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 = atomic64_read(&next_event->count);
1152 value = atomic64_xchg(&event->count, value);
1153 atomic64_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 (atomic64_read(&hwc->period_left) > 8*sample_period) {
1545 perf_event_stop(event);
1546 atomic64_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 = atomic64_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 u64 perf_event_read(struct perf_event *event)
1742 * If event is enabled and currently active on a CPU, update the
1743 * value in the event structure:
1745 if (event->state == PERF_EVENT_STATE_ACTIVE) {
1746 smp_call_function_single(event->oncpu,
1747 __perf_event_read, event, 1);
1748 } else if (event->state == PERF_EVENT_STATE_INACTIVE) {
1749 struct perf_event_context *ctx = event->ctx;
1750 unsigned long flags;
1752 raw_spin_lock_irqsave(&ctx->lock, flags);
1753 update_context_time(ctx);
1754 update_event_times(event);
1755 raw_spin_unlock_irqrestore(&ctx->lock, flags);
1758 return atomic64_read(&event->count);
1762 * Initialize the perf_event context in a task_struct:
1765 __perf_event_init_context(struct perf_event_context *ctx,
1766 struct task_struct *task)
1768 raw_spin_lock_init(&ctx->lock);
1769 mutex_init(&ctx->mutex);
1770 INIT_LIST_HEAD(&ctx->pinned_groups);
1771 INIT_LIST_HEAD(&ctx->flexible_groups);
1772 INIT_LIST_HEAD(&ctx->event_list);
1773 atomic_set(&ctx->refcount, 1);
1777 static struct perf_event_context *find_get_context(pid_t pid, int cpu)
1779 struct perf_event_context *ctx;
1780 struct perf_cpu_context *cpuctx;
1781 struct task_struct *task;
1782 unsigned long flags;
1785 if (pid == -1 && cpu != -1) {
1786 /* Must be root to operate on a CPU event: */
1787 if (perf_paranoid_cpu() && !capable(CAP_SYS_ADMIN))
1788 return ERR_PTR(-EACCES);
1790 if (cpu < 0 || cpu >= nr_cpumask_bits)
1791 return ERR_PTR(-EINVAL);
1794 * We could be clever and allow to attach a event to an
1795 * offline CPU and activate it when the CPU comes up, but
1798 if (!cpu_online(cpu))
1799 return ERR_PTR(-ENODEV);
1801 cpuctx = &per_cpu(perf_cpu_context, cpu);
1812 task = find_task_by_vpid(pid);
1814 get_task_struct(task);
1818 return ERR_PTR(-ESRCH);
1821 * Can't attach events to a dying task.
1824 if (task->flags & PF_EXITING)
1827 /* Reuse ptrace permission checks for now. */
1829 if (!ptrace_may_access(task, PTRACE_MODE_READ))
1833 ctx = perf_lock_task_context(task, &flags);
1836 raw_spin_unlock_irqrestore(&ctx->lock, flags);
1840 ctx = kzalloc(sizeof(struct perf_event_context), GFP_KERNEL);
1844 __perf_event_init_context(ctx, task);
1846 if (cmpxchg(&task->perf_event_ctxp, NULL, ctx)) {
1848 * We raced with some other task; use
1849 * the context they set.
1854 get_task_struct(task);
1857 put_task_struct(task);
1861 put_task_struct(task);
1862 return ERR_PTR(err);
1865 static void perf_event_free_filter(struct perf_event *event);
1867 static void free_event_rcu(struct rcu_head *head)
1869 struct perf_event *event;
1871 event = container_of(head, struct perf_event, rcu_head);
1873 put_pid_ns(event->ns);
1874 perf_event_free_filter(event);
1878 static void perf_pending_sync(struct perf_event *event);
1879 static void perf_mmap_data_put(struct perf_mmap_data *data);
1881 static void free_event(struct perf_event *event)
1883 perf_pending_sync(event);
1885 if (!event->parent) {
1886 atomic_dec(&nr_events);
1887 if (event->attr.mmap || event->attr.mmap_data)
1888 atomic_dec(&nr_mmap_events);
1889 if (event->attr.comm)
1890 atomic_dec(&nr_comm_events);
1891 if (event->attr.task)
1892 atomic_dec(&nr_task_events);
1896 perf_mmap_data_put(event->data);
1901 event->destroy(event);
1903 put_ctx(event->ctx);
1904 call_rcu(&event->rcu_head, free_event_rcu);
1907 int perf_event_release_kernel(struct perf_event *event)
1909 struct perf_event_context *ctx = event->ctx;
1912 * Remove from the PMU, can't get re-enabled since we got
1913 * here because the last ref went.
1915 perf_event_disable(event);
1917 WARN_ON_ONCE(ctx->parent_ctx);
1919 * There are two ways this annotation is useful:
1921 * 1) there is a lock recursion from perf_event_exit_task
1922 * see the comment there.
1924 * 2) there is a lock-inversion with mmap_sem through
1925 * perf_event_read_group(), which takes faults while
1926 * holding ctx->mutex, however this is called after
1927 * the last filedesc died, so there is no possibility
1928 * to trigger the AB-BA case.
1930 mutex_lock_nested(&ctx->mutex, SINGLE_DEPTH_NESTING);
1931 raw_spin_lock_irq(&ctx->lock);
1932 perf_group_detach(event);
1933 list_del_event(event, ctx);
1934 raw_spin_unlock_irq(&ctx->lock);
1935 mutex_unlock(&ctx->mutex);
1937 mutex_lock(&event->owner->perf_event_mutex);
1938 list_del_init(&event->owner_entry);
1939 mutex_unlock(&event->owner->perf_event_mutex);
1940 put_task_struct(event->owner);
1946 EXPORT_SYMBOL_GPL(perf_event_release_kernel);
1949 * Called when the last reference to the file is gone.
1951 static int perf_release(struct inode *inode, struct file *file)
1953 struct perf_event *event = file->private_data;
1955 file->private_data = NULL;
1957 return perf_event_release_kernel(event);
1960 static int perf_event_read_size(struct perf_event *event)
1962 int entry = sizeof(u64); /* value */
1966 if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1967 size += sizeof(u64);
1969 if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1970 size += sizeof(u64);
1972 if (event->attr.read_format & PERF_FORMAT_ID)
1973 entry += sizeof(u64);
1975 if (event->attr.read_format & PERF_FORMAT_GROUP) {
1976 nr += event->group_leader->nr_siblings;
1977 size += sizeof(u64);
1985 u64 perf_event_read_value(struct perf_event *event, u64 *enabled, u64 *running)
1987 struct perf_event *child;
1993 mutex_lock(&event->child_mutex);
1994 total += perf_event_read(event);
1995 *enabled += event->total_time_enabled +
1996 atomic64_read(&event->child_total_time_enabled);
1997 *running += event->total_time_running +
1998 atomic64_read(&event->child_total_time_running);
2000 list_for_each_entry(child, &event->child_list, child_list) {
2001 total += perf_event_read(child);
2002 *enabled += child->total_time_enabled;
2003 *running += child->total_time_running;
2005 mutex_unlock(&event->child_mutex);
2009 EXPORT_SYMBOL_GPL(perf_event_read_value);
2011 static int perf_event_read_group(struct perf_event *event,
2012 u64 read_format, char __user *buf)
2014 struct perf_event *leader = event->group_leader, *sub;
2015 int n = 0, size = 0, ret = -EFAULT;
2016 struct perf_event_context *ctx = leader->ctx;
2018 u64 count, enabled, running;
2020 mutex_lock(&ctx->mutex);
2021 count = perf_event_read_value(leader, &enabled, &running);
2023 values[n++] = 1 + leader->nr_siblings;
2024 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
2025 values[n++] = enabled;
2026 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
2027 values[n++] = running;
2028 values[n++] = count;
2029 if (read_format & PERF_FORMAT_ID)
2030 values[n++] = primary_event_id(leader);
2032 size = n * sizeof(u64);
2034 if (copy_to_user(buf, values, size))
2039 list_for_each_entry(sub, &leader->sibling_list, group_entry) {
2042 values[n++] = perf_event_read_value(sub, &enabled, &running);
2043 if (read_format & PERF_FORMAT_ID)
2044 values[n++] = primary_event_id(sub);
2046 size = n * sizeof(u64);
2048 if (copy_to_user(buf + ret, values, size)) {
2056 mutex_unlock(&ctx->mutex);
2061 static int perf_event_read_one(struct perf_event *event,
2062 u64 read_format, char __user *buf)
2064 u64 enabled, running;
2068 values[n++] = perf_event_read_value(event, &enabled, &running);
2069 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
2070 values[n++] = enabled;
2071 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
2072 values[n++] = running;
2073 if (read_format & PERF_FORMAT_ID)
2074 values[n++] = primary_event_id(event);
2076 if (copy_to_user(buf, values, n * sizeof(u64)))
2079 return n * sizeof(u64);
2083 * Read the performance event - simple non blocking version for now
2086 perf_read_hw(struct perf_event *event, char __user *buf, size_t count)
2088 u64 read_format = event->attr.read_format;
2092 * Return end-of-file for a read on a event that is in
2093 * error state (i.e. because it was pinned but it couldn't be
2094 * scheduled on to the CPU at some point).
2096 if (event->state == PERF_EVENT_STATE_ERROR)
2099 if (count < perf_event_read_size(event))
2102 WARN_ON_ONCE(event->ctx->parent_ctx);
2103 if (read_format & PERF_FORMAT_GROUP)
2104 ret = perf_event_read_group(event, read_format, buf);
2106 ret = perf_event_read_one(event, read_format, buf);
2112 perf_read(struct file *file, char __user *buf, size_t count, loff_t *ppos)
2114 struct perf_event *event = file->private_data;
2116 return perf_read_hw(event, buf, count);
2119 static unsigned int perf_poll(struct file *file, poll_table *wait)
2121 struct perf_event *event = file->private_data;
2122 struct perf_mmap_data *data;
2123 unsigned int events = POLL_HUP;
2126 data = rcu_dereference(event->data);
2128 events = atomic_xchg(&data->poll, 0);
2131 poll_wait(file, &event->waitq, wait);
2136 static void perf_event_reset(struct perf_event *event)
2138 (void)perf_event_read(event);
2139 atomic64_set(&event->count, 0);
2140 perf_event_update_userpage(event);
2144 * Holding the top-level event's child_mutex means that any
2145 * descendant process that has inherited this event will block
2146 * in sync_child_event if it goes to exit, thus satisfying the
2147 * task existence requirements of perf_event_enable/disable.
2149 static void perf_event_for_each_child(struct perf_event *event,
2150 void (*func)(struct perf_event *))
2152 struct perf_event *child;
2154 WARN_ON_ONCE(event->ctx->parent_ctx);
2155 mutex_lock(&event->child_mutex);
2157 list_for_each_entry(child, &event->child_list, child_list)
2159 mutex_unlock(&event->child_mutex);
2162 static void perf_event_for_each(struct perf_event *event,
2163 void (*func)(struct perf_event *))
2165 struct perf_event_context *ctx = event->ctx;
2166 struct perf_event *sibling;
2168 WARN_ON_ONCE(ctx->parent_ctx);
2169 mutex_lock(&ctx->mutex);
2170 event = event->group_leader;
2172 perf_event_for_each_child(event, func);
2174 list_for_each_entry(sibling, &event->sibling_list, group_entry)
2175 perf_event_for_each_child(event, func);
2176 mutex_unlock(&ctx->mutex);
2179 static int perf_event_period(struct perf_event *event, u64 __user *arg)
2181 struct perf_event_context *ctx = event->ctx;
2186 if (!event->attr.sample_period)
2189 size = copy_from_user(&value, arg, sizeof(value));
2190 if (size != sizeof(value))
2196 raw_spin_lock_irq(&ctx->lock);
2197 if (event->attr.freq) {
2198 if (value > sysctl_perf_event_sample_rate) {
2203 event->attr.sample_freq = value;
2205 event->attr.sample_period = value;
2206 event->hw.sample_period = value;
2209 raw_spin_unlock_irq(&ctx->lock);
2214 static const struct file_operations perf_fops;
2216 static struct perf_event *perf_fget_light(int fd, int *fput_needed)
2220 file = fget_light(fd, fput_needed);
2222 return ERR_PTR(-EBADF);
2224 if (file->f_op != &perf_fops) {
2225 fput_light(file, *fput_needed);
2227 return ERR_PTR(-EBADF);
2230 return file->private_data;
2233 static int perf_event_set_output(struct perf_event *event,
2234 struct perf_event *output_event);
2235 static int perf_event_set_filter(struct perf_event *event, void __user *arg);
2237 static long perf_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
2239 struct perf_event *event = file->private_data;
2240 void (*func)(struct perf_event *);
2244 case PERF_EVENT_IOC_ENABLE:
2245 func = perf_event_enable;
2247 case PERF_EVENT_IOC_DISABLE:
2248 func = perf_event_disable;
2250 case PERF_EVENT_IOC_RESET:
2251 func = perf_event_reset;
2254 case PERF_EVENT_IOC_REFRESH:
2255 return perf_event_refresh(event, arg);
2257 case PERF_EVENT_IOC_PERIOD:
2258 return perf_event_period(event, (u64 __user *)arg);
2260 case PERF_EVENT_IOC_SET_OUTPUT:
2262 struct perf_event *output_event = NULL;
2263 int fput_needed = 0;
2267 output_event = perf_fget_light(arg, &fput_needed);
2268 if (IS_ERR(output_event))
2269 return PTR_ERR(output_event);
2272 ret = perf_event_set_output(event, output_event);
2274 fput_light(output_event->filp, fput_needed);
2279 case PERF_EVENT_IOC_SET_FILTER:
2280 return perf_event_set_filter(event, (void __user *)arg);
2286 if (flags & PERF_IOC_FLAG_GROUP)
2287 perf_event_for_each(event, func);
2289 perf_event_for_each_child(event, func);
2294 int perf_event_task_enable(void)
2296 struct perf_event *event;
2298 mutex_lock(¤t->perf_event_mutex);
2299 list_for_each_entry(event, ¤t->perf_event_list, owner_entry)
2300 perf_event_for_each_child(event, perf_event_enable);
2301 mutex_unlock(¤t->perf_event_mutex);
2306 int perf_event_task_disable(void)
2308 struct perf_event *event;
2310 mutex_lock(¤t->perf_event_mutex);
2311 list_for_each_entry(event, ¤t->perf_event_list, owner_entry)
2312 perf_event_for_each_child(event, perf_event_disable);
2313 mutex_unlock(¤t->perf_event_mutex);
2318 #ifndef PERF_EVENT_INDEX_OFFSET
2319 # define PERF_EVENT_INDEX_OFFSET 0
2322 static int perf_event_index(struct perf_event *event)
2324 if (event->state != PERF_EVENT_STATE_ACTIVE)
2327 return event->hw.idx + 1 - PERF_EVENT_INDEX_OFFSET;
2331 * Callers need to ensure there can be no nesting of this function, otherwise
2332 * the seqlock logic goes bad. We can not serialize this because the arch
2333 * code calls this from NMI context.
2335 void perf_event_update_userpage(struct perf_event *event)
2337 struct perf_event_mmap_page *userpg;
2338 struct perf_mmap_data *data;
2341 data = rcu_dereference(event->data);
2345 userpg = data->user_page;
2348 * Disable preemption so as to not let the corresponding user-space
2349 * spin too long if we get preempted.
2354 userpg->index = perf_event_index(event);
2355 userpg->offset = atomic64_read(&event->count);
2356 if (event->state == PERF_EVENT_STATE_ACTIVE)
2357 userpg->offset -= atomic64_read(&event->hw.prev_count);
2359 userpg->time_enabled = event->total_time_enabled +
2360 atomic64_read(&event->child_total_time_enabled);
2362 userpg->time_running = event->total_time_running +
2363 atomic64_read(&event->child_total_time_running);
2372 #ifndef CONFIG_PERF_USE_VMALLOC
2375 * Back perf_mmap() with regular GFP_KERNEL-0 pages.
2378 static struct page *
2379 perf_mmap_to_page(struct perf_mmap_data *data, unsigned long pgoff)
2381 if (pgoff > data->nr_pages)
2385 return virt_to_page(data->user_page);
2387 return virt_to_page(data->data_pages[pgoff - 1]);
2390 static void *perf_mmap_alloc_page(int cpu)
2395 node = (cpu == -1) ? cpu : cpu_to_node(cpu);
2396 page = alloc_pages_node(node, GFP_KERNEL | __GFP_ZERO, 0);
2400 return page_address(page);
2403 static struct perf_mmap_data *
2404 perf_mmap_data_alloc(struct perf_event *event, int nr_pages)
2406 struct perf_mmap_data *data;
2410 size = sizeof(struct perf_mmap_data);
2411 size += nr_pages * sizeof(void *);
2413 data = kzalloc(size, GFP_KERNEL);
2417 data->user_page = perf_mmap_alloc_page(event->cpu);
2418 if (!data->user_page)
2419 goto fail_user_page;
2421 for (i = 0; i < nr_pages; i++) {
2422 data->data_pages[i] = perf_mmap_alloc_page(event->cpu);
2423 if (!data->data_pages[i])
2424 goto fail_data_pages;
2427 data->nr_pages = nr_pages;
2432 for (i--; i >= 0; i--)
2433 free_page((unsigned long)data->data_pages[i]);
2435 free_page((unsigned long)data->user_page);
2444 static void perf_mmap_free_page(unsigned long addr)
2446 struct page *page = virt_to_page((void *)addr);
2448 page->mapping = NULL;
2452 static void perf_mmap_data_free(struct perf_mmap_data *data)
2456 perf_mmap_free_page((unsigned long)data->user_page);
2457 for (i = 0; i < data->nr_pages; i++)
2458 perf_mmap_free_page((unsigned long)data->data_pages[i]);
2462 static inline int page_order(struct perf_mmap_data *data)
2470 * Back perf_mmap() with vmalloc memory.
2472 * Required for architectures that have d-cache aliasing issues.
2475 static inline int page_order(struct perf_mmap_data *data)
2477 return data->page_order;
2480 static struct page *
2481 perf_mmap_to_page(struct perf_mmap_data *data, unsigned long pgoff)
2483 if (pgoff > (1UL << page_order(data)))
2486 return vmalloc_to_page((void *)data->user_page + pgoff * PAGE_SIZE);
2489 static void perf_mmap_unmark_page(void *addr)
2491 struct page *page = vmalloc_to_page(addr);
2493 page->mapping = NULL;
2496 static void perf_mmap_data_free_work(struct work_struct *work)
2498 struct perf_mmap_data *data;
2502 data = container_of(work, struct perf_mmap_data, work);
2503 nr = 1 << page_order(data);
2505 base = data->user_page;
2506 for (i = 0; i < nr + 1; i++)
2507 perf_mmap_unmark_page(base + (i * PAGE_SIZE));
2513 static void perf_mmap_data_free(struct perf_mmap_data *data)
2515 schedule_work(&data->work);
2518 static struct perf_mmap_data *
2519 perf_mmap_data_alloc(struct perf_event *event, int nr_pages)
2521 struct perf_mmap_data *data;
2525 size = sizeof(struct perf_mmap_data);
2526 size += sizeof(void *);
2528 data = kzalloc(size, GFP_KERNEL);
2532 INIT_WORK(&data->work, perf_mmap_data_free_work);
2534 all_buf = vmalloc_user((nr_pages + 1) * PAGE_SIZE);
2538 data->user_page = all_buf;
2539 data->data_pages[0] = all_buf + PAGE_SIZE;
2540 data->page_order = ilog2(nr_pages);
2554 static unsigned long perf_data_size(struct perf_mmap_data *data)
2556 return data->nr_pages << (PAGE_SHIFT + page_order(data));
2559 static int perf_mmap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
2561 struct perf_event *event = vma->vm_file->private_data;
2562 struct perf_mmap_data *data;
2563 int ret = VM_FAULT_SIGBUS;
2565 if (vmf->flags & FAULT_FLAG_MKWRITE) {
2566 if (vmf->pgoff == 0)
2572 data = rcu_dereference(event->data);
2576 if (vmf->pgoff && (vmf->flags & FAULT_FLAG_WRITE))
2579 vmf->page = perf_mmap_to_page(data, vmf->pgoff);
2583 get_page(vmf->page);
2584 vmf->page->mapping = vma->vm_file->f_mapping;
2585 vmf->page->index = vmf->pgoff;
2595 perf_mmap_data_init(struct perf_event *event, struct perf_mmap_data *data)
2597 long max_size = perf_data_size(data);
2599 if (event->attr.watermark) {
2600 data->watermark = min_t(long, max_size,
2601 event->attr.wakeup_watermark);
2604 if (!data->watermark)
2605 data->watermark = max_size / 2;
2607 atomic_set(&data->refcount, 1);
2608 rcu_assign_pointer(event->data, data);
2611 static void perf_mmap_data_free_rcu(struct rcu_head *rcu_head)
2613 struct perf_mmap_data *data;
2615 data = container_of(rcu_head, struct perf_mmap_data, rcu_head);
2616 perf_mmap_data_free(data);
2619 static struct perf_mmap_data *perf_mmap_data_get(struct perf_event *event)
2621 struct perf_mmap_data *data;
2624 data = rcu_dereference(event->data);
2626 if (!atomic_inc_not_zero(&data->refcount))
2634 static void perf_mmap_data_put(struct perf_mmap_data *data)
2636 if (!atomic_dec_and_test(&data->refcount))
2639 call_rcu(&data->rcu_head, perf_mmap_data_free_rcu);
2642 static void perf_mmap_open(struct vm_area_struct *vma)
2644 struct perf_event *event = vma->vm_file->private_data;
2646 atomic_inc(&event->mmap_count);
2649 static void perf_mmap_close(struct vm_area_struct *vma)
2651 struct perf_event *event = vma->vm_file->private_data;
2653 if (atomic_dec_and_mutex_lock(&event->mmap_count, &event->mmap_mutex)) {
2654 unsigned long size = perf_data_size(event->data);
2655 struct user_struct *user = event->mmap_user;
2656 struct perf_mmap_data *data = event->data;
2658 atomic_long_sub((size >> PAGE_SHIFT) + 1, &user->locked_vm);
2659 vma->vm_mm->locked_vm -= event->mmap_locked;
2660 rcu_assign_pointer(event->data, NULL);
2661 mutex_unlock(&event->mmap_mutex);
2663 perf_mmap_data_put(data);
2668 static const struct vm_operations_struct perf_mmap_vmops = {
2669 .open = perf_mmap_open,
2670 .close = perf_mmap_close,
2671 .fault = perf_mmap_fault,
2672 .page_mkwrite = perf_mmap_fault,
2675 static int perf_mmap(struct file *file, struct vm_area_struct *vma)
2677 struct perf_event *event = file->private_data;
2678 unsigned long user_locked, user_lock_limit;
2679 struct user_struct *user = current_user();
2680 unsigned long locked, lock_limit;
2681 struct perf_mmap_data *data;
2682 unsigned long vma_size;
2683 unsigned long nr_pages;
2684 long user_extra, extra;
2688 * Don't allow mmap() of inherited per-task counters. This would
2689 * create a performance issue due to all children writing to the
2692 if (event->cpu == -1 && event->attr.inherit)
2695 if (!(vma->vm_flags & VM_SHARED))
2698 vma_size = vma->vm_end - vma->vm_start;
2699 nr_pages = (vma_size / PAGE_SIZE) - 1;
2702 * If we have data pages ensure they're a power-of-two number, so we
2703 * can do bitmasks instead of modulo.
2705 if (nr_pages != 0 && !is_power_of_2(nr_pages))
2708 if (vma_size != PAGE_SIZE * (1 + nr_pages))
2711 if (vma->vm_pgoff != 0)
2714 WARN_ON_ONCE(event->ctx->parent_ctx);
2715 mutex_lock(&event->mmap_mutex);
2717 if (event->data->nr_pages == nr_pages)
2718 atomic_inc(&event->data->refcount);
2724 user_extra = nr_pages + 1;
2725 user_lock_limit = sysctl_perf_event_mlock >> (PAGE_SHIFT - 10);
2728 * Increase the limit linearly with more CPUs:
2730 user_lock_limit *= num_online_cpus();
2732 user_locked = atomic_long_read(&user->locked_vm) + user_extra;
2735 if (user_locked > user_lock_limit)
2736 extra = user_locked - user_lock_limit;
2738 lock_limit = rlimit(RLIMIT_MEMLOCK);
2739 lock_limit >>= PAGE_SHIFT;
2740 locked = vma->vm_mm->locked_vm + extra;
2742 if ((locked > lock_limit) && perf_paranoid_tracepoint_raw() &&
2743 !capable(CAP_IPC_LOCK)) {
2748 WARN_ON(event->data);
2750 data = perf_mmap_data_alloc(event, nr_pages);
2756 perf_mmap_data_init(event, data);
2757 if (vma->vm_flags & VM_WRITE)
2758 event->data->writable = 1;
2760 atomic_long_add(user_extra, &user->locked_vm);
2761 event->mmap_locked = extra;
2762 event->mmap_user = get_current_user();
2763 vma->vm_mm->locked_vm += event->mmap_locked;
2767 atomic_inc(&event->mmap_count);
2768 mutex_unlock(&event->mmap_mutex);
2770 vma->vm_flags |= VM_RESERVED;
2771 vma->vm_ops = &perf_mmap_vmops;
2776 static int perf_fasync(int fd, struct file *filp, int on)
2778 struct inode *inode = filp->f_path.dentry->d_inode;
2779 struct perf_event *event = filp->private_data;
2782 mutex_lock(&inode->i_mutex);
2783 retval = fasync_helper(fd, filp, on, &event->fasync);
2784 mutex_unlock(&inode->i_mutex);
2792 static const struct file_operations perf_fops = {
2793 .llseek = no_llseek,
2794 .release = perf_release,
2797 .unlocked_ioctl = perf_ioctl,
2798 .compat_ioctl = perf_ioctl,
2800 .fasync = perf_fasync,
2806 * If there's data, ensure we set the poll() state and publish everything
2807 * to user-space before waking everybody up.
2810 void perf_event_wakeup(struct perf_event *event)
2812 wake_up_all(&event->waitq);
2814 if (event->pending_kill) {
2815 kill_fasync(&event->fasync, SIGIO, event->pending_kill);
2816 event->pending_kill = 0;
2823 * Handle the case where we need to wakeup up from NMI (or rq->lock) context.
2825 * The NMI bit means we cannot possibly take locks. Therefore, maintain a
2826 * single linked list and use cmpxchg() to add entries lockless.
2829 static void perf_pending_event(struct perf_pending_entry *entry)
2831 struct perf_event *event = container_of(entry,
2832 struct perf_event, pending);
2834 if (event->pending_disable) {
2835 event->pending_disable = 0;
2836 __perf_event_disable(event);
2839 if (event->pending_wakeup) {
2840 event->pending_wakeup = 0;
2841 perf_event_wakeup(event);
2845 #define PENDING_TAIL ((struct perf_pending_entry *)-1UL)
2847 static DEFINE_PER_CPU(struct perf_pending_entry *, perf_pending_head) = {
2851 static void perf_pending_queue(struct perf_pending_entry *entry,
2852 void (*func)(struct perf_pending_entry *))
2854 struct perf_pending_entry **head;
2856 if (cmpxchg(&entry->next, NULL, PENDING_TAIL) != NULL)
2861 head = &get_cpu_var(perf_pending_head);
2864 entry->next = *head;
2865 } while (cmpxchg(head, entry->next, entry) != entry->next);
2867 set_perf_event_pending();
2869 put_cpu_var(perf_pending_head);
2872 static int __perf_pending_run(void)
2874 struct perf_pending_entry *list;
2877 list = xchg(&__get_cpu_var(perf_pending_head), PENDING_TAIL);
2878 while (list != PENDING_TAIL) {
2879 void (*func)(struct perf_pending_entry *);
2880 struct perf_pending_entry *entry = list;
2887 * Ensure we observe the unqueue before we issue the wakeup,
2888 * so that we won't be waiting forever.
2889 * -- see perf_not_pending().
2900 static inline int perf_not_pending(struct perf_event *event)
2903 * If we flush on whatever cpu we run, there is a chance we don't
2907 __perf_pending_run();
2911 * Ensure we see the proper queue state before going to sleep
2912 * so that we do not miss the wakeup. -- see perf_pending_handle()
2915 return event->pending.next == NULL;
2918 static void perf_pending_sync(struct perf_event *event)
2920 wait_event(event->waitq, perf_not_pending(event));
2923 void perf_event_do_pending(void)
2925 __perf_pending_run();
2929 * Callchain support -- arch specific
2932 __weak struct perf_callchain_entry *perf_callchain(struct pt_regs *regs)
2938 void perf_arch_fetch_caller_regs(struct pt_regs *regs, unsigned long ip, int skip)
2944 * We assume there is only KVM supporting the callbacks.
2945 * Later on, we might change it to a list if there is
2946 * another virtualization implementation supporting the callbacks.
2948 struct perf_guest_info_callbacks *perf_guest_cbs;
2950 int perf_register_guest_info_callbacks(struct perf_guest_info_callbacks *cbs)
2952 perf_guest_cbs = cbs;
2955 EXPORT_SYMBOL_GPL(perf_register_guest_info_callbacks);
2957 int perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks *cbs)
2959 perf_guest_cbs = NULL;
2962 EXPORT_SYMBOL_GPL(perf_unregister_guest_info_callbacks);
2967 static bool perf_output_space(struct perf_mmap_data *data, unsigned long tail,
2968 unsigned long offset, unsigned long head)
2972 if (!data->writable)
2975 mask = perf_data_size(data) - 1;
2977 offset = (offset - tail) & mask;
2978 head = (head - tail) & mask;
2980 if ((int)(head - offset) < 0)
2986 static void perf_output_wakeup(struct perf_output_handle *handle)
2988 atomic_set(&handle->data->poll, POLL_IN);
2991 handle->event->pending_wakeup = 1;
2992 perf_pending_queue(&handle->event->pending,
2993 perf_pending_event);
2995 perf_event_wakeup(handle->event);
2999 * We need to ensure a later event_id doesn't publish a head when a former
3000 * event isn't done writing. However since we need to deal with NMIs we
3001 * cannot fully serialize things.
3003 * We only publish the head (and generate a wakeup) when the outer-most
3006 static void perf_output_get_handle(struct perf_output_handle *handle)
3008 struct perf_mmap_data *data = handle->data;
3011 local_inc(&data->nest);
3012 handle->wakeup = local_read(&data->wakeup);
3015 static void perf_output_put_handle(struct perf_output_handle *handle)
3017 struct perf_mmap_data *data = handle->data;
3021 head = local_read(&data->head);
3024 * IRQ/NMI can happen here, which means we can miss a head update.
3027 if (!local_dec_and_test(&data->nest))
3031 * Publish the known good head. Rely on the full barrier implied
3032 * by atomic_dec_and_test() order the data->head read and this
3035 data->user_page->data_head = head;
3038 * Now check if we missed an update, rely on the (compiler)
3039 * barrier in atomic_dec_and_test() to re-read data->head.
3041 if (unlikely(head != local_read(&data->head))) {
3042 local_inc(&data->nest);
3046 if (handle->wakeup != local_read(&data->wakeup))
3047 perf_output_wakeup(handle);
3053 __always_inline void perf_output_copy(struct perf_output_handle *handle,
3054 const void *buf, unsigned int len)
3057 unsigned long size = min_t(unsigned long, handle->size, len);
3059 memcpy(handle->addr, buf, size);
3062 handle->addr += size;
3064 handle->size -= size;
3065 if (!handle->size) {
3066 struct perf_mmap_data *data = handle->data;
3069 handle->page &= data->nr_pages - 1;
3070 handle->addr = data->data_pages[handle->page];
3071 handle->size = PAGE_SIZE << page_order(data);
3076 int perf_output_begin(struct perf_output_handle *handle,
3077 struct perf_event *event, unsigned int size,
3078 int nmi, int sample)
3080 struct perf_mmap_data *data;
3081 unsigned long tail, offset, head;
3084 struct perf_event_header header;
3091 * For inherited events we send all the output towards the parent.
3094 event = event->parent;
3096 data = rcu_dereference(event->data);
3100 handle->data = data;
3101 handle->event = event;
3103 handle->sample = sample;
3105 if (!data->nr_pages)
3108 have_lost = local_read(&data->lost);
3110 size += sizeof(lost_event);
3112 perf_output_get_handle(handle);
3116 * Userspace could choose to issue a mb() before updating the
3117 * tail pointer. So that all reads will be completed before the
3120 tail = ACCESS_ONCE(data->user_page->data_tail);
3122 offset = head = local_read(&data->head);
3124 if (unlikely(!perf_output_space(data, tail, offset, head)))
3126 } while (local_cmpxchg(&data->head, offset, head) != offset);
3128 if (head - local_read(&data->wakeup) > data->watermark)
3129 local_add(data->watermark, &data->wakeup);
3131 handle->page = offset >> (PAGE_SHIFT + page_order(data));
3132 handle->page &= data->nr_pages - 1;
3133 handle->size = offset & ((PAGE_SIZE << page_order(data)) - 1);
3134 handle->addr = data->data_pages[handle->page];
3135 handle->addr += handle->size;
3136 handle->size = (PAGE_SIZE << page_order(data)) - handle->size;
3139 lost_event.header.type = PERF_RECORD_LOST;
3140 lost_event.header.misc = 0;
3141 lost_event.header.size = sizeof(lost_event);
3142 lost_event.id = event->id;
3143 lost_event.lost = local_xchg(&data->lost, 0);
3145 perf_output_put(handle, lost_event);
3151 local_inc(&data->lost);
3152 perf_output_put_handle(handle);
3159 void perf_output_end(struct perf_output_handle *handle)
3161 struct perf_event *event = handle->event;
3162 struct perf_mmap_data *data = handle->data;
3164 int wakeup_events = event->attr.wakeup_events;
3166 if (handle->sample && wakeup_events) {
3167 int events = local_inc_return(&data->events);
3168 if (events >= wakeup_events) {
3169 local_sub(wakeup_events, &data->events);
3170 local_inc(&data->wakeup);
3174 perf_output_put_handle(handle);
3178 static u32 perf_event_pid(struct perf_event *event, struct task_struct *p)
3181 * only top level events have the pid namespace they were created in
3184 event = event->parent;
3186 return task_tgid_nr_ns(p, event->ns);
3189 static u32 perf_event_tid(struct perf_event *event, struct task_struct *p)
3192 * only top level events have the pid namespace they were created in
3195 event = event->parent;
3197 return task_pid_nr_ns(p, event->ns);
3200 static void perf_output_read_one(struct perf_output_handle *handle,
3201 struct perf_event *event)
3203 u64 read_format = event->attr.read_format;
3207 values[n++] = atomic64_read(&event->count);
3208 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
3209 values[n++] = event->total_time_enabled +
3210 atomic64_read(&event->child_total_time_enabled);
3212 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
3213 values[n++] = event->total_time_running +
3214 atomic64_read(&event->child_total_time_running);
3216 if (read_format & PERF_FORMAT_ID)
3217 values[n++] = primary_event_id(event);
3219 perf_output_copy(handle, values, n * sizeof(u64));
3223 * XXX PERF_FORMAT_GROUP vs inherited events seems difficult.
3225 static void perf_output_read_group(struct perf_output_handle *handle,
3226 struct perf_event *event)
3228 struct perf_event *leader = event->group_leader, *sub;
3229 u64 read_format = event->attr.read_format;
3233 values[n++] = 1 + leader->nr_siblings;
3235 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
3236 values[n++] = leader->total_time_enabled;
3238 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
3239 values[n++] = leader->total_time_running;
3241 if (leader != event)
3242 leader->pmu->read(leader);
3244 values[n++] = atomic64_read(&leader->count);
3245 if (read_format & PERF_FORMAT_ID)
3246 values[n++] = primary_event_id(leader);
3248 perf_output_copy(handle, values, n * sizeof(u64));
3250 list_for_each_entry(sub, &leader->sibling_list, group_entry) {
3254 sub->pmu->read(sub);
3256 values[n++] = atomic64_read(&sub->count);
3257 if (read_format & PERF_FORMAT_ID)
3258 values[n++] = primary_event_id(sub);
3260 perf_output_copy(handle, values, n * sizeof(u64));
3264 static void perf_output_read(struct perf_output_handle *handle,
3265 struct perf_event *event)
3267 if (event->attr.read_format & PERF_FORMAT_GROUP)
3268 perf_output_read_group(handle, event);
3270 perf_output_read_one(handle, event);
3273 void perf_output_sample(struct perf_output_handle *handle,
3274 struct perf_event_header *header,
3275 struct perf_sample_data *data,
3276 struct perf_event *event)
3278 u64 sample_type = data->type;
3280 perf_output_put(handle, *header);
3282 if (sample_type & PERF_SAMPLE_IP)
3283 perf_output_put(handle, data->ip);
3285 if (sample_type & PERF_SAMPLE_TID)
3286 perf_output_put(handle, data->tid_entry);
3288 if (sample_type & PERF_SAMPLE_TIME)
3289 perf_output_put(handle, data->time);
3291 if (sample_type & PERF_SAMPLE_ADDR)
3292 perf_output_put(handle, data->addr);
3294 if (sample_type & PERF_SAMPLE_ID)
3295 perf_output_put(handle, data->id);
3297 if (sample_type & PERF_SAMPLE_STREAM_ID)
3298 perf_output_put(handle, data->stream_id);
3300 if (sample_type & PERF_SAMPLE_CPU)
3301 perf_output_put(handle, data->cpu_entry);
3303 if (sample_type & PERF_SAMPLE_PERIOD)
3304 perf_output_put(handle, data->period);
3306 if (sample_type & PERF_SAMPLE_READ)
3307 perf_output_read(handle, event);
3309 if (sample_type & PERF_SAMPLE_CALLCHAIN) {
3310 if (data->callchain) {
3313 if (data->callchain)
3314 size += data->callchain->nr;
3316 size *= sizeof(u64);
3318 perf_output_copy(handle, data->callchain, size);
3321 perf_output_put(handle, nr);
3325 if (sample_type & PERF_SAMPLE_RAW) {
3327 perf_output_put(handle, data->raw->size);
3328 perf_output_copy(handle, data->raw->data,
3335 .size = sizeof(u32),
3338 perf_output_put(handle, raw);
3343 void perf_prepare_sample(struct perf_event_header *header,
3344 struct perf_sample_data *data,
3345 struct perf_event *event,
3346 struct pt_regs *regs)
3348 u64 sample_type = event->attr.sample_type;
3350 data->type = sample_type;
3352 header->type = PERF_RECORD_SAMPLE;
3353 header->size = sizeof(*header);
3356 header->misc |= perf_misc_flags(regs);
3358 if (sample_type & PERF_SAMPLE_IP) {
3359 data->ip = perf_instruction_pointer(regs);
3361 header->size += sizeof(data->ip);
3364 if (sample_type & PERF_SAMPLE_TID) {
3365 /* namespace issues */
3366 data->tid_entry.pid = perf_event_pid(event, current);
3367 data->tid_entry.tid = perf_event_tid(event, current);
3369 header->size += sizeof(data->tid_entry);
3372 if (sample_type & PERF_SAMPLE_TIME) {
3373 data->time = perf_clock();
3375 header->size += sizeof(data->time);
3378 if (sample_type & PERF_SAMPLE_ADDR)
3379 header->size += sizeof(data->addr);
3381 if (sample_type & PERF_SAMPLE_ID) {
3382 data->id = primary_event_id(event);
3384 header->size += sizeof(data->id);
3387 if (sample_type & PERF_SAMPLE_STREAM_ID) {
3388 data->stream_id = event->id;
3390 header->size += sizeof(data->stream_id);
3393 if (sample_type & PERF_SAMPLE_CPU) {
3394 data->cpu_entry.cpu = raw_smp_processor_id();
3395 data->cpu_entry.reserved = 0;
3397 header->size += sizeof(data->cpu_entry);
3400 if (sample_type & PERF_SAMPLE_PERIOD)
3401 header->size += sizeof(data->period);
3403 if (sample_type & PERF_SAMPLE_READ)
3404 header->size += perf_event_read_size(event);
3406 if (sample_type & PERF_SAMPLE_CALLCHAIN) {
3409 data->callchain = perf_callchain(regs);
3411 if (data->callchain)
3412 size += data->callchain->nr;
3414 header->size += size * sizeof(u64);
3417 if (sample_type & PERF_SAMPLE_RAW) {
3418 int size = sizeof(u32);
3421 size += data->raw->size;
3423 size += sizeof(u32);
3425 WARN_ON_ONCE(size & (sizeof(u64)-1));
3426 header->size += size;
3430 static void perf_event_output(struct perf_event *event, int nmi,
3431 struct perf_sample_data *data,
3432 struct pt_regs *regs)
3434 struct perf_output_handle handle;
3435 struct perf_event_header header;
3437 perf_prepare_sample(&header, data, event, regs);
3439 if (perf_output_begin(&handle, event, header.size, nmi, 1))
3442 perf_output_sample(&handle, &header, data, event);
3444 perf_output_end(&handle);
3451 struct perf_read_event {
3452 struct perf_event_header header;
3459 perf_event_read_event(struct perf_event *event,
3460 struct task_struct *task)
3462 struct perf_output_handle handle;
3463 struct perf_read_event read_event = {
3465 .type = PERF_RECORD_READ,
3467 .size = sizeof(read_event) + perf_event_read_size(event),
3469 .pid = perf_event_pid(event, task),
3470 .tid = perf_event_tid(event, task),
3474 ret = perf_output_begin(&handle, event, read_event.header.size, 0, 0);
3478 perf_output_put(&handle, read_event);
3479 perf_output_read(&handle, event);
3481 perf_output_end(&handle);
3485 * task tracking -- fork/exit
3487 * enabled by: attr.comm | attr.mmap | attr.mmap_data | attr.task
3490 struct perf_task_event {
3491 struct task_struct *task;
3492 struct perf_event_context *task_ctx;
3495 struct perf_event_header header;
3505 static void perf_event_task_output(struct perf_event *event,
3506 struct perf_task_event *task_event)
3508 struct perf_output_handle handle;
3509 struct task_struct *task = task_event->task;
3512 size = task_event->event_id.header.size;
3513 ret = perf_output_begin(&handle, event, size, 0, 0);
3518 task_event->event_id.pid = perf_event_pid(event, task);
3519 task_event->event_id.ppid = perf_event_pid(event, current);
3521 task_event->event_id.tid = perf_event_tid(event, task);
3522 task_event->event_id.ptid = perf_event_tid(event, current);
3524 perf_output_put(&handle, task_event->event_id);
3526 perf_output_end(&handle);
3529 static int perf_event_task_match(struct perf_event *event)
3531 if (event->state < PERF_EVENT_STATE_INACTIVE)
3534 if (event->cpu != -1 && event->cpu != smp_processor_id())
3537 if (event->attr.comm || event->attr.mmap ||
3538 event->attr.mmap_data || event->attr.task)
3544 static void perf_event_task_ctx(struct perf_event_context *ctx,
3545 struct perf_task_event *task_event)
3547 struct perf_event *event;
3549 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
3550 if (perf_event_task_match(event))
3551 perf_event_task_output(event, task_event);
3555 static void perf_event_task_event(struct perf_task_event *task_event)
3557 struct perf_cpu_context *cpuctx;
3558 struct perf_event_context *ctx = task_event->task_ctx;
3561 cpuctx = &get_cpu_var(perf_cpu_context);
3562 perf_event_task_ctx(&cpuctx->ctx, task_event);
3564 ctx = rcu_dereference(current->perf_event_ctxp);
3566 perf_event_task_ctx(ctx, task_event);
3567 put_cpu_var(perf_cpu_context);
3571 static void perf_event_task(struct task_struct *task,
3572 struct perf_event_context *task_ctx,
3575 struct perf_task_event task_event;
3577 if (!atomic_read(&nr_comm_events) &&
3578 !atomic_read(&nr_mmap_events) &&
3579 !atomic_read(&nr_task_events))
3582 task_event = (struct perf_task_event){
3584 .task_ctx = task_ctx,
3587 .type = new ? PERF_RECORD_FORK : PERF_RECORD_EXIT,
3589 .size = sizeof(task_event.event_id),
3595 .time = perf_clock(),
3599 perf_event_task_event(&task_event);
3602 void perf_event_fork(struct task_struct *task)
3604 perf_event_task(task, NULL, 1);
3611 struct perf_comm_event {
3612 struct task_struct *task;
3617 struct perf_event_header header;
3624 static void perf_event_comm_output(struct perf_event *event,
3625 struct perf_comm_event *comm_event)
3627 struct perf_output_handle handle;
3628 int size = comm_event->event_id.header.size;
3629 int ret = perf_output_begin(&handle, event, size, 0, 0);
3634 comm_event->event_id.pid = perf_event_pid(event, comm_event->task);
3635 comm_event->event_id.tid = perf_event_tid(event, comm_event->task);
3637 perf_output_put(&handle, comm_event->event_id);
3638 perf_output_copy(&handle, comm_event->comm,
3639 comm_event->comm_size);
3640 perf_output_end(&handle);
3643 static int perf_event_comm_match(struct perf_event *event)
3645 if (event->state < PERF_EVENT_STATE_INACTIVE)
3648 if (event->cpu != -1 && event->cpu != smp_processor_id())
3651 if (event->attr.comm)
3657 static void perf_event_comm_ctx(struct perf_event_context *ctx,
3658 struct perf_comm_event *comm_event)
3660 struct perf_event *event;
3662 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
3663 if (perf_event_comm_match(event))
3664 perf_event_comm_output(event, comm_event);
3668 static void perf_event_comm_event(struct perf_comm_event *comm_event)
3670 struct perf_cpu_context *cpuctx;
3671 struct perf_event_context *ctx;
3673 char comm[TASK_COMM_LEN];
3675 memset(comm, 0, sizeof(comm));
3676 strlcpy(comm, comm_event->task->comm, sizeof(comm));
3677 size = ALIGN(strlen(comm)+1, sizeof(u64));
3679 comm_event->comm = comm;
3680 comm_event->comm_size = size;
3682 comm_event->event_id.header.size = sizeof(comm_event->event_id) + size;
3685 cpuctx = &get_cpu_var(perf_cpu_context);
3686 perf_event_comm_ctx(&cpuctx->ctx, comm_event);
3687 ctx = rcu_dereference(current->perf_event_ctxp);
3689 perf_event_comm_ctx(ctx, comm_event);
3690 put_cpu_var(perf_cpu_context);
3694 void perf_event_comm(struct task_struct *task)
3696 struct perf_comm_event comm_event;
3698 if (task->perf_event_ctxp)
3699 perf_event_enable_on_exec(task);
3701 if (!atomic_read(&nr_comm_events))
3704 comm_event = (struct perf_comm_event){
3710 .type = PERF_RECORD_COMM,
3719 perf_event_comm_event(&comm_event);
3726 struct perf_mmap_event {
3727 struct vm_area_struct *vma;
3729 const char *file_name;
3733 struct perf_event_header header;
3743 static void perf_event_mmap_output(struct perf_event *event,
3744 struct perf_mmap_event *mmap_event)
3746 struct perf_output_handle handle;
3747 int size = mmap_event->event_id.header.size;
3748 int ret = perf_output_begin(&handle, event, size, 0, 0);
3753 mmap_event->event_id.pid = perf_event_pid(event, current);
3754 mmap_event->event_id.tid = perf_event_tid(event, current);
3756 perf_output_put(&handle, mmap_event->event_id);
3757 perf_output_copy(&handle, mmap_event->file_name,
3758 mmap_event->file_size);
3759 perf_output_end(&handle);
3762 static int perf_event_mmap_match(struct perf_event *event,
3763 struct perf_mmap_event *mmap_event,
3766 if (event->state < PERF_EVENT_STATE_INACTIVE)
3769 if (event->cpu != -1 && event->cpu != smp_processor_id())
3772 if ((!executable && event->attr.mmap_data) ||
3773 (executable && event->attr.mmap))
3779 static void perf_event_mmap_ctx(struct perf_event_context *ctx,
3780 struct perf_mmap_event *mmap_event,
3783 struct perf_event *event;
3785 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
3786 if (perf_event_mmap_match(event, mmap_event, executable))
3787 perf_event_mmap_output(event, mmap_event);
3791 static void perf_event_mmap_event(struct perf_mmap_event *mmap_event)
3793 struct perf_cpu_context *cpuctx;
3794 struct perf_event_context *ctx;
3795 struct vm_area_struct *vma = mmap_event->vma;
3796 struct file *file = vma->vm_file;
3802 memset(tmp, 0, sizeof(tmp));
3806 * d_path works from the end of the buffer backwards, so we
3807 * need to add enough zero bytes after the string to handle
3808 * the 64bit alignment we do later.
3810 buf = kzalloc(PATH_MAX + sizeof(u64), GFP_KERNEL);
3812 name = strncpy(tmp, "//enomem", sizeof(tmp));
3815 name = d_path(&file->f_path, buf, PATH_MAX);
3817 name = strncpy(tmp, "//toolong", sizeof(tmp));
3821 if (arch_vma_name(mmap_event->vma)) {
3822 name = strncpy(tmp, arch_vma_name(mmap_event->vma),
3828 name = strncpy(tmp, "[vdso]", sizeof(tmp));
3830 } else if (vma->vm_start <= vma->vm_mm->start_brk &&
3831 vma->vm_end >= vma->vm_mm->brk) {
3832 name = strncpy(tmp, "[heap]", sizeof(tmp));
3834 } else if (vma->vm_start <= vma->vm_mm->start_stack &&
3835 vma->vm_end >= vma->vm_mm->start_stack) {
3836 name = strncpy(tmp, "[stack]", sizeof(tmp));
3840 name = strncpy(tmp, "//anon", sizeof(tmp));
3845 size = ALIGN(strlen(name)+1, sizeof(u64));
3847 mmap_event->file_name = name;
3848 mmap_event->file_size = size;
3850 mmap_event->event_id.header.size = sizeof(mmap_event->event_id) + size;
3853 cpuctx = &get_cpu_var(perf_cpu_context);
3854 perf_event_mmap_ctx(&cpuctx->ctx, mmap_event, vma->vm_flags & VM_EXEC);
3855 ctx = rcu_dereference(current->perf_event_ctxp);
3857 perf_event_mmap_ctx(ctx, mmap_event, vma->vm_flags & VM_EXEC);
3858 put_cpu_var(perf_cpu_context);
3864 void perf_event_mmap(struct vm_area_struct *vma)
3866 struct perf_mmap_event mmap_event;
3868 if (!atomic_read(&nr_mmap_events))
3871 mmap_event = (struct perf_mmap_event){
3877 .type = PERF_RECORD_MMAP,
3878 .misc = PERF_RECORD_MISC_USER,
3883 .start = vma->vm_start,
3884 .len = vma->vm_end - vma->vm_start,
3885 .pgoff = (u64)vma->vm_pgoff << PAGE_SHIFT,
3889 perf_event_mmap_event(&mmap_event);
3893 * IRQ throttle logging
3896 static void perf_log_throttle(struct perf_event *event, int enable)
3898 struct perf_output_handle handle;
3902 struct perf_event_header header;
3906 } throttle_event = {
3908 .type = PERF_RECORD_THROTTLE,
3910 .size = sizeof(throttle_event),
3912 .time = perf_clock(),
3913 .id = primary_event_id(event),
3914 .stream_id = event->id,
3918 throttle_event.header.type = PERF_RECORD_UNTHROTTLE;
3920 ret = perf_output_begin(&handle, event, sizeof(throttle_event), 1, 0);
3924 perf_output_put(&handle, throttle_event);
3925 perf_output_end(&handle);
3929 * Generic event overflow handling, sampling.
3932 static int __perf_event_overflow(struct perf_event *event, int nmi,
3933 int throttle, struct perf_sample_data *data,
3934 struct pt_regs *regs)
3936 int events = atomic_read(&event->event_limit);
3937 struct hw_perf_event *hwc = &event->hw;
3940 throttle = (throttle && event->pmu->unthrottle != NULL);
3945 if (hwc->interrupts != MAX_INTERRUPTS) {
3947 if (HZ * hwc->interrupts >
3948 (u64)sysctl_perf_event_sample_rate) {
3949 hwc->interrupts = MAX_INTERRUPTS;
3950 perf_log_throttle(event, 0);
3955 * Keep re-disabling events even though on the previous
3956 * pass we disabled it - just in case we raced with a
3957 * sched-in and the event got enabled again:
3963 if (event->attr.freq) {
3964 u64 now = perf_clock();
3965 s64 delta = now - hwc->freq_time_stamp;
3967 hwc->freq_time_stamp = now;
3969 if (delta > 0 && delta < 2*TICK_NSEC)
3970 perf_adjust_period(event, delta, hwc->last_period);
3974 * XXX event_limit might not quite work as expected on inherited
3978 event->pending_kill = POLL_IN;
3979 if (events && atomic_dec_and_test(&event->event_limit)) {
3981 event->pending_kill = POLL_HUP;
3983 event->pending_disable = 1;
3984 perf_pending_queue(&event->pending,
3985 perf_pending_event);
3987 perf_event_disable(event);
3990 if (event->overflow_handler)
3991 event->overflow_handler(event, nmi, data, regs);
3993 perf_event_output(event, nmi, data, regs);
3998 int perf_event_overflow(struct perf_event *event, int nmi,
3999 struct perf_sample_data *data,
4000 struct pt_regs *regs)
4002 return __perf_event_overflow(event, nmi, 1, data, regs);
4006 * Generic software event infrastructure
4010 * We directly increment event->count and keep a second value in
4011 * event->hw.period_left to count intervals. This period event
4012 * is kept in the range [-sample_period, 0] so that we can use the
4016 static u64 perf_swevent_set_period(struct perf_event *event)
4018 struct hw_perf_event *hwc = &event->hw;
4019 u64 period = hwc->last_period;
4023 hwc->last_period = hwc->sample_period;
4026 old = val = atomic64_read(&hwc->period_left);
4030 nr = div64_u64(period + val, period);
4031 offset = nr * period;
4033 if (atomic64_cmpxchg(&hwc->period_left, old, val) != old)
4039 static void perf_swevent_overflow(struct perf_event *event, u64 overflow,
4040 int nmi, struct perf_sample_data *data,
4041 struct pt_regs *regs)
4043 struct hw_perf_event *hwc = &event->hw;
4046 data->period = event->hw.last_period;
4048 overflow = perf_swevent_set_period(event);
4050 if (hwc->interrupts == MAX_INTERRUPTS)
4053 for (; overflow; overflow--) {
4054 if (__perf_event_overflow(event, nmi, throttle,
4057 * We inhibit the overflow from happening when
4058 * hwc->interrupts == MAX_INTERRUPTS.
4066 static void perf_swevent_add(struct perf_event *event, u64 nr,
4067 int nmi, struct perf_sample_data *data,
4068 struct pt_regs *regs)
4070 struct hw_perf_event *hwc = &event->hw;
4072 atomic64_add(nr, &event->count);
4077 if (!hwc->sample_period)
4080 if (nr == 1 && hwc->sample_period == 1 && !event->attr.freq)
4081 return perf_swevent_overflow(event, 1, nmi, data, regs);
4083 if (atomic64_add_negative(nr, &hwc->period_left))
4086 perf_swevent_overflow(event, 0, nmi, data, regs);
4089 static int perf_exclude_event(struct perf_event *event,
4090 struct pt_regs *regs)
4093 if (event->attr.exclude_user && user_mode(regs))
4096 if (event->attr.exclude_kernel && !user_mode(regs))
4103 static int perf_swevent_match(struct perf_event *event,
4104 enum perf_type_id type,
4106 struct perf_sample_data *data,
4107 struct pt_regs *regs)
4109 if (event->attr.type != type)
4112 if (event->attr.config != event_id)
4115 if (perf_exclude_event(event, regs))
4121 static inline u64 swevent_hash(u64 type, u32 event_id)
4123 u64 val = event_id | (type << 32);
4125 return hash_64(val, SWEVENT_HLIST_BITS);
4128 static inline struct hlist_head *
4129 __find_swevent_head(struct swevent_hlist *hlist, u64 type, u32 event_id)
4131 u64 hash = swevent_hash(type, event_id);
4133 return &hlist->heads[hash];
4136 /* For the read side: events when they trigger */
4137 static inline struct hlist_head *
4138 find_swevent_head_rcu(struct perf_cpu_context *ctx, u64 type, u32 event_id)
4140 struct swevent_hlist *hlist;
4142 hlist = rcu_dereference(ctx->swevent_hlist);
4146 return __find_swevent_head(hlist, type, event_id);
4149 /* For the event head insertion and removal in the hlist */
4150 static inline struct hlist_head *
4151 find_swevent_head(struct perf_cpu_context *ctx, struct perf_event *event)
4153 struct swevent_hlist *hlist;
4154 u32 event_id = event->attr.config;
4155 u64 type = event->attr.type;
4158 * Event scheduling is always serialized against hlist allocation
4159 * and release. Which makes the protected version suitable here.
4160 * The context lock guarantees that.
4162 hlist = rcu_dereference_protected(ctx->swevent_hlist,
4163 lockdep_is_held(&event->ctx->lock));
4167 return __find_swevent_head(hlist, type, event_id);
4170 static void do_perf_sw_event(enum perf_type_id type, u32 event_id,
4172 struct perf_sample_data *data,
4173 struct pt_regs *regs)
4175 struct perf_cpu_context *cpuctx;
4176 struct perf_event *event;
4177 struct hlist_node *node;
4178 struct hlist_head *head;
4180 cpuctx = &__get_cpu_var(perf_cpu_context);
4184 head = find_swevent_head_rcu(cpuctx, type, event_id);
4189 hlist_for_each_entry_rcu(event, node, head, hlist_entry) {
4190 if (perf_swevent_match(event, type, event_id, data, regs))
4191 perf_swevent_add(event, nr, nmi, data, regs);
4197 int perf_swevent_get_recursion_context(void)
4199 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
4206 else if (in_softirq())
4211 if (cpuctx->recursion[rctx])
4214 cpuctx->recursion[rctx]++;
4219 EXPORT_SYMBOL_GPL(perf_swevent_get_recursion_context);
4221 void inline perf_swevent_put_recursion_context(int rctx)
4223 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
4225 cpuctx->recursion[rctx]--;
4228 void __perf_sw_event(u32 event_id, u64 nr, int nmi,
4229 struct pt_regs *regs, u64 addr)
4231 struct perf_sample_data data;
4234 preempt_disable_notrace();
4235 rctx = perf_swevent_get_recursion_context();
4239 perf_sample_data_init(&data, addr);
4241 do_perf_sw_event(PERF_TYPE_SOFTWARE, event_id, nr, nmi, &data, regs);
4243 perf_swevent_put_recursion_context(rctx);
4244 preempt_enable_notrace();
4247 static void perf_swevent_read(struct perf_event *event)
4251 static int perf_swevent_enable(struct perf_event *event)
4253 struct hw_perf_event *hwc = &event->hw;
4254 struct perf_cpu_context *cpuctx;
4255 struct hlist_head *head;
4257 cpuctx = &__get_cpu_var(perf_cpu_context);
4259 if (hwc->sample_period) {
4260 hwc->last_period = hwc->sample_period;
4261 perf_swevent_set_period(event);
4264 head = find_swevent_head(cpuctx, event);
4265 if (WARN_ON_ONCE(!head))
4268 hlist_add_head_rcu(&event->hlist_entry, head);
4273 static void perf_swevent_disable(struct perf_event *event)
4275 hlist_del_rcu(&event->hlist_entry);
4278 static void perf_swevent_void(struct perf_event *event)
4282 static int perf_swevent_int(struct perf_event *event)
4287 static const struct pmu perf_ops_generic = {
4288 .enable = perf_swevent_enable,
4289 .disable = perf_swevent_disable,
4290 .start = perf_swevent_int,
4291 .stop = perf_swevent_void,
4292 .read = perf_swevent_read,
4293 .unthrottle = perf_swevent_void, /* hwc->interrupts already reset */
4297 * hrtimer based swevent callback
4300 static enum hrtimer_restart perf_swevent_hrtimer(struct hrtimer *hrtimer)
4302 enum hrtimer_restart ret = HRTIMER_RESTART;
4303 struct perf_sample_data data;
4304 struct pt_regs *regs;
4305 struct perf_event *event;
4308 event = container_of(hrtimer, struct perf_event, hw.hrtimer);
4309 event->pmu->read(event);
4311 perf_sample_data_init(&data, 0);
4312 data.period = event->hw.last_period;
4313 regs = get_irq_regs();
4315 if (regs && !perf_exclude_event(event, regs)) {
4316 if (!(event->attr.exclude_idle && current->pid == 0))
4317 if (perf_event_overflow(event, 0, &data, regs))
4318 ret = HRTIMER_NORESTART;
4321 period = max_t(u64, 10000, event->hw.sample_period);
4322 hrtimer_forward_now(hrtimer, ns_to_ktime(period));
4327 static void perf_swevent_start_hrtimer(struct perf_event *event)
4329 struct hw_perf_event *hwc = &event->hw;
4331 hrtimer_init(&hwc->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
4332 hwc->hrtimer.function = perf_swevent_hrtimer;
4333 if (hwc->sample_period) {
4336 if (hwc->remaining) {
4337 if (hwc->remaining < 0)
4340 period = hwc->remaining;
4343 period = max_t(u64, 10000, hwc->sample_period);
4345 __hrtimer_start_range_ns(&hwc->hrtimer,
4346 ns_to_ktime(period), 0,
4347 HRTIMER_MODE_REL, 0);
4351 static void perf_swevent_cancel_hrtimer(struct perf_event *event)
4353 struct hw_perf_event *hwc = &event->hw;
4355 if (hwc->sample_period) {
4356 ktime_t remaining = hrtimer_get_remaining(&hwc->hrtimer);
4357 hwc->remaining = ktime_to_ns(remaining);
4359 hrtimer_cancel(&hwc->hrtimer);
4364 * Software event: cpu wall time clock
4367 static void cpu_clock_perf_event_update(struct perf_event *event)
4369 int cpu = raw_smp_processor_id();
4373 now = cpu_clock(cpu);
4374 prev = atomic64_xchg(&event->hw.prev_count, now);
4375 atomic64_add(now - prev, &event->count);
4378 static int cpu_clock_perf_event_enable(struct perf_event *event)
4380 struct hw_perf_event *hwc = &event->hw;
4381 int cpu = raw_smp_processor_id();
4383 atomic64_set(&hwc->prev_count, cpu_clock(cpu));
4384 perf_swevent_start_hrtimer(event);
4389 static void cpu_clock_perf_event_disable(struct perf_event *event)
4391 perf_swevent_cancel_hrtimer(event);
4392 cpu_clock_perf_event_update(event);
4395 static void cpu_clock_perf_event_read(struct perf_event *event)
4397 cpu_clock_perf_event_update(event);
4400 static const struct pmu perf_ops_cpu_clock = {
4401 .enable = cpu_clock_perf_event_enable,
4402 .disable = cpu_clock_perf_event_disable,
4403 .read = cpu_clock_perf_event_read,
4407 * Software event: task time clock
4410 static void task_clock_perf_event_update(struct perf_event *event, u64 now)
4415 prev = atomic64_xchg(&event->hw.prev_count, now);
4417 atomic64_add(delta, &event->count);
4420 static int task_clock_perf_event_enable(struct perf_event *event)
4422 struct hw_perf_event *hwc = &event->hw;
4425 now = event->ctx->time;
4427 atomic64_set(&hwc->prev_count, now);
4429 perf_swevent_start_hrtimer(event);
4434 static void task_clock_perf_event_disable(struct perf_event *event)
4436 perf_swevent_cancel_hrtimer(event);
4437 task_clock_perf_event_update(event, event->ctx->time);
4441 static void task_clock_perf_event_read(struct perf_event *event)
4446 update_context_time(event->ctx);
4447 time = event->ctx->time;
4449 u64 now = perf_clock();
4450 u64 delta = now - event->ctx->timestamp;
4451 time = event->ctx->time + delta;
4454 task_clock_perf_event_update(event, time);
4457 static const struct pmu perf_ops_task_clock = {
4458 .enable = task_clock_perf_event_enable,
4459 .disable = task_clock_perf_event_disable,
4460 .read = task_clock_perf_event_read,
4463 /* Deref the hlist from the update side */
4464 static inline struct swevent_hlist *
4465 swevent_hlist_deref(struct perf_cpu_context *cpuctx)
4467 return rcu_dereference_protected(cpuctx->swevent_hlist,
4468 lockdep_is_held(&cpuctx->hlist_mutex));
4471 static void swevent_hlist_release_rcu(struct rcu_head *rcu_head)
4473 struct swevent_hlist *hlist;
4475 hlist = container_of(rcu_head, struct swevent_hlist, rcu_head);
4479 static void swevent_hlist_release(struct perf_cpu_context *cpuctx)
4481 struct swevent_hlist *hlist = swevent_hlist_deref(cpuctx);
4486 rcu_assign_pointer(cpuctx->swevent_hlist, NULL);
4487 call_rcu(&hlist->rcu_head, swevent_hlist_release_rcu);
4490 static void swevent_hlist_put_cpu(struct perf_event *event, int cpu)
4492 struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
4494 mutex_lock(&cpuctx->hlist_mutex);
4496 if (!--cpuctx->hlist_refcount)
4497 swevent_hlist_release(cpuctx);
4499 mutex_unlock(&cpuctx->hlist_mutex);
4502 static void swevent_hlist_put(struct perf_event *event)
4506 if (event->cpu != -1) {
4507 swevent_hlist_put_cpu(event, event->cpu);
4511 for_each_possible_cpu(cpu)
4512 swevent_hlist_put_cpu(event, cpu);
4515 static int swevent_hlist_get_cpu(struct perf_event *event, int cpu)
4517 struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
4520 mutex_lock(&cpuctx->hlist_mutex);
4522 if (!swevent_hlist_deref(cpuctx) && cpu_online(cpu)) {
4523 struct swevent_hlist *hlist;
4525 hlist = kzalloc(sizeof(*hlist), GFP_KERNEL);
4530 rcu_assign_pointer(cpuctx->swevent_hlist, hlist);
4532 cpuctx->hlist_refcount++;
4534 mutex_unlock(&cpuctx->hlist_mutex);
4539 static int swevent_hlist_get(struct perf_event *event)
4542 int cpu, failed_cpu;
4544 if (event->cpu != -1)
4545 return swevent_hlist_get_cpu(event, event->cpu);
4548 for_each_possible_cpu(cpu) {
4549 err = swevent_hlist_get_cpu(event, cpu);
4559 for_each_possible_cpu(cpu) {
4560 if (cpu == failed_cpu)
4562 swevent_hlist_put_cpu(event, cpu);
4569 #ifdef CONFIG_EVENT_TRACING
4571 static const struct pmu perf_ops_tracepoint = {
4572 .enable = perf_trace_enable,
4573 .disable = perf_trace_disable,
4574 .start = perf_swevent_int,
4575 .stop = perf_swevent_void,
4576 .read = perf_swevent_read,
4577 .unthrottle = perf_swevent_void,
4580 static int perf_tp_filter_match(struct perf_event *event,
4581 struct perf_sample_data *data)
4583 void *record = data->raw->data;
4585 if (likely(!event->filter) || filter_match_preds(event->filter, record))
4590 static int perf_tp_event_match(struct perf_event *event,
4591 struct perf_sample_data *data,
4592 struct pt_regs *regs)
4595 * All tracepoints are from kernel-space.
4597 if (event->attr.exclude_kernel)
4600 if (!perf_tp_filter_match(event, data))
4606 void perf_tp_event(u64 addr, u64 count, void *record, int entry_size,
4607 struct pt_regs *regs, struct hlist_head *head, int rctx)
4609 struct perf_sample_data data;
4610 struct perf_event *event;
4611 struct hlist_node *node;
4613 struct perf_raw_record raw = {
4618 perf_sample_data_init(&data, addr);
4621 hlist_for_each_entry_rcu(event, node, head, hlist_entry) {
4622 if (perf_tp_event_match(event, &data, regs))
4623 perf_swevent_add(event, count, 1, &data, regs);
4626 perf_swevent_put_recursion_context(rctx);
4628 EXPORT_SYMBOL_GPL(perf_tp_event);
4630 static void tp_perf_event_destroy(struct perf_event *event)
4632 perf_trace_destroy(event);
4635 static const struct pmu *tp_perf_event_init(struct perf_event *event)
4640 * Raw tracepoint data is a severe data leak, only allow root to
4643 if ((event->attr.sample_type & PERF_SAMPLE_RAW) &&
4644 perf_paranoid_tracepoint_raw() &&
4645 !capable(CAP_SYS_ADMIN))
4646 return ERR_PTR(-EPERM);
4648 err = perf_trace_init(event);
4652 event->destroy = tp_perf_event_destroy;
4654 return &perf_ops_tracepoint;
4657 static int perf_event_set_filter(struct perf_event *event, void __user *arg)
4662 if (event->attr.type != PERF_TYPE_TRACEPOINT)
4665 filter_str = strndup_user(arg, PAGE_SIZE);
4666 if (IS_ERR(filter_str))
4667 return PTR_ERR(filter_str);
4669 ret = ftrace_profile_set_filter(event, event->attr.config, filter_str);
4675 static void perf_event_free_filter(struct perf_event *event)
4677 ftrace_profile_free_filter(event);
4682 static const struct pmu *tp_perf_event_init(struct perf_event *event)
4687 static int perf_event_set_filter(struct perf_event *event, void __user *arg)
4692 static void perf_event_free_filter(struct perf_event *event)
4696 #endif /* CONFIG_EVENT_TRACING */
4698 #ifdef CONFIG_HAVE_HW_BREAKPOINT
4699 static void bp_perf_event_destroy(struct perf_event *event)
4701 release_bp_slot(event);
4704 static const struct pmu *bp_perf_event_init(struct perf_event *bp)
4708 err = register_perf_hw_breakpoint(bp);
4710 return ERR_PTR(err);
4712 bp->destroy = bp_perf_event_destroy;
4714 return &perf_ops_bp;
4717 void perf_bp_event(struct perf_event *bp, void *data)
4719 struct perf_sample_data sample;
4720 struct pt_regs *regs = data;
4722 perf_sample_data_init(&sample, bp->attr.bp_addr);
4724 if (!perf_exclude_event(bp, regs))
4725 perf_swevent_add(bp, 1, 1, &sample, regs);
4728 static const struct pmu *bp_perf_event_init(struct perf_event *bp)
4733 void perf_bp_event(struct perf_event *bp, void *regs)
4738 atomic_t perf_swevent_enabled[PERF_COUNT_SW_MAX];
4740 static void sw_perf_event_destroy(struct perf_event *event)
4742 u64 event_id = event->attr.config;
4744 WARN_ON(event->parent);
4746 atomic_dec(&perf_swevent_enabled[event_id]);
4747 swevent_hlist_put(event);
4750 static const struct pmu *sw_perf_event_init(struct perf_event *event)
4752 const struct pmu *pmu = NULL;
4753 u64 event_id = event->attr.config;
4756 * Software events (currently) can't in general distinguish
4757 * between user, kernel and hypervisor events.
4758 * However, context switches and cpu migrations are considered
4759 * to be kernel events, and page faults are never hypervisor
4763 case PERF_COUNT_SW_CPU_CLOCK:
4764 pmu = &perf_ops_cpu_clock;
4767 case PERF_COUNT_SW_TASK_CLOCK:
4769 * If the user instantiates this as a per-cpu event,
4770 * use the cpu_clock event instead.
4772 if (event->ctx->task)
4773 pmu = &perf_ops_task_clock;
4775 pmu = &perf_ops_cpu_clock;
4778 case PERF_COUNT_SW_PAGE_FAULTS:
4779 case PERF_COUNT_SW_PAGE_FAULTS_MIN:
4780 case PERF_COUNT_SW_PAGE_FAULTS_MAJ:
4781 case PERF_COUNT_SW_CONTEXT_SWITCHES:
4782 case PERF_COUNT_SW_CPU_MIGRATIONS:
4783 case PERF_COUNT_SW_ALIGNMENT_FAULTS:
4784 case PERF_COUNT_SW_EMULATION_FAULTS:
4785 if (!event->parent) {
4788 err = swevent_hlist_get(event);
4790 return ERR_PTR(err);
4792 atomic_inc(&perf_swevent_enabled[event_id]);
4793 event->destroy = sw_perf_event_destroy;
4795 pmu = &perf_ops_generic;
4803 * Allocate and initialize a event structure
4805 static struct perf_event *
4806 perf_event_alloc(struct perf_event_attr *attr,
4808 struct perf_event_context *ctx,
4809 struct perf_event *group_leader,
4810 struct perf_event *parent_event,
4811 perf_overflow_handler_t overflow_handler,
4814 const struct pmu *pmu;
4815 struct perf_event *event;
4816 struct hw_perf_event *hwc;
4819 event = kzalloc(sizeof(*event), gfpflags);
4821 return ERR_PTR(-ENOMEM);
4824 * Single events are their own group leaders, with an
4825 * empty sibling list:
4828 group_leader = event;
4830 mutex_init(&event->child_mutex);
4831 INIT_LIST_HEAD(&event->child_list);
4833 INIT_LIST_HEAD(&event->group_entry);
4834 INIT_LIST_HEAD(&event->event_entry);
4835 INIT_LIST_HEAD(&event->sibling_list);
4836 init_waitqueue_head(&event->waitq);
4838 mutex_init(&event->mmap_mutex);
4841 event->attr = *attr;
4842 event->group_leader = group_leader;
4847 event->parent = parent_event;
4849 event->ns = get_pid_ns(current->nsproxy->pid_ns);
4850 event->id = atomic64_inc_return(&perf_event_id);
4852 event->state = PERF_EVENT_STATE_INACTIVE;
4854 if (!overflow_handler && parent_event)
4855 overflow_handler = parent_event->overflow_handler;
4857 event->overflow_handler = overflow_handler;
4860 event->state = PERF_EVENT_STATE_OFF;
4865 hwc->sample_period = attr->sample_period;
4866 if (attr->freq && attr->sample_freq)
4867 hwc->sample_period = 1;
4868 hwc->last_period = hwc->sample_period;
4870 atomic64_set(&hwc->period_left, hwc->sample_period);
4873 * we currently do not support PERF_FORMAT_GROUP on inherited events
4875 if (attr->inherit && (attr->read_format & PERF_FORMAT_GROUP))
4878 switch (attr->type) {
4880 case PERF_TYPE_HARDWARE:
4881 case PERF_TYPE_HW_CACHE:
4882 pmu = hw_perf_event_init(event);
4885 case PERF_TYPE_SOFTWARE:
4886 pmu = sw_perf_event_init(event);
4889 case PERF_TYPE_TRACEPOINT:
4890 pmu = tp_perf_event_init(event);
4893 case PERF_TYPE_BREAKPOINT:
4894 pmu = bp_perf_event_init(event);
4905 else if (IS_ERR(pmu))
4910 put_pid_ns(event->ns);
4912 return ERR_PTR(err);
4917 if (!event->parent) {
4918 atomic_inc(&nr_events);
4919 if (event->attr.mmap || event->attr.mmap_data)
4920 atomic_inc(&nr_mmap_events);
4921 if (event->attr.comm)
4922 atomic_inc(&nr_comm_events);
4923 if (event->attr.task)
4924 atomic_inc(&nr_task_events);
4930 static int perf_copy_attr(struct perf_event_attr __user *uattr,
4931 struct perf_event_attr *attr)
4936 if (!access_ok(VERIFY_WRITE, uattr, PERF_ATTR_SIZE_VER0))
4940 * zero the full structure, so that a short copy will be nice.
4942 memset(attr, 0, sizeof(*attr));
4944 ret = get_user(size, &uattr->size);
4948 if (size > PAGE_SIZE) /* silly large */
4951 if (!size) /* abi compat */
4952 size = PERF_ATTR_SIZE_VER0;
4954 if (size < PERF_ATTR_SIZE_VER0)
4958 * If we're handed a bigger struct than we know of,
4959 * ensure all the unknown bits are 0 - i.e. new
4960 * user-space does not rely on any kernel feature
4961 * extensions we dont know about yet.
4963 if (size > sizeof(*attr)) {
4964 unsigned char __user *addr;
4965 unsigned char __user *end;
4968 addr = (void __user *)uattr + sizeof(*attr);
4969 end = (void __user *)uattr + size;
4971 for (; addr < end; addr++) {
4972 ret = get_user(val, addr);
4978 size = sizeof(*attr);
4981 ret = copy_from_user(attr, uattr, size);
4986 * If the type exists, the corresponding creation will verify
4989 if (attr->type >= PERF_TYPE_MAX)
4992 if (attr->__reserved_1)
4995 if (attr->sample_type & ~(PERF_SAMPLE_MAX-1))
4998 if (attr->read_format & ~(PERF_FORMAT_MAX-1))
5005 put_user(sizeof(*attr), &uattr->size);
5011 perf_event_set_output(struct perf_event *event, struct perf_event *output_event)
5013 struct perf_mmap_data *data = NULL, *old_data = NULL;
5019 /* don't allow circular references */
5020 if (event == output_event)
5024 * Don't allow cross-cpu buffers
5026 if (output_event->cpu != event->cpu)
5030 * If its not a per-cpu buffer, it must be the same task.
5032 if (output_event->cpu == -1 && output_event->ctx != event->ctx)
5036 mutex_lock(&event->mmap_mutex);
5037 /* Can't redirect output if we've got an active mmap() */
5038 if (atomic_read(&event->mmap_count))
5042 /* get the buffer we want to redirect to */
5043 data = perf_mmap_data_get(output_event);
5048 old_data = event->data;
5049 rcu_assign_pointer(event->data, data);
5052 mutex_unlock(&event->mmap_mutex);
5055 perf_mmap_data_put(old_data);
5061 * sys_perf_event_open - open a performance event, associate it to a task/cpu
5063 * @attr_uptr: event_id type attributes for monitoring/sampling
5066 * @group_fd: group leader event fd
5068 SYSCALL_DEFINE5(perf_event_open,
5069 struct perf_event_attr __user *, attr_uptr,
5070 pid_t, pid, int, cpu, int, group_fd, unsigned long, flags)
5072 struct perf_event *event, *group_leader = NULL, *output_event = NULL;
5073 struct perf_event_attr attr;
5074 struct perf_event_context *ctx;
5075 struct file *event_file = NULL;
5076 struct file *group_file = NULL;
5078 int fput_needed = 0;
5081 /* for future expandability... */
5082 if (flags & ~(PERF_FLAG_FD_NO_GROUP | PERF_FLAG_FD_OUTPUT))
5085 err = perf_copy_attr(attr_uptr, &attr);
5089 if (!attr.exclude_kernel) {
5090 if (perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN))
5095 if (attr.sample_freq > sysctl_perf_event_sample_rate)
5099 event_fd = get_unused_fd_flags(O_RDWR);
5104 * Get the target context (task or percpu):
5106 ctx = find_get_context(pid, cpu);
5112 if (group_fd != -1) {
5113 group_leader = perf_fget_light(group_fd, &fput_needed);
5114 if (IS_ERR(group_leader)) {
5115 err = PTR_ERR(group_leader);
5116 goto err_put_context;
5118 group_file = group_leader->filp;
5119 if (flags & PERF_FLAG_FD_OUTPUT)
5120 output_event = group_leader;
5121 if (flags & PERF_FLAG_FD_NO_GROUP)
5122 group_leader = NULL;
5126 * Look up the group leader (we will attach this event to it):
5132 * Do not allow a recursive hierarchy (this new sibling
5133 * becoming part of another group-sibling):
5135 if (group_leader->group_leader != group_leader)
5136 goto err_put_context;
5138 * Do not allow to attach to a group in a different
5139 * task or CPU context:
5141 if (group_leader->ctx != ctx)
5142 goto err_put_context;
5144 * Only a group leader can be exclusive or pinned
5146 if (attr.exclusive || attr.pinned)
5147 goto err_put_context;
5150 event = perf_event_alloc(&attr, cpu, ctx, group_leader,
5151 NULL, NULL, GFP_KERNEL);
5152 if (IS_ERR(event)) {
5153 err = PTR_ERR(event);
5154 goto err_put_context;
5158 err = perf_event_set_output(event, output_event);
5160 goto err_free_put_context;
5163 event_file = anon_inode_getfile("[perf_event]", &perf_fops, event, O_RDWR);
5164 if (IS_ERR(event_file)) {
5165 err = PTR_ERR(event_file);
5166 goto err_free_put_context;
5169 event->filp = event_file;
5170 WARN_ON_ONCE(ctx->parent_ctx);
5171 mutex_lock(&ctx->mutex);
5172 perf_install_in_context(ctx, event, cpu);
5174 mutex_unlock(&ctx->mutex);
5176 event->owner = current;
5177 get_task_struct(current);
5178 mutex_lock(¤t->perf_event_mutex);
5179 list_add_tail(&event->owner_entry, ¤t->perf_event_list);
5180 mutex_unlock(¤t->perf_event_mutex);
5183 * Drop the reference on the group_event after placing the
5184 * new event on the sibling_list. This ensures destruction
5185 * of the group leader will find the pointer to itself in
5186 * perf_group_detach().
5188 fput_light(group_file, fput_needed);
5189 fd_install(event_fd, event_file);
5192 err_free_put_context:
5195 fput_light(group_file, fput_needed);
5198 put_unused_fd(event_fd);
5203 * perf_event_create_kernel_counter
5205 * @attr: attributes of the counter to create
5206 * @cpu: cpu in which the counter is bound
5207 * @pid: task to profile
5210 perf_event_create_kernel_counter(struct perf_event_attr *attr, int cpu,
5212 perf_overflow_handler_t overflow_handler)
5214 struct perf_event *event;
5215 struct perf_event_context *ctx;
5219 * Get the target context (task or percpu):
5222 ctx = find_get_context(pid, cpu);
5228 event = perf_event_alloc(attr, cpu, ctx, NULL,
5229 NULL, overflow_handler, GFP_KERNEL);
5230 if (IS_ERR(event)) {
5231 err = PTR_ERR(event);
5232 goto err_put_context;
5236 WARN_ON_ONCE(ctx->parent_ctx);
5237 mutex_lock(&ctx->mutex);
5238 perf_install_in_context(ctx, event, cpu);
5240 mutex_unlock(&ctx->mutex);
5242 event->owner = current;
5243 get_task_struct(current);
5244 mutex_lock(¤t->perf_event_mutex);
5245 list_add_tail(&event->owner_entry, ¤t->perf_event_list);
5246 mutex_unlock(¤t->perf_event_mutex);
5253 return ERR_PTR(err);
5255 EXPORT_SYMBOL_GPL(perf_event_create_kernel_counter);
5258 * inherit a event from parent task to child task:
5260 static struct perf_event *
5261 inherit_event(struct perf_event *parent_event,
5262 struct task_struct *parent,
5263 struct perf_event_context *parent_ctx,
5264 struct task_struct *child,
5265 struct perf_event *group_leader,
5266 struct perf_event_context *child_ctx)
5268 struct perf_event *child_event;
5271 * Instead of creating recursive hierarchies of events,
5272 * we link inherited events back to the original parent,
5273 * which has a filp for sure, which we use as the reference
5276 if (parent_event->parent)
5277 parent_event = parent_event->parent;
5279 child_event = perf_event_alloc(&parent_event->attr,
5280 parent_event->cpu, child_ctx,
5281 group_leader, parent_event,
5283 if (IS_ERR(child_event))
5288 * Make the child state follow the state of the parent event,
5289 * not its attr.disabled bit. We hold the parent's mutex,
5290 * so we won't race with perf_event_{en, dis}able_family.
5292 if (parent_event->state >= PERF_EVENT_STATE_INACTIVE)
5293 child_event->state = PERF_EVENT_STATE_INACTIVE;
5295 child_event->state = PERF_EVENT_STATE_OFF;
5297 if (parent_event->attr.freq) {
5298 u64 sample_period = parent_event->hw.sample_period;
5299 struct hw_perf_event *hwc = &child_event->hw;
5301 hwc->sample_period = sample_period;
5302 hwc->last_period = sample_period;
5304 atomic64_set(&hwc->period_left, sample_period);
5307 child_event->overflow_handler = parent_event->overflow_handler;
5310 * Link it up in the child's context:
5312 add_event_to_ctx(child_event, child_ctx);
5315 * Get a reference to the parent filp - we will fput it
5316 * when the child event exits. This is safe to do because
5317 * we are in the parent and we know that the filp still
5318 * exists and has a nonzero count:
5320 atomic_long_inc(&parent_event->filp->f_count);
5323 * Link this into the parent event's child list
5325 WARN_ON_ONCE(parent_event->ctx->parent_ctx);
5326 mutex_lock(&parent_event->child_mutex);
5327 list_add_tail(&child_event->child_list, &parent_event->child_list);
5328 mutex_unlock(&parent_event->child_mutex);
5333 static int inherit_group(struct perf_event *parent_event,
5334 struct task_struct *parent,
5335 struct perf_event_context *parent_ctx,
5336 struct task_struct *child,
5337 struct perf_event_context *child_ctx)
5339 struct perf_event *leader;
5340 struct perf_event *sub;
5341 struct perf_event *child_ctr;
5343 leader = inherit_event(parent_event, parent, parent_ctx,
5344 child, NULL, child_ctx);
5346 return PTR_ERR(leader);
5347 list_for_each_entry(sub, &parent_event->sibling_list, group_entry) {
5348 child_ctr = inherit_event(sub, parent, parent_ctx,
5349 child, leader, child_ctx);
5350 if (IS_ERR(child_ctr))
5351 return PTR_ERR(child_ctr);
5356 static void sync_child_event(struct perf_event *child_event,
5357 struct task_struct *child)
5359 struct perf_event *parent_event = child_event->parent;
5362 if (child_event->attr.inherit_stat)
5363 perf_event_read_event(child_event, child);
5365 child_val = atomic64_read(&child_event->count);
5368 * Add back the child's count to the parent's count:
5370 atomic64_add(child_val, &parent_event->count);
5371 atomic64_add(child_event->total_time_enabled,
5372 &parent_event->child_total_time_enabled);
5373 atomic64_add(child_event->total_time_running,
5374 &parent_event->child_total_time_running);
5377 * Remove this event from the parent's list
5379 WARN_ON_ONCE(parent_event->ctx->parent_ctx);
5380 mutex_lock(&parent_event->child_mutex);
5381 list_del_init(&child_event->child_list);
5382 mutex_unlock(&parent_event->child_mutex);
5385 * Release the parent event, if this was the last
5388 fput(parent_event->filp);
5392 __perf_event_exit_task(struct perf_event *child_event,
5393 struct perf_event_context *child_ctx,
5394 struct task_struct *child)
5396 struct perf_event *parent_event;
5398 perf_event_remove_from_context(child_event);
5400 parent_event = child_event->parent;
5402 * It can happen that parent exits first, and has events
5403 * that are still around due to the child reference. These
5404 * events need to be zapped - but otherwise linger.
5407 sync_child_event(child_event, child);
5408 free_event(child_event);
5413 * When a child task exits, feed back event values to parent events.
5415 void perf_event_exit_task(struct task_struct *child)
5417 struct perf_event *child_event, *tmp;
5418 struct perf_event_context *child_ctx;
5419 unsigned long flags;
5421 if (likely(!child->perf_event_ctxp)) {
5422 perf_event_task(child, NULL, 0);
5426 local_irq_save(flags);
5428 * We can't reschedule here because interrupts are disabled,
5429 * and either child is current or it is a task that can't be
5430 * scheduled, so we are now safe from rescheduling changing
5433 child_ctx = child->perf_event_ctxp;
5434 __perf_event_task_sched_out(child_ctx);
5437 * Take the context lock here so that if find_get_context is
5438 * reading child->perf_event_ctxp, we wait until it has
5439 * incremented the context's refcount before we do put_ctx below.
5441 raw_spin_lock(&child_ctx->lock);
5442 child->perf_event_ctxp = NULL;
5444 * If this context is a clone; unclone it so it can't get
5445 * swapped to another process while we're removing all
5446 * the events from it.
5448 unclone_ctx(child_ctx);
5449 update_context_time(child_ctx);
5450 raw_spin_unlock_irqrestore(&child_ctx->lock, flags);
5453 * Report the task dead after unscheduling the events so that we
5454 * won't get any samples after PERF_RECORD_EXIT. We can however still
5455 * get a few PERF_RECORD_READ events.
5457 perf_event_task(child, child_ctx, 0);
5460 * We can recurse on the same lock type through:
5462 * __perf_event_exit_task()
5463 * sync_child_event()
5464 * fput(parent_event->filp)
5466 * mutex_lock(&ctx->mutex)
5468 * But since its the parent context it won't be the same instance.
5470 mutex_lock(&child_ctx->mutex);
5473 list_for_each_entry_safe(child_event, tmp, &child_ctx->pinned_groups,
5475 __perf_event_exit_task(child_event, child_ctx, child);
5477 list_for_each_entry_safe(child_event, tmp, &child_ctx->flexible_groups,
5479 __perf_event_exit_task(child_event, child_ctx, child);
5482 * If the last event was a group event, it will have appended all
5483 * its siblings to the list, but we obtained 'tmp' before that which
5484 * will still point to the list head terminating the iteration.
5486 if (!list_empty(&child_ctx->pinned_groups) ||
5487 !list_empty(&child_ctx->flexible_groups))
5490 mutex_unlock(&child_ctx->mutex);
5495 static void perf_free_event(struct perf_event *event,
5496 struct perf_event_context *ctx)
5498 struct perf_event *parent = event->parent;
5500 if (WARN_ON_ONCE(!parent))
5503 mutex_lock(&parent->child_mutex);
5504 list_del_init(&event->child_list);
5505 mutex_unlock(&parent->child_mutex);
5509 perf_group_detach(event);
5510 list_del_event(event, ctx);
5515 * free an unexposed, unused context as created by inheritance by
5516 * init_task below, used by fork() in case of fail.
5518 void perf_event_free_task(struct task_struct *task)
5520 struct perf_event_context *ctx = task->perf_event_ctxp;
5521 struct perf_event *event, *tmp;
5526 mutex_lock(&ctx->mutex);
5528 list_for_each_entry_safe(event, tmp, &ctx->pinned_groups, group_entry)
5529 perf_free_event(event, ctx);
5531 list_for_each_entry_safe(event, tmp, &ctx->flexible_groups,
5533 perf_free_event(event, ctx);
5535 if (!list_empty(&ctx->pinned_groups) ||
5536 !list_empty(&ctx->flexible_groups))
5539 mutex_unlock(&ctx->mutex);
5545 inherit_task_group(struct perf_event *event, struct task_struct *parent,
5546 struct perf_event_context *parent_ctx,
5547 struct task_struct *child,
5551 struct perf_event_context *child_ctx = child->perf_event_ctxp;
5553 if (!event->attr.inherit) {
5560 * This is executed from the parent task context, so
5561 * inherit events that have been marked for cloning.
5562 * First allocate and initialize a context for the
5566 child_ctx = kzalloc(sizeof(struct perf_event_context),
5571 __perf_event_init_context(child_ctx, child);
5572 child->perf_event_ctxp = child_ctx;
5573 get_task_struct(child);
5576 ret = inherit_group(event, parent, parent_ctx,
5587 * Initialize the perf_event context in task_struct
5589 int perf_event_init_task(struct task_struct *child)
5591 struct perf_event_context *child_ctx, *parent_ctx;
5592 struct perf_event_context *cloned_ctx;
5593 struct perf_event *event;
5594 struct task_struct *parent = current;
5595 int inherited_all = 1;
5598 child->perf_event_ctxp = NULL;
5600 mutex_init(&child->perf_event_mutex);
5601 INIT_LIST_HEAD(&child->perf_event_list);
5603 if (likely(!parent->perf_event_ctxp))
5607 * If the parent's context is a clone, pin it so it won't get
5610 parent_ctx = perf_pin_task_context(parent);
5613 * No need to check if parent_ctx != NULL here; since we saw
5614 * it non-NULL earlier, the only reason for it to become NULL
5615 * is if we exit, and since we're currently in the middle of
5616 * a fork we can't be exiting at the same time.
5620 * Lock the parent list. No need to lock the child - not PID
5621 * hashed yet and not running, so nobody can access it.
5623 mutex_lock(&parent_ctx->mutex);
5626 * We dont have to disable NMIs - we are only looking at
5627 * the list, not manipulating it:
5629 list_for_each_entry(event, &parent_ctx->pinned_groups, group_entry) {
5630 ret = inherit_task_group(event, parent, parent_ctx, child,
5636 list_for_each_entry(event, &parent_ctx->flexible_groups, group_entry) {
5637 ret = inherit_task_group(event, parent, parent_ctx, child,
5643 child_ctx = child->perf_event_ctxp;
5645 if (child_ctx && inherited_all) {
5647 * Mark the child context as a clone of the parent
5648 * context, or of whatever the parent is a clone of.
5649 * Note that if the parent is a clone, it could get
5650 * uncloned at any point, but that doesn't matter
5651 * because the list of events and the generation
5652 * count can't have changed since we took the mutex.
5654 cloned_ctx = rcu_dereference(parent_ctx->parent_ctx);
5656 child_ctx->parent_ctx = cloned_ctx;
5657 child_ctx->parent_gen = parent_ctx->parent_gen;
5659 child_ctx->parent_ctx = parent_ctx;
5660 child_ctx->parent_gen = parent_ctx->generation;
5662 get_ctx(child_ctx->parent_ctx);
5665 mutex_unlock(&parent_ctx->mutex);
5667 perf_unpin_context(parent_ctx);
5672 static void __init perf_event_init_all_cpus(void)
5675 struct perf_cpu_context *cpuctx;
5677 for_each_possible_cpu(cpu) {
5678 cpuctx = &per_cpu(perf_cpu_context, cpu);
5679 mutex_init(&cpuctx->hlist_mutex);
5680 __perf_event_init_context(&cpuctx->ctx, NULL);
5684 static void __cpuinit perf_event_init_cpu(int cpu)
5686 struct perf_cpu_context *cpuctx;
5688 cpuctx = &per_cpu(perf_cpu_context, cpu);
5690 spin_lock(&perf_resource_lock);
5691 cpuctx->max_pertask = perf_max_events - perf_reserved_percpu;
5692 spin_unlock(&perf_resource_lock);
5694 mutex_lock(&cpuctx->hlist_mutex);
5695 if (cpuctx->hlist_refcount > 0) {
5696 struct swevent_hlist *hlist;
5698 hlist = kzalloc(sizeof(*hlist), GFP_KERNEL);
5699 WARN_ON_ONCE(!hlist);
5700 rcu_assign_pointer(cpuctx->swevent_hlist, hlist);
5702 mutex_unlock(&cpuctx->hlist_mutex);
5705 #ifdef CONFIG_HOTPLUG_CPU
5706 static void __perf_event_exit_cpu(void *info)
5708 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
5709 struct perf_event_context *ctx = &cpuctx->ctx;
5710 struct perf_event *event, *tmp;
5712 list_for_each_entry_safe(event, tmp, &ctx->pinned_groups, group_entry)
5713 __perf_event_remove_from_context(event);
5714 list_for_each_entry_safe(event, tmp, &ctx->flexible_groups, group_entry)
5715 __perf_event_remove_from_context(event);
5717 static void perf_event_exit_cpu(int cpu)
5719 struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
5720 struct perf_event_context *ctx = &cpuctx->ctx;
5722 mutex_lock(&cpuctx->hlist_mutex);
5723 swevent_hlist_release(cpuctx);
5724 mutex_unlock(&cpuctx->hlist_mutex);
5726 mutex_lock(&ctx->mutex);
5727 smp_call_function_single(cpu, __perf_event_exit_cpu, NULL, 1);
5728 mutex_unlock(&ctx->mutex);
5731 static inline void perf_event_exit_cpu(int cpu) { }
5734 static int __cpuinit
5735 perf_cpu_notify(struct notifier_block *self, unsigned long action, void *hcpu)
5737 unsigned int cpu = (long)hcpu;
5741 case CPU_UP_PREPARE:
5742 case CPU_UP_PREPARE_FROZEN:
5743 perf_event_init_cpu(cpu);
5746 case CPU_DOWN_PREPARE:
5747 case CPU_DOWN_PREPARE_FROZEN:
5748 perf_event_exit_cpu(cpu);
5759 * This has to have a higher priority than migration_notifier in sched.c.
5761 static struct notifier_block __cpuinitdata perf_cpu_nb = {
5762 .notifier_call = perf_cpu_notify,
5766 void __init perf_event_init(void)
5768 perf_event_init_all_cpus();
5769 perf_cpu_notify(&perf_cpu_nb, (unsigned long)CPU_UP_PREPARE,
5770 (void *)(long)smp_processor_id());
5771 perf_cpu_notify(&perf_cpu_nb, (unsigned long)CPU_ONLINE,
5772 (void *)(long)smp_processor_id());
5773 register_cpu_notifier(&perf_cpu_nb);
5776 static ssize_t perf_show_reserve_percpu(struct sysdev_class *class,
5777 struct sysdev_class_attribute *attr,
5780 return sprintf(buf, "%d\n", perf_reserved_percpu);
5784 perf_set_reserve_percpu(struct sysdev_class *class,
5785 struct sysdev_class_attribute *attr,
5789 struct perf_cpu_context *cpuctx;
5793 err = strict_strtoul(buf, 10, &val);
5796 if (val > perf_max_events)
5799 spin_lock(&perf_resource_lock);
5800 perf_reserved_percpu = val;
5801 for_each_online_cpu(cpu) {
5802 cpuctx = &per_cpu(perf_cpu_context, cpu);
5803 raw_spin_lock_irq(&cpuctx->ctx.lock);
5804 mpt = min(perf_max_events - cpuctx->ctx.nr_events,
5805 perf_max_events - perf_reserved_percpu);
5806 cpuctx->max_pertask = mpt;
5807 raw_spin_unlock_irq(&cpuctx->ctx.lock);
5809 spin_unlock(&perf_resource_lock);
5814 static ssize_t perf_show_overcommit(struct sysdev_class *class,
5815 struct sysdev_class_attribute *attr,
5818 return sprintf(buf, "%d\n", perf_overcommit);
5822 perf_set_overcommit(struct sysdev_class *class,
5823 struct sysdev_class_attribute *attr,
5824 const char *buf, size_t count)
5829 err = strict_strtoul(buf, 10, &val);
5835 spin_lock(&perf_resource_lock);
5836 perf_overcommit = val;
5837 spin_unlock(&perf_resource_lock);
5842 static SYSDEV_CLASS_ATTR(
5845 perf_show_reserve_percpu,
5846 perf_set_reserve_percpu
5849 static SYSDEV_CLASS_ATTR(
5852 perf_show_overcommit,
5856 static struct attribute *perfclass_attrs[] = {
5857 &attr_reserve_percpu.attr,
5858 &attr_overcommit.attr,
5862 static struct attribute_group perfclass_attr_group = {
5863 .attrs = perfclass_attrs,
5864 .name = "perf_events",
5867 static int __init perf_event_sysfs_init(void)
5869 return sysfs_create_group(&cpu_sysdev_class.kset.kobj,
5870 &perfclass_attr_group);
5872 device_initcall(perf_event_sysfs_init);