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(); }
87 hw_perf_group_sched_in(struct perf_event *group_leader,
88 struct perf_cpu_context *cpuctx,
89 struct perf_event_context *ctx)
94 void __weak perf_event_print_debug(void) { }
96 static DEFINE_PER_CPU(int, perf_disable_count);
98 void perf_disable(void)
100 if (!__get_cpu_var(perf_disable_count)++)
104 void perf_enable(void)
106 if (!--__get_cpu_var(perf_disable_count))
110 static void get_ctx(struct perf_event_context *ctx)
112 WARN_ON(!atomic_inc_not_zero(&ctx->refcount));
115 static void free_ctx(struct rcu_head *head)
117 struct perf_event_context *ctx;
119 ctx = container_of(head, struct perf_event_context, rcu_head);
123 static void put_ctx(struct perf_event_context *ctx)
125 if (atomic_dec_and_test(&ctx->refcount)) {
127 put_ctx(ctx->parent_ctx);
129 put_task_struct(ctx->task);
130 call_rcu(&ctx->rcu_head, free_ctx);
134 static void unclone_ctx(struct perf_event_context *ctx)
136 if (ctx->parent_ctx) {
137 put_ctx(ctx->parent_ctx);
138 ctx->parent_ctx = NULL;
143 * If we inherit events we want to return the parent event id
146 static u64 primary_event_id(struct perf_event *event)
151 id = event->parent->id;
157 * Get the perf_event_context for a task and lock it.
158 * This has to cope with with the fact that until it is locked,
159 * the context could get moved to another task.
161 static struct perf_event_context *
162 perf_lock_task_context(struct task_struct *task, unsigned long *flags)
164 struct perf_event_context *ctx;
168 ctx = rcu_dereference(task->perf_event_ctxp);
171 * If this context is a clone of another, it might
172 * get swapped for another underneath us by
173 * perf_event_task_sched_out, though the
174 * rcu_read_lock() protects us from any context
175 * getting freed. Lock the context and check if it
176 * got swapped before we could get the lock, and retry
177 * if so. If we locked the right context, then it
178 * can't get swapped on us any more.
180 raw_spin_lock_irqsave(&ctx->lock, *flags);
181 if (ctx != rcu_dereference(task->perf_event_ctxp)) {
182 raw_spin_unlock_irqrestore(&ctx->lock, *flags);
186 if (!atomic_inc_not_zero(&ctx->refcount)) {
187 raw_spin_unlock_irqrestore(&ctx->lock, *flags);
196 * Get the context for a task and increment its pin_count so it
197 * can't get swapped to another task. This also increments its
198 * reference count so that the context can't get freed.
200 static struct perf_event_context *perf_pin_task_context(struct task_struct *task)
202 struct perf_event_context *ctx;
205 ctx = perf_lock_task_context(task, &flags);
208 raw_spin_unlock_irqrestore(&ctx->lock, flags);
213 static void perf_unpin_context(struct perf_event_context *ctx)
217 raw_spin_lock_irqsave(&ctx->lock, flags);
219 raw_spin_unlock_irqrestore(&ctx->lock, flags);
223 static inline u64 perf_clock(void)
225 return cpu_clock(raw_smp_processor_id());
229 * Update the record of the current time in a context.
231 static void update_context_time(struct perf_event_context *ctx)
233 u64 now = perf_clock();
235 ctx->time += now - ctx->timestamp;
236 ctx->timestamp = now;
240 * Update the total_time_enabled and total_time_running fields for a event.
242 static void update_event_times(struct perf_event *event)
244 struct perf_event_context *ctx = event->ctx;
247 if (event->state < PERF_EVENT_STATE_INACTIVE ||
248 event->group_leader->state < PERF_EVENT_STATE_INACTIVE)
254 run_end = event->tstamp_stopped;
256 event->total_time_enabled = run_end - event->tstamp_enabled;
258 if (event->state == PERF_EVENT_STATE_INACTIVE)
259 run_end = event->tstamp_stopped;
263 event->total_time_running = run_end - event->tstamp_running;
266 static struct list_head *
267 ctx_group_list(struct perf_event *event, struct perf_event_context *ctx)
269 if (event->attr.pinned)
270 return &ctx->pinned_groups;
272 return &ctx->flexible_groups;
276 * Add a event from the lists for its context.
277 * Must be called with ctx->mutex and ctx->lock held.
280 list_add_event(struct perf_event *event, struct perf_event_context *ctx)
282 struct perf_event *group_leader = event->group_leader;
285 * Depending on whether it is a standalone or sibling event,
286 * add it straight to the context's event list, or to the group
287 * leader's sibling list:
289 if (group_leader == event) {
290 struct list_head *list;
292 if (is_software_event(event))
293 event->group_flags |= PERF_GROUP_SOFTWARE;
295 list = ctx_group_list(event, ctx);
296 list_add_tail(&event->group_entry, list);
298 if (group_leader->group_flags & PERF_GROUP_SOFTWARE &&
299 !is_software_event(event))
300 group_leader->group_flags &= ~PERF_GROUP_SOFTWARE;
302 list_add_tail(&event->group_entry, &group_leader->sibling_list);
303 group_leader->nr_siblings++;
306 list_add_rcu(&event->event_entry, &ctx->event_list);
308 if (event->attr.inherit_stat)
313 * Remove a event from the lists for its context.
314 * Must be called with ctx->mutex and ctx->lock held.
317 list_del_event(struct perf_event *event, struct perf_event_context *ctx)
319 struct perf_event *sibling, *tmp;
321 if (list_empty(&event->group_entry))
324 if (event->attr.inherit_stat)
327 list_del_init(&event->group_entry);
328 list_del_rcu(&event->event_entry);
330 if (event->group_leader != event)
331 event->group_leader->nr_siblings--;
333 update_event_times(event);
336 * If event was in error state, then keep it
337 * that way, otherwise bogus counts will be
338 * returned on read(). The only way to get out
339 * of error state is by explicit re-enabling
342 if (event->state > PERF_EVENT_STATE_OFF)
343 event->state = PERF_EVENT_STATE_OFF;
345 if (event->state > PERF_EVENT_STATE_FREE)
349 * If this was a group event with sibling events then
350 * upgrade the siblings to singleton events by adding them
351 * to the context list directly:
353 list_for_each_entry_safe(sibling, tmp, &event->sibling_list, group_entry) {
354 struct list_head *list;
356 list = ctx_group_list(event, ctx);
357 list_move_tail(&sibling->group_entry, list);
358 sibling->group_leader = sibling;
360 /* Inherit group flags from the previous leader */
361 sibling->group_flags = event->group_flags;
366 event_sched_out(struct perf_event *event,
367 struct perf_cpu_context *cpuctx,
368 struct perf_event_context *ctx)
370 if (event->state != PERF_EVENT_STATE_ACTIVE)
373 event->state = PERF_EVENT_STATE_INACTIVE;
374 if (event->pending_disable) {
375 event->pending_disable = 0;
376 event->state = PERF_EVENT_STATE_OFF;
378 event->tstamp_stopped = ctx->time;
379 event->pmu->disable(event);
382 if (!is_software_event(event))
383 cpuctx->active_oncpu--;
385 if (event->attr.exclusive || !cpuctx->active_oncpu)
386 cpuctx->exclusive = 0;
390 group_sched_out(struct perf_event *group_event,
391 struct perf_cpu_context *cpuctx,
392 struct perf_event_context *ctx)
394 struct perf_event *event;
396 if (group_event->state != PERF_EVENT_STATE_ACTIVE)
399 event_sched_out(group_event, cpuctx, ctx);
402 * Schedule out siblings (if any):
404 list_for_each_entry(event, &group_event->sibling_list, group_entry)
405 event_sched_out(event, cpuctx, ctx);
407 if (group_event->attr.exclusive)
408 cpuctx->exclusive = 0;
412 * Cross CPU call to remove a performance event
414 * We disable the event on the hardware level first. After that we
415 * remove it from the context list.
417 static void __perf_event_remove_from_context(void *info)
419 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
420 struct perf_event *event = info;
421 struct perf_event_context *ctx = event->ctx;
424 * If this is a task context, we need to check whether it is
425 * the current task context of this cpu. If not it has been
426 * scheduled out before the smp call arrived.
428 if (ctx->task && cpuctx->task_ctx != ctx)
431 raw_spin_lock(&ctx->lock);
433 * Protect the list operation against NMI by disabling the
434 * events on a global level.
438 event_sched_out(event, cpuctx, ctx);
440 list_del_event(event, ctx);
444 * Allow more per task events with respect to the
447 cpuctx->max_pertask =
448 min(perf_max_events - ctx->nr_events,
449 perf_max_events - perf_reserved_percpu);
453 raw_spin_unlock(&ctx->lock);
458 * Remove the event from a task's (or a CPU's) list of events.
460 * Must be called with ctx->mutex held.
462 * CPU events are removed with a smp call. For task events we only
463 * call when the task is on a CPU.
465 * If event->ctx is a cloned context, callers must make sure that
466 * every task struct that event->ctx->task could possibly point to
467 * remains valid. This is OK when called from perf_release since
468 * that only calls us on the top-level context, which can't be a clone.
469 * When called from perf_event_exit_task, it's OK because the
470 * context has been detached from its task.
472 static void perf_event_remove_from_context(struct perf_event *event)
474 struct perf_event_context *ctx = event->ctx;
475 struct task_struct *task = ctx->task;
479 * Per cpu events are removed via an smp call and
480 * the removal is always successful.
482 smp_call_function_single(event->cpu,
483 __perf_event_remove_from_context,
489 task_oncpu_function_call(task, __perf_event_remove_from_context,
492 raw_spin_lock_irq(&ctx->lock);
494 * If the context is active we need to retry the smp call.
496 if (ctx->nr_active && !list_empty(&event->group_entry)) {
497 raw_spin_unlock_irq(&ctx->lock);
502 * The lock prevents that this context is scheduled in so we
503 * can remove the event safely, if the call above did not
506 if (!list_empty(&event->group_entry))
507 list_del_event(event, ctx);
508 raw_spin_unlock_irq(&ctx->lock);
512 * Update total_time_enabled and total_time_running for all events in a group.
514 static void update_group_times(struct perf_event *leader)
516 struct perf_event *event;
518 update_event_times(leader);
519 list_for_each_entry(event, &leader->sibling_list, group_entry)
520 update_event_times(event);
524 * Cross CPU call to disable a performance event
526 static void __perf_event_disable(void *info)
528 struct perf_event *event = info;
529 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
530 struct perf_event_context *ctx = event->ctx;
533 * If this is a per-task event, need to check whether this
534 * event's task is the current task on this cpu.
536 if (ctx->task && cpuctx->task_ctx != ctx)
539 raw_spin_lock(&ctx->lock);
542 * If the event is on, turn it off.
543 * If it is in error state, leave it in error state.
545 if (event->state >= PERF_EVENT_STATE_INACTIVE) {
546 update_context_time(ctx);
547 update_group_times(event);
548 if (event == event->group_leader)
549 group_sched_out(event, cpuctx, ctx);
551 event_sched_out(event, cpuctx, ctx);
552 event->state = PERF_EVENT_STATE_OFF;
555 raw_spin_unlock(&ctx->lock);
561 * If event->ctx is a cloned context, callers must make sure that
562 * every task struct that event->ctx->task could possibly point to
563 * remains valid. This condition is satisifed when called through
564 * perf_event_for_each_child or perf_event_for_each because they
565 * hold the top-level event's child_mutex, so any descendant that
566 * goes to exit will block in sync_child_event.
567 * When called from perf_pending_event it's OK because event->ctx
568 * is the current context on this CPU and preemption is disabled,
569 * hence we can't get into perf_event_task_sched_out for this context.
571 void perf_event_disable(struct perf_event *event)
573 struct perf_event_context *ctx = event->ctx;
574 struct task_struct *task = ctx->task;
578 * Disable the event on the cpu that it's on
580 smp_call_function_single(event->cpu, __perf_event_disable,
586 task_oncpu_function_call(task, __perf_event_disable, event);
588 raw_spin_lock_irq(&ctx->lock);
590 * If the event is still active, we need to retry the cross-call.
592 if (event->state == PERF_EVENT_STATE_ACTIVE) {
593 raw_spin_unlock_irq(&ctx->lock);
598 * Since we have the lock this context can't be scheduled
599 * in, so we can change the state safely.
601 if (event->state == PERF_EVENT_STATE_INACTIVE) {
602 update_group_times(event);
603 event->state = PERF_EVENT_STATE_OFF;
606 raw_spin_unlock_irq(&ctx->lock);
610 event_sched_in(struct perf_event *event,
611 struct perf_cpu_context *cpuctx,
612 struct perf_event_context *ctx)
614 if (event->state <= PERF_EVENT_STATE_OFF)
617 event->state = PERF_EVENT_STATE_ACTIVE;
618 event->oncpu = smp_processor_id();
620 * The new state must be visible before we turn it on in the hardware:
624 if (event->pmu->enable(event)) {
625 event->state = PERF_EVENT_STATE_INACTIVE;
630 event->tstamp_running += ctx->time - event->tstamp_stopped;
632 if (!is_software_event(event))
633 cpuctx->active_oncpu++;
636 if (event->attr.exclusive)
637 cpuctx->exclusive = 1;
643 group_sched_in(struct perf_event *group_event,
644 struct perf_cpu_context *cpuctx,
645 struct perf_event_context *ctx)
647 struct perf_event *event, *partial_group;
650 if (group_event->state == PERF_EVENT_STATE_OFF)
653 ret = hw_perf_group_sched_in(group_event, cpuctx, ctx);
655 return ret < 0 ? ret : 0;
657 if (event_sched_in(group_event, cpuctx, ctx))
661 * Schedule in siblings as one group (if any):
663 list_for_each_entry(event, &group_event->sibling_list, group_entry) {
664 if (event_sched_in(event, cpuctx, ctx)) {
665 partial_group = event;
674 * Groups can be scheduled in as one unit only, so undo any
675 * partial group before returning:
677 list_for_each_entry(event, &group_event->sibling_list, group_entry) {
678 if (event == partial_group)
680 event_sched_out(event, cpuctx, ctx);
682 event_sched_out(group_event, cpuctx, ctx);
688 * Work out whether we can put this event group on the CPU now.
690 static int group_can_go_on(struct perf_event *event,
691 struct perf_cpu_context *cpuctx,
695 * Groups consisting entirely of software events can always go on.
697 if (event->group_flags & PERF_GROUP_SOFTWARE)
700 * If an exclusive group is already on, no other hardware
703 if (cpuctx->exclusive)
706 * If this group is exclusive and there are already
707 * events on the CPU, it can't go on.
709 if (event->attr.exclusive && cpuctx->active_oncpu)
712 * Otherwise, try to add it if all previous groups were able
718 static void add_event_to_ctx(struct perf_event *event,
719 struct perf_event_context *ctx)
721 list_add_event(event, ctx);
722 event->tstamp_enabled = ctx->time;
723 event->tstamp_running = ctx->time;
724 event->tstamp_stopped = ctx->time;
728 * Cross CPU call to install and enable a performance event
730 * Must be called with ctx->mutex held
732 static void __perf_install_in_context(void *info)
734 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
735 struct perf_event *event = info;
736 struct perf_event_context *ctx = event->ctx;
737 struct perf_event *leader = event->group_leader;
741 * If this is a task context, we need to check whether it is
742 * the current task context of this cpu. If not it has been
743 * scheduled out before the smp call arrived.
744 * Or possibly this is the right context but it isn't
745 * on this cpu because it had no events.
747 if (ctx->task && cpuctx->task_ctx != ctx) {
748 if (cpuctx->task_ctx || ctx->task != current)
750 cpuctx->task_ctx = ctx;
753 raw_spin_lock(&ctx->lock);
755 update_context_time(ctx);
758 * Protect the list operation against NMI by disabling the
759 * events on a global level. NOP for non NMI based events.
763 add_event_to_ctx(event, ctx);
765 if (event->cpu != -1 && event->cpu != smp_processor_id())
769 * Don't put the event on if it is disabled or if
770 * it is in a group and the group isn't on.
772 if (event->state != PERF_EVENT_STATE_INACTIVE ||
773 (leader != event && leader->state != PERF_EVENT_STATE_ACTIVE))
777 * An exclusive event can't go on if there are already active
778 * hardware events, and no hardware event can go on if there
779 * is already an exclusive event on.
781 if (!group_can_go_on(event, cpuctx, 1))
784 err = event_sched_in(event, cpuctx, ctx);
788 * This event couldn't go on. If it is in a group
789 * then we have to pull the whole group off.
790 * If the event group is pinned then put it in error state.
793 group_sched_out(leader, cpuctx, ctx);
794 if (leader->attr.pinned) {
795 update_group_times(leader);
796 leader->state = PERF_EVENT_STATE_ERROR;
800 if (!err && !ctx->task && cpuctx->max_pertask)
801 cpuctx->max_pertask--;
806 raw_spin_unlock(&ctx->lock);
810 * Attach a performance event to a context
812 * First we add the event to the list with the hardware enable bit
813 * in event->hw_config cleared.
815 * If the event is attached to a task which is on a CPU we use a smp
816 * call to enable it in the task context. The task might have been
817 * scheduled away, but we check this in the smp call again.
819 * Must be called with ctx->mutex held.
822 perf_install_in_context(struct perf_event_context *ctx,
823 struct perf_event *event,
826 struct task_struct *task = ctx->task;
830 * Per cpu events are installed via an smp call and
831 * the install is always successful.
833 smp_call_function_single(cpu, __perf_install_in_context,
839 task_oncpu_function_call(task, __perf_install_in_context,
842 raw_spin_lock_irq(&ctx->lock);
844 * we need to retry the smp call.
846 if (ctx->is_active && list_empty(&event->group_entry)) {
847 raw_spin_unlock_irq(&ctx->lock);
852 * The lock prevents that this context is scheduled in so we
853 * can add the event safely, if it the call above did not
856 if (list_empty(&event->group_entry))
857 add_event_to_ctx(event, ctx);
858 raw_spin_unlock_irq(&ctx->lock);
862 * Put a event into inactive state and update time fields.
863 * Enabling the leader of a group effectively enables all
864 * the group members that aren't explicitly disabled, so we
865 * have to update their ->tstamp_enabled also.
866 * Note: this works for group members as well as group leaders
867 * since the non-leader members' sibling_lists will be empty.
869 static void __perf_event_mark_enabled(struct perf_event *event,
870 struct perf_event_context *ctx)
872 struct perf_event *sub;
874 event->state = PERF_EVENT_STATE_INACTIVE;
875 event->tstamp_enabled = ctx->time - event->total_time_enabled;
876 list_for_each_entry(sub, &event->sibling_list, group_entry)
877 if (sub->state >= PERF_EVENT_STATE_INACTIVE)
878 sub->tstamp_enabled =
879 ctx->time - sub->total_time_enabled;
883 * Cross CPU call to enable a performance event
885 static void __perf_event_enable(void *info)
887 struct perf_event *event = info;
888 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
889 struct perf_event_context *ctx = event->ctx;
890 struct perf_event *leader = event->group_leader;
894 * If this is a per-task event, need to check whether this
895 * event's task is the current task on this cpu.
897 if (ctx->task && cpuctx->task_ctx != ctx) {
898 if (cpuctx->task_ctx || ctx->task != current)
900 cpuctx->task_ctx = ctx;
903 raw_spin_lock(&ctx->lock);
905 update_context_time(ctx);
907 if (event->state >= PERF_EVENT_STATE_INACTIVE)
909 __perf_event_mark_enabled(event, ctx);
911 if (event->cpu != -1 && event->cpu != smp_processor_id())
915 * If the event is in a group and isn't the group leader,
916 * then don't put it on unless the group is on.
918 if (leader != event && leader->state != PERF_EVENT_STATE_ACTIVE)
921 if (!group_can_go_on(event, cpuctx, 1)) {
926 err = group_sched_in(event, cpuctx, ctx);
928 err = event_sched_in(event, cpuctx, ctx);
934 * If this event can't go on and it's part of a
935 * group, then the whole group has to come off.
938 group_sched_out(leader, cpuctx, ctx);
939 if (leader->attr.pinned) {
940 update_group_times(leader);
941 leader->state = PERF_EVENT_STATE_ERROR;
946 raw_spin_unlock(&ctx->lock);
952 * If event->ctx is a cloned context, callers must make sure that
953 * every task struct that event->ctx->task could possibly point to
954 * remains valid. This condition is satisfied when called through
955 * perf_event_for_each_child or perf_event_for_each as described
956 * for perf_event_disable.
958 void perf_event_enable(struct perf_event *event)
960 struct perf_event_context *ctx = event->ctx;
961 struct task_struct *task = ctx->task;
965 * Enable the event on the cpu that it's on
967 smp_call_function_single(event->cpu, __perf_event_enable,
972 raw_spin_lock_irq(&ctx->lock);
973 if (event->state >= PERF_EVENT_STATE_INACTIVE)
977 * If the event is in error state, clear that first.
978 * That way, if we see the event in error state below, we
979 * know that it has gone back into error state, as distinct
980 * from the task having been scheduled away before the
981 * cross-call arrived.
983 if (event->state == PERF_EVENT_STATE_ERROR)
984 event->state = PERF_EVENT_STATE_OFF;
987 raw_spin_unlock_irq(&ctx->lock);
988 task_oncpu_function_call(task, __perf_event_enable, event);
990 raw_spin_lock_irq(&ctx->lock);
993 * If the context is active and the event is still off,
994 * we need to retry the cross-call.
996 if (ctx->is_active && event->state == PERF_EVENT_STATE_OFF)
1000 * Since we have the lock this context can't be scheduled
1001 * in, so we can change the state safely.
1003 if (event->state == PERF_EVENT_STATE_OFF)
1004 __perf_event_mark_enabled(event, ctx);
1007 raw_spin_unlock_irq(&ctx->lock);
1010 static int perf_event_refresh(struct perf_event *event, int refresh)
1013 * not supported on inherited events
1015 if (event->attr.inherit)
1018 atomic_add(refresh, &event->event_limit);
1019 perf_event_enable(event);
1025 EVENT_FLEXIBLE = 0x1,
1027 EVENT_ALL = EVENT_FLEXIBLE | EVENT_PINNED,
1030 static void ctx_sched_out(struct perf_event_context *ctx,
1031 struct perf_cpu_context *cpuctx,
1032 enum event_type_t event_type)
1034 struct perf_event *event;
1036 raw_spin_lock(&ctx->lock);
1038 if (likely(!ctx->nr_events))
1040 update_context_time(ctx);
1043 if (!ctx->nr_active)
1046 if (event_type & EVENT_PINNED)
1047 list_for_each_entry(event, &ctx->pinned_groups, group_entry)
1048 group_sched_out(event, cpuctx, ctx);
1050 if (event_type & EVENT_FLEXIBLE)
1051 list_for_each_entry(event, &ctx->flexible_groups, group_entry)
1052 group_sched_out(event, cpuctx, ctx);
1057 raw_spin_unlock(&ctx->lock);
1061 * Test whether two contexts are equivalent, i.e. whether they
1062 * have both been cloned from the same version of the same context
1063 * and they both have the same number of enabled events.
1064 * If the number of enabled events is the same, then the set
1065 * of enabled events should be the same, because these are both
1066 * inherited contexts, therefore we can't access individual events
1067 * in them directly with an fd; we can only enable/disable all
1068 * events via prctl, or enable/disable all events in a family
1069 * via ioctl, which will have the same effect on both contexts.
1071 static int context_equiv(struct perf_event_context *ctx1,
1072 struct perf_event_context *ctx2)
1074 return ctx1->parent_ctx && ctx1->parent_ctx == ctx2->parent_ctx
1075 && ctx1->parent_gen == ctx2->parent_gen
1076 && !ctx1->pin_count && !ctx2->pin_count;
1079 static void __perf_event_sync_stat(struct perf_event *event,
1080 struct perf_event *next_event)
1084 if (!event->attr.inherit_stat)
1088 * Update the event value, we cannot use perf_event_read()
1089 * because we're in the middle of a context switch and have IRQs
1090 * disabled, which upsets smp_call_function_single(), however
1091 * we know the event must be on the current CPU, therefore we
1092 * don't need to use it.
1094 switch (event->state) {
1095 case PERF_EVENT_STATE_ACTIVE:
1096 event->pmu->read(event);
1099 case PERF_EVENT_STATE_INACTIVE:
1100 update_event_times(event);
1108 * In order to keep per-task stats reliable we need to flip the event
1109 * values when we flip the contexts.
1111 value = atomic64_read(&next_event->count);
1112 value = atomic64_xchg(&event->count, value);
1113 atomic64_set(&next_event->count, value);
1115 swap(event->total_time_enabled, next_event->total_time_enabled);
1116 swap(event->total_time_running, next_event->total_time_running);
1119 * Since we swizzled the values, update the user visible data too.
1121 perf_event_update_userpage(event);
1122 perf_event_update_userpage(next_event);
1125 #define list_next_entry(pos, member) \
1126 list_entry(pos->member.next, typeof(*pos), member)
1128 static void perf_event_sync_stat(struct perf_event_context *ctx,
1129 struct perf_event_context *next_ctx)
1131 struct perf_event *event, *next_event;
1136 update_context_time(ctx);
1138 event = list_first_entry(&ctx->event_list,
1139 struct perf_event, event_entry);
1141 next_event = list_first_entry(&next_ctx->event_list,
1142 struct perf_event, event_entry);
1144 while (&event->event_entry != &ctx->event_list &&
1145 &next_event->event_entry != &next_ctx->event_list) {
1147 __perf_event_sync_stat(event, next_event);
1149 event = list_next_entry(event, event_entry);
1150 next_event = list_next_entry(next_event, event_entry);
1155 * Called from scheduler to remove the events of the current task,
1156 * with interrupts disabled.
1158 * We stop each event and update the event value in event->count.
1160 * This does not protect us against NMI, but disable()
1161 * sets the disabled bit in the control field of event _before_
1162 * accessing the event control register. If a NMI hits, then it will
1163 * not restart the event.
1165 void perf_event_task_sched_out(struct task_struct *task,
1166 struct task_struct *next)
1168 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
1169 struct perf_event_context *ctx = task->perf_event_ctxp;
1170 struct perf_event_context *next_ctx;
1171 struct perf_event_context *parent;
1174 perf_sw_event(PERF_COUNT_SW_CONTEXT_SWITCHES, 1, 1, NULL, 0);
1176 if (likely(!ctx || !cpuctx->task_ctx))
1180 parent = rcu_dereference(ctx->parent_ctx);
1181 next_ctx = next->perf_event_ctxp;
1182 if (parent && next_ctx &&
1183 rcu_dereference(next_ctx->parent_ctx) == parent) {
1185 * Looks like the two contexts are clones, so we might be
1186 * able to optimize the context switch. We lock both
1187 * contexts and check that they are clones under the
1188 * lock (including re-checking that neither has been
1189 * uncloned in the meantime). It doesn't matter which
1190 * order we take the locks because no other cpu could
1191 * be trying to lock both of these tasks.
1193 raw_spin_lock(&ctx->lock);
1194 raw_spin_lock_nested(&next_ctx->lock, SINGLE_DEPTH_NESTING);
1195 if (context_equiv(ctx, next_ctx)) {
1197 * XXX do we need a memory barrier of sorts
1198 * wrt to rcu_dereference() of perf_event_ctxp
1200 task->perf_event_ctxp = next_ctx;
1201 next->perf_event_ctxp = ctx;
1203 next_ctx->task = task;
1206 perf_event_sync_stat(ctx, next_ctx);
1208 raw_spin_unlock(&next_ctx->lock);
1209 raw_spin_unlock(&ctx->lock);
1214 ctx_sched_out(ctx, cpuctx, EVENT_ALL);
1215 cpuctx->task_ctx = NULL;
1219 static void task_ctx_sched_out(struct perf_event_context *ctx,
1220 enum event_type_t event_type)
1222 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
1224 if (!cpuctx->task_ctx)
1227 if (WARN_ON_ONCE(ctx != cpuctx->task_ctx))
1230 ctx_sched_out(ctx, cpuctx, event_type);
1231 cpuctx->task_ctx = NULL;
1235 * Called with IRQs disabled
1237 static void __perf_event_task_sched_out(struct perf_event_context *ctx)
1239 task_ctx_sched_out(ctx, EVENT_ALL);
1243 * Called with IRQs disabled
1245 static void cpu_ctx_sched_out(struct perf_cpu_context *cpuctx,
1246 enum event_type_t event_type)
1248 ctx_sched_out(&cpuctx->ctx, cpuctx, event_type);
1252 ctx_pinned_sched_in(struct perf_event_context *ctx,
1253 struct perf_cpu_context *cpuctx)
1255 struct perf_event *event;
1257 list_for_each_entry(event, &ctx->pinned_groups, group_entry) {
1258 if (event->state <= PERF_EVENT_STATE_OFF)
1260 if (event->cpu != -1 && event->cpu != smp_processor_id())
1263 if (group_can_go_on(event, cpuctx, 1))
1264 group_sched_in(event, cpuctx, ctx);
1267 * If this pinned group hasn't been scheduled,
1268 * put it in error state.
1270 if (event->state == PERF_EVENT_STATE_INACTIVE) {
1271 update_group_times(event);
1272 event->state = PERF_EVENT_STATE_ERROR;
1278 ctx_flexible_sched_in(struct perf_event_context *ctx,
1279 struct perf_cpu_context *cpuctx)
1281 struct perf_event *event;
1284 list_for_each_entry(event, &ctx->flexible_groups, group_entry) {
1285 /* Ignore events in OFF or ERROR state */
1286 if (event->state <= PERF_EVENT_STATE_OFF)
1289 * Listen to the 'cpu' scheduling filter constraint
1292 if (event->cpu != -1 && event->cpu != smp_processor_id())
1295 if (group_can_go_on(event, cpuctx, can_add_hw))
1296 if (group_sched_in(event, cpuctx, ctx))
1302 ctx_sched_in(struct perf_event_context *ctx,
1303 struct perf_cpu_context *cpuctx,
1304 enum event_type_t event_type)
1306 raw_spin_lock(&ctx->lock);
1308 if (likely(!ctx->nr_events))
1311 ctx->timestamp = perf_clock();
1316 * First go through the list and put on any pinned groups
1317 * in order to give them the best chance of going on.
1319 if (event_type & EVENT_PINNED)
1320 ctx_pinned_sched_in(ctx, cpuctx);
1322 /* Then walk through the lower prio flexible groups */
1323 if (event_type & EVENT_FLEXIBLE)
1324 ctx_flexible_sched_in(ctx, cpuctx);
1328 raw_spin_unlock(&ctx->lock);
1331 static void cpu_ctx_sched_in(struct perf_cpu_context *cpuctx,
1332 enum event_type_t event_type)
1334 struct perf_event_context *ctx = &cpuctx->ctx;
1336 ctx_sched_in(ctx, cpuctx, event_type);
1339 static void task_ctx_sched_in(struct task_struct *task,
1340 enum event_type_t event_type)
1342 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
1343 struct perf_event_context *ctx = task->perf_event_ctxp;
1347 if (cpuctx->task_ctx == ctx)
1349 ctx_sched_in(ctx, cpuctx, event_type);
1350 cpuctx->task_ctx = ctx;
1353 * Called from scheduler to add the events of the current task
1354 * with interrupts disabled.
1356 * We restore the event value and then enable it.
1358 * This does not protect us against NMI, but enable()
1359 * sets the enabled bit in the control field of event _before_
1360 * accessing the event control register. If a NMI hits, then it will
1361 * keep the event running.
1363 void perf_event_task_sched_in(struct task_struct *task)
1365 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
1366 struct perf_event_context *ctx = task->perf_event_ctxp;
1371 if (cpuctx->task_ctx == ctx)
1377 * We want to keep the following priority order:
1378 * cpu pinned (that don't need to move), task pinned,
1379 * cpu flexible, task flexible.
1381 cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE);
1383 ctx_sched_in(ctx, cpuctx, EVENT_PINNED);
1384 cpu_ctx_sched_in(cpuctx, EVENT_FLEXIBLE);
1385 ctx_sched_in(ctx, cpuctx, EVENT_FLEXIBLE);
1387 cpuctx->task_ctx = ctx;
1392 #define MAX_INTERRUPTS (~0ULL)
1394 static void perf_log_throttle(struct perf_event *event, int enable);
1396 static u64 perf_calculate_period(struct perf_event *event, u64 nsec, u64 count)
1398 u64 frequency = event->attr.sample_freq;
1399 u64 sec = NSEC_PER_SEC;
1400 u64 divisor, dividend;
1402 int count_fls, nsec_fls, frequency_fls, sec_fls;
1404 count_fls = fls64(count);
1405 nsec_fls = fls64(nsec);
1406 frequency_fls = fls64(frequency);
1410 * We got @count in @nsec, with a target of sample_freq HZ
1411 * the target period becomes:
1414 * period = -------------------
1415 * @nsec * sample_freq
1420 * Reduce accuracy by one bit such that @a and @b converge
1421 * to a similar magnitude.
1423 #define REDUCE_FLS(a, b) \
1425 if (a##_fls > b##_fls) { \
1435 * Reduce accuracy until either term fits in a u64, then proceed with
1436 * the other, so that finally we can do a u64/u64 division.
1438 while (count_fls + sec_fls > 64 && nsec_fls + frequency_fls > 64) {
1439 REDUCE_FLS(nsec, frequency);
1440 REDUCE_FLS(sec, count);
1443 if (count_fls + sec_fls > 64) {
1444 divisor = nsec * frequency;
1446 while (count_fls + sec_fls > 64) {
1447 REDUCE_FLS(count, sec);
1451 dividend = count * sec;
1453 dividend = count * sec;
1455 while (nsec_fls + frequency_fls > 64) {
1456 REDUCE_FLS(nsec, frequency);
1460 divisor = nsec * frequency;
1463 return div64_u64(dividend, divisor);
1466 static void perf_event_stop(struct perf_event *event)
1468 if (!event->pmu->stop)
1469 return event->pmu->disable(event);
1471 return event->pmu->stop(event);
1474 static int perf_event_start(struct perf_event *event)
1476 if (!event->pmu->start)
1477 return event->pmu->enable(event);
1479 return event->pmu->start(event);
1482 static void perf_adjust_period(struct perf_event *event, u64 nsec, u64 count)
1484 struct hw_perf_event *hwc = &event->hw;
1485 u64 period, sample_period;
1488 period = perf_calculate_period(event, nsec, count);
1490 delta = (s64)(period - hwc->sample_period);
1491 delta = (delta + 7) / 8; /* low pass filter */
1493 sample_period = hwc->sample_period + delta;
1498 hwc->sample_period = sample_period;
1500 if (atomic64_read(&hwc->period_left) > 8*sample_period) {
1502 perf_event_stop(event);
1503 atomic64_set(&hwc->period_left, 0);
1504 perf_event_start(event);
1509 static void perf_ctx_adjust_freq(struct perf_event_context *ctx)
1511 struct perf_event *event;
1512 struct hw_perf_event *hwc;
1513 u64 interrupts, now;
1516 raw_spin_lock(&ctx->lock);
1517 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
1518 if (event->state != PERF_EVENT_STATE_ACTIVE)
1521 if (event->cpu != -1 && event->cpu != smp_processor_id())
1526 interrupts = hwc->interrupts;
1527 hwc->interrupts = 0;
1530 * unthrottle events on the tick
1532 if (interrupts == MAX_INTERRUPTS) {
1533 perf_log_throttle(event, 1);
1535 event->pmu->unthrottle(event);
1539 if (!event->attr.freq || !event->attr.sample_freq)
1543 event->pmu->read(event);
1544 now = atomic64_read(&event->count);
1545 delta = now - hwc->freq_count_stamp;
1546 hwc->freq_count_stamp = now;
1549 perf_adjust_period(event, TICK_NSEC, delta);
1552 raw_spin_unlock(&ctx->lock);
1556 * Round-robin a context's events:
1558 static void rotate_ctx(struct perf_event_context *ctx)
1560 raw_spin_lock(&ctx->lock);
1562 /* Rotate the first entry last of non-pinned groups */
1563 list_rotate_left(&ctx->flexible_groups);
1565 raw_spin_unlock(&ctx->lock);
1568 void perf_event_task_tick(struct task_struct *curr)
1570 struct perf_cpu_context *cpuctx;
1571 struct perf_event_context *ctx;
1574 if (!atomic_read(&nr_events))
1577 cpuctx = &__get_cpu_var(perf_cpu_context);
1578 if (cpuctx->ctx.nr_events &&
1579 cpuctx->ctx.nr_events != cpuctx->ctx.nr_active)
1582 ctx = curr->perf_event_ctxp;
1583 if (ctx && ctx->nr_events && ctx->nr_events != ctx->nr_active)
1586 perf_ctx_adjust_freq(&cpuctx->ctx);
1588 perf_ctx_adjust_freq(ctx);
1594 cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE);
1596 task_ctx_sched_out(ctx, EVENT_FLEXIBLE);
1598 rotate_ctx(&cpuctx->ctx);
1602 cpu_ctx_sched_in(cpuctx, EVENT_FLEXIBLE);
1604 task_ctx_sched_in(curr, EVENT_FLEXIBLE);
1608 static int event_enable_on_exec(struct perf_event *event,
1609 struct perf_event_context *ctx)
1611 if (!event->attr.enable_on_exec)
1614 event->attr.enable_on_exec = 0;
1615 if (event->state >= PERF_EVENT_STATE_INACTIVE)
1618 __perf_event_mark_enabled(event, ctx);
1624 * Enable all of a task's events that have been marked enable-on-exec.
1625 * This expects task == current.
1627 static void perf_event_enable_on_exec(struct task_struct *task)
1629 struct perf_event_context *ctx;
1630 struct perf_event *event;
1631 unsigned long flags;
1635 local_irq_save(flags);
1636 ctx = task->perf_event_ctxp;
1637 if (!ctx || !ctx->nr_events)
1640 __perf_event_task_sched_out(ctx);
1642 raw_spin_lock(&ctx->lock);
1644 list_for_each_entry(event, &ctx->pinned_groups, group_entry) {
1645 ret = event_enable_on_exec(event, ctx);
1650 list_for_each_entry(event, &ctx->flexible_groups, group_entry) {
1651 ret = event_enable_on_exec(event, ctx);
1657 * Unclone this context if we enabled any event.
1662 raw_spin_unlock(&ctx->lock);
1664 perf_event_task_sched_in(task);
1666 local_irq_restore(flags);
1670 * Cross CPU call to read the hardware event
1672 static void __perf_event_read(void *info)
1674 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
1675 struct perf_event *event = info;
1676 struct perf_event_context *ctx = event->ctx;
1679 * If this is a task context, we need to check whether it is
1680 * the current task context of this cpu. If not it has been
1681 * scheduled out before the smp call arrived. In that case
1682 * event->count would have been updated to a recent sample
1683 * when the event was scheduled out.
1685 if (ctx->task && cpuctx->task_ctx != ctx)
1688 raw_spin_lock(&ctx->lock);
1689 update_context_time(ctx);
1690 update_event_times(event);
1691 raw_spin_unlock(&ctx->lock);
1693 event->pmu->read(event);
1696 static u64 perf_event_read(struct perf_event *event)
1699 * If event is enabled and currently active on a CPU, update the
1700 * value in the event structure:
1702 if (event->state == PERF_EVENT_STATE_ACTIVE) {
1703 smp_call_function_single(event->oncpu,
1704 __perf_event_read, event, 1);
1705 } else if (event->state == PERF_EVENT_STATE_INACTIVE) {
1706 struct perf_event_context *ctx = event->ctx;
1707 unsigned long flags;
1709 raw_spin_lock_irqsave(&ctx->lock, flags);
1710 update_context_time(ctx);
1711 update_event_times(event);
1712 raw_spin_unlock_irqrestore(&ctx->lock, flags);
1715 return atomic64_read(&event->count);
1719 * Initialize the perf_event context in a task_struct:
1722 __perf_event_init_context(struct perf_event_context *ctx,
1723 struct task_struct *task)
1725 raw_spin_lock_init(&ctx->lock);
1726 mutex_init(&ctx->mutex);
1727 INIT_LIST_HEAD(&ctx->pinned_groups);
1728 INIT_LIST_HEAD(&ctx->flexible_groups);
1729 INIT_LIST_HEAD(&ctx->event_list);
1730 atomic_set(&ctx->refcount, 1);
1734 static struct perf_event_context *find_get_context(pid_t pid, int cpu)
1736 struct perf_event_context *ctx;
1737 struct perf_cpu_context *cpuctx;
1738 struct task_struct *task;
1739 unsigned long flags;
1742 if (pid == -1 && cpu != -1) {
1743 /* Must be root to operate on a CPU event: */
1744 if (perf_paranoid_cpu() && !capable(CAP_SYS_ADMIN))
1745 return ERR_PTR(-EACCES);
1747 if (cpu < 0 || cpu >= nr_cpumask_bits)
1748 return ERR_PTR(-EINVAL);
1751 * We could be clever and allow to attach a event to an
1752 * offline CPU and activate it when the CPU comes up, but
1755 if (!cpu_online(cpu))
1756 return ERR_PTR(-ENODEV);
1758 cpuctx = &per_cpu(perf_cpu_context, cpu);
1769 task = find_task_by_vpid(pid);
1771 get_task_struct(task);
1775 return ERR_PTR(-ESRCH);
1778 * Can't attach events to a dying task.
1781 if (task->flags & PF_EXITING)
1784 /* Reuse ptrace permission checks for now. */
1786 if (!ptrace_may_access(task, PTRACE_MODE_READ))
1790 ctx = perf_lock_task_context(task, &flags);
1793 raw_spin_unlock_irqrestore(&ctx->lock, flags);
1797 ctx = kzalloc(sizeof(struct perf_event_context), GFP_KERNEL);
1801 __perf_event_init_context(ctx, task);
1803 if (cmpxchg(&task->perf_event_ctxp, NULL, ctx)) {
1805 * We raced with some other task; use
1806 * the context they set.
1811 get_task_struct(task);
1814 put_task_struct(task);
1818 put_task_struct(task);
1819 return ERR_PTR(err);
1822 static void perf_event_free_filter(struct perf_event *event);
1824 static void free_event_rcu(struct rcu_head *head)
1826 struct perf_event *event;
1828 event = container_of(head, struct perf_event, rcu_head);
1830 put_pid_ns(event->ns);
1831 perf_event_free_filter(event);
1835 static void perf_pending_sync(struct perf_event *event);
1837 static void free_event(struct perf_event *event)
1839 perf_pending_sync(event);
1841 if (!event->parent) {
1842 atomic_dec(&nr_events);
1843 if (event->attr.mmap)
1844 atomic_dec(&nr_mmap_events);
1845 if (event->attr.comm)
1846 atomic_dec(&nr_comm_events);
1847 if (event->attr.task)
1848 atomic_dec(&nr_task_events);
1851 if (event->output) {
1852 fput(event->output->filp);
1853 event->output = NULL;
1857 event->destroy(event);
1859 put_ctx(event->ctx);
1860 call_rcu(&event->rcu_head, free_event_rcu);
1863 int perf_event_release_kernel(struct perf_event *event)
1865 struct perf_event_context *ctx = event->ctx;
1867 event->state = PERF_EVENT_STATE_FREE;
1869 WARN_ON_ONCE(ctx->parent_ctx);
1871 * There are two ways this annotation is useful:
1873 * 1) there is a lock recursion from perf_event_exit_task
1874 * see the comment there.
1876 * 2) there is a lock-inversion with mmap_sem through
1877 * perf_event_read_group(), which takes faults while
1878 * holding ctx->mutex, however this is called after
1879 * the last filedesc died, so there is no possibility
1880 * to trigger the AB-BA case.
1882 mutex_lock_nested(&ctx->mutex, SINGLE_DEPTH_NESTING);
1883 perf_event_remove_from_context(event);
1884 mutex_unlock(&ctx->mutex);
1886 mutex_lock(&event->owner->perf_event_mutex);
1887 list_del_init(&event->owner_entry);
1888 mutex_unlock(&event->owner->perf_event_mutex);
1889 put_task_struct(event->owner);
1895 EXPORT_SYMBOL_GPL(perf_event_release_kernel);
1898 * Called when the last reference to the file is gone.
1900 static int perf_release(struct inode *inode, struct file *file)
1902 struct perf_event *event = file->private_data;
1904 file->private_data = NULL;
1906 return perf_event_release_kernel(event);
1909 static int perf_event_read_size(struct perf_event *event)
1911 int entry = sizeof(u64); /* value */
1915 if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1916 size += sizeof(u64);
1918 if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1919 size += sizeof(u64);
1921 if (event->attr.read_format & PERF_FORMAT_ID)
1922 entry += sizeof(u64);
1924 if (event->attr.read_format & PERF_FORMAT_GROUP) {
1925 nr += event->group_leader->nr_siblings;
1926 size += sizeof(u64);
1934 u64 perf_event_read_value(struct perf_event *event, u64 *enabled, u64 *running)
1936 struct perf_event *child;
1942 mutex_lock(&event->child_mutex);
1943 total += perf_event_read(event);
1944 *enabled += event->total_time_enabled +
1945 atomic64_read(&event->child_total_time_enabled);
1946 *running += event->total_time_running +
1947 atomic64_read(&event->child_total_time_running);
1949 list_for_each_entry(child, &event->child_list, child_list) {
1950 total += perf_event_read(child);
1951 *enabled += child->total_time_enabled;
1952 *running += child->total_time_running;
1954 mutex_unlock(&event->child_mutex);
1958 EXPORT_SYMBOL_GPL(perf_event_read_value);
1960 static int perf_event_read_group(struct perf_event *event,
1961 u64 read_format, char __user *buf)
1963 struct perf_event *leader = event->group_leader, *sub;
1964 int n = 0, size = 0, ret = -EFAULT;
1965 struct perf_event_context *ctx = leader->ctx;
1967 u64 count, enabled, running;
1969 mutex_lock(&ctx->mutex);
1970 count = perf_event_read_value(leader, &enabled, &running);
1972 values[n++] = 1 + leader->nr_siblings;
1973 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1974 values[n++] = enabled;
1975 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1976 values[n++] = running;
1977 values[n++] = count;
1978 if (read_format & PERF_FORMAT_ID)
1979 values[n++] = primary_event_id(leader);
1981 size = n * sizeof(u64);
1983 if (copy_to_user(buf, values, size))
1988 list_for_each_entry(sub, &leader->sibling_list, group_entry) {
1991 values[n++] = perf_event_read_value(sub, &enabled, &running);
1992 if (read_format & PERF_FORMAT_ID)
1993 values[n++] = primary_event_id(sub);
1995 size = n * sizeof(u64);
1997 if (copy_to_user(buf + ret, values, size)) {
2005 mutex_unlock(&ctx->mutex);
2010 static int perf_event_read_one(struct perf_event *event,
2011 u64 read_format, char __user *buf)
2013 u64 enabled, running;
2017 values[n++] = perf_event_read_value(event, &enabled, &running);
2018 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
2019 values[n++] = enabled;
2020 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
2021 values[n++] = running;
2022 if (read_format & PERF_FORMAT_ID)
2023 values[n++] = primary_event_id(event);
2025 if (copy_to_user(buf, values, n * sizeof(u64)))
2028 return n * sizeof(u64);
2032 * Read the performance event - simple non blocking version for now
2035 perf_read_hw(struct perf_event *event, char __user *buf, size_t count)
2037 u64 read_format = event->attr.read_format;
2041 * Return end-of-file for a read on a event that is in
2042 * error state (i.e. because it was pinned but it couldn't be
2043 * scheduled on to the CPU at some point).
2045 if (event->state == PERF_EVENT_STATE_ERROR)
2048 if (count < perf_event_read_size(event))
2051 WARN_ON_ONCE(event->ctx->parent_ctx);
2052 if (read_format & PERF_FORMAT_GROUP)
2053 ret = perf_event_read_group(event, read_format, buf);
2055 ret = perf_event_read_one(event, read_format, buf);
2061 perf_read(struct file *file, char __user *buf, size_t count, loff_t *ppos)
2063 struct perf_event *event = file->private_data;
2065 return perf_read_hw(event, buf, count);
2068 static unsigned int perf_poll(struct file *file, poll_table *wait)
2070 struct perf_event *event = file->private_data;
2071 struct perf_mmap_data *data;
2072 unsigned int events = POLL_HUP;
2075 data = rcu_dereference(event->data);
2077 events = atomic_xchg(&data->poll, 0);
2080 poll_wait(file, &event->waitq, wait);
2085 static void perf_event_reset(struct perf_event *event)
2087 (void)perf_event_read(event);
2088 atomic64_set(&event->count, 0);
2089 perf_event_update_userpage(event);
2093 * Holding the top-level event's child_mutex means that any
2094 * descendant process that has inherited this event will block
2095 * in sync_child_event if it goes to exit, thus satisfying the
2096 * task existence requirements of perf_event_enable/disable.
2098 static void perf_event_for_each_child(struct perf_event *event,
2099 void (*func)(struct perf_event *))
2101 struct perf_event *child;
2103 WARN_ON_ONCE(event->ctx->parent_ctx);
2104 mutex_lock(&event->child_mutex);
2106 list_for_each_entry(child, &event->child_list, child_list)
2108 mutex_unlock(&event->child_mutex);
2111 static void perf_event_for_each(struct perf_event *event,
2112 void (*func)(struct perf_event *))
2114 struct perf_event_context *ctx = event->ctx;
2115 struct perf_event *sibling;
2117 WARN_ON_ONCE(ctx->parent_ctx);
2118 mutex_lock(&ctx->mutex);
2119 event = event->group_leader;
2121 perf_event_for_each_child(event, func);
2123 list_for_each_entry(sibling, &event->sibling_list, group_entry)
2124 perf_event_for_each_child(event, func);
2125 mutex_unlock(&ctx->mutex);
2128 static int perf_event_period(struct perf_event *event, u64 __user *arg)
2130 struct perf_event_context *ctx = event->ctx;
2135 if (!event->attr.sample_period)
2138 size = copy_from_user(&value, arg, sizeof(value));
2139 if (size != sizeof(value))
2145 raw_spin_lock_irq(&ctx->lock);
2146 if (event->attr.freq) {
2147 if (value > sysctl_perf_event_sample_rate) {
2152 event->attr.sample_freq = value;
2154 event->attr.sample_period = value;
2155 event->hw.sample_period = value;
2158 raw_spin_unlock_irq(&ctx->lock);
2163 static int perf_event_set_output(struct perf_event *event, int output_fd);
2164 static int perf_event_set_filter(struct perf_event *event, void __user *arg);
2166 static long perf_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
2168 struct perf_event *event = file->private_data;
2169 void (*func)(struct perf_event *);
2173 case PERF_EVENT_IOC_ENABLE:
2174 func = perf_event_enable;
2176 case PERF_EVENT_IOC_DISABLE:
2177 func = perf_event_disable;
2179 case PERF_EVENT_IOC_RESET:
2180 func = perf_event_reset;
2183 case PERF_EVENT_IOC_REFRESH:
2184 return perf_event_refresh(event, arg);
2186 case PERF_EVENT_IOC_PERIOD:
2187 return perf_event_period(event, (u64 __user *)arg);
2189 case PERF_EVENT_IOC_SET_OUTPUT:
2190 return perf_event_set_output(event, arg);
2192 case PERF_EVENT_IOC_SET_FILTER:
2193 return perf_event_set_filter(event, (void __user *)arg);
2199 if (flags & PERF_IOC_FLAG_GROUP)
2200 perf_event_for_each(event, func);
2202 perf_event_for_each_child(event, func);
2207 int perf_event_task_enable(void)
2209 struct perf_event *event;
2211 mutex_lock(¤t->perf_event_mutex);
2212 list_for_each_entry(event, ¤t->perf_event_list, owner_entry)
2213 perf_event_for_each_child(event, perf_event_enable);
2214 mutex_unlock(¤t->perf_event_mutex);
2219 int perf_event_task_disable(void)
2221 struct perf_event *event;
2223 mutex_lock(¤t->perf_event_mutex);
2224 list_for_each_entry(event, ¤t->perf_event_list, owner_entry)
2225 perf_event_for_each_child(event, perf_event_disable);
2226 mutex_unlock(¤t->perf_event_mutex);
2231 #ifndef PERF_EVENT_INDEX_OFFSET
2232 # define PERF_EVENT_INDEX_OFFSET 0
2235 static int perf_event_index(struct perf_event *event)
2237 if (event->state != PERF_EVENT_STATE_ACTIVE)
2240 return event->hw.idx + 1 - PERF_EVENT_INDEX_OFFSET;
2244 * Callers need to ensure there can be no nesting of this function, otherwise
2245 * the seqlock logic goes bad. We can not serialize this because the arch
2246 * code calls this from NMI context.
2248 void perf_event_update_userpage(struct perf_event *event)
2250 struct perf_event_mmap_page *userpg;
2251 struct perf_mmap_data *data;
2254 data = rcu_dereference(event->data);
2258 userpg = data->user_page;
2261 * Disable preemption so as to not let the corresponding user-space
2262 * spin too long if we get preempted.
2267 userpg->index = perf_event_index(event);
2268 userpg->offset = atomic64_read(&event->count);
2269 if (event->state == PERF_EVENT_STATE_ACTIVE)
2270 userpg->offset -= atomic64_read(&event->hw.prev_count);
2272 userpg->time_enabled = event->total_time_enabled +
2273 atomic64_read(&event->child_total_time_enabled);
2275 userpg->time_running = event->total_time_running +
2276 atomic64_read(&event->child_total_time_running);
2285 static unsigned long perf_data_size(struct perf_mmap_data *data)
2287 return data->nr_pages << (PAGE_SHIFT + data->data_order);
2290 #ifndef CONFIG_PERF_USE_VMALLOC
2293 * Back perf_mmap() with regular GFP_KERNEL-0 pages.
2296 static struct page *
2297 perf_mmap_to_page(struct perf_mmap_data *data, unsigned long pgoff)
2299 if (pgoff > data->nr_pages)
2303 return virt_to_page(data->user_page);
2305 return virt_to_page(data->data_pages[pgoff - 1]);
2308 static struct perf_mmap_data *
2309 perf_mmap_data_alloc(struct perf_event *event, int nr_pages)
2311 struct perf_mmap_data *data;
2315 WARN_ON(atomic_read(&event->mmap_count));
2317 size = sizeof(struct perf_mmap_data);
2318 size += nr_pages * sizeof(void *);
2320 data = kzalloc(size, GFP_KERNEL);
2324 data->user_page = (void *)get_zeroed_page(GFP_KERNEL);
2325 if (!data->user_page)
2326 goto fail_user_page;
2328 for (i = 0; i < nr_pages; i++) {
2329 data->data_pages[i] = (void *)get_zeroed_page(GFP_KERNEL);
2330 if (!data->data_pages[i])
2331 goto fail_data_pages;
2334 data->data_order = 0;
2335 data->nr_pages = nr_pages;
2340 for (i--; i >= 0; i--)
2341 free_page((unsigned long)data->data_pages[i]);
2343 free_page((unsigned long)data->user_page);
2352 static void perf_mmap_free_page(unsigned long addr)
2354 struct page *page = virt_to_page((void *)addr);
2356 page->mapping = NULL;
2360 static void perf_mmap_data_free(struct perf_mmap_data *data)
2364 perf_mmap_free_page((unsigned long)data->user_page);
2365 for (i = 0; i < data->nr_pages; i++)
2366 perf_mmap_free_page((unsigned long)data->data_pages[i]);
2373 * Back perf_mmap() with vmalloc memory.
2375 * Required for architectures that have d-cache aliasing issues.
2378 static struct page *
2379 perf_mmap_to_page(struct perf_mmap_data *data, unsigned long pgoff)
2381 if (pgoff > (1UL << data->data_order))
2384 return vmalloc_to_page((void *)data->user_page + pgoff * PAGE_SIZE);
2387 static void perf_mmap_unmark_page(void *addr)
2389 struct page *page = vmalloc_to_page(addr);
2391 page->mapping = NULL;
2394 static void perf_mmap_data_free_work(struct work_struct *work)
2396 struct perf_mmap_data *data;
2400 data = container_of(work, struct perf_mmap_data, work);
2401 nr = 1 << data->data_order;
2403 base = data->user_page;
2404 for (i = 0; i < nr + 1; i++)
2405 perf_mmap_unmark_page(base + (i * PAGE_SIZE));
2411 static void perf_mmap_data_free(struct perf_mmap_data *data)
2413 schedule_work(&data->work);
2416 static struct perf_mmap_data *
2417 perf_mmap_data_alloc(struct perf_event *event, int nr_pages)
2419 struct perf_mmap_data *data;
2423 WARN_ON(atomic_read(&event->mmap_count));
2425 size = sizeof(struct perf_mmap_data);
2426 size += sizeof(void *);
2428 data = kzalloc(size, GFP_KERNEL);
2432 INIT_WORK(&data->work, perf_mmap_data_free_work);
2434 all_buf = vmalloc_user((nr_pages + 1) * PAGE_SIZE);
2438 data->user_page = all_buf;
2439 data->data_pages[0] = all_buf + PAGE_SIZE;
2440 data->data_order = ilog2(nr_pages);
2454 static int perf_mmap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
2456 struct perf_event *event = vma->vm_file->private_data;
2457 struct perf_mmap_data *data;
2458 int ret = VM_FAULT_SIGBUS;
2460 if (vmf->flags & FAULT_FLAG_MKWRITE) {
2461 if (vmf->pgoff == 0)
2467 data = rcu_dereference(event->data);
2471 if (vmf->pgoff && (vmf->flags & FAULT_FLAG_WRITE))
2474 vmf->page = perf_mmap_to_page(data, vmf->pgoff);
2478 get_page(vmf->page);
2479 vmf->page->mapping = vma->vm_file->f_mapping;
2480 vmf->page->index = vmf->pgoff;
2490 perf_mmap_data_init(struct perf_event *event, struct perf_mmap_data *data)
2492 long max_size = perf_data_size(data);
2494 atomic_set(&data->lock, -1);
2496 if (event->attr.watermark) {
2497 data->watermark = min_t(long, max_size,
2498 event->attr.wakeup_watermark);
2501 if (!data->watermark)
2502 data->watermark = max_size / 2;
2505 rcu_assign_pointer(event->data, data);
2508 static void perf_mmap_data_free_rcu(struct rcu_head *rcu_head)
2510 struct perf_mmap_data *data;
2512 data = container_of(rcu_head, struct perf_mmap_data, rcu_head);
2513 perf_mmap_data_free(data);
2516 static void perf_mmap_data_release(struct perf_event *event)
2518 struct perf_mmap_data *data = event->data;
2520 WARN_ON(atomic_read(&event->mmap_count));
2522 rcu_assign_pointer(event->data, NULL);
2523 call_rcu(&data->rcu_head, perf_mmap_data_free_rcu);
2526 static void perf_mmap_open(struct vm_area_struct *vma)
2528 struct perf_event *event = vma->vm_file->private_data;
2530 atomic_inc(&event->mmap_count);
2533 static void perf_mmap_close(struct vm_area_struct *vma)
2535 struct perf_event *event = vma->vm_file->private_data;
2537 WARN_ON_ONCE(event->ctx->parent_ctx);
2538 if (atomic_dec_and_mutex_lock(&event->mmap_count, &event->mmap_mutex)) {
2539 unsigned long size = perf_data_size(event->data);
2540 struct user_struct *user = current_user();
2542 atomic_long_sub((size >> PAGE_SHIFT) + 1, &user->locked_vm);
2543 vma->vm_mm->locked_vm -= event->data->nr_locked;
2544 perf_mmap_data_release(event);
2545 mutex_unlock(&event->mmap_mutex);
2549 static const struct vm_operations_struct perf_mmap_vmops = {
2550 .open = perf_mmap_open,
2551 .close = perf_mmap_close,
2552 .fault = perf_mmap_fault,
2553 .page_mkwrite = perf_mmap_fault,
2556 static int perf_mmap(struct file *file, struct vm_area_struct *vma)
2558 struct perf_event *event = file->private_data;
2559 unsigned long user_locked, user_lock_limit;
2560 struct user_struct *user = current_user();
2561 unsigned long locked, lock_limit;
2562 struct perf_mmap_data *data;
2563 unsigned long vma_size;
2564 unsigned long nr_pages;
2565 long user_extra, extra;
2568 if (!(vma->vm_flags & VM_SHARED))
2571 vma_size = vma->vm_end - vma->vm_start;
2572 nr_pages = (vma_size / PAGE_SIZE) - 1;
2575 * If we have data pages ensure they're a power-of-two number, so we
2576 * can do bitmasks instead of modulo.
2578 if (nr_pages != 0 && !is_power_of_2(nr_pages))
2581 if (vma_size != PAGE_SIZE * (1 + nr_pages))
2584 if (vma->vm_pgoff != 0)
2587 WARN_ON_ONCE(event->ctx->parent_ctx);
2588 mutex_lock(&event->mmap_mutex);
2589 if (event->output) {
2594 if (atomic_inc_not_zero(&event->mmap_count)) {
2595 if (nr_pages != event->data->nr_pages)
2600 user_extra = nr_pages + 1;
2601 user_lock_limit = sysctl_perf_event_mlock >> (PAGE_SHIFT - 10);
2604 * Increase the limit linearly with more CPUs:
2606 user_lock_limit *= num_online_cpus();
2608 user_locked = atomic_long_read(&user->locked_vm) + user_extra;
2611 if (user_locked > user_lock_limit)
2612 extra = user_locked - user_lock_limit;
2614 lock_limit = rlimit(RLIMIT_MEMLOCK);
2615 lock_limit >>= PAGE_SHIFT;
2616 locked = vma->vm_mm->locked_vm + extra;
2618 if ((locked > lock_limit) && perf_paranoid_tracepoint_raw() &&
2619 !capable(CAP_IPC_LOCK)) {
2624 WARN_ON(event->data);
2626 data = perf_mmap_data_alloc(event, nr_pages);
2632 perf_mmap_data_init(event, data);
2634 atomic_set(&event->mmap_count, 1);
2635 atomic_long_add(user_extra, &user->locked_vm);
2636 vma->vm_mm->locked_vm += extra;
2637 event->data->nr_locked = extra;
2638 if (vma->vm_flags & VM_WRITE)
2639 event->data->writable = 1;
2642 mutex_unlock(&event->mmap_mutex);
2644 vma->vm_flags |= VM_RESERVED;
2645 vma->vm_ops = &perf_mmap_vmops;
2650 static int perf_fasync(int fd, struct file *filp, int on)
2652 struct inode *inode = filp->f_path.dentry->d_inode;
2653 struct perf_event *event = filp->private_data;
2656 mutex_lock(&inode->i_mutex);
2657 retval = fasync_helper(fd, filp, on, &event->fasync);
2658 mutex_unlock(&inode->i_mutex);
2666 static const struct file_operations perf_fops = {
2667 .llseek = no_llseek,
2668 .release = perf_release,
2671 .unlocked_ioctl = perf_ioctl,
2672 .compat_ioctl = perf_ioctl,
2674 .fasync = perf_fasync,
2680 * If there's data, ensure we set the poll() state and publish everything
2681 * to user-space before waking everybody up.
2684 void perf_event_wakeup(struct perf_event *event)
2686 wake_up_all(&event->waitq);
2688 if (event->pending_kill) {
2689 kill_fasync(&event->fasync, SIGIO, event->pending_kill);
2690 event->pending_kill = 0;
2697 * Handle the case where we need to wakeup up from NMI (or rq->lock) context.
2699 * The NMI bit means we cannot possibly take locks. Therefore, maintain a
2700 * single linked list and use cmpxchg() to add entries lockless.
2703 static void perf_pending_event(struct perf_pending_entry *entry)
2705 struct perf_event *event = container_of(entry,
2706 struct perf_event, pending);
2708 if (event->pending_disable) {
2709 event->pending_disable = 0;
2710 __perf_event_disable(event);
2713 if (event->pending_wakeup) {
2714 event->pending_wakeup = 0;
2715 perf_event_wakeup(event);
2719 #define PENDING_TAIL ((struct perf_pending_entry *)-1UL)
2721 static DEFINE_PER_CPU(struct perf_pending_entry *, perf_pending_head) = {
2725 static void perf_pending_queue(struct perf_pending_entry *entry,
2726 void (*func)(struct perf_pending_entry *))
2728 struct perf_pending_entry **head;
2730 if (cmpxchg(&entry->next, NULL, PENDING_TAIL) != NULL)
2735 head = &get_cpu_var(perf_pending_head);
2738 entry->next = *head;
2739 } while (cmpxchg(head, entry->next, entry) != entry->next);
2741 set_perf_event_pending();
2743 put_cpu_var(perf_pending_head);
2746 static int __perf_pending_run(void)
2748 struct perf_pending_entry *list;
2751 list = xchg(&__get_cpu_var(perf_pending_head), PENDING_TAIL);
2752 while (list != PENDING_TAIL) {
2753 void (*func)(struct perf_pending_entry *);
2754 struct perf_pending_entry *entry = list;
2761 * Ensure we observe the unqueue before we issue the wakeup,
2762 * so that we won't be waiting forever.
2763 * -- see perf_not_pending().
2774 static inline int perf_not_pending(struct perf_event *event)
2777 * If we flush on whatever cpu we run, there is a chance we don't
2781 __perf_pending_run();
2785 * Ensure we see the proper queue state before going to sleep
2786 * so that we do not miss the wakeup. -- see perf_pending_handle()
2789 return event->pending.next == NULL;
2792 static void perf_pending_sync(struct perf_event *event)
2794 wait_event(event->waitq, perf_not_pending(event));
2797 void perf_event_do_pending(void)
2799 __perf_pending_run();
2803 * Callchain support -- arch specific
2806 __weak struct perf_callchain_entry *perf_callchain(struct pt_regs *regs)
2812 void perf_arch_fetch_caller_regs(struct pt_regs *regs, unsigned long ip, int skip)
2818 * We assume there is only KVM supporting the callbacks.
2819 * Later on, we might change it to a list if there is
2820 * another virtualization implementation supporting the callbacks.
2822 struct perf_guest_info_callbacks *perf_guest_cbs;
2824 int perf_register_guest_info_callbacks(struct perf_guest_info_callbacks *cbs)
2826 perf_guest_cbs = cbs;
2829 EXPORT_SYMBOL_GPL(perf_register_guest_info_callbacks);
2831 int perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks *cbs)
2833 perf_guest_cbs = NULL;
2836 EXPORT_SYMBOL_GPL(perf_unregister_guest_info_callbacks);
2841 static bool perf_output_space(struct perf_mmap_data *data, unsigned long tail,
2842 unsigned long offset, unsigned long head)
2846 if (!data->writable)
2849 mask = perf_data_size(data) - 1;
2851 offset = (offset - tail) & mask;
2852 head = (head - tail) & mask;
2854 if ((int)(head - offset) < 0)
2860 static void perf_output_wakeup(struct perf_output_handle *handle)
2862 atomic_set(&handle->data->poll, POLL_IN);
2865 handle->event->pending_wakeup = 1;
2866 perf_pending_queue(&handle->event->pending,
2867 perf_pending_event);
2869 perf_event_wakeup(handle->event);
2873 * Curious locking construct.
2875 * We need to ensure a later event_id doesn't publish a head when a former
2876 * event_id isn't done writing. However since we need to deal with NMIs we
2877 * cannot fully serialize things.
2879 * What we do is serialize between CPUs so we only have to deal with NMI
2880 * nesting on a single CPU.
2882 * We only publish the head (and generate a wakeup) when the outer-most
2883 * event_id completes.
2885 static void perf_output_lock(struct perf_output_handle *handle)
2887 struct perf_mmap_data *data = handle->data;
2888 int cur, cpu = get_cpu();
2893 cur = atomic_cmpxchg(&data->lock, -1, cpu);
2905 static void perf_output_unlock(struct perf_output_handle *handle)
2907 struct perf_mmap_data *data = handle->data;
2911 data->done_head = data->head;
2913 if (!handle->locked)
2918 * The xchg implies a full barrier that ensures all writes are done
2919 * before we publish the new head, matched by a rmb() in userspace when
2920 * reading this position.
2922 while ((head = atomic_long_xchg(&data->done_head, 0)))
2923 data->user_page->data_head = head;
2926 * NMI can happen here, which means we can miss a done_head update.
2929 cpu = atomic_xchg(&data->lock, -1);
2930 WARN_ON_ONCE(cpu != smp_processor_id());
2933 * Therefore we have to validate we did not indeed do so.
2935 if (unlikely(atomic_long_read(&data->done_head))) {
2937 * Since we had it locked, we can lock it again.
2939 while (atomic_cmpxchg(&data->lock, -1, cpu) != -1)
2945 if (atomic_xchg(&data->wakeup, 0))
2946 perf_output_wakeup(handle);
2951 void perf_output_copy(struct perf_output_handle *handle,
2952 const void *buf, unsigned int len)
2954 unsigned int pages_mask;
2955 unsigned long offset;
2959 offset = handle->offset;
2960 pages_mask = handle->data->nr_pages - 1;
2961 pages = handle->data->data_pages;
2964 unsigned long page_offset;
2965 unsigned long page_size;
2968 nr = (offset >> PAGE_SHIFT) & pages_mask;
2969 page_size = 1UL << (handle->data->data_order + PAGE_SHIFT);
2970 page_offset = offset & (page_size - 1);
2971 size = min_t(unsigned int, page_size - page_offset, len);
2973 memcpy(pages[nr] + page_offset, buf, size);
2980 handle->offset = offset;
2983 * Check we didn't copy past our reservation window, taking the
2984 * possible unsigned int wrap into account.
2986 WARN_ON_ONCE(((long)(handle->head - handle->offset)) < 0);
2989 int perf_output_begin(struct perf_output_handle *handle,
2990 struct perf_event *event, unsigned int size,
2991 int nmi, int sample)
2993 struct perf_event *output_event;
2994 struct perf_mmap_data *data;
2995 unsigned long tail, offset, head;
2998 struct perf_event_header header;
3005 * For inherited events we send all the output towards the parent.
3008 event = event->parent;
3010 output_event = rcu_dereference(event->output);
3012 event = output_event;
3014 data = rcu_dereference(event->data);
3018 handle->data = data;
3019 handle->event = event;
3021 handle->sample = sample;
3023 if (!data->nr_pages)
3026 have_lost = atomic_read(&data->lost);
3028 size += sizeof(lost_event);
3030 perf_output_lock(handle);
3034 * Userspace could choose to issue a mb() before updating the
3035 * tail pointer. So that all reads will be completed before the
3038 tail = ACCESS_ONCE(data->user_page->data_tail);
3040 offset = head = atomic_long_read(&data->head);
3042 if (unlikely(!perf_output_space(data, tail, offset, head)))
3044 } while (atomic_long_cmpxchg(&data->head, offset, head) != offset);
3046 handle->offset = offset;
3047 handle->head = head;
3049 if (head - tail > data->watermark)
3050 atomic_set(&data->wakeup, 1);
3053 lost_event.header.type = PERF_RECORD_LOST;
3054 lost_event.header.misc = 0;
3055 lost_event.header.size = sizeof(lost_event);
3056 lost_event.id = event->id;
3057 lost_event.lost = atomic_xchg(&data->lost, 0);
3059 perf_output_put(handle, lost_event);
3065 atomic_inc(&data->lost);
3066 perf_output_unlock(handle);
3073 void perf_output_end(struct perf_output_handle *handle)
3075 struct perf_event *event = handle->event;
3076 struct perf_mmap_data *data = handle->data;
3078 int wakeup_events = event->attr.wakeup_events;
3080 if (handle->sample && wakeup_events) {
3081 int events = atomic_inc_return(&data->events);
3082 if (events >= wakeup_events) {
3083 atomic_sub(wakeup_events, &data->events);
3084 atomic_set(&data->wakeup, 1);
3088 perf_output_unlock(handle);
3092 static u32 perf_event_pid(struct perf_event *event, struct task_struct *p)
3095 * only top level events have the pid namespace they were created in
3098 event = event->parent;
3100 return task_tgid_nr_ns(p, event->ns);
3103 static u32 perf_event_tid(struct perf_event *event, struct task_struct *p)
3106 * only top level events have the pid namespace they were created in
3109 event = event->parent;
3111 return task_pid_nr_ns(p, event->ns);
3114 static void perf_output_read_one(struct perf_output_handle *handle,
3115 struct perf_event *event)
3117 u64 read_format = event->attr.read_format;
3121 values[n++] = atomic64_read(&event->count);
3122 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
3123 values[n++] = event->total_time_enabled +
3124 atomic64_read(&event->child_total_time_enabled);
3126 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
3127 values[n++] = event->total_time_running +
3128 atomic64_read(&event->child_total_time_running);
3130 if (read_format & PERF_FORMAT_ID)
3131 values[n++] = primary_event_id(event);
3133 perf_output_copy(handle, values, n * sizeof(u64));
3137 * XXX PERF_FORMAT_GROUP vs inherited events seems difficult.
3139 static void perf_output_read_group(struct perf_output_handle *handle,
3140 struct perf_event *event)
3142 struct perf_event *leader = event->group_leader, *sub;
3143 u64 read_format = event->attr.read_format;
3147 values[n++] = 1 + leader->nr_siblings;
3149 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
3150 values[n++] = leader->total_time_enabled;
3152 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
3153 values[n++] = leader->total_time_running;
3155 if (leader != event)
3156 leader->pmu->read(leader);
3158 values[n++] = atomic64_read(&leader->count);
3159 if (read_format & PERF_FORMAT_ID)
3160 values[n++] = primary_event_id(leader);
3162 perf_output_copy(handle, values, n * sizeof(u64));
3164 list_for_each_entry(sub, &leader->sibling_list, group_entry) {
3168 sub->pmu->read(sub);
3170 values[n++] = atomic64_read(&sub->count);
3171 if (read_format & PERF_FORMAT_ID)
3172 values[n++] = primary_event_id(sub);
3174 perf_output_copy(handle, values, n * sizeof(u64));
3178 static void perf_output_read(struct perf_output_handle *handle,
3179 struct perf_event *event)
3181 if (event->attr.read_format & PERF_FORMAT_GROUP)
3182 perf_output_read_group(handle, event);
3184 perf_output_read_one(handle, event);
3187 void perf_output_sample(struct perf_output_handle *handle,
3188 struct perf_event_header *header,
3189 struct perf_sample_data *data,
3190 struct perf_event *event)
3192 u64 sample_type = data->type;
3194 perf_output_put(handle, *header);
3196 if (sample_type & PERF_SAMPLE_IP)
3197 perf_output_put(handle, data->ip);
3199 if (sample_type & PERF_SAMPLE_TID)
3200 perf_output_put(handle, data->tid_entry);
3202 if (sample_type & PERF_SAMPLE_TIME)
3203 perf_output_put(handle, data->time);
3205 if (sample_type & PERF_SAMPLE_ADDR)
3206 perf_output_put(handle, data->addr);
3208 if (sample_type & PERF_SAMPLE_ID)
3209 perf_output_put(handle, data->id);
3211 if (sample_type & PERF_SAMPLE_STREAM_ID)
3212 perf_output_put(handle, data->stream_id);
3214 if (sample_type & PERF_SAMPLE_CPU)
3215 perf_output_put(handle, data->cpu_entry);
3217 if (sample_type & PERF_SAMPLE_PERIOD)
3218 perf_output_put(handle, data->period);
3220 if (sample_type & PERF_SAMPLE_READ)
3221 perf_output_read(handle, event);
3223 if (sample_type & PERF_SAMPLE_CALLCHAIN) {
3224 if (data->callchain) {
3227 if (data->callchain)
3228 size += data->callchain->nr;
3230 size *= sizeof(u64);
3232 perf_output_copy(handle, data->callchain, size);
3235 perf_output_put(handle, nr);
3239 if (sample_type & PERF_SAMPLE_RAW) {
3241 perf_output_put(handle, data->raw->size);
3242 perf_output_copy(handle, data->raw->data,
3249 .size = sizeof(u32),
3252 perf_output_put(handle, raw);
3257 void perf_prepare_sample(struct perf_event_header *header,
3258 struct perf_sample_data *data,
3259 struct perf_event *event,
3260 struct pt_regs *regs)
3262 u64 sample_type = event->attr.sample_type;
3264 data->type = sample_type;
3266 header->type = PERF_RECORD_SAMPLE;
3267 header->size = sizeof(*header);
3270 header->misc |= perf_misc_flags(regs);
3272 if (sample_type & PERF_SAMPLE_IP) {
3273 data->ip = perf_instruction_pointer(regs);
3275 header->size += sizeof(data->ip);
3278 if (sample_type & PERF_SAMPLE_TID) {
3279 /* namespace issues */
3280 data->tid_entry.pid = perf_event_pid(event, current);
3281 data->tid_entry.tid = perf_event_tid(event, current);
3283 header->size += sizeof(data->tid_entry);
3286 if (sample_type & PERF_SAMPLE_TIME) {
3287 data->time = perf_clock();
3289 header->size += sizeof(data->time);
3292 if (sample_type & PERF_SAMPLE_ADDR)
3293 header->size += sizeof(data->addr);
3295 if (sample_type & PERF_SAMPLE_ID) {
3296 data->id = primary_event_id(event);
3298 header->size += sizeof(data->id);
3301 if (sample_type & PERF_SAMPLE_STREAM_ID) {
3302 data->stream_id = event->id;
3304 header->size += sizeof(data->stream_id);
3307 if (sample_type & PERF_SAMPLE_CPU) {
3308 data->cpu_entry.cpu = raw_smp_processor_id();
3309 data->cpu_entry.reserved = 0;
3311 header->size += sizeof(data->cpu_entry);
3314 if (sample_type & PERF_SAMPLE_PERIOD)
3315 header->size += sizeof(data->period);
3317 if (sample_type & PERF_SAMPLE_READ)
3318 header->size += perf_event_read_size(event);
3320 if (sample_type & PERF_SAMPLE_CALLCHAIN) {
3323 data->callchain = perf_callchain(regs);
3325 if (data->callchain)
3326 size += data->callchain->nr;
3328 header->size += size * sizeof(u64);
3331 if (sample_type & PERF_SAMPLE_RAW) {
3332 int size = sizeof(u32);
3335 size += data->raw->size;
3337 size += sizeof(u32);
3339 WARN_ON_ONCE(size & (sizeof(u64)-1));
3340 header->size += size;
3344 static void perf_event_output(struct perf_event *event, int nmi,
3345 struct perf_sample_data *data,
3346 struct pt_regs *regs)
3348 struct perf_output_handle handle;
3349 struct perf_event_header header;
3351 perf_prepare_sample(&header, data, event, regs);
3353 if (perf_output_begin(&handle, event, header.size, nmi, 1))
3356 perf_output_sample(&handle, &header, data, event);
3358 perf_output_end(&handle);
3365 struct perf_read_event {
3366 struct perf_event_header header;
3373 perf_event_read_event(struct perf_event *event,
3374 struct task_struct *task)
3376 struct perf_output_handle handle;
3377 struct perf_read_event read_event = {
3379 .type = PERF_RECORD_READ,
3381 .size = sizeof(read_event) + perf_event_read_size(event),
3383 .pid = perf_event_pid(event, task),
3384 .tid = perf_event_tid(event, task),
3388 ret = perf_output_begin(&handle, event, read_event.header.size, 0, 0);
3392 perf_output_put(&handle, read_event);
3393 perf_output_read(&handle, event);
3395 perf_output_end(&handle);
3399 * task tracking -- fork/exit
3401 * enabled by: attr.comm | attr.mmap | attr.task
3404 struct perf_task_event {
3405 struct task_struct *task;
3406 struct perf_event_context *task_ctx;
3409 struct perf_event_header header;
3419 static void perf_event_task_output(struct perf_event *event,
3420 struct perf_task_event *task_event)
3422 struct perf_output_handle handle;
3423 struct task_struct *task = task_event->task;
3424 unsigned long flags;
3428 * If this CPU attempts to acquire an rq lock held by a CPU spinning
3429 * in perf_output_lock() from interrupt context, it's game over.
3431 local_irq_save(flags);
3433 size = task_event->event_id.header.size;
3434 ret = perf_output_begin(&handle, event, size, 0, 0);
3437 local_irq_restore(flags);
3441 task_event->event_id.pid = perf_event_pid(event, task);
3442 task_event->event_id.ppid = perf_event_pid(event, current);
3444 task_event->event_id.tid = perf_event_tid(event, task);
3445 task_event->event_id.ptid = perf_event_tid(event, current);
3447 perf_output_put(&handle, task_event->event_id);
3449 perf_output_end(&handle);
3450 local_irq_restore(flags);
3453 static int perf_event_task_match(struct perf_event *event)
3455 if (event->state < PERF_EVENT_STATE_INACTIVE)
3458 if (event->cpu != -1 && event->cpu != smp_processor_id())
3461 if (event->attr.comm || event->attr.mmap || event->attr.task)
3467 static void perf_event_task_ctx(struct perf_event_context *ctx,
3468 struct perf_task_event *task_event)
3470 struct perf_event *event;
3472 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
3473 if (perf_event_task_match(event))
3474 perf_event_task_output(event, task_event);
3478 static void perf_event_task_event(struct perf_task_event *task_event)
3480 struct perf_cpu_context *cpuctx;
3481 struct perf_event_context *ctx = task_event->task_ctx;
3484 cpuctx = &get_cpu_var(perf_cpu_context);
3485 perf_event_task_ctx(&cpuctx->ctx, task_event);
3487 ctx = rcu_dereference(current->perf_event_ctxp);
3489 perf_event_task_ctx(ctx, task_event);
3490 put_cpu_var(perf_cpu_context);
3494 static void perf_event_task(struct task_struct *task,
3495 struct perf_event_context *task_ctx,
3498 struct perf_task_event task_event;
3500 if (!atomic_read(&nr_comm_events) &&
3501 !atomic_read(&nr_mmap_events) &&
3502 !atomic_read(&nr_task_events))
3505 task_event = (struct perf_task_event){
3507 .task_ctx = task_ctx,
3510 .type = new ? PERF_RECORD_FORK : PERF_RECORD_EXIT,
3512 .size = sizeof(task_event.event_id),
3518 .time = perf_clock(),
3522 perf_event_task_event(&task_event);
3525 void perf_event_fork(struct task_struct *task)
3527 perf_event_task(task, NULL, 1);
3534 struct perf_comm_event {
3535 struct task_struct *task;
3540 struct perf_event_header header;
3547 static void perf_event_comm_output(struct perf_event *event,
3548 struct perf_comm_event *comm_event)
3550 struct perf_output_handle handle;
3551 int size = comm_event->event_id.header.size;
3552 int ret = perf_output_begin(&handle, event, size, 0, 0);
3557 comm_event->event_id.pid = perf_event_pid(event, comm_event->task);
3558 comm_event->event_id.tid = perf_event_tid(event, comm_event->task);
3560 perf_output_put(&handle, comm_event->event_id);
3561 perf_output_copy(&handle, comm_event->comm,
3562 comm_event->comm_size);
3563 perf_output_end(&handle);
3566 static int perf_event_comm_match(struct perf_event *event)
3568 if (event->state < PERF_EVENT_STATE_INACTIVE)
3571 if (event->cpu != -1 && event->cpu != smp_processor_id())
3574 if (event->attr.comm)
3580 static void perf_event_comm_ctx(struct perf_event_context *ctx,
3581 struct perf_comm_event *comm_event)
3583 struct perf_event *event;
3585 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
3586 if (perf_event_comm_match(event))
3587 perf_event_comm_output(event, comm_event);
3591 static void perf_event_comm_event(struct perf_comm_event *comm_event)
3593 struct perf_cpu_context *cpuctx;
3594 struct perf_event_context *ctx;
3596 char comm[TASK_COMM_LEN];
3598 memset(comm, 0, sizeof(comm));
3599 strlcpy(comm, comm_event->task->comm, sizeof(comm));
3600 size = ALIGN(strlen(comm)+1, sizeof(u64));
3602 comm_event->comm = comm;
3603 comm_event->comm_size = size;
3605 comm_event->event_id.header.size = sizeof(comm_event->event_id) + size;
3608 cpuctx = &get_cpu_var(perf_cpu_context);
3609 perf_event_comm_ctx(&cpuctx->ctx, comm_event);
3610 ctx = rcu_dereference(current->perf_event_ctxp);
3612 perf_event_comm_ctx(ctx, comm_event);
3613 put_cpu_var(perf_cpu_context);
3617 void perf_event_comm(struct task_struct *task)
3619 struct perf_comm_event comm_event;
3621 if (task->perf_event_ctxp)
3622 perf_event_enable_on_exec(task);
3624 if (!atomic_read(&nr_comm_events))
3627 comm_event = (struct perf_comm_event){
3633 .type = PERF_RECORD_COMM,
3642 perf_event_comm_event(&comm_event);
3649 struct perf_mmap_event {
3650 struct vm_area_struct *vma;
3652 const char *file_name;
3656 struct perf_event_header header;
3666 static void perf_event_mmap_output(struct perf_event *event,
3667 struct perf_mmap_event *mmap_event)
3669 struct perf_output_handle handle;
3670 int size = mmap_event->event_id.header.size;
3671 int ret = perf_output_begin(&handle, event, size, 0, 0);
3676 mmap_event->event_id.pid = perf_event_pid(event, current);
3677 mmap_event->event_id.tid = perf_event_tid(event, current);
3679 perf_output_put(&handle, mmap_event->event_id);
3680 perf_output_copy(&handle, mmap_event->file_name,
3681 mmap_event->file_size);
3682 perf_output_end(&handle);
3685 static int perf_event_mmap_match(struct perf_event *event,
3686 struct perf_mmap_event *mmap_event)
3688 if (event->state < PERF_EVENT_STATE_INACTIVE)
3691 if (event->cpu != -1 && event->cpu != smp_processor_id())
3694 if (event->attr.mmap)
3700 static void perf_event_mmap_ctx(struct perf_event_context *ctx,
3701 struct perf_mmap_event *mmap_event)
3703 struct perf_event *event;
3705 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
3706 if (perf_event_mmap_match(event, mmap_event))
3707 perf_event_mmap_output(event, mmap_event);
3711 static void perf_event_mmap_event(struct perf_mmap_event *mmap_event)
3713 struct perf_cpu_context *cpuctx;
3714 struct perf_event_context *ctx;
3715 struct vm_area_struct *vma = mmap_event->vma;
3716 struct file *file = vma->vm_file;
3722 memset(tmp, 0, sizeof(tmp));
3726 * d_path works from the end of the buffer backwards, so we
3727 * need to add enough zero bytes after the string to handle
3728 * the 64bit alignment we do later.
3730 buf = kzalloc(PATH_MAX + sizeof(u64), GFP_KERNEL);
3732 name = strncpy(tmp, "//enomem", sizeof(tmp));
3735 name = d_path(&file->f_path, buf, PATH_MAX);
3737 name = strncpy(tmp, "//toolong", sizeof(tmp));
3741 if (arch_vma_name(mmap_event->vma)) {
3742 name = strncpy(tmp, arch_vma_name(mmap_event->vma),
3748 name = strncpy(tmp, "[vdso]", sizeof(tmp));
3752 name = strncpy(tmp, "//anon", sizeof(tmp));
3757 size = ALIGN(strlen(name)+1, sizeof(u64));
3759 mmap_event->file_name = name;
3760 mmap_event->file_size = size;
3762 mmap_event->event_id.header.size = sizeof(mmap_event->event_id) + size;
3765 cpuctx = &get_cpu_var(perf_cpu_context);
3766 perf_event_mmap_ctx(&cpuctx->ctx, mmap_event);
3767 ctx = rcu_dereference(current->perf_event_ctxp);
3769 perf_event_mmap_ctx(ctx, mmap_event);
3770 put_cpu_var(perf_cpu_context);
3776 void __perf_event_mmap(struct vm_area_struct *vma)
3778 struct perf_mmap_event mmap_event;
3780 if (!atomic_read(&nr_mmap_events))
3783 mmap_event = (struct perf_mmap_event){
3789 .type = PERF_RECORD_MMAP,
3790 .misc = PERF_RECORD_MISC_USER,
3795 .start = vma->vm_start,
3796 .len = vma->vm_end - vma->vm_start,
3797 .pgoff = (u64)vma->vm_pgoff << PAGE_SHIFT,
3801 perf_event_mmap_event(&mmap_event);
3805 * IRQ throttle logging
3808 static void perf_log_throttle(struct perf_event *event, int enable)
3810 struct perf_output_handle handle;
3814 struct perf_event_header header;
3818 } throttle_event = {
3820 .type = PERF_RECORD_THROTTLE,
3822 .size = sizeof(throttle_event),
3824 .time = perf_clock(),
3825 .id = primary_event_id(event),
3826 .stream_id = event->id,
3830 throttle_event.header.type = PERF_RECORD_UNTHROTTLE;
3832 ret = perf_output_begin(&handle, event, sizeof(throttle_event), 1, 0);
3836 perf_output_put(&handle, throttle_event);
3837 perf_output_end(&handle);
3841 * Generic event overflow handling, sampling.
3844 static int __perf_event_overflow(struct perf_event *event, int nmi,
3845 int throttle, struct perf_sample_data *data,
3846 struct pt_regs *regs)
3848 int events = atomic_read(&event->event_limit);
3849 struct hw_perf_event *hwc = &event->hw;
3852 throttle = (throttle && event->pmu->unthrottle != NULL);
3857 if (hwc->interrupts != MAX_INTERRUPTS) {
3859 if (HZ * hwc->interrupts >
3860 (u64)sysctl_perf_event_sample_rate) {
3861 hwc->interrupts = MAX_INTERRUPTS;
3862 perf_log_throttle(event, 0);
3867 * Keep re-disabling events even though on the previous
3868 * pass we disabled it - just in case we raced with a
3869 * sched-in and the event got enabled again:
3875 if (event->attr.freq) {
3876 u64 now = perf_clock();
3877 s64 delta = now - hwc->freq_time_stamp;
3879 hwc->freq_time_stamp = now;
3881 if (delta > 0 && delta < 2*TICK_NSEC)
3882 perf_adjust_period(event, delta, hwc->last_period);
3886 * XXX event_limit might not quite work as expected on inherited
3890 event->pending_kill = POLL_IN;
3891 if (events && atomic_dec_and_test(&event->event_limit)) {
3893 event->pending_kill = POLL_HUP;
3895 event->pending_disable = 1;
3896 perf_pending_queue(&event->pending,
3897 perf_pending_event);
3899 perf_event_disable(event);
3902 if (event->overflow_handler)
3903 event->overflow_handler(event, nmi, data, regs);
3905 perf_event_output(event, nmi, data, regs);
3910 int perf_event_overflow(struct perf_event *event, int nmi,
3911 struct perf_sample_data *data,
3912 struct pt_regs *regs)
3914 return __perf_event_overflow(event, nmi, 1, data, regs);
3918 * Generic software event infrastructure
3922 * We directly increment event->count and keep a second value in
3923 * event->hw.period_left to count intervals. This period event
3924 * is kept in the range [-sample_period, 0] so that we can use the
3928 static u64 perf_swevent_set_period(struct perf_event *event)
3930 struct hw_perf_event *hwc = &event->hw;
3931 u64 period = hwc->last_period;
3935 hwc->last_period = hwc->sample_period;
3938 old = val = atomic64_read(&hwc->period_left);
3942 nr = div64_u64(period + val, period);
3943 offset = nr * period;
3945 if (atomic64_cmpxchg(&hwc->period_left, old, val) != old)
3951 static void perf_swevent_overflow(struct perf_event *event, u64 overflow,
3952 int nmi, struct perf_sample_data *data,
3953 struct pt_regs *regs)
3955 struct hw_perf_event *hwc = &event->hw;
3958 data->period = event->hw.last_period;
3960 overflow = perf_swevent_set_period(event);
3962 if (hwc->interrupts == MAX_INTERRUPTS)
3965 for (; overflow; overflow--) {
3966 if (__perf_event_overflow(event, nmi, throttle,
3969 * We inhibit the overflow from happening when
3970 * hwc->interrupts == MAX_INTERRUPTS.
3978 static void perf_swevent_unthrottle(struct perf_event *event)
3981 * Nothing to do, we already reset hwc->interrupts.
3985 static void perf_swevent_add(struct perf_event *event, u64 nr,
3986 int nmi, struct perf_sample_data *data,
3987 struct pt_regs *regs)
3989 struct hw_perf_event *hwc = &event->hw;
3991 atomic64_add(nr, &event->count);
3996 if (!hwc->sample_period)
3999 if (nr == 1 && hwc->sample_period == 1 && !event->attr.freq)
4000 return perf_swevent_overflow(event, 1, nmi, data, regs);
4002 if (atomic64_add_negative(nr, &hwc->period_left))
4005 perf_swevent_overflow(event, 0, nmi, data, regs);
4008 static int perf_tp_event_match(struct perf_event *event,
4009 struct perf_sample_data *data);
4011 static int perf_exclude_event(struct perf_event *event,
4012 struct pt_regs *regs)
4015 if (event->attr.exclude_user && user_mode(regs))
4018 if (event->attr.exclude_kernel && !user_mode(regs))
4025 static int perf_swevent_match(struct perf_event *event,
4026 enum perf_type_id type,
4028 struct perf_sample_data *data,
4029 struct pt_regs *regs)
4031 if (event->attr.type != type)
4034 if (event->attr.config != event_id)
4037 if (perf_exclude_event(event, regs))
4040 if (event->attr.type == PERF_TYPE_TRACEPOINT &&
4041 !perf_tp_event_match(event, data))
4047 static inline u64 swevent_hash(u64 type, u32 event_id)
4049 u64 val = event_id | (type << 32);
4051 return hash_64(val, SWEVENT_HLIST_BITS);
4054 static struct hlist_head *
4055 find_swevent_head(struct perf_cpu_context *ctx, u64 type, u32 event_id)
4058 struct swevent_hlist *hlist;
4060 hash = swevent_hash(type, event_id);
4062 hlist = rcu_dereference(ctx->swevent_hlist);
4066 return &hlist->heads[hash];
4069 static void do_perf_sw_event(enum perf_type_id type, u32 event_id,
4071 struct perf_sample_data *data,
4072 struct pt_regs *regs)
4074 struct perf_cpu_context *cpuctx;
4075 struct perf_event *event;
4076 struct hlist_node *node;
4077 struct hlist_head *head;
4079 cpuctx = &__get_cpu_var(perf_cpu_context);
4083 head = find_swevent_head(cpuctx, type, event_id);
4088 hlist_for_each_entry_rcu(event, node, head, hlist_entry) {
4089 if (perf_swevent_match(event, type, event_id, data, regs))
4090 perf_swevent_add(event, nr, nmi, data, regs);
4096 int perf_swevent_get_recursion_context(void)
4098 struct perf_cpu_context *cpuctx = &get_cpu_var(perf_cpu_context);
4105 else if (in_softirq())
4110 if (cpuctx->recursion[rctx]) {
4111 put_cpu_var(perf_cpu_context);
4115 cpuctx->recursion[rctx]++;
4120 EXPORT_SYMBOL_GPL(perf_swevent_get_recursion_context);
4122 void perf_swevent_put_recursion_context(int rctx)
4124 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
4126 cpuctx->recursion[rctx]--;
4127 put_cpu_var(perf_cpu_context);
4129 EXPORT_SYMBOL_GPL(perf_swevent_put_recursion_context);
4132 void __perf_sw_event(u32 event_id, u64 nr, int nmi,
4133 struct pt_regs *regs, u64 addr)
4135 struct perf_sample_data data;
4138 rctx = perf_swevent_get_recursion_context();
4142 perf_sample_data_init(&data, addr);
4144 do_perf_sw_event(PERF_TYPE_SOFTWARE, event_id, nr, nmi, &data, regs);
4146 perf_swevent_put_recursion_context(rctx);
4149 static void perf_swevent_read(struct perf_event *event)
4153 static int perf_swevent_enable(struct perf_event *event)
4155 struct hw_perf_event *hwc = &event->hw;
4156 struct perf_cpu_context *cpuctx;
4157 struct hlist_head *head;
4159 cpuctx = &__get_cpu_var(perf_cpu_context);
4161 if (hwc->sample_period) {
4162 hwc->last_period = hwc->sample_period;
4163 perf_swevent_set_period(event);
4166 head = find_swevent_head(cpuctx, event->attr.type, event->attr.config);
4167 if (WARN_ON_ONCE(!head))
4170 hlist_add_head_rcu(&event->hlist_entry, head);
4175 static void perf_swevent_disable(struct perf_event *event)
4177 hlist_del_rcu(&event->hlist_entry);
4180 static const struct pmu perf_ops_generic = {
4181 .enable = perf_swevent_enable,
4182 .disable = perf_swevent_disable,
4183 .read = perf_swevent_read,
4184 .unthrottle = perf_swevent_unthrottle,
4188 * hrtimer based swevent callback
4191 static enum hrtimer_restart perf_swevent_hrtimer(struct hrtimer *hrtimer)
4193 enum hrtimer_restart ret = HRTIMER_RESTART;
4194 struct perf_sample_data data;
4195 struct pt_regs *regs;
4196 struct perf_event *event;
4199 event = container_of(hrtimer, struct perf_event, hw.hrtimer);
4200 event->pmu->read(event);
4202 perf_sample_data_init(&data, 0);
4203 data.period = event->hw.last_period;
4204 regs = get_irq_regs();
4206 if (regs && !perf_exclude_event(event, regs)) {
4207 if (!(event->attr.exclude_idle && current->pid == 0))
4208 if (perf_event_overflow(event, 0, &data, regs))
4209 ret = HRTIMER_NORESTART;
4212 period = max_t(u64, 10000, event->hw.sample_period);
4213 hrtimer_forward_now(hrtimer, ns_to_ktime(period));
4218 static void perf_swevent_start_hrtimer(struct perf_event *event)
4220 struct hw_perf_event *hwc = &event->hw;
4222 hrtimer_init(&hwc->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
4223 hwc->hrtimer.function = perf_swevent_hrtimer;
4224 if (hwc->sample_period) {
4227 if (hwc->remaining) {
4228 if (hwc->remaining < 0)
4231 period = hwc->remaining;
4234 period = max_t(u64, 10000, hwc->sample_period);
4236 __hrtimer_start_range_ns(&hwc->hrtimer,
4237 ns_to_ktime(period), 0,
4238 HRTIMER_MODE_REL, 0);
4242 static void perf_swevent_cancel_hrtimer(struct perf_event *event)
4244 struct hw_perf_event *hwc = &event->hw;
4246 if (hwc->sample_period) {
4247 ktime_t remaining = hrtimer_get_remaining(&hwc->hrtimer);
4248 hwc->remaining = ktime_to_ns(remaining);
4250 hrtimer_cancel(&hwc->hrtimer);
4255 * Software event: cpu wall time clock
4258 static void cpu_clock_perf_event_update(struct perf_event *event)
4260 int cpu = raw_smp_processor_id();
4264 now = cpu_clock(cpu);
4265 prev = atomic64_xchg(&event->hw.prev_count, now);
4266 atomic64_add(now - prev, &event->count);
4269 static int cpu_clock_perf_event_enable(struct perf_event *event)
4271 struct hw_perf_event *hwc = &event->hw;
4272 int cpu = raw_smp_processor_id();
4274 atomic64_set(&hwc->prev_count, cpu_clock(cpu));
4275 perf_swevent_start_hrtimer(event);
4280 static void cpu_clock_perf_event_disable(struct perf_event *event)
4282 perf_swevent_cancel_hrtimer(event);
4283 cpu_clock_perf_event_update(event);
4286 static void cpu_clock_perf_event_read(struct perf_event *event)
4288 cpu_clock_perf_event_update(event);
4291 static const struct pmu perf_ops_cpu_clock = {
4292 .enable = cpu_clock_perf_event_enable,
4293 .disable = cpu_clock_perf_event_disable,
4294 .read = cpu_clock_perf_event_read,
4298 * Software event: task time clock
4301 static void task_clock_perf_event_update(struct perf_event *event, u64 now)
4306 prev = atomic64_xchg(&event->hw.prev_count, now);
4308 atomic64_add(delta, &event->count);
4311 static int task_clock_perf_event_enable(struct perf_event *event)
4313 struct hw_perf_event *hwc = &event->hw;
4316 now = event->ctx->time;
4318 atomic64_set(&hwc->prev_count, now);
4320 perf_swevent_start_hrtimer(event);
4325 static void task_clock_perf_event_disable(struct perf_event *event)
4327 perf_swevent_cancel_hrtimer(event);
4328 task_clock_perf_event_update(event, event->ctx->time);
4332 static void task_clock_perf_event_read(struct perf_event *event)
4337 update_context_time(event->ctx);
4338 time = event->ctx->time;
4340 u64 now = perf_clock();
4341 u64 delta = now - event->ctx->timestamp;
4342 time = event->ctx->time + delta;
4345 task_clock_perf_event_update(event, time);
4348 static const struct pmu perf_ops_task_clock = {
4349 .enable = task_clock_perf_event_enable,
4350 .disable = task_clock_perf_event_disable,
4351 .read = task_clock_perf_event_read,
4354 static void swevent_hlist_release_rcu(struct rcu_head *rcu_head)
4356 struct swevent_hlist *hlist;
4358 hlist = container_of(rcu_head, struct swevent_hlist, rcu_head);
4362 static void swevent_hlist_release(struct perf_cpu_context *cpuctx)
4364 struct swevent_hlist *hlist;
4366 if (!cpuctx->swevent_hlist)
4369 hlist = cpuctx->swevent_hlist;
4370 rcu_assign_pointer(cpuctx->swevent_hlist, NULL);
4371 call_rcu(&hlist->rcu_head, swevent_hlist_release_rcu);
4374 static void swevent_hlist_put_cpu(struct perf_event *event, int cpu)
4376 struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
4378 mutex_lock(&cpuctx->hlist_mutex);
4380 if (!--cpuctx->hlist_refcount)
4381 swevent_hlist_release(cpuctx);
4383 mutex_unlock(&cpuctx->hlist_mutex);
4386 static void swevent_hlist_put(struct perf_event *event)
4390 if (event->cpu != -1) {
4391 swevent_hlist_put_cpu(event, event->cpu);
4395 for_each_possible_cpu(cpu)
4396 swevent_hlist_put_cpu(event, cpu);
4399 static int swevent_hlist_get_cpu(struct perf_event *event, int cpu)
4401 struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
4404 mutex_lock(&cpuctx->hlist_mutex);
4406 if (!cpuctx->swevent_hlist && cpu_online(cpu)) {
4407 struct swevent_hlist *hlist;
4409 hlist = kzalloc(sizeof(*hlist), GFP_KERNEL);
4414 rcu_assign_pointer(cpuctx->swevent_hlist, hlist);
4416 cpuctx->hlist_refcount++;
4418 mutex_unlock(&cpuctx->hlist_mutex);
4423 static int swevent_hlist_get(struct perf_event *event)
4426 int cpu, failed_cpu;
4428 if (event->cpu != -1)
4429 return swevent_hlist_get_cpu(event, event->cpu);
4432 for_each_possible_cpu(cpu) {
4433 err = swevent_hlist_get_cpu(event, cpu);
4443 for_each_possible_cpu(cpu) {
4444 if (cpu == failed_cpu)
4446 swevent_hlist_put_cpu(event, cpu);
4453 #ifdef CONFIG_EVENT_TRACING
4455 void perf_tp_event(int event_id, u64 addr, u64 count, void *record,
4456 int entry_size, struct pt_regs *regs)
4458 struct perf_sample_data data;
4459 struct perf_raw_record raw = {
4464 perf_sample_data_init(&data, addr);
4467 /* Trace events already protected against recursion */
4468 do_perf_sw_event(PERF_TYPE_TRACEPOINT, event_id, count, 1,
4471 EXPORT_SYMBOL_GPL(perf_tp_event);
4473 static int perf_tp_event_match(struct perf_event *event,
4474 struct perf_sample_data *data)
4476 void *record = data->raw->data;
4478 if (likely(!event->filter) || filter_match_preds(event->filter, record))
4483 static void tp_perf_event_destroy(struct perf_event *event)
4485 perf_trace_disable(event->attr.config);
4486 swevent_hlist_put(event);
4489 static const struct pmu *tp_perf_event_init(struct perf_event *event)
4494 * Raw tracepoint data is a severe data leak, only allow root to
4497 if ((event->attr.sample_type & PERF_SAMPLE_RAW) &&
4498 perf_paranoid_tracepoint_raw() &&
4499 !capable(CAP_SYS_ADMIN))
4500 return ERR_PTR(-EPERM);
4502 if (perf_trace_enable(event->attr.config))
4505 event->destroy = tp_perf_event_destroy;
4506 err = swevent_hlist_get(event);
4508 perf_trace_disable(event->attr.config);
4509 return ERR_PTR(err);
4512 return &perf_ops_generic;
4515 static int perf_event_set_filter(struct perf_event *event, void __user *arg)
4520 if (event->attr.type != PERF_TYPE_TRACEPOINT)
4523 filter_str = strndup_user(arg, PAGE_SIZE);
4524 if (IS_ERR(filter_str))
4525 return PTR_ERR(filter_str);
4527 ret = ftrace_profile_set_filter(event, event->attr.config, filter_str);
4533 static void perf_event_free_filter(struct perf_event *event)
4535 ftrace_profile_free_filter(event);
4540 static int perf_tp_event_match(struct perf_event *event,
4541 struct perf_sample_data *data)
4546 static const struct pmu *tp_perf_event_init(struct perf_event *event)
4551 static int perf_event_set_filter(struct perf_event *event, void __user *arg)
4556 static void perf_event_free_filter(struct perf_event *event)
4560 #endif /* CONFIG_EVENT_TRACING */
4562 #ifdef CONFIG_HAVE_HW_BREAKPOINT
4563 static void bp_perf_event_destroy(struct perf_event *event)
4565 release_bp_slot(event);
4568 static const struct pmu *bp_perf_event_init(struct perf_event *bp)
4572 err = register_perf_hw_breakpoint(bp);
4574 return ERR_PTR(err);
4576 bp->destroy = bp_perf_event_destroy;
4578 return &perf_ops_bp;
4581 void perf_bp_event(struct perf_event *bp, void *data)
4583 struct perf_sample_data sample;
4584 struct pt_regs *regs = data;
4586 perf_sample_data_init(&sample, bp->attr.bp_addr);
4588 if (!perf_exclude_event(bp, regs))
4589 perf_swevent_add(bp, 1, 1, &sample, regs);
4592 static const struct pmu *bp_perf_event_init(struct perf_event *bp)
4597 void perf_bp_event(struct perf_event *bp, void *regs)
4602 atomic_t perf_swevent_enabled[PERF_COUNT_SW_MAX];
4604 static void sw_perf_event_destroy(struct perf_event *event)
4606 u64 event_id = event->attr.config;
4608 WARN_ON(event->parent);
4610 atomic_dec(&perf_swevent_enabled[event_id]);
4611 swevent_hlist_put(event);
4614 static const struct pmu *sw_perf_event_init(struct perf_event *event)
4616 const struct pmu *pmu = NULL;
4617 u64 event_id = event->attr.config;
4620 * Software events (currently) can't in general distinguish
4621 * between user, kernel and hypervisor events.
4622 * However, context switches and cpu migrations are considered
4623 * to be kernel events, and page faults are never hypervisor
4627 case PERF_COUNT_SW_CPU_CLOCK:
4628 pmu = &perf_ops_cpu_clock;
4631 case PERF_COUNT_SW_TASK_CLOCK:
4633 * If the user instantiates this as a per-cpu event,
4634 * use the cpu_clock event instead.
4636 if (event->ctx->task)
4637 pmu = &perf_ops_task_clock;
4639 pmu = &perf_ops_cpu_clock;
4642 case PERF_COUNT_SW_PAGE_FAULTS:
4643 case PERF_COUNT_SW_PAGE_FAULTS_MIN:
4644 case PERF_COUNT_SW_PAGE_FAULTS_MAJ:
4645 case PERF_COUNT_SW_CONTEXT_SWITCHES:
4646 case PERF_COUNT_SW_CPU_MIGRATIONS:
4647 case PERF_COUNT_SW_ALIGNMENT_FAULTS:
4648 case PERF_COUNT_SW_EMULATION_FAULTS:
4649 if (!event->parent) {
4652 err = swevent_hlist_get(event);
4654 return ERR_PTR(err);
4656 atomic_inc(&perf_swevent_enabled[event_id]);
4657 event->destroy = sw_perf_event_destroy;
4659 pmu = &perf_ops_generic;
4667 * Allocate and initialize a event structure
4669 static struct perf_event *
4670 perf_event_alloc(struct perf_event_attr *attr,
4672 struct perf_event_context *ctx,
4673 struct perf_event *group_leader,
4674 struct perf_event *parent_event,
4675 perf_overflow_handler_t overflow_handler,
4678 const struct pmu *pmu;
4679 struct perf_event *event;
4680 struct hw_perf_event *hwc;
4683 event = kzalloc(sizeof(*event), gfpflags);
4685 return ERR_PTR(-ENOMEM);
4688 * Single events are their own group leaders, with an
4689 * empty sibling list:
4692 group_leader = event;
4694 mutex_init(&event->child_mutex);
4695 INIT_LIST_HEAD(&event->child_list);
4697 INIT_LIST_HEAD(&event->group_entry);
4698 INIT_LIST_HEAD(&event->event_entry);
4699 INIT_LIST_HEAD(&event->sibling_list);
4700 init_waitqueue_head(&event->waitq);
4702 mutex_init(&event->mmap_mutex);
4705 event->attr = *attr;
4706 event->group_leader = group_leader;
4711 event->parent = parent_event;
4713 event->ns = get_pid_ns(current->nsproxy->pid_ns);
4714 event->id = atomic64_inc_return(&perf_event_id);
4716 event->state = PERF_EVENT_STATE_INACTIVE;
4718 if (!overflow_handler && parent_event)
4719 overflow_handler = parent_event->overflow_handler;
4721 event->overflow_handler = overflow_handler;
4724 event->state = PERF_EVENT_STATE_OFF;
4729 hwc->sample_period = attr->sample_period;
4730 if (attr->freq && attr->sample_freq)
4731 hwc->sample_period = 1;
4732 hwc->last_period = hwc->sample_period;
4734 atomic64_set(&hwc->period_left, hwc->sample_period);
4737 * we currently do not support PERF_FORMAT_GROUP on inherited events
4739 if (attr->inherit && (attr->read_format & PERF_FORMAT_GROUP))
4742 switch (attr->type) {
4744 case PERF_TYPE_HARDWARE:
4745 case PERF_TYPE_HW_CACHE:
4746 pmu = hw_perf_event_init(event);
4749 case PERF_TYPE_SOFTWARE:
4750 pmu = sw_perf_event_init(event);
4753 case PERF_TYPE_TRACEPOINT:
4754 pmu = tp_perf_event_init(event);
4757 case PERF_TYPE_BREAKPOINT:
4758 pmu = bp_perf_event_init(event);
4769 else if (IS_ERR(pmu))
4774 put_pid_ns(event->ns);
4776 return ERR_PTR(err);
4781 if (!event->parent) {
4782 atomic_inc(&nr_events);
4783 if (event->attr.mmap)
4784 atomic_inc(&nr_mmap_events);
4785 if (event->attr.comm)
4786 atomic_inc(&nr_comm_events);
4787 if (event->attr.task)
4788 atomic_inc(&nr_task_events);
4794 static int perf_copy_attr(struct perf_event_attr __user *uattr,
4795 struct perf_event_attr *attr)
4800 if (!access_ok(VERIFY_WRITE, uattr, PERF_ATTR_SIZE_VER0))
4804 * zero the full structure, so that a short copy will be nice.
4806 memset(attr, 0, sizeof(*attr));
4808 ret = get_user(size, &uattr->size);
4812 if (size > PAGE_SIZE) /* silly large */
4815 if (!size) /* abi compat */
4816 size = PERF_ATTR_SIZE_VER0;
4818 if (size < PERF_ATTR_SIZE_VER0)
4822 * If we're handed a bigger struct than we know of,
4823 * ensure all the unknown bits are 0 - i.e. new
4824 * user-space does not rely on any kernel feature
4825 * extensions we dont know about yet.
4827 if (size > sizeof(*attr)) {
4828 unsigned char __user *addr;
4829 unsigned char __user *end;
4832 addr = (void __user *)uattr + sizeof(*attr);
4833 end = (void __user *)uattr + size;
4835 for (; addr < end; addr++) {
4836 ret = get_user(val, addr);
4842 size = sizeof(*attr);
4845 ret = copy_from_user(attr, uattr, size);
4850 * If the type exists, the corresponding creation will verify
4853 if (attr->type >= PERF_TYPE_MAX)
4856 if (attr->__reserved_1)
4859 if (attr->sample_type & ~(PERF_SAMPLE_MAX-1))
4862 if (attr->read_format & ~(PERF_FORMAT_MAX-1))
4869 put_user(sizeof(*attr), &uattr->size);
4874 static int perf_event_set_output(struct perf_event *event, int output_fd)
4876 struct perf_event *output_event = NULL;
4877 struct file *output_file = NULL;
4878 struct perf_event *old_output;
4879 int fput_needed = 0;
4885 output_file = fget_light(output_fd, &fput_needed);
4889 if (output_file->f_op != &perf_fops)
4892 output_event = output_file->private_data;
4894 /* Don't chain output fds */
4895 if (output_event->output)
4898 /* Don't set an output fd when we already have an output channel */
4902 atomic_long_inc(&output_file->f_count);
4905 mutex_lock(&event->mmap_mutex);
4906 old_output = event->output;
4907 rcu_assign_pointer(event->output, output_event);
4908 mutex_unlock(&event->mmap_mutex);
4912 * we need to make sure no existing perf_output_*()
4913 * is still referencing this event.
4916 fput(old_output->filp);
4921 fput_light(output_file, fput_needed);
4926 * sys_perf_event_open - open a performance event, associate it to a task/cpu
4928 * @attr_uptr: event_id type attributes for monitoring/sampling
4931 * @group_fd: group leader event fd
4933 SYSCALL_DEFINE5(perf_event_open,
4934 struct perf_event_attr __user *, attr_uptr,
4935 pid_t, pid, int, cpu, int, group_fd, unsigned long, flags)
4937 struct perf_event *event, *group_leader;
4938 struct perf_event_attr attr;
4939 struct perf_event_context *ctx;
4940 struct file *event_file = NULL;
4941 struct file *group_file = NULL;
4942 int fput_needed = 0;
4943 int fput_needed2 = 0;
4946 /* for future expandability... */
4947 if (flags & ~(PERF_FLAG_FD_NO_GROUP | PERF_FLAG_FD_OUTPUT))
4950 err = perf_copy_attr(attr_uptr, &attr);
4954 if (!attr.exclude_kernel) {
4955 if (perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN))
4960 if (attr.sample_freq > sysctl_perf_event_sample_rate)
4965 * Get the target context (task or percpu):
4967 ctx = find_get_context(pid, cpu);
4969 return PTR_ERR(ctx);
4972 * Look up the group leader (we will attach this event to it):
4974 group_leader = NULL;
4975 if (group_fd != -1 && !(flags & PERF_FLAG_FD_NO_GROUP)) {
4977 group_file = fget_light(group_fd, &fput_needed);
4979 goto err_put_context;
4980 if (group_file->f_op != &perf_fops)
4981 goto err_put_context;
4983 group_leader = group_file->private_data;
4985 * Do not allow a recursive hierarchy (this new sibling
4986 * becoming part of another group-sibling):
4988 if (group_leader->group_leader != group_leader)
4989 goto err_put_context;
4991 * Do not allow to attach to a group in a different
4992 * task or CPU context:
4994 if (group_leader->ctx != ctx)
4995 goto err_put_context;
4997 * Only a group leader can be exclusive or pinned
4999 if (attr.exclusive || attr.pinned)
5000 goto err_put_context;
5003 event = perf_event_alloc(&attr, cpu, ctx, group_leader,
5004 NULL, NULL, GFP_KERNEL);
5005 err = PTR_ERR(event);
5007 goto err_put_context;
5009 err = anon_inode_getfd("[perf_event]", &perf_fops, event, O_RDWR);
5011 goto err_free_put_context;
5013 event_file = fget_light(err, &fput_needed2);
5015 goto err_free_put_context;
5017 if (flags & PERF_FLAG_FD_OUTPUT) {
5018 err = perf_event_set_output(event, group_fd);
5020 goto err_fput_free_put_context;
5023 event->filp = event_file;
5024 WARN_ON_ONCE(ctx->parent_ctx);
5025 mutex_lock(&ctx->mutex);
5026 perf_install_in_context(ctx, event, cpu);
5028 mutex_unlock(&ctx->mutex);
5030 event->owner = current;
5031 get_task_struct(current);
5032 mutex_lock(¤t->perf_event_mutex);
5033 list_add_tail(&event->owner_entry, ¤t->perf_event_list);
5034 mutex_unlock(¤t->perf_event_mutex);
5036 err_fput_free_put_context:
5037 fput_light(event_file, fput_needed2);
5039 err_free_put_context:
5047 fput_light(group_file, fput_needed);
5053 * perf_event_create_kernel_counter
5055 * @attr: attributes of the counter to create
5056 * @cpu: cpu in which the counter is bound
5057 * @pid: task to profile
5060 perf_event_create_kernel_counter(struct perf_event_attr *attr, int cpu,
5062 perf_overflow_handler_t overflow_handler)
5064 struct perf_event *event;
5065 struct perf_event_context *ctx;
5069 * Get the target context (task or percpu):
5072 ctx = find_get_context(pid, cpu);
5078 event = perf_event_alloc(attr, cpu, ctx, NULL,
5079 NULL, overflow_handler, GFP_KERNEL);
5080 if (IS_ERR(event)) {
5081 err = PTR_ERR(event);
5082 goto err_put_context;
5086 WARN_ON_ONCE(ctx->parent_ctx);
5087 mutex_lock(&ctx->mutex);
5088 perf_install_in_context(ctx, event, cpu);
5090 mutex_unlock(&ctx->mutex);
5092 event->owner = current;
5093 get_task_struct(current);
5094 mutex_lock(¤t->perf_event_mutex);
5095 list_add_tail(&event->owner_entry, ¤t->perf_event_list);
5096 mutex_unlock(¤t->perf_event_mutex);
5103 return ERR_PTR(err);
5105 EXPORT_SYMBOL_GPL(perf_event_create_kernel_counter);
5108 * inherit a event from parent task to child task:
5110 static struct perf_event *
5111 inherit_event(struct perf_event *parent_event,
5112 struct task_struct *parent,
5113 struct perf_event_context *parent_ctx,
5114 struct task_struct *child,
5115 struct perf_event *group_leader,
5116 struct perf_event_context *child_ctx)
5118 struct perf_event *child_event;
5121 * Instead of creating recursive hierarchies of events,
5122 * we link inherited events back to the original parent,
5123 * which has a filp for sure, which we use as the reference
5126 if (parent_event->parent)
5127 parent_event = parent_event->parent;
5129 child_event = perf_event_alloc(&parent_event->attr,
5130 parent_event->cpu, child_ctx,
5131 group_leader, parent_event,
5133 if (IS_ERR(child_event))
5138 * Make the child state follow the state of the parent event,
5139 * not its attr.disabled bit. We hold the parent's mutex,
5140 * so we won't race with perf_event_{en, dis}able_family.
5142 if (parent_event->state >= PERF_EVENT_STATE_INACTIVE)
5143 child_event->state = PERF_EVENT_STATE_INACTIVE;
5145 child_event->state = PERF_EVENT_STATE_OFF;
5147 if (parent_event->attr.freq) {
5148 u64 sample_period = parent_event->hw.sample_period;
5149 struct hw_perf_event *hwc = &child_event->hw;
5151 hwc->sample_period = sample_period;
5152 hwc->last_period = sample_period;
5154 atomic64_set(&hwc->period_left, sample_period);
5157 child_event->overflow_handler = parent_event->overflow_handler;
5160 * Link it up in the child's context:
5162 add_event_to_ctx(child_event, child_ctx);
5165 * Get a reference to the parent filp - we will fput it
5166 * when the child event exits. This is safe to do because
5167 * we are in the parent and we know that the filp still
5168 * exists and has a nonzero count:
5170 atomic_long_inc(&parent_event->filp->f_count);
5173 * Link this into the parent event's child list
5175 WARN_ON_ONCE(parent_event->ctx->parent_ctx);
5176 mutex_lock(&parent_event->child_mutex);
5177 list_add_tail(&child_event->child_list, &parent_event->child_list);
5178 mutex_unlock(&parent_event->child_mutex);
5183 static int inherit_group(struct perf_event *parent_event,
5184 struct task_struct *parent,
5185 struct perf_event_context *parent_ctx,
5186 struct task_struct *child,
5187 struct perf_event_context *child_ctx)
5189 struct perf_event *leader;
5190 struct perf_event *sub;
5191 struct perf_event *child_ctr;
5193 leader = inherit_event(parent_event, parent, parent_ctx,
5194 child, NULL, child_ctx);
5196 return PTR_ERR(leader);
5197 list_for_each_entry(sub, &parent_event->sibling_list, group_entry) {
5198 child_ctr = inherit_event(sub, parent, parent_ctx,
5199 child, leader, child_ctx);
5200 if (IS_ERR(child_ctr))
5201 return PTR_ERR(child_ctr);
5206 static void sync_child_event(struct perf_event *child_event,
5207 struct task_struct *child)
5209 struct perf_event *parent_event = child_event->parent;
5212 if (child_event->attr.inherit_stat)
5213 perf_event_read_event(child_event, child);
5215 child_val = atomic64_read(&child_event->count);
5218 * Add back the child's count to the parent's count:
5220 atomic64_add(child_val, &parent_event->count);
5221 atomic64_add(child_event->total_time_enabled,
5222 &parent_event->child_total_time_enabled);
5223 atomic64_add(child_event->total_time_running,
5224 &parent_event->child_total_time_running);
5227 * Remove this event from the parent's list
5229 WARN_ON_ONCE(parent_event->ctx->parent_ctx);
5230 mutex_lock(&parent_event->child_mutex);
5231 list_del_init(&child_event->child_list);
5232 mutex_unlock(&parent_event->child_mutex);
5235 * Release the parent event, if this was the last
5238 fput(parent_event->filp);
5242 __perf_event_exit_task(struct perf_event *child_event,
5243 struct perf_event_context *child_ctx,
5244 struct task_struct *child)
5246 struct perf_event *parent_event;
5248 perf_event_remove_from_context(child_event);
5250 parent_event = child_event->parent;
5252 * It can happen that parent exits first, and has events
5253 * that are still around due to the child reference. These
5254 * events need to be zapped - but otherwise linger.
5257 sync_child_event(child_event, child);
5258 free_event(child_event);
5263 * When a child task exits, feed back event values to parent events.
5265 void perf_event_exit_task(struct task_struct *child)
5267 struct perf_event *child_event, *tmp;
5268 struct perf_event_context *child_ctx;
5269 unsigned long flags;
5271 if (likely(!child->perf_event_ctxp)) {
5272 perf_event_task(child, NULL, 0);
5276 local_irq_save(flags);
5278 * We can't reschedule here because interrupts are disabled,
5279 * and either child is current or it is a task that can't be
5280 * scheduled, so we are now safe from rescheduling changing
5283 child_ctx = child->perf_event_ctxp;
5284 __perf_event_task_sched_out(child_ctx);
5287 * Take the context lock here so that if find_get_context is
5288 * reading child->perf_event_ctxp, we wait until it has
5289 * incremented the context's refcount before we do put_ctx below.
5291 raw_spin_lock(&child_ctx->lock);
5292 child->perf_event_ctxp = NULL;
5294 * If this context is a clone; unclone it so it can't get
5295 * swapped to another process while we're removing all
5296 * the events from it.
5298 unclone_ctx(child_ctx);
5299 update_context_time(child_ctx);
5300 raw_spin_unlock_irqrestore(&child_ctx->lock, flags);
5303 * Report the task dead after unscheduling the events so that we
5304 * won't get any samples after PERF_RECORD_EXIT. We can however still
5305 * get a few PERF_RECORD_READ events.
5307 perf_event_task(child, child_ctx, 0);
5310 * We can recurse on the same lock type through:
5312 * __perf_event_exit_task()
5313 * sync_child_event()
5314 * fput(parent_event->filp)
5316 * mutex_lock(&ctx->mutex)
5318 * But since its the parent context it won't be the same instance.
5320 mutex_lock(&child_ctx->mutex);
5323 list_for_each_entry_safe(child_event, tmp, &child_ctx->pinned_groups,
5325 __perf_event_exit_task(child_event, child_ctx, child);
5327 list_for_each_entry_safe(child_event, tmp, &child_ctx->flexible_groups,
5329 __perf_event_exit_task(child_event, child_ctx, child);
5332 * If the last event was a group event, it will have appended all
5333 * its siblings to the list, but we obtained 'tmp' before that which
5334 * will still point to the list head terminating the iteration.
5336 if (!list_empty(&child_ctx->pinned_groups) ||
5337 !list_empty(&child_ctx->flexible_groups))
5340 mutex_unlock(&child_ctx->mutex);
5345 static void perf_free_event(struct perf_event *event,
5346 struct perf_event_context *ctx)
5348 struct perf_event *parent = event->parent;
5350 if (WARN_ON_ONCE(!parent))
5353 mutex_lock(&parent->child_mutex);
5354 list_del_init(&event->child_list);
5355 mutex_unlock(&parent->child_mutex);
5359 list_del_event(event, ctx);
5364 * free an unexposed, unused context as created by inheritance by
5365 * init_task below, used by fork() in case of fail.
5367 void perf_event_free_task(struct task_struct *task)
5369 struct perf_event_context *ctx = task->perf_event_ctxp;
5370 struct perf_event *event, *tmp;
5375 mutex_lock(&ctx->mutex);
5377 list_for_each_entry_safe(event, tmp, &ctx->pinned_groups, group_entry)
5378 perf_free_event(event, ctx);
5380 list_for_each_entry_safe(event, tmp, &ctx->flexible_groups,
5382 perf_free_event(event, ctx);
5384 if (!list_empty(&ctx->pinned_groups) ||
5385 !list_empty(&ctx->flexible_groups))
5388 mutex_unlock(&ctx->mutex);
5394 inherit_task_group(struct perf_event *event, struct task_struct *parent,
5395 struct perf_event_context *parent_ctx,
5396 struct task_struct *child,
5400 struct perf_event_context *child_ctx = child->perf_event_ctxp;
5402 if (!event->attr.inherit) {
5409 * This is executed from the parent task context, so
5410 * inherit events that have been marked for cloning.
5411 * First allocate and initialize a context for the
5415 child_ctx = kzalloc(sizeof(struct perf_event_context),
5420 __perf_event_init_context(child_ctx, child);
5421 child->perf_event_ctxp = child_ctx;
5422 get_task_struct(child);
5425 ret = inherit_group(event, parent, parent_ctx,
5436 * Initialize the perf_event context in task_struct
5438 int perf_event_init_task(struct task_struct *child)
5440 struct perf_event_context *child_ctx, *parent_ctx;
5441 struct perf_event_context *cloned_ctx;
5442 struct perf_event *event;
5443 struct task_struct *parent = current;
5444 int inherited_all = 1;
5447 child->perf_event_ctxp = NULL;
5449 mutex_init(&child->perf_event_mutex);
5450 INIT_LIST_HEAD(&child->perf_event_list);
5452 if (likely(!parent->perf_event_ctxp))
5456 * If the parent's context is a clone, pin it so it won't get
5459 parent_ctx = perf_pin_task_context(parent);
5462 * No need to check if parent_ctx != NULL here; since we saw
5463 * it non-NULL earlier, the only reason for it to become NULL
5464 * is if we exit, and since we're currently in the middle of
5465 * a fork we can't be exiting at the same time.
5469 * Lock the parent list. No need to lock the child - not PID
5470 * hashed yet and not running, so nobody can access it.
5472 mutex_lock(&parent_ctx->mutex);
5475 * We dont have to disable NMIs - we are only looking at
5476 * the list, not manipulating it:
5478 list_for_each_entry(event, &parent_ctx->pinned_groups, group_entry) {
5479 ret = inherit_task_group(event, parent, parent_ctx, child,
5485 list_for_each_entry(event, &parent_ctx->flexible_groups, group_entry) {
5486 ret = inherit_task_group(event, parent, parent_ctx, child,
5492 child_ctx = child->perf_event_ctxp;
5494 if (child_ctx && inherited_all) {
5496 * Mark the child context as a clone of the parent
5497 * context, or of whatever the parent is a clone of.
5498 * Note that if the parent is a clone, it could get
5499 * uncloned at any point, but that doesn't matter
5500 * because the list of events and the generation
5501 * count can't have changed since we took the mutex.
5503 cloned_ctx = rcu_dereference(parent_ctx->parent_ctx);
5505 child_ctx->parent_ctx = cloned_ctx;
5506 child_ctx->parent_gen = parent_ctx->parent_gen;
5508 child_ctx->parent_ctx = parent_ctx;
5509 child_ctx->parent_gen = parent_ctx->generation;
5511 get_ctx(child_ctx->parent_ctx);
5514 mutex_unlock(&parent_ctx->mutex);
5516 perf_unpin_context(parent_ctx);
5521 static void __init perf_event_init_all_cpus(void)
5524 struct perf_cpu_context *cpuctx;
5526 for_each_possible_cpu(cpu) {
5527 cpuctx = &per_cpu(perf_cpu_context, cpu);
5528 mutex_init(&cpuctx->hlist_mutex);
5529 __perf_event_init_context(&cpuctx->ctx, NULL);
5533 static void __cpuinit perf_event_init_cpu(int cpu)
5535 struct perf_cpu_context *cpuctx;
5537 cpuctx = &per_cpu(perf_cpu_context, cpu);
5539 spin_lock(&perf_resource_lock);
5540 cpuctx->max_pertask = perf_max_events - perf_reserved_percpu;
5541 spin_unlock(&perf_resource_lock);
5543 mutex_lock(&cpuctx->hlist_mutex);
5544 if (cpuctx->hlist_refcount > 0) {
5545 struct swevent_hlist *hlist;
5547 hlist = kzalloc(sizeof(*hlist), GFP_KERNEL);
5548 WARN_ON_ONCE(!hlist);
5549 rcu_assign_pointer(cpuctx->swevent_hlist, hlist);
5551 mutex_unlock(&cpuctx->hlist_mutex);
5554 #ifdef CONFIG_HOTPLUG_CPU
5555 static void __perf_event_exit_cpu(void *info)
5557 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
5558 struct perf_event_context *ctx = &cpuctx->ctx;
5559 struct perf_event *event, *tmp;
5561 list_for_each_entry_safe(event, tmp, &ctx->pinned_groups, group_entry)
5562 __perf_event_remove_from_context(event);
5563 list_for_each_entry_safe(event, tmp, &ctx->flexible_groups, group_entry)
5564 __perf_event_remove_from_context(event);
5566 static void perf_event_exit_cpu(int cpu)
5568 struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
5569 struct perf_event_context *ctx = &cpuctx->ctx;
5571 mutex_lock(&cpuctx->hlist_mutex);
5572 swevent_hlist_release(cpuctx);
5573 mutex_unlock(&cpuctx->hlist_mutex);
5575 mutex_lock(&ctx->mutex);
5576 smp_call_function_single(cpu, __perf_event_exit_cpu, NULL, 1);
5577 mutex_unlock(&ctx->mutex);
5580 static inline void perf_event_exit_cpu(int cpu) { }
5583 static int __cpuinit
5584 perf_cpu_notify(struct notifier_block *self, unsigned long action, void *hcpu)
5586 unsigned int cpu = (long)hcpu;
5590 case CPU_UP_PREPARE:
5591 case CPU_UP_PREPARE_FROZEN:
5592 perf_event_init_cpu(cpu);
5595 case CPU_DOWN_PREPARE:
5596 case CPU_DOWN_PREPARE_FROZEN:
5597 perf_event_exit_cpu(cpu);
5608 * This has to have a higher priority than migration_notifier in sched.c.
5610 static struct notifier_block __cpuinitdata perf_cpu_nb = {
5611 .notifier_call = perf_cpu_notify,
5615 void __init perf_event_init(void)
5617 perf_event_init_all_cpus();
5618 perf_cpu_notify(&perf_cpu_nb, (unsigned long)CPU_UP_PREPARE,
5619 (void *)(long)smp_processor_id());
5620 perf_cpu_notify(&perf_cpu_nb, (unsigned long)CPU_ONLINE,
5621 (void *)(long)smp_processor_id());
5622 register_cpu_notifier(&perf_cpu_nb);
5625 static ssize_t perf_show_reserve_percpu(struct sysdev_class *class,
5626 struct sysdev_class_attribute *attr,
5629 return sprintf(buf, "%d\n", perf_reserved_percpu);
5633 perf_set_reserve_percpu(struct sysdev_class *class,
5634 struct sysdev_class_attribute *attr,
5638 struct perf_cpu_context *cpuctx;
5642 err = strict_strtoul(buf, 10, &val);
5645 if (val > perf_max_events)
5648 spin_lock(&perf_resource_lock);
5649 perf_reserved_percpu = val;
5650 for_each_online_cpu(cpu) {
5651 cpuctx = &per_cpu(perf_cpu_context, cpu);
5652 raw_spin_lock_irq(&cpuctx->ctx.lock);
5653 mpt = min(perf_max_events - cpuctx->ctx.nr_events,
5654 perf_max_events - perf_reserved_percpu);
5655 cpuctx->max_pertask = mpt;
5656 raw_spin_unlock_irq(&cpuctx->ctx.lock);
5658 spin_unlock(&perf_resource_lock);
5663 static ssize_t perf_show_overcommit(struct sysdev_class *class,
5664 struct sysdev_class_attribute *attr,
5667 return sprintf(buf, "%d\n", perf_overcommit);
5671 perf_set_overcommit(struct sysdev_class *class,
5672 struct sysdev_class_attribute *attr,
5673 const char *buf, size_t count)
5678 err = strict_strtoul(buf, 10, &val);
5684 spin_lock(&perf_resource_lock);
5685 perf_overcommit = val;
5686 spin_unlock(&perf_resource_lock);
5691 static SYSDEV_CLASS_ATTR(
5694 perf_show_reserve_percpu,
5695 perf_set_reserve_percpu
5698 static SYSDEV_CLASS_ATTR(
5701 perf_show_overcommit,
5705 static struct attribute *perfclass_attrs[] = {
5706 &attr_reserve_percpu.attr,
5707 &attr_overcommit.attr,
5711 static struct attribute_group perfclass_attr_group = {
5712 .attrs = perfclass_attrs,
5713 .name = "perf_events",
5716 static int __init perf_event_sysfs_init(void)
5718 return sysfs_create_group(&cpu_sysdev_class.kset.kobj,
5719 &perfclass_attr_group);
5721 device_initcall(perf_event_sysfs_init);