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, unsigned long *flags)
153 struct perf_event_context *ctx;
157 ctx = rcu_dereference(task->perf_event_ctxp);
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)) {
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 *perf_pin_task_context(struct task_struct *task)
191 struct perf_event_context *ctx;
194 ctx = perf_lock_task_context(task, &flags);
197 raw_spin_unlock_irqrestore(&ctx->lock, flags);
202 static void perf_unpin_context(struct perf_event_context *ctx)
206 raw_spin_lock_irqsave(&ctx->lock, flags);
208 raw_spin_unlock_irqrestore(&ctx->lock, flags);
212 static inline u64 perf_clock(void)
214 return local_clock();
218 * Update the record of the current time in a context.
220 static void update_context_time(struct perf_event_context *ctx)
222 u64 now = perf_clock();
224 ctx->time += now - ctx->timestamp;
225 ctx->timestamp = now;
229 * Update the total_time_enabled and total_time_running fields for a event.
231 static void update_event_times(struct perf_event *event)
233 struct perf_event_context *ctx = event->ctx;
236 if (event->state < PERF_EVENT_STATE_INACTIVE ||
237 event->group_leader->state < PERF_EVENT_STATE_INACTIVE)
243 run_end = event->tstamp_stopped;
245 event->total_time_enabled = run_end - event->tstamp_enabled;
247 if (event->state == PERF_EVENT_STATE_INACTIVE)
248 run_end = event->tstamp_stopped;
252 event->total_time_running = run_end - event->tstamp_running;
256 * Update total_time_enabled and total_time_running for all events in a group.
258 static void update_group_times(struct perf_event *leader)
260 struct perf_event *event;
262 update_event_times(leader);
263 list_for_each_entry(event, &leader->sibling_list, group_entry)
264 update_event_times(event);
267 static struct list_head *
268 ctx_group_list(struct perf_event *event, struct perf_event_context *ctx)
270 if (event->attr.pinned)
271 return &ctx->pinned_groups;
273 return &ctx->flexible_groups;
277 * Add a event from the lists for its context.
278 * Must be called with ctx->mutex and ctx->lock held.
281 list_add_event(struct perf_event *event, struct perf_event_context *ctx)
283 WARN_ON_ONCE(event->attach_state & PERF_ATTACH_CONTEXT);
284 event->attach_state |= PERF_ATTACH_CONTEXT;
287 * If we're a stand alone event or group leader, we go to the context
288 * list, group events are kept attached to the group so that
289 * perf_group_detach can, at all times, locate all siblings.
291 if (event->group_leader == event) {
292 struct list_head *list;
294 if (is_software_event(event))
295 event->group_flags |= PERF_GROUP_SOFTWARE;
297 list = ctx_group_list(event, ctx);
298 list_add_tail(&event->group_entry, list);
301 list_add_rcu(&event->event_entry, &ctx->event_list);
303 perf_pmu_rotate_start(ctx->pmu);
305 if (event->attr.inherit_stat)
309 static void perf_group_attach(struct perf_event *event)
311 struct perf_event *group_leader = event->group_leader;
313 WARN_ON_ONCE(event->attach_state & PERF_ATTACH_GROUP);
314 event->attach_state |= PERF_ATTACH_GROUP;
316 if (group_leader == event)
319 if (group_leader->group_flags & PERF_GROUP_SOFTWARE &&
320 !is_software_event(event))
321 group_leader->group_flags &= ~PERF_GROUP_SOFTWARE;
323 list_add_tail(&event->group_entry, &group_leader->sibling_list);
324 group_leader->nr_siblings++;
328 * Remove a event from the lists for its context.
329 * Must be called with ctx->mutex and ctx->lock held.
332 list_del_event(struct perf_event *event, struct perf_event_context *ctx)
335 * We can have double detach due to exit/hot-unplug + close.
337 if (!(event->attach_state & PERF_ATTACH_CONTEXT))
340 event->attach_state &= ~PERF_ATTACH_CONTEXT;
343 if (event->attr.inherit_stat)
346 list_del_rcu(&event->event_entry);
348 if (event->group_leader == event)
349 list_del_init(&event->group_entry);
351 update_group_times(event);
354 * If event was in error state, then keep it
355 * that way, otherwise bogus counts will be
356 * returned on read(). The only way to get out
357 * of error state is by explicit re-enabling
360 if (event->state > PERF_EVENT_STATE_OFF)
361 event->state = PERF_EVENT_STATE_OFF;
364 static void perf_group_detach(struct perf_event *event)
366 struct perf_event *sibling, *tmp;
367 struct list_head *list = NULL;
370 * We can have double detach due to exit/hot-unplug + close.
372 if (!(event->attach_state & PERF_ATTACH_GROUP))
375 event->attach_state &= ~PERF_ATTACH_GROUP;
378 * If this is a sibling, remove it from its group.
380 if (event->group_leader != event) {
381 list_del_init(&event->group_entry);
382 event->group_leader->nr_siblings--;
386 if (!list_empty(&event->group_entry))
387 list = &event->group_entry;
390 * If this was a group event with sibling events then
391 * upgrade the siblings to singleton events by adding them
392 * to whatever list we are on.
394 list_for_each_entry_safe(sibling, tmp, &event->sibling_list, group_entry) {
396 list_move_tail(&sibling->group_entry, list);
397 sibling->group_leader = sibling;
399 /* Inherit group flags from the previous leader */
400 sibling->group_flags = event->group_flags;
405 event_filter_match(struct perf_event *event)
407 return event->cpu == -1 || event->cpu == smp_processor_id();
411 event_sched_out(struct perf_event *event,
412 struct perf_cpu_context *cpuctx,
413 struct perf_event_context *ctx)
417 * An event which could not be activated because of
418 * filter mismatch still needs to have its timings
419 * maintained, otherwise bogus information is return
420 * via read() for time_enabled, time_running:
422 if (event->state == PERF_EVENT_STATE_INACTIVE
423 && !event_filter_match(event)) {
424 delta = ctx->time - event->tstamp_stopped;
425 event->tstamp_running += delta;
426 event->tstamp_stopped = ctx->time;
429 if (event->state != PERF_EVENT_STATE_ACTIVE)
432 event->state = PERF_EVENT_STATE_INACTIVE;
433 if (event->pending_disable) {
434 event->pending_disable = 0;
435 event->state = PERF_EVENT_STATE_OFF;
437 event->tstamp_stopped = ctx->time;
438 event->pmu->del(event, 0);
441 if (!is_software_event(event))
442 cpuctx->active_oncpu--;
444 if (event->attr.exclusive || !cpuctx->active_oncpu)
445 cpuctx->exclusive = 0;
449 group_sched_out(struct perf_event *group_event,
450 struct perf_cpu_context *cpuctx,
451 struct perf_event_context *ctx)
453 struct perf_event *event;
454 int state = group_event->state;
456 event_sched_out(group_event, cpuctx, ctx);
459 * Schedule out siblings (if any):
461 list_for_each_entry(event, &group_event->sibling_list, group_entry)
462 event_sched_out(event, cpuctx, ctx);
464 if (state == PERF_EVENT_STATE_ACTIVE && group_event->attr.exclusive)
465 cpuctx->exclusive = 0;
468 static inline struct perf_cpu_context *
469 __get_cpu_context(struct perf_event_context *ctx)
471 return this_cpu_ptr(ctx->pmu->pmu_cpu_context);
475 * Cross CPU call to remove a performance event
477 * We disable the event on the hardware level first. After that we
478 * remove it from the context list.
480 static void __perf_event_remove_from_context(void *info)
482 struct perf_event *event = info;
483 struct perf_event_context *ctx = event->ctx;
484 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
487 * If this is a task context, we need to check whether it is
488 * the current task context of this cpu. If not it has been
489 * scheduled out before the smp call arrived.
491 if (ctx->task && cpuctx->task_ctx != ctx)
494 raw_spin_lock(&ctx->lock);
496 event_sched_out(event, cpuctx, ctx);
498 list_del_event(event, ctx);
500 raw_spin_unlock(&ctx->lock);
505 * Remove the event from a task's (or a CPU's) list of events.
507 * Must be called with ctx->mutex held.
509 * CPU events are removed with a smp call. For task events we only
510 * call when the task is on a CPU.
512 * If event->ctx is a cloned context, callers must make sure that
513 * every task struct that event->ctx->task could possibly point to
514 * remains valid. This is OK when called from perf_release since
515 * that only calls us on the top-level context, which can't be a clone.
516 * When called from perf_event_exit_task, it's OK because the
517 * context has been detached from its task.
519 static void perf_event_remove_from_context(struct perf_event *event)
521 struct perf_event_context *ctx = event->ctx;
522 struct task_struct *task = ctx->task;
526 * Per cpu events are removed via an smp call and
527 * the removal is always successful.
529 smp_call_function_single(event->cpu,
530 __perf_event_remove_from_context,
536 task_oncpu_function_call(task, __perf_event_remove_from_context,
539 raw_spin_lock_irq(&ctx->lock);
541 * If the context is active we need to retry the smp call.
543 if (ctx->nr_active && !list_empty(&event->group_entry)) {
544 raw_spin_unlock_irq(&ctx->lock);
549 * The lock prevents that this context is scheduled in so we
550 * can remove the event safely, if the call above did not
553 if (!list_empty(&event->group_entry))
554 list_del_event(event, ctx);
555 raw_spin_unlock_irq(&ctx->lock);
559 * Cross CPU call to disable a performance event
561 static void __perf_event_disable(void *info)
563 struct perf_event *event = info;
564 struct perf_event_context *ctx = event->ctx;
565 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
568 * If this is a per-task event, need to check whether this
569 * event's task is the current task on this cpu.
571 if (ctx->task && cpuctx->task_ctx != ctx)
574 raw_spin_lock(&ctx->lock);
577 * If the event is on, turn it off.
578 * If it is in error state, leave it in error state.
580 if (event->state >= PERF_EVENT_STATE_INACTIVE) {
581 update_context_time(ctx);
582 update_group_times(event);
583 if (event == event->group_leader)
584 group_sched_out(event, cpuctx, ctx);
586 event_sched_out(event, cpuctx, ctx);
587 event->state = PERF_EVENT_STATE_OFF;
590 raw_spin_unlock(&ctx->lock);
596 * If event->ctx is a cloned context, callers must make sure that
597 * every task struct that event->ctx->task could possibly point to
598 * remains valid. This condition is satisifed when called through
599 * perf_event_for_each_child or perf_event_for_each because they
600 * hold the top-level event's child_mutex, so any descendant that
601 * goes to exit will block in sync_child_event.
602 * When called from perf_pending_event it's OK because event->ctx
603 * is the current context on this CPU and preemption is disabled,
604 * hence we can't get into perf_event_task_sched_out for this context.
606 void perf_event_disable(struct perf_event *event)
608 struct perf_event_context *ctx = event->ctx;
609 struct task_struct *task = ctx->task;
613 * Disable the event on the cpu that it's on
615 smp_call_function_single(event->cpu, __perf_event_disable,
621 task_oncpu_function_call(task, __perf_event_disable, event);
623 raw_spin_lock_irq(&ctx->lock);
625 * If the event is still active, we need to retry the cross-call.
627 if (event->state == PERF_EVENT_STATE_ACTIVE) {
628 raw_spin_unlock_irq(&ctx->lock);
633 * Since we have the lock this context can't be scheduled
634 * in, so we can change the state safely.
636 if (event->state == PERF_EVENT_STATE_INACTIVE) {
637 update_group_times(event);
638 event->state = PERF_EVENT_STATE_OFF;
641 raw_spin_unlock_irq(&ctx->lock);
645 event_sched_in(struct perf_event *event,
646 struct perf_cpu_context *cpuctx,
647 struct perf_event_context *ctx)
649 if (event->state <= PERF_EVENT_STATE_OFF)
652 event->state = PERF_EVENT_STATE_ACTIVE;
653 event->oncpu = smp_processor_id();
655 * The new state must be visible before we turn it on in the hardware:
659 if (event->pmu->add(event, PERF_EF_START)) {
660 event->state = PERF_EVENT_STATE_INACTIVE;
665 event->tstamp_running += ctx->time - event->tstamp_stopped;
667 if (!is_software_event(event))
668 cpuctx->active_oncpu++;
671 if (event->attr.exclusive)
672 cpuctx->exclusive = 1;
678 group_sched_in(struct perf_event *group_event,
679 struct perf_cpu_context *cpuctx,
680 struct perf_event_context *ctx)
682 struct perf_event *event, *partial_group = NULL;
683 struct pmu *pmu = group_event->pmu;
685 if (group_event->state == PERF_EVENT_STATE_OFF)
690 if (event_sched_in(group_event, cpuctx, ctx)) {
691 pmu->cancel_txn(pmu);
696 * Schedule in siblings as one group (if any):
698 list_for_each_entry(event, &group_event->sibling_list, group_entry) {
699 if (event_sched_in(event, cpuctx, ctx)) {
700 partial_group = event;
705 if (!pmu->commit_txn(pmu))
710 * Groups can be scheduled in as one unit only, so undo any
711 * partial group before returning:
713 list_for_each_entry(event, &group_event->sibling_list, group_entry) {
714 if (event == partial_group)
716 event_sched_out(event, cpuctx, ctx);
718 event_sched_out(group_event, cpuctx, ctx);
720 pmu->cancel_txn(pmu);
726 * Work out whether we can put this event group on the CPU now.
728 static int group_can_go_on(struct perf_event *event,
729 struct perf_cpu_context *cpuctx,
733 * Groups consisting entirely of software events can always go on.
735 if (event->group_flags & PERF_GROUP_SOFTWARE)
738 * If an exclusive group is already on, no other hardware
741 if (cpuctx->exclusive)
744 * If this group is exclusive and there are already
745 * events on the CPU, it can't go on.
747 if (event->attr.exclusive && cpuctx->active_oncpu)
750 * Otherwise, try to add it if all previous groups were able
756 static void add_event_to_ctx(struct perf_event *event,
757 struct perf_event_context *ctx)
759 list_add_event(event, ctx);
760 perf_group_attach(event);
761 event->tstamp_enabled = ctx->time;
762 event->tstamp_running = ctx->time;
763 event->tstamp_stopped = ctx->time;
767 * Cross CPU call to install and enable a performance event
769 * Must be called with ctx->mutex held
771 static void __perf_install_in_context(void *info)
773 struct perf_event *event = info;
774 struct perf_event_context *ctx = event->ctx;
775 struct perf_event *leader = event->group_leader;
776 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
780 * If this is a task context, we need to check whether it is
781 * the current task context of this cpu. If not it has been
782 * scheduled out before the smp call arrived.
783 * Or possibly this is the right context but it isn't
784 * on this cpu because it had no events.
786 if (ctx->task && cpuctx->task_ctx != ctx) {
787 if (cpuctx->task_ctx || ctx->task != current)
789 cpuctx->task_ctx = ctx;
792 raw_spin_lock(&ctx->lock);
794 update_context_time(ctx);
796 add_event_to_ctx(event, ctx);
798 if (event->cpu != -1 && event->cpu != smp_processor_id())
802 * Don't put the event on if it is disabled or if
803 * it is in a group and the group isn't on.
805 if (event->state != PERF_EVENT_STATE_INACTIVE ||
806 (leader != event && leader->state != PERF_EVENT_STATE_ACTIVE))
810 * An exclusive event can't go on if there are already active
811 * hardware events, and no hardware event can go on if there
812 * is already an exclusive event on.
814 if (!group_can_go_on(event, cpuctx, 1))
817 err = event_sched_in(event, cpuctx, ctx);
821 * This event couldn't go on. If it is in a group
822 * then we have to pull the whole group off.
823 * If the event group is pinned then put it in error state.
826 group_sched_out(leader, cpuctx, ctx);
827 if (leader->attr.pinned) {
828 update_group_times(leader);
829 leader->state = PERF_EVENT_STATE_ERROR;
834 raw_spin_unlock(&ctx->lock);
838 * Attach a performance event to a context
840 * First we add the event to the list with the hardware enable bit
841 * in event->hw_config cleared.
843 * If the event is attached to a task which is on a CPU we use a smp
844 * call to enable it in the task context. The task might have been
845 * scheduled away, but we check this in the smp call again.
847 * Must be called with ctx->mutex held.
850 perf_install_in_context(struct perf_event_context *ctx,
851 struct perf_event *event,
854 struct task_struct *task = ctx->task;
860 * Per cpu events are installed via an smp call and
861 * the install is always successful.
863 smp_call_function_single(cpu, __perf_install_in_context,
869 task_oncpu_function_call(task, __perf_install_in_context,
872 raw_spin_lock_irq(&ctx->lock);
874 * we need to retry the smp call.
876 if (ctx->is_active && list_empty(&event->group_entry)) {
877 raw_spin_unlock_irq(&ctx->lock);
882 * The lock prevents that this context is scheduled in so we
883 * can add the event safely, if it the call above did not
886 if (list_empty(&event->group_entry))
887 add_event_to_ctx(event, ctx);
888 raw_spin_unlock_irq(&ctx->lock);
892 * Put a event into inactive state and update time fields.
893 * Enabling the leader of a group effectively enables all
894 * the group members that aren't explicitly disabled, so we
895 * have to update their ->tstamp_enabled also.
896 * Note: this works for group members as well as group leaders
897 * since the non-leader members' sibling_lists will be empty.
899 static void __perf_event_mark_enabled(struct perf_event *event,
900 struct perf_event_context *ctx)
902 struct perf_event *sub;
904 event->state = PERF_EVENT_STATE_INACTIVE;
905 event->tstamp_enabled = ctx->time - event->total_time_enabled;
906 list_for_each_entry(sub, &event->sibling_list, group_entry) {
907 if (sub->state >= PERF_EVENT_STATE_INACTIVE) {
908 sub->tstamp_enabled =
909 ctx->time - sub->total_time_enabled;
915 * Cross CPU call to enable a performance event
917 static void __perf_event_enable(void *info)
919 struct perf_event *event = info;
920 struct perf_event_context *ctx = event->ctx;
921 struct perf_event *leader = event->group_leader;
922 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
926 * If this is a per-task event, need to check whether this
927 * event's task is the current task on this cpu.
929 if (ctx->task && cpuctx->task_ctx != ctx) {
930 if (cpuctx->task_ctx || ctx->task != current)
932 cpuctx->task_ctx = ctx;
935 raw_spin_lock(&ctx->lock);
937 update_context_time(ctx);
939 if (event->state >= PERF_EVENT_STATE_INACTIVE)
941 __perf_event_mark_enabled(event, ctx);
943 if (event->cpu != -1 && event->cpu != smp_processor_id())
947 * If the event is in a group and isn't the group leader,
948 * then don't put it on unless the group is on.
950 if (leader != event && leader->state != PERF_EVENT_STATE_ACTIVE)
953 if (!group_can_go_on(event, cpuctx, 1)) {
957 err = group_sched_in(event, cpuctx, ctx);
959 err = event_sched_in(event, cpuctx, ctx);
964 * If this event can't go on and it's part of a
965 * group, then the whole group has to come off.
968 group_sched_out(leader, cpuctx, ctx);
969 if (leader->attr.pinned) {
970 update_group_times(leader);
971 leader->state = PERF_EVENT_STATE_ERROR;
976 raw_spin_unlock(&ctx->lock);
982 * If event->ctx is a cloned context, callers must make sure that
983 * every task struct that event->ctx->task could possibly point to
984 * remains valid. This condition is satisfied when called through
985 * perf_event_for_each_child or perf_event_for_each as described
986 * for perf_event_disable.
988 void perf_event_enable(struct perf_event *event)
990 struct perf_event_context *ctx = event->ctx;
991 struct task_struct *task = ctx->task;
995 * Enable the event on the cpu that it's on
997 smp_call_function_single(event->cpu, __perf_event_enable,
1002 raw_spin_lock_irq(&ctx->lock);
1003 if (event->state >= PERF_EVENT_STATE_INACTIVE)
1007 * If the event is in error state, clear that first.
1008 * That way, if we see the event in error state below, we
1009 * know that it has gone back into error state, as distinct
1010 * from the task having been scheduled away before the
1011 * cross-call arrived.
1013 if (event->state == PERF_EVENT_STATE_ERROR)
1014 event->state = PERF_EVENT_STATE_OFF;
1017 raw_spin_unlock_irq(&ctx->lock);
1018 task_oncpu_function_call(task, __perf_event_enable, event);
1020 raw_spin_lock_irq(&ctx->lock);
1023 * If the context is active and the event is still off,
1024 * we need to retry the cross-call.
1026 if (ctx->is_active && event->state == PERF_EVENT_STATE_OFF)
1030 * Since we have the lock this context can't be scheduled
1031 * in, so we can change the state safely.
1033 if (event->state == PERF_EVENT_STATE_OFF)
1034 __perf_event_mark_enabled(event, ctx);
1037 raw_spin_unlock_irq(&ctx->lock);
1040 static int perf_event_refresh(struct perf_event *event, int refresh)
1043 * not supported on inherited events
1045 if (event->attr.inherit)
1048 atomic_add(refresh, &event->event_limit);
1049 perf_event_enable(event);
1055 EVENT_FLEXIBLE = 0x1,
1057 EVENT_ALL = EVENT_FLEXIBLE | EVENT_PINNED,
1060 static void ctx_sched_out(struct perf_event_context *ctx,
1061 struct perf_cpu_context *cpuctx,
1062 enum event_type_t event_type)
1064 struct perf_event *event;
1066 raw_spin_lock(&ctx->lock);
1068 if (likely(!ctx->nr_events))
1070 update_context_time(ctx);
1072 if (!ctx->nr_active)
1075 if (event_type & EVENT_PINNED) {
1076 list_for_each_entry(event, &ctx->pinned_groups, group_entry)
1077 group_sched_out(event, cpuctx, ctx);
1080 if (event_type & EVENT_FLEXIBLE) {
1081 list_for_each_entry(event, &ctx->flexible_groups, group_entry)
1082 group_sched_out(event, cpuctx, ctx);
1085 raw_spin_unlock(&ctx->lock);
1089 * Test whether two contexts are equivalent, i.e. whether they
1090 * have both been cloned from the same version of the same context
1091 * and they both have the same number of enabled events.
1092 * If the number of enabled events is the same, then the set
1093 * of enabled events should be the same, because these are both
1094 * inherited contexts, therefore we can't access individual events
1095 * in them directly with an fd; we can only enable/disable all
1096 * events via prctl, or enable/disable all events in a family
1097 * via ioctl, which will have the same effect on both contexts.
1099 static int context_equiv(struct perf_event_context *ctx1,
1100 struct perf_event_context *ctx2)
1102 return ctx1->parent_ctx && ctx1->parent_ctx == ctx2->parent_ctx
1103 && ctx1->parent_gen == ctx2->parent_gen
1104 && !ctx1->pin_count && !ctx2->pin_count;
1107 static void __perf_event_sync_stat(struct perf_event *event,
1108 struct perf_event *next_event)
1112 if (!event->attr.inherit_stat)
1116 * Update the event value, we cannot use perf_event_read()
1117 * because we're in the middle of a context switch and have IRQs
1118 * disabled, which upsets smp_call_function_single(), however
1119 * we know the event must be on the current CPU, therefore we
1120 * don't need to use it.
1122 switch (event->state) {
1123 case PERF_EVENT_STATE_ACTIVE:
1124 event->pmu->read(event);
1127 case PERF_EVENT_STATE_INACTIVE:
1128 update_event_times(event);
1136 * In order to keep per-task stats reliable we need to flip the event
1137 * values when we flip the contexts.
1139 value = local64_read(&next_event->count);
1140 value = local64_xchg(&event->count, value);
1141 local64_set(&next_event->count, value);
1143 swap(event->total_time_enabled, next_event->total_time_enabled);
1144 swap(event->total_time_running, next_event->total_time_running);
1147 * Since we swizzled the values, update the user visible data too.
1149 perf_event_update_userpage(event);
1150 perf_event_update_userpage(next_event);
1153 #define list_next_entry(pos, member) \
1154 list_entry(pos->member.next, typeof(*pos), member)
1156 static void perf_event_sync_stat(struct perf_event_context *ctx,
1157 struct perf_event_context *next_ctx)
1159 struct perf_event *event, *next_event;
1164 update_context_time(ctx);
1166 event = list_first_entry(&ctx->event_list,
1167 struct perf_event, event_entry);
1169 next_event = list_first_entry(&next_ctx->event_list,
1170 struct perf_event, event_entry);
1172 while (&event->event_entry != &ctx->event_list &&
1173 &next_event->event_entry != &next_ctx->event_list) {
1175 __perf_event_sync_stat(event, next_event);
1177 event = list_next_entry(event, event_entry);
1178 next_event = list_next_entry(next_event, event_entry);
1183 * Called from scheduler to remove the events of the current task,
1184 * with interrupts disabled.
1186 * We stop each event and update the event value in event->count.
1188 * This does not protect us against NMI, but disable()
1189 * sets the disabled bit in the control field of event _before_
1190 * accessing the event control register. If a NMI hits, then it will
1191 * not restart the event.
1193 void perf_event_task_sched_out(struct task_struct *task,
1194 struct task_struct *next)
1196 struct perf_event_context *ctx = task->perf_event_ctxp;
1197 struct perf_event_context *next_ctx;
1198 struct perf_event_context *parent;
1199 struct perf_cpu_context *cpuctx;
1202 perf_sw_event(PERF_COUNT_SW_CONTEXT_SWITCHES, 1, 1, NULL, 0);
1207 cpuctx = __get_cpu_context(ctx);
1208 if (!cpuctx->task_ctx)
1212 parent = rcu_dereference(ctx->parent_ctx);
1213 next_ctx = next->perf_event_ctxp;
1214 if (parent && next_ctx &&
1215 rcu_dereference(next_ctx->parent_ctx) == parent) {
1217 * Looks like the two contexts are clones, so we might be
1218 * able to optimize the context switch. We lock both
1219 * contexts and check that they are clones under the
1220 * lock (including re-checking that neither has been
1221 * uncloned in the meantime). It doesn't matter which
1222 * order we take the locks because no other cpu could
1223 * be trying to lock both of these tasks.
1225 raw_spin_lock(&ctx->lock);
1226 raw_spin_lock_nested(&next_ctx->lock, SINGLE_DEPTH_NESTING);
1227 if (context_equiv(ctx, next_ctx)) {
1229 * XXX do we need a memory barrier of sorts
1230 * wrt to rcu_dereference() of perf_event_ctxp
1232 task->perf_event_ctxp = next_ctx;
1233 next->perf_event_ctxp = ctx;
1235 next_ctx->task = task;
1238 perf_event_sync_stat(ctx, next_ctx);
1240 raw_spin_unlock(&next_ctx->lock);
1241 raw_spin_unlock(&ctx->lock);
1246 ctx_sched_out(ctx, cpuctx, EVENT_ALL);
1247 cpuctx->task_ctx = NULL;
1251 static void task_ctx_sched_out(struct perf_event_context *ctx,
1252 enum event_type_t event_type)
1254 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
1256 if (!cpuctx->task_ctx)
1259 if (WARN_ON_ONCE(ctx != cpuctx->task_ctx))
1262 ctx_sched_out(ctx, cpuctx, event_type);
1263 cpuctx->task_ctx = NULL;
1267 * Called with IRQs disabled
1269 static void __perf_event_task_sched_out(struct perf_event_context *ctx)
1271 task_ctx_sched_out(ctx, EVENT_ALL);
1275 * Called with IRQs disabled
1277 static void cpu_ctx_sched_out(struct perf_cpu_context *cpuctx,
1278 enum event_type_t event_type)
1280 ctx_sched_out(&cpuctx->ctx, cpuctx, event_type);
1284 ctx_pinned_sched_in(struct perf_event_context *ctx,
1285 struct perf_cpu_context *cpuctx)
1287 struct perf_event *event;
1289 list_for_each_entry(event, &ctx->pinned_groups, group_entry) {
1290 if (event->state <= PERF_EVENT_STATE_OFF)
1292 if (event->cpu != -1 && event->cpu != smp_processor_id())
1295 if (group_can_go_on(event, cpuctx, 1))
1296 group_sched_in(event, cpuctx, ctx);
1299 * If this pinned group hasn't been scheduled,
1300 * put it in error state.
1302 if (event->state == PERF_EVENT_STATE_INACTIVE) {
1303 update_group_times(event);
1304 event->state = PERF_EVENT_STATE_ERROR;
1310 ctx_flexible_sched_in(struct perf_event_context *ctx,
1311 struct perf_cpu_context *cpuctx)
1313 struct perf_event *event;
1316 list_for_each_entry(event, &ctx->flexible_groups, group_entry) {
1317 /* Ignore events in OFF or ERROR state */
1318 if (event->state <= PERF_EVENT_STATE_OFF)
1321 * Listen to the 'cpu' scheduling filter constraint
1324 if (event->cpu != -1 && event->cpu != smp_processor_id())
1327 if (group_can_go_on(event, cpuctx, can_add_hw)) {
1328 if (group_sched_in(event, cpuctx, ctx))
1335 ctx_sched_in(struct perf_event_context *ctx,
1336 struct perf_cpu_context *cpuctx,
1337 enum event_type_t event_type)
1339 raw_spin_lock(&ctx->lock);
1341 if (likely(!ctx->nr_events))
1344 ctx->timestamp = perf_clock();
1347 * First go through the list and put on any pinned groups
1348 * in order to give them the best chance of going on.
1350 if (event_type & EVENT_PINNED)
1351 ctx_pinned_sched_in(ctx, cpuctx);
1353 /* Then walk through the lower prio flexible groups */
1354 if (event_type & EVENT_FLEXIBLE)
1355 ctx_flexible_sched_in(ctx, cpuctx);
1358 raw_spin_unlock(&ctx->lock);
1361 static void cpu_ctx_sched_in(struct perf_cpu_context *cpuctx,
1362 enum event_type_t event_type)
1364 struct perf_event_context *ctx = &cpuctx->ctx;
1366 ctx_sched_in(ctx, cpuctx, event_type);
1369 static void task_ctx_sched_in(struct task_struct *task,
1370 enum event_type_t event_type)
1372 struct perf_event_context *ctx = task->perf_event_ctxp;
1373 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
1377 if (cpuctx->task_ctx == ctx)
1379 ctx_sched_in(ctx, cpuctx, event_type);
1380 cpuctx->task_ctx = ctx;
1383 * Called from scheduler to add the events of the current task
1384 * with interrupts disabled.
1386 * We restore the event value and then enable it.
1388 * This does not protect us against NMI, but enable()
1389 * sets the enabled bit in the control field of event _before_
1390 * accessing the event control register. If a NMI hits, then it will
1391 * keep the event running.
1393 void perf_event_task_sched_in(struct task_struct *task)
1395 struct perf_event_context *ctx = task->perf_event_ctxp;
1396 struct perf_cpu_context *cpuctx;
1401 cpuctx = __get_cpu_context(ctx);
1402 if (cpuctx->task_ctx == ctx)
1406 * We want to keep the following priority order:
1407 * cpu pinned (that don't need to move), task pinned,
1408 * cpu flexible, task flexible.
1410 cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE);
1412 ctx_sched_in(ctx, cpuctx, EVENT_PINNED);
1413 cpu_ctx_sched_in(cpuctx, EVENT_FLEXIBLE);
1414 ctx_sched_in(ctx, cpuctx, EVENT_FLEXIBLE);
1416 cpuctx->task_ctx = ctx;
1419 * Since these rotations are per-cpu, we need to ensure the
1420 * cpu-context we got scheduled on is actually rotating.
1422 perf_pmu_rotate_start(ctx->pmu);
1425 #define MAX_INTERRUPTS (~0ULL)
1427 static void perf_log_throttle(struct perf_event *event, int enable);
1429 static u64 perf_calculate_period(struct perf_event *event, u64 nsec, u64 count)
1431 u64 frequency = event->attr.sample_freq;
1432 u64 sec = NSEC_PER_SEC;
1433 u64 divisor, dividend;
1435 int count_fls, nsec_fls, frequency_fls, sec_fls;
1437 count_fls = fls64(count);
1438 nsec_fls = fls64(nsec);
1439 frequency_fls = fls64(frequency);
1443 * We got @count in @nsec, with a target of sample_freq HZ
1444 * the target period becomes:
1447 * period = -------------------
1448 * @nsec * sample_freq
1453 * Reduce accuracy by one bit such that @a and @b converge
1454 * to a similar magnitude.
1456 #define REDUCE_FLS(a, b) \
1458 if (a##_fls > b##_fls) { \
1468 * Reduce accuracy until either term fits in a u64, then proceed with
1469 * the other, so that finally we can do a u64/u64 division.
1471 while (count_fls + sec_fls > 64 && nsec_fls + frequency_fls > 64) {
1472 REDUCE_FLS(nsec, frequency);
1473 REDUCE_FLS(sec, count);
1476 if (count_fls + sec_fls > 64) {
1477 divisor = nsec * frequency;
1479 while (count_fls + sec_fls > 64) {
1480 REDUCE_FLS(count, sec);
1484 dividend = count * sec;
1486 dividend = count * sec;
1488 while (nsec_fls + frequency_fls > 64) {
1489 REDUCE_FLS(nsec, frequency);
1493 divisor = nsec * frequency;
1499 return div64_u64(dividend, divisor);
1502 static void perf_adjust_period(struct perf_event *event, u64 nsec, u64 count)
1504 struct hw_perf_event *hwc = &event->hw;
1505 s64 period, sample_period;
1508 period = perf_calculate_period(event, nsec, count);
1510 delta = (s64)(period - hwc->sample_period);
1511 delta = (delta + 7) / 8; /* low pass filter */
1513 sample_period = hwc->sample_period + delta;
1518 hwc->sample_period = sample_period;
1520 if (local64_read(&hwc->period_left) > 8*sample_period) {
1521 event->pmu->stop(event, PERF_EF_UPDATE);
1522 local64_set(&hwc->period_left, 0);
1523 event->pmu->start(event, PERF_EF_RELOAD);
1527 static void perf_ctx_adjust_freq(struct perf_event_context *ctx, u64 period)
1529 struct perf_event *event;
1530 struct hw_perf_event *hwc;
1531 u64 interrupts, now;
1534 raw_spin_lock(&ctx->lock);
1535 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
1536 if (event->state != PERF_EVENT_STATE_ACTIVE)
1539 if (event->cpu != -1 && event->cpu != smp_processor_id())
1544 interrupts = hwc->interrupts;
1545 hwc->interrupts = 0;
1548 * unthrottle events on the tick
1550 if (interrupts == MAX_INTERRUPTS) {
1551 perf_log_throttle(event, 1);
1552 event->pmu->start(event, 0);
1555 if (!event->attr.freq || !event->attr.sample_freq)
1558 event->pmu->read(event);
1559 now = local64_read(&event->count);
1560 delta = now - hwc->freq_count_stamp;
1561 hwc->freq_count_stamp = now;
1564 perf_adjust_period(event, period, delta);
1566 raw_spin_unlock(&ctx->lock);
1570 * Round-robin a context's events:
1572 static void rotate_ctx(struct perf_event_context *ctx)
1574 raw_spin_lock(&ctx->lock);
1576 /* Rotate the first entry last of non-pinned groups */
1577 list_rotate_left(&ctx->flexible_groups);
1579 raw_spin_unlock(&ctx->lock);
1583 * Cannot race with ->pmu_rotate_start() because this is ran from hardirq
1584 * context, and ->pmu_rotate_start() is called with irqs disabled (both are
1585 * cpu affine, so there are no SMP races).
1587 static enum hrtimer_restart perf_event_context_tick(struct hrtimer *timer)
1589 enum hrtimer_restart restart = HRTIMER_NORESTART;
1590 struct perf_cpu_context *cpuctx;
1591 struct perf_event_context *ctx;
1594 cpuctx = container_of(timer, struct perf_cpu_context, timer);
1596 if (cpuctx->ctx.nr_events) {
1597 restart = HRTIMER_RESTART;
1598 if (cpuctx->ctx.nr_events != cpuctx->ctx.nr_active)
1602 ctx = current->perf_event_ctxp;
1603 if (ctx && ctx->nr_events) {
1604 restart = HRTIMER_RESTART;
1605 if (ctx->nr_events != ctx->nr_active)
1609 perf_ctx_adjust_freq(&cpuctx->ctx, cpuctx->timer_interval);
1611 perf_ctx_adjust_freq(ctx, cpuctx->timer_interval);
1616 cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE);
1618 task_ctx_sched_out(ctx, EVENT_FLEXIBLE);
1620 rotate_ctx(&cpuctx->ctx);
1624 cpu_ctx_sched_in(cpuctx, EVENT_FLEXIBLE);
1626 task_ctx_sched_in(current, EVENT_FLEXIBLE);
1629 hrtimer_forward_now(timer, ns_to_ktime(cpuctx->timer_interval));
1634 static int event_enable_on_exec(struct perf_event *event,
1635 struct perf_event_context *ctx)
1637 if (!event->attr.enable_on_exec)
1640 event->attr.enable_on_exec = 0;
1641 if (event->state >= PERF_EVENT_STATE_INACTIVE)
1644 __perf_event_mark_enabled(event, ctx);
1650 * Enable all of a task's events that have been marked enable-on-exec.
1651 * This expects task == current.
1653 static void perf_event_enable_on_exec(struct task_struct *task)
1655 struct perf_event_context *ctx;
1656 struct perf_event *event;
1657 unsigned long flags;
1661 local_irq_save(flags);
1662 ctx = task->perf_event_ctxp;
1663 if (!ctx || !ctx->nr_events)
1666 __perf_event_task_sched_out(ctx);
1668 raw_spin_lock(&ctx->lock);
1670 list_for_each_entry(event, &ctx->pinned_groups, group_entry) {
1671 ret = event_enable_on_exec(event, ctx);
1676 list_for_each_entry(event, &ctx->flexible_groups, group_entry) {
1677 ret = event_enable_on_exec(event, ctx);
1683 * Unclone this context if we enabled any event.
1688 raw_spin_unlock(&ctx->lock);
1690 perf_event_task_sched_in(task);
1692 local_irq_restore(flags);
1696 * Cross CPU call to read the hardware event
1698 static void __perf_event_read(void *info)
1700 struct perf_event *event = info;
1701 struct perf_event_context *ctx = event->ctx;
1702 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
1705 * If this is a task context, we need to check whether it is
1706 * the current task context of this cpu. If not it has been
1707 * scheduled out before the smp call arrived. In that case
1708 * event->count would have been updated to a recent sample
1709 * when the event was scheduled out.
1711 if (ctx->task && cpuctx->task_ctx != ctx)
1714 raw_spin_lock(&ctx->lock);
1715 update_context_time(ctx);
1716 update_event_times(event);
1717 raw_spin_unlock(&ctx->lock);
1719 event->pmu->read(event);
1722 static inline u64 perf_event_count(struct perf_event *event)
1724 return local64_read(&event->count) + atomic64_read(&event->child_count);
1727 static u64 perf_event_read(struct perf_event *event)
1730 * If event is enabled and currently active on a CPU, update the
1731 * value in the event structure:
1733 if (event->state == PERF_EVENT_STATE_ACTIVE) {
1734 smp_call_function_single(event->oncpu,
1735 __perf_event_read, event, 1);
1736 } else if (event->state == PERF_EVENT_STATE_INACTIVE) {
1737 struct perf_event_context *ctx = event->ctx;
1738 unsigned long flags;
1740 raw_spin_lock_irqsave(&ctx->lock, flags);
1741 update_context_time(ctx);
1742 update_event_times(event);
1743 raw_spin_unlock_irqrestore(&ctx->lock, flags);
1746 return perf_event_count(event);
1753 struct callchain_cpus_entries {
1754 struct rcu_head rcu_head;
1755 struct perf_callchain_entry *cpu_entries[0];
1758 static DEFINE_PER_CPU(int, callchain_recursion[PERF_NR_CONTEXTS]);
1759 static atomic_t nr_callchain_events;
1760 static DEFINE_MUTEX(callchain_mutex);
1761 struct callchain_cpus_entries *callchain_cpus_entries;
1764 __weak void perf_callchain_kernel(struct perf_callchain_entry *entry,
1765 struct pt_regs *regs)
1769 __weak void perf_callchain_user(struct perf_callchain_entry *entry,
1770 struct pt_regs *regs)
1774 static void release_callchain_buffers_rcu(struct rcu_head *head)
1776 struct callchain_cpus_entries *entries;
1779 entries = container_of(head, struct callchain_cpus_entries, rcu_head);
1781 for_each_possible_cpu(cpu)
1782 kfree(entries->cpu_entries[cpu]);
1787 static void release_callchain_buffers(void)
1789 struct callchain_cpus_entries *entries;
1791 entries = callchain_cpus_entries;
1792 rcu_assign_pointer(callchain_cpus_entries, NULL);
1793 call_rcu(&entries->rcu_head, release_callchain_buffers_rcu);
1796 static int alloc_callchain_buffers(void)
1800 struct callchain_cpus_entries *entries;
1803 * We can't use the percpu allocation API for data that can be
1804 * accessed from NMI. Use a temporary manual per cpu allocation
1805 * until that gets sorted out.
1807 size = sizeof(*entries) + sizeof(struct perf_callchain_entry *) *
1808 num_possible_cpus();
1810 entries = kzalloc(size, GFP_KERNEL);
1814 size = sizeof(struct perf_callchain_entry) * PERF_NR_CONTEXTS;
1816 for_each_possible_cpu(cpu) {
1817 entries->cpu_entries[cpu] = kmalloc_node(size, GFP_KERNEL,
1819 if (!entries->cpu_entries[cpu])
1823 rcu_assign_pointer(callchain_cpus_entries, entries);
1828 for_each_possible_cpu(cpu)
1829 kfree(entries->cpu_entries[cpu]);
1835 static int get_callchain_buffers(void)
1840 mutex_lock(&callchain_mutex);
1842 count = atomic_inc_return(&nr_callchain_events);
1843 if (WARN_ON_ONCE(count < 1)) {
1849 /* If the allocation failed, give up */
1850 if (!callchain_cpus_entries)
1855 err = alloc_callchain_buffers();
1857 release_callchain_buffers();
1859 mutex_unlock(&callchain_mutex);
1864 static void put_callchain_buffers(void)
1866 if (atomic_dec_and_mutex_lock(&nr_callchain_events, &callchain_mutex)) {
1867 release_callchain_buffers();
1868 mutex_unlock(&callchain_mutex);
1872 static int get_recursion_context(int *recursion)
1880 else if (in_softirq())
1885 if (recursion[rctx])
1894 static inline void put_recursion_context(int *recursion, int rctx)
1900 static struct perf_callchain_entry *get_callchain_entry(int *rctx)
1903 struct callchain_cpus_entries *entries;
1905 *rctx = get_recursion_context(__get_cpu_var(callchain_recursion));
1909 entries = rcu_dereference(callchain_cpus_entries);
1913 cpu = smp_processor_id();
1915 return &entries->cpu_entries[cpu][*rctx];
1919 put_callchain_entry(int rctx)
1921 put_recursion_context(__get_cpu_var(callchain_recursion), rctx);
1924 static struct perf_callchain_entry *perf_callchain(struct pt_regs *regs)
1927 struct perf_callchain_entry *entry;
1930 entry = get_callchain_entry(&rctx);
1939 if (!user_mode(regs)) {
1940 perf_callchain_store(entry, PERF_CONTEXT_KERNEL);
1941 perf_callchain_kernel(entry, regs);
1943 regs = task_pt_regs(current);
1949 perf_callchain_store(entry, PERF_CONTEXT_USER);
1950 perf_callchain_user(entry, regs);
1954 put_callchain_entry(rctx);
1960 * Initialize the perf_event context in a task_struct:
1962 static void __perf_event_init_context(struct perf_event_context *ctx)
1964 raw_spin_lock_init(&ctx->lock);
1965 mutex_init(&ctx->mutex);
1966 INIT_LIST_HEAD(&ctx->pinned_groups);
1967 INIT_LIST_HEAD(&ctx->flexible_groups);
1968 INIT_LIST_HEAD(&ctx->event_list);
1969 atomic_set(&ctx->refcount, 1);
1972 static struct perf_event_context *
1973 alloc_perf_context(struct pmu *pmu, struct task_struct *task)
1975 struct perf_event_context *ctx;
1977 ctx = kzalloc(sizeof(struct perf_event_context), GFP_KERNEL);
1981 __perf_event_init_context(ctx);
1984 get_task_struct(task);
1991 static struct perf_event_context *
1992 find_get_context(struct pmu *pmu, pid_t pid, int cpu)
1994 struct perf_event_context *ctx;
1995 struct perf_cpu_context *cpuctx;
1996 struct task_struct *task;
1997 unsigned long flags;
2000 if (pid == -1 && cpu != -1) {
2001 /* Must be root to operate on a CPU event: */
2002 if (perf_paranoid_cpu() && !capable(CAP_SYS_ADMIN))
2003 return ERR_PTR(-EACCES);
2005 if (cpu < 0 || cpu >= nr_cpumask_bits)
2006 return ERR_PTR(-EINVAL);
2009 * We could be clever and allow to attach a event to an
2010 * offline CPU and activate it when the CPU comes up, but
2013 if (!cpu_online(cpu))
2014 return ERR_PTR(-ENODEV);
2016 cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu);
2027 task = find_task_by_vpid(pid);
2029 get_task_struct(task);
2033 return ERR_PTR(-ESRCH);
2036 * Can't attach events to a dying task.
2039 if (task->flags & PF_EXITING)
2042 /* Reuse ptrace permission checks for now. */
2044 if (!ptrace_may_access(task, PTRACE_MODE_READ))
2048 ctx = perf_lock_task_context(task, &flags);
2051 raw_spin_unlock_irqrestore(&ctx->lock, flags);
2055 ctx = alloc_perf_context(pmu, task);
2062 if (cmpxchg(&task->perf_event_ctxp, NULL, ctx)) {
2064 * We raced with some other task; use
2065 * the context they set.
2067 put_task_struct(task);
2073 put_task_struct(task);
2077 put_task_struct(task);
2078 return ERR_PTR(err);
2081 static void perf_event_free_filter(struct perf_event *event);
2083 static void free_event_rcu(struct rcu_head *head)
2085 struct perf_event *event;
2087 event = container_of(head, struct perf_event, rcu_head);
2089 put_pid_ns(event->ns);
2090 perf_event_free_filter(event);
2094 static void perf_pending_sync(struct perf_event *event);
2095 static void perf_buffer_put(struct perf_buffer *buffer);
2097 static void free_event(struct perf_event *event)
2099 perf_pending_sync(event);
2101 if (!event->parent) {
2102 atomic_dec(&nr_events);
2103 if (event->attr.mmap || event->attr.mmap_data)
2104 atomic_dec(&nr_mmap_events);
2105 if (event->attr.comm)
2106 atomic_dec(&nr_comm_events);
2107 if (event->attr.task)
2108 atomic_dec(&nr_task_events);
2109 if (event->attr.sample_type & PERF_SAMPLE_CALLCHAIN)
2110 put_callchain_buffers();
2113 if (event->buffer) {
2114 perf_buffer_put(event->buffer);
2115 event->buffer = NULL;
2119 event->destroy(event);
2121 put_ctx(event->ctx);
2122 call_rcu(&event->rcu_head, free_event_rcu);
2125 int perf_event_release_kernel(struct perf_event *event)
2127 struct perf_event_context *ctx = event->ctx;
2130 * Remove from the PMU, can't get re-enabled since we got
2131 * here because the last ref went.
2133 perf_event_disable(event);
2135 WARN_ON_ONCE(ctx->parent_ctx);
2137 * There are two ways this annotation is useful:
2139 * 1) there is a lock recursion from perf_event_exit_task
2140 * see the comment there.
2142 * 2) there is a lock-inversion with mmap_sem through
2143 * perf_event_read_group(), which takes faults while
2144 * holding ctx->mutex, however this is called after
2145 * the last filedesc died, so there is no possibility
2146 * to trigger the AB-BA case.
2148 mutex_lock_nested(&ctx->mutex, SINGLE_DEPTH_NESTING);
2149 raw_spin_lock_irq(&ctx->lock);
2150 perf_group_detach(event);
2151 list_del_event(event, ctx);
2152 raw_spin_unlock_irq(&ctx->lock);
2153 mutex_unlock(&ctx->mutex);
2155 mutex_lock(&event->owner->perf_event_mutex);
2156 list_del_init(&event->owner_entry);
2157 mutex_unlock(&event->owner->perf_event_mutex);
2158 put_task_struct(event->owner);
2164 EXPORT_SYMBOL_GPL(perf_event_release_kernel);
2167 * Called when the last reference to the file is gone.
2169 static int perf_release(struct inode *inode, struct file *file)
2171 struct perf_event *event = file->private_data;
2173 file->private_data = NULL;
2175 return perf_event_release_kernel(event);
2178 static int perf_event_read_size(struct perf_event *event)
2180 int entry = sizeof(u64); /* value */
2184 if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
2185 size += sizeof(u64);
2187 if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
2188 size += sizeof(u64);
2190 if (event->attr.read_format & PERF_FORMAT_ID)
2191 entry += sizeof(u64);
2193 if (event->attr.read_format & PERF_FORMAT_GROUP) {
2194 nr += event->group_leader->nr_siblings;
2195 size += sizeof(u64);
2203 u64 perf_event_read_value(struct perf_event *event, u64 *enabled, u64 *running)
2205 struct perf_event *child;
2211 mutex_lock(&event->child_mutex);
2212 total += perf_event_read(event);
2213 *enabled += event->total_time_enabled +
2214 atomic64_read(&event->child_total_time_enabled);
2215 *running += event->total_time_running +
2216 atomic64_read(&event->child_total_time_running);
2218 list_for_each_entry(child, &event->child_list, child_list) {
2219 total += perf_event_read(child);
2220 *enabled += child->total_time_enabled;
2221 *running += child->total_time_running;
2223 mutex_unlock(&event->child_mutex);
2227 EXPORT_SYMBOL_GPL(perf_event_read_value);
2229 static int perf_event_read_group(struct perf_event *event,
2230 u64 read_format, char __user *buf)
2232 struct perf_event *leader = event->group_leader, *sub;
2233 int n = 0, size = 0, ret = -EFAULT;
2234 struct perf_event_context *ctx = leader->ctx;
2236 u64 count, enabled, running;
2238 mutex_lock(&ctx->mutex);
2239 count = perf_event_read_value(leader, &enabled, &running);
2241 values[n++] = 1 + leader->nr_siblings;
2242 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
2243 values[n++] = enabled;
2244 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
2245 values[n++] = running;
2246 values[n++] = count;
2247 if (read_format & PERF_FORMAT_ID)
2248 values[n++] = primary_event_id(leader);
2250 size = n * sizeof(u64);
2252 if (copy_to_user(buf, values, size))
2257 list_for_each_entry(sub, &leader->sibling_list, group_entry) {
2260 values[n++] = perf_event_read_value(sub, &enabled, &running);
2261 if (read_format & PERF_FORMAT_ID)
2262 values[n++] = primary_event_id(sub);
2264 size = n * sizeof(u64);
2266 if (copy_to_user(buf + ret, values, size)) {
2274 mutex_unlock(&ctx->mutex);
2279 static int perf_event_read_one(struct perf_event *event,
2280 u64 read_format, char __user *buf)
2282 u64 enabled, running;
2286 values[n++] = perf_event_read_value(event, &enabled, &running);
2287 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
2288 values[n++] = enabled;
2289 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
2290 values[n++] = running;
2291 if (read_format & PERF_FORMAT_ID)
2292 values[n++] = primary_event_id(event);
2294 if (copy_to_user(buf, values, n * sizeof(u64)))
2297 return n * sizeof(u64);
2301 * Read the performance event - simple non blocking version for now
2304 perf_read_hw(struct perf_event *event, char __user *buf, size_t count)
2306 u64 read_format = event->attr.read_format;
2310 * Return end-of-file for a read on a event that is in
2311 * error state (i.e. because it was pinned but it couldn't be
2312 * scheduled on to the CPU at some point).
2314 if (event->state == PERF_EVENT_STATE_ERROR)
2317 if (count < perf_event_read_size(event))
2320 WARN_ON_ONCE(event->ctx->parent_ctx);
2321 if (read_format & PERF_FORMAT_GROUP)
2322 ret = perf_event_read_group(event, read_format, buf);
2324 ret = perf_event_read_one(event, read_format, buf);
2330 perf_read(struct file *file, char __user *buf, size_t count, loff_t *ppos)
2332 struct perf_event *event = file->private_data;
2334 return perf_read_hw(event, buf, count);
2337 static unsigned int perf_poll(struct file *file, poll_table *wait)
2339 struct perf_event *event = file->private_data;
2340 struct perf_buffer *buffer;
2341 unsigned int events = POLL_HUP;
2344 buffer = rcu_dereference(event->buffer);
2346 events = atomic_xchg(&buffer->poll, 0);
2349 poll_wait(file, &event->waitq, wait);
2354 static void perf_event_reset(struct perf_event *event)
2356 (void)perf_event_read(event);
2357 local64_set(&event->count, 0);
2358 perf_event_update_userpage(event);
2362 * Holding the top-level event's child_mutex means that any
2363 * descendant process that has inherited this event will block
2364 * in sync_child_event if it goes to exit, thus satisfying the
2365 * task existence requirements of perf_event_enable/disable.
2367 static void perf_event_for_each_child(struct perf_event *event,
2368 void (*func)(struct perf_event *))
2370 struct perf_event *child;
2372 WARN_ON_ONCE(event->ctx->parent_ctx);
2373 mutex_lock(&event->child_mutex);
2375 list_for_each_entry(child, &event->child_list, child_list)
2377 mutex_unlock(&event->child_mutex);
2380 static void perf_event_for_each(struct perf_event *event,
2381 void (*func)(struct perf_event *))
2383 struct perf_event_context *ctx = event->ctx;
2384 struct perf_event *sibling;
2386 WARN_ON_ONCE(ctx->parent_ctx);
2387 mutex_lock(&ctx->mutex);
2388 event = event->group_leader;
2390 perf_event_for_each_child(event, func);
2392 list_for_each_entry(sibling, &event->sibling_list, group_entry)
2393 perf_event_for_each_child(event, func);
2394 mutex_unlock(&ctx->mutex);
2397 static int perf_event_period(struct perf_event *event, u64 __user *arg)
2399 struct perf_event_context *ctx = event->ctx;
2404 if (!event->attr.sample_period)
2407 size = copy_from_user(&value, arg, sizeof(value));
2408 if (size != sizeof(value))
2414 raw_spin_lock_irq(&ctx->lock);
2415 if (event->attr.freq) {
2416 if (value > sysctl_perf_event_sample_rate) {
2421 event->attr.sample_freq = value;
2423 event->attr.sample_period = value;
2424 event->hw.sample_period = value;
2427 raw_spin_unlock_irq(&ctx->lock);
2432 static const struct file_operations perf_fops;
2434 static struct perf_event *perf_fget_light(int fd, int *fput_needed)
2438 file = fget_light(fd, fput_needed);
2440 return ERR_PTR(-EBADF);
2442 if (file->f_op != &perf_fops) {
2443 fput_light(file, *fput_needed);
2445 return ERR_PTR(-EBADF);
2448 return file->private_data;
2451 static int perf_event_set_output(struct perf_event *event,
2452 struct perf_event *output_event);
2453 static int perf_event_set_filter(struct perf_event *event, void __user *arg);
2455 static long perf_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
2457 struct perf_event *event = file->private_data;
2458 void (*func)(struct perf_event *);
2462 case PERF_EVENT_IOC_ENABLE:
2463 func = perf_event_enable;
2465 case PERF_EVENT_IOC_DISABLE:
2466 func = perf_event_disable;
2468 case PERF_EVENT_IOC_RESET:
2469 func = perf_event_reset;
2472 case PERF_EVENT_IOC_REFRESH:
2473 return perf_event_refresh(event, arg);
2475 case PERF_EVENT_IOC_PERIOD:
2476 return perf_event_period(event, (u64 __user *)arg);
2478 case PERF_EVENT_IOC_SET_OUTPUT:
2480 struct perf_event *output_event = NULL;
2481 int fput_needed = 0;
2485 output_event = perf_fget_light(arg, &fput_needed);
2486 if (IS_ERR(output_event))
2487 return PTR_ERR(output_event);
2490 ret = perf_event_set_output(event, output_event);
2492 fput_light(output_event->filp, fput_needed);
2497 case PERF_EVENT_IOC_SET_FILTER:
2498 return perf_event_set_filter(event, (void __user *)arg);
2504 if (flags & PERF_IOC_FLAG_GROUP)
2505 perf_event_for_each(event, func);
2507 perf_event_for_each_child(event, func);
2512 int perf_event_task_enable(void)
2514 struct perf_event *event;
2516 mutex_lock(¤t->perf_event_mutex);
2517 list_for_each_entry(event, ¤t->perf_event_list, owner_entry)
2518 perf_event_for_each_child(event, perf_event_enable);
2519 mutex_unlock(¤t->perf_event_mutex);
2524 int perf_event_task_disable(void)
2526 struct perf_event *event;
2528 mutex_lock(¤t->perf_event_mutex);
2529 list_for_each_entry(event, ¤t->perf_event_list, owner_entry)
2530 perf_event_for_each_child(event, perf_event_disable);
2531 mutex_unlock(¤t->perf_event_mutex);
2536 #ifndef PERF_EVENT_INDEX_OFFSET
2537 # define PERF_EVENT_INDEX_OFFSET 0
2540 static int perf_event_index(struct perf_event *event)
2542 if (event->hw.state & PERF_HES_STOPPED)
2545 if (event->state != PERF_EVENT_STATE_ACTIVE)
2548 return event->hw.idx + 1 - PERF_EVENT_INDEX_OFFSET;
2552 * Callers need to ensure there can be no nesting of this function, otherwise
2553 * the seqlock logic goes bad. We can not serialize this because the arch
2554 * code calls this from NMI context.
2556 void perf_event_update_userpage(struct perf_event *event)
2558 struct perf_event_mmap_page *userpg;
2559 struct perf_buffer *buffer;
2562 buffer = rcu_dereference(event->buffer);
2566 userpg = buffer->user_page;
2569 * Disable preemption so as to not let the corresponding user-space
2570 * spin too long if we get preempted.
2575 userpg->index = perf_event_index(event);
2576 userpg->offset = perf_event_count(event);
2577 if (event->state == PERF_EVENT_STATE_ACTIVE)
2578 userpg->offset -= local64_read(&event->hw.prev_count);
2580 userpg->time_enabled = event->total_time_enabled +
2581 atomic64_read(&event->child_total_time_enabled);
2583 userpg->time_running = event->total_time_running +
2584 atomic64_read(&event->child_total_time_running);
2593 static unsigned long perf_data_size(struct perf_buffer *buffer);
2596 perf_buffer_init(struct perf_buffer *buffer, long watermark, int flags)
2598 long max_size = perf_data_size(buffer);
2601 buffer->watermark = min(max_size, watermark);
2603 if (!buffer->watermark)
2604 buffer->watermark = max_size / 2;
2606 if (flags & PERF_BUFFER_WRITABLE)
2607 buffer->writable = 1;
2609 atomic_set(&buffer->refcount, 1);
2612 #ifndef CONFIG_PERF_USE_VMALLOC
2615 * Back perf_mmap() with regular GFP_KERNEL-0 pages.
2618 static struct page *
2619 perf_mmap_to_page(struct perf_buffer *buffer, unsigned long pgoff)
2621 if (pgoff > buffer->nr_pages)
2625 return virt_to_page(buffer->user_page);
2627 return virt_to_page(buffer->data_pages[pgoff - 1]);
2630 static void *perf_mmap_alloc_page(int cpu)
2635 node = (cpu == -1) ? cpu : cpu_to_node(cpu);
2636 page = alloc_pages_node(node, GFP_KERNEL | __GFP_ZERO, 0);
2640 return page_address(page);
2643 static struct perf_buffer *
2644 perf_buffer_alloc(int nr_pages, long watermark, int cpu, int flags)
2646 struct perf_buffer *buffer;
2650 size = sizeof(struct perf_buffer);
2651 size += nr_pages * sizeof(void *);
2653 buffer = kzalloc(size, GFP_KERNEL);
2657 buffer->user_page = perf_mmap_alloc_page(cpu);
2658 if (!buffer->user_page)
2659 goto fail_user_page;
2661 for (i = 0; i < nr_pages; i++) {
2662 buffer->data_pages[i] = perf_mmap_alloc_page(cpu);
2663 if (!buffer->data_pages[i])
2664 goto fail_data_pages;
2667 buffer->nr_pages = nr_pages;
2669 perf_buffer_init(buffer, watermark, flags);
2674 for (i--; i >= 0; i--)
2675 free_page((unsigned long)buffer->data_pages[i]);
2677 free_page((unsigned long)buffer->user_page);
2686 static void perf_mmap_free_page(unsigned long addr)
2688 struct page *page = virt_to_page((void *)addr);
2690 page->mapping = NULL;
2694 static void perf_buffer_free(struct perf_buffer *buffer)
2698 perf_mmap_free_page((unsigned long)buffer->user_page);
2699 for (i = 0; i < buffer->nr_pages; i++)
2700 perf_mmap_free_page((unsigned long)buffer->data_pages[i]);
2704 static inline int page_order(struct perf_buffer *buffer)
2712 * Back perf_mmap() with vmalloc memory.
2714 * Required for architectures that have d-cache aliasing issues.
2717 static inline int page_order(struct perf_buffer *buffer)
2719 return buffer->page_order;
2722 static struct page *
2723 perf_mmap_to_page(struct perf_buffer *buffer, unsigned long pgoff)
2725 if (pgoff > (1UL << page_order(buffer)))
2728 return vmalloc_to_page((void *)buffer->user_page + pgoff * PAGE_SIZE);
2731 static void perf_mmap_unmark_page(void *addr)
2733 struct page *page = vmalloc_to_page(addr);
2735 page->mapping = NULL;
2738 static void perf_buffer_free_work(struct work_struct *work)
2740 struct perf_buffer *buffer;
2744 buffer = container_of(work, struct perf_buffer, work);
2745 nr = 1 << page_order(buffer);
2747 base = buffer->user_page;
2748 for (i = 0; i < nr + 1; i++)
2749 perf_mmap_unmark_page(base + (i * PAGE_SIZE));
2755 static void perf_buffer_free(struct perf_buffer *buffer)
2757 schedule_work(&buffer->work);
2760 static struct perf_buffer *
2761 perf_buffer_alloc(int nr_pages, long watermark, int cpu, int flags)
2763 struct perf_buffer *buffer;
2767 size = sizeof(struct perf_buffer);
2768 size += sizeof(void *);
2770 buffer = kzalloc(size, GFP_KERNEL);
2774 INIT_WORK(&buffer->work, perf_buffer_free_work);
2776 all_buf = vmalloc_user((nr_pages + 1) * PAGE_SIZE);
2780 buffer->user_page = all_buf;
2781 buffer->data_pages[0] = all_buf + PAGE_SIZE;
2782 buffer->page_order = ilog2(nr_pages);
2783 buffer->nr_pages = 1;
2785 perf_buffer_init(buffer, watermark, flags);
2798 static unsigned long perf_data_size(struct perf_buffer *buffer)
2800 return buffer->nr_pages << (PAGE_SHIFT + page_order(buffer));
2803 static int perf_mmap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
2805 struct perf_event *event = vma->vm_file->private_data;
2806 struct perf_buffer *buffer;
2807 int ret = VM_FAULT_SIGBUS;
2809 if (vmf->flags & FAULT_FLAG_MKWRITE) {
2810 if (vmf->pgoff == 0)
2816 buffer = rcu_dereference(event->buffer);
2820 if (vmf->pgoff && (vmf->flags & FAULT_FLAG_WRITE))
2823 vmf->page = perf_mmap_to_page(buffer, vmf->pgoff);
2827 get_page(vmf->page);
2828 vmf->page->mapping = vma->vm_file->f_mapping;
2829 vmf->page->index = vmf->pgoff;
2838 static void perf_buffer_free_rcu(struct rcu_head *rcu_head)
2840 struct perf_buffer *buffer;
2842 buffer = container_of(rcu_head, struct perf_buffer, rcu_head);
2843 perf_buffer_free(buffer);
2846 static struct perf_buffer *perf_buffer_get(struct perf_event *event)
2848 struct perf_buffer *buffer;
2851 buffer = rcu_dereference(event->buffer);
2853 if (!atomic_inc_not_zero(&buffer->refcount))
2861 static void perf_buffer_put(struct perf_buffer *buffer)
2863 if (!atomic_dec_and_test(&buffer->refcount))
2866 call_rcu(&buffer->rcu_head, perf_buffer_free_rcu);
2869 static void perf_mmap_open(struct vm_area_struct *vma)
2871 struct perf_event *event = vma->vm_file->private_data;
2873 atomic_inc(&event->mmap_count);
2876 static void perf_mmap_close(struct vm_area_struct *vma)
2878 struct perf_event *event = vma->vm_file->private_data;
2880 if (atomic_dec_and_mutex_lock(&event->mmap_count, &event->mmap_mutex)) {
2881 unsigned long size = perf_data_size(event->buffer);
2882 struct user_struct *user = event->mmap_user;
2883 struct perf_buffer *buffer = event->buffer;
2885 atomic_long_sub((size >> PAGE_SHIFT) + 1, &user->locked_vm);
2886 vma->vm_mm->locked_vm -= event->mmap_locked;
2887 rcu_assign_pointer(event->buffer, NULL);
2888 mutex_unlock(&event->mmap_mutex);
2890 perf_buffer_put(buffer);
2895 static const struct vm_operations_struct perf_mmap_vmops = {
2896 .open = perf_mmap_open,
2897 .close = perf_mmap_close,
2898 .fault = perf_mmap_fault,
2899 .page_mkwrite = perf_mmap_fault,
2902 static int perf_mmap(struct file *file, struct vm_area_struct *vma)
2904 struct perf_event *event = file->private_data;
2905 unsigned long user_locked, user_lock_limit;
2906 struct user_struct *user = current_user();
2907 unsigned long locked, lock_limit;
2908 struct perf_buffer *buffer;
2909 unsigned long vma_size;
2910 unsigned long nr_pages;
2911 long user_extra, extra;
2912 int ret = 0, flags = 0;
2915 * Don't allow mmap() of inherited per-task counters. This would
2916 * create a performance issue due to all children writing to the
2919 if (event->cpu == -1 && event->attr.inherit)
2922 if (!(vma->vm_flags & VM_SHARED))
2925 vma_size = vma->vm_end - vma->vm_start;
2926 nr_pages = (vma_size / PAGE_SIZE) - 1;
2929 * If we have buffer pages ensure they're a power-of-two number, so we
2930 * can do bitmasks instead of modulo.
2932 if (nr_pages != 0 && !is_power_of_2(nr_pages))
2935 if (vma_size != PAGE_SIZE * (1 + nr_pages))
2938 if (vma->vm_pgoff != 0)
2941 WARN_ON_ONCE(event->ctx->parent_ctx);
2942 mutex_lock(&event->mmap_mutex);
2943 if (event->buffer) {
2944 if (event->buffer->nr_pages == nr_pages)
2945 atomic_inc(&event->buffer->refcount);
2951 user_extra = nr_pages + 1;
2952 user_lock_limit = sysctl_perf_event_mlock >> (PAGE_SHIFT - 10);
2955 * Increase the limit linearly with more CPUs:
2957 user_lock_limit *= num_online_cpus();
2959 user_locked = atomic_long_read(&user->locked_vm) + user_extra;
2962 if (user_locked > user_lock_limit)
2963 extra = user_locked - user_lock_limit;
2965 lock_limit = rlimit(RLIMIT_MEMLOCK);
2966 lock_limit >>= PAGE_SHIFT;
2967 locked = vma->vm_mm->locked_vm + extra;
2969 if ((locked > lock_limit) && perf_paranoid_tracepoint_raw() &&
2970 !capable(CAP_IPC_LOCK)) {
2975 WARN_ON(event->buffer);
2977 if (vma->vm_flags & VM_WRITE)
2978 flags |= PERF_BUFFER_WRITABLE;
2980 buffer = perf_buffer_alloc(nr_pages, event->attr.wakeup_watermark,
2986 rcu_assign_pointer(event->buffer, buffer);
2988 atomic_long_add(user_extra, &user->locked_vm);
2989 event->mmap_locked = extra;
2990 event->mmap_user = get_current_user();
2991 vma->vm_mm->locked_vm += event->mmap_locked;
2995 atomic_inc(&event->mmap_count);
2996 mutex_unlock(&event->mmap_mutex);
2998 vma->vm_flags |= VM_RESERVED;
2999 vma->vm_ops = &perf_mmap_vmops;
3004 static int perf_fasync(int fd, struct file *filp, int on)
3006 struct inode *inode = filp->f_path.dentry->d_inode;
3007 struct perf_event *event = filp->private_data;
3010 mutex_lock(&inode->i_mutex);
3011 retval = fasync_helper(fd, filp, on, &event->fasync);
3012 mutex_unlock(&inode->i_mutex);
3020 static const struct file_operations perf_fops = {
3021 .llseek = no_llseek,
3022 .release = perf_release,
3025 .unlocked_ioctl = perf_ioctl,
3026 .compat_ioctl = perf_ioctl,
3028 .fasync = perf_fasync,
3034 * If there's data, ensure we set the poll() state and publish everything
3035 * to user-space before waking everybody up.
3038 void perf_event_wakeup(struct perf_event *event)
3040 wake_up_all(&event->waitq);
3042 if (event->pending_kill) {
3043 kill_fasync(&event->fasync, SIGIO, event->pending_kill);
3044 event->pending_kill = 0;
3051 * Handle the case where we need to wakeup up from NMI (or rq->lock) context.
3053 * The NMI bit means we cannot possibly take locks. Therefore, maintain a
3054 * single linked list and use cmpxchg() to add entries lockless.
3057 static void perf_pending_event(struct perf_pending_entry *entry)
3059 struct perf_event *event = container_of(entry,
3060 struct perf_event, pending);
3062 if (event->pending_disable) {
3063 event->pending_disable = 0;
3064 __perf_event_disable(event);
3067 if (event->pending_wakeup) {
3068 event->pending_wakeup = 0;
3069 perf_event_wakeup(event);
3073 #define PENDING_TAIL ((struct perf_pending_entry *)-1UL)
3075 static DEFINE_PER_CPU(struct perf_pending_entry *, perf_pending_head) = {
3079 static void perf_pending_queue(struct perf_pending_entry *entry,
3080 void (*func)(struct perf_pending_entry *))
3082 struct perf_pending_entry **head;
3084 if (cmpxchg(&entry->next, NULL, PENDING_TAIL) != NULL)
3089 head = &get_cpu_var(perf_pending_head);
3092 entry->next = *head;
3093 } while (cmpxchg(head, entry->next, entry) != entry->next);
3095 set_perf_event_pending();
3097 put_cpu_var(perf_pending_head);
3100 static int __perf_pending_run(void)
3102 struct perf_pending_entry *list;
3105 list = xchg(&__get_cpu_var(perf_pending_head), PENDING_TAIL);
3106 while (list != PENDING_TAIL) {
3107 void (*func)(struct perf_pending_entry *);
3108 struct perf_pending_entry *entry = list;
3115 * Ensure we observe the unqueue before we issue the wakeup,
3116 * so that we won't be waiting forever.
3117 * -- see perf_not_pending().
3128 static inline int perf_not_pending(struct perf_event *event)
3131 * If we flush on whatever cpu we run, there is a chance we don't
3135 __perf_pending_run();
3139 * Ensure we see the proper queue state before going to sleep
3140 * so that we do not miss the wakeup. -- see perf_pending_handle()
3143 return event->pending.next == NULL;
3146 static void perf_pending_sync(struct perf_event *event)
3148 wait_event(event->waitq, perf_not_pending(event));
3151 void perf_event_do_pending(void)
3153 __perf_pending_run();
3157 * We assume there is only KVM supporting the callbacks.
3158 * Later on, we might change it to a list if there is
3159 * another virtualization implementation supporting the callbacks.
3161 struct perf_guest_info_callbacks *perf_guest_cbs;
3163 int perf_register_guest_info_callbacks(struct perf_guest_info_callbacks *cbs)
3165 perf_guest_cbs = cbs;
3168 EXPORT_SYMBOL_GPL(perf_register_guest_info_callbacks);
3170 int perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks *cbs)
3172 perf_guest_cbs = NULL;
3175 EXPORT_SYMBOL_GPL(perf_unregister_guest_info_callbacks);
3180 static bool perf_output_space(struct perf_buffer *buffer, unsigned long tail,
3181 unsigned long offset, unsigned long head)
3185 if (!buffer->writable)
3188 mask = perf_data_size(buffer) - 1;
3190 offset = (offset - tail) & mask;
3191 head = (head - tail) & mask;
3193 if ((int)(head - offset) < 0)
3199 static void perf_output_wakeup(struct perf_output_handle *handle)
3201 atomic_set(&handle->buffer->poll, POLL_IN);
3204 handle->event->pending_wakeup = 1;
3205 perf_pending_queue(&handle->event->pending,
3206 perf_pending_event);
3208 perf_event_wakeup(handle->event);
3212 * We need to ensure a later event_id doesn't publish a head when a former
3213 * event isn't done writing. However since we need to deal with NMIs we
3214 * cannot fully serialize things.
3216 * We only publish the head (and generate a wakeup) when the outer-most
3219 static void perf_output_get_handle(struct perf_output_handle *handle)
3221 struct perf_buffer *buffer = handle->buffer;
3224 local_inc(&buffer->nest);
3225 handle->wakeup = local_read(&buffer->wakeup);
3228 static void perf_output_put_handle(struct perf_output_handle *handle)
3230 struct perf_buffer *buffer = handle->buffer;
3234 head = local_read(&buffer->head);
3237 * IRQ/NMI can happen here, which means we can miss a head update.
3240 if (!local_dec_and_test(&buffer->nest))
3244 * Publish the known good head. Rely on the full barrier implied
3245 * by atomic_dec_and_test() order the buffer->head read and this
3248 buffer->user_page->data_head = head;
3251 * Now check if we missed an update, rely on the (compiler)
3252 * barrier in atomic_dec_and_test() to re-read buffer->head.
3254 if (unlikely(head != local_read(&buffer->head))) {
3255 local_inc(&buffer->nest);
3259 if (handle->wakeup != local_read(&buffer->wakeup))
3260 perf_output_wakeup(handle);
3266 __always_inline void perf_output_copy(struct perf_output_handle *handle,
3267 const void *buf, unsigned int len)
3270 unsigned long size = min_t(unsigned long, handle->size, len);
3272 memcpy(handle->addr, buf, size);
3275 handle->addr += size;
3277 handle->size -= size;
3278 if (!handle->size) {
3279 struct perf_buffer *buffer = handle->buffer;
3282 handle->page &= buffer->nr_pages - 1;
3283 handle->addr = buffer->data_pages[handle->page];
3284 handle->size = PAGE_SIZE << page_order(buffer);
3289 int perf_output_begin(struct perf_output_handle *handle,
3290 struct perf_event *event, unsigned int size,
3291 int nmi, int sample)
3293 struct perf_buffer *buffer;
3294 unsigned long tail, offset, head;
3297 struct perf_event_header header;
3304 * For inherited events we send all the output towards the parent.
3307 event = event->parent;
3309 buffer = rcu_dereference(event->buffer);
3313 handle->buffer = buffer;
3314 handle->event = event;
3316 handle->sample = sample;
3318 if (!buffer->nr_pages)
3321 have_lost = local_read(&buffer->lost);
3323 size += sizeof(lost_event);
3325 perf_output_get_handle(handle);
3329 * Userspace could choose to issue a mb() before updating the
3330 * tail pointer. So that all reads will be completed before the
3333 tail = ACCESS_ONCE(buffer->user_page->data_tail);
3335 offset = head = local_read(&buffer->head);
3337 if (unlikely(!perf_output_space(buffer, tail, offset, head)))
3339 } while (local_cmpxchg(&buffer->head, offset, head) != offset);
3341 if (head - local_read(&buffer->wakeup) > buffer->watermark)
3342 local_add(buffer->watermark, &buffer->wakeup);
3344 handle->page = offset >> (PAGE_SHIFT + page_order(buffer));
3345 handle->page &= buffer->nr_pages - 1;
3346 handle->size = offset & ((PAGE_SIZE << page_order(buffer)) - 1);
3347 handle->addr = buffer->data_pages[handle->page];
3348 handle->addr += handle->size;
3349 handle->size = (PAGE_SIZE << page_order(buffer)) - handle->size;
3352 lost_event.header.type = PERF_RECORD_LOST;
3353 lost_event.header.misc = 0;
3354 lost_event.header.size = sizeof(lost_event);
3355 lost_event.id = event->id;
3356 lost_event.lost = local_xchg(&buffer->lost, 0);
3358 perf_output_put(handle, lost_event);
3364 local_inc(&buffer->lost);
3365 perf_output_put_handle(handle);
3372 void perf_output_end(struct perf_output_handle *handle)
3374 struct perf_event *event = handle->event;
3375 struct perf_buffer *buffer = handle->buffer;
3377 int wakeup_events = event->attr.wakeup_events;
3379 if (handle->sample && wakeup_events) {
3380 int events = local_inc_return(&buffer->events);
3381 if (events >= wakeup_events) {
3382 local_sub(wakeup_events, &buffer->events);
3383 local_inc(&buffer->wakeup);
3387 perf_output_put_handle(handle);
3391 static u32 perf_event_pid(struct perf_event *event, struct task_struct *p)
3394 * only top level events have the pid namespace they were created in
3397 event = event->parent;
3399 return task_tgid_nr_ns(p, event->ns);
3402 static u32 perf_event_tid(struct perf_event *event, struct task_struct *p)
3405 * only top level events have the pid namespace they were created in
3408 event = event->parent;
3410 return task_pid_nr_ns(p, event->ns);
3413 static void perf_output_read_one(struct perf_output_handle *handle,
3414 struct perf_event *event)
3416 u64 read_format = event->attr.read_format;
3420 values[n++] = perf_event_count(event);
3421 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
3422 values[n++] = event->total_time_enabled +
3423 atomic64_read(&event->child_total_time_enabled);
3425 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
3426 values[n++] = event->total_time_running +
3427 atomic64_read(&event->child_total_time_running);
3429 if (read_format & PERF_FORMAT_ID)
3430 values[n++] = primary_event_id(event);
3432 perf_output_copy(handle, values, n * sizeof(u64));
3436 * XXX PERF_FORMAT_GROUP vs inherited events seems difficult.
3438 static void perf_output_read_group(struct perf_output_handle *handle,
3439 struct perf_event *event)
3441 struct perf_event *leader = event->group_leader, *sub;
3442 u64 read_format = event->attr.read_format;
3446 values[n++] = 1 + leader->nr_siblings;
3448 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
3449 values[n++] = leader->total_time_enabled;
3451 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
3452 values[n++] = leader->total_time_running;
3454 if (leader != event)
3455 leader->pmu->read(leader);
3457 values[n++] = perf_event_count(leader);
3458 if (read_format & PERF_FORMAT_ID)
3459 values[n++] = primary_event_id(leader);
3461 perf_output_copy(handle, values, n * sizeof(u64));
3463 list_for_each_entry(sub, &leader->sibling_list, group_entry) {
3467 sub->pmu->read(sub);
3469 values[n++] = perf_event_count(sub);
3470 if (read_format & PERF_FORMAT_ID)
3471 values[n++] = primary_event_id(sub);
3473 perf_output_copy(handle, values, n * sizeof(u64));
3477 static void perf_output_read(struct perf_output_handle *handle,
3478 struct perf_event *event)
3480 if (event->attr.read_format & PERF_FORMAT_GROUP)
3481 perf_output_read_group(handle, event);
3483 perf_output_read_one(handle, event);
3486 void perf_output_sample(struct perf_output_handle *handle,
3487 struct perf_event_header *header,
3488 struct perf_sample_data *data,
3489 struct perf_event *event)
3491 u64 sample_type = data->type;
3493 perf_output_put(handle, *header);
3495 if (sample_type & PERF_SAMPLE_IP)
3496 perf_output_put(handle, data->ip);
3498 if (sample_type & PERF_SAMPLE_TID)
3499 perf_output_put(handle, data->tid_entry);
3501 if (sample_type & PERF_SAMPLE_TIME)
3502 perf_output_put(handle, data->time);
3504 if (sample_type & PERF_SAMPLE_ADDR)
3505 perf_output_put(handle, data->addr);
3507 if (sample_type & PERF_SAMPLE_ID)
3508 perf_output_put(handle, data->id);
3510 if (sample_type & PERF_SAMPLE_STREAM_ID)
3511 perf_output_put(handle, data->stream_id);
3513 if (sample_type & PERF_SAMPLE_CPU)
3514 perf_output_put(handle, data->cpu_entry);
3516 if (sample_type & PERF_SAMPLE_PERIOD)
3517 perf_output_put(handle, data->period);
3519 if (sample_type & PERF_SAMPLE_READ)
3520 perf_output_read(handle, event);
3522 if (sample_type & PERF_SAMPLE_CALLCHAIN) {
3523 if (data->callchain) {
3526 if (data->callchain)
3527 size += data->callchain->nr;
3529 size *= sizeof(u64);
3531 perf_output_copy(handle, data->callchain, size);
3534 perf_output_put(handle, nr);
3538 if (sample_type & PERF_SAMPLE_RAW) {
3540 perf_output_put(handle, data->raw->size);
3541 perf_output_copy(handle, data->raw->data,
3548 .size = sizeof(u32),
3551 perf_output_put(handle, raw);
3556 void perf_prepare_sample(struct perf_event_header *header,
3557 struct perf_sample_data *data,
3558 struct perf_event *event,
3559 struct pt_regs *regs)
3561 u64 sample_type = event->attr.sample_type;
3563 data->type = sample_type;
3565 header->type = PERF_RECORD_SAMPLE;
3566 header->size = sizeof(*header);
3569 header->misc |= perf_misc_flags(regs);
3571 if (sample_type & PERF_SAMPLE_IP) {
3572 data->ip = perf_instruction_pointer(regs);
3574 header->size += sizeof(data->ip);
3577 if (sample_type & PERF_SAMPLE_TID) {
3578 /* namespace issues */
3579 data->tid_entry.pid = perf_event_pid(event, current);
3580 data->tid_entry.tid = perf_event_tid(event, current);
3582 header->size += sizeof(data->tid_entry);
3585 if (sample_type & PERF_SAMPLE_TIME) {
3586 data->time = perf_clock();
3588 header->size += sizeof(data->time);
3591 if (sample_type & PERF_SAMPLE_ADDR)
3592 header->size += sizeof(data->addr);
3594 if (sample_type & PERF_SAMPLE_ID) {
3595 data->id = primary_event_id(event);
3597 header->size += sizeof(data->id);
3600 if (sample_type & PERF_SAMPLE_STREAM_ID) {
3601 data->stream_id = event->id;
3603 header->size += sizeof(data->stream_id);
3606 if (sample_type & PERF_SAMPLE_CPU) {
3607 data->cpu_entry.cpu = raw_smp_processor_id();
3608 data->cpu_entry.reserved = 0;
3610 header->size += sizeof(data->cpu_entry);
3613 if (sample_type & PERF_SAMPLE_PERIOD)
3614 header->size += sizeof(data->period);
3616 if (sample_type & PERF_SAMPLE_READ)
3617 header->size += perf_event_read_size(event);
3619 if (sample_type & PERF_SAMPLE_CALLCHAIN) {
3622 data->callchain = perf_callchain(regs);
3624 if (data->callchain)
3625 size += data->callchain->nr;
3627 header->size += size * sizeof(u64);
3630 if (sample_type & PERF_SAMPLE_RAW) {
3631 int size = sizeof(u32);
3634 size += data->raw->size;
3636 size += sizeof(u32);
3638 WARN_ON_ONCE(size & (sizeof(u64)-1));
3639 header->size += size;
3643 static void perf_event_output(struct perf_event *event, int nmi,
3644 struct perf_sample_data *data,
3645 struct pt_regs *regs)
3647 struct perf_output_handle handle;
3648 struct perf_event_header header;
3650 /* protect the callchain buffers */
3653 perf_prepare_sample(&header, data, event, regs);
3655 if (perf_output_begin(&handle, event, header.size, nmi, 1))
3658 perf_output_sample(&handle, &header, data, event);
3660 perf_output_end(&handle);
3670 struct perf_read_event {
3671 struct perf_event_header header;
3678 perf_event_read_event(struct perf_event *event,
3679 struct task_struct *task)
3681 struct perf_output_handle handle;
3682 struct perf_read_event read_event = {
3684 .type = PERF_RECORD_READ,
3686 .size = sizeof(read_event) + perf_event_read_size(event),
3688 .pid = perf_event_pid(event, task),
3689 .tid = perf_event_tid(event, task),
3693 ret = perf_output_begin(&handle, event, read_event.header.size, 0, 0);
3697 perf_output_put(&handle, read_event);
3698 perf_output_read(&handle, event);
3700 perf_output_end(&handle);
3704 * task tracking -- fork/exit
3706 * enabled by: attr.comm | attr.mmap | attr.mmap_data | attr.task
3709 struct perf_task_event {
3710 struct task_struct *task;
3711 struct perf_event_context *task_ctx;
3714 struct perf_event_header header;
3724 static void perf_event_task_output(struct perf_event *event,
3725 struct perf_task_event *task_event)
3727 struct perf_output_handle handle;
3728 struct task_struct *task = task_event->task;
3731 size = task_event->event_id.header.size;
3732 ret = perf_output_begin(&handle, event, size, 0, 0);
3737 task_event->event_id.pid = perf_event_pid(event, task);
3738 task_event->event_id.ppid = perf_event_pid(event, current);
3740 task_event->event_id.tid = perf_event_tid(event, task);
3741 task_event->event_id.ptid = perf_event_tid(event, current);
3743 perf_output_put(&handle, task_event->event_id);
3745 perf_output_end(&handle);
3748 static int perf_event_task_match(struct perf_event *event)
3750 if (event->state < PERF_EVENT_STATE_INACTIVE)
3753 if (event->cpu != -1 && event->cpu != smp_processor_id())
3756 if (event->attr.comm || event->attr.mmap ||
3757 event->attr.mmap_data || event->attr.task)
3763 static void perf_event_task_ctx(struct perf_event_context *ctx,
3764 struct perf_task_event *task_event)
3766 struct perf_event *event;
3768 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
3769 if (perf_event_task_match(event))
3770 perf_event_task_output(event, task_event);
3774 static void perf_event_task_event(struct perf_task_event *task_event)
3776 struct perf_event_context *ctx = task_event->task_ctx;
3777 struct perf_cpu_context *cpuctx;
3780 rcu_read_lock_sched();
3781 list_for_each_entry_rcu(pmu, &pmus, entry) {
3782 cpuctx = this_cpu_ptr(pmu->pmu_cpu_context);
3783 perf_event_task_ctx(&cpuctx->ctx, task_event);
3786 ctx = rcu_dereference(current->perf_event_ctxp);
3788 perf_event_task_ctx(ctx, task_event);
3789 rcu_read_unlock_sched();
3792 static void perf_event_task(struct task_struct *task,
3793 struct perf_event_context *task_ctx,
3796 struct perf_task_event task_event;
3798 if (!atomic_read(&nr_comm_events) &&
3799 !atomic_read(&nr_mmap_events) &&
3800 !atomic_read(&nr_task_events))
3803 task_event = (struct perf_task_event){
3805 .task_ctx = task_ctx,
3808 .type = new ? PERF_RECORD_FORK : PERF_RECORD_EXIT,
3810 .size = sizeof(task_event.event_id),
3816 .time = perf_clock(),
3820 perf_event_task_event(&task_event);
3823 void perf_event_fork(struct task_struct *task)
3825 perf_event_task(task, NULL, 1);
3832 struct perf_comm_event {
3833 struct task_struct *task;
3838 struct perf_event_header header;
3845 static void perf_event_comm_output(struct perf_event *event,
3846 struct perf_comm_event *comm_event)
3848 struct perf_output_handle handle;
3849 int size = comm_event->event_id.header.size;
3850 int ret = perf_output_begin(&handle, event, size, 0, 0);
3855 comm_event->event_id.pid = perf_event_pid(event, comm_event->task);
3856 comm_event->event_id.tid = perf_event_tid(event, comm_event->task);
3858 perf_output_put(&handle, comm_event->event_id);
3859 perf_output_copy(&handle, comm_event->comm,
3860 comm_event->comm_size);
3861 perf_output_end(&handle);
3864 static int perf_event_comm_match(struct perf_event *event)
3866 if (event->state < PERF_EVENT_STATE_INACTIVE)
3869 if (event->cpu != -1 && event->cpu != smp_processor_id())
3872 if (event->attr.comm)
3878 static void perf_event_comm_ctx(struct perf_event_context *ctx,
3879 struct perf_comm_event *comm_event)
3881 struct perf_event *event;
3883 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
3884 if (perf_event_comm_match(event))
3885 perf_event_comm_output(event, comm_event);
3889 static void perf_event_comm_event(struct perf_comm_event *comm_event)
3891 struct perf_cpu_context *cpuctx;
3892 struct perf_event_context *ctx;
3895 char comm[TASK_COMM_LEN];
3897 memset(comm, 0, sizeof(comm));
3898 strlcpy(comm, comm_event->task->comm, sizeof(comm));
3899 size = ALIGN(strlen(comm)+1, sizeof(u64));
3901 comm_event->comm = comm;
3902 comm_event->comm_size = size;
3904 comm_event->event_id.header.size = sizeof(comm_event->event_id) + size;
3906 rcu_read_lock_sched();
3907 list_for_each_entry_rcu(pmu, &pmus, entry) {
3908 cpuctx = this_cpu_ptr(pmu->pmu_cpu_context);
3909 perf_event_comm_ctx(&cpuctx->ctx, comm_event);
3911 ctx = rcu_dereference(current->perf_event_ctxp);
3913 perf_event_comm_ctx(ctx, comm_event);
3914 rcu_read_unlock_sched();
3917 void perf_event_comm(struct task_struct *task)
3919 struct perf_comm_event comm_event;
3921 if (task->perf_event_ctxp)
3922 perf_event_enable_on_exec(task);
3924 if (!atomic_read(&nr_comm_events))
3927 comm_event = (struct perf_comm_event){
3933 .type = PERF_RECORD_COMM,
3942 perf_event_comm_event(&comm_event);
3949 struct perf_mmap_event {
3950 struct vm_area_struct *vma;
3952 const char *file_name;
3956 struct perf_event_header header;
3966 static void perf_event_mmap_output(struct perf_event *event,
3967 struct perf_mmap_event *mmap_event)
3969 struct perf_output_handle handle;
3970 int size = mmap_event->event_id.header.size;
3971 int ret = perf_output_begin(&handle, event, size, 0, 0);
3976 mmap_event->event_id.pid = perf_event_pid(event, current);
3977 mmap_event->event_id.tid = perf_event_tid(event, current);
3979 perf_output_put(&handle, mmap_event->event_id);
3980 perf_output_copy(&handle, mmap_event->file_name,
3981 mmap_event->file_size);
3982 perf_output_end(&handle);
3985 static int perf_event_mmap_match(struct perf_event *event,
3986 struct perf_mmap_event *mmap_event,
3989 if (event->state < PERF_EVENT_STATE_INACTIVE)
3992 if (event->cpu != -1 && event->cpu != smp_processor_id())
3995 if ((!executable && event->attr.mmap_data) ||
3996 (executable && event->attr.mmap))
4002 static void perf_event_mmap_ctx(struct perf_event_context *ctx,
4003 struct perf_mmap_event *mmap_event,
4006 struct perf_event *event;
4008 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
4009 if (perf_event_mmap_match(event, mmap_event, executable))
4010 perf_event_mmap_output(event, mmap_event);
4014 static void perf_event_mmap_event(struct perf_mmap_event *mmap_event)
4016 struct perf_cpu_context *cpuctx;
4017 struct perf_event_context *ctx;
4018 struct vm_area_struct *vma = mmap_event->vma;
4019 struct file *file = vma->vm_file;
4026 memset(tmp, 0, sizeof(tmp));
4030 * d_path works from the end of the buffer backwards, so we
4031 * need to add enough zero bytes after the string to handle
4032 * the 64bit alignment we do later.
4034 buf = kzalloc(PATH_MAX + sizeof(u64), GFP_KERNEL);
4036 name = strncpy(tmp, "//enomem", sizeof(tmp));
4039 name = d_path(&file->f_path, buf, PATH_MAX);
4041 name = strncpy(tmp, "//toolong", sizeof(tmp));
4045 if (arch_vma_name(mmap_event->vma)) {
4046 name = strncpy(tmp, arch_vma_name(mmap_event->vma),
4052 name = strncpy(tmp, "[vdso]", sizeof(tmp));
4054 } else if (vma->vm_start <= vma->vm_mm->start_brk &&
4055 vma->vm_end >= vma->vm_mm->brk) {
4056 name = strncpy(tmp, "[heap]", sizeof(tmp));
4058 } else if (vma->vm_start <= vma->vm_mm->start_stack &&
4059 vma->vm_end >= vma->vm_mm->start_stack) {
4060 name = strncpy(tmp, "[stack]", sizeof(tmp));
4064 name = strncpy(tmp, "//anon", sizeof(tmp));
4069 size = ALIGN(strlen(name)+1, sizeof(u64));
4071 mmap_event->file_name = name;
4072 mmap_event->file_size = size;
4074 mmap_event->event_id.header.size = sizeof(mmap_event->event_id) + size;
4076 rcu_read_lock_sched();
4077 list_for_each_entry_rcu(pmu, &pmus, entry) {
4078 cpuctx = this_cpu_ptr(pmu->pmu_cpu_context);
4079 perf_event_mmap_ctx(&cpuctx->ctx, mmap_event,
4080 vma->vm_flags & VM_EXEC);
4082 ctx = rcu_dereference(current->perf_event_ctxp);
4084 perf_event_mmap_ctx(ctx, mmap_event, vma->vm_flags & VM_EXEC);
4085 rcu_read_unlock_sched();
4090 void perf_event_mmap(struct vm_area_struct *vma)
4092 struct perf_mmap_event mmap_event;
4094 if (!atomic_read(&nr_mmap_events))
4097 mmap_event = (struct perf_mmap_event){
4103 .type = PERF_RECORD_MMAP,
4104 .misc = PERF_RECORD_MISC_USER,
4109 .start = vma->vm_start,
4110 .len = vma->vm_end - vma->vm_start,
4111 .pgoff = (u64)vma->vm_pgoff << PAGE_SHIFT,
4115 perf_event_mmap_event(&mmap_event);
4119 * IRQ throttle logging
4122 static void perf_log_throttle(struct perf_event *event, int enable)
4124 struct perf_output_handle handle;
4128 struct perf_event_header header;
4132 } throttle_event = {
4134 .type = PERF_RECORD_THROTTLE,
4136 .size = sizeof(throttle_event),
4138 .time = perf_clock(),
4139 .id = primary_event_id(event),
4140 .stream_id = event->id,
4144 throttle_event.header.type = PERF_RECORD_UNTHROTTLE;
4146 ret = perf_output_begin(&handle, event, sizeof(throttle_event), 1, 0);
4150 perf_output_put(&handle, throttle_event);
4151 perf_output_end(&handle);
4155 * Generic event overflow handling, sampling.
4158 static int __perf_event_overflow(struct perf_event *event, int nmi,
4159 int throttle, struct perf_sample_data *data,
4160 struct pt_regs *regs)
4162 int events = atomic_read(&event->event_limit);
4163 struct hw_perf_event *hwc = &event->hw;
4169 if (hwc->interrupts != MAX_INTERRUPTS) {
4171 if (HZ * hwc->interrupts >
4172 (u64)sysctl_perf_event_sample_rate) {
4173 hwc->interrupts = MAX_INTERRUPTS;
4174 perf_log_throttle(event, 0);
4179 * Keep re-disabling events even though on the previous
4180 * pass we disabled it - just in case we raced with a
4181 * sched-in and the event got enabled again:
4187 if (event->attr.freq) {
4188 u64 now = perf_clock();
4189 s64 delta = now - hwc->freq_time_stamp;
4191 hwc->freq_time_stamp = now;
4193 if (delta > 0 && delta < 2*TICK_NSEC)
4194 perf_adjust_period(event, delta, hwc->last_period);
4198 * XXX event_limit might not quite work as expected on inherited
4202 event->pending_kill = POLL_IN;
4203 if (events && atomic_dec_and_test(&event->event_limit)) {
4205 event->pending_kill = POLL_HUP;
4207 event->pending_disable = 1;
4208 perf_pending_queue(&event->pending,
4209 perf_pending_event);
4211 perf_event_disable(event);
4214 if (event->overflow_handler)
4215 event->overflow_handler(event, nmi, data, regs);
4217 perf_event_output(event, nmi, data, regs);
4222 int perf_event_overflow(struct perf_event *event, int nmi,
4223 struct perf_sample_data *data,
4224 struct pt_regs *regs)
4226 return __perf_event_overflow(event, nmi, 1, data, regs);
4230 * Generic software event infrastructure
4233 struct swevent_htable {
4234 struct swevent_hlist *swevent_hlist;
4235 struct mutex hlist_mutex;
4238 /* Recursion avoidance in each contexts */
4239 int recursion[PERF_NR_CONTEXTS];
4242 static DEFINE_PER_CPU(struct swevent_htable, swevent_htable);
4245 * We directly increment event->count and keep a second value in
4246 * event->hw.period_left to count intervals. This period event
4247 * is kept in the range [-sample_period, 0] so that we can use the
4251 static u64 perf_swevent_set_period(struct perf_event *event)
4253 struct hw_perf_event *hwc = &event->hw;
4254 u64 period = hwc->last_period;
4258 hwc->last_period = hwc->sample_period;
4261 old = val = local64_read(&hwc->period_left);
4265 nr = div64_u64(period + val, period);
4266 offset = nr * period;
4268 if (local64_cmpxchg(&hwc->period_left, old, val) != old)
4274 static void perf_swevent_overflow(struct perf_event *event, u64 overflow,
4275 int nmi, struct perf_sample_data *data,
4276 struct pt_regs *regs)
4278 struct hw_perf_event *hwc = &event->hw;
4281 data->period = event->hw.last_period;
4283 overflow = perf_swevent_set_period(event);
4285 if (hwc->interrupts == MAX_INTERRUPTS)
4288 for (; overflow; overflow--) {
4289 if (__perf_event_overflow(event, nmi, throttle,
4292 * We inhibit the overflow from happening when
4293 * hwc->interrupts == MAX_INTERRUPTS.
4301 static void perf_swevent_event(struct perf_event *event, u64 nr,
4302 int nmi, struct perf_sample_data *data,
4303 struct pt_regs *regs)
4305 struct hw_perf_event *hwc = &event->hw;
4307 local64_add(nr, &event->count);
4312 if (!hwc->sample_period)
4315 if (nr == 1 && hwc->sample_period == 1 && !event->attr.freq)
4316 return perf_swevent_overflow(event, 1, nmi, data, regs);
4318 if (local64_add_negative(nr, &hwc->period_left))
4321 perf_swevent_overflow(event, 0, nmi, data, regs);
4324 static int perf_exclude_event(struct perf_event *event,
4325 struct pt_regs *regs)
4327 if (event->hw.state & PERF_HES_STOPPED)
4331 if (event->attr.exclude_user && user_mode(regs))
4334 if (event->attr.exclude_kernel && !user_mode(regs))
4341 static int perf_swevent_match(struct perf_event *event,
4342 enum perf_type_id type,
4344 struct perf_sample_data *data,
4345 struct pt_regs *regs)
4347 if (event->attr.type != type)
4350 if (event->attr.config != event_id)
4353 if (perf_exclude_event(event, regs))
4359 static inline u64 swevent_hash(u64 type, u32 event_id)
4361 u64 val = event_id | (type << 32);
4363 return hash_64(val, SWEVENT_HLIST_BITS);
4366 static inline struct hlist_head *
4367 __find_swevent_head(struct swevent_hlist *hlist, u64 type, u32 event_id)
4369 u64 hash = swevent_hash(type, event_id);
4371 return &hlist->heads[hash];
4374 /* For the read side: events when they trigger */
4375 static inline struct hlist_head *
4376 find_swevent_head_rcu(struct swevent_htable *swhash, u64 type, u32 event_id)
4378 struct swevent_hlist *hlist;
4380 hlist = rcu_dereference(swhash->swevent_hlist);
4384 return __find_swevent_head(hlist, type, event_id);
4387 /* For the event head insertion and removal in the hlist */
4388 static inline struct hlist_head *
4389 find_swevent_head(struct swevent_htable *swhash, struct perf_event *event)
4391 struct swevent_hlist *hlist;
4392 u32 event_id = event->attr.config;
4393 u64 type = event->attr.type;
4396 * Event scheduling is always serialized against hlist allocation
4397 * and release. Which makes the protected version suitable here.
4398 * The context lock guarantees that.
4400 hlist = rcu_dereference_protected(swhash->swevent_hlist,
4401 lockdep_is_held(&event->ctx->lock));
4405 return __find_swevent_head(hlist, type, event_id);
4408 static void do_perf_sw_event(enum perf_type_id type, u32 event_id,
4410 struct perf_sample_data *data,
4411 struct pt_regs *regs)
4413 struct swevent_htable *swhash = &__get_cpu_var(swevent_htable);
4414 struct perf_event *event;
4415 struct hlist_node *node;
4416 struct hlist_head *head;
4419 head = find_swevent_head_rcu(swhash, type, event_id);
4423 hlist_for_each_entry_rcu(event, node, head, hlist_entry) {
4424 if (perf_swevent_match(event, type, event_id, data, regs))
4425 perf_swevent_event(event, nr, nmi, data, regs);
4431 int perf_swevent_get_recursion_context(void)
4433 struct swevent_htable *swhash = &__get_cpu_var(swevent_htable);
4435 return get_recursion_context(swhash->recursion);
4437 EXPORT_SYMBOL_GPL(perf_swevent_get_recursion_context);
4439 void inline perf_swevent_put_recursion_context(int rctx)
4441 struct swevent_htable *swhash = &__get_cpu_var(swevent_htable);
4443 put_recursion_context(swhash->recursion, rctx);
4446 void __perf_sw_event(u32 event_id, u64 nr, int nmi,
4447 struct pt_regs *regs, u64 addr)
4449 struct perf_sample_data data;
4452 preempt_disable_notrace();
4453 rctx = perf_swevent_get_recursion_context();
4457 perf_sample_data_init(&data, addr);
4459 do_perf_sw_event(PERF_TYPE_SOFTWARE, event_id, nr, nmi, &data, regs);
4461 perf_swevent_put_recursion_context(rctx);
4462 preempt_enable_notrace();
4465 static void perf_swevent_read(struct perf_event *event)
4469 static int perf_swevent_add(struct perf_event *event, int flags)
4471 struct swevent_htable *swhash = &__get_cpu_var(swevent_htable);
4472 struct hw_perf_event *hwc = &event->hw;
4473 struct hlist_head *head;
4475 if (hwc->sample_period) {
4476 hwc->last_period = hwc->sample_period;
4477 perf_swevent_set_period(event);
4480 hwc->state = !(flags & PERF_EF_START);
4482 head = find_swevent_head(swhash, event);
4483 if (WARN_ON_ONCE(!head))
4486 hlist_add_head_rcu(&event->hlist_entry, head);
4491 static void perf_swevent_del(struct perf_event *event, int flags)
4493 hlist_del_rcu(&event->hlist_entry);
4496 static void perf_swevent_start(struct perf_event *event, int flags)
4498 event->hw.state = 0;
4501 static void perf_swevent_stop(struct perf_event *event, int flags)
4503 event->hw.state = PERF_HES_STOPPED;
4506 /* Deref the hlist from the update side */
4507 static inline struct swevent_hlist *
4508 swevent_hlist_deref(struct swevent_htable *swhash)
4510 return rcu_dereference_protected(swhash->swevent_hlist,
4511 lockdep_is_held(&swhash->hlist_mutex));
4514 static void swevent_hlist_release_rcu(struct rcu_head *rcu_head)
4516 struct swevent_hlist *hlist;
4518 hlist = container_of(rcu_head, struct swevent_hlist, rcu_head);
4522 static void swevent_hlist_release(struct swevent_htable *swhash)
4524 struct swevent_hlist *hlist = swevent_hlist_deref(swhash);
4529 rcu_assign_pointer(swhash->swevent_hlist, NULL);
4530 call_rcu(&hlist->rcu_head, swevent_hlist_release_rcu);
4533 static void swevent_hlist_put_cpu(struct perf_event *event, int cpu)
4535 struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
4537 mutex_lock(&swhash->hlist_mutex);
4539 if (!--swhash->hlist_refcount)
4540 swevent_hlist_release(swhash);
4542 mutex_unlock(&swhash->hlist_mutex);
4545 static void swevent_hlist_put(struct perf_event *event)
4549 if (event->cpu != -1) {
4550 swevent_hlist_put_cpu(event, event->cpu);
4554 for_each_possible_cpu(cpu)
4555 swevent_hlist_put_cpu(event, cpu);
4558 static int swevent_hlist_get_cpu(struct perf_event *event, int cpu)
4560 struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
4563 mutex_lock(&swhash->hlist_mutex);
4565 if (!swevent_hlist_deref(swhash) && cpu_online(cpu)) {
4566 struct swevent_hlist *hlist;
4568 hlist = kzalloc(sizeof(*hlist), GFP_KERNEL);
4573 rcu_assign_pointer(swhash->swevent_hlist, hlist);
4575 swhash->hlist_refcount++;
4577 mutex_unlock(&swhash->hlist_mutex);
4582 static int swevent_hlist_get(struct perf_event *event)
4585 int cpu, failed_cpu;
4587 if (event->cpu != -1)
4588 return swevent_hlist_get_cpu(event, event->cpu);
4591 for_each_possible_cpu(cpu) {
4592 err = swevent_hlist_get_cpu(event, cpu);
4602 for_each_possible_cpu(cpu) {
4603 if (cpu == failed_cpu)
4605 swevent_hlist_put_cpu(event, cpu);
4612 atomic_t perf_swevent_enabled[PERF_COUNT_SW_MAX];
4614 static void sw_perf_event_destroy(struct perf_event *event)
4616 u64 event_id = event->attr.config;
4618 WARN_ON(event->parent);
4620 atomic_dec(&perf_swevent_enabled[event_id]);
4621 swevent_hlist_put(event);
4624 static int perf_swevent_init(struct perf_event *event)
4626 int event_id = event->attr.config;
4628 if (event->attr.type != PERF_TYPE_SOFTWARE)
4632 case PERF_COUNT_SW_CPU_CLOCK:
4633 case PERF_COUNT_SW_TASK_CLOCK:
4640 if (event_id > PERF_COUNT_SW_MAX)
4643 if (!event->parent) {
4646 err = swevent_hlist_get(event);
4650 atomic_inc(&perf_swevent_enabled[event_id]);
4651 event->destroy = sw_perf_event_destroy;
4657 static struct pmu perf_swevent = {
4658 .event_init = perf_swevent_init,
4659 .add = perf_swevent_add,
4660 .del = perf_swevent_del,
4661 .start = perf_swevent_start,
4662 .stop = perf_swevent_stop,
4663 .read = perf_swevent_read,
4666 #ifdef CONFIG_EVENT_TRACING
4668 static int perf_tp_filter_match(struct perf_event *event,
4669 struct perf_sample_data *data)
4671 void *record = data->raw->data;
4673 if (likely(!event->filter) || filter_match_preds(event->filter, record))
4678 static int perf_tp_event_match(struct perf_event *event,
4679 struct perf_sample_data *data,
4680 struct pt_regs *regs)
4683 * All tracepoints are from kernel-space.
4685 if (event->attr.exclude_kernel)
4688 if (!perf_tp_filter_match(event, data))
4694 void perf_tp_event(u64 addr, u64 count, void *record, int entry_size,
4695 struct pt_regs *regs, struct hlist_head *head, int rctx)
4697 struct perf_sample_data data;
4698 struct perf_event *event;
4699 struct hlist_node *node;
4701 struct perf_raw_record raw = {
4706 perf_sample_data_init(&data, addr);
4709 hlist_for_each_entry_rcu(event, node, head, hlist_entry) {
4710 if (perf_tp_event_match(event, &data, regs))
4711 perf_swevent_event(event, count, 1, &data, regs);
4714 perf_swevent_put_recursion_context(rctx);
4716 EXPORT_SYMBOL_GPL(perf_tp_event);
4718 static void tp_perf_event_destroy(struct perf_event *event)
4720 perf_trace_destroy(event);
4723 static int perf_tp_event_init(struct perf_event *event)
4727 if (event->attr.type != PERF_TYPE_TRACEPOINT)
4731 * Raw tracepoint data is a severe data leak, only allow root to
4734 if ((event->attr.sample_type & PERF_SAMPLE_RAW) &&
4735 perf_paranoid_tracepoint_raw() &&
4736 !capable(CAP_SYS_ADMIN))
4739 err = perf_trace_init(event);
4743 event->destroy = tp_perf_event_destroy;
4748 static struct pmu perf_tracepoint = {
4749 .event_init = perf_tp_event_init,
4750 .add = perf_trace_add,
4751 .del = perf_trace_del,
4752 .start = perf_swevent_start,
4753 .stop = perf_swevent_stop,
4754 .read = perf_swevent_read,
4757 static inline void perf_tp_register(void)
4759 perf_pmu_register(&perf_tracepoint);
4762 static int perf_event_set_filter(struct perf_event *event, void __user *arg)
4767 if (event->attr.type != PERF_TYPE_TRACEPOINT)
4770 filter_str = strndup_user(arg, PAGE_SIZE);
4771 if (IS_ERR(filter_str))
4772 return PTR_ERR(filter_str);
4774 ret = ftrace_profile_set_filter(event, event->attr.config, filter_str);
4780 static void perf_event_free_filter(struct perf_event *event)
4782 ftrace_profile_free_filter(event);
4787 static inline void perf_tp_register(void)
4791 static int perf_event_set_filter(struct perf_event *event, void __user *arg)
4796 static void perf_event_free_filter(struct perf_event *event)
4800 #endif /* CONFIG_EVENT_TRACING */
4802 #ifdef CONFIG_HAVE_HW_BREAKPOINT
4803 void perf_bp_event(struct perf_event *bp, void *data)
4805 struct perf_sample_data sample;
4806 struct pt_regs *regs = data;
4808 perf_sample_data_init(&sample, bp->attr.bp_addr);
4810 if (!bp->hw.state && !perf_exclude_event(bp, regs))
4811 perf_swevent_event(bp, 1, 1, &sample, regs);
4816 * hrtimer based swevent callback
4819 static enum hrtimer_restart perf_swevent_hrtimer(struct hrtimer *hrtimer)
4821 enum hrtimer_restart ret = HRTIMER_RESTART;
4822 struct perf_sample_data data;
4823 struct pt_regs *regs;
4824 struct perf_event *event;
4827 event = container_of(hrtimer, struct perf_event, hw.hrtimer);
4828 event->pmu->read(event);
4830 perf_sample_data_init(&data, 0);
4831 data.period = event->hw.last_period;
4832 regs = get_irq_regs();
4834 if (regs && !perf_exclude_event(event, regs)) {
4835 if (!(event->attr.exclude_idle && current->pid == 0))
4836 if (perf_event_overflow(event, 0, &data, regs))
4837 ret = HRTIMER_NORESTART;
4840 period = max_t(u64, 10000, event->hw.sample_period);
4841 hrtimer_forward_now(hrtimer, ns_to_ktime(period));
4846 static void perf_swevent_start_hrtimer(struct perf_event *event)
4848 struct hw_perf_event *hwc = &event->hw;
4850 hrtimer_init(&hwc->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
4851 hwc->hrtimer.function = perf_swevent_hrtimer;
4852 if (hwc->sample_period) {
4853 s64 period = local64_read(&hwc->period_left);
4859 local64_set(&hwc->period_left, 0);
4861 period = max_t(u64, 10000, hwc->sample_period);
4863 __hrtimer_start_range_ns(&hwc->hrtimer,
4864 ns_to_ktime(period), 0,
4865 HRTIMER_MODE_REL_PINNED, 0);
4869 static void perf_swevent_cancel_hrtimer(struct perf_event *event)
4871 struct hw_perf_event *hwc = &event->hw;
4873 if (hwc->sample_period) {
4874 ktime_t remaining = hrtimer_get_remaining(&hwc->hrtimer);
4875 local64_set(&hwc->period_left, ktime_to_ns(remaining));
4877 hrtimer_cancel(&hwc->hrtimer);
4882 * Software event: cpu wall time clock
4885 static void cpu_clock_event_update(struct perf_event *event)
4890 now = local_clock();
4891 prev = local64_xchg(&event->hw.prev_count, now);
4892 local64_add(now - prev, &event->count);
4895 static void cpu_clock_event_start(struct perf_event *event, int flags)
4897 local64_set(&event->hw.prev_count, local_clock());
4898 perf_swevent_start_hrtimer(event);
4901 static void cpu_clock_event_stop(struct perf_event *event, int flags)
4903 perf_swevent_cancel_hrtimer(event);
4904 cpu_clock_event_update(event);
4907 static int cpu_clock_event_add(struct perf_event *event, int flags)
4909 if (flags & PERF_EF_START)
4910 cpu_clock_event_start(event, flags);
4915 static void cpu_clock_event_del(struct perf_event *event, int flags)
4917 cpu_clock_event_stop(event, flags);
4920 static void cpu_clock_event_read(struct perf_event *event)
4922 cpu_clock_event_update(event);
4925 static int cpu_clock_event_init(struct perf_event *event)
4927 if (event->attr.type != PERF_TYPE_SOFTWARE)
4930 if (event->attr.config != PERF_COUNT_SW_CPU_CLOCK)
4936 static struct pmu perf_cpu_clock = {
4937 .event_init = cpu_clock_event_init,
4938 .add = cpu_clock_event_add,
4939 .del = cpu_clock_event_del,
4940 .start = cpu_clock_event_start,
4941 .stop = cpu_clock_event_stop,
4942 .read = cpu_clock_event_read,
4946 * Software event: task time clock
4949 static void task_clock_event_update(struct perf_event *event, u64 now)
4954 prev = local64_xchg(&event->hw.prev_count, now);
4956 local64_add(delta, &event->count);
4959 static void task_clock_event_start(struct perf_event *event, int flags)
4961 local64_set(&event->hw.prev_count, event->ctx->time);
4962 perf_swevent_start_hrtimer(event);
4965 static void task_clock_event_stop(struct perf_event *event, int flags)
4967 perf_swevent_cancel_hrtimer(event);
4968 task_clock_event_update(event, event->ctx->time);
4971 static int task_clock_event_add(struct perf_event *event, int flags)
4973 if (flags & PERF_EF_START)
4974 task_clock_event_start(event, flags);
4979 static void task_clock_event_del(struct perf_event *event, int flags)
4981 task_clock_event_stop(event, PERF_EF_UPDATE);
4984 static void task_clock_event_read(struct perf_event *event)
4989 update_context_time(event->ctx);
4990 time = event->ctx->time;
4992 u64 now = perf_clock();
4993 u64 delta = now - event->ctx->timestamp;
4994 time = event->ctx->time + delta;
4997 task_clock_event_update(event, time);
5000 static int task_clock_event_init(struct perf_event *event)
5002 if (event->attr.type != PERF_TYPE_SOFTWARE)
5005 if (event->attr.config != PERF_COUNT_SW_TASK_CLOCK)
5011 static struct pmu perf_task_clock = {
5012 .event_init = task_clock_event_init,
5013 .add = task_clock_event_add,
5014 .del = task_clock_event_del,
5015 .start = task_clock_event_start,
5016 .stop = task_clock_event_stop,
5017 .read = task_clock_event_read,
5020 static void perf_pmu_nop_void(struct pmu *pmu)
5024 static int perf_pmu_nop_int(struct pmu *pmu)
5029 static void perf_pmu_start_txn(struct pmu *pmu)
5031 perf_pmu_disable(pmu);
5034 static int perf_pmu_commit_txn(struct pmu *pmu)
5036 perf_pmu_enable(pmu);
5040 static void perf_pmu_cancel_txn(struct pmu *pmu)
5042 perf_pmu_enable(pmu);
5045 int perf_pmu_register(struct pmu *pmu)
5049 mutex_lock(&pmus_lock);
5051 pmu->pmu_disable_count = alloc_percpu(int);
5052 if (!pmu->pmu_disable_count)
5055 pmu->pmu_cpu_context = alloc_percpu(struct perf_cpu_context);
5056 if (!pmu->pmu_cpu_context)
5059 for_each_possible_cpu(cpu) {
5060 struct perf_cpu_context *cpuctx;
5062 cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu);
5063 __perf_event_init_context(&cpuctx->ctx);
5064 cpuctx->ctx.pmu = pmu;
5065 cpuctx->timer_interval = TICK_NSEC;
5066 hrtimer_init(&cpuctx->timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
5067 cpuctx->timer.function = perf_event_context_tick;
5070 if (!pmu->start_txn) {
5071 if (pmu->pmu_enable) {
5073 * If we have pmu_enable/pmu_disable calls, install
5074 * transaction stubs that use that to try and batch
5075 * hardware accesses.
5077 pmu->start_txn = perf_pmu_start_txn;
5078 pmu->commit_txn = perf_pmu_commit_txn;
5079 pmu->cancel_txn = perf_pmu_cancel_txn;
5081 pmu->start_txn = perf_pmu_nop_void;
5082 pmu->commit_txn = perf_pmu_nop_int;
5083 pmu->cancel_txn = perf_pmu_nop_void;
5087 if (!pmu->pmu_enable) {
5088 pmu->pmu_enable = perf_pmu_nop_void;
5089 pmu->pmu_disable = perf_pmu_nop_void;
5092 list_add_rcu(&pmu->entry, &pmus);
5095 mutex_unlock(&pmus_lock);
5100 free_percpu(pmu->pmu_disable_count);
5104 void perf_pmu_unregister(struct pmu *pmu)
5106 mutex_lock(&pmus_lock);
5107 list_del_rcu(&pmu->entry);
5108 mutex_unlock(&pmus_lock);
5111 * We use the pmu list either under SRCU or preempt_disable,
5112 * synchronize_srcu() implies synchronize_sched() so we're good.
5114 synchronize_srcu(&pmus_srcu);
5116 free_percpu(pmu->pmu_disable_count);
5117 free_percpu(pmu->pmu_cpu_context);
5120 struct pmu *perf_init_event(struct perf_event *event)
5122 struct pmu *pmu = NULL;
5125 idx = srcu_read_lock(&pmus_srcu);
5126 list_for_each_entry_rcu(pmu, &pmus, entry) {
5127 int ret = pmu->event_init(event);
5130 if (ret != -ENOENT) {
5135 srcu_read_unlock(&pmus_srcu, idx);
5141 * Allocate and initialize a event structure
5143 static struct perf_event *
5144 perf_event_alloc(struct perf_event_attr *attr, int cpu,
5145 struct perf_event *group_leader,
5146 struct perf_event *parent_event,
5147 perf_overflow_handler_t overflow_handler)
5150 struct perf_event *event;
5151 struct hw_perf_event *hwc;
5154 event = kzalloc(sizeof(*event), GFP_KERNEL);
5156 return ERR_PTR(-ENOMEM);
5159 * Single events are their own group leaders, with an
5160 * empty sibling list:
5163 group_leader = event;
5165 mutex_init(&event->child_mutex);
5166 INIT_LIST_HEAD(&event->child_list);
5168 INIT_LIST_HEAD(&event->group_entry);
5169 INIT_LIST_HEAD(&event->event_entry);
5170 INIT_LIST_HEAD(&event->sibling_list);
5171 init_waitqueue_head(&event->waitq);
5173 mutex_init(&event->mmap_mutex);
5176 event->attr = *attr;
5177 event->group_leader = group_leader;
5181 event->parent = parent_event;
5183 event->ns = get_pid_ns(current->nsproxy->pid_ns);
5184 event->id = atomic64_inc_return(&perf_event_id);
5186 event->state = PERF_EVENT_STATE_INACTIVE;
5188 if (!overflow_handler && parent_event)
5189 overflow_handler = parent_event->overflow_handler;
5191 event->overflow_handler = overflow_handler;
5194 event->state = PERF_EVENT_STATE_OFF;
5199 hwc->sample_period = attr->sample_period;
5200 if (attr->freq && attr->sample_freq)
5201 hwc->sample_period = 1;
5202 hwc->last_period = hwc->sample_period;
5204 local64_set(&hwc->period_left, hwc->sample_period);
5207 * we currently do not support PERF_FORMAT_GROUP on inherited events
5209 if (attr->inherit && (attr->read_format & PERF_FORMAT_GROUP))
5212 pmu = perf_init_event(event);
5218 else if (IS_ERR(pmu))
5223 put_pid_ns(event->ns);
5225 return ERR_PTR(err);
5230 if (!event->parent) {
5231 atomic_inc(&nr_events);
5232 if (event->attr.mmap || event->attr.mmap_data)
5233 atomic_inc(&nr_mmap_events);
5234 if (event->attr.comm)
5235 atomic_inc(&nr_comm_events);
5236 if (event->attr.task)
5237 atomic_inc(&nr_task_events);
5238 if (event->attr.sample_type & PERF_SAMPLE_CALLCHAIN) {
5239 err = get_callchain_buffers();
5242 return ERR_PTR(err);
5250 static int perf_copy_attr(struct perf_event_attr __user *uattr,
5251 struct perf_event_attr *attr)
5256 if (!access_ok(VERIFY_WRITE, uattr, PERF_ATTR_SIZE_VER0))
5260 * zero the full structure, so that a short copy will be nice.
5262 memset(attr, 0, sizeof(*attr));
5264 ret = get_user(size, &uattr->size);
5268 if (size > PAGE_SIZE) /* silly large */
5271 if (!size) /* abi compat */
5272 size = PERF_ATTR_SIZE_VER0;
5274 if (size < PERF_ATTR_SIZE_VER0)
5278 * If we're handed a bigger struct than we know of,
5279 * ensure all the unknown bits are 0 - i.e. new
5280 * user-space does not rely on any kernel feature
5281 * extensions we dont know about yet.
5283 if (size > sizeof(*attr)) {
5284 unsigned char __user *addr;
5285 unsigned char __user *end;
5288 addr = (void __user *)uattr + sizeof(*attr);
5289 end = (void __user *)uattr + size;
5291 for (; addr < end; addr++) {
5292 ret = get_user(val, addr);
5298 size = sizeof(*attr);
5301 ret = copy_from_user(attr, uattr, size);
5306 * If the type exists, the corresponding creation will verify
5309 if (attr->type >= PERF_TYPE_MAX)
5312 if (attr->__reserved_1)
5315 if (attr->sample_type & ~(PERF_SAMPLE_MAX-1))
5318 if (attr->read_format & ~(PERF_FORMAT_MAX-1))
5325 put_user(sizeof(*attr), &uattr->size);
5331 perf_event_set_output(struct perf_event *event, struct perf_event *output_event)
5333 struct perf_buffer *buffer = NULL, *old_buffer = NULL;
5339 /* don't allow circular references */
5340 if (event == output_event)
5344 * Don't allow cross-cpu buffers
5346 if (output_event->cpu != event->cpu)
5350 * If its not a per-cpu buffer, it must be the same task.
5352 if (output_event->cpu == -1 && output_event->ctx != event->ctx)
5356 mutex_lock(&event->mmap_mutex);
5357 /* Can't redirect output if we've got an active mmap() */
5358 if (atomic_read(&event->mmap_count))
5362 /* get the buffer we want to redirect to */
5363 buffer = perf_buffer_get(output_event);
5368 old_buffer = event->buffer;
5369 rcu_assign_pointer(event->buffer, buffer);
5372 mutex_unlock(&event->mmap_mutex);
5375 perf_buffer_put(old_buffer);
5381 * sys_perf_event_open - open a performance event, associate it to a task/cpu
5383 * @attr_uptr: event_id type attributes for monitoring/sampling
5386 * @group_fd: group leader event fd
5388 SYSCALL_DEFINE5(perf_event_open,
5389 struct perf_event_attr __user *, attr_uptr,
5390 pid_t, pid, int, cpu, int, group_fd, unsigned long, flags)
5392 struct perf_event *event, *group_leader = NULL, *output_event = NULL;
5393 struct perf_event_attr attr;
5394 struct perf_event_context *ctx;
5395 struct file *event_file = NULL;
5396 struct file *group_file = NULL;
5398 int fput_needed = 0;
5401 /* for future expandability... */
5402 if (flags & ~(PERF_FLAG_FD_NO_GROUP | PERF_FLAG_FD_OUTPUT))
5405 err = perf_copy_attr(attr_uptr, &attr);
5409 if (!attr.exclude_kernel) {
5410 if (perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN))
5415 if (attr.sample_freq > sysctl_perf_event_sample_rate)
5419 event_fd = get_unused_fd_flags(O_RDWR);
5423 event = perf_event_alloc(&attr, cpu, group_leader, NULL, NULL);
5424 if (IS_ERR(event)) {
5425 err = PTR_ERR(event);
5430 * Get the target context (task or percpu):
5432 ctx = find_get_context(event->pmu, pid, cpu);
5438 if (group_fd != -1) {
5439 group_leader = perf_fget_light(group_fd, &fput_needed);
5440 if (IS_ERR(group_leader)) {
5441 err = PTR_ERR(group_leader);
5444 group_file = group_leader->filp;
5445 if (flags & PERF_FLAG_FD_OUTPUT)
5446 output_event = group_leader;
5447 if (flags & PERF_FLAG_FD_NO_GROUP)
5448 group_leader = NULL;
5452 * Look up the group leader (we will attach this event to it):
5458 * Do not allow a recursive hierarchy (this new sibling
5459 * becoming part of another group-sibling):
5461 if (group_leader->group_leader != group_leader)
5464 * Do not allow to attach to a group in a different
5465 * task or CPU context:
5467 if (group_leader->ctx != ctx)
5470 * Only a group leader can be exclusive or pinned
5472 if (attr.exclusive || attr.pinned)
5477 err = perf_event_set_output(event, output_event);
5482 event_file = anon_inode_getfile("[perf_event]", &perf_fops, event, O_RDWR);
5483 if (IS_ERR(event_file)) {
5484 err = PTR_ERR(event_file);
5488 event->filp = event_file;
5489 WARN_ON_ONCE(ctx->parent_ctx);
5490 mutex_lock(&ctx->mutex);
5491 perf_install_in_context(ctx, event, cpu);
5493 mutex_unlock(&ctx->mutex);
5495 event->owner = current;
5496 get_task_struct(current);
5497 mutex_lock(¤t->perf_event_mutex);
5498 list_add_tail(&event->owner_entry, ¤t->perf_event_list);
5499 mutex_unlock(¤t->perf_event_mutex);
5502 * Drop the reference on the group_event after placing the
5503 * new event on the sibling_list. This ensures destruction
5504 * of the group leader will find the pointer to itself in
5505 * perf_group_detach().
5507 fput_light(group_file, fput_needed);
5508 fd_install(event_fd, event_file);
5512 fput_light(group_file, fput_needed);
5517 put_unused_fd(event_fd);
5522 * perf_event_create_kernel_counter
5524 * @attr: attributes of the counter to create
5525 * @cpu: cpu in which the counter is bound
5526 * @pid: task to profile
5529 perf_event_create_kernel_counter(struct perf_event_attr *attr, int cpu,
5531 perf_overflow_handler_t overflow_handler)
5533 struct perf_event_context *ctx;
5534 struct perf_event *event;
5538 * Get the target context (task or percpu):
5541 event = perf_event_alloc(attr, cpu, NULL, NULL, overflow_handler);
5542 if (IS_ERR(event)) {
5543 err = PTR_ERR(event);
5547 ctx = find_get_context(event->pmu, pid, cpu);
5554 WARN_ON_ONCE(ctx->parent_ctx);
5555 mutex_lock(&ctx->mutex);
5556 perf_install_in_context(ctx, event, cpu);
5558 mutex_unlock(&ctx->mutex);
5560 event->owner = current;
5561 get_task_struct(current);
5562 mutex_lock(¤t->perf_event_mutex);
5563 list_add_tail(&event->owner_entry, ¤t->perf_event_list);
5564 mutex_unlock(¤t->perf_event_mutex);
5571 return ERR_PTR(err);
5573 EXPORT_SYMBOL_GPL(perf_event_create_kernel_counter);
5575 static void sync_child_event(struct perf_event *child_event,
5576 struct task_struct *child)
5578 struct perf_event *parent_event = child_event->parent;
5581 if (child_event->attr.inherit_stat)
5582 perf_event_read_event(child_event, child);
5584 child_val = perf_event_count(child_event);
5587 * Add back the child's count to the parent's count:
5589 atomic64_add(child_val, &parent_event->child_count);
5590 atomic64_add(child_event->total_time_enabled,
5591 &parent_event->child_total_time_enabled);
5592 atomic64_add(child_event->total_time_running,
5593 &parent_event->child_total_time_running);
5596 * Remove this event from the parent's list
5598 WARN_ON_ONCE(parent_event->ctx->parent_ctx);
5599 mutex_lock(&parent_event->child_mutex);
5600 list_del_init(&child_event->child_list);
5601 mutex_unlock(&parent_event->child_mutex);
5604 * Release the parent event, if this was the last
5607 fput(parent_event->filp);
5611 __perf_event_exit_task(struct perf_event *child_event,
5612 struct perf_event_context *child_ctx,
5613 struct task_struct *child)
5615 struct perf_event *parent_event;
5617 perf_event_remove_from_context(child_event);
5619 parent_event = child_event->parent;
5621 * It can happen that parent exits first, and has events
5622 * that are still around due to the child reference. These
5623 * events need to be zapped - but otherwise linger.
5626 sync_child_event(child_event, child);
5627 free_event(child_event);
5632 * When a child task exits, feed back event values to parent events.
5634 void perf_event_exit_task(struct task_struct *child)
5636 struct perf_event *child_event, *tmp;
5637 struct perf_event_context *child_ctx;
5638 unsigned long flags;
5640 if (likely(!child->perf_event_ctxp)) {
5641 perf_event_task(child, NULL, 0);
5645 local_irq_save(flags);
5647 * We can't reschedule here because interrupts are disabled,
5648 * and either child is current or it is a task that can't be
5649 * scheduled, so we are now safe from rescheduling changing
5652 child_ctx = child->perf_event_ctxp;
5653 __perf_event_task_sched_out(child_ctx);
5656 * Take the context lock here so that if find_get_context is
5657 * reading child->perf_event_ctxp, we wait until it has
5658 * incremented the context's refcount before we do put_ctx below.
5660 raw_spin_lock(&child_ctx->lock);
5661 child->perf_event_ctxp = NULL;
5663 * If this context is a clone; unclone it so it can't get
5664 * swapped to another process while we're removing all
5665 * the events from it.
5667 unclone_ctx(child_ctx);
5668 update_context_time(child_ctx);
5669 raw_spin_unlock_irqrestore(&child_ctx->lock, flags);
5672 * Report the task dead after unscheduling the events so that we
5673 * won't get any samples after PERF_RECORD_EXIT. We can however still
5674 * get a few PERF_RECORD_READ events.
5676 perf_event_task(child, child_ctx, 0);
5679 * We can recurse on the same lock type through:
5681 * __perf_event_exit_task()
5682 * sync_child_event()
5683 * fput(parent_event->filp)
5685 * mutex_lock(&ctx->mutex)
5687 * But since its the parent context it won't be the same instance.
5689 mutex_lock(&child_ctx->mutex);
5692 list_for_each_entry_safe(child_event, tmp, &child_ctx->pinned_groups,
5694 __perf_event_exit_task(child_event, child_ctx, child);
5696 list_for_each_entry_safe(child_event, tmp, &child_ctx->flexible_groups,
5698 __perf_event_exit_task(child_event, child_ctx, child);
5701 * If the last event was a group event, it will have appended all
5702 * its siblings to the list, but we obtained 'tmp' before that which
5703 * will still point to the list head terminating the iteration.
5705 if (!list_empty(&child_ctx->pinned_groups) ||
5706 !list_empty(&child_ctx->flexible_groups))
5709 mutex_unlock(&child_ctx->mutex);
5714 static void perf_free_event(struct perf_event *event,
5715 struct perf_event_context *ctx)
5717 struct perf_event *parent = event->parent;
5719 if (WARN_ON_ONCE(!parent))
5722 mutex_lock(&parent->child_mutex);
5723 list_del_init(&event->child_list);
5724 mutex_unlock(&parent->child_mutex);
5728 perf_group_detach(event);
5729 list_del_event(event, ctx);
5734 * free an unexposed, unused context as created by inheritance by
5735 * init_task below, used by fork() in case of fail.
5737 void perf_event_free_task(struct task_struct *task)
5739 struct perf_event_context *ctx = task->perf_event_ctxp;
5740 struct perf_event *event, *tmp;
5745 mutex_lock(&ctx->mutex);
5747 list_for_each_entry_safe(event, tmp, &ctx->pinned_groups, group_entry)
5748 perf_free_event(event, ctx);
5750 list_for_each_entry_safe(event, tmp, &ctx->flexible_groups,
5752 perf_free_event(event, ctx);
5754 if (!list_empty(&ctx->pinned_groups) ||
5755 !list_empty(&ctx->flexible_groups))
5758 mutex_unlock(&ctx->mutex);
5764 * inherit a event from parent task to child task:
5766 static struct perf_event *
5767 inherit_event(struct perf_event *parent_event,
5768 struct task_struct *parent,
5769 struct perf_event_context *parent_ctx,
5770 struct task_struct *child,
5771 struct perf_event *group_leader,
5772 struct perf_event_context *child_ctx)
5774 struct perf_event *child_event;
5777 * Instead of creating recursive hierarchies of events,
5778 * we link inherited events back to the original parent,
5779 * which has a filp for sure, which we use as the reference
5782 if (parent_event->parent)
5783 parent_event = parent_event->parent;
5785 child_event = perf_event_alloc(&parent_event->attr,
5787 group_leader, parent_event,
5789 if (IS_ERR(child_event))
5794 * Make the child state follow the state of the parent event,
5795 * not its attr.disabled bit. We hold the parent's mutex,
5796 * so we won't race with perf_event_{en, dis}able_family.
5798 if (parent_event->state >= PERF_EVENT_STATE_INACTIVE)
5799 child_event->state = PERF_EVENT_STATE_INACTIVE;
5801 child_event->state = PERF_EVENT_STATE_OFF;
5803 if (parent_event->attr.freq) {
5804 u64 sample_period = parent_event->hw.sample_period;
5805 struct hw_perf_event *hwc = &child_event->hw;
5807 hwc->sample_period = sample_period;
5808 hwc->last_period = sample_period;
5810 local64_set(&hwc->period_left, sample_period);
5813 child_event->ctx = child_ctx;
5814 child_event->overflow_handler = parent_event->overflow_handler;
5817 * Link it up in the child's context:
5819 add_event_to_ctx(child_event, child_ctx);
5822 * Get a reference to the parent filp - we will fput it
5823 * when the child event exits. This is safe to do because
5824 * we are in the parent and we know that the filp still
5825 * exists and has a nonzero count:
5827 atomic_long_inc(&parent_event->filp->f_count);
5830 * Link this into the parent event's child list
5832 WARN_ON_ONCE(parent_event->ctx->parent_ctx);
5833 mutex_lock(&parent_event->child_mutex);
5834 list_add_tail(&child_event->child_list, &parent_event->child_list);
5835 mutex_unlock(&parent_event->child_mutex);
5840 static int inherit_group(struct perf_event *parent_event,
5841 struct task_struct *parent,
5842 struct perf_event_context *parent_ctx,
5843 struct task_struct *child,
5844 struct perf_event_context *child_ctx)
5846 struct perf_event *leader;
5847 struct perf_event *sub;
5848 struct perf_event *child_ctr;
5850 leader = inherit_event(parent_event, parent, parent_ctx,
5851 child, NULL, child_ctx);
5853 return PTR_ERR(leader);
5854 list_for_each_entry(sub, &parent_event->sibling_list, group_entry) {
5855 child_ctr = inherit_event(sub, parent, parent_ctx,
5856 child, leader, child_ctx);
5857 if (IS_ERR(child_ctr))
5858 return PTR_ERR(child_ctr);
5864 inherit_task_group(struct perf_event *event, struct task_struct *parent,
5865 struct perf_event_context *parent_ctx,
5866 struct task_struct *child,
5870 struct perf_event_context *child_ctx = child->perf_event_ctxp;
5872 if (!event->attr.inherit) {
5879 * This is executed from the parent task context, so
5880 * inherit events that have been marked for cloning.
5881 * First allocate and initialize a context for the
5885 child_ctx = alloc_perf_context(event->pmu, child);
5889 child->perf_event_ctxp = child_ctx;
5892 ret = inherit_group(event, parent, parent_ctx,
5902 * Initialize the perf_event context in task_struct
5904 int perf_event_init_task(struct task_struct *child)
5906 struct perf_event_context *child_ctx, *parent_ctx;
5907 struct perf_event_context *cloned_ctx;
5908 struct perf_event *event;
5909 struct task_struct *parent = current;
5910 int inherited_all = 1;
5913 child->perf_event_ctxp = NULL;
5915 mutex_init(&child->perf_event_mutex);
5916 INIT_LIST_HEAD(&child->perf_event_list);
5918 if (likely(!parent->perf_event_ctxp))
5922 * If the parent's context is a clone, pin it so it won't get
5925 parent_ctx = perf_pin_task_context(parent);
5928 * No need to check if parent_ctx != NULL here; since we saw
5929 * it non-NULL earlier, the only reason for it to become NULL
5930 * is if we exit, and since we're currently in the middle of
5931 * a fork we can't be exiting at the same time.
5935 * Lock the parent list. No need to lock the child - not PID
5936 * hashed yet and not running, so nobody can access it.
5938 mutex_lock(&parent_ctx->mutex);
5941 * We dont have to disable NMIs - we are only looking at
5942 * the list, not manipulating it:
5944 list_for_each_entry(event, &parent_ctx->pinned_groups, group_entry) {
5945 ret = inherit_task_group(event, parent, parent_ctx, child,
5951 list_for_each_entry(event, &parent_ctx->flexible_groups, group_entry) {
5952 ret = inherit_task_group(event, parent, parent_ctx, child,
5958 child_ctx = child->perf_event_ctxp;
5960 if (child_ctx && inherited_all) {
5962 * Mark the child context as a clone of the parent
5963 * context, or of whatever the parent is a clone of.
5964 * Note that if the parent is a clone, it could get
5965 * uncloned at any point, but that doesn't matter
5966 * because the list of events and the generation
5967 * count can't have changed since we took the mutex.
5969 cloned_ctx = rcu_dereference(parent_ctx->parent_ctx);
5971 child_ctx->parent_ctx = cloned_ctx;
5972 child_ctx->parent_gen = parent_ctx->parent_gen;
5974 child_ctx->parent_ctx = parent_ctx;
5975 child_ctx->parent_gen = parent_ctx->generation;
5977 get_ctx(child_ctx->parent_ctx);
5980 mutex_unlock(&parent_ctx->mutex);
5982 perf_unpin_context(parent_ctx);
5987 static void __init perf_event_init_all_cpus(void)
5989 struct swevent_htable *swhash;
5992 for_each_possible_cpu(cpu) {
5993 swhash = &per_cpu(swevent_htable, cpu);
5994 mutex_init(&swhash->hlist_mutex);
5998 static void __cpuinit perf_event_init_cpu(int cpu)
6000 struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
6002 mutex_lock(&swhash->hlist_mutex);
6003 if (swhash->hlist_refcount > 0) {
6004 struct swevent_hlist *hlist;
6006 hlist = kzalloc_node(sizeof(*hlist), GFP_KERNEL, cpu_to_node(cpu));
6008 rcu_assign_pointer(swhash->swevent_hlist, hlist);
6010 mutex_unlock(&swhash->hlist_mutex);
6013 #ifdef CONFIG_HOTPLUG_CPU
6014 static void __perf_event_exit_context(void *__info)
6016 struct perf_event_context *ctx = __info;
6017 struct perf_event *event, *tmp;
6019 perf_pmu_rotate_stop(ctx->pmu);
6021 list_for_each_entry_safe(event, tmp, &ctx->pinned_groups, group_entry)
6022 __perf_event_remove_from_context(event);
6023 list_for_each_entry_safe(event, tmp, &ctx->flexible_groups, group_entry)
6024 __perf_event_remove_from_context(event);
6027 static void perf_event_exit_cpu_context(int cpu)
6029 struct perf_event_context *ctx;
6033 idx = srcu_read_lock(&pmus_srcu);
6034 list_for_each_entry_rcu(pmu, &pmus, entry) {
6035 ctx = &this_cpu_ptr(pmu->pmu_cpu_context)->ctx;
6037 mutex_lock(&ctx->mutex);
6038 smp_call_function_single(cpu, __perf_event_exit_context, ctx, 1);
6039 mutex_unlock(&ctx->mutex);
6041 srcu_read_unlock(&pmus_srcu, idx);
6045 static void perf_event_exit_cpu(int cpu)
6047 struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
6049 mutex_lock(&swhash->hlist_mutex);
6050 swevent_hlist_release(swhash);
6051 mutex_unlock(&swhash->hlist_mutex);
6053 perf_event_exit_cpu_context(cpu);
6056 static inline void perf_event_exit_cpu(int cpu) { }
6059 static int __cpuinit
6060 perf_cpu_notify(struct notifier_block *self, unsigned long action, void *hcpu)
6062 unsigned int cpu = (long)hcpu;
6064 switch (action & ~CPU_TASKS_FROZEN) {
6066 case CPU_UP_PREPARE:
6067 case CPU_DOWN_FAILED:
6068 perf_event_init_cpu(cpu);
6071 case CPU_UP_CANCELED:
6072 case CPU_DOWN_PREPARE:
6073 perf_event_exit_cpu(cpu);
6083 void __init perf_event_init(void)
6085 perf_event_init_all_cpus();
6086 init_srcu_struct(&pmus_srcu);
6087 perf_pmu_register(&perf_swevent);
6088 perf_pmu_register(&perf_cpu_clock);
6089 perf_pmu_register(&perf_task_clock);
6091 perf_cpu_notifier(perf_cpu_notify);