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:
1963 __perf_event_init_context(struct perf_event_context *ctx,
1964 struct task_struct *task)
1966 raw_spin_lock_init(&ctx->lock);
1967 mutex_init(&ctx->mutex);
1968 INIT_LIST_HEAD(&ctx->pinned_groups);
1969 INIT_LIST_HEAD(&ctx->flexible_groups);
1970 INIT_LIST_HEAD(&ctx->event_list);
1971 atomic_set(&ctx->refcount, 1);
1975 static struct perf_event_context *
1976 find_get_context(struct pmu *pmu, pid_t pid, int cpu)
1978 struct perf_event_context *ctx;
1979 struct perf_cpu_context *cpuctx;
1980 struct task_struct *task;
1981 unsigned long flags;
1984 if (pid == -1 && cpu != -1) {
1985 /* Must be root to operate on a CPU event: */
1986 if (perf_paranoid_cpu() && !capable(CAP_SYS_ADMIN))
1987 return ERR_PTR(-EACCES);
1989 if (cpu < 0 || cpu >= nr_cpumask_bits)
1990 return ERR_PTR(-EINVAL);
1993 * We could be clever and allow to attach a event to an
1994 * offline CPU and activate it when the CPU comes up, but
1997 if (!cpu_online(cpu))
1998 return ERR_PTR(-ENODEV);
2000 cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu);
2011 task = find_task_by_vpid(pid);
2013 get_task_struct(task);
2017 return ERR_PTR(-ESRCH);
2020 * Can't attach events to a dying task.
2023 if (task->flags & PF_EXITING)
2026 /* Reuse ptrace permission checks for now. */
2028 if (!ptrace_may_access(task, PTRACE_MODE_READ))
2032 ctx = perf_lock_task_context(task, &flags);
2035 raw_spin_unlock_irqrestore(&ctx->lock, flags);
2039 ctx = kzalloc(sizeof(struct perf_event_context), GFP_KERNEL);
2043 __perf_event_init_context(ctx, task);
2046 if (cmpxchg(&task->perf_event_ctxp, NULL, ctx)) {
2048 * We raced with some other task; use
2049 * the context they set.
2054 get_task_struct(task);
2057 put_task_struct(task);
2061 put_task_struct(task);
2062 return ERR_PTR(err);
2065 static void perf_event_free_filter(struct perf_event *event);
2067 static void free_event_rcu(struct rcu_head *head)
2069 struct perf_event *event;
2071 event = container_of(head, struct perf_event, rcu_head);
2073 put_pid_ns(event->ns);
2074 perf_event_free_filter(event);
2078 static void perf_pending_sync(struct perf_event *event);
2079 static void perf_buffer_put(struct perf_buffer *buffer);
2081 static void free_event(struct perf_event *event)
2083 perf_pending_sync(event);
2085 if (!event->parent) {
2086 atomic_dec(&nr_events);
2087 if (event->attr.mmap || event->attr.mmap_data)
2088 atomic_dec(&nr_mmap_events);
2089 if (event->attr.comm)
2090 atomic_dec(&nr_comm_events);
2091 if (event->attr.task)
2092 atomic_dec(&nr_task_events);
2093 if (event->attr.sample_type & PERF_SAMPLE_CALLCHAIN)
2094 put_callchain_buffers();
2097 if (event->buffer) {
2098 perf_buffer_put(event->buffer);
2099 event->buffer = NULL;
2103 event->destroy(event);
2105 put_ctx(event->ctx);
2106 call_rcu(&event->rcu_head, free_event_rcu);
2109 int perf_event_release_kernel(struct perf_event *event)
2111 struct perf_event_context *ctx = event->ctx;
2114 * Remove from the PMU, can't get re-enabled since we got
2115 * here because the last ref went.
2117 perf_event_disable(event);
2119 WARN_ON_ONCE(ctx->parent_ctx);
2121 * There are two ways this annotation is useful:
2123 * 1) there is a lock recursion from perf_event_exit_task
2124 * see the comment there.
2126 * 2) there is a lock-inversion with mmap_sem through
2127 * perf_event_read_group(), which takes faults while
2128 * holding ctx->mutex, however this is called after
2129 * the last filedesc died, so there is no possibility
2130 * to trigger the AB-BA case.
2132 mutex_lock_nested(&ctx->mutex, SINGLE_DEPTH_NESTING);
2133 raw_spin_lock_irq(&ctx->lock);
2134 perf_group_detach(event);
2135 list_del_event(event, ctx);
2136 raw_spin_unlock_irq(&ctx->lock);
2137 mutex_unlock(&ctx->mutex);
2139 mutex_lock(&event->owner->perf_event_mutex);
2140 list_del_init(&event->owner_entry);
2141 mutex_unlock(&event->owner->perf_event_mutex);
2142 put_task_struct(event->owner);
2148 EXPORT_SYMBOL_GPL(perf_event_release_kernel);
2151 * Called when the last reference to the file is gone.
2153 static int perf_release(struct inode *inode, struct file *file)
2155 struct perf_event *event = file->private_data;
2157 file->private_data = NULL;
2159 return perf_event_release_kernel(event);
2162 static int perf_event_read_size(struct perf_event *event)
2164 int entry = sizeof(u64); /* value */
2168 if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
2169 size += sizeof(u64);
2171 if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
2172 size += sizeof(u64);
2174 if (event->attr.read_format & PERF_FORMAT_ID)
2175 entry += sizeof(u64);
2177 if (event->attr.read_format & PERF_FORMAT_GROUP) {
2178 nr += event->group_leader->nr_siblings;
2179 size += sizeof(u64);
2187 u64 perf_event_read_value(struct perf_event *event, u64 *enabled, u64 *running)
2189 struct perf_event *child;
2195 mutex_lock(&event->child_mutex);
2196 total += perf_event_read(event);
2197 *enabled += event->total_time_enabled +
2198 atomic64_read(&event->child_total_time_enabled);
2199 *running += event->total_time_running +
2200 atomic64_read(&event->child_total_time_running);
2202 list_for_each_entry(child, &event->child_list, child_list) {
2203 total += perf_event_read(child);
2204 *enabled += child->total_time_enabled;
2205 *running += child->total_time_running;
2207 mutex_unlock(&event->child_mutex);
2211 EXPORT_SYMBOL_GPL(perf_event_read_value);
2213 static int perf_event_read_group(struct perf_event *event,
2214 u64 read_format, char __user *buf)
2216 struct perf_event *leader = event->group_leader, *sub;
2217 int n = 0, size = 0, ret = -EFAULT;
2218 struct perf_event_context *ctx = leader->ctx;
2220 u64 count, enabled, running;
2222 mutex_lock(&ctx->mutex);
2223 count = perf_event_read_value(leader, &enabled, &running);
2225 values[n++] = 1 + leader->nr_siblings;
2226 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
2227 values[n++] = enabled;
2228 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
2229 values[n++] = running;
2230 values[n++] = count;
2231 if (read_format & PERF_FORMAT_ID)
2232 values[n++] = primary_event_id(leader);
2234 size = n * sizeof(u64);
2236 if (copy_to_user(buf, values, size))
2241 list_for_each_entry(sub, &leader->sibling_list, group_entry) {
2244 values[n++] = perf_event_read_value(sub, &enabled, &running);
2245 if (read_format & PERF_FORMAT_ID)
2246 values[n++] = primary_event_id(sub);
2248 size = n * sizeof(u64);
2250 if (copy_to_user(buf + ret, values, size)) {
2258 mutex_unlock(&ctx->mutex);
2263 static int perf_event_read_one(struct perf_event *event,
2264 u64 read_format, char __user *buf)
2266 u64 enabled, running;
2270 values[n++] = perf_event_read_value(event, &enabled, &running);
2271 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
2272 values[n++] = enabled;
2273 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
2274 values[n++] = running;
2275 if (read_format & PERF_FORMAT_ID)
2276 values[n++] = primary_event_id(event);
2278 if (copy_to_user(buf, values, n * sizeof(u64)))
2281 return n * sizeof(u64);
2285 * Read the performance event - simple non blocking version for now
2288 perf_read_hw(struct perf_event *event, char __user *buf, size_t count)
2290 u64 read_format = event->attr.read_format;
2294 * Return end-of-file for a read on a event that is in
2295 * error state (i.e. because it was pinned but it couldn't be
2296 * scheduled on to the CPU at some point).
2298 if (event->state == PERF_EVENT_STATE_ERROR)
2301 if (count < perf_event_read_size(event))
2304 WARN_ON_ONCE(event->ctx->parent_ctx);
2305 if (read_format & PERF_FORMAT_GROUP)
2306 ret = perf_event_read_group(event, read_format, buf);
2308 ret = perf_event_read_one(event, read_format, buf);
2314 perf_read(struct file *file, char __user *buf, size_t count, loff_t *ppos)
2316 struct perf_event *event = file->private_data;
2318 return perf_read_hw(event, buf, count);
2321 static unsigned int perf_poll(struct file *file, poll_table *wait)
2323 struct perf_event *event = file->private_data;
2324 struct perf_buffer *buffer;
2325 unsigned int events = POLL_HUP;
2328 buffer = rcu_dereference(event->buffer);
2330 events = atomic_xchg(&buffer->poll, 0);
2333 poll_wait(file, &event->waitq, wait);
2338 static void perf_event_reset(struct perf_event *event)
2340 (void)perf_event_read(event);
2341 local64_set(&event->count, 0);
2342 perf_event_update_userpage(event);
2346 * Holding the top-level event's child_mutex means that any
2347 * descendant process that has inherited this event will block
2348 * in sync_child_event if it goes to exit, thus satisfying the
2349 * task existence requirements of perf_event_enable/disable.
2351 static void perf_event_for_each_child(struct perf_event *event,
2352 void (*func)(struct perf_event *))
2354 struct perf_event *child;
2356 WARN_ON_ONCE(event->ctx->parent_ctx);
2357 mutex_lock(&event->child_mutex);
2359 list_for_each_entry(child, &event->child_list, child_list)
2361 mutex_unlock(&event->child_mutex);
2364 static void perf_event_for_each(struct perf_event *event,
2365 void (*func)(struct perf_event *))
2367 struct perf_event_context *ctx = event->ctx;
2368 struct perf_event *sibling;
2370 WARN_ON_ONCE(ctx->parent_ctx);
2371 mutex_lock(&ctx->mutex);
2372 event = event->group_leader;
2374 perf_event_for_each_child(event, func);
2376 list_for_each_entry(sibling, &event->sibling_list, group_entry)
2377 perf_event_for_each_child(event, func);
2378 mutex_unlock(&ctx->mutex);
2381 static int perf_event_period(struct perf_event *event, u64 __user *arg)
2383 struct perf_event_context *ctx = event->ctx;
2388 if (!event->attr.sample_period)
2391 size = copy_from_user(&value, arg, sizeof(value));
2392 if (size != sizeof(value))
2398 raw_spin_lock_irq(&ctx->lock);
2399 if (event->attr.freq) {
2400 if (value > sysctl_perf_event_sample_rate) {
2405 event->attr.sample_freq = value;
2407 event->attr.sample_period = value;
2408 event->hw.sample_period = value;
2411 raw_spin_unlock_irq(&ctx->lock);
2416 static const struct file_operations perf_fops;
2418 static struct perf_event *perf_fget_light(int fd, int *fput_needed)
2422 file = fget_light(fd, fput_needed);
2424 return ERR_PTR(-EBADF);
2426 if (file->f_op != &perf_fops) {
2427 fput_light(file, *fput_needed);
2429 return ERR_PTR(-EBADF);
2432 return file->private_data;
2435 static int perf_event_set_output(struct perf_event *event,
2436 struct perf_event *output_event);
2437 static int perf_event_set_filter(struct perf_event *event, void __user *arg);
2439 static long perf_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
2441 struct perf_event *event = file->private_data;
2442 void (*func)(struct perf_event *);
2446 case PERF_EVENT_IOC_ENABLE:
2447 func = perf_event_enable;
2449 case PERF_EVENT_IOC_DISABLE:
2450 func = perf_event_disable;
2452 case PERF_EVENT_IOC_RESET:
2453 func = perf_event_reset;
2456 case PERF_EVENT_IOC_REFRESH:
2457 return perf_event_refresh(event, arg);
2459 case PERF_EVENT_IOC_PERIOD:
2460 return perf_event_period(event, (u64 __user *)arg);
2462 case PERF_EVENT_IOC_SET_OUTPUT:
2464 struct perf_event *output_event = NULL;
2465 int fput_needed = 0;
2469 output_event = perf_fget_light(arg, &fput_needed);
2470 if (IS_ERR(output_event))
2471 return PTR_ERR(output_event);
2474 ret = perf_event_set_output(event, output_event);
2476 fput_light(output_event->filp, fput_needed);
2481 case PERF_EVENT_IOC_SET_FILTER:
2482 return perf_event_set_filter(event, (void __user *)arg);
2488 if (flags & PERF_IOC_FLAG_GROUP)
2489 perf_event_for_each(event, func);
2491 perf_event_for_each_child(event, func);
2496 int perf_event_task_enable(void)
2498 struct perf_event *event;
2500 mutex_lock(¤t->perf_event_mutex);
2501 list_for_each_entry(event, ¤t->perf_event_list, owner_entry)
2502 perf_event_for_each_child(event, perf_event_enable);
2503 mutex_unlock(¤t->perf_event_mutex);
2508 int perf_event_task_disable(void)
2510 struct perf_event *event;
2512 mutex_lock(¤t->perf_event_mutex);
2513 list_for_each_entry(event, ¤t->perf_event_list, owner_entry)
2514 perf_event_for_each_child(event, perf_event_disable);
2515 mutex_unlock(¤t->perf_event_mutex);
2520 #ifndef PERF_EVENT_INDEX_OFFSET
2521 # define PERF_EVENT_INDEX_OFFSET 0
2524 static int perf_event_index(struct perf_event *event)
2526 if (event->hw.state & PERF_HES_STOPPED)
2529 if (event->state != PERF_EVENT_STATE_ACTIVE)
2532 return event->hw.idx + 1 - PERF_EVENT_INDEX_OFFSET;
2536 * Callers need to ensure there can be no nesting of this function, otherwise
2537 * the seqlock logic goes bad. We can not serialize this because the arch
2538 * code calls this from NMI context.
2540 void perf_event_update_userpage(struct perf_event *event)
2542 struct perf_event_mmap_page *userpg;
2543 struct perf_buffer *buffer;
2546 buffer = rcu_dereference(event->buffer);
2550 userpg = buffer->user_page;
2553 * Disable preemption so as to not let the corresponding user-space
2554 * spin too long if we get preempted.
2559 userpg->index = perf_event_index(event);
2560 userpg->offset = perf_event_count(event);
2561 if (event->state == PERF_EVENT_STATE_ACTIVE)
2562 userpg->offset -= local64_read(&event->hw.prev_count);
2564 userpg->time_enabled = event->total_time_enabled +
2565 atomic64_read(&event->child_total_time_enabled);
2567 userpg->time_running = event->total_time_running +
2568 atomic64_read(&event->child_total_time_running);
2577 static unsigned long perf_data_size(struct perf_buffer *buffer);
2580 perf_buffer_init(struct perf_buffer *buffer, long watermark, int flags)
2582 long max_size = perf_data_size(buffer);
2585 buffer->watermark = min(max_size, watermark);
2587 if (!buffer->watermark)
2588 buffer->watermark = max_size / 2;
2590 if (flags & PERF_BUFFER_WRITABLE)
2591 buffer->writable = 1;
2593 atomic_set(&buffer->refcount, 1);
2596 #ifndef CONFIG_PERF_USE_VMALLOC
2599 * Back perf_mmap() with regular GFP_KERNEL-0 pages.
2602 static struct page *
2603 perf_mmap_to_page(struct perf_buffer *buffer, unsigned long pgoff)
2605 if (pgoff > buffer->nr_pages)
2609 return virt_to_page(buffer->user_page);
2611 return virt_to_page(buffer->data_pages[pgoff - 1]);
2614 static void *perf_mmap_alloc_page(int cpu)
2619 node = (cpu == -1) ? cpu : cpu_to_node(cpu);
2620 page = alloc_pages_node(node, GFP_KERNEL | __GFP_ZERO, 0);
2624 return page_address(page);
2627 static struct perf_buffer *
2628 perf_buffer_alloc(int nr_pages, long watermark, int cpu, int flags)
2630 struct perf_buffer *buffer;
2634 size = sizeof(struct perf_buffer);
2635 size += nr_pages * sizeof(void *);
2637 buffer = kzalloc(size, GFP_KERNEL);
2641 buffer->user_page = perf_mmap_alloc_page(cpu);
2642 if (!buffer->user_page)
2643 goto fail_user_page;
2645 for (i = 0; i < nr_pages; i++) {
2646 buffer->data_pages[i] = perf_mmap_alloc_page(cpu);
2647 if (!buffer->data_pages[i])
2648 goto fail_data_pages;
2651 buffer->nr_pages = nr_pages;
2653 perf_buffer_init(buffer, watermark, flags);
2658 for (i--; i >= 0; i--)
2659 free_page((unsigned long)buffer->data_pages[i]);
2661 free_page((unsigned long)buffer->user_page);
2670 static void perf_mmap_free_page(unsigned long addr)
2672 struct page *page = virt_to_page((void *)addr);
2674 page->mapping = NULL;
2678 static void perf_buffer_free(struct perf_buffer *buffer)
2682 perf_mmap_free_page((unsigned long)buffer->user_page);
2683 for (i = 0; i < buffer->nr_pages; i++)
2684 perf_mmap_free_page((unsigned long)buffer->data_pages[i]);
2688 static inline int page_order(struct perf_buffer *buffer)
2696 * Back perf_mmap() with vmalloc memory.
2698 * Required for architectures that have d-cache aliasing issues.
2701 static inline int page_order(struct perf_buffer *buffer)
2703 return buffer->page_order;
2706 static struct page *
2707 perf_mmap_to_page(struct perf_buffer *buffer, unsigned long pgoff)
2709 if (pgoff > (1UL << page_order(buffer)))
2712 return vmalloc_to_page((void *)buffer->user_page + pgoff * PAGE_SIZE);
2715 static void perf_mmap_unmark_page(void *addr)
2717 struct page *page = vmalloc_to_page(addr);
2719 page->mapping = NULL;
2722 static void perf_buffer_free_work(struct work_struct *work)
2724 struct perf_buffer *buffer;
2728 buffer = container_of(work, struct perf_buffer, work);
2729 nr = 1 << page_order(buffer);
2731 base = buffer->user_page;
2732 for (i = 0; i < nr + 1; i++)
2733 perf_mmap_unmark_page(base + (i * PAGE_SIZE));
2739 static void perf_buffer_free(struct perf_buffer *buffer)
2741 schedule_work(&buffer->work);
2744 static struct perf_buffer *
2745 perf_buffer_alloc(int nr_pages, long watermark, int cpu, int flags)
2747 struct perf_buffer *buffer;
2751 size = sizeof(struct perf_buffer);
2752 size += sizeof(void *);
2754 buffer = kzalloc(size, GFP_KERNEL);
2758 INIT_WORK(&buffer->work, perf_buffer_free_work);
2760 all_buf = vmalloc_user((nr_pages + 1) * PAGE_SIZE);
2764 buffer->user_page = all_buf;
2765 buffer->data_pages[0] = all_buf + PAGE_SIZE;
2766 buffer->page_order = ilog2(nr_pages);
2767 buffer->nr_pages = 1;
2769 perf_buffer_init(buffer, watermark, flags);
2782 static unsigned long perf_data_size(struct perf_buffer *buffer)
2784 return buffer->nr_pages << (PAGE_SHIFT + page_order(buffer));
2787 static int perf_mmap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
2789 struct perf_event *event = vma->vm_file->private_data;
2790 struct perf_buffer *buffer;
2791 int ret = VM_FAULT_SIGBUS;
2793 if (vmf->flags & FAULT_FLAG_MKWRITE) {
2794 if (vmf->pgoff == 0)
2800 buffer = rcu_dereference(event->buffer);
2804 if (vmf->pgoff && (vmf->flags & FAULT_FLAG_WRITE))
2807 vmf->page = perf_mmap_to_page(buffer, vmf->pgoff);
2811 get_page(vmf->page);
2812 vmf->page->mapping = vma->vm_file->f_mapping;
2813 vmf->page->index = vmf->pgoff;
2822 static void perf_buffer_free_rcu(struct rcu_head *rcu_head)
2824 struct perf_buffer *buffer;
2826 buffer = container_of(rcu_head, struct perf_buffer, rcu_head);
2827 perf_buffer_free(buffer);
2830 static struct perf_buffer *perf_buffer_get(struct perf_event *event)
2832 struct perf_buffer *buffer;
2835 buffer = rcu_dereference(event->buffer);
2837 if (!atomic_inc_not_zero(&buffer->refcount))
2845 static void perf_buffer_put(struct perf_buffer *buffer)
2847 if (!atomic_dec_and_test(&buffer->refcount))
2850 call_rcu(&buffer->rcu_head, perf_buffer_free_rcu);
2853 static void perf_mmap_open(struct vm_area_struct *vma)
2855 struct perf_event *event = vma->vm_file->private_data;
2857 atomic_inc(&event->mmap_count);
2860 static void perf_mmap_close(struct vm_area_struct *vma)
2862 struct perf_event *event = vma->vm_file->private_data;
2864 if (atomic_dec_and_mutex_lock(&event->mmap_count, &event->mmap_mutex)) {
2865 unsigned long size = perf_data_size(event->buffer);
2866 struct user_struct *user = event->mmap_user;
2867 struct perf_buffer *buffer = event->buffer;
2869 atomic_long_sub((size >> PAGE_SHIFT) + 1, &user->locked_vm);
2870 vma->vm_mm->locked_vm -= event->mmap_locked;
2871 rcu_assign_pointer(event->buffer, NULL);
2872 mutex_unlock(&event->mmap_mutex);
2874 perf_buffer_put(buffer);
2879 static const struct vm_operations_struct perf_mmap_vmops = {
2880 .open = perf_mmap_open,
2881 .close = perf_mmap_close,
2882 .fault = perf_mmap_fault,
2883 .page_mkwrite = perf_mmap_fault,
2886 static int perf_mmap(struct file *file, struct vm_area_struct *vma)
2888 struct perf_event *event = file->private_data;
2889 unsigned long user_locked, user_lock_limit;
2890 struct user_struct *user = current_user();
2891 unsigned long locked, lock_limit;
2892 struct perf_buffer *buffer;
2893 unsigned long vma_size;
2894 unsigned long nr_pages;
2895 long user_extra, extra;
2896 int ret = 0, flags = 0;
2899 * Don't allow mmap() of inherited per-task counters. This would
2900 * create a performance issue due to all children writing to the
2903 if (event->cpu == -1 && event->attr.inherit)
2906 if (!(vma->vm_flags & VM_SHARED))
2909 vma_size = vma->vm_end - vma->vm_start;
2910 nr_pages = (vma_size / PAGE_SIZE) - 1;
2913 * If we have buffer pages ensure they're a power-of-two number, so we
2914 * can do bitmasks instead of modulo.
2916 if (nr_pages != 0 && !is_power_of_2(nr_pages))
2919 if (vma_size != PAGE_SIZE * (1 + nr_pages))
2922 if (vma->vm_pgoff != 0)
2925 WARN_ON_ONCE(event->ctx->parent_ctx);
2926 mutex_lock(&event->mmap_mutex);
2927 if (event->buffer) {
2928 if (event->buffer->nr_pages == nr_pages)
2929 atomic_inc(&event->buffer->refcount);
2935 user_extra = nr_pages + 1;
2936 user_lock_limit = sysctl_perf_event_mlock >> (PAGE_SHIFT - 10);
2939 * Increase the limit linearly with more CPUs:
2941 user_lock_limit *= num_online_cpus();
2943 user_locked = atomic_long_read(&user->locked_vm) + user_extra;
2946 if (user_locked > user_lock_limit)
2947 extra = user_locked - user_lock_limit;
2949 lock_limit = rlimit(RLIMIT_MEMLOCK);
2950 lock_limit >>= PAGE_SHIFT;
2951 locked = vma->vm_mm->locked_vm + extra;
2953 if ((locked > lock_limit) && perf_paranoid_tracepoint_raw() &&
2954 !capable(CAP_IPC_LOCK)) {
2959 WARN_ON(event->buffer);
2961 if (vma->vm_flags & VM_WRITE)
2962 flags |= PERF_BUFFER_WRITABLE;
2964 buffer = perf_buffer_alloc(nr_pages, event->attr.wakeup_watermark,
2970 rcu_assign_pointer(event->buffer, buffer);
2972 atomic_long_add(user_extra, &user->locked_vm);
2973 event->mmap_locked = extra;
2974 event->mmap_user = get_current_user();
2975 vma->vm_mm->locked_vm += event->mmap_locked;
2979 atomic_inc(&event->mmap_count);
2980 mutex_unlock(&event->mmap_mutex);
2982 vma->vm_flags |= VM_RESERVED;
2983 vma->vm_ops = &perf_mmap_vmops;
2988 static int perf_fasync(int fd, struct file *filp, int on)
2990 struct inode *inode = filp->f_path.dentry->d_inode;
2991 struct perf_event *event = filp->private_data;
2994 mutex_lock(&inode->i_mutex);
2995 retval = fasync_helper(fd, filp, on, &event->fasync);
2996 mutex_unlock(&inode->i_mutex);
3004 static const struct file_operations perf_fops = {
3005 .llseek = no_llseek,
3006 .release = perf_release,
3009 .unlocked_ioctl = perf_ioctl,
3010 .compat_ioctl = perf_ioctl,
3012 .fasync = perf_fasync,
3018 * If there's data, ensure we set the poll() state and publish everything
3019 * to user-space before waking everybody up.
3022 void perf_event_wakeup(struct perf_event *event)
3024 wake_up_all(&event->waitq);
3026 if (event->pending_kill) {
3027 kill_fasync(&event->fasync, SIGIO, event->pending_kill);
3028 event->pending_kill = 0;
3035 * Handle the case where we need to wakeup up from NMI (or rq->lock) context.
3037 * The NMI bit means we cannot possibly take locks. Therefore, maintain a
3038 * single linked list and use cmpxchg() to add entries lockless.
3041 static void perf_pending_event(struct perf_pending_entry *entry)
3043 struct perf_event *event = container_of(entry,
3044 struct perf_event, pending);
3046 if (event->pending_disable) {
3047 event->pending_disable = 0;
3048 __perf_event_disable(event);
3051 if (event->pending_wakeup) {
3052 event->pending_wakeup = 0;
3053 perf_event_wakeup(event);
3057 #define PENDING_TAIL ((struct perf_pending_entry *)-1UL)
3059 static DEFINE_PER_CPU(struct perf_pending_entry *, perf_pending_head) = {
3063 static void perf_pending_queue(struct perf_pending_entry *entry,
3064 void (*func)(struct perf_pending_entry *))
3066 struct perf_pending_entry **head;
3068 if (cmpxchg(&entry->next, NULL, PENDING_TAIL) != NULL)
3073 head = &get_cpu_var(perf_pending_head);
3076 entry->next = *head;
3077 } while (cmpxchg(head, entry->next, entry) != entry->next);
3079 set_perf_event_pending();
3081 put_cpu_var(perf_pending_head);
3084 static int __perf_pending_run(void)
3086 struct perf_pending_entry *list;
3089 list = xchg(&__get_cpu_var(perf_pending_head), PENDING_TAIL);
3090 while (list != PENDING_TAIL) {
3091 void (*func)(struct perf_pending_entry *);
3092 struct perf_pending_entry *entry = list;
3099 * Ensure we observe the unqueue before we issue the wakeup,
3100 * so that we won't be waiting forever.
3101 * -- see perf_not_pending().
3112 static inline int perf_not_pending(struct perf_event *event)
3115 * If we flush on whatever cpu we run, there is a chance we don't
3119 __perf_pending_run();
3123 * Ensure we see the proper queue state before going to sleep
3124 * so that we do not miss the wakeup. -- see perf_pending_handle()
3127 return event->pending.next == NULL;
3130 static void perf_pending_sync(struct perf_event *event)
3132 wait_event(event->waitq, perf_not_pending(event));
3135 void perf_event_do_pending(void)
3137 __perf_pending_run();
3141 * We assume there is only KVM supporting the callbacks.
3142 * Later on, we might change it to a list if there is
3143 * another virtualization implementation supporting the callbacks.
3145 struct perf_guest_info_callbacks *perf_guest_cbs;
3147 int perf_register_guest_info_callbacks(struct perf_guest_info_callbacks *cbs)
3149 perf_guest_cbs = cbs;
3152 EXPORT_SYMBOL_GPL(perf_register_guest_info_callbacks);
3154 int perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks *cbs)
3156 perf_guest_cbs = NULL;
3159 EXPORT_SYMBOL_GPL(perf_unregister_guest_info_callbacks);
3164 static bool perf_output_space(struct perf_buffer *buffer, unsigned long tail,
3165 unsigned long offset, unsigned long head)
3169 if (!buffer->writable)
3172 mask = perf_data_size(buffer) - 1;
3174 offset = (offset - tail) & mask;
3175 head = (head - tail) & mask;
3177 if ((int)(head - offset) < 0)
3183 static void perf_output_wakeup(struct perf_output_handle *handle)
3185 atomic_set(&handle->buffer->poll, POLL_IN);
3188 handle->event->pending_wakeup = 1;
3189 perf_pending_queue(&handle->event->pending,
3190 perf_pending_event);
3192 perf_event_wakeup(handle->event);
3196 * We need to ensure a later event_id doesn't publish a head when a former
3197 * event isn't done writing. However since we need to deal with NMIs we
3198 * cannot fully serialize things.
3200 * We only publish the head (and generate a wakeup) when the outer-most
3203 static void perf_output_get_handle(struct perf_output_handle *handle)
3205 struct perf_buffer *buffer = handle->buffer;
3208 local_inc(&buffer->nest);
3209 handle->wakeup = local_read(&buffer->wakeup);
3212 static void perf_output_put_handle(struct perf_output_handle *handle)
3214 struct perf_buffer *buffer = handle->buffer;
3218 head = local_read(&buffer->head);
3221 * IRQ/NMI can happen here, which means we can miss a head update.
3224 if (!local_dec_and_test(&buffer->nest))
3228 * Publish the known good head. Rely on the full barrier implied
3229 * by atomic_dec_and_test() order the buffer->head read and this
3232 buffer->user_page->data_head = head;
3235 * Now check if we missed an update, rely on the (compiler)
3236 * barrier in atomic_dec_and_test() to re-read buffer->head.
3238 if (unlikely(head != local_read(&buffer->head))) {
3239 local_inc(&buffer->nest);
3243 if (handle->wakeup != local_read(&buffer->wakeup))
3244 perf_output_wakeup(handle);
3250 __always_inline void perf_output_copy(struct perf_output_handle *handle,
3251 const void *buf, unsigned int len)
3254 unsigned long size = min_t(unsigned long, handle->size, len);
3256 memcpy(handle->addr, buf, size);
3259 handle->addr += size;
3261 handle->size -= size;
3262 if (!handle->size) {
3263 struct perf_buffer *buffer = handle->buffer;
3266 handle->page &= buffer->nr_pages - 1;
3267 handle->addr = buffer->data_pages[handle->page];
3268 handle->size = PAGE_SIZE << page_order(buffer);
3273 int perf_output_begin(struct perf_output_handle *handle,
3274 struct perf_event *event, unsigned int size,
3275 int nmi, int sample)
3277 struct perf_buffer *buffer;
3278 unsigned long tail, offset, head;
3281 struct perf_event_header header;
3288 * For inherited events we send all the output towards the parent.
3291 event = event->parent;
3293 buffer = rcu_dereference(event->buffer);
3297 handle->buffer = buffer;
3298 handle->event = event;
3300 handle->sample = sample;
3302 if (!buffer->nr_pages)
3305 have_lost = local_read(&buffer->lost);
3307 size += sizeof(lost_event);
3309 perf_output_get_handle(handle);
3313 * Userspace could choose to issue a mb() before updating the
3314 * tail pointer. So that all reads will be completed before the
3317 tail = ACCESS_ONCE(buffer->user_page->data_tail);
3319 offset = head = local_read(&buffer->head);
3321 if (unlikely(!perf_output_space(buffer, tail, offset, head)))
3323 } while (local_cmpxchg(&buffer->head, offset, head) != offset);
3325 if (head - local_read(&buffer->wakeup) > buffer->watermark)
3326 local_add(buffer->watermark, &buffer->wakeup);
3328 handle->page = offset >> (PAGE_SHIFT + page_order(buffer));
3329 handle->page &= buffer->nr_pages - 1;
3330 handle->size = offset & ((PAGE_SIZE << page_order(buffer)) - 1);
3331 handle->addr = buffer->data_pages[handle->page];
3332 handle->addr += handle->size;
3333 handle->size = (PAGE_SIZE << page_order(buffer)) - handle->size;
3336 lost_event.header.type = PERF_RECORD_LOST;
3337 lost_event.header.misc = 0;
3338 lost_event.header.size = sizeof(lost_event);
3339 lost_event.id = event->id;
3340 lost_event.lost = local_xchg(&buffer->lost, 0);
3342 perf_output_put(handle, lost_event);
3348 local_inc(&buffer->lost);
3349 perf_output_put_handle(handle);
3356 void perf_output_end(struct perf_output_handle *handle)
3358 struct perf_event *event = handle->event;
3359 struct perf_buffer *buffer = handle->buffer;
3361 int wakeup_events = event->attr.wakeup_events;
3363 if (handle->sample && wakeup_events) {
3364 int events = local_inc_return(&buffer->events);
3365 if (events >= wakeup_events) {
3366 local_sub(wakeup_events, &buffer->events);
3367 local_inc(&buffer->wakeup);
3371 perf_output_put_handle(handle);
3375 static u32 perf_event_pid(struct perf_event *event, struct task_struct *p)
3378 * only top level events have the pid namespace they were created in
3381 event = event->parent;
3383 return task_tgid_nr_ns(p, event->ns);
3386 static u32 perf_event_tid(struct perf_event *event, struct task_struct *p)
3389 * only top level events have the pid namespace they were created in
3392 event = event->parent;
3394 return task_pid_nr_ns(p, event->ns);
3397 static void perf_output_read_one(struct perf_output_handle *handle,
3398 struct perf_event *event)
3400 u64 read_format = event->attr.read_format;
3404 values[n++] = perf_event_count(event);
3405 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
3406 values[n++] = event->total_time_enabled +
3407 atomic64_read(&event->child_total_time_enabled);
3409 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
3410 values[n++] = event->total_time_running +
3411 atomic64_read(&event->child_total_time_running);
3413 if (read_format & PERF_FORMAT_ID)
3414 values[n++] = primary_event_id(event);
3416 perf_output_copy(handle, values, n * sizeof(u64));
3420 * XXX PERF_FORMAT_GROUP vs inherited events seems difficult.
3422 static void perf_output_read_group(struct perf_output_handle *handle,
3423 struct perf_event *event)
3425 struct perf_event *leader = event->group_leader, *sub;
3426 u64 read_format = event->attr.read_format;
3430 values[n++] = 1 + leader->nr_siblings;
3432 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
3433 values[n++] = leader->total_time_enabled;
3435 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
3436 values[n++] = leader->total_time_running;
3438 if (leader != event)
3439 leader->pmu->read(leader);
3441 values[n++] = perf_event_count(leader);
3442 if (read_format & PERF_FORMAT_ID)
3443 values[n++] = primary_event_id(leader);
3445 perf_output_copy(handle, values, n * sizeof(u64));
3447 list_for_each_entry(sub, &leader->sibling_list, group_entry) {
3451 sub->pmu->read(sub);
3453 values[n++] = perf_event_count(sub);
3454 if (read_format & PERF_FORMAT_ID)
3455 values[n++] = primary_event_id(sub);
3457 perf_output_copy(handle, values, n * sizeof(u64));
3461 static void perf_output_read(struct perf_output_handle *handle,
3462 struct perf_event *event)
3464 if (event->attr.read_format & PERF_FORMAT_GROUP)
3465 perf_output_read_group(handle, event);
3467 perf_output_read_one(handle, event);
3470 void perf_output_sample(struct perf_output_handle *handle,
3471 struct perf_event_header *header,
3472 struct perf_sample_data *data,
3473 struct perf_event *event)
3475 u64 sample_type = data->type;
3477 perf_output_put(handle, *header);
3479 if (sample_type & PERF_SAMPLE_IP)
3480 perf_output_put(handle, data->ip);
3482 if (sample_type & PERF_SAMPLE_TID)
3483 perf_output_put(handle, data->tid_entry);
3485 if (sample_type & PERF_SAMPLE_TIME)
3486 perf_output_put(handle, data->time);
3488 if (sample_type & PERF_SAMPLE_ADDR)
3489 perf_output_put(handle, data->addr);
3491 if (sample_type & PERF_SAMPLE_ID)
3492 perf_output_put(handle, data->id);
3494 if (sample_type & PERF_SAMPLE_STREAM_ID)
3495 perf_output_put(handle, data->stream_id);
3497 if (sample_type & PERF_SAMPLE_CPU)
3498 perf_output_put(handle, data->cpu_entry);
3500 if (sample_type & PERF_SAMPLE_PERIOD)
3501 perf_output_put(handle, data->period);
3503 if (sample_type & PERF_SAMPLE_READ)
3504 perf_output_read(handle, event);
3506 if (sample_type & PERF_SAMPLE_CALLCHAIN) {
3507 if (data->callchain) {
3510 if (data->callchain)
3511 size += data->callchain->nr;
3513 size *= sizeof(u64);
3515 perf_output_copy(handle, data->callchain, size);
3518 perf_output_put(handle, nr);
3522 if (sample_type & PERF_SAMPLE_RAW) {
3524 perf_output_put(handle, data->raw->size);
3525 perf_output_copy(handle, data->raw->data,
3532 .size = sizeof(u32),
3535 perf_output_put(handle, raw);
3540 void perf_prepare_sample(struct perf_event_header *header,
3541 struct perf_sample_data *data,
3542 struct perf_event *event,
3543 struct pt_regs *regs)
3545 u64 sample_type = event->attr.sample_type;
3547 data->type = sample_type;
3549 header->type = PERF_RECORD_SAMPLE;
3550 header->size = sizeof(*header);
3553 header->misc |= perf_misc_flags(regs);
3555 if (sample_type & PERF_SAMPLE_IP) {
3556 data->ip = perf_instruction_pointer(regs);
3558 header->size += sizeof(data->ip);
3561 if (sample_type & PERF_SAMPLE_TID) {
3562 /* namespace issues */
3563 data->tid_entry.pid = perf_event_pid(event, current);
3564 data->tid_entry.tid = perf_event_tid(event, current);
3566 header->size += sizeof(data->tid_entry);
3569 if (sample_type & PERF_SAMPLE_TIME) {
3570 data->time = perf_clock();
3572 header->size += sizeof(data->time);
3575 if (sample_type & PERF_SAMPLE_ADDR)
3576 header->size += sizeof(data->addr);
3578 if (sample_type & PERF_SAMPLE_ID) {
3579 data->id = primary_event_id(event);
3581 header->size += sizeof(data->id);
3584 if (sample_type & PERF_SAMPLE_STREAM_ID) {
3585 data->stream_id = event->id;
3587 header->size += sizeof(data->stream_id);
3590 if (sample_type & PERF_SAMPLE_CPU) {
3591 data->cpu_entry.cpu = raw_smp_processor_id();
3592 data->cpu_entry.reserved = 0;
3594 header->size += sizeof(data->cpu_entry);
3597 if (sample_type & PERF_SAMPLE_PERIOD)
3598 header->size += sizeof(data->period);
3600 if (sample_type & PERF_SAMPLE_READ)
3601 header->size += perf_event_read_size(event);
3603 if (sample_type & PERF_SAMPLE_CALLCHAIN) {
3606 data->callchain = perf_callchain(regs);
3608 if (data->callchain)
3609 size += data->callchain->nr;
3611 header->size += size * sizeof(u64);
3614 if (sample_type & PERF_SAMPLE_RAW) {
3615 int size = sizeof(u32);
3618 size += data->raw->size;
3620 size += sizeof(u32);
3622 WARN_ON_ONCE(size & (sizeof(u64)-1));
3623 header->size += size;
3627 static void perf_event_output(struct perf_event *event, int nmi,
3628 struct perf_sample_data *data,
3629 struct pt_regs *regs)
3631 struct perf_output_handle handle;
3632 struct perf_event_header header;
3634 /* protect the callchain buffers */
3637 perf_prepare_sample(&header, data, event, regs);
3639 if (perf_output_begin(&handle, event, header.size, nmi, 1))
3642 perf_output_sample(&handle, &header, data, event);
3644 perf_output_end(&handle);
3654 struct perf_read_event {
3655 struct perf_event_header header;
3662 perf_event_read_event(struct perf_event *event,
3663 struct task_struct *task)
3665 struct perf_output_handle handle;
3666 struct perf_read_event read_event = {
3668 .type = PERF_RECORD_READ,
3670 .size = sizeof(read_event) + perf_event_read_size(event),
3672 .pid = perf_event_pid(event, task),
3673 .tid = perf_event_tid(event, task),
3677 ret = perf_output_begin(&handle, event, read_event.header.size, 0, 0);
3681 perf_output_put(&handle, read_event);
3682 perf_output_read(&handle, event);
3684 perf_output_end(&handle);
3688 * task tracking -- fork/exit
3690 * enabled by: attr.comm | attr.mmap | attr.mmap_data | attr.task
3693 struct perf_task_event {
3694 struct task_struct *task;
3695 struct perf_event_context *task_ctx;
3698 struct perf_event_header header;
3708 static void perf_event_task_output(struct perf_event *event,
3709 struct perf_task_event *task_event)
3711 struct perf_output_handle handle;
3712 struct task_struct *task = task_event->task;
3715 size = task_event->event_id.header.size;
3716 ret = perf_output_begin(&handle, event, size, 0, 0);
3721 task_event->event_id.pid = perf_event_pid(event, task);
3722 task_event->event_id.ppid = perf_event_pid(event, current);
3724 task_event->event_id.tid = perf_event_tid(event, task);
3725 task_event->event_id.ptid = perf_event_tid(event, current);
3727 perf_output_put(&handle, task_event->event_id);
3729 perf_output_end(&handle);
3732 static int perf_event_task_match(struct perf_event *event)
3734 if (event->state < PERF_EVENT_STATE_INACTIVE)
3737 if (event->cpu != -1 && event->cpu != smp_processor_id())
3740 if (event->attr.comm || event->attr.mmap ||
3741 event->attr.mmap_data || event->attr.task)
3747 static void perf_event_task_ctx(struct perf_event_context *ctx,
3748 struct perf_task_event *task_event)
3750 struct perf_event *event;
3752 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
3753 if (perf_event_task_match(event))
3754 perf_event_task_output(event, task_event);
3758 static void perf_event_task_event(struct perf_task_event *task_event)
3760 struct perf_event_context *ctx = task_event->task_ctx;
3761 struct perf_cpu_context *cpuctx;
3764 rcu_read_lock_sched();
3765 list_for_each_entry_rcu(pmu, &pmus, entry) {
3766 cpuctx = this_cpu_ptr(pmu->pmu_cpu_context);
3767 perf_event_task_ctx(&cpuctx->ctx, task_event);
3770 ctx = rcu_dereference(current->perf_event_ctxp);
3772 perf_event_task_ctx(ctx, task_event);
3773 rcu_read_unlock_sched();
3776 static void perf_event_task(struct task_struct *task,
3777 struct perf_event_context *task_ctx,
3780 struct perf_task_event task_event;
3782 if (!atomic_read(&nr_comm_events) &&
3783 !atomic_read(&nr_mmap_events) &&
3784 !atomic_read(&nr_task_events))
3787 task_event = (struct perf_task_event){
3789 .task_ctx = task_ctx,
3792 .type = new ? PERF_RECORD_FORK : PERF_RECORD_EXIT,
3794 .size = sizeof(task_event.event_id),
3800 .time = perf_clock(),
3804 perf_event_task_event(&task_event);
3807 void perf_event_fork(struct task_struct *task)
3809 perf_event_task(task, NULL, 1);
3816 struct perf_comm_event {
3817 struct task_struct *task;
3822 struct perf_event_header header;
3829 static void perf_event_comm_output(struct perf_event *event,
3830 struct perf_comm_event *comm_event)
3832 struct perf_output_handle handle;
3833 int size = comm_event->event_id.header.size;
3834 int ret = perf_output_begin(&handle, event, size, 0, 0);
3839 comm_event->event_id.pid = perf_event_pid(event, comm_event->task);
3840 comm_event->event_id.tid = perf_event_tid(event, comm_event->task);
3842 perf_output_put(&handle, comm_event->event_id);
3843 perf_output_copy(&handle, comm_event->comm,
3844 comm_event->comm_size);
3845 perf_output_end(&handle);
3848 static int perf_event_comm_match(struct perf_event *event)
3850 if (event->state < PERF_EVENT_STATE_INACTIVE)
3853 if (event->cpu != -1 && event->cpu != smp_processor_id())
3856 if (event->attr.comm)
3862 static void perf_event_comm_ctx(struct perf_event_context *ctx,
3863 struct perf_comm_event *comm_event)
3865 struct perf_event *event;
3867 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
3868 if (perf_event_comm_match(event))
3869 perf_event_comm_output(event, comm_event);
3873 static void perf_event_comm_event(struct perf_comm_event *comm_event)
3875 struct perf_cpu_context *cpuctx;
3876 struct perf_event_context *ctx;
3879 char comm[TASK_COMM_LEN];
3881 memset(comm, 0, sizeof(comm));
3882 strlcpy(comm, comm_event->task->comm, sizeof(comm));
3883 size = ALIGN(strlen(comm)+1, sizeof(u64));
3885 comm_event->comm = comm;
3886 comm_event->comm_size = size;
3888 comm_event->event_id.header.size = sizeof(comm_event->event_id) + size;
3890 rcu_read_lock_sched();
3891 list_for_each_entry_rcu(pmu, &pmus, entry) {
3892 cpuctx = this_cpu_ptr(pmu->pmu_cpu_context);
3893 perf_event_comm_ctx(&cpuctx->ctx, comm_event);
3895 ctx = rcu_dereference(current->perf_event_ctxp);
3897 perf_event_comm_ctx(ctx, comm_event);
3898 rcu_read_unlock_sched();
3901 void perf_event_comm(struct task_struct *task)
3903 struct perf_comm_event comm_event;
3905 if (task->perf_event_ctxp)
3906 perf_event_enable_on_exec(task);
3908 if (!atomic_read(&nr_comm_events))
3911 comm_event = (struct perf_comm_event){
3917 .type = PERF_RECORD_COMM,
3926 perf_event_comm_event(&comm_event);
3933 struct perf_mmap_event {
3934 struct vm_area_struct *vma;
3936 const char *file_name;
3940 struct perf_event_header header;
3950 static void perf_event_mmap_output(struct perf_event *event,
3951 struct perf_mmap_event *mmap_event)
3953 struct perf_output_handle handle;
3954 int size = mmap_event->event_id.header.size;
3955 int ret = perf_output_begin(&handle, event, size, 0, 0);
3960 mmap_event->event_id.pid = perf_event_pid(event, current);
3961 mmap_event->event_id.tid = perf_event_tid(event, current);
3963 perf_output_put(&handle, mmap_event->event_id);
3964 perf_output_copy(&handle, mmap_event->file_name,
3965 mmap_event->file_size);
3966 perf_output_end(&handle);
3969 static int perf_event_mmap_match(struct perf_event *event,
3970 struct perf_mmap_event *mmap_event,
3973 if (event->state < PERF_EVENT_STATE_INACTIVE)
3976 if (event->cpu != -1 && event->cpu != smp_processor_id())
3979 if ((!executable && event->attr.mmap_data) ||
3980 (executable && event->attr.mmap))
3986 static void perf_event_mmap_ctx(struct perf_event_context *ctx,
3987 struct perf_mmap_event *mmap_event,
3990 struct perf_event *event;
3992 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
3993 if (perf_event_mmap_match(event, mmap_event, executable))
3994 perf_event_mmap_output(event, mmap_event);
3998 static void perf_event_mmap_event(struct perf_mmap_event *mmap_event)
4000 struct perf_cpu_context *cpuctx;
4001 struct perf_event_context *ctx;
4002 struct vm_area_struct *vma = mmap_event->vma;
4003 struct file *file = vma->vm_file;
4010 memset(tmp, 0, sizeof(tmp));
4014 * d_path works from the end of the buffer backwards, so we
4015 * need to add enough zero bytes after the string to handle
4016 * the 64bit alignment we do later.
4018 buf = kzalloc(PATH_MAX + sizeof(u64), GFP_KERNEL);
4020 name = strncpy(tmp, "//enomem", sizeof(tmp));
4023 name = d_path(&file->f_path, buf, PATH_MAX);
4025 name = strncpy(tmp, "//toolong", sizeof(tmp));
4029 if (arch_vma_name(mmap_event->vma)) {
4030 name = strncpy(tmp, arch_vma_name(mmap_event->vma),
4036 name = strncpy(tmp, "[vdso]", sizeof(tmp));
4038 } else if (vma->vm_start <= vma->vm_mm->start_brk &&
4039 vma->vm_end >= vma->vm_mm->brk) {
4040 name = strncpy(tmp, "[heap]", sizeof(tmp));
4042 } else if (vma->vm_start <= vma->vm_mm->start_stack &&
4043 vma->vm_end >= vma->vm_mm->start_stack) {
4044 name = strncpy(tmp, "[stack]", sizeof(tmp));
4048 name = strncpy(tmp, "//anon", sizeof(tmp));
4053 size = ALIGN(strlen(name)+1, sizeof(u64));
4055 mmap_event->file_name = name;
4056 mmap_event->file_size = size;
4058 mmap_event->event_id.header.size = sizeof(mmap_event->event_id) + size;
4060 rcu_read_lock_sched();
4061 list_for_each_entry_rcu(pmu, &pmus, entry) {
4062 cpuctx = this_cpu_ptr(pmu->pmu_cpu_context);
4063 perf_event_mmap_ctx(&cpuctx->ctx, mmap_event,
4064 vma->vm_flags & VM_EXEC);
4066 ctx = rcu_dereference(current->perf_event_ctxp);
4068 perf_event_mmap_ctx(ctx, mmap_event, vma->vm_flags & VM_EXEC);
4069 rcu_read_unlock_sched();
4074 void perf_event_mmap(struct vm_area_struct *vma)
4076 struct perf_mmap_event mmap_event;
4078 if (!atomic_read(&nr_mmap_events))
4081 mmap_event = (struct perf_mmap_event){
4087 .type = PERF_RECORD_MMAP,
4088 .misc = PERF_RECORD_MISC_USER,
4093 .start = vma->vm_start,
4094 .len = vma->vm_end - vma->vm_start,
4095 .pgoff = (u64)vma->vm_pgoff << PAGE_SHIFT,
4099 perf_event_mmap_event(&mmap_event);
4103 * IRQ throttle logging
4106 static void perf_log_throttle(struct perf_event *event, int enable)
4108 struct perf_output_handle handle;
4112 struct perf_event_header header;
4116 } throttle_event = {
4118 .type = PERF_RECORD_THROTTLE,
4120 .size = sizeof(throttle_event),
4122 .time = perf_clock(),
4123 .id = primary_event_id(event),
4124 .stream_id = event->id,
4128 throttle_event.header.type = PERF_RECORD_UNTHROTTLE;
4130 ret = perf_output_begin(&handle, event, sizeof(throttle_event), 1, 0);
4134 perf_output_put(&handle, throttle_event);
4135 perf_output_end(&handle);
4139 * Generic event overflow handling, sampling.
4142 static int __perf_event_overflow(struct perf_event *event, int nmi,
4143 int throttle, struct perf_sample_data *data,
4144 struct pt_regs *regs)
4146 int events = atomic_read(&event->event_limit);
4147 struct hw_perf_event *hwc = &event->hw;
4153 if (hwc->interrupts != MAX_INTERRUPTS) {
4155 if (HZ * hwc->interrupts >
4156 (u64)sysctl_perf_event_sample_rate) {
4157 hwc->interrupts = MAX_INTERRUPTS;
4158 perf_log_throttle(event, 0);
4163 * Keep re-disabling events even though on the previous
4164 * pass we disabled it - just in case we raced with a
4165 * sched-in and the event got enabled again:
4171 if (event->attr.freq) {
4172 u64 now = perf_clock();
4173 s64 delta = now - hwc->freq_time_stamp;
4175 hwc->freq_time_stamp = now;
4177 if (delta > 0 && delta < 2*TICK_NSEC)
4178 perf_adjust_period(event, delta, hwc->last_period);
4182 * XXX event_limit might not quite work as expected on inherited
4186 event->pending_kill = POLL_IN;
4187 if (events && atomic_dec_and_test(&event->event_limit)) {
4189 event->pending_kill = POLL_HUP;
4191 event->pending_disable = 1;
4192 perf_pending_queue(&event->pending,
4193 perf_pending_event);
4195 perf_event_disable(event);
4198 if (event->overflow_handler)
4199 event->overflow_handler(event, nmi, data, regs);
4201 perf_event_output(event, nmi, data, regs);
4206 int perf_event_overflow(struct perf_event *event, int nmi,
4207 struct perf_sample_data *data,
4208 struct pt_regs *regs)
4210 return __perf_event_overflow(event, nmi, 1, data, regs);
4214 * Generic software event infrastructure
4217 struct swevent_htable {
4218 struct swevent_hlist *swevent_hlist;
4219 struct mutex hlist_mutex;
4222 /* Recursion avoidance in each contexts */
4223 int recursion[PERF_NR_CONTEXTS];
4226 static DEFINE_PER_CPU(struct swevent_htable, swevent_htable);
4229 * We directly increment event->count and keep a second value in
4230 * event->hw.period_left to count intervals. This period event
4231 * is kept in the range [-sample_period, 0] so that we can use the
4235 static u64 perf_swevent_set_period(struct perf_event *event)
4237 struct hw_perf_event *hwc = &event->hw;
4238 u64 period = hwc->last_period;
4242 hwc->last_period = hwc->sample_period;
4245 old = val = local64_read(&hwc->period_left);
4249 nr = div64_u64(period + val, period);
4250 offset = nr * period;
4252 if (local64_cmpxchg(&hwc->period_left, old, val) != old)
4258 static void perf_swevent_overflow(struct perf_event *event, u64 overflow,
4259 int nmi, struct perf_sample_data *data,
4260 struct pt_regs *regs)
4262 struct hw_perf_event *hwc = &event->hw;
4265 data->period = event->hw.last_period;
4267 overflow = perf_swevent_set_period(event);
4269 if (hwc->interrupts == MAX_INTERRUPTS)
4272 for (; overflow; overflow--) {
4273 if (__perf_event_overflow(event, nmi, throttle,
4276 * We inhibit the overflow from happening when
4277 * hwc->interrupts == MAX_INTERRUPTS.
4285 static void perf_swevent_event(struct perf_event *event, u64 nr,
4286 int nmi, struct perf_sample_data *data,
4287 struct pt_regs *regs)
4289 struct hw_perf_event *hwc = &event->hw;
4291 local64_add(nr, &event->count);
4296 if (!hwc->sample_period)
4299 if (nr == 1 && hwc->sample_period == 1 && !event->attr.freq)
4300 return perf_swevent_overflow(event, 1, nmi, data, regs);
4302 if (local64_add_negative(nr, &hwc->period_left))
4305 perf_swevent_overflow(event, 0, nmi, data, regs);
4308 static int perf_exclude_event(struct perf_event *event,
4309 struct pt_regs *regs)
4311 if (event->hw.state & PERF_HES_STOPPED)
4315 if (event->attr.exclude_user && user_mode(regs))
4318 if (event->attr.exclude_kernel && !user_mode(regs))
4325 static int perf_swevent_match(struct perf_event *event,
4326 enum perf_type_id type,
4328 struct perf_sample_data *data,
4329 struct pt_regs *regs)
4331 if (event->attr.type != type)
4334 if (event->attr.config != event_id)
4337 if (perf_exclude_event(event, regs))
4343 static inline u64 swevent_hash(u64 type, u32 event_id)
4345 u64 val = event_id | (type << 32);
4347 return hash_64(val, SWEVENT_HLIST_BITS);
4350 static inline struct hlist_head *
4351 __find_swevent_head(struct swevent_hlist *hlist, u64 type, u32 event_id)
4353 u64 hash = swevent_hash(type, event_id);
4355 return &hlist->heads[hash];
4358 /* For the read side: events when they trigger */
4359 static inline struct hlist_head *
4360 find_swevent_head_rcu(struct swevent_htable *swhash, u64 type, u32 event_id)
4362 struct swevent_hlist *hlist;
4364 hlist = rcu_dereference(swhash->swevent_hlist);
4368 return __find_swevent_head(hlist, type, event_id);
4371 /* For the event head insertion and removal in the hlist */
4372 static inline struct hlist_head *
4373 find_swevent_head(struct swevent_htable *swhash, struct perf_event *event)
4375 struct swevent_hlist *hlist;
4376 u32 event_id = event->attr.config;
4377 u64 type = event->attr.type;
4380 * Event scheduling is always serialized against hlist allocation
4381 * and release. Which makes the protected version suitable here.
4382 * The context lock guarantees that.
4384 hlist = rcu_dereference_protected(swhash->swevent_hlist,
4385 lockdep_is_held(&event->ctx->lock));
4389 return __find_swevent_head(hlist, type, event_id);
4392 static void do_perf_sw_event(enum perf_type_id type, u32 event_id,
4394 struct perf_sample_data *data,
4395 struct pt_regs *regs)
4397 struct swevent_htable *swhash = &__get_cpu_var(swevent_htable);
4398 struct perf_event *event;
4399 struct hlist_node *node;
4400 struct hlist_head *head;
4403 head = find_swevent_head_rcu(swhash, type, event_id);
4407 hlist_for_each_entry_rcu(event, node, head, hlist_entry) {
4408 if (perf_swevent_match(event, type, event_id, data, regs))
4409 perf_swevent_event(event, nr, nmi, data, regs);
4415 int perf_swevent_get_recursion_context(void)
4417 struct swevent_htable *swhash = &__get_cpu_var(swevent_htable);
4419 return get_recursion_context(swhash->recursion);
4421 EXPORT_SYMBOL_GPL(perf_swevent_get_recursion_context);
4423 void inline perf_swevent_put_recursion_context(int rctx)
4425 struct swevent_htable *swhash = &__get_cpu_var(swevent_htable);
4427 put_recursion_context(swhash->recursion, rctx);
4430 void __perf_sw_event(u32 event_id, u64 nr, int nmi,
4431 struct pt_regs *regs, u64 addr)
4433 struct perf_sample_data data;
4436 preempt_disable_notrace();
4437 rctx = perf_swevent_get_recursion_context();
4441 perf_sample_data_init(&data, addr);
4443 do_perf_sw_event(PERF_TYPE_SOFTWARE, event_id, nr, nmi, &data, regs);
4445 perf_swevent_put_recursion_context(rctx);
4446 preempt_enable_notrace();
4449 static void perf_swevent_read(struct perf_event *event)
4453 static int perf_swevent_add(struct perf_event *event, int flags)
4455 struct swevent_htable *swhash = &__get_cpu_var(swevent_htable);
4456 struct hw_perf_event *hwc = &event->hw;
4457 struct hlist_head *head;
4459 if (hwc->sample_period) {
4460 hwc->last_period = hwc->sample_period;
4461 perf_swevent_set_period(event);
4464 hwc->state = !(flags & PERF_EF_START);
4466 head = find_swevent_head(swhash, event);
4467 if (WARN_ON_ONCE(!head))
4470 hlist_add_head_rcu(&event->hlist_entry, head);
4475 static void perf_swevent_del(struct perf_event *event, int flags)
4477 hlist_del_rcu(&event->hlist_entry);
4480 static void perf_swevent_start(struct perf_event *event, int flags)
4482 event->hw.state = 0;
4485 static void perf_swevent_stop(struct perf_event *event, int flags)
4487 event->hw.state = PERF_HES_STOPPED;
4490 /* Deref the hlist from the update side */
4491 static inline struct swevent_hlist *
4492 swevent_hlist_deref(struct swevent_htable *swhash)
4494 return rcu_dereference_protected(swhash->swevent_hlist,
4495 lockdep_is_held(&swhash->hlist_mutex));
4498 static void swevent_hlist_release_rcu(struct rcu_head *rcu_head)
4500 struct swevent_hlist *hlist;
4502 hlist = container_of(rcu_head, struct swevent_hlist, rcu_head);
4506 static void swevent_hlist_release(struct swevent_htable *swhash)
4508 struct swevent_hlist *hlist = swevent_hlist_deref(swhash);
4513 rcu_assign_pointer(swhash->swevent_hlist, NULL);
4514 call_rcu(&hlist->rcu_head, swevent_hlist_release_rcu);
4517 static void swevent_hlist_put_cpu(struct perf_event *event, int cpu)
4519 struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
4521 mutex_lock(&swhash->hlist_mutex);
4523 if (!--swhash->hlist_refcount)
4524 swevent_hlist_release(swhash);
4526 mutex_unlock(&swhash->hlist_mutex);
4529 static void swevent_hlist_put(struct perf_event *event)
4533 if (event->cpu != -1) {
4534 swevent_hlist_put_cpu(event, event->cpu);
4538 for_each_possible_cpu(cpu)
4539 swevent_hlist_put_cpu(event, cpu);
4542 static int swevent_hlist_get_cpu(struct perf_event *event, int cpu)
4544 struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
4547 mutex_lock(&swhash->hlist_mutex);
4549 if (!swevent_hlist_deref(swhash) && cpu_online(cpu)) {
4550 struct swevent_hlist *hlist;
4552 hlist = kzalloc(sizeof(*hlist), GFP_KERNEL);
4557 rcu_assign_pointer(swhash->swevent_hlist, hlist);
4559 swhash->hlist_refcount++;
4561 mutex_unlock(&swhash->hlist_mutex);
4566 static int swevent_hlist_get(struct perf_event *event)
4569 int cpu, failed_cpu;
4571 if (event->cpu != -1)
4572 return swevent_hlist_get_cpu(event, event->cpu);
4575 for_each_possible_cpu(cpu) {
4576 err = swevent_hlist_get_cpu(event, cpu);
4586 for_each_possible_cpu(cpu) {
4587 if (cpu == failed_cpu)
4589 swevent_hlist_put_cpu(event, cpu);
4596 atomic_t perf_swevent_enabled[PERF_COUNT_SW_MAX];
4598 static void sw_perf_event_destroy(struct perf_event *event)
4600 u64 event_id = event->attr.config;
4602 WARN_ON(event->parent);
4604 atomic_dec(&perf_swevent_enabled[event_id]);
4605 swevent_hlist_put(event);
4608 static int perf_swevent_init(struct perf_event *event)
4610 int event_id = event->attr.config;
4612 if (event->attr.type != PERF_TYPE_SOFTWARE)
4616 case PERF_COUNT_SW_CPU_CLOCK:
4617 case PERF_COUNT_SW_TASK_CLOCK:
4624 if (event_id > PERF_COUNT_SW_MAX)
4627 if (!event->parent) {
4630 err = swevent_hlist_get(event);
4634 atomic_inc(&perf_swevent_enabled[event_id]);
4635 event->destroy = sw_perf_event_destroy;
4641 static struct pmu perf_swevent = {
4642 .event_init = perf_swevent_init,
4643 .add = perf_swevent_add,
4644 .del = perf_swevent_del,
4645 .start = perf_swevent_start,
4646 .stop = perf_swevent_stop,
4647 .read = perf_swevent_read,
4650 #ifdef CONFIG_EVENT_TRACING
4652 static int perf_tp_filter_match(struct perf_event *event,
4653 struct perf_sample_data *data)
4655 void *record = data->raw->data;
4657 if (likely(!event->filter) || filter_match_preds(event->filter, record))
4662 static int perf_tp_event_match(struct perf_event *event,
4663 struct perf_sample_data *data,
4664 struct pt_regs *regs)
4667 * All tracepoints are from kernel-space.
4669 if (event->attr.exclude_kernel)
4672 if (!perf_tp_filter_match(event, data))
4678 void perf_tp_event(u64 addr, u64 count, void *record, int entry_size,
4679 struct pt_regs *regs, struct hlist_head *head, int rctx)
4681 struct perf_sample_data data;
4682 struct perf_event *event;
4683 struct hlist_node *node;
4685 struct perf_raw_record raw = {
4690 perf_sample_data_init(&data, addr);
4693 hlist_for_each_entry_rcu(event, node, head, hlist_entry) {
4694 if (perf_tp_event_match(event, &data, regs))
4695 perf_swevent_event(event, count, 1, &data, regs);
4698 perf_swevent_put_recursion_context(rctx);
4700 EXPORT_SYMBOL_GPL(perf_tp_event);
4702 static void tp_perf_event_destroy(struct perf_event *event)
4704 perf_trace_destroy(event);
4707 static int perf_tp_event_init(struct perf_event *event)
4711 if (event->attr.type != PERF_TYPE_TRACEPOINT)
4715 * Raw tracepoint data is a severe data leak, only allow root to
4718 if ((event->attr.sample_type & PERF_SAMPLE_RAW) &&
4719 perf_paranoid_tracepoint_raw() &&
4720 !capable(CAP_SYS_ADMIN))
4723 err = perf_trace_init(event);
4727 event->destroy = tp_perf_event_destroy;
4732 static struct pmu perf_tracepoint = {
4733 .event_init = perf_tp_event_init,
4734 .add = perf_trace_add,
4735 .del = perf_trace_del,
4736 .start = perf_swevent_start,
4737 .stop = perf_swevent_stop,
4738 .read = perf_swevent_read,
4741 static inline void perf_tp_register(void)
4743 perf_pmu_register(&perf_tracepoint);
4746 static int perf_event_set_filter(struct perf_event *event, void __user *arg)
4751 if (event->attr.type != PERF_TYPE_TRACEPOINT)
4754 filter_str = strndup_user(arg, PAGE_SIZE);
4755 if (IS_ERR(filter_str))
4756 return PTR_ERR(filter_str);
4758 ret = ftrace_profile_set_filter(event, event->attr.config, filter_str);
4764 static void perf_event_free_filter(struct perf_event *event)
4766 ftrace_profile_free_filter(event);
4771 static inline void perf_tp_register(void)
4775 static int perf_event_set_filter(struct perf_event *event, void __user *arg)
4780 static void perf_event_free_filter(struct perf_event *event)
4784 #endif /* CONFIG_EVENT_TRACING */
4786 #ifdef CONFIG_HAVE_HW_BREAKPOINT
4787 void perf_bp_event(struct perf_event *bp, void *data)
4789 struct perf_sample_data sample;
4790 struct pt_regs *regs = data;
4792 perf_sample_data_init(&sample, bp->attr.bp_addr);
4794 if (!bp->hw.state && !perf_exclude_event(bp, regs))
4795 perf_swevent_event(bp, 1, 1, &sample, regs);
4800 * hrtimer based swevent callback
4803 static enum hrtimer_restart perf_swevent_hrtimer(struct hrtimer *hrtimer)
4805 enum hrtimer_restart ret = HRTIMER_RESTART;
4806 struct perf_sample_data data;
4807 struct pt_regs *regs;
4808 struct perf_event *event;
4811 event = container_of(hrtimer, struct perf_event, hw.hrtimer);
4812 event->pmu->read(event);
4814 perf_sample_data_init(&data, 0);
4815 data.period = event->hw.last_period;
4816 regs = get_irq_regs();
4818 if (regs && !perf_exclude_event(event, regs)) {
4819 if (!(event->attr.exclude_idle && current->pid == 0))
4820 if (perf_event_overflow(event, 0, &data, regs))
4821 ret = HRTIMER_NORESTART;
4824 period = max_t(u64, 10000, event->hw.sample_period);
4825 hrtimer_forward_now(hrtimer, ns_to_ktime(period));
4830 static void perf_swevent_start_hrtimer(struct perf_event *event)
4832 struct hw_perf_event *hwc = &event->hw;
4834 hrtimer_init(&hwc->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
4835 hwc->hrtimer.function = perf_swevent_hrtimer;
4836 if (hwc->sample_period) {
4837 s64 period = local64_read(&hwc->period_left);
4843 local64_set(&hwc->period_left, 0);
4845 period = max_t(u64, 10000, hwc->sample_period);
4847 __hrtimer_start_range_ns(&hwc->hrtimer,
4848 ns_to_ktime(period), 0,
4849 HRTIMER_MODE_REL_PINNED, 0);
4853 static void perf_swevent_cancel_hrtimer(struct perf_event *event)
4855 struct hw_perf_event *hwc = &event->hw;
4857 if (hwc->sample_period) {
4858 ktime_t remaining = hrtimer_get_remaining(&hwc->hrtimer);
4859 local64_set(&hwc->period_left, ktime_to_ns(remaining));
4861 hrtimer_cancel(&hwc->hrtimer);
4866 * Software event: cpu wall time clock
4869 static void cpu_clock_event_update(struct perf_event *event)
4874 now = local_clock();
4875 prev = local64_xchg(&event->hw.prev_count, now);
4876 local64_add(now - prev, &event->count);
4879 static void cpu_clock_event_start(struct perf_event *event, int flags)
4881 local64_set(&event->hw.prev_count, local_clock());
4882 perf_swevent_start_hrtimer(event);
4885 static void cpu_clock_event_stop(struct perf_event *event, int flags)
4887 perf_swevent_cancel_hrtimer(event);
4888 cpu_clock_event_update(event);
4891 static int cpu_clock_event_add(struct perf_event *event, int flags)
4893 if (flags & PERF_EF_START)
4894 cpu_clock_event_start(event, flags);
4899 static void cpu_clock_event_del(struct perf_event *event, int flags)
4901 cpu_clock_event_stop(event, flags);
4904 static void cpu_clock_event_read(struct perf_event *event)
4906 cpu_clock_event_update(event);
4909 static int cpu_clock_event_init(struct perf_event *event)
4911 if (event->attr.type != PERF_TYPE_SOFTWARE)
4914 if (event->attr.config != PERF_COUNT_SW_CPU_CLOCK)
4920 static struct pmu perf_cpu_clock = {
4921 .event_init = cpu_clock_event_init,
4922 .add = cpu_clock_event_add,
4923 .del = cpu_clock_event_del,
4924 .start = cpu_clock_event_start,
4925 .stop = cpu_clock_event_stop,
4926 .read = cpu_clock_event_read,
4930 * Software event: task time clock
4933 static void task_clock_event_update(struct perf_event *event, u64 now)
4938 prev = local64_xchg(&event->hw.prev_count, now);
4940 local64_add(delta, &event->count);
4943 static void task_clock_event_start(struct perf_event *event, int flags)
4945 local64_set(&event->hw.prev_count, event->ctx->time);
4946 perf_swevent_start_hrtimer(event);
4949 static void task_clock_event_stop(struct perf_event *event, int flags)
4951 perf_swevent_cancel_hrtimer(event);
4952 task_clock_event_update(event, event->ctx->time);
4955 static int task_clock_event_add(struct perf_event *event, int flags)
4957 if (flags & PERF_EF_START)
4958 task_clock_event_start(event, flags);
4963 static void task_clock_event_del(struct perf_event *event, int flags)
4965 task_clock_event_stop(event, PERF_EF_UPDATE);
4968 static void task_clock_event_read(struct perf_event *event)
4973 update_context_time(event->ctx);
4974 time = event->ctx->time;
4976 u64 now = perf_clock();
4977 u64 delta = now - event->ctx->timestamp;
4978 time = event->ctx->time + delta;
4981 task_clock_event_update(event, time);
4984 static int task_clock_event_init(struct perf_event *event)
4986 if (event->attr.type != PERF_TYPE_SOFTWARE)
4989 if (event->attr.config != PERF_COUNT_SW_TASK_CLOCK)
4995 static struct pmu perf_task_clock = {
4996 .event_init = task_clock_event_init,
4997 .add = task_clock_event_add,
4998 .del = task_clock_event_del,
4999 .start = task_clock_event_start,
5000 .stop = task_clock_event_stop,
5001 .read = task_clock_event_read,
5004 static void perf_pmu_nop_void(struct pmu *pmu)
5008 static int perf_pmu_nop_int(struct pmu *pmu)
5013 static void perf_pmu_start_txn(struct pmu *pmu)
5015 perf_pmu_disable(pmu);
5018 static int perf_pmu_commit_txn(struct pmu *pmu)
5020 perf_pmu_enable(pmu);
5024 static void perf_pmu_cancel_txn(struct pmu *pmu)
5026 perf_pmu_enable(pmu);
5029 int perf_pmu_register(struct pmu *pmu)
5033 mutex_lock(&pmus_lock);
5035 pmu->pmu_disable_count = alloc_percpu(int);
5036 if (!pmu->pmu_disable_count)
5039 pmu->pmu_cpu_context = alloc_percpu(struct perf_cpu_context);
5040 if (!pmu->pmu_cpu_context)
5043 for_each_possible_cpu(cpu) {
5044 struct perf_cpu_context *cpuctx;
5046 cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu);
5047 __perf_event_init_context(&cpuctx->ctx, NULL);
5048 cpuctx->ctx.pmu = pmu;
5049 cpuctx->timer_interval = TICK_NSEC;
5050 hrtimer_init(&cpuctx->timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
5051 cpuctx->timer.function = perf_event_context_tick;
5054 if (!pmu->start_txn) {
5055 if (pmu->pmu_enable) {
5057 * If we have pmu_enable/pmu_disable calls, install
5058 * transaction stubs that use that to try and batch
5059 * hardware accesses.
5061 pmu->start_txn = perf_pmu_start_txn;
5062 pmu->commit_txn = perf_pmu_commit_txn;
5063 pmu->cancel_txn = perf_pmu_cancel_txn;
5065 pmu->start_txn = perf_pmu_nop_void;
5066 pmu->commit_txn = perf_pmu_nop_int;
5067 pmu->cancel_txn = perf_pmu_nop_void;
5071 if (!pmu->pmu_enable) {
5072 pmu->pmu_enable = perf_pmu_nop_void;
5073 pmu->pmu_disable = perf_pmu_nop_void;
5076 list_add_rcu(&pmu->entry, &pmus);
5079 mutex_unlock(&pmus_lock);
5084 free_percpu(pmu->pmu_disable_count);
5088 void perf_pmu_unregister(struct pmu *pmu)
5090 mutex_lock(&pmus_lock);
5091 list_del_rcu(&pmu->entry);
5092 mutex_unlock(&pmus_lock);
5095 * We use the pmu list either under SRCU or preempt_disable,
5096 * synchronize_srcu() implies synchronize_sched() so we're good.
5098 synchronize_srcu(&pmus_srcu);
5100 free_percpu(pmu->pmu_disable_count);
5101 free_percpu(pmu->pmu_cpu_context);
5104 struct pmu *perf_init_event(struct perf_event *event)
5106 struct pmu *pmu = NULL;
5109 idx = srcu_read_lock(&pmus_srcu);
5110 list_for_each_entry_rcu(pmu, &pmus, entry) {
5111 int ret = pmu->event_init(event);
5114 if (ret != -ENOENT) {
5119 srcu_read_unlock(&pmus_srcu, idx);
5125 * Allocate and initialize a event structure
5127 static struct perf_event *
5128 perf_event_alloc(struct perf_event_attr *attr, int cpu,
5129 struct perf_event *group_leader,
5130 struct perf_event *parent_event,
5131 perf_overflow_handler_t overflow_handler)
5134 struct perf_event *event;
5135 struct hw_perf_event *hwc;
5138 event = kzalloc(sizeof(*event), GFP_KERNEL);
5140 return ERR_PTR(-ENOMEM);
5143 * Single events are their own group leaders, with an
5144 * empty sibling list:
5147 group_leader = event;
5149 mutex_init(&event->child_mutex);
5150 INIT_LIST_HEAD(&event->child_list);
5152 INIT_LIST_HEAD(&event->group_entry);
5153 INIT_LIST_HEAD(&event->event_entry);
5154 INIT_LIST_HEAD(&event->sibling_list);
5155 init_waitqueue_head(&event->waitq);
5157 mutex_init(&event->mmap_mutex);
5160 event->attr = *attr;
5161 event->group_leader = group_leader;
5165 event->parent = parent_event;
5167 event->ns = get_pid_ns(current->nsproxy->pid_ns);
5168 event->id = atomic64_inc_return(&perf_event_id);
5170 event->state = PERF_EVENT_STATE_INACTIVE;
5172 if (!overflow_handler && parent_event)
5173 overflow_handler = parent_event->overflow_handler;
5175 event->overflow_handler = overflow_handler;
5178 event->state = PERF_EVENT_STATE_OFF;
5183 hwc->sample_period = attr->sample_period;
5184 if (attr->freq && attr->sample_freq)
5185 hwc->sample_period = 1;
5186 hwc->last_period = hwc->sample_period;
5188 local64_set(&hwc->period_left, hwc->sample_period);
5191 * we currently do not support PERF_FORMAT_GROUP on inherited events
5193 if (attr->inherit && (attr->read_format & PERF_FORMAT_GROUP))
5196 pmu = perf_init_event(event);
5202 else if (IS_ERR(pmu))
5207 put_pid_ns(event->ns);
5209 return ERR_PTR(err);
5214 if (!event->parent) {
5215 atomic_inc(&nr_events);
5216 if (event->attr.mmap || event->attr.mmap_data)
5217 atomic_inc(&nr_mmap_events);
5218 if (event->attr.comm)
5219 atomic_inc(&nr_comm_events);
5220 if (event->attr.task)
5221 atomic_inc(&nr_task_events);
5222 if (event->attr.sample_type & PERF_SAMPLE_CALLCHAIN) {
5223 err = get_callchain_buffers();
5226 return ERR_PTR(err);
5234 static int perf_copy_attr(struct perf_event_attr __user *uattr,
5235 struct perf_event_attr *attr)
5240 if (!access_ok(VERIFY_WRITE, uattr, PERF_ATTR_SIZE_VER0))
5244 * zero the full structure, so that a short copy will be nice.
5246 memset(attr, 0, sizeof(*attr));
5248 ret = get_user(size, &uattr->size);
5252 if (size > PAGE_SIZE) /* silly large */
5255 if (!size) /* abi compat */
5256 size = PERF_ATTR_SIZE_VER0;
5258 if (size < PERF_ATTR_SIZE_VER0)
5262 * If we're handed a bigger struct than we know of,
5263 * ensure all the unknown bits are 0 - i.e. new
5264 * user-space does not rely on any kernel feature
5265 * extensions we dont know about yet.
5267 if (size > sizeof(*attr)) {
5268 unsigned char __user *addr;
5269 unsigned char __user *end;
5272 addr = (void __user *)uattr + sizeof(*attr);
5273 end = (void __user *)uattr + size;
5275 for (; addr < end; addr++) {
5276 ret = get_user(val, addr);
5282 size = sizeof(*attr);
5285 ret = copy_from_user(attr, uattr, size);
5290 * If the type exists, the corresponding creation will verify
5293 if (attr->type >= PERF_TYPE_MAX)
5296 if (attr->__reserved_1)
5299 if (attr->sample_type & ~(PERF_SAMPLE_MAX-1))
5302 if (attr->read_format & ~(PERF_FORMAT_MAX-1))
5309 put_user(sizeof(*attr), &uattr->size);
5315 perf_event_set_output(struct perf_event *event, struct perf_event *output_event)
5317 struct perf_buffer *buffer = NULL, *old_buffer = NULL;
5323 /* don't allow circular references */
5324 if (event == output_event)
5328 * Don't allow cross-cpu buffers
5330 if (output_event->cpu != event->cpu)
5334 * If its not a per-cpu buffer, it must be the same task.
5336 if (output_event->cpu == -1 && output_event->ctx != event->ctx)
5340 mutex_lock(&event->mmap_mutex);
5341 /* Can't redirect output if we've got an active mmap() */
5342 if (atomic_read(&event->mmap_count))
5346 /* get the buffer we want to redirect to */
5347 buffer = perf_buffer_get(output_event);
5352 old_buffer = event->buffer;
5353 rcu_assign_pointer(event->buffer, buffer);
5356 mutex_unlock(&event->mmap_mutex);
5359 perf_buffer_put(old_buffer);
5365 * sys_perf_event_open - open a performance event, associate it to a task/cpu
5367 * @attr_uptr: event_id type attributes for monitoring/sampling
5370 * @group_fd: group leader event fd
5372 SYSCALL_DEFINE5(perf_event_open,
5373 struct perf_event_attr __user *, attr_uptr,
5374 pid_t, pid, int, cpu, int, group_fd, unsigned long, flags)
5376 struct perf_event *event, *group_leader = NULL, *output_event = NULL;
5377 struct perf_event_attr attr;
5378 struct perf_event_context *ctx;
5379 struct file *event_file = NULL;
5380 struct file *group_file = NULL;
5382 int fput_needed = 0;
5385 /* for future expandability... */
5386 if (flags & ~(PERF_FLAG_FD_NO_GROUP | PERF_FLAG_FD_OUTPUT))
5389 err = perf_copy_attr(attr_uptr, &attr);
5393 if (!attr.exclude_kernel) {
5394 if (perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN))
5399 if (attr.sample_freq > sysctl_perf_event_sample_rate)
5403 event_fd = get_unused_fd_flags(O_RDWR);
5407 event = perf_event_alloc(&attr, cpu, group_leader, NULL, NULL);
5408 if (IS_ERR(event)) {
5409 err = PTR_ERR(event);
5414 * Get the target context (task or percpu):
5416 ctx = find_get_context(event->pmu, pid, cpu);
5422 if (group_fd != -1) {
5423 group_leader = perf_fget_light(group_fd, &fput_needed);
5424 if (IS_ERR(group_leader)) {
5425 err = PTR_ERR(group_leader);
5428 group_file = group_leader->filp;
5429 if (flags & PERF_FLAG_FD_OUTPUT)
5430 output_event = group_leader;
5431 if (flags & PERF_FLAG_FD_NO_GROUP)
5432 group_leader = NULL;
5436 * Look up the group leader (we will attach this event to it):
5442 * Do not allow a recursive hierarchy (this new sibling
5443 * becoming part of another group-sibling):
5445 if (group_leader->group_leader != group_leader)
5448 * Do not allow to attach to a group in a different
5449 * task or CPU context:
5451 if (group_leader->ctx != ctx)
5454 * Only a group leader can be exclusive or pinned
5456 if (attr.exclusive || attr.pinned)
5461 err = perf_event_set_output(event, output_event);
5466 event_file = anon_inode_getfile("[perf_event]", &perf_fops, event, O_RDWR);
5467 if (IS_ERR(event_file)) {
5468 err = PTR_ERR(event_file);
5472 event->filp = event_file;
5473 WARN_ON_ONCE(ctx->parent_ctx);
5474 mutex_lock(&ctx->mutex);
5475 perf_install_in_context(ctx, event, cpu);
5477 mutex_unlock(&ctx->mutex);
5479 event->owner = current;
5480 get_task_struct(current);
5481 mutex_lock(¤t->perf_event_mutex);
5482 list_add_tail(&event->owner_entry, ¤t->perf_event_list);
5483 mutex_unlock(¤t->perf_event_mutex);
5486 * Drop the reference on the group_event after placing the
5487 * new event on the sibling_list. This ensures destruction
5488 * of the group leader will find the pointer to itself in
5489 * perf_group_detach().
5491 fput_light(group_file, fput_needed);
5492 fd_install(event_fd, event_file);
5496 fput_light(group_file, fput_needed);
5501 put_unused_fd(event_fd);
5506 * perf_event_create_kernel_counter
5508 * @attr: attributes of the counter to create
5509 * @cpu: cpu in which the counter is bound
5510 * @pid: task to profile
5513 perf_event_create_kernel_counter(struct perf_event_attr *attr, int cpu,
5515 perf_overflow_handler_t overflow_handler)
5517 struct perf_event_context *ctx;
5518 struct perf_event *event;
5522 * Get the target context (task or percpu):
5525 event = perf_event_alloc(attr, cpu, NULL, NULL, overflow_handler);
5526 if (IS_ERR(event)) {
5527 err = PTR_ERR(event);
5531 ctx = find_get_context(event->pmu, pid, cpu);
5538 WARN_ON_ONCE(ctx->parent_ctx);
5539 mutex_lock(&ctx->mutex);
5540 perf_install_in_context(ctx, event, cpu);
5542 mutex_unlock(&ctx->mutex);
5544 event->owner = current;
5545 get_task_struct(current);
5546 mutex_lock(¤t->perf_event_mutex);
5547 list_add_tail(&event->owner_entry, ¤t->perf_event_list);
5548 mutex_unlock(¤t->perf_event_mutex);
5555 return ERR_PTR(err);
5557 EXPORT_SYMBOL_GPL(perf_event_create_kernel_counter);
5559 static void sync_child_event(struct perf_event *child_event,
5560 struct task_struct *child)
5562 struct perf_event *parent_event = child_event->parent;
5565 if (child_event->attr.inherit_stat)
5566 perf_event_read_event(child_event, child);
5568 child_val = perf_event_count(child_event);
5571 * Add back the child's count to the parent's count:
5573 atomic64_add(child_val, &parent_event->child_count);
5574 atomic64_add(child_event->total_time_enabled,
5575 &parent_event->child_total_time_enabled);
5576 atomic64_add(child_event->total_time_running,
5577 &parent_event->child_total_time_running);
5580 * Remove this event from the parent's list
5582 WARN_ON_ONCE(parent_event->ctx->parent_ctx);
5583 mutex_lock(&parent_event->child_mutex);
5584 list_del_init(&child_event->child_list);
5585 mutex_unlock(&parent_event->child_mutex);
5588 * Release the parent event, if this was the last
5591 fput(parent_event->filp);
5595 __perf_event_exit_task(struct perf_event *child_event,
5596 struct perf_event_context *child_ctx,
5597 struct task_struct *child)
5599 struct perf_event *parent_event;
5601 perf_event_remove_from_context(child_event);
5603 parent_event = child_event->parent;
5605 * It can happen that parent exits first, and has events
5606 * that are still around due to the child reference. These
5607 * events need to be zapped - but otherwise linger.
5610 sync_child_event(child_event, child);
5611 free_event(child_event);
5616 * When a child task exits, feed back event values to parent events.
5618 void perf_event_exit_task(struct task_struct *child)
5620 struct perf_event *child_event, *tmp;
5621 struct perf_event_context *child_ctx;
5622 unsigned long flags;
5624 if (likely(!child->perf_event_ctxp)) {
5625 perf_event_task(child, NULL, 0);
5629 local_irq_save(flags);
5631 * We can't reschedule here because interrupts are disabled,
5632 * and either child is current or it is a task that can't be
5633 * scheduled, so we are now safe from rescheduling changing
5636 child_ctx = child->perf_event_ctxp;
5637 __perf_event_task_sched_out(child_ctx);
5640 * Take the context lock here so that if find_get_context is
5641 * reading child->perf_event_ctxp, we wait until it has
5642 * incremented the context's refcount before we do put_ctx below.
5644 raw_spin_lock(&child_ctx->lock);
5645 child->perf_event_ctxp = NULL;
5647 * If this context is a clone; unclone it so it can't get
5648 * swapped to another process while we're removing all
5649 * the events from it.
5651 unclone_ctx(child_ctx);
5652 update_context_time(child_ctx);
5653 raw_spin_unlock_irqrestore(&child_ctx->lock, flags);
5656 * Report the task dead after unscheduling the events so that we
5657 * won't get any samples after PERF_RECORD_EXIT. We can however still
5658 * get a few PERF_RECORD_READ events.
5660 perf_event_task(child, child_ctx, 0);
5663 * We can recurse on the same lock type through:
5665 * __perf_event_exit_task()
5666 * sync_child_event()
5667 * fput(parent_event->filp)
5669 * mutex_lock(&ctx->mutex)
5671 * But since its the parent context it won't be the same instance.
5673 mutex_lock(&child_ctx->mutex);
5676 list_for_each_entry_safe(child_event, tmp, &child_ctx->pinned_groups,
5678 __perf_event_exit_task(child_event, child_ctx, child);
5680 list_for_each_entry_safe(child_event, tmp, &child_ctx->flexible_groups,
5682 __perf_event_exit_task(child_event, child_ctx, child);
5685 * If the last event was a group event, it will have appended all
5686 * its siblings to the list, but we obtained 'tmp' before that which
5687 * will still point to the list head terminating the iteration.
5689 if (!list_empty(&child_ctx->pinned_groups) ||
5690 !list_empty(&child_ctx->flexible_groups))
5693 mutex_unlock(&child_ctx->mutex);
5698 static void perf_free_event(struct perf_event *event,
5699 struct perf_event_context *ctx)
5701 struct perf_event *parent = event->parent;
5703 if (WARN_ON_ONCE(!parent))
5706 mutex_lock(&parent->child_mutex);
5707 list_del_init(&event->child_list);
5708 mutex_unlock(&parent->child_mutex);
5712 perf_group_detach(event);
5713 list_del_event(event, ctx);
5718 * free an unexposed, unused context as created by inheritance by
5719 * init_task below, used by fork() in case of fail.
5721 void perf_event_free_task(struct task_struct *task)
5723 struct perf_event_context *ctx = task->perf_event_ctxp;
5724 struct perf_event *event, *tmp;
5729 mutex_lock(&ctx->mutex);
5731 list_for_each_entry_safe(event, tmp, &ctx->pinned_groups, group_entry)
5732 perf_free_event(event, ctx);
5734 list_for_each_entry_safe(event, tmp, &ctx->flexible_groups,
5736 perf_free_event(event, ctx);
5738 if (!list_empty(&ctx->pinned_groups) ||
5739 !list_empty(&ctx->flexible_groups))
5742 mutex_unlock(&ctx->mutex);
5748 * inherit a event from parent task to child task:
5750 static struct perf_event *
5751 inherit_event(struct perf_event *parent_event,
5752 struct task_struct *parent,
5753 struct perf_event_context *parent_ctx,
5754 struct task_struct *child,
5755 struct perf_event *group_leader,
5756 struct perf_event_context *child_ctx)
5758 struct perf_event *child_event;
5761 * Instead of creating recursive hierarchies of events,
5762 * we link inherited events back to the original parent,
5763 * which has a filp for sure, which we use as the reference
5766 if (parent_event->parent)
5767 parent_event = parent_event->parent;
5769 child_event = perf_event_alloc(&parent_event->attr,
5771 group_leader, parent_event,
5773 if (IS_ERR(child_event))
5778 * Make the child state follow the state of the parent event,
5779 * not its attr.disabled bit. We hold the parent's mutex,
5780 * so we won't race with perf_event_{en, dis}able_family.
5782 if (parent_event->state >= PERF_EVENT_STATE_INACTIVE)
5783 child_event->state = PERF_EVENT_STATE_INACTIVE;
5785 child_event->state = PERF_EVENT_STATE_OFF;
5787 if (parent_event->attr.freq) {
5788 u64 sample_period = parent_event->hw.sample_period;
5789 struct hw_perf_event *hwc = &child_event->hw;
5791 hwc->sample_period = sample_period;
5792 hwc->last_period = sample_period;
5794 local64_set(&hwc->period_left, sample_period);
5797 child_event->ctx = child_ctx;
5798 child_event->overflow_handler = parent_event->overflow_handler;
5801 * Link it up in the child's context:
5803 add_event_to_ctx(child_event, child_ctx);
5806 * Get a reference to the parent filp - we will fput it
5807 * when the child event exits. This is safe to do because
5808 * we are in the parent and we know that the filp still
5809 * exists and has a nonzero count:
5811 atomic_long_inc(&parent_event->filp->f_count);
5814 * Link this into the parent event's child list
5816 WARN_ON_ONCE(parent_event->ctx->parent_ctx);
5817 mutex_lock(&parent_event->child_mutex);
5818 list_add_tail(&child_event->child_list, &parent_event->child_list);
5819 mutex_unlock(&parent_event->child_mutex);
5824 static int inherit_group(struct perf_event *parent_event,
5825 struct task_struct *parent,
5826 struct perf_event_context *parent_ctx,
5827 struct task_struct *child,
5828 struct perf_event_context *child_ctx)
5830 struct perf_event *leader;
5831 struct perf_event *sub;
5832 struct perf_event *child_ctr;
5834 leader = inherit_event(parent_event, parent, parent_ctx,
5835 child, NULL, child_ctx);
5837 return PTR_ERR(leader);
5838 list_for_each_entry(sub, &parent_event->sibling_list, group_entry) {
5839 child_ctr = inherit_event(sub, parent, parent_ctx,
5840 child, leader, child_ctx);
5841 if (IS_ERR(child_ctr))
5842 return PTR_ERR(child_ctr);
5848 inherit_task_group(struct perf_event *event, struct task_struct *parent,
5849 struct perf_event_context *parent_ctx,
5850 struct task_struct *child,
5854 struct perf_event_context *child_ctx = child->perf_event_ctxp;
5856 if (!event->attr.inherit) {
5863 * This is executed from the parent task context, so
5864 * inherit events that have been marked for cloning.
5865 * First allocate and initialize a context for the
5869 child_ctx = kzalloc(sizeof(struct perf_event_context),
5874 __perf_event_init_context(child_ctx, child);
5875 child_ctx->pmu = event->pmu;
5876 child->perf_event_ctxp = child_ctx;
5877 get_task_struct(child);
5880 ret = inherit_group(event, parent, parent_ctx,
5891 * Initialize the perf_event context in task_struct
5893 int perf_event_init_task(struct task_struct *child)
5895 struct perf_event_context *child_ctx, *parent_ctx;
5896 struct perf_event_context *cloned_ctx;
5897 struct perf_event *event;
5898 struct task_struct *parent = current;
5899 int inherited_all = 1;
5902 child->perf_event_ctxp = NULL;
5904 mutex_init(&child->perf_event_mutex);
5905 INIT_LIST_HEAD(&child->perf_event_list);
5907 if (likely(!parent->perf_event_ctxp))
5911 * If the parent's context is a clone, pin it so it won't get
5914 parent_ctx = perf_pin_task_context(parent);
5917 * No need to check if parent_ctx != NULL here; since we saw
5918 * it non-NULL earlier, the only reason for it to become NULL
5919 * is if we exit, and since we're currently in the middle of
5920 * a fork we can't be exiting at the same time.
5924 * Lock the parent list. No need to lock the child - not PID
5925 * hashed yet and not running, so nobody can access it.
5927 mutex_lock(&parent_ctx->mutex);
5930 * We dont have to disable NMIs - we are only looking at
5931 * the list, not manipulating it:
5933 list_for_each_entry(event, &parent_ctx->pinned_groups, group_entry) {
5934 ret = inherit_task_group(event, parent, parent_ctx, child,
5940 list_for_each_entry(event, &parent_ctx->flexible_groups, group_entry) {
5941 ret = inherit_task_group(event, parent, parent_ctx, child,
5947 child_ctx = child->perf_event_ctxp;
5949 if (child_ctx && inherited_all) {
5951 * Mark the child context as a clone of the parent
5952 * context, or of whatever the parent is a clone of.
5953 * Note that if the parent is a clone, it could get
5954 * uncloned at any point, but that doesn't matter
5955 * because the list of events and the generation
5956 * count can't have changed since we took the mutex.
5958 cloned_ctx = rcu_dereference(parent_ctx->parent_ctx);
5960 child_ctx->parent_ctx = cloned_ctx;
5961 child_ctx->parent_gen = parent_ctx->parent_gen;
5963 child_ctx->parent_ctx = parent_ctx;
5964 child_ctx->parent_gen = parent_ctx->generation;
5966 get_ctx(child_ctx->parent_ctx);
5969 mutex_unlock(&parent_ctx->mutex);
5971 perf_unpin_context(parent_ctx);
5976 static void __init perf_event_init_all_cpus(void)
5978 struct swevent_htable *swhash;
5981 for_each_possible_cpu(cpu) {
5982 swhash = &per_cpu(swevent_htable, cpu);
5983 mutex_init(&swhash->hlist_mutex);
5987 static void __cpuinit perf_event_init_cpu(int cpu)
5989 struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
5991 mutex_lock(&swhash->hlist_mutex);
5992 if (swhash->hlist_refcount > 0) {
5993 struct swevent_hlist *hlist;
5995 hlist = kzalloc_node(sizeof(*hlist), GFP_KERNEL, cpu_to_node(cpu));
5997 rcu_assign_pointer(swhash->swevent_hlist, hlist);
5999 mutex_unlock(&swhash->hlist_mutex);
6002 #ifdef CONFIG_HOTPLUG_CPU
6003 static void __perf_event_exit_context(void *__info)
6005 struct perf_event_context *ctx = __info;
6006 struct perf_event *event, *tmp;
6008 perf_pmu_rotate_stop(ctx->pmu);
6010 list_for_each_entry_safe(event, tmp, &ctx->pinned_groups, group_entry)
6011 __perf_event_remove_from_context(event);
6012 list_for_each_entry_safe(event, tmp, &ctx->flexible_groups, group_entry)
6013 __perf_event_remove_from_context(event);
6016 static void perf_event_exit_cpu_context(int cpu)
6018 struct perf_event_context *ctx;
6022 idx = srcu_read_lock(&pmus_srcu);
6023 list_for_each_entry_rcu(pmu, &pmus, entry) {
6024 ctx = &this_cpu_ptr(pmu->pmu_cpu_context)->ctx;
6026 mutex_lock(&ctx->mutex);
6027 smp_call_function_single(cpu, __perf_event_exit_context, ctx, 1);
6028 mutex_unlock(&ctx->mutex);
6030 srcu_read_unlock(&pmus_srcu, idx);
6034 static void perf_event_exit_cpu(int cpu)
6036 struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
6038 mutex_lock(&swhash->hlist_mutex);
6039 swevent_hlist_release(swhash);
6040 mutex_unlock(&swhash->hlist_mutex);
6042 perf_event_exit_cpu_context(cpu);
6045 static inline void perf_event_exit_cpu(int cpu) { }
6048 static int __cpuinit
6049 perf_cpu_notify(struct notifier_block *self, unsigned long action, void *hcpu)
6051 unsigned int cpu = (long)hcpu;
6053 switch (action & ~CPU_TASKS_FROZEN) {
6055 case CPU_UP_PREPARE:
6056 case CPU_DOWN_FAILED:
6057 perf_event_init_cpu(cpu);
6060 case CPU_UP_CANCELED:
6061 case CPU_DOWN_PREPARE:
6062 perf_event_exit_cpu(cpu);
6072 void __init perf_event_init(void)
6074 perf_event_init_all_cpus();
6075 init_srcu_struct(&pmus_srcu);
6076 perf_pmu_register(&perf_swevent);
6077 perf_pmu_register(&perf_cpu_clock);
6078 perf_pmu_register(&perf_task_clock);
6080 perf_cpu_notifier(perf_cpu_notify);