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;
680 if (group_event->state == PERF_EVENT_STATE_OFF)
683 /* Check if group transaction availabe */
690 if (event_sched_in(group_event, cpuctx, ctx)) {
692 pmu->cancel_txn(pmu);
697 * Schedule in siblings as one group (if any):
699 list_for_each_entry(event, &group_event->sibling_list, group_entry) {
700 if (event_sched_in(event, cpuctx, ctx)) {
701 partial_group = event;
709 ret = pmu->commit_txn(pmu);
711 pmu->cancel_txn(pmu);
717 * Groups can be scheduled in as one unit only, so undo any
718 * partial group before returning:
720 list_for_each_entry(event, &group_event->sibling_list, group_entry) {
721 if (event == partial_group)
723 event_sched_out(event, cpuctx, ctx);
725 event_sched_out(group_event, cpuctx, ctx);
728 pmu->cancel_txn(pmu);
734 * Work out whether we can put this event group on the CPU now.
736 static int group_can_go_on(struct perf_event *event,
737 struct perf_cpu_context *cpuctx,
741 * Groups consisting entirely of software events can always go on.
743 if (event->group_flags & PERF_GROUP_SOFTWARE)
746 * If an exclusive group is already on, no other hardware
749 if (cpuctx->exclusive)
752 * If this group is exclusive and there are already
753 * events on the CPU, it can't go on.
755 if (event->attr.exclusive && cpuctx->active_oncpu)
758 * Otherwise, try to add it if all previous groups were able
764 static void add_event_to_ctx(struct perf_event *event,
765 struct perf_event_context *ctx)
767 list_add_event(event, ctx);
768 perf_group_attach(event);
769 event->tstamp_enabled = ctx->time;
770 event->tstamp_running = ctx->time;
771 event->tstamp_stopped = ctx->time;
775 * Cross CPU call to install and enable a performance event
777 * Must be called with ctx->mutex held
779 static void __perf_install_in_context(void *info)
781 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
782 struct perf_event *event = info;
783 struct perf_event_context *ctx = event->ctx;
784 struct perf_event *leader = event->group_leader;
788 * If this is a task context, we need to check whether it is
789 * the current task context of this cpu. If not it has been
790 * scheduled out before the smp call arrived.
791 * Or possibly this is the right context but it isn't
792 * on this cpu because it had no events.
794 if (ctx->task && cpuctx->task_ctx != ctx) {
795 if (cpuctx->task_ctx || ctx->task != current)
797 cpuctx->task_ctx = ctx;
800 raw_spin_lock(&ctx->lock);
802 update_context_time(ctx);
805 * Protect the list operation against NMI by disabling the
806 * events on a global level. NOP for non NMI based events.
810 add_event_to_ctx(event, ctx);
812 if (event->cpu != -1 && event->cpu != smp_processor_id())
816 * Don't put the event on if it is disabled or if
817 * it is in a group and the group isn't on.
819 if (event->state != PERF_EVENT_STATE_INACTIVE ||
820 (leader != event && leader->state != PERF_EVENT_STATE_ACTIVE))
824 * An exclusive event can't go on if there are already active
825 * hardware events, and no hardware event can go on if there
826 * is already an exclusive event on.
828 if (!group_can_go_on(event, cpuctx, 1))
831 err = event_sched_in(event, cpuctx, ctx);
835 * This event couldn't go on. If it is in a group
836 * then we have to pull the whole group off.
837 * If the event group is pinned then put it in error state.
840 group_sched_out(leader, cpuctx, ctx);
841 if (leader->attr.pinned) {
842 update_group_times(leader);
843 leader->state = PERF_EVENT_STATE_ERROR;
847 if (!err && !ctx->task && cpuctx->max_pertask)
848 cpuctx->max_pertask--;
853 raw_spin_unlock(&ctx->lock);
857 * Attach a performance event to a context
859 * First we add the event to the list with the hardware enable bit
860 * in event->hw_config cleared.
862 * If the event is attached to a task which is on a CPU we use a smp
863 * call to enable it in the task context. The task might have been
864 * scheduled away, but we check this in the smp call again.
866 * Must be called with ctx->mutex held.
869 perf_install_in_context(struct perf_event_context *ctx,
870 struct perf_event *event,
873 struct task_struct *task = ctx->task;
877 * Per cpu events are installed via an smp call and
878 * the install is always successful.
880 smp_call_function_single(cpu, __perf_install_in_context,
886 task_oncpu_function_call(task, __perf_install_in_context,
889 raw_spin_lock_irq(&ctx->lock);
891 * we need to retry the smp call.
893 if (ctx->is_active && list_empty(&event->group_entry)) {
894 raw_spin_unlock_irq(&ctx->lock);
899 * The lock prevents that this context is scheduled in so we
900 * can add the event safely, if it the call above did not
903 if (list_empty(&event->group_entry))
904 add_event_to_ctx(event, ctx);
905 raw_spin_unlock_irq(&ctx->lock);
909 * Put a event into inactive state and update time fields.
910 * Enabling the leader of a group effectively enables all
911 * the group members that aren't explicitly disabled, so we
912 * have to update their ->tstamp_enabled also.
913 * Note: this works for group members as well as group leaders
914 * since the non-leader members' sibling_lists will be empty.
916 static void __perf_event_mark_enabled(struct perf_event *event,
917 struct perf_event_context *ctx)
919 struct perf_event *sub;
921 event->state = PERF_EVENT_STATE_INACTIVE;
922 event->tstamp_enabled = ctx->time - event->total_time_enabled;
923 list_for_each_entry(sub, &event->sibling_list, group_entry)
924 if (sub->state >= PERF_EVENT_STATE_INACTIVE)
925 sub->tstamp_enabled =
926 ctx->time - sub->total_time_enabled;
930 * Cross CPU call to enable a performance event
932 static void __perf_event_enable(void *info)
934 struct perf_event *event = info;
935 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
936 struct perf_event_context *ctx = event->ctx;
937 struct perf_event *leader = event->group_leader;
941 * If this is a per-task event, need to check whether this
942 * event's task is the current task on this cpu.
944 if (ctx->task && cpuctx->task_ctx != ctx) {
945 if (cpuctx->task_ctx || ctx->task != current)
947 cpuctx->task_ctx = ctx;
950 raw_spin_lock(&ctx->lock);
952 update_context_time(ctx);
954 if (event->state >= PERF_EVENT_STATE_INACTIVE)
956 __perf_event_mark_enabled(event, ctx);
958 if (event->cpu != -1 && event->cpu != smp_processor_id())
962 * If the event is in a group and isn't the group leader,
963 * then don't put it on unless the group is on.
965 if (leader != event && leader->state != PERF_EVENT_STATE_ACTIVE)
968 if (!group_can_go_on(event, cpuctx, 1)) {
973 err = group_sched_in(event, cpuctx, ctx);
975 err = event_sched_in(event, cpuctx, ctx);
981 * If this event can't go on and it's part of a
982 * group, then the whole group has to come off.
985 group_sched_out(leader, cpuctx, ctx);
986 if (leader->attr.pinned) {
987 update_group_times(leader);
988 leader->state = PERF_EVENT_STATE_ERROR;
993 raw_spin_unlock(&ctx->lock);
999 * If event->ctx is a cloned context, callers must make sure that
1000 * every task struct that event->ctx->task could possibly point to
1001 * remains valid. This condition is satisfied when called through
1002 * perf_event_for_each_child or perf_event_for_each as described
1003 * for perf_event_disable.
1005 void perf_event_enable(struct perf_event *event)
1007 struct perf_event_context *ctx = event->ctx;
1008 struct task_struct *task = ctx->task;
1012 * Enable the event on the cpu that it's on
1014 smp_call_function_single(event->cpu, __perf_event_enable,
1019 raw_spin_lock_irq(&ctx->lock);
1020 if (event->state >= PERF_EVENT_STATE_INACTIVE)
1024 * If the event is in error state, clear that first.
1025 * That way, if we see the event in error state below, we
1026 * know that it has gone back into error state, as distinct
1027 * from the task having been scheduled away before the
1028 * cross-call arrived.
1030 if (event->state == PERF_EVENT_STATE_ERROR)
1031 event->state = PERF_EVENT_STATE_OFF;
1034 raw_spin_unlock_irq(&ctx->lock);
1035 task_oncpu_function_call(task, __perf_event_enable, event);
1037 raw_spin_lock_irq(&ctx->lock);
1040 * If the context is active and the event is still off,
1041 * we need to retry the cross-call.
1043 if (ctx->is_active && event->state == PERF_EVENT_STATE_OFF)
1047 * Since we have the lock this context can't be scheduled
1048 * in, so we can change the state safely.
1050 if (event->state == PERF_EVENT_STATE_OFF)
1051 __perf_event_mark_enabled(event, ctx);
1054 raw_spin_unlock_irq(&ctx->lock);
1057 static int perf_event_refresh(struct perf_event *event, int refresh)
1060 * not supported on inherited events
1062 if (event->attr.inherit)
1065 atomic_add(refresh, &event->event_limit);
1066 perf_event_enable(event);
1072 EVENT_FLEXIBLE = 0x1,
1074 EVENT_ALL = EVENT_FLEXIBLE | EVENT_PINNED,
1077 static void ctx_sched_out(struct perf_event_context *ctx,
1078 struct perf_cpu_context *cpuctx,
1079 enum event_type_t event_type)
1081 struct perf_event *event;
1083 raw_spin_lock(&ctx->lock);
1085 if (likely(!ctx->nr_events))
1087 update_context_time(ctx);
1090 if (!ctx->nr_active)
1093 if (event_type & EVENT_PINNED)
1094 list_for_each_entry(event, &ctx->pinned_groups, group_entry)
1095 group_sched_out(event, cpuctx, ctx);
1097 if (event_type & EVENT_FLEXIBLE)
1098 list_for_each_entry(event, &ctx->flexible_groups, group_entry)
1099 group_sched_out(event, cpuctx, ctx);
1104 raw_spin_unlock(&ctx->lock);
1108 * Test whether two contexts are equivalent, i.e. whether they
1109 * have both been cloned from the same version of the same context
1110 * and they both have the same number of enabled events.
1111 * If the number of enabled events is the same, then the set
1112 * of enabled events should be the same, because these are both
1113 * inherited contexts, therefore we can't access individual events
1114 * in them directly with an fd; we can only enable/disable all
1115 * events via prctl, or enable/disable all events in a family
1116 * via ioctl, which will have the same effect on both contexts.
1118 static int context_equiv(struct perf_event_context *ctx1,
1119 struct perf_event_context *ctx2)
1121 return ctx1->parent_ctx && ctx1->parent_ctx == ctx2->parent_ctx
1122 && ctx1->parent_gen == ctx2->parent_gen
1123 && !ctx1->pin_count && !ctx2->pin_count;
1126 static void __perf_event_sync_stat(struct perf_event *event,
1127 struct perf_event *next_event)
1131 if (!event->attr.inherit_stat)
1135 * Update the event value, we cannot use perf_event_read()
1136 * because we're in the middle of a context switch and have IRQs
1137 * disabled, which upsets smp_call_function_single(), however
1138 * we know the event must be on the current CPU, therefore we
1139 * don't need to use it.
1141 switch (event->state) {
1142 case PERF_EVENT_STATE_ACTIVE:
1143 event->pmu->read(event);
1146 case PERF_EVENT_STATE_INACTIVE:
1147 update_event_times(event);
1155 * In order to keep per-task stats reliable we need to flip the event
1156 * values when we flip the contexts.
1158 value = atomic64_read(&next_event->count);
1159 value = atomic64_xchg(&event->count, value);
1160 atomic64_set(&next_event->count, value);
1162 swap(event->total_time_enabled, next_event->total_time_enabled);
1163 swap(event->total_time_running, next_event->total_time_running);
1166 * Since we swizzled the values, update the user visible data too.
1168 perf_event_update_userpage(event);
1169 perf_event_update_userpage(next_event);
1172 #define list_next_entry(pos, member) \
1173 list_entry(pos->member.next, typeof(*pos), member)
1175 static void perf_event_sync_stat(struct perf_event_context *ctx,
1176 struct perf_event_context *next_ctx)
1178 struct perf_event *event, *next_event;
1183 update_context_time(ctx);
1185 event = list_first_entry(&ctx->event_list,
1186 struct perf_event, event_entry);
1188 next_event = list_first_entry(&next_ctx->event_list,
1189 struct perf_event, event_entry);
1191 while (&event->event_entry != &ctx->event_list &&
1192 &next_event->event_entry != &next_ctx->event_list) {
1194 __perf_event_sync_stat(event, next_event);
1196 event = list_next_entry(event, event_entry);
1197 next_event = list_next_entry(next_event, event_entry);
1202 * Called from scheduler to remove the events of the current task,
1203 * with interrupts disabled.
1205 * We stop each event and update the event value in event->count.
1207 * This does not protect us against NMI, but disable()
1208 * sets the disabled bit in the control field of event _before_
1209 * accessing the event control register. If a NMI hits, then it will
1210 * not restart the event.
1212 void perf_event_task_sched_out(struct task_struct *task,
1213 struct task_struct *next)
1215 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
1216 struct perf_event_context *ctx = task->perf_event_ctxp;
1217 struct perf_event_context *next_ctx;
1218 struct perf_event_context *parent;
1221 perf_sw_event(PERF_COUNT_SW_CONTEXT_SWITCHES, 1, 1, NULL, 0);
1223 if (likely(!ctx || !cpuctx->task_ctx))
1227 parent = rcu_dereference(ctx->parent_ctx);
1228 next_ctx = next->perf_event_ctxp;
1229 if (parent && next_ctx &&
1230 rcu_dereference(next_ctx->parent_ctx) == parent) {
1232 * Looks like the two contexts are clones, so we might be
1233 * able to optimize the context switch. We lock both
1234 * contexts and check that they are clones under the
1235 * lock (including re-checking that neither has been
1236 * uncloned in the meantime). It doesn't matter which
1237 * order we take the locks because no other cpu could
1238 * be trying to lock both of these tasks.
1240 raw_spin_lock(&ctx->lock);
1241 raw_spin_lock_nested(&next_ctx->lock, SINGLE_DEPTH_NESTING);
1242 if (context_equiv(ctx, next_ctx)) {
1244 * XXX do we need a memory barrier of sorts
1245 * wrt to rcu_dereference() of perf_event_ctxp
1247 task->perf_event_ctxp = next_ctx;
1248 next->perf_event_ctxp = ctx;
1250 next_ctx->task = task;
1253 perf_event_sync_stat(ctx, next_ctx);
1255 raw_spin_unlock(&next_ctx->lock);
1256 raw_spin_unlock(&ctx->lock);
1261 ctx_sched_out(ctx, cpuctx, EVENT_ALL);
1262 cpuctx->task_ctx = NULL;
1266 static void task_ctx_sched_out(struct perf_event_context *ctx,
1267 enum event_type_t event_type)
1269 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
1271 if (!cpuctx->task_ctx)
1274 if (WARN_ON_ONCE(ctx != cpuctx->task_ctx))
1277 ctx_sched_out(ctx, cpuctx, event_type);
1278 cpuctx->task_ctx = NULL;
1282 * Called with IRQs disabled
1284 static void __perf_event_task_sched_out(struct perf_event_context *ctx)
1286 task_ctx_sched_out(ctx, EVENT_ALL);
1290 * Called with IRQs disabled
1292 static void cpu_ctx_sched_out(struct perf_cpu_context *cpuctx,
1293 enum event_type_t event_type)
1295 ctx_sched_out(&cpuctx->ctx, cpuctx, event_type);
1299 ctx_pinned_sched_in(struct perf_event_context *ctx,
1300 struct perf_cpu_context *cpuctx)
1302 struct perf_event *event;
1304 list_for_each_entry(event, &ctx->pinned_groups, group_entry) {
1305 if (event->state <= PERF_EVENT_STATE_OFF)
1307 if (event->cpu != -1 && event->cpu != smp_processor_id())
1310 if (group_can_go_on(event, cpuctx, 1))
1311 group_sched_in(event, cpuctx, ctx);
1314 * If this pinned group hasn't been scheduled,
1315 * put it in error state.
1317 if (event->state == PERF_EVENT_STATE_INACTIVE) {
1318 update_group_times(event);
1319 event->state = PERF_EVENT_STATE_ERROR;
1325 ctx_flexible_sched_in(struct perf_event_context *ctx,
1326 struct perf_cpu_context *cpuctx)
1328 struct perf_event *event;
1331 list_for_each_entry(event, &ctx->flexible_groups, group_entry) {
1332 /* Ignore events in OFF or ERROR state */
1333 if (event->state <= PERF_EVENT_STATE_OFF)
1336 * Listen to the 'cpu' scheduling filter constraint
1339 if (event->cpu != -1 && event->cpu != smp_processor_id())
1342 if (group_can_go_on(event, cpuctx, can_add_hw))
1343 if (group_sched_in(event, cpuctx, ctx))
1349 ctx_sched_in(struct perf_event_context *ctx,
1350 struct perf_cpu_context *cpuctx,
1351 enum event_type_t event_type)
1353 raw_spin_lock(&ctx->lock);
1355 if (likely(!ctx->nr_events))
1358 ctx->timestamp = perf_clock();
1363 * First go through the list and put on any pinned groups
1364 * in order to give them the best chance of going on.
1366 if (event_type & EVENT_PINNED)
1367 ctx_pinned_sched_in(ctx, cpuctx);
1369 /* Then walk through the lower prio flexible groups */
1370 if (event_type & EVENT_FLEXIBLE)
1371 ctx_flexible_sched_in(ctx, cpuctx);
1375 raw_spin_unlock(&ctx->lock);
1378 static void cpu_ctx_sched_in(struct perf_cpu_context *cpuctx,
1379 enum event_type_t event_type)
1381 struct perf_event_context *ctx = &cpuctx->ctx;
1383 ctx_sched_in(ctx, cpuctx, event_type);
1386 static void task_ctx_sched_in(struct task_struct *task,
1387 enum event_type_t event_type)
1389 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
1390 struct perf_event_context *ctx = task->perf_event_ctxp;
1394 if (cpuctx->task_ctx == ctx)
1396 ctx_sched_in(ctx, cpuctx, event_type);
1397 cpuctx->task_ctx = ctx;
1400 * Called from scheduler to add the events of the current task
1401 * with interrupts disabled.
1403 * We restore the event value and then enable it.
1405 * This does not protect us against NMI, but enable()
1406 * sets the enabled bit in the control field of event _before_
1407 * accessing the event control register. If a NMI hits, then it will
1408 * keep the event running.
1410 void perf_event_task_sched_in(struct task_struct *task)
1412 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
1413 struct perf_event_context *ctx = task->perf_event_ctxp;
1418 if (cpuctx->task_ctx == ctx)
1424 * We want to keep the following priority order:
1425 * cpu pinned (that don't need to move), task pinned,
1426 * cpu flexible, task flexible.
1428 cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE);
1430 ctx_sched_in(ctx, cpuctx, EVENT_PINNED);
1431 cpu_ctx_sched_in(cpuctx, EVENT_FLEXIBLE);
1432 ctx_sched_in(ctx, cpuctx, EVENT_FLEXIBLE);
1434 cpuctx->task_ctx = ctx;
1439 #define MAX_INTERRUPTS (~0ULL)
1441 static void perf_log_throttle(struct perf_event *event, int enable);
1443 static u64 perf_calculate_period(struct perf_event *event, u64 nsec, u64 count)
1445 u64 frequency = event->attr.sample_freq;
1446 u64 sec = NSEC_PER_SEC;
1447 u64 divisor, dividend;
1449 int count_fls, nsec_fls, frequency_fls, sec_fls;
1451 count_fls = fls64(count);
1452 nsec_fls = fls64(nsec);
1453 frequency_fls = fls64(frequency);
1457 * We got @count in @nsec, with a target of sample_freq HZ
1458 * the target period becomes:
1461 * period = -------------------
1462 * @nsec * sample_freq
1467 * Reduce accuracy by one bit such that @a and @b converge
1468 * to a similar magnitude.
1470 #define REDUCE_FLS(a, b) \
1472 if (a##_fls > b##_fls) { \
1482 * Reduce accuracy until either term fits in a u64, then proceed with
1483 * the other, so that finally we can do a u64/u64 division.
1485 while (count_fls + sec_fls > 64 && nsec_fls + frequency_fls > 64) {
1486 REDUCE_FLS(nsec, frequency);
1487 REDUCE_FLS(sec, count);
1490 if (count_fls + sec_fls > 64) {
1491 divisor = nsec * frequency;
1493 while (count_fls + sec_fls > 64) {
1494 REDUCE_FLS(count, sec);
1498 dividend = count * sec;
1500 dividend = count * sec;
1502 while (nsec_fls + frequency_fls > 64) {
1503 REDUCE_FLS(nsec, frequency);
1507 divisor = nsec * frequency;
1510 return div64_u64(dividend, divisor);
1513 static void perf_event_stop(struct perf_event *event)
1515 if (!event->pmu->stop)
1516 return event->pmu->disable(event);
1518 return event->pmu->stop(event);
1521 static int perf_event_start(struct perf_event *event)
1523 if (!event->pmu->start)
1524 return event->pmu->enable(event);
1526 return event->pmu->start(event);
1529 static void perf_adjust_period(struct perf_event *event, u64 nsec, u64 count)
1531 struct hw_perf_event *hwc = &event->hw;
1532 u64 period, sample_period;
1535 period = perf_calculate_period(event, nsec, count);
1537 delta = (s64)(period - hwc->sample_period);
1538 delta = (delta + 7) / 8; /* low pass filter */
1540 sample_period = hwc->sample_period + delta;
1545 hwc->sample_period = sample_period;
1547 if (atomic64_read(&hwc->period_left) > 8*sample_period) {
1549 perf_event_stop(event);
1550 atomic64_set(&hwc->period_left, 0);
1551 perf_event_start(event);
1556 static void perf_ctx_adjust_freq(struct perf_event_context *ctx)
1558 struct perf_event *event;
1559 struct hw_perf_event *hwc;
1560 u64 interrupts, now;
1563 raw_spin_lock(&ctx->lock);
1564 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
1565 if (event->state != PERF_EVENT_STATE_ACTIVE)
1568 if (event->cpu != -1 && event->cpu != smp_processor_id())
1573 interrupts = hwc->interrupts;
1574 hwc->interrupts = 0;
1577 * unthrottle events on the tick
1579 if (interrupts == MAX_INTERRUPTS) {
1580 perf_log_throttle(event, 1);
1582 event->pmu->unthrottle(event);
1586 if (!event->attr.freq || !event->attr.sample_freq)
1590 event->pmu->read(event);
1591 now = atomic64_read(&event->count);
1592 delta = now - hwc->freq_count_stamp;
1593 hwc->freq_count_stamp = now;
1596 perf_adjust_period(event, TICK_NSEC, delta);
1599 raw_spin_unlock(&ctx->lock);
1603 * Round-robin a context's events:
1605 static void rotate_ctx(struct perf_event_context *ctx)
1607 raw_spin_lock(&ctx->lock);
1609 /* Rotate the first entry last of non-pinned groups */
1610 list_rotate_left(&ctx->flexible_groups);
1612 raw_spin_unlock(&ctx->lock);
1615 void perf_event_task_tick(struct task_struct *curr)
1617 struct perf_cpu_context *cpuctx;
1618 struct perf_event_context *ctx;
1621 if (!atomic_read(&nr_events))
1624 cpuctx = &__get_cpu_var(perf_cpu_context);
1625 if (cpuctx->ctx.nr_events &&
1626 cpuctx->ctx.nr_events != cpuctx->ctx.nr_active)
1629 ctx = curr->perf_event_ctxp;
1630 if (ctx && ctx->nr_events && ctx->nr_events != ctx->nr_active)
1633 perf_ctx_adjust_freq(&cpuctx->ctx);
1635 perf_ctx_adjust_freq(ctx);
1641 cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE);
1643 task_ctx_sched_out(ctx, EVENT_FLEXIBLE);
1645 rotate_ctx(&cpuctx->ctx);
1649 cpu_ctx_sched_in(cpuctx, EVENT_FLEXIBLE);
1651 task_ctx_sched_in(curr, EVENT_FLEXIBLE);
1655 static int event_enable_on_exec(struct perf_event *event,
1656 struct perf_event_context *ctx)
1658 if (!event->attr.enable_on_exec)
1661 event->attr.enable_on_exec = 0;
1662 if (event->state >= PERF_EVENT_STATE_INACTIVE)
1665 __perf_event_mark_enabled(event, ctx);
1671 * Enable all of a task's events that have been marked enable-on-exec.
1672 * This expects task == current.
1674 static void perf_event_enable_on_exec(struct task_struct *task)
1676 struct perf_event_context *ctx;
1677 struct perf_event *event;
1678 unsigned long flags;
1682 local_irq_save(flags);
1683 ctx = task->perf_event_ctxp;
1684 if (!ctx || !ctx->nr_events)
1687 __perf_event_task_sched_out(ctx);
1689 raw_spin_lock(&ctx->lock);
1691 list_for_each_entry(event, &ctx->pinned_groups, group_entry) {
1692 ret = event_enable_on_exec(event, ctx);
1697 list_for_each_entry(event, &ctx->flexible_groups, group_entry) {
1698 ret = event_enable_on_exec(event, ctx);
1704 * Unclone this context if we enabled any event.
1709 raw_spin_unlock(&ctx->lock);
1711 perf_event_task_sched_in(task);
1713 local_irq_restore(flags);
1717 * Cross CPU call to read the hardware event
1719 static void __perf_event_read(void *info)
1721 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
1722 struct perf_event *event = info;
1723 struct perf_event_context *ctx = event->ctx;
1726 * If this is a task context, we need to check whether it is
1727 * the current task context of this cpu. If not it has been
1728 * scheduled out before the smp call arrived. In that case
1729 * event->count would have been updated to a recent sample
1730 * when the event was scheduled out.
1732 if (ctx->task && cpuctx->task_ctx != ctx)
1735 raw_spin_lock(&ctx->lock);
1736 update_context_time(ctx);
1737 update_event_times(event);
1738 raw_spin_unlock(&ctx->lock);
1740 event->pmu->read(event);
1743 static u64 perf_event_read(struct perf_event *event)
1746 * If event is enabled and currently active on a CPU, update the
1747 * value in the event structure:
1749 if (event->state == PERF_EVENT_STATE_ACTIVE) {
1750 smp_call_function_single(event->oncpu,
1751 __perf_event_read, event, 1);
1752 } else if (event->state == PERF_EVENT_STATE_INACTIVE) {
1753 struct perf_event_context *ctx = event->ctx;
1754 unsigned long flags;
1756 raw_spin_lock_irqsave(&ctx->lock, flags);
1757 update_context_time(ctx);
1758 update_event_times(event);
1759 raw_spin_unlock_irqrestore(&ctx->lock, flags);
1762 return atomic64_read(&event->count);
1766 * Initialize the perf_event context in a task_struct:
1769 __perf_event_init_context(struct perf_event_context *ctx,
1770 struct task_struct *task)
1772 raw_spin_lock_init(&ctx->lock);
1773 mutex_init(&ctx->mutex);
1774 INIT_LIST_HEAD(&ctx->pinned_groups);
1775 INIT_LIST_HEAD(&ctx->flexible_groups);
1776 INIT_LIST_HEAD(&ctx->event_list);
1777 atomic_set(&ctx->refcount, 1);
1781 static struct perf_event_context *find_get_context(pid_t pid, int cpu)
1783 struct perf_event_context *ctx;
1784 struct perf_cpu_context *cpuctx;
1785 struct task_struct *task;
1786 unsigned long flags;
1789 if (pid == -1 && cpu != -1) {
1790 /* Must be root to operate on a CPU event: */
1791 if (perf_paranoid_cpu() && !capable(CAP_SYS_ADMIN))
1792 return ERR_PTR(-EACCES);
1794 if (cpu < 0 || cpu >= nr_cpumask_bits)
1795 return ERR_PTR(-EINVAL);
1798 * We could be clever and allow to attach a event to an
1799 * offline CPU and activate it when the CPU comes up, but
1802 if (!cpu_online(cpu))
1803 return ERR_PTR(-ENODEV);
1805 cpuctx = &per_cpu(perf_cpu_context, cpu);
1816 task = find_task_by_vpid(pid);
1818 get_task_struct(task);
1822 return ERR_PTR(-ESRCH);
1825 * Can't attach events to a dying task.
1828 if (task->flags & PF_EXITING)
1831 /* Reuse ptrace permission checks for now. */
1833 if (!ptrace_may_access(task, PTRACE_MODE_READ))
1837 ctx = perf_lock_task_context(task, &flags);
1840 raw_spin_unlock_irqrestore(&ctx->lock, flags);
1844 ctx = kzalloc(sizeof(struct perf_event_context), GFP_KERNEL);
1848 __perf_event_init_context(ctx, task);
1850 if (cmpxchg(&task->perf_event_ctxp, NULL, ctx)) {
1852 * We raced with some other task; use
1853 * the context they set.
1858 get_task_struct(task);
1861 put_task_struct(task);
1865 put_task_struct(task);
1866 return ERR_PTR(err);
1869 static void perf_event_free_filter(struct perf_event *event);
1871 static void free_event_rcu(struct rcu_head *head)
1873 struct perf_event *event;
1875 event = container_of(head, struct perf_event, rcu_head);
1877 put_pid_ns(event->ns);
1878 perf_event_free_filter(event);
1882 static void perf_pending_sync(struct perf_event *event);
1883 static void perf_mmap_data_put(struct perf_mmap_data *data);
1885 static void free_event(struct perf_event *event)
1887 perf_pending_sync(event);
1889 if (!event->parent) {
1890 atomic_dec(&nr_events);
1891 if (event->attr.mmap)
1892 atomic_dec(&nr_mmap_events);
1893 if (event->attr.comm)
1894 atomic_dec(&nr_comm_events);
1895 if (event->attr.task)
1896 atomic_dec(&nr_task_events);
1900 perf_mmap_data_put(event->data);
1905 event->destroy(event);
1907 put_ctx(event->ctx);
1908 call_rcu(&event->rcu_head, free_event_rcu);
1911 int perf_event_release_kernel(struct perf_event *event)
1913 struct perf_event_context *ctx = event->ctx;
1916 * Remove from the PMU, can't get re-enabled since we got
1917 * here because the last ref went.
1919 perf_event_disable(event);
1921 WARN_ON_ONCE(ctx->parent_ctx);
1923 * There are two ways this annotation is useful:
1925 * 1) there is a lock recursion from perf_event_exit_task
1926 * see the comment there.
1928 * 2) there is a lock-inversion with mmap_sem through
1929 * perf_event_read_group(), which takes faults while
1930 * holding ctx->mutex, however this is called after
1931 * the last filedesc died, so there is no possibility
1932 * to trigger the AB-BA case.
1934 mutex_lock_nested(&ctx->mutex, SINGLE_DEPTH_NESTING);
1935 raw_spin_lock_irq(&ctx->lock);
1936 perf_group_detach(event);
1937 list_del_event(event, ctx);
1938 raw_spin_unlock_irq(&ctx->lock);
1939 mutex_unlock(&ctx->mutex);
1941 mutex_lock(&event->owner->perf_event_mutex);
1942 list_del_init(&event->owner_entry);
1943 mutex_unlock(&event->owner->perf_event_mutex);
1944 put_task_struct(event->owner);
1950 EXPORT_SYMBOL_GPL(perf_event_release_kernel);
1953 * Called when the last reference to the file is gone.
1955 static int perf_release(struct inode *inode, struct file *file)
1957 struct perf_event *event = file->private_data;
1959 file->private_data = NULL;
1961 return perf_event_release_kernel(event);
1964 static int perf_event_read_size(struct perf_event *event)
1966 int entry = sizeof(u64); /* value */
1970 if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1971 size += sizeof(u64);
1973 if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1974 size += sizeof(u64);
1976 if (event->attr.read_format & PERF_FORMAT_ID)
1977 entry += sizeof(u64);
1979 if (event->attr.read_format & PERF_FORMAT_GROUP) {
1980 nr += event->group_leader->nr_siblings;
1981 size += sizeof(u64);
1989 u64 perf_event_read_value(struct perf_event *event, u64 *enabled, u64 *running)
1991 struct perf_event *child;
1997 mutex_lock(&event->child_mutex);
1998 total += perf_event_read(event);
1999 *enabled += event->total_time_enabled +
2000 atomic64_read(&event->child_total_time_enabled);
2001 *running += event->total_time_running +
2002 atomic64_read(&event->child_total_time_running);
2004 list_for_each_entry(child, &event->child_list, child_list) {
2005 total += perf_event_read(child);
2006 *enabled += child->total_time_enabled;
2007 *running += child->total_time_running;
2009 mutex_unlock(&event->child_mutex);
2013 EXPORT_SYMBOL_GPL(perf_event_read_value);
2015 static int perf_event_read_group(struct perf_event *event,
2016 u64 read_format, char __user *buf)
2018 struct perf_event *leader = event->group_leader, *sub;
2019 int n = 0, size = 0, ret = -EFAULT;
2020 struct perf_event_context *ctx = leader->ctx;
2022 u64 count, enabled, running;
2024 mutex_lock(&ctx->mutex);
2025 count = perf_event_read_value(leader, &enabled, &running);
2027 values[n++] = 1 + leader->nr_siblings;
2028 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
2029 values[n++] = enabled;
2030 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
2031 values[n++] = running;
2032 values[n++] = count;
2033 if (read_format & PERF_FORMAT_ID)
2034 values[n++] = primary_event_id(leader);
2036 size = n * sizeof(u64);
2038 if (copy_to_user(buf, values, size))
2043 list_for_each_entry(sub, &leader->sibling_list, group_entry) {
2046 values[n++] = perf_event_read_value(sub, &enabled, &running);
2047 if (read_format & PERF_FORMAT_ID)
2048 values[n++] = primary_event_id(sub);
2050 size = n * sizeof(u64);
2052 if (copy_to_user(buf + ret, values, size)) {
2060 mutex_unlock(&ctx->mutex);
2065 static int perf_event_read_one(struct perf_event *event,
2066 u64 read_format, char __user *buf)
2068 u64 enabled, running;
2072 values[n++] = perf_event_read_value(event, &enabled, &running);
2073 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
2074 values[n++] = enabled;
2075 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
2076 values[n++] = running;
2077 if (read_format & PERF_FORMAT_ID)
2078 values[n++] = primary_event_id(event);
2080 if (copy_to_user(buf, values, n * sizeof(u64)))
2083 return n * sizeof(u64);
2087 * Read the performance event - simple non blocking version for now
2090 perf_read_hw(struct perf_event *event, char __user *buf, size_t count)
2092 u64 read_format = event->attr.read_format;
2096 * Return end-of-file for a read on a event that is in
2097 * error state (i.e. because it was pinned but it couldn't be
2098 * scheduled on to the CPU at some point).
2100 if (event->state == PERF_EVENT_STATE_ERROR)
2103 if (count < perf_event_read_size(event))
2106 WARN_ON_ONCE(event->ctx->parent_ctx);
2107 if (read_format & PERF_FORMAT_GROUP)
2108 ret = perf_event_read_group(event, read_format, buf);
2110 ret = perf_event_read_one(event, read_format, buf);
2116 perf_read(struct file *file, char __user *buf, size_t count, loff_t *ppos)
2118 struct perf_event *event = file->private_data;
2120 return perf_read_hw(event, buf, count);
2123 static unsigned int perf_poll(struct file *file, poll_table *wait)
2125 struct perf_event *event = file->private_data;
2126 struct perf_mmap_data *data;
2127 unsigned int events = POLL_HUP;
2130 data = rcu_dereference(event->data);
2132 events = atomic_xchg(&data->poll, 0);
2135 poll_wait(file, &event->waitq, wait);
2140 static void perf_event_reset(struct perf_event *event)
2142 (void)perf_event_read(event);
2143 atomic64_set(&event->count, 0);
2144 perf_event_update_userpage(event);
2148 * Holding the top-level event's child_mutex means that any
2149 * descendant process that has inherited this event will block
2150 * in sync_child_event if it goes to exit, thus satisfying the
2151 * task existence requirements of perf_event_enable/disable.
2153 static void perf_event_for_each_child(struct perf_event *event,
2154 void (*func)(struct perf_event *))
2156 struct perf_event *child;
2158 WARN_ON_ONCE(event->ctx->parent_ctx);
2159 mutex_lock(&event->child_mutex);
2161 list_for_each_entry(child, &event->child_list, child_list)
2163 mutex_unlock(&event->child_mutex);
2166 static void perf_event_for_each(struct perf_event *event,
2167 void (*func)(struct perf_event *))
2169 struct perf_event_context *ctx = event->ctx;
2170 struct perf_event *sibling;
2172 WARN_ON_ONCE(ctx->parent_ctx);
2173 mutex_lock(&ctx->mutex);
2174 event = event->group_leader;
2176 perf_event_for_each_child(event, func);
2178 list_for_each_entry(sibling, &event->sibling_list, group_entry)
2179 perf_event_for_each_child(event, func);
2180 mutex_unlock(&ctx->mutex);
2183 static int perf_event_period(struct perf_event *event, u64 __user *arg)
2185 struct perf_event_context *ctx = event->ctx;
2190 if (!event->attr.sample_period)
2193 size = copy_from_user(&value, arg, sizeof(value));
2194 if (size != sizeof(value))
2200 raw_spin_lock_irq(&ctx->lock);
2201 if (event->attr.freq) {
2202 if (value > sysctl_perf_event_sample_rate) {
2207 event->attr.sample_freq = value;
2209 event->attr.sample_period = value;
2210 event->hw.sample_period = value;
2213 raw_spin_unlock_irq(&ctx->lock);
2218 static const struct file_operations perf_fops;
2220 static struct perf_event *perf_fget_light(int fd, int *fput_needed)
2224 file = fget_light(fd, fput_needed);
2226 return ERR_PTR(-EBADF);
2228 if (file->f_op != &perf_fops) {
2229 fput_light(file, *fput_needed);
2231 return ERR_PTR(-EBADF);
2234 return file->private_data;
2237 static int perf_event_set_output(struct perf_event *event,
2238 struct perf_event *output_event);
2239 static int perf_event_set_filter(struct perf_event *event, void __user *arg);
2241 static long perf_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
2243 struct perf_event *event = file->private_data;
2244 void (*func)(struct perf_event *);
2248 case PERF_EVENT_IOC_ENABLE:
2249 func = perf_event_enable;
2251 case PERF_EVENT_IOC_DISABLE:
2252 func = perf_event_disable;
2254 case PERF_EVENT_IOC_RESET:
2255 func = perf_event_reset;
2258 case PERF_EVENT_IOC_REFRESH:
2259 return perf_event_refresh(event, arg);
2261 case PERF_EVENT_IOC_PERIOD:
2262 return perf_event_period(event, (u64 __user *)arg);
2264 case PERF_EVENT_IOC_SET_OUTPUT:
2266 struct perf_event *output_event = NULL;
2267 int fput_needed = 0;
2271 output_event = perf_fget_light(arg, &fput_needed);
2272 if (IS_ERR(output_event))
2273 return PTR_ERR(output_event);
2276 ret = perf_event_set_output(event, output_event);
2278 fput_light(output_event->filp, fput_needed);
2283 case PERF_EVENT_IOC_SET_FILTER:
2284 return perf_event_set_filter(event, (void __user *)arg);
2290 if (flags & PERF_IOC_FLAG_GROUP)
2291 perf_event_for_each(event, func);
2293 perf_event_for_each_child(event, func);
2298 int perf_event_task_enable(void)
2300 struct perf_event *event;
2302 mutex_lock(¤t->perf_event_mutex);
2303 list_for_each_entry(event, ¤t->perf_event_list, owner_entry)
2304 perf_event_for_each_child(event, perf_event_enable);
2305 mutex_unlock(¤t->perf_event_mutex);
2310 int perf_event_task_disable(void)
2312 struct perf_event *event;
2314 mutex_lock(¤t->perf_event_mutex);
2315 list_for_each_entry(event, ¤t->perf_event_list, owner_entry)
2316 perf_event_for_each_child(event, perf_event_disable);
2317 mutex_unlock(¤t->perf_event_mutex);
2322 #ifndef PERF_EVENT_INDEX_OFFSET
2323 # define PERF_EVENT_INDEX_OFFSET 0
2326 static int perf_event_index(struct perf_event *event)
2328 if (event->state != PERF_EVENT_STATE_ACTIVE)
2331 return event->hw.idx + 1 - PERF_EVENT_INDEX_OFFSET;
2335 * Callers need to ensure there can be no nesting of this function, otherwise
2336 * the seqlock logic goes bad. We can not serialize this because the arch
2337 * code calls this from NMI context.
2339 void perf_event_update_userpage(struct perf_event *event)
2341 struct perf_event_mmap_page *userpg;
2342 struct perf_mmap_data *data;
2345 data = rcu_dereference(event->data);
2349 userpg = data->user_page;
2352 * Disable preemption so as to not let the corresponding user-space
2353 * spin too long if we get preempted.
2358 userpg->index = perf_event_index(event);
2359 userpg->offset = atomic64_read(&event->count);
2360 if (event->state == PERF_EVENT_STATE_ACTIVE)
2361 userpg->offset -= atomic64_read(&event->hw.prev_count);
2363 userpg->time_enabled = event->total_time_enabled +
2364 atomic64_read(&event->child_total_time_enabled);
2366 userpg->time_running = event->total_time_running +
2367 atomic64_read(&event->child_total_time_running);
2376 #ifndef CONFIG_PERF_USE_VMALLOC
2379 * Back perf_mmap() with regular GFP_KERNEL-0 pages.
2382 static struct page *
2383 perf_mmap_to_page(struct perf_mmap_data *data, unsigned long pgoff)
2385 if (pgoff > data->nr_pages)
2389 return virt_to_page(data->user_page);
2391 return virt_to_page(data->data_pages[pgoff - 1]);
2394 static void *perf_mmap_alloc_page(int cpu)
2399 node = (cpu == -1) ? cpu : cpu_to_node(cpu);
2400 page = alloc_pages_node(node, GFP_KERNEL | __GFP_ZERO, 0);
2404 return page_address(page);
2407 static struct perf_mmap_data *
2408 perf_mmap_data_alloc(struct perf_event *event, int nr_pages)
2410 struct perf_mmap_data *data;
2414 size = sizeof(struct perf_mmap_data);
2415 size += nr_pages * sizeof(void *);
2417 data = kzalloc(size, GFP_KERNEL);
2421 data->user_page = perf_mmap_alloc_page(event->cpu);
2422 if (!data->user_page)
2423 goto fail_user_page;
2425 for (i = 0; i < nr_pages; i++) {
2426 data->data_pages[i] = perf_mmap_alloc_page(event->cpu);
2427 if (!data->data_pages[i])
2428 goto fail_data_pages;
2431 data->nr_pages = nr_pages;
2436 for (i--; i >= 0; i--)
2437 free_page((unsigned long)data->data_pages[i]);
2439 free_page((unsigned long)data->user_page);
2448 static void perf_mmap_free_page(unsigned long addr)
2450 struct page *page = virt_to_page((void *)addr);
2452 page->mapping = NULL;
2456 static void perf_mmap_data_free(struct perf_mmap_data *data)
2460 perf_mmap_free_page((unsigned long)data->user_page);
2461 for (i = 0; i < data->nr_pages; i++)
2462 perf_mmap_free_page((unsigned long)data->data_pages[i]);
2466 static inline int page_order(struct perf_mmap_data *data)
2474 * Back perf_mmap() with vmalloc memory.
2476 * Required for architectures that have d-cache aliasing issues.
2479 static inline int page_order(struct perf_mmap_data *data)
2481 return data->page_order;
2484 static struct page *
2485 perf_mmap_to_page(struct perf_mmap_data *data, unsigned long pgoff)
2487 if (pgoff > (1UL << page_order(data)))
2490 return vmalloc_to_page((void *)data->user_page + pgoff * PAGE_SIZE);
2493 static void perf_mmap_unmark_page(void *addr)
2495 struct page *page = vmalloc_to_page(addr);
2497 page->mapping = NULL;
2500 static void perf_mmap_data_free_work(struct work_struct *work)
2502 struct perf_mmap_data *data;
2506 data = container_of(work, struct perf_mmap_data, work);
2507 nr = 1 << page_order(data);
2509 base = data->user_page;
2510 for (i = 0; i < nr + 1; i++)
2511 perf_mmap_unmark_page(base + (i * PAGE_SIZE));
2517 static void perf_mmap_data_free(struct perf_mmap_data *data)
2519 schedule_work(&data->work);
2522 static struct perf_mmap_data *
2523 perf_mmap_data_alloc(struct perf_event *event, int nr_pages)
2525 struct perf_mmap_data *data;
2529 size = sizeof(struct perf_mmap_data);
2530 size += sizeof(void *);
2532 data = kzalloc(size, GFP_KERNEL);
2536 INIT_WORK(&data->work, perf_mmap_data_free_work);
2538 all_buf = vmalloc_user((nr_pages + 1) * PAGE_SIZE);
2542 data->user_page = all_buf;
2543 data->data_pages[0] = all_buf + PAGE_SIZE;
2544 data->page_order = ilog2(nr_pages);
2558 static unsigned long perf_data_size(struct perf_mmap_data *data)
2560 return data->nr_pages << (PAGE_SHIFT + page_order(data));
2563 static int perf_mmap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
2565 struct perf_event *event = vma->vm_file->private_data;
2566 struct perf_mmap_data *data;
2567 int ret = VM_FAULT_SIGBUS;
2569 if (vmf->flags & FAULT_FLAG_MKWRITE) {
2570 if (vmf->pgoff == 0)
2576 data = rcu_dereference(event->data);
2580 if (vmf->pgoff && (vmf->flags & FAULT_FLAG_WRITE))
2583 vmf->page = perf_mmap_to_page(data, vmf->pgoff);
2587 get_page(vmf->page);
2588 vmf->page->mapping = vma->vm_file->f_mapping;
2589 vmf->page->index = vmf->pgoff;
2599 perf_mmap_data_init(struct perf_event *event, struct perf_mmap_data *data)
2601 long max_size = perf_data_size(data);
2603 if (event->attr.watermark) {
2604 data->watermark = min_t(long, max_size,
2605 event->attr.wakeup_watermark);
2608 if (!data->watermark)
2609 data->watermark = max_size / 2;
2611 atomic_set(&data->refcount, 1);
2612 rcu_assign_pointer(event->data, data);
2615 static void perf_mmap_data_free_rcu(struct rcu_head *rcu_head)
2617 struct perf_mmap_data *data;
2619 data = container_of(rcu_head, struct perf_mmap_data, rcu_head);
2620 perf_mmap_data_free(data);
2623 static struct perf_mmap_data *perf_mmap_data_get(struct perf_event *event)
2625 struct perf_mmap_data *data;
2628 data = rcu_dereference(event->data);
2630 if (!atomic_inc_not_zero(&data->refcount))
2638 static void perf_mmap_data_put(struct perf_mmap_data *data)
2640 if (!atomic_dec_and_test(&data->refcount))
2643 call_rcu(&data->rcu_head, perf_mmap_data_free_rcu);
2646 static void perf_mmap_open(struct vm_area_struct *vma)
2648 struct perf_event *event = vma->vm_file->private_data;
2650 atomic_inc(&event->mmap_count);
2653 static void perf_mmap_close(struct vm_area_struct *vma)
2655 struct perf_event *event = vma->vm_file->private_data;
2657 if (atomic_dec_and_mutex_lock(&event->mmap_count, &event->mmap_mutex)) {
2658 unsigned long size = perf_data_size(event->data);
2659 struct user_struct *user = event->mmap_user;
2660 struct perf_mmap_data *data = event->data;
2662 atomic_long_sub((size >> PAGE_SHIFT) + 1, &user->locked_vm);
2663 vma->vm_mm->locked_vm -= event->mmap_locked;
2664 rcu_assign_pointer(event->data, NULL);
2665 mutex_unlock(&event->mmap_mutex);
2667 perf_mmap_data_put(data);
2672 static const struct vm_operations_struct perf_mmap_vmops = {
2673 .open = perf_mmap_open,
2674 .close = perf_mmap_close,
2675 .fault = perf_mmap_fault,
2676 .page_mkwrite = perf_mmap_fault,
2679 static int perf_mmap(struct file *file, struct vm_area_struct *vma)
2681 struct perf_event *event = file->private_data;
2682 unsigned long user_locked, user_lock_limit;
2683 struct user_struct *user = current_user();
2684 unsigned long locked, lock_limit;
2685 struct perf_mmap_data *data;
2686 unsigned long vma_size;
2687 unsigned long nr_pages;
2688 long user_extra, extra;
2692 * Don't allow mmap() of inherited per-task counters. This would
2693 * create a performance issue due to all children writing to the
2696 if (event->cpu == -1 && event->attr.inherit)
2699 if (!(vma->vm_flags & VM_SHARED))
2702 vma_size = vma->vm_end - vma->vm_start;
2703 nr_pages = (vma_size / PAGE_SIZE) - 1;
2706 * If we have data pages ensure they're a power-of-two number, so we
2707 * can do bitmasks instead of modulo.
2709 if (nr_pages != 0 && !is_power_of_2(nr_pages))
2712 if (vma_size != PAGE_SIZE * (1 + nr_pages))
2715 if (vma->vm_pgoff != 0)
2718 WARN_ON_ONCE(event->ctx->parent_ctx);
2719 mutex_lock(&event->mmap_mutex);
2721 if (event->data->nr_pages == nr_pages)
2722 atomic_inc(&event->data->refcount);
2728 user_extra = nr_pages + 1;
2729 user_lock_limit = sysctl_perf_event_mlock >> (PAGE_SHIFT - 10);
2732 * Increase the limit linearly with more CPUs:
2734 user_lock_limit *= num_online_cpus();
2736 user_locked = atomic_long_read(&user->locked_vm) + user_extra;
2739 if (user_locked > user_lock_limit)
2740 extra = user_locked - user_lock_limit;
2742 lock_limit = rlimit(RLIMIT_MEMLOCK);
2743 lock_limit >>= PAGE_SHIFT;
2744 locked = vma->vm_mm->locked_vm + extra;
2746 if ((locked > lock_limit) && perf_paranoid_tracepoint_raw() &&
2747 !capable(CAP_IPC_LOCK)) {
2752 WARN_ON(event->data);
2754 data = perf_mmap_data_alloc(event, nr_pages);
2760 perf_mmap_data_init(event, data);
2761 if (vma->vm_flags & VM_WRITE)
2762 event->data->writable = 1;
2764 atomic_long_add(user_extra, &user->locked_vm);
2765 event->mmap_locked = extra;
2766 event->mmap_user = get_current_user();
2767 vma->vm_mm->locked_vm += event->mmap_locked;
2771 atomic_inc(&event->mmap_count);
2772 mutex_unlock(&event->mmap_mutex);
2774 vma->vm_flags |= VM_RESERVED;
2775 vma->vm_ops = &perf_mmap_vmops;
2780 static int perf_fasync(int fd, struct file *filp, int on)
2782 struct inode *inode = filp->f_path.dentry->d_inode;
2783 struct perf_event *event = filp->private_data;
2786 mutex_lock(&inode->i_mutex);
2787 retval = fasync_helper(fd, filp, on, &event->fasync);
2788 mutex_unlock(&inode->i_mutex);
2796 static const struct file_operations perf_fops = {
2797 .llseek = no_llseek,
2798 .release = perf_release,
2801 .unlocked_ioctl = perf_ioctl,
2802 .compat_ioctl = perf_ioctl,
2804 .fasync = perf_fasync,
2810 * If there's data, ensure we set the poll() state and publish everything
2811 * to user-space before waking everybody up.
2814 void perf_event_wakeup(struct perf_event *event)
2816 wake_up_all(&event->waitq);
2818 if (event->pending_kill) {
2819 kill_fasync(&event->fasync, SIGIO, event->pending_kill);
2820 event->pending_kill = 0;
2827 * Handle the case where we need to wakeup up from NMI (or rq->lock) context.
2829 * The NMI bit means we cannot possibly take locks. Therefore, maintain a
2830 * single linked list and use cmpxchg() to add entries lockless.
2833 static void perf_pending_event(struct perf_pending_entry *entry)
2835 struct perf_event *event = container_of(entry,
2836 struct perf_event, pending);
2838 if (event->pending_disable) {
2839 event->pending_disable = 0;
2840 __perf_event_disable(event);
2843 if (event->pending_wakeup) {
2844 event->pending_wakeup = 0;
2845 perf_event_wakeup(event);
2849 #define PENDING_TAIL ((struct perf_pending_entry *)-1UL)
2851 static DEFINE_PER_CPU(struct perf_pending_entry *, perf_pending_head) = {
2855 static void perf_pending_queue(struct perf_pending_entry *entry,
2856 void (*func)(struct perf_pending_entry *))
2858 struct perf_pending_entry **head;
2860 if (cmpxchg(&entry->next, NULL, PENDING_TAIL) != NULL)
2865 head = &get_cpu_var(perf_pending_head);
2868 entry->next = *head;
2869 } while (cmpxchg(head, entry->next, entry) != entry->next);
2871 set_perf_event_pending();
2873 put_cpu_var(perf_pending_head);
2876 static int __perf_pending_run(void)
2878 struct perf_pending_entry *list;
2881 list = xchg(&__get_cpu_var(perf_pending_head), PENDING_TAIL);
2882 while (list != PENDING_TAIL) {
2883 void (*func)(struct perf_pending_entry *);
2884 struct perf_pending_entry *entry = list;
2891 * Ensure we observe the unqueue before we issue the wakeup,
2892 * so that we won't be waiting forever.
2893 * -- see perf_not_pending().
2904 static inline int perf_not_pending(struct perf_event *event)
2907 * If we flush on whatever cpu we run, there is a chance we don't
2911 __perf_pending_run();
2915 * Ensure we see the proper queue state before going to sleep
2916 * so that we do not miss the wakeup. -- see perf_pending_handle()
2919 return event->pending.next == NULL;
2922 static void perf_pending_sync(struct perf_event *event)
2924 wait_event(event->waitq, perf_not_pending(event));
2927 void perf_event_do_pending(void)
2929 __perf_pending_run();
2933 * Callchain support -- arch specific
2936 __weak struct perf_callchain_entry *perf_callchain(struct pt_regs *regs)
2942 void perf_arch_fetch_caller_regs(struct pt_regs *regs, unsigned long ip, int skip)
2948 * We assume there is only KVM supporting the callbacks.
2949 * Later on, we might change it to a list if there is
2950 * another virtualization implementation supporting the callbacks.
2952 struct perf_guest_info_callbacks *perf_guest_cbs;
2954 int perf_register_guest_info_callbacks(struct perf_guest_info_callbacks *cbs)
2956 perf_guest_cbs = cbs;
2959 EXPORT_SYMBOL_GPL(perf_register_guest_info_callbacks);
2961 int perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks *cbs)
2963 perf_guest_cbs = NULL;
2966 EXPORT_SYMBOL_GPL(perf_unregister_guest_info_callbacks);
2971 static bool perf_output_space(struct perf_mmap_data *data, unsigned long tail,
2972 unsigned long offset, unsigned long head)
2976 if (!data->writable)
2979 mask = perf_data_size(data) - 1;
2981 offset = (offset - tail) & mask;
2982 head = (head - tail) & mask;
2984 if ((int)(head - offset) < 0)
2990 static void perf_output_wakeup(struct perf_output_handle *handle)
2992 atomic_set(&handle->data->poll, POLL_IN);
2995 handle->event->pending_wakeup = 1;
2996 perf_pending_queue(&handle->event->pending,
2997 perf_pending_event);
2999 perf_event_wakeup(handle->event);
3003 * We need to ensure a later event_id doesn't publish a head when a former
3004 * event isn't done writing. However since we need to deal with NMIs we
3005 * cannot fully serialize things.
3007 * We only publish the head (and generate a wakeup) when the outer-most
3010 static void perf_output_get_handle(struct perf_output_handle *handle)
3012 struct perf_mmap_data *data = handle->data;
3015 local_inc(&data->nest);
3016 handle->wakeup = local_read(&data->wakeup);
3019 static void perf_output_put_handle(struct perf_output_handle *handle)
3021 struct perf_mmap_data *data = handle->data;
3025 head = local_read(&data->head);
3028 * IRQ/NMI can happen here, which means we can miss a head update.
3031 if (!local_dec_and_test(&data->nest))
3035 * Publish the known good head. Rely on the full barrier implied
3036 * by atomic_dec_and_test() order the data->head read and this
3039 data->user_page->data_head = head;
3042 * Now check if we missed an update, rely on the (compiler)
3043 * barrier in atomic_dec_and_test() to re-read data->head.
3045 if (unlikely(head != local_read(&data->head))) {
3046 local_inc(&data->nest);
3050 if (handle->wakeup != local_read(&data->wakeup))
3051 perf_output_wakeup(handle);
3057 __always_inline void perf_output_copy(struct perf_output_handle *handle,
3058 const void *buf, unsigned int len)
3061 unsigned long size = min_t(unsigned long, handle->size, len);
3063 memcpy(handle->addr, buf, size);
3066 handle->addr += size;
3067 handle->size -= size;
3068 if (!handle->size) {
3069 struct perf_mmap_data *data = handle->data;
3072 handle->page &= data->nr_pages - 1;
3073 handle->addr = data->data_pages[handle->page];
3074 handle->size = PAGE_SIZE << page_order(data);
3079 int perf_output_begin(struct perf_output_handle *handle,
3080 struct perf_event *event, unsigned int size,
3081 int nmi, int sample)
3083 struct perf_mmap_data *data;
3084 unsigned long tail, offset, head;
3087 struct perf_event_header header;
3094 * For inherited events we send all the output towards the parent.
3097 event = event->parent;
3099 data = rcu_dereference(event->data);
3103 handle->data = data;
3104 handle->event = event;
3106 handle->sample = sample;
3108 if (!data->nr_pages)
3111 have_lost = local_read(&data->lost);
3113 size += sizeof(lost_event);
3115 perf_output_get_handle(handle);
3119 * Userspace could choose to issue a mb() before updating the
3120 * tail pointer. So that all reads will be completed before the
3123 tail = ACCESS_ONCE(data->user_page->data_tail);
3125 offset = head = local_read(&data->head);
3127 if (unlikely(!perf_output_space(data, tail, offset, head)))
3129 } while (local_cmpxchg(&data->head, offset, head) != offset);
3131 if (head - local_read(&data->wakeup) > data->watermark)
3132 local_add(data->watermark, &data->wakeup);
3134 handle->page = offset >> (PAGE_SHIFT + page_order(data));
3135 handle->page &= data->nr_pages - 1;
3136 handle->size = offset & ((PAGE_SIZE << page_order(data)) - 1);
3137 handle->addr = data->data_pages[handle->page];
3138 handle->addr += handle->size;
3139 handle->size = (PAGE_SIZE << page_order(data)) - handle->size;
3142 lost_event.header.type = PERF_RECORD_LOST;
3143 lost_event.header.misc = 0;
3144 lost_event.header.size = sizeof(lost_event);
3145 lost_event.id = event->id;
3146 lost_event.lost = local_xchg(&data->lost, 0);
3148 perf_output_put(handle, lost_event);
3154 local_inc(&data->lost);
3155 perf_output_put_handle(handle);
3162 void perf_output_end(struct perf_output_handle *handle)
3164 struct perf_event *event = handle->event;
3165 struct perf_mmap_data *data = handle->data;
3167 int wakeup_events = event->attr.wakeup_events;
3169 if (handle->sample && wakeup_events) {
3170 int events = local_inc_return(&data->events);
3171 if (events >= wakeup_events) {
3172 local_sub(wakeup_events, &data->events);
3173 local_inc(&data->wakeup);
3177 perf_output_put_handle(handle);
3181 static u32 perf_event_pid(struct perf_event *event, struct task_struct *p)
3184 * only top level events have the pid namespace they were created in
3187 event = event->parent;
3189 return task_tgid_nr_ns(p, event->ns);
3192 static u32 perf_event_tid(struct perf_event *event, struct task_struct *p)
3195 * only top level events have the pid namespace they were created in
3198 event = event->parent;
3200 return task_pid_nr_ns(p, event->ns);
3203 static void perf_output_read_one(struct perf_output_handle *handle,
3204 struct perf_event *event)
3206 u64 read_format = event->attr.read_format;
3210 values[n++] = atomic64_read(&event->count);
3211 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
3212 values[n++] = event->total_time_enabled +
3213 atomic64_read(&event->child_total_time_enabled);
3215 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
3216 values[n++] = event->total_time_running +
3217 atomic64_read(&event->child_total_time_running);
3219 if (read_format & PERF_FORMAT_ID)
3220 values[n++] = primary_event_id(event);
3222 perf_output_copy(handle, values, n * sizeof(u64));
3226 * XXX PERF_FORMAT_GROUP vs inherited events seems difficult.
3228 static void perf_output_read_group(struct perf_output_handle *handle,
3229 struct perf_event *event)
3231 struct perf_event *leader = event->group_leader, *sub;
3232 u64 read_format = event->attr.read_format;
3236 values[n++] = 1 + leader->nr_siblings;
3238 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
3239 values[n++] = leader->total_time_enabled;
3241 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
3242 values[n++] = leader->total_time_running;
3244 if (leader != event)
3245 leader->pmu->read(leader);
3247 values[n++] = atomic64_read(&leader->count);
3248 if (read_format & PERF_FORMAT_ID)
3249 values[n++] = primary_event_id(leader);
3251 perf_output_copy(handle, values, n * sizeof(u64));
3253 list_for_each_entry(sub, &leader->sibling_list, group_entry) {
3257 sub->pmu->read(sub);
3259 values[n++] = atomic64_read(&sub->count);
3260 if (read_format & PERF_FORMAT_ID)
3261 values[n++] = primary_event_id(sub);
3263 perf_output_copy(handle, values, n * sizeof(u64));
3267 static void perf_output_read(struct perf_output_handle *handle,
3268 struct perf_event *event)
3270 if (event->attr.read_format & PERF_FORMAT_GROUP)
3271 perf_output_read_group(handle, event);
3273 perf_output_read_one(handle, event);
3276 void perf_output_sample(struct perf_output_handle *handle,
3277 struct perf_event_header *header,
3278 struct perf_sample_data *data,
3279 struct perf_event *event)
3281 u64 sample_type = data->type;
3283 perf_output_put(handle, *header);
3285 if (sample_type & PERF_SAMPLE_IP)
3286 perf_output_put(handle, data->ip);
3288 if (sample_type & PERF_SAMPLE_TID)
3289 perf_output_put(handle, data->tid_entry);
3291 if (sample_type & PERF_SAMPLE_TIME)
3292 perf_output_put(handle, data->time);
3294 if (sample_type & PERF_SAMPLE_ADDR)
3295 perf_output_put(handle, data->addr);
3297 if (sample_type & PERF_SAMPLE_ID)
3298 perf_output_put(handle, data->id);
3300 if (sample_type & PERF_SAMPLE_STREAM_ID)
3301 perf_output_put(handle, data->stream_id);
3303 if (sample_type & PERF_SAMPLE_CPU)
3304 perf_output_put(handle, data->cpu_entry);
3306 if (sample_type & PERF_SAMPLE_PERIOD)
3307 perf_output_put(handle, data->period);
3309 if (sample_type & PERF_SAMPLE_READ)
3310 perf_output_read(handle, event);
3312 if (sample_type & PERF_SAMPLE_CALLCHAIN) {
3313 if (data->callchain) {
3316 if (data->callchain)
3317 size += data->callchain->nr;
3319 size *= sizeof(u64);
3321 perf_output_copy(handle, data->callchain, size);
3324 perf_output_put(handle, nr);
3328 if (sample_type & PERF_SAMPLE_RAW) {
3330 perf_output_put(handle, data->raw->size);
3331 perf_output_copy(handle, data->raw->data,
3338 .size = sizeof(u32),
3341 perf_output_put(handle, raw);
3346 void perf_prepare_sample(struct perf_event_header *header,
3347 struct perf_sample_data *data,
3348 struct perf_event *event,
3349 struct pt_regs *regs)
3351 u64 sample_type = event->attr.sample_type;
3353 data->type = sample_type;
3355 header->type = PERF_RECORD_SAMPLE;
3356 header->size = sizeof(*header);
3359 header->misc |= perf_misc_flags(regs);
3361 if (sample_type & PERF_SAMPLE_IP) {
3362 data->ip = perf_instruction_pointer(regs);
3364 header->size += sizeof(data->ip);
3367 if (sample_type & PERF_SAMPLE_TID) {
3368 /* namespace issues */
3369 data->tid_entry.pid = perf_event_pid(event, current);
3370 data->tid_entry.tid = perf_event_tid(event, current);
3372 header->size += sizeof(data->tid_entry);
3375 if (sample_type & PERF_SAMPLE_TIME) {
3376 data->time = perf_clock();
3378 header->size += sizeof(data->time);
3381 if (sample_type & PERF_SAMPLE_ADDR)
3382 header->size += sizeof(data->addr);
3384 if (sample_type & PERF_SAMPLE_ID) {
3385 data->id = primary_event_id(event);
3387 header->size += sizeof(data->id);
3390 if (sample_type & PERF_SAMPLE_STREAM_ID) {
3391 data->stream_id = event->id;
3393 header->size += sizeof(data->stream_id);
3396 if (sample_type & PERF_SAMPLE_CPU) {
3397 data->cpu_entry.cpu = raw_smp_processor_id();
3398 data->cpu_entry.reserved = 0;
3400 header->size += sizeof(data->cpu_entry);
3403 if (sample_type & PERF_SAMPLE_PERIOD)
3404 header->size += sizeof(data->period);
3406 if (sample_type & PERF_SAMPLE_READ)
3407 header->size += perf_event_read_size(event);
3409 if (sample_type & PERF_SAMPLE_CALLCHAIN) {
3412 data->callchain = perf_callchain(regs);
3414 if (data->callchain)
3415 size += data->callchain->nr;
3417 header->size += size * sizeof(u64);
3420 if (sample_type & PERF_SAMPLE_RAW) {
3421 int size = sizeof(u32);
3424 size += data->raw->size;
3426 size += sizeof(u32);
3428 WARN_ON_ONCE(size & (sizeof(u64)-1));
3429 header->size += size;
3433 static void perf_event_output(struct perf_event *event, int nmi,
3434 struct perf_sample_data *data,
3435 struct pt_regs *regs)
3437 struct perf_output_handle handle;
3438 struct perf_event_header header;
3440 perf_prepare_sample(&header, data, event, regs);
3442 if (perf_output_begin(&handle, event, header.size, nmi, 1))
3445 perf_output_sample(&handle, &header, data, event);
3447 perf_output_end(&handle);
3454 struct perf_read_event {
3455 struct perf_event_header header;
3462 perf_event_read_event(struct perf_event *event,
3463 struct task_struct *task)
3465 struct perf_output_handle handle;
3466 struct perf_read_event read_event = {
3468 .type = PERF_RECORD_READ,
3470 .size = sizeof(read_event) + perf_event_read_size(event),
3472 .pid = perf_event_pid(event, task),
3473 .tid = perf_event_tid(event, task),
3477 ret = perf_output_begin(&handle, event, read_event.header.size, 0, 0);
3481 perf_output_put(&handle, read_event);
3482 perf_output_read(&handle, event);
3484 perf_output_end(&handle);
3488 * task tracking -- fork/exit
3490 * enabled by: attr.comm | attr.mmap | attr.task
3493 struct perf_task_event {
3494 struct task_struct *task;
3495 struct perf_event_context *task_ctx;
3498 struct perf_event_header header;
3508 static void perf_event_task_output(struct perf_event *event,
3509 struct perf_task_event *task_event)
3511 struct perf_output_handle handle;
3512 struct task_struct *task = task_event->task;
3515 size = task_event->event_id.header.size;
3516 ret = perf_output_begin(&handle, event, size, 0, 0);
3521 task_event->event_id.pid = perf_event_pid(event, task);
3522 task_event->event_id.ppid = perf_event_pid(event, current);
3524 task_event->event_id.tid = perf_event_tid(event, task);
3525 task_event->event_id.ptid = perf_event_tid(event, current);
3527 perf_output_put(&handle, task_event->event_id);
3529 perf_output_end(&handle);
3532 static int perf_event_task_match(struct perf_event *event)
3534 if (event->state < PERF_EVENT_STATE_INACTIVE)
3537 if (event->cpu != -1 && event->cpu != smp_processor_id())
3540 if (event->attr.comm || event->attr.mmap || event->attr.task)
3546 static void perf_event_task_ctx(struct perf_event_context *ctx,
3547 struct perf_task_event *task_event)
3549 struct perf_event *event;
3551 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
3552 if (perf_event_task_match(event))
3553 perf_event_task_output(event, task_event);
3557 static void perf_event_task_event(struct perf_task_event *task_event)
3559 struct perf_cpu_context *cpuctx;
3560 struct perf_event_context *ctx = task_event->task_ctx;
3563 cpuctx = &get_cpu_var(perf_cpu_context);
3564 perf_event_task_ctx(&cpuctx->ctx, task_event);
3566 ctx = rcu_dereference(current->perf_event_ctxp);
3568 perf_event_task_ctx(ctx, task_event);
3569 put_cpu_var(perf_cpu_context);
3573 static void perf_event_task(struct task_struct *task,
3574 struct perf_event_context *task_ctx,
3577 struct perf_task_event task_event;
3579 if (!atomic_read(&nr_comm_events) &&
3580 !atomic_read(&nr_mmap_events) &&
3581 !atomic_read(&nr_task_events))
3584 task_event = (struct perf_task_event){
3586 .task_ctx = task_ctx,
3589 .type = new ? PERF_RECORD_FORK : PERF_RECORD_EXIT,
3591 .size = sizeof(task_event.event_id),
3597 .time = perf_clock(),
3601 perf_event_task_event(&task_event);
3604 void perf_event_fork(struct task_struct *task)
3606 perf_event_task(task, NULL, 1);
3613 struct perf_comm_event {
3614 struct task_struct *task;
3619 struct perf_event_header header;
3626 static void perf_event_comm_output(struct perf_event *event,
3627 struct perf_comm_event *comm_event)
3629 struct perf_output_handle handle;
3630 int size = comm_event->event_id.header.size;
3631 int ret = perf_output_begin(&handle, event, size, 0, 0);
3636 comm_event->event_id.pid = perf_event_pid(event, comm_event->task);
3637 comm_event->event_id.tid = perf_event_tid(event, comm_event->task);
3639 perf_output_put(&handle, comm_event->event_id);
3640 perf_output_copy(&handle, comm_event->comm,
3641 comm_event->comm_size);
3642 perf_output_end(&handle);
3645 static int perf_event_comm_match(struct perf_event *event)
3647 if (event->state < PERF_EVENT_STATE_INACTIVE)
3650 if (event->cpu != -1 && event->cpu != smp_processor_id())
3653 if (event->attr.comm)
3659 static void perf_event_comm_ctx(struct perf_event_context *ctx,
3660 struct perf_comm_event *comm_event)
3662 struct perf_event *event;
3664 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
3665 if (perf_event_comm_match(event))
3666 perf_event_comm_output(event, comm_event);
3670 static void perf_event_comm_event(struct perf_comm_event *comm_event)
3672 struct perf_cpu_context *cpuctx;
3673 struct perf_event_context *ctx;
3675 char comm[TASK_COMM_LEN];
3677 memset(comm, 0, sizeof(comm));
3678 strlcpy(comm, comm_event->task->comm, sizeof(comm));
3679 size = ALIGN(strlen(comm)+1, sizeof(u64));
3681 comm_event->comm = comm;
3682 comm_event->comm_size = size;
3684 comm_event->event_id.header.size = sizeof(comm_event->event_id) + size;
3687 cpuctx = &get_cpu_var(perf_cpu_context);
3688 perf_event_comm_ctx(&cpuctx->ctx, comm_event);
3689 ctx = rcu_dereference(current->perf_event_ctxp);
3691 perf_event_comm_ctx(ctx, comm_event);
3692 put_cpu_var(perf_cpu_context);
3696 void perf_event_comm(struct task_struct *task)
3698 struct perf_comm_event comm_event;
3700 if (task->perf_event_ctxp)
3701 perf_event_enable_on_exec(task);
3703 if (!atomic_read(&nr_comm_events))
3706 comm_event = (struct perf_comm_event){
3712 .type = PERF_RECORD_COMM,
3721 perf_event_comm_event(&comm_event);
3728 struct perf_mmap_event {
3729 struct vm_area_struct *vma;
3731 const char *file_name;
3735 struct perf_event_header header;
3745 static void perf_event_mmap_output(struct perf_event *event,
3746 struct perf_mmap_event *mmap_event)
3748 struct perf_output_handle handle;
3749 int size = mmap_event->event_id.header.size;
3750 int ret = perf_output_begin(&handle, event, size, 0, 0);
3755 mmap_event->event_id.pid = perf_event_pid(event, current);
3756 mmap_event->event_id.tid = perf_event_tid(event, current);
3758 perf_output_put(&handle, mmap_event->event_id);
3759 perf_output_copy(&handle, mmap_event->file_name,
3760 mmap_event->file_size);
3761 perf_output_end(&handle);
3764 static int perf_event_mmap_match(struct perf_event *event,
3765 struct perf_mmap_event *mmap_event)
3767 if (event->state < PERF_EVENT_STATE_INACTIVE)
3770 if (event->cpu != -1 && event->cpu != smp_processor_id())
3773 if (event->attr.mmap)
3779 static void perf_event_mmap_ctx(struct perf_event_context *ctx,
3780 struct perf_mmap_event *mmap_event)
3782 struct perf_event *event;
3784 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
3785 if (perf_event_mmap_match(event, mmap_event))
3786 perf_event_mmap_output(event, mmap_event);
3790 static void perf_event_mmap_event(struct perf_mmap_event *mmap_event)
3792 struct perf_cpu_context *cpuctx;
3793 struct perf_event_context *ctx;
3794 struct vm_area_struct *vma = mmap_event->vma;
3795 struct file *file = vma->vm_file;
3801 memset(tmp, 0, sizeof(tmp));
3805 * d_path works from the end of the buffer backwards, so we
3806 * need to add enough zero bytes after the string to handle
3807 * the 64bit alignment we do later.
3809 buf = kzalloc(PATH_MAX + sizeof(u64), GFP_KERNEL);
3811 name = strncpy(tmp, "//enomem", sizeof(tmp));
3814 name = d_path(&file->f_path, buf, PATH_MAX);
3816 name = strncpy(tmp, "//toolong", sizeof(tmp));
3820 if (arch_vma_name(mmap_event->vma)) {
3821 name = strncpy(tmp, arch_vma_name(mmap_event->vma),
3827 name = strncpy(tmp, "[vdso]", sizeof(tmp));
3831 name = strncpy(tmp, "//anon", sizeof(tmp));
3836 size = ALIGN(strlen(name)+1, sizeof(u64));
3838 mmap_event->file_name = name;
3839 mmap_event->file_size = size;
3841 mmap_event->event_id.header.size = sizeof(mmap_event->event_id) + size;
3844 cpuctx = &get_cpu_var(perf_cpu_context);
3845 perf_event_mmap_ctx(&cpuctx->ctx, mmap_event);
3846 ctx = rcu_dereference(current->perf_event_ctxp);
3848 perf_event_mmap_ctx(ctx, mmap_event);
3849 put_cpu_var(perf_cpu_context);
3855 void __perf_event_mmap(struct vm_area_struct *vma)
3857 struct perf_mmap_event mmap_event;
3859 if (!atomic_read(&nr_mmap_events))
3862 mmap_event = (struct perf_mmap_event){
3868 .type = PERF_RECORD_MMAP,
3869 .misc = PERF_RECORD_MISC_USER,
3874 .start = vma->vm_start,
3875 .len = vma->vm_end - vma->vm_start,
3876 .pgoff = (u64)vma->vm_pgoff << PAGE_SHIFT,
3880 perf_event_mmap_event(&mmap_event);
3884 * IRQ throttle logging
3887 static void perf_log_throttle(struct perf_event *event, int enable)
3889 struct perf_output_handle handle;
3893 struct perf_event_header header;
3897 } throttle_event = {
3899 .type = PERF_RECORD_THROTTLE,
3901 .size = sizeof(throttle_event),
3903 .time = perf_clock(),
3904 .id = primary_event_id(event),
3905 .stream_id = event->id,
3909 throttle_event.header.type = PERF_RECORD_UNTHROTTLE;
3911 ret = perf_output_begin(&handle, event, sizeof(throttle_event), 1, 0);
3915 perf_output_put(&handle, throttle_event);
3916 perf_output_end(&handle);
3920 * Generic event overflow handling, sampling.
3923 static int __perf_event_overflow(struct perf_event *event, int nmi,
3924 int throttle, struct perf_sample_data *data,
3925 struct pt_regs *regs)
3927 int events = atomic_read(&event->event_limit);
3928 struct hw_perf_event *hwc = &event->hw;
3931 throttle = (throttle && event->pmu->unthrottle != NULL);
3936 if (hwc->interrupts != MAX_INTERRUPTS) {
3938 if (HZ * hwc->interrupts >
3939 (u64)sysctl_perf_event_sample_rate) {
3940 hwc->interrupts = MAX_INTERRUPTS;
3941 perf_log_throttle(event, 0);
3946 * Keep re-disabling events even though on the previous
3947 * pass we disabled it - just in case we raced with a
3948 * sched-in and the event got enabled again:
3954 if (event->attr.freq) {
3955 u64 now = perf_clock();
3956 s64 delta = now - hwc->freq_time_stamp;
3958 hwc->freq_time_stamp = now;
3960 if (delta > 0 && delta < 2*TICK_NSEC)
3961 perf_adjust_period(event, delta, hwc->last_period);
3965 * XXX event_limit might not quite work as expected on inherited
3969 event->pending_kill = POLL_IN;
3970 if (events && atomic_dec_and_test(&event->event_limit)) {
3972 event->pending_kill = POLL_HUP;
3974 event->pending_disable = 1;
3975 perf_pending_queue(&event->pending,
3976 perf_pending_event);
3978 perf_event_disable(event);
3981 if (event->overflow_handler)
3982 event->overflow_handler(event, nmi, data, regs);
3984 perf_event_output(event, nmi, data, regs);
3989 int perf_event_overflow(struct perf_event *event, int nmi,
3990 struct perf_sample_data *data,
3991 struct pt_regs *regs)
3993 return __perf_event_overflow(event, nmi, 1, data, regs);
3997 * Generic software event infrastructure
4001 * We directly increment event->count and keep a second value in
4002 * event->hw.period_left to count intervals. This period event
4003 * is kept in the range [-sample_period, 0] so that we can use the
4007 static u64 perf_swevent_set_period(struct perf_event *event)
4009 struct hw_perf_event *hwc = &event->hw;
4010 u64 period = hwc->last_period;
4014 hwc->last_period = hwc->sample_period;
4017 old = val = atomic64_read(&hwc->period_left);
4021 nr = div64_u64(period + val, period);
4022 offset = nr * period;
4024 if (atomic64_cmpxchg(&hwc->period_left, old, val) != old)
4030 static void perf_swevent_overflow(struct perf_event *event, u64 overflow,
4031 int nmi, struct perf_sample_data *data,
4032 struct pt_regs *regs)
4034 struct hw_perf_event *hwc = &event->hw;
4037 data->period = event->hw.last_period;
4039 overflow = perf_swevent_set_period(event);
4041 if (hwc->interrupts == MAX_INTERRUPTS)
4044 for (; overflow; overflow--) {
4045 if (__perf_event_overflow(event, nmi, throttle,
4048 * We inhibit the overflow from happening when
4049 * hwc->interrupts == MAX_INTERRUPTS.
4057 static void perf_swevent_unthrottle(struct perf_event *event)
4060 * Nothing to do, we already reset hwc->interrupts.
4064 static void perf_swevent_add(struct perf_event *event, u64 nr,
4065 int nmi, struct perf_sample_data *data,
4066 struct pt_regs *regs)
4068 struct hw_perf_event *hwc = &event->hw;
4070 atomic64_add(nr, &event->count);
4075 if (!hwc->sample_period)
4078 if (nr == 1 && hwc->sample_period == 1 && !event->attr.freq)
4079 return perf_swevent_overflow(event, 1, nmi, data, regs);
4081 if (atomic64_add_negative(nr, &hwc->period_left))
4084 perf_swevent_overflow(event, 0, nmi, data, regs);
4087 static int perf_exclude_event(struct perf_event *event,
4088 struct pt_regs *regs)
4091 if (event->attr.exclude_user && user_mode(regs))
4094 if (event->attr.exclude_kernel && !user_mode(regs))
4101 static int perf_swevent_match(struct perf_event *event,
4102 enum perf_type_id type,
4104 struct perf_sample_data *data,
4105 struct pt_regs *regs)
4107 if (event->attr.type != type)
4110 if (event->attr.config != event_id)
4113 if (perf_exclude_event(event, regs))
4119 static inline u64 swevent_hash(u64 type, u32 event_id)
4121 u64 val = event_id | (type << 32);
4123 return hash_64(val, SWEVENT_HLIST_BITS);
4126 static inline struct hlist_head *
4127 __find_swevent_head(struct swevent_hlist *hlist, u64 type, u32 event_id)
4129 u64 hash = swevent_hash(type, event_id);
4131 return &hlist->heads[hash];
4134 /* For the read side: events when they trigger */
4135 static inline struct hlist_head *
4136 find_swevent_head_rcu(struct perf_cpu_context *ctx, u64 type, u32 event_id)
4138 struct swevent_hlist *hlist;
4140 hlist = rcu_dereference(ctx->swevent_hlist);
4144 return __find_swevent_head(hlist, type, event_id);
4147 /* For the event head insertion and removal in the hlist */
4148 static inline struct hlist_head *
4149 find_swevent_head(struct perf_cpu_context *ctx, struct perf_event *event)
4151 struct swevent_hlist *hlist;
4152 u32 event_id = event->attr.config;
4153 u64 type = event->attr.type;
4156 * Event scheduling is always serialized against hlist allocation
4157 * and release. Which makes the protected version suitable here.
4158 * The context lock guarantees that.
4160 hlist = rcu_dereference_protected(ctx->swevent_hlist,
4161 lockdep_is_held(&event->ctx->lock));
4165 return __find_swevent_head(hlist, type, event_id);
4168 static void do_perf_sw_event(enum perf_type_id type, u32 event_id,
4170 struct perf_sample_data *data,
4171 struct pt_regs *regs)
4173 struct perf_cpu_context *cpuctx;
4174 struct perf_event *event;
4175 struct hlist_node *node;
4176 struct hlist_head *head;
4178 cpuctx = &__get_cpu_var(perf_cpu_context);
4182 head = find_swevent_head_rcu(cpuctx, type, event_id);
4187 hlist_for_each_entry_rcu(event, node, head, hlist_entry) {
4188 if (perf_swevent_match(event, type, event_id, data, regs))
4189 perf_swevent_add(event, nr, nmi, data, regs);
4195 int perf_swevent_get_recursion_context(void)
4197 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
4204 else if (in_softirq())
4209 if (cpuctx->recursion[rctx])
4212 cpuctx->recursion[rctx]++;
4217 EXPORT_SYMBOL_GPL(perf_swevent_get_recursion_context);
4219 void perf_swevent_put_recursion_context(int rctx)
4221 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
4223 cpuctx->recursion[rctx]--;
4225 EXPORT_SYMBOL_GPL(perf_swevent_put_recursion_context);
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 const struct pmu perf_ops_generic = {
4279 .enable = perf_swevent_enable,
4280 .disable = perf_swevent_disable,
4281 .read = perf_swevent_read,
4282 .unthrottle = perf_swevent_unthrottle,
4286 * hrtimer based swevent callback
4289 static enum hrtimer_restart perf_swevent_hrtimer(struct hrtimer *hrtimer)
4291 enum hrtimer_restart ret = HRTIMER_RESTART;
4292 struct perf_sample_data data;
4293 struct pt_regs *regs;
4294 struct perf_event *event;
4297 event = container_of(hrtimer, struct perf_event, hw.hrtimer);
4298 event->pmu->read(event);
4300 perf_sample_data_init(&data, 0);
4301 data.period = event->hw.last_period;
4302 regs = get_irq_regs();
4304 if (regs && !perf_exclude_event(event, regs)) {
4305 if (!(event->attr.exclude_idle && current->pid == 0))
4306 if (perf_event_overflow(event, 0, &data, regs))
4307 ret = HRTIMER_NORESTART;
4310 period = max_t(u64, 10000, event->hw.sample_period);
4311 hrtimer_forward_now(hrtimer, ns_to_ktime(period));
4316 static void perf_swevent_start_hrtimer(struct perf_event *event)
4318 struct hw_perf_event *hwc = &event->hw;
4320 hrtimer_init(&hwc->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
4321 hwc->hrtimer.function = perf_swevent_hrtimer;
4322 if (hwc->sample_period) {
4325 if (hwc->remaining) {
4326 if (hwc->remaining < 0)
4329 period = hwc->remaining;
4332 period = max_t(u64, 10000, hwc->sample_period);
4334 __hrtimer_start_range_ns(&hwc->hrtimer,
4335 ns_to_ktime(period), 0,
4336 HRTIMER_MODE_REL, 0);
4340 static void perf_swevent_cancel_hrtimer(struct perf_event *event)
4342 struct hw_perf_event *hwc = &event->hw;
4344 if (hwc->sample_period) {
4345 ktime_t remaining = hrtimer_get_remaining(&hwc->hrtimer);
4346 hwc->remaining = ktime_to_ns(remaining);
4348 hrtimer_cancel(&hwc->hrtimer);
4353 * Software event: cpu wall time clock
4356 static void cpu_clock_perf_event_update(struct perf_event *event)
4358 int cpu = raw_smp_processor_id();
4362 now = cpu_clock(cpu);
4363 prev = atomic64_xchg(&event->hw.prev_count, now);
4364 atomic64_add(now - prev, &event->count);
4367 static int cpu_clock_perf_event_enable(struct perf_event *event)
4369 struct hw_perf_event *hwc = &event->hw;
4370 int cpu = raw_smp_processor_id();
4372 atomic64_set(&hwc->prev_count, cpu_clock(cpu));
4373 perf_swevent_start_hrtimer(event);
4378 static void cpu_clock_perf_event_disable(struct perf_event *event)
4380 perf_swevent_cancel_hrtimer(event);
4381 cpu_clock_perf_event_update(event);
4384 static void cpu_clock_perf_event_read(struct perf_event *event)
4386 cpu_clock_perf_event_update(event);
4389 static const struct pmu perf_ops_cpu_clock = {
4390 .enable = cpu_clock_perf_event_enable,
4391 .disable = cpu_clock_perf_event_disable,
4392 .read = cpu_clock_perf_event_read,
4396 * Software event: task time clock
4399 static void task_clock_perf_event_update(struct perf_event *event, u64 now)
4404 prev = atomic64_xchg(&event->hw.prev_count, now);
4406 atomic64_add(delta, &event->count);
4409 static int task_clock_perf_event_enable(struct perf_event *event)
4411 struct hw_perf_event *hwc = &event->hw;
4414 now = event->ctx->time;
4416 atomic64_set(&hwc->prev_count, now);
4418 perf_swevent_start_hrtimer(event);
4423 static void task_clock_perf_event_disable(struct perf_event *event)
4425 perf_swevent_cancel_hrtimer(event);
4426 task_clock_perf_event_update(event, event->ctx->time);
4430 static void task_clock_perf_event_read(struct perf_event *event)
4435 update_context_time(event->ctx);
4436 time = event->ctx->time;
4438 u64 now = perf_clock();
4439 u64 delta = now - event->ctx->timestamp;
4440 time = event->ctx->time + delta;
4443 task_clock_perf_event_update(event, time);
4446 static const struct pmu perf_ops_task_clock = {
4447 .enable = task_clock_perf_event_enable,
4448 .disable = task_clock_perf_event_disable,
4449 .read = task_clock_perf_event_read,
4452 /* Deref the hlist from the update side */
4453 static inline struct swevent_hlist *
4454 swevent_hlist_deref(struct perf_cpu_context *cpuctx)
4456 return rcu_dereference_protected(cpuctx->swevent_hlist,
4457 lockdep_is_held(&cpuctx->hlist_mutex));
4460 static void swevent_hlist_release_rcu(struct rcu_head *rcu_head)
4462 struct swevent_hlist *hlist;
4464 hlist = container_of(rcu_head, struct swevent_hlist, rcu_head);
4468 static void swevent_hlist_release(struct perf_cpu_context *cpuctx)
4470 struct swevent_hlist *hlist = swevent_hlist_deref(cpuctx);
4475 rcu_assign_pointer(cpuctx->swevent_hlist, NULL);
4476 call_rcu(&hlist->rcu_head, swevent_hlist_release_rcu);
4479 static void swevent_hlist_put_cpu(struct perf_event *event, int cpu)
4481 struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
4483 mutex_lock(&cpuctx->hlist_mutex);
4485 if (!--cpuctx->hlist_refcount)
4486 swevent_hlist_release(cpuctx);
4488 mutex_unlock(&cpuctx->hlist_mutex);
4491 static void swevent_hlist_put(struct perf_event *event)
4495 if (event->cpu != -1) {
4496 swevent_hlist_put_cpu(event, event->cpu);
4500 for_each_possible_cpu(cpu)
4501 swevent_hlist_put_cpu(event, cpu);
4504 static int swevent_hlist_get_cpu(struct perf_event *event, int cpu)
4506 struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
4509 mutex_lock(&cpuctx->hlist_mutex);
4511 if (!swevent_hlist_deref(cpuctx) && cpu_online(cpu)) {
4512 struct swevent_hlist *hlist;
4514 hlist = kzalloc(sizeof(*hlist), GFP_KERNEL);
4519 rcu_assign_pointer(cpuctx->swevent_hlist, hlist);
4521 cpuctx->hlist_refcount++;
4523 mutex_unlock(&cpuctx->hlist_mutex);
4528 static int swevent_hlist_get(struct perf_event *event)
4531 int cpu, failed_cpu;
4533 if (event->cpu != -1)
4534 return swevent_hlist_get_cpu(event, event->cpu);
4537 for_each_possible_cpu(cpu) {
4538 err = swevent_hlist_get_cpu(event, cpu);
4548 for_each_possible_cpu(cpu) {
4549 if (cpu == failed_cpu)
4551 swevent_hlist_put_cpu(event, cpu);
4558 #ifdef CONFIG_EVENT_TRACING
4560 static const struct pmu perf_ops_tracepoint = {
4561 .enable = perf_trace_enable,
4562 .disable = perf_trace_disable,
4563 .read = perf_swevent_read,
4564 .unthrottle = perf_swevent_unthrottle,
4567 static int perf_tp_filter_match(struct perf_event *event,
4568 struct perf_sample_data *data)
4570 void *record = data->raw->data;
4572 if (likely(!event->filter) || filter_match_preds(event->filter, record))
4577 static int perf_tp_event_match(struct perf_event *event,
4578 struct perf_sample_data *data,
4579 struct pt_regs *regs)
4582 * All tracepoints are from kernel-space.
4584 if (event->attr.exclude_kernel)
4587 if (!perf_tp_filter_match(event, data))
4593 void perf_tp_event(u64 addr, u64 count, void *record, int entry_size,
4594 struct pt_regs *regs, struct hlist_head *head)
4596 struct perf_sample_data data;
4597 struct perf_event *event;
4598 struct hlist_node *node;
4600 struct perf_raw_record raw = {
4605 perf_sample_data_init(&data, addr);
4609 hlist_for_each_entry_rcu(event, node, head, hlist_entry) {
4610 if (perf_tp_event_match(event, &data, regs))
4611 perf_swevent_add(event, count, 1, &data, regs);
4615 EXPORT_SYMBOL_GPL(perf_tp_event);
4617 static void tp_perf_event_destroy(struct perf_event *event)
4619 perf_trace_destroy(event);
4622 static const struct pmu *tp_perf_event_init(struct perf_event *event)
4627 * Raw tracepoint data is a severe data leak, only allow root to
4630 if ((event->attr.sample_type & PERF_SAMPLE_RAW) &&
4631 perf_paranoid_tracepoint_raw() &&
4632 !capable(CAP_SYS_ADMIN))
4633 return ERR_PTR(-EPERM);
4635 err = perf_trace_init(event);
4639 event->destroy = tp_perf_event_destroy;
4641 return &perf_ops_tracepoint;
4644 static int perf_event_set_filter(struct perf_event *event, void __user *arg)
4649 if (event->attr.type != PERF_TYPE_TRACEPOINT)
4652 filter_str = strndup_user(arg, PAGE_SIZE);
4653 if (IS_ERR(filter_str))
4654 return PTR_ERR(filter_str);
4656 ret = ftrace_profile_set_filter(event, event->attr.config, filter_str);
4662 static void perf_event_free_filter(struct perf_event *event)
4664 ftrace_profile_free_filter(event);
4669 static const struct pmu *tp_perf_event_init(struct perf_event *event)
4674 static int perf_event_set_filter(struct perf_event *event, void __user *arg)
4679 static void perf_event_free_filter(struct perf_event *event)
4683 #endif /* CONFIG_EVENT_TRACING */
4685 #ifdef CONFIG_HAVE_HW_BREAKPOINT
4686 static void bp_perf_event_destroy(struct perf_event *event)
4688 release_bp_slot(event);
4691 static const struct pmu *bp_perf_event_init(struct perf_event *bp)
4695 err = register_perf_hw_breakpoint(bp);
4697 return ERR_PTR(err);
4699 bp->destroy = bp_perf_event_destroy;
4701 return &perf_ops_bp;
4704 void perf_bp_event(struct perf_event *bp, void *data)
4706 struct perf_sample_data sample;
4707 struct pt_regs *regs = data;
4709 perf_sample_data_init(&sample, bp->attr.bp_addr);
4711 if (!perf_exclude_event(bp, regs))
4712 perf_swevent_add(bp, 1, 1, &sample, regs);
4715 static const struct pmu *bp_perf_event_init(struct perf_event *bp)
4720 void perf_bp_event(struct perf_event *bp, void *regs)
4725 atomic_t perf_swevent_enabled[PERF_COUNT_SW_MAX];
4727 static void sw_perf_event_destroy(struct perf_event *event)
4729 u64 event_id = event->attr.config;
4731 WARN_ON(event->parent);
4733 atomic_dec(&perf_swevent_enabled[event_id]);
4734 swevent_hlist_put(event);
4737 static const struct pmu *sw_perf_event_init(struct perf_event *event)
4739 const struct pmu *pmu = NULL;
4740 u64 event_id = event->attr.config;
4743 * Software events (currently) can't in general distinguish
4744 * between user, kernel and hypervisor events.
4745 * However, context switches and cpu migrations are considered
4746 * to be kernel events, and page faults are never hypervisor
4750 case PERF_COUNT_SW_CPU_CLOCK:
4751 pmu = &perf_ops_cpu_clock;
4754 case PERF_COUNT_SW_TASK_CLOCK:
4756 * If the user instantiates this as a per-cpu event,
4757 * use the cpu_clock event instead.
4759 if (event->ctx->task)
4760 pmu = &perf_ops_task_clock;
4762 pmu = &perf_ops_cpu_clock;
4765 case PERF_COUNT_SW_PAGE_FAULTS:
4766 case PERF_COUNT_SW_PAGE_FAULTS_MIN:
4767 case PERF_COUNT_SW_PAGE_FAULTS_MAJ:
4768 case PERF_COUNT_SW_CONTEXT_SWITCHES:
4769 case PERF_COUNT_SW_CPU_MIGRATIONS:
4770 case PERF_COUNT_SW_ALIGNMENT_FAULTS:
4771 case PERF_COUNT_SW_EMULATION_FAULTS:
4772 if (!event->parent) {
4775 err = swevent_hlist_get(event);
4777 return ERR_PTR(err);
4779 atomic_inc(&perf_swevent_enabled[event_id]);
4780 event->destroy = sw_perf_event_destroy;
4782 pmu = &perf_ops_generic;
4790 * Allocate and initialize a event structure
4792 static struct perf_event *
4793 perf_event_alloc(struct perf_event_attr *attr,
4795 struct perf_event_context *ctx,
4796 struct perf_event *group_leader,
4797 struct perf_event *parent_event,
4798 perf_overflow_handler_t overflow_handler,
4801 const struct pmu *pmu;
4802 struct perf_event *event;
4803 struct hw_perf_event *hwc;
4806 event = kzalloc(sizeof(*event), gfpflags);
4808 return ERR_PTR(-ENOMEM);
4811 * Single events are their own group leaders, with an
4812 * empty sibling list:
4815 group_leader = event;
4817 mutex_init(&event->child_mutex);
4818 INIT_LIST_HEAD(&event->child_list);
4820 INIT_LIST_HEAD(&event->group_entry);
4821 INIT_LIST_HEAD(&event->event_entry);
4822 INIT_LIST_HEAD(&event->sibling_list);
4823 init_waitqueue_head(&event->waitq);
4825 mutex_init(&event->mmap_mutex);
4828 event->attr = *attr;
4829 event->group_leader = group_leader;
4834 event->parent = parent_event;
4836 event->ns = get_pid_ns(current->nsproxy->pid_ns);
4837 event->id = atomic64_inc_return(&perf_event_id);
4839 event->state = PERF_EVENT_STATE_INACTIVE;
4841 if (!overflow_handler && parent_event)
4842 overflow_handler = parent_event->overflow_handler;
4844 event->overflow_handler = overflow_handler;
4847 event->state = PERF_EVENT_STATE_OFF;
4852 hwc->sample_period = attr->sample_period;
4853 if (attr->freq && attr->sample_freq)
4854 hwc->sample_period = 1;
4855 hwc->last_period = hwc->sample_period;
4857 atomic64_set(&hwc->period_left, hwc->sample_period);
4860 * we currently do not support PERF_FORMAT_GROUP on inherited events
4862 if (attr->inherit && (attr->read_format & PERF_FORMAT_GROUP))
4865 switch (attr->type) {
4867 case PERF_TYPE_HARDWARE:
4868 case PERF_TYPE_HW_CACHE:
4869 pmu = hw_perf_event_init(event);
4872 case PERF_TYPE_SOFTWARE:
4873 pmu = sw_perf_event_init(event);
4876 case PERF_TYPE_TRACEPOINT:
4877 pmu = tp_perf_event_init(event);
4880 case PERF_TYPE_BREAKPOINT:
4881 pmu = bp_perf_event_init(event);
4892 else if (IS_ERR(pmu))
4897 put_pid_ns(event->ns);
4899 return ERR_PTR(err);
4904 if (!event->parent) {
4905 atomic_inc(&nr_events);
4906 if (event->attr.mmap)
4907 atomic_inc(&nr_mmap_events);
4908 if (event->attr.comm)
4909 atomic_inc(&nr_comm_events);
4910 if (event->attr.task)
4911 atomic_inc(&nr_task_events);
4917 static int perf_copy_attr(struct perf_event_attr __user *uattr,
4918 struct perf_event_attr *attr)
4923 if (!access_ok(VERIFY_WRITE, uattr, PERF_ATTR_SIZE_VER0))
4927 * zero the full structure, so that a short copy will be nice.
4929 memset(attr, 0, sizeof(*attr));
4931 ret = get_user(size, &uattr->size);
4935 if (size > PAGE_SIZE) /* silly large */
4938 if (!size) /* abi compat */
4939 size = PERF_ATTR_SIZE_VER0;
4941 if (size < PERF_ATTR_SIZE_VER0)
4945 * If we're handed a bigger struct than we know of,
4946 * ensure all the unknown bits are 0 - i.e. new
4947 * user-space does not rely on any kernel feature
4948 * extensions we dont know about yet.
4950 if (size > sizeof(*attr)) {
4951 unsigned char __user *addr;
4952 unsigned char __user *end;
4955 addr = (void __user *)uattr + sizeof(*attr);
4956 end = (void __user *)uattr + size;
4958 for (; addr < end; addr++) {
4959 ret = get_user(val, addr);
4965 size = sizeof(*attr);
4968 ret = copy_from_user(attr, uattr, size);
4973 * If the type exists, the corresponding creation will verify
4976 if (attr->type >= PERF_TYPE_MAX)
4979 if (attr->__reserved_1)
4982 if (attr->sample_type & ~(PERF_SAMPLE_MAX-1))
4985 if (attr->read_format & ~(PERF_FORMAT_MAX-1))
4992 put_user(sizeof(*attr), &uattr->size);
4998 perf_event_set_output(struct perf_event *event, struct perf_event *output_event)
5000 struct perf_mmap_data *data = NULL, *old_data = NULL;
5006 /* don't allow circular references */
5007 if (event == output_event)
5011 * Don't allow cross-cpu buffers
5013 if (output_event->cpu != event->cpu)
5017 * If its not a per-cpu buffer, it must be the same task.
5019 if (output_event->cpu == -1 && output_event->ctx != event->ctx)
5023 mutex_lock(&event->mmap_mutex);
5024 /* Can't redirect output if we've got an active mmap() */
5025 if (atomic_read(&event->mmap_count))
5029 /* get the buffer we want to redirect to */
5030 data = perf_mmap_data_get(output_event);
5035 old_data = event->data;
5036 rcu_assign_pointer(event->data, data);
5039 mutex_unlock(&event->mmap_mutex);
5042 perf_mmap_data_put(old_data);
5048 * sys_perf_event_open - open a performance event, associate it to a task/cpu
5050 * @attr_uptr: event_id type attributes for monitoring/sampling
5053 * @group_fd: group leader event fd
5055 SYSCALL_DEFINE5(perf_event_open,
5056 struct perf_event_attr __user *, attr_uptr,
5057 pid_t, pid, int, cpu, int, group_fd, unsigned long, flags)
5059 struct perf_event *event, *group_leader = NULL, *output_event = NULL;
5060 struct perf_event_attr attr;
5061 struct perf_event_context *ctx;
5062 struct file *event_file = NULL;
5063 struct file *group_file = NULL;
5065 int fput_needed = 0;
5068 /* for future expandability... */
5069 if (flags & ~(PERF_FLAG_FD_NO_GROUP | PERF_FLAG_FD_OUTPUT))
5072 err = perf_copy_attr(attr_uptr, &attr);
5076 if (!attr.exclude_kernel) {
5077 if (perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN))
5082 if (attr.sample_freq > sysctl_perf_event_sample_rate)
5086 event_fd = get_unused_fd_flags(O_RDWR);
5091 * Get the target context (task or percpu):
5093 ctx = find_get_context(pid, cpu);
5099 if (group_fd != -1) {
5100 group_leader = perf_fget_light(group_fd, &fput_needed);
5101 if (IS_ERR(group_leader)) {
5102 err = PTR_ERR(group_leader);
5103 goto err_put_context;
5105 group_file = group_leader->filp;
5106 if (flags & PERF_FLAG_FD_OUTPUT)
5107 output_event = group_leader;
5108 if (flags & PERF_FLAG_FD_NO_GROUP)
5109 group_leader = NULL;
5113 * Look up the group leader (we will attach this event to it):
5119 * Do not allow a recursive hierarchy (this new sibling
5120 * becoming part of another group-sibling):
5122 if (group_leader->group_leader != group_leader)
5123 goto err_put_context;
5125 * Do not allow to attach to a group in a different
5126 * task or CPU context:
5128 if (group_leader->ctx != ctx)
5129 goto err_put_context;
5131 * Only a group leader can be exclusive or pinned
5133 if (attr.exclusive || attr.pinned)
5134 goto err_put_context;
5137 event = perf_event_alloc(&attr, cpu, ctx, group_leader,
5138 NULL, NULL, GFP_KERNEL);
5139 if (IS_ERR(event)) {
5140 err = PTR_ERR(event);
5141 goto err_put_context;
5145 err = perf_event_set_output(event, output_event);
5147 goto err_free_put_context;
5150 event_file = anon_inode_getfile("[perf_event]", &perf_fops, event, O_RDWR);
5151 if (IS_ERR(event_file)) {
5152 err = PTR_ERR(event_file);
5153 goto err_free_put_context;
5156 event->filp = event_file;
5157 WARN_ON_ONCE(ctx->parent_ctx);
5158 mutex_lock(&ctx->mutex);
5159 perf_install_in_context(ctx, event, cpu);
5161 mutex_unlock(&ctx->mutex);
5163 event->owner = current;
5164 get_task_struct(current);
5165 mutex_lock(¤t->perf_event_mutex);
5166 list_add_tail(&event->owner_entry, ¤t->perf_event_list);
5167 mutex_unlock(¤t->perf_event_mutex);
5170 * Drop the reference on the group_event after placing the
5171 * new event on the sibling_list. This ensures destruction
5172 * of the group leader will find the pointer to itself in
5173 * perf_group_detach().
5175 fput_light(group_file, fput_needed);
5176 fd_install(event_fd, event_file);
5179 err_free_put_context:
5182 fput_light(group_file, fput_needed);
5185 put_unused_fd(event_fd);
5190 * perf_event_create_kernel_counter
5192 * @attr: attributes of the counter to create
5193 * @cpu: cpu in which the counter is bound
5194 * @pid: task to profile
5197 perf_event_create_kernel_counter(struct perf_event_attr *attr, int cpu,
5199 perf_overflow_handler_t overflow_handler)
5201 struct perf_event *event;
5202 struct perf_event_context *ctx;
5206 * Get the target context (task or percpu):
5209 ctx = find_get_context(pid, cpu);
5215 event = perf_event_alloc(attr, cpu, ctx, NULL,
5216 NULL, overflow_handler, GFP_KERNEL);
5217 if (IS_ERR(event)) {
5218 err = PTR_ERR(event);
5219 goto err_put_context;
5223 WARN_ON_ONCE(ctx->parent_ctx);
5224 mutex_lock(&ctx->mutex);
5225 perf_install_in_context(ctx, event, cpu);
5227 mutex_unlock(&ctx->mutex);
5229 event->owner = current;
5230 get_task_struct(current);
5231 mutex_lock(¤t->perf_event_mutex);
5232 list_add_tail(&event->owner_entry, ¤t->perf_event_list);
5233 mutex_unlock(¤t->perf_event_mutex);
5240 return ERR_PTR(err);
5242 EXPORT_SYMBOL_GPL(perf_event_create_kernel_counter);
5245 * inherit a event from parent task to child task:
5247 static struct perf_event *
5248 inherit_event(struct perf_event *parent_event,
5249 struct task_struct *parent,
5250 struct perf_event_context *parent_ctx,
5251 struct task_struct *child,
5252 struct perf_event *group_leader,
5253 struct perf_event_context *child_ctx)
5255 struct perf_event *child_event;
5258 * Instead of creating recursive hierarchies of events,
5259 * we link inherited events back to the original parent,
5260 * which has a filp for sure, which we use as the reference
5263 if (parent_event->parent)
5264 parent_event = parent_event->parent;
5266 child_event = perf_event_alloc(&parent_event->attr,
5267 parent_event->cpu, child_ctx,
5268 group_leader, parent_event,
5270 if (IS_ERR(child_event))
5275 * Make the child state follow the state of the parent event,
5276 * not its attr.disabled bit. We hold the parent's mutex,
5277 * so we won't race with perf_event_{en, dis}able_family.
5279 if (parent_event->state >= PERF_EVENT_STATE_INACTIVE)
5280 child_event->state = PERF_EVENT_STATE_INACTIVE;
5282 child_event->state = PERF_EVENT_STATE_OFF;
5284 if (parent_event->attr.freq) {
5285 u64 sample_period = parent_event->hw.sample_period;
5286 struct hw_perf_event *hwc = &child_event->hw;
5288 hwc->sample_period = sample_period;
5289 hwc->last_period = sample_period;
5291 atomic64_set(&hwc->period_left, sample_period);
5294 child_event->overflow_handler = parent_event->overflow_handler;
5297 * Link it up in the child's context:
5299 add_event_to_ctx(child_event, child_ctx);
5302 * Get a reference to the parent filp - we will fput it
5303 * when the child event exits. This is safe to do because
5304 * we are in the parent and we know that the filp still
5305 * exists and has a nonzero count:
5307 atomic_long_inc(&parent_event->filp->f_count);
5310 * Link this into the parent event's child list
5312 WARN_ON_ONCE(parent_event->ctx->parent_ctx);
5313 mutex_lock(&parent_event->child_mutex);
5314 list_add_tail(&child_event->child_list, &parent_event->child_list);
5315 mutex_unlock(&parent_event->child_mutex);
5320 static int inherit_group(struct perf_event *parent_event,
5321 struct task_struct *parent,
5322 struct perf_event_context *parent_ctx,
5323 struct task_struct *child,
5324 struct perf_event_context *child_ctx)
5326 struct perf_event *leader;
5327 struct perf_event *sub;
5328 struct perf_event *child_ctr;
5330 leader = inherit_event(parent_event, parent, parent_ctx,
5331 child, NULL, child_ctx);
5333 return PTR_ERR(leader);
5334 list_for_each_entry(sub, &parent_event->sibling_list, group_entry) {
5335 child_ctr = inherit_event(sub, parent, parent_ctx,
5336 child, leader, child_ctx);
5337 if (IS_ERR(child_ctr))
5338 return PTR_ERR(child_ctr);
5343 static void sync_child_event(struct perf_event *child_event,
5344 struct task_struct *child)
5346 struct perf_event *parent_event = child_event->parent;
5349 if (child_event->attr.inherit_stat)
5350 perf_event_read_event(child_event, child);
5352 child_val = atomic64_read(&child_event->count);
5355 * Add back the child's count to the parent's count:
5357 atomic64_add(child_val, &parent_event->count);
5358 atomic64_add(child_event->total_time_enabled,
5359 &parent_event->child_total_time_enabled);
5360 atomic64_add(child_event->total_time_running,
5361 &parent_event->child_total_time_running);
5364 * Remove this event from the parent's list
5366 WARN_ON_ONCE(parent_event->ctx->parent_ctx);
5367 mutex_lock(&parent_event->child_mutex);
5368 list_del_init(&child_event->child_list);
5369 mutex_unlock(&parent_event->child_mutex);
5372 * Release the parent event, if this was the last
5375 fput(parent_event->filp);
5379 __perf_event_exit_task(struct perf_event *child_event,
5380 struct perf_event_context *child_ctx,
5381 struct task_struct *child)
5383 struct perf_event *parent_event;
5385 perf_event_remove_from_context(child_event);
5387 parent_event = child_event->parent;
5389 * It can happen that parent exits first, and has events
5390 * that are still around due to the child reference. These
5391 * events need to be zapped - but otherwise linger.
5394 sync_child_event(child_event, child);
5395 free_event(child_event);
5400 * When a child task exits, feed back event values to parent events.
5402 void perf_event_exit_task(struct task_struct *child)
5404 struct perf_event *child_event, *tmp;
5405 struct perf_event_context *child_ctx;
5406 unsigned long flags;
5408 if (likely(!child->perf_event_ctxp)) {
5409 perf_event_task(child, NULL, 0);
5413 local_irq_save(flags);
5415 * We can't reschedule here because interrupts are disabled,
5416 * and either child is current or it is a task that can't be
5417 * scheduled, so we are now safe from rescheduling changing
5420 child_ctx = child->perf_event_ctxp;
5421 __perf_event_task_sched_out(child_ctx);
5424 * Take the context lock here so that if find_get_context is
5425 * reading child->perf_event_ctxp, we wait until it has
5426 * incremented the context's refcount before we do put_ctx below.
5428 raw_spin_lock(&child_ctx->lock);
5429 child->perf_event_ctxp = NULL;
5431 * If this context is a clone; unclone it so it can't get
5432 * swapped to another process while we're removing all
5433 * the events from it.
5435 unclone_ctx(child_ctx);
5436 update_context_time(child_ctx);
5437 raw_spin_unlock_irqrestore(&child_ctx->lock, flags);
5440 * Report the task dead after unscheduling the events so that we
5441 * won't get any samples after PERF_RECORD_EXIT. We can however still
5442 * get a few PERF_RECORD_READ events.
5444 perf_event_task(child, child_ctx, 0);
5447 * We can recurse on the same lock type through:
5449 * __perf_event_exit_task()
5450 * sync_child_event()
5451 * fput(parent_event->filp)
5453 * mutex_lock(&ctx->mutex)
5455 * But since its the parent context it won't be the same instance.
5457 mutex_lock(&child_ctx->mutex);
5460 list_for_each_entry_safe(child_event, tmp, &child_ctx->pinned_groups,
5462 __perf_event_exit_task(child_event, child_ctx, child);
5464 list_for_each_entry_safe(child_event, tmp, &child_ctx->flexible_groups,
5466 __perf_event_exit_task(child_event, child_ctx, child);
5469 * If the last event was a group event, it will have appended all
5470 * its siblings to the list, but we obtained 'tmp' before that which
5471 * will still point to the list head terminating the iteration.
5473 if (!list_empty(&child_ctx->pinned_groups) ||
5474 !list_empty(&child_ctx->flexible_groups))
5477 mutex_unlock(&child_ctx->mutex);
5482 static void perf_free_event(struct perf_event *event,
5483 struct perf_event_context *ctx)
5485 struct perf_event *parent = event->parent;
5487 if (WARN_ON_ONCE(!parent))
5490 mutex_lock(&parent->child_mutex);
5491 list_del_init(&event->child_list);
5492 mutex_unlock(&parent->child_mutex);
5496 perf_group_detach(event);
5497 list_del_event(event, ctx);
5502 * free an unexposed, unused context as created by inheritance by
5503 * init_task below, used by fork() in case of fail.
5505 void perf_event_free_task(struct task_struct *task)
5507 struct perf_event_context *ctx = task->perf_event_ctxp;
5508 struct perf_event *event, *tmp;
5513 mutex_lock(&ctx->mutex);
5515 list_for_each_entry_safe(event, tmp, &ctx->pinned_groups, group_entry)
5516 perf_free_event(event, ctx);
5518 list_for_each_entry_safe(event, tmp, &ctx->flexible_groups,
5520 perf_free_event(event, ctx);
5522 if (!list_empty(&ctx->pinned_groups) ||
5523 !list_empty(&ctx->flexible_groups))
5526 mutex_unlock(&ctx->mutex);
5532 inherit_task_group(struct perf_event *event, struct task_struct *parent,
5533 struct perf_event_context *parent_ctx,
5534 struct task_struct *child,
5538 struct perf_event_context *child_ctx = child->perf_event_ctxp;
5540 if (!event->attr.inherit) {
5547 * This is executed from the parent task context, so
5548 * inherit events that have been marked for cloning.
5549 * First allocate and initialize a context for the
5553 child_ctx = kzalloc(sizeof(struct perf_event_context),
5558 __perf_event_init_context(child_ctx, child);
5559 child->perf_event_ctxp = child_ctx;
5560 get_task_struct(child);
5563 ret = inherit_group(event, parent, parent_ctx,
5574 * Initialize the perf_event context in task_struct
5576 int perf_event_init_task(struct task_struct *child)
5578 struct perf_event_context *child_ctx, *parent_ctx;
5579 struct perf_event_context *cloned_ctx;
5580 struct perf_event *event;
5581 struct task_struct *parent = current;
5582 int inherited_all = 1;
5585 child->perf_event_ctxp = NULL;
5587 mutex_init(&child->perf_event_mutex);
5588 INIT_LIST_HEAD(&child->perf_event_list);
5590 if (likely(!parent->perf_event_ctxp))
5594 * If the parent's context is a clone, pin it so it won't get
5597 parent_ctx = perf_pin_task_context(parent);
5600 * No need to check if parent_ctx != NULL here; since we saw
5601 * it non-NULL earlier, the only reason for it to become NULL
5602 * is if we exit, and since we're currently in the middle of
5603 * a fork we can't be exiting at the same time.
5607 * Lock the parent list. No need to lock the child - not PID
5608 * hashed yet and not running, so nobody can access it.
5610 mutex_lock(&parent_ctx->mutex);
5613 * We dont have to disable NMIs - we are only looking at
5614 * the list, not manipulating it:
5616 list_for_each_entry(event, &parent_ctx->pinned_groups, group_entry) {
5617 ret = inherit_task_group(event, parent, parent_ctx, child,
5623 list_for_each_entry(event, &parent_ctx->flexible_groups, group_entry) {
5624 ret = inherit_task_group(event, parent, parent_ctx, child,
5630 child_ctx = child->perf_event_ctxp;
5632 if (child_ctx && inherited_all) {
5634 * Mark the child context as a clone of the parent
5635 * context, or of whatever the parent is a clone of.
5636 * Note that if the parent is a clone, it could get
5637 * uncloned at any point, but that doesn't matter
5638 * because the list of events and the generation
5639 * count can't have changed since we took the mutex.
5641 cloned_ctx = rcu_dereference(parent_ctx->parent_ctx);
5643 child_ctx->parent_ctx = cloned_ctx;
5644 child_ctx->parent_gen = parent_ctx->parent_gen;
5646 child_ctx->parent_ctx = parent_ctx;
5647 child_ctx->parent_gen = parent_ctx->generation;
5649 get_ctx(child_ctx->parent_ctx);
5652 mutex_unlock(&parent_ctx->mutex);
5654 perf_unpin_context(parent_ctx);
5659 static void __init perf_event_init_all_cpus(void)
5662 struct perf_cpu_context *cpuctx;
5664 for_each_possible_cpu(cpu) {
5665 cpuctx = &per_cpu(perf_cpu_context, cpu);
5666 mutex_init(&cpuctx->hlist_mutex);
5667 __perf_event_init_context(&cpuctx->ctx, NULL);
5671 static void __cpuinit perf_event_init_cpu(int cpu)
5673 struct perf_cpu_context *cpuctx;
5675 cpuctx = &per_cpu(perf_cpu_context, cpu);
5677 spin_lock(&perf_resource_lock);
5678 cpuctx->max_pertask = perf_max_events - perf_reserved_percpu;
5679 spin_unlock(&perf_resource_lock);
5681 mutex_lock(&cpuctx->hlist_mutex);
5682 if (cpuctx->hlist_refcount > 0) {
5683 struct swevent_hlist *hlist;
5685 hlist = kzalloc(sizeof(*hlist), GFP_KERNEL);
5686 WARN_ON_ONCE(!hlist);
5687 rcu_assign_pointer(cpuctx->swevent_hlist, hlist);
5689 mutex_unlock(&cpuctx->hlist_mutex);
5692 #ifdef CONFIG_HOTPLUG_CPU
5693 static void __perf_event_exit_cpu(void *info)
5695 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
5696 struct perf_event_context *ctx = &cpuctx->ctx;
5697 struct perf_event *event, *tmp;
5699 list_for_each_entry_safe(event, tmp, &ctx->pinned_groups, group_entry)
5700 __perf_event_remove_from_context(event);
5701 list_for_each_entry_safe(event, tmp, &ctx->flexible_groups, group_entry)
5702 __perf_event_remove_from_context(event);
5704 static void perf_event_exit_cpu(int cpu)
5706 struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
5707 struct perf_event_context *ctx = &cpuctx->ctx;
5709 mutex_lock(&cpuctx->hlist_mutex);
5710 swevent_hlist_release(cpuctx);
5711 mutex_unlock(&cpuctx->hlist_mutex);
5713 mutex_lock(&ctx->mutex);
5714 smp_call_function_single(cpu, __perf_event_exit_cpu, NULL, 1);
5715 mutex_unlock(&ctx->mutex);
5718 static inline void perf_event_exit_cpu(int cpu) { }
5721 static int __cpuinit
5722 perf_cpu_notify(struct notifier_block *self, unsigned long action, void *hcpu)
5724 unsigned int cpu = (long)hcpu;
5728 case CPU_UP_PREPARE:
5729 case CPU_UP_PREPARE_FROZEN:
5730 perf_event_init_cpu(cpu);
5733 case CPU_DOWN_PREPARE:
5734 case CPU_DOWN_PREPARE_FROZEN:
5735 perf_event_exit_cpu(cpu);
5746 * This has to have a higher priority than migration_notifier in sched.c.
5748 static struct notifier_block __cpuinitdata perf_cpu_nb = {
5749 .notifier_call = perf_cpu_notify,
5753 void __init perf_event_init(void)
5755 perf_event_init_all_cpus();
5756 perf_cpu_notify(&perf_cpu_nb, (unsigned long)CPU_UP_PREPARE,
5757 (void *)(long)smp_processor_id());
5758 perf_cpu_notify(&perf_cpu_nb, (unsigned long)CPU_ONLINE,
5759 (void *)(long)smp_processor_id());
5760 register_cpu_notifier(&perf_cpu_nb);
5763 static ssize_t perf_show_reserve_percpu(struct sysdev_class *class,
5764 struct sysdev_class_attribute *attr,
5767 return sprintf(buf, "%d\n", perf_reserved_percpu);
5771 perf_set_reserve_percpu(struct sysdev_class *class,
5772 struct sysdev_class_attribute *attr,
5776 struct perf_cpu_context *cpuctx;
5780 err = strict_strtoul(buf, 10, &val);
5783 if (val > perf_max_events)
5786 spin_lock(&perf_resource_lock);
5787 perf_reserved_percpu = val;
5788 for_each_online_cpu(cpu) {
5789 cpuctx = &per_cpu(perf_cpu_context, cpu);
5790 raw_spin_lock_irq(&cpuctx->ctx.lock);
5791 mpt = min(perf_max_events - cpuctx->ctx.nr_events,
5792 perf_max_events - perf_reserved_percpu);
5793 cpuctx->max_pertask = mpt;
5794 raw_spin_unlock_irq(&cpuctx->ctx.lock);
5796 spin_unlock(&perf_resource_lock);
5801 static ssize_t perf_show_overcommit(struct sysdev_class *class,
5802 struct sysdev_class_attribute *attr,
5805 return sprintf(buf, "%d\n", perf_overcommit);
5809 perf_set_overcommit(struct sysdev_class *class,
5810 struct sysdev_class_attribute *attr,
5811 const char *buf, size_t count)
5816 err = strict_strtoul(buf, 10, &val);
5822 spin_lock(&perf_resource_lock);
5823 perf_overcommit = val;
5824 spin_unlock(&perf_resource_lock);
5829 static SYSDEV_CLASS_ATTR(
5832 perf_show_reserve_percpu,
5833 perf_set_reserve_percpu
5836 static SYSDEV_CLASS_ATTR(
5839 perf_show_overcommit,
5843 static struct attribute *perfclass_attrs[] = {
5844 &attr_reserve_percpu.attr,
5845 &attr_overcommit.attr,
5849 static struct attribute_group perfclass_attr_group = {
5850 .attrs = perfclass_attrs,
5851 .name = "perf_events",
5854 static int __init perf_event_sysfs_init(void)
5856 return sysfs_create_group(&cpu_sysdev_class.kset.kobj,
5857 &perfclass_attr_group);
5859 device_initcall(perf_event_sysfs_init);