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>
34 #include <linux/hw_breakpoint.h>
36 #include <asm/irq_regs.h>
39 * Each CPU has a list of per CPU events:
41 static DEFINE_PER_CPU(struct perf_cpu_context, perf_cpu_context);
43 int perf_max_events __read_mostly = 1;
44 static int perf_reserved_percpu __read_mostly;
45 static int perf_overcommit __read_mostly = 1;
47 static atomic_t nr_events __read_mostly;
48 static atomic_t nr_mmap_events __read_mostly;
49 static atomic_t nr_comm_events __read_mostly;
50 static atomic_t nr_task_events __read_mostly;
53 * perf event paranoia level:
54 * -1 - not paranoid at all
55 * 0 - disallow raw tracepoint access for unpriv
56 * 1 - disallow cpu events for unpriv
57 * 2 - disallow kernel profiling for unpriv
59 int sysctl_perf_event_paranoid __read_mostly = 1;
61 int sysctl_perf_event_mlock __read_mostly = 512; /* 'free' kb per user */
64 * max perf event sample rate
66 int sysctl_perf_event_sample_rate __read_mostly = 100000;
68 static atomic64_t perf_event_id;
71 * Lock for (sysadmin-configurable) event reservations:
73 static DEFINE_SPINLOCK(perf_resource_lock);
76 * Architecture provided APIs - weak aliases:
78 extern __weak const struct pmu *hw_perf_event_init(struct perf_event *event)
83 void __weak hw_perf_disable(void) { barrier(); }
84 void __weak hw_perf_enable(void) { barrier(); }
86 void __weak perf_event_print_debug(void) { }
88 static DEFINE_PER_CPU(int, perf_disable_count);
90 void perf_disable(void)
92 if (!__get_cpu_var(perf_disable_count)++)
96 void perf_enable(void)
98 if (!--__get_cpu_var(perf_disable_count))
102 static void get_ctx(struct perf_event_context *ctx)
104 WARN_ON(!atomic_inc_not_zero(&ctx->refcount));
107 static void free_ctx(struct rcu_head *head)
109 struct perf_event_context *ctx;
111 ctx = container_of(head, struct perf_event_context, rcu_head);
115 static void put_ctx(struct perf_event_context *ctx)
117 if (atomic_dec_and_test(&ctx->refcount)) {
119 put_ctx(ctx->parent_ctx);
121 put_task_struct(ctx->task);
122 call_rcu(&ctx->rcu_head, free_ctx);
126 static void unclone_ctx(struct perf_event_context *ctx)
128 if (ctx->parent_ctx) {
129 put_ctx(ctx->parent_ctx);
130 ctx->parent_ctx = NULL;
135 * If we inherit events we want to return the parent event id
138 static u64 primary_event_id(struct perf_event *event)
143 id = event->parent->id;
149 * Get the perf_event_context for a task and lock it.
150 * This has to cope with with the fact that until it is locked,
151 * the context could get moved to another task.
153 static struct perf_event_context *
154 perf_lock_task_context(struct task_struct *task, unsigned long *flags)
156 struct perf_event_context *ctx;
160 ctx = rcu_dereference(task->perf_event_ctxp);
163 * If this context is a clone of another, it might
164 * get swapped for another underneath us by
165 * perf_event_task_sched_out, though the
166 * rcu_read_lock() protects us from any context
167 * getting freed. Lock the context and check if it
168 * got swapped before we could get the lock, and retry
169 * if so. If we locked the right context, then it
170 * can't get swapped on us any more.
172 raw_spin_lock_irqsave(&ctx->lock, *flags);
173 if (ctx != rcu_dereference(task->perf_event_ctxp)) {
174 raw_spin_unlock_irqrestore(&ctx->lock, *flags);
178 if (!atomic_inc_not_zero(&ctx->refcount)) {
179 raw_spin_unlock_irqrestore(&ctx->lock, *flags);
188 * Get the context for a task and increment its pin_count so it
189 * can't get swapped to another task. This also increments its
190 * reference count so that the context can't get freed.
192 static struct perf_event_context *perf_pin_task_context(struct task_struct *task)
194 struct perf_event_context *ctx;
197 ctx = perf_lock_task_context(task, &flags);
200 raw_spin_unlock_irqrestore(&ctx->lock, flags);
205 static void perf_unpin_context(struct perf_event_context *ctx)
209 raw_spin_lock_irqsave(&ctx->lock, flags);
211 raw_spin_unlock_irqrestore(&ctx->lock, flags);
215 static inline u64 perf_clock(void)
217 return cpu_clock(raw_smp_processor_id());
221 * Update the record of the current time in a context.
223 static void update_context_time(struct perf_event_context *ctx)
225 u64 now = perf_clock();
227 ctx->time += now - ctx->timestamp;
228 ctx->timestamp = now;
232 * Update the total_time_enabled and total_time_running fields for a event.
234 static void update_event_times(struct perf_event *event)
236 struct perf_event_context *ctx = event->ctx;
239 if (event->state < PERF_EVENT_STATE_INACTIVE ||
240 event->group_leader->state < PERF_EVENT_STATE_INACTIVE)
246 run_end = event->tstamp_stopped;
248 event->total_time_enabled = run_end - event->tstamp_enabled;
250 if (event->state == PERF_EVENT_STATE_INACTIVE)
251 run_end = event->tstamp_stopped;
255 event->total_time_running = run_end - event->tstamp_running;
259 * Update total_time_enabled and total_time_running for all events in a group.
261 static void update_group_times(struct perf_event *leader)
263 struct perf_event *event;
265 update_event_times(leader);
266 list_for_each_entry(event, &leader->sibling_list, group_entry)
267 update_event_times(event);
270 static struct list_head *
271 ctx_group_list(struct perf_event *event, struct perf_event_context *ctx)
273 if (event->attr.pinned)
274 return &ctx->pinned_groups;
276 return &ctx->flexible_groups;
280 * Add a event from the lists for its context.
281 * Must be called with ctx->mutex and ctx->lock held.
284 list_add_event(struct perf_event *event, struct perf_event_context *ctx)
286 WARN_ON_ONCE(event->attach_state & PERF_ATTACH_CONTEXT);
287 event->attach_state |= PERF_ATTACH_CONTEXT;
290 * If we're a stand alone event or group leader, we go to the context
291 * list, group events are kept attached to the group so that
292 * perf_group_detach can, at all times, locate all siblings.
294 if (event->group_leader == event) {
295 struct list_head *list;
297 if (is_software_event(event))
298 event->group_flags |= PERF_GROUP_SOFTWARE;
300 list = ctx_group_list(event, ctx);
301 list_add_tail(&event->group_entry, list);
304 list_add_rcu(&event->event_entry, &ctx->event_list);
306 if (event->attr.inherit_stat)
310 static void perf_group_attach(struct perf_event *event)
312 struct perf_event *group_leader = event->group_leader;
314 WARN_ON_ONCE(event->attach_state & PERF_ATTACH_GROUP);
315 event->attach_state |= PERF_ATTACH_GROUP;
317 if (group_leader == event)
320 if (group_leader->group_flags & PERF_GROUP_SOFTWARE &&
321 !is_software_event(event))
322 group_leader->group_flags &= ~PERF_GROUP_SOFTWARE;
324 list_add_tail(&event->group_entry, &group_leader->sibling_list);
325 group_leader->nr_siblings++;
329 * Remove a event from the lists for its context.
330 * Must be called with ctx->mutex and ctx->lock held.
333 list_del_event(struct perf_event *event, struct perf_event_context *ctx)
336 * We can have double detach due to exit/hot-unplug + close.
338 if (!(event->attach_state & PERF_ATTACH_CONTEXT))
341 event->attach_state &= ~PERF_ATTACH_CONTEXT;
344 if (event->attr.inherit_stat)
347 list_del_rcu(&event->event_entry);
349 if (event->group_leader == event)
350 list_del_init(&event->group_entry);
352 update_group_times(event);
355 * If event was in error state, then keep it
356 * that way, otherwise bogus counts will be
357 * returned on read(). The only way to get out
358 * of error state is by explicit re-enabling
361 if (event->state > PERF_EVENT_STATE_OFF)
362 event->state = PERF_EVENT_STATE_OFF;
365 static void perf_group_detach(struct perf_event *event)
367 struct perf_event *sibling, *tmp;
368 struct list_head *list = NULL;
371 * We can have double detach due to exit/hot-unplug + close.
373 if (!(event->attach_state & PERF_ATTACH_GROUP))
376 event->attach_state &= ~PERF_ATTACH_GROUP;
379 * If this is a sibling, remove it from its group.
381 if (event->group_leader != event) {
382 list_del_init(&event->group_entry);
383 event->group_leader->nr_siblings--;
387 if (!list_empty(&event->group_entry))
388 list = &event->group_entry;
391 * If this was a group event with sibling events then
392 * upgrade the siblings to singleton events by adding them
393 * to whatever list we are on.
395 list_for_each_entry_safe(sibling, tmp, &event->sibling_list, group_entry) {
397 list_move_tail(&sibling->group_entry, list);
398 sibling->group_leader = sibling;
400 /* Inherit group flags from the previous leader */
401 sibling->group_flags = event->group_flags;
406 event_sched_out(struct perf_event *event,
407 struct perf_cpu_context *cpuctx,
408 struct perf_event_context *ctx)
410 if (event->state != PERF_EVENT_STATE_ACTIVE)
413 event->state = PERF_EVENT_STATE_INACTIVE;
414 if (event->pending_disable) {
415 event->pending_disable = 0;
416 event->state = PERF_EVENT_STATE_OFF;
418 event->tstamp_stopped = ctx->time;
419 event->pmu->disable(event);
422 if (!is_software_event(event))
423 cpuctx->active_oncpu--;
425 if (event->attr.exclusive || !cpuctx->active_oncpu)
426 cpuctx->exclusive = 0;
430 group_sched_out(struct perf_event *group_event,
431 struct perf_cpu_context *cpuctx,
432 struct perf_event_context *ctx)
434 struct perf_event *event;
436 if (group_event->state != PERF_EVENT_STATE_ACTIVE)
439 event_sched_out(group_event, cpuctx, ctx);
442 * Schedule out siblings (if any):
444 list_for_each_entry(event, &group_event->sibling_list, group_entry)
445 event_sched_out(event, cpuctx, ctx);
447 if (group_event->attr.exclusive)
448 cpuctx->exclusive = 0;
452 * Cross CPU call to remove a performance event
454 * We disable the event on the hardware level first. After that we
455 * remove it from the context list.
457 static void __perf_event_remove_from_context(void *info)
459 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
460 struct perf_event *event = info;
461 struct perf_event_context *ctx = event->ctx;
464 * If this is a task context, we need to check whether it is
465 * the current task context of this cpu. If not it has been
466 * scheduled out before the smp call arrived.
468 if (ctx->task && cpuctx->task_ctx != ctx)
471 raw_spin_lock(&ctx->lock);
473 * Protect the list operation against NMI by disabling the
474 * events on a global level.
478 event_sched_out(event, cpuctx, ctx);
480 list_del_event(event, ctx);
484 * Allow more per task events with respect to the
487 cpuctx->max_pertask =
488 min(perf_max_events - ctx->nr_events,
489 perf_max_events - perf_reserved_percpu);
493 raw_spin_unlock(&ctx->lock);
498 * Remove the event from a task's (or a CPU's) list of events.
500 * Must be called with ctx->mutex held.
502 * CPU events are removed with a smp call. For task events we only
503 * call when the task is on a CPU.
505 * If event->ctx is a cloned context, callers must make sure that
506 * every task struct that event->ctx->task could possibly point to
507 * remains valid. This is OK when called from perf_release since
508 * that only calls us on the top-level context, which can't be a clone.
509 * When called from perf_event_exit_task, it's OK because the
510 * context has been detached from its task.
512 static void perf_event_remove_from_context(struct perf_event *event)
514 struct perf_event_context *ctx = event->ctx;
515 struct task_struct *task = ctx->task;
519 * Per cpu events are removed via an smp call and
520 * the removal is always successful.
522 smp_call_function_single(event->cpu,
523 __perf_event_remove_from_context,
529 task_oncpu_function_call(task, __perf_event_remove_from_context,
532 raw_spin_lock_irq(&ctx->lock);
534 * If the context is active we need to retry the smp call.
536 if (ctx->nr_active && !list_empty(&event->group_entry)) {
537 raw_spin_unlock_irq(&ctx->lock);
542 * The lock prevents that this context is scheduled in so we
543 * can remove the event safely, if the call above did not
546 if (!list_empty(&event->group_entry))
547 list_del_event(event, ctx);
548 raw_spin_unlock_irq(&ctx->lock);
552 * Cross CPU call to disable a performance event
554 static void __perf_event_disable(void *info)
556 struct perf_event *event = info;
557 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
558 struct perf_event_context *ctx = event->ctx;
561 * If this is a per-task event, need to check whether this
562 * event's task is the current task on this cpu.
564 if (ctx->task && cpuctx->task_ctx != ctx)
567 raw_spin_lock(&ctx->lock);
570 * If the event is on, turn it off.
571 * If it is in error state, leave it in error state.
573 if (event->state >= PERF_EVENT_STATE_INACTIVE) {
574 update_context_time(ctx);
575 update_group_times(event);
576 if (event == event->group_leader)
577 group_sched_out(event, cpuctx, ctx);
579 event_sched_out(event, cpuctx, ctx);
580 event->state = PERF_EVENT_STATE_OFF;
583 raw_spin_unlock(&ctx->lock);
589 * If event->ctx is a cloned context, callers must make sure that
590 * every task struct that event->ctx->task could possibly point to
591 * remains valid. This condition is satisifed when called through
592 * perf_event_for_each_child or perf_event_for_each because they
593 * hold the top-level event's child_mutex, so any descendant that
594 * goes to exit will block in sync_child_event.
595 * When called from perf_pending_event it's OK because event->ctx
596 * is the current context on this CPU and preemption is disabled,
597 * hence we can't get into perf_event_task_sched_out for this context.
599 void perf_event_disable(struct perf_event *event)
601 struct perf_event_context *ctx = event->ctx;
602 struct task_struct *task = ctx->task;
606 * Disable the event on the cpu that it's on
608 smp_call_function_single(event->cpu, __perf_event_disable,
614 task_oncpu_function_call(task, __perf_event_disable, event);
616 raw_spin_lock_irq(&ctx->lock);
618 * If the event is still active, we need to retry the cross-call.
620 if (event->state == PERF_EVENT_STATE_ACTIVE) {
621 raw_spin_unlock_irq(&ctx->lock);
626 * Since we have the lock this context can't be scheduled
627 * in, so we can change the state safely.
629 if (event->state == PERF_EVENT_STATE_INACTIVE) {
630 update_group_times(event);
631 event->state = PERF_EVENT_STATE_OFF;
634 raw_spin_unlock_irq(&ctx->lock);
638 event_sched_in(struct perf_event *event,
639 struct perf_cpu_context *cpuctx,
640 struct perf_event_context *ctx)
642 if (event->state <= PERF_EVENT_STATE_OFF)
645 event->state = PERF_EVENT_STATE_ACTIVE;
646 event->oncpu = smp_processor_id();
648 * The new state must be visible before we turn it on in the hardware:
652 if (event->pmu->enable(event)) {
653 event->state = PERF_EVENT_STATE_INACTIVE;
658 event->tstamp_running += ctx->time - event->tstamp_stopped;
660 if (!is_software_event(event))
661 cpuctx->active_oncpu++;
664 if (event->attr.exclusive)
665 cpuctx->exclusive = 1;
671 group_sched_in(struct perf_event *group_event,
672 struct perf_cpu_context *cpuctx,
673 struct perf_event_context *ctx)
675 struct perf_event *event, *partial_group = NULL;
676 const struct pmu *pmu = group_event->pmu;
679 if (group_event->state == PERF_EVENT_STATE_OFF)
682 /* Check if group transaction availabe */
689 if (event_sched_in(group_event, cpuctx, ctx)) {
691 pmu->cancel_txn(pmu);
696 * Schedule in siblings as one group (if any):
698 list_for_each_entry(event, &group_event->sibling_list, group_entry) {
699 if (event_sched_in(event, cpuctx, ctx)) {
700 partial_group = event;
705 if (!txn || !pmu->commit_txn(pmu))
710 * Groups can be scheduled in as one unit only, so undo any
711 * partial group before returning:
713 list_for_each_entry(event, &group_event->sibling_list, group_entry) {
714 if (event == partial_group)
716 event_sched_out(event, cpuctx, ctx);
718 event_sched_out(group_event, cpuctx, ctx);
721 pmu->cancel_txn(pmu);
727 * Work out whether we can put this event group on the CPU now.
729 static int group_can_go_on(struct perf_event *event,
730 struct perf_cpu_context *cpuctx,
734 * Groups consisting entirely of software events can always go on.
736 if (event->group_flags & PERF_GROUP_SOFTWARE)
739 * If an exclusive group is already on, no other hardware
742 if (cpuctx->exclusive)
745 * If this group is exclusive and there are already
746 * events on the CPU, it can't go on.
748 if (event->attr.exclusive && cpuctx->active_oncpu)
751 * Otherwise, try to add it if all previous groups were able
757 static void add_event_to_ctx(struct perf_event *event,
758 struct perf_event_context *ctx)
760 list_add_event(event, ctx);
761 perf_group_attach(event);
762 event->tstamp_enabled = ctx->time;
763 event->tstamp_running = ctx->time;
764 event->tstamp_stopped = ctx->time;
768 * Cross CPU call to install and enable a performance event
770 * Must be called with ctx->mutex held
772 static void __perf_install_in_context(void *info)
774 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
775 struct perf_event *event = info;
776 struct perf_event_context *ctx = event->ctx;
777 struct perf_event *leader = event->group_leader;
781 * If this is a task context, we need to check whether it is
782 * the current task context of this cpu. If not it has been
783 * scheduled out before the smp call arrived.
784 * Or possibly this is the right context but it isn't
785 * on this cpu because it had no events.
787 if (ctx->task && cpuctx->task_ctx != ctx) {
788 if (cpuctx->task_ctx || ctx->task != current)
790 cpuctx->task_ctx = ctx;
793 raw_spin_lock(&ctx->lock);
795 update_context_time(ctx);
798 * Protect the list operation against NMI by disabling the
799 * events on a global level. NOP for non NMI based events.
803 add_event_to_ctx(event, ctx);
805 if (event->cpu != -1 && event->cpu != smp_processor_id())
809 * Don't put the event on if it is disabled or if
810 * it is in a group and the group isn't on.
812 if (event->state != PERF_EVENT_STATE_INACTIVE ||
813 (leader != event && leader->state != PERF_EVENT_STATE_ACTIVE))
817 * An exclusive event can't go on if there are already active
818 * hardware events, and no hardware event can go on if there
819 * is already an exclusive event on.
821 if (!group_can_go_on(event, cpuctx, 1))
824 err = event_sched_in(event, cpuctx, ctx);
828 * This event couldn't go on. If it is in a group
829 * then we have to pull the whole group off.
830 * If the event group is pinned then put it in error state.
833 group_sched_out(leader, cpuctx, ctx);
834 if (leader->attr.pinned) {
835 update_group_times(leader);
836 leader->state = PERF_EVENT_STATE_ERROR;
840 if (!err && !ctx->task && cpuctx->max_pertask)
841 cpuctx->max_pertask--;
846 raw_spin_unlock(&ctx->lock);
850 * Attach a performance event to a context
852 * First we add the event to the list with the hardware enable bit
853 * in event->hw_config cleared.
855 * If the event is attached to a task which is on a CPU we use a smp
856 * call to enable it in the task context. The task might have been
857 * scheduled away, but we check this in the smp call again.
859 * Must be called with ctx->mutex held.
862 perf_install_in_context(struct perf_event_context *ctx,
863 struct perf_event *event,
866 struct task_struct *task = ctx->task;
870 * Per cpu events are installed via an smp call and
871 * the install is always successful.
873 smp_call_function_single(cpu, __perf_install_in_context,
879 task_oncpu_function_call(task, __perf_install_in_context,
882 raw_spin_lock_irq(&ctx->lock);
884 * we need to retry the smp call.
886 if (ctx->is_active && list_empty(&event->group_entry)) {
887 raw_spin_unlock_irq(&ctx->lock);
892 * The lock prevents that this context is scheduled in so we
893 * can add the event safely, if it the call above did not
896 if (list_empty(&event->group_entry))
897 add_event_to_ctx(event, ctx);
898 raw_spin_unlock_irq(&ctx->lock);
902 * Put a event into inactive state and update time fields.
903 * Enabling the leader of a group effectively enables all
904 * the group members that aren't explicitly disabled, so we
905 * have to update their ->tstamp_enabled also.
906 * Note: this works for group members as well as group leaders
907 * since the non-leader members' sibling_lists will be empty.
909 static void __perf_event_mark_enabled(struct perf_event *event,
910 struct perf_event_context *ctx)
912 struct perf_event *sub;
914 event->state = PERF_EVENT_STATE_INACTIVE;
915 event->tstamp_enabled = ctx->time - event->total_time_enabled;
916 list_for_each_entry(sub, &event->sibling_list, group_entry)
917 if (sub->state >= PERF_EVENT_STATE_INACTIVE)
918 sub->tstamp_enabled =
919 ctx->time - sub->total_time_enabled;
923 * Cross CPU call to enable a performance event
925 static void __perf_event_enable(void *info)
927 struct perf_event *event = info;
928 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
929 struct perf_event_context *ctx = event->ctx;
930 struct perf_event *leader = event->group_leader;
934 * If this is a per-task event, need to check whether this
935 * event's task is the current task on this cpu.
937 if (ctx->task && cpuctx->task_ctx != ctx) {
938 if (cpuctx->task_ctx || ctx->task != current)
940 cpuctx->task_ctx = ctx;
943 raw_spin_lock(&ctx->lock);
945 update_context_time(ctx);
947 if (event->state >= PERF_EVENT_STATE_INACTIVE)
949 __perf_event_mark_enabled(event, ctx);
951 if (event->cpu != -1 && event->cpu != smp_processor_id())
955 * If the event is in a group and isn't the group leader,
956 * then don't put it on unless the group is on.
958 if (leader != event && leader->state != PERF_EVENT_STATE_ACTIVE)
961 if (!group_can_go_on(event, cpuctx, 1)) {
966 err = group_sched_in(event, cpuctx, ctx);
968 err = event_sched_in(event, cpuctx, ctx);
974 * If this event can't go on and it's part of a
975 * group, then the whole group has to come off.
978 group_sched_out(leader, cpuctx, ctx);
979 if (leader->attr.pinned) {
980 update_group_times(leader);
981 leader->state = PERF_EVENT_STATE_ERROR;
986 raw_spin_unlock(&ctx->lock);
992 * If event->ctx is a cloned context, callers must make sure that
993 * every task struct that event->ctx->task could possibly point to
994 * remains valid. This condition is satisfied when called through
995 * perf_event_for_each_child or perf_event_for_each as described
996 * for perf_event_disable.
998 void perf_event_enable(struct perf_event *event)
1000 struct perf_event_context *ctx = event->ctx;
1001 struct task_struct *task = ctx->task;
1005 * Enable the event on the cpu that it's on
1007 smp_call_function_single(event->cpu, __perf_event_enable,
1012 raw_spin_lock_irq(&ctx->lock);
1013 if (event->state >= PERF_EVENT_STATE_INACTIVE)
1017 * If the event is in error state, clear that first.
1018 * That way, if we see the event in error state below, we
1019 * know that it has gone back into error state, as distinct
1020 * from the task having been scheduled away before the
1021 * cross-call arrived.
1023 if (event->state == PERF_EVENT_STATE_ERROR)
1024 event->state = PERF_EVENT_STATE_OFF;
1027 raw_spin_unlock_irq(&ctx->lock);
1028 task_oncpu_function_call(task, __perf_event_enable, event);
1030 raw_spin_lock_irq(&ctx->lock);
1033 * If the context is active and the event is still off,
1034 * we need to retry the cross-call.
1036 if (ctx->is_active && event->state == PERF_EVENT_STATE_OFF)
1040 * Since we have the lock this context can't be scheduled
1041 * in, so we can change the state safely.
1043 if (event->state == PERF_EVENT_STATE_OFF)
1044 __perf_event_mark_enabled(event, ctx);
1047 raw_spin_unlock_irq(&ctx->lock);
1050 static int perf_event_refresh(struct perf_event *event, int refresh)
1053 * not supported on inherited events
1055 if (event->attr.inherit)
1058 atomic_add(refresh, &event->event_limit);
1059 perf_event_enable(event);
1065 EVENT_FLEXIBLE = 0x1,
1067 EVENT_ALL = EVENT_FLEXIBLE | EVENT_PINNED,
1070 static void ctx_sched_out(struct perf_event_context *ctx,
1071 struct perf_cpu_context *cpuctx,
1072 enum event_type_t event_type)
1074 struct perf_event *event;
1076 raw_spin_lock(&ctx->lock);
1078 if (likely(!ctx->nr_events))
1080 update_context_time(ctx);
1083 if (!ctx->nr_active)
1086 if (event_type & EVENT_PINNED)
1087 list_for_each_entry(event, &ctx->pinned_groups, group_entry)
1088 group_sched_out(event, cpuctx, ctx);
1090 if (event_type & EVENT_FLEXIBLE)
1091 list_for_each_entry(event, &ctx->flexible_groups, group_entry)
1092 group_sched_out(event, cpuctx, ctx);
1097 raw_spin_unlock(&ctx->lock);
1101 * Test whether two contexts are equivalent, i.e. whether they
1102 * have both been cloned from the same version of the same context
1103 * and they both have the same number of enabled events.
1104 * If the number of enabled events is the same, then the set
1105 * of enabled events should be the same, because these are both
1106 * inherited contexts, therefore we can't access individual events
1107 * in them directly with an fd; we can only enable/disable all
1108 * events via prctl, or enable/disable all events in a family
1109 * via ioctl, which will have the same effect on both contexts.
1111 static int context_equiv(struct perf_event_context *ctx1,
1112 struct perf_event_context *ctx2)
1114 return ctx1->parent_ctx && ctx1->parent_ctx == ctx2->parent_ctx
1115 && ctx1->parent_gen == ctx2->parent_gen
1116 && !ctx1->pin_count && !ctx2->pin_count;
1119 static void __perf_event_sync_stat(struct perf_event *event,
1120 struct perf_event *next_event)
1124 if (!event->attr.inherit_stat)
1128 * Update the event value, we cannot use perf_event_read()
1129 * because we're in the middle of a context switch and have IRQs
1130 * disabled, which upsets smp_call_function_single(), however
1131 * we know the event must be on the current CPU, therefore we
1132 * don't need to use it.
1134 switch (event->state) {
1135 case PERF_EVENT_STATE_ACTIVE:
1136 event->pmu->read(event);
1139 case PERF_EVENT_STATE_INACTIVE:
1140 update_event_times(event);
1148 * In order to keep per-task stats reliable we need to flip the event
1149 * values when we flip the contexts.
1151 value = local64_read(&next_event->count);
1152 value = local64_xchg(&event->count, value);
1153 local64_set(&next_event->count, value);
1155 swap(event->total_time_enabled, next_event->total_time_enabled);
1156 swap(event->total_time_running, next_event->total_time_running);
1159 * Since we swizzled the values, update the user visible data too.
1161 perf_event_update_userpage(event);
1162 perf_event_update_userpage(next_event);
1165 #define list_next_entry(pos, member) \
1166 list_entry(pos->member.next, typeof(*pos), member)
1168 static void perf_event_sync_stat(struct perf_event_context *ctx,
1169 struct perf_event_context *next_ctx)
1171 struct perf_event *event, *next_event;
1176 update_context_time(ctx);
1178 event = list_first_entry(&ctx->event_list,
1179 struct perf_event, event_entry);
1181 next_event = list_first_entry(&next_ctx->event_list,
1182 struct perf_event, event_entry);
1184 while (&event->event_entry != &ctx->event_list &&
1185 &next_event->event_entry != &next_ctx->event_list) {
1187 __perf_event_sync_stat(event, next_event);
1189 event = list_next_entry(event, event_entry);
1190 next_event = list_next_entry(next_event, event_entry);
1195 * Called from scheduler to remove the events of the current task,
1196 * with interrupts disabled.
1198 * We stop each event and update the event value in event->count.
1200 * This does not protect us against NMI, but disable()
1201 * sets the disabled bit in the control field of event _before_
1202 * accessing the event control register. If a NMI hits, then it will
1203 * not restart the event.
1205 void perf_event_task_sched_out(struct task_struct *task,
1206 struct task_struct *next)
1208 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
1209 struct perf_event_context *ctx = task->perf_event_ctxp;
1210 struct perf_event_context *next_ctx;
1211 struct perf_event_context *parent;
1214 perf_sw_event(PERF_COUNT_SW_CONTEXT_SWITCHES, 1, 1, NULL, 0);
1216 if (likely(!ctx || !cpuctx->task_ctx))
1220 parent = rcu_dereference(ctx->parent_ctx);
1221 next_ctx = next->perf_event_ctxp;
1222 if (parent && next_ctx &&
1223 rcu_dereference(next_ctx->parent_ctx) == parent) {
1225 * Looks like the two contexts are clones, so we might be
1226 * able to optimize the context switch. We lock both
1227 * contexts and check that they are clones under the
1228 * lock (including re-checking that neither has been
1229 * uncloned in the meantime). It doesn't matter which
1230 * order we take the locks because no other cpu could
1231 * be trying to lock both of these tasks.
1233 raw_spin_lock(&ctx->lock);
1234 raw_spin_lock_nested(&next_ctx->lock, SINGLE_DEPTH_NESTING);
1235 if (context_equiv(ctx, next_ctx)) {
1237 * XXX do we need a memory barrier of sorts
1238 * wrt to rcu_dereference() of perf_event_ctxp
1240 task->perf_event_ctxp = next_ctx;
1241 next->perf_event_ctxp = ctx;
1243 next_ctx->task = task;
1246 perf_event_sync_stat(ctx, next_ctx);
1248 raw_spin_unlock(&next_ctx->lock);
1249 raw_spin_unlock(&ctx->lock);
1254 ctx_sched_out(ctx, cpuctx, EVENT_ALL);
1255 cpuctx->task_ctx = NULL;
1259 static void task_ctx_sched_out(struct perf_event_context *ctx,
1260 enum event_type_t event_type)
1262 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
1264 if (!cpuctx->task_ctx)
1267 if (WARN_ON_ONCE(ctx != cpuctx->task_ctx))
1270 ctx_sched_out(ctx, cpuctx, event_type);
1271 cpuctx->task_ctx = NULL;
1275 * Called with IRQs disabled
1277 static void __perf_event_task_sched_out(struct perf_event_context *ctx)
1279 task_ctx_sched_out(ctx, EVENT_ALL);
1283 * Called with IRQs disabled
1285 static void cpu_ctx_sched_out(struct perf_cpu_context *cpuctx,
1286 enum event_type_t event_type)
1288 ctx_sched_out(&cpuctx->ctx, cpuctx, event_type);
1292 ctx_pinned_sched_in(struct perf_event_context *ctx,
1293 struct perf_cpu_context *cpuctx)
1295 struct perf_event *event;
1297 list_for_each_entry(event, &ctx->pinned_groups, group_entry) {
1298 if (event->state <= PERF_EVENT_STATE_OFF)
1300 if (event->cpu != -1 && event->cpu != smp_processor_id())
1303 if (group_can_go_on(event, cpuctx, 1))
1304 group_sched_in(event, cpuctx, ctx);
1307 * If this pinned group hasn't been scheduled,
1308 * put it in error state.
1310 if (event->state == PERF_EVENT_STATE_INACTIVE) {
1311 update_group_times(event);
1312 event->state = PERF_EVENT_STATE_ERROR;
1318 ctx_flexible_sched_in(struct perf_event_context *ctx,
1319 struct perf_cpu_context *cpuctx)
1321 struct perf_event *event;
1324 list_for_each_entry(event, &ctx->flexible_groups, group_entry) {
1325 /* Ignore events in OFF or ERROR state */
1326 if (event->state <= PERF_EVENT_STATE_OFF)
1329 * Listen to the 'cpu' scheduling filter constraint
1332 if (event->cpu != -1 && event->cpu != smp_processor_id())
1335 if (group_can_go_on(event, cpuctx, can_add_hw))
1336 if (group_sched_in(event, cpuctx, ctx))
1342 ctx_sched_in(struct perf_event_context *ctx,
1343 struct perf_cpu_context *cpuctx,
1344 enum event_type_t event_type)
1346 raw_spin_lock(&ctx->lock);
1348 if (likely(!ctx->nr_events))
1351 ctx->timestamp = perf_clock();
1356 * First go through the list and put on any pinned groups
1357 * in order to give them the best chance of going on.
1359 if (event_type & EVENT_PINNED)
1360 ctx_pinned_sched_in(ctx, cpuctx);
1362 /* Then walk through the lower prio flexible groups */
1363 if (event_type & EVENT_FLEXIBLE)
1364 ctx_flexible_sched_in(ctx, cpuctx);
1368 raw_spin_unlock(&ctx->lock);
1371 static void cpu_ctx_sched_in(struct perf_cpu_context *cpuctx,
1372 enum event_type_t event_type)
1374 struct perf_event_context *ctx = &cpuctx->ctx;
1376 ctx_sched_in(ctx, cpuctx, event_type);
1379 static void task_ctx_sched_in(struct task_struct *task,
1380 enum event_type_t event_type)
1382 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
1383 struct perf_event_context *ctx = task->perf_event_ctxp;
1387 if (cpuctx->task_ctx == ctx)
1389 ctx_sched_in(ctx, cpuctx, event_type);
1390 cpuctx->task_ctx = ctx;
1393 * Called from scheduler to add the events of the current task
1394 * with interrupts disabled.
1396 * We restore the event value and then enable it.
1398 * This does not protect us against NMI, but enable()
1399 * sets the enabled bit in the control field of event _before_
1400 * accessing the event control register. If a NMI hits, then it will
1401 * keep the event running.
1403 void perf_event_task_sched_in(struct task_struct *task)
1405 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
1406 struct perf_event_context *ctx = task->perf_event_ctxp;
1411 if (cpuctx->task_ctx == ctx)
1417 * We want to keep the following priority order:
1418 * cpu pinned (that don't need to move), task pinned,
1419 * cpu flexible, task flexible.
1421 cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE);
1423 ctx_sched_in(ctx, cpuctx, EVENT_PINNED);
1424 cpu_ctx_sched_in(cpuctx, EVENT_FLEXIBLE);
1425 ctx_sched_in(ctx, cpuctx, EVENT_FLEXIBLE);
1427 cpuctx->task_ctx = ctx;
1432 #define MAX_INTERRUPTS (~0ULL)
1434 static void perf_log_throttle(struct perf_event *event, int enable);
1436 static u64 perf_calculate_period(struct perf_event *event, u64 nsec, u64 count)
1438 u64 frequency = event->attr.sample_freq;
1439 u64 sec = NSEC_PER_SEC;
1440 u64 divisor, dividend;
1442 int count_fls, nsec_fls, frequency_fls, sec_fls;
1444 count_fls = fls64(count);
1445 nsec_fls = fls64(nsec);
1446 frequency_fls = fls64(frequency);
1450 * We got @count in @nsec, with a target of sample_freq HZ
1451 * the target period becomes:
1454 * period = -------------------
1455 * @nsec * sample_freq
1460 * Reduce accuracy by one bit such that @a and @b converge
1461 * to a similar magnitude.
1463 #define REDUCE_FLS(a, b) \
1465 if (a##_fls > b##_fls) { \
1475 * Reduce accuracy until either term fits in a u64, then proceed with
1476 * the other, so that finally we can do a u64/u64 division.
1478 while (count_fls + sec_fls > 64 && nsec_fls + frequency_fls > 64) {
1479 REDUCE_FLS(nsec, frequency);
1480 REDUCE_FLS(sec, count);
1483 if (count_fls + sec_fls > 64) {
1484 divisor = nsec * frequency;
1486 while (count_fls + sec_fls > 64) {
1487 REDUCE_FLS(count, sec);
1491 dividend = count * sec;
1493 dividend = count * sec;
1495 while (nsec_fls + frequency_fls > 64) {
1496 REDUCE_FLS(nsec, frequency);
1500 divisor = nsec * frequency;
1506 return div64_u64(dividend, divisor);
1509 static void perf_event_stop(struct perf_event *event)
1511 if (!event->pmu->stop)
1512 return event->pmu->disable(event);
1514 return event->pmu->stop(event);
1517 static int perf_event_start(struct perf_event *event)
1519 if (!event->pmu->start)
1520 return event->pmu->enable(event);
1522 return event->pmu->start(event);
1525 static void perf_adjust_period(struct perf_event *event, u64 nsec, u64 count)
1527 struct hw_perf_event *hwc = &event->hw;
1528 s64 period, sample_period;
1531 period = perf_calculate_period(event, nsec, count);
1533 delta = (s64)(period - hwc->sample_period);
1534 delta = (delta + 7) / 8; /* low pass filter */
1536 sample_period = hwc->sample_period + delta;
1541 hwc->sample_period = sample_period;
1543 if (local64_read(&hwc->period_left) > 8*sample_period) {
1545 perf_event_stop(event);
1546 local64_set(&hwc->period_left, 0);
1547 perf_event_start(event);
1552 static void perf_ctx_adjust_freq(struct perf_event_context *ctx)
1554 struct perf_event *event;
1555 struct hw_perf_event *hwc;
1556 u64 interrupts, now;
1559 raw_spin_lock(&ctx->lock);
1560 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
1561 if (event->state != PERF_EVENT_STATE_ACTIVE)
1564 if (event->cpu != -1 && event->cpu != smp_processor_id())
1569 interrupts = hwc->interrupts;
1570 hwc->interrupts = 0;
1573 * unthrottle events on the tick
1575 if (interrupts == MAX_INTERRUPTS) {
1576 perf_log_throttle(event, 1);
1578 event->pmu->unthrottle(event);
1582 if (!event->attr.freq || !event->attr.sample_freq)
1586 event->pmu->read(event);
1587 now = local64_read(&event->count);
1588 delta = now - hwc->freq_count_stamp;
1589 hwc->freq_count_stamp = now;
1592 perf_adjust_period(event, TICK_NSEC, delta);
1595 raw_spin_unlock(&ctx->lock);
1599 * Round-robin a context's events:
1601 static void rotate_ctx(struct perf_event_context *ctx)
1603 raw_spin_lock(&ctx->lock);
1605 /* Rotate the first entry last of non-pinned groups */
1606 list_rotate_left(&ctx->flexible_groups);
1608 raw_spin_unlock(&ctx->lock);
1611 void perf_event_task_tick(struct task_struct *curr)
1613 struct perf_cpu_context *cpuctx;
1614 struct perf_event_context *ctx;
1617 if (!atomic_read(&nr_events))
1620 cpuctx = &__get_cpu_var(perf_cpu_context);
1621 if (cpuctx->ctx.nr_events &&
1622 cpuctx->ctx.nr_events != cpuctx->ctx.nr_active)
1625 ctx = curr->perf_event_ctxp;
1626 if (ctx && ctx->nr_events && ctx->nr_events != ctx->nr_active)
1629 perf_ctx_adjust_freq(&cpuctx->ctx);
1631 perf_ctx_adjust_freq(ctx);
1637 cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE);
1639 task_ctx_sched_out(ctx, EVENT_FLEXIBLE);
1641 rotate_ctx(&cpuctx->ctx);
1645 cpu_ctx_sched_in(cpuctx, EVENT_FLEXIBLE);
1647 task_ctx_sched_in(curr, EVENT_FLEXIBLE);
1651 static int event_enable_on_exec(struct perf_event *event,
1652 struct perf_event_context *ctx)
1654 if (!event->attr.enable_on_exec)
1657 event->attr.enable_on_exec = 0;
1658 if (event->state >= PERF_EVENT_STATE_INACTIVE)
1661 __perf_event_mark_enabled(event, ctx);
1667 * Enable all of a task's events that have been marked enable-on-exec.
1668 * This expects task == current.
1670 static void perf_event_enable_on_exec(struct task_struct *task)
1672 struct perf_event_context *ctx;
1673 struct perf_event *event;
1674 unsigned long flags;
1678 local_irq_save(flags);
1679 ctx = task->perf_event_ctxp;
1680 if (!ctx || !ctx->nr_events)
1683 __perf_event_task_sched_out(ctx);
1685 raw_spin_lock(&ctx->lock);
1687 list_for_each_entry(event, &ctx->pinned_groups, group_entry) {
1688 ret = event_enable_on_exec(event, ctx);
1693 list_for_each_entry(event, &ctx->flexible_groups, group_entry) {
1694 ret = event_enable_on_exec(event, ctx);
1700 * Unclone this context if we enabled any event.
1705 raw_spin_unlock(&ctx->lock);
1707 perf_event_task_sched_in(task);
1709 local_irq_restore(flags);
1713 * Cross CPU call to read the hardware event
1715 static void __perf_event_read(void *info)
1717 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
1718 struct perf_event *event = info;
1719 struct perf_event_context *ctx = event->ctx;
1722 * If this is a task context, we need to check whether it is
1723 * the current task context of this cpu. If not it has been
1724 * scheduled out before the smp call arrived. In that case
1725 * event->count would have been updated to a recent sample
1726 * when the event was scheduled out.
1728 if (ctx->task && cpuctx->task_ctx != ctx)
1731 raw_spin_lock(&ctx->lock);
1732 update_context_time(ctx);
1733 update_event_times(event);
1734 raw_spin_unlock(&ctx->lock);
1736 event->pmu->read(event);
1739 static inline u64 perf_event_count(struct perf_event *event)
1741 return local64_read(&event->count) + atomic64_read(&event->child_count);
1744 static u64 perf_event_read(struct perf_event *event)
1747 * If event is enabled and currently active on a CPU, update the
1748 * value in the event structure:
1750 if (event->state == PERF_EVENT_STATE_ACTIVE) {
1751 smp_call_function_single(event->oncpu,
1752 __perf_event_read, event, 1);
1753 } else if (event->state == PERF_EVENT_STATE_INACTIVE) {
1754 struct perf_event_context *ctx = event->ctx;
1755 unsigned long flags;
1757 raw_spin_lock_irqsave(&ctx->lock, flags);
1758 update_context_time(ctx);
1759 update_event_times(event);
1760 raw_spin_unlock_irqrestore(&ctx->lock, flags);
1763 return perf_event_count(event);
1767 * Initialize the perf_event context in a task_struct:
1770 __perf_event_init_context(struct perf_event_context *ctx,
1771 struct task_struct *task)
1773 raw_spin_lock_init(&ctx->lock);
1774 mutex_init(&ctx->mutex);
1775 INIT_LIST_HEAD(&ctx->pinned_groups);
1776 INIT_LIST_HEAD(&ctx->flexible_groups);
1777 INIT_LIST_HEAD(&ctx->event_list);
1778 atomic_set(&ctx->refcount, 1);
1782 static struct perf_event_context *find_get_context(pid_t pid, int cpu)
1784 struct perf_event_context *ctx;
1785 struct perf_cpu_context *cpuctx;
1786 struct task_struct *task;
1787 unsigned long flags;
1790 if (pid == -1 && cpu != -1) {
1791 /* Must be root to operate on a CPU event: */
1792 if (perf_paranoid_cpu() && !capable(CAP_SYS_ADMIN))
1793 return ERR_PTR(-EACCES);
1795 if (cpu < 0 || cpu >= nr_cpumask_bits)
1796 return ERR_PTR(-EINVAL);
1799 * We could be clever and allow to attach a event to an
1800 * offline CPU and activate it when the CPU comes up, but
1803 if (!cpu_online(cpu))
1804 return ERR_PTR(-ENODEV);
1806 cpuctx = &per_cpu(perf_cpu_context, cpu);
1817 task = find_task_by_vpid(pid);
1819 get_task_struct(task);
1823 return ERR_PTR(-ESRCH);
1826 * Can't attach events to a dying task.
1829 if (task->flags & PF_EXITING)
1832 /* Reuse ptrace permission checks for now. */
1834 if (!ptrace_may_access(task, PTRACE_MODE_READ))
1838 ctx = perf_lock_task_context(task, &flags);
1841 raw_spin_unlock_irqrestore(&ctx->lock, flags);
1845 ctx = kzalloc(sizeof(struct perf_event_context), GFP_KERNEL);
1849 __perf_event_init_context(ctx, task);
1851 if (cmpxchg(&task->perf_event_ctxp, NULL, ctx)) {
1853 * We raced with some other task; use
1854 * the context they set.
1859 get_task_struct(task);
1862 put_task_struct(task);
1866 put_task_struct(task);
1867 return ERR_PTR(err);
1870 static void perf_event_free_filter(struct perf_event *event);
1872 static void free_event_rcu(struct rcu_head *head)
1874 struct perf_event *event;
1876 event = container_of(head, struct perf_event, rcu_head);
1878 put_pid_ns(event->ns);
1879 perf_event_free_filter(event);
1883 static void perf_pending_sync(struct perf_event *event);
1884 static void perf_buffer_put(struct perf_buffer *buffer);
1886 static void free_event(struct perf_event *event)
1888 perf_pending_sync(event);
1890 if (!event->parent) {
1891 atomic_dec(&nr_events);
1892 if (event->attr.mmap || event->attr.mmap_data)
1893 atomic_dec(&nr_mmap_events);
1894 if (event->attr.comm)
1895 atomic_dec(&nr_comm_events);
1896 if (event->attr.task)
1897 atomic_dec(&nr_task_events);
1900 if (event->buffer) {
1901 perf_buffer_put(event->buffer);
1902 event->buffer = NULL;
1906 event->destroy(event);
1908 put_ctx(event->ctx);
1909 call_rcu(&event->rcu_head, free_event_rcu);
1912 int perf_event_release_kernel(struct perf_event *event)
1914 struct perf_event_context *ctx = event->ctx;
1917 * Remove from the PMU, can't get re-enabled since we got
1918 * here because the last ref went.
1920 perf_event_disable(event);
1922 WARN_ON_ONCE(ctx->parent_ctx);
1924 * There are two ways this annotation is useful:
1926 * 1) there is a lock recursion from perf_event_exit_task
1927 * see the comment there.
1929 * 2) there is a lock-inversion with mmap_sem through
1930 * perf_event_read_group(), which takes faults while
1931 * holding ctx->mutex, however this is called after
1932 * the last filedesc died, so there is no possibility
1933 * to trigger the AB-BA case.
1935 mutex_lock_nested(&ctx->mutex, SINGLE_DEPTH_NESTING);
1936 raw_spin_lock_irq(&ctx->lock);
1937 perf_group_detach(event);
1938 list_del_event(event, ctx);
1939 raw_spin_unlock_irq(&ctx->lock);
1940 mutex_unlock(&ctx->mutex);
1942 mutex_lock(&event->owner->perf_event_mutex);
1943 list_del_init(&event->owner_entry);
1944 mutex_unlock(&event->owner->perf_event_mutex);
1945 put_task_struct(event->owner);
1951 EXPORT_SYMBOL_GPL(perf_event_release_kernel);
1954 * Called when the last reference to the file is gone.
1956 static int perf_release(struct inode *inode, struct file *file)
1958 struct perf_event *event = file->private_data;
1960 file->private_data = NULL;
1962 return perf_event_release_kernel(event);
1965 static int perf_event_read_size(struct perf_event *event)
1967 int entry = sizeof(u64); /* value */
1971 if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1972 size += sizeof(u64);
1974 if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1975 size += sizeof(u64);
1977 if (event->attr.read_format & PERF_FORMAT_ID)
1978 entry += sizeof(u64);
1980 if (event->attr.read_format & PERF_FORMAT_GROUP) {
1981 nr += event->group_leader->nr_siblings;
1982 size += sizeof(u64);
1990 u64 perf_event_read_value(struct perf_event *event, u64 *enabled, u64 *running)
1992 struct perf_event *child;
1998 mutex_lock(&event->child_mutex);
1999 total += perf_event_read(event);
2000 *enabled += event->total_time_enabled +
2001 atomic64_read(&event->child_total_time_enabled);
2002 *running += event->total_time_running +
2003 atomic64_read(&event->child_total_time_running);
2005 list_for_each_entry(child, &event->child_list, child_list) {
2006 total += perf_event_read(child);
2007 *enabled += child->total_time_enabled;
2008 *running += child->total_time_running;
2010 mutex_unlock(&event->child_mutex);
2014 EXPORT_SYMBOL_GPL(perf_event_read_value);
2016 static int perf_event_read_group(struct perf_event *event,
2017 u64 read_format, char __user *buf)
2019 struct perf_event *leader = event->group_leader, *sub;
2020 int n = 0, size = 0, ret = -EFAULT;
2021 struct perf_event_context *ctx = leader->ctx;
2023 u64 count, enabled, running;
2025 mutex_lock(&ctx->mutex);
2026 count = perf_event_read_value(leader, &enabled, &running);
2028 values[n++] = 1 + leader->nr_siblings;
2029 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
2030 values[n++] = enabled;
2031 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
2032 values[n++] = running;
2033 values[n++] = count;
2034 if (read_format & PERF_FORMAT_ID)
2035 values[n++] = primary_event_id(leader);
2037 size = n * sizeof(u64);
2039 if (copy_to_user(buf, values, size))
2044 list_for_each_entry(sub, &leader->sibling_list, group_entry) {
2047 values[n++] = perf_event_read_value(sub, &enabled, &running);
2048 if (read_format & PERF_FORMAT_ID)
2049 values[n++] = primary_event_id(sub);
2051 size = n * sizeof(u64);
2053 if (copy_to_user(buf + ret, values, size)) {
2061 mutex_unlock(&ctx->mutex);
2066 static int perf_event_read_one(struct perf_event *event,
2067 u64 read_format, char __user *buf)
2069 u64 enabled, running;
2073 values[n++] = perf_event_read_value(event, &enabled, &running);
2074 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
2075 values[n++] = enabled;
2076 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
2077 values[n++] = running;
2078 if (read_format & PERF_FORMAT_ID)
2079 values[n++] = primary_event_id(event);
2081 if (copy_to_user(buf, values, n * sizeof(u64)))
2084 return n * sizeof(u64);
2088 * Read the performance event - simple non blocking version for now
2091 perf_read_hw(struct perf_event *event, char __user *buf, size_t count)
2093 u64 read_format = event->attr.read_format;
2097 * Return end-of-file for a read on a event that is in
2098 * error state (i.e. because it was pinned but it couldn't be
2099 * scheduled on to the CPU at some point).
2101 if (event->state == PERF_EVENT_STATE_ERROR)
2104 if (count < perf_event_read_size(event))
2107 WARN_ON_ONCE(event->ctx->parent_ctx);
2108 if (read_format & PERF_FORMAT_GROUP)
2109 ret = perf_event_read_group(event, read_format, buf);
2111 ret = perf_event_read_one(event, read_format, buf);
2117 perf_read(struct file *file, char __user *buf, size_t count, loff_t *ppos)
2119 struct perf_event *event = file->private_data;
2121 return perf_read_hw(event, buf, count);
2124 static unsigned int perf_poll(struct file *file, poll_table *wait)
2126 struct perf_event *event = file->private_data;
2127 struct perf_buffer *buffer;
2128 unsigned int events = POLL_HUP;
2131 buffer = rcu_dereference(event->buffer);
2133 events = atomic_xchg(&buffer->poll, 0);
2136 poll_wait(file, &event->waitq, wait);
2141 static void perf_event_reset(struct perf_event *event)
2143 (void)perf_event_read(event);
2144 local64_set(&event->count, 0);
2145 perf_event_update_userpage(event);
2149 * Holding the top-level event's child_mutex means that any
2150 * descendant process that has inherited this event will block
2151 * in sync_child_event if it goes to exit, thus satisfying the
2152 * task existence requirements of perf_event_enable/disable.
2154 static void perf_event_for_each_child(struct perf_event *event,
2155 void (*func)(struct perf_event *))
2157 struct perf_event *child;
2159 WARN_ON_ONCE(event->ctx->parent_ctx);
2160 mutex_lock(&event->child_mutex);
2162 list_for_each_entry(child, &event->child_list, child_list)
2164 mutex_unlock(&event->child_mutex);
2167 static void perf_event_for_each(struct perf_event *event,
2168 void (*func)(struct perf_event *))
2170 struct perf_event_context *ctx = event->ctx;
2171 struct perf_event *sibling;
2173 WARN_ON_ONCE(ctx->parent_ctx);
2174 mutex_lock(&ctx->mutex);
2175 event = event->group_leader;
2177 perf_event_for_each_child(event, func);
2179 list_for_each_entry(sibling, &event->sibling_list, group_entry)
2180 perf_event_for_each_child(event, func);
2181 mutex_unlock(&ctx->mutex);
2184 static int perf_event_period(struct perf_event *event, u64 __user *arg)
2186 struct perf_event_context *ctx = event->ctx;
2191 if (!event->attr.sample_period)
2194 size = copy_from_user(&value, arg, sizeof(value));
2195 if (size != sizeof(value))
2201 raw_spin_lock_irq(&ctx->lock);
2202 if (event->attr.freq) {
2203 if (value > sysctl_perf_event_sample_rate) {
2208 event->attr.sample_freq = value;
2210 event->attr.sample_period = value;
2211 event->hw.sample_period = value;
2214 raw_spin_unlock_irq(&ctx->lock);
2219 static const struct file_operations perf_fops;
2221 static struct perf_event *perf_fget_light(int fd, int *fput_needed)
2225 file = fget_light(fd, fput_needed);
2227 return ERR_PTR(-EBADF);
2229 if (file->f_op != &perf_fops) {
2230 fput_light(file, *fput_needed);
2232 return ERR_PTR(-EBADF);
2235 return file->private_data;
2238 static int perf_event_set_output(struct perf_event *event,
2239 struct perf_event *output_event);
2240 static int perf_event_set_filter(struct perf_event *event, void __user *arg);
2242 static long perf_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
2244 struct perf_event *event = file->private_data;
2245 void (*func)(struct perf_event *);
2249 case PERF_EVENT_IOC_ENABLE:
2250 func = perf_event_enable;
2252 case PERF_EVENT_IOC_DISABLE:
2253 func = perf_event_disable;
2255 case PERF_EVENT_IOC_RESET:
2256 func = perf_event_reset;
2259 case PERF_EVENT_IOC_REFRESH:
2260 return perf_event_refresh(event, arg);
2262 case PERF_EVENT_IOC_PERIOD:
2263 return perf_event_period(event, (u64 __user *)arg);
2265 case PERF_EVENT_IOC_SET_OUTPUT:
2267 struct perf_event *output_event = NULL;
2268 int fput_needed = 0;
2272 output_event = perf_fget_light(arg, &fput_needed);
2273 if (IS_ERR(output_event))
2274 return PTR_ERR(output_event);
2277 ret = perf_event_set_output(event, output_event);
2279 fput_light(output_event->filp, fput_needed);
2284 case PERF_EVENT_IOC_SET_FILTER:
2285 return perf_event_set_filter(event, (void __user *)arg);
2291 if (flags & PERF_IOC_FLAG_GROUP)
2292 perf_event_for_each(event, func);
2294 perf_event_for_each_child(event, func);
2299 int perf_event_task_enable(void)
2301 struct perf_event *event;
2303 mutex_lock(¤t->perf_event_mutex);
2304 list_for_each_entry(event, ¤t->perf_event_list, owner_entry)
2305 perf_event_for_each_child(event, perf_event_enable);
2306 mutex_unlock(¤t->perf_event_mutex);
2311 int perf_event_task_disable(void)
2313 struct perf_event *event;
2315 mutex_lock(¤t->perf_event_mutex);
2316 list_for_each_entry(event, ¤t->perf_event_list, owner_entry)
2317 perf_event_for_each_child(event, perf_event_disable);
2318 mutex_unlock(¤t->perf_event_mutex);
2323 #ifndef PERF_EVENT_INDEX_OFFSET
2324 # define PERF_EVENT_INDEX_OFFSET 0
2327 static int perf_event_index(struct perf_event *event)
2329 if (event->state != PERF_EVENT_STATE_ACTIVE)
2332 return event->hw.idx + 1 - PERF_EVENT_INDEX_OFFSET;
2336 * Callers need to ensure there can be no nesting of this function, otherwise
2337 * the seqlock logic goes bad. We can not serialize this because the arch
2338 * code calls this from NMI context.
2340 void perf_event_update_userpage(struct perf_event *event)
2342 struct perf_event_mmap_page *userpg;
2343 struct perf_buffer *buffer;
2346 buffer = rcu_dereference(event->buffer);
2350 userpg = buffer->user_page;
2353 * Disable preemption so as to not let the corresponding user-space
2354 * spin too long if we get preempted.
2359 userpg->index = perf_event_index(event);
2360 userpg->offset = perf_event_count(event);
2361 if (event->state == PERF_EVENT_STATE_ACTIVE)
2362 userpg->offset -= local64_read(&event->hw.prev_count);
2364 userpg->time_enabled = event->total_time_enabled +
2365 atomic64_read(&event->child_total_time_enabled);
2367 userpg->time_running = event->total_time_running +
2368 atomic64_read(&event->child_total_time_running);
2377 static unsigned long perf_data_size(struct perf_buffer *buffer);
2380 perf_buffer_init(struct perf_buffer *buffer, long watermark, int flags)
2382 long max_size = perf_data_size(buffer);
2385 buffer->watermark = min(max_size, watermark);
2387 if (!buffer->watermark)
2388 buffer->watermark = max_size / 2;
2390 if (flags & PERF_BUFFER_WRITABLE)
2391 buffer->writable = 1;
2393 atomic_set(&buffer->refcount, 1);
2396 #ifndef CONFIG_PERF_USE_VMALLOC
2399 * Back perf_mmap() with regular GFP_KERNEL-0 pages.
2402 static struct page *
2403 perf_mmap_to_page(struct perf_buffer *buffer, unsigned long pgoff)
2405 if (pgoff > buffer->nr_pages)
2409 return virt_to_page(buffer->user_page);
2411 return virt_to_page(buffer->data_pages[pgoff - 1]);
2414 static void *perf_mmap_alloc_page(int cpu)
2419 node = (cpu == -1) ? cpu : cpu_to_node(cpu);
2420 page = alloc_pages_node(node, GFP_KERNEL | __GFP_ZERO, 0);
2424 return page_address(page);
2427 static struct perf_buffer *
2428 perf_buffer_alloc(int nr_pages, long watermark, int cpu, int flags)
2430 struct perf_buffer *buffer;
2434 size = sizeof(struct perf_buffer);
2435 size += nr_pages * sizeof(void *);
2437 buffer = kzalloc(size, GFP_KERNEL);
2441 buffer->user_page = perf_mmap_alloc_page(cpu);
2442 if (!buffer->user_page)
2443 goto fail_user_page;
2445 for (i = 0; i < nr_pages; i++) {
2446 buffer->data_pages[i] = perf_mmap_alloc_page(cpu);
2447 if (!buffer->data_pages[i])
2448 goto fail_data_pages;
2451 buffer->nr_pages = nr_pages;
2453 perf_buffer_init(buffer, watermark, flags);
2458 for (i--; i >= 0; i--)
2459 free_page((unsigned long)buffer->data_pages[i]);
2461 free_page((unsigned long)buffer->user_page);
2470 static void perf_mmap_free_page(unsigned long addr)
2472 struct page *page = virt_to_page((void *)addr);
2474 page->mapping = NULL;
2478 static void perf_buffer_free(struct perf_buffer *buffer)
2482 perf_mmap_free_page((unsigned long)buffer->user_page);
2483 for (i = 0; i < buffer->nr_pages; i++)
2484 perf_mmap_free_page((unsigned long)buffer->data_pages[i]);
2488 static inline int page_order(struct perf_buffer *buffer)
2496 * Back perf_mmap() with vmalloc memory.
2498 * Required for architectures that have d-cache aliasing issues.
2501 static inline int page_order(struct perf_buffer *buffer)
2503 return buffer->page_order;
2506 static struct page *
2507 perf_mmap_to_page(struct perf_buffer *buffer, unsigned long pgoff)
2509 if (pgoff > (1UL << page_order(buffer)))
2512 return vmalloc_to_page((void *)buffer->user_page + pgoff * PAGE_SIZE);
2515 static void perf_mmap_unmark_page(void *addr)
2517 struct page *page = vmalloc_to_page(addr);
2519 page->mapping = NULL;
2522 static void perf_buffer_free_work(struct work_struct *work)
2524 struct perf_buffer *buffer;
2528 buffer = container_of(work, struct perf_buffer, work);
2529 nr = 1 << page_order(buffer);
2531 base = buffer->user_page;
2532 for (i = 0; i < nr + 1; i++)
2533 perf_mmap_unmark_page(base + (i * PAGE_SIZE));
2539 static void perf_buffer_free(struct perf_buffer *buffer)
2541 schedule_work(&buffer->work);
2544 static struct perf_buffer *
2545 perf_buffer_alloc(int nr_pages, long watermark, int cpu, int flags)
2547 struct perf_buffer *buffer;
2551 size = sizeof(struct perf_buffer);
2552 size += sizeof(void *);
2554 buffer = kzalloc(size, GFP_KERNEL);
2558 INIT_WORK(&buffer->work, perf_buffer_free_work);
2560 all_buf = vmalloc_user((nr_pages + 1) * PAGE_SIZE);
2564 buffer->user_page = all_buf;
2565 buffer->data_pages[0] = all_buf + PAGE_SIZE;
2566 buffer->page_order = ilog2(nr_pages);
2567 buffer->nr_pages = 1;
2569 perf_buffer_init(buffer, watermark, flags);
2582 static unsigned long perf_data_size(struct perf_buffer *buffer)
2584 return buffer->nr_pages << (PAGE_SHIFT + page_order(buffer));
2587 static int perf_mmap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
2589 struct perf_event *event = vma->vm_file->private_data;
2590 struct perf_buffer *buffer;
2591 int ret = VM_FAULT_SIGBUS;
2593 if (vmf->flags & FAULT_FLAG_MKWRITE) {
2594 if (vmf->pgoff == 0)
2600 buffer = rcu_dereference(event->buffer);
2604 if (vmf->pgoff && (vmf->flags & FAULT_FLAG_WRITE))
2607 vmf->page = perf_mmap_to_page(buffer, vmf->pgoff);
2611 get_page(vmf->page);
2612 vmf->page->mapping = vma->vm_file->f_mapping;
2613 vmf->page->index = vmf->pgoff;
2622 static void perf_buffer_free_rcu(struct rcu_head *rcu_head)
2624 struct perf_buffer *buffer;
2626 buffer = container_of(rcu_head, struct perf_buffer, rcu_head);
2627 perf_buffer_free(buffer);
2630 static struct perf_buffer *perf_buffer_get(struct perf_event *event)
2632 struct perf_buffer *buffer;
2635 buffer = rcu_dereference(event->buffer);
2637 if (!atomic_inc_not_zero(&buffer->refcount))
2645 static void perf_buffer_put(struct perf_buffer *buffer)
2647 if (!atomic_dec_and_test(&buffer->refcount))
2650 call_rcu(&buffer->rcu_head, perf_buffer_free_rcu);
2653 static void perf_mmap_open(struct vm_area_struct *vma)
2655 struct perf_event *event = vma->vm_file->private_data;
2657 atomic_inc(&event->mmap_count);
2660 static void perf_mmap_close(struct vm_area_struct *vma)
2662 struct perf_event *event = vma->vm_file->private_data;
2664 if (atomic_dec_and_mutex_lock(&event->mmap_count, &event->mmap_mutex)) {
2665 unsigned long size = perf_data_size(event->buffer);
2666 struct user_struct *user = event->mmap_user;
2667 struct perf_buffer *buffer = event->buffer;
2669 atomic_long_sub((size >> PAGE_SHIFT) + 1, &user->locked_vm);
2670 vma->vm_mm->locked_vm -= event->mmap_locked;
2671 rcu_assign_pointer(event->buffer, NULL);
2672 mutex_unlock(&event->mmap_mutex);
2674 perf_buffer_put(buffer);
2679 static const struct vm_operations_struct perf_mmap_vmops = {
2680 .open = perf_mmap_open,
2681 .close = perf_mmap_close,
2682 .fault = perf_mmap_fault,
2683 .page_mkwrite = perf_mmap_fault,
2686 static int perf_mmap(struct file *file, struct vm_area_struct *vma)
2688 struct perf_event *event = file->private_data;
2689 unsigned long user_locked, user_lock_limit;
2690 struct user_struct *user = current_user();
2691 unsigned long locked, lock_limit;
2692 struct perf_buffer *buffer;
2693 unsigned long vma_size;
2694 unsigned long nr_pages;
2695 long user_extra, extra;
2696 int ret = 0, flags = 0;
2699 * Don't allow mmap() of inherited per-task counters. This would
2700 * create a performance issue due to all children writing to the
2703 if (event->cpu == -1 && event->attr.inherit)
2706 if (!(vma->vm_flags & VM_SHARED))
2709 vma_size = vma->vm_end - vma->vm_start;
2710 nr_pages = (vma_size / PAGE_SIZE) - 1;
2713 * If we have buffer pages ensure they're a power-of-two number, so we
2714 * can do bitmasks instead of modulo.
2716 if (nr_pages != 0 && !is_power_of_2(nr_pages))
2719 if (vma_size != PAGE_SIZE * (1 + nr_pages))
2722 if (vma->vm_pgoff != 0)
2725 WARN_ON_ONCE(event->ctx->parent_ctx);
2726 mutex_lock(&event->mmap_mutex);
2727 if (event->buffer) {
2728 if (event->buffer->nr_pages == nr_pages)
2729 atomic_inc(&event->buffer->refcount);
2735 user_extra = nr_pages + 1;
2736 user_lock_limit = sysctl_perf_event_mlock >> (PAGE_SHIFT - 10);
2739 * Increase the limit linearly with more CPUs:
2741 user_lock_limit *= num_online_cpus();
2743 user_locked = atomic_long_read(&user->locked_vm) + user_extra;
2746 if (user_locked > user_lock_limit)
2747 extra = user_locked - user_lock_limit;
2749 lock_limit = rlimit(RLIMIT_MEMLOCK);
2750 lock_limit >>= PAGE_SHIFT;
2751 locked = vma->vm_mm->locked_vm + extra;
2753 if ((locked > lock_limit) && perf_paranoid_tracepoint_raw() &&
2754 !capable(CAP_IPC_LOCK)) {
2759 WARN_ON(event->buffer);
2761 if (vma->vm_flags & VM_WRITE)
2762 flags |= PERF_BUFFER_WRITABLE;
2764 buffer = perf_buffer_alloc(nr_pages, event->attr.wakeup_watermark,
2770 rcu_assign_pointer(event->buffer, buffer);
2772 atomic_long_add(user_extra, &user->locked_vm);
2773 event->mmap_locked = extra;
2774 event->mmap_user = get_current_user();
2775 vma->vm_mm->locked_vm += event->mmap_locked;
2779 atomic_inc(&event->mmap_count);
2780 mutex_unlock(&event->mmap_mutex);
2782 vma->vm_flags |= VM_RESERVED;
2783 vma->vm_ops = &perf_mmap_vmops;
2788 static int perf_fasync(int fd, struct file *filp, int on)
2790 struct inode *inode = filp->f_path.dentry->d_inode;
2791 struct perf_event *event = filp->private_data;
2794 mutex_lock(&inode->i_mutex);
2795 retval = fasync_helper(fd, filp, on, &event->fasync);
2796 mutex_unlock(&inode->i_mutex);
2804 static const struct file_operations perf_fops = {
2805 .llseek = no_llseek,
2806 .release = perf_release,
2809 .unlocked_ioctl = perf_ioctl,
2810 .compat_ioctl = perf_ioctl,
2812 .fasync = perf_fasync,
2818 * If there's data, ensure we set the poll() state and publish everything
2819 * to user-space before waking everybody up.
2822 void perf_event_wakeup(struct perf_event *event)
2824 wake_up_all(&event->waitq);
2826 if (event->pending_kill) {
2827 kill_fasync(&event->fasync, SIGIO, event->pending_kill);
2828 event->pending_kill = 0;
2835 * Handle the case where we need to wakeup up from NMI (or rq->lock) context.
2837 * The NMI bit means we cannot possibly take locks. Therefore, maintain a
2838 * single linked list and use cmpxchg() to add entries lockless.
2841 static void perf_pending_event(struct perf_pending_entry *entry)
2843 struct perf_event *event = container_of(entry,
2844 struct perf_event, pending);
2846 if (event->pending_disable) {
2847 event->pending_disable = 0;
2848 __perf_event_disable(event);
2851 if (event->pending_wakeup) {
2852 event->pending_wakeup = 0;
2853 perf_event_wakeup(event);
2857 #define PENDING_TAIL ((struct perf_pending_entry *)-1UL)
2859 static DEFINE_PER_CPU(struct perf_pending_entry *, perf_pending_head) = {
2863 static void perf_pending_queue(struct perf_pending_entry *entry,
2864 void (*func)(struct perf_pending_entry *))
2866 struct perf_pending_entry **head;
2868 if (cmpxchg(&entry->next, NULL, PENDING_TAIL) != NULL)
2873 head = &get_cpu_var(perf_pending_head);
2876 entry->next = *head;
2877 } while (cmpxchg(head, entry->next, entry) != entry->next);
2879 set_perf_event_pending();
2881 put_cpu_var(perf_pending_head);
2884 static int __perf_pending_run(void)
2886 struct perf_pending_entry *list;
2889 list = xchg(&__get_cpu_var(perf_pending_head), PENDING_TAIL);
2890 while (list != PENDING_TAIL) {
2891 void (*func)(struct perf_pending_entry *);
2892 struct perf_pending_entry *entry = list;
2899 * Ensure we observe the unqueue before we issue the wakeup,
2900 * so that we won't be waiting forever.
2901 * -- see perf_not_pending().
2912 static inline int perf_not_pending(struct perf_event *event)
2915 * If we flush on whatever cpu we run, there is a chance we don't
2919 __perf_pending_run();
2923 * Ensure we see the proper queue state before going to sleep
2924 * so that we do not miss the wakeup. -- see perf_pending_handle()
2927 return event->pending.next == NULL;
2930 static void perf_pending_sync(struct perf_event *event)
2932 wait_event(event->waitq, perf_not_pending(event));
2935 void perf_event_do_pending(void)
2937 __perf_pending_run();
2940 DEFINE_PER_CPU(struct perf_callchain_entry, perf_callchain_entry);
2943 * Callchain support -- arch specific
2946 __weak struct perf_callchain_entry *perf_callchain_buffer(void)
2948 return &__get_cpu_var(perf_callchain_entry);
2951 __weak void perf_callchain_kernel(struct perf_callchain_entry *entry,
2952 struct pt_regs *regs)
2956 __weak void perf_callchain_user(struct perf_callchain_entry *entry,
2957 struct pt_regs *regs)
2961 static struct perf_callchain_entry *perf_callchain(struct pt_regs *regs)
2963 struct perf_callchain_entry *entry;
2965 entry = perf_callchain_buffer();
2971 if (!user_mode(regs)) {
2972 perf_callchain_kernel(entry, regs);
2974 regs = task_pt_regs(current);
2980 perf_callchain_user(entry, regs);
2987 * We assume there is only KVM supporting the callbacks.
2988 * Later on, we might change it to a list if there is
2989 * another virtualization implementation supporting the callbacks.
2991 struct perf_guest_info_callbacks *perf_guest_cbs;
2993 int perf_register_guest_info_callbacks(struct perf_guest_info_callbacks *cbs)
2995 perf_guest_cbs = cbs;
2998 EXPORT_SYMBOL_GPL(perf_register_guest_info_callbacks);
3000 int perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks *cbs)
3002 perf_guest_cbs = NULL;
3005 EXPORT_SYMBOL_GPL(perf_unregister_guest_info_callbacks);
3010 static bool perf_output_space(struct perf_buffer *buffer, unsigned long tail,
3011 unsigned long offset, unsigned long head)
3015 if (!buffer->writable)
3018 mask = perf_data_size(buffer) - 1;
3020 offset = (offset - tail) & mask;
3021 head = (head - tail) & mask;
3023 if ((int)(head - offset) < 0)
3029 static void perf_output_wakeup(struct perf_output_handle *handle)
3031 atomic_set(&handle->buffer->poll, POLL_IN);
3034 handle->event->pending_wakeup = 1;
3035 perf_pending_queue(&handle->event->pending,
3036 perf_pending_event);
3038 perf_event_wakeup(handle->event);
3042 * We need to ensure a later event_id doesn't publish a head when a former
3043 * event isn't done writing. However since we need to deal with NMIs we
3044 * cannot fully serialize things.
3046 * We only publish the head (and generate a wakeup) when the outer-most
3049 static void perf_output_get_handle(struct perf_output_handle *handle)
3051 struct perf_buffer *buffer = handle->buffer;
3054 local_inc(&buffer->nest);
3055 handle->wakeup = local_read(&buffer->wakeup);
3058 static void perf_output_put_handle(struct perf_output_handle *handle)
3060 struct perf_buffer *buffer = handle->buffer;
3064 head = local_read(&buffer->head);
3067 * IRQ/NMI can happen here, which means we can miss a head update.
3070 if (!local_dec_and_test(&buffer->nest))
3074 * Publish the known good head. Rely on the full barrier implied
3075 * by atomic_dec_and_test() order the buffer->head read and this
3078 buffer->user_page->data_head = head;
3081 * Now check if we missed an update, rely on the (compiler)
3082 * barrier in atomic_dec_and_test() to re-read buffer->head.
3084 if (unlikely(head != local_read(&buffer->head))) {
3085 local_inc(&buffer->nest);
3089 if (handle->wakeup != local_read(&buffer->wakeup))
3090 perf_output_wakeup(handle);
3096 __always_inline void perf_output_copy(struct perf_output_handle *handle,
3097 const void *buf, unsigned int len)
3100 unsigned long size = min_t(unsigned long, handle->size, len);
3102 memcpy(handle->addr, buf, size);
3105 handle->addr += size;
3107 handle->size -= size;
3108 if (!handle->size) {
3109 struct perf_buffer *buffer = handle->buffer;
3112 handle->page &= buffer->nr_pages - 1;
3113 handle->addr = buffer->data_pages[handle->page];
3114 handle->size = PAGE_SIZE << page_order(buffer);
3119 int perf_output_begin(struct perf_output_handle *handle,
3120 struct perf_event *event, unsigned int size,
3121 int nmi, int sample)
3123 struct perf_buffer *buffer;
3124 unsigned long tail, offset, head;
3127 struct perf_event_header header;
3134 * For inherited events we send all the output towards the parent.
3137 event = event->parent;
3139 buffer = rcu_dereference(event->buffer);
3143 handle->buffer = buffer;
3144 handle->event = event;
3146 handle->sample = sample;
3148 if (!buffer->nr_pages)
3151 have_lost = local_read(&buffer->lost);
3153 size += sizeof(lost_event);
3155 perf_output_get_handle(handle);
3159 * Userspace could choose to issue a mb() before updating the
3160 * tail pointer. So that all reads will be completed before the
3163 tail = ACCESS_ONCE(buffer->user_page->data_tail);
3165 offset = head = local_read(&buffer->head);
3167 if (unlikely(!perf_output_space(buffer, tail, offset, head)))
3169 } while (local_cmpxchg(&buffer->head, offset, head) != offset);
3171 if (head - local_read(&buffer->wakeup) > buffer->watermark)
3172 local_add(buffer->watermark, &buffer->wakeup);
3174 handle->page = offset >> (PAGE_SHIFT + page_order(buffer));
3175 handle->page &= buffer->nr_pages - 1;
3176 handle->size = offset & ((PAGE_SIZE << page_order(buffer)) - 1);
3177 handle->addr = buffer->data_pages[handle->page];
3178 handle->addr += handle->size;
3179 handle->size = (PAGE_SIZE << page_order(buffer)) - handle->size;
3182 lost_event.header.type = PERF_RECORD_LOST;
3183 lost_event.header.misc = 0;
3184 lost_event.header.size = sizeof(lost_event);
3185 lost_event.id = event->id;
3186 lost_event.lost = local_xchg(&buffer->lost, 0);
3188 perf_output_put(handle, lost_event);
3194 local_inc(&buffer->lost);
3195 perf_output_put_handle(handle);
3202 void perf_output_end(struct perf_output_handle *handle)
3204 struct perf_event *event = handle->event;
3205 struct perf_buffer *buffer = handle->buffer;
3207 int wakeup_events = event->attr.wakeup_events;
3209 if (handle->sample && wakeup_events) {
3210 int events = local_inc_return(&buffer->events);
3211 if (events >= wakeup_events) {
3212 local_sub(wakeup_events, &buffer->events);
3213 local_inc(&buffer->wakeup);
3217 perf_output_put_handle(handle);
3221 static u32 perf_event_pid(struct perf_event *event, struct task_struct *p)
3224 * only top level events have the pid namespace they were created in
3227 event = event->parent;
3229 return task_tgid_nr_ns(p, event->ns);
3232 static u32 perf_event_tid(struct perf_event *event, struct task_struct *p)
3235 * only top level events have the pid namespace they were created in
3238 event = event->parent;
3240 return task_pid_nr_ns(p, event->ns);
3243 static void perf_output_read_one(struct perf_output_handle *handle,
3244 struct perf_event *event)
3246 u64 read_format = event->attr.read_format;
3250 values[n++] = perf_event_count(event);
3251 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
3252 values[n++] = event->total_time_enabled +
3253 atomic64_read(&event->child_total_time_enabled);
3255 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
3256 values[n++] = event->total_time_running +
3257 atomic64_read(&event->child_total_time_running);
3259 if (read_format & PERF_FORMAT_ID)
3260 values[n++] = primary_event_id(event);
3262 perf_output_copy(handle, values, n * sizeof(u64));
3266 * XXX PERF_FORMAT_GROUP vs inherited events seems difficult.
3268 static void perf_output_read_group(struct perf_output_handle *handle,
3269 struct perf_event *event)
3271 struct perf_event *leader = event->group_leader, *sub;
3272 u64 read_format = event->attr.read_format;
3276 values[n++] = 1 + leader->nr_siblings;
3278 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
3279 values[n++] = leader->total_time_enabled;
3281 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
3282 values[n++] = leader->total_time_running;
3284 if (leader != event)
3285 leader->pmu->read(leader);
3287 values[n++] = perf_event_count(leader);
3288 if (read_format & PERF_FORMAT_ID)
3289 values[n++] = primary_event_id(leader);
3291 perf_output_copy(handle, values, n * sizeof(u64));
3293 list_for_each_entry(sub, &leader->sibling_list, group_entry) {
3297 sub->pmu->read(sub);
3299 values[n++] = perf_event_count(sub);
3300 if (read_format & PERF_FORMAT_ID)
3301 values[n++] = primary_event_id(sub);
3303 perf_output_copy(handle, values, n * sizeof(u64));
3307 static void perf_output_read(struct perf_output_handle *handle,
3308 struct perf_event *event)
3310 if (event->attr.read_format & PERF_FORMAT_GROUP)
3311 perf_output_read_group(handle, event);
3313 perf_output_read_one(handle, event);
3316 void perf_output_sample(struct perf_output_handle *handle,
3317 struct perf_event_header *header,
3318 struct perf_sample_data *data,
3319 struct perf_event *event)
3321 u64 sample_type = data->type;
3323 perf_output_put(handle, *header);
3325 if (sample_type & PERF_SAMPLE_IP)
3326 perf_output_put(handle, data->ip);
3328 if (sample_type & PERF_SAMPLE_TID)
3329 perf_output_put(handle, data->tid_entry);
3331 if (sample_type & PERF_SAMPLE_TIME)
3332 perf_output_put(handle, data->time);
3334 if (sample_type & PERF_SAMPLE_ADDR)
3335 perf_output_put(handle, data->addr);
3337 if (sample_type & PERF_SAMPLE_ID)
3338 perf_output_put(handle, data->id);
3340 if (sample_type & PERF_SAMPLE_STREAM_ID)
3341 perf_output_put(handle, data->stream_id);
3343 if (sample_type & PERF_SAMPLE_CPU)
3344 perf_output_put(handle, data->cpu_entry);
3346 if (sample_type & PERF_SAMPLE_PERIOD)
3347 perf_output_put(handle, data->period);
3349 if (sample_type & PERF_SAMPLE_READ)
3350 perf_output_read(handle, event);
3352 if (sample_type & PERF_SAMPLE_CALLCHAIN) {
3353 if (data->callchain) {
3356 if (data->callchain)
3357 size += data->callchain->nr;
3359 size *= sizeof(u64);
3361 perf_output_copy(handle, data->callchain, size);
3364 perf_output_put(handle, nr);
3368 if (sample_type & PERF_SAMPLE_RAW) {
3370 perf_output_put(handle, data->raw->size);
3371 perf_output_copy(handle, data->raw->data,
3378 .size = sizeof(u32),
3381 perf_output_put(handle, raw);
3386 void perf_prepare_sample(struct perf_event_header *header,
3387 struct perf_sample_data *data,
3388 struct perf_event *event,
3389 struct pt_regs *regs)
3391 u64 sample_type = event->attr.sample_type;
3393 data->type = sample_type;
3395 header->type = PERF_RECORD_SAMPLE;
3396 header->size = sizeof(*header);
3399 header->misc |= perf_misc_flags(regs);
3401 if (sample_type & PERF_SAMPLE_IP) {
3402 data->ip = perf_instruction_pointer(regs);
3404 header->size += sizeof(data->ip);
3407 if (sample_type & PERF_SAMPLE_TID) {
3408 /* namespace issues */
3409 data->tid_entry.pid = perf_event_pid(event, current);
3410 data->tid_entry.tid = perf_event_tid(event, current);
3412 header->size += sizeof(data->tid_entry);
3415 if (sample_type & PERF_SAMPLE_TIME) {
3416 data->time = perf_clock();
3418 header->size += sizeof(data->time);
3421 if (sample_type & PERF_SAMPLE_ADDR)
3422 header->size += sizeof(data->addr);
3424 if (sample_type & PERF_SAMPLE_ID) {
3425 data->id = primary_event_id(event);
3427 header->size += sizeof(data->id);
3430 if (sample_type & PERF_SAMPLE_STREAM_ID) {
3431 data->stream_id = event->id;
3433 header->size += sizeof(data->stream_id);
3436 if (sample_type & PERF_SAMPLE_CPU) {
3437 data->cpu_entry.cpu = raw_smp_processor_id();
3438 data->cpu_entry.reserved = 0;
3440 header->size += sizeof(data->cpu_entry);
3443 if (sample_type & PERF_SAMPLE_PERIOD)
3444 header->size += sizeof(data->period);
3446 if (sample_type & PERF_SAMPLE_READ)
3447 header->size += perf_event_read_size(event);
3449 if (sample_type & PERF_SAMPLE_CALLCHAIN) {
3452 data->callchain = perf_callchain(regs);
3454 if (data->callchain)
3455 size += data->callchain->nr;
3457 header->size += size * sizeof(u64);
3460 if (sample_type & PERF_SAMPLE_RAW) {
3461 int size = sizeof(u32);
3464 size += data->raw->size;
3466 size += sizeof(u32);
3468 WARN_ON_ONCE(size & (sizeof(u64)-1));
3469 header->size += size;
3473 static void perf_event_output(struct perf_event *event, int nmi,
3474 struct perf_sample_data *data,
3475 struct pt_regs *regs)
3477 struct perf_output_handle handle;
3478 struct perf_event_header header;
3480 perf_prepare_sample(&header, data, event, regs);
3482 if (perf_output_begin(&handle, event, header.size, nmi, 1))
3485 perf_output_sample(&handle, &header, data, event);
3487 perf_output_end(&handle);
3494 struct perf_read_event {
3495 struct perf_event_header header;
3502 perf_event_read_event(struct perf_event *event,
3503 struct task_struct *task)
3505 struct perf_output_handle handle;
3506 struct perf_read_event read_event = {
3508 .type = PERF_RECORD_READ,
3510 .size = sizeof(read_event) + perf_event_read_size(event),
3512 .pid = perf_event_pid(event, task),
3513 .tid = perf_event_tid(event, task),
3517 ret = perf_output_begin(&handle, event, read_event.header.size, 0, 0);
3521 perf_output_put(&handle, read_event);
3522 perf_output_read(&handle, event);
3524 perf_output_end(&handle);
3528 * task tracking -- fork/exit
3530 * enabled by: attr.comm | attr.mmap | attr.mmap_data | attr.task
3533 struct perf_task_event {
3534 struct task_struct *task;
3535 struct perf_event_context *task_ctx;
3538 struct perf_event_header header;
3548 static void perf_event_task_output(struct perf_event *event,
3549 struct perf_task_event *task_event)
3551 struct perf_output_handle handle;
3552 struct task_struct *task = task_event->task;
3555 size = task_event->event_id.header.size;
3556 ret = perf_output_begin(&handle, event, size, 0, 0);
3561 task_event->event_id.pid = perf_event_pid(event, task);
3562 task_event->event_id.ppid = perf_event_pid(event, current);
3564 task_event->event_id.tid = perf_event_tid(event, task);
3565 task_event->event_id.ptid = perf_event_tid(event, current);
3567 perf_output_put(&handle, task_event->event_id);
3569 perf_output_end(&handle);
3572 static int perf_event_task_match(struct perf_event *event)
3574 if (event->state < PERF_EVENT_STATE_INACTIVE)
3577 if (event->cpu != -1 && event->cpu != smp_processor_id())
3580 if (event->attr.comm || event->attr.mmap ||
3581 event->attr.mmap_data || event->attr.task)
3587 static void perf_event_task_ctx(struct perf_event_context *ctx,
3588 struct perf_task_event *task_event)
3590 struct perf_event *event;
3592 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
3593 if (perf_event_task_match(event))
3594 perf_event_task_output(event, task_event);
3598 static void perf_event_task_event(struct perf_task_event *task_event)
3600 struct perf_cpu_context *cpuctx;
3601 struct perf_event_context *ctx = task_event->task_ctx;
3604 cpuctx = &get_cpu_var(perf_cpu_context);
3605 perf_event_task_ctx(&cpuctx->ctx, task_event);
3607 ctx = rcu_dereference(current->perf_event_ctxp);
3609 perf_event_task_ctx(ctx, task_event);
3610 put_cpu_var(perf_cpu_context);
3614 static void perf_event_task(struct task_struct *task,
3615 struct perf_event_context *task_ctx,
3618 struct perf_task_event task_event;
3620 if (!atomic_read(&nr_comm_events) &&
3621 !atomic_read(&nr_mmap_events) &&
3622 !atomic_read(&nr_task_events))
3625 task_event = (struct perf_task_event){
3627 .task_ctx = task_ctx,
3630 .type = new ? PERF_RECORD_FORK : PERF_RECORD_EXIT,
3632 .size = sizeof(task_event.event_id),
3638 .time = perf_clock(),
3642 perf_event_task_event(&task_event);
3645 void perf_event_fork(struct task_struct *task)
3647 perf_event_task(task, NULL, 1);
3654 struct perf_comm_event {
3655 struct task_struct *task;
3660 struct perf_event_header header;
3667 static void perf_event_comm_output(struct perf_event *event,
3668 struct perf_comm_event *comm_event)
3670 struct perf_output_handle handle;
3671 int size = comm_event->event_id.header.size;
3672 int ret = perf_output_begin(&handle, event, size, 0, 0);
3677 comm_event->event_id.pid = perf_event_pid(event, comm_event->task);
3678 comm_event->event_id.tid = perf_event_tid(event, comm_event->task);
3680 perf_output_put(&handle, comm_event->event_id);
3681 perf_output_copy(&handle, comm_event->comm,
3682 comm_event->comm_size);
3683 perf_output_end(&handle);
3686 static int perf_event_comm_match(struct perf_event *event)
3688 if (event->state < PERF_EVENT_STATE_INACTIVE)
3691 if (event->cpu != -1 && event->cpu != smp_processor_id())
3694 if (event->attr.comm)
3700 static void perf_event_comm_ctx(struct perf_event_context *ctx,
3701 struct perf_comm_event *comm_event)
3703 struct perf_event *event;
3705 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
3706 if (perf_event_comm_match(event))
3707 perf_event_comm_output(event, comm_event);
3711 static void perf_event_comm_event(struct perf_comm_event *comm_event)
3713 struct perf_cpu_context *cpuctx;
3714 struct perf_event_context *ctx;
3716 char comm[TASK_COMM_LEN];
3718 memset(comm, 0, sizeof(comm));
3719 strlcpy(comm, comm_event->task->comm, sizeof(comm));
3720 size = ALIGN(strlen(comm)+1, sizeof(u64));
3722 comm_event->comm = comm;
3723 comm_event->comm_size = size;
3725 comm_event->event_id.header.size = sizeof(comm_event->event_id) + size;
3728 cpuctx = &get_cpu_var(perf_cpu_context);
3729 perf_event_comm_ctx(&cpuctx->ctx, comm_event);
3730 ctx = rcu_dereference(current->perf_event_ctxp);
3732 perf_event_comm_ctx(ctx, comm_event);
3733 put_cpu_var(perf_cpu_context);
3737 void perf_event_comm(struct task_struct *task)
3739 struct perf_comm_event comm_event;
3741 if (task->perf_event_ctxp)
3742 perf_event_enable_on_exec(task);
3744 if (!atomic_read(&nr_comm_events))
3747 comm_event = (struct perf_comm_event){
3753 .type = PERF_RECORD_COMM,
3762 perf_event_comm_event(&comm_event);
3769 struct perf_mmap_event {
3770 struct vm_area_struct *vma;
3772 const char *file_name;
3776 struct perf_event_header header;
3786 static void perf_event_mmap_output(struct perf_event *event,
3787 struct perf_mmap_event *mmap_event)
3789 struct perf_output_handle handle;
3790 int size = mmap_event->event_id.header.size;
3791 int ret = perf_output_begin(&handle, event, size, 0, 0);
3796 mmap_event->event_id.pid = perf_event_pid(event, current);
3797 mmap_event->event_id.tid = perf_event_tid(event, current);
3799 perf_output_put(&handle, mmap_event->event_id);
3800 perf_output_copy(&handle, mmap_event->file_name,
3801 mmap_event->file_size);
3802 perf_output_end(&handle);
3805 static int perf_event_mmap_match(struct perf_event *event,
3806 struct perf_mmap_event *mmap_event,
3809 if (event->state < PERF_EVENT_STATE_INACTIVE)
3812 if (event->cpu != -1 && event->cpu != smp_processor_id())
3815 if ((!executable && event->attr.mmap_data) ||
3816 (executable && event->attr.mmap))
3822 static void perf_event_mmap_ctx(struct perf_event_context *ctx,
3823 struct perf_mmap_event *mmap_event,
3826 struct perf_event *event;
3828 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
3829 if (perf_event_mmap_match(event, mmap_event, executable))
3830 perf_event_mmap_output(event, mmap_event);
3834 static void perf_event_mmap_event(struct perf_mmap_event *mmap_event)
3836 struct perf_cpu_context *cpuctx;
3837 struct perf_event_context *ctx;
3838 struct vm_area_struct *vma = mmap_event->vma;
3839 struct file *file = vma->vm_file;
3845 memset(tmp, 0, sizeof(tmp));
3849 * d_path works from the end of the buffer backwards, so we
3850 * need to add enough zero bytes after the string to handle
3851 * the 64bit alignment we do later.
3853 buf = kzalloc(PATH_MAX + sizeof(u64), GFP_KERNEL);
3855 name = strncpy(tmp, "//enomem", sizeof(tmp));
3858 name = d_path(&file->f_path, buf, PATH_MAX);
3860 name = strncpy(tmp, "//toolong", sizeof(tmp));
3864 if (arch_vma_name(mmap_event->vma)) {
3865 name = strncpy(tmp, arch_vma_name(mmap_event->vma),
3871 name = strncpy(tmp, "[vdso]", sizeof(tmp));
3873 } else if (vma->vm_start <= vma->vm_mm->start_brk &&
3874 vma->vm_end >= vma->vm_mm->brk) {
3875 name = strncpy(tmp, "[heap]", sizeof(tmp));
3877 } else if (vma->vm_start <= vma->vm_mm->start_stack &&
3878 vma->vm_end >= vma->vm_mm->start_stack) {
3879 name = strncpy(tmp, "[stack]", sizeof(tmp));
3883 name = strncpy(tmp, "//anon", sizeof(tmp));
3888 size = ALIGN(strlen(name)+1, sizeof(u64));
3890 mmap_event->file_name = name;
3891 mmap_event->file_size = size;
3893 mmap_event->event_id.header.size = sizeof(mmap_event->event_id) + size;
3896 cpuctx = &get_cpu_var(perf_cpu_context);
3897 perf_event_mmap_ctx(&cpuctx->ctx, mmap_event, vma->vm_flags & VM_EXEC);
3898 ctx = rcu_dereference(current->perf_event_ctxp);
3900 perf_event_mmap_ctx(ctx, mmap_event, vma->vm_flags & VM_EXEC);
3901 put_cpu_var(perf_cpu_context);
3907 void perf_event_mmap(struct vm_area_struct *vma)
3909 struct perf_mmap_event mmap_event;
3911 if (!atomic_read(&nr_mmap_events))
3914 mmap_event = (struct perf_mmap_event){
3920 .type = PERF_RECORD_MMAP,
3921 .misc = PERF_RECORD_MISC_USER,
3926 .start = vma->vm_start,
3927 .len = vma->vm_end - vma->vm_start,
3928 .pgoff = (u64)vma->vm_pgoff << PAGE_SHIFT,
3932 perf_event_mmap_event(&mmap_event);
3936 * IRQ throttle logging
3939 static void perf_log_throttle(struct perf_event *event, int enable)
3941 struct perf_output_handle handle;
3945 struct perf_event_header header;
3949 } throttle_event = {
3951 .type = PERF_RECORD_THROTTLE,
3953 .size = sizeof(throttle_event),
3955 .time = perf_clock(),
3956 .id = primary_event_id(event),
3957 .stream_id = event->id,
3961 throttle_event.header.type = PERF_RECORD_UNTHROTTLE;
3963 ret = perf_output_begin(&handle, event, sizeof(throttle_event), 1, 0);
3967 perf_output_put(&handle, throttle_event);
3968 perf_output_end(&handle);
3972 * Generic event overflow handling, sampling.
3975 static int __perf_event_overflow(struct perf_event *event, int nmi,
3976 int throttle, struct perf_sample_data *data,
3977 struct pt_regs *regs)
3979 int events = atomic_read(&event->event_limit);
3980 struct hw_perf_event *hwc = &event->hw;
3983 throttle = (throttle && event->pmu->unthrottle != NULL);
3988 if (hwc->interrupts != MAX_INTERRUPTS) {
3990 if (HZ * hwc->interrupts >
3991 (u64)sysctl_perf_event_sample_rate) {
3992 hwc->interrupts = MAX_INTERRUPTS;
3993 perf_log_throttle(event, 0);
3998 * Keep re-disabling events even though on the previous
3999 * pass we disabled it - just in case we raced with a
4000 * sched-in and the event got enabled again:
4006 if (event->attr.freq) {
4007 u64 now = perf_clock();
4008 s64 delta = now - hwc->freq_time_stamp;
4010 hwc->freq_time_stamp = now;
4012 if (delta > 0 && delta < 2*TICK_NSEC)
4013 perf_adjust_period(event, delta, hwc->last_period);
4017 * XXX event_limit might not quite work as expected on inherited
4021 event->pending_kill = POLL_IN;
4022 if (events && atomic_dec_and_test(&event->event_limit)) {
4024 event->pending_kill = POLL_HUP;
4026 event->pending_disable = 1;
4027 perf_pending_queue(&event->pending,
4028 perf_pending_event);
4030 perf_event_disable(event);
4033 if (event->overflow_handler)
4034 event->overflow_handler(event, nmi, data, regs);
4036 perf_event_output(event, nmi, data, regs);
4041 int perf_event_overflow(struct perf_event *event, int nmi,
4042 struct perf_sample_data *data,
4043 struct pt_regs *regs)
4045 return __perf_event_overflow(event, nmi, 1, data, regs);
4049 * Generic software event infrastructure
4053 * We directly increment event->count and keep a second value in
4054 * event->hw.period_left to count intervals. This period event
4055 * is kept in the range [-sample_period, 0] so that we can use the
4059 static u64 perf_swevent_set_period(struct perf_event *event)
4061 struct hw_perf_event *hwc = &event->hw;
4062 u64 period = hwc->last_period;
4066 hwc->last_period = hwc->sample_period;
4069 old = val = local64_read(&hwc->period_left);
4073 nr = div64_u64(period + val, period);
4074 offset = nr * period;
4076 if (local64_cmpxchg(&hwc->period_left, old, val) != old)
4082 static void perf_swevent_overflow(struct perf_event *event, u64 overflow,
4083 int nmi, struct perf_sample_data *data,
4084 struct pt_regs *regs)
4086 struct hw_perf_event *hwc = &event->hw;
4089 data->period = event->hw.last_period;
4091 overflow = perf_swevent_set_period(event);
4093 if (hwc->interrupts == MAX_INTERRUPTS)
4096 for (; overflow; overflow--) {
4097 if (__perf_event_overflow(event, nmi, throttle,
4100 * We inhibit the overflow from happening when
4101 * hwc->interrupts == MAX_INTERRUPTS.
4109 static void perf_swevent_add(struct perf_event *event, u64 nr,
4110 int nmi, struct perf_sample_data *data,
4111 struct pt_regs *regs)
4113 struct hw_perf_event *hwc = &event->hw;
4115 local64_add(nr, &event->count);
4120 if (!hwc->sample_period)
4123 if (nr == 1 && hwc->sample_period == 1 && !event->attr.freq)
4124 return perf_swevent_overflow(event, 1, nmi, data, regs);
4126 if (local64_add_negative(nr, &hwc->period_left))
4129 perf_swevent_overflow(event, 0, nmi, data, regs);
4132 static int perf_exclude_event(struct perf_event *event,
4133 struct pt_regs *regs)
4136 if (event->attr.exclude_user && user_mode(regs))
4139 if (event->attr.exclude_kernel && !user_mode(regs))
4146 static int perf_swevent_match(struct perf_event *event,
4147 enum perf_type_id type,
4149 struct perf_sample_data *data,
4150 struct pt_regs *regs)
4152 if (event->attr.type != type)
4155 if (event->attr.config != event_id)
4158 if (perf_exclude_event(event, regs))
4164 static inline u64 swevent_hash(u64 type, u32 event_id)
4166 u64 val = event_id | (type << 32);
4168 return hash_64(val, SWEVENT_HLIST_BITS);
4171 static inline struct hlist_head *
4172 __find_swevent_head(struct swevent_hlist *hlist, u64 type, u32 event_id)
4174 u64 hash = swevent_hash(type, event_id);
4176 return &hlist->heads[hash];
4179 /* For the read side: events when they trigger */
4180 static inline struct hlist_head *
4181 find_swevent_head_rcu(struct perf_cpu_context *ctx, u64 type, u32 event_id)
4183 struct swevent_hlist *hlist;
4185 hlist = rcu_dereference(ctx->swevent_hlist);
4189 return __find_swevent_head(hlist, type, event_id);
4192 /* For the event head insertion and removal in the hlist */
4193 static inline struct hlist_head *
4194 find_swevent_head(struct perf_cpu_context *ctx, struct perf_event *event)
4196 struct swevent_hlist *hlist;
4197 u32 event_id = event->attr.config;
4198 u64 type = event->attr.type;
4201 * Event scheduling is always serialized against hlist allocation
4202 * and release. Which makes the protected version suitable here.
4203 * The context lock guarantees that.
4205 hlist = rcu_dereference_protected(ctx->swevent_hlist,
4206 lockdep_is_held(&event->ctx->lock));
4210 return __find_swevent_head(hlist, type, event_id);
4213 static void do_perf_sw_event(enum perf_type_id type, u32 event_id,
4215 struct perf_sample_data *data,
4216 struct pt_regs *regs)
4218 struct perf_cpu_context *cpuctx;
4219 struct perf_event *event;
4220 struct hlist_node *node;
4221 struct hlist_head *head;
4223 cpuctx = &__get_cpu_var(perf_cpu_context);
4227 head = find_swevent_head_rcu(cpuctx, type, event_id);
4232 hlist_for_each_entry_rcu(event, node, head, hlist_entry) {
4233 if (perf_swevent_match(event, type, event_id, data, regs))
4234 perf_swevent_add(event, nr, nmi, data, regs);
4240 int perf_swevent_get_recursion_context(void)
4242 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
4249 else if (in_softirq())
4254 if (cpuctx->recursion[rctx])
4257 cpuctx->recursion[rctx]++;
4262 EXPORT_SYMBOL_GPL(perf_swevent_get_recursion_context);
4264 void inline perf_swevent_put_recursion_context(int rctx)
4266 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
4268 cpuctx->recursion[rctx]--;
4271 void __perf_sw_event(u32 event_id, u64 nr, int nmi,
4272 struct pt_regs *regs, u64 addr)
4274 struct perf_sample_data data;
4277 preempt_disable_notrace();
4278 rctx = perf_swevent_get_recursion_context();
4282 perf_sample_data_init(&data, addr);
4284 do_perf_sw_event(PERF_TYPE_SOFTWARE, event_id, nr, nmi, &data, regs);
4286 perf_swevent_put_recursion_context(rctx);
4287 preempt_enable_notrace();
4290 static void perf_swevent_read(struct perf_event *event)
4294 static int perf_swevent_enable(struct perf_event *event)
4296 struct hw_perf_event *hwc = &event->hw;
4297 struct perf_cpu_context *cpuctx;
4298 struct hlist_head *head;
4300 cpuctx = &__get_cpu_var(perf_cpu_context);
4302 if (hwc->sample_period) {
4303 hwc->last_period = hwc->sample_period;
4304 perf_swevent_set_period(event);
4307 head = find_swevent_head(cpuctx, event);
4308 if (WARN_ON_ONCE(!head))
4311 hlist_add_head_rcu(&event->hlist_entry, head);
4316 static void perf_swevent_disable(struct perf_event *event)
4318 hlist_del_rcu(&event->hlist_entry);
4321 static void perf_swevent_void(struct perf_event *event)
4325 static int perf_swevent_int(struct perf_event *event)
4330 static const struct pmu perf_ops_generic = {
4331 .enable = perf_swevent_enable,
4332 .disable = perf_swevent_disable,
4333 .start = perf_swevent_int,
4334 .stop = perf_swevent_void,
4335 .read = perf_swevent_read,
4336 .unthrottle = perf_swevent_void, /* hwc->interrupts already reset */
4340 * hrtimer based swevent callback
4343 static enum hrtimer_restart perf_swevent_hrtimer(struct hrtimer *hrtimer)
4345 enum hrtimer_restart ret = HRTIMER_RESTART;
4346 struct perf_sample_data data;
4347 struct pt_regs *regs;
4348 struct perf_event *event;
4351 event = container_of(hrtimer, struct perf_event, hw.hrtimer);
4352 event->pmu->read(event);
4354 perf_sample_data_init(&data, 0);
4355 data.period = event->hw.last_period;
4356 regs = get_irq_regs();
4358 if (regs && !perf_exclude_event(event, regs)) {
4359 if (!(event->attr.exclude_idle && current->pid == 0))
4360 if (perf_event_overflow(event, 0, &data, regs))
4361 ret = HRTIMER_NORESTART;
4364 period = max_t(u64, 10000, event->hw.sample_period);
4365 hrtimer_forward_now(hrtimer, ns_to_ktime(period));
4370 static void perf_swevent_start_hrtimer(struct perf_event *event)
4372 struct hw_perf_event *hwc = &event->hw;
4374 hrtimer_init(&hwc->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
4375 hwc->hrtimer.function = perf_swevent_hrtimer;
4376 if (hwc->sample_period) {
4379 if (hwc->remaining) {
4380 if (hwc->remaining < 0)
4383 period = hwc->remaining;
4386 period = max_t(u64, 10000, hwc->sample_period);
4388 __hrtimer_start_range_ns(&hwc->hrtimer,
4389 ns_to_ktime(period), 0,
4390 HRTIMER_MODE_REL, 0);
4394 static void perf_swevent_cancel_hrtimer(struct perf_event *event)
4396 struct hw_perf_event *hwc = &event->hw;
4398 if (hwc->sample_period) {
4399 ktime_t remaining = hrtimer_get_remaining(&hwc->hrtimer);
4400 hwc->remaining = ktime_to_ns(remaining);
4402 hrtimer_cancel(&hwc->hrtimer);
4407 * Software event: cpu wall time clock
4410 static void cpu_clock_perf_event_update(struct perf_event *event)
4412 int cpu = raw_smp_processor_id();
4416 now = cpu_clock(cpu);
4417 prev = local64_xchg(&event->hw.prev_count, now);
4418 local64_add(now - prev, &event->count);
4421 static int cpu_clock_perf_event_enable(struct perf_event *event)
4423 struct hw_perf_event *hwc = &event->hw;
4424 int cpu = raw_smp_processor_id();
4426 local64_set(&hwc->prev_count, cpu_clock(cpu));
4427 perf_swevent_start_hrtimer(event);
4432 static void cpu_clock_perf_event_disable(struct perf_event *event)
4434 perf_swevent_cancel_hrtimer(event);
4435 cpu_clock_perf_event_update(event);
4438 static void cpu_clock_perf_event_read(struct perf_event *event)
4440 cpu_clock_perf_event_update(event);
4443 static const struct pmu perf_ops_cpu_clock = {
4444 .enable = cpu_clock_perf_event_enable,
4445 .disable = cpu_clock_perf_event_disable,
4446 .read = cpu_clock_perf_event_read,
4450 * Software event: task time clock
4453 static void task_clock_perf_event_update(struct perf_event *event, u64 now)
4458 prev = local64_xchg(&event->hw.prev_count, now);
4460 local64_add(delta, &event->count);
4463 static int task_clock_perf_event_enable(struct perf_event *event)
4465 struct hw_perf_event *hwc = &event->hw;
4468 now = event->ctx->time;
4470 local64_set(&hwc->prev_count, now);
4472 perf_swevent_start_hrtimer(event);
4477 static void task_clock_perf_event_disable(struct perf_event *event)
4479 perf_swevent_cancel_hrtimer(event);
4480 task_clock_perf_event_update(event, event->ctx->time);
4484 static void task_clock_perf_event_read(struct perf_event *event)
4489 update_context_time(event->ctx);
4490 time = event->ctx->time;
4492 u64 now = perf_clock();
4493 u64 delta = now - event->ctx->timestamp;
4494 time = event->ctx->time + delta;
4497 task_clock_perf_event_update(event, time);
4500 static const struct pmu perf_ops_task_clock = {
4501 .enable = task_clock_perf_event_enable,
4502 .disable = task_clock_perf_event_disable,
4503 .read = task_clock_perf_event_read,
4506 /* Deref the hlist from the update side */
4507 static inline struct swevent_hlist *
4508 swevent_hlist_deref(struct perf_cpu_context *cpuctx)
4510 return rcu_dereference_protected(cpuctx->swevent_hlist,
4511 lockdep_is_held(&cpuctx->hlist_mutex));
4514 static void swevent_hlist_release_rcu(struct rcu_head *rcu_head)
4516 struct swevent_hlist *hlist;
4518 hlist = container_of(rcu_head, struct swevent_hlist, rcu_head);
4522 static void swevent_hlist_release(struct perf_cpu_context *cpuctx)
4524 struct swevent_hlist *hlist = swevent_hlist_deref(cpuctx);
4529 rcu_assign_pointer(cpuctx->swevent_hlist, NULL);
4530 call_rcu(&hlist->rcu_head, swevent_hlist_release_rcu);
4533 static void swevent_hlist_put_cpu(struct perf_event *event, int cpu)
4535 struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
4537 mutex_lock(&cpuctx->hlist_mutex);
4539 if (!--cpuctx->hlist_refcount)
4540 swevent_hlist_release(cpuctx);
4542 mutex_unlock(&cpuctx->hlist_mutex);
4545 static void swevent_hlist_put(struct perf_event *event)
4549 if (event->cpu != -1) {
4550 swevent_hlist_put_cpu(event, event->cpu);
4554 for_each_possible_cpu(cpu)
4555 swevent_hlist_put_cpu(event, cpu);
4558 static int swevent_hlist_get_cpu(struct perf_event *event, int cpu)
4560 struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
4563 mutex_lock(&cpuctx->hlist_mutex);
4565 if (!swevent_hlist_deref(cpuctx) && cpu_online(cpu)) {
4566 struct swevent_hlist *hlist;
4568 hlist = kzalloc(sizeof(*hlist), GFP_KERNEL);
4573 rcu_assign_pointer(cpuctx->swevent_hlist, hlist);
4575 cpuctx->hlist_refcount++;
4577 mutex_unlock(&cpuctx->hlist_mutex);
4582 static int swevent_hlist_get(struct perf_event *event)
4585 int cpu, failed_cpu;
4587 if (event->cpu != -1)
4588 return swevent_hlist_get_cpu(event, event->cpu);
4591 for_each_possible_cpu(cpu) {
4592 err = swevent_hlist_get_cpu(event, cpu);
4602 for_each_possible_cpu(cpu) {
4603 if (cpu == failed_cpu)
4605 swevent_hlist_put_cpu(event, cpu);
4612 #ifdef CONFIG_EVENT_TRACING
4614 static const struct pmu perf_ops_tracepoint = {
4615 .enable = perf_trace_enable,
4616 .disable = perf_trace_disable,
4617 .start = perf_swevent_int,
4618 .stop = perf_swevent_void,
4619 .read = perf_swevent_read,
4620 .unthrottle = perf_swevent_void,
4623 static int perf_tp_filter_match(struct perf_event *event,
4624 struct perf_sample_data *data)
4626 void *record = data->raw->data;
4628 if (likely(!event->filter) || filter_match_preds(event->filter, record))
4633 static int perf_tp_event_match(struct perf_event *event,
4634 struct perf_sample_data *data,
4635 struct pt_regs *regs)
4638 * All tracepoints are from kernel-space.
4640 if (event->attr.exclude_kernel)
4643 if (!perf_tp_filter_match(event, data))
4649 void perf_tp_event(u64 addr, u64 count, void *record, int entry_size,
4650 struct pt_regs *regs, struct hlist_head *head, int rctx)
4652 struct perf_sample_data data;
4653 struct perf_event *event;
4654 struct hlist_node *node;
4656 struct perf_raw_record raw = {
4661 perf_sample_data_init(&data, addr);
4664 hlist_for_each_entry_rcu(event, node, head, hlist_entry) {
4665 if (perf_tp_event_match(event, &data, regs))
4666 perf_swevent_add(event, count, 1, &data, regs);
4669 perf_swevent_put_recursion_context(rctx);
4671 EXPORT_SYMBOL_GPL(perf_tp_event);
4673 static void tp_perf_event_destroy(struct perf_event *event)
4675 perf_trace_destroy(event);
4678 static const struct pmu *tp_perf_event_init(struct perf_event *event)
4683 * Raw tracepoint data is a severe data leak, only allow root to
4686 if ((event->attr.sample_type & PERF_SAMPLE_RAW) &&
4687 perf_paranoid_tracepoint_raw() &&
4688 !capable(CAP_SYS_ADMIN))
4689 return ERR_PTR(-EPERM);
4691 err = perf_trace_init(event);
4695 event->destroy = tp_perf_event_destroy;
4697 return &perf_ops_tracepoint;
4700 static int perf_event_set_filter(struct perf_event *event, void __user *arg)
4705 if (event->attr.type != PERF_TYPE_TRACEPOINT)
4708 filter_str = strndup_user(arg, PAGE_SIZE);
4709 if (IS_ERR(filter_str))
4710 return PTR_ERR(filter_str);
4712 ret = ftrace_profile_set_filter(event, event->attr.config, filter_str);
4718 static void perf_event_free_filter(struct perf_event *event)
4720 ftrace_profile_free_filter(event);
4725 static const struct pmu *tp_perf_event_init(struct perf_event *event)
4730 static int perf_event_set_filter(struct perf_event *event, void __user *arg)
4735 static void perf_event_free_filter(struct perf_event *event)
4739 #endif /* CONFIG_EVENT_TRACING */
4741 #ifdef CONFIG_HAVE_HW_BREAKPOINT
4742 static void bp_perf_event_destroy(struct perf_event *event)
4744 release_bp_slot(event);
4747 static const struct pmu *bp_perf_event_init(struct perf_event *bp)
4751 err = register_perf_hw_breakpoint(bp);
4753 return ERR_PTR(err);
4755 bp->destroy = bp_perf_event_destroy;
4757 return &perf_ops_bp;
4760 void perf_bp_event(struct perf_event *bp, void *data)
4762 struct perf_sample_data sample;
4763 struct pt_regs *regs = data;
4765 perf_sample_data_init(&sample, bp->attr.bp_addr);
4767 if (!perf_exclude_event(bp, regs))
4768 perf_swevent_add(bp, 1, 1, &sample, regs);
4771 static const struct pmu *bp_perf_event_init(struct perf_event *bp)
4776 void perf_bp_event(struct perf_event *bp, void *regs)
4781 atomic_t perf_swevent_enabled[PERF_COUNT_SW_MAX];
4783 static void sw_perf_event_destroy(struct perf_event *event)
4785 u64 event_id = event->attr.config;
4787 WARN_ON(event->parent);
4789 atomic_dec(&perf_swevent_enabled[event_id]);
4790 swevent_hlist_put(event);
4793 static const struct pmu *sw_perf_event_init(struct perf_event *event)
4795 const struct pmu *pmu = NULL;
4796 u64 event_id = event->attr.config;
4799 * Software events (currently) can't in general distinguish
4800 * between user, kernel and hypervisor events.
4801 * However, context switches and cpu migrations are considered
4802 * to be kernel events, and page faults are never hypervisor
4806 case PERF_COUNT_SW_CPU_CLOCK:
4807 pmu = &perf_ops_cpu_clock;
4810 case PERF_COUNT_SW_TASK_CLOCK:
4812 * If the user instantiates this as a per-cpu event,
4813 * use the cpu_clock event instead.
4815 if (event->ctx->task)
4816 pmu = &perf_ops_task_clock;
4818 pmu = &perf_ops_cpu_clock;
4821 case PERF_COUNT_SW_PAGE_FAULTS:
4822 case PERF_COUNT_SW_PAGE_FAULTS_MIN:
4823 case PERF_COUNT_SW_PAGE_FAULTS_MAJ:
4824 case PERF_COUNT_SW_CONTEXT_SWITCHES:
4825 case PERF_COUNT_SW_CPU_MIGRATIONS:
4826 case PERF_COUNT_SW_ALIGNMENT_FAULTS:
4827 case PERF_COUNT_SW_EMULATION_FAULTS:
4828 if (!event->parent) {
4831 err = swevent_hlist_get(event);
4833 return ERR_PTR(err);
4835 atomic_inc(&perf_swevent_enabled[event_id]);
4836 event->destroy = sw_perf_event_destroy;
4838 pmu = &perf_ops_generic;
4846 * Allocate and initialize a event structure
4848 static struct perf_event *
4849 perf_event_alloc(struct perf_event_attr *attr,
4851 struct perf_event_context *ctx,
4852 struct perf_event *group_leader,
4853 struct perf_event *parent_event,
4854 perf_overflow_handler_t overflow_handler,
4857 const struct pmu *pmu;
4858 struct perf_event *event;
4859 struct hw_perf_event *hwc;
4862 event = kzalloc(sizeof(*event), gfpflags);
4864 return ERR_PTR(-ENOMEM);
4867 * Single events are their own group leaders, with an
4868 * empty sibling list:
4871 group_leader = event;
4873 mutex_init(&event->child_mutex);
4874 INIT_LIST_HEAD(&event->child_list);
4876 INIT_LIST_HEAD(&event->group_entry);
4877 INIT_LIST_HEAD(&event->event_entry);
4878 INIT_LIST_HEAD(&event->sibling_list);
4879 init_waitqueue_head(&event->waitq);
4881 mutex_init(&event->mmap_mutex);
4884 event->attr = *attr;
4885 event->group_leader = group_leader;
4890 event->parent = parent_event;
4892 event->ns = get_pid_ns(current->nsproxy->pid_ns);
4893 event->id = atomic64_inc_return(&perf_event_id);
4895 event->state = PERF_EVENT_STATE_INACTIVE;
4897 if (!overflow_handler && parent_event)
4898 overflow_handler = parent_event->overflow_handler;
4900 event->overflow_handler = overflow_handler;
4903 event->state = PERF_EVENT_STATE_OFF;
4908 hwc->sample_period = attr->sample_period;
4909 if (attr->freq && attr->sample_freq)
4910 hwc->sample_period = 1;
4911 hwc->last_period = hwc->sample_period;
4913 local64_set(&hwc->period_left, hwc->sample_period);
4916 * we currently do not support PERF_FORMAT_GROUP on inherited events
4918 if (attr->inherit && (attr->read_format & PERF_FORMAT_GROUP))
4921 switch (attr->type) {
4923 case PERF_TYPE_HARDWARE:
4924 case PERF_TYPE_HW_CACHE:
4925 pmu = hw_perf_event_init(event);
4928 case PERF_TYPE_SOFTWARE:
4929 pmu = sw_perf_event_init(event);
4932 case PERF_TYPE_TRACEPOINT:
4933 pmu = tp_perf_event_init(event);
4936 case PERF_TYPE_BREAKPOINT:
4937 pmu = bp_perf_event_init(event);
4948 else if (IS_ERR(pmu))
4953 put_pid_ns(event->ns);
4955 return ERR_PTR(err);
4960 if (!event->parent) {
4961 atomic_inc(&nr_events);
4962 if (event->attr.mmap || event->attr.mmap_data)
4963 atomic_inc(&nr_mmap_events);
4964 if (event->attr.comm)
4965 atomic_inc(&nr_comm_events);
4966 if (event->attr.task)
4967 atomic_inc(&nr_task_events);
4973 static int perf_copy_attr(struct perf_event_attr __user *uattr,
4974 struct perf_event_attr *attr)
4979 if (!access_ok(VERIFY_WRITE, uattr, PERF_ATTR_SIZE_VER0))
4983 * zero the full structure, so that a short copy will be nice.
4985 memset(attr, 0, sizeof(*attr));
4987 ret = get_user(size, &uattr->size);
4991 if (size > PAGE_SIZE) /* silly large */
4994 if (!size) /* abi compat */
4995 size = PERF_ATTR_SIZE_VER0;
4997 if (size < PERF_ATTR_SIZE_VER0)
5001 * If we're handed a bigger struct than we know of,
5002 * ensure all the unknown bits are 0 - i.e. new
5003 * user-space does not rely on any kernel feature
5004 * extensions we dont know about yet.
5006 if (size > sizeof(*attr)) {
5007 unsigned char __user *addr;
5008 unsigned char __user *end;
5011 addr = (void __user *)uattr + sizeof(*attr);
5012 end = (void __user *)uattr + size;
5014 for (; addr < end; addr++) {
5015 ret = get_user(val, addr);
5021 size = sizeof(*attr);
5024 ret = copy_from_user(attr, uattr, size);
5029 * If the type exists, the corresponding creation will verify
5032 if (attr->type >= PERF_TYPE_MAX)
5035 if (attr->__reserved_1)
5038 if (attr->sample_type & ~(PERF_SAMPLE_MAX-1))
5041 if (attr->read_format & ~(PERF_FORMAT_MAX-1))
5048 put_user(sizeof(*attr), &uattr->size);
5054 perf_event_set_output(struct perf_event *event, struct perf_event *output_event)
5056 struct perf_buffer *buffer = NULL, *old_buffer = NULL;
5062 /* don't allow circular references */
5063 if (event == output_event)
5067 * Don't allow cross-cpu buffers
5069 if (output_event->cpu != event->cpu)
5073 * If its not a per-cpu buffer, it must be the same task.
5075 if (output_event->cpu == -1 && output_event->ctx != event->ctx)
5079 mutex_lock(&event->mmap_mutex);
5080 /* Can't redirect output if we've got an active mmap() */
5081 if (atomic_read(&event->mmap_count))
5085 /* get the buffer we want to redirect to */
5086 buffer = perf_buffer_get(output_event);
5091 old_buffer = event->buffer;
5092 rcu_assign_pointer(event->buffer, buffer);
5095 mutex_unlock(&event->mmap_mutex);
5098 perf_buffer_put(old_buffer);
5104 * sys_perf_event_open - open a performance event, associate it to a task/cpu
5106 * @attr_uptr: event_id type attributes for monitoring/sampling
5109 * @group_fd: group leader event fd
5111 SYSCALL_DEFINE5(perf_event_open,
5112 struct perf_event_attr __user *, attr_uptr,
5113 pid_t, pid, int, cpu, int, group_fd, unsigned long, flags)
5115 struct perf_event *event, *group_leader = NULL, *output_event = NULL;
5116 struct perf_event_attr attr;
5117 struct perf_event_context *ctx;
5118 struct file *event_file = NULL;
5119 struct file *group_file = NULL;
5121 int fput_needed = 0;
5124 /* for future expandability... */
5125 if (flags & ~(PERF_FLAG_FD_NO_GROUP | PERF_FLAG_FD_OUTPUT))
5128 err = perf_copy_attr(attr_uptr, &attr);
5132 if (!attr.exclude_kernel) {
5133 if (perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN))
5138 if (attr.sample_freq > sysctl_perf_event_sample_rate)
5142 event_fd = get_unused_fd_flags(O_RDWR);
5147 * Get the target context (task or percpu):
5149 ctx = find_get_context(pid, cpu);
5155 if (group_fd != -1) {
5156 group_leader = perf_fget_light(group_fd, &fput_needed);
5157 if (IS_ERR(group_leader)) {
5158 err = PTR_ERR(group_leader);
5159 goto err_put_context;
5161 group_file = group_leader->filp;
5162 if (flags & PERF_FLAG_FD_OUTPUT)
5163 output_event = group_leader;
5164 if (flags & PERF_FLAG_FD_NO_GROUP)
5165 group_leader = NULL;
5169 * Look up the group leader (we will attach this event to it):
5175 * Do not allow a recursive hierarchy (this new sibling
5176 * becoming part of another group-sibling):
5178 if (group_leader->group_leader != group_leader)
5179 goto err_put_context;
5181 * Do not allow to attach to a group in a different
5182 * task or CPU context:
5184 if (group_leader->ctx != ctx)
5185 goto err_put_context;
5187 * Only a group leader can be exclusive or pinned
5189 if (attr.exclusive || attr.pinned)
5190 goto err_put_context;
5193 event = perf_event_alloc(&attr, cpu, ctx, group_leader,
5194 NULL, NULL, GFP_KERNEL);
5195 if (IS_ERR(event)) {
5196 err = PTR_ERR(event);
5197 goto err_put_context;
5201 err = perf_event_set_output(event, output_event);
5203 goto err_free_put_context;
5206 event_file = anon_inode_getfile("[perf_event]", &perf_fops, event, O_RDWR);
5207 if (IS_ERR(event_file)) {
5208 err = PTR_ERR(event_file);
5209 goto err_free_put_context;
5212 event->filp = event_file;
5213 WARN_ON_ONCE(ctx->parent_ctx);
5214 mutex_lock(&ctx->mutex);
5215 perf_install_in_context(ctx, event, cpu);
5217 mutex_unlock(&ctx->mutex);
5219 event->owner = current;
5220 get_task_struct(current);
5221 mutex_lock(¤t->perf_event_mutex);
5222 list_add_tail(&event->owner_entry, ¤t->perf_event_list);
5223 mutex_unlock(¤t->perf_event_mutex);
5226 * Drop the reference on the group_event after placing the
5227 * new event on the sibling_list. This ensures destruction
5228 * of the group leader will find the pointer to itself in
5229 * perf_group_detach().
5231 fput_light(group_file, fput_needed);
5232 fd_install(event_fd, event_file);
5235 err_free_put_context:
5238 fput_light(group_file, fput_needed);
5241 put_unused_fd(event_fd);
5246 * perf_event_create_kernel_counter
5248 * @attr: attributes of the counter to create
5249 * @cpu: cpu in which the counter is bound
5250 * @pid: task to profile
5253 perf_event_create_kernel_counter(struct perf_event_attr *attr, int cpu,
5255 perf_overflow_handler_t overflow_handler)
5257 struct perf_event *event;
5258 struct perf_event_context *ctx;
5262 * Get the target context (task or percpu):
5265 ctx = find_get_context(pid, cpu);
5271 event = perf_event_alloc(attr, cpu, ctx, NULL,
5272 NULL, overflow_handler, GFP_KERNEL);
5273 if (IS_ERR(event)) {
5274 err = PTR_ERR(event);
5275 goto err_put_context;
5279 WARN_ON_ONCE(ctx->parent_ctx);
5280 mutex_lock(&ctx->mutex);
5281 perf_install_in_context(ctx, event, cpu);
5283 mutex_unlock(&ctx->mutex);
5285 event->owner = current;
5286 get_task_struct(current);
5287 mutex_lock(¤t->perf_event_mutex);
5288 list_add_tail(&event->owner_entry, ¤t->perf_event_list);
5289 mutex_unlock(¤t->perf_event_mutex);
5296 return ERR_PTR(err);
5298 EXPORT_SYMBOL_GPL(perf_event_create_kernel_counter);
5301 * inherit a event from parent task to child task:
5303 static struct perf_event *
5304 inherit_event(struct perf_event *parent_event,
5305 struct task_struct *parent,
5306 struct perf_event_context *parent_ctx,
5307 struct task_struct *child,
5308 struct perf_event *group_leader,
5309 struct perf_event_context *child_ctx)
5311 struct perf_event *child_event;
5314 * Instead of creating recursive hierarchies of events,
5315 * we link inherited events back to the original parent,
5316 * which has a filp for sure, which we use as the reference
5319 if (parent_event->parent)
5320 parent_event = parent_event->parent;
5322 child_event = perf_event_alloc(&parent_event->attr,
5323 parent_event->cpu, child_ctx,
5324 group_leader, parent_event,
5326 if (IS_ERR(child_event))
5331 * Make the child state follow the state of the parent event,
5332 * not its attr.disabled bit. We hold the parent's mutex,
5333 * so we won't race with perf_event_{en, dis}able_family.
5335 if (parent_event->state >= PERF_EVENT_STATE_INACTIVE)
5336 child_event->state = PERF_EVENT_STATE_INACTIVE;
5338 child_event->state = PERF_EVENT_STATE_OFF;
5340 if (parent_event->attr.freq) {
5341 u64 sample_period = parent_event->hw.sample_period;
5342 struct hw_perf_event *hwc = &child_event->hw;
5344 hwc->sample_period = sample_period;
5345 hwc->last_period = sample_period;
5347 local64_set(&hwc->period_left, sample_period);
5350 child_event->overflow_handler = parent_event->overflow_handler;
5353 * Link it up in the child's context:
5355 add_event_to_ctx(child_event, child_ctx);
5358 * Get a reference to the parent filp - we will fput it
5359 * when the child event exits. This is safe to do because
5360 * we are in the parent and we know that the filp still
5361 * exists and has a nonzero count:
5363 atomic_long_inc(&parent_event->filp->f_count);
5366 * Link this into the parent event's child list
5368 WARN_ON_ONCE(parent_event->ctx->parent_ctx);
5369 mutex_lock(&parent_event->child_mutex);
5370 list_add_tail(&child_event->child_list, &parent_event->child_list);
5371 mutex_unlock(&parent_event->child_mutex);
5376 static int inherit_group(struct perf_event *parent_event,
5377 struct task_struct *parent,
5378 struct perf_event_context *parent_ctx,
5379 struct task_struct *child,
5380 struct perf_event_context *child_ctx)
5382 struct perf_event *leader;
5383 struct perf_event *sub;
5384 struct perf_event *child_ctr;
5386 leader = inherit_event(parent_event, parent, parent_ctx,
5387 child, NULL, child_ctx);
5389 return PTR_ERR(leader);
5390 list_for_each_entry(sub, &parent_event->sibling_list, group_entry) {
5391 child_ctr = inherit_event(sub, parent, parent_ctx,
5392 child, leader, child_ctx);
5393 if (IS_ERR(child_ctr))
5394 return PTR_ERR(child_ctr);
5399 static void sync_child_event(struct perf_event *child_event,
5400 struct task_struct *child)
5402 struct perf_event *parent_event = child_event->parent;
5405 if (child_event->attr.inherit_stat)
5406 perf_event_read_event(child_event, child);
5408 child_val = perf_event_count(child_event);
5411 * Add back the child's count to the parent's count:
5413 atomic64_add(child_val, &parent_event->child_count);
5414 atomic64_add(child_event->total_time_enabled,
5415 &parent_event->child_total_time_enabled);
5416 atomic64_add(child_event->total_time_running,
5417 &parent_event->child_total_time_running);
5420 * Remove this event from the parent's list
5422 WARN_ON_ONCE(parent_event->ctx->parent_ctx);
5423 mutex_lock(&parent_event->child_mutex);
5424 list_del_init(&child_event->child_list);
5425 mutex_unlock(&parent_event->child_mutex);
5428 * Release the parent event, if this was the last
5431 fput(parent_event->filp);
5435 __perf_event_exit_task(struct perf_event *child_event,
5436 struct perf_event_context *child_ctx,
5437 struct task_struct *child)
5439 struct perf_event *parent_event;
5441 perf_event_remove_from_context(child_event);
5443 parent_event = child_event->parent;
5445 * It can happen that parent exits first, and has events
5446 * that are still around due to the child reference. These
5447 * events need to be zapped - but otherwise linger.
5450 sync_child_event(child_event, child);
5451 free_event(child_event);
5456 * When a child task exits, feed back event values to parent events.
5458 void perf_event_exit_task(struct task_struct *child)
5460 struct perf_event *child_event, *tmp;
5461 struct perf_event_context *child_ctx;
5462 unsigned long flags;
5464 if (likely(!child->perf_event_ctxp)) {
5465 perf_event_task(child, NULL, 0);
5469 local_irq_save(flags);
5471 * We can't reschedule here because interrupts are disabled,
5472 * and either child is current or it is a task that can't be
5473 * scheduled, so we are now safe from rescheduling changing
5476 child_ctx = child->perf_event_ctxp;
5477 __perf_event_task_sched_out(child_ctx);
5480 * Take the context lock here so that if find_get_context is
5481 * reading child->perf_event_ctxp, we wait until it has
5482 * incremented the context's refcount before we do put_ctx below.
5484 raw_spin_lock(&child_ctx->lock);
5485 child->perf_event_ctxp = NULL;
5487 * If this context is a clone; unclone it so it can't get
5488 * swapped to another process while we're removing all
5489 * the events from it.
5491 unclone_ctx(child_ctx);
5492 update_context_time(child_ctx);
5493 raw_spin_unlock_irqrestore(&child_ctx->lock, flags);
5496 * Report the task dead after unscheduling the events so that we
5497 * won't get any samples after PERF_RECORD_EXIT. We can however still
5498 * get a few PERF_RECORD_READ events.
5500 perf_event_task(child, child_ctx, 0);
5503 * We can recurse on the same lock type through:
5505 * __perf_event_exit_task()
5506 * sync_child_event()
5507 * fput(parent_event->filp)
5509 * mutex_lock(&ctx->mutex)
5511 * But since its the parent context it won't be the same instance.
5513 mutex_lock(&child_ctx->mutex);
5516 list_for_each_entry_safe(child_event, tmp, &child_ctx->pinned_groups,
5518 __perf_event_exit_task(child_event, child_ctx, child);
5520 list_for_each_entry_safe(child_event, tmp, &child_ctx->flexible_groups,
5522 __perf_event_exit_task(child_event, child_ctx, child);
5525 * If the last event was a group event, it will have appended all
5526 * its siblings to the list, but we obtained 'tmp' before that which
5527 * will still point to the list head terminating the iteration.
5529 if (!list_empty(&child_ctx->pinned_groups) ||
5530 !list_empty(&child_ctx->flexible_groups))
5533 mutex_unlock(&child_ctx->mutex);
5538 static void perf_free_event(struct perf_event *event,
5539 struct perf_event_context *ctx)
5541 struct perf_event *parent = event->parent;
5543 if (WARN_ON_ONCE(!parent))
5546 mutex_lock(&parent->child_mutex);
5547 list_del_init(&event->child_list);
5548 mutex_unlock(&parent->child_mutex);
5552 perf_group_detach(event);
5553 list_del_event(event, ctx);
5558 * free an unexposed, unused context as created by inheritance by
5559 * init_task below, used by fork() in case of fail.
5561 void perf_event_free_task(struct task_struct *task)
5563 struct perf_event_context *ctx = task->perf_event_ctxp;
5564 struct perf_event *event, *tmp;
5569 mutex_lock(&ctx->mutex);
5571 list_for_each_entry_safe(event, tmp, &ctx->pinned_groups, group_entry)
5572 perf_free_event(event, ctx);
5574 list_for_each_entry_safe(event, tmp, &ctx->flexible_groups,
5576 perf_free_event(event, ctx);
5578 if (!list_empty(&ctx->pinned_groups) ||
5579 !list_empty(&ctx->flexible_groups))
5582 mutex_unlock(&ctx->mutex);
5588 inherit_task_group(struct perf_event *event, struct task_struct *parent,
5589 struct perf_event_context *parent_ctx,
5590 struct task_struct *child,
5594 struct perf_event_context *child_ctx = child->perf_event_ctxp;
5596 if (!event->attr.inherit) {
5603 * This is executed from the parent task context, so
5604 * inherit events that have been marked for cloning.
5605 * First allocate and initialize a context for the
5609 child_ctx = kzalloc(sizeof(struct perf_event_context),
5614 __perf_event_init_context(child_ctx, child);
5615 child->perf_event_ctxp = child_ctx;
5616 get_task_struct(child);
5619 ret = inherit_group(event, parent, parent_ctx,
5630 * Initialize the perf_event context in task_struct
5632 int perf_event_init_task(struct task_struct *child)
5634 struct perf_event_context *child_ctx, *parent_ctx;
5635 struct perf_event_context *cloned_ctx;
5636 struct perf_event *event;
5637 struct task_struct *parent = current;
5638 int inherited_all = 1;
5641 child->perf_event_ctxp = NULL;
5643 mutex_init(&child->perf_event_mutex);
5644 INIT_LIST_HEAD(&child->perf_event_list);
5646 if (likely(!parent->perf_event_ctxp))
5650 * If the parent's context is a clone, pin it so it won't get
5653 parent_ctx = perf_pin_task_context(parent);
5656 * No need to check if parent_ctx != NULL here; since we saw
5657 * it non-NULL earlier, the only reason for it to become NULL
5658 * is if we exit, and since we're currently in the middle of
5659 * a fork we can't be exiting at the same time.
5663 * Lock the parent list. No need to lock the child - not PID
5664 * hashed yet and not running, so nobody can access it.
5666 mutex_lock(&parent_ctx->mutex);
5669 * We dont have to disable NMIs - we are only looking at
5670 * the list, not manipulating it:
5672 list_for_each_entry(event, &parent_ctx->pinned_groups, group_entry) {
5673 ret = inherit_task_group(event, parent, parent_ctx, child,
5679 list_for_each_entry(event, &parent_ctx->flexible_groups, group_entry) {
5680 ret = inherit_task_group(event, parent, parent_ctx, child,
5686 child_ctx = child->perf_event_ctxp;
5688 if (child_ctx && inherited_all) {
5690 * Mark the child context as a clone of the parent
5691 * context, or of whatever the parent is a clone of.
5692 * Note that if the parent is a clone, it could get
5693 * uncloned at any point, but that doesn't matter
5694 * because the list of events and the generation
5695 * count can't have changed since we took the mutex.
5697 cloned_ctx = rcu_dereference(parent_ctx->parent_ctx);
5699 child_ctx->parent_ctx = cloned_ctx;
5700 child_ctx->parent_gen = parent_ctx->parent_gen;
5702 child_ctx->parent_ctx = parent_ctx;
5703 child_ctx->parent_gen = parent_ctx->generation;
5705 get_ctx(child_ctx->parent_ctx);
5708 mutex_unlock(&parent_ctx->mutex);
5710 perf_unpin_context(parent_ctx);
5715 static void __init perf_event_init_all_cpus(void)
5718 struct perf_cpu_context *cpuctx;
5720 for_each_possible_cpu(cpu) {
5721 cpuctx = &per_cpu(perf_cpu_context, cpu);
5722 mutex_init(&cpuctx->hlist_mutex);
5723 __perf_event_init_context(&cpuctx->ctx, NULL);
5727 static void __cpuinit perf_event_init_cpu(int cpu)
5729 struct perf_cpu_context *cpuctx;
5731 cpuctx = &per_cpu(perf_cpu_context, cpu);
5733 spin_lock(&perf_resource_lock);
5734 cpuctx->max_pertask = perf_max_events - perf_reserved_percpu;
5735 spin_unlock(&perf_resource_lock);
5737 mutex_lock(&cpuctx->hlist_mutex);
5738 if (cpuctx->hlist_refcount > 0) {
5739 struct swevent_hlist *hlist;
5741 hlist = kzalloc(sizeof(*hlist), GFP_KERNEL);
5742 WARN_ON_ONCE(!hlist);
5743 rcu_assign_pointer(cpuctx->swevent_hlist, hlist);
5745 mutex_unlock(&cpuctx->hlist_mutex);
5748 #ifdef CONFIG_HOTPLUG_CPU
5749 static void __perf_event_exit_cpu(void *info)
5751 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
5752 struct perf_event_context *ctx = &cpuctx->ctx;
5753 struct perf_event *event, *tmp;
5755 list_for_each_entry_safe(event, tmp, &ctx->pinned_groups, group_entry)
5756 __perf_event_remove_from_context(event);
5757 list_for_each_entry_safe(event, tmp, &ctx->flexible_groups, group_entry)
5758 __perf_event_remove_from_context(event);
5760 static void perf_event_exit_cpu(int cpu)
5762 struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
5763 struct perf_event_context *ctx = &cpuctx->ctx;
5765 mutex_lock(&cpuctx->hlist_mutex);
5766 swevent_hlist_release(cpuctx);
5767 mutex_unlock(&cpuctx->hlist_mutex);
5769 mutex_lock(&ctx->mutex);
5770 smp_call_function_single(cpu, __perf_event_exit_cpu, NULL, 1);
5771 mutex_unlock(&ctx->mutex);
5774 static inline void perf_event_exit_cpu(int cpu) { }
5777 static int __cpuinit
5778 perf_cpu_notify(struct notifier_block *self, unsigned long action, void *hcpu)
5780 unsigned int cpu = (long)hcpu;
5784 case CPU_UP_PREPARE:
5785 case CPU_UP_PREPARE_FROZEN:
5786 perf_event_init_cpu(cpu);
5789 case CPU_DOWN_PREPARE:
5790 case CPU_DOWN_PREPARE_FROZEN:
5791 perf_event_exit_cpu(cpu);
5802 * This has to have a higher priority than migration_notifier in sched.c.
5804 static struct notifier_block __cpuinitdata perf_cpu_nb = {
5805 .notifier_call = perf_cpu_notify,
5809 void __init perf_event_init(void)
5811 perf_event_init_all_cpus();
5812 perf_cpu_notify(&perf_cpu_nb, (unsigned long)CPU_UP_PREPARE,
5813 (void *)(long)smp_processor_id());
5814 perf_cpu_notify(&perf_cpu_nb, (unsigned long)CPU_ONLINE,
5815 (void *)(long)smp_processor_id());
5816 register_cpu_notifier(&perf_cpu_nb);
5819 static ssize_t perf_show_reserve_percpu(struct sysdev_class *class,
5820 struct sysdev_class_attribute *attr,
5823 return sprintf(buf, "%d\n", perf_reserved_percpu);
5827 perf_set_reserve_percpu(struct sysdev_class *class,
5828 struct sysdev_class_attribute *attr,
5832 struct perf_cpu_context *cpuctx;
5836 err = strict_strtoul(buf, 10, &val);
5839 if (val > perf_max_events)
5842 spin_lock(&perf_resource_lock);
5843 perf_reserved_percpu = val;
5844 for_each_online_cpu(cpu) {
5845 cpuctx = &per_cpu(perf_cpu_context, cpu);
5846 raw_spin_lock_irq(&cpuctx->ctx.lock);
5847 mpt = min(perf_max_events - cpuctx->ctx.nr_events,
5848 perf_max_events - perf_reserved_percpu);
5849 cpuctx->max_pertask = mpt;
5850 raw_spin_unlock_irq(&cpuctx->ctx.lock);
5852 spin_unlock(&perf_resource_lock);
5857 static ssize_t perf_show_overcommit(struct sysdev_class *class,
5858 struct sysdev_class_attribute *attr,
5861 return sprintf(buf, "%d\n", perf_overcommit);
5865 perf_set_overcommit(struct sysdev_class *class,
5866 struct sysdev_class_attribute *attr,
5867 const char *buf, size_t count)
5872 err = strict_strtoul(buf, 10, &val);
5878 spin_lock(&perf_resource_lock);
5879 perf_overcommit = val;
5880 spin_unlock(&perf_resource_lock);
5885 static SYSDEV_CLASS_ATTR(
5888 perf_show_reserve_percpu,
5889 perf_set_reserve_percpu
5892 static SYSDEV_CLASS_ATTR(
5895 perf_show_overcommit,
5899 static struct attribute *perfclass_attrs[] = {
5900 &attr_reserve_percpu.attr,
5901 &attr_overcommit.attr,
5905 static struct attribute_group perfclass_attr_group = {
5906 .attrs = perfclass_attrs,
5907 .name = "perf_events",
5910 static int __init perf_event_sysfs_init(void)
5912 return sysfs_create_group(&cpu_sysdev_class.kset.kobj,
5913 &perfclass_attr_group);
5915 device_initcall(perf_event_sysfs_init);