 int fput_needed
= 0;
5519 /* for future expandability... */
5520 if (flags
& ~(PERF_FLAG_FD_NO_GROUP
| PERF_FLAG_FD_OUTPUT
))
5523 err
= perf_copy_attr(attr_uptr
, &attr
);
5527 if (!attr
.exclude_kernel
) {
5528 if (perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN
))
5533 if (attr
.sample_freq
> sysctl_perf_event_sample_rate
)
5537 event_fd
= get_unused_fd_flags(O_RDWR
);
5541 if (group_fd
!= -1) {
5542 group_leader
= perf_fget_light(group_fd
, &fput_needed
);
5543 if (IS_ERR(group_leader
)) {
5544 err
= PTR_ERR(group_leader
);
5547 group_file
= group_leader
->filp
;
5548 if (flags
& PERF_FLAG_FD_OUTPUT
)
5549 output_event
= group_leader
;
5550 if (flags
& PERF_FLAG_FD_NO_GROUP
)
5551 group_leader
= NULL
;
5555 task
= find_lively_task_by_vpid(pid
);
5557 err
= PTR_ERR(task
);
5562 event
= perf_event_alloc(&attr
, cpu
, group_leader
, NULL
, NULL
);
5563 if (IS_ERR(event
)) {
5564 err
= PTR_ERR(event
);
5569 * Special case software events and allow them to be part of
5570 * any hardware group.
5575 (is_software_event(event
) != is_software_event(group_leader
))) {
5576 if (is_software_event(event
)) {
5578 * If event and group_leader are not both a software
5579 * event, and event is, then group leader is not.
5581 * Allow the addition of software events to !software
5582 * groups, this is safe because software events never
5585 pmu
= group_leader
->pmu
;
5586 } else if (is_software_event(group_leader
) &&
5587 (group_leader
->group_flags
& PERF_GROUP_SOFTWARE
)) {
5589 * In case the group is a pure software group, and we
5590 * try to add a hardware event, move the whole group to
5591 * the hardware context.
5598 * Get the target context (task or percpu):
5600 ctx
= find_get_context(pmu
, task
, cpu
);
5607 * Look up the group leader (we will attach this event to it):
5613 * Do not allow a recursive hierarchy (this new sibling
5614 * becoming part of another group-sibling):
5616 if (group_leader
->group_leader
!= group_leader
)
5619 * Do not allow to attach to a group in a different
5620 * task or CPU context:
5623 if (group_leader
->ctx
->type
!= ctx
->type
)
5626 if (group_leader
->ctx
!= ctx
)
5631 * Only a group leader can be exclusive or pinned
5633 if (attr
.exclusive
|| attr
.pinned
)
5638 err
= perf_event_set_output(event
, output_event
);
5643 event_file
= anon_inode_getfile("[perf_event]", &perf_fops
, event
, O_RDWR
);
5644 if (IS_ERR(event_file
)) {
5645 err
= PTR_ERR(event_file
);
5650 struct perf_event_context
*gctx
= group_leader
->ctx
;
5652 mutex_lock(&gctx
->mutex
);
5653 perf_event_remove_from_context(group_leader
);
5654 list_for_each_entry(sibling
, &group_leader
->sibling_list
,
5656 perf_event_remove_from_context(sibling
);
5659 mutex_unlock(&gctx
->mutex
);
5663 event
->filp
= event_file
;
5664 WARN_ON_ONCE(ctx
->parent_ctx
);
5665 mutex_lock(&ctx
->mutex
);
5668 perf_install_in_context(ctx
, group_leader
, cpu
);
5670 list_for_each_entry(sibling
, &group_leader
->sibling_list
,
5672 perf_install_in_context(ctx
, sibling
, cpu
);
5677 perf_install_in_context(ctx
, event
, cpu
);
5679 mutex_unlock(&ctx
->mutex
);
5681 event
->owner
= current
;
5682 get_task_struct(current
);
5683 mutex_lock(¤t
->perf_event_mutex
);
5684 list_add_tail(&event
->owner_entry
, ¤t
->perf_event_list
);
5685 mutex_unlock(¤t
->perf_event_mutex
);
5688 * Drop the reference on the group_event after placing the
5689 * new event on the sibling_list. This ensures destruction
5690 * of the group leader will find the pointer to itself in
5691 * perf_group_detach().
5693 fput_light(group_file
, fput_needed
);
5694 fd_install(event_fd
, event_file
);
5703 put_task_struct(task
);
5705 fput_light(group_file
, fput_needed
);
5707 put_unused_fd(event_fd
);
5712 * perf_event_create_kernel_counter
5714 * @attr: attributes of the counter to create
5715 * @cpu: cpu in which the counter is bound
5716 * @task: task to profile (NULL for percpu)
5719 perf_event_create_kernel_counter(struct perf_event_attr
*attr
, int cpu
,
5720 struct task_struct
*task
,
5721 perf_overflow_handler_t overflow_handler
)
5723 struct perf_event_context
*ctx
;
5724 struct perf_event
*event
;
5728 * Get the target context (task or percpu):
5731 event
= perf_event_alloc(attr
, cpu
, NULL
, NULL
, overflow_handler
);
5732 if (IS_ERR(event
)) {
5733 err
= PTR_ERR(event
);
5737 ctx
= find_get_context(event
->pmu
, task
, cpu
);
5744 WARN_ON_ONCE(ctx
->parent_ctx
);
5745 mutex_lock(&ctx
->mutex
);
5746 perf_install_in_context(ctx
, event
, cpu
);
5748 mutex_unlock(&ctx
->mutex
);
5750 event
->owner
= current
;
5751 get_task_struct(current
);
5752 mutex_lock(¤t
->perf_event_mutex
);
5753 list_add_tail(&event
->owner_entry
, ¤t
->perf_event_list
);
5754 mutex_unlock(¤t
->perf_event_mutex
);
5761 return ERR_PTR(err
);
5763 EXPORT_SYMBOL_GPL(perf_event_create_kernel_counter
);
5765 static void sync_child_event(struct perf_event
*child_event
,
5766 struct task_struct
*child
)
5768 struct perf_event
*parent_event
= child_event
->parent
;
5771 if (child_event
->attr
.inherit_stat
)
5772 perf_event_read_event(child_event
, child
);
5774 child_val
= perf_event_count(child_event
);
5777 * Add back the child's count to the parent's count:
5779 atomic64_add(child_val
, &parent_event
->child_count
);
5780 atomic64_add(child_event
->total_time_enabled
,
5781 &parent_event
->child_total_time_enabled
);
5782 atomic64_add(child_event
->total_time_running
,
5783 &parent_event
->child_total_time_running
);
5786 * Remove this event from the parent's list
5788 WARN_ON_ONCE(parent_event
->ctx
->parent_ctx
);
5789 mutex_lock(&parent_event
->child_mutex
);
5790 list_del_init(&child_event
->child_list
);
5791 mutex_unlock(&parent_event
->child_mutex
);
5794 * Release the parent event, if this was the last
5797 fput(parent_event
->filp
);
5801 __perf_event_exit_task(struct perf_event
*child_event
,
5802 struct perf_event_context
*child_ctx
,
5803 struct task_struct
*child
)
5805 struct perf_event
*parent_event
;
5807 perf_event_remove_from_context(child_event
);
5809 parent_event
= child_event
->parent
;
5811 * It can happen that parent exits first, and has events
5812 * that are still around due to the child reference. These
5813 * events need to be zapped - but otherwise linger.
5816 sync_child_event(child_event
, child
);
5817 free_event(child_event
);
5821 static void perf_event_exit_task_context(struct task_struct
*child
, int ctxn
)
5823 struct perf_event
*child_event
, *tmp
;
5824 struct perf_event_context
*child_ctx
;
5825 unsigned long flags
;
5827 if (likely(!child
->perf_event_ctxp
[ctxn
])) {
5828 perf_event_task(child
, NULL
, 0);
5832 local_irq_save(flags
);
5834 * We can't reschedule here because interrupts are disabled,
5835 * and either child is current or it is a task that can't be
5836 * scheduled, so we are now safe from rescheduling changing
5839 child_ctx
= child
->perf_event_ctxp
[ctxn
];
5840 __perf_event_task_sched_out(child_ctx
);
5843 * Take the context lock here so that if find_get_context is
5844 * reading child->perf_event_ctxp, we wait until it has
5845 * incremented the context's refcount before we do put_ctx below.
5847 raw_spin_lock(&child_ctx
->lock
);
5848 child
->perf_event_ctxp
[ctxn
] = NULL
;
5850 * If this context is a clone; unclone it so it can't get
5851 * swapped to another process while we're removing all
5852 * the events from it.
5854 unclone_ctx(child_ctx
);
5855 update_context_time(child_ctx
);
5856 raw_spin_unlock_irqrestore(&child_ctx
->lock
, flags
);
5859 * Report the task dead after unscheduling the events so that we
5860 * won't get any samples after PERF_RECORD_EXIT. We can however still
5861 * get a few PERF_RECORD_READ events.
5863 perf_event_task(child
, child_ctx
, 0);
5866 * We can recurse on the same lock type through:
5868 * __perf_event_exit_task()
5869 * sync_child_event()
5870 * fput(parent_event->filp)
5872 * mutex_lock(&ctx->mutex)
5874 * But since its the parent context it won't be the same instance.
5876 mutex_lock(&child_ctx
->mutex
);
5879 list_for_each_entry_safe(child_event
, tmp
, &child_ctx
->pinned_groups
,
5881 __perf_event_exit_task(child_event
, child_ctx
, child
);
5883 list_for_each_entry_safe(child_event
, tmp
, &child_ctx
->flexible_groups
,
5885 __perf_event_exit_task(child_event
, child_ctx
, child
);
5888 * If the last event was a group event, it will have appended all
5889 * its siblings to the list, but we obtained 'tmp' before that which
5890 * will still point to the list head terminating the iteration.
5892 if (!list_empty(&child_ctx
->pinned_groups
) ||
5893 !list_empty(&child_ctx
->flexible_groups
))
5896 mutex_unlock(&child_ctx
->mutex
);
5902 * When a child task exits, feed back event values to parent events.
5904 void perf_event_exit_task(struct task_struct
*child
)
5908 for_each_task_context_nr(ctxn
)
5909 perf_event_exit_task_context(child
, ctxn
);
5912 static void perf_free_event(struct perf_event
*event
,
5913 struct perf_event_context
*ctx
)
5915 struct perf_event
*parent
= event
->parent
;
5917 if (WARN_ON_ONCE(!parent
))
5920 mutex_lock(&parent
->child_mutex
);
5921 list_del_init(&event
->child_list
);
5922 mutex_unlock(&parent
->child_mutex
);
5926 perf_group_detach(event
);
5927 list_del_event(event
, ctx
);
5932 * free an unexposed, unused context as created by inheritance by
5933 * perf_event_init_task below, used by fork() in case of fail.
5935 void perf_event_free_task(struct task_struct
*task
)
5937 struct perf_event_context
*ctx
;
5938 struct perf_event
*event
, *tmp
;
5941 for_each_task_context_nr(ctxn
) {
5942 ctx
= task
->perf_event_ctxp
[ctxn
];
5946 mutex_lock(&ctx
->mutex
);
5948 list_for_each_entry_safe(event
, tmp
, &ctx
->pinned_groups
,
5950 perf_free_event(event
, ctx
);
5952 list_for_each_entry_safe(event
, tmp
, &ctx
->flexible_groups
,
5954 perf_free_event(event
, ctx
);
5956 if (!list_empty(&ctx
->pinned_groups
) ||
5957 !list_empty(&ctx
->flexible_groups
))
5960 mutex_unlock(&ctx
->mutex
);
5966 void perf_event_delayed_put(struct task_struct
*task
)
5970 for_each_task_context_nr(ctxn
)
5971 WARN_ON_ONCE(task
->perf_event_ctxp
[ctxn
]);
5975 * inherit a event from parent task to child task:
5977 static struct perf_event
*
5978 inherit_event(struct perf_event
*parent_event
,
5979 struct task_struct
*parent
,
5980 struct perf_event_context
*parent_ctx
,
5981 struct task_struct
*child
,
5982 struct perf_event
*group_leader
,
5983 struct perf_event_context
*child_ctx
)
5985 struct perf_event
*child_event
;
5986 unsigned long flags
;
5989 * Instead of creating recursive hierarchies of events,
5990 * we link inherited events back to the original parent,
5991 * which has a filp for sure, which we use as the reference
5994 if (parent_event
->parent
)
5995 parent_event
= parent_event
->parent
;
5997 child_event
= perf_event_alloc(&parent_event
->attr
,
5999 group_leader
, parent_event
,
6001 if (IS_ERR(child_event
))
6006 * Make the child state follow the state of the parent event,
6007 * not its attr.disabled bit. We hold the parent's mutex,
6008 * so we won't race with perf_event_{en, dis}able_family.
6010 if (parent_event
->state
>= PERF_EVENT_STATE_INACTIVE
)
6011 child_event
->state
= PERF_EVENT_STATE_INACTIVE
;
6013 child_event
->state
= PERF_EVENT_STATE_OFF
;
6015 if (parent_event
->attr
.freq
) {
6016 u64 sample_period
= parent_event
->hw
.sample_period
;
6017 struct hw_perf_event
*hwc
= &child_event
->hw
;
6019 hwc
->sample_period
= sample_period
;
6020 hwc
->last_period
= sample_period
;
6022 local64_set(&hwc
->period_left
, sample_period
);
6025 child_event
->ctx
= child_ctx
;
6026 child_event
->overflow_handler
= parent_event
->overflow_handler
;
6029 * Link it up in the child's context:
6031 raw_spin_lock_irqsave(&child_ctx
->lock
, flags
);
6032 add_event_to_ctx(child_event
, child_ctx
);
6033 raw_spin_unlock_irqrestore(&child_ctx
->lock
, flags
);
6036 * Get a reference to the parent filp - we will fput it
6037 * when the child event exits. This is safe to do because
6038 * we are in the parent and we know that the filp still
6039 * exists and has a nonzero count:
6041 atomic_long_inc(&parent_event
->filp
->f_count
);
6044 * Link this into the parent event's child list
6046 WARN_ON_ONCE(parent_event
->ctx
->parent_ctx
);
6047 mutex_lock(&parent_event
->child_mutex
);
6048 list_add_tail(&child_event
->child_list
, &parent_event
->child_list
);
6049 mutex_unlock(&parent_event
->child_mutex
);
6054 static int inherit_group(struct perf_event
*parent_event
,
6055 struct task_struct
*parent
,
6056 struct perf_event_context
*parent_ctx
,
6057 struct task_struct
*child
,
6058 struct perf_event_context
*child_ctx
)
6060 struct perf_event
*leader
;
6061 struct perf_event
*sub
;
6062 struct perf_event
*child_ctr
;
6064 leader
= inherit_event(parent_event
, parent
, parent_ctx
,
6065 child
, NULL
, child_ctx
);
6067 return PTR_ERR(leader
);
6068 list_for_each_entry(sub
, &parent_event
->sibling_list
, group_entry
) {
6069 child_ctr
= inherit_event(sub
, parent
, parent_ctx
,
6070 child
, leader
, child_ctx
);
6071 if (IS_ERR(child_ctr
))
6072 return PTR_ERR(child_ctr
);
6078 inherit_task_group(struct perf_event
*event
, struct task_struct
*parent
,
6079 struct perf_event_context
*parent_ctx
,
6080 struct task_struct
*child
, int ctxn
,
6084 struct perf_event_context
*child_ctx
;
6086 if (!event
->attr
.inherit
) {
6091 child_ctx
= child
->perf_event_ctxp
[ctxn
];
6094 * This is executed from the parent task context, so
6095 * inherit events that have been marked for cloning.
6096 * First allocate and initialize a context for the
6100 child_ctx
= alloc_perf_context(event
->pmu
, child
);
6104 child
->perf_event_ctxp
[ctxn
] = child_ctx
;
6107 ret
= inherit_group(event
, parent
, parent_ctx
,
6117 * Initialize the perf_event context in task_struct
6119 int perf_event_init_context(struct task_struct
*child
, int ctxn
)
6121 struct perf_event_context
*child_ctx
, *parent_ctx
;
6122 struct perf_event_context
*cloned_ctx
;
6123 struct perf_event
*event
;
6124 struct task_struct
*parent
= current
;
6125 int inherited_all
= 1;
6128 child
->perf_event_ctxp
[ctxn
] = NULL
;
6130 mutex_init(&child
->perf_event_mutex
);
6131 INIT_LIST_HEAD(&child
->perf_event_list
);
6133 if (likely(!parent
->perf_event_ctxp
[ctxn
]))
6137 * If the parent's context is a clone, pin it so it won't get
6140 parent_ctx
= perf_pin_task_context(parent
, ctxn
);
6143 * No need to check if parent_ctx != NULL here; since we saw
6144 * it non-NULL earlier, the only reason for it to become NULL
6145 * is if we exit, and since we're currently in the middle of
6146 * a fork we can't be exiting at the same time.
6150 * Lock the parent list. No need to lock the child - not PID
6151 * hashed yet and not running, so nobody can access it.
6153 mutex_lock(&parent_ctx
->mutex
);
6156 * We dont have to disable NMIs - we are only looking at
6157 * the list, not manipulating it:
6159 list_for_each_entry(event
, &parent_ctx
->pinned_groups
, group_entry
) {
6160 ret
= inherit_task_group(event
, parent
, parent_ctx
,
6161 child
, ctxn
, &inherited_all
);
6166 list_for_each_entry(event
, &parent_ctx
->flexible_groups
, group_entry
) {
6167 ret
= inherit_task_group(event
, parent
, parent_ctx
,
6168 child
, ctxn
, &inherited_all
);
6173 child_ctx
= child
->perf_event_ctxp
[ctxn
];
6175 if (child_ctx
&& inherited_all
) {
6177 * Mark the child context as a clone of the parent
6178 * context, or of whatever the parent is a clone of.
6179 * Note that if the parent is a clone, it could get
6180 * uncloned at any point, but that doesn't matter
6181 * because the list of events and the generation
6182 * count can't have changed since we took the mutex.
6184 cloned_ctx
= rcu_dereference(parent_ctx
->parent_ctx
);
6186 child_ctx
->parent_ctx
= cloned_ctx
;
6187 child_ctx
->parent_gen
= parent_ctx
->parent_gen
;
6189 child_ctx
->parent_ctx
= parent_ctx
;
6190 child_ctx
->parent_gen
= parent_ctx
->generation
;
6192 get_ctx(child_ctx
->parent_ctx
);
6195 mutex_unlock(&parent_ctx
->mutex
);
6197 perf_unpin_context(parent_ctx
);
6203 * Initialize the perf_event context in task_struct
6205 int perf_event_init_task(struct task_struct
*child
)
6209 for_each_task_context_nr(ctxn
) {
6210 ret
= perf_event_init_context(child
, ctxn
);
6218 static void __init
perf_event_init_all_cpus(void)
6220 struct swevent_htable
*swhash
;
6223 for_each_possible_cpu(cpu
) {
6224 swhash
= &per_cpu(swevent_htable
, cpu
);
6225 mutex_init(&swhash
->hlist_mutex
);
6226 INIT_LIST_HEAD(&per_cpu(rotation_list
, cpu
));
6230 static void __cpuinit
perf_event_init_cpu(int cpu
)
6232 struct swevent_htable
*swhash
= &per_cpu(swevent_htable
, cpu
);
6234 mutex_lock(&swhash
->hlist_mutex
);
6235 if (swhash
->hlist_refcount
> 0) {
6236 struct swevent_hlist
*hlist
;
6238 hlist
= kzalloc_node(sizeof(*hlist
), GFP_KERNEL
, cpu_to_node(cpu
));
6240 rcu_assign_pointer(swhash
->swevent_hlist
, hlist
);
6242 mutex_unlock(&swhash
->hlist_mutex
);
6245 #ifdef CONFIG_HOTPLUG_CPU
6246 static void perf_pmu_rotate_stop(struct pmu
*pmu
)
6248 struct perf_cpu_context
*cpuctx
= this_cpu_ptr(pmu
->pmu_cpu_context
);
6250 WARN_ON(!irqs_disabled());
6252 list_del_init(&cpuctx
->rotation_list
);
6255 static void __perf_event_exit_context(void *__info
)
6257 struct perf_event_context
*ctx
= __info
;
6258 struct perf_event
*event
, *tmp
;
6260 perf_pmu_rotate_stop(ctx
->pmu
);
6262 list_for_each_entry_safe(event
, tmp
, &ctx
->pinned_groups
, group_entry
)
6263 __perf_event_remove_from_context(event
);
6264 list_for_each_entry_safe(event
, tmp
, &ctx
->flexible_groups
, group_entry
)
6265 __perf_event_remove_from_context(event
);
6268 static void perf_event_exit_cpu_context(int cpu
)
6270 struct perf_event_context
*ctx
;
6274 idx
= srcu_read_lock(&pmus_srcu
);
6275 list_for_each_entry_rcu(pmu
, &pmus
, entry
) {
6276 ctx
= &per_cpu_ptr(pmu
->pmu_cpu_context
, cpu
)->ctx
;
6278 mutex_lock(&ctx
->mutex
);
6279 smp_call_function_single(cpu
, __perf_event_exit_context
, ctx
, 1);
6280 mutex_unlock(&ctx
->mutex
);
6282 srcu_read_unlock(&pmus_srcu
, idx
);
6285 static void perf_event_exit_cpu(int cpu
)
6287 struct swevent_htable
*swhash
= &per_cpu(swevent_htable
, cpu
);
6289 mutex_lock(&swhash
->hlist_mutex
);
6290 swevent_hlist_release(swhash
);
6291 mutex_unlock(&swhash
->hlist_mutex
);
6293 perf_event_exit_cpu_context(cpu
);
6296 static inline void perf_event_exit_cpu(int cpu
) { }
6299 static int __cpuinit
6300 perf_cpu_notify(struct notifier_block
*self
, unsigned long action
, void *hcpu
)
6302 unsigned int cpu
= (long)hcpu
;
6304 switch (action
& ~CPU_TASKS_FROZEN
) {
6306 case CPU_UP_PREPARE
:
6307 case CPU_DOWN_FAILED
:
6308 perf_event_init_cpu(cpu
);
6311 case CPU_UP_CANCELED
:
6312 case CPU_DOWN_PREPARE
:
6313 perf_event_exit_cpu(cpu
);
6323 void __init
perf_event_init(void)
6325 perf_event_init_all_cpus();
6326 init_srcu_struct(&pmus_srcu
);
6327 perf_pmu_register(&perf_swevent
);
6328 perf_pmu_register(&perf_cpu_clock
);
6329 perf_pmu_register(&perf_task_clock
);
6331 perf_cpu_notifier(perf_cpu_notify
);