2 * Performance events core code:
4 * Copyright (C) 2008 Thomas Gleixner <tglx@linutronix.de>
5 * Copyright (C) 2008-2009 Red Hat, Inc., Ingo Molnar
6 * Copyright (C) 2008-2009 Red Hat, Inc., Peter Zijlstra <pzijlstr@redhat.com>
7 * Copyright © 2009 Paul Mackerras, IBM Corp. <paulus@au1.ibm.com>
9 * For licensing details see kernel-base/COPYING
14 #include <linux/cpu.h>
15 #include <linux/smp.h>
16 #include <linux/file.h>
17 #include <linux/poll.h>
18 #include <linux/slab.h>
19 #include <linux/hash.h>
20 #include <linux/sysfs.h>
21 #include <linux/dcache.h>
22 #include <linux/percpu.h>
23 #include <linux/ptrace.h>
24 #include <linux/vmstat.h>
25 #include <linux/vmalloc.h>
26 #include <linux/hardirq.h>
27 #include <linux/rculist.h>
28 #include <linux/uaccess.h>
29 #include <linux/syscalls.h>
30 #include <linux/anon_inodes.h>
31 #include <linux/kernel_stat.h>
32 #include <linux/perf_event.h>
33 #include <linux/ftrace_event.h>
35 #include <asm/irq_regs.h>
37 static atomic_t nr_events __read_mostly;
38 static atomic_t nr_mmap_events __read_mostly;
39 static atomic_t nr_comm_events __read_mostly;
40 static atomic_t nr_task_events __read_mostly;
42 static LIST_HEAD(pmus);
43 static DEFINE_MUTEX(pmus_lock);
44 static struct srcu_struct pmus_srcu;
47 * perf event paranoia level:
48 * -1 - not paranoid at all
49 * 0 - disallow raw tracepoint access for unpriv
50 * 1 - disallow cpu events for unpriv
51 * 2 - disallow kernel profiling for unpriv
53 int sysctl_perf_event_paranoid __read_mostly = 1;
55 int sysctl_perf_event_mlock __read_mostly = 512; /* 'free' kb per user */
58 * max perf event sample rate
60 int sysctl_perf_event_sample_rate __read_mostly = 100000;
62 static atomic64_t perf_event_id;
64 void __weak perf_event_print_debug(void) { }
66 extern __weak const char *perf_pmu_name(void)
71 void perf_pmu_disable(struct pmu *pmu)
73 int *count = this_cpu_ptr(pmu->pmu_disable_count);
75 pmu->pmu_disable(pmu);
78 void perf_pmu_enable(struct pmu *pmu)
80 int *count = this_cpu_ptr(pmu->pmu_disable_count);
85 static DEFINE_PER_CPU(struct list_head, rotation_list);
88 * perf_pmu_rotate_start() and perf_rotate_context() are fully serialized
89 * because they're strictly cpu affine and rotate_start is called with IRQs
90 * disabled, while rotate_context is called from IRQ context.
92 static void perf_pmu_rotate_start(struct pmu *pmu)
94 struct perf_cpu_context *cpuctx = this_cpu_ptr(pmu->pmu_cpu_context);
95 struct list_head *head = &__get_cpu_var(rotation_list);
97 WARN_ON(!irqs_disabled());
99 if (list_empty(&cpuctx->rotation_list))
100 list_add(&cpuctx->rotation_list, head);
103 static void get_ctx(struct perf_event_context *ctx)
105 WARN_ON(!atomic_inc_not_zero(&ctx->refcount));
108 static void free_ctx(struct rcu_head *head)
110 struct perf_event_context *ctx;
112 ctx = container_of(head, struct perf_event_context, rcu_head);
116 static void put_ctx(struct perf_event_context *ctx)
118 if (atomic_dec_and_test(&ctx->refcount)) {
120 put_ctx(ctx->parent_ctx);
122 put_task_struct(ctx->task);
123 call_rcu(&ctx->rcu_head, free_ctx);
127 static void unclone_ctx(struct perf_event_context *ctx)
129 if (ctx->parent_ctx) {
130 put_ctx(ctx->parent_ctx);
131 ctx->parent_ctx = NULL;
136 * If we inherit events we want to return the parent event id
139 static u64 primary_event_id(struct perf_event *event)
144 id = event->parent->id;
150 * Get the perf_event_context for a task and lock it.
151 * This has to cope with with the fact that until it is locked,
152 * the context could get moved to another task.
154 static struct perf_event_context *
155 perf_lock_task_context(struct task_struct *task, int ctxn, unsigned long *flags)
157 struct perf_event_context *ctx;
161 ctx = rcu_dereference(task->perf_event_ctxp[ctxn]);
164 * If this context is a clone of another, it might
165 * get swapped for another underneath us by
166 * perf_event_task_sched_out, though the
167 * rcu_read_lock() protects us from any context
168 * getting freed. Lock the context and check if it
169 * got swapped before we could get the lock, and retry
170 * if so. If we locked the right context, then it
171 * can't get swapped on us any more.
173 raw_spin_lock_irqsave(&ctx->lock, *flags);
174 if (ctx != rcu_dereference(task->perf_event_ctxp[ctxn])) {
175 raw_spin_unlock_irqrestore(&ctx->lock, *flags);
179 if (!atomic_inc_not_zero(&ctx->refcount)) {
180 raw_spin_unlock_irqrestore(&ctx->lock, *flags);
189 * Get the context for a task and increment its pin_count so it
190 * can't get swapped to another task. This also increments its
191 * reference count so that the context can't get freed.
193 static struct perf_event_context *
194 perf_pin_task_context(struct task_struct *task, int ctxn)
196 struct perf_event_context *ctx;
199 ctx = perf_lock_task_context(task, ctxn, &flags);
202 raw_spin_unlock_irqrestore(&ctx->lock, flags);
207 static void perf_unpin_context(struct perf_event_context *ctx)
211 raw_spin_lock_irqsave(&ctx->lock, flags);
213 raw_spin_unlock_irqrestore(&ctx->lock, flags);
217 static inline u64 perf_clock(void)
219 return local_clock();
223 * Update the record of the current time in a context.
225 static void update_context_time(struct perf_event_context *ctx)
227 u64 now = perf_clock();
229 ctx->time += now - ctx->timestamp;
230 ctx->timestamp = now;
234 * Update the total_time_enabled and total_time_running fields for a event.
236 static void update_event_times(struct perf_event *event)
238 struct perf_event_context *ctx = event->ctx;
241 if (event->state < PERF_EVENT_STATE_INACTIVE ||
242 event->group_leader->state < PERF_EVENT_STATE_INACTIVE)
248 run_end = event->tstamp_stopped;
250 event->total_time_enabled = run_end - event->tstamp_enabled;
252 if (event->state == PERF_EVENT_STATE_INACTIVE)
253 run_end = event->tstamp_stopped;
257 event->total_time_running = run_end - event->tstamp_running;
261 * Update total_time_enabled and total_time_running for all events in a group.
263 static void update_group_times(struct perf_event *leader)
265 struct perf_event *event;
267 update_event_times(leader);
268 list_for_each_entry(event, &leader->sibling_list, group_entry)
269 update_event_times(event);
272 static struct list_head *
273 ctx_group_list(struct perf_event *event, struct perf_event_context *ctx)
275 if (event->attr.pinned)
276 return &ctx->pinned_groups;
278 return &ctx->flexible_groups;
282 * Add a event from the lists for its context.
283 * Must be called with ctx->mutex and ctx->lock held.
286 list_add_event(struct perf_event *event, struct perf_event_context *ctx)
288 WARN_ON_ONCE(event->attach_state & PERF_ATTACH_CONTEXT);
289 event->attach_state |= PERF_ATTACH_CONTEXT;
292 * If we're a stand alone event or group leader, we go to the context
293 * list, group events are kept attached to the group so that
294 * perf_group_detach can, at all times, locate all siblings.
296 if (event->group_leader == event) {
297 struct list_head *list;
299 if (is_software_event(event))
300 event->group_flags |= PERF_GROUP_SOFTWARE;
302 list = ctx_group_list(event, ctx);
303 list_add_tail(&event->group_entry, list);
306 list_add_rcu(&event->event_entry, &ctx->event_list);
308 perf_pmu_rotate_start(ctx->pmu);
310 if (event->attr.inherit_stat)
314 static void perf_group_attach(struct perf_event *event)
316 struct perf_event *group_leader = event->group_leader;
318 WARN_ON_ONCE(event->attach_state & PERF_ATTACH_GROUP);
319 event->attach_state |= PERF_ATTACH_GROUP;
321 if (group_leader == event)
324 if (group_leader->group_flags & PERF_GROUP_SOFTWARE &&
325 !is_software_event(event))
326 group_leader->group_flags &= ~PERF_GROUP_SOFTWARE;
328 list_add_tail(&event->group_entry, &group_leader->sibling_list);
329 group_leader->nr_siblings++;
333 * Remove a event from the lists for its context.
334 * Must be called with ctx->mutex and ctx->lock held.
337 list_del_event(struct perf_event *event, struct perf_event_context *ctx)
340 * We can have double detach due to exit/hot-unplug + close.
342 if (!(event->attach_state & PERF_ATTACH_CONTEXT))
345 event->attach_state &= ~PERF_ATTACH_CONTEXT;
348 if (event->attr.inherit_stat)
351 list_del_rcu(&event->event_entry);
353 if (event->group_leader == event)
354 list_del_init(&event->group_entry);
356 update_group_times(event);
359 * If event was in error state, then keep it
360 * that way, otherwise bogus counts will be
361 * returned on read(). The only way to get out
362 * of error state is by explicit re-enabling
365 if (event->state > PERF_EVENT_STATE_OFF)
366 event->state = PERF_EVENT_STATE_OFF;
369 static void perf_group_detach(struct perf_event *event)
371 struct perf_event *sibling, *tmp;
372 struct list_head *list = NULL;
375 * We can have double detach due to exit/hot-unplug + close.
377 if (!(event->attach_state & PERF_ATTACH_GROUP))
380 event->attach_state &= ~PERF_ATTACH_GROUP;
383 * If this is a sibling, remove it from its group.
385 if (event->group_leader != event) {
386 list_del_init(&event->group_entry);
387 event->group_leader->nr_siblings--;
391 if (!list_empty(&event->group_entry))
392 list = &event->group_entry;
395 * If this was a group event with sibling events then
396 * upgrade the siblings to singleton events by adding them
397 * to whatever list we are on.
399 list_for_each_entry_safe(sibling, tmp, &event->sibling_list, group_entry) {
401 list_move_tail(&sibling->group_entry, list);
402 sibling->group_leader = sibling;
404 /* Inherit group flags from the previous leader */
405 sibling->group_flags = event->group_flags;
410 event_filter_match(struct perf_event *event)
412 return event->cpu == -1 || event->cpu == smp_processor_id();
416 event_sched_out(struct perf_event *event,
417 struct perf_cpu_context *cpuctx,
418 struct perf_event_context *ctx)
422 * An event which could not be activated because of
423 * filter mismatch still needs to have its timings
424 * maintained, otherwise bogus information is return
425 * via read() for time_enabled, time_running:
427 if (event->state == PERF_EVENT_STATE_INACTIVE
428 && !event_filter_match(event)) {
429 delta = ctx->time - event->tstamp_stopped;
430 event->tstamp_running += delta;
431 event->tstamp_stopped = ctx->time;
434 if (event->state != PERF_EVENT_STATE_ACTIVE)
437 event->state = PERF_EVENT_STATE_INACTIVE;
438 if (event->pending_disable) {
439 event->pending_disable = 0;
440 event->state = PERF_EVENT_STATE_OFF;
442 event->tstamp_stopped = ctx->time;
443 event->pmu->del(event, 0);
446 if (!is_software_event(event))
447 cpuctx->active_oncpu--;
449 if (event->attr.exclusive || !cpuctx->active_oncpu)
450 cpuctx->exclusive = 0;
454 group_sched_out(struct perf_event *group_event,
455 struct perf_cpu_context *cpuctx,
456 struct perf_event_context *ctx)
458 struct perf_event *event;
459 int state = group_event->state;
461 event_sched_out(group_event, cpuctx, ctx);
464 * Schedule out siblings (if any):
466 list_for_each_entry(event, &group_event->sibling_list, group_entry)
467 event_sched_out(event, cpuctx, ctx);
469 if (state == PERF_EVENT_STATE_ACTIVE && group_event->attr.exclusive)
470 cpuctx->exclusive = 0;
473 static inline struct perf_cpu_context *
474 __get_cpu_context(struct perf_event_context *ctx)
476 return this_cpu_ptr(ctx->pmu->pmu_cpu_context);
480 * Cross CPU call to remove a performance event
482 * We disable the event on the hardware level first. After that we
483 * remove it from the context list.
485 static void __perf_event_remove_from_context(void *info)
487 struct perf_event *event = info;
488 struct perf_event_context *ctx = event->ctx;
489 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
492 * If this is a task context, we need to check whether it is
493 * the current task context of this cpu. If not it has been
494 * scheduled out before the smp call arrived.
496 if (ctx->task && cpuctx->task_ctx != ctx)
499 raw_spin_lock(&ctx->lock);
501 event_sched_out(event, cpuctx, ctx);
503 list_del_event(event, ctx);
505 raw_spin_unlock(&ctx->lock);
510 * Remove the event from a task's (or a CPU's) list of events.
512 * Must be called with ctx->mutex held.
514 * CPU events are removed with a smp call. For task events we only
515 * call when the task is on a CPU.
517 * If event->ctx is a cloned context, callers must make sure that
518 * every task struct that event->ctx->task could possibly point to
519 * remains valid. This is OK when called from perf_release since
520 * that only calls us on the top-level context, which can't be a clone.
521 * When called from perf_event_exit_task, it's OK because the
522 * context has been detached from its task.
524 static void perf_event_remove_from_context(struct perf_event *event)
526 struct perf_event_context *ctx = event->ctx;
527 struct task_struct *task = ctx->task;
531 * Per cpu events are removed via an smp call and
532 * the removal is always successful.
534 smp_call_function_single(event->cpu,
535 __perf_event_remove_from_context,
541 task_oncpu_function_call(task, __perf_event_remove_from_context,
544 raw_spin_lock_irq(&ctx->lock);
546 * If the context is active we need to retry the smp call.
548 if (ctx->nr_active && !list_empty(&event->group_entry)) {
549 raw_spin_unlock_irq(&ctx->lock);
554 * The lock prevents that this context is scheduled in so we
555 * can remove the event safely, if the call above did not
558 if (!list_empty(&event->group_entry))
559 list_del_event(event, ctx);
560 raw_spin_unlock_irq(&ctx->lock);
564 * Cross CPU call to disable a performance event
566 static void __perf_event_disable(void *info)
568 struct perf_event *event = info;
569 struct perf_event_context *ctx = event->ctx;
570 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
573 * If this is a per-task event, need to check whether this
574 * event's task is the current task on this cpu.
576 if (ctx->task && cpuctx->task_ctx != ctx)
579 raw_spin_lock(&ctx->lock);
582 * If the event is on, turn it off.
583 * If it is in error state, leave it in error state.
585 if (event->state >= PERF_EVENT_STATE_INACTIVE) {
586 update_context_time(ctx);
587 update_group_times(event);
588 if (event == event->group_leader)
589 group_sched_out(event, cpuctx, ctx);
591 event_sched_out(event, cpuctx, ctx);
592 event->state = PERF_EVENT_STATE_OFF;
595 raw_spin_unlock(&ctx->lock);
601 * If event->ctx is a cloned context, callers must make sure that
602 * every task struct that event->ctx->task could possibly point to
603 * remains valid. This condition is satisifed when called through
604 * perf_event_for_each_child or perf_event_for_each because they
605 * hold the top-level event's child_mutex, so any descendant that
606 * goes to exit will block in sync_child_event.
607 * When called from perf_pending_event it's OK because event->ctx
608 * is the current context on this CPU and preemption is disabled,
609 * hence we can't get into perf_event_task_sched_out for this context.
611 void perf_event_disable(struct perf_event *event)
613 struct perf_event_context *ctx = event->ctx;
614 struct task_struct *task = ctx->task;
618 * Disable the event on the cpu that it's on
620 smp_call_function_single(event->cpu, __perf_event_disable,
626 task_oncpu_function_call(task, __perf_event_disable, event);
628 raw_spin_lock_irq(&ctx->lock);
630 * If the event is still active, we need to retry the cross-call.
632 if (event->state == PERF_EVENT_STATE_ACTIVE) {
633 raw_spin_unlock_irq(&ctx->lock);
638 * Since we have the lock this context can't be scheduled
639 * in, so we can change the state safely.
641 if (event->state == PERF_EVENT_STATE_INACTIVE) {
642 update_group_times(event);
643 event->state = PERF_EVENT_STATE_OFF;
646 raw_spin_unlock_irq(&ctx->lock);
650 event_sched_in(struct perf_event *event,
651 struct perf_cpu_context *cpuctx,
652 struct perf_event_context *ctx)
654 if (event->state <= PERF_EVENT_STATE_OFF)
657 event->state = PERF_EVENT_STATE_ACTIVE;
658 event->oncpu = smp_processor_id();
660 * The new state must be visible before we turn it on in the hardware:
664 if (event->pmu->add(event, PERF_EF_START)) {
665 event->state = PERF_EVENT_STATE_INACTIVE;
670 event->tstamp_running += ctx->time - event->tstamp_stopped;
672 if (!is_software_event(event))
673 cpuctx->active_oncpu++;
676 if (event->attr.exclusive)
677 cpuctx->exclusive = 1;
683 group_sched_in(struct perf_event *group_event,
684 struct perf_cpu_context *cpuctx,
685 struct perf_event_context *ctx)
687 struct perf_event *event, *partial_group = NULL;
688 struct pmu *pmu = group_event->pmu;
690 if (group_event->state == PERF_EVENT_STATE_OFF)
695 if (event_sched_in(group_event, cpuctx, ctx)) {
696 pmu->cancel_txn(pmu);
701 * Schedule in siblings as one group (if any):
703 list_for_each_entry(event, &group_event->sibling_list, group_entry) {
704 if (event_sched_in(event, cpuctx, ctx)) {
705 partial_group = event;
710 if (!pmu->commit_txn(pmu))
715 * Groups can be scheduled in as one unit only, so undo any
716 * partial group before returning:
718 list_for_each_entry(event, &group_event->sibling_list, group_entry) {
719 if (event == partial_group)
721 event_sched_out(event, cpuctx, ctx);
723 event_sched_out(group_event, cpuctx, ctx);
725 pmu->cancel_txn(pmu);
731 * Work out whether we can put this event group on the CPU now.
733 static int group_can_go_on(struct perf_event *event,
734 struct perf_cpu_context *cpuctx,
738 * Groups consisting entirely of software events can always go on.
740 if (event->group_flags & PERF_GROUP_SOFTWARE)
743 * If an exclusive group is already on, no other hardware
746 if (cpuctx->exclusive)
749 * If this group is exclusive and there are already
750 * events on the CPU, it can't go on.
752 if (event->attr.exclusive && cpuctx->active_oncpu)
755 * Otherwise, try to add it if all previous groups were able
761 static void add_event_to_ctx(struct perf_event *event,
762 struct perf_event_context *ctx)
764 list_add_event(event, ctx);
765 perf_group_attach(event);
766 event->tstamp_enabled = ctx->time;
767 event->tstamp_running = ctx->time;
768 event->tstamp_stopped = ctx->time;
772 * Cross CPU call to install and enable a performance event
774 * Must be called with ctx->mutex held
776 static void __perf_install_in_context(void *info)
778 struct perf_event *event = info;
779 struct perf_event_context *ctx = event->ctx;
780 struct perf_event *leader = event->group_leader;
781 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
785 * If this is a task context, we need to check whether it is
786 * the current task context of this cpu. If not it has been
787 * scheduled out before the smp call arrived.
788 * Or possibly this is the right context but it isn't
789 * on this cpu because it had no events.
791 if (ctx->task && cpuctx->task_ctx != ctx) {
792 if (cpuctx->task_ctx || ctx->task != current)
794 cpuctx->task_ctx = ctx;
797 raw_spin_lock(&ctx->lock);
799 update_context_time(ctx);
801 add_event_to_ctx(event, ctx);
803 if (event->cpu != -1 && event->cpu != smp_processor_id())
807 * Don't put the event on if it is disabled or if
808 * it is in a group and the group isn't on.
810 if (event->state != PERF_EVENT_STATE_INACTIVE ||
811 (leader != event && leader->state != PERF_EVENT_STATE_ACTIVE))
815 * An exclusive event can't go on if there are already active
816 * hardware events, and no hardware event can go on if there
817 * is already an exclusive event on.
819 if (!group_can_go_on(event, cpuctx, 1))
822 err = event_sched_in(event, cpuctx, ctx);
826 * This event couldn't go on. If it is in a group
827 * then we have to pull the whole group off.
828 * If the event group is pinned then put it in error state.
831 group_sched_out(leader, cpuctx, ctx);
832 if (leader->attr.pinned) {
833 update_group_times(leader);
834 leader->state = PERF_EVENT_STATE_ERROR;
839 raw_spin_unlock(&ctx->lock);
843 * Attach a performance event to a context
845 * First we add the event to the list with the hardware enable bit
846 * in event->hw_config cleared.
848 * If the event is attached to a task which is on a CPU we use a smp
849 * call to enable it in the task context. The task might have been
850 * scheduled away, but we check this in the smp call again.
852 * Must be called with ctx->mutex held.
855 perf_install_in_context(struct perf_event_context *ctx,
856 struct perf_event *event,
859 struct task_struct *task = ctx->task;
865 * Per cpu events are installed via an smp call and
866 * the install is always successful.
868 smp_call_function_single(cpu, __perf_install_in_context,
874 task_oncpu_function_call(task, __perf_install_in_context,
877 raw_spin_lock_irq(&ctx->lock);
879 * we need to retry the smp call.
881 if (ctx->is_active && list_empty(&event->group_entry)) {
882 raw_spin_unlock_irq(&ctx->lock);
887 * The lock prevents that this context is scheduled in so we
888 * can add the event safely, if it the call above did not
891 if (list_empty(&event->group_entry))
892 add_event_to_ctx(event, ctx);
893 raw_spin_unlock_irq(&ctx->lock);
897 * Put a event into inactive state and update time fields.
898 * Enabling the leader of a group effectively enables all
899 * the group members that aren't explicitly disabled, so we
900 * have to update their ->tstamp_enabled also.
901 * Note: this works for group members as well as group leaders
902 * since the non-leader members' sibling_lists will be empty.
904 static void __perf_event_mark_enabled(struct perf_event *event,
905 struct perf_event_context *ctx)
907 struct perf_event *sub;
909 event->state = PERF_EVENT_STATE_INACTIVE;
910 event->tstamp_enabled = ctx->time - event->total_time_enabled;
911 list_for_each_entry(sub, &event->sibling_list, group_entry) {
912 if (sub->state >= PERF_EVENT_STATE_INACTIVE) {
913 sub->tstamp_enabled =
914 ctx->time - sub->total_time_enabled;
920 * Cross CPU call to enable a performance event
922 static void __perf_event_enable(void *info)
924 struct perf_event *event = info;
925 struct perf_event_context *ctx = event->ctx;
926 struct perf_event *leader = event->group_leader;
927 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
931 * If this is a per-task event, need to check whether this
932 * event's task is the current task on this cpu.
934 if (ctx->task && cpuctx->task_ctx != ctx) {
935 if (cpuctx->task_ctx || ctx->task != current)
937 cpuctx->task_ctx = ctx;
940 raw_spin_lock(&ctx->lock);
942 update_context_time(ctx);
944 if (event->state >= PERF_EVENT_STATE_INACTIVE)
946 __perf_event_mark_enabled(event, ctx);
948 if (event->cpu != -1 && event->cpu != smp_processor_id())
952 * If the event is in a group and isn't the group leader,
953 * then don't put it on unless the group is on.
955 if (leader != event && leader->state != PERF_EVENT_STATE_ACTIVE)
958 if (!group_can_go_on(event, cpuctx, 1)) {
962 err = group_sched_in(event, cpuctx, ctx);
964 err = event_sched_in(event, cpuctx, ctx);
969 * If this event can't go on and it's part of a
970 * group, then the whole group has to come off.
973 group_sched_out(leader, cpuctx, ctx);
974 if (leader->attr.pinned) {
975 update_group_times(leader);
976 leader->state = PERF_EVENT_STATE_ERROR;
981 raw_spin_unlock(&ctx->lock);
987 * If event->ctx is a cloned context, callers must make sure that
988 * every task struct that event->ctx->task could possibly point to
989 * remains valid. This condition is satisfied when called through
990 * perf_event_for_each_child or perf_event_for_each as described
991 * for perf_event_disable.
993 void perf_event_enable(struct perf_event *event)
995 struct perf_event_context *ctx = event->ctx;
996 struct task_struct *task = ctx->task;
1000 * Enable the event on the cpu that it's on
1002 smp_call_function_single(event->cpu, __perf_event_enable,
1007 raw_spin_lock_irq(&ctx->lock);
1008 if (event->state >= PERF_EVENT_STATE_INACTIVE)
1012 * If the event is in error state, clear that first.
1013 * That way, if we see the event in error state below, we
1014 * know that it has gone back into error state, as distinct
1015 * from the task having been scheduled away before the
1016 * cross-call arrived.
1018 if (event->state == PERF_EVENT_STATE_ERROR)
1019 event->state = PERF_EVENT_STATE_OFF;
1022 raw_spin_unlock_irq(&ctx->lock);
1023 task_oncpu_function_call(task, __perf_event_enable, event);
1025 raw_spin_lock_irq(&ctx->lock);
1028 * If the context is active and the event is still off,
1029 * we need to retry the cross-call.
1031 if (ctx->is_active && event->state == PERF_EVENT_STATE_OFF)
1035 * Since we have the lock this context can't be scheduled
1036 * in, so we can change the state safely.
1038 if (event->state == PERF_EVENT_STATE_OFF)
1039 __perf_event_mark_enabled(event, ctx);
1042 raw_spin_unlock_irq(&ctx->lock);
1045 static int perf_event_refresh(struct perf_event *event, int refresh)
1048 * not supported on inherited events
1050 if (event->attr.inherit)
1053 atomic_add(refresh, &event->event_limit);
1054 perf_event_enable(event);
1060 EVENT_FLEXIBLE = 0x1,
1062 EVENT_ALL = EVENT_FLEXIBLE | EVENT_PINNED,
1065 static void ctx_sched_out(struct perf_event_context *ctx,
1066 struct perf_cpu_context *cpuctx,
1067 enum event_type_t event_type)
1069 struct perf_event *event;
1071 raw_spin_lock(&ctx->lock);
1072 perf_pmu_disable(ctx->pmu);
1074 if (likely(!ctx->nr_events))
1076 update_context_time(ctx);
1078 if (!ctx->nr_active)
1081 if (event_type & EVENT_PINNED) {
1082 list_for_each_entry(event, &ctx->pinned_groups, group_entry)
1083 group_sched_out(event, cpuctx, ctx);
1086 if (event_type & EVENT_FLEXIBLE) {
1087 list_for_each_entry(event, &ctx->flexible_groups, group_entry)
1088 group_sched_out(event, cpuctx, ctx);
1091 perf_pmu_enable(ctx->pmu);
1092 raw_spin_unlock(&ctx->lock);
1096 * Test whether two contexts are equivalent, i.e. whether they
1097 * have both been cloned from the same version of the same context
1098 * and they both have the same number of enabled events.
1099 * If the number of enabled events is the same, then the set
1100 * of enabled events should be the same, because these are both
1101 * inherited contexts, therefore we can't access individual events
1102 * in them directly with an fd; we can only enable/disable all
1103 * events via prctl, or enable/disable all events in a family
1104 * via ioctl, which will have the same effect on both contexts.
1106 static int context_equiv(struct perf_event_context *ctx1,
1107 struct perf_event_context *ctx2)
1109 return ctx1->parent_ctx && ctx1->parent_ctx == ctx2->parent_ctx
1110 && ctx1->parent_gen == ctx2->parent_gen
1111 && !ctx1->pin_count && !ctx2->pin_count;
1114 static void __perf_event_sync_stat(struct perf_event *event,
1115 struct perf_event *next_event)
1119 if (!event->attr.inherit_stat)
1123 * Update the event value, we cannot use perf_event_read()
1124 * because we're in the middle of a context switch and have IRQs
1125 * disabled, which upsets smp_call_function_single(), however
1126 * we know the event must be on the current CPU, therefore we
1127 * don't need to use it.
1129 switch (event->state) {
1130 case PERF_EVENT_STATE_ACTIVE:
1131 event->pmu->read(event);
1134 case PERF_EVENT_STATE_INACTIVE:
1135 update_event_times(event);
1143 * In order to keep per-task stats reliable we need to flip the event
1144 * values when we flip the contexts.
1146 value = local64_read(&next_event->count);
1147 value = local64_xchg(&event->count, value);
1148 local64_set(&next_event->count, value);
1150 swap(event->total_time_enabled, next_event->total_time_enabled);
1151 swap(event->total_time_running, next_event->total_time_running);
1154 * Since we swizzled the values, update the user visible data too.
1156 perf_event_update_userpage(event);
1157 perf_event_update_userpage(next_event);
1160 #define list_next_entry(pos, member) \
1161 list_entry(pos->member.next, typeof(*pos), member)
1163 static void perf_event_sync_stat(struct perf_event_context *ctx,
1164 struct perf_event_context *next_ctx)
1166 struct perf_event *event, *next_event;
1171 update_context_time(ctx);
1173 event = list_first_entry(&ctx->event_list,
1174 struct perf_event, event_entry);
1176 next_event = list_first_entry(&next_ctx->event_list,
1177 struct perf_event, event_entry);
1179 while (&event->event_entry != &ctx->event_list &&
1180 &next_event->event_entry != &next_ctx->event_list) {
1182 __perf_event_sync_stat(event, next_event);
1184 event = list_next_entry(event, event_entry);
1185 next_event = list_next_entry(next_event, event_entry);
1189 void perf_event_context_sched_out(struct task_struct *task, int ctxn,
1190 struct task_struct *next)
1192 struct perf_event_context *ctx = task->perf_event_ctxp[ctxn];
1193 struct perf_event_context *next_ctx;
1194 struct perf_event_context *parent;
1195 struct perf_cpu_context *cpuctx;
1201 cpuctx = __get_cpu_context(ctx);
1202 if (!cpuctx->task_ctx)
1206 parent = rcu_dereference(ctx->parent_ctx);
1207 next_ctx = next->perf_event_ctxp[ctxn];
1208 if (parent && next_ctx &&
1209 rcu_dereference(next_ctx->parent_ctx) == parent) {
1211 * Looks like the two contexts are clones, so we might be
1212 * able to optimize the context switch. We lock both
1213 * contexts and check that they are clones under the
1214 * lock (including re-checking that neither has been
1215 * uncloned in the meantime). It doesn't matter which
1216 * order we take the locks because no other cpu could
1217 * be trying to lock both of these tasks.
1219 raw_spin_lock(&ctx->lock);
1220 raw_spin_lock_nested(&next_ctx->lock, SINGLE_DEPTH_NESTING);
1221 if (context_equiv(ctx, next_ctx)) {
1223 * XXX do we need a memory barrier of sorts
1224 * wrt to rcu_dereference() of perf_event_ctxp
1226 task->perf_event_ctxp[ctxn] = next_ctx;
1227 next->perf_event_ctxp[ctxn] = ctx;
1229 next_ctx->task = task;
1232 perf_event_sync_stat(ctx, next_ctx);
1234 raw_spin_unlock(&next_ctx->lock);
1235 raw_spin_unlock(&ctx->lock);
1240 ctx_sched_out(ctx, cpuctx, EVENT_ALL);
1241 cpuctx->task_ctx = NULL;
1245 #define for_each_task_context_nr(ctxn) \
1246 for ((ctxn) = 0; (ctxn) < perf_nr_task_contexts; (ctxn)++)
1249 * Called from scheduler to remove the events of the current task,
1250 * with interrupts disabled.
1252 * We stop each event and update the event value in event->count.
1254 * This does not protect us against NMI, but disable()
1255 * sets the disabled bit in the control field of event _before_
1256 * accessing the event control register. If a NMI hits, then it will
1257 * not restart the event.
1259 void perf_event_task_sched_out(struct task_struct *task,
1260 struct task_struct *next)
1264 perf_sw_event(PERF_COUNT_SW_CONTEXT_SWITCHES, 1, 1, NULL, 0);
1266 for_each_task_context_nr(ctxn)
1267 perf_event_context_sched_out(task, ctxn, next);
1270 static void task_ctx_sched_out(struct perf_event_context *ctx,
1271 enum event_type_t event_type)
1273 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
1275 if (!cpuctx->task_ctx)
1278 if (WARN_ON_ONCE(ctx != cpuctx->task_ctx))
1281 ctx_sched_out(ctx, cpuctx, event_type);
1282 cpuctx->task_ctx = NULL;
1286 * Called with IRQs disabled
1288 static void __perf_event_task_sched_out(struct perf_event_context *ctx)
1290 task_ctx_sched_out(ctx, EVENT_ALL);
1294 * Called with IRQs disabled
1296 static void cpu_ctx_sched_out(struct perf_cpu_context *cpuctx,
1297 enum event_type_t event_type)
1299 ctx_sched_out(&cpuctx->ctx, cpuctx, event_type);
1303 ctx_pinned_sched_in(struct perf_event_context *ctx,
1304 struct perf_cpu_context *cpuctx)
1306 struct perf_event *event;
1308 list_for_each_entry(event, &ctx->pinned_groups, group_entry) {
1309 if (event->state <= PERF_EVENT_STATE_OFF)
1311 if (event->cpu != -1 && event->cpu != smp_processor_id())
1314 if (group_can_go_on(event, cpuctx, 1))
1315 group_sched_in(event, cpuctx, ctx);
1318 * If this pinned group hasn't been scheduled,
1319 * put it in error state.
1321 if (event->state == PERF_EVENT_STATE_INACTIVE) {
1322 update_group_times(event);
1323 event->state = PERF_EVENT_STATE_ERROR;
1329 ctx_flexible_sched_in(struct perf_event_context *ctx,
1330 struct perf_cpu_context *cpuctx)
1332 struct perf_event *event;
1335 list_for_each_entry(event, &ctx->flexible_groups, group_entry) {
1336 /* Ignore events in OFF or ERROR state */
1337 if (event->state <= PERF_EVENT_STATE_OFF)
1340 * Listen to the 'cpu' scheduling filter constraint
1343 if (event->cpu != -1 && event->cpu != smp_processor_id())
1346 if (group_can_go_on(event, cpuctx, can_add_hw)) {
1347 if (group_sched_in(event, cpuctx, ctx))
1354 ctx_sched_in(struct perf_event_context *ctx,
1355 struct perf_cpu_context *cpuctx,
1356 enum event_type_t event_type)
1358 raw_spin_lock(&ctx->lock);
1360 if (likely(!ctx->nr_events))
1363 ctx->timestamp = perf_clock();
1366 * First go through the list and put on any pinned groups
1367 * in order to give them the best chance of going on.
1369 if (event_type & EVENT_PINNED)
1370 ctx_pinned_sched_in(ctx, cpuctx);
1372 /* Then walk through the lower prio flexible groups */
1373 if (event_type & EVENT_FLEXIBLE)
1374 ctx_flexible_sched_in(ctx, cpuctx);
1377 raw_spin_unlock(&ctx->lock);
1380 static void cpu_ctx_sched_in(struct perf_cpu_context *cpuctx,
1381 enum event_type_t event_type)
1383 struct perf_event_context *ctx = &cpuctx->ctx;
1385 ctx_sched_in(ctx, cpuctx, event_type);
1388 static void task_ctx_sched_in(struct perf_event_context *ctx,
1389 enum event_type_t event_type)
1391 struct perf_cpu_context *cpuctx;
1393 cpuctx = __get_cpu_context(ctx);
1394 if (cpuctx->task_ctx == ctx)
1397 ctx_sched_in(ctx, cpuctx, event_type);
1398 cpuctx->task_ctx = ctx;
1401 void perf_event_context_sched_in(struct perf_event_context *ctx)
1403 struct perf_cpu_context *cpuctx;
1405 cpuctx = __get_cpu_context(ctx);
1406 if (cpuctx->task_ctx == ctx)
1409 perf_pmu_disable(ctx->pmu);
1411 * We want to keep the following priority order:
1412 * cpu pinned (that don't need to move), task pinned,
1413 * cpu flexible, task flexible.
1415 cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE);
1417 ctx_sched_in(ctx, cpuctx, EVENT_PINNED);
1418 cpu_ctx_sched_in(cpuctx, EVENT_FLEXIBLE);
1419 ctx_sched_in(ctx, cpuctx, EVENT_FLEXIBLE);
1421 cpuctx->task_ctx = ctx;
1424 * Since these rotations are per-cpu, we need to ensure the
1425 * cpu-context we got scheduled on is actually rotating.
1427 perf_pmu_rotate_start(ctx->pmu);
1428 perf_pmu_enable(ctx->pmu);
1432 * Called from scheduler to add the events of the current task
1433 * with interrupts disabled.
1435 * We restore the event value and then enable it.
1437 * This does not protect us against NMI, but enable()
1438 * sets the enabled bit in the control field of event _before_
1439 * accessing the event control register. If a NMI hits, then it will
1440 * keep the event running.
1442 void perf_event_task_sched_in(struct task_struct *task)
1444 struct perf_event_context *ctx;
1447 for_each_task_context_nr(ctxn) {
1448 ctx = task->perf_event_ctxp[ctxn];
1452 perf_event_context_sched_in(ctx);
1456 #define MAX_INTERRUPTS (~0ULL)
1458 static void perf_log_throttle(struct perf_event *event, int enable);
1460 static u64 perf_calculate_period(struct perf_event *event, u64 nsec, u64 count)
1462 u64 frequency = event->attr.sample_freq;
1463 u64 sec = NSEC_PER_SEC;
1464 u64 divisor, dividend;
1466 int count_fls, nsec_fls, frequency_fls, sec_fls;
1468 count_fls = fls64(count);
1469 nsec_fls = fls64(nsec);
1470 frequency_fls = fls64(frequency);
1474 * We got @count in @nsec, with a target of sample_freq HZ
1475 * the target period becomes:
1478 * period = -------------------
1479 * @nsec * sample_freq
1484 * Reduce accuracy by one bit such that @a and @b converge
1485 * to a similar magnitude.
1487 #define REDUCE_FLS(a, b) \
1489 if (a##_fls > b##_fls) { \
1499 * Reduce accuracy until either term fits in a u64, then proceed with
1500 * the other, so that finally we can do a u64/u64 division.
1502 while (count_fls + sec_fls > 64 && nsec_fls + frequency_fls > 64) {
1503 REDUCE_FLS(nsec, frequency);
1504 REDUCE_FLS(sec, count);
1507 if (count_fls + sec_fls > 64) {
1508 divisor = nsec * frequency;
1510 while (count_fls + sec_fls > 64) {
1511 REDUCE_FLS(count, sec);
1515 dividend = count * sec;
1517 dividend = count * sec;
1519 while (nsec_fls + frequency_fls > 64) {
1520 REDUCE_FLS(nsec, frequency);
1524 divisor = nsec * frequency;
1530 return div64_u64(dividend, divisor);
1533 static void perf_adjust_period(struct perf_event *event, u64 nsec, u64 count)
1535 struct hw_perf_event *hwc = &event->hw;
1536 s64 period, sample_period;
1539 period = perf_calculate_period(event, nsec, count);
1541 delta = (s64)(period - hwc->sample_period);
1542 delta = (delta + 7) / 8; /* low pass filter */
1544 sample_period = hwc->sample_period + delta;
1549 hwc->sample_period = sample_period;
1551 if (local64_read(&hwc->period_left) > 8*sample_period) {
1552 event->pmu->stop(event, PERF_EF_UPDATE);
1553 local64_set(&hwc->period_left, 0);
1554 event->pmu->start(event, PERF_EF_RELOAD);
1558 static void perf_ctx_adjust_freq(struct perf_event_context *ctx, u64 period)
1560 struct perf_event *event;
1561 struct hw_perf_event *hwc;
1562 u64 interrupts, now;
1565 raw_spin_lock(&ctx->lock);
1566 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
1567 if (event->state != PERF_EVENT_STATE_ACTIVE)
1570 if (event->cpu != -1 && event->cpu != smp_processor_id())
1575 interrupts = hwc->interrupts;
1576 hwc->interrupts = 0;
1579 * unthrottle events on the tick
1581 if (interrupts == MAX_INTERRUPTS) {
1582 perf_log_throttle(event, 1);
1583 event->pmu->start(event, 0);
1586 if (!event->attr.freq || !event->attr.sample_freq)
1589 event->pmu->read(event);
1590 now = local64_read(&event->count);
1591 delta = now - hwc->freq_count_stamp;
1592 hwc->freq_count_stamp = now;
1595 perf_adjust_period(event, period, delta);
1597 raw_spin_unlock(&ctx->lock);
1601 * Round-robin a context's events:
1603 static void rotate_ctx(struct perf_event_context *ctx)
1605 raw_spin_lock(&ctx->lock);
1607 /* Rotate the first entry last of non-pinned groups */
1608 list_rotate_left(&ctx->flexible_groups);
1610 raw_spin_unlock(&ctx->lock);
1614 * perf_pmu_rotate_start() and perf_rotate_context() are fully serialized
1615 * because they're strictly cpu affine and rotate_start is called with IRQs
1616 * disabled, while rotate_context is called from IRQ context.
1618 static void perf_rotate_context(struct perf_cpu_context *cpuctx)
1620 u64 interval = (u64)cpuctx->jiffies_interval * TICK_NSEC;
1621 struct perf_event_context *ctx = NULL;
1622 int rotate = 0, remove = 1;
1624 if (cpuctx->ctx.nr_events) {
1626 if (cpuctx->ctx.nr_events != cpuctx->ctx.nr_active)
1630 ctx = cpuctx->task_ctx;
1631 if (ctx && ctx->nr_events) {
1633 if (ctx->nr_events != ctx->nr_active)
1637 perf_pmu_disable(cpuctx->ctx.pmu);
1638 perf_ctx_adjust_freq(&cpuctx->ctx, interval);
1640 perf_ctx_adjust_freq(ctx, interval);
1645 cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE);
1647 task_ctx_sched_out(ctx, EVENT_FLEXIBLE);
1649 rotate_ctx(&cpuctx->ctx);
1653 cpu_ctx_sched_in(cpuctx, EVENT_FLEXIBLE);
1655 task_ctx_sched_in(ctx, EVENT_FLEXIBLE);
1659 list_del_init(&cpuctx->rotation_list);
1661 perf_pmu_enable(cpuctx->ctx.pmu);
1664 void perf_event_task_tick(void)
1666 struct list_head *head = &__get_cpu_var(rotation_list);
1667 struct perf_cpu_context *cpuctx, *tmp;
1669 WARN_ON(!irqs_disabled());
1671 list_for_each_entry_safe(cpuctx, tmp, head, rotation_list) {
1672 if (cpuctx->jiffies_interval == 1 ||
1673 !(jiffies % cpuctx->jiffies_interval))
1674 perf_rotate_context(cpuctx);
1678 static int event_enable_on_exec(struct perf_event *event,
1679 struct perf_event_context *ctx)
1681 if (!event->attr.enable_on_exec)
1684 event->attr.enable_on_exec = 0;
1685 if (event->state >= PERF_EVENT_STATE_INACTIVE)
1688 __perf_event_mark_enabled(event, ctx);
1694 * Enable all of a task's events that have been marked enable-on-exec.
1695 * This expects task == current.
1697 static void perf_event_enable_on_exec(struct perf_event_context *ctx)
1699 struct perf_event *event;
1700 unsigned long flags;
1704 local_irq_save(flags);
1705 if (!ctx || !ctx->nr_events)
1708 task_ctx_sched_out(ctx, EVENT_ALL);
1710 raw_spin_lock(&ctx->lock);
1712 list_for_each_entry(event, &ctx->pinned_groups, group_entry) {
1713 ret = event_enable_on_exec(event, ctx);
1718 list_for_each_entry(event, &ctx->flexible_groups, group_entry) {
1719 ret = event_enable_on_exec(event, ctx);
1725 * Unclone this context if we enabled any event.
1730 raw_spin_unlock(&ctx->lock);
1732 perf_event_context_sched_in(ctx);
1734 local_irq_restore(flags);
1738 * Cross CPU call to read the hardware event
1740 static void __perf_event_read(void *info)
1742 struct perf_event *event = info;
1743 struct perf_event_context *ctx = event->ctx;
1744 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
1747 * If this is a task context, we need to check whether it is
1748 * the current task context of this cpu. If not it has been
1749 * scheduled out before the smp call arrived. In that case
1750 * event->count would have been updated to a recent sample
1751 * when the event was scheduled out.
1753 if (ctx->task && cpuctx->task_ctx != ctx)
1756 raw_spin_lock(&ctx->lock);
1757 update_context_time(ctx);
1758 update_event_times(event);
1759 raw_spin_unlock(&ctx->lock);
1761 event->pmu->read(event);
1764 static inline u64 perf_event_count(struct perf_event *event)
1766 return local64_read(&event->count) + atomic64_read(&event->child_count);
1769 static u64 perf_event_read(struct perf_event *event)
1772 * If event is enabled and currently active on a CPU, update the
1773 * value in the event structure:
1775 if (event->state == PERF_EVENT_STATE_ACTIVE) {
1776 smp_call_function_single(event->oncpu,
1777 __perf_event_read, event, 1);
1778 } else if (event->state == PERF_EVENT_STATE_INACTIVE) {
1779 struct perf_event_context *ctx = event->ctx;
1780 unsigned long flags;
1782 raw_spin_lock_irqsave(&ctx->lock, flags);
1783 update_context_time(ctx);
1784 update_event_times(event);
1785 raw_spin_unlock_irqrestore(&ctx->lock, flags);
1788 return perf_event_count(event);
1795 struct callchain_cpus_entries {
1796 struct rcu_head rcu_head;
1797 struct perf_callchain_entry *cpu_entries[0];
1800 static DEFINE_PER_CPU(int, callchain_recursion[PERF_NR_CONTEXTS]);
1801 static atomic_t nr_callchain_events;
1802 static DEFINE_MUTEX(callchain_mutex);
1803 struct callchain_cpus_entries *callchain_cpus_entries;
1806 __weak void perf_callchain_kernel(struct perf_callchain_entry *entry,
1807 struct pt_regs *regs)
1811 __weak void perf_callchain_user(struct perf_callchain_entry *entry,
1812 struct pt_regs *regs)
1816 static void release_callchain_buffers_rcu(struct rcu_head *head)
1818 struct callchain_cpus_entries *entries;
1821 entries = container_of(head, struct callchain_cpus_entries, rcu_head);
1823 for_each_possible_cpu(cpu)
1824 kfree(entries->cpu_entries[cpu]);
1829 static void release_callchain_buffers(void)
1831 struct callchain_cpus_entries *entries;
1833 entries = callchain_cpus_entries;
1834 rcu_assign_pointer(callchain_cpus_entries, NULL);
1835 call_rcu(&entries->rcu_head, release_callchain_buffers_rcu);
1838 static int alloc_callchain_buffers(void)
1842 struct callchain_cpus_entries *entries;
1845 * We can't use the percpu allocation API for data that can be
1846 * accessed from NMI. Use a temporary manual per cpu allocation
1847 * until that gets sorted out.
1849 size = sizeof(*entries) + sizeof(struct perf_callchain_entry *) *
1850 num_possible_cpus();
1852 entries = kzalloc(size, GFP_KERNEL);
1856 size = sizeof(struct perf_callchain_entry) * PERF_NR_CONTEXTS;
1858 for_each_possible_cpu(cpu) {
1859 entries->cpu_entries[cpu] = kmalloc_node(size, GFP_KERNEL,
1861 if (!entries->cpu_entries[cpu])
1865 rcu_assign_pointer(callchain_cpus_entries, entries);
1870 for_each_possible_cpu(cpu)
1871 kfree(entries->cpu_entries[cpu]);
1877 static int get_callchain_buffers(void)
1882 mutex_lock(&callchain_mutex);
1884 count = atomic_inc_return(&nr_callchain_events);
1885 if (WARN_ON_ONCE(count < 1)) {
1891 /* If the allocation failed, give up */
1892 if (!callchain_cpus_entries)
1897 err = alloc_callchain_buffers();
1899 release_callchain_buffers();
1901 mutex_unlock(&callchain_mutex);
1906 static void put_callchain_buffers(void)
1908 if (atomic_dec_and_mutex_lock(&nr_callchain_events, &callchain_mutex)) {
1909 release_callchain_buffers();
1910 mutex_unlock(&callchain_mutex);
1914 static int get_recursion_context(int *recursion)
1922 else if (in_softirq())
1927 if (recursion[rctx])
1936 static inline void put_recursion_context(int *recursion, int rctx)
1942 static struct perf_callchain_entry *get_callchain_entry(int *rctx)
1945 struct callchain_cpus_entries *entries;
1947 *rctx = get_recursion_context(__get_cpu_var(callchain_recursion));
1951 entries = rcu_dereference(callchain_cpus_entries);
1955 cpu = smp_processor_id();
1957 return &entries->cpu_entries[cpu][*rctx];
1961 put_callchain_entry(int rctx)
1963 put_recursion_context(__get_cpu_var(callchain_recursion), rctx);
1966 static struct perf_callchain_entry *perf_callchain(struct pt_regs *regs)
1969 struct perf_callchain_entry *entry;
1972 entry = get_callchain_entry(&rctx);
1981 if (!user_mode(regs)) {
1982 perf_callchain_store(entry, PERF_CONTEXT_KERNEL);
1983 perf_callchain_kernel(entry, regs);
1985 regs = task_pt_regs(current);
1991 perf_callchain_store(entry, PERF_CONTEXT_USER);
1992 perf_callchain_user(entry, regs);
1996 put_callchain_entry(rctx);
2002 * Initialize the perf_event context in a task_struct:
2004 static void __perf_event_init_context(struct perf_event_context *ctx)
2006 raw_spin_lock_init(&ctx->lock);
2007 mutex_init(&ctx->mutex);
2008 INIT_LIST_HEAD(&ctx->pinned_groups);
2009 INIT_LIST_HEAD(&ctx->flexible_groups);
2010 INIT_LIST_HEAD(&ctx->event_list);
2011 atomic_set(&ctx->refcount, 1);
2014 static struct perf_event_context *
2015 alloc_perf_context(struct pmu *pmu, struct task_struct *task)
2017 struct perf_event_context *ctx;
2019 ctx = kzalloc(sizeof(struct perf_event_context), GFP_KERNEL);
2023 __perf_event_init_context(ctx);
2026 get_task_struct(task);
2033 static struct task_struct *
2034 find_lively_task_by_vpid(pid_t vpid)
2036 struct task_struct *task;
2043 task = find_task_by_vpid(vpid);
2045 get_task_struct(task);
2049 return ERR_PTR(-ESRCH);
2052 * Can't attach events to a dying task.
2055 if (task->flags & PF_EXITING)
2058 /* Reuse ptrace permission checks for now. */
2060 if (!ptrace_may_access(task, PTRACE_MODE_READ))
2065 put_task_struct(task);
2066 return ERR_PTR(err);
2070 static struct perf_event_context *
2071 find_get_context(struct pmu *pmu, struct task_struct *task, int cpu)
2073 struct perf_event_context *ctx;
2074 struct perf_cpu_context *cpuctx;
2075 unsigned long flags;
2078 if (!task && cpu != -1) {
2079 /* Must be root to operate on a CPU event: */
2080 if (perf_paranoid_cpu() && !capable(CAP_SYS_ADMIN))
2081 return ERR_PTR(-EACCES);
2083 if (cpu < 0 || cpu >= nr_cpumask_bits)
2084 return ERR_PTR(-EINVAL);
2087 * We could be clever and allow to attach a event to an
2088 * offline CPU and activate it when the CPU comes up, but
2091 if (!cpu_online(cpu))
2092 return ERR_PTR(-ENODEV);
2094 cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu);
2102 ctxn = pmu->task_ctx_nr;
2107 ctx = perf_lock_task_context(task, ctxn, &flags);
2110 raw_spin_unlock_irqrestore(&ctx->lock, flags);
2114 ctx = alloc_perf_context(pmu, task);
2121 if (cmpxchg(&task->perf_event_ctxp[ctxn], NULL, ctx)) {
2123 * We raced with some other task; use
2124 * the context they set.
2126 put_task_struct(task);
2132 put_task_struct(task);
2136 put_task_struct(task);
2137 return ERR_PTR(err);
2140 static void perf_event_free_filter(struct perf_event *event);
2142 static void free_event_rcu(struct rcu_head *head)
2144 struct perf_event *event;
2146 event = container_of(head, struct perf_event, rcu_head);
2148 put_pid_ns(event->ns);
2149 perf_event_free_filter(event);
2153 static void perf_pending_sync(struct perf_event *event);
2154 static void perf_buffer_put(struct perf_buffer *buffer);
2156 static void free_event(struct perf_event *event)
2158 perf_pending_sync(event);
2160 if (!event->parent) {
2161 atomic_dec(&nr_events);
2162 if (event->attr.mmap || event->attr.mmap_data)
2163 atomic_dec(&nr_mmap_events);
2164 if (event->attr.comm)
2165 atomic_dec(&nr_comm_events);
2166 if (event->attr.task)
2167 atomic_dec(&nr_task_events);
2168 if (event->attr.sample_type & PERF_SAMPLE_CALLCHAIN)
2169 put_callchain_buffers();
2172 if (event->buffer) {
2173 perf_buffer_put(event->buffer);
2174 event->buffer = NULL;
2178 event->destroy(event);
2181 put_ctx(event->ctx);
2183 call_rcu(&event->rcu_head, free_event_rcu);
2186 int perf_event_release_kernel(struct perf_event *event)
2188 struct perf_event_context *ctx = event->ctx;
2191 * Remove from the PMU, can't get re-enabled since we got
2192 * here because the last ref went.
2194 perf_event_disable(event);
2196 WARN_ON_ONCE(ctx->parent_ctx);
2198 * There are two ways this annotation is useful:
2200 * 1) there is a lock recursion from perf_event_exit_task
2201 * see the comment there.
2203 * 2) there is a lock-inversion with mmap_sem through
2204 * perf_event_read_group(), which takes faults while
2205 * holding ctx->mutex, however this is called after
2206 * the last filedesc died, so there is no possibility
2207 * to trigger the AB-BA case.
2209 mutex_lock_nested(&ctx->mutex, SINGLE_DEPTH_NESTING);
2210 raw_spin_lock_irq(&ctx->lock);
2211 perf_group_detach(event);
2212 list_del_event(event, ctx);
2213 raw_spin_unlock_irq(&ctx->lock);
2214 mutex_unlock(&ctx->mutex);
2216 mutex_lock(&event->owner->perf_event_mutex);
2217 list_del_init(&event->owner_entry);
2218 mutex_unlock(&event->owner->perf_event_mutex);
2219 put_task_struct(event->owner);
2225 EXPORT_SYMBOL_GPL(perf_event_release_kernel);
2228 * Called when the last reference to the file is gone.
2230 static int perf_release(struct inode *inode, struct file *file)
2232 struct perf_event *event = file->private_data;
2234 file->private_data = NULL;
2236 return perf_event_release_kernel(event);
2239 static int perf_event_read_size(struct perf_event *event)
2241 int entry = sizeof(u64); /* value */
2245 if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
2246 size += sizeof(u64);
2248 if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
2249 size += sizeof(u64);
2251 if (event->attr.read_format & PERF_FORMAT_ID)
2252 entry += sizeof(u64);
2254 if (event->attr.read_format & PERF_FORMAT_GROUP) {
2255 nr += event->group_leader->nr_siblings;
2256 size += sizeof(u64);
2264 u64 perf_event_read_value(struct perf_event *event, u64 *enabled, u64 *running)
2266 struct perf_event *child;
2272 mutex_lock(&event->child_mutex);
2273 total += perf_event_read(event);
2274 *enabled += event->total_time_enabled +
2275 atomic64_read(&event->child_total_time_enabled);
2276 *running += event->total_time_running +
2277 atomic64_read(&event->child_total_time_running);
2279 list_for_each_entry(child, &event->child_list, child_list) {
2280 total += perf_event_read(child);
2281 *enabled += child->total_time_enabled;
2282 *running += child->total_time_running;
2284 mutex_unlock(&event->child_mutex);
2288 EXPORT_SYMBOL_GPL(perf_event_read_value);
2290 static int perf_event_read_group(struct perf_event *event,
2291 u64 read_format, char __user *buf)
2293 struct perf_event *leader = event->group_leader, *sub;
2294 int n = 0, size = 0, ret = -EFAULT;
2295 struct perf_event_context *ctx = leader->ctx;
2297 u64 count, enabled, running;
2299 mutex_lock(&ctx->mutex);
2300 count = perf_event_read_value(leader, &enabled, &running);
2302 values[n++] = 1 + leader->nr_siblings;
2303 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
2304 values[n++] = enabled;
2305 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
2306 values[n++] = running;
2307 values[n++] = count;
2308 if (read_format & PERF_FORMAT_ID)
2309 values[n++] = primary_event_id(leader);
2311 size = n * sizeof(u64);
2313 if (copy_to_user(buf, values, size))
2318 list_for_each_entry(sub, &leader->sibling_list, group_entry) {
2321 values[n++] = perf_event_read_value(sub, &enabled, &running);
2322 if (read_format & PERF_FORMAT_ID)
2323 values[n++] = primary_event_id(sub);
2325 size = n * sizeof(u64);
2327 if (copy_to_user(buf + ret, values, size)) {
2335 mutex_unlock(&ctx->mutex);
2340 static int perf_event_read_one(struct perf_event *event,
2341 u64 read_format, char __user *buf)
2343 u64 enabled, running;
2347 values[n++] = perf_event_read_value(event, &enabled, &running);
2348 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
2349 values[n++] = enabled;
2350 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
2351 values[n++] = running;
2352 if (read_format & PERF_FORMAT_ID)
2353 values[n++] = primary_event_id(event);
2355 if (copy_to_user(buf, values, n * sizeof(u64)))
2358 return n * sizeof(u64);
2362 * Read the performance event - simple non blocking version for now
2365 perf_read_hw(struct perf_event *event, char __user *buf, size_t count)
2367 u64 read_format = event->attr.read_format;
2371 * Return end-of-file for a read on a event that is in
2372 * error state (i.e. because it was pinned but it couldn't be
2373 * scheduled on to the CPU at some point).
2375 if (event->state == PERF_EVENT_STATE_ERROR)
2378 if (count < perf_event_read_size(event))
2381 WARN_ON_ONCE(event->ctx->parent_ctx);
2382 if (read_format & PERF_FORMAT_GROUP)
2383 ret = perf_event_read_group(event, read_format, buf);
2385 ret = perf_event_read_one(event, read_format, buf);
2391 perf_read(struct file *file, char __user *buf, size_t count, loff_t *ppos)
2393 struct perf_event *event = file->private_data;
2395 return perf_read_hw(event, buf, count);
2398 static unsigned int perf_poll(struct file *file, poll_table *wait)
2400 struct perf_event *event = file->private_data;
2401 struct perf_buffer *buffer;
2402 unsigned int events = POLL_HUP;
2405 buffer = rcu_dereference(event->buffer);
2407 events = atomic_xchg(&buffer->poll, 0);
2410 poll_wait(file, &event->waitq, wait);
2415 static void perf_event_reset(struct perf_event *event)
2417 (void)perf_event_read(event);
2418 local64_set(&event->count, 0);
2419 perf_event_update_userpage(event);
2423 * Holding the top-level event's child_mutex means that any
2424 * descendant process that has inherited this event will block
2425 * in sync_child_event if it goes to exit, thus satisfying the
2426 * task existence requirements of perf_event_enable/disable.
2428 static void perf_event_for_each_child(struct perf_event *event,
2429 void (*func)(struct perf_event *))
2431 struct perf_event *child;
2433 WARN_ON_ONCE(event->ctx->parent_ctx);
2434 mutex_lock(&event->child_mutex);
2436 list_for_each_entry(child, &event->child_list, child_list)
2438 mutex_unlock(&event->child_mutex);
2441 static void perf_event_for_each(struct perf_event *event,
2442 void (*func)(struct perf_event *))
2444 struct perf_event_context *ctx = event->ctx;
2445 struct perf_event *sibling;
2447 WARN_ON_ONCE(ctx->parent_ctx);
2448 mutex_lock(&ctx->mutex);
2449 event = event->group_leader;
2451 perf_event_for_each_child(event, func);
2453 list_for_each_entry(sibling, &event->sibling_list, group_entry)
2454 perf_event_for_each_child(event, func);
2455 mutex_unlock(&ctx->mutex);
2458 static int perf_event_period(struct perf_event *event, u64 __user *arg)
2460 struct perf_event_context *ctx = event->ctx;
2465 if (!event->attr.sample_period)
2468 size = copy_from_user(&value, arg, sizeof(value));
2469 if (size != sizeof(value))
2475 raw_spin_lock_irq(&ctx->lock);
2476 if (event->attr.freq) {
2477 if (value > sysctl_perf_event_sample_rate) {
2482 event->attr.sample_freq = value;
2484 event->attr.sample_period = value;
2485 event->hw.sample_period = value;
2488 raw_spin_unlock_irq(&ctx->lock);
2493 static const struct file_operations perf_fops;
2495 static struct perf_event *perf_fget_light(int fd, int *fput_needed)
2499 file = fget_light(fd, fput_needed);
2501 return ERR_PTR(-EBADF);
2503 if (file->f_op != &perf_fops) {
2504 fput_light(file, *fput_needed);
2506 return ERR_PTR(-EBADF);
2509 return file->private_data;
2512 static int perf_event_set_output(struct perf_event *event,
2513 struct perf_event *output_event);
2514 static int perf_event_set_filter(struct perf_event *event, void __user *arg);
2516 static long perf_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
2518 struct perf_event *event = file->private_data;
2519 void (*func)(struct perf_event *);
2523 case PERF_EVENT_IOC_ENABLE:
2524 func = perf_event_enable;
2526 case PERF_EVENT_IOC_DISABLE:
2527 func = perf_event_disable;
2529 case PERF_EVENT_IOC_RESET:
2530 func = perf_event_reset;
2533 case PERF_EVENT_IOC_REFRESH:
2534 return perf_event_refresh(event, arg);
2536 case PERF_EVENT_IOC_PERIOD:
2537 return perf_event_period(event, (u64 __user *)arg);
2539 case PERF_EVENT_IOC_SET_OUTPUT:
2541 struct perf_event *output_event = NULL;
2542 int fput_needed = 0;
2546 output_event = perf_fget_light(arg, &fput_needed);
2547 if (IS_ERR(output_event))
2548 return PTR_ERR(output_event);
2551 ret = perf_event_set_output(event, output_event);
2553 fput_light(output_event->filp, fput_needed);
2558 case PERF_EVENT_IOC_SET_FILTER:
2559 return perf_event_set_filter(event, (void __user *)arg);
2565 if (flags & PERF_IOC_FLAG_GROUP)
2566 perf_event_for_each(event, func);
2568 perf_event_for_each_child(event, func);
2573 int perf_event_task_enable(void)
2575 struct perf_event *event;
2577 mutex_lock(¤t->perf_event_mutex);
2578 list_for_each_entry(event, ¤t->perf_event_list, owner_entry)
2579 perf_event_for_each_child(event, perf_event_enable);
2580 mutex_unlock(¤t->perf_event_mutex);
2585 int perf_event_task_disable(void)
2587 struct perf_event *event;
2589 mutex_lock(¤t->perf_event_mutex);
2590 list_for_each_entry(event, ¤t->perf_event_list, owner_entry)
2591 perf_event_for_each_child(event, perf_event_disable);
2592 mutex_unlock(¤t->perf_event_mutex);
2597 #ifndef PERF_EVENT_INDEX_OFFSET
2598 # define PERF_EVENT_INDEX_OFFSET 0
2601 static int perf_event_index(struct perf_event *event)
2603 if (event->hw.state & PERF_HES_STOPPED)
2606 if (event->state != PERF_EVENT_STATE_ACTIVE)
2609 return event->hw.idx + 1 - PERF_EVENT_INDEX_OFFSET;
2613 * Callers need to ensure there can be no nesting of this function, otherwise
2614 * the seqlock logic goes bad. We can not serialize this because the arch
2615 * code calls this from NMI context.
2617 void perf_event_update_userpage(struct perf_event *event)
2619 struct perf_event_mmap_page *userpg;
2620 struct perf_buffer *buffer;
2623 buffer = rcu_dereference(event->buffer);
2627 userpg = buffer->user_page;
2630 * Disable preemption so as to not let the corresponding user-space
2631 * spin too long if we get preempted.
2636 userpg->index = perf_event_index(event);
2637 userpg->offset = perf_event_count(event);
2638 if (event->state == PERF_EVENT_STATE_ACTIVE)
2639 userpg->offset -= local64_read(&event->hw.prev_count);
2641 userpg->time_enabled = event->total_time_enabled +
2642 atomic64_read(&event->child_total_time_enabled);
2644 userpg->time_running = event->total_time_running +
2645 atomic64_read(&event->child_total_time_running);
2654 static unsigned long perf_data_size(struct perf_buffer *buffer);
2657 perf_buffer_init(struct perf_buffer *buffer, long watermark, int flags)
2659 long max_size = perf_data_size(buffer);
2662 buffer->watermark = min(max_size, watermark);
2664 if (!buffer->watermark)
2665 buffer->watermark = max_size / 2;
2667 if (flags & PERF_BUFFER_WRITABLE)
2668 buffer->writable = 1;
2670 atomic_set(&buffer->refcount, 1);
2673 #ifndef CONFIG_PERF_USE_VMALLOC
2676 * Back perf_mmap() with regular GFP_KERNEL-0 pages.
2679 static struct page *
2680 perf_mmap_to_page(struct perf_buffer *buffer, unsigned long pgoff)
2682 if (pgoff > buffer->nr_pages)
2686 return virt_to_page(buffer->user_page);
2688 return virt_to_page(buffer->data_pages[pgoff - 1]);
2691 static void *perf_mmap_alloc_page(int cpu)
2696 node = (cpu == -1) ? cpu : cpu_to_node(cpu);
2697 page = alloc_pages_node(node, GFP_KERNEL | __GFP_ZERO, 0);
2701 return page_address(page);
2704 static struct perf_buffer *
2705 perf_buffer_alloc(int nr_pages, long watermark, int cpu, int flags)
2707 struct perf_buffer *buffer;
2711 size = sizeof(struct perf_buffer);
2712 size += nr_pages * sizeof(void *);
2714 buffer = kzalloc(size, GFP_KERNEL);
2718 buffer->user_page = perf_mmap_alloc_page(cpu);
2719 if (!buffer->user_page)
2720 goto fail_user_page;
2722 for (i = 0; i < nr_pages; i++) {
2723 buffer->data_pages[i] = perf_mmap_alloc_page(cpu);
2724 if (!buffer->data_pages[i])
2725 goto fail_data_pages;
2728 buffer->nr_pages = nr_pages;
2730 perf_buffer_init(buffer, watermark, flags);
2735 for (i--; i >= 0; i--)
2736 free_page((unsigned long)buffer->data_pages[i]);
2738 free_page((unsigned long)buffer->user_page);
2747 static void perf_mmap_free_page(unsigned long addr)
2749 struct page *page = virt_to_page((void *)addr);
2751 page->mapping = NULL;
2755 static void perf_buffer_free(struct perf_buffer *buffer)
2759 perf_mmap_free_page((unsigned long)buffer->user_page);
2760 for (i = 0; i < buffer->nr_pages; i++)
2761 perf_mmap_free_page((unsigned long)buffer->data_pages[i]);
2765 static inline int page_order(struct perf_buffer *buffer)
2773 * Back perf_mmap() with vmalloc memory.
2775 * Required for architectures that have d-cache aliasing issues.
2778 static inline int page_order(struct perf_buffer *buffer)
2780 return buffer->page_order;
2783 static struct page *
2784 perf_mmap_to_page(struct perf_buffer *buffer, unsigned long pgoff)
2786 if (pgoff > (1UL << page_order(buffer)))
2789 return vmalloc_to_page((void *)buffer->user_page + pgoff * PAGE_SIZE);
2792 static void perf_mmap_unmark_page(void *addr)
2794 struct page *page = vmalloc_to_page(addr);
2796 page->mapping = NULL;
2799 static void perf_buffer_free_work(struct work_struct *work)
2801 struct perf_buffer *buffer;
2805 buffer = container_of(work, struct perf_buffer, work);
2806 nr = 1 << page_order(buffer);
2808 base = buffer->user_page;
2809 for (i = 0; i < nr + 1; i++)
2810 perf_mmap_unmark_page(base + (i * PAGE_SIZE));
2816 static void perf_buffer_free(struct perf_buffer *buffer)
2818 schedule_work(&buffer->work);
2821 static struct perf_buffer *
2822 perf_buffer_alloc(int nr_pages, long watermark, int cpu, int flags)
2824 struct perf_buffer *buffer;
2828 size = sizeof(struct perf_buffer);
2829 size += sizeof(void *);
2831 buffer = kzalloc(size, GFP_KERNEL);
2835 INIT_WORK(&buffer->work, perf_buffer_free_work);
2837 all_buf = vmalloc_user((nr_pages + 1) * PAGE_SIZE);
2841 buffer->user_page = all_buf;
2842 buffer->data_pages[0] = all_buf + PAGE_SIZE;
2843 buffer->page_order = ilog2(nr_pages);
2844 buffer->nr_pages = 1;
2846 perf_buffer_init(buffer, watermark, flags);
2859 static unsigned long perf_data_size(struct perf_buffer *buffer)
2861 return buffer->nr_pages << (PAGE_SHIFT + page_order(buffer));
2864 static int perf_mmap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
2866 struct perf_event *event = vma->vm_file->private_data;
2867 struct perf_buffer *buffer;
2868 int ret = VM_FAULT_SIGBUS;
2870 if (vmf->flags & FAULT_FLAG_MKWRITE) {
2871 if (vmf->pgoff == 0)
2877 buffer = rcu_dereference(event->buffer);
2881 if (vmf->pgoff && (vmf->flags & FAULT_FLAG_WRITE))
2884 vmf->page = perf_mmap_to_page(buffer, vmf->pgoff);
2888 get_page(vmf->page);
2889 vmf->page->mapping = vma->vm_file->f_mapping;
2890 vmf->page->index = vmf->pgoff;
2899 static void perf_buffer_free_rcu(struct rcu_head *rcu_head)
2901 struct perf_buffer *buffer;
2903 buffer = container_of(rcu_head, struct perf_buffer, rcu_head);
2904 perf_buffer_free(buffer);
2907 static struct perf_buffer *perf_buffer_get(struct perf_event *event)
2909 struct perf_buffer *buffer;
2912 buffer = rcu_dereference(event->buffer);
2914 if (!atomic_inc_not_zero(&buffer->refcount))
2922 static void perf_buffer_put(struct perf_buffer *buffer)
2924 if (!atomic_dec_and_test(&buffer->refcount))
2927 call_rcu(&buffer->rcu_head, perf_buffer_free_rcu);
2930 static void perf_mmap_open(struct vm_area_struct *vma)
2932 struct perf_event *event = vma->vm_file->private_data;
2934 atomic_inc(&event->mmap_count);
2937 static void perf_mmap_close(struct vm_area_struct *vma)
2939 struct perf_event *event = vma->vm_file->private_data;
2941 if (atomic_dec_and_mutex_lock(&event->mmap_count, &event->mmap_mutex)) {
2942 unsigned long size = perf_data_size(event->buffer);
2943 struct user_struct *user = event->mmap_user;
2944 struct perf_buffer *buffer = event->buffer;
2946 atomic_long_sub((size >> PAGE_SHIFT) + 1, &user->locked_vm);
2947 vma->vm_mm->locked_vm -= event->mmap_locked;
2948 rcu_assign_pointer(event->buffer, NULL);
2949 mutex_unlock(&event->mmap_mutex);
2951 perf_buffer_put(buffer);
2956 static const struct vm_operations_struct perf_mmap_vmops = {
2957 .open = perf_mmap_open,
2958 .close = perf_mmap_close,
2959 .fault = perf_mmap_fault,
2960 .page_mkwrite = perf_mmap_fault,
2963 static int perf_mmap(struct file *file, struct vm_area_struct *vma)
2965 struct perf_event *event = file->private_data;
2966 unsigned long user_locked, user_lock_limit;
2967 struct user_struct *user = current_user();
2968 unsigned long locked, lock_limit;
2969 struct perf_buffer *buffer;
2970 unsigned long vma_size;
2971 unsigned long nr_pages;
2972 long user_extra, extra;
2973 int ret = 0, flags = 0;
2976 * Don't allow mmap() of inherited per-task counters. This would
2977 * create a performance issue due to all children writing to the
2980 if (event->cpu == -1 && event->attr.inherit)
2983 if (!(vma->vm_flags & VM_SHARED))
2986 vma_size = vma->vm_end - vma->vm_start;
2987 nr_pages = (vma_size / PAGE_SIZE) - 1;
2990 * If we have buffer pages ensure they're a power-of-two number, so we
2991 * can do bitmasks instead of modulo.
2993 if (nr_pages != 0 && !is_power_of_2(nr_pages))
2996 if (vma_size != PAGE_SIZE * (1 + nr_pages))
2999 if (vma->vm_pgoff != 0)
3002 WARN_ON_ONCE(event->ctx->parent_ctx);
3003 mutex_lock(&event->mmap_mutex);
3004 if (event->buffer) {
3005 if (event->buffer->nr_pages == nr_pages)
3006 atomic_inc(&event->buffer->refcount);
3012 user_extra = nr_pages + 1;
3013 user_lock_limit = sysctl_perf_event_mlock >> (PAGE_SHIFT - 10);
3016 * Increase the limit linearly with more CPUs:
3018 user_lock_limit *= num_online_cpus();
3020 user_locked = atomic_long_read(&user->locked_vm) + user_extra;
3023 if (user_locked > user_lock_limit)
3024 extra = user_locked - user_lock_limit;
3026 lock_limit = rlimit(RLIMIT_MEMLOCK);
3027 lock_limit >>= PAGE_SHIFT;
3028 locked = vma->vm_mm->locked_vm + extra;
3030 if ((locked > lock_limit) && perf_paranoid_tracepoint_raw() &&
3031 !capable(CAP_IPC_LOCK)) {
3036 WARN_ON(event->buffer);
3038 if (vma->vm_flags & VM_WRITE)
3039 flags |= PERF_BUFFER_WRITABLE;
3041 buffer = perf_buffer_alloc(nr_pages, event->attr.wakeup_watermark,
3047 rcu_assign_pointer(event->buffer, buffer);
3049 atomic_long_add(user_extra, &user->locked_vm);
3050 event->mmap_locked = extra;
3051 event->mmap_user = get_current_user();
3052 vma->vm_mm->locked_vm += event->mmap_locked;
3056 atomic_inc(&event->mmap_count);
3057 mutex_unlock(&event->mmap_mutex);
3059 vma->vm_flags |= VM_RESERVED;
3060 vma->vm_ops = &perf_mmap_vmops;
3065 static int perf_fasync(int fd, struct file *filp, int on)
3067 struct inode *inode = filp->f_path.dentry->d_inode;
3068 struct perf_event *event = filp->private_data;
3071 mutex_lock(&inode->i_mutex);
3072 retval = fasync_helper(fd, filp, on, &event->fasync);
3073 mutex_unlock(&inode->i_mutex);
3081 static const struct file_operations perf_fops = {
3082 .llseek = no_llseek,
3083 .release = perf_release,
3086 .unlocked_ioctl = perf_ioctl,
3087 .compat_ioctl = perf_ioctl,
3089 .fasync = perf_fasync,
3095 * If there's data, ensure we set the poll() state and publish everything
3096 * to user-space before waking everybody up.
3099 void perf_event_wakeup(struct perf_event *event)
3101 wake_up_all(&event->waitq);
3103 if (event->pending_kill) {
3104 kill_fasync(&event->fasync, SIGIO, event->pending_kill);
3105 event->pending_kill = 0;
3112 * Handle the case where we need to wakeup up from NMI (or rq->lock) context.
3114 * The NMI bit means we cannot possibly take locks. Therefore, maintain a
3115 * single linked list and use cmpxchg() to add entries lockless.
3118 static void perf_pending_event(struct perf_pending_entry *entry)
3120 struct perf_event *event = container_of(entry,
3121 struct perf_event, pending);
3123 if (event->pending_disable) {
3124 event->pending_disable = 0;
3125 __perf_event_disable(event);
3128 if (event->pending_wakeup) {
3129 event->pending_wakeup = 0;
3130 perf_event_wakeup(event);
3134 #define PENDING_TAIL ((struct perf_pending_entry *)-1UL)
3136 static DEFINE_PER_CPU(struct perf_pending_entry *, perf_pending_head) = {
3140 static void perf_pending_queue(struct perf_pending_entry *entry,
3141 void (*func)(struct perf_pending_entry *))
3143 struct perf_pending_entry **head;
3145 if (cmpxchg(&entry->next, NULL, PENDING_TAIL) != NULL)
3150 head = &get_cpu_var(perf_pending_head);
3153 entry->next = *head;
3154 } while (cmpxchg(head, entry->next, entry) != entry->next);
3156 set_perf_event_pending();
3158 put_cpu_var(perf_pending_head);
3161 static int __perf_pending_run(void)
3163 struct perf_pending_entry *list;
3166 list = xchg(&__get_cpu_var(perf_pending_head), PENDING_TAIL);
3167 while (list != PENDING_TAIL) {
3168 void (*func)(struct perf_pending_entry *);
3169 struct perf_pending_entry *entry = list;
3176 * Ensure we observe the unqueue before we issue the wakeup,
3177 * so that we won't be waiting forever.
3178 * -- see perf_not_pending().
3189 static inline int perf_not_pending(struct perf_event *event)
3192 * If we flush on whatever cpu we run, there is a chance we don't
3196 __perf_pending_run();
3200 * Ensure we see the proper queue state before going to sleep
3201 * so that we do not miss the wakeup. -- see perf_pending_handle()
3204 return event->pending.next == NULL;
3207 static void perf_pending_sync(struct perf_event *event)
3209 wait_event(event->waitq, perf_not_pending(event));
3212 void perf_event_do_pending(void)
3214 __perf_pending_run();
3218 * We assume there is only KVM supporting the callbacks.
3219 * Later on, we might change it to a list if there is
3220 * another virtualization implementation supporting the callbacks.
3222 struct perf_guest_info_callbacks *perf_guest_cbs;
3224 int perf_register_guest_info_callbacks(struct perf_guest_info_callbacks *cbs)
3226 perf_guest_cbs = cbs;
3229 EXPORT_SYMBOL_GPL(perf_register_guest_info_callbacks);
3231 int perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks *cbs)
3233 perf_guest_cbs = NULL;
3236 EXPORT_SYMBOL_GPL(perf_unregister_guest_info_callbacks);
3241 static bool perf_output_space(struct perf_buffer *buffer, unsigned long tail,
3242 unsigned long offset, unsigned long head)
3246 if (!buffer->writable)
3249 mask = perf_data_size(buffer) - 1;
3251 offset = (offset - tail) & mask;
3252 head = (head - tail) & mask;
3254 if ((int)(head - offset) < 0)
3260 static void perf_output_wakeup(struct perf_output_handle *handle)
3262 atomic_set(&handle->buffer->poll, POLL_IN);
3265 handle->event->pending_wakeup = 1;
3266 perf_pending_queue(&handle->event->pending,
3267 perf_pending_event);
3269 perf_event_wakeup(handle->event);
3273 * We need to ensure a later event_id doesn't publish a head when a former
3274 * event isn't done writing. However since we need to deal with NMIs we
3275 * cannot fully serialize things.
3277 * We only publish the head (and generate a wakeup) when the outer-most
3280 static void perf_output_get_handle(struct perf_output_handle *handle)
3282 struct perf_buffer *buffer = handle->buffer;
3285 local_inc(&buffer->nest);
3286 handle->wakeup = local_read(&buffer->wakeup);
3289 static void perf_output_put_handle(struct perf_output_handle *handle)
3291 struct perf_buffer *buffer = handle->buffer;
3295 head = local_read(&buffer->head);
3298 * IRQ/NMI can happen here, which means we can miss a head update.
3301 if (!local_dec_and_test(&buffer->nest))
3305 * Publish the known good head. Rely on the full barrier implied
3306 * by atomic_dec_and_test() order the buffer->head read and this
3309 buffer->user_page->data_head = head;
3312 * Now check if we missed an update, rely on the (compiler)
3313 * barrier in atomic_dec_and_test() to re-read buffer->head.
3315 if (unlikely(head != local_read(&buffer->head))) {
3316 local_inc(&buffer->nest);
3320 if (handle->wakeup != local_read(&buffer->wakeup))
3321 perf_output_wakeup(handle);
3327 __always_inline void perf_output_copy(struct perf_output_handle *handle,
3328 const void *buf, unsigned int len)
3331 unsigned long size = min_t(unsigned long, handle->size, len);
3333 memcpy(handle->addr, buf, size);
3336 handle->addr += size;
3338 handle->size -= size;
3339 if (!handle->size) {
3340 struct perf_buffer *buffer = handle->buffer;
3343 handle->page &= buffer->nr_pages - 1;
3344 handle->addr = buffer->data_pages[handle->page];
3345 handle->size = PAGE_SIZE << page_order(buffer);
3350 int perf_output_begin(struct perf_output_handle *handle,
3351 struct perf_event *event, unsigned int size,
3352 int nmi, int sample)
3354 struct perf_buffer *buffer;
3355 unsigned long tail, offset, head;
3358 struct perf_event_header header;
3365 * For inherited events we send all the output towards the parent.
3368 event = event->parent;
3370 buffer = rcu_dereference(event->buffer);
3374 handle->buffer = buffer;
3375 handle->event = event;
3377 handle->sample = sample;
3379 if (!buffer->nr_pages)
3382 have_lost = local_read(&buffer->lost);
3384 size += sizeof(lost_event);
3386 perf_output_get_handle(handle);
3390 * Userspace could choose to issue a mb() before updating the
3391 * tail pointer. So that all reads will be completed before the
3394 tail = ACCESS_ONCE(buffer->user_page->data_tail);
3396 offset = head = local_read(&buffer->head);
3398 if (unlikely(!perf_output_space(buffer, tail, offset, head)))
3400 } while (local_cmpxchg(&buffer->head, offset, head) != offset);
3402 if (head - local_read(&buffer->wakeup) > buffer->watermark)
3403 local_add(buffer->watermark, &buffer->wakeup);
3405 handle->page = offset >> (PAGE_SHIFT + page_order(buffer));
3406 handle->page &= buffer->nr_pages - 1;
3407 handle->size = offset & ((PAGE_SIZE << page_order(buffer)) - 1);
3408 handle->addr = buffer->data_pages[handle->page];
3409 handle->addr += handle->size;
3410 handle->size = (PAGE_SIZE << page_order(buffer)) - handle->size;
3413 lost_event.header.type = PERF_RECORD_LOST;
3414 lost_event.header.misc = 0;
3415 lost_event.header.size = sizeof(lost_event);
3416 lost_event.id = event->id;
3417 lost_event.lost = local_xchg(&buffer->lost, 0);
3419 perf_output_put(handle, lost_event);
3425 local_inc(&buffer->lost);
3426 perf_output_put_handle(handle);
3433 void perf_output_end(struct perf_output_handle *handle)
3435 struct perf_event *event = handle->event;
3436 struct perf_buffer *buffer = handle->buffer;
3438 int wakeup_events = event->attr.wakeup_events;
3440 if (handle->sample && wakeup_events) {
3441 int events = local_inc_return(&buffer->events);
3442 if (events >= wakeup_events) {
3443 local_sub(wakeup_events, &buffer->events);
3444 local_inc(&buffer->wakeup);
3448 perf_output_put_handle(handle);
3452 static u32 perf_event_pid(struct perf_event *event, struct task_struct *p)
3455 * only top level events have the pid namespace they were created in
3458 event = event->parent;
3460 return task_tgid_nr_ns(p, event->ns);
3463 static u32 perf_event_tid(struct perf_event *event, struct task_struct *p)
3466 * only top level events have the pid namespace they were created in
3469 event = event->parent;
3471 return task_pid_nr_ns(p, event->ns);
3474 static void perf_output_read_one(struct perf_output_handle *handle,
3475 struct perf_event *event)
3477 u64 read_format = event->attr.read_format;
3481 values[n++] = perf_event_count(event);
3482 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
3483 values[n++] = event->total_time_enabled +
3484 atomic64_read(&event->child_total_time_enabled);
3486 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
3487 values[n++] = event->total_time_running +
3488 atomic64_read(&event->child_total_time_running);
3490 if (read_format & PERF_FORMAT_ID)
3491 values[n++] = primary_event_id(event);
3493 perf_output_copy(handle, values, n * sizeof(u64));
3497 * XXX PERF_FORMAT_GROUP vs inherited events seems difficult.
3499 static void perf_output_read_group(struct perf_output_handle *handle,
3500 struct perf_event *event)
3502 struct perf_event *leader = event->group_leader, *sub;
3503 u64 read_format = event->attr.read_format;
3507 values[n++] = 1 + leader->nr_siblings;
3509 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
3510 values[n++] = leader->total_time_enabled;
3512 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
3513 values[n++] = leader->total_time_running;
3515 if (leader != event)
3516 leader->pmu->read(leader);
3518 values[n++] = perf_event_count(leader);
3519 if (read_format & PERF_FORMAT_ID)
3520 values[n++] = primary_event_id(leader);
3522 perf_output_copy(handle, values, n * sizeof(u64));
3524 list_for_each_entry(sub, &leader->sibling_list, group_entry) {
3528 sub->pmu->read(sub);
3530 values[n++] = perf_event_count(sub);
3531 if (read_format & PERF_FORMAT_ID)
3532 values[n++] = primary_event_id(sub);
3534 perf_output_copy(handle, values, n * sizeof(u64));
3538 static void perf_output_read(struct perf_output_handle *handle,
3539 struct perf_event *event)
3541 if (event->attr.read_format & PERF_FORMAT_GROUP)
3542 perf_output_read_group(handle, event);
3544 perf_output_read_one(handle, event);
3547 void perf_output_sample(struct perf_output_handle *handle,
3548 struct perf_event_header *header,
3549 struct perf_sample_data *data,
3550 struct perf_event *event)
3552 u64 sample_type = data->type;
3554 perf_output_put(handle, *header);
3556 if (sample_type & PERF_SAMPLE_IP)
3557 perf_output_put(handle, data->ip);
3559 if (sample_type & PERF_SAMPLE_TID)
3560 perf_output_put(handle, data->tid_entry);
3562 if (sample_type & PERF_SAMPLE_TIME)
3563 perf_output_put(handle, data->time);
3565 if (sample_type & PERF_SAMPLE_ADDR)
3566 perf_output_put(handle, data->addr);
3568 if (sample_type & PERF_SAMPLE_ID)
3569 perf_output_put(handle, data->id);
3571 if (sample_type & PERF_SAMPLE_STREAM_ID)
3572 perf_output_put(handle, data->stream_id);
3574 if (sample_type & PERF_SAMPLE_CPU)
3575 perf_output_put(handle, data->cpu_entry);
3577 if (sample_type & PERF_SAMPLE_PERIOD)
3578 perf_output_put(handle, data->period);
3580 if (sample_type & PERF_SAMPLE_READ)
3581 perf_output_read(handle, event);
3583 if (sample_type & PERF_SAMPLE_CALLCHAIN) {
3584 if (data->callchain) {
3587 if (data->callchain)
3588 size += data->callchain->nr;
3590 size *= sizeof(u64);
3592 perf_output_copy(handle, data->callchain, size);
3595 perf_output_put(handle, nr);
3599 if (sample_type & PERF_SAMPLE_RAW) {
3601 perf_output_put(handle, data->raw->size);
3602 perf_output_copy(handle, data->raw->data,
3609 .size = sizeof(u32),
3612 perf_output_put(handle, raw);
3617 void perf_prepare_sample(struct perf_event_header *header,
3618 struct perf_sample_data *data,
3619 struct perf_event *event,
3620 struct pt_regs *regs)
3622 u64 sample_type = event->attr.sample_type;
3624 data->type = sample_type;
3626 header->type = PERF_RECORD_SAMPLE;
3627 header->size = sizeof(*header);
3630 header->misc |= perf_misc_flags(regs);
3632 if (sample_type & PERF_SAMPLE_IP) {
3633 data->ip = perf_instruction_pointer(regs);
3635 header->size += sizeof(data->ip);
3638 if (sample_type & PERF_SAMPLE_TID) {
3639 /* namespace issues */
3640 data->tid_entry.pid = perf_event_pid(event, current);
3641 data->tid_entry.tid = perf_event_tid(event, current);
3643 header->size += sizeof(data->tid_entry);
3646 if (sample_type & PERF_SAMPLE_TIME) {
3647 data->time = perf_clock();
3649 header->size += sizeof(data->time);
3652 if (sample_type & PERF_SAMPLE_ADDR)
3653 header->size += sizeof(data->addr);
3655 if (sample_type & PERF_SAMPLE_ID) {
3656 data->id = primary_event_id(event);
3658 header->size += sizeof(data->id);
3661 if (sample_type & PERF_SAMPLE_STREAM_ID) {
3662 data->stream_id = event->id;
3664 header->size += sizeof(data->stream_id);
3667 if (sample_type & PERF_SAMPLE_CPU) {
3668 data->cpu_entry.cpu = raw_smp_processor_id();
3669 data->cpu_entry.reserved = 0;
3671 header->size += sizeof(data->cpu_entry);
3674 if (sample_type & PERF_SAMPLE_PERIOD)
3675 header->size += sizeof(data->period);
3677 if (sample_type & PERF_SAMPLE_READ)
3678 header->size += perf_event_read_size(event);
3680 if (sample_type & PERF_SAMPLE_CALLCHAIN) {
3683 data->callchain = perf_callchain(regs);
3685 if (data->callchain)
3686 size += data->callchain->nr;
3688 header->size += size * sizeof(u64);
3691 if (sample_type & PERF_SAMPLE_RAW) {
3692 int size = sizeof(u32);
3695 size += data->raw->size;
3697 size += sizeof(u32);
3699 WARN_ON_ONCE(size & (sizeof(u64)-1));
3700 header->size += size;
3704 static void perf_event_output(struct perf_event *event, int nmi,
3705 struct perf_sample_data *data,
3706 struct pt_regs *regs)
3708 struct perf_output_handle handle;
3709 struct perf_event_header header;
3711 /* protect the callchain buffers */
3714 perf_prepare_sample(&header, data, event, regs);
3716 if (perf_output_begin(&handle, event, header.size, nmi, 1))
3719 perf_output_sample(&handle, &header, data, event);
3721 perf_output_end(&handle);
3731 struct perf_read_event {
3732 struct perf_event_header header;
3739 perf_event_read_event(struct perf_event *event,
3740 struct task_struct *task)
3742 struct perf_output_handle handle;
3743 struct perf_read_event read_event = {
3745 .type = PERF_RECORD_READ,
3747 .size = sizeof(read_event) + perf_event_read_size(event),
3749 .pid = perf_event_pid(event, task),
3750 .tid = perf_event_tid(event, task),
3754 ret = perf_output_begin(&handle, event, read_event.header.size, 0, 0);
3758 perf_output_put(&handle, read_event);
3759 perf_output_read(&handle, event);
3761 perf_output_end(&handle);
3765 * task tracking -- fork/exit
3767 * enabled by: attr.comm | attr.mmap | attr.mmap_data | attr.task
3770 struct perf_task_event {
3771 struct task_struct *task;
3772 struct perf_event_context *task_ctx;
3775 struct perf_event_header header;
3785 static void perf_event_task_output(struct perf_event *event,
3786 struct perf_task_event *task_event)
3788 struct perf_output_handle handle;
3789 struct task_struct *task = task_event->task;
3792 size = task_event->event_id.header.size;
3793 ret = perf_output_begin(&handle, event, size, 0, 0);
3798 task_event->event_id.pid = perf_event_pid(event, task);
3799 task_event->event_id.ppid = perf_event_pid(event, current);
3801 task_event->event_id.tid = perf_event_tid(event, task);
3802 task_event->event_id.ptid = perf_event_tid(event, current);
3804 perf_output_put(&handle, task_event->event_id);
3806 perf_output_end(&handle);
3809 static int perf_event_task_match(struct perf_event *event)
3811 if (event->state < PERF_EVENT_STATE_INACTIVE)
3814 if (event->cpu != -1 && event->cpu != smp_processor_id())
3817 if (event->attr.comm || event->attr.mmap ||
3818 event->attr.mmap_data || event->attr.task)
3824 static void perf_event_task_ctx(struct perf_event_context *ctx,
3825 struct perf_task_event *task_event)
3827 struct perf_event *event;
3829 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
3830 if (perf_event_task_match(event))
3831 perf_event_task_output(event, task_event);
3835 static void perf_event_task_event(struct perf_task_event *task_event)
3837 struct perf_cpu_context *cpuctx;
3838 struct perf_event_context *ctx;
3843 list_for_each_entry_rcu(pmu, &pmus, entry) {
3844 cpuctx = get_cpu_ptr(pmu->pmu_cpu_context);
3845 perf_event_task_ctx(&cpuctx->ctx, task_event);
3847 ctx = task_event->task_ctx;
3849 ctxn = pmu->task_ctx_nr;
3852 ctx = rcu_dereference(current->perf_event_ctxp[ctxn]);
3855 perf_event_task_ctx(ctx, task_event);
3857 put_cpu_ptr(pmu->pmu_cpu_context);
3862 static void perf_event_task(struct task_struct *task,
3863 struct perf_event_context *task_ctx,
3866 struct perf_task_event task_event;
3868 if (!atomic_read(&nr_comm_events) &&
3869 !atomic_read(&nr_mmap_events) &&
3870 !atomic_read(&nr_task_events))
3873 task_event = (struct perf_task_event){
3875 .task_ctx = task_ctx,
3878 .type = new ? PERF_RECORD_FORK : PERF_RECORD_EXIT,
3880 .size = sizeof(task_event.event_id),
3886 .time = perf_clock(),
3890 perf_event_task_event(&task_event);
3893 void perf_event_fork(struct task_struct *task)
3895 perf_event_task(task, NULL, 1);
3902 struct perf_comm_event {
3903 struct task_struct *task;
3908 struct perf_event_header header;
3915 static void perf_event_comm_output(struct perf_event *event,
3916 struct perf_comm_event *comm_event)
3918 struct perf_output_handle handle;
3919 int size = comm_event->event_id.header.size;
3920 int ret = perf_output_begin(&handle, event, size, 0, 0);
3925 comm_event->event_id.pid = perf_event_pid(event, comm_event->task);
3926 comm_event->event_id.tid = perf_event_tid(event, comm_event->task);
3928 perf_output_put(&handle, comm_event->event_id);
3929 perf_output_copy(&handle, comm_event->comm,
3930 comm_event->comm_size);
3931 perf_output_end(&handle);
3934 static int perf_event_comm_match(struct perf_event *event)
3936 if (event->state < PERF_EVENT_STATE_INACTIVE)
3939 if (event->cpu != -1 && event->cpu != smp_processor_id())
3942 if (event->attr.comm)
3948 static void perf_event_comm_ctx(struct perf_event_context *ctx,
3949 struct perf_comm_event *comm_event)
3951 struct perf_event *event;
3953 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
3954 if (perf_event_comm_match(event))
3955 perf_event_comm_output(event, comm_event);
3959 static void perf_event_comm_event(struct perf_comm_event *comm_event)
3961 struct perf_cpu_context *cpuctx;
3962 struct perf_event_context *ctx;
3963 char comm[TASK_COMM_LEN];
3968 memset(comm, 0, sizeof(comm));
3969 strlcpy(comm, comm_event->task->comm, sizeof(comm));
3970 size = ALIGN(strlen(comm)+1, sizeof(u64));
3972 comm_event->comm = comm;
3973 comm_event->comm_size = size;
3975 comm_event->event_id.header.size = sizeof(comm_event->event_id) + size;
3978 list_for_each_entry_rcu(pmu, &pmus, entry) {
3979 cpuctx = get_cpu_ptr(pmu->pmu_cpu_context);
3980 perf_event_comm_ctx(&cpuctx->ctx, comm_event);
3982 ctxn = pmu->task_ctx_nr;
3986 ctx = rcu_dereference(current->perf_event_ctxp[ctxn]);
3988 perf_event_comm_ctx(ctx, comm_event);
3990 put_cpu_ptr(pmu->pmu_cpu_context);
3995 void perf_event_comm(struct task_struct *task)
3997 struct perf_comm_event comm_event;
3998 struct perf_event_context *ctx;
4001 for_each_task_context_nr(ctxn) {
4002 ctx = task->perf_event_ctxp[ctxn];
4006 perf_event_enable_on_exec(ctx);
4009 if (!atomic_read(&nr_comm_events))
4012 comm_event = (struct perf_comm_event){
4018 .type = PERF_RECORD_COMM,
4027 perf_event_comm_event(&comm_event);
4034 struct perf_mmap_event {
4035 struct vm_area_struct *vma;
4037 const char *file_name;
4041 struct perf_event_header header;
4051 static void perf_event_mmap_output(struct perf_event *event,
4052 struct perf_mmap_event *mmap_event)
4054 struct perf_output_handle handle;
4055 int size = mmap_event->event_id.header.size;
4056 int ret = perf_output_begin(&handle, event, size, 0, 0);
4061 mmap_event->event_id.pid = perf_event_pid(event, current);
4062 mmap_event->event_id.tid = perf_event_tid(event, current);
4064 perf_output_put(&handle, mmap_event->event_id);
4065 perf_output_copy(&handle, mmap_event->file_name,
4066 mmap_event->file_size);
4067 perf_output_end(&handle);
4070 static int perf_event_mmap_match(struct perf_event *event,
4071 struct perf_mmap_event *mmap_event,
4074 if (event->state < PERF_EVENT_STATE_INACTIVE)
4077 if (event->cpu != -1 && event->cpu != smp_processor_id())
4080 if ((!executable && event->attr.mmap_data) ||
4081 (executable && event->attr.mmap))
4087 static void perf_event_mmap_ctx(struct perf_event_context *ctx,
4088 struct perf_mmap_event *mmap_event,
4091 struct perf_event *event;
4093 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
4094 if (perf_event_mmap_match(event, mmap_event, executable))
4095 perf_event_mmap_output(event, mmap_event);
4099 static void perf_event_mmap_event(struct perf_mmap_event *mmap_event)
4101 struct perf_cpu_context *cpuctx;
4102 struct perf_event_context *ctx;
4103 struct vm_area_struct *vma = mmap_event->vma;
4104 struct file *file = vma->vm_file;
4112 memset(tmp, 0, sizeof(tmp));
4116 * d_path works from the end of the buffer backwards, so we
4117 * need to add enough zero bytes after the string to handle
4118 * the 64bit alignment we do later.
4120 buf = kzalloc(PATH_MAX + sizeof(u64), GFP_KERNEL);
4122 name = strncpy(tmp, "//enomem", sizeof(tmp));
4125 name = d_path(&file->f_path, buf, PATH_MAX);
4127 name = strncpy(tmp, "//toolong", sizeof(tmp));
4131 if (arch_vma_name(mmap_event->vma)) {
4132 name = strncpy(tmp, arch_vma_name(mmap_event->vma),
4138 name = strncpy(tmp, "[vdso]", sizeof(tmp));
4140 } else if (vma->vm_start <= vma->vm_mm->start_brk &&
4141 vma->vm_end >= vma->vm_mm->brk) {
4142 name = strncpy(tmp, "[heap]", sizeof(tmp));
4144 } else if (vma->vm_start <= vma->vm_mm->start_stack &&
4145 vma->vm_end >= vma->vm_mm->start_stack) {
4146 name = strncpy(tmp, "[stack]", sizeof(tmp));
4150 name = strncpy(tmp, "//anon", sizeof(tmp));
4155 size = ALIGN(strlen(name)+1, sizeof(u64));
4157 mmap_event->file_name = name;
4158 mmap_event->file_size = size;
4160 mmap_event->event_id.header.size = sizeof(mmap_event->event_id) + size;
4163 list_for_each_entry_rcu(pmu, &pmus, entry) {
4164 cpuctx = get_cpu_ptr(pmu->pmu_cpu_context);
4165 perf_event_mmap_ctx(&cpuctx->ctx, mmap_event,
4166 vma->vm_flags & VM_EXEC);
4168 ctxn = pmu->task_ctx_nr;
4172 ctx = rcu_dereference(current->perf_event_ctxp[ctxn]);
4174 perf_event_mmap_ctx(ctx, mmap_event,
4175 vma->vm_flags & VM_EXEC);
4178 put_cpu_ptr(pmu->pmu_cpu_context);
4185 void perf_event_mmap(struct vm_area_struct *vma)
4187 struct perf_mmap_event mmap_event;
4189 if (!atomic_read(&nr_mmap_events))
4192 mmap_event = (struct perf_mmap_event){
4198 .type = PERF_RECORD_MMAP,
4199 .misc = PERF_RECORD_MISC_USER,
4204 .start = vma->vm_start,
4205 .len = vma->vm_end - vma->vm_start,
4206 .pgoff = (u64)vma->vm_pgoff << PAGE_SHIFT,
4210 perf_event_mmap_event(&mmap_event);
4214 * IRQ throttle logging
4217 static void perf_log_throttle(struct perf_event *event, int enable)
4219 struct perf_output_handle handle;
4223 struct perf_event_header header;
4227 } throttle_event = {
4229 .type = PERF_RECORD_THROTTLE,
4231 .size = sizeof(throttle_event),
4233 .time = perf_clock(),
4234 .id = primary_event_id(event),
4235 .stream_id = event->id,
4239 throttle_event.header.type = PERF_RECORD_UNTHROTTLE;
4241 ret = perf_output_begin(&handle, event, sizeof(throttle_event), 1, 0);
4245 perf_output_put(&handle, throttle_event);
4246 perf_output_end(&handle);
4250 * Generic event overflow handling, sampling.
4253 static int __perf_event_overflow(struct perf_event *event, int nmi,
4254 int throttle, struct perf_sample_data *data,
4255 struct pt_regs *regs)
4257 int events = atomic_read(&event->event_limit);
4258 struct hw_perf_event *hwc = &event->hw;
4264 if (hwc->interrupts != MAX_INTERRUPTS) {
4266 if (HZ * hwc->interrupts >
4267 (u64)sysctl_perf_event_sample_rate) {
4268 hwc->interrupts = MAX_INTERRUPTS;
4269 perf_log_throttle(event, 0);
4274 * Keep re-disabling events even though on the previous
4275 * pass we disabled it - just in case we raced with a
4276 * sched-in and the event got enabled again:
4282 if (event->attr.freq) {
4283 u64 now = perf_clock();
4284 s64 delta = now - hwc->freq_time_stamp;
4286 hwc->freq_time_stamp = now;
4288 if (delta > 0 && delta < 2*TICK_NSEC)
4289 perf_adjust_period(event, delta, hwc->last_period);
4293 * XXX event_limit might not quite work as expected on inherited
4297 event->pending_kill = POLL_IN;
4298 if (events && atomic_dec_and_test(&event->event_limit)) {
4300 event->pending_kill = POLL_HUP;
4302 event->pending_disable = 1;
4303 perf_pending_queue(&event->pending,
4304 perf_pending_event);
4306 perf_event_disable(event);
4309 if (event->overflow_handler)
4310 event->overflow_handler(event, nmi, data, regs);
4312 perf_event_output(event, nmi, data, regs);
4317 int perf_event_overflow(struct perf_event *event, int nmi,
4318 struct perf_sample_data *data,
4319 struct pt_regs *regs)
4321 return __perf_event_overflow(event, nmi, 1, data, regs);
4325 * Generic software event infrastructure
4328 struct swevent_htable {
4329 struct swevent_hlist *swevent_hlist;
4330 struct mutex hlist_mutex;
4333 /* Recursion avoidance in each contexts */
4334 int recursion[PERF_NR_CONTEXTS];
4337 static DEFINE_PER_CPU(struct swevent_htable, swevent_htable);
4340 * We directly increment event->count and keep a second value in
4341 * event->hw.period_left to count intervals. This period event
4342 * is kept in the range [-sample_period, 0] so that we can use the
4346 static u64 perf_swevent_set_period(struct perf_event *event)
4348 struct hw_perf_event *hwc = &event->hw;
4349 u64 period = hwc->last_period;
4353 hwc->last_period = hwc->sample_period;
4356 old = val = local64_read(&hwc->period_left);
4360 nr = div64_u64(period + val, period);
4361 offset = nr * period;
4363 if (local64_cmpxchg(&hwc->period_left, old, val) != old)
4369 static void perf_swevent_overflow(struct perf_event *event, u64 overflow,
4370 int nmi, struct perf_sample_data *data,
4371 struct pt_regs *regs)
4373 struct hw_perf_event *hwc = &event->hw;
4376 data->period = event->hw.last_period;
4378 overflow = perf_swevent_set_period(event);
4380 if (hwc->interrupts == MAX_INTERRUPTS)
4383 for (; overflow; overflow--) {
4384 if (__perf_event_overflow(event, nmi, throttle,
4387 * We inhibit the overflow from happening when
4388 * hwc->interrupts == MAX_INTERRUPTS.
4396 static void perf_swevent_event(struct perf_event *event, u64 nr,
4397 int nmi, struct perf_sample_data *data,
4398 struct pt_regs *regs)
4400 struct hw_perf_event *hwc = &event->hw;
4402 local64_add(nr, &event->count);
4407 if (!hwc->sample_period)
4410 if (nr == 1 && hwc->sample_period == 1 && !event->attr.freq)
4411 return perf_swevent_overflow(event, 1, nmi, data, regs);
4413 if (local64_add_negative(nr, &hwc->period_left))
4416 perf_swevent_overflow(event, 0, nmi, data, regs);
4419 static int perf_exclude_event(struct perf_event *event,
4420 struct pt_regs *regs)
4422 if (event->hw.state & PERF_HES_STOPPED)
4426 if (event->attr.exclude_user && user_mode(regs))
4429 if (event->attr.exclude_kernel && !user_mode(regs))
4436 static int perf_swevent_match(struct perf_event *event,
4437 enum perf_type_id type,
4439 struct perf_sample_data *data,
4440 struct pt_regs *regs)
4442 if (event->attr.type != type)
4445 if (event->attr.config != event_id)
4448 if (perf_exclude_event(event, regs))
4454 static inline u64 swevent_hash(u64 type, u32 event_id)
4456 u64 val = event_id | (type << 32);
4458 return hash_64(val, SWEVENT_HLIST_BITS);
4461 static inline struct hlist_head *
4462 __find_swevent_head(struct swevent_hlist *hlist, u64 type, u32 event_id)
4464 u64 hash = swevent_hash(type, event_id);
4466 return &hlist->heads[hash];
4469 /* For the read side: events when they trigger */
4470 static inline struct hlist_head *
4471 find_swevent_head_rcu(struct swevent_htable *swhash, u64 type, u32 event_id)
4473 struct swevent_hlist *hlist;
4475 hlist = rcu_dereference(swhash->swevent_hlist);
4479 return __find_swevent_head(hlist, type, event_id);
4482 /* For the event head insertion and removal in the hlist */
4483 static inline struct hlist_head *
4484 find_swevent_head(struct swevent_htable *swhash, struct perf_event *event)
4486 struct swevent_hlist *hlist;
4487 u32 event_id = event->attr.config;
4488 u64 type = event->attr.type;
4491 * Event scheduling is always serialized against hlist allocation
4492 * and release. Which makes the protected version suitable here.
4493 * The context lock guarantees that.
4495 hlist = rcu_dereference_protected(swhash->swevent_hlist,
4496 lockdep_is_held(&event->ctx->lock));
4500 return __find_swevent_head(hlist, type, event_id);
4503 static void do_perf_sw_event(enum perf_type_id type, u32 event_id,
4505 struct perf_sample_data *data,
4506 struct pt_regs *regs)
4508 struct swevent_htable *swhash = &__get_cpu_var(swevent_htable);
4509 struct perf_event *event;
4510 struct hlist_node *node;
4511 struct hlist_head *head;
4514 head = find_swevent_head_rcu(swhash, type, event_id);
4518 hlist_for_each_entry_rcu(event, node, head, hlist_entry) {
4519 if (perf_swevent_match(event, type, event_id, data, regs))
4520 perf_swevent_event(event, nr, nmi, data, regs);
4526 int perf_swevent_get_recursion_context(void)
4528 struct swevent_htable *swhash = &__get_cpu_var(swevent_htable);
4530 return get_recursion_context(swhash->recursion);
4532 EXPORT_SYMBOL_GPL(perf_swevent_get_recursion_context);
4534 void inline perf_swevent_put_recursion_context(int rctx)
4536 struct swevent_htable *swhash = &__get_cpu_var(swevent_htable);
4538 put_recursion_context(swhash->recursion, rctx);
4541 void __perf_sw_event(u32 event_id, u64 nr, int nmi,
4542 struct pt_regs *regs, u64 addr)
4544 struct perf_sample_data data;
4547 preempt_disable_notrace();
4548 rctx = perf_swevent_get_recursion_context();
4552 perf_sample_data_init(&data, addr);
4554 do_perf_sw_event(PERF_TYPE_SOFTWARE, event_id, nr, nmi, &data, regs);
4556 perf_swevent_put_recursion_context(rctx);
4557 preempt_enable_notrace();
4560 static void perf_swevent_read(struct perf_event *event)
4564 static int perf_swevent_add(struct perf_event *event, int flags)
4566 struct swevent_htable *swhash = &__get_cpu_var(swevent_htable);
4567 struct hw_perf_event *hwc = &event->hw;
4568 struct hlist_head *head;
4570 if (hwc->sample_period) {
4571 hwc->last_period = hwc->sample_period;
4572 perf_swevent_set_period(event);
4575 hwc->state = !(flags & PERF_EF_START);
4577 head = find_swevent_head(swhash, event);
4578 if (WARN_ON_ONCE(!head))
4581 hlist_add_head_rcu(&event->hlist_entry, head);
4586 static void perf_swevent_del(struct perf_event *event, int flags)
4588 hlist_del_rcu(&event->hlist_entry);
4591 static void perf_swevent_start(struct perf_event *event, int flags)
4593 event->hw.state = 0;
4596 static void perf_swevent_stop(struct perf_event *event, int flags)
4598 event->hw.state = PERF_HES_STOPPED;
4601 /* Deref the hlist from the update side */
4602 static inline struct swevent_hlist *
4603 swevent_hlist_deref(struct swevent_htable *swhash)
4605 return rcu_dereference_protected(swhash->swevent_hlist,
4606 lockdep_is_held(&swhash->hlist_mutex));
4609 static void swevent_hlist_release_rcu(struct rcu_head *rcu_head)
4611 struct swevent_hlist *hlist;
4613 hlist = container_of(rcu_head, struct swevent_hlist, rcu_head);
4617 static void swevent_hlist_release(struct swevent_htable *swhash)
4619 struct swevent_hlist *hlist = swevent_hlist_deref(swhash);
4624 rcu_assign_pointer(swhash->swevent_hlist, NULL);
4625 call_rcu(&hlist->rcu_head, swevent_hlist_release_rcu);
4628 static void swevent_hlist_put_cpu(struct perf_event *event, int cpu)
4630 struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
4632 mutex_lock(&swhash->hlist_mutex);
4634 if (!--swhash->hlist_refcount)
4635 swevent_hlist_release(swhash);
4637 mutex_unlock(&swhash->hlist_mutex);
4640 static void swevent_hlist_put(struct perf_event *event)
4644 if (event->cpu != -1) {
4645 swevent_hlist_put_cpu(event, event->cpu);
4649 for_each_possible_cpu(cpu)
4650 swevent_hlist_put_cpu(event, cpu);
4653 static int swevent_hlist_get_cpu(struct perf_event *event, int cpu)
4655 struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
4658 mutex_lock(&swhash->hlist_mutex);
4660 if (!swevent_hlist_deref(swhash) && cpu_online(cpu)) {
4661 struct swevent_hlist *hlist;
4663 hlist = kzalloc(sizeof(*hlist), GFP_KERNEL);
4668 rcu_assign_pointer(swhash->swevent_hlist, hlist);
4670 swhash->hlist_refcount++;
4672 mutex_unlock(&swhash->hlist_mutex);
4677 static int swevent_hlist_get(struct perf_event *event)
4680 int cpu, failed_cpu;
4682 if (event->cpu != -1)
4683 return swevent_hlist_get_cpu(event, event->cpu);
4686 for_each_possible_cpu(cpu) {
4687 err = swevent_hlist_get_cpu(event, cpu);
4697 for_each_possible_cpu(cpu) {
4698 if (cpu == failed_cpu)
4700 swevent_hlist_put_cpu(event, cpu);
4707 atomic_t perf_swevent_enabled[PERF_COUNT_SW_MAX];
4709 static void sw_perf_event_destroy(struct perf_event *event)
4711 u64 event_id = event->attr.config;
4713 WARN_ON(event->parent);
4715 atomic_dec(&perf_swevent_enabled[event_id]);
4716 swevent_hlist_put(event);
4719 static int perf_swevent_init(struct perf_event *event)
4721 int event_id = event->attr.config;
4723 if (event->attr.type != PERF_TYPE_SOFTWARE)
4727 case PERF_COUNT_SW_CPU_CLOCK:
4728 case PERF_COUNT_SW_TASK_CLOCK:
4735 if (event_id > PERF_COUNT_SW_MAX)
4738 if (!event->parent) {
4741 err = swevent_hlist_get(event);
4745 atomic_inc(&perf_swevent_enabled[event_id]);
4746 event->destroy = sw_perf_event_destroy;
4752 static struct pmu perf_swevent = {
4753 .task_ctx_nr = perf_sw_context,
4755 .event_init = perf_swevent_init,
4756 .add = perf_swevent_add,
4757 .del = perf_swevent_del,
4758 .start = perf_swevent_start,
4759 .stop = perf_swevent_stop,
4760 .read = perf_swevent_read,
4763 #ifdef CONFIG_EVENT_TRACING
4765 static int perf_tp_filter_match(struct perf_event *event,
4766 struct perf_sample_data *data)
4768 void *record = data->raw->data;
4770 if (likely(!event->filter) || filter_match_preds(event->filter, record))
4775 static int perf_tp_event_match(struct perf_event *event,
4776 struct perf_sample_data *data,
4777 struct pt_regs *regs)
4780 * All tracepoints are from kernel-space.
4782 if (event->attr.exclude_kernel)
4785 if (!perf_tp_filter_match(event, data))
4791 void perf_tp_event(u64 addr, u64 count, void *record, int entry_size,
4792 struct pt_regs *regs, struct hlist_head *head, int rctx)
4794 struct perf_sample_data data;
4795 struct perf_event *event;
4796 struct hlist_node *node;
4798 struct perf_raw_record raw = {
4803 perf_sample_data_init(&data, addr);
4806 hlist_for_each_entry_rcu(event, node, head, hlist_entry) {
4807 if (perf_tp_event_match(event, &data, regs))
4808 perf_swevent_event(event, count, 1, &data, regs);
4811 perf_swevent_put_recursion_context(rctx);
4813 EXPORT_SYMBOL_GPL(perf_tp_event);
4815 static void tp_perf_event_destroy(struct perf_event *event)
4817 perf_trace_destroy(event);
4820 static int perf_tp_event_init(struct perf_event *event)
4824 if (event->attr.type != PERF_TYPE_TRACEPOINT)
4828 * Raw tracepoint data is a severe data leak, only allow root to
4831 if ((event->attr.sample_type & PERF_SAMPLE_RAW) &&
4832 perf_paranoid_tracepoint_raw() &&
4833 !capable(CAP_SYS_ADMIN))
4836 err = perf_trace_init(event);
4840 event->destroy = tp_perf_event_destroy;
4845 static struct pmu perf_tracepoint = {
4846 .task_ctx_nr = perf_sw_context,
4848 .event_init = perf_tp_event_init,
4849 .add = perf_trace_add,
4850 .del = perf_trace_del,
4851 .start = perf_swevent_start,
4852 .stop = perf_swevent_stop,
4853 .read = perf_swevent_read,
4856 static inline void perf_tp_register(void)
4858 perf_pmu_register(&perf_tracepoint);
4861 static int perf_event_set_filter(struct perf_event *event, void __user *arg)
4866 if (event->attr.type != PERF_TYPE_TRACEPOINT)
4869 filter_str = strndup_user(arg, PAGE_SIZE);
4870 if (IS_ERR(filter_str))
4871 return PTR_ERR(filter_str);
4873 ret = ftrace_profile_set_filter(event, event->attr.config, filter_str);
4879 static void perf_event_free_filter(struct perf_event *event)
4881 ftrace_profile_free_filter(event);
4886 static inline void perf_tp_register(void)
4890 static int perf_event_set_filter(struct perf_event *event, void __user *arg)
4895 static void perf_event_free_filter(struct perf_event *event)
4899 #endif /* CONFIG_EVENT_TRACING */
4901 #ifdef CONFIG_HAVE_HW_BREAKPOINT
4902 void perf_bp_event(struct perf_event *bp, void *data)
4904 struct perf_sample_data sample;
4905 struct pt_regs *regs = data;
4907 perf_sample_data_init(&sample, bp->attr.bp_addr);
4909 if (!bp->hw.state && !perf_exclude_event(bp, regs))
4910 perf_swevent_event(bp, 1, 1, &sample, regs);
4915 * hrtimer based swevent callback
4918 static enum hrtimer_restart perf_swevent_hrtimer(struct hrtimer *hrtimer)
4920 enum hrtimer_restart ret = HRTIMER_RESTART;
4921 struct perf_sample_data data;
4922 struct pt_regs *regs;
4923 struct perf_event *event;
4926 event = container_of(hrtimer, struct perf_event, hw.hrtimer);
4927 event->pmu->read(event);
4929 perf_sample_data_init(&data, 0);
4930 data.period = event->hw.last_period;
4931 regs = get_irq_regs();
4933 if (regs && !perf_exclude_event(event, regs)) {
4934 if (!(event->attr.exclude_idle && current->pid == 0))
4935 if (perf_event_overflow(event, 0, &data, regs))
4936 ret = HRTIMER_NORESTART;
4939 period = max_t(u64, 10000, event->hw.sample_period);
4940 hrtimer_forward_now(hrtimer, ns_to_ktime(period));
4945 static void perf_swevent_start_hrtimer(struct perf_event *event)
4947 struct hw_perf_event *hwc = &event->hw;
4949 hrtimer_init(&hwc->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
4950 hwc->hrtimer.function = perf_swevent_hrtimer;
4951 if (hwc->sample_period) {
4952 s64 period = local64_read(&hwc->period_left);
4958 local64_set(&hwc->period_left, 0);
4960 period = max_t(u64, 10000, hwc->sample_period);
4962 __hrtimer_start_range_ns(&hwc->hrtimer,
4963 ns_to_ktime(period), 0,
4964 HRTIMER_MODE_REL_PINNED, 0);
4968 static void perf_swevent_cancel_hrtimer(struct perf_event *event)
4970 struct hw_perf_event *hwc = &event->hw;
4972 if (hwc->sample_period) {
4973 ktime_t remaining = hrtimer_get_remaining(&hwc->hrtimer);
4974 local64_set(&hwc->period_left, ktime_to_ns(remaining));
4976 hrtimer_cancel(&hwc->hrtimer);
4981 * Software event: cpu wall time clock
4984 static void cpu_clock_event_update(struct perf_event *event)
4989 now = local_clock();
4990 prev = local64_xchg(&event->hw.prev_count, now);
4991 local64_add(now - prev, &event->count);
4994 static void cpu_clock_event_start(struct perf_event *event, int flags)
4996 local64_set(&event->hw.prev_count, local_clock());
4997 perf_swevent_start_hrtimer(event);
5000 static void cpu_clock_event_stop(struct perf_event *event, int flags)
5002 perf_swevent_cancel_hrtimer(event);
5003 cpu_clock_event_update(event);
5006 static int cpu_clock_event_add(struct perf_event *event, int flags)
5008 if (flags & PERF_EF_START)
5009 cpu_clock_event_start(event, flags);
5014 static void cpu_clock_event_del(struct perf_event *event, int flags)
5016 cpu_clock_event_stop(event, flags);
5019 static void cpu_clock_event_read(struct perf_event *event)
5021 cpu_clock_event_update(event);
5024 static int cpu_clock_event_init(struct perf_event *event)
5026 if (event->attr.type != PERF_TYPE_SOFTWARE)
5029 if (event->attr.config != PERF_COUNT_SW_CPU_CLOCK)
5035 static struct pmu perf_cpu_clock = {
5036 .task_ctx_nr = perf_sw_context,
5038 .event_init = cpu_clock_event_init,
5039 .add = cpu_clock_event_add,
5040 .del = cpu_clock_event_del,
5041 .start = cpu_clock_event_start,
5042 .stop = cpu_clock_event_stop,
5043 .read = cpu_clock_event_read,
5047 * Software event: task time clock
5050 static void task_clock_event_update(struct perf_event *event, u64 now)
5055 prev = local64_xchg(&event->hw.prev_count, now);
5057 local64_add(delta, &event->count);
5060 static void task_clock_event_start(struct perf_event *event, int flags)
5062 local64_set(&event->hw.prev_count, event->ctx->time);
5063 perf_swevent_start_hrtimer(event);
5066 static void task_clock_event_stop(struct perf_event *event, int flags)
5068 perf_swevent_cancel_hrtimer(event);
5069 task_clock_event_update(event, event->ctx->time);
5072 static int task_clock_event_add(struct perf_event *event, int flags)
5074 if (flags & PERF_EF_START)
5075 task_clock_event_start(event, flags);
5080 static void task_clock_event_del(struct perf_event *event, int flags)
5082 task_clock_event_stop(event, PERF_EF_UPDATE);
5085 static void task_clock_event_read(struct perf_event *event)
5090 update_context_time(event->ctx);
5091 time = event->ctx->time;
5093 u64 now = perf_clock();
5094 u64 delta = now - event->ctx->timestamp;
5095 time = event->ctx->time + delta;
5098 task_clock_event_update(event, time);
5101 static int task_clock_event_init(struct perf_event *event)
5103 if (event->attr.type != PERF_TYPE_SOFTWARE)
5106 if (event->attr.config != PERF_COUNT_SW_TASK_CLOCK)
5112 static struct pmu perf_task_clock = {
5113 .task_ctx_nr = perf_sw_context,
5115 .event_init = task_clock_event_init,
5116 .add = task_clock_event_add,
5117 .del = task_clock_event_del,
5118 .start = task_clock_event_start,
5119 .stop = task_clock_event_stop,
5120 .read = task_clock_event_read,
5123 static void perf_pmu_nop_void(struct pmu *pmu)
5127 static int perf_pmu_nop_int(struct pmu *pmu)
5132 static void perf_pmu_start_txn(struct pmu *pmu)
5134 perf_pmu_disable(pmu);
5137 static int perf_pmu_commit_txn(struct pmu *pmu)
5139 perf_pmu_enable(pmu);
5143 static void perf_pmu_cancel_txn(struct pmu *pmu)
5145 perf_pmu_enable(pmu);
5149 * Ensures all contexts with the same task_ctx_nr have the same
5150 * pmu_cpu_context too.
5152 static void *find_pmu_context(int ctxn)
5159 list_for_each_entry(pmu, &pmus, entry) {
5160 if (pmu->task_ctx_nr == ctxn)
5161 return pmu->pmu_cpu_context;
5167 static void free_pmu_context(void * __percpu cpu_context)
5171 mutex_lock(&pmus_lock);
5173 * Like a real lame refcount.
5175 list_for_each_entry(pmu, &pmus, entry) {
5176 if (pmu->pmu_cpu_context == cpu_context)
5180 free_percpu(cpu_context);
5182 mutex_unlock(&pmus_lock);
5185 int perf_pmu_register(struct pmu *pmu)
5189 mutex_lock(&pmus_lock);
5191 pmu->pmu_disable_count = alloc_percpu(int);
5192 if (!pmu->pmu_disable_count)
5195 pmu->pmu_cpu_context = find_pmu_context(pmu->task_ctx_nr);
5196 if (pmu->pmu_cpu_context)
5197 goto got_cpu_context;
5199 pmu->pmu_cpu_context = alloc_percpu(struct perf_cpu_context);
5200 if (!pmu->pmu_cpu_context)
5203 for_each_possible_cpu(cpu) {
5204 struct perf_cpu_context *cpuctx;
5206 cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu);
5207 __perf_event_init_context(&cpuctx->ctx);
5208 cpuctx->ctx.type = cpu_context;
5209 cpuctx->ctx.pmu = pmu;
5210 cpuctx->jiffies_interval = 1;
5211 INIT_LIST_HEAD(&cpuctx->rotation_list);
5215 if (!pmu->start_txn) {
5216 if (pmu->pmu_enable) {
5218 * If we have pmu_enable/pmu_disable calls, install
5219 * transaction stubs that use that to try and batch
5220 * hardware accesses.
5222 pmu->start_txn = perf_pmu_start_txn;
5223 pmu->commit_txn = perf_pmu_commit_txn;
5224 pmu->cancel_txn = perf_pmu_cancel_txn;
5226 pmu->start_txn = perf_pmu_nop_void;
5227 pmu->commit_txn = perf_pmu_nop_int;
5228 pmu->cancel_txn = perf_pmu_nop_void;
5232 if (!pmu->pmu_enable) {
5233 pmu->pmu_enable = perf_pmu_nop_void;
5234 pmu->pmu_disable = perf_pmu_nop_void;
5237 list_add_rcu(&pmu->entry, &pmus);
5240 mutex_unlock(&pmus_lock);
5245 free_percpu(pmu->pmu_disable_count);
5249 void perf_pmu_unregister(struct pmu *pmu)
5251 mutex_lock(&pmus_lock);
5252 list_del_rcu(&pmu->entry);
5253 mutex_unlock(&pmus_lock);
5256 * We dereference the pmu list under both SRCU and regular RCU, so
5257 * synchronize against both of those.
5259 synchronize_srcu(&pmus_srcu);
5262 free_percpu(pmu->pmu_disable_count);
5263 free_pmu_context(pmu->pmu_cpu_context);
5266 struct pmu *perf_init_event(struct perf_event *event)
5268 struct pmu *pmu = NULL;
5271 idx = srcu_read_lock(&pmus_srcu);
5272 list_for_each_entry_rcu(pmu, &pmus, entry) {
5273 int ret = pmu->event_init(event);
5277 if (ret != -ENOENT) {
5282 pmu = ERR_PTR(-ENOENT);
5284 srcu_read_unlock(&pmus_srcu, idx);
5290 * Allocate and initialize a event structure
5292 static struct perf_event *
5293 perf_event_alloc(struct perf_event_attr *attr, int cpu,
5294 struct perf_event *group_leader,
5295 struct perf_event *parent_event,
5296 perf_overflow_handler_t overflow_handler)
5299 struct perf_event *event;
5300 struct hw_perf_event *hwc;
5303 event = kzalloc(sizeof(*event), GFP_KERNEL);
5305 return ERR_PTR(-ENOMEM);
5308 * Single events are their own group leaders, with an
5309 * empty sibling list:
5312 group_leader = event;
5314 mutex_init(&event->child_mutex);
5315 INIT_LIST_HEAD(&event->child_list);
5317 INIT_LIST_HEAD(&event->group_entry);
5318 INIT_LIST_HEAD(&event->event_entry);
5319 INIT_LIST_HEAD(&event->sibling_list);
5320 init_waitqueue_head(&event->waitq);
5322 mutex_init(&event->mmap_mutex);
5325 event->attr = *attr;
5326 event->group_leader = group_leader;
5330 event->parent = parent_event;
5332 event->ns = get_pid_ns(current->nsproxy->pid_ns);
5333 event->id = atomic64_inc_return(&perf_event_id);
5335 event->state = PERF_EVENT_STATE_INACTIVE;
5337 if (!overflow_handler && parent_event)
5338 overflow_handler = parent_event->overflow_handler;
5340 event->overflow_handler = overflow_handler;
5343 event->state = PERF_EVENT_STATE_OFF;
5348 hwc->sample_period = attr->sample_period;
5349 if (attr->freq && attr->sample_freq)
5350 hwc->sample_period = 1;
5351 hwc->last_period = hwc->sample_period;
5353 local64_set(&hwc->period_left, hwc->sample_period);
5356 * we currently do not support PERF_FORMAT_GROUP on inherited events
5358 if (attr->inherit && (attr->read_format & PERF_FORMAT_GROUP))
5361 pmu = perf_init_event(event);
5367 else if (IS_ERR(pmu))
5372 put_pid_ns(event->ns);
5374 return ERR_PTR(err);
5379 if (!event->parent) {
5380 atomic_inc(&nr_events);
5381 if (event->attr.mmap || event->attr.mmap_data)
5382 atomic_inc(&nr_mmap_events);
5383 if (event->attr.comm)
5384 atomic_inc(&nr_comm_events);
5385 if (event->attr.task)
5386 atomic_inc(&nr_task_events);
5387 if (event->attr.sample_type & PERF_SAMPLE_CALLCHAIN) {
5388 err = get_callchain_buffers();
5391 return ERR_PTR(err);
5399 static int perf_copy_attr(struct perf_event_attr __user *uattr,
5400 struct perf_event_attr *attr)
5405 if (!access_ok(VERIFY_WRITE, uattr, PERF_ATTR_SIZE_VER0))
5409 * zero the full structure, so that a short copy will be nice.
5411 memset(attr, 0, sizeof(*attr));
5413 ret = get_user(size, &uattr->size);
5417 if (size > PAGE_SIZE) /* silly large */
5420 if (!size) /* abi compat */
5421 size = PERF_ATTR_SIZE_VER0;
5423 if (size < PERF_ATTR_SIZE_VER0)
5427 * If we're handed a bigger struct than we know of,
5428 * ensure all the unknown bits are 0 - i.e. new
5429 * user-space does not rely on any kernel feature
5430 * extensions we dont know about yet.
5432 if (size > sizeof(*attr)) {
5433 unsigned char __user *addr;
5434 unsigned char __user *end;
5437 addr = (void __user *)uattr + sizeof(*attr);
5438 end = (void __user *)uattr + size;
5440 for (; addr < end; addr++) {
5441 ret = get_user(val, addr);
5447 size = sizeof(*attr);
5450 ret = copy_from_user(attr, uattr, size);
5455 * If the type exists, the corresponding creation will verify
5458 if (attr->type >= PERF_TYPE_MAX)
5461 if (attr->__reserved_1)
5464 if (attr->sample_type & ~(PERF_SAMPLE_MAX-1))
5467 if (attr->read_format & ~(PERF_FORMAT_MAX-1))
5474 put_user(sizeof(*attr), &uattr->size);
5480 perf_event_set_output(struct perf_event *event, struct perf_event *output_event)
5482 struct perf_buffer *buffer = NULL, *old_buffer = NULL;
5488 /* don't allow circular references */
5489 if (event == output_event)
5493 * Don't allow cross-cpu buffers
5495 if (output_event->cpu != event->cpu)
5499 * If its not a per-cpu buffer, it must be the same task.
5501 if (output_event->cpu == -1 && output_event->ctx != event->ctx)
5505 mutex_lock(&event->mmap_mutex);
5506 /* Can't redirect output if we've got an active mmap() */
5507 if (atomic_read(&event->mmap_count))
5511 /* get the buffer we want to redirect to */
5512 buffer = perf_buffer_get(output_event);
5517 old_buffer = event->buffer;
5518 rcu_assign_pointer(event->buffer, buffer);
5521 mutex_unlock(&event->mmap_mutex);
5524 perf_buffer_put(old_buffer);
5530 * sys_perf_event_open - open a performance event, associate it to a task/cpu
5532 * @attr_uptr: event_id type attributes for monitoring/sampling
5535 * @group_fd: group leader event fd
5537 SYSCALL_DEFINE5(perf_event_open,
5538 struct perf_event_attr __user *, attr_uptr,
5539 pid_t, pid, int, cpu, int, group_fd, unsigned long, flags)
5541 struct perf_event *group_leader = NULL, *output_event = NULL;
5542 struct perf_event *event, *sibling;
5543 struct perf_event_attr attr;
5544 struct perf_event_context *ctx;
5545 struct file *event_file = NULL;
5546 struct file *group_file = NULL;
5547 struct task_struct *task = NULL;
5551 int fput_needed = 0;
5554 /* for future expandability... */
5555 if (flags & ~(PERF_FLAG_FD_NO_GROUP | PERF_FLAG_FD_OUTPUT))
5558 err = perf_copy_attr(attr_uptr, &attr);
5562 if (!attr.exclude_kernel) {
5563 if (perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN))
5568 if (attr.sample_freq > sysctl_perf_event_sample_rate)
5572 event_fd = get_unused_fd_flags(O_RDWR);
5576 if (group_fd != -1) {
5577 group_leader = perf_fget_light(group_fd, &fput_needed);
5578 if (IS_ERR(group_leader)) {
5579 err = PTR_ERR(group_leader);
5582 group_file = group_leader->filp;
5583 if (flags & PERF_FLAG_FD_OUTPUT)
5584 output_event = group_leader;
5585 if (flags & PERF_FLAG_FD_NO_GROUP)
5586 group_leader = NULL;
5589 event = perf_event_alloc(&attr, cpu, group_leader, NULL, NULL);
5590 if (IS_ERR(event)) {
5591 err = PTR_ERR(event);
5596 * Special case software events and allow them to be part of
5597 * any hardware group.
5602 (is_software_event(event) != is_software_event(group_leader))) {
5603 if (is_software_event(event)) {
5605 * If event and group_leader are not both a software
5606 * event, and event is, then group leader is not.
5608 * Allow the addition of software events to !software
5609 * groups, this is safe because software events never
5612 pmu = group_leader->pmu;
5613 } else if (is_software_event(group_leader) &&
5614 (group_leader->group_flags & PERF_GROUP_SOFTWARE)) {
5616 * In case the group is a pure software group, and we
5617 * try to add a hardware event, move the whole group to
5618 * the hardware context.
5625 task = find_lively_task_by_vpid(pid);
5627 err = PTR_ERR(task);
5633 * Get the target context (task or percpu):
5635 ctx = find_get_context(pmu, task, cpu);
5642 * Look up the group leader (we will attach this event to it):
5648 * Do not allow a recursive hierarchy (this new sibling
5649 * becoming part of another group-sibling):
5651 if (group_leader->group_leader != group_leader)
5654 * Do not allow to attach to a group in a different
5655 * task or CPU context:
5658 if (group_leader->ctx->type != ctx->type)
5661 if (group_leader->ctx != ctx)
5666 * Only a group leader can be exclusive or pinned
5668 if (attr.exclusive || attr.pinned)
5673 err = perf_event_set_output(event, output_event);
5678 event_file = anon_inode_getfile("[perf_event]", &perf_fops, event, O_RDWR);
5679 if (IS_ERR(event_file)) {
5680 err = PTR_ERR(event_file);
5685 struct perf_event_context *gctx = group_leader->ctx;
5687 mutex_lock(&gctx->mutex);
5688 perf_event_remove_from_context(group_leader);
5689 list_for_each_entry(sibling, &group_leader->sibling_list,
5691 perf_event_remove_from_context(sibling);
5694 mutex_unlock(&gctx->mutex);
5698 event->filp = event_file;
5699 WARN_ON_ONCE(ctx->parent_ctx);
5700 mutex_lock(&ctx->mutex);
5703 perf_install_in_context(ctx, group_leader, cpu);
5705 list_for_each_entry(sibling, &group_leader->sibling_list,
5707 perf_install_in_context(ctx, sibling, cpu);
5712 perf_install_in_context(ctx, event, cpu);
5714 mutex_unlock(&ctx->mutex);
5716 event->owner = current;
5717 get_task_struct(current);
5718 mutex_lock(¤t->perf_event_mutex);
5719 list_add_tail(&event->owner_entry, ¤t->perf_event_list);
5720 mutex_unlock(¤t->perf_event_mutex);
5723 * Drop the reference on the group_event after placing the
5724 * new event on the sibling_list. This ensures destruction
5725 * of the group leader will find the pointer to itself in
5726 * perf_group_detach().
5728 fput_light(group_file, fput_needed);
5729 fd_install(event_fd, event_file);
5735 fput_light(group_file, fput_needed);
5738 put_unused_fd(event_fd);
5743 * perf_event_create_kernel_counter
5745 * @attr: attributes of the counter to create
5746 * @cpu: cpu in which the counter is bound
5747 * @task: task to profile (NULL for percpu)
5750 perf_event_create_kernel_counter(struct perf_event_attr *attr, int cpu,
5751 struct task_struct *task,
5752 perf_overflow_handler_t overflow_handler)
5754 struct perf_event_context *ctx;
5755 struct perf_event *event;
5759 * Get the target context (task or percpu):
5762 event = perf_event_alloc(attr, cpu, NULL, NULL, overflow_handler);
5763 if (IS_ERR(event)) {
5764 err = PTR_ERR(event);
5768 ctx = find_get_context(event->pmu, task, cpu);
5775 WARN_ON_ONCE(ctx->parent_ctx);
5776 mutex_lock(&ctx->mutex);
5777 perf_install_in_context(ctx, event, cpu);
5779 mutex_unlock(&ctx->mutex);
5781 event->owner = current;
5782 get_task_struct(current);
5783 mutex_lock(¤t->perf_event_mutex);
5784 list_add_tail(&event->owner_entry, ¤t->perf_event_list);
5785 mutex_unlock(¤t->perf_event_mutex);
5792 return ERR_PTR(err);
5794 EXPORT_SYMBOL_GPL(perf_event_create_kernel_counter);
5796 static void sync_child_event(struct perf_event *child_event,
5797 struct task_struct *child)
5799 struct perf_event *parent_event = child_event->parent;
5802 if (child_event->attr.inherit_stat)
5803 perf_event_read_event(child_event, child);
5805 child_val = perf_event_count(child_event);
5808 * Add back the child's count to the parent's count:
5810 atomic64_add(child_val, &parent_event->child_count);
5811 atomic64_add(child_event->total_time_enabled,
5812 &parent_event->child_total_time_enabled);
5813 atomic64_add(child_event->total_time_running,
5814 &parent_event->child_total_time_running);
5817 * Remove this event from the parent's list
5819 WARN_ON_ONCE(parent_event->ctx->parent_ctx);
5820 mutex_lock(&parent_event->child_mutex);
5821 list_del_init(&child_event->child_list);
5822 mutex_unlock(&parent_event->child_mutex);
5825 * Release the parent event, if this was the last
5828 fput(parent_event->filp);
5832 __perf_event_exit_task(struct perf_event *child_event,
5833 struct perf_event_context *child_ctx,
5834 struct task_struct *child)
5836 struct perf_event *parent_event;
5838 perf_event_remove_from_context(child_event);
5840 parent_event = child_event->parent;
5842 * It can happen that parent exits first, and has events
5843 * that are still around due to the child reference. These
5844 * events need to be zapped - but otherwise linger.
5847 sync_child_event(child_event, child);
5848 free_event(child_event);
5852 static void perf_event_exit_task_context(struct task_struct *child, int ctxn)
5854 struct perf_event *child_event, *tmp;
5855 struct perf_event_context *child_ctx;
5856 unsigned long flags;
5858 if (likely(!child->perf_event_ctxp[ctxn])) {
5859 perf_event_task(child, NULL, 0);
5863 local_irq_save(flags);
5865 * We can't reschedule here because interrupts are disabled,
5866 * and either child is current or it is a task that can't be
5867 * scheduled, so we are now safe from rescheduling changing
5870 child_ctx = child->perf_event_ctxp[ctxn];
5871 __perf_event_task_sched_out(child_ctx);
5874 * Take the context lock here so that if find_get_context is
5875 * reading child->perf_event_ctxp, we wait until it has
5876 * incremented the context's refcount before we do put_ctx below.
5878 raw_spin_lock(&child_ctx->lock);
5879 child->perf_event_ctxp[ctxn] = NULL;
5881 * If this context is a clone; unclone it so it can't get
5882 * swapped to another process while we're removing all
5883 * the events from it.
5885 unclone_ctx(child_ctx);
5886 update_context_time(child_ctx);
5887 raw_spin_unlock_irqrestore(&child_ctx->lock, flags);
5890 * Report the task dead after unscheduling the events so that we
5891 * won't get any samples after PERF_RECORD_EXIT. We can however still
5892 * get a few PERF_RECORD_READ events.
5894 perf_event_task(child, child_ctx, 0);
5897 * We can recurse on the same lock type through:
5899 * __perf_event_exit_task()
5900 * sync_child_event()
5901 * fput(parent_event->filp)
5903 * mutex_lock(&ctx->mutex)
5905 * But since its the parent context it won't be the same instance.
5907 mutex_lock(&child_ctx->mutex);
5910 list_for_each_entry_safe(child_event, tmp, &child_ctx->pinned_groups,
5912 __perf_event_exit_task(child_event, child_ctx, child);
5914 list_for_each_entry_safe(child_event, tmp, &child_ctx->flexible_groups,
5916 __perf_event_exit_task(child_event, child_ctx, child);
5919 * If the last event was a group event, it will have appended all
5920 * its siblings to the list, but we obtained 'tmp' before that which
5921 * will still point to the list head terminating the iteration.
5923 if (!list_empty(&child_ctx->pinned_groups) ||
5924 !list_empty(&child_ctx->flexible_groups))
5927 mutex_unlock(&child_ctx->mutex);
5933 * When a child task exits, feed back event values to parent events.
5935 void perf_event_exit_task(struct task_struct *child)
5939 for_each_task_context_nr(ctxn)
5940 perf_event_exit_task_context(child, ctxn);
5943 static void perf_free_event(struct perf_event *event,
5944 struct perf_event_context *ctx)
5946 struct perf_event *parent = event->parent;
5948 if (WARN_ON_ONCE(!parent))
5951 mutex_lock(&parent->child_mutex);
5952 list_del_init(&event->child_list);
5953 mutex_unlock(&parent->child_mutex);
5957 perf_group_detach(event);
5958 list_del_event(event, ctx);
5963 * free an unexposed, unused context as created by inheritance by
5964 * perf_event_init_task below, used by fork() in case of fail.
5966 void perf_event_free_task(struct task_struct *task)
5968 struct perf_event_context *ctx;
5969 struct perf_event *event, *tmp;
5972 for_each_task_context_nr(ctxn) {
5973 ctx = task->perf_event_ctxp[ctxn];
5977 mutex_lock(&ctx->mutex);
5979 list_for_each_entry_safe(event, tmp, &ctx->pinned_groups,
5981 perf_free_event(event, ctx);
5983 list_for_each_entry_safe(event, tmp, &ctx->flexible_groups,
5985 perf_free_event(event, ctx);
5987 if (!list_empty(&ctx->pinned_groups) ||
5988 !list_empty(&ctx->flexible_groups))
5991 mutex_unlock(&ctx->mutex);
5997 void perf_event_delayed_put(struct task_struct *task)
6001 for_each_task_context_nr(ctxn)
6002 WARN_ON_ONCE(task->perf_event_ctxp[ctxn]);
6006 * inherit a event from parent task to child task:
6008 static struct perf_event *
6009 inherit_event(struct perf_event *parent_event,
6010 struct task_struct *parent,
6011 struct perf_event_context *parent_ctx,
6012 struct task_struct *child,
6013 struct perf_event *group_leader,
6014 struct perf_event_context *child_ctx)
6016 struct perf_event *child_event;
6017 unsigned long flags;
6020 * Instead of creating recursive hierarchies of events,
6021 * we link inherited events back to the original parent,
6022 * which has a filp for sure, which we use as the reference
6025 if (parent_event->parent)
6026 parent_event = parent_event->parent;
6028 child_event = perf_event_alloc(&parent_event->attr,
6030 group_leader, parent_event,
6032 if (IS_ERR(child_event))
6037 * Make the child state follow the state of the parent event,
6038 * not its attr.disabled bit. We hold the parent's mutex,
6039 * so we won't race with perf_event_{en, dis}able_family.
6041 if (parent_event->state >= PERF_EVENT_STATE_INACTIVE)
6042 child_event->state = PERF_EVENT_STATE_INACTIVE;
6044 child_event->state = PERF_EVENT_STATE_OFF;
6046 if (parent_event->attr.freq) {
6047 u64 sample_period = parent_event->hw.sample_period;
6048 struct hw_perf_event *hwc = &child_event->hw;
6050 hwc->sample_period = sample_period;
6051 hwc->last_period = sample_period;
6053 local64_set(&hwc->period_left, sample_period);
6056 child_event->ctx = child_ctx;
6057 child_event->overflow_handler = parent_event->overflow_handler;
6060 * Link it up in the child's context:
6062 raw_spin_lock_irqsave(&child_ctx->lock, flags);
6063 add_event_to_ctx(child_event, child_ctx);
6064 raw_spin_unlock_irqrestore(&child_ctx->lock, flags);
6067 * Get a reference to the parent filp - we will fput it
6068 * when the child event exits. This is safe to do because
6069 * we are in the parent and we know that the filp still
6070 * exists and has a nonzero count:
6072 atomic_long_inc(&parent_event->filp->f_count);
6075 * Link this into the parent event's child list
6077 WARN_ON_ONCE(parent_event->ctx->parent_ctx);
6078 mutex_lock(&parent_event->child_mutex);
6079 list_add_tail(&child_event->child_list, &parent_event->child_list);
6080 mutex_unlock(&parent_event->child_mutex);
6085 static int inherit_group(struct perf_event *parent_event,
6086 struct task_struct *parent,
6087 struct perf_event_context *parent_ctx,
6088 struct task_struct *child,
6089 struct perf_event_context *child_ctx)
6091 struct perf_event *leader;
6092 struct perf_event *sub;
6093 struct perf_event *child_ctr;
6095 leader = inherit_event(parent_event, parent, parent_ctx,
6096 child, NULL, child_ctx);
6098 return PTR_ERR(leader);
6099 list_for_each_entry(sub, &parent_event->sibling_list, group_entry) {
6100 child_ctr = inherit_event(sub, parent, parent_ctx,
6101 child, leader, child_ctx);
6102 if (IS_ERR(child_ctr))
6103 return PTR_ERR(child_ctr);
6109 inherit_task_group(struct perf_event *event, struct task_struct *parent,
6110 struct perf_event_context *parent_ctx,
6111 struct task_struct *child, int ctxn,
6115 struct perf_event_context *child_ctx;
6117 if (!event->attr.inherit) {
6122 child_ctx = child->perf_event_ctxp[ctxn];
6125 * This is executed from the parent task context, so
6126 * inherit events that have been marked for cloning.
6127 * First allocate and initialize a context for the
6131 child_ctx = alloc_perf_context(event->pmu, child);
6135 child->perf_event_ctxp[ctxn] = child_ctx;
6138 ret = inherit_group(event, parent, parent_ctx,
6148 * Initialize the perf_event context in task_struct
6150 int perf_event_init_context(struct task_struct *child, int ctxn)
6152 struct perf_event_context *child_ctx, *parent_ctx;
6153 struct perf_event_context *cloned_ctx;
6154 struct perf_event *event;
6155 struct task_struct *parent = current;
6156 int inherited_all = 1;
6159 child->perf_event_ctxp[ctxn] = NULL;
6161 mutex_init(&child->perf_event_mutex);
6162 INIT_LIST_HEAD(&child->perf_event_list);
6164 if (likely(!parent->perf_event_ctxp[ctxn]))
6168 * If the parent's context is a clone, pin it so it won't get
6171 parent_ctx = perf_pin_task_context(parent, ctxn);
6174 * No need to check if parent_ctx != NULL here; since we saw
6175 * it non-NULL earlier, the only reason for it to become NULL
6176 * is if we exit, and since we're currently in the middle of
6177 * a fork we can't be exiting at the same time.
6181 * Lock the parent list. No need to lock the child - not PID
6182 * hashed yet and not running, so nobody can access it.
6184 mutex_lock(&parent_ctx->mutex);
6187 * We dont have to disable NMIs - we are only looking at
6188 * the list, not manipulating it:
6190 list_for_each_entry(event, &parent_ctx->pinned_groups, group_entry) {
6191 ret = inherit_task_group(event, parent, parent_ctx,
6192 child, ctxn, &inherited_all);
6197 list_for_each_entry(event, &parent_ctx->flexible_groups, group_entry) {
6198 ret = inherit_task_group(event, parent, parent_ctx,
6199 child, ctxn, &inherited_all);
6204 child_ctx = child->perf_event_ctxp[ctxn];
6206 if (child_ctx && inherited_all) {
6208 * Mark the child context as a clone of the parent
6209 * context, or of whatever the parent is a clone of.
6210 * Note that if the parent is a clone, it could get
6211 * uncloned at any point, but that doesn't matter
6212 * because the list of events and the generation
6213 * count can't have changed since we took the mutex.
6215 cloned_ctx = rcu_dereference(parent_ctx->parent_ctx);
6217 child_ctx->parent_ctx = cloned_ctx;
6218 child_ctx->parent_gen = parent_ctx->parent_gen;
6220 child_ctx->parent_ctx = parent_ctx;
6221 child_ctx->parent_gen = parent_ctx->generation;
6223 get_ctx(child_ctx->parent_ctx);
6226 mutex_unlock(&parent_ctx->mutex);
6228 perf_unpin_context(parent_ctx);
6234 * Initialize the perf_event context in task_struct
6236 int perf_event_init_task(struct task_struct *child)
6240 for_each_task_context_nr(ctxn) {
6241 ret = perf_event_init_context(child, ctxn);
6249 static void __init perf_event_init_all_cpus(void)
6251 struct swevent_htable *swhash;
6254 for_each_possible_cpu(cpu) {
6255 swhash = &per_cpu(swevent_htable, cpu);
6256 mutex_init(&swhash->hlist_mutex);
6257 INIT_LIST_HEAD(&per_cpu(rotation_list, cpu));
6261 static void __cpuinit perf_event_init_cpu(int cpu)
6263 struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
6265 mutex_lock(&swhash->hlist_mutex);
6266 if (swhash->hlist_refcount > 0) {
6267 struct swevent_hlist *hlist;
6269 hlist = kzalloc_node(sizeof(*hlist), GFP_KERNEL, cpu_to_node(cpu));
6271 rcu_assign_pointer(swhash->swevent_hlist, hlist);
6273 mutex_unlock(&swhash->hlist_mutex);
6276 #ifdef CONFIG_HOTPLUG_CPU
6277 static void perf_pmu_rotate_stop(struct pmu *pmu)
6279 struct perf_cpu_context *cpuctx = this_cpu_ptr(pmu->pmu_cpu_context);
6281 WARN_ON(!irqs_disabled());
6283 list_del_init(&cpuctx->rotation_list);
6286 static void __perf_event_exit_context(void *__info)
6288 struct perf_event_context *ctx = __info;
6289 struct perf_event *event, *tmp;
6291 perf_pmu_rotate_stop(ctx->pmu);
6293 list_for_each_entry_safe(event, tmp, &ctx->pinned_groups, group_entry)
6294 __perf_event_remove_from_context(event);
6295 list_for_each_entry_safe(event, tmp, &ctx->flexible_groups, group_entry)
6296 __perf_event_remove_from_context(event);
6299 static void perf_event_exit_cpu_context(int cpu)
6301 struct perf_event_context *ctx;
6305 idx = srcu_read_lock(&pmus_srcu);
6306 list_for_each_entry_rcu(pmu, &pmus, entry) {
6307 ctx = &per_cpu_ptr(pmu->pmu_cpu_context, cpu)->ctx;
6309 mutex_lock(&ctx->mutex);
6310 smp_call_function_single(cpu, __perf_event_exit_context, ctx, 1);
6311 mutex_unlock(&ctx->mutex);
6313 srcu_read_unlock(&pmus_srcu, idx);
6316 static void perf_event_exit_cpu(int cpu)
6318 struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
6320 mutex_lock(&swhash->hlist_mutex);
6321 swevent_hlist_release(swhash);
6322 mutex_unlock(&swhash->hlist_mutex);
6324 perf_event_exit_cpu_context(cpu);
6327 static inline void perf_event_exit_cpu(int cpu) { }
6330 static int __cpuinit
6331 perf_cpu_notify(struct notifier_block *self, unsigned long action, void *hcpu)
6333 unsigned int cpu = (long)hcpu;
6335 switch (action & ~CPU_TASKS_FROZEN) {
6337 case CPU_UP_PREPARE:
6338 case CPU_DOWN_FAILED:
6339 perf_event_init_cpu(cpu);
6342 case CPU_UP_CANCELED:
6343 case CPU_DOWN_PREPARE:
6344 perf_event_exit_cpu(cpu);
6354 void __init perf_event_init(void)
6356 perf_event_init_all_cpus();
6357 init_srcu_struct(&pmus_srcu);
6358 perf_pmu_register(&perf_swevent);
6359 perf_pmu_register(&perf_cpu_clock);
6360 perf_pmu_register(&perf_task_clock);
6362 perf_cpu_notifier(perf_cpu_notify);