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 local_clock();
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_filter_match(struct perf_event *event)
408 return event->cpu == -1 || event->cpu == smp_processor_id();
412 event_sched_out(struct perf_event *event,
413 struct perf_cpu_context *cpuctx,
414 struct perf_event_context *ctx)
418 * An event which could not be activated because of
419 * filter mismatch still needs to have its timings
420 * maintained, otherwise bogus information is return
421 * via read() for time_enabled, time_running:
423 if (event->state == PERF_EVENT_STATE_INACTIVE
424 && !event_filter_match(event)) {
425 delta = ctx->time - event->tstamp_stopped;
426 event->tstamp_running += delta;
427 event->tstamp_stopped = ctx->time;
430 if (event->state != PERF_EVENT_STATE_ACTIVE)
433 event->state = PERF_EVENT_STATE_INACTIVE;
434 if (event->pending_disable) {
435 event->pending_disable = 0;
436 event->state = PERF_EVENT_STATE_OFF;
438 event->tstamp_stopped = ctx->time;
439 event->pmu->disable(event);
442 if (!is_software_event(event))
443 cpuctx->active_oncpu--;
445 if (event->attr.exclusive || !cpuctx->active_oncpu)
446 cpuctx->exclusive = 0;
450 group_sched_out(struct perf_event *group_event,
451 struct perf_cpu_context *cpuctx,
452 struct perf_event_context *ctx)
454 struct perf_event *event;
455 int state = group_event->state;
457 event_sched_out(group_event, cpuctx, ctx);
460 * Schedule out siblings (if any):
462 list_for_each_entry(event, &group_event->sibling_list, group_entry)
463 event_sched_out(event, cpuctx, ctx);
465 if (state == PERF_EVENT_STATE_ACTIVE && group_event->attr.exclusive)
466 cpuctx->exclusive = 0;
470 * Cross CPU call to remove a performance event
472 * We disable the event on the hardware level first. After that we
473 * remove it from the context list.
475 static void __perf_event_remove_from_context(void *info)
477 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
478 struct perf_event *event = info;
479 struct perf_event_context *ctx = event->ctx;
482 * If this is a task context, we need to check whether it is
483 * the current task context of this cpu. If not it has been
484 * scheduled out before the smp call arrived.
486 if (ctx->task && cpuctx->task_ctx != ctx)
489 raw_spin_lock(&ctx->lock);
491 * Protect the list operation against NMI by disabling the
492 * events on a global level.
496 event_sched_out(event, cpuctx, ctx);
498 list_del_event(event, ctx);
502 * Allow more per task events with respect to the
505 cpuctx->max_pertask =
506 min(perf_max_events - ctx->nr_events,
507 perf_max_events - perf_reserved_percpu);
511 raw_spin_unlock(&ctx->lock);
516 * Remove the event from a task's (or a CPU's) list of events.
518 * Must be called with ctx->mutex held.
520 * CPU events are removed with a smp call. For task events we only
521 * call when the task is on a CPU.
523 * If event->ctx is a cloned context, callers must make sure that
524 * every task struct that event->ctx->task could possibly point to
525 * remains valid. This is OK when called from perf_release since
526 * that only calls us on the top-level context, which can't be a clone.
527 * When called from perf_event_exit_task, it's OK because the
528 * context has been detached from its task.
530 static void perf_event_remove_from_context(struct perf_event *event)
532 struct perf_event_context *ctx = event->ctx;
533 struct task_struct *task = ctx->task;
537 * Per cpu events are removed via an smp call and
538 * the removal is always successful.
540 smp_call_function_single(event->cpu,
541 __perf_event_remove_from_context,
547 task_oncpu_function_call(task, __perf_event_remove_from_context,
550 raw_spin_lock_irq(&ctx->lock);
552 * If the context is active we need to retry the smp call.
554 if (ctx->nr_active && !list_empty(&event->group_entry)) {
555 raw_spin_unlock_irq(&ctx->lock);
560 * The lock prevents that this context is scheduled in so we
561 * can remove the event safely, if the call above did not
564 if (!list_empty(&event->group_entry))
565 list_del_event(event, ctx);
566 raw_spin_unlock_irq(&ctx->lock);
570 * Cross CPU call to disable a performance event
572 static void __perf_event_disable(void *info)
574 struct perf_event *event = info;
575 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
576 struct perf_event_context *ctx = event->ctx;
579 * If this is a per-task event, need to check whether this
580 * event's task is the current task on this cpu.
582 if (ctx->task && cpuctx->task_ctx != ctx)
585 raw_spin_lock(&ctx->lock);
588 * If the event is on, turn it off.
589 * If it is in error state, leave it in error state.
591 if (event->state >= PERF_EVENT_STATE_INACTIVE) {
592 update_context_time(ctx);
593 update_group_times(event);
594 if (event == event->group_leader)
595 group_sched_out(event, cpuctx, ctx);
597 event_sched_out(event, cpuctx, ctx);
598 event->state = PERF_EVENT_STATE_OFF;
601 raw_spin_unlock(&ctx->lock);
607 * If event->ctx is a cloned context, callers must make sure that
608 * every task struct that event->ctx->task could possibly point to
609 * remains valid. This condition is satisifed when called through
610 * perf_event_for_each_child or perf_event_for_each because they
611 * hold the top-level event's child_mutex, so any descendant that
612 * goes to exit will block in sync_child_event.
613 * When called from perf_pending_event it's OK because event->ctx
614 * is the current context on this CPU and preemption is disabled,
615 * hence we can't get into perf_event_task_sched_out for this context.
617 void perf_event_disable(struct perf_event *event)
619 struct perf_event_context *ctx = event->ctx;
620 struct task_struct *task = ctx->task;
624 * Disable the event on the cpu that it's on
626 smp_call_function_single(event->cpu, __perf_event_disable,
632 task_oncpu_function_call(task, __perf_event_disable, event);
634 raw_spin_lock_irq(&ctx->lock);
636 * If the event is still active, we need to retry the cross-call.
638 if (event->state == PERF_EVENT_STATE_ACTIVE) {
639 raw_spin_unlock_irq(&ctx->lock);
644 * Since we have the lock this context can't be scheduled
645 * in, so we can change the state safely.
647 if (event->state == PERF_EVENT_STATE_INACTIVE) {
648 update_group_times(event);
649 event->state = PERF_EVENT_STATE_OFF;
652 raw_spin_unlock_irq(&ctx->lock);
656 event_sched_in(struct perf_event *event,
657 struct perf_cpu_context *cpuctx,
658 struct perf_event_context *ctx)
660 if (event->state <= PERF_EVENT_STATE_OFF)
663 event->state = PERF_EVENT_STATE_ACTIVE;
664 event->oncpu = smp_processor_id();
666 * The new state must be visible before we turn it on in the hardware:
670 if (event->pmu->enable(event)) {
671 event->state = PERF_EVENT_STATE_INACTIVE;
676 event->tstamp_running += ctx->time - event->tstamp_stopped;
678 if (!is_software_event(event))
679 cpuctx->active_oncpu++;
682 if (event->attr.exclusive)
683 cpuctx->exclusive = 1;
689 group_sched_in(struct perf_event *group_event,
690 struct perf_cpu_context *cpuctx,
691 struct perf_event_context *ctx)
693 struct perf_event *event, *partial_group = NULL;
694 const struct pmu *pmu = group_event->pmu;
697 if (group_event->state == PERF_EVENT_STATE_OFF)
700 /* Check if group transaction availabe */
707 if (event_sched_in(group_event, cpuctx, ctx)) {
709 pmu->cancel_txn(pmu);
714 * Schedule in siblings as one group (if any):
716 list_for_each_entry(event, &group_event->sibling_list, group_entry) {
717 if (event_sched_in(event, cpuctx, ctx)) {
718 partial_group = event;
723 if (!txn || !pmu->commit_txn(pmu))
728 * Groups can be scheduled in as one unit only, so undo any
729 * partial group before returning:
731 list_for_each_entry(event, &group_event->sibling_list, group_entry) {
732 if (event == partial_group)
734 event_sched_out(event, cpuctx, ctx);
736 event_sched_out(group_event, cpuctx, ctx);
739 pmu->cancel_txn(pmu);
745 * Work out whether we can put this event group on the CPU now.
747 static int group_can_go_on(struct perf_event *event,
748 struct perf_cpu_context *cpuctx,
752 * Groups consisting entirely of software events can always go on.
754 if (event->group_flags & PERF_GROUP_SOFTWARE)
757 * If an exclusive group is already on, no other hardware
760 if (cpuctx->exclusive)
763 * If this group is exclusive and there are already
764 * events on the CPU, it can't go on.
766 if (event->attr.exclusive && cpuctx->active_oncpu)
769 * Otherwise, try to add it if all previous groups were able
775 static void add_event_to_ctx(struct perf_event *event,
776 struct perf_event_context *ctx)
778 list_add_event(event, ctx);
779 perf_group_attach(event);
780 event->tstamp_enabled = ctx->time;
781 event->tstamp_running = ctx->time;
782 event->tstamp_stopped = ctx->time;
786 * Cross CPU call to install and enable a performance event
788 * Must be called with ctx->mutex held
790 static void __perf_install_in_context(void *info)
792 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
793 struct perf_event *event = info;
794 struct perf_event_context *ctx = event->ctx;
795 struct perf_event *leader = event->group_leader;
799 * If this is a task context, we need to check whether it is
800 * the current task context of this cpu. If not it has been
801 * scheduled out before the smp call arrived.
802 * Or possibly this is the right context but it isn't
803 * on this cpu because it had no events.
805 if (ctx->task && cpuctx->task_ctx != ctx) {
806 if (cpuctx->task_ctx || ctx->task != current)
808 cpuctx->task_ctx = ctx;
811 raw_spin_lock(&ctx->lock);
813 update_context_time(ctx);
816 * Protect the list operation against NMI by disabling the
817 * events on a global level. NOP for non NMI based events.
821 add_event_to_ctx(event, ctx);
823 if (event->cpu != -1 && event->cpu != smp_processor_id())
827 * Don't put the event on if it is disabled or if
828 * it is in a group and the group isn't on.
830 if (event->state != PERF_EVENT_STATE_INACTIVE ||
831 (leader != event && leader->state != PERF_EVENT_STATE_ACTIVE))
835 * An exclusive event can't go on if there are already active
836 * hardware events, and no hardware event can go on if there
837 * is already an exclusive event on.
839 if (!group_can_go_on(event, cpuctx, 1))
842 err = event_sched_in(event, cpuctx, ctx);
846 * This event couldn't go on. If it is in a group
847 * then we have to pull the whole group off.
848 * If the event group is pinned then put it in error state.
851 group_sched_out(leader, cpuctx, ctx);
852 if (leader->attr.pinned) {
853 update_group_times(leader);
854 leader->state = PERF_EVENT_STATE_ERROR;
858 if (!err && !ctx->task && cpuctx->max_pertask)
859 cpuctx->max_pertask--;
864 raw_spin_unlock(&ctx->lock);
868 * Attach a performance event to a context
870 * First we add the event to the list with the hardware enable bit
871 * in event->hw_config cleared.
873 * If the event is attached to a task which is on a CPU we use a smp
874 * call to enable it in the task context. The task might have been
875 * scheduled away, but we check this in the smp call again.
877 * Must be called with ctx->mutex held.
880 perf_install_in_context(struct perf_event_context *ctx,
881 struct perf_event *event,
884 struct task_struct *task = ctx->task;
888 * Per cpu events are installed via an smp call and
889 * the install is always successful.
891 smp_call_function_single(cpu, __perf_install_in_context,
897 task_oncpu_function_call(task, __perf_install_in_context,
900 raw_spin_lock_irq(&ctx->lock);
902 * we need to retry the smp call.
904 if (ctx->is_active && list_empty(&event->group_entry)) {
905 raw_spin_unlock_irq(&ctx->lock);
910 * The lock prevents that this context is scheduled in so we
911 * can add the event safely, if it the call above did not
914 if (list_empty(&event->group_entry))
915 add_event_to_ctx(event, ctx);
916 raw_spin_unlock_irq(&ctx->lock);
920 * Put a event into inactive state and update time fields.
921 * Enabling the leader of a group effectively enables all
922 * the group members that aren't explicitly disabled, so we
923 * have to update their ->tstamp_enabled also.
924 * Note: this works for group members as well as group leaders
925 * since the non-leader members' sibling_lists will be empty.
927 static void __perf_event_mark_enabled(struct perf_event *event,
928 struct perf_event_context *ctx)
930 struct perf_event *sub;
932 event->state = PERF_EVENT_STATE_INACTIVE;
933 event->tstamp_enabled = ctx->time - event->total_time_enabled;
934 list_for_each_entry(sub, &event->sibling_list, group_entry)
935 if (sub->state >= PERF_EVENT_STATE_INACTIVE)
936 sub->tstamp_enabled =
937 ctx->time - sub->total_time_enabled;
941 * Cross CPU call to enable a performance event
943 static void __perf_event_enable(void *info)
945 struct perf_event *event = info;
946 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
947 struct perf_event_context *ctx = event->ctx;
948 struct perf_event *leader = event->group_leader;
952 * If this is a per-task event, need to check whether this
953 * event's task is the current task on this cpu.
955 if (ctx->task && cpuctx->task_ctx != ctx) {
956 if (cpuctx->task_ctx || ctx->task != current)
958 cpuctx->task_ctx = ctx;
961 raw_spin_lock(&ctx->lock);
963 update_context_time(ctx);
965 if (event->state >= PERF_EVENT_STATE_INACTIVE)
967 __perf_event_mark_enabled(event, ctx);
969 if (event->cpu != -1 && event->cpu != smp_processor_id())
973 * If the event is in a group and isn't the group leader,
974 * then don't put it on unless the group is on.
976 if (leader != event && leader->state != PERF_EVENT_STATE_ACTIVE)
979 if (!group_can_go_on(event, cpuctx, 1)) {
984 err = group_sched_in(event, cpuctx, ctx);
986 err = event_sched_in(event, cpuctx, ctx);
992 * If this event can't go on and it's part of a
993 * group, then the whole group has to come off.
996 group_sched_out(leader, cpuctx, ctx);
997 if (leader->attr.pinned) {
998 update_group_times(leader);
999 leader->state = PERF_EVENT_STATE_ERROR;
1004 raw_spin_unlock(&ctx->lock);
1010 * If event->ctx is a cloned context, callers must make sure that
1011 * every task struct that event->ctx->task could possibly point to
1012 * remains valid. This condition is satisfied when called through
1013 * perf_event_for_each_child or perf_event_for_each as described
1014 * for perf_event_disable.
1016 void perf_event_enable(struct perf_event *event)
1018 struct perf_event_context *ctx = event->ctx;
1019 struct task_struct *task = ctx->task;
1023 * Enable the event on the cpu that it's on
1025 smp_call_function_single(event->cpu, __perf_event_enable,
1030 raw_spin_lock_irq(&ctx->lock);
1031 if (event->state >= PERF_EVENT_STATE_INACTIVE)
1035 * If the event is in error state, clear that first.
1036 * That way, if we see the event in error state below, we
1037 * know that it has gone back into error state, as distinct
1038 * from the task having been scheduled away before the
1039 * cross-call arrived.
1041 if (event->state == PERF_EVENT_STATE_ERROR)
1042 event->state = PERF_EVENT_STATE_OFF;
1045 raw_spin_unlock_irq(&ctx->lock);
1046 task_oncpu_function_call(task, __perf_event_enable, event);
1048 raw_spin_lock_irq(&ctx->lock);
1051 * If the context is active and the event is still off,
1052 * we need to retry the cross-call.
1054 if (ctx->is_active && event->state == PERF_EVENT_STATE_OFF)
1058 * Since we have the lock this context can't be scheduled
1059 * in, so we can change the state safely.
1061 if (event->state == PERF_EVENT_STATE_OFF)
1062 __perf_event_mark_enabled(event, ctx);
1065 raw_spin_unlock_irq(&ctx->lock);
1068 static int perf_event_refresh(struct perf_event *event, int refresh)
1071 * not supported on inherited events
1073 if (event->attr.inherit)
1076 atomic_add(refresh, &event->event_limit);
1077 perf_event_enable(event);
1083 EVENT_FLEXIBLE = 0x1,
1085 EVENT_ALL = EVENT_FLEXIBLE | EVENT_PINNED,
1088 static void ctx_sched_out(struct perf_event_context *ctx,
1089 struct perf_cpu_context *cpuctx,
1090 enum event_type_t event_type)
1092 struct perf_event *event;
1094 raw_spin_lock(&ctx->lock);
1096 if (likely(!ctx->nr_events))
1098 update_context_time(ctx);
1101 if (!ctx->nr_active)
1104 if (event_type & EVENT_PINNED)
1105 list_for_each_entry(event, &ctx->pinned_groups, group_entry)
1106 group_sched_out(event, cpuctx, ctx);
1108 if (event_type & EVENT_FLEXIBLE)
1109 list_for_each_entry(event, &ctx->flexible_groups, group_entry)
1110 group_sched_out(event, cpuctx, ctx);
1115 raw_spin_unlock(&ctx->lock);
1119 * Test whether two contexts are equivalent, i.e. whether they
1120 * have both been cloned from the same version of the same context
1121 * and they both have the same number of enabled events.
1122 * If the number of enabled events is the same, then the set
1123 * of enabled events should be the same, because these are both
1124 * inherited contexts, therefore we can't access individual events
1125 * in them directly with an fd; we can only enable/disable all
1126 * events via prctl, or enable/disable all events in a family
1127 * via ioctl, which will have the same effect on both contexts.
1129 static int context_equiv(struct perf_event_context *ctx1,
1130 struct perf_event_context *ctx2)
1132 return ctx1->parent_ctx && ctx1->parent_ctx == ctx2->parent_ctx
1133 && ctx1->parent_gen == ctx2->parent_gen
1134 && !ctx1->pin_count && !ctx2->pin_count;
1137 static void __perf_event_sync_stat(struct perf_event *event,
1138 struct perf_event *next_event)
1142 if (!event->attr.inherit_stat)
1146 * Update the event value, we cannot use perf_event_read()
1147 * because we're in the middle of a context switch and have IRQs
1148 * disabled, which upsets smp_call_function_single(), however
1149 * we know the event must be on the current CPU, therefore we
1150 * don't need to use it.
1152 switch (event->state) {
1153 case PERF_EVENT_STATE_ACTIVE:
1154 event->pmu->read(event);
1157 case PERF_EVENT_STATE_INACTIVE:
1158 update_event_times(event);
1166 * In order to keep per-task stats reliable we need to flip the event
1167 * values when we flip the contexts.
1169 value = local64_read(&next_event->count);
1170 value = local64_xchg(&event->count, value);
1171 local64_set(&next_event->count, value);
1173 swap(event->total_time_enabled, next_event->total_time_enabled);
1174 swap(event->total_time_running, next_event->total_time_running);
1177 * Since we swizzled the values, update the user visible data too.
1179 perf_event_update_userpage(event);
1180 perf_event_update_userpage(next_event);
1183 #define list_next_entry(pos, member) \
1184 list_entry(pos->member.next, typeof(*pos), member)
1186 static void perf_event_sync_stat(struct perf_event_context *ctx,
1187 struct perf_event_context *next_ctx)
1189 struct perf_event *event, *next_event;
1194 update_context_time(ctx);
1196 event = list_first_entry(&ctx->event_list,
1197 struct perf_event, event_entry);
1199 next_event = list_first_entry(&next_ctx->event_list,
1200 struct perf_event, event_entry);
1202 while (&event->event_entry != &ctx->event_list &&
1203 &next_event->event_entry != &next_ctx->event_list) {
1205 __perf_event_sync_stat(event, next_event);
1207 event = list_next_entry(event, event_entry);
1208 next_event = list_next_entry(next_event, event_entry);
1213 * Called from scheduler to remove the events of the current task,
1214 * with interrupts disabled.
1216 * We stop each event and update the event value in event->count.
1218 * This does not protect us against NMI, but disable()
1219 * sets the disabled bit in the control field of event _before_
1220 * accessing the event control register. If a NMI hits, then it will
1221 * not restart the event.
1223 void perf_event_task_sched_out(struct task_struct *task,
1224 struct task_struct *next)
1226 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
1227 struct perf_event_context *ctx = task->perf_event_ctxp;
1228 struct perf_event_context *next_ctx;
1229 struct perf_event_context *parent;
1232 perf_sw_event(PERF_COUNT_SW_CONTEXT_SWITCHES, 1, 1, NULL, 0);
1234 if (likely(!ctx || !cpuctx->task_ctx))
1238 parent = rcu_dereference(ctx->parent_ctx);
1239 next_ctx = next->perf_event_ctxp;
1240 if (parent && next_ctx &&
1241 rcu_dereference(next_ctx->parent_ctx) == parent) {
1243 * Looks like the two contexts are clones, so we might be
1244 * able to optimize the context switch. We lock both
1245 * contexts and check that they are clones under the
1246 * lock (including re-checking that neither has been
1247 * uncloned in the meantime). It doesn't matter which
1248 * order we take the locks because no other cpu could
1249 * be trying to lock both of these tasks.
1251 raw_spin_lock(&ctx->lock);
1252 raw_spin_lock_nested(&next_ctx->lock, SINGLE_DEPTH_NESTING);
1253 if (context_equiv(ctx, next_ctx)) {
1255 * XXX do we need a memory barrier of sorts
1256 * wrt to rcu_dereference() of perf_event_ctxp
1258 task->perf_event_ctxp = next_ctx;
1259 next->perf_event_ctxp = ctx;
1261 next_ctx->task = task;
1264 perf_event_sync_stat(ctx, next_ctx);
1266 raw_spin_unlock(&next_ctx->lock);
1267 raw_spin_unlock(&ctx->lock);
1272 ctx_sched_out(ctx, cpuctx, EVENT_ALL);
1273 cpuctx->task_ctx = NULL;
1277 static void task_ctx_sched_out(struct perf_event_context *ctx,
1278 enum event_type_t event_type)
1280 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
1282 if (!cpuctx->task_ctx)
1285 if (WARN_ON_ONCE(ctx != cpuctx->task_ctx))
1288 ctx_sched_out(ctx, cpuctx, event_type);
1289 cpuctx->task_ctx = NULL;
1293 * Called with IRQs disabled
1295 static void __perf_event_task_sched_out(struct perf_event_context *ctx)
1297 task_ctx_sched_out(ctx, EVENT_ALL);
1301 * Called with IRQs disabled
1303 static void cpu_ctx_sched_out(struct perf_cpu_context *cpuctx,
1304 enum event_type_t event_type)
1306 ctx_sched_out(&cpuctx->ctx, cpuctx, event_type);
1310 ctx_pinned_sched_in(struct perf_event_context *ctx,
1311 struct perf_cpu_context *cpuctx)
1313 struct perf_event *event;
1315 list_for_each_entry(event, &ctx->pinned_groups, group_entry) {
1316 if (event->state <= PERF_EVENT_STATE_OFF)
1318 if (event->cpu != -1 && event->cpu != smp_processor_id())
1321 if (group_can_go_on(event, cpuctx, 1))
1322 group_sched_in(event, cpuctx, ctx);
1325 * If this pinned group hasn't been scheduled,
1326 * put it in error state.
1328 if (event->state == PERF_EVENT_STATE_INACTIVE) {
1329 update_group_times(event);
1330 event->state = PERF_EVENT_STATE_ERROR;
1336 ctx_flexible_sched_in(struct perf_event_context *ctx,
1337 struct perf_cpu_context *cpuctx)
1339 struct perf_event *event;
1342 list_for_each_entry(event, &ctx->flexible_groups, group_entry) {
1343 /* Ignore events in OFF or ERROR state */
1344 if (event->state <= PERF_EVENT_STATE_OFF)
1347 * Listen to the 'cpu' scheduling filter constraint
1350 if (event->cpu != -1 && event->cpu != smp_processor_id())
1353 if (group_can_go_on(event, cpuctx, can_add_hw))
1354 if (group_sched_in(event, cpuctx, ctx))
1360 ctx_sched_in(struct perf_event_context *ctx,
1361 struct perf_cpu_context *cpuctx,
1362 enum event_type_t event_type)
1364 raw_spin_lock(&ctx->lock);
1366 if (likely(!ctx->nr_events))
1369 ctx->timestamp = perf_clock();
1374 * First go through the list and put on any pinned groups
1375 * in order to give them the best chance of going on.
1377 if (event_type & EVENT_PINNED)
1378 ctx_pinned_sched_in(ctx, cpuctx);
1380 /* Then walk through the lower prio flexible groups */
1381 if (event_type & EVENT_FLEXIBLE)
1382 ctx_flexible_sched_in(ctx, cpuctx);
1386 raw_spin_unlock(&ctx->lock);
1389 static void cpu_ctx_sched_in(struct perf_cpu_context *cpuctx,
1390 enum event_type_t event_type)
1392 struct perf_event_context *ctx = &cpuctx->ctx;
1394 ctx_sched_in(ctx, cpuctx, event_type);
1397 static void task_ctx_sched_in(struct task_struct *task,
1398 enum event_type_t event_type)
1400 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
1401 struct perf_event_context *ctx = task->perf_event_ctxp;
1405 if (cpuctx->task_ctx == ctx)
1407 ctx_sched_in(ctx, cpuctx, event_type);
1408 cpuctx->task_ctx = ctx;
1411 * Called from scheduler to add the events of the current task
1412 * with interrupts disabled.
1414 * We restore the event value and then enable it.
1416 * This does not protect us against NMI, but enable()
1417 * sets the enabled bit in the control field of event _before_
1418 * accessing the event control register. If a NMI hits, then it will
1419 * keep the event running.
1421 void perf_event_task_sched_in(struct task_struct *task)
1423 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
1424 struct perf_event_context *ctx = task->perf_event_ctxp;
1429 if (cpuctx->task_ctx == ctx)
1435 * We want to keep the following priority order:
1436 * cpu pinned (that don't need to move), task pinned,
1437 * cpu flexible, task flexible.
1439 cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE);
1441 ctx_sched_in(ctx, cpuctx, EVENT_PINNED);
1442 cpu_ctx_sched_in(cpuctx, EVENT_FLEXIBLE);
1443 ctx_sched_in(ctx, cpuctx, EVENT_FLEXIBLE);
1445 cpuctx->task_ctx = ctx;
1450 #define MAX_INTERRUPTS (~0ULL)
1452 static void perf_log_throttle(struct perf_event *event, int enable);
1454 static u64 perf_calculate_period(struct perf_event *event, u64 nsec, u64 count)
1456 u64 frequency = event->attr.sample_freq;
1457 u64 sec = NSEC_PER_SEC;
1458 u64 divisor, dividend;
1460 int count_fls, nsec_fls, frequency_fls, sec_fls;
1462 count_fls = fls64(count);
1463 nsec_fls = fls64(nsec);
1464 frequency_fls = fls64(frequency);
1468 * We got @count in @nsec, with a target of sample_freq HZ
1469 * the target period becomes:
1472 * period = -------------------
1473 * @nsec * sample_freq
1478 * Reduce accuracy by one bit such that @a and @b converge
1479 * to a similar magnitude.
1481 #define REDUCE_FLS(a, b) \
1483 if (a##_fls > b##_fls) { \
1493 * Reduce accuracy until either term fits in a u64, then proceed with
1494 * the other, so that finally we can do a u64/u64 division.
1496 while (count_fls + sec_fls > 64 && nsec_fls + frequency_fls > 64) {
1497 REDUCE_FLS(nsec, frequency);
1498 REDUCE_FLS(sec, count);
1501 if (count_fls + sec_fls > 64) {
1502 divisor = nsec * frequency;
1504 while (count_fls + sec_fls > 64) {
1505 REDUCE_FLS(count, sec);
1509 dividend = count * sec;
1511 dividend = count * sec;
1513 while (nsec_fls + frequency_fls > 64) {
1514 REDUCE_FLS(nsec, frequency);
1518 divisor = nsec * frequency;
1524 return div64_u64(dividend, divisor);
1527 static void perf_event_stop(struct perf_event *event)
1529 if (!event->pmu->stop)
1530 return event->pmu->disable(event);
1532 return event->pmu->stop(event);
1535 static int perf_event_start(struct perf_event *event)
1537 if (!event->pmu->start)
1538 return event->pmu->enable(event);
1540 return event->pmu->start(event);
1543 static void perf_adjust_period(struct perf_event *event, u64 nsec, u64 count)
1545 struct hw_perf_event *hwc = &event->hw;
1546 s64 period, sample_period;
1549 period = perf_calculate_period(event, nsec, count);
1551 delta = (s64)(period - hwc->sample_period);
1552 delta = (delta + 7) / 8; /* low pass filter */
1554 sample_period = hwc->sample_period + delta;
1559 hwc->sample_period = sample_period;
1561 if (local64_read(&hwc->period_left) > 8*sample_period) {
1563 perf_event_stop(event);
1564 local64_set(&hwc->period_left, 0);
1565 perf_event_start(event);
1570 static void perf_ctx_adjust_freq(struct perf_event_context *ctx)
1572 struct perf_event *event;
1573 struct hw_perf_event *hwc;
1574 u64 interrupts, now;
1577 raw_spin_lock(&ctx->lock);
1578 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
1579 if (event->state != PERF_EVENT_STATE_ACTIVE)
1582 if (event->cpu != -1 && event->cpu != smp_processor_id())
1587 interrupts = hwc->interrupts;
1588 hwc->interrupts = 0;
1591 * unthrottle events on the tick
1593 if (interrupts == MAX_INTERRUPTS) {
1594 perf_log_throttle(event, 1);
1596 event->pmu->unthrottle(event);
1600 if (!event->attr.freq || !event->attr.sample_freq)
1604 event->pmu->read(event);
1605 now = local64_read(&event->count);
1606 delta = now - hwc->freq_count_stamp;
1607 hwc->freq_count_stamp = now;
1610 perf_adjust_period(event, TICK_NSEC, delta);
1613 raw_spin_unlock(&ctx->lock);
1617 * Round-robin a context's events:
1619 static void rotate_ctx(struct perf_event_context *ctx)
1621 raw_spin_lock(&ctx->lock);
1623 /* Rotate the first entry last of non-pinned groups */
1624 list_rotate_left(&ctx->flexible_groups);
1626 raw_spin_unlock(&ctx->lock);
1629 void perf_event_task_tick(struct task_struct *curr)
1631 struct perf_cpu_context *cpuctx;
1632 struct perf_event_context *ctx;
1635 if (!atomic_read(&nr_events))
1638 cpuctx = &__get_cpu_var(perf_cpu_context);
1639 if (cpuctx->ctx.nr_events &&
1640 cpuctx->ctx.nr_events != cpuctx->ctx.nr_active)
1643 ctx = curr->perf_event_ctxp;
1644 if (ctx && ctx->nr_events && ctx->nr_events != ctx->nr_active)
1647 perf_ctx_adjust_freq(&cpuctx->ctx);
1649 perf_ctx_adjust_freq(ctx);
1655 cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE);
1657 task_ctx_sched_out(ctx, EVENT_FLEXIBLE);
1659 rotate_ctx(&cpuctx->ctx);
1663 cpu_ctx_sched_in(cpuctx, EVENT_FLEXIBLE);
1665 task_ctx_sched_in(curr, EVENT_FLEXIBLE);
1669 static int event_enable_on_exec(struct perf_event *event,
1670 struct perf_event_context *ctx)
1672 if (!event->attr.enable_on_exec)
1675 event->attr.enable_on_exec = 0;
1676 if (event->state >= PERF_EVENT_STATE_INACTIVE)
1679 __perf_event_mark_enabled(event, ctx);
1685 * Enable all of a task's events that have been marked enable-on-exec.
1686 * This expects task == current.
1688 static void perf_event_enable_on_exec(struct task_struct *task)
1690 struct perf_event_context *ctx;
1691 struct perf_event *event;
1692 unsigned long flags;
1696 local_irq_save(flags);
1697 ctx = task->perf_event_ctxp;
1698 if (!ctx || !ctx->nr_events)
1701 __perf_event_task_sched_out(ctx);
1703 raw_spin_lock(&ctx->lock);
1705 list_for_each_entry(event, &ctx->pinned_groups, group_entry) {
1706 ret = event_enable_on_exec(event, ctx);
1711 list_for_each_entry(event, &ctx->flexible_groups, group_entry) {
1712 ret = event_enable_on_exec(event, ctx);
1718 * Unclone this context if we enabled any event.
1723 raw_spin_unlock(&ctx->lock);
1725 perf_event_task_sched_in(task);
1727 local_irq_restore(flags);
1731 * Cross CPU call to read the hardware event
1733 static void __perf_event_read(void *info)
1735 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
1736 struct perf_event *event = info;
1737 struct perf_event_context *ctx = event->ctx;
1740 * If this is a task context, we need to check whether it is
1741 * the current task context of this cpu. If not it has been
1742 * scheduled out before the smp call arrived. In that case
1743 * event->count would have been updated to a recent sample
1744 * when the event was scheduled out.
1746 if (ctx->task && cpuctx->task_ctx != ctx)
1749 raw_spin_lock(&ctx->lock);
1750 update_context_time(ctx);
1751 update_event_times(event);
1752 raw_spin_unlock(&ctx->lock);
1754 event->pmu->read(event);
1757 static inline u64 perf_event_count(struct perf_event *event)
1759 return local64_read(&event->count) + atomic64_read(&event->child_count);
1762 static u64 perf_event_read(struct perf_event *event)
1765 * If event is enabled and currently active on a CPU, update the
1766 * value in the event structure:
1768 if (event->state == PERF_EVENT_STATE_ACTIVE) {
1769 smp_call_function_single(event->oncpu,
1770 __perf_event_read, event, 1);
1771 } else if (event->state == PERF_EVENT_STATE_INACTIVE) {
1772 struct perf_event_context *ctx = event->ctx;
1773 unsigned long flags;
1775 raw_spin_lock_irqsave(&ctx->lock, flags);
1776 update_context_time(ctx);
1777 update_event_times(event);
1778 raw_spin_unlock_irqrestore(&ctx->lock, flags);
1781 return perf_event_count(event);
1788 struct callchain_cpus_entries {
1789 struct rcu_head rcu_head;
1790 struct perf_callchain_entry *cpu_entries[0];
1793 static DEFINE_PER_CPU(int, callchain_recursion[PERF_NR_CONTEXTS]);
1794 static atomic_t nr_callchain_events;
1795 static DEFINE_MUTEX(callchain_mutex);
1796 struct callchain_cpus_entries *callchain_cpus_entries;
1799 __weak void perf_callchain_kernel(struct perf_callchain_entry *entry,
1800 struct pt_regs *regs)
1804 __weak void perf_callchain_user(struct perf_callchain_entry *entry,
1805 struct pt_regs *regs)
1809 static void release_callchain_buffers_rcu(struct rcu_head *head)
1811 struct callchain_cpus_entries *entries;
1814 entries = container_of(head, struct callchain_cpus_entries, rcu_head);
1816 for_each_possible_cpu(cpu)
1817 kfree(entries->cpu_entries[cpu]);
1822 static void release_callchain_buffers(void)
1824 struct callchain_cpus_entries *entries;
1826 entries = callchain_cpus_entries;
1827 rcu_assign_pointer(callchain_cpus_entries, NULL);
1828 call_rcu(&entries->rcu_head, release_callchain_buffers_rcu);
1831 static int alloc_callchain_buffers(void)
1835 struct callchain_cpus_entries *entries;
1838 * We can't use the percpu allocation API for data that can be
1839 * accessed from NMI. Use a temporary manual per cpu allocation
1840 * until that gets sorted out.
1842 size = sizeof(*entries) + sizeof(struct perf_callchain_entry *) *
1843 num_possible_cpus();
1845 entries = kzalloc(size, GFP_KERNEL);
1849 size = sizeof(struct perf_callchain_entry) * PERF_NR_CONTEXTS;
1851 for_each_possible_cpu(cpu) {
1852 entries->cpu_entries[cpu] = kmalloc_node(size, GFP_KERNEL,
1854 if (!entries->cpu_entries[cpu])
1858 rcu_assign_pointer(callchain_cpus_entries, entries);
1863 for_each_possible_cpu(cpu)
1864 kfree(entries->cpu_entries[cpu]);
1870 static int get_callchain_buffers(void)
1875 mutex_lock(&callchain_mutex);
1877 count = atomic_inc_return(&nr_callchain_events);
1878 if (WARN_ON_ONCE(count < 1)) {
1884 /* If the allocation failed, give up */
1885 if (!callchain_cpus_entries)
1890 err = alloc_callchain_buffers();
1892 release_callchain_buffers();
1894 mutex_unlock(&callchain_mutex);
1899 static void put_callchain_buffers(void)
1901 if (atomic_dec_and_mutex_lock(&nr_callchain_events, &callchain_mutex)) {
1902 release_callchain_buffers();
1903 mutex_unlock(&callchain_mutex);
1907 static int get_recursion_context(int *recursion)
1915 else if (in_softirq())
1920 if (recursion[rctx])
1929 static inline void put_recursion_context(int *recursion, int rctx)
1935 static struct perf_callchain_entry *get_callchain_entry(int *rctx)
1938 struct callchain_cpus_entries *entries;
1940 *rctx = get_recursion_context(__get_cpu_var(callchain_recursion));
1944 entries = rcu_dereference(callchain_cpus_entries);
1948 cpu = smp_processor_id();
1950 return &entries->cpu_entries[cpu][*rctx];
1954 put_callchain_entry(int rctx)
1956 put_recursion_context(__get_cpu_var(callchain_recursion), rctx);
1959 static struct perf_callchain_entry *perf_callchain(struct pt_regs *regs)
1962 struct perf_callchain_entry *entry;
1965 entry = get_callchain_entry(&rctx);
1974 if (!user_mode(regs)) {
1975 perf_callchain_store(entry, PERF_CONTEXT_KERNEL);
1976 perf_callchain_kernel(entry, regs);
1978 regs = task_pt_regs(current);
1984 perf_callchain_store(entry, PERF_CONTEXT_USER);
1985 perf_callchain_user(entry, regs);
1989 put_callchain_entry(rctx);
1995 * Initialize the perf_event context in a task_struct:
1998 __perf_event_init_context(struct perf_event_context *ctx,
1999 struct task_struct *task)
2001 raw_spin_lock_init(&ctx->lock);
2002 mutex_init(&ctx->mutex);
2003 INIT_LIST_HEAD(&ctx->pinned_groups);
2004 INIT_LIST_HEAD(&ctx->flexible_groups);
2005 INIT_LIST_HEAD(&ctx->event_list);
2006 atomic_set(&ctx->refcount, 1);
2010 static struct perf_event_context *find_get_context(pid_t pid, int cpu)
2012 struct perf_event_context *ctx;
2013 struct perf_cpu_context *cpuctx;
2014 struct task_struct *task;
2015 unsigned long flags;
2018 if (pid == -1 && cpu != -1) {
2019 /* Must be root to operate on a CPU event: */
2020 if (perf_paranoid_cpu() && !capable(CAP_SYS_ADMIN))
2021 return ERR_PTR(-EACCES);
2023 if (cpu < 0 || cpu >= nr_cpumask_bits)
2024 return ERR_PTR(-EINVAL);
2027 * We could be clever and allow to attach a event to an
2028 * offline CPU and activate it when the CPU comes up, but
2031 if (!cpu_online(cpu))
2032 return ERR_PTR(-ENODEV);
2034 cpuctx = &per_cpu(perf_cpu_context, cpu);
2045 task = find_task_by_vpid(pid);
2047 get_task_struct(task);
2051 return ERR_PTR(-ESRCH);
2054 * Can't attach events to a dying task.
2057 if (task->flags & PF_EXITING)
2060 /* Reuse ptrace permission checks for now. */
2062 if (!ptrace_may_access(task, PTRACE_MODE_READ))
2066 ctx = perf_lock_task_context(task, &flags);
2069 raw_spin_unlock_irqrestore(&ctx->lock, flags);
2073 ctx = kzalloc(sizeof(struct perf_event_context), GFP_KERNEL);
2077 __perf_event_init_context(ctx, task);
2079 if (cmpxchg(&task->perf_event_ctxp, NULL, ctx)) {
2081 * We raced with some other task; use
2082 * the context they set.
2087 get_task_struct(task);
2090 put_task_struct(task);
2094 put_task_struct(task);
2095 return ERR_PTR(err);
2098 static void perf_event_free_filter(struct perf_event *event);
2100 static void free_event_rcu(struct rcu_head *head)
2102 struct perf_event *event;
2104 event = container_of(head, struct perf_event, rcu_head);
2106 put_pid_ns(event->ns);
2107 perf_event_free_filter(event);
2111 static void perf_pending_sync(struct perf_event *event);
2112 static void perf_buffer_put(struct perf_buffer *buffer);
2114 static void free_event(struct perf_event *event)
2116 perf_pending_sync(event);
2118 if (!event->parent) {
2119 atomic_dec(&nr_events);
2120 if (event->attr.mmap || event->attr.mmap_data)
2121 atomic_dec(&nr_mmap_events);
2122 if (event->attr.comm)
2123 atomic_dec(&nr_comm_events);
2124 if (event->attr.task)
2125 atomic_dec(&nr_task_events);
2126 if (event->attr.sample_type & PERF_SAMPLE_CALLCHAIN)
2127 put_callchain_buffers();
2130 if (event->buffer) {
2131 perf_buffer_put(event->buffer);
2132 event->buffer = NULL;
2136 event->destroy(event);
2138 put_ctx(event->ctx);
2139 call_rcu(&event->rcu_head, free_event_rcu);
2142 int perf_event_release_kernel(struct perf_event *event)
2144 struct perf_event_context *ctx = event->ctx;
2147 * Remove from the PMU, can't get re-enabled since we got
2148 * here because the last ref went.
2150 perf_event_disable(event);
2152 WARN_ON_ONCE(ctx->parent_ctx);
2154 * There are two ways this annotation is useful:
2156 * 1) there is a lock recursion from perf_event_exit_task
2157 * see the comment there.
2159 * 2) there is a lock-inversion with mmap_sem through
2160 * perf_event_read_group(), which takes faults while
2161 * holding ctx->mutex, however this is called after
2162 * the last filedesc died, so there is no possibility
2163 * to trigger the AB-BA case.
2165 mutex_lock_nested(&ctx->mutex, SINGLE_DEPTH_NESTING);
2166 raw_spin_lock_irq(&ctx->lock);
2167 perf_group_detach(event);
2168 list_del_event(event, ctx);
2169 raw_spin_unlock_irq(&ctx->lock);
2170 mutex_unlock(&ctx->mutex);
2172 mutex_lock(&event->owner->perf_event_mutex);
2173 list_del_init(&event->owner_entry);
2174 mutex_unlock(&event->owner->perf_event_mutex);
2175 put_task_struct(event->owner);
2181 EXPORT_SYMBOL_GPL(perf_event_release_kernel);
2184 * Called when the last reference to the file is gone.
2186 static int perf_release(struct inode *inode, struct file *file)
2188 struct perf_event *event = file->private_data;
2190 file->private_data = NULL;
2192 return perf_event_release_kernel(event);
2195 static int perf_event_read_size(struct perf_event *event)
2197 int entry = sizeof(u64); /* value */
2201 if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
2202 size += sizeof(u64);
2204 if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
2205 size += sizeof(u64);
2207 if (event->attr.read_format & PERF_FORMAT_ID)
2208 entry += sizeof(u64);
2210 if (event->attr.read_format & PERF_FORMAT_GROUP) {
2211 nr += event->group_leader->nr_siblings;
2212 size += sizeof(u64);
2220 u64 perf_event_read_value(struct perf_event *event, u64 *enabled, u64 *running)
2222 struct perf_event *child;
2228 mutex_lock(&event->child_mutex);
2229 total += perf_event_read(event);
2230 *enabled += event->total_time_enabled +
2231 atomic64_read(&event->child_total_time_enabled);
2232 *running += event->total_time_running +
2233 atomic64_read(&event->child_total_time_running);
2235 list_for_each_entry(child, &event->child_list, child_list) {
2236 total += perf_event_read(child);
2237 *enabled += child->total_time_enabled;
2238 *running += child->total_time_running;
2240 mutex_unlock(&event->child_mutex);
2244 EXPORT_SYMBOL_GPL(perf_event_read_value);
2246 static int perf_event_read_group(struct perf_event *event,
2247 u64 read_format, char __user *buf)
2249 struct perf_event *leader = event->group_leader, *sub;
2250 int n = 0, size = 0, ret = -EFAULT;
2251 struct perf_event_context *ctx = leader->ctx;
2253 u64 count, enabled, running;
2255 mutex_lock(&ctx->mutex);
2256 count = perf_event_read_value(leader, &enabled, &running);
2258 values[n++] = 1 + leader->nr_siblings;
2259 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
2260 values[n++] = enabled;
2261 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
2262 values[n++] = running;
2263 values[n++] = count;
2264 if (read_format & PERF_FORMAT_ID)
2265 values[n++] = primary_event_id(leader);
2267 size = n * sizeof(u64);
2269 if (copy_to_user(buf, values, size))
2274 list_for_each_entry(sub, &leader->sibling_list, group_entry) {
2277 values[n++] = perf_event_read_value(sub, &enabled, &running);
2278 if (read_format & PERF_FORMAT_ID)
2279 values[n++] = primary_event_id(sub);
2281 size = n * sizeof(u64);
2283 if (copy_to_user(buf + ret, values, size)) {
2291 mutex_unlock(&ctx->mutex);
2296 static int perf_event_read_one(struct perf_event *event,
2297 u64 read_format, char __user *buf)
2299 u64 enabled, running;
2303 values[n++] = perf_event_read_value(event, &enabled, &running);
2304 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
2305 values[n++] = enabled;
2306 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
2307 values[n++] = running;
2308 if (read_format & PERF_FORMAT_ID)
2309 values[n++] = primary_event_id(event);
2311 if (copy_to_user(buf, values, n * sizeof(u64)))
2314 return n * sizeof(u64);
2318 * Read the performance event - simple non blocking version for now
2321 perf_read_hw(struct perf_event *event, char __user *buf, size_t count)
2323 u64 read_format = event->attr.read_format;
2327 * Return end-of-file for a read on a event that is in
2328 * error state (i.e. because it was pinned but it couldn't be
2329 * scheduled on to the CPU at some point).
2331 if (event->state == PERF_EVENT_STATE_ERROR)
2334 if (count < perf_event_read_size(event))
2337 WARN_ON_ONCE(event->ctx->parent_ctx);
2338 if (read_format & PERF_FORMAT_GROUP)
2339 ret = perf_event_read_group(event, read_format, buf);
2341 ret = perf_event_read_one(event, read_format, buf);
2347 perf_read(struct file *file, char __user *buf, size_t count, loff_t *ppos)
2349 struct perf_event *event = file->private_data;
2351 return perf_read_hw(event, buf, count);
2354 static unsigned int perf_poll(struct file *file, poll_table *wait)
2356 struct perf_event *event = file->private_data;
2357 struct perf_buffer *buffer;
2358 unsigned int events = POLL_HUP;
2361 buffer = rcu_dereference(event->buffer);
2363 events = atomic_xchg(&buffer->poll, 0);
2366 poll_wait(file, &event->waitq, wait);
2371 static void perf_event_reset(struct perf_event *event)
2373 (void)perf_event_read(event);
2374 local64_set(&event->count, 0);
2375 perf_event_update_userpage(event);
2379 * Holding the top-level event's child_mutex means that any
2380 * descendant process that has inherited this event will block
2381 * in sync_child_event if it goes to exit, thus satisfying the
2382 * task existence requirements of perf_event_enable/disable.
2384 static void perf_event_for_each_child(struct perf_event *event,
2385 void (*func)(struct perf_event *))
2387 struct perf_event *child;
2389 WARN_ON_ONCE(event->ctx->parent_ctx);
2390 mutex_lock(&event->child_mutex);
2392 list_for_each_entry(child, &event->child_list, child_list)
2394 mutex_unlock(&event->child_mutex);
2397 static void perf_event_for_each(struct perf_event *event,
2398 void (*func)(struct perf_event *))
2400 struct perf_event_context *ctx = event->ctx;
2401 struct perf_event *sibling;
2403 WARN_ON_ONCE(ctx->parent_ctx);
2404 mutex_lock(&ctx->mutex);
2405 event = event->group_leader;
2407 perf_event_for_each_child(event, func);
2409 list_for_each_entry(sibling, &event->sibling_list, group_entry)
2410 perf_event_for_each_child(event, func);
2411 mutex_unlock(&ctx->mutex);
2414 static int perf_event_period(struct perf_event *event, u64 __user *arg)
2416 struct perf_event_context *ctx = event->ctx;
2421 if (!event->attr.sample_period)
2424 size = copy_from_user(&value, arg, sizeof(value));
2425 if (size != sizeof(value))
2431 raw_spin_lock_irq(&ctx->lock);
2432 if (event->attr.freq) {
2433 if (value > sysctl_perf_event_sample_rate) {
2438 event->attr.sample_freq = value;
2440 event->attr.sample_period = value;
2441 event->hw.sample_period = value;
2444 raw_spin_unlock_irq(&ctx->lock);
2449 static const struct file_operations perf_fops;
2451 static struct perf_event *perf_fget_light(int fd, int *fput_needed)
2455 file = fget_light(fd, fput_needed);
2457 return ERR_PTR(-EBADF);
2459 if (file->f_op != &perf_fops) {
2460 fput_light(file, *fput_needed);
2462 return ERR_PTR(-EBADF);
2465 return file->private_data;
2468 static int perf_event_set_output(struct perf_event *event,
2469 struct perf_event *output_event);
2470 static int perf_event_set_filter(struct perf_event *event, void __user *arg);
2472 static long perf_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
2474 struct perf_event *event = file->private_data;
2475 void (*func)(struct perf_event *);
2479 case PERF_EVENT_IOC_ENABLE:
2480 func = perf_event_enable;
2482 case PERF_EVENT_IOC_DISABLE:
2483 func = perf_event_disable;
2485 case PERF_EVENT_IOC_RESET:
2486 func = perf_event_reset;
2489 case PERF_EVENT_IOC_REFRESH:
2490 return perf_event_refresh(event, arg);
2492 case PERF_EVENT_IOC_PERIOD:
2493 return perf_event_period(event, (u64 __user *)arg);
2495 case PERF_EVENT_IOC_SET_OUTPUT:
2497 struct perf_event *output_event = NULL;
2498 int fput_needed = 0;
2502 output_event = perf_fget_light(arg, &fput_needed);
2503 if (IS_ERR(output_event))
2504 return PTR_ERR(output_event);
2507 ret = perf_event_set_output(event, output_event);
2509 fput_light(output_event->filp, fput_needed);
2514 case PERF_EVENT_IOC_SET_FILTER:
2515 return perf_event_set_filter(event, (void __user *)arg);
2521 if (flags & PERF_IOC_FLAG_GROUP)
2522 perf_event_for_each(event, func);
2524 perf_event_for_each_child(event, func);
2529 int perf_event_task_enable(void)
2531 struct perf_event *event;
2533 mutex_lock(¤t->perf_event_mutex);
2534 list_for_each_entry(event, ¤t->perf_event_list, owner_entry)
2535 perf_event_for_each_child(event, perf_event_enable);
2536 mutex_unlock(¤t->perf_event_mutex);
2541 int perf_event_task_disable(void)
2543 struct perf_event *event;
2545 mutex_lock(¤t->perf_event_mutex);
2546 list_for_each_entry(event, ¤t->perf_event_list, owner_entry)
2547 perf_event_for_each_child(event, perf_event_disable);
2548 mutex_unlock(¤t->perf_event_mutex);
2553 #ifndef PERF_EVENT_INDEX_OFFSET
2554 # define PERF_EVENT_INDEX_OFFSET 0
2557 static int perf_event_index(struct perf_event *event)
2559 if (event->state != PERF_EVENT_STATE_ACTIVE)
2562 return event->hw.idx + 1 - PERF_EVENT_INDEX_OFFSET;
2566 * Callers need to ensure there can be no nesting of this function, otherwise
2567 * the seqlock logic goes bad. We can not serialize this because the arch
2568 * code calls this from NMI context.
2570 void perf_event_update_userpage(struct perf_event *event)
2572 struct perf_event_mmap_page *userpg;
2573 struct perf_buffer *buffer;
2576 buffer = rcu_dereference(event->buffer);
2580 userpg = buffer->user_page;
2583 * Disable preemption so as to not let the corresponding user-space
2584 * spin too long if we get preempted.
2589 userpg->index = perf_event_index(event);
2590 userpg->offset = perf_event_count(event);
2591 if (event->state == PERF_EVENT_STATE_ACTIVE)
2592 userpg->offset -= local64_read(&event->hw.prev_count);
2594 userpg->time_enabled = event->total_time_enabled +
2595 atomic64_read(&event->child_total_time_enabled);
2597 userpg->time_running = event->total_time_running +
2598 atomic64_read(&event->child_total_time_running);
2607 static unsigned long perf_data_size(struct perf_buffer *buffer);
2610 perf_buffer_init(struct perf_buffer *buffer, long watermark, int flags)
2612 long max_size = perf_data_size(buffer);
2615 buffer->watermark = min(max_size, watermark);
2617 if (!buffer->watermark)
2618 buffer->watermark = max_size / 2;
2620 if (flags & PERF_BUFFER_WRITABLE)
2621 buffer->writable = 1;
2623 atomic_set(&buffer->refcount, 1);
2626 #ifndef CONFIG_PERF_USE_VMALLOC
2629 * Back perf_mmap() with regular GFP_KERNEL-0 pages.
2632 static struct page *
2633 perf_mmap_to_page(struct perf_buffer *buffer, unsigned long pgoff)
2635 if (pgoff > buffer->nr_pages)
2639 return virt_to_page(buffer->user_page);
2641 return virt_to_page(buffer->data_pages[pgoff - 1]);
2644 static void *perf_mmap_alloc_page(int cpu)
2649 node = (cpu == -1) ? cpu : cpu_to_node(cpu);
2650 page = alloc_pages_node(node, GFP_KERNEL | __GFP_ZERO, 0);
2654 return page_address(page);
2657 static struct perf_buffer *
2658 perf_buffer_alloc(int nr_pages, long watermark, int cpu, int flags)
2660 struct perf_buffer *buffer;
2664 size = sizeof(struct perf_buffer);
2665 size += nr_pages * sizeof(void *);
2667 buffer = kzalloc(size, GFP_KERNEL);
2671 buffer->user_page = perf_mmap_alloc_page(cpu);
2672 if (!buffer->user_page)
2673 goto fail_user_page;
2675 for (i = 0; i < nr_pages; i++) {
2676 buffer->data_pages[i] = perf_mmap_alloc_page(cpu);
2677 if (!buffer->data_pages[i])
2678 goto fail_data_pages;
2681 buffer->nr_pages = nr_pages;
2683 perf_buffer_init(buffer, watermark, flags);
2688 for (i--; i >= 0; i--)
2689 free_page((unsigned long)buffer->data_pages[i]);
2691 free_page((unsigned long)buffer->user_page);
2700 static void perf_mmap_free_page(unsigned long addr)
2702 struct page *page = virt_to_page((void *)addr);
2704 page->mapping = NULL;
2708 static void perf_buffer_free(struct perf_buffer *buffer)
2712 perf_mmap_free_page((unsigned long)buffer->user_page);
2713 for (i = 0; i < buffer->nr_pages; i++)
2714 perf_mmap_free_page((unsigned long)buffer->data_pages[i]);
2718 static inline int page_order(struct perf_buffer *buffer)
2726 * Back perf_mmap() with vmalloc memory.
2728 * Required for architectures that have d-cache aliasing issues.
2731 static inline int page_order(struct perf_buffer *buffer)
2733 return buffer->page_order;
2736 static struct page *
2737 perf_mmap_to_page(struct perf_buffer *buffer, unsigned long pgoff)
2739 if (pgoff > (1UL << page_order(buffer)))
2742 return vmalloc_to_page((void *)buffer->user_page + pgoff * PAGE_SIZE);
2745 static void perf_mmap_unmark_page(void *addr)
2747 struct page *page = vmalloc_to_page(addr);
2749 page->mapping = NULL;
2752 static void perf_buffer_free_work(struct work_struct *work)
2754 struct perf_buffer *buffer;
2758 buffer = container_of(work, struct perf_buffer, work);
2759 nr = 1 << page_order(buffer);
2761 base = buffer->user_page;
2762 for (i = 0; i < nr + 1; i++)
2763 perf_mmap_unmark_page(base + (i * PAGE_SIZE));
2769 static void perf_buffer_free(struct perf_buffer *buffer)
2771 schedule_work(&buffer->work);
2774 static struct perf_buffer *
2775 perf_buffer_alloc(int nr_pages, long watermark, int cpu, int flags)
2777 struct perf_buffer *buffer;
2781 size = sizeof(struct perf_buffer);
2782 size += sizeof(void *);
2784 buffer = kzalloc(size, GFP_KERNEL);
2788 INIT_WORK(&buffer->work, perf_buffer_free_work);
2790 all_buf = vmalloc_user((nr_pages + 1) * PAGE_SIZE);
2794 buffer->user_page = all_buf;
2795 buffer->data_pages[0] = all_buf + PAGE_SIZE;
2796 buffer->page_order = ilog2(nr_pages);
2797 buffer->nr_pages = 1;
2799 perf_buffer_init(buffer, watermark, flags);
2812 static unsigned long perf_data_size(struct perf_buffer *buffer)
2814 return buffer->nr_pages << (PAGE_SHIFT + page_order(buffer));
2817 static int perf_mmap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
2819 struct perf_event *event = vma->vm_file->private_data;
2820 struct perf_buffer *buffer;
2821 int ret = VM_FAULT_SIGBUS;
2823 if (vmf->flags & FAULT_FLAG_MKWRITE) {
2824 if (vmf->pgoff == 0)
2830 buffer = rcu_dereference(event->buffer);
2834 if (vmf->pgoff && (vmf->flags & FAULT_FLAG_WRITE))
2837 vmf->page = perf_mmap_to_page(buffer, vmf->pgoff);
2841 get_page(vmf->page);
2842 vmf->page->mapping = vma->vm_file->f_mapping;
2843 vmf->page->index = vmf->pgoff;
2852 static void perf_buffer_free_rcu(struct rcu_head *rcu_head)
2854 struct perf_buffer *buffer;
2856 buffer = container_of(rcu_head, struct perf_buffer, rcu_head);
2857 perf_buffer_free(buffer);
2860 static struct perf_buffer *perf_buffer_get(struct perf_event *event)
2862 struct perf_buffer *buffer;
2865 buffer = rcu_dereference(event->buffer);
2867 if (!atomic_inc_not_zero(&buffer->refcount))
2875 static void perf_buffer_put(struct perf_buffer *buffer)
2877 if (!atomic_dec_and_test(&buffer->refcount))
2880 call_rcu(&buffer->rcu_head, perf_buffer_free_rcu);
2883 static void perf_mmap_open(struct vm_area_struct *vma)
2885 struct perf_event *event = vma->vm_file->private_data;
2887 atomic_inc(&event->mmap_count);
2890 static void perf_mmap_close(struct vm_area_struct *vma)
2892 struct perf_event *event = vma->vm_file->private_data;
2894 if (atomic_dec_and_mutex_lock(&event->mmap_count, &event->mmap_mutex)) {
2895 unsigned long size = perf_data_size(event->buffer);
2896 struct user_struct *user = event->mmap_user;
2897 struct perf_buffer *buffer = event->buffer;
2899 atomic_long_sub((size >> PAGE_SHIFT) + 1, &user->locked_vm);
2900 vma->vm_mm->locked_vm -= event->mmap_locked;
2901 rcu_assign_pointer(event->buffer, NULL);
2902 mutex_unlock(&event->mmap_mutex);
2904 perf_buffer_put(buffer);
2909 static const struct vm_operations_struct perf_mmap_vmops = {
2910 .open = perf_mmap_open,
2911 .close = perf_mmap_close,
2912 .fault = perf_mmap_fault,
2913 .page_mkwrite = perf_mmap_fault,
2916 static int perf_mmap(struct file *file, struct vm_area_struct *vma)
2918 struct perf_event *event = file->private_data;
2919 unsigned long user_locked, user_lock_limit;
2920 struct user_struct *user = current_user();
2921 unsigned long locked, lock_limit;
2922 struct perf_buffer *buffer;
2923 unsigned long vma_size;
2924 unsigned long nr_pages;
2925 long user_extra, extra;
2926 int ret = 0, flags = 0;
2929 * Don't allow mmap() of inherited per-task counters. This would
2930 * create a performance issue due to all children writing to the
2933 if (event->cpu == -1 && event->attr.inherit)
2936 if (!(vma->vm_flags & VM_SHARED))
2939 vma_size = vma->vm_end - vma->vm_start;
2940 nr_pages = (vma_size / PAGE_SIZE) - 1;
2943 * If we have buffer pages ensure they're a power-of-two number, so we
2944 * can do bitmasks instead of modulo.
2946 if (nr_pages != 0 && !is_power_of_2(nr_pages))
2949 if (vma_size != PAGE_SIZE * (1 + nr_pages))
2952 if (vma->vm_pgoff != 0)
2955 WARN_ON_ONCE(event->ctx->parent_ctx);
2956 mutex_lock(&event->mmap_mutex);
2957 if (event->buffer) {
2958 if (event->buffer->nr_pages == nr_pages)
2959 atomic_inc(&event->buffer->refcount);
2965 user_extra = nr_pages + 1;
2966 user_lock_limit = sysctl_perf_event_mlock >> (PAGE_SHIFT - 10);
2969 * Increase the limit linearly with more CPUs:
2971 user_lock_limit *= num_online_cpus();
2973 user_locked = atomic_long_read(&user->locked_vm) + user_extra;
2976 if (user_locked > user_lock_limit)
2977 extra = user_locked - user_lock_limit;
2979 lock_limit = rlimit(RLIMIT_MEMLOCK);
2980 lock_limit >>= PAGE_SHIFT;
2981 locked = vma->vm_mm->locked_vm + extra;
2983 if ((locked > lock_limit) && perf_paranoid_tracepoint_raw() &&
2984 !capable(CAP_IPC_LOCK)) {
2989 WARN_ON(event->buffer);
2991 if (vma->vm_flags & VM_WRITE)
2992 flags |= PERF_BUFFER_WRITABLE;
2994 buffer = perf_buffer_alloc(nr_pages, event->attr.wakeup_watermark,
3000 rcu_assign_pointer(event->buffer, buffer);
3002 atomic_long_add(user_extra, &user->locked_vm);
3003 event->mmap_locked = extra;
3004 event->mmap_user = get_current_user();
3005 vma->vm_mm->locked_vm += event->mmap_locked;
3009 atomic_inc(&event->mmap_count);
3010 mutex_unlock(&event->mmap_mutex);
3012 vma->vm_flags |= VM_RESERVED;
3013 vma->vm_ops = &perf_mmap_vmops;
3018 static int perf_fasync(int fd, struct file *filp, int on)
3020 struct inode *inode = filp->f_path.dentry->d_inode;
3021 struct perf_event *event = filp->private_data;
3024 mutex_lock(&inode->i_mutex);
3025 retval = fasync_helper(fd, filp, on, &event->fasync);
3026 mutex_unlock(&inode->i_mutex);
3034 static const struct file_operations perf_fops = {
3035 .llseek = no_llseek,
3036 .release = perf_release,
3039 .unlocked_ioctl = perf_ioctl,
3040 .compat_ioctl = perf_ioctl,
3042 .fasync = perf_fasync,
3048 * If there's data, ensure we set the poll() state and publish everything
3049 * to user-space before waking everybody up.
3052 void perf_event_wakeup(struct perf_event *event)
3054 wake_up_all(&event->waitq);
3056 if (event->pending_kill) {
3057 kill_fasync(&event->fasync, SIGIO, event->pending_kill);
3058 event->pending_kill = 0;
3065 * Handle the case where we need to wakeup up from NMI (or rq->lock) context.
3067 * The NMI bit means we cannot possibly take locks. Therefore, maintain a
3068 * single linked list and use cmpxchg() to add entries lockless.
3071 static void perf_pending_event(struct perf_pending_entry *entry)
3073 struct perf_event *event = container_of(entry,
3074 struct perf_event, pending);
3076 if (event->pending_disable) {
3077 event->pending_disable = 0;
3078 __perf_event_disable(event);
3081 if (event->pending_wakeup) {
3082 event->pending_wakeup = 0;
3083 perf_event_wakeup(event);
3087 #define PENDING_TAIL ((struct perf_pending_entry *)-1UL)
3089 static DEFINE_PER_CPU(struct perf_pending_entry *, perf_pending_head) = {
3093 static void perf_pending_queue(struct perf_pending_entry *entry,
3094 void (*func)(struct perf_pending_entry *))
3096 struct perf_pending_entry **head;
3098 if (cmpxchg(&entry->next, NULL, PENDING_TAIL) != NULL)
3103 head = &get_cpu_var(perf_pending_head);
3106 entry->next = *head;
3107 } while (cmpxchg(head, entry->next, entry) != entry->next);
3109 set_perf_event_pending();
3111 put_cpu_var(perf_pending_head);
3114 static int __perf_pending_run(void)
3116 struct perf_pending_entry *list;
3119 list = xchg(&__get_cpu_var(perf_pending_head), PENDING_TAIL);
3120 while (list != PENDING_TAIL) {
3121 void (*func)(struct perf_pending_entry *);
3122 struct perf_pending_entry *entry = list;
3129 * Ensure we observe the unqueue before we issue the wakeup,
3130 * so that we won't be waiting forever.
3131 * -- see perf_not_pending().
3142 static inline int perf_not_pending(struct perf_event *event)
3145 * If we flush on whatever cpu we run, there is a chance we don't
3149 __perf_pending_run();
3153 * Ensure we see the proper queue state before going to sleep
3154 * so that we do not miss the wakeup. -- see perf_pending_handle()
3157 return event->pending.next == NULL;
3160 static void perf_pending_sync(struct perf_event *event)
3162 wait_event(event->waitq, perf_not_pending(event));
3165 void perf_event_do_pending(void)
3167 __perf_pending_run();
3171 * We assume there is only KVM supporting the callbacks.
3172 * Later on, we might change it to a list if there is
3173 * another virtualization implementation supporting the callbacks.
3175 struct perf_guest_info_callbacks *perf_guest_cbs;
3177 int perf_register_guest_info_callbacks(struct perf_guest_info_callbacks *cbs)
3179 perf_guest_cbs = cbs;
3182 EXPORT_SYMBOL_GPL(perf_register_guest_info_callbacks);
3184 int perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks *cbs)
3186 perf_guest_cbs = NULL;
3189 EXPORT_SYMBOL_GPL(perf_unregister_guest_info_callbacks);
3194 static bool perf_output_space(struct perf_buffer *buffer, unsigned long tail,
3195 unsigned long offset, unsigned long head)
3199 if (!buffer->writable)
3202 mask = perf_data_size(buffer) - 1;
3204 offset = (offset - tail) & mask;
3205 head = (head - tail) & mask;
3207 if ((int)(head - offset) < 0)
3213 static void perf_output_wakeup(struct perf_output_handle *handle)
3215 atomic_set(&handle->buffer->poll, POLL_IN);
3218 handle->event->pending_wakeup = 1;
3219 perf_pending_queue(&handle->event->pending,
3220 perf_pending_event);
3222 perf_event_wakeup(handle->event);
3226 * We need to ensure a later event_id doesn't publish a head when a former
3227 * event isn't done writing. However since we need to deal with NMIs we
3228 * cannot fully serialize things.
3230 * We only publish the head (and generate a wakeup) when the outer-most
3233 static void perf_output_get_handle(struct perf_output_handle *handle)
3235 struct perf_buffer *buffer = handle->buffer;
3238 local_inc(&buffer->nest);
3239 handle->wakeup = local_read(&buffer->wakeup);
3242 static void perf_output_put_handle(struct perf_output_handle *handle)
3244 struct perf_buffer *buffer = handle->buffer;
3248 head = local_read(&buffer->head);
3251 * IRQ/NMI can happen here, which means we can miss a head update.
3254 if (!local_dec_and_test(&buffer->nest))
3258 * Publish the known good head. Rely on the full barrier implied
3259 * by atomic_dec_and_test() order the buffer->head read and this
3262 buffer->user_page->data_head = head;
3265 * Now check if we missed an update, rely on the (compiler)
3266 * barrier in atomic_dec_and_test() to re-read buffer->head.
3268 if (unlikely(head != local_read(&buffer->head))) {
3269 local_inc(&buffer->nest);
3273 if (handle->wakeup != local_read(&buffer->wakeup))
3274 perf_output_wakeup(handle);
3280 __always_inline void perf_output_copy(struct perf_output_handle *handle,
3281 const void *buf, unsigned int len)
3284 unsigned long size = min_t(unsigned long, handle->size, len);
3286 memcpy(handle->addr, buf, size);
3289 handle->addr += size;
3291 handle->size -= size;
3292 if (!handle->size) {
3293 struct perf_buffer *buffer = handle->buffer;
3296 handle->page &= buffer->nr_pages - 1;
3297 handle->addr = buffer->data_pages[handle->page];
3298 handle->size = PAGE_SIZE << page_order(buffer);
3303 int perf_output_begin(struct perf_output_handle *handle,
3304 struct perf_event *event, unsigned int size,
3305 int nmi, int sample)
3307 struct perf_buffer *buffer;
3308 unsigned long tail, offset, head;
3311 struct perf_event_header header;
3318 * For inherited events we send all the output towards the parent.
3321 event = event->parent;
3323 buffer = rcu_dereference(event->buffer);
3327 handle->buffer = buffer;
3328 handle->event = event;
3330 handle->sample = sample;
3332 if (!buffer->nr_pages)
3335 have_lost = local_read(&buffer->lost);
3337 size += sizeof(lost_event);
3339 perf_output_get_handle(handle);
3343 * Userspace could choose to issue a mb() before updating the
3344 * tail pointer. So that all reads will be completed before the
3347 tail = ACCESS_ONCE(buffer->user_page->data_tail);
3349 offset = head = local_read(&buffer->head);
3351 if (unlikely(!perf_output_space(buffer, tail, offset, head)))
3353 } while (local_cmpxchg(&buffer->head, offset, head) != offset);
3355 if (head - local_read(&buffer->wakeup) > buffer->watermark)
3356 local_add(buffer->watermark, &buffer->wakeup);
3358 handle->page = offset >> (PAGE_SHIFT + page_order(buffer));
3359 handle->page &= buffer->nr_pages - 1;
3360 handle->size = offset & ((PAGE_SIZE << page_order(buffer)) - 1);
3361 handle->addr = buffer->data_pages[handle->page];
3362 handle->addr += handle->size;
3363 handle->size = (PAGE_SIZE << page_order(buffer)) - handle->size;
3366 lost_event.header.type = PERF_RECORD_LOST;
3367 lost_event.header.misc = 0;
3368 lost_event.header.size = sizeof(lost_event);
3369 lost_event.id = event->id;
3370 lost_event.lost = local_xchg(&buffer->lost, 0);
3372 perf_output_put(handle, lost_event);
3378 local_inc(&buffer->lost);
3379 perf_output_put_handle(handle);
3386 void perf_output_end(struct perf_output_handle *handle)
3388 struct perf_event *event = handle->event;
3389 struct perf_buffer *buffer = handle->buffer;
3391 int wakeup_events = event->attr.wakeup_events;
3393 if (handle->sample && wakeup_events) {
3394 int events = local_inc_return(&buffer->events);
3395 if (events >= wakeup_events) {
3396 local_sub(wakeup_events, &buffer->events);
3397 local_inc(&buffer->wakeup);
3401 perf_output_put_handle(handle);
3405 static u32 perf_event_pid(struct perf_event *event, struct task_struct *p)
3408 * only top level events have the pid namespace they were created in
3411 event = event->parent;
3413 return task_tgid_nr_ns(p, event->ns);
3416 static u32 perf_event_tid(struct perf_event *event, struct task_struct *p)
3419 * only top level events have the pid namespace they were created in
3422 event = event->parent;
3424 return task_pid_nr_ns(p, event->ns);
3427 static void perf_output_read_one(struct perf_output_handle *handle,
3428 struct perf_event *event)
3430 u64 read_format = event->attr.read_format;
3434 values[n++] = perf_event_count(event);
3435 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
3436 values[n++] = event->total_time_enabled +
3437 atomic64_read(&event->child_total_time_enabled);
3439 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
3440 values[n++] = event->total_time_running +
3441 atomic64_read(&event->child_total_time_running);
3443 if (read_format & PERF_FORMAT_ID)
3444 values[n++] = primary_event_id(event);
3446 perf_output_copy(handle, values, n * sizeof(u64));
3450 * XXX PERF_FORMAT_GROUP vs inherited events seems difficult.
3452 static void perf_output_read_group(struct perf_output_handle *handle,
3453 struct perf_event *event)
3455 struct perf_event *leader = event->group_leader, *sub;
3456 u64 read_format = event->attr.read_format;
3460 values[n++] = 1 + leader->nr_siblings;
3462 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
3463 values[n++] = leader->total_time_enabled;
3465 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
3466 values[n++] = leader->total_time_running;
3468 if (leader != event)
3469 leader->pmu->read(leader);
3471 values[n++] = perf_event_count(leader);
3472 if (read_format & PERF_FORMAT_ID)
3473 values[n++] = primary_event_id(leader);
3475 perf_output_copy(handle, values, n * sizeof(u64));
3477 list_for_each_entry(sub, &leader->sibling_list, group_entry) {
3481 sub->pmu->read(sub);
3483 values[n++] = perf_event_count(sub);
3484 if (read_format & PERF_FORMAT_ID)
3485 values[n++] = primary_event_id(sub);
3487 perf_output_copy(handle, values, n * sizeof(u64));
3491 static void perf_output_read(struct perf_output_handle *handle,
3492 struct perf_event *event)
3494 if (event->attr.read_format & PERF_FORMAT_GROUP)
3495 perf_output_read_group(handle, event);
3497 perf_output_read_one(handle, event);
3500 void perf_output_sample(struct perf_output_handle *handle,
3501 struct perf_event_header *header,
3502 struct perf_sample_data *data,
3503 struct perf_event *event)
3505 u64 sample_type = data->type;
3507 perf_output_put(handle, *header);
3509 if (sample_type & PERF_SAMPLE_IP)
3510 perf_output_put(handle, data->ip);
3512 if (sample_type & PERF_SAMPLE_TID)
3513 perf_output_put(handle, data->tid_entry);
3515 if (sample_type & PERF_SAMPLE_TIME)
3516 perf_output_put(handle, data->time);
3518 if (sample_type & PERF_SAMPLE_ADDR)
3519 perf_output_put(handle, data->addr);
3521 if (sample_type & PERF_SAMPLE_ID)
3522 perf_output_put(handle, data->id);
3524 if (sample_type & PERF_SAMPLE_STREAM_ID)
3525 perf_output_put(handle, data->stream_id);
3527 if (sample_type & PERF_SAMPLE_CPU)
3528 perf_output_put(handle, data->cpu_entry);
3530 if (sample_type & PERF_SAMPLE_PERIOD)
3531 perf_output_put(handle, data->period);
3533 if (sample_type & PERF_SAMPLE_READ)
3534 perf_output_read(handle, event);
3536 if (sample_type & PERF_SAMPLE_CALLCHAIN) {
3537 if (data->callchain) {
3540 if (data->callchain)
3541 size += data->callchain->nr;
3543 size *= sizeof(u64);
3545 perf_output_copy(handle, data->callchain, size);
3548 perf_output_put(handle, nr);
3552 if (sample_type & PERF_SAMPLE_RAW) {
3554 perf_output_put(handle, data->raw->size);
3555 perf_output_copy(handle, data->raw->data,
3562 .size = sizeof(u32),
3565 perf_output_put(handle, raw);
3570 void perf_prepare_sample(struct perf_event_header *header,
3571 struct perf_sample_data *data,
3572 struct perf_event *event,
3573 struct pt_regs *regs)
3575 u64 sample_type = event->attr.sample_type;
3577 data->type = sample_type;
3579 header->type = PERF_RECORD_SAMPLE;
3580 header->size = sizeof(*header);
3583 header->misc |= perf_misc_flags(regs);
3585 if (sample_type & PERF_SAMPLE_IP) {
3586 data->ip = perf_instruction_pointer(regs);
3588 header->size += sizeof(data->ip);
3591 if (sample_type & PERF_SAMPLE_TID) {
3592 /* namespace issues */
3593 data->tid_entry.pid = perf_event_pid(event, current);
3594 data->tid_entry.tid = perf_event_tid(event, current);
3596 header->size += sizeof(data->tid_entry);
3599 if (sample_type & PERF_SAMPLE_TIME) {
3600 data->time = perf_clock();
3602 header->size += sizeof(data->time);
3605 if (sample_type & PERF_SAMPLE_ADDR)
3606 header->size += sizeof(data->addr);
3608 if (sample_type & PERF_SAMPLE_ID) {
3609 data->id = primary_event_id(event);
3611 header->size += sizeof(data->id);
3614 if (sample_type & PERF_SAMPLE_STREAM_ID) {
3615 data->stream_id = event->id;
3617 header->size += sizeof(data->stream_id);
3620 if (sample_type & PERF_SAMPLE_CPU) {
3621 data->cpu_entry.cpu = raw_smp_processor_id();
3622 data->cpu_entry.reserved = 0;
3624 header->size += sizeof(data->cpu_entry);
3627 if (sample_type & PERF_SAMPLE_PERIOD)
3628 header->size += sizeof(data->period);
3630 if (sample_type & PERF_SAMPLE_READ)
3631 header->size += perf_event_read_size(event);
3633 if (sample_type & PERF_SAMPLE_CALLCHAIN) {
3636 data->callchain = perf_callchain(regs);
3638 if (data->callchain)
3639 size += data->callchain->nr;
3641 header->size += size * sizeof(u64);
3644 if (sample_type & PERF_SAMPLE_RAW) {
3645 int size = sizeof(u32);
3648 size += data->raw->size;
3650 size += sizeof(u32);
3652 WARN_ON_ONCE(size & (sizeof(u64)-1));
3653 header->size += size;
3657 static void perf_event_output(struct perf_event *event, int nmi,
3658 struct perf_sample_data *data,
3659 struct pt_regs *regs)
3661 struct perf_output_handle handle;
3662 struct perf_event_header header;
3664 /* protect the callchain buffers */
3667 perf_prepare_sample(&header, data, event, regs);
3669 if (perf_output_begin(&handle, event, header.size, nmi, 1))
3672 perf_output_sample(&handle, &header, data, event);
3674 perf_output_end(&handle);
3684 struct perf_read_event {
3685 struct perf_event_header header;
3692 perf_event_read_event(struct perf_event *event,
3693 struct task_struct *task)
3695 struct perf_output_handle handle;
3696 struct perf_read_event read_event = {
3698 .type = PERF_RECORD_READ,
3700 .size = sizeof(read_event) + perf_event_read_size(event),
3702 .pid = perf_event_pid(event, task),
3703 .tid = perf_event_tid(event, task),
3707 ret = perf_output_begin(&handle, event, read_event.header.size, 0, 0);
3711 perf_output_put(&handle, read_event);
3712 perf_output_read(&handle, event);
3714 perf_output_end(&handle);
3718 * task tracking -- fork/exit
3720 * enabled by: attr.comm | attr.mmap | attr.mmap_data | attr.task
3723 struct perf_task_event {
3724 struct task_struct *task;
3725 struct perf_event_context *task_ctx;
3728 struct perf_event_header header;
3738 static void perf_event_task_output(struct perf_event *event,
3739 struct perf_task_event *task_event)
3741 struct perf_output_handle handle;
3742 struct task_struct *task = task_event->task;
3745 size = task_event->event_id.header.size;
3746 ret = perf_output_begin(&handle, event, size, 0, 0);
3751 task_event->event_id.pid = perf_event_pid(event, task);
3752 task_event->event_id.ppid = perf_event_pid(event, current);
3754 task_event->event_id.tid = perf_event_tid(event, task);
3755 task_event->event_id.ptid = perf_event_tid(event, current);
3757 perf_output_put(&handle, task_event->event_id);
3759 perf_output_end(&handle);
3762 static int perf_event_task_match(struct perf_event *event)
3764 if (event->state < PERF_EVENT_STATE_INACTIVE)
3767 if (event->cpu != -1 && event->cpu != smp_processor_id())
3770 if (event->attr.comm || event->attr.mmap ||
3771 event->attr.mmap_data || event->attr.task)
3777 static void perf_event_task_ctx(struct perf_event_context *ctx,
3778 struct perf_task_event *task_event)
3780 struct perf_event *event;
3782 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
3783 if (perf_event_task_match(event))
3784 perf_event_task_output(event, task_event);
3788 static void perf_event_task_event(struct perf_task_event *task_event)
3790 struct perf_cpu_context *cpuctx;
3791 struct perf_event_context *ctx = task_event->task_ctx;
3794 cpuctx = &get_cpu_var(perf_cpu_context);
3795 perf_event_task_ctx(&cpuctx->ctx, task_event);
3797 ctx = rcu_dereference(current->perf_event_ctxp);
3799 perf_event_task_ctx(ctx, task_event);
3800 put_cpu_var(perf_cpu_context);
3804 static void perf_event_task(struct task_struct *task,
3805 struct perf_event_context *task_ctx,
3808 struct perf_task_event task_event;
3810 if (!atomic_read(&nr_comm_events) &&
3811 !atomic_read(&nr_mmap_events) &&
3812 !atomic_read(&nr_task_events))
3815 task_event = (struct perf_task_event){
3817 .task_ctx = task_ctx,
3820 .type = new ? PERF_RECORD_FORK : PERF_RECORD_EXIT,
3822 .size = sizeof(task_event.event_id),
3828 .time = perf_clock(),
3832 perf_event_task_event(&task_event);
3835 void perf_event_fork(struct task_struct *task)
3837 perf_event_task(task, NULL, 1);
3844 struct perf_comm_event {
3845 struct task_struct *task;
3850 struct perf_event_header header;
3857 static void perf_event_comm_output(struct perf_event *event,
3858 struct perf_comm_event *comm_event)
3860 struct perf_output_handle handle;
3861 int size = comm_event->event_id.header.size;
3862 int ret = perf_output_begin(&handle, event, size, 0, 0);
3867 comm_event->event_id.pid = perf_event_pid(event, comm_event->task);
3868 comm_event->event_id.tid = perf_event_tid(event, comm_event->task);
3870 perf_output_put(&handle, comm_event->event_id);
3871 perf_output_copy(&handle, comm_event->comm,
3872 comm_event->comm_size);
3873 perf_output_end(&handle);
3876 static int perf_event_comm_match(struct perf_event *event)
3878 if (event->state < PERF_EVENT_STATE_INACTIVE)
3881 if (event->cpu != -1 && event->cpu != smp_processor_id())
3884 if (event->attr.comm)
3890 static void perf_event_comm_ctx(struct perf_event_context *ctx,
3891 struct perf_comm_event *comm_event)
3893 struct perf_event *event;
3895 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
3896 if (perf_event_comm_match(event))
3897 perf_event_comm_output(event, comm_event);
3901 static void perf_event_comm_event(struct perf_comm_event *comm_event)
3903 struct perf_cpu_context *cpuctx;
3904 struct perf_event_context *ctx;
3906 char comm[TASK_COMM_LEN];
3908 memset(comm, 0, sizeof(comm));
3909 strlcpy(comm, comm_event->task->comm, sizeof(comm));
3910 size = ALIGN(strlen(comm)+1, sizeof(u64));
3912 comm_event->comm = comm;
3913 comm_event->comm_size = size;
3915 comm_event->event_id.header.size = sizeof(comm_event->event_id) + size;
3918 cpuctx = &get_cpu_var(perf_cpu_context);
3919 perf_event_comm_ctx(&cpuctx->ctx, comm_event);
3920 ctx = rcu_dereference(current->perf_event_ctxp);
3922 perf_event_comm_ctx(ctx, comm_event);
3923 put_cpu_var(perf_cpu_context);
3927 void perf_event_comm(struct task_struct *task)
3929 struct perf_comm_event comm_event;
3931 if (task->perf_event_ctxp)
3932 perf_event_enable_on_exec(task);
3934 if (!atomic_read(&nr_comm_events))
3937 comm_event = (struct perf_comm_event){
3943 .type = PERF_RECORD_COMM,
3952 perf_event_comm_event(&comm_event);
3959 struct perf_mmap_event {
3960 struct vm_area_struct *vma;
3962 const char *file_name;
3966 struct perf_event_header header;
3976 static void perf_event_mmap_output(struct perf_event *event,
3977 struct perf_mmap_event *mmap_event)
3979 struct perf_output_handle handle;
3980 int size = mmap_event->event_id.header.size;
3981 int ret = perf_output_begin(&handle, event, size, 0, 0);
3986 mmap_event->event_id.pid = perf_event_pid(event, current);
3987 mmap_event->event_id.tid = perf_event_tid(event, current);
3989 perf_output_put(&handle, mmap_event->event_id);
3990 perf_output_copy(&handle, mmap_event->file_name,
3991 mmap_event->file_size);
3992 perf_output_end(&handle);
3995 static int perf_event_mmap_match(struct perf_event *event,
3996 struct perf_mmap_event *mmap_event,
3999 if (event->state < PERF_EVENT_STATE_INACTIVE)
4002 if (event->cpu != -1 && event->cpu != smp_processor_id())
4005 if ((!executable && event->attr.mmap_data) ||
4006 (executable && event->attr.mmap))
4012 static void perf_event_mmap_ctx(struct perf_event_context *ctx,
4013 struct perf_mmap_event *mmap_event,
4016 struct perf_event *event;
4018 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
4019 if (perf_event_mmap_match(event, mmap_event, executable))
4020 perf_event_mmap_output(event, mmap_event);
4024 static void perf_event_mmap_event(struct perf_mmap_event *mmap_event)
4026 struct perf_cpu_context *cpuctx;
4027 struct perf_event_context *ctx;
4028 struct vm_area_struct *vma = mmap_event->vma;
4029 struct file *file = vma->vm_file;
4035 memset(tmp, 0, sizeof(tmp));
4039 * d_path works from the end of the buffer backwards, so we
4040 * need to add enough zero bytes after the string to handle
4041 * the 64bit alignment we do later.
4043 buf = kzalloc(PATH_MAX + sizeof(u64), GFP_KERNEL);
4045 name = strncpy(tmp, "//enomem", sizeof(tmp));
4048 name = d_path(&file->f_path, buf, PATH_MAX);
4050 name = strncpy(tmp, "//toolong", sizeof(tmp));
4054 if (arch_vma_name(mmap_event->vma)) {
4055 name = strncpy(tmp, arch_vma_name(mmap_event->vma),
4061 name = strncpy(tmp, "[vdso]", sizeof(tmp));
4063 } else if (vma->vm_start <= vma->vm_mm->start_brk &&
4064 vma->vm_end >= vma->vm_mm->brk) {
4065 name = strncpy(tmp, "[heap]", sizeof(tmp));
4067 } else if (vma->vm_start <= vma->vm_mm->start_stack &&
4068 vma->vm_end >= vma->vm_mm->start_stack) {
4069 name = strncpy(tmp, "[stack]", sizeof(tmp));
4073 name = strncpy(tmp, "//anon", sizeof(tmp));
4078 size = ALIGN(strlen(name)+1, sizeof(u64));
4080 mmap_event->file_name = name;
4081 mmap_event->file_size = size;
4083 mmap_event->event_id.header.size = sizeof(mmap_event->event_id) + size;
4086 cpuctx = &get_cpu_var(perf_cpu_context);
4087 perf_event_mmap_ctx(&cpuctx->ctx, mmap_event, vma->vm_flags & VM_EXEC);
4088 ctx = rcu_dereference(current->perf_event_ctxp);
4090 perf_event_mmap_ctx(ctx, mmap_event, vma->vm_flags & VM_EXEC);
4091 put_cpu_var(perf_cpu_context);
4097 void perf_event_mmap(struct vm_area_struct *vma)
4099 struct perf_mmap_event mmap_event;
4101 if (!atomic_read(&nr_mmap_events))
4104 mmap_event = (struct perf_mmap_event){
4110 .type = PERF_RECORD_MMAP,
4111 .misc = PERF_RECORD_MISC_USER,
4116 .start = vma->vm_start,
4117 .len = vma->vm_end - vma->vm_start,
4118 .pgoff = (u64)vma->vm_pgoff << PAGE_SHIFT,
4122 perf_event_mmap_event(&mmap_event);
4126 * IRQ throttle logging
4129 static void perf_log_throttle(struct perf_event *event, int enable)
4131 struct perf_output_handle handle;
4135 struct perf_event_header header;
4139 } throttle_event = {
4141 .type = PERF_RECORD_THROTTLE,
4143 .size = sizeof(throttle_event),
4145 .time = perf_clock(),
4146 .id = primary_event_id(event),
4147 .stream_id = event->id,
4151 throttle_event.header.type = PERF_RECORD_UNTHROTTLE;
4153 ret = perf_output_begin(&handle, event, sizeof(throttle_event), 1, 0);
4157 perf_output_put(&handle, throttle_event);
4158 perf_output_end(&handle);
4162 * Generic event overflow handling, sampling.
4165 static int __perf_event_overflow(struct perf_event *event, int nmi,
4166 int throttle, struct perf_sample_data *data,
4167 struct pt_regs *regs)
4169 int events = atomic_read(&event->event_limit);
4170 struct hw_perf_event *hwc = &event->hw;
4173 throttle = (throttle && event->pmu->unthrottle != NULL);
4178 if (hwc->interrupts != MAX_INTERRUPTS) {
4180 if (HZ * hwc->interrupts >
4181 (u64)sysctl_perf_event_sample_rate) {
4182 hwc->interrupts = MAX_INTERRUPTS;
4183 perf_log_throttle(event, 0);
4188 * Keep re-disabling events even though on the previous
4189 * pass we disabled it - just in case we raced with a
4190 * sched-in and the event got enabled again:
4196 if (event->attr.freq) {
4197 u64 now = perf_clock();
4198 s64 delta = now - hwc->freq_time_stamp;
4200 hwc->freq_time_stamp = now;
4202 if (delta > 0 && delta < 2*TICK_NSEC)
4203 perf_adjust_period(event, delta, hwc->last_period);
4207 * XXX event_limit might not quite work as expected on inherited
4211 event->pending_kill = POLL_IN;
4212 if (events && atomic_dec_and_test(&event->event_limit)) {
4214 event->pending_kill = POLL_HUP;
4216 event->pending_disable = 1;
4217 perf_pending_queue(&event->pending,
4218 perf_pending_event);
4220 perf_event_disable(event);
4223 if (event->overflow_handler)
4224 event->overflow_handler(event, nmi, data, regs);
4226 perf_event_output(event, nmi, data, regs);
4231 int perf_event_overflow(struct perf_event *event, int nmi,
4232 struct perf_sample_data *data,
4233 struct pt_regs *regs)
4235 return __perf_event_overflow(event, nmi, 1, data, regs);
4239 * Generic software event infrastructure
4243 * We directly increment event->count and keep a second value in
4244 * event->hw.period_left to count intervals. This period event
4245 * is kept in the range [-sample_period, 0] so that we can use the
4249 static u64 perf_swevent_set_period(struct perf_event *event)
4251 struct hw_perf_event *hwc = &event->hw;
4252 u64 period = hwc->last_period;
4256 hwc->last_period = hwc->sample_period;
4259 old = val = local64_read(&hwc->period_left);
4263 nr = div64_u64(period + val, period);
4264 offset = nr * period;
4266 if (local64_cmpxchg(&hwc->period_left, old, val) != old)
4272 static void perf_swevent_overflow(struct perf_event *event, u64 overflow,
4273 int nmi, struct perf_sample_data *data,
4274 struct pt_regs *regs)
4276 struct hw_perf_event *hwc = &event->hw;
4279 data->period = event->hw.last_period;
4281 overflow = perf_swevent_set_period(event);
4283 if (hwc->interrupts == MAX_INTERRUPTS)
4286 for (; overflow; overflow--) {
4287 if (__perf_event_overflow(event, nmi, throttle,
4290 * We inhibit the overflow from happening when
4291 * hwc->interrupts == MAX_INTERRUPTS.
4299 static void perf_swevent_add(struct perf_event *event, u64 nr,
4300 int nmi, struct perf_sample_data *data,
4301 struct pt_regs *regs)
4303 struct hw_perf_event *hwc = &event->hw;
4305 local64_add(nr, &event->count);
4310 if (!hwc->sample_period)
4313 if (nr == 1 && hwc->sample_period == 1 && !event->attr.freq)
4314 return perf_swevent_overflow(event, 1, nmi, data, regs);
4316 if (local64_add_negative(nr, &hwc->period_left))
4319 perf_swevent_overflow(event, 0, nmi, data, regs);
4322 static int perf_exclude_event(struct perf_event *event,
4323 struct pt_regs *regs)
4326 if (event->attr.exclude_user && user_mode(regs))
4329 if (event->attr.exclude_kernel && !user_mode(regs))
4336 static int perf_swevent_match(struct perf_event *event,
4337 enum perf_type_id type,
4339 struct perf_sample_data *data,
4340 struct pt_regs *regs)
4342 if (event->attr.type != type)
4345 if (event->attr.config != event_id)
4348 if (perf_exclude_event(event, regs))
4354 static inline u64 swevent_hash(u64 type, u32 event_id)
4356 u64 val = event_id | (type << 32);
4358 return hash_64(val, SWEVENT_HLIST_BITS);
4361 static inline struct hlist_head *
4362 __find_swevent_head(struct swevent_hlist *hlist, u64 type, u32 event_id)
4364 u64 hash = swevent_hash(type, event_id);
4366 return &hlist->heads[hash];
4369 /* For the read side: events when they trigger */
4370 static inline struct hlist_head *
4371 find_swevent_head_rcu(struct perf_cpu_context *ctx, u64 type, u32 event_id)
4373 struct swevent_hlist *hlist;
4375 hlist = rcu_dereference(ctx->swevent_hlist);
4379 return __find_swevent_head(hlist, type, event_id);
4382 /* For the event head insertion and removal in the hlist */
4383 static inline struct hlist_head *
4384 find_swevent_head(struct perf_cpu_context *ctx, struct perf_event *event)
4386 struct swevent_hlist *hlist;
4387 u32 event_id = event->attr.config;
4388 u64 type = event->attr.type;
4391 * Event scheduling is always serialized against hlist allocation
4392 * and release. Which makes the protected version suitable here.
4393 * The context lock guarantees that.
4395 hlist = rcu_dereference_protected(ctx->swevent_hlist,
4396 lockdep_is_held(&event->ctx->lock));
4400 return __find_swevent_head(hlist, type, event_id);
4403 static void do_perf_sw_event(enum perf_type_id type, u32 event_id,
4405 struct perf_sample_data *data,
4406 struct pt_regs *regs)
4408 struct perf_cpu_context *cpuctx;
4409 struct perf_event *event;
4410 struct hlist_node *node;
4411 struct hlist_head *head;
4413 cpuctx = &__get_cpu_var(perf_cpu_context);
4417 head = find_swevent_head_rcu(cpuctx, type, event_id);
4422 hlist_for_each_entry_rcu(event, node, head, hlist_entry) {
4423 if (perf_swevent_match(event, type, event_id, data, regs))
4424 perf_swevent_add(event, nr, nmi, data, regs);
4430 int perf_swevent_get_recursion_context(void)
4432 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
4434 return get_recursion_context(cpuctx->recursion);
4436 EXPORT_SYMBOL_GPL(perf_swevent_get_recursion_context);
4438 void inline perf_swevent_put_recursion_context(int rctx)
4440 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
4442 put_recursion_context(cpuctx->recursion, rctx);
4445 void __perf_sw_event(u32 event_id, u64 nr, int nmi,
4446 struct pt_regs *regs, u64 addr)
4448 struct perf_sample_data data;
4451 preempt_disable_notrace();
4452 rctx = perf_swevent_get_recursion_context();
4456 perf_sample_data_init(&data, addr);
4458 do_perf_sw_event(PERF_TYPE_SOFTWARE, event_id, nr, nmi, &data, regs);
4460 perf_swevent_put_recursion_context(rctx);
4461 preempt_enable_notrace();
4464 static void perf_swevent_read(struct perf_event *event)
4468 static int perf_swevent_enable(struct perf_event *event)
4470 struct hw_perf_event *hwc = &event->hw;
4471 struct perf_cpu_context *cpuctx;
4472 struct hlist_head *head;
4474 cpuctx = &__get_cpu_var(perf_cpu_context);
4476 if (hwc->sample_period) {
4477 hwc->last_period = hwc->sample_period;
4478 perf_swevent_set_period(event);
4481 head = find_swevent_head(cpuctx, event);
4482 if (WARN_ON_ONCE(!head))
4485 hlist_add_head_rcu(&event->hlist_entry, head);
4490 static void perf_swevent_disable(struct perf_event *event)
4492 hlist_del_rcu(&event->hlist_entry);
4495 static void perf_swevent_void(struct perf_event *event)
4499 static int perf_swevent_int(struct perf_event *event)
4504 static const struct pmu perf_ops_generic = {
4505 .enable = perf_swevent_enable,
4506 .disable = perf_swevent_disable,
4507 .start = perf_swevent_int,
4508 .stop = perf_swevent_void,
4509 .read = perf_swevent_read,
4510 .unthrottle = perf_swevent_void, /* hwc->interrupts already reset */
4514 * hrtimer based swevent callback
4517 static enum hrtimer_restart perf_swevent_hrtimer(struct hrtimer *hrtimer)
4519 enum hrtimer_restart ret = HRTIMER_RESTART;
4520 struct perf_sample_data data;
4521 struct pt_regs *regs;
4522 struct perf_event *event;
4525 event = container_of(hrtimer, struct perf_event, hw.hrtimer);
4526 event->pmu->read(event);
4528 perf_sample_data_init(&data, 0);
4529 data.period = event->hw.last_period;
4530 regs = get_irq_regs();
4532 if (regs && !perf_exclude_event(event, regs)) {
4533 if (!(event->attr.exclude_idle && current->pid == 0))
4534 if (perf_event_overflow(event, 0, &data, regs))
4535 ret = HRTIMER_NORESTART;
4538 period = max_t(u64, 10000, event->hw.sample_period);
4539 hrtimer_forward_now(hrtimer, ns_to_ktime(period));
4544 static void perf_swevent_start_hrtimer(struct perf_event *event)
4546 struct hw_perf_event *hwc = &event->hw;
4548 hrtimer_init(&hwc->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
4549 hwc->hrtimer.function = perf_swevent_hrtimer;
4550 if (hwc->sample_period) {
4553 if (hwc->remaining) {
4554 if (hwc->remaining < 0)
4557 period = hwc->remaining;
4560 period = max_t(u64, 10000, hwc->sample_period);
4562 __hrtimer_start_range_ns(&hwc->hrtimer,
4563 ns_to_ktime(period), 0,
4564 HRTIMER_MODE_REL, 0);
4568 static void perf_swevent_cancel_hrtimer(struct perf_event *event)
4570 struct hw_perf_event *hwc = &event->hw;
4572 if (hwc->sample_period) {
4573 ktime_t remaining = hrtimer_get_remaining(&hwc->hrtimer);
4574 hwc->remaining = ktime_to_ns(remaining);
4576 hrtimer_cancel(&hwc->hrtimer);
4581 * Software event: cpu wall time clock
4584 static void cpu_clock_perf_event_update(struct perf_event *event)
4586 int cpu = raw_smp_processor_id();
4590 now = cpu_clock(cpu);
4591 prev = local64_xchg(&event->hw.prev_count, now);
4592 local64_add(now - prev, &event->count);
4595 static int cpu_clock_perf_event_enable(struct perf_event *event)
4597 struct hw_perf_event *hwc = &event->hw;
4598 int cpu = raw_smp_processor_id();
4600 local64_set(&hwc->prev_count, cpu_clock(cpu));
4601 perf_swevent_start_hrtimer(event);
4606 static void cpu_clock_perf_event_disable(struct perf_event *event)
4608 perf_swevent_cancel_hrtimer(event);
4609 cpu_clock_perf_event_update(event);
4612 static void cpu_clock_perf_event_read(struct perf_event *event)
4614 cpu_clock_perf_event_update(event);
4617 static const struct pmu perf_ops_cpu_clock = {
4618 .enable = cpu_clock_perf_event_enable,
4619 .disable = cpu_clock_perf_event_disable,
4620 .read = cpu_clock_perf_event_read,
4624 * Software event: task time clock
4627 static void task_clock_perf_event_update(struct perf_event *event, u64 now)
4632 prev = local64_xchg(&event->hw.prev_count, now);
4634 local64_add(delta, &event->count);
4637 static int task_clock_perf_event_enable(struct perf_event *event)
4639 struct hw_perf_event *hwc = &event->hw;
4642 now = event->ctx->time;
4644 local64_set(&hwc->prev_count, now);
4646 perf_swevent_start_hrtimer(event);
4651 static void task_clock_perf_event_disable(struct perf_event *event)
4653 perf_swevent_cancel_hrtimer(event);
4654 task_clock_perf_event_update(event, event->ctx->time);
4658 static void task_clock_perf_event_read(struct perf_event *event)
4663 update_context_time(event->ctx);
4664 time = event->ctx->time;
4666 u64 now = perf_clock();
4667 u64 delta = now - event->ctx->timestamp;
4668 time = event->ctx->time + delta;
4671 task_clock_perf_event_update(event, time);
4674 static const struct pmu perf_ops_task_clock = {
4675 .enable = task_clock_perf_event_enable,
4676 .disable = task_clock_perf_event_disable,
4677 .read = task_clock_perf_event_read,
4680 /* Deref the hlist from the update side */
4681 static inline struct swevent_hlist *
4682 swevent_hlist_deref(struct perf_cpu_context *cpuctx)
4684 return rcu_dereference_protected(cpuctx->swevent_hlist,
4685 lockdep_is_held(&cpuctx->hlist_mutex));
4688 static void swevent_hlist_release_rcu(struct rcu_head *rcu_head)
4690 struct swevent_hlist *hlist;
4692 hlist = container_of(rcu_head, struct swevent_hlist, rcu_head);
4696 static void swevent_hlist_release(struct perf_cpu_context *cpuctx)
4698 struct swevent_hlist *hlist = swevent_hlist_deref(cpuctx);
4703 rcu_assign_pointer(cpuctx->swevent_hlist, NULL);
4704 call_rcu(&hlist->rcu_head, swevent_hlist_release_rcu);
4707 static void swevent_hlist_put_cpu(struct perf_event *event, int cpu)
4709 struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
4711 mutex_lock(&cpuctx->hlist_mutex);
4713 if (!--cpuctx->hlist_refcount)
4714 swevent_hlist_release(cpuctx);
4716 mutex_unlock(&cpuctx->hlist_mutex);
4719 static void swevent_hlist_put(struct perf_event *event)
4723 if (event->cpu != -1) {
4724 swevent_hlist_put_cpu(event, event->cpu);
4728 for_each_possible_cpu(cpu)
4729 swevent_hlist_put_cpu(event, cpu);
4732 static int swevent_hlist_get_cpu(struct perf_event *event, int cpu)
4734 struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
4737 mutex_lock(&cpuctx->hlist_mutex);
4739 if (!swevent_hlist_deref(cpuctx) && cpu_online(cpu)) {
4740 struct swevent_hlist *hlist;
4742 hlist = kzalloc(sizeof(*hlist), GFP_KERNEL);
4747 rcu_assign_pointer(cpuctx->swevent_hlist, hlist);
4749 cpuctx->hlist_refcount++;
4751 mutex_unlock(&cpuctx->hlist_mutex);
4756 static int swevent_hlist_get(struct perf_event *event)
4759 int cpu, failed_cpu;
4761 if (event->cpu != -1)
4762 return swevent_hlist_get_cpu(event, event->cpu);
4765 for_each_possible_cpu(cpu) {
4766 err = swevent_hlist_get_cpu(event, cpu);
4776 for_each_possible_cpu(cpu) {
4777 if (cpu == failed_cpu)
4779 swevent_hlist_put_cpu(event, cpu);
4786 #ifdef CONFIG_EVENT_TRACING
4788 static const struct pmu perf_ops_tracepoint = {
4789 .enable = perf_trace_enable,
4790 .disable = perf_trace_disable,
4791 .start = perf_swevent_int,
4792 .stop = perf_swevent_void,
4793 .read = perf_swevent_read,
4794 .unthrottle = perf_swevent_void,
4797 static int perf_tp_filter_match(struct perf_event *event,
4798 struct perf_sample_data *data)
4800 void *record = data->raw->data;
4802 if (likely(!event->filter) || filter_match_preds(event->filter, record))
4807 static int perf_tp_event_match(struct perf_event *event,
4808 struct perf_sample_data *data,
4809 struct pt_regs *regs)
4812 * All tracepoints are from kernel-space.
4814 if (event->attr.exclude_kernel)
4817 if (!perf_tp_filter_match(event, data))
4823 void perf_tp_event(u64 addr, u64 count, void *record, int entry_size,
4824 struct pt_regs *regs, struct hlist_head *head, int rctx)
4826 struct perf_sample_data data;
4827 struct perf_event *event;
4828 struct hlist_node *node;
4830 struct perf_raw_record raw = {
4835 perf_sample_data_init(&data, addr);
4838 hlist_for_each_entry_rcu(event, node, head, hlist_entry) {
4839 if (perf_tp_event_match(event, &data, regs))
4840 perf_swevent_add(event, count, 1, &data, regs);
4843 perf_swevent_put_recursion_context(rctx);
4845 EXPORT_SYMBOL_GPL(perf_tp_event);
4847 static void tp_perf_event_destroy(struct perf_event *event)
4849 perf_trace_destroy(event);
4852 static const struct pmu *tp_perf_event_init(struct perf_event *event)
4857 * Raw tracepoint data is a severe data leak, only allow root to
4860 if ((event->attr.sample_type & PERF_SAMPLE_RAW) &&
4861 perf_paranoid_tracepoint_raw() &&
4862 !capable(CAP_SYS_ADMIN))
4863 return ERR_PTR(-EPERM);
4865 err = perf_trace_init(event);
4869 event->destroy = tp_perf_event_destroy;
4871 return &perf_ops_tracepoint;
4874 static int perf_event_set_filter(struct perf_event *event, void __user *arg)
4879 if (event->attr.type != PERF_TYPE_TRACEPOINT)
4882 filter_str = strndup_user(arg, PAGE_SIZE);
4883 if (IS_ERR(filter_str))
4884 return PTR_ERR(filter_str);
4886 ret = ftrace_profile_set_filter(event, event->attr.config, filter_str);
4892 static void perf_event_free_filter(struct perf_event *event)
4894 ftrace_profile_free_filter(event);
4899 static const struct pmu *tp_perf_event_init(struct perf_event *event)
4904 static int perf_event_set_filter(struct perf_event *event, void __user *arg)
4909 static void perf_event_free_filter(struct perf_event *event)
4913 #endif /* CONFIG_EVENT_TRACING */
4915 #ifdef CONFIG_HAVE_HW_BREAKPOINT
4916 static void bp_perf_event_destroy(struct perf_event *event)
4918 release_bp_slot(event);
4921 static const struct pmu *bp_perf_event_init(struct perf_event *bp)
4925 err = register_perf_hw_breakpoint(bp);
4927 return ERR_PTR(err);
4929 bp->destroy = bp_perf_event_destroy;
4931 return &perf_ops_bp;
4934 void perf_bp_event(struct perf_event *bp, void *data)
4936 struct perf_sample_data sample;
4937 struct pt_regs *regs = data;
4939 perf_sample_data_init(&sample, bp->attr.bp_addr);
4941 if (!perf_exclude_event(bp, regs))
4942 perf_swevent_add(bp, 1, 1, &sample, regs);
4945 static const struct pmu *bp_perf_event_init(struct perf_event *bp)
4950 void perf_bp_event(struct perf_event *bp, void *regs)
4955 atomic_t perf_swevent_enabled[PERF_COUNT_SW_MAX];
4957 static void sw_perf_event_destroy(struct perf_event *event)
4959 u64 event_id = event->attr.config;
4961 WARN_ON(event->parent);
4963 atomic_dec(&perf_swevent_enabled[event_id]);
4964 swevent_hlist_put(event);
4967 static const struct pmu *sw_perf_event_init(struct perf_event *event)
4969 const struct pmu *pmu = NULL;
4970 u64 event_id = event->attr.config;
4973 * Software events (currently) can't in general distinguish
4974 * between user, kernel and hypervisor events.
4975 * However, context switches and cpu migrations are considered
4976 * to be kernel events, and page faults are never hypervisor
4980 case PERF_COUNT_SW_CPU_CLOCK:
4981 pmu = &perf_ops_cpu_clock;
4984 case PERF_COUNT_SW_TASK_CLOCK:
4986 * If the user instantiates this as a per-cpu event,
4987 * use the cpu_clock event instead.
4989 if (event->ctx->task)
4990 pmu = &perf_ops_task_clock;
4992 pmu = &perf_ops_cpu_clock;
4995 case PERF_COUNT_SW_PAGE_FAULTS:
4996 case PERF_COUNT_SW_PAGE_FAULTS_MIN:
4997 case PERF_COUNT_SW_PAGE_FAULTS_MAJ:
4998 case PERF_COUNT_SW_CONTEXT_SWITCHES:
4999 case PERF_COUNT_SW_CPU_MIGRATIONS:
5000 case PERF_COUNT_SW_ALIGNMENT_FAULTS:
5001 case PERF_COUNT_SW_EMULATION_FAULTS:
5002 if (!event->parent) {
5005 err = swevent_hlist_get(event);
5007 return ERR_PTR(err);
5009 atomic_inc(&perf_swevent_enabled[event_id]);
5010 event->destroy = sw_perf_event_destroy;
5012 pmu = &perf_ops_generic;
5020 * Allocate and initialize a event structure
5022 static struct perf_event *
5023 perf_event_alloc(struct perf_event_attr *attr,
5025 struct perf_event_context *ctx,
5026 struct perf_event *group_leader,
5027 struct perf_event *parent_event,
5028 perf_overflow_handler_t overflow_handler,
5031 const struct pmu *pmu;
5032 struct perf_event *event;
5033 struct hw_perf_event *hwc;
5036 event = kzalloc(sizeof(*event), gfpflags);
5038 return ERR_PTR(-ENOMEM);
5041 * Single events are their own group leaders, with an
5042 * empty sibling list:
5045 group_leader = event;
5047 mutex_init(&event->child_mutex);
5048 INIT_LIST_HEAD(&event->child_list);
5050 INIT_LIST_HEAD(&event->group_entry);
5051 INIT_LIST_HEAD(&event->event_entry);
5052 INIT_LIST_HEAD(&event->sibling_list);
5053 init_waitqueue_head(&event->waitq);
5055 mutex_init(&event->mmap_mutex);
5058 event->attr = *attr;
5059 event->group_leader = group_leader;
5064 event->parent = parent_event;
5066 event->ns = get_pid_ns(current->nsproxy->pid_ns);
5067 event->id = atomic64_inc_return(&perf_event_id);
5069 event->state = PERF_EVENT_STATE_INACTIVE;
5071 if (!overflow_handler && parent_event)
5072 overflow_handler = parent_event->overflow_handler;
5074 event->overflow_handler = overflow_handler;
5077 event->state = PERF_EVENT_STATE_OFF;
5082 hwc->sample_period = attr->sample_period;
5083 if (attr->freq && attr->sample_freq)
5084 hwc->sample_period = 1;
5085 hwc->last_period = hwc->sample_period;
5087 local64_set(&hwc->period_left, hwc->sample_period);
5090 * we currently do not support PERF_FORMAT_GROUP on inherited events
5092 if (attr->inherit && (attr->read_format & PERF_FORMAT_GROUP))
5095 switch (attr->type) {
5097 case PERF_TYPE_HARDWARE:
5098 case PERF_TYPE_HW_CACHE:
5099 pmu = hw_perf_event_init(event);
5102 case PERF_TYPE_SOFTWARE:
5103 pmu = sw_perf_event_init(event);
5106 case PERF_TYPE_TRACEPOINT:
5107 pmu = tp_perf_event_init(event);
5110 case PERF_TYPE_BREAKPOINT:
5111 pmu = bp_perf_event_init(event);
5122 else if (IS_ERR(pmu))
5127 put_pid_ns(event->ns);
5129 return ERR_PTR(err);
5134 if (!event->parent) {
5135 atomic_inc(&nr_events);
5136 if (event->attr.mmap || event->attr.mmap_data)
5137 atomic_inc(&nr_mmap_events);
5138 if (event->attr.comm)
5139 atomic_inc(&nr_comm_events);
5140 if (event->attr.task)
5141 atomic_inc(&nr_task_events);
5142 if (event->attr.sample_type & PERF_SAMPLE_CALLCHAIN) {
5143 err = get_callchain_buffers();
5146 return ERR_PTR(err);
5154 static int perf_copy_attr(struct perf_event_attr __user *uattr,
5155 struct perf_event_attr *attr)
5160 if (!access_ok(VERIFY_WRITE, uattr, PERF_ATTR_SIZE_VER0))
5164 * zero the full structure, so that a short copy will be nice.
5166 memset(attr, 0, sizeof(*attr));
5168 ret = get_user(size, &uattr->size);
5172 if (size > PAGE_SIZE) /* silly large */
5175 if (!size) /* abi compat */
5176 size = PERF_ATTR_SIZE_VER0;
5178 if (size < PERF_ATTR_SIZE_VER0)
5182 * If we're handed a bigger struct than we know of,
5183 * ensure all the unknown bits are 0 - i.e. new
5184 * user-space does not rely on any kernel feature
5185 * extensions we dont know about yet.
5187 if (size > sizeof(*attr)) {
5188 unsigned char __user *addr;
5189 unsigned char __user *end;
5192 addr = (void __user *)uattr + sizeof(*attr);
5193 end = (void __user *)uattr + size;
5195 for (; addr < end; addr++) {
5196 ret = get_user(val, addr);
5202 size = sizeof(*attr);
5205 ret = copy_from_user(attr, uattr, size);
5210 * If the type exists, the corresponding creation will verify
5213 if (attr->type >= PERF_TYPE_MAX)
5216 if (attr->__reserved_1)
5219 if (attr->sample_type & ~(PERF_SAMPLE_MAX-1))
5222 if (attr->read_format & ~(PERF_FORMAT_MAX-1))
5229 put_user(sizeof(*attr), &uattr->size);
5235 perf_event_set_output(struct perf_event *event, struct perf_event *output_event)
5237 struct perf_buffer *buffer = NULL, *old_buffer = NULL;
5243 /* don't allow circular references */
5244 if (event == output_event)
5248 * Don't allow cross-cpu buffers
5250 if (output_event->cpu != event->cpu)
5254 * If its not a per-cpu buffer, it must be the same task.
5256 if (output_event->cpu == -1 && output_event->ctx != event->ctx)
5260 mutex_lock(&event->mmap_mutex);
5261 /* Can't redirect output if we've got an active mmap() */
5262 if (atomic_read(&event->mmap_count))
5266 /* get the buffer we want to redirect to */
5267 buffer = perf_buffer_get(output_event);
5272 old_buffer = event->buffer;
5273 rcu_assign_pointer(event->buffer, buffer);
5276 mutex_unlock(&event->mmap_mutex);
5279 perf_buffer_put(old_buffer);
5285 * sys_perf_event_open - open a performance event, associate it to a task/cpu
5287 * @attr_uptr: event_id type attributes for monitoring/sampling
5290 * @group_fd: group leader event fd
5292 SYSCALL_DEFINE5(perf_event_open,
5293 struct perf_event_attr __user *, attr_uptr,
5294 pid_t, pid, int, cpu, int, group_fd, unsigned long, flags)
5296 struct perf_event *event, *group_leader = NULL, *output_event = NULL;
5297 struct perf_event_attr attr;
5298 struct perf_event_context *ctx;
5299 struct file *event_file = NULL;
5300 struct file *group_file = NULL;
5302 int fput_needed = 0;
5305 /* for future expandability... */
5306 if (flags & ~(PERF_FLAG_FD_NO_GROUP | PERF_FLAG_FD_OUTPUT))
5309 err = perf_copy_attr(attr_uptr, &attr);
5313 if (!attr.exclude_kernel) {
5314 if (perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN))
5319 if (attr.sample_freq > sysctl_perf_event_sample_rate)
5323 event_fd = get_unused_fd_flags(O_RDWR);
5328 * Get the target context (task or percpu):
5330 ctx = find_get_context(pid, cpu);
5336 if (group_fd != -1) {
5337 group_leader = perf_fget_light(group_fd, &fput_needed);
5338 if (IS_ERR(group_leader)) {
5339 err = PTR_ERR(group_leader);
5340 goto err_put_context;
5342 group_file = group_leader->filp;
5343 if (flags & PERF_FLAG_FD_OUTPUT)
5344 output_event = group_leader;
5345 if (flags & PERF_FLAG_FD_NO_GROUP)
5346 group_leader = NULL;
5350 * Look up the group leader (we will attach this event to it):
5356 * Do not allow a recursive hierarchy (this new sibling
5357 * becoming part of another group-sibling):
5359 if (group_leader->group_leader != group_leader)
5360 goto err_put_context;
5362 * Do not allow to attach to a group in a different
5363 * task or CPU context:
5365 if (group_leader->ctx != ctx)
5366 goto err_put_context;
5368 * Only a group leader can be exclusive or pinned
5370 if (attr.exclusive || attr.pinned)
5371 goto err_put_context;
5374 event = perf_event_alloc(&attr, cpu, ctx, group_leader,
5375 NULL, NULL, GFP_KERNEL);
5376 if (IS_ERR(event)) {
5377 err = PTR_ERR(event);
5378 goto err_put_context;
5382 err = perf_event_set_output(event, output_event);
5384 goto err_free_put_context;
5387 event_file = anon_inode_getfile("[perf_event]", &perf_fops, event, O_RDWR);
5388 if (IS_ERR(event_file)) {
5389 err = PTR_ERR(event_file);
5390 goto err_free_put_context;
5393 event->filp = event_file;
5394 WARN_ON_ONCE(ctx->parent_ctx);
5395 mutex_lock(&ctx->mutex);
5396 perf_install_in_context(ctx, event, cpu);
5398 mutex_unlock(&ctx->mutex);
5400 event->owner = current;
5401 get_task_struct(current);
5402 mutex_lock(¤t->perf_event_mutex);
5403 list_add_tail(&event->owner_entry, ¤t->perf_event_list);
5404 mutex_unlock(¤t->perf_event_mutex);
5407 * Drop the reference on the group_event after placing the
5408 * new event on the sibling_list. This ensures destruction
5409 * of the group leader will find the pointer to itself in
5410 * perf_group_detach().
5412 fput_light(group_file, fput_needed);
5413 fd_install(event_fd, event_file);
5416 err_free_put_context:
5419 fput_light(group_file, fput_needed);
5422 put_unused_fd(event_fd);
5427 * perf_event_create_kernel_counter
5429 * @attr: attributes of the counter to create
5430 * @cpu: cpu in which the counter is bound
5431 * @pid: task to profile
5434 perf_event_create_kernel_counter(struct perf_event_attr *attr, int cpu,
5436 perf_overflow_handler_t overflow_handler)
5438 struct perf_event *event;
5439 struct perf_event_context *ctx;
5443 * Get the target context (task or percpu):
5446 ctx = find_get_context(pid, cpu);
5452 event = perf_event_alloc(attr, cpu, ctx, NULL,
5453 NULL, overflow_handler, GFP_KERNEL);
5454 if (IS_ERR(event)) {
5455 err = PTR_ERR(event);
5456 goto err_put_context;
5460 WARN_ON_ONCE(ctx->parent_ctx);
5461 mutex_lock(&ctx->mutex);
5462 perf_install_in_context(ctx, event, cpu);
5464 mutex_unlock(&ctx->mutex);
5466 event->owner = current;
5467 get_task_struct(current);
5468 mutex_lock(¤t->perf_event_mutex);
5469 list_add_tail(&event->owner_entry, ¤t->perf_event_list);
5470 mutex_unlock(¤t->perf_event_mutex);
5477 return ERR_PTR(err);
5479 EXPORT_SYMBOL_GPL(perf_event_create_kernel_counter);
5482 * inherit a event from parent task to child task:
5484 static struct perf_event *
5485 inherit_event(struct perf_event *parent_event,
5486 struct task_struct *parent,
5487 struct perf_event_context *parent_ctx,
5488 struct task_struct *child,
5489 struct perf_event *group_leader,
5490 struct perf_event_context *child_ctx)
5492 struct perf_event *child_event;
5495 * Instead of creating recursive hierarchies of events,
5496 * we link inherited events back to the original parent,
5497 * which has a filp for sure, which we use as the reference
5500 if (parent_event->parent)
5501 parent_event = parent_event->parent;
5503 child_event = perf_event_alloc(&parent_event->attr,
5504 parent_event->cpu, child_ctx,
5505 group_leader, parent_event,
5507 if (IS_ERR(child_event))
5512 * Make the child state follow the state of the parent event,
5513 * not its attr.disabled bit. We hold the parent's mutex,
5514 * so we won't race with perf_event_{en, dis}able_family.
5516 if (parent_event->state >= PERF_EVENT_STATE_INACTIVE)
5517 child_event->state = PERF_EVENT_STATE_INACTIVE;
5519 child_event->state = PERF_EVENT_STATE_OFF;
5521 if (parent_event->attr.freq) {
5522 u64 sample_period = parent_event->hw.sample_period;
5523 struct hw_perf_event *hwc = &child_event->hw;
5525 hwc->sample_period = sample_period;
5526 hwc->last_period = sample_period;
5528 local64_set(&hwc->period_left, sample_period);
5531 child_event->overflow_handler = parent_event->overflow_handler;
5534 * Link it up in the child's context:
5536 add_event_to_ctx(child_event, child_ctx);
5539 * Get a reference to the parent filp - we will fput it
5540 * when the child event exits. This is safe to do because
5541 * we are in the parent and we know that the filp still
5542 * exists and has a nonzero count:
5544 atomic_long_inc(&parent_event->filp->f_count);
5547 * Link this into the parent event's child list
5549 WARN_ON_ONCE(parent_event->ctx->parent_ctx);
5550 mutex_lock(&parent_event->child_mutex);
5551 list_add_tail(&child_event->child_list, &parent_event->child_list);
5552 mutex_unlock(&parent_event->child_mutex);
5557 static int inherit_group(struct perf_event *parent_event,
5558 struct task_struct *parent,
5559 struct perf_event_context *parent_ctx,
5560 struct task_struct *child,
5561 struct perf_event_context *child_ctx)
5563 struct perf_event *leader;
5564 struct perf_event *sub;
5565 struct perf_event *child_ctr;
5567 leader = inherit_event(parent_event, parent, parent_ctx,
5568 child, NULL, child_ctx);
5570 return PTR_ERR(leader);
5571 list_for_each_entry(sub, &parent_event->sibling_list, group_entry) {
5572 child_ctr = inherit_event(sub, parent, parent_ctx,
5573 child, leader, child_ctx);
5574 if (IS_ERR(child_ctr))
5575 return PTR_ERR(child_ctr);
5580 static void sync_child_event(struct perf_event *child_event,
5581 struct task_struct *child)
5583 struct perf_event *parent_event = child_event->parent;
5586 if (child_event->attr.inherit_stat)
5587 perf_event_read_event(child_event, child);
5589 child_val = perf_event_count(child_event);
5592 * Add back the child's count to the parent's count:
5594 atomic64_add(child_val, &parent_event->child_count);
5595 atomic64_add(child_event->total_time_enabled,
5596 &parent_event->child_total_time_enabled);
5597 atomic64_add(child_event->total_time_running,
5598 &parent_event->child_total_time_running);
5601 * Remove this event from the parent's list
5603 WARN_ON_ONCE(parent_event->ctx->parent_ctx);
5604 mutex_lock(&parent_event->child_mutex);
5605 list_del_init(&child_event->child_list);
5606 mutex_unlock(&parent_event->child_mutex);
5609 * Release the parent event, if this was the last
5612 fput(parent_event->filp);
5616 __perf_event_exit_task(struct perf_event *child_event,
5617 struct perf_event_context *child_ctx,
5618 struct task_struct *child)
5620 struct perf_event *parent_event;
5622 perf_event_remove_from_context(child_event);
5624 parent_event = child_event->parent;
5626 * It can happen that parent exits first, and has events
5627 * that are still around due to the child reference. These
5628 * events need to be zapped - but otherwise linger.
5631 sync_child_event(child_event, child);
5632 free_event(child_event);
5637 * When a child task exits, feed back event values to parent events.
5639 void perf_event_exit_task(struct task_struct *child)
5641 struct perf_event *child_event, *tmp;
5642 struct perf_event_context *child_ctx;
5643 unsigned long flags;
5645 if (likely(!child->perf_event_ctxp)) {
5646 perf_event_task(child, NULL, 0);
5650 local_irq_save(flags);
5652 * We can't reschedule here because interrupts are disabled,
5653 * and either child is current or it is a task that can't be
5654 * scheduled, so we are now safe from rescheduling changing
5657 child_ctx = child->perf_event_ctxp;
5658 __perf_event_task_sched_out(child_ctx);
5661 * Take the context lock here so that if find_get_context is
5662 * reading child->perf_event_ctxp, we wait until it has
5663 * incremented the context's refcount before we do put_ctx below.
5665 raw_spin_lock(&child_ctx->lock);
5666 child->perf_event_ctxp = NULL;
5668 * If this context is a clone; unclone it so it can't get
5669 * swapped to another process while we're removing all
5670 * the events from it.
5672 unclone_ctx(child_ctx);
5673 update_context_time(child_ctx);
5674 raw_spin_unlock_irqrestore(&child_ctx->lock, flags);
5677 * Report the task dead after unscheduling the events so that we
5678 * won't get any samples after PERF_RECORD_EXIT. We can however still
5679 * get a few PERF_RECORD_READ events.
5681 perf_event_task(child, child_ctx, 0);
5684 * We can recurse on the same lock type through:
5686 * __perf_event_exit_task()
5687 * sync_child_event()
5688 * fput(parent_event->filp)
5690 * mutex_lock(&ctx->mutex)
5692 * But since its the parent context it won't be the same instance.
5694 mutex_lock(&child_ctx->mutex);
5697 list_for_each_entry_safe(child_event, tmp, &child_ctx->pinned_groups,
5699 __perf_event_exit_task(child_event, child_ctx, child);
5701 list_for_each_entry_safe(child_event, tmp, &child_ctx->flexible_groups,
5703 __perf_event_exit_task(child_event, child_ctx, child);
5706 * If the last event was a group event, it will have appended all
5707 * its siblings to the list, but we obtained 'tmp' before that which
5708 * will still point to the list head terminating the iteration.
5710 if (!list_empty(&child_ctx->pinned_groups) ||
5711 !list_empty(&child_ctx->flexible_groups))
5714 mutex_unlock(&child_ctx->mutex);
5719 static void perf_free_event(struct perf_event *event,
5720 struct perf_event_context *ctx)
5722 struct perf_event *parent = event->parent;
5724 if (WARN_ON_ONCE(!parent))
5727 mutex_lock(&parent->child_mutex);
5728 list_del_init(&event->child_list);
5729 mutex_unlock(&parent->child_mutex);
5733 perf_group_detach(event);
5734 list_del_event(event, ctx);
5739 * free an unexposed, unused context as created by inheritance by
5740 * init_task below, used by fork() in case of fail.
5742 void perf_event_free_task(struct task_struct *task)
5744 struct perf_event_context *ctx = task->perf_event_ctxp;
5745 struct perf_event *event, *tmp;
5750 mutex_lock(&ctx->mutex);
5752 list_for_each_entry_safe(event, tmp, &ctx->pinned_groups, group_entry)
5753 perf_free_event(event, ctx);
5755 list_for_each_entry_safe(event, tmp, &ctx->flexible_groups,
5757 perf_free_event(event, ctx);
5759 if (!list_empty(&ctx->pinned_groups) ||
5760 !list_empty(&ctx->flexible_groups))
5763 mutex_unlock(&ctx->mutex);
5769 inherit_task_group(struct perf_event *event, struct task_struct *parent,
5770 struct perf_event_context *parent_ctx,
5771 struct task_struct *child,
5775 struct perf_event_context *child_ctx = child->perf_event_ctxp;
5777 if (!event->attr.inherit) {
5784 * This is executed from the parent task context, so
5785 * inherit events that have been marked for cloning.
5786 * First allocate and initialize a context for the
5790 child_ctx = kzalloc(sizeof(struct perf_event_context),
5795 __perf_event_init_context(child_ctx, child);
5796 child->perf_event_ctxp = child_ctx;
5797 get_task_struct(child);
5800 ret = inherit_group(event, parent, parent_ctx,
5811 * Initialize the perf_event context in task_struct
5813 int perf_event_init_task(struct task_struct *child)
5815 struct perf_event_context *child_ctx, *parent_ctx;
5816 struct perf_event_context *cloned_ctx;
5817 struct perf_event *event;
5818 struct task_struct *parent = current;
5819 int inherited_all = 1;
5822 child->perf_event_ctxp = NULL;
5824 mutex_init(&child->perf_event_mutex);
5825 INIT_LIST_HEAD(&child->perf_event_list);
5827 if (likely(!parent->perf_event_ctxp))
5831 * If the parent's context is a clone, pin it so it won't get
5834 parent_ctx = perf_pin_task_context(parent);
5837 * No need to check if parent_ctx != NULL here; since we saw
5838 * it non-NULL earlier, the only reason for it to become NULL
5839 * is if we exit, and since we're currently in the middle of
5840 * a fork we can't be exiting at the same time.
5844 * Lock the parent list. No need to lock the child - not PID
5845 * hashed yet and not running, so nobody can access it.
5847 mutex_lock(&parent_ctx->mutex);
5850 * We dont have to disable NMIs - we are only looking at
5851 * the list, not manipulating it:
5853 list_for_each_entry(event, &parent_ctx->pinned_groups, group_entry) {
5854 ret = inherit_task_group(event, parent, parent_ctx, child,
5860 list_for_each_entry(event, &parent_ctx->flexible_groups, group_entry) {
5861 ret = inherit_task_group(event, parent, parent_ctx, child,
5867 child_ctx = child->perf_event_ctxp;
5869 if (child_ctx && inherited_all) {
5871 * Mark the child context as a clone of the parent
5872 * context, or of whatever the parent is a clone of.
5873 * Note that if the parent is a clone, it could get
5874 * uncloned at any point, but that doesn't matter
5875 * because the list of events and the generation
5876 * count can't have changed since we took the mutex.
5878 cloned_ctx = rcu_dereference(parent_ctx->parent_ctx);
5880 child_ctx->parent_ctx = cloned_ctx;
5881 child_ctx->parent_gen = parent_ctx->parent_gen;
5883 child_ctx->parent_ctx = parent_ctx;
5884 child_ctx->parent_gen = parent_ctx->generation;
5886 get_ctx(child_ctx->parent_ctx);
5889 mutex_unlock(&parent_ctx->mutex);
5891 perf_unpin_context(parent_ctx);
5896 static void __init perf_event_init_all_cpus(void)
5899 struct perf_cpu_context *cpuctx;
5901 for_each_possible_cpu(cpu) {
5902 cpuctx = &per_cpu(perf_cpu_context, cpu);
5903 mutex_init(&cpuctx->hlist_mutex);
5904 __perf_event_init_context(&cpuctx->ctx, NULL);
5908 static void __cpuinit perf_event_init_cpu(int cpu)
5910 struct perf_cpu_context *cpuctx;
5912 cpuctx = &per_cpu(perf_cpu_context, cpu);
5914 spin_lock(&perf_resource_lock);
5915 cpuctx->max_pertask = perf_max_events - perf_reserved_percpu;
5916 spin_unlock(&perf_resource_lock);
5918 mutex_lock(&cpuctx->hlist_mutex);
5919 if (cpuctx->hlist_refcount > 0) {
5920 struct swevent_hlist *hlist;
5922 hlist = kzalloc(sizeof(*hlist), GFP_KERNEL);
5923 WARN_ON_ONCE(!hlist);
5924 rcu_assign_pointer(cpuctx->swevent_hlist, hlist);
5926 mutex_unlock(&cpuctx->hlist_mutex);
5929 #ifdef CONFIG_HOTPLUG_CPU
5930 static void __perf_event_exit_cpu(void *info)
5932 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
5933 struct perf_event_context *ctx = &cpuctx->ctx;
5934 struct perf_event *event, *tmp;
5936 list_for_each_entry_safe(event, tmp, &ctx->pinned_groups, group_entry)
5937 __perf_event_remove_from_context(event);
5938 list_for_each_entry_safe(event, tmp, &ctx->flexible_groups, group_entry)
5939 __perf_event_remove_from_context(event);
5941 static void perf_event_exit_cpu(int cpu)
5943 struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
5944 struct perf_event_context *ctx = &cpuctx->ctx;
5946 mutex_lock(&cpuctx->hlist_mutex);
5947 swevent_hlist_release(cpuctx);
5948 mutex_unlock(&cpuctx->hlist_mutex);
5950 mutex_lock(&ctx->mutex);
5951 smp_call_function_single(cpu, __perf_event_exit_cpu, NULL, 1);
5952 mutex_unlock(&ctx->mutex);
5955 static inline void perf_event_exit_cpu(int cpu) { }
5958 static int __cpuinit
5959 perf_cpu_notify(struct notifier_block *self, unsigned long action, void *hcpu)
5961 unsigned int cpu = (long)hcpu;
5963 switch (action & ~CPU_TASKS_FROZEN) {
5965 case CPU_UP_PREPARE:
5966 case CPU_DOWN_FAILED:
5967 perf_event_init_cpu(cpu);
5970 case CPU_UP_CANCELED:
5971 case CPU_DOWN_PREPARE:
5972 perf_event_exit_cpu(cpu);
5983 * This has to have a higher priority than migration_notifier in sched.c.
5985 static struct notifier_block __cpuinitdata perf_cpu_nb = {
5986 .notifier_call = perf_cpu_notify,
5990 void __init perf_event_init(void)
5992 perf_event_init_all_cpus();
5993 perf_cpu_notify(&perf_cpu_nb, (unsigned long)CPU_UP_PREPARE,
5994 (void *)(long)smp_processor_id());
5995 perf_cpu_notify(&perf_cpu_nb, (unsigned long)CPU_ONLINE,
5996 (void *)(long)smp_processor_id());
5997 register_cpu_notifier(&perf_cpu_nb);
6000 static ssize_t perf_show_reserve_percpu(struct sysdev_class *class,
6001 struct sysdev_class_attribute *attr,
6004 return sprintf(buf, "%d\n", perf_reserved_percpu);
6008 perf_set_reserve_percpu(struct sysdev_class *class,
6009 struct sysdev_class_attribute *attr,
6013 struct perf_cpu_context *cpuctx;
6017 err = strict_strtoul(buf, 10, &val);
6020 if (val > perf_max_events)
6023 spin_lock(&perf_resource_lock);
6024 perf_reserved_percpu = val;
6025 for_each_online_cpu(cpu) {
6026 cpuctx = &per_cpu(perf_cpu_context, cpu);
6027 raw_spin_lock_irq(&cpuctx->ctx.lock);
6028 mpt = min(perf_max_events - cpuctx->ctx.nr_events,
6029 perf_max_events - perf_reserved_percpu);
6030 cpuctx->max_pertask = mpt;
6031 raw_spin_unlock_irq(&cpuctx->ctx.lock);
6033 spin_unlock(&perf_resource_lock);
6038 static ssize_t perf_show_overcommit(struct sysdev_class *class,
6039 struct sysdev_class_attribute *attr,
6042 return sprintf(buf, "%d\n", perf_overcommit);
6046 perf_set_overcommit(struct sysdev_class *class,
6047 struct sysdev_class_attribute *attr,
6048 const char *buf, size_t count)
6053 err = strict_strtoul(buf, 10, &val);
6059 spin_lock(&perf_resource_lock);
6060 perf_overcommit = val;
6061 spin_unlock(&perf_resource_lock);
6066 static SYSDEV_CLASS_ATTR(
6069 perf_show_reserve_percpu,
6070 perf_set_reserve_percpu
6073 static SYSDEV_CLASS_ATTR(
6076 perf_show_overcommit,
6080 static struct attribute *perfclass_attrs[] = {
6081 &attr_reserve_percpu.attr,
6082 &attr_overcommit.attr,
6086 static struct attribute_group perfclass_attr_group = {
6087 .attrs = perfclass_attrs,
6088 .name = "perf_events",
6091 static int __init perf_event_sysfs_init(void)
6093 return sysfs_create_group(&cpu_sysdev_class.kset.kobj,
6094 &perfclass_attr_group);
6096 device_initcall(perf_event_sysfs_init);