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 void perf_pmu_disable(struct pmu *pmu)
68 int *count = this_cpu_ptr(pmu->pmu_disable_count);
70 pmu->pmu_disable(pmu);
73 void perf_pmu_enable(struct pmu *pmu)
75 int *count = this_cpu_ptr(pmu->pmu_disable_count);
80 static void perf_pmu_rotate_start(struct pmu *pmu)
82 struct perf_cpu_context *cpuctx = this_cpu_ptr(pmu->pmu_cpu_context);
84 if (hrtimer_active(&cpuctx->timer))
87 __hrtimer_start_range_ns(&cpuctx->timer,
88 ns_to_ktime(cpuctx->timer_interval), 0,
89 HRTIMER_MODE_REL_PINNED, 0);
92 static void perf_pmu_rotate_stop(struct pmu *pmu)
94 struct perf_cpu_context *cpuctx = this_cpu_ptr(pmu->pmu_cpu_context);
96 hrtimer_cancel(&cpuctx->timer);
99 static void get_ctx(struct perf_event_context *ctx)
101 WARN_ON(!atomic_inc_not_zero(&ctx->refcount));
104 static void free_ctx(struct rcu_head *head)
106 struct perf_event_context *ctx;
108 ctx = container_of(head, struct perf_event_context, rcu_head);
112 static void put_ctx(struct perf_event_context *ctx)
114 if (atomic_dec_and_test(&ctx->refcount)) {
116 put_ctx(ctx->parent_ctx);
118 put_task_struct(ctx->task);
119 call_rcu(&ctx->rcu_head, free_ctx);
123 static void unclone_ctx(struct perf_event_context *ctx)
125 if (ctx->parent_ctx) {
126 put_ctx(ctx->parent_ctx);
127 ctx->parent_ctx = NULL;
132 * If we inherit events we want to return the parent event id
135 static u64 primary_event_id(struct perf_event *event)
140 id = event->parent->id;
146 * Get the perf_event_context for a task and lock it.
147 * This has to cope with with the fact that until it is locked,
148 * the context could get moved to another task.
150 static struct perf_event_context *
151 perf_lock_task_context(struct task_struct *task, int ctxn, unsigned long *flags)
153 struct perf_event_context *ctx;
157 ctx = rcu_dereference(task->perf_event_ctxp[ctxn]);
160 * If this context is a clone of another, it might
161 * get swapped for another underneath us by
162 * perf_event_task_sched_out, though the
163 * rcu_read_lock() protects us from any context
164 * getting freed. Lock the context and check if it
165 * got swapped before we could get the lock, and retry
166 * if so. If we locked the right context, then it
167 * can't get swapped on us any more.
169 raw_spin_lock_irqsave(&ctx->lock, *flags);
170 if (ctx != rcu_dereference(task->perf_event_ctxp[ctxn])) {
171 raw_spin_unlock_irqrestore(&ctx->lock, *flags);
175 if (!atomic_inc_not_zero(&ctx->refcount)) {
176 raw_spin_unlock_irqrestore(&ctx->lock, *flags);
185 * Get the context for a task and increment its pin_count so it
186 * can't get swapped to another task. This also increments its
187 * reference count so that the context can't get freed.
189 static struct perf_event_context *
190 perf_pin_task_context(struct task_struct *task, int ctxn)
192 struct perf_event_context *ctx;
195 ctx = perf_lock_task_context(task, ctxn, &flags);
198 raw_spin_unlock_irqrestore(&ctx->lock, flags);
203 static void perf_unpin_context(struct perf_event_context *ctx)
207 raw_spin_lock_irqsave(&ctx->lock, flags);
209 raw_spin_unlock_irqrestore(&ctx->lock, flags);
213 static inline u64 perf_clock(void)
215 return local_clock();
219 * Update the record of the current time in a context.
221 static void update_context_time(struct perf_event_context *ctx)
223 u64 now = perf_clock();
225 ctx->time += now - ctx->timestamp;
226 ctx->timestamp = now;
230 * Update the total_time_enabled and total_time_running fields for a event.
232 static void update_event_times(struct perf_event *event)
234 struct perf_event_context *ctx = event->ctx;
237 if (event->state < PERF_EVENT_STATE_INACTIVE ||
238 event->group_leader->state < PERF_EVENT_STATE_INACTIVE)
244 run_end = event->tstamp_stopped;
246 event->total_time_enabled = run_end - event->tstamp_enabled;
248 if (event->state == PERF_EVENT_STATE_INACTIVE)
249 run_end = event->tstamp_stopped;
253 event->total_time_running = run_end - event->tstamp_running;
257 * Update total_time_enabled and total_time_running for all events in a group.
259 static void update_group_times(struct perf_event *leader)
261 struct perf_event *event;
263 update_event_times(leader);
264 list_for_each_entry(event, &leader->sibling_list, group_entry)
265 update_event_times(event);
268 static struct list_head *
269 ctx_group_list(struct perf_event *event, struct perf_event_context *ctx)
271 if (event->attr.pinned)
272 return &ctx->pinned_groups;
274 return &ctx->flexible_groups;
278 * Add a event from the lists for its context.
279 * Must be called with ctx->mutex and ctx->lock held.
282 list_add_event(struct perf_event *event, struct perf_event_context *ctx)
284 WARN_ON_ONCE(event->attach_state & PERF_ATTACH_CONTEXT);
285 event->attach_state |= PERF_ATTACH_CONTEXT;
288 * If we're a stand alone event or group leader, we go to the context
289 * list, group events are kept attached to the group so that
290 * perf_group_detach can, at all times, locate all siblings.
292 if (event->group_leader == event) {
293 struct list_head *list;
295 if (is_software_event(event))
296 event->group_flags |= PERF_GROUP_SOFTWARE;
298 list = ctx_group_list(event, ctx);
299 list_add_tail(&event->group_entry, list);
302 list_add_rcu(&event->event_entry, &ctx->event_list);
304 perf_pmu_rotate_start(ctx->pmu);
306 if (event->attr.inherit_stat)
310 static void perf_group_attach(struct perf_event *event)
312 struct perf_event *group_leader = event->group_leader;
314 WARN_ON_ONCE(event->attach_state & PERF_ATTACH_GROUP);
315 event->attach_state |= PERF_ATTACH_GROUP;
317 if (group_leader == event)
320 if (group_leader->group_flags & PERF_GROUP_SOFTWARE &&
321 !is_software_event(event))
322 group_leader->group_flags &= ~PERF_GROUP_SOFTWARE;
324 list_add_tail(&event->group_entry, &group_leader->sibling_list);
325 group_leader->nr_siblings++;
329 * Remove a event from the lists for its context.
330 * Must be called with ctx->mutex and ctx->lock held.
333 list_del_event(struct perf_event *event, struct perf_event_context *ctx)
336 * We can have double detach due to exit/hot-unplug + close.
338 if (!(event->attach_state & PERF_ATTACH_CONTEXT))
341 event->attach_state &= ~PERF_ATTACH_CONTEXT;
344 if (event->attr.inherit_stat)
347 list_del_rcu(&event->event_entry);
349 if (event->group_leader == event)
350 list_del_init(&event->group_entry);
352 update_group_times(event);
355 * If event was in error state, then keep it
356 * that way, otherwise bogus counts will be
357 * returned on read(). The only way to get out
358 * of error state is by explicit re-enabling
361 if (event->state > PERF_EVENT_STATE_OFF)
362 event->state = PERF_EVENT_STATE_OFF;
365 static void perf_group_detach(struct perf_event *event)
367 struct perf_event *sibling, *tmp;
368 struct list_head *list = NULL;
371 * We can have double detach due to exit/hot-unplug + close.
373 if (!(event->attach_state & PERF_ATTACH_GROUP))
376 event->attach_state &= ~PERF_ATTACH_GROUP;
379 * If this is a sibling, remove it from its group.
381 if (event->group_leader != event) {
382 list_del_init(&event->group_entry);
383 event->group_leader->nr_siblings--;
387 if (!list_empty(&event->group_entry))
388 list = &event->group_entry;
391 * If this was a group event with sibling events then
392 * upgrade the siblings to singleton events by adding them
393 * to whatever list we are on.
395 list_for_each_entry_safe(sibling, tmp, &event->sibling_list, group_entry) {
397 list_move_tail(&sibling->group_entry, list);
398 sibling->group_leader = sibling;
400 /* Inherit group flags from the previous leader */
401 sibling->group_flags = event->group_flags;
406 event_filter_match(struct perf_event *event)
408 return event->cpu == -1 || event->cpu == smp_processor_id();
412 event_sched_out(struct perf_event *event,
413 struct perf_cpu_context *cpuctx,
414 struct perf_event_context *ctx)
418 * An event which could not be activated because of
419 * filter mismatch still needs to have its timings
420 * maintained, otherwise bogus information is return
421 * via read() for time_enabled, time_running:
423 if (event->state == PERF_EVENT_STATE_INACTIVE
424 && !event_filter_match(event)) {
425 delta = ctx->time - event->tstamp_stopped;
426 event->tstamp_running += delta;
427 event->tstamp_stopped = ctx->time;
430 if (event->state != PERF_EVENT_STATE_ACTIVE)
433 event->state = PERF_EVENT_STATE_INACTIVE;
434 if (event->pending_disable) {
435 event->pending_disable = 0;
436 event->state = PERF_EVENT_STATE_OFF;
438 event->tstamp_stopped = ctx->time;
439 event->pmu->del(event, 0);
442 if (!is_software_event(event))
443 cpuctx->active_oncpu--;
445 if (event->attr.exclusive || !cpuctx->active_oncpu)
446 cpuctx->exclusive = 0;
450 group_sched_out(struct perf_event *group_event,
451 struct perf_cpu_context *cpuctx,
452 struct perf_event_context *ctx)
454 struct perf_event *event;
455 int state = group_event->state;
457 event_sched_out(group_event, cpuctx, ctx);
460 * Schedule out siblings (if any):
462 list_for_each_entry(event, &group_event->sibling_list, group_entry)
463 event_sched_out(event, cpuctx, ctx);
465 if (state == PERF_EVENT_STATE_ACTIVE && group_event->attr.exclusive)
466 cpuctx->exclusive = 0;
469 static inline struct perf_cpu_context *
470 __get_cpu_context(struct perf_event_context *ctx)
472 return this_cpu_ptr(ctx->pmu->pmu_cpu_context);
476 * Cross CPU call to remove a performance event
478 * We disable the event on the hardware level first. After that we
479 * remove it from the context list.
481 static void __perf_event_remove_from_context(void *info)
483 struct perf_event *event = info;
484 struct perf_event_context *ctx = event->ctx;
485 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
488 * If this is a task context, we need to check whether it is
489 * the current task context of this cpu. If not it has been
490 * scheduled out before the smp call arrived.
492 if (ctx->task && cpuctx->task_ctx != ctx)
495 raw_spin_lock(&ctx->lock);
497 event_sched_out(event, cpuctx, ctx);
499 list_del_event(event, ctx);
501 raw_spin_unlock(&ctx->lock);
506 * Remove the event from a task's (or a CPU's) list of events.
508 * Must be called with ctx->mutex held.
510 * CPU events are removed with a smp call. For task events we only
511 * call when the task is on a CPU.
513 * If event->ctx is a cloned context, callers must make sure that
514 * every task struct that event->ctx->task could possibly point to
515 * remains valid. This is OK when called from perf_release since
516 * that only calls us on the top-level context, which can't be a clone.
517 * When called from perf_event_exit_task, it's OK because the
518 * context has been detached from its task.
520 static void perf_event_remove_from_context(struct perf_event *event)
522 struct perf_event_context *ctx = event->ctx;
523 struct task_struct *task = ctx->task;
527 * Per cpu events are removed via an smp call and
528 * the removal is always successful.
530 smp_call_function_single(event->cpu,
531 __perf_event_remove_from_context,
537 task_oncpu_function_call(task, __perf_event_remove_from_context,
540 raw_spin_lock_irq(&ctx->lock);
542 * If the context is active we need to retry the smp call.
544 if (ctx->nr_active && !list_empty(&event->group_entry)) {
545 raw_spin_unlock_irq(&ctx->lock);
550 * The lock prevents that this context is scheduled in so we
551 * can remove the event safely, if the call above did not
554 if (!list_empty(&event->group_entry))
555 list_del_event(event, ctx);
556 raw_spin_unlock_irq(&ctx->lock);
560 * Cross CPU call to disable a performance event
562 static void __perf_event_disable(void *info)
564 struct perf_event *event = info;
565 struct perf_event_context *ctx = event->ctx;
566 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
569 * If this is a per-task event, need to check whether this
570 * event's task is the current task on this cpu.
572 if (ctx->task && cpuctx->task_ctx != ctx)
575 raw_spin_lock(&ctx->lock);
578 * If the event is on, turn it off.
579 * If it is in error state, leave it in error state.
581 if (event->state >= PERF_EVENT_STATE_INACTIVE) {
582 update_context_time(ctx);
583 update_group_times(event);
584 if (event == event->group_leader)
585 group_sched_out(event, cpuctx, ctx);
587 event_sched_out(event, cpuctx, ctx);
588 event->state = PERF_EVENT_STATE_OFF;
591 raw_spin_unlock(&ctx->lock);
597 * If event->ctx is a cloned context, callers must make sure that
598 * every task struct that event->ctx->task could possibly point to
599 * remains valid. This condition is satisifed when called through
600 * perf_event_for_each_child or perf_event_for_each because they
601 * hold the top-level event's child_mutex, so any descendant that
602 * goes to exit will block in sync_child_event.
603 * When called from perf_pending_event it's OK because event->ctx
604 * is the current context on this CPU and preemption is disabled,
605 * hence we can't get into perf_event_task_sched_out for this context.
607 void perf_event_disable(struct perf_event *event)
609 struct perf_event_context *ctx = event->ctx;
610 struct task_struct *task = ctx->task;
614 * Disable the event on the cpu that it's on
616 smp_call_function_single(event->cpu, __perf_event_disable,
622 task_oncpu_function_call(task, __perf_event_disable, event);
624 raw_spin_lock_irq(&ctx->lock);
626 * If the event is still active, we need to retry the cross-call.
628 if (event->state == PERF_EVENT_STATE_ACTIVE) {
629 raw_spin_unlock_irq(&ctx->lock);
634 * Since we have the lock this context can't be scheduled
635 * in, so we can change the state safely.
637 if (event->state == PERF_EVENT_STATE_INACTIVE) {
638 update_group_times(event);
639 event->state = PERF_EVENT_STATE_OFF;
642 raw_spin_unlock_irq(&ctx->lock);
646 event_sched_in(struct perf_event *event,
647 struct perf_cpu_context *cpuctx,
648 struct perf_event_context *ctx)
650 if (event->state <= PERF_EVENT_STATE_OFF)
653 event->state = PERF_EVENT_STATE_ACTIVE;
654 event->oncpu = smp_processor_id();
656 * The new state must be visible before we turn it on in the hardware:
660 if (event->pmu->add(event, PERF_EF_START)) {
661 event->state = PERF_EVENT_STATE_INACTIVE;
666 event->tstamp_running += ctx->time - event->tstamp_stopped;
668 if (!is_software_event(event))
669 cpuctx->active_oncpu++;
672 if (event->attr.exclusive)
673 cpuctx->exclusive = 1;
679 group_sched_in(struct perf_event *group_event,
680 struct perf_cpu_context *cpuctx,
681 struct perf_event_context *ctx)
683 struct perf_event *event, *partial_group = NULL;
684 struct pmu *pmu = group_event->pmu;
686 if (group_event->state == PERF_EVENT_STATE_OFF)
691 if (event_sched_in(group_event, cpuctx, ctx)) {
692 pmu->cancel_txn(pmu);
697 * Schedule in siblings as one group (if any):
699 list_for_each_entry(event, &group_event->sibling_list, group_entry) {
700 if (event_sched_in(event, cpuctx, ctx)) {
701 partial_group = event;
706 if (!pmu->commit_txn(pmu))
711 * Groups can be scheduled in as one unit only, so undo any
712 * partial group before returning:
714 list_for_each_entry(event, &group_event->sibling_list, group_entry) {
715 if (event == partial_group)
717 event_sched_out(event, cpuctx, ctx);
719 event_sched_out(group_event, cpuctx, ctx);
721 pmu->cancel_txn(pmu);
727 * Work out whether we can put this event group on the CPU now.
729 static int group_can_go_on(struct perf_event *event,
730 struct perf_cpu_context *cpuctx,
734 * Groups consisting entirely of software events can always go on.
736 if (event->group_flags & PERF_GROUP_SOFTWARE)
739 * If an exclusive group is already on, no other hardware
742 if (cpuctx->exclusive)
745 * If this group is exclusive and there are already
746 * events on the CPU, it can't go on.
748 if (event->attr.exclusive && cpuctx->active_oncpu)
751 * Otherwise, try to add it if all previous groups were able
757 static void add_event_to_ctx(struct perf_event *event,
758 struct perf_event_context *ctx)
760 list_add_event(event, ctx);
761 perf_group_attach(event);
762 event->tstamp_enabled = ctx->time;
763 event->tstamp_running = ctx->time;
764 event->tstamp_stopped = ctx->time;
768 * Cross CPU call to install and enable a performance event
770 * Must be called with ctx->mutex held
772 static void __perf_install_in_context(void *info)
774 struct perf_event *event = info;
775 struct perf_event_context *ctx = event->ctx;
776 struct perf_event *leader = event->group_leader;
777 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
781 * If this is a task context, we need to check whether it is
782 * the current task context of this cpu. If not it has been
783 * scheduled out before the smp call arrived.
784 * Or possibly this is the right context but it isn't
785 * on this cpu because it had no events.
787 if (ctx->task && cpuctx->task_ctx != ctx) {
788 if (cpuctx->task_ctx || ctx->task != current)
790 cpuctx->task_ctx = ctx;
793 raw_spin_lock(&ctx->lock);
795 update_context_time(ctx);
797 add_event_to_ctx(event, ctx);
799 if (event->cpu != -1 && event->cpu != smp_processor_id())
803 * Don't put the event on if it is disabled or if
804 * it is in a group and the group isn't on.
806 if (event->state != PERF_EVENT_STATE_INACTIVE ||
807 (leader != event && leader->state != PERF_EVENT_STATE_ACTIVE))
811 * An exclusive event can't go on if there are already active
812 * hardware events, and no hardware event can go on if there
813 * is already an exclusive event on.
815 if (!group_can_go_on(event, cpuctx, 1))
818 err = event_sched_in(event, cpuctx, ctx);
822 * This event couldn't go on. If it is in a group
823 * then we have to pull the whole group off.
824 * If the event group is pinned then put it in error state.
827 group_sched_out(leader, cpuctx, ctx);
828 if (leader->attr.pinned) {
829 update_group_times(leader);
830 leader->state = PERF_EVENT_STATE_ERROR;
835 raw_spin_unlock(&ctx->lock);
839 * Attach a performance event to a context
841 * First we add the event to the list with the hardware enable bit
842 * in event->hw_config cleared.
844 * If the event is attached to a task which is on a CPU we use a smp
845 * call to enable it in the task context. The task might have been
846 * scheduled away, but we check this in the smp call again.
848 * Must be called with ctx->mutex held.
851 perf_install_in_context(struct perf_event_context *ctx,
852 struct perf_event *event,
855 struct task_struct *task = ctx->task;
861 * Per cpu events are installed via an smp call and
862 * the install is always successful.
864 smp_call_function_single(cpu, __perf_install_in_context,
870 task_oncpu_function_call(task, __perf_install_in_context,
873 raw_spin_lock_irq(&ctx->lock);
875 * we need to retry the smp call.
877 if (ctx->is_active && list_empty(&event->group_entry)) {
878 raw_spin_unlock_irq(&ctx->lock);
883 * The lock prevents that this context is scheduled in so we
884 * can add the event safely, if it the call above did not
887 if (list_empty(&event->group_entry))
888 add_event_to_ctx(event, ctx);
889 raw_spin_unlock_irq(&ctx->lock);
893 * Put a event into inactive state and update time fields.
894 * Enabling the leader of a group effectively enables all
895 * the group members that aren't explicitly disabled, so we
896 * have to update their ->tstamp_enabled also.
897 * Note: this works for group members as well as group leaders
898 * since the non-leader members' sibling_lists will be empty.
900 static void __perf_event_mark_enabled(struct perf_event *event,
901 struct perf_event_context *ctx)
903 struct perf_event *sub;
905 event->state = PERF_EVENT_STATE_INACTIVE;
906 event->tstamp_enabled = ctx->time - event->total_time_enabled;
907 list_for_each_entry(sub, &event->sibling_list, group_entry) {
908 if (sub->state >= PERF_EVENT_STATE_INACTIVE) {
909 sub->tstamp_enabled =
910 ctx->time - sub->total_time_enabled;
916 * Cross CPU call to enable a performance event
918 static void __perf_event_enable(void *info)
920 struct perf_event *event = info;
921 struct perf_event_context *ctx = event->ctx;
922 struct perf_event *leader = event->group_leader;
923 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
927 * If this is a per-task event, need to check whether this
928 * event's task is the current task on this cpu.
930 if (ctx->task && cpuctx->task_ctx != ctx) {
931 if (cpuctx->task_ctx || ctx->task != current)
933 cpuctx->task_ctx = ctx;
936 raw_spin_lock(&ctx->lock);
938 update_context_time(ctx);
940 if (event->state >= PERF_EVENT_STATE_INACTIVE)
942 __perf_event_mark_enabled(event, ctx);
944 if (event->cpu != -1 && event->cpu != smp_processor_id())
948 * If the event is in a group and isn't the group leader,
949 * then don't put it on unless the group is on.
951 if (leader != event && leader->state != PERF_EVENT_STATE_ACTIVE)
954 if (!group_can_go_on(event, cpuctx, 1)) {
958 err = group_sched_in(event, cpuctx, ctx);
960 err = event_sched_in(event, cpuctx, ctx);
965 * If this event can't go on and it's part of a
966 * group, then the whole group has to come off.
969 group_sched_out(leader, cpuctx, ctx);
970 if (leader->attr.pinned) {
971 update_group_times(leader);
972 leader->state = PERF_EVENT_STATE_ERROR;
977 raw_spin_unlock(&ctx->lock);
983 * If event->ctx is a cloned context, callers must make sure that
984 * every task struct that event->ctx->task could possibly point to
985 * remains valid. This condition is satisfied when called through
986 * perf_event_for_each_child or perf_event_for_each as described
987 * for perf_event_disable.
989 void perf_event_enable(struct perf_event *event)
991 struct perf_event_context *ctx = event->ctx;
992 struct task_struct *task = ctx->task;
996 * Enable the event on the cpu that it's on
998 smp_call_function_single(event->cpu, __perf_event_enable,
1003 raw_spin_lock_irq(&ctx->lock);
1004 if (event->state >= PERF_EVENT_STATE_INACTIVE)
1008 * If the event is in error state, clear that first.
1009 * That way, if we see the event in error state below, we
1010 * know that it has gone back into error state, as distinct
1011 * from the task having been scheduled away before the
1012 * cross-call arrived.
1014 if (event->state == PERF_EVENT_STATE_ERROR)
1015 event->state = PERF_EVENT_STATE_OFF;
1018 raw_spin_unlock_irq(&ctx->lock);
1019 task_oncpu_function_call(task, __perf_event_enable, event);
1021 raw_spin_lock_irq(&ctx->lock);
1024 * If the context is active and the event is still off,
1025 * we need to retry the cross-call.
1027 if (ctx->is_active && event->state == PERF_EVENT_STATE_OFF)
1031 * Since we have the lock this context can't be scheduled
1032 * in, so we can change the state safely.
1034 if (event->state == PERF_EVENT_STATE_OFF)
1035 __perf_event_mark_enabled(event, ctx);
1038 raw_spin_unlock_irq(&ctx->lock);
1041 static int perf_event_refresh(struct perf_event *event, int refresh)
1044 * not supported on inherited events
1046 if (event->attr.inherit)
1049 atomic_add(refresh, &event->event_limit);
1050 perf_event_enable(event);
1056 EVENT_FLEXIBLE = 0x1,
1058 EVENT_ALL = EVENT_FLEXIBLE | EVENT_PINNED,
1061 static void ctx_sched_out(struct perf_event_context *ctx,
1062 struct perf_cpu_context *cpuctx,
1063 enum event_type_t event_type)
1065 struct perf_event *event;
1067 raw_spin_lock(&ctx->lock);
1069 if (likely(!ctx->nr_events))
1071 update_context_time(ctx);
1073 if (!ctx->nr_active)
1076 if (event_type & EVENT_PINNED) {
1077 list_for_each_entry(event, &ctx->pinned_groups, group_entry)
1078 group_sched_out(event, cpuctx, ctx);
1081 if (event_type & EVENT_FLEXIBLE) {
1082 list_for_each_entry(event, &ctx->flexible_groups, group_entry)
1083 group_sched_out(event, cpuctx, ctx);
1086 raw_spin_unlock(&ctx->lock);
1090 * Test whether two contexts are equivalent, i.e. whether they
1091 * have both been cloned from the same version of the same context
1092 * and they both have the same number of enabled events.
1093 * If the number of enabled events is the same, then the set
1094 * of enabled events should be the same, because these are both
1095 * inherited contexts, therefore we can't access individual events
1096 * in them directly with an fd; we can only enable/disable all
1097 * events via prctl, or enable/disable all events in a family
1098 * via ioctl, which will have the same effect on both contexts.
1100 static int context_equiv(struct perf_event_context *ctx1,
1101 struct perf_event_context *ctx2)
1103 return ctx1->parent_ctx && ctx1->parent_ctx == ctx2->parent_ctx
1104 && ctx1->parent_gen == ctx2->parent_gen
1105 && !ctx1->pin_count && !ctx2->pin_count;
1108 static void __perf_event_sync_stat(struct perf_event *event,
1109 struct perf_event *next_event)
1113 if (!event->attr.inherit_stat)
1117 * Update the event value, we cannot use perf_event_read()
1118 * because we're in the middle of a context switch and have IRQs
1119 * disabled, which upsets smp_call_function_single(), however
1120 * we know the event must be on the current CPU, therefore we
1121 * don't need to use it.
1123 switch (event->state) {
1124 case PERF_EVENT_STATE_ACTIVE:
1125 event->pmu->read(event);
1128 case PERF_EVENT_STATE_INACTIVE:
1129 update_event_times(event);
1137 * In order to keep per-task stats reliable we need to flip the event
1138 * values when we flip the contexts.
1140 value = local64_read(&next_event->count);
1141 value = local64_xchg(&event->count, value);
1142 local64_set(&next_event->count, value);
1144 swap(event->total_time_enabled, next_event->total_time_enabled);
1145 swap(event->total_time_running, next_event->total_time_running);
1148 * Since we swizzled the values, update the user visible data too.
1150 perf_event_update_userpage(event);
1151 perf_event_update_userpage(next_event);
1154 #define list_next_entry(pos, member) \
1155 list_entry(pos->member.next, typeof(*pos), member)
1157 static void perf_event_sync_stat(struct perf_event_context *ctx,
1158 struct perf_event_context *next_ctx)
1160 struct perf_event *event, *next_event;
1165 update_context_time(ctx);
1167 event = list_first_entry(&ctx->event_list,
1168 struct perf_event, event_entry);
1170 next_event = list_first_entry(&next_ctx->event_list,
1171 struct perf_event, event_entry);
1173 while (&event->event_entry != &ctx->event_list &&
1174 &next_event->event_entry != &next_ctx->event_list) {
1176 __perf_event_sync_stat(event, next_event);
1178 event = list_next_entry(event, event_entry);
1179 next_event = list_next_entry(next_event, event_entry);
1183 void perf_event_context_sched_out(struct task_struct *task, int ctxn,
1184 struct task_struct *next)
1186 struct perf_event_context *ctx = task->perf_event_ctxp[ctxn];
1187 struct perf_event_context *next_ctx;
1188 struct perf_event_context *parent;
1189 struct perf_cpu_context *cpuctx;
1195 cpuctx = __get_cpu_context(ctx);
1196 if (!cpuctx->task_ctx)
1200 parent = rcu_dereference(ctx->parent_ctx);
1201 next_ctx = next->perf_event_ctxp[ctxn];
1202 if (parent && next_ctx &&
1203 rcu_dereference(next_ctx->parent_ctx) == parent) {
1205 * Looks like the two contexts are clones, so we might be
1206 * able to optimize the context switch. We lock both
1207 * contexts and check that they are clones under the
1208 * lock (including re-checking that neither has been
1209 * uncloned in the meantime). It doesn't matter which
1210 * order we take the locks because no other cpu could
1211 * be trying to lock both of these tasks.
1213 raw_spin_lock(&ctx->lock);
1214 raw_spin_lock_nested(&next_ctx->lock, SINGLE_DEPTH_NESTING);
1215 if (context_equiv(ctx, next_ctx)) {
1217 * XXX do we need a memory barrier of sorts
1218 * wrt to rcu_dereference() of perf_event_ctxp
1220 task->perf_event_ctxp[ctxn] = next_ctx;
1221 next->perf_event_ctxp[ctxn] = ctx;
1223 next_ctx->task = task;
1226 perf_event_sync_stat(ctx, next_ctx);
1228 raw_spin_unlock(&next_ctx->lock);
1229 raw_spin_unlock(&ctx->lock);
1234 ctx_sched_out(ctx, cpuctx, EVENT_ALL);
1235 cpuctx->task_ctx = NULL;
1239 #define for_each_task_context_nr(ctxn) \
1240 for ((ctxn) = 0; (ctxn) < perf_nr_task_contexts; (ctxn)++)
1243 * Called from scheduler to remove the events of the current task,
1244 * with interrupts disabled.
1246 * We stop each event and update the event value in event->count.
1248 * This does not protect us against NMI, but disable()
1249 * sets the disabled bit in the control field of event _before_
1250 * accessing the event control register. If a NMI hits, then it will
1251 * not restart the event.
1253 void perf_event_task_sched_out(struct task_struct *task,
1254 struct task_struct *next)
1258 perf_sw_event(PERF_COUNT_SW_CONTEXT_SWITCHES, 1, 1, NULL, 0);
1260 for_each_task_context_nr(ctxn)
1261 perf_event_context_sched_out(task, ctxn, next);
1264 static void task_ctx_sched_out(struct perf_event_context *ctx,
1265 enum event_type_t event_type)
1267 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
1269 if (!cpuctx->task_ctx)
1272 if (WARN_ON_ONCE(ctx != cpuctx->task_ctx))
1275 ctx_sched_out(ctx, cpuctx, event_type);
1276 cpuctx->task_ctx = NULL;
1280 * Called with IRQs disabled
1282 static void __perf_event_task_sched_out(struct perf_event_context *ctx)
1284 task_ctx_sched_out(ctx, EVENT_ALL);
1288 * Called with IRQs disabled
1290 static void cpu_ctx_sched_out(struct perf_cpu_context *cpuctx,
1291 enum event_type_t event_type)
1293 ctx_sched_out(&cpuctx->ctx, cpuctx, event_type);
1297 ctx_pinned_sched_in(struct perf_event_context *ctx,
1298 struct perf_cpu_context *cpuctx)
1300 struct perf_event *event;
1302 list_for_each_entry(event, &ctx->pinned_groups, group_entry) {
1303 if (event->state <= PERF_EVENT_STATE_OFF)
1305 if (event->cpu != -1 && event->cpu != smp_processor_id())
1308 if (group_can_go_on(event, cpuctx, 1))
1309 group_sched_in(event, cpuctx, ctx);
1312 * If this pinned group hasn't been scheduled,
1313 * put it in error state.
1315 if (event->state == PERF_EVENT_STATE_INACTIVE) {
1316 update_group_times(event);
1317 event->state = PERF_EVENT_STATE_ERROR;
1323 ctx_flexible_sched_in(struct perf_event_context *ctx,
1324 struct perf_cpu_context *cpuctx)
1326 struct perf_event *event;
1329 list_for_each_entry(event, &ctx->flexible_groups, group_entry) {
1330 /* Ignore events in OFF or ERROR state */
1331 if (event->state <= PERF_EVENT_STATE_OFF)
1334 * Listen to the 'cpu' scheduling filter constraint
1337 if (event->cpu != -1 && event->cpu != smp_processor_id())
1340 if (group_can_go_on(event, cpuctx, can_add_hw)) {
1341 if (group_sched_in(event, cpuctx, ctx))
1348 ctx_sched_in(struct perf_event_context *ctx,
1349 struct perf_cpu_context *cpuctx,
1350 enum event_type_t event_type)
1352 raw_spin_lock(&ctx->lock);
1354 if (likely(!ctx->nr_events))
1357 ctx->timestamp = perf_clock();
1360 * First go through the list and put on any pinned groups
1361 * in order to give them the best chance of going on.
1363 if (event_type & EVENT_PINNED)
1364 ctx_pinned_sched_in(ctx, cpuctx);
1366 /* Then walk through the lower prio flexible groups */
1367 if (event_type & EVENT_FLEXIBLE)
1368 ctx_flexible_sched_in(ctx, cpuctx);
1371 raw_spin_unlock(&ctx->lock);
1374 static void cpu_ctx_sched_in(struct perf_cpu_context *cpuctx,
1375 enum event_type_t event_type)
1377 struct perf_event_context *ctx = &cpuctx->ctx;
1379 ctx_sched_in(ctx, cpuctx, event_type);
1382 static void task_ctx_sched_in(struct perf_event_context *ctx,
1383 enum event_type_t event_type)
1385 struct perf_cpu_context *cpuctx;
1387 cpuctx = __get_cpu_context(ctx);
1388 if (cpuctx->task_ctx == ctx)
1391 ctx_sched_in(ctx, cpuctx, event_type);
1392 cpuctx->task_ctx = ctx;
1395 void perf_event_context_sched_in(struct perf_event_context *ctx)
1397 struct perf_cpu_context *cpuctx;
1399 cpuctx = __get_cpu_context(ctx);
1400 if (cpuctx->task_ctx == ctx)
1404 * We want to keep the following priority order:
1405 * cpu pinned (that don't need to move), task pinned,
1406 * cpu flexible, task flexible.
1408 cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE);
1410 ctx_sched_in(ctx, cpuctx, EVENT_PINNED);
1411 cpu_ctx_sched_in(cpuctx, EVENT_FLEXIBLE);
1412 ctx_sched_in(ctx, cpuctx, EVENT_FLEXIBLE);
1414 cpuctx->task_ctx = ctx;
1417 * Since these rotations are per-cpu, we need to ensure the
1418 * cpu-context we got scheduled on is actually rotating.
1420 perf_pmu_rotate_start(ctx->pmu);
1424 * Called from scheduler to add the events of the current task
1425 * with interrupts disabled.
1427 * We restore the event value and then enable it.
1429 * This does not protect us against NMI, but enable()
1430 * sets the enabled bit in the control field of event _before_
1431 * accessing the event control register. If a NMI hits, then it will
1432 * keep the event running.
1434 void perf_event_task_sched_in(struct task_struct *task)
1436 struct perf_event_context *ctx;
1439 for_each_task_context_nr(ctxn) {
1440 ctx = task->perf_event_ctxp[ctxn];
1444 perf_event_context_sched_in(ctx);
1448 #define MAX_INTERRUPTS (~0ULL)
1450 static void perf_log_throttle(struct perf_event *event, int enable);
1452 static u64 perf_calculate_period(struct perf_event *event, u64 nsec, u64 count)
1454 u64 frequency = event->attr.sample_freq;
1455 u64 sec = NSEC_PER_SEC;
1456 u64 divisor, dividend;
1458 int count_fls, nsec_fls, frequency_fls, sec_fls;
1460 count_fls = fls64(count);
1461 nsec_fls = fls64(nsec);
1462 frequency_fls = fls64(frequency);
1466 * We got @count in @nsec, with a target of sample_freq HZ
1467 * the target period becomes:
1470 * period = -------------------
1471 * @nsec * sample_freq
1476 * Reduce accuracy by one bit such that @a and @b converge
1477 * to a similar magnitude.
1479 #define REDUCE_FLS(a, b) \
1481 if (a##_fls > b##_fls) { \
1491 * Reduce accuracy until either term fits in a u64, then proceed with
1492 * the other, so that finally we can do a u64/u64 division.
1494 while (count_fls + sec_fls > 64 && nsec_fls + frequency_fls > 64) {
1495 REDUCE_FLS(nsec, frequency);
1496 REDUCE_FLS(sec, count);
1499 if (count_fls + sec_fls > 64) {
1500 divisor = nsec * frequency;
1502 while (count_fls + sec_fls > 64) {
1503 REDUCE_FLS(count, sec);
1507 dividend = count * sec;
1509 dividend = count * sec;
1511 while (nsec_fls + frequency_fls > 64) {
1512 REDUCE_FLS(nsec, frequency);
1516 divisor = nsec * frequency;
1522 return div64_u64(dividend, divisor);
1525 static void perf_adjust_period(struct perf_event *event, u64 nsec, u64 count)
1527 struct hw_perf_event *hwc = &event->hw;
1528 s64 period, sample_period;
1531 period = perf_calculate_period(event, nsec, count);
1533 delta = (s64)(period - hwc->sample_period);
1534 delta = (delta + 7) / 8; /* low pass filter */
1536 sample_period = hwc->sample_period + delta;
1541 hwc->sample_period = sample_period;
1543 if (local64_read(&hwc->period_left) > 8*sample_period) {
1544 event->pmu->stop(event, PERF_EF_UPDATE);
1545 local64_set(&hwc->period_left, 0);
1546 event->pmu->start(event, PERF_EF_RELOAD);
1550 static void perf_ctx_adjust_freq(struct perf_event_context *ctx, u64 period)
1552 struct perf_event *event;
1553 struct hw_perf_event *hwc;
1554 u64 interrupts, now;
1557 raw_spin_lock(&ctx->lock);
1558 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
1559 if (event->state != PERF_EVENT_STATE_ACTIVE)
1562 if (event->cpu != -1 && event->cpu != smp_processor_id())
1567 interrupts = hwc->interrupts;
1568 hwc->interrupts = 0;
1571 * unthrottle events on the tick
1573 if (interrupts == MAX_INTERRUPTS) {
1574 perf_log_throttle(event, 1);
1575 event->pmu->start(event, 0);
1578 if (!event->attr.freq || !event->attr.sample_freq)
1581 event->pmu->read(event);
1582 now = local64_read(&event->count);
1583 delta = now - hwc->freq_count_stamp;
1584 hwc->freq_count_stamp = now;
1587 perf_adjust_period(event, period, delta);
1589 raw_spin_unlock(&ctx->lock);
1593 * Round-robin a context's events:
1595 static void rotate_ctx(struct perf_event_context *ctx)
1597 raw_spin_lock(&ctx->lock);
1599 /* Rotate the first entry last of non-pinned groups */
1600 list_rotate_left(&ctx->flexible_groups);
1602 raw_spin_unlock(&ctx->lock);
1606 * Cannot race with ->pmu_rotate_start() because this is ran from hardirq
1607 * context, and ->pmu_rotate_start() is called with irqs disabled (both are
1608 * cpu affine, so there are no SMP races).
1610 static enum hrtimer_restart perf_event_context_tick(struct hrtimer *timer)
1612 enum hrtimer_restart restart = HRTIMER_NORESTART;
1613 struct perf_cpu_context *cpuctx;
1614 struct perf_event_context *ctx = NULL;
1617 cpuctx = container_of(timer, struct perf_cpu_context, timer);
1619 if (cpuctx->ctx.nr_events) {
1620 restart = HRTIMER_RESTART;
1621 if (cpuctx->ctx.nr_events != cpuctx->ctx.nr_active)
1625 ctx = cpuctx->task_ctx;
1626 if (ctx && ctx->nr_events) {
1627 restart = HRTIMER_RESTART;
1628 if (ctx->nr_events != ctx->nr_active)
1632 perf_ctx_adjust_freq(&cpuctx->ctx, cpuctx->timer_interval);
1634 perf_ctx_adjust_freq(ctx, cpuctx->timer_interval);
1639 cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE);
1641 task_ctx_sched_out(ctx, EVENT_FLEXIBLE);
1643 rotate_ctx(&cpuctx->ctx);
1647 cpu_ctx_sched_in(cpuctx, EVENT_FLEXIBLE);
1649 task_ctx_sched_in(ctx, EVENT_FLEXIBLE);
1652 hrtimer_forward_now(timer, ns_to_ktime(cpuctx->timer_interval));
1657 static int event_enable_on_exec(struct perf_event *event,
1658 struct perf_event_context *ctx)
1660 if (!event->attr.enable_on_exec)
1663 event->attr.enable_on_exec = 0;
1664 if (event->state >= PERF_EVENT_STATE_INACTIVE)
1667 __perf_event_mark_enabled(event, ctx);
1673 * Enable all of a task's events that have been marked enable-on-exec.
1674 * This expects task == current.
1676 static void perf_event_enable_on_exec(struct perf_event_context *ctx)
1678 struct perf_event *event;
1679 unsigned long flags;
1683 local_irq_save(flags);
1684 if (!ctx || !ctx->nr_events)
1687 task_ctx_sched_out(ctx, EVENT_ALL);
1689 raw_spin_lock(&ctx->lock);
1691 list_for_each_entry(event, &ctx->pinned_groups, group_entry) {
1692 ret = event_enable_on_exec(event, ctx);
1697 list_for_each_entry(event, &ctx->flexible_groups, group_entry) {
1698 ret = event_enable_on_exec(event, ctx);
1704 * Unclone this context if we enabled any event.
1709 raw_spin_unlock(&ctx->lock);
1711 perf_event_context_sched_in(ctx);
1713 local_irq_restore(flags);
1717 * Cross CPU call to read the hardware event
1719 static void __perf_event_read(void *info)
1721 struct perf_event *event = info;
1722 struct perf_event_context *ctx = event->ctx;
1723 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
1726 * If this is a task context, we need to check whether it is
1727 * the current task context of this cpu. If not it has been
1728 * scheduled out before the smp call arrived. In that case
1729 * event->count would have been updated to a recent sample
1730 * when the event was scheduled out.
1732 if (ctx->task && cpuctx->task_ctx != ctx)
1735 raw_spin_lock(&ctx->lock);
1736 update_context_time(ctx);
1737 update_event_times(event);
1738 raw_spin_unlock(&ctx->lock);
1740 event->pmu->read(event);
1743 static inline u64 perf_event_count(struct perf_event *event)
1745 return local64_read(&event->count) + atomic64_read(&event->child_count);
1748 static u64 perf_event_read(struct perf_event *event)
1751 * If event is enabled and currently active on a CPU, update the
1752 * value in the event structure:
1754 if (event->state == PERF_EVENT_STATE_ACTIVE) {
1755 smp_call_function_single(event->oncpu,
1756 __perf_event_read, event, 1);
1757 } else if (event->state == PERF_EVENT_STATE_INACTIVE) {
1758 struct perf_event_context *ctx = event->ctx;
1759 unsigned long flags;
1761 raw_spin_lock_irqsave(&ctx->lock, flags);
1762 update_context_time(ctx);
1763 update_event_times(event);
1764 raw_spin_unlock_irqrestore(&ctx->lock, flags);
1767 return perf_event_count(event);
1774 struct callchain_cpus_entries {
1775 struct rcu_head rcu_head;
1776 struct perf_callchain_entry *cpu_entries[0];
1779 static DEFINE_PER_CPU(int, callchain_recursion[PERF_NR_CONTEXTS]);
1780 static atomic_t nr_callchain_events;
1781 static DEFINE_MUTEX(callchain_mutex);
1782 struct callchain_cpus_entries *callchain_cpus_entries;
1785 __weak void perf_callchain_kernel(struct perf_callchain_entry *entry,
1786 struct pt_regs *regs)
1790 __weak void perf_callchain_user(struct perf_callchain_entry *entry,
1791 struct pt_regs *regs)
1795 static void release_callchain_buffers_rcu(struct rcu_head *head)
1797 struct callchain_cpus_entries *entries;
1800 entries = container_of(head, struct callchain_cpus_entries, rcu_head);
1802 for_each_possible_cpu(cpu)
1803 kfree(entries->cpu_entries[cpu]);
1808 static void release_callchain_buffers(void)
1810 struct callchain_cpus_entries *entries;
1812 entries = callchain_cpus_entries;
1813 rcu_assign_pointer(callchain_cpus_entries, NULL);
1814 call_rcu(&entries->rcu_head, release_callchain_buffers_rcu);
1817 static int alloc_callchain_buffers(void)
1821 struct callchain_cpus_entries *entries;
1824 * We can't use the percpu allocation API for data that can be
1825 * accessed from NMI. Use a temporary manual per cpu allocation
1826 * until that gets sorted out.
1828 size = sizeof(*entries) + sizeof(struct perf_callchain_entry *) *
1829 num_possible_cpus();
1831 entries = kzalloc(size, GFP_KERNEL);
1835 size = sizeof(struct perf_callchain_entry) * PERF_NR_CONTEXTS;
1837 for_each_possible_cpu(cpu) {
1838 entries->cpu_entries[cpu] = kmalloc_node(size, GFP_KERNEL,
1840 if (!entries->cpu_entries[cpu])
1844 rcu_assign_pointer(callchain_cpus_entries, entries);
1849 for_each_possible_cpu(cpu)
1850 kfree(entries->cpu_entries[cpu]);
1856 static int get_callchain_buffers(void)
1861 mutex_lock(&callchain_mutex);
1863 count = atomic_inc_return(&nr_callchain_events);
1864 if (WARN_ON_ONCE(count < 1)) {
1870 /* If the allocation failed, give up */
1871 if (!callchain_cpus_entries)
1876 err = alloc_callchain_buffers();
1878 release_callchain_buffers();
1880 mutex_unlock(&callchain_mutex);
1885 static void put_callchain_buffers(void)
1887 if (atomic_dec_and_mutex_lock(&nr_callchain_events, &callchain_mutex)) {
1888 release_callchain_buffers();
1889 mutex_unlock(&callchain_mutex);
1893 static int get_recursion_context(int *recursion)
1901 else if (in_softirq())
1906 if (recursion[rctx])
1915 static inline void put_recursion_context(int *recursion, int rctx)
1921 static struct perf_callchain_entry *get_callchain_entry(int *rctx)
1924 struct callchain_cpus_entries *entries;
1926 *rctx = get_recursion_context(__get_cpu_var(callchain_recursion));
1930 entries = rcu_dereference(callchain_cpus_entries);
1934 cpu = smp_processor_id();
1936 return &entries->cpu_entries[cpu][*rctx];
1940 put_callchain_entry(int rctx)
1942 put_recursion_context(__get_cpu_var(callchain_recursion), rctx);
1945 static struct perf_callchain_entry *perf_callchain(struct pt_regs *regs)
1948 struct perf_callchain_entry *entry;
1951 entry = get_callchain_entry(&rctx);
1960 if (!user_mode(regs)) {
1961 perf_callchain_store(entry, PERF_CONTEXT_KERNEL);
1962 perf_callchain_kernel(entry, regs);
1964 regs = task_pt_regs(current);
1970 perf_callchain_store(entry, PERF_CONTEXT_USER);
1971 perf_callchain_user(entry, regs);
1975 put_callchain_entry(rctx);
1981 * Initialize the perf_event context in a task_struct:
1983 static void __perf_event_init_context(struct perf_event_context *ctx)
1985 raw_spin_lock_init(&ctx->lock);
1986 mutex_init(&ctx->mutex);
1987 INIT_LIST_HEAD(&ctx->pinned_groups);
1988 INIT_LIST_HEAD(&ctx->flexible_groups);
1989 INIT_LIST_HEAD(&ctx->event_list);
1990 atomic_set(&ctx->refcount, 1);
1993 static struct perf_event_context *
1994 alloc_perf_context(struct pmu *pmu, struct task_struct *task)
1996 struct perf_event_context *ctx;
1998 ctx = kzalloc(sizeof(struct perf_event_context), GFP_KERNEL);
2002 __perf_event_init_context(ctx);
2005 get_task_struct(task);
2012 static struct perf_event_context *
2013 find_get_context(struct pmu *pmu, pid_t pid, int cpu)
2015 struct perf_event_context *ctx;
2016 struct perf_cpu_context *cpuctx;
2017 struct task_struct *task;
2018 unsigned long flags;
2021 if (pid == -1 && cpu != -1) {
2022 /* Must be root to operate on a CPU event: */
2023 if (perf_paranoid_cpu() && !capable(CAP_SYS_ADMIN))
2024 return ERR_PTR(-EACCES);
2026 if (cpu < 0 || cpu >= nr_cpumask_bits)
2027 return ERR_PTR(-EINVAL);
2030 * We could be clever and allow to attach a event to an
2031 * offline CPU and activate it when the CPU comes up, but
2034 if (!cpu_online(cpu))
2035 return ERR_PTR(-ENODEV);
2037 cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu);
2048 task = find_task_by_vpid(pid);
2050 get_task_struct(task);
2054 return ERR_PTR(-ESRCH);
2057 * Can't attach events to a dying task.
2060 if (task->flags & PF_EXITING)
2063 /* Reuse ptrace permission checks for now. */
2065 if (!ptrace_may_access(task, PTRACE_MODE_READ))
2069 ctxn = pmu->task_ctx_nr;
2074 ctx = perf_lock_task_context(task, ctxn, &flags);
2077 raw_spin_unlock_irqrestore(&ctx->lock, flags);
2081 ctx = alloc_perf_context(pmu, task);
2088 if (cmpxchg(&task->perf_event_ctxp[ctxn], NULL, ctx)) {
2090 * We raced with some other task; use
2091 * the context they set.
2093 put_task_struct(task);
2099 put_task_struct(task);
2103 put_task_struct(task);
2104 return ERR_PTR(err);
2107 static void perf_event_free_filter(struct perf_event *event);
2109 static void free_event_rcu(struct rcu_head *head)
2111 struct perf_event *event;
2113 event = container_of(head, struct perf_event, rcu_head);
2115 put_pid_ns(event->ns);
2116 perf_event_free_filter(event);
2120 static void perf_pending_sync(struct perf_event *event);
2121 static void perf_buffer_put(struct perf_buffer *buffer);
2123 static void free_event(struct perf_event *event)
2125 perf_pending_sync(event);
2127 if (!event->parent) {
2128 atomic_dec(&nr_events);
2129 if (event->attr.mmap || event->attr.mmap_data)
2130 atomic_dec(&nr_mmap_events);
2131 if (event->attr.comm)
2132 atomic_dec(&nr_comm_events);
2133 if (event->attr.task)
2134 atomic_dec(&nr_task_events);
2135 if (event->attr.sample_type & PERF_SAMPLE_CALLCHAIN)
2136 put_callchain_buffers();
2139 if (event->buffer) {
2140 perf_buffer_put(event->buffer);
2141 event->buffer = NULL;
2145 event->destroy(event);
2147 put_ctx(event->ctx);
2148 call_rcu(&event->rcu_head, free_event_rcu);
2151 int perf_event_release_kernel(struct perf_event *event)
2153 struct perf_event_context *ctx = event->ctx;
2156 * Remove from the PMU, can't get re-enabled since we got
2157 * here because the last ref went.
2159 perf_event_disable(event);
2161 WARN_ON_ONCE(ctx->parent_ctx);
2163 * There are two ways this annotation is useful:
2165 * 1) there is a lock recursion from perf_event_exit_task
2166 * see the comment there.
2168 * 2) there is a lock-inversion with mmap_sem through
2169 * perf_event_read_group(), which takes faults while
2170 * holding ctx->mutex, however this is called after
2171 * the last filedesc died, so there is no possibility
2172 * to trigger the AB-BA case.
2174 mutex_lock_nested(&ctx->mutex, SINGLE_DEPTH_NESTING);
2175 raw_spin_lock_irq(&ctx->lock);
2176 perf_group_detach(event);
2177 list_del_event(event, ctx);
2178 raw_spin_unlock_irq(&ctx->lock);
2179 mutex_unlock(&ctx->mutex);
2181 mutex_lock(&event->owner->perf_event_mutex);
2182 list_del_init(&event->owner_entry);
2183 mutex_unlock(&event->owner->perf_event_mutex);
2184 put_task_struct(event->owner);
2190 EXPORT_SYMBOL_GPL(perf_event_release_kernel);
2193 * Called when the last reference to the file is gone.
2195 static int perf_release(struct inode *inode, struct file *file)
2197 struct perf_event *event = file->private_data;
2199 file->private_data = NULL;
2201 return perf_event_release_kernel(event);
2204 static int perf_event_read_size(struct perf_event *event)
2206 int entry = sizeof(u64); /* value */
2210 if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
2211 size += sizeof(u64);
2213 if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
2214 size += sizeof(u64);
2216 if (event->attr.read_format & PERF_FORMAT_ID)
2217 entry += sizeof(u64);
2219 if (event->attr.read_format & PERF_FORMAT_GROUP) {
2220 nr += event->group_leader->nr_siblings;
2221 size += sizeof(u64);
2229 u64 perf_event_read_value(struct perf_event *event, u64 *enabled, u64 *running)
2231 struct perf_event *child;
2237 mutex_lock(&event->child_mutex);
2238 total += perf_event_read(event);
2239 *enabled += event->total_time_enabled +
2240 atomic64_read(&event->child_total_time_enabled);
2241 *running += event->total_time_running +
2242 atomic64_read(&event->child_total_time_running);
2244 list_for_each_entry(child, &event->child_list, child_list) {
2245 total += perf_event_read(child);
2246 *enabled += child->total_time_enabled;
2247 *running += child->total_time_running;
2249 mutex_unlock(&event->child_mutex);
2253 EXPORT_SYMBOL_GPL(perf_event_read_value);
2255 static int perf_event_read_group(struct perf_event *event,
2256 u64 read_format, char __user *buf)
2258 struct perf_event *leader = event->group_leader, *sub;
2259 int n = 0, size = 0, ret = -EFAULT;
2260 struct perf_event_context *ctx = leader->ctx;
2262 u64 count, enabled, running;
2264 mutex_lock(&ctx->mutex);
2265 count = perf_event_read_value(leader, &enabled, &running);
2267 values[n++] = 1 + leader->nr_siblings;
2268 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
2269 values[n++] = enabled;
2270 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
2271 values[n++] = running;
2272 values[n++] = count;
2273 if (read_format & PERF_FORMAT_ID)
2274 values[n++] = primary_event_id(leader);
2276 size = n * sizeof(u64);
2278 if (copy_to_user(buf, values, size))
2283 list_for_each_entry(sub, &leader->sibling_list, group_entry) {
2286 values[n++] = perf_event_read_value(sub, &enabled, &running);
2287 if (read_format & PERF_FORMAT_ID)
2288 values[n++] = primary_event_id(sub);
2290 size = n * sizeof(u64);
2292 if (copy_to_user(buf + ret, values, size)) {
2300 mutex_unlock(&ctx->mutex);
2305 static int perf_event_read_one(struct perf_event *event,
2306 u64 read_format, char __user *buf)
2308 u64 enabled, running;
2312 values[n++] = perf_event_read_value(event, &enabled, &running);
2313 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
2314 values[n++] = enabled;
2315 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
2316 values[n++] = running;
2317 if (read_format & PERF_FORMAT_ID)
2318 values[n++] = primary_event_id(event);
2320 if (copy_to_user(buf, values, n * sizeof(u64)))
2323 return n * sizeof(u64);
2327 * Read the performance event - simple non blocking version for now
2330 perf_read_hw(struct perf_event *event, char __user *buf, size_t count)
2332 u64 read_format = event->attr.read_format;
2336 * Return end-of-file for a read on a event that is in
2337 * error state (i.e. because it was pinned but it couldn't be
2338 * scheduled on to the CPU at some point).
2340 if (event->state == PERF_EVENT_STATE_ERROR)
2343 if (count < perf_event_read_size(event))
2346 WARN_ON_ONCE(event->ctx->parent_ctx);
2347 if (read_format & PERF_FORMAT_GROUP)
2348 ret = perf_event_read_group(event, read_format, buf);
2350 ret = perf_event_read_one(event, read_format, buf);
2356 perf_read(struct file *file, char __user *buf, size_t count, loff_t *ppos)
2358 struct perf_event *event = file->private_data;
2360 return perf_read_hw(event, buf, count);
2363 static unsigned int perf_poll(struct file *file, poll_table *wait)
2365 struct perf_event *event = file->private_data;
2366 struct perf_buffer *buffer;
2367 unsigned int events = POLL_HUP;
2370 buffer = rcu_dereference(event->buffer);
2372 events = atomic_xchg(&buffer->poll, 0);
2375 poll_wait(file, &event->waitq, wait);
2380 static void perf_event_reset(struct perf_event *event)
2382 (void)perf_event_read(event);
2383 local64_set(&event->count, 0);
2384 perf_event_update_userpage(event);
2388 * Holding the top-level event's child_mutex means that any
2389 * descendant process that has inherited this event will block
2390 * in sync_child_event if it goes to exit, thus satisfying the
2391 * task existence requirements of perf_event_enable/disable.
2393 static void perf_event_for_each_child(struct perf_event *event,
2394 void (*func)(struct perf_event *))
2396 struct perf_event *child;
2398 WARN_ON_ONCE(event->ctx->parent_ctx);
2399 mutex_lock(&event->child_mutex);
2401 list_for_each_entry(child, &event->child_list, child_list)
2403 mutex_unlock(&event->child_mutex);
2406 static void perf_event_for_each(struct perf_event *event,
2407 void (*func)(struct perf_event *))
2409 struct perf_event_context *ctx = event->ctx;
2410 struct perf_event *sibling;
2412 WARN_ON_ONCE(ctx->parent_ctx);
2413 mutex_lock(&ctx->mutex);
2414 event = event->group_leader;
2416 perf_event_for_each_child(event, func);
2418 list_for_each_entry(sibling, &event->sibling_list, group_entry)
2419 perf_event_for_each_child(event, func);
2420 mutex_unlock(&ctx->mutex);
2423 static int perf_event_period(struct perf_event *event, u64 __user *arg)
2425 struct perf_event_context *ctx = event->ctx;
2430 if (!event->attr.sample_period)
2433 size = copy_from_user(&value, arg, sizeof(value));
2434 if (size != sizeof(value))
2440 raw_spin_lock_irq(&ctx->lock);
2441 if (event->attr.freq) {
2442 if (value > sysctl_perf_event_sample_rate) {
2447 event->attr.sample_freq = value;
2449 event->attr.sample_period = value;
2450 event->hw.sample_period = value;
2453 raw_spin_unlock_irq(&ctx->lock);
2458 static const struct file_operations perf_fops;
2460 static struct perf_event *perf_fget_light(int fd, int *fput_needed)
2464 file = fget_light(fd, fput_needed);
2466 return ERR_PTR(-EBADF);
2468 if (file->f_op != &perf_fops) {
2469 fput_light(file, *fput_needed);
2471 return ERR_PTR(-EBADF);
2474 return file->private_data;
2477 static int perf_event_set_output(struct perf_event *event,
2478 struct perf_event *output_event);
2479 static int perf_event_set_filter(struct perf_event *event, void __user *arg);
2481 static long perf_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
2483 struct perf_event *event = file->private_data;
2484 void (*func)(struct perf_event *);
2488 case PERF_EVENT_IOC_ENABLE:
2489 func = perf_event_enable;
2491 case PERF_EVENT_IOC_DISABLE:
2492 func = perf_event_disable;
2494 case PERF_EVENT_IOC_RESET:
2495 func = perf_event_reset;
2498 case PERF_EVENT_IOC_REFRESH:
2499 return perf_event_refresh(event, arg);
2501 case PERF_EVENT_IOC_PERIOD:
2502 return perf_event_period(event, (u64 __user *)arg);
2504 case PERF_EVENT_IOC_SET_OUTPUT:
2506 struct perf_event *output_event = NULL;
2507 int fput_needed = 0;
2511 output_event = perf_fget_light(arg, &fput_needed);
2512 if (IS_ERR(output_event))
2513 return PTR_ERR(output_event);
2516 ret = perf_event_set_output(event, output_event);
2518 fput_light(output_event->filp, fput_needed);
2523 case PERF_EVENT_IOC_SET_FILTER:
2524 return perf_event_set_filter(event, (void __user *)arg);
2530 if (flags & PERF_IOC_FLAG_GROUP)
2531 perf_event_for_each(event, func);
2533 perf_event_for_each_child(event, func);
2538 int perf_event_task_enable(void)
2540 struct perf_event *event;
2542 mutex_lock(¤t->perf_event_mutex);
2543 list_for_each_entry(event, ¤t->perf_event_list, owner_entry)
2544 perf_event_for_each_child(event, perf_event_enable);
2545 mutex_unlock(¤t->perf_event_mutex);
2550 int perf_event_task_disable(void)
2552 struct perf_event *event;
2554 mutex_lock(¤t->perf_event_mutex);
2555 list_for_each_entry(event, ¤t->perf_event_list, owner_entry)
2556 perf_event_for_each_child(event, perf_event_disable);
2557 mutex_unlock(¤t->perf_event_mutex);
2562 #ifndef PERF_EVENT_INDEX_OFFSET
2563 # define PERF_EVENT_INDEX_OFFSET 0
2566 static int perf_event_index(struct perf_event *event)
2568 if (event->hw.state & PERF_HES_STOPPED)
2571 if (event->state != PERF_EVENT_STATE_ACTIVE)
2574 return event->hw.idx + 1 - PERF_EVENT_INDEX_OFFSET;
2578 * Callers need to ensure there can be no nesting of this function, otherwise
2579 * the seqlock logic goes bad. We can not serialize this because the arch
2580 * code calls this from NMI context.
2582 void perf_event_update_userpage(struct perf_event *event)
2584 struct perf_event_mmap_page *userpg;
2585 struct perf_buffer *buffer;
2588 buffer = rcu_dereference(event->buffer);
2592 userpg = buffer->user_page;
2595 * Disable preemption so as to not let the corresponding user-space
2596 * spin too long if we get preempted.
2601 userpg->index = perf_event_index(event);
2602 userpg->offset = perf_event_count(event);
2603 if (event->state == PERF_EVENT_STATE_ACTIVE)
2604 userpg->offset -= local64_read(&event->hw.prev_count);
2606 userpg->time_enabled = event->total_time_enabled +
2607 atomic64_read(&event->child_total_time_enabled);
2609 userpg->time_running = event->total_time_running +
2610 atomic64_read(&event->child_total_time_running);
2619 static unsigned long perf_data_size(struct perf_buffer *buffer);
2622 perf_buffer_init(struct perf_buffer *buffer, long watermark, int flags)
2624 long max_size = perf_data_size(buffer);
2627 buffer->watermark = min(max_size, watermark);
2629 if (!buffer->watermark)
2630 buffer->watermark = max_size / 2;
2632 if (flags & PERF_BUFFER_WRITABLE)
2633 buffer->writable = 1;
2635 atomic_set(&buffer->refcount, 1);
2638 #ifndef CONFIG_PERF_USE_VMALLOC
2641 * Back perf_mmap() with regular GFP_KERNEL-0 pages.
2644 static struct page *
2645 perf_mmap_to_page(struct perf_buffer *buffer, unsigned long pgoff)
2647 if (pgoff > buffer->nr_pages)
2651 return virt_to_page(buffer->user_page);
2653 return virt_to_page(buffer->data_pages[pgoff - 1]);
2656 static void *perf_mmap_alloc_page(int cpu)
2661 node = (cpu == -1) ? cpu : cpu_to_node(cpu);
2662 page = alloc_pages_node(node, GFP_KERNEL | __GFP_ZERO, 0);
2666 return page_address(page);
2669 static struct perf_buffer *
2670 perf_buffer_alloc(int nr_pages, long watermark, int cpu, int flags)
2672 struct perf_buffer *buffer;
2676 size = sizeof(struct perf_buffer);
2677 size += nr_pages * sizeof(void *);
2679 buffer = kzalloc(size, GFP_KERNEL);
2683 buffer->user_page = perf_mmap_alloc_page(cpu);
2684 if (!buffer->user_page)
2685 goto fail_user_page;
2687 for (i = 0; i < nr_pages; i++) {
2688 buffer->data_pages[i] = perf_mmap_alloc_page(cpu);
2689 if (!buffer->data_pages[i])
2690 goto fail_data_pages;
2693 buffer->nr_pages = nr_pages;
2695 perf_buffer_init(buffer, watermark, flags);
2700 for (i--; i >= 0; i--)
2701 free_page((unsigned long)buffer->data_pages[i]);
2703 free_page((unsigned long)buffer->user_page);
2712 static void perf_mmap_free_page(unsigned long addr)
2714 struct page *page = virt_to_page((void *)addr);
2716 page->mapping = NULL;
2720 static void perf_buffer_free(struct perf_buffer *buffer)
2724 perf_mmap_free_page((unsigned long)buffer->user_page);
2725 for (i = 0; i < buffer->nr_pages; i++)
2726 perf_mmap_free_page((unsigned long)buffer->data_pages[i]);
2730 static inline int page_order(struct perf_buffer *buffer)
2738 * Back perf_mmap() with vmalloc memory.
2740 * Required for architectures that have d-cache aliasing issues.
2743 static inline int page_order(struct perf_buffer *buffer)
2745 return buffer->page_order;
2748 static struct page *
2749 perf_mmap_to_page(struct perf_buffer *buffer, unsigned long pgoff)
2751 if (pgoff > (1UL << page_order(buffer)))
2754 return vmalloc_to_page((void *)buffer->user_page + pgoff * PAGE_SIZE);
2757 static void perf_mmap_unmark_page(void *addr)
2759 struct page *page = vmalloc_to_page(addr);
2761 page->mapping = NULL;
2764 static void perf_buffer_free_work(struct work_struct *work)
2766 struct perf_buffer *buffer;
2770 buffer = container_of(work, struct perf_buffer, work);
2771 nr = 1 << page_order(buffer);
2773 base = buffer->user_page;
2774 for (i = 0; i < nr + 1; i++)
2775 perf_mmap_unmark_page(base + (i * PAGE_SIZE));
2781 static void perf_buffer_free(struct perf_buffer *buffer)
2783 schedule_work(&buffer->work);
2786 static struct perf_buffer *
2787 perf_buffer_alloc(int nr_pages, long watermark, int cpu, int flags)
2789 struct perf_buffer *buffer;
2793 size = sizeof(struct perf_buffer);
2794 size += sizeof(void *);
2796 buffer = kzalloc(size, GFP_KERNEL);
2800 INIT_WORK(&buffer->work, perf_buffer_free_work);
2802 all_buf = vmalloc_user((nr_pages + 1) * PAGE_SIZE);
2806 buffer->user_page = all_buf;
2807 buffer->data_pages[0] = all_buf + PAGE_SIZE;
2808 buffer->page_order = ilog2(nr_pages);
2809 buffer->nr_pages = 1;
2811 perf_buffer_init(buffer, watermark, flags);
2824 static unsigned long perf_data_size(struct perf_buffer *buffer)
2826 return buffer->nr_pages << (PAGE_SHIFT + page_order(buffer));
2829 static int perf_mmap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
2831 struct perf_event *event = vma->vm_file->private_data;
2832 struct perf_buffer *buffer;
2833 int ret = VM_FAULT_SIGBUS;
2835 if (vmf->flags & FAULT_FLAG_MKWRITE) {
2836 if (vmf->pgoff == 0)
2842 buffer = rcu_dereference(event->buffer);
2846 if (vmf->pgoff && (vmf->flags & FAULT_FLAG_WRITE))
2849 vmf->page = perf_mmap_to_page(buffer, vmf->pgoff);
2853 get_page(vmf->page);
2854 vmf->page->mapping = vma->vm_file->f_mapping;
2855 vmf->page->index = vmf->pgoff;
2864 static void perf_buffer_free_rcu(struct rcu_head *rcu_head)
2866 struct perf_buffer *buffer;
2868 buffer = container_of(rcu_head, struct perf_buffer, rcu_head);
2869 perf_buffer_free(buffer);
2872 static struct perf_buffer *perf_buffer_get(struct perf_event *event)
2874 struct perf_buffer *buffer;
2877 buffer = rcu_dereference(event->buffer);
2879 if (!atomic_inc_not_zero(&buffer->refcount))
2887 static void perf_buffer_put(struct perf_buffer *buffer)
2889 if (!atomic_dec_and_test(&buffer->refcount))
2892 call_rcu(&buffer->rcu_head, perf_buffer_free_rcu);
2895 static void perf_mmap_open(struct vm_area_struct *vma)
2897 struct perf_event *event = vma->vm_file->private_data;
2899 atomic_inc(&event->mmap_count);
2902 static void perf_mmap_close(struct vm_area_struct *vma)
2904 struct perf_event *event = vma->vm_file->private_data;
2906 if (atomic_dec_and_mutex_lock(&event->mmap_count, &event->mmap_mutex)) {
2907 unsigned long size = perf_data_size(event->buffer);
2908 struct user_struct *user = event->mmap_user;
2909 struct perf_buffer *buffer = event->buffer;
2911 atomic_long_sub((size >> PAGE_SHIFT) + 1, &user->locked_vm);
2912 vma->vm_mm->locked_vm -= event->mmap_locked;
2913 rcu_assign_pointer(event->buffer, NULL);
2914 mutex_unlock(&event->mmap_mutex);
2916 perf_buffer_put(buffer);
2921 static const struct vm_operations_struct perf_mmap_vmops = {
2922 .open = perf_mmap_open,
2923 .close = perf_mmap_close,
2924 .fault = perf_mmap_fault,
2925 .page_mkwrite = perf_mmap_fault,
2928 static int perf_mmap(struct file *file, struct vm_area_struct *vma)
2930 struct perf_event *event = file->private_data;
2931 unsigned long user_locked, user_lock_limit;
2932 struct user_struct *user = current_user();
2933 unsigned long locked, lock_limit;
2934 struct perf_buffer *buffer;
2935 unsigned long vma_size;
2936 unsigned long nr_pages;
2937 long user_extra, extra;
2938 int ret = 0, flags = 0;
2941 * Don't allow mmap() of inherited per-task counters. This would
2942 * create a performance issue due to all children writing to the
2945 if (event->cpu == -1 && event->attr.inherit)
2948 if (!(vma->vm_flags & VM_SHARED))
2951 vma_size = vma->vm_end - vma->vm_start;
2952 nr_pages = (vma_size / PAGE_SIZE) - 1;
2955 * If we have buffer pages ensure they're a power-of-two number, so we
2956 * can do bitmasks instead of modulo.
2958 if (nr_pages != 0 && !is_power_of_2(nr_pages))
2961 if (vma_size != PAGE_SIZE * (1 + nr_pages))
2964 if (vma->vm_pgoff != 0)
2967 WARN_ON_ONCE(event->ctx->parent_ctx);
2968 mutex_lock(&event->mmap_mutex);
2969 if (event->buffer) {
2970 if (event->buffer->nr_pages == nr_pages)
2971 atomic_inc(&event->buffer->refcount);
2977 user_extra = nr_pages + 1;
2978 user_lock_limit = sysctl_perf_event_mlock >> (PAGE_SHIFT - 10);
2981 * Increase the limit linearly with more CPUs:
2983 user_lock_limit *= num_online_cpus();
2985 user_locked = atomic_long_read(&user->locked_vm) + user_extra;
2988 if (user_locked > user_lock_limit)
2989 extra = user_locked - user_lock_limit;
2991 lock_limit = rlimit(RLIMIT_MEMLOCK);
2992 lock_limit >>= PAGE_SHIFT;
2993 locked = vma->vm_mm->locked_vm + extra;
2995 if ((locked > lock_limit) && perf_paranoid_tracepoint_raw() &&
2996 !capable(CAP_IPC_LOCK)) {
3001 WARN_ON(event->buffer);
3003 if (vma->vm_flags & VM_WRITE)
3004 flags |= PERF_BUFFER_WRITABLE;
3006 buffer = perf_buffer_alloc(nr_pages, event->attr.wakeup_watermark,
3012 rcu_assign_pointer(event->buffer, buffer);
3014 atomic_long_add(user_extra, &user->locked_vm);
3015 event->mmap_locked = extra;
3016 event->mmap_user = get_current_user();
3017 vma->vm_mm->locked_vm += event->mmap_locked;
3021 atomic_inc(&event->mmap_count);
3022 mutex_unlock(&event->mmap_mutex);
3024 vma->vm_flags |= VM_RESERVED;
3025 vma->vm_ops = &perf_mmap_vmops;
3030 static int perf_fasync(int fd, struct file *filp, int on)
3032 struct inode *inode = filp->f_path.dentry->d_inode;
3033 struct perf_event *event = filp->private_data;
3036 mutex_lock(&inode->i_mutex);
3037 retval = fasync_helper(fd, filp, on, &event->fasync);
3038 mutex_unlock(&inode->i_mutex);
3046 static const struct file_operations perf_fops = {
3047 .llseek = no_llseek,
3048 .release = perf_release,
3051 .unlocked_ioctl = perf_ioctl,
3052 .compat_ioctl = perf_ioctl,
3054 .fasync = perf_fasync,
3060 * If there's data, ensure we set the poll() state and publish everything
3061 * to user-space before waking everybody up.
3064 void perf_event_wakeup(struct perf_event *event)
3066 wake_up_all(&event->waitq);
3068 if (event->pending_kill) {
3069 kill_fasync(&event->fasync, SIGIO, event->pending_kill);
3070 event->pending_kill = 0;
3077 * Handle the case where we need to wakeup up from NMI (or rq->lock) context.
3079 * The NMI bit means we cannot possibly take locks. Therefore, maintain a
3080 * single linked list and use cmpxchg() to add entries lockless.
3083 static void perf_pending_event(struct perf_pending_entry *entry)
3085 struct perf_event *event = container_of(entry,
3086 struct perf_event, pending);
3088 if (event->pending_disable) {
3089 event->pending_disable = 0;
3090 __perf_event_disable(event);
3093 if (event->pending_wakeup) {
3094 event->pending_wakeup = 0;
3095 perf_event_wakeup(event);
3099 #define PENDING_TAIL ((struct perf_pending_entry *)-1UL)
3101 static DEFINE_PER_CPU(struct perf_pending_entry *, perf_pending_head) = {
3105 static void perf_pending_queue(struct perf_pending_entry *entry,
3106 void (*func)(struct perf_pending_entry *))
3108 struct perf_pending_entry **head;
3110 if (cmpxchg(&entry->next, NULL, PENDING_TAIL) != NULL)
3115 head = &get_cpu_var(perf_pending_head);
3118 entry->next = *head;
3119 } while (cmpxchg(head, entry->next, entry) != entry->next);
3121 set_perf_event_pending();
3123 put_cpu_var(perf_pending_head);
3126 static int __perf_pending_run(void)
3128 struct perf_pending_entry *list;
3131 list = xchg(&__get_cpu_var(perf_pending_head), PENDING_TAIL);
3132 while (list != PENDING_TAIL) {
3133 void (*func)(struct perf_pending_entry *);
3134 struct perf_pending_entry *entry = list;
3141 * Ensure we observe the unqueue before we issue the wakeup,
3142 * so that we won't be waiting forever.
3143 * -- see perf_not_pending().
3154 static inline int perf_not_pending(struct perf_event *event)
3157 * If we flush on whatever cpu we run, there is a chance we don't
3161 __perf_pending_run();
3165 * Ensure we see the proper queue state before going to sleep
3166 * so that we do not miss the wakeup. -- see perf_pending_handle()
3169 return event->pending.next == NULL;
3172 static void perf_pending_sync(struct perf_event *event)
3174 wait_event(event->waitq, perf_not_pending(event));
3177 void perf_event_do_pending(void)
3179 __perf_pending_run();
3183 * We assume there is only KVM supporting the callbacks.
3184 * Later on, we might change it to a list if there is
3185 * another virtualization implementation supporting the callbacks.
3187 struct perf_guest_info_callbacks *perf_guest_cbs;
3189 int perf_register_guest_info_callbacks(struct perf_guest_info_callbacks *cbs)
3191 perf_guest_cbs = cbs;
3194 EXPORT_SYMBOL_GPL(perf_register_guest_info_callbacks);
3196 int perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks *cbs)
3198 perf_guest_cbs = NULL;
3201 EXPORT_SYMBOL_GPL(perf_unregister_guest_info_callbacks);
3206 static bool perf_output_space(struct perf_buffer *buffer, unsigned long tail,
3207 unsigned long offset, unsigned long head)
3211 if (!buffer->writable)
3214 mask = perf_data_size(buffer) - 1;
3216 offset = (offset - tail) & mask;
3217 head = (head - tail) & mask;
3219 if ((int)(head - offset) < 0)
3225 static void perf_output_wakeup(struct perf_output_handle *handle)
3227 atomic_set(&handle->buffer->poll, POLL_IN);
3230 handle->event->pending_wakeup = 1;
3231 perf_pending_queue(&handle->event->pending,
3232 perf_pending_event);
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;
3807 rcu_read_lock_sched();
3808 list_for_each_entry_rcu(pmu, &pmus, entry) {
3809 cpuctx = this_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 rcu_read_unlock_sched();
3825 static void perf_event_task(struct task_struct *task,
3826 struct perf_event_context *task_ctx,
3829 struct perf_task_event task_event;
3831 if (!atomic_read(&nr_comm_events) &&
3832 !atomic_read(&nr_mmap_events) &&
3833 !atomic_read(&nr_task_events))
3836 task_event = (struct perf_task_event){
3838 .task_ctx = task_ctx,
3841 .type = new ? PERF_RECORD_FORK : PERF_RECORD_EXIT,
3843 .size = sizeof(task_event.event_id),
3849 .time = perf_clock(),
3853 perf_event_task_event(&task_event);
3856 void perf_event_fork(struct task_struct *task)
3858 perf_event_task(task, NULL, 1);
3865 struct perf_comm_event {
3866 struct task_struct *task;
3871 struct perf_event_header header;
3878 static void perf_event_comm_output(struct perf_event *event,
3879 struct perf_comm_event *comm_event)
3881 struct perf_output_handle handle;
3882 int size = comm_event->event_id.header.size;
3883 int ret = perf_output_begin(&handle, event, size, 0, 0);
3888 comm_event->event_id.pid = perf_event_pid(event, comm_event->task);
3889 comm_event->event_id.tid = perf_event_tid(event, comm_event->task);
3891 perf_output_put(&handle, comm_event->event_id);
3892 perf_output_copy(&handle, comm_event->comm,
3893 comm_event->comm_size);
3894 perf_output_end(&handle);
3897 static int perf_event_comm_match(struct perf_event *event)
3899 if (event->state < PERF_EVENT_STATE_INACTIVE)
3902 if (event->cpu != -1 && event->cpu != smp_processor_id())
3905 if (event->attr.comm)
3911 static void perf_event_comm_ctx(struct perf_event_context *ctx,
3912 struct perf_comm_event *comm_event)
3914 struct perf_event *event;
3916 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
3917 if (perf_event_comm_match(event))
3918 perf_event_comm_output(event, comm_event);
3922 static void perf_event_comm_event(struct perf_comm_event *comm_event)
3924 struct perf_cpu_context *cpuctx;
3925 struct perf_event_context *ctx;
3926 char comm[TASK_COMM_LEN];
3931 memset(comm, 0, sizeof(comm));
3932 strlcpy(comm, comm_event->task->comm, sizeof(comm));
3933 size = ALIGN(strlen(comm)+1, sizeof(u64));
3935 comm_event->comm = comm;
3936 comm_event->comm_size = size;
3938 comm_event->event_id.header.size = sizeof(comm_event->event_id) + size;
3940 rcu_read_lock_sched();
3941 list_for_each_entry_rcu(pmu, &pmus, entry) {
3942 cpuctx = this_cpu_ptr(pmu->pmu_cpu_context);
3943 perf_event_comm_ctx(&cpuctx->ctx, comm_event);
3945 ctxn = pmu->task_ctx_nr;
3949 ctx = rcu_dereference(current->perf_event_ctxp[ctxn]);
3951 perf_event_comm_ctx(ctx, comm_event);
3953 rcu_read_unlock_sched();
3956 void perf_event_comm(struct task_struct *task)
3958 struct perf_comm_event comm_event;
3959 struct perf_event_context *ctx;
3962 for_each_task_context_nr(ctxn) {
3963 ctx = task->perf_event_ctxp[ctxn];
3967 perf_event_enable_on_exec(ctx);
3970 if (!atomic_read(&nr_comm_events))
3973 comm_event = (struct perf_comm_event){
3979 .type = PERF_RECORD_COMM,
3988 perf_event_comm_event(&comm_event);
3995 struct perf_mmap_event {
3996 struct vm_area_struct *vma;
3998 const char *file_name;
4002 struct perf_event_header header;
4012 static void perf_event_mmap_output(struct perf_event *event,
4013 struct perf_mmap_event *mmap_event)
4015 struct perf_output_handle handle;
4016 int size = mmap_event->event_id.header.size;
4017 int ret = perf_output_begin(&handle, event, size, 0, 0);
4022 mmap_event->event_id.pid = perf_event_pid(event, current);
4023 mmap_event->event_id.tid = perf_event_tid(event, current);
4025 perf_output_put(&handle, mmap_event->event_id);
4026 perf_output_copy(&handle, mmap_event->file_name,
4027 mmap_event->file_size);
4028 perf_output_end(&handle);
4031 static int perf_event_mmap_match(struct perf_event *event,
4032 struct perf_mmap_event *mmap_event,
4035 if (event->state < PERF_EVENT_STATE_INACTIVE)
4038 if (event->cpu != -1 && event->cpu != smp_processor_id())
4041 if ((!executable && event->attr.mmap_data) ||
4042 (executable && event->attr.mmap))
4048 static void perf_event_mmap_ctx(struct perf_event_context *ctx,
4049 struct perf_mmap_event *mmap_event,
4052 struct perf_event *event;
4054 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
4055 if (perf_event_mmap_match(event, mmap_event, executable))
4056 perf_event_mmap_output(event, mmap_event);
4060 static void perf_event_mmap_event(struct perf_mmap_event *mmap_event)
4062 struct perf_cpu_context *cpuctx;
4063 struct perf_event_context *ctx;
4064 struct vm_area_struct *vma = mmap_event->vma;
4065 struct file *file = vma->vm_file;
4073 memset(tmp, 0, sizeof(tmp));
4077 * d_path works from the end of the buffer backwards, so we
4078 * need to add enough zero bytes after the string to handle
4079 * the 64bit alignment we do later.
4081 buf = kzalloc(PATH_MAX + sizeof(u64), GFP_KERNEL);
4083 name = strncpy(tmp, "//enomem", sizeof(tmp));
4086 name = d_path(&file->f_path, buf, PATH_MAX);
4088 name = strncpy(tmp, "//toolong", sizeof(tmp));
4092 if (arch_vma_name(mmap_event->vma)) {
4093 name = strncpy(tmp, arch_vma_name(mmap_event->vma),
4099 name = strncpy(tmp, "[vdso]", sizeof(tmp));
4101 } else if (vma->vm_start <= vma->vm_mm->start_brk &&
4102 vma->vm_end >= vma->vm_mm->brk) {
4103 name = strncpy(tmp, "[heap]", sizeof(tmp));
4105 } else if (vma->vm_start <= vma->vm_mm->start_stack &&
4106 vma->vm_end >= vma->vm_mm->start_stack) {
4107 name = strncpy(tmp, "[stack]", sizeof(tmp));
4111 name = strncpy(tmp, "//anon", sizeof(tmp));
4116 size = ALIGN(strlen(name)+1, sizeof(u64));
4118 mmap_event->file_name = name;
4119 mmap_event->file_size = size;
4121 mmap_event->event_id.header.size = sizeof(mmap_event->event_id) + size;
4123 rcu_read_lock_sched();
4124 list_for_each_entry_rcu(pmu, &pmus, entry) {
4125 cpuctx = this_cpu_ptr(pmu->pmu_cpu_context);
4126 perf_event_mmap_ctx(&cpuctx->ctx, mmap_event,
4127 vma->vm_flags & VM_EXEC);
4129 ctxn = pmu->task_ctx_nr;
4133 ctx = rcu_dereference(current->perf_event_ctxp[ctxn]);
4135 perf_event_mmap_ctx(ctx, mmap_event,
4136 vma->vm_flags & VM_EXEC);
4139 rcu_read_unlock_sched();
4144 void perf_event_mmap(struct vm_area_struct *vma)
4146 struct perf_mmap_event mmap_event;
4148 if (!atomic_read(&nr_mmap_events))
4151 mmap_event = (struct perf_mmap_event){
4157 .type = PERF_RECORD_MMAP,
4158 .misc = PERF_RECORD_MISC_USER,
4163 .start = vma->vm_start,
4164 .len = vma->vm_end - vma->vm_start,
4165 .pgoff = (u64)vma->vm_pgoff << PAGE_SHIFT,
4169 perf_event_mmap_event(&mmap_event);
4173 * IRQ throttle logging
4176 static void perf_log_throttle(struct perf_event *event, int enable)
4178 struct perf_output_handle handle;
4182 struct perf_event_header header;
4186 } throttle_event = {
4188 .type = PERF_RECORD_THROTTLE,
4190 .size = sizeof(throttle_event),
4192 .time = perf_clock(),
4193 .id = primary_event_id(event),
4194 .stream_id = event->id,
4198 throttle_event.header.type = PERF_RECORD_UNTHROTTLE;
4200 ret = perf_output_begin(&handle, event, sizeof(throttle_event), 1, 0);
4204 perf_output_put(&handle, throttle_event);
4205 perf_output_end(&handle);
4209 * Generic event overflow handling, sampling.
4212 static int __perf_event_overflow(struct perf_event *event, int nmi,
4213 int throttle, struct perf_sample_data *data,
4214 struct pt_regs *regs)
4216 int events = atomic_read(&event->event_limit);
4217 struct hw_perf_event *hwc = &event->hw;
4223 if (hwc->interrupts != MAX_INTERRUPTS) {
4225 if (HZ * hwc->interrupts >
4226 (u64)sysctl_perf_event_sample_rate) {
4227 hwc->interrupts = MAX_INTERRUPTS;
4228 perf_log_throttle(event, 0);
4233 * Keep re-disabling events even though on the previous
4234 * pass we disabled it - just in case we raced with a
4235 * sched-in and the event got enabled again:
4241 if (event->attr.freq) {
4242 u64 now = perf_clock();
4243 s64 delta = now - hwc->freq_time_stamp;
4245 hwc->freq_time_stamp = now;
4247 if (delta > 0 && delta < 2*TICK_NSEC)
4248 perf_adjust_period(event, delta, hwc->last_period);
4252 * XXX event_limit might not quite work as expected on inherited
4256 event->pending_kill = POLL_IN;
4257 if (events && atomic_dec_and_test(&event->event_limit)) {
4259 event->pending_kill = POLL_HUP;
4261 event->pending_disable = 1;
4262 perf_pending_queue(&event->pending,
4263 perf_pending_event);
4265 perf_event_disable(event);
4268 if (event->overflow_handler)
4269 event->overflow_handler(event, nmi, data, regs);
4271 perf_event_output(event, nmi, data, regs);
4276 int perf_event_overflow(struct perf_event *event, int nmi,
4277 struct perf_sample_data *data,
4278 struct pt_regs *regs)
4280 return __perf_event_overflow(event, nmi, 1, data, regs);
4284 * Generic software event infrastructure
4287 struct swevent_htable {
4288 struct swevent_hlist *swevent_hlist;
4289 struct mutex hlist_mutex;
4292 /* Recursion avoidance in each contexts */
4293 int recursion[PERF_NR_CONTEXTS];
4296 static DEFINE_PER_CPU(struct swevent_htable, swevent_htable);
4299 * We directly increment event->count and keep a second value in
4300 * event->hw.period_left to count intervals. This period event
4301 * is kept in the range [-sample_period, 0] so that we can use the
4305 static u64 perf_swevent_set_period(struct perf_event *event)
4307 struct hw_perf_event *hwc = &event->hw;
4308 u64 period = hwc->last_period;
4312 hwc->last_period = hwc->sample_period;
4315 old = val = local64_read(&hwc->period_left);
4319 nr = div64_u64(period + val, period);
4320 offset = nr * period;
4322 if (local64_cmpxchg(&hwc->period_left, old, val) != old)
4328 static void perf_swevent_overflow(struct perf_event *event, u64 overflow,
4329 int nmi, struct perf_sample_data *data,
4330 struct pt_regs *regs)
4332 struct hw_perf_event *hwc = &event->hw;
4335 data->period = event->hw.last_period;
4337 overflow = perf_swevent_set_period(event);
4339 if (hwc->interrupts == MAX_INTERRUPTS)
4342 for (; overflow; overflow--) {
4343 if (__perf_event_overflow(event, nmi, throttle,
4346 * We inhibit the overflow from happening when
4347 * hwc->interrupts == MAX_INTERRUPTS.
4355 static void perf_swevent_event(struct perf_event *event, u64 nr,
4356 int nmi, struct perf_sample_data *data,
4357 struct pt_regs *regs)
4359 struct hw_perf_event *hwc = &event->hw;
4361 local64_add(nr, &event->count);
4366 if (!hwc->sample_period)
4369 if (nr == 1 && hwc->sample_period == 1 && !event->attr.freq)
4370 return perf_swevent_overflow(event, 1, nmi, data, regs);
4372 if (local64_add_negative(nr, &hwc->period_left))
4375 perf_swevent_overflow(event, 0, nmi, data, regs);
4378 static int perf_exclude_event(struct perf_event *event,
4379 struct pt_regs *regs)
4381 if (event->hw.state & PERF_HES_STOPPED)
4385 if (event->attr.exclude_user && user_mode(regs))
4388 if (event->attr.exclude_kernel && !user_mode(regs))
4395 static int perf_swevent_match(struct perf_event *event,
4396 enum perf_type_id type,
4398 struct perf_sample_data *data,
4399 struct pt_regs *regs)
4401 if (event->attr.type != type)
4404 if (event->attr.config != event_id)
4407 if (perf_exclude_event(event, regs))
4413 static inline u64 swevent_hash(u64 type, u32 event_id)
4415 u64 val = event_id | (type << 32);
4417 return hash_64(val, SWEVENT_HLIST_BITS);
4420 static inline struct hlist_head *
4421 __find_swevent_head(struct swevent_hlist *hlist, u64 type, u32 event_id)
4423 u64 hash = swevent_hash(type, event_id);
4425 return &hlist->heads[hash];
4428 /* For the read side: events when they trigger */
4429 static inline struct hlist_head *
4430 find_swevent_head_rcu(struct swevent_htable *swhash, u64 type, u32 event_id)
4432 struct swevent_hlist *hlist;
4434 hlist = rcu_dereference(swhash->swevent_hlist);
4438 return __find_swevent_head(hlist, type, event_id);
4441 /* For the event head insertion and removal in the hlist */
4442 static inline struct hlist_head *
4443 find_swevent_head(struct swevent_htable *swhash, struct perf_event *event)
4445 struct swevent_hlist *hlist;
4446 u32 event_id = event->attr.config;
4447 u64 type = event->attr.type;
4450 * Event scheduling is always serialized against hlist allocation
4451 * and release. Which makes the protected version suitable here.
4452 * The context lock guarantees that.
4454 hlist = rcu_dereference_protected(swhash->swevent_hlist,
4455 lockdep_is_held(&event->ctx->lock));
4459 return __find_swevent_head(hlist, type, event_id);
4462 static void do_perf_sw_event(enum perf_type_id type, u32 event_id,
4464 struct perf_sample_data *data,
4465 struct pt_regs *regs)
4467 struct swevent_htable *swhash = &__get_cpu_var(swevent_htable);
4468 struct perf_event *event;
4469 struct hlist_node *node;
4470 struct hlist_head *head;
4473 head = find_swevent_head_rcu(swhash, type, event_id);
4477 hlist_for_each_entry_rcu(event, node, head, hlist_entry) {
4478 if (perf_swevent_match(event, type, event_id, data, regs))
4479 perf_swevent_event(event, nr, nmi, data, regs);
4485 int perf_swevent_get_recursion_context(void)
4487 struct swevent_htable *swhash = &__get_cpu_var(swevent_htable);
4489 return get_recursion_context(swhash->recursion);
4491 EXPORT_SYMBOL_GPL(perf_swevent_get_recursion_context);
4493 void inline perf_swevent_put_recursion_context(int rctx)
4495 struct swevent_htable *swhash = &__get_cpu_var(swevent_htable);
4497 put_recursion_context(swhash->recursion, rctx);
4500 void __perf_sw_event(u32 event_id, u64 nr, int nmi,
4501 struct pt_regs *regs, u64 addr)
4503 struct perf_sample_data data;
4506 preempt_disable_notrace();
4507 rctx = perf_swevent_get_recursion_context();
4511 perf_sample_data_init(&data, addr);
4513 do_perf_sw_event(PERF_TYPE_SOFTWARE, event_id, nr, nmi, &data, regs);
4515 perf_swevent_put_recursion_context(rctx);
4516 preempt_enable_notrace();
4519 static void perf_swevent_read(struct perf_event *event)
4523 static int perf_swevent_add(struct perf_event *event, int flags)
4525 struct swevent_htable *swhash = &__get_cpu_var(swevent_htable);
4526 struct hw_perf_event *hwc = &event->hw;
4527 struct hlist_head *head;
4529 if (hwc->sample_period) {
4530 hwc->last_period = hwc->sample_period;
4531 perf_swevent_set_period(event);
4534 hwc->state = !(flags & PERF_EF_START);
4536 head = find_swevent_head(swhash, event);
4537 if (WARN_ON_ONCE(!head))
4540 hlist_add_head_rcu(&event->hlist_entry, head);
4545 static void perf_swevent_del(struct perf_event *event, int flags)
4547 hlist_del_rcu(&event->hlist_entry);
4550 static void perf_swevent_start(struct perf_event *event, int flags)
4552 event->hw.state = 0;
4555 static void perf_swevent_stop(struct perf_event *event, int flags)
4557 event->hw.state = PERF_HES_STOPPED;
4560 /* Deref the hlist from the update side */
4561 static inline struct swevent_hlist *
4562 swevent_hlist_deref(struct swevent_htable *swhash)
4564 return rcu_dereference_protected(swhash->swevent_hlist,
4565 lockdep_is_held(&swhash->hlist_mutex));
4568 static void swevent_hlist_release_rcu(struct rcu_head *rcu_head)
4570 struct swevent_hlist *hlist;
4572 hlist = container_of(rcu_head, struct swevent_hlist, rcu_head);
4576 static void swevent_hlist_release(struct swevent_htable *swhash)
4578 struct swevent_hlist *hlist = swevent_hlist_deref(swhash);
4583 rcu_assign_pointer(swhash->swevent_hlist, NULL);
4584 call_rcu(&hlist->rcu_head, swevent_hlist_release_rcu);
4587 static void swevent_hlist_put_cpu(struct perf_event *event, int cpu)
4589 struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
4591 mutex_lock(&swhash->hlist_mutex);
4593 if (!--swhash->hlist_refcount)
4594 swevent_hlist_release(swhash);
4596 mutex_unlock(&swhash->hlist_mutex);
4599 static void swevent_hlist_put(struct perf_event *event)
4603 if (event->cpu != -1) {
4604 swevent_hlist_put_cpu(event, event->cpu);
4608 for_each_possible_cpu(cpu)
4609 swevent_hlist_put_cpu(event, cpu);
4612 static int swevent_hlist_get_cpu(struct perf_event *event, int cpu)
4614 struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
4617 mutex_lock(&swhash->hlist_mutex);
4619 if (!swevent_hlist_deref(swhash) && cpu_online(cpu)) {
4620 struct swevent_hlist *hlist;
4622 hlist = kzalloc(sizeof(*hlist), GFP_KERNEL);
4627 rcu_assign_pointer(swhash->swevent_hlist, hlist);
4629 swhash->hlist_refcount++;
4631 mutex_unlock(&swhash->hlist_mutex);
4636 static int swevent_hlist_get(struct perf_event *event)
4639 int cpu, failed_cpu;
4641 if (event->cpu != -1)
4642 return swevent_hlist_get_cpu(event, event->cpu);
4645 for_each_possible_cpu(cpu) {
4646 err = swevent_hlist_get_cpu(event, cpu);
4656 for_each_possible_cpu(cpu) {
4657 if (cpu == failed_cpu)
4659 swevent_hlist_put_cpu(event, cpu);
4666 atomic_t perf_swevent_enabled[PERF_COUNT_SW_MAX];
4668 static void sw_perf_event_destroy(struct perf_event *event)
4670 u64 event_id = event->attr.config;
4672 WARN_ON(event->parent);
4674 atomic_dec(&perf_swevent_enabled[event_id]);
4675 swevent_hlist_put(event);
4678 static int perf_swevent_init(struct perf_event *event)
4680 int event_id = event->attr.config;
4682 if (event->attr.type != PERF_TYPE_SOFTWARE)
4686 case PERF_COUNT_SW_CPU_CLOCK:
4687 case PERF_COUNT_SW_TASK_CLOCK:
4694 if (event_id > PERF_COUNT_SW_MAX)
4697 if (!event->parent) {
4700 err = swevent_hlist_get(event);
4704 atomic_inc(&perf_swevent_enabled[event_id]);
4705 event->destroy = sw_perf_event_destroy;
4711 static struct pmu perf_swevent = {
4712 .task_ctx_nr = perf_sw_context,
4714 .event_init = perf_swevent_init,
4715 .add = perf_swevent_add,
4716 .del = perf_swevent_del,
4717 .start = perf_swevent_start,
4718 .stop = perf_swevent_stop,
4719 .read = perf_swevent_read,
4722 #ifdef CONFIG_EVENT_TRACING
4724 static int perf_tp_filter_match(struct perf_event *event,
4725 struct perf_sample_data *data)
4727 void *record = data->raw->data;
4729 if (likely(!event->filter) || filter_match_preds(event->filter, record))
4734 static int perf_tp_event_match(struct perf_event *event,
4735 struct perf_sample_data *data,
4736 struct pt_regs *regs)
4739 * All tracepoints are from kernel-space.
4741 if (event->attr.exclude_kernel)
4744 if (!perf_tp_filter_match(event, data))
4750 void perf_tp_event(u64 addr, u64 count, void *record, int entry_size,
4751 struct pt_regs *regs, struct hlist_head *head, int rctx)
4753 struct perf_sample_data data;
4754 struct perf_event *event;
4755 struct hlist_node *node;
4757 struct perf_raw_record raw = {
4762 perf_sample_data_init(&data, addr);
4765 hlist_for_each_entry_rcu(event, node, head, hlist_entry) {
4766 if (perf_tp_event_match(event, &data, regs))
4767 perf_swevent_event(event, count, 1, &data, regs);
4770 perf_swevent_put_recursion_context(rctx);
4772 EXPORT_SYMBOL_GPL(perf_tp_event);
4774 static void tp_perf_event_destroy(struct perf_event *event)
4776 perf_trace_destroy(event);
4779 static int perf_tp_event_init(struct perf_event *event)
4783 if (event->attr.type != PERF_TYPE_TRACEPOINT)
4787 * Raw tracepoint data is a severe data leak, only allow root to
4790 if ((event->attr.sample_type & PERF_SAMPLE_RAW) &&
4791 perf_paranoid_tracepoint_raw() &&
4792 !capable(CAP_SYS_ADMIN))
4795 err = perf_trace_init(event);
4799 event->destroy = tp_perf_event_destroy;
4804 static struct pmu perf_tracepoint = {
4805 .task_ctx_nr = perf_sw_context,
4807 .event_init = perf_tp_event_init,
4808 .add = perf_trace_add,
4809 .del = perf_trace_del,
4810 .start = perf_swevent_start,
4811 .stop = perf_swevent_stop,
4812 .read = perf_swevent_read,
4815 static inline void perf_tp_register(void)
4817 perf_pmu_register(&perf_tracepoint);
4820 static int perf_event_set_filter(struct perf_event *event, void __user *arg)
4825 if (event->attr.type != PERF_TYPE_TRACEPOINT)
4828 filter_str = strndup_user(arg, PAGE_SIZE);
4829 if (IS_ERR(filter_str))
4830 return PTR_ERR(filter_str);
4832 ret = ftrace_profile_set_filter(event, event->attr.config, filter_str);
4838 static void perf_event_free_filter(struct perf_event *event)
4840 ftrace_profile_free_filter(event);
4845 static inline void perf_tp_register(void)
4849 static int perf_event_set_filter(struct perf_event *event, void __user *arg)
4854 static void perf_event_free_filter(struct perf_event *event)
4858 #endif /* CONFIG_EVENT_TRACING */
4860 #ifdef CONFIG_HAVE_HW_BREAKPOINT
4861 void perf_bp_event(struct perf_event *bp, void *data)
4863 struct perf_sample_data sample;
4864 struct pt_regs *regs = data;
4866 perf_sample_data_init(&sample, bp->attr.bp_addr);
4868 if (!bp->hw.state && !perf_exclude_event(bp, regs))
4869 perf_swevent_event(bp, 1, 1, &sample, regs);
4874 * hrtimer based swevent callback
4877 static enum hrtimer_restart perf_swevent_hrtimer(struct hrtimer *hrtimer)
4879 enum hrtimer_restart ret = HRTIMER_RESTART;
4880 struct perf_sample_data data;
4881 struct pt_regs *regs;
4882 struct perf_event *event;
4885 event = container_of(hrtimer, struct perf_event, hw.hrtimer);
4886 event->pmu->read(event);
4888 perf_sample_data_init(&data, 0);
4889 data.period = event->hw.last_period;
4890 regs = get_irq_regs();
4892 if (regs && !perf_exclude_event(event, regs)) {
4893 if (!(event->attr.exclude_idle && current->pid == 0))
4894 if (perf_event_overflow(event, 0, &data, regs))
4895 ret = HRTIMER_NORESTART;
4898 period = max_t(u64, 10000, event->hw.sample_period);
4899 hrtimer_forward_now(hrtimer, ns_to_ktime(period));
4904 static void perf_swevent_start_hrtimer(struct perf_event *event)
4906 struct hw_perf_event *hwc = &event->hw;
4908 hrtimer_init(&hwc->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
4909 hwc->hrtimer.function = perf_swevent_hrtimer;
4910 if (hwc->sample_period) {
4911 s64 period = local64_read(&hwc->period_left);
4917 local64_set(&hwc->period_left, 0);
4919 period = max_t(u64, 10000, hwc->sample_period);
4921 __hrtimer_start_range_ns(&hwc->hrtimer,
4922 ns_to_ktime(period), 0,
4923 HRTIMER_MODE_REL_PINNED, 0);
4927 static void perf_swevent_cancel_hrtimer(struct perf_event *event)
4929 struct hw_perf_event *hwc = &event->hw;
4931 if (hwc->sample_period) {
4932 ktime_t remaining = hrtimer_get_remaining(&hwc->hrtimer);
4933 local64_set(&hwc->period_left, ktime_to_ns(remaining));
4935 hrtimer_cancel(&hwc->hrtimer);
4940 * Software event: cpu wall time clock
4943 static void cpu_clock_event_update(struct perf_event *event)
4948 now = local_clock();
4949 prev = local64_xchg(&event->hw.prev_count, now);
4950 local64_add(now - prev, &event->count);
4953 static void cpu_clock_event_start(struct perf_event *event, int flags)
4955 local64_set(&event->hw.prev_count, local_clock());
4956 perf_swevent_start_hrtimer(event);
4959 static void cpu_clock_event_stop(struct perf_event *event, int flags)
4961 perf_swevent_cancel_hrtimer(event);
4962 cpu_clock_event_update(event);
4965 static int cpu_clock_event_add(struct perf_event *event, int flags)
4967 if (flags & PERF_EF_START)
4968 cpu_clock_event_start(event, flags);
4973 static void cpu_clock_event_del(struct perf_event *event, int flags)
4975 cpu_clock_event_stop(event, flags);
4978 static void cpu_clock_event_read(struct perf_event *event)
4980 cpu_clock_event_update(event);
4983 static int cpu_clock_event_init(struct perf_event *event)
4985 if (event->attr.type != PERF_TYPE_SOFTWARE)
4988 if (event->attr.config != PERF_COUNT_SW_CPU_CLOCK)
4994 static struct pmu perf_cpu_clock = {
4995 .task_ctx_nr = perf_sw_context,
4997 .event_init = cpu_clock_event_init,
4998 .add = cpu_clock_event_add,
4999 .del = cpu_clock_event_del,
5000 .start = cpu_clock_event_start,
5001 .stop = cpu_clock_event_stop,
5002 .read = cpu_clock_event_read,
5006 * Software event: task time clock
5009 static void task_clock_event_update(struct perf_event *event, u64 now)
5014 prev = local64_xchg(&event->hw.prev_count, now);
5016 local64_add(delta, &event->count);
5019 static void task_clock_event_start(struct perf_event *event, int flags)
5021 local64_set(&event->hw.prev_count, event->ctx->time);
5022 perf_swevent_start_hrtimer(event);
5025 static void task_clock_event_stop(struct perf_event *event, int flags)
5027 perf_swevent_cancel_hrtimer(event);
5028 task_clock_event_update(event, event->ctx->time);
5031 static int task_clock_event_add(struct perf_event *event, int flags)
5033 if (flags & PERF_EF_START)
5034 task_clock_event_start(event, flags);
5039 static void task_clock_event_del(struct perf_event *event, int flags)
5041 task_clock_event_stop(event, PERF_EF_UPDATE);
5044 static void task_clock_event_read(struct perf_event *event)
5049 update_context_time(event->ctx);
5050 time = event->ctx->time;
5052 u64 now = perf_clock();
5053 u64 delta = now - event->ctx->timestamp;
5054 time = event->ctx->time + delta;
5057 task_clock_event_update(event, time);
5060 static int task_clock_event_init(struct perf_event *event)
5062 if (event->attr.type != PERF_TYPE_SOFTWARE)
5065 if (event->attr.config != PERF_COUNT_SW_TASK_CLOCK)
5071 static struct pmu perf_task_clock = {
5072 .task_ctx_nr = perf_sw_context,
5074 .event_init = task_clock_event_init,
5075 .add = task_clock_event_add,
5076 .del = task_clock_event_del,
5077 .start = task_clock_event_start,
5078 .stop = task_clock_event_stop,
5079 .read = task_clock_event_read,
5082 static void perf_pmu_nop_void(struct pmu *pmu)
5086 static int perf_pmu_nop_int(struct pmu *pmu)
5091 static void perf_pmu_start_txn(struct pmu *pmu)
5093 perf_pmu_disable(pmu);
5096 static int perf_pmu_commit_txn(struct pmu *pmu)
5098 perf_pmu_enable(pmu);
5102 static void perf_pmu_cancel_txn(struct pmu *pmu)
5104 perf_pmu_enable(pmu);
5108 * Ensures all contexts with the same task_ctx_nr have the same
5109 * pmu_cpu_context too.
5111 static void *find_pmu_context(int ctxn)
5118 list_for_each_entry(pmu, &pmus, entry) {
5119 if (pmu->task_ctx_nr == ctxn)
5120 return pmu->pmu_cpu_context;
5126 static void free_pmu_context(void * __percpu cpu_context)
5130 mutex_lock(&pmus_lock);
5132 * Like a real lame refcount.
5134 list_for_each_entry(pmu, &pmus, entry) {
5135 if (pmu->pmu_cpu_context == cpu_context)
5139 free_percpu(cpu_context);
5141 mutex_unlock(&pmus_lock);
5144 int perf_pmu_register(struct pmu *pmu)
5148 mutex_lock(&pmus_lock);
5150 pmu->pmu_disable_count = alloc_percpu(int);
5151 if (!pmu->pmu_disable_count)
5154 pmu->pmu_cpu_context = find_pmu_context(pmu->task_ctx_nr);
5155 if (pmu->pmu_cpu_context)
5156 goto got_cpu_context;
5158 pmu->pmu_cpu_context = alloc_percpu(struct perf_cpu_context);
5159 if (!pmu->pmu_cpu_context)
5162 for_each_possible_cpu(cpu) {
5163 struct perf_cpu_context *cpuctx;
5165 cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu);
5166 __perf_event_init_context(&cpuctx->ctx);
5167 cpuctx->ctx.pmu = pmu;
5168 cpuctx->timer_interval = TICK_NSEC;
5169 hrtimer_init(&cpuctx->timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
5170 cpuctx->timer.function = perf_event_context_tick;
5174 if (!pmu->start_txn) {
5175 if (pmu->pmu_enable) {
5177 * If we have pmu_enable/pmu_disable calls, install
5178 * transaction stubs that use that to try and batch
5179 * hardware accesses.
5181 pmu->start_txn = perf_pmu_start_txn;
5182 pmu->commit_txn = perf_pmu_commit_txn;
5183 pmu->cancel_txn = perf_pmu_cancel_txn;
5185 pmu->start_txn = perf_pmu_nop_void;
5186 pmu->commit_txn = perf_pmu_nop_int;
5187 pmu->cancel_txn = perf_pmu_nop_void;
5191 if (!pmu->pmu_enable) {
5192 pmu->pmu_enable = perf_pmu_nop_void;
5193 pmu->pmu_disable = perf_pmu_nop_void;
5196 list_add_rcu(&pmu->entry, &pmus);
5199 mutex_unlock(&pmus_lock);
5204 free_percpu(pmu->pmu_disable_count);
5208 void perf_pmu_unregister(struct pmu *pmu)
5210 mutex_lock(&pmus_lock);
5211 list_del_rcu(&pmu->entry);
5212 mutex_unlock(&pmus_lock);
5215 * We use the pmu list either under SRCU or preempt_disable,
5216 * synchronize_srcu() implies synchronize_sched() so we're good.
5218 synchronize_srcu(&pmus_srcu);
5220 free_percpu(pmu->pmu_disable_count);
5221 free_pmu_context(pmu->pmu_cpu_context);
5224 struct pmu *perf_init_event(struct perf_event *event)
5226 struct pmu *pmu = NULL;
5229 idx = srcu_read_lock(&pmus_srcu);
5230 list_for_each_entry_rcu(pmu, &pmus, entry) {
5231 int ret = pmu->event_init(event);
5234 if (ret != -ENOENT) {
5239 srcu_read_unlock(&pmus_srcu, idx);
5245 * Allocate and initialize a event structure
5247 static struct perf_event *
5248 perf_event_alloc(struct perf_event_attr *attr, int cpu,
5249 struct perf_event *group_leader,
5250 struct perf_event *parent_event,
5251 perf_overflow_handler_t overflow_handler)
5254 struct perf_event *event;
5255 struct hw_perf_event *hwc;
5258 event = kzalloc(sizeof(*event), GFP_KERNEL);
5260 return ERR_PTR(-ENOMEM);
5263 * Single events are their own group leaders, with an
5264 * empty sibling list:
5267 group_leader = event;
5269 mutex_init(&event->child_mutex);
5270 INIT_LIST_HEAD(&event->child_list);
5272 INIT_LIST_HEAD(&event->group_entry);
5273 INIT_LIST_HEAD(&event->event_entry);
5274 INIT_LIST_HEAD(&event->sibling_list);
5275 init_waitqueue_head(&event->waitq);
5277 mutex_init(&event->mmap_mutex);
5280 event->attr = *attr;
5281 event->group_leader = group_leader;
5285 event->parent = parent_event;
5287 event->ns = get_pid_ns(current->nsproxy->pid_ns);
5288 event->id = atomic64_inc_return(&perf_event_id);
5290 event->state = PERF_EVENT_STATE_INACTIVE;
5292 if (!overflow_handler && parent_event)
5293 overflow_handler = parent_event->overflow_handler;
5295 event->overflow_handler = overflow_handler;
5298 event->state = PERF_EVENT_STATE_OFF;
5303 hwc->sample_period = attr->sample_period;
5304 if (attr->freq && attr->sample_freq)
5305 hwc->sample_period = 1;
5306 hwc->last_period = hwc->sample_period;
5308 local64_set(&hwc->period_left, hwc->sample_period);
5311 * we currently do not support PERF_FORMAT_GROUP on inherited events
5313 if (attr->inherit && (attr->read_format & PERF_FORMAT_GROUP))
5316 pmu = perf_init_event(event);
5322 else if (IS_ERR(pmu))
5327 put_pid_ns(event->ns);
5329 return ERR_PTR(err);
5334 if (!event->parent) {
5335 atomic_inc(&nr_events);
5336 if (event->attr.mmap || event->attr.mmap_data)
5337 atomic_inc(&nr_mmap_events);
5338 if (event->attr.comm)
5339 atomic_inc(&nr_comm_events);
5340 if (event->attr.task)
5341 atomic_inc(&nr_task_events);
5342 if (event->attr.sample_type & PERF_SAMPLE_CALLCHAIN) {
5343 err = get_callchain_buffers();
5346 return ERR_PTR(err);
5354 static int perf_copy_attr(struct perf_event_attr __user *uattr,
5355 struct perf_event_attr *attr)
5360 if (!access_ok(VERIFY_WRITE, uattr, PERF_ATTR_SIZE_VER0))
5364 * zero the full structure, so that a short copy will be nice.
5366 memset(attr, 0, sizeof(*attr));
5368 ret = get_user(size, &uattr->size);
5372 if (size > PAGE_SIZE) /* silly large */
5375 if (!size) /* abi compat */
5376 size = PERF_ATTR_SIZE_VER0;
5378 if (size < PERF_ATTR_SIZE_VER0)
5382 * If we're handed a bigger struct than we know of,
5383 * ensure all the unknown bits are 0 - i.e. new
5384 * user-space does not rely on any kernel feature
5385 * extensions we dont know about yet.
5387 if (size > sizeof(*attr)) {
5388 unsigned char __user *addr;
5389 unsigned char __user *end;
5392 addr = (void __user *)uattr + sizeof(*attr);
5393 end = (void __user *)uattr + size;
5395 for (; addr < end; addr++) {
5396 ret = get_user(val, addr);
5402 size = sizeof(*attr);
5405 ret = copy_from_user(attr, uattr, size);
5410 * If the type exists, the corresponding creation will verify
5413 if (attr->type >= PERF_TYPE_MAX)
5416 if (attr->__reserved_1)
5419 if (attr->sample_type & ~(PERF_SAMPLE_MAX-1))
5422 if (attr->read_format & ~(PERF_FORMAT_MAX-1))
5429 put_user(sizeof(*attr), &uattr->size);
5435 perf_event_set_output(struct perf_event *event, struct perf_event *output_event)
5437 struct perf_buffer *buffer = NULL, *old_buffer = NULL;
5443 /* don't allow circular references */
5444 if (event == output_event)
5448 * Don't allow cross-cpu buffers
5450 if (output_event->cpu != event->cpu)
5454 * If its not a per-cpu buffer, it must be the same task.
5456 if (output_event->cpu == -1 && output_event->ctx != event->ctx)
5460 mutex_lock(&event->mmap_mutex);
5461 /* Can't redirect output if we've got an active mmap() */
5462 if (atomic_read(&event->mmap_count))
5466 /* get the buffer we want to redirect to */
5467 buffer = perf_buffer_get(output_event);
5472 old_buffer = event->buffer;
5473 rcu_assign_pointer(event->buffer, buffer);
5476 mutex_unlock(&event->mmap_mutex);
5479 perf_buffer_put(old_buffer);
5485 * sys_perf_event_open - open a performance event, associate it to a task/cpu
5487 * @attr_uptr: event_id type attributes for monitoring/sampling
5490 * @group_fd: group leader event fd
5492 SYSCALL_DEFINE5(perf_event_open,
5493 struct perf_event_attr __user *, attr_uptr,
5494 pid_t, pid, int, cpu, int, group_fd, unsigned long, flags)
5496 struct perf_event *event, *group_leader = NULL, *output_event = NULL;
5497 struct perf_event_attr attr;
5498 struct perf_event_context *ctx;
5499 struct file *event_file = NULL;
5500 struct file *group_file = NULL;
5503 int fput_needed = 0;
5506 /* for future expandability... */
5507 if (flags & ~(PERF_FLAG_FD_NO_GROUP | PERF_FLAG_FD_OUTPUT))
5510 err = perf_copy_attr(attr_uptr, &attr);
5514 if (!attr.exclude_kernel) {
5515 if (perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN))
5520 if (attr.sample_freq > sysctl_perf_event_sample_rate)
5524 event_fd = get_unused_fd_flags(O_RDWR);
5528 event = perf_event_alloc(&attr, cpu, group_leader, NULL, NULL);
5529 if (IS_ERR(event)) {
5530 err = PTR_ERR(event);
5534 if (group_fd != -1) {
5535 group_leader = perf_fget_light(group_fd, &fput_needed);
5536 if (IS_ERR(group_leader)) {
5537 err = PTR_ERR(group_leader);
5540 group_file = group_leader->filp;
5541 if (flags & PERF_FLAG_FD_OUTPUT)
5542 output_event = group_leader;
5543 if (flags & PERF_FLAG_FD_NO_GROUP)
5544 group_leader = NULL;
5548 * Special case software events and allow them to be part of
5549 * any hardware group.
5552 if ((pmu->task_ctx_nr == perf_sw_context) && group_leader)
5553 pmu = group_leader->pmu;
5556 * Get the target context (task or percpu):
5558 ctx = find_get_context(pmu, pid, cpu);
5565 * Look up the group leader (we will attach this event to it):
5571 * Do not allow a recursive hierarchy (this new sibling
5572 * becoming part of another group-sibling):
5574 if (group_leader->group_leader != group_leader)
5577 * Do not allow to attach to a group in a different
5578 * task or CPU context:
5580 if (group_leader->ctx != ctx)
5583 * Only a group leader can be exclusive or pinned
5585 if (attr.exclusive || attr.pinned)
5590 err = perf_event_set_output(event, output_event);
5595 event_file = anon_inode_getfile("[perf_event]", &perf_fops, event, O_RDWR);
5596 if (IS_ERR(event_file)) {
5597 err = PTR_ERR(event_file);
5601 event->filp = event_file;
5602 WARN_ON_ONCE(ctx->parent_ctx);
5603 mutex_lock(&ctx->mutex);
5604 perf_install_in_context(ctx, event, cpu);
5606 mutex_unlock(&ctx->mutex);
5608 event->owner = current;
5609 get_task_struct(current);
5610 mutex_lock(¤t->perf_event_mutex);
5611 list_add_tail(&event->owner_entry, ¤t->perf_event_list);
5612 mutex_unlock(¤t->perf_event_mutex);
5615 * Drop the reference on the group_event after placing the
5616 * new event on the sibling_list. This ensures destruction
5617 * of the group leader will find the pointer to itself in
5618 * perf_group_detach().
5620 fput_light(group_file, fput_needed);
5621 fd_install(event_fd, event_file);
5627 fput_light(group_file, fput_needed);
5631 put_unused_fd(event_fd);
5636 * perf_event_create_kernel_counter
5638 * @attr: attributes of the counter to create
5639 * @cpu: cpu in which the counter is bound
5640 * @pid: task to profile
5643 perf_event_create_kernel_counter(struct perf_event_attr *attr, int cpu,
5645 perf_overflow_handler_t overflow_handler)
5647 struct perf_event_context *ctx;
5648 struct perf_event *event;
5652 * Get the target context (task or percpu):
5655 event = perf_event_alloc(attr, cpu, NULL, NULL, overflow_handler);
5656 if (IS_ERR(event)) {
5657 err = PTR_ERR(event);
5661 ctx = find_get_context(event->pmu, pid, cpu);
5668 WARN_ON_ONCE(ctx->parent_ctx);
5669 mutex_lock(&ctx->mutex);
5670 perf_install_in_context(ctx, event, cpu);
5672 mutex_unlock(&ctx->mutex);
5674 event->owner = current;
5675 get_task_struct(current);
5676 mutex_lock(¤t->perf_event_mutex);
5677 list_add_tail(&event->owner_entry, ¤t->perf_event_list);
5678 mutex_unlock(¤t->perf_event_mutex);
5685 return ERR_PTR(err);
5687 EXPORT_SYMBOL_GPL(perf_event_create_kernel_counter);
5689 static void sync_child_event(struct perf_event *child_event,
5690 struct task_struct *child)
5692 struct perf_event *parent_event = child_event->parent;
5695 if (child_event->attr.inherit_stat)
5696 perf_event_read_event(child_event, child);
5698 child_val = perf_event_count(child_event);
5701 * Add back the child's count to the parent's count:
5703 atomic64_add(child_val, &parent_event->child_count);
5704 atomic64_add(child_event->total_time_enabled,
5705 &parent_event->child_total_time_enabled);
5706 atomic64_add(child_event->total_time_running,
5707 &parent_event->child_total_time_running);
5710 * Remove this event from the parent's list
5712 WARN_ON_ONCE(parent_event->ctx->parent_ctx);
5713 mutex_lock(&parent_event->child_mutex);
5714 list_del_init(&child_event->child_list);
5715 mutex_unlock(&parent_event->child_mutex);
5718 * Release the parent event, if this was the last
5721 fput(parent_event->filp);
5725 __perf_event_exit_task(struct perf_event *child_event,
5726 struct perf_event_context *child_ctx,
5727 struct task_struct *child)
5729 struct perf_event *parent_event;
5731 perf_event_remove_from_context(child_event);
5733 parent_event = child_event->parent;
5735 * It can happen that parent exits first, and has events
5736 * that are still around due to the child reference. These
5737 * events need to be zapped - but otherwise linger.
5740 sync_child_event(child_event, child);
5741 free_event(child_event);
5745 static void perf_event_exit_task_context(struct task_struct *child, int ctxn)
5747 struct perf_event *child_event, *tmp;
5748 struct perf_event_context *child_ctx;
5749 unsigned long flags;
5751 if (likely(!child->perf_event_ctxp[ctxn])) {
5752 perf_event_task(child, NULL, 0);
5756 local_irq_save(flags);
5758 * We can't reschedule here because interrupts are disabled,
5759 * and either child is current or it is a task that can't be
5760 * scheduled, so we are now safe from rescheduling changing
5763 child_ctx = child->perf_event_ctxp[ctxn];
5764 __perf_event_task_sched_out(child_ctx);
5767 * Take the context lock here so that if find_get_context is
5768 * reading child->perf_event_ctxp, we wait until it has
5769 * incremented the context's refcount before we do put_ctx below.
5771 raw_spin_lock(&child_ctx->lock);
5772 child->perf_event_ctxp[ctxn] = NULL;
5774 * If this context is a clone; unclone it so it can't get
5775 * swapped to another process while we're removing all
5776 * the events from it.
5778 unclone_ctx(child_ctx);
5779 update_context_time(child_ctx);
5780 raw_spin_unlock_irqrestore(&child_ctx->lock, flags);
5783 * Report the task dead after unscheduling the events so that we
5784 * won't get any samples after PERF_RECORD_EXIT. We can however still
5785 * get a few PERF_RECORD_READ events.
5787 perf_event_task(child, child_ctx, 0);
5790 * We can recurse on the same lock type through:
5792 * __perf_event_exit_task()
5793 * sync_child_event()
5794 * fput(parent_event->filp)
5796 * mutex_lock(&ctx->mutex)
5798 * But since its the parent context it won't be the same instance.
5800 mutex_lock(&child_ctx->mutex);
5803 list_for_each_entry_safe(child_event, tmp, &child_ctx->pinned_groups,
5805 __perf_event_exit_task(child_event, child_ctx, child);
5807 list_for_each_entry_safe(child_event, tmp, &child_ctx->flexible_groups,
5809 __perf_event_exit_task(child_event, child_ctx, child);
5812 * If the last event was a group event, it will have appended all
5813 * its siblings to the list, but we obtained 'tmp' before that which
5814 * will still point to the list head terminating the iteration.
5816 if (!list_empty(&child_ctx->pinned_groups) ||
5817 !list_empty(&child_ctx->flexible_groups))
5820 mutex_unlock(&child_ctx->mutex);
5826 * When a child task exits, feed back event values to parent events.
5828 void perf_event_exit_task(struct task_struct *child)
5832 for_each_task_context_nr(ctxn)
5833 perf_event_exit_task_context(child, ctxn);
5836 static void perf_free_event(struct perf_event *event,
5837 struct perf_event_context *ctx)
5839 struct perf_event *parent = event->parent;
5841 if (WARN_ON_ONCE(!parent))
5844 mutex_lock(&parent->child_mutex);
5845 list_del_init(&event->child_list);
5846 mutex_unlock(&parent->child_mutex);
5850 perf_group_detach(event);
5851 list_del_event(event, ctx);
5856 * free an unexposed, unused context as created by inheritance by
5857 * perf_event_init_task below, used by fork() in case of fail.
5859 void perf_event_free_task(struct task_struct *task)
5861 struct perf_event_context *ctx;
5862 struct perf_event *event, *tmp;
5865 for_each_task_context_nr(ctxn) {
5866 ctx = task->perf_event_ctxp[ctxn];
5870 mutex_lock(&ctx->mutex);
5872 list_for_each_entry_safe(event, tmp, &ctx->pinned_groups,
5874 perf_free_event(event, ctx);
5876 list_for_each_entry_safe(event, tmp, &ctx->flexible_groups,
5878 perf_free_event(event, ctx);
5880 if (!list_empty(&ctx->pinned_groups) ||
5881 !list_empty(&ctx->flexible_groups))
5884 mutex_unlock(&ctx->mutex);
5891 * inherit a event from parent task to child task:
5893 static struct perf_event *
5894 inherit_event(struct perf_event *parent_event,
5895 struct task_struct *parent,
5896 struct perf_event_context *parent_ctx,
5897 struct task_struct *child,
5898 struct perf_event *group_leader,
5899 struct perf_event_context *child_ctx)
5901 struct perf_event *child_event;
5904 * Instead of creating recursive hierarchies of events,
5905 * we link inherited events back to the original parent,
5906 * which has a filp for sure, which we use as the reference
5909 if (parent_event->parent)
5910 parent_event = parent_event->parent;
5912 child_event = perf_event_alloc(&parent_event->attr,
5914 group_leader, parent_event,
5916 if (IS_ERR(child_event))
5921 * Make the child state follow the state of the parent event,
5922 * not its attr.disabled bit. We hold the parent's mutex,
5923 * so we won't race with perf_event_{en, dis}able_family.
5925 if (parent_event->state >= PERF_EVENT_STATE_INACTIVE)
5926 child_event->state = PERF_EVENT_STATE_INACTIVE;
5928 child_event->state = PERF_EVENT_STATE_OFF;
5930 if (parent_event->attr.freq) {
5931 u64 sample_period = parent_event->hw.sample_period;
5932 struct hw_perf_event *hwc = &child_event->hw;
5934 hwc->sample_period = sample_period;
5935 hwc->last_period = sample_period;
5937 local64_set(&hwc->period_left, sample_period);
5940 child_event->ctx = child_ctx;
5941 child_event->overflow_handler = parent_event->overflow_handler;
5944 * Link it up in the child's context:
5946 add_event_to_ctx(child_event, child_ctx);
5949 * Get a reference to the parent filp - we will fput it
5950 * when the child event exits. This is safe to do because
5951 * we are in the parent and we know that the filp still
5952 * exists and has a nonzero count:
5954 atomic_long_inc(&parent_event->filp->f_count);
5957 * Link this into the parent event's child list
5959 WARN_ON_ONCE(parent_event->ctx->parent_ctx);
5960 mutex_lock(&parent_event->child_mutex);
5961 list_add_tail(&child_event->child_list, &parent_event->child_list);
5962 mutex_unlock(&parent_event->child_mutex);
5967 static int inherit_group(struct perf_event *parent_event,
5968 struct task_struct *parent,
5969 struct perf_event_context *parent_ctx,
5970 struct task_struct *child,
5971 struct perf_event_context *child_ctx)
5973 struct perf_event *leader;
5974 struct perf_event *sub;
5975 struct perf_event *child_ctr;
5977 leader = inherit_event(parent_event, parent, parent_ctx,
5978 child, NULL, child_ctx);
5980 return PTR_ERR(leader);
5981 list_for_each_entry(sub, &parent_event->sibling_list, group_entry) {
5982 child_ctr = inherit_event(sub, parent, parent_ctx,
5983 child, leader, child_ctx);
5984 if (IS_ERR(child_ctr))
5985 return PTR_ERR(child_ctr);
5991 inherit_task_group(struct perf_event *event, struct task_struct *parent,
5992 struct perf_event_context *parent_ctx,
5993 struct task_struct *child, int ctxn,
5997 struct perf_event_context *child_ctx;
5999 if (!event->attr.inherit) {
6004 child_ctx = child->perf_event_ctxp[ctxn];
6007 * This is executed from the parent task context, so
6008 * inherit events that have been marked for cloning.
6009 * First allocate and initialize a context for the
6013 child_ctx = alloc_perf_context(event->pmu, child);
6017 child->perf_event_ctxp[ctxn] = child_ctx;
6020 ret = inherit_group(event, parent, parent_ctx,
6030 * Initialize the perf_event context in task_struct
6032 int perf_event_init_context(struct task_struct *child, int ctxn)
6034 struct perf_event_context *child_ctx, *parent_ctx;
6035 struct perf_event_context *cloned_ctx;
6036 struct perf_event *event;
6037 struct task_struct *parent = current;
6038 int inherited_all = 1;
6041 child->perf_event_ctxp[ctxn] = NULL;
6043 mutex_init(&child->perf_event_mutex);
6044 INIT_LIST_HEAD(&child->perf_event_list);
6046 if (likely(!parent->perf_event_ctxp[ctxn]))
6050 * If the parent's context is a clone, pin it so it won't get
6053 parent_ctx = perf_pin_task_context(parent, ctxn);
6056 * No need to check if parent_ctx != NULL here; since we saw
6057 * it non-NULL earlier, the only reason for it to become NULL
6058 * is if we exit, and since we're currently in the middle of
6059 * a fork we can't be exiting at the same time.
6063 * Lock the parent list. No need to lock the child - not PID
6064 * hashed yet and not running, so nobody can access it.
6066 mutex_lock(&parent_ctx->mutex);
6069 * We dont have to disable NMIs - we are only looking at
6070 * the list, not manipulating it:
6072 list_for_each_entry(event, &parent_ctx->pinned_groups, group_entry) {
6073 ret = inherit_task_group(event, parent, parent_ctx,
6074 child, ctxn, &inherited_all);
6079 list_for_each_entry(event, &parent_ctx->flexible_groups, group_entry) {
6080 ret = inherit_task_group(event, parent, parent_ctx,
6081 child, ctxn, &inherited_all);
6086 child_ctx = child->perf_event_ctxp[ctxn];
6088 if (child_ctx && inherited_all) {
6090 * Mark the child context as a clone of the parent
6091 * context, or of whatever the parent is a clone of.
6092 * Note that if the parent is a clone, it could get
6093 * uncloned at any point, but that doesn't matter
6094 * because the list of events and the generation
6095 * count can't have changed since we took the mutex.
6097 cloned_ctx = rcu_dereference(parent_ctx->parent_ctx);
6099 child_ctx->parent_ctx = cloned_ctx;
6100 child_ctx->parent_gen = parent_ctx->parent_gen;
6102 child_ctx->parent_ctx = parent_ctx;
6103 child_ctx->parent_gen = parent_ctx->generation;
6105 get_ctx(child_ctx->parent_ctx);
6108 mutex_unlock(&parent_ctx->mutex);
6110 perf_unpin_context(parent_ctx);
6116 * Initialize the perf_event context in task_struct
6118 int perf_event_init_task(struct task_struct *child)
6122 for_each_task_context_nr(ctxn) {
6123 ret = perf_event_init_context(child, ctxn);
6131 static void __init perf_event_init_all_cpus(void)
6133 struct swevent_htable *swhash;
6136 for_each_possible_cpu(cpu) {
6137 swhash = &per_cpu(swevent_htable, cpu);
6138 mutex_init(&swhash->hlist_mutex);
6142 static void __cpuinit perf_event_init_cpu(int cpu)
6144 struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
6146 mutex_lock(&swhash->hlist_mutex);
6147 if (swhash->hlist_refcount > 0) {
6148 struct swevent_hlist *hlist;
6150 hlist = kzalloc_node(sizeof(*hlist), GFP_KERNEL, cpu_to_node(cpu));
6152 rcu_assign_pointer(swhash->swevent_hlist, hlist);
6154 mutex_unlock(&swhash->hlist_mutex);
6157 #ifdef CONFIG_HOTPLUG_CPU
6158 static void __perf_event_exit_context(void *__info)
6160 struct perf_event_context *ctx = __info;
6161 struct perf_event *event, *tmp;
6163 perf_pmu_rotate_stop(ctx->pmu);
6165 list_for_each_entry_safe(event, tmp, &ctx->pinned_groups, group_entry)
6166 __perf_event_remove_from_context(event);
6167 list_for_each_entry_safe(event, tmp, &ctx->flexible_groups, group_entry)
6168 __perf_event_remove_from_context(event);
6171 static void perf_event_exit_cpu_context(int cpu)
6173 struct perf_event_context *ctx;
6177 idx = srcu_read_lock(&pmus_srcu);
6178 list_for_each_entry_rcu(pmu, &pmus, entry) {
6179 ctx = &this_cpu_ptr(pmu->pmu_cpu_context)->ctx;
6181 mutex_lock(&ctx->mutex);
6182 smp_call_function_single(cpu, __perf_event_exit_context, ctx, 1);
6183 mutex_unlock(&ctx->mutex);
6185 srcu_read_unlock(&pmus_srcu, idx);
6189 static void perf_event_exit_cpu(int cpu)
6191 struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
6193 mutex_lock(&swhash->hlist_mutex);
6194 swevent_hlist_release(swhash);
6195 mutex_unlock(&swhash->hlist_mutex);
6197 perf_event_exit_cpu_context(cpu);
6200 static inline void perf_event_exit_cpu(int cpu) { }
6203 static int __cpuinit
6204 perf_cpu_notify(struct notifier_block *self, unsigned long action, void *hcpu)
6206 unsigned int cpu = (long)hcpu;
6208 switch (action & ~CPU_TASKS_FROZEN) {
6210 case CPU_UP_PREPARE:
6211 case CPU_DOWN_FAILED:
6212 perf_event_init_cpu(cpu);
6215 case CPU_UP_CANCELED:
6216 case CPU_DOWN_PREPARE:
6217 perf_event_exit_cpu(cpu);
6227 void __init perf_event_init(void)
6229 perf_event_init_all_cpus();
6230 init_srcu_struct(&pmus_srcu);
6231 perf_pmu_register(&perf_swevent);
6232 perf_pmu_register(&perf_cpu_clock);
6233 perf_pmu_register(&perf_task_clock);
6235 perf_cpu_notifier(perf_cpu_notify);