2 * Performance events core code:
4 * Copyright (C) 2008 Thomas Gleixner <tglx@linutronix.de>
5 * Copyright (C) 2008-2009 Red Hat, Inc., Ingo Molnar
6 * Copyright (C) 2008-2009 Red Hat, Inc., Peter Zijlstra <pzijlstr@redhat.com>
7 * Copyright © 2009 Paul Mackerras, IBM Corp. <paulus@au1.ibm.com>
9 * For licensing details see kernel-base/COPYING
14 #include <linux/cpu.h>
15 #include <linux/smp.h>
16 #include <linux/file.h>
17 #include <linux/poll.h>
18 #include <linux/slab.h>
19 #include <linux/hash.h>
20 #include <linux/sysfs.h>
21 #include <linux/dcache.h>
22 #include <linux/percpu.h>
23 #include <linux/ptrace.h>
24 #include <linux/vmstat.h>
25 #include <linux/vmalloc.h>
26 #include <linux/hardirq.h>
27 #include <linux/rculist.h>
28 #include <linux/uaccess.h>
29 #include <linux/syscalls.h>
30 #include <linux/anon_inodes.h>
31 #include <linux/kernel_stat.h>
32 #include <linux/perf_event.h>
33 #include <linux/ftrace_event.h>
34 #include <linux/hw_breakpoint.h>
36 #include <asm/irq_regs.h>
39 * Each CPU has a list of per CPU events:
41 static DEFINE_PER_CPU(struct perf_cpu_context, perf_cpu_context);
43 int perf_max_events __read_mostly = 1;
44 static int perf_reserved_percpu __read_mostly;
45 static int perf_overcommit __read_mostly = 1;
47 static atomic_t nr_events __read_mostly;
48 static atomic_t nr_mmap_events __read_mostly;
49 static atomic_t nr_comm_events __read_mostly;
50 static atomic_t nr_task_events __read_mostly;
53 * perf event paranoia level:
54 * -1 - not paranoid at all
55 * 0 - disallow raw tracepoint access for unpriv
56 * 1 - disallow cpu events for unpriv
57 * 2 - disallow kernel profiling for unpriv
59 int sysctl_perf_event_paranoid __read_mostly = 1;
61 int sysctl_perf_event_mlock __read_mostly = 512; /* 'free' kb per user */
64 * max perf event sample rate
66 int sysctl_perf_event_sample_rate __read_mostly = 100000;
68 static atomic64_t perf_event_id;
71 * Lock for (sysadmin-configurable) event reservations:
73 static DEFINE_SPINLOCK(perf_resource_lock);
76 * Architecture provided APIs - weak aliases:
78 extern __weak const struct pmu *hw_perf_event_init(struct perf_event *event)
83 void __weak hw_perf_disable(void) { barrier(); }
84 void __weak hw_perf_enable(void) { barrier(); }
86 void __weak perf_event_print_debug(void) { }
88 static DEFINE_PER_CPU(int, perf_disable_count);
90 void perf_disable(void)
92 if (!__get_cpu_var(perf_disable_count)++)
96 void perf_enable(void)
98 if (!--__get_cpu_var(perf_disable_count))
102 static void get_ctx(struct perf_event_context *ctx)
104 WARN_ON(!atomic_inc_not_zero(&ctx->refcount));
107 static void free_ctx(struct rcu_head *head)
109 struct perf_event_context *ctx;
111 ctx = container_of(head, struct perf_event_context, rcu_head);
115 static void put_ctx(struct perf_event_context *ctx)
117 if (atomic_dec_and_test(&ctx->refcount)) {
119 put_ctx(ctx->parent_ctx);
121 put_task_struct(ctx->task);
122 call_rcu(&ctx->rcu_head, free_ctx);
126 static void unclone_ctx(struct perf_event_context *ctx)
128 if (ctx->parent_ctx) {
129 put_ctx(ctx->parent_ctx);
130 ctx->parent_ctx = NULL;
135 * If we inherit events we want to return the parent event id
138 static u64 primary_event_id(struct perf_event *event)
143 id = event->parent->id;
149 * Get the perf_event_context for a task and lock it.
150 * This has to cope with with the fact that until it is locked,
151 * the context could get moved to another task.
153 static struct perf_event_context *
154 perf_lock_task_context(struct task_struct *task, unsigned long *flags)
156 struct perf_event_context *ctx;
160 ctx = rcu_dereference(task->perf_event_ctxp);
163 * If this context is a clone of another, it might
164 * get swapped for another underneath us by
165 * perf_event_task_sched_out, though the
166 * rcu_read_lock() protects us from any context
167 * getting freed. Lock the context and check if it
168 * got swapped before we could get the lock, and retry
169 * if so. If we locked the right context, then it
170 * can't get swapped on us any more.
172 raw_spin_lock_irqsave(&ctx->lock, *flags);
173 if (ctx != rcu_dereference(task->perf_event_ctxp)) {
174 raw_spin_unlock_irqrestore(&ctx->lock, *flags);
178 if (!atomic_inc_not_zero(&ctx->refcount)) {
179 raw_spin_unlock_irqrestore(&ctx->lock, *flags);
188 * Get the context for a task and increment its pin_count so it
189 * can't get swapped to another task. This also increments its
190 * reference count so that the context can't get freed.
192 static struct perf_event_context *perf_pin_task_context(struct task_struct *task)
194 struct perf_event_context *ctx;
197 ctx = perf_lock_task_context(task, &flags);
200 raw_spin_unlock_irqrestore(&ctx->lock, flags);
205 static void perf_unpin_context(struct perf_event_context *ctx)
209 raw_spin_lock_irqsave(&ctx->lock, flags);
211 raw_spin_unlock_irqrestore(&ctx->lock, flags);
215 static inline u64 perf_clock(void)
217 return cpu_clock(raw_smp_processor_id());
221 * Update the record of the current time in a context.
223 static void update_context_time(struct perf_event_context *ctx)
225 u64 now = perf_clock();
227 ctx->time += now - ctx->timestamp;
228 ctx->timestamp = now;
232 * Update the total_time_enabled and total_time_running fields for a event.
234 static void update_event_times(struct perf_event *event)
236 struct perf_event_context *ctx = event->ctx;
239 if (event->state < PERF_EVENT_STATE_INACTIVE ||
240 event->group_leader->state < PERF_EVENT_STATE_INACTIVE)
246 run_end = event->tstamp_stopped;
248 event->total_time_enabled = run_end - event->tstamp_enabled;
250 if (event->state == PERF_EVENT_STATE_INACTIVE)
251 run_end = event->tstamp_stopped;
255 event->total_time_running = run_end - event->tstamp_running;
258 static struct list_head *
259 ctx_group_list(struct perf_event *event, struct perf_event_context *ctx)
261 if (event->attr.pinned)
262 return &ctx->pinned_groups;
264 return &ctx->flexible_groups;
268 * Add a event from the lists for its context.
269 * Must be called with ctx->mutex and ctx->lock held.
272 list_add_event(struct perf_event *event, struct perf_event_context *ctx)
274 struct perf_event *group_leader = event->group_leader;
277 * Depending on whether it is a standalone or sibling event,
278 * add it straight to the context's event list, or to the group
279 * leader's sibling list:
281 if (group_leader == event) {
282 struct list_head *list;
284 if (is_software_event(event))
285 event->group_flags |= PERF_GROUP_SOFTWARE;
287 list = ctx_group_list(event, ctx);
288 list_add_tail(&event->group_entry, list);
290 if (group_leader->group_flags & PERF_GROUP_SOFTWARE &&
291 !is_software_event(event))
292 group_leader->group_flags &= ~PERF_GROUP_SOFTWARE;
294 list_add_tail(&event->group_entry, &group_leader->sibling_list);
295 group_leader->nr_siblings++;
298 list_add_rcu(&event->event_entry, &ctx->event_list);
300 if (event->attr.inherit_stat)
305 * Remove a event from the lists for its context.
306 * Must be called with ctx->mutex and ctx->lock held.
309 list_del_event(struct perf_event *event, struct perf_event_context *ctx)
311 struct perf_event *sibling, *tmp;
313 if (list_empty(&event->group_entry))
316 if (event->attr.inherit_stat)
319 list_del_init(&event->group_entry);
320 list_del_rcu(&event->event_entry);
322 if (event->group_leader != event)
323 event->group_leader->nr_siblings--;
325 update_event_times(event);
328 * If event was in error state, then keep it
329 * that way, otherwise bogus counts will be
330 * returned on read(). The only way to get out
331 * of error state is by explicit re-enabling
334 if (event->state > PERF_EVENT_STATE_OFF)
335 event->state = PERF_EVENT_STATE_OFF;
337 if (event->state > PERF_EVENT_STATE_FREE)
341 * If this was a group event with sibling events then
342 * upgrade the siblings to singleton events by adding them
343 * to the context list directly:
345 list_for_each_entry_safe(sibling, tmp, &event->sibling_list, group_entry) {
346 struct list_head *list;
348 list = ctx_group_list(event, ctx);
349 list_move_tail(&sibling->group_entry, list);
350 sibling->group_leader = sibling;
352 /* Inherit group flags from the previous leader */
353 sibling->group_flags = event->group_flags;
358 event_sched_out(struct perf_event *event,
359 struct perf_cpu_context *cpuctx,
360 struct perf_event_context *ctx)
362 if (event->state != PERF_EVENT_STATE_ACTIVE)
365 event->state = PERF_EVENT_STATE_INACTIVE;
366 if (event->pending_disable) {
367 event->pending_disable = 0;
368 event->state = PERF_EVENT_STATE_OFF;
370 event->tstamp_stopped = ctx->time;
371 event->pmu->disable(event);
374 if (!is_software_event(event))
375 cpuctx->active_oncpu--;
377 if (event->attr.exclusive || !cpuctx->active_oncpu)
378 cpuctx->exclusive = 0;
382 group_sched_out(struct perf_event *group_event,
383 struct perf_cpu_context *cpuctx,
384 struct perf_event_context *ctx)
386 struct perf_event *event;
388 if (group_event->state != PERF_EVENT_STATE_ACTIVE)
391 event_sched_out(group_event, cpuctx, ctx);
394 * Schedule out siblings (if any):
396 list_for_each_entry(event, &group_event->sibling_list, group_entry)
397 event_sched_out(event, cpuctx, ctx);
399 if (group_event->attr.exclusive)
400 cpuctx->exclusive = 0;
404 * Cross CPU call to remove a performance event
406 * We disable the event on the hardware level first. After that we
407 * remove it from the context list.
409 static void __perf_event_remove_from_context(void *info)
411 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
412 struct perf_event *event = info;
413 struct perf_event_context *ctx = event->ctx;
416 * If this is a task context, we need to check whether it is
417 * the current task context of this cpu. If not it has been
418 * scheduled out before the smp call arrived.
420 if (ctx->task && cpuctx->task_ctx != ctx)
423 raw_spin_lock(&ctx->lock);
425 * Protect the list operation against NMI by disabling the
426 * events on a global level.
430 event_sched_out(event, cpuctx, ctx);
432 list_del_event(event, ctx);
436 * Allow more per task events with respect to the
439 cpuctx->max_pertask =
440 min(perf_max_events - ctx->nr_events,
441 perf_max_events - perf_reserved_percpu);
445 raw_spin_unlock(&ctx->lock);
450 * Remove the event from a task's (or a CPU's) list of events.
452 * Must be called with ctx->mutex held.
454 * CPU events are removed with a smp call. For task events we only
455 * call when the task is on a CPU.
457 * If event->ctx is a cloned context, callers must make sure that
458 * every task struct that event->ctx->task could possibly point to
459 * remains valid. This is OK when called from perf_release since
460 * that only calls us on the top-level context, which can't be a clone.
461 * When called from perf_event_exit_task, it's OK because the
462 * context has been detached from its task.
464 static void perf_event_remove_from_context(struct perf_event *event)
466 struct perf_event_context *ctx = event->ctx;
467 struct task_struct *task = ctx->task;
471 * Per cpu events are removed via an smp call and
472 * the removal is always successful.
474 smp_call_function_single(event->cpu,
475 __perf_event_remove_from_context,
481 task_oncpu_function_call(task, __perf_event_remove_from_context,
484 raw_spin_lock_irq(&ctx->lock);
486 * If the context is active we need to retry the smp call.
488 if (ctx->nr_active && !list_empty(&event->group_entry)) {
489 raw_spin_unlock_irq(&ctx->lock);
494 * The lock prevents that this context is scheduled in so we
495 * can remove the event safely, if the call above did not
498 if (!list_empty(&event->group_entry))
499 list_del_event(event, ctx);
500 raw_spin_unlock_irq(&ctx->lock);
504 * Update total_time_enabled and total_time_running for all events in a group.
506 static void update_group_times(struct perf_event *leader)
508 struct perf_event *event;
510 update_event_times(leader);
511 list_for_each_entry(event, &leader->sibling_list, group_entry)
512 update_event_times(event);
516 * Cross CPU call to disable a performance event
518 static void __perf_event_disable(void *info)
520 struct perf_event *event = info;
521 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
522 struct perf_event_context *ctx = event->ctx;
525 * If this is a per-task event, need to check whether this
526 * event's task is the current task on this cpu.
528 if (ctx->task && cpuctx->task_ctx != ctx)
531 raw_spin_lock(&ctx->lock);
534 * If the event is on, turn it off.
535 * If it is in error state, leave it in error state.
537 if (event->state >= PERF_EVENT_STATE_INACTIVE) {
538 update_context_time(ctx);
539 update_group_times(event);
540 if (event == event->group_leader)
541 group_sched_out(event, cpuctx, ctx);
543 event_sched_out(event, cpuctx, ctx);
544 event->state = PERF_EVENT_STATE_OFF;
547 raw_spin_unlock(&ctx->lock);
553 * If event->ctx is a cloned context, callers must make sure that
554 * every task struct that event->ctx->task could possibly point to
555 * remains valid. This condition is satisifed when called through
556 * perf_event_for_each_child or perf_event_for_each because they
557 * hold the top-level event's child_mutex, so any descendant that
558 * goes to exit will block in sync_child_event.
559 * When called from perf_pending_event it's OK because event->ctx
560 * is the current context on this CPU and preemption is disabled,
561 * hence we can't get into perf_event_task_sched_out for this context.
563 void perf_event_disable(struct perf_event *event)
565 struct perf_event_context *ctx = event->ctx;
566 struct task_struct *task = ctx->task;
570 * Disable the event on the cpu that it's on
572 smp_call_function_single(event->cpu, __perf_event_disable,
578 task_oncpu_function_call(task, __perf_event_disable, event);
580 raw_spin_lock_irq(&ctx->lock);
582 * If the event is still active, we need to retry the cross-call.
584 if (event->state == PERF_EVENT_STATE_ACTIVE) {
585 raw_spin_unlock_irq(&ctx->lock);
590 * Since we have the lock this context can't be scheduled
591 * in, so we can change the state safely.
593 if (event->state == PERF_EVENT_STATE_INACTIVE) {
594 update_group_times(event);
595 event->state = PERF_EVENT_STATE_OFF;
598 raw_spin_unlock_irq(&ctx->lock);
602 event_sched_in(struct perf_event *event,
603 struct perf_cpu_context *cpuctx,
604 struct perf_event_context *ctx)
606 if (event->state <= PERF_EVENT_STATE_OFF)
609 event->state = PERF_EVENT_STATE_ACTIVE;
610 event->oncpu = smp_processor_id();
612 * The new state must be visible before we turn it on in the hardware:
616 if (event->pmu->enable(event)) {
617 event->state = PERF_EVENT_STATE_INACTIVE;
622 event->tstamp_running += ctx->time - event->tstamp_stopped;
624 if (!is_software_event(event))
625 cpuctx->active_oncpu++;
628 if (event->attr.exclusive)
629 cpuctx->exclusive = 1;
635 group_sched_in(struct perf_event *group_event,
636 struct perf_cpu_context *cpuctx,
637 struct perf_event_context *ctx)
639 struct perf_event *event, *partial_group = NULL;
640 const struct pmu *pmu = group_event->pmu;
644 if (group_event->state == PERF_EVENT_STATE_OFF)
647 /* Check if group transaction availabe */
654 if (event_sched_in(group_event, cpuctx, ctx))
658 * Schedule in siblings as one group (if any):
660 list_for_each_entry(event, &group_event->sibling_list, group_entry) {
661 if (event_sched_in(event, cpuctx, ctx)) {
662 partial_group = event;
670 ret = pmu->commit_txn(pmu);
672 pmu->cancel_txn(pmu);
678 pmu->cancel_txn(pmu);
681 * Groups can be scheduled in as one unit only, so undo any
682 * partial group before returning:
684 list_for_each_entry(event, &group_event->sibling_list, group_entry) {
685 if (event == partial_group)
687 event_sched_out(event, cpuctx, ctx);
689 event_sched_out(group_event, cpuctx, ctx);
695 * Work out whether we can put this event group on the CPU now.
697 static int group_can_go_on(struct perf_event *event,
698 struct perf_cpu_context *cpuctx,
702 * Groups consisting entirely of software events can always go on.
704 if (event->group_flags & PERF_GROUP_SOFTWARE)
707 * If an exclusive group is already on, no other hardware
710 if (cpuctx->exclusive)
713 * If this group is exclusive and there are already
714 * events on the CPU, it can't go on.
716 if (event->attr.exclusive && cpuctx->active_oncpu)
719 * Otherwise, try to add it if all previous groups were able
725 static void add_event_to_ctx(struct perf_event *event,
726 struct perf_event_context *ctx)
728 list_add_event(event, ctx);
729 event->tstamp_enabled = ctx->time;
730 event->tstamp_running = ctx->time;
731 event->tstamp_stopped = ctx->time;
735 * Cross CPU call to install and enable a performance event
737 * Must be called with ctx->mutex held
739 static void __perf_install_in_context(void *info)
741 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
742 struct perf_event *event = info;
743 struct perf_event_context *ctx = event->ctx;
744 struct perf_event *leader = event->group_leader;
748 * If this is a task context, we need to check whether it is
749 * the current task context of this cpu. If not it has been
750 * scheduled out before the smp call arrived.
751 * Or possibly this is the right context but it isn't
752 * on this cpu because it had no events.
754 if (ctx->task && cpuctx->task_ctx != ctx) {
755 if (cpuctx->task_ctx || ctx->task != current)
757 cpuctx->task_ctx = ctx;
760 raw_spin_lock(&ctx->lock);
762 update_context_time(ctx);
765 * Protect the list operation against NMI by disabling the
766 * events on a global level. NOP for non NMI based events.
770 add_event_to_ctx(event, ctx);
772 if (event->cpu != -1 && event->cpu != smp_processor_id())
776 * Don't put the event on if it is disabled or if
777 * it is in a group and the group isn't on.
779 if (event->state != PERF_EVENT_STATE_INACTIVE ||
780 (leader != event && leader->state != PERF_EVENT_STATE_ACTIVE))
784 * An exclusive event can't go on if there are already active
785 * hardware events, and no hardware event can go on if there
786 * is already an exclusive event on.
788 if (!group_can_go_on(event, cpuctx, 1))
791 err = event_sched_in(event, cpuctx, ctx);
795 * This event couldn't go on. If it is in a group
796 * then we have to pull the whole group off.
797 * If the event group is pinned then put it in error state.
800 group_sched_out(leader, cpuctx, ctx);
801 if (leader->attr.pinned) {
802 update_group_times(leader);
803 leader->state = PERF_EVENT_STATE_ERROR;
807 if (!err && !ctx->task && cpuctx->max_pertask)
808 cpuctx->max_pertask--;
813 raw_spin_unlock(&ctx->lock);
817 * Attach a performance event to a context
819 * First we add the event to the list with the hardware enable bit
820 * in event->hw_config cleared.
822 * If the event is attached to a task which is on a CPU we use a smp
823 * call to enable it in the task context. The task might have been
824 * scheduled away, but we check this in the smp call again.
826 * Must be called with ctx->mutex held.
829 perf_install_in_context(struct perf_event_context *ctx,
830 struct perf_event *event,
833 struct task_struct *task = ctx->task;
837 * Per cpu events are installed via an smp call and
838 * the install is always successful.
840 smp_call_function_single(cpu, __perf_install_in_context,
846 task_oncpu_function_call(task, __perf_install_in_context,
849 raw_spin_lock_irq(&ctx->lock);
851 * we need to retry the smp call.
853 if (ctx->is_active && list_empty(&event->group_entry)) {
854 raw_spin_unlock_irq(&ctx->lock);
859 * The lock prevents that this context is scheduled in so we
860 * can add the event safely, if it the call above did not
863 if (list_empty(&event->group_entry))
864 add_event_to_ctx(event, ctx);
865 raw_spin_unlock_irq(&ctx->lock);
869 * Put a event into inactive state and update time fields.
870 * Enabling the leader of a group effectively enables all
871 * the group members that aren't explicitly disabled, so we
872 * have to update their ->tstamp_enabled also.
873 * Note: this works for group members as well as group leaders
874 * since the non-leader members' sibling_lists will be empty.
876 static void __perf_event_mark_enabled(struct perf_event *event,
877 struct perf_event_context *ctx)
879 struct perf_event *sub;
881 event->state = PERF_EVENT_STATE_INACTIVE;
882 event->tstamp_enabled = ctx->time - event->total_time_enabled;
883 list_for_each_entry(sub, &event->sibling_list, group_entry)
884 if (sub->state >= PERF_EVENT_STATE_INACTIVE)
885 sub->tstamp_enabled =
886 ctx->time - sub->total_time_enabled;
890 * Cross CPU call to enable a performance event
892 static void __perf_event_enable(void *info)
894 struct perf_event *event = info;
895 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
896 struct perf_event_context *ctx = event->ctx;
897 struct perf_event *leader = event->group_leader;
901 * If this is a per-task event, need to check whether this
902 * event's task is the current task on this cpu.
904 if (ctx->task && cpuctx->task_ctx != ctx) {
905 if (cpuctx->task_ctx || ctx->task != current)
907 cpuctx->task_ctx = ctx;
910 raw_spin_lock(&ctx->lock);
912 update_context_time(ctx);
914 if (event->state >= PERF_EVENT_STATE_INACTIVE)
916 __perf_event_mark_enabled(event, ctx);
918 if (event->cpu != -1 && event->cpu != smp_processor_id())
922 * If the event is in a group and isn't the group leader,
923 * then don't put it on unless the group is on.
925 if (leader != event && leader->state != PERF_EVENT_STATE_ACTIVE)
928 if (!group_can_go_on(event, cpuctx, 1)) {
933 err = group_sched_in(event, cpuctx, ctx);
935 err = event_sched_in(event, cpuctx, ctx);
941 * If this event can't go on and it's part of a
942 * group, then the whole group has to come off.
945 group_sched_out(leader, cpuctx, ctx);
946 if (leader->attr.pinned) {
947 update_group_times(leader);
948 leader->state = PERF_EVENT_STATE_ERROR;
953 raw_spin_unlock(&ctx->lock);
959 * If event->ctx is a cloned context, callers must make sure that
960 * every task struct that event->ctx->task could possibly point to
961 * remains valid. This condition is satisfied when called through
962 * perf_event_for_each_child or perf_event_for_each as described
963 * for perf_event_disable.
965 void perf_event_enable(struct perf_event *event)
967 struct perf_event_context *ctx = event->ctx;
968 struct task_struct *task = ctx->task;
972 * Enable the event on the cpu that it's on
974 smp_call_function_single(event->cpu, __perf_event_enable,
979 raw_spin_lock_irq(&ctx->lock);
980 if (event->state >= PERF_EVENT_STATE_INACTIVE)
984 * If the event is in error state, clear that first.
985 * That way, if we see the event in error state below, we
986 * know that it has gone back into error state, as distinct
987 * from the task having been scheduled away before the
988 * cross-call arrived.
990 if (event->state == PERF_EVENT_STATE_ERROR)
991 event->state = PERF_EVENT_STATE_OFF;
994 raw_spin_unlock_irq(&ctx->lock);
995 task_oncpu_function_call(task, __perf_event_enable, event);
997 raw_spin_lock_irq(&ctx->lock);
1000 * If the context is active and the event is still off,
1001 * we need to retry the cross-call.
1003 if (ctx->is_active && event->state == PERF_EVENT_STATE_OFF)
1007 * Since we have the lock this context can't be scheduled
1008 * in, so we can change the state safely.
1010 if (event->state == PERF_EVENT_STATE_OFF)
1011 __perf_event_mark_enabled(event, ctx);
1014 raw_spin_unlock_irq(&ctx->lock);
1017 static int perf_event_refresh(struct perf_event *event, int refresh)
1020 * not supported on inherited events
1022 if (event->attr.inherit)
1025 atomic_add(refresh, &event->event_limit);
1026 perf_event_enable(event);
1032 EVENT_FLEXIBLE = 0x1,
1034 EVENT_ALL = EVENT_FLEXIBLE | EVENT_PINNED,
1037 static void ctx_sched_out(struct perf_event_context *ctx,
1038 struct perf_cpu_context *cpuctx,
1039 enum event_type_t event_type)
1041 struct perf_event *event;
1043 raw_spin_lock(&ctx->lock);
1045 if (likely(!ctx->nr_events))
1047 update_context_time(ctx);
1050 if (!ctx->nr_active)
1053 if (event_type & EVENT_PINNED)
1054 list_for_each_entry(event, &ctx->pinned_groups, group_entry)
1055 group_sched_out(event, cpuctx, ctx);
1057 if (event_type & EVENT_FLEXIBLE)
1058 list_for_each_entry(event, &ctx->flexible_groups, group_entry)
1059 group_sched_out(event, cpuctx, ctx);
1064 raw_spin_unlock(&ctx->lock);
1068 * Test whether two contexts are equivalent, i.e. whether they
1069 * have both been cloned from the same version of the same context
1070 * and they both have the same number of enabled events.
1071 * If the number of enabled events is the same, then the set
1072 * of enabled events should be the same, because these are both
1073 * inherited contexts, therefore we can't access individual events
1074 * in them directly with an fd; we can only enable/disable all
1075 * events via prctl, or enable/disable all events in a family
1076 * via ioctl, which will have the same effect on both contexts.
1078 static int context_equiv(struct perf_event_context *ctx1,
1079 struct perf_event_context *ctx2)
1081 return ctx1->parent_ctx && ctx1->parent_ctx == ctx2->parent_ctx
1082 && ctx1->parent_gen == ctx2->parent_gen
1083 && !ctx1->pin_count && !ctx2->pin_count;
1086 static void __perf_event_sync_stat(struct perf_event *event,
1087 struct perf_event *next_event)
1091 if (!event->attr.inherit_stat)
1095 * Update the event value, we cannot use perf_event_read()
1096 * because we're in the middle of a context switch and have IRQs
1097 * disabled, which upsets smp_call_function_single(), however
1098 * we know the event must be on the current CPU, therefore we
1099 * don't need to use it.
1101 switch (event->state) {
1102 case PERF_EVENT_STATE_ACTIVE:
1103 event->pmu->read(event);
1106 case PERF_EVENT_STATE_INACTIVE:
1107 update_event_times(event);
1115 * In order to keep per-task stats reliable we need to flip the event
1116 * values when we flip the contexts.
1118 value = atomic64_read(&next_event->count);
1119 value = atomic64_xchg(&event->count, value);
1120 atomic64_set(&next_event->count, value);
1122 swap(event->total_time_enabled, next_event->total_time_enabled);
1123 swap(event->total_time_running, next_event->total_time_running);
1126 * Since we swizzled the values, update the user visible data too.
1128 perf_event_update_userpage(event);
1129 perf_event_update_userpage(next_event);
1132 #define list_next_entry(pos, member) \
1133 list_entry(pos->member.next, typeof(*pos), member)
1135 static void perf_event_sync_stat(struct perf_event_context *ctx,
1136 struct perf_event_context *next_ctx)
1138 struct perf_event *event, *next_event;
1143 update_context_time(ctx);
1145 event = list_first_entry(&ctx->event_list,
1146 struct perf_event, event_entry);
1148 next_event = list_first_entry(&next_ctx->event_list,
1149 struct perf_event, event_entry);
1151 while (&event->event_entry != &ctx->event_list &&
1152 &next_event->event_entry != &next_ctx->event_list) {
1154 __perf_event_sync_stat(event, next_event);
1156 event = list_next_entry(event, event_entry);
1157 next_event = list_next_entry(next_event, event_entry);
1162 * Called from scheduler to remove the events of the current task,
1163 * with interrupts disabled.
1165 * We stop each event and update the event value in event->count.
1167 * This does not protect us against NMI, but disable()
1168 * sets the disabled bit in the control field of event _before_
1169 * accessing the event control register. If a NMI hits, then it will
1170 * not restart the event.
1172 void perf_event_task_sched_out(struct task_struct *task,
1173 struct task_struct *next)
1175 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
1176 struct perf_event_context *ctx = task->perf_event_ctxp;
1177 struct perf_event_context *next_ctx;
1178 struct perf_event_context *parent;
1181 perf_sw_event(PERF_COUNT_SW_CONTEXT_SWITCHES, 1, 1, NULL, 0);
1183 if (likely(!ctx || !cpuctx->task_ctx))
1187 parent = rcu_dereference(ctx->parent_ctx);
1188 next_ctx = next->perf_event_ctxp;
1189 if (parent && next_ctx &&
1190 rcu_dereference(next_ctx->parent_ctx) == parent) {
1192 * Looks like the two contexts are clones, so we might be
1193 * able to optimize the context switch. We lock both
1194 * contexts and check that they are clones under the
1195 * lock (including re-checking that neither has been
1196 * uncloned in the meantime). It doesn't matter which
1197 * order we take the locks because no other cpu could
1198 * be trying to lock both of these tasks.
1200 raw_spin_lock(&ctx->lock);
1201 raw_spin_lock_nested(&next_ctx->lock, SINGLE_DEPTH_NESTING);
1202 if (context_equiv(ctx, next_ctx)) {
1204 * XXX do we need a memory barrier of sorts
1205 * wrt to rcu_dereference() of perf_event_ctxp
1207 task->perf_event_ctxp = next_ctx;
1208 next->perf_event_ctxp = ctx;
1210 next_ctx->task = task;
1213 perf_event_sync_stat(ctx, next_ctx);
1215 raw_spin_unlock(&next_ctx->lock);
1216 raw_spin_unlock(&ctx->lock);
1221 ctx_sched_out(ctx, cpuctx, EVENT_ALL);
1222 cpuctx->task_ctx = NULL;
1226 static void task_ctx_sched_out(struct perf_event_context *ctx,
1227 enum event_type_t event_type)
1229 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
1231 if (!cpuctx->task_ctx)
1234 if (WARN_ON_ONCE(ctx != cpuctx->task_ctx))
1237 ctx_sched_out(ctx, cpuctx, event_type);
1238 cpuctx->task_ctx = NULL;
1242 * Called with IRQs disabled
1244 static void __perf_event_task_sched_out(struct perf_event_context *ctx)
1246 task_ctx_sched_out(ctx, EVENT_ALL);
1250 * Called with IRQs disabled
1252 static void cpu_ctx_sched_out(struct perf_cpu_context *cpuctx,
1253 enum event_type_t event_type)
1255 ctx_sched_out(&cpuctx->ctx, cpuctx, event_type);
1259 ctx_pinned_sched_in(struct perf_event_context *ctx,
1260 struct perf_cpu_context *cpuctx)
1262 struct perf_event *event;
1264 list_for_each_entry(event, &ctx->pinned_groups, group_entry) {
1265 if (event->state <= PERF_EVENT_STATE_OFF)
1267 if (event->cpu != -1 && event->cpu != smp_processor_id())
1270 if (group_can_go_on(event, cpuctx, 1))
1271 group_sched_in(event, cpuctx, ctx);
1274 * If this pinned group hasn't been scheduled,
1275 * put it in error state.
1277 if (event->state == PERF_EVENT_STATE_INACTIVE) {
1278 update_group_times(event);
1279 event->state = PERF_EVENT_STATE_ERROR;
1285 ctx_flexible_sched_in(struct perf_event_context *ctx,
1286 struct perf_cpu_context *cpuctx)
1288 struct perf_event *event;
1291 list_for_each_entry(event, &ctx->flexible_groups, group_entry) {
1292 /* Ignore events in OFF or ERROR state */
1293 if (event->state <= PERF_EVENT_STATE_OFF)
1296 * Listen to the 'cpu' scheduling filter constraint
1299 if (event->cpu != -1 && event->cpu != smp_processor_id())
1302 if (group_can_go_on(event, cpuctx, can_add_hw))
1303 if (group_sched_in(event, cpuctx, ctx))
1309 ctx_sched_in(struct perf_event_context *ctx,
1310 struct perf_cpu_context *cpuctx,
1311 enum event_type_t event_type)
1313 raw_spin_lock(&ctx->lock);
1315 if (likely(!ctx->nr_events))
1318 ctx->timestamp = perf_clock();
1323 * First go through the list and put on any pinned groups
1324 * in order to give them the best chance of going on.
1326 if (event_type & EVENT_PINNED)
1327 ctx_pinned_sched_in(ctx, cpuctx);
1329 /* Then walk through the lower prio flexible groups */
1330 if (event_type & EVENT_FLEXIBLE)
1331 ctx_flexible_sched_in(ctx, cpuctx);
1335 raw_spin_unlock(&ctx->lock);
1338 static void cpu_ctx_sched_in(struct perf_cpu_context *cpuctx,
1339 enum event_type_t event_type)
1341 struct perf_event_context *ctx = &cpuctx->ctx;
1343 ctx_sched_in(ctx, cpuctx, event_type);
1346 static void task_ctx_sched_in(struct task_struct *task,
1347 enum event_type_t event_type)
1349 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
1350 struct perf_event_context *ctx = task->perf_event_ctxp;
1354 if (cpuctx->task_ctx == ctx)
1356 ctx_sched_in(ctx, cpuctx, event_type);
1357 cpuctx->task_ctx = ctx;
1360 * Called from scheduler to add the events of the current task
1361 * with interrupts disabled.
1363 * We restore the event value and then enable it.
1365 * This does not protect us against NMI, but enable()
1366 * sets the enabled bit in the control field of event _before_
1367 * accessing the event control register. If a NMI hits, then it will
1368 * keep the event running.
1370 void perf_event_task_sched_in(struct task_struct *task)
1372 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
1373 struct perf_event_context *ctx = task->perf_event_ctxp;
1378 if (cpuctx->task_ctx == ctx)
1384 * We want to keep the following priority order:
1385 * cpu pinned (that don't need to move), task pinned,
1386 * cpu flexible, task flexible.
1388 cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE);
1390 ctx_sched_in(ctx, cpuctx, EVENT_PINNED);
1391 cpu_ctx_sched_in(cpuctx, EVENT_FLEXIBLE);
1392 ctx_sched_in(ctx, cpuctx, EVENT_FLEXIBLE);
1394 cpuctx->task_ctx = ctx;
1399 #define MAX_INTERRUPTS (~0ULL)
1401 static void perf_log_throttle(struct perf_event *event, int enable);
1403 static u64 perf_calculate_period(struct perf_event *event, u64 nsec, u64 count)
1405 u64 frequency = event->attr.sample_freq;
1406 u64 sec = NSEC_PER_SEC;
1407 u64 divisor, dividend;
1409 int count_fls, nsec_fls, frequency_fls, sec_fls;
1411 count_fls = fls64(count);
1412 nsec_fls = fls64(nsec);
1413 frequency_fls = fls64(frequency);
1417 * We got @count in @nsec, with a target of sample_freq HZ
1418 * the target period becomes:
1421 * period = -------------------
1422 * @nsec * sample_freq
1427 * Reduce accuracy by one bit such that @a and @b converge
1428 * to a similar magnitude.
1430 #define REDUCE_FLS(a, b) \
1432 if (a##_fls > b##_fls) { \
1442 * Reduce accuracy until either term fits in a u64, then proceed with
1443 * the other, so that finally we can do a u64/u64 division.
1445 while (count_fls + sec_fls > 64 && nsec_fls + frequency_fls > 64) {
1446 REDUCE_FLS(nsec, frequency);
1447 REDUCE_FLS(sec, count);
1450 if (count_fls + sec_fls > 64) {
1451 divisor = nsec * frequency;
1453 while (count_fls + sec_fls > 64) {
1454 REDUCE_FLS(count, sec);
1458 dividend = count * sec;
1460 dividend = count * sec;
1462 while (nsec_fls + frequency_fls > 64) {
1463 REDUCE_FLS(nsec, frequency);
1467 divisor = nsec * frequency;
1470 return div64_u64(dividend, divisor);
1473 static void perf_event_stop(struct perf_event *event)
1475 if (!event->pmu->stop)
1476 return event->pmu->disable(event);
1478 return event->pmu->stop(event);
1481 static int perf_event_start(struct perf_event *event)
1483 if (!event->pmu->start)
1484 return event->pmu->enable(event);
1486 return event->pmu->start(event);
1489 static void perf_adjust_period(struct perf_event *event, u64 nsec, u64 count)
1491 struct hw_perf_event *hwc = &event->hw;
1492 u64 period, sample_period;
1495 period = perf_calculate_period(event, nsec, count);
1497 delta = (s64)(period - hwc->sample_period);
1498 delta = (delta + 7) / 8; /* low pass filter */
1500 sample_period = hwc->sample_period + delta;
1505 hwc->sample_period = sample_period;
1507 if (atomic64_read(&hwc->period_left) > 8*sample_period) {
1509 perf_event_stop(event);
1510 atomic64_set(&hwc->period_left, 0);
1511 perf_event_start(event);
1516 static void perf_ctx_adjust_freq(struct perf_event_context *ctx)
1518 struct perf_event *event;
1519 struct hw_perf_event *hwc;
1520 u64 interrupts, now;
1523 raw_spin_lock(&ctx->lock);
1524 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
1525 if (event->state != PERF_EVENT_STATE_ACTIVE)
1528 if (event->cpu != -1 && event->cpu != smp_processor_id())
1533 interrupts = hwc->interrupts;
1534 hwc->interrupts = 0;
1537 * unthrottle events on the tick
1539 if (interrupts == MAX_INTERRUPTS) {
1540 perf_log_throttle(event, 1);
1542 event->pmu->unthrottle(event);
1546 if (!event->attr.freq || !event->attr.sample_freq)
1550 event->pmu->read(event);
1551 now = atomic64_read(&event->count);
1552 delta = now - hwc->freq_count_stamp;
1553 hwc->freq_count_stamp = now;
1556 perf_adjust_period(event, TICK_NSEC, delta);
1559 raw_spin_unlock(&ctx->lock);
1563 * Round-robin a context's events:
1565 static void rotate_ctx(struct perf_event_context *ctx)
1567 raw_spin_lock(&ctx->lock);
1569 /* Rotate the first entry last of non-pinned groups */
1570 list_rotate_left(&ctx->flexible_groups);
1572 raw_spin_unlock(&ctx->lock);
1575 void perf_event_task_tick(struct task_struct *curr)
1577 struct perf_cpu_context *cpuctx;
1578 struct perf_event_context *ctx;
1581 if (!atomic_read(&nr_events))
1584 cpuctx = &__get_cpu_var(perf_cpu_context);
1585 if (cpuctx->ctx.nr_events &&
1586 cpuctx->ctx.nr_events != cpuctx->ctx.nr_active)
1589 ctx = curr->perf_event_ctxp;
1590 if (ctx && ctx->nr_events && ctx->nr_events != ctx->nr_active)
1593 perf_ctx_adjust_freq(&cpuctx->ctx);
1595 perf_ctx_adjust_freq(ctx);
1601 cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE);
1603 task_ctx_sched_out(ctx, EVENT_FLEXIBLE);
1605 rotate_ctx(&cpuctx->ctx);
1609 cpu_ctx_sched_in(cpuctx, EVENT_FLEXIBLE);
1611 task_ctx_sched_in(curr, EVENT_FLEXIBLE);
1615 static int event_enable_on_exec(struct perf_event *event,
1616 struct perf_event_context *ctx)
1618 if (!event->attr.enable_on_exec)
1621 event->attr.enable_on_exec = 0;
1622 if (event->state >= PERF_EVENT_STATE_INACTIVE)
1625 __perf_event_mark_enabled(event, ctx);
1631 * Enable all of a task's events that have been marked enable-on-exec.
1632 * This expects task == current.
1634 static void perf_event_enable_on_exec(struct task_struct *task)
1636 struct perf_event_context *ctx;
1637 struct perf_event *event;
1638 unsigned long flags;
1642 local_irq_save(flags);
1643 ctx = task->perf_event_ctxp;
1644 if (!ctx || !ctx->nr_events)
1647 __perf_event_task_sched_out(ctx);
1649 raw_spin_lock(&ctx->lock);
1651 list_for_each_entry(event, &ctx->pinned_groups, group_entry) {
1652 ret = event_enable_on_exec(event, ctx);
1657 list_for_each_entry(event, &ctx->flexible_groups, group_entry) {
1658 ret = event_enable_on_exec(event, ctx);
1664 * Unclone this context if we enabled any event.
1669 raw_spin_unlock(&ctx->lock);
1671 perf_event_task_sched_in(task);
1673 local_irq_restore(flags);
1677 * Cross CPU call to read the hardware event
1679 static void __perf_event_read(void *info)
1681 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
1682 struct perf_event *event = info;
1683 struct perf_event_context *ctx = event->ctx;
1686 * If this is a task context, we need to check whether it is
1687 * the current task context of this cpu. If not it has been
1688 * scheduled out before the smp call arrived. In that case
1689 * event->count would have been updated to a recent sample
1690 * when the event was scheduled out.
1692 if (ctx->task && cpuctx->task_ctx != ctx)
1695 raw_spin_lock(&ctx->lock);
1696 update_context_time(ctx);
1697 update_event_times(event);
1698 raw_spin_unlock(&ctx->lock);
1700 event->pmu->read(event);
1703 static u64 perf_event_read(struct perf_event *event)
1706 * If event is enabled and currently active on a CPU, update the
1707 * value in the event structure:
1709 if (event->state == PERF_EVENT_STATE_ACTIVE) {
1710 smp_call_function_single(event->oncpu,
1711 __perf_event_read, event, 1);
1712 } else if (event->state == PERF_EVENT_STATE_INACTIVE) {
1713 struct perf_event_context *ctx = event->ctx;
1714 unsigned long flags;
1716 raw_spin_lock_irqsave(&ctx->lock, flags);
1717 update_context_time(ctx);
1718 update_event_times(event);
1719 raw_spin_unlock_irqrestore(&ctx->lock, flags);
1722 return atomic64_read(&event->count);
1726 * Initialize the perf_event context in a task_struct:
1729 __perf_event_init_context(struct perf_event_context *ctx,
1730 struct task_struct *task)
1732 raw_spin_lock_init(&ctx->lock);
1733 mutex_init(&ctx->mutex);
1734 INIT_LIST_HEAD(&ctx->pinned_groups);
1735 INIT_LIST_HEAD(&ctx->flexible_groups);
1736 INIT_LIST_HEAD(&ctx->event_list);
1737 atomic_set(&ctx->refcount, 1);
1741 static struct perf_event_context *find_get_context(pid_t pid, int cpu)
1743 struct perf_event_context *ctx;
1744 struct perf_cpu_context *cpuctx;
1745 struct task_struct *task;
1746 unsigned long flags;
1749 if (pid == -1 && cpu != -1) {
1750 /* Must be root to operate on a CPU event: */
1751 if (perf_paranoid_cpu() && !capable(CAP_SYS_ADMIN))
1752 return ERR_PTR(-EACCES);
1754 if (cpu < 0 || cpu >= nr_cpumask_bits)
1755 return ERR_PTR(-EINVAL);
1758 * We could be clever and allow to attach a event to an
1759 * offline CPU and activate it when the CPU comes up, but
1762 if (!cpu_online(cpu))
1763 return ERR_PTR(-ENODEV);
1765 cpuctx = &per_cpu(perf_cpu_context, cpu);
1776 task = find_task_by_vpid(pid);
1778 get_task_struct(task);
1782 return ERR_PTR(-ESRCH);
1785 * Can't attach events to a dying task.
1788 if (task->flags & PF_EXITING)
1791 /* Reuse ptrace permission checks for now. */
1793 if (!ptrace_may_access(task, PTRACE_MODE_READ))
1797 ctx = perf_lock_task_context(task, &flags);
1800 raw_spin_unlock_irqrestore(&ctx->lock, flags);
1804 ctx = kzalloc(sizeof(struct perf_event_context), GFP_KERNEL);
1808 __perf_event_init_context(ctx, task);
1810 if (cmpxchg(&task->perf_event_ctxp, NULL, ctx)) {
1812 * We raced with some other task; use
1813 * the context they set.
1818 get_task_struct(task);
1821 put_task_struct(task);
1825 put_task_struct(task);
1826 return ERR_PTR(err);
1829 static void perf_event_free_filter(struct perf_event *event);
1831 static void free_event_rcu(struct rcu_head *head)
1833 struct perf_event *event;
1835 event = container_of(head, struct perf_event, rcu_head);
1837 put_pid_ns(event->ns);
1838 perf_event_free_filter(event);
1842 static void perf_pending_sync(struct perf_event *event);
1844 static void free_event(struct perf_event *event)
1846 perf_pending_sync(event);
1848 if (!event->parent) {
1849 atomic_dec(&nr_events);
1850 if (event->attr.mmap)
1851 atomic_dec(&nr_mmap_events);
1852 if (event->attr.comm)
1853 atomic_dec(&nr_comm_events);
1854 if (event->attr.task)
1855 atomic_dec(&nr_task_events);
1858 if (event->output) {
1859 fput(event->output->filp);
1860 event->output = NULL;
1864 event->destroy(event);
1866 put_ctx(event->ctx);
1867 call_rcu(&event->rcu_head, free_event_rcu);
1870 int perf_event_release_kernel(struct perf_event *event)
1872 struct perf_event_context *ctx = event->ctx;
1874 event->state = PERF_EVENT_STATE_FREE;
1876 WARN_ON_ONCE(ctx->parent_ctx);
1878 * There are two ways this annotation is useful:
1880 * 1) there is a lock recursion from perf_event_exit_task
1881 * see the comment there.
1883 * 2) there is a lock-inversion with mmap_sem through
1884 * perf_event_read_group(), which takes faults while
1885 * holding ctx->mutex, however this is called after
1886 * the last filedesc died, so there is no possibility
1887 * to trigger the AB-BA case.
1889 mutex_lock_nested(&ctx->mutex, SINGLE_DEPTH_NESTING);
1890 perf_event_remove_from_context(event);
1891 mutex_unlock(&ctx->mutex);
1893 mutex_lock(&event->owner->perf_event_mutex);
1894 list_del_init(&event->owner_entry);
1895 mutex_unlock(&event->owner->perf_event_mutex);
1896 put_task_struct(event->owner);
1902 EXPORT_SYMBOL_GPL(perf_event_release_kernel);
1905 * Called when the last reference to the file is gone.
1907 static int perf_release(struct inode *inode, struct file *file)
1909 struct perf_event *event = file->private_data;
1911 file->private_data = NULL;
1913 return perf_event_release_kernel(event);
1916 static int perf_event_read_size(struct perf_event *event)
1918 int entry = sizeof(u64); /* value */
1922 if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1923 size += sizeof(u64);
1925 if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1926 size += sizeof(u64);
1928 if (event->attr.read_format & PERF_FORMAT_ID)
1929 entry += sizeof(u64);
1931 if (event->attr.read_format & PERF_FORMAT_GROUP) {
1932 nr += event->group_leader->nr_siblings;
1933 size += sizeof(u64);
1941 u64 perf_event_read_value(struct perf_event *event, u64 *enabled, u64 *running)
1943 struct perf_event *child;
1949 mutex_lock(&event->child_mutex);
1950 total += perf_event_read(event);
1951 *enabled += event->total_time_enabled +
1952 atomic64_read(&event->child_total_time_enabled);
1953 *running += event->total_time_running +
1954 atomic64_read(&event->child_total_time_running);
1956 list_for_each_entry(child, &event->child_list, child_list) {
1957 total += perf_event_read(child);
1958 *enabled += child->total_time_enabled;
1959 *running += child->total_time_running;
1961 mutex_unlock(&event->child_mutex);
1965 EXPORT_SYMBOL_GPL(perf_event_read_value);
1967 static int perf_event_read_group(struct perf_event *event,
1968 u64 read_format, char __user *buf)
1970 struct perf_event *leader = event->group_leader, *sub;
1971 int n = 0, size = 0, ret = -EFAULT;
1972 struct perf_event_context *ctx = leader->ctx;
1974 u64 count, enabled, running;
1976 mutex_lock(&ctx->mutex);
1977 count = perf_event_read_value(leader, &enabled, &running);
1979 values[n++] = 1 + leader->nr_siblings;
1980 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1981 values[n++] = enabled;
1982 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1983 values[n++] = running;
1984 values[n++] = count;
1985 if (read_format & PERF_FORMAT_ID)
1986 values[n++] = primary_event_id(leader);
1988 size = n * sizeof(u64);
1990 if (copy_to_user(buf, values, size))
1995 list_for_each_entry(sub, &leader->sibling_list, group_entry) {
1998 values[n++] = perf_event_read_value(sub, &enabled, &running);
1999 if (read_format & PERF_FORMAT_ID)
2000 values[n++] = primary_event_id(sub);
2002 size = n * sizeof(u64);
2004 if (copy_to_user(buf + ret, values, size)) {
2012 mutex_unlock(&ctx->mutex);
2017 static int perf_event_read_one(struct perf_event *event,
2018 u64 read_format, char __user *buf)
2020 u64 enabled, running;
2024 values[n++] = perf_event_read_value(event, &enabled, &running);
2025 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
2026 values[n++] = enabled;
2027 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
2028 values[n++] = running;
2029 if (read_format & PERF_FORMAT_ID)
2030 values[n++] = primary_event_id(event);
2032 if (copy_to_user(buf, values, n * sizeof(u64)))
2035 return n * sizeof(u64);
2039 * Read the performance event - simple non blocking version for now
2042 perf_read_hw(struct perf_event *event, char __user *buf, size_t count)
2044 u64 read_format = event->attr.read_format;
2048 * Return end-of-file for a read on a event that is in
2049 * error state (i.e. because it was pinned but it couldn't be
2050 * scheduled on to the CPU at some point).
2052 if (event->state == PERF_EVENT_STATE_ERROR)
2055 if (count < perf_event_read_size(event))
2058 WARN_ON_ONCE(event->ctx->parent_ctx);
2059 if (read_format & PERF_FORMAT_GROUP)
2060 ret = perf_event_read_group(event, read_format, buf);
2062 ret = perf_event_read_one(event, read_format, buf);
2068 perf_read(struct file *file, char __user *buf, size_t count, loff_t *ppos)
2070 struct perf_event *event = file->private_data;
2072 return perf_read_hw(event, buf, count);
2075 static unsigned int perf_poll(struct file *file, poll_table *wait)
2077 struct perf_event *event = file->private_data;
2078 struct perf_mmap_data *data;
2079 unsigned int events = POLL_HUP;
2082 data = rcu_dereference(event->data);
2084 events = atomic_xchg(&data->poll, 0);
2087 poll_wait(file, &event->waitq, wait);
2092 static void perf_event_reset(struct perf_event *event)
2094 (void)perf_event_read(event);
2095 atomic64_set(&event->count, 0);
2096 perf_event_update_userpage(event);
2100 * Holding the top-level event's child_mutex means that any
2101 * descendant process that has inherited this event will block
2102 * in sync_child_event if it goes to exit, thus satisfying the
2103 * task existence requirements of perf_event_enable/disable.
2105 static void perf_event_for_each_child(struct perf_event *event,
2106 void (*func)(struct perf_event *))
2108 struct perf_event *child;
2110 WARN_ON_ONCE(event->ctx->parent_ctx);
2111 mutex_lock(&event->child_mutex);
2113 list_for_each_entry(child, &event->child_list, child_list)
2115 mutex_unlock(&event->child_mutex);
2118 static void perf_event_for_each(struct perf_event *event,
2119 void (*func)(struct perf_event *))
2121 struct perf_event_context *ctx = event->ctx;
2122 struct perf_event *sibling;
2124 WARN_ON_ONCE(ctx->parent_ctx);
2125 mutex_lock(&ctx->mutex);
2126 event = event->group_leader;
2128 perf_event_for_each_child(event, func);
2130 list_for_each_entry(sibling, &event->sibling_list, group_entry)
2131 perf_event_for_each_child(event, func);
2132 mutex_unlock(&ctx->mutex);
2135 static int perf_event_period(struct perf_event *event, u64 __user *arg)
2137 struct perf_event_context *ctx = event->ctx;
2142 if (!event->attr.sample_period)
2145 size = copy_from_user(&value, arg, sizeof(value));
2146 if (size != sizeof(value))
2152 raw_spin_lock_irq(&ctx->lock);
2153 if (event->attr.freq) {
2154 if (value > sysctl_perf_event_sample_rate) {
2159 event->attr.sample_freq = value;
2161 event->attr.sample_period = value;
2162 event->hw.sample_period = value;
2165 raw_spin_unlock_irq(&ctx->lock);
2170 static int perf_event_set_output(struct perf_event *event, int output_fd);
2171 static int perf_event_set_filter(struct perf_event *event, void __user *arg);
2173 static long perf_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
2175 struct perf_event *event = file->private_data;
2176 void (*func)(struct perf_event *);
2180 case PERF_EVENT_IOC_ENABLE:
2181 func = perf_event_enable;
2183 case PERF_EVENT_IOC_DISABLE:
2184 func = perf_event_disable;
2186 case PERF_EVENT_IOC_RESET:
2187 func = perf_event_reset;
2190 case PERF_EVENT_IOC_REFRESH:
2191 return perf_event_refresh(event, arg);
2193 case PERF_EVENT_IOC_PERIOD:
2194 return perf_event_period(event, (u64 __user *)arg);
2196 case PERF_EVENT_IOC_SET_OUTPUT:
2197 return perf_event_set_output(event, arg);
2199 case PERF_EVENT_IOC_SET_FILTER:
2200 return perf_event_set_filter(event, (void __user *)arg);
2206 if (flags & PERF_IOC_FLAG_GROUP)
2207 perf_event_for_each(event, func);
2209 perf_event_for_each_child(event, func);
2214 int perf_event_task_enable(void)
2216 struct perf_event *event;
2218 mutex_lock(¤t->perf_event_mutex);
2219 list_for_each_entry(event, ¤t->perf_event_list, owner_entry)
2220 perf_event_for_each_child(event, perf_event_enable);
2221 mutex_unlock(¤t->perf_event_mutex);
2226 int perf_event_task_disable(void)
2228 struct perf_event *event;
2230 mutex_lock(¤t->perf_event_mutex);
2231 list_for_each_entry(event, ¤t->perf_event_list, owner_entry)
2232 perf_event_for_each_child(event, perf_event_disable);
2233 mutex_unlock(¤t->perf_event_mutex);
2238 #ifndef PERF_EVENT_INDEX_OFFSET
2239 # define PERF_EVENT_INDEX_OFFSET 0
2242 static int perf_event_index(struct perf_event *event)
2244 if (event->state != PERF_EVENT_STATE_ACTIVE)
2247 return event->hw.idx + 1 - PERF_EVENT_INDEX_OFFSET;
2251 * Callers need to ensure there can be no nesting of this function, otherwise
2252 * the seqlock logic goes bad. We can not serialize this because the arch
2253 * code calls this from NMI context.
2255 void perf_event_update_userpage(struct perf_event *event)
2257 struct perf_event_mmap_page *userpg;
2258 struct perf_mmap_data *data;
2261 data = rcu_dereference(event->data);
2265 userpg = data->user_page;
2268 * Disable preemption so as to not let the corresponding user-space
2269 * spin too long if we get preempted.
2274 userpg->index = perf_event_index(event);
2275 userpg->offset = atomic64_read(&event->count);
2276 if (event->state == PERF_EVENT_STATE_ACTIVE)
2277 userpg->offset -= atomic64_read(&event->hw.prev_count);
2279 userpg->time_enabled = event->total_time_enabled +
2280 atomic64_read(&event->child_total_time_enabled);
2282 userpg->time_running = event->total_time_running +
2283 atomic64_read(&event->child_total_time_running);
2292 static unsigned long perf_data_size(struct perf_mmap_data *data)
2294 return data->nr_pages << (PAGE_SHIFT + data->data_order);
2297 #ifndef CONFIG_PERF_USE_VMALLOC
2300 * Back perf_mmap() with regular GFP_KERNEL-0 pages.
2303 static struct page *
2304 perf_mmap_to_page(struct perf_mmap_data *data, unsigned long pgoff)
2306 if (pgoff > data->nr_pages)
2310 return virt_to_page(data->user_page);
2312 return virt_to_page(data->data_pages[pgoff - 1]);
2315 static struct perf_mmap_data *
2316 perf_mmap_data_alloc(struct perf_event *event, int nr_pages)
2318 struct perf_mmap_data *data;
2322 WARN_ON(atomic_read(&event->mmap_count));
2324 size = sizeof(struct perf_mmap_data);
2325 size += nr_pages * sizeof(void *);
2327 data = kzalloc(size, GFP_KERNEL);
2331 data->user_page = (void *)get_zeroed_page(GFP_KERNEL);
2332 if (!data->user_page)
2333 goto fail_user_page;
2335 for (i = 0; i < nr_pages; i++) {
2336 data->data_pages[i] = (void *)get_zeroed_page(GFP_KERNEL);
2337 if (!data->data_pages[i])
2338 goto fail_data_pages;
2341 data->data_order = 0;
2342 data->nr_pages = nr_pages;
2347 for (i--; i >= 0; i--)
2348 free_page((unsigned long)data->data_pages[i]);
2350 free_page((unsigned long)data->user_page);
2359 static void perf_mmap_free_page(unsigned long addr)
2361 struct page *page = virt_to_page((void *)addr);
2363 page->mapping = NULL;
2367 static void perf_mmap_data_free(struct perf_mmap_data *data)
2371 perf_mmap_free_page((unsigned long)data->user_page);
2372 for (i = 0; i < data->nr_pages; i++)
2373 perf_mmap_free_page((unsigned long)data->data_pages[i]);
2380 * Back perf_mmap() with vmalloc memory.
2382 * Required for architectures that have d-cache aliasing issues.
2385 static struct page *
2386 perf_mmap_to_page(struct perf_mmap_data *data, unsigned long pgoff)
2388 if (pgoff > (1UL << data->data_order))
2391 return vmalloc_to_page((void *)data->user_page + pgoff * PAGE_SIZE);
2394 static void perf_mmap_unmark_page(void *addr)
2396 struct page *page = vmalloc_to_page(addr);
2398 page->mapping = NULL;
2401 static void perf_mmap_data_free_work(struct work_struct *work)
2403 struct perf_mmap_data *data;
2407 data = container_of(work, struct perf_mmap_data, work);
2408 nr = 1 << data->data_order;
2410 base = data->user_page;
2411 for (i = 0; i < nr + 1; i++)
2412 perf_mmap_unmark_page(base + (i * PAGE_SIZE));
2418 static void perf_mmap_data_free(struct perf_mmap_data *data)
2420 schedule_work(&data->work);
2423 static struct perf_mmap_data *
2424 perf_mmap_data_alloc(struct perf_event *event, int nr_pages)
2426 struct perf_mmap_data *data;
2430 WARN_ON(atomic_read(&event->mmap_count));
2432 size = sizeof(struct perf_mmap_data);
2433 size += sizeof(void *);
2435 data = kzalloc(size, GFP_KERNEL);
2439 INIT_WORK(&data->work, perf_mmap_data_free_work);
2441 all_buf = vmalloc_user((nr_pages + 1) * PAGE_SIZE);
2445 data->user_page = all_buf;
2446 data->data_pages[0] = all_buf + PAGE_SIZE;
2447 data->data_order = ilog2(nr_pages);
2461 static int perf_mmap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
2463 struct perf_event *event = vma->vm_file->private_data;
2464 struct perf_mmap_data *data;
2465 int ret = VM_FAULT_SIGBUS;
2467 if (vmf->flags & FAULT_FLAG_MKWRITE) {
2468 if (vmf->pgoff == 0)
2474 data = rcu_dereference(event->data);
2478 if (vmf->pgoff && (vmf->flags & FAULT_FLAG_WRITE))
2481 vmf->page = perf_mmap_to_page(data, vmf->pgoff);
2485 get_page(vmf->page);
2486 vmf->page->mapping = vma->vm_file->f_mapping;
2487 vmf->page->index = vmf->pgoff;
2497 perf_mmap_data_init(struct perf_event *event, struct perf_mmap_data *data)
2499 long max_size = perf_data_size(data);
2501 atomic_set(&data->lock, -1);
2503 if (event->attr.watermark) {
2504 data->watermark = min_t(long, max_size,
2505 event->attr.wakeup_watermark);
2508 if (!data->watermark)
2509 data->watermark = max_size / 2;
2512 rcu_assign_pointer(event->data, data);
2515 static void perf_mmap_data_free_rcu(struct rcu_head *rcu_head)
2517 struct perf_mmap_data *data;
2519 data = container_of(rcu_head, struct perf_mmap_data, rcu_head);
2520 perf_mmap_data_free(data);
2523 static void perf_mmap_data_release(struct perf_event *event)
2525 struct perf_mmap_data *data = event->data;
2527 WARN_ON(atomic_read(&event->mmap_count));
2529 rcu_assign_pointer(event->data, NULL);
2530 call_rcu(&data->rcu_head, perf_mmap_data_free_rcu);
2533 static void perf_mmap_open(struct vm_area_struct *vma)
2535 struct perf_event *event = vma->vm_file->private_data;
2537 atomic_inc(&event->mmap_count);
2540 static void perf_mmap_close(struct vm_area_struct *vma)
2542 struct perf_event *event = vma->vm_file->private_data;
2544 WARN_ON_ONCE(event->ctx->parent_ctx);
2545 if (atomic_dec_and_mutex_lock(&event->mmap_count, &event->mmap_mutex)) {
2546 unsigned long size = perf_data_size(event->data);
2547 struct user_struct *user = current_user();
2549 atomic_long_sub((size >> PAGE_SHIFT) + 1, &user->locked_vm);
2550 vma->vm_mm->locked_vm -= event->data->nr_locked;
2551 perf_mmap_data_release(event);
2552 mutex_unlock(&event->mmap_mutex);
2556 static const struct vm_operations_struct perf_mmap_vmops = {
2557 .open = perf_mmap_open,
2558 .close = perf_mmap_close,
2559 .fault = perf_mmap_fault,
2560 .page_mkwrite = perf_mmap_fault,
2563 static int perf_mmap(struct file *file, struct vm_area_struct *vma)
2565 struct perf_event *event = file->private_data;
2566 unsigned long user_locked, user_lock_limit;
2567 struct user_struct *user = current_user();
2568 unsigned long locked, lock_limit;
2569 struct perf_mmap_data *data;
2570 unsigned long vma_size;
2571 unsigned long nr_pages;
2572 long user_extra, extra;
2575 if (!(vma->vm_flags & VM_SHARED))
2578 vma_size = vma->vm_end - vma->vm_start;
2579 nr_pages = (vma_size / PAGE_SIZE) - 1;
2582 * If we have data pages ensure they're a power-of-two number, so we
2583 * can do bitmasks instead of modulo.
2585 if (nr_pages != 0 && !is_power_of_2(nr_pages))
2588 if (vma_size != PAGE_SIZE * (1 + nr_pages))
2591 if (vma->vm_pgoff != 0)
2594 WARN_ON_ONCE(event->ctx->parent_ctx);
2595 mutex_lock(&event->mmap_mutex);
2596 if (event->output) {
2601 if (atomic_inc_not_zero(&event->mmap_count)) {
2602 if (nr_pages != event->data->nr_pages)
2607 user_extra = nr_pages + 1;
2608 user_lock_limit = sysctl_perf_event_mlock >> (PAGE_SHIFT - 10);
2611 * Increase the limit linearly with more CPUs:
2613 user_lock_limit *= num_online_cpus();
2615 user_locked = atomic_long_read(&user->locked_vm) + user_extra;
2618 if (user_locked > user_lock_limit)
2619 extra = user_locked - user_lock_limit;
2621 lock_limit = rlimit(RLIMIT_MEMLOCK);
2622 lock_limit >>= PAGE_SHIFT;
2623 locked = vma->vm_mm->locked_vm + extra;
2625 if ((locked > lock_limit) && perf_paranoid_tracepoint_raw() &&
2626 !capable(CAP_IPC_LOCK)) {
2631 WARN_ON(event->data);
2633 data = perf_mmap_data_alloc(event, nr_pages);
2639 perf_mmap_data_init(event, data);
2641 atomic_set(&event->mmap_count, 1);
2642 atomic_long_add(user_extra, &user->locked_vm);
2643 vma->vm_mm->locked_vm += extra;
2644 event->data->nr_locked = extra;
2645 if (vma->vm_flags & VM_WRITE)
2646 event->data->writable = 1;
2649 mutex_unlock(&event->mmap_mutex);
2651 vma->vm_flags |= VM_RESERVED;
2652 vma->vm_ops = &perf_mmap_vmops;
2657 static int perf_fasync(int fd, struct file *filp, int on)
2659 struct inode *inode = filp->f_path.dentry->d_inode;
2660 struct perf_event *event = filp->private_data;
2663 mutex_lock(&inode->i_mutex);
2664 retval = fasync_helper(fd, filp, on, &event->fasync);
2665 mutex_unlock(&inode->i_mutex);
2673 static const struct file_operations perf_fops = {
2674 .llseek = no_llseek,
2675 .release = perf_release,
2678 .unlocked_ioctl = perf_ioctl,
2679 .compat_ioctl = perf_ioctl,
2681 .fasync = perf_fasync,
2687 * If there's data, ensure we set the poll() state and publish everything
2688 * to user-space before waking everybody up.
2691 void perf_event_wakeup(struct perf_event *event)
2693 wake_up_all(&event->waitq);
2695 if (event->pending_kill) {
2696 kill_fasync(&event->fasync, SIGIO, event->pending_kill);
2697 event->pending_kill = 0;
2704 * Handle the case where we need to wakeup up from NMI (or rq->lock) context.
2706 * The NMI bit means we cannot possibly take locks. Therefore, maintain a
2707 * single linked list and use cmpxchg() to add entries lockless.
2710 static void perf_pending_event(struct perf_pending_entry *entry)
2712 struct perf_event *event = container_of(entry,
2713 struct perf_event, pending);
2715 if (event->pending_disable) {
2716 event->pending_disable = 0;
2717 __perf_event_disable(event);
2720 if (event->pending_wakeup) {
2721 event->pending_wakeup = 0;
2722 perf_event_wakeup(event);
2726 #define PENDING_TAIL ((struct perf_pending_entry *)-1UL)
2728 static DEFINE_PER_CPU(struct perf_pending_entry *, perf_pending_head) = {
2732 static void perf_pending_queue(struct perf_pending_entry *entry,
2733 void (*func)(struct perf_pending_entry *))
2735 struct perf_pending_entry **head;
2737 if (cmpxchg(&entry->next, NULL, PENDING_TAIL) != NULL)
2742 head = &get_cpu_var(perf_pending_head);
2745 entry->next = *head;
2746 } while (cmpxchg(head, entry->next, entry) != entry->next);
2748 set_perf_event_pending();
2750 put_cpu_var(perf_pending_head);
2753 static int __perf_pending_run(void)
2755 struct perf_pending_entry *list;
2758 list = xchg(&__get_cpu_var(perf_pending_head), PENDING_TAIL);
2759 while (list != PENDING_TAIL) {
2760 void (*func)(struct perf_pending_entry *);
2761 struct perf_pending_entry *entry = list;
2768 * Ensure we observe the unqueue before we issue the wakeup,
2769 * so that we won't be waiting forever.
2770 * -- see perf_not_pending().
2781 static inline int perf_not_pending(struct perf_event *event)
2784 * If we flush on whatever cpu we run, there is a chance we don't
2788 __perf_pending_run();
2792 * Ensure we see the proper queue state before going to sleep
2793 * so that we do not miss the wakeup. -- see perf_pending_handle()
2796 return event->pending.next == NULL;
2799 static void perf_pending_sync(struct perf_event *event)
2801 wait_event(event->waitq, perf_not_pending(event));
2804 void perf_event_do_pending(void)
2806 __perf_pending_run();
2810 * Callchain support -- arch specific
2813 __weak struct perf_callchain_entry *perf_callchain(struct pt_regs *regs)
2819 void perf_arch_fetch_caller_regs(struct pt_regs *regs, unsigned long ip, int skip)
2825 * We assume there is only KVM supporting the callbacks.
2826 * Later on, we might change it to a list if there is
2827 * another virtualization implementation supporting the callbacks.
2829 struct perf_guest_info_callbacks *perf_guest_cbs;
2831 int perf_register_guest_info_callbacks(struct perf_guest_info_callbacks *cbs)
2833 perf_guest_cbs = cbs;
2836 EXPORT_SYMBOL_GPL(perf_register_guest_info_callbacks);
2838 int perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks *cbs)
2840 perf_guest_cbs = NULL;
2843 EXPORT_SYMBOL_GPL(perf_unregister_guest_info_callbacks);
2848 static bool perf_output_space(struct perf_mmap_data *data, unsigned long tail,
2849 unsigned long offset, unsigned long head)
2853 if (!data->writable)
2856 mask = perf_data_size(data) - 1;
2858 offset = (offset - tail) & mask;
2859 head = (head - tail) & mask;
2861 if ((int)(head - offset) < 0)
2867 static void perf_output_wakeup(struct perf_output_handle *handle)
2869 atomic_set(&handle->data->poll, POLL_IN);
2872 handle->event->pending_wakeup = 1;
2873 perf_pending_queue(&handle->event->pending,
2874 perf_pending_event);
2876 perf_event_wakeup(handle->event);
2880 * Curious locking construct.
2882 * We need to ensure a later event_id doesn't publish a head when a former
2883 * event_id isn't done writing. However since we need to deal with NMIs we
2884 * cannot fully serialize things.
2886 * What we do is serialize between CPUs so we only have to deal with NMI
2887 * nesting on a single CPU.
2889 * We only publish the head (and generate a wakeup) when the outer-most
2890 * event_id completes.
2892 static void perf_output_lock(struct perf_output_handle *handle)
2894 struct perf_mmap_data *data = handle->data;
2895 int cur, cpu = get_cpu();
2900 cur = atomic_cmpxchg(&data->lock, -1, cpu);
2912 static void perf_output_unlock(struct perf_output_handle *handle)
2914 struct perf_mmap_data *data = handle->data;
2918 data->done_head = data->head;
2920 if (!handle->locked)
2925 * The xchg implies a full barrier that ensures all writes are done
2926 * before we publish the new head, matched by a rmb() in userspace when
2927 * reading this position.
2929 while ((head = atomic_long_xchg(&data->done_head, 0)))
2930 data->user_page->data_head = head;
2933 * NMI can happen here, which means we can miss a done_head update.
2936 cpu = atomic_xchg(&data->lock, -1);
2937 WARN_ON_ONCE(cpu != smp_processor_id());
2940 * Therefore we have to validate we did not indeed do so.
2942 if (unlikely(atomic_long_read(&data->done_head))) {
2944 * Since we had it locked, we can lock it again.
2946 while (atomic_cmpxchg(&data->lock, -1, cpu) != -1)
2952 if (atomic_xchg(&data->wakeup, 0))
2953 perf_output_wakeup(handle);
2958 void perf_output_copy(struct perf_output_handle *handle,
2959 const void *buf, unsigned int len)
2961 unsigned int pages_mask;
2962 unsigned long offset;
2966 offset = handle->offset;
2967 pages_mask = handle->data->nr_pages - 1;
2968 pages = handle->data->data_pages;
2971 unsigned long page_offset;
2972 unsigned long page_size;
2975 nr = (offset >> PAGE_SHIFT) & pages_mask;
2976 page_size = 1UL << (handle->data->data_order + PAGE_SHIFT);
2977 page_offset = offset & (page_size - 1);
2978 size = min_t(unsigned int, page_size - page_offset, len);
2980 memcpy(pages[nr] + page_offset, buf, size);
2987 handle->offset = offset;
2990 * Check we didn't copy past our reservation window, taking the
2991 * possible unsigned int wrap into account.
2993 WARN_ON_ONCE(((long)(handle->head - handle->offset)) < 0);
2996 int perf_output_begin(struct perf_output_handle *handle,
2997 struct perf_event *event, unsigned int size,
2998 int nmi, int sample)
3000 struct perf_event *output_event;
3001 struct perf_mmap_data *data;
3002 unsigned long tail, offset, head;
3005 struct perf_event_header header;
3012 * For inherited events we send all the output towards the parent.
3015 event = event->parent;
3017 output_event = rcu_dereference(event->output);
3019 event = output_event;
3021 data = rcu_dereference(event->data);
3025 handle->data = data;
3026 handle->event = event;
3028 handle->sample = sample;
3030 if (!data->nr_pages)
3033 have_lost = atomic_read(&data->lost);
3035 size += sizeof(lost_event);
3037 perf_output_lock(handle);
3041 * Userspace could choose to issue a mb() before updating the
3042 * tail pointer. So that all reads will be completed before the
3045 tail = ACCESS_ONCE(data->user_page->data_tail);
3047 offset = head = atomic_long_read(&data->head);
3049 if (unlikely(!perf_output_space(data, tail, offset, head)))
3051 } while (atomic_long_cmpxchg(&data->head, offset, head) != offset);
3053 handle->offset = offset;
3054 handle->head = head;
3056 if (head - tail > data->watermark)
3057 atomic_set(&data->wakeup, 1);
3060 lost_event.header.type = PERF_RECORD_LOST;
3061 lost_event.header.misc = 0;
3062 lost_event.header.size = sizeof(lost_event);
3063 lost_event.id = event->id;
3064 lost_event.lost = atomic_xchg(&data->lost, 0);
3066 perf_output_put(handle, lost_event);
3072 atomic_inc(&data->lost);
3073 perf_output_unlock(handle);
3080 void perf_output_end(struct perf_output_handle *handle)
3082 struct perf_event *event = handle->event;
3083 struct perf_mmap_data *data = handle->data;
3085 int wakeup_events = event->attr.wakeup_events;
3087 if (handle->sample && wakeup_events) {
3088 int events = atomic_inc_return(&data->events);
3089 if (events >= wakeup_events) {
3090 atomic_sub(wakeup_events, &data->events);
3091 atomic_set(&data->wakeup, 1);
3095 perf_output_unlock(handle);
3099 static u32 perf_event_pid(struct perf_event *event, struct task_struct *p)
3102 * only top level events have the pid namespace they were created in
3105 event = event->parent;
3107 return task_tgid_nr_ns(p, event->ns);
3110 static u32 perf_event_tid(struct perf_event *event, struct task_struct *p)
3113 * only top level events have the pid namespace they were created in
3116 event = event->parent;
3118 return task_pid_nr_ns(p, event->ns);
3121 static void perf_output_read_one(struct perf_output_handle *handle,
3122 struct perf_event *event)
3124 u64 read_format = event->attr.read_format;
3128 values[n++] = atomic64_read(&event->count);
3129 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
3130 values[n++] = event->total_time_enabled +
3131 atomic64_read(&event->child_total_time_enabled);
3133 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
3134 values[n++] = event->total_time_running +
3135 atomic64_read(&event->child_total_time_running);
3137 if (read_format & PERF_FORMAT_ID)
3138 values[n++] = primary_event_id(event);
3140 perf_output_copy(handle, values, n * sizeof(u64));
3144 * XXX PERF_FORMAT_GROUP vs inherited events seems difficult.
3146 static void perf_output_read_group(struct perf_output_handle *handle,
3147 struct perf_event *event)
3149 struct perf_event *leader = event->group_leader, *sub;
3150 u64 read_format = event->attr.read_format;
3154 values[n++] = 1 + leader->nr_siblings;
3156 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
3157 values[n++] = leader->total_time_enabled;
3159 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
3160 values[n++] = leader->total_time_running;
3162 if (leader != event)
3163 leader->pmu->read(leader);
3165 values[n++] = atomic64_read(&leader->count);
3166 if (read_format & PERF_FORMAT_ID)
3167 values[n++] = primary_event_id(leader);
3169 perf_output_copy(handle, values, n * sizeof(u64));
3171 list_for_each_entry(sub, &leader->sibling_list, group_entry) {
3175 sub->pmu->read(sub);
3177 values[n++] = atomic64_read(&sub->count);
3178 if (read_format & PERF_FORMAT_ID)
3179 values[n++] = primary_event_id(sub);
3181 perf_output_copy(handle, values, n * sizeof(u64));
3185 static void perf_output_read(struct perf_output_handle *handle,
3186 struct perf_event *event)
3188 if (event->attr.read_format & PERF_FORMAT_GROUP)
3189 perf_output_read_group(handle, event);
3191 perf_output_read_one(handle, event);
3194 void perf_output_sample(struct perf_output_handle *handle,
3195 struct perf_event_header *header,
3196 struct perf_sample_data *data,
3197 struct perf_event *event)
3199 u64 sample_type = data->type;
3201 perf_output_put(handle, *header);
3203 if (sample_type & PERF_SAMPLE_IP)
3204 perf_output_put(handle, data->ip);
3206 if (sample_type & PERF_SAMPLE_TID)
3207 perf_output_put(handle, data->tid_entry);
3209 if (sample_type & PERF_SAMPLE_TIME)
3210 perf_output_put(handle, data->time);
3212 if (sample_type & PERF_SAMPLE_ADDR)
3213 perf_output_put(handle, data->addr);
3215 if (sample_type & PERF_SAMPLE_ID)
3216 perf_output_put(handle, data->id);
3218 if (sample_type & PERF_SAMPLE_STREAM_ID)
3219 perf_output_put(handle, data->stream_id);
3221 if (sample_type & PERF_SAMPLE_CPU)
3222 perf_output_put(handle, data->cpu_entry);
3224 if (sample_type & PERF_SAMPLE_PERIOD)
3225 perf_output_put(handle, data->period);
3227 if (sample_type & PERF_SAMPLE_READ)
3228 perf_output_read(handle, event);
3230 if (sample_type & PERF_SAMPLE_CALLCHAIN) {
3231 if (data->callchain) {
3234 if (data->callchain)
3235 size += data->callchain->nr;
3237 size *= sizeof(u64);
3239 perf_output_copy(handle, data->callchain, size);
3242 perf_output_put(handle, nr);
3246 if (sample_type & PERF_SAMPLE_RAW) {
3248 perf_output_put(handle, data->raw->size);
3249 perf_output_copy(handle, data->raw->data,
3256 .size = sizeof(u32),
3259 perf_output_put(handle, raw);
3264 void perf_prepare_sample(struct perf_event_header *header,
3265 struct perf_sample_data *data,
3266 struct perf_event *event,
3267 struct pt_regs *regs)
3269 u64 sample_type = event->attr.sample_type;
3271 data->type = sample_type;
3273 header->type = PERF_RECORD_SAMPLE;
3274 header->size = sizeof(*header);
3277 header->misc |= perf_misc_flags(regs);
3279 if (sample_type & PERF_SAMPLE_IP) {
3280 data->ip = perf_instruction_pointer(regs);
3282 header->size += sizeof(data->ip);
3285 if (sample_type & PERF_SAMPLE_TID) {
3286 /* namespace issues */
3287 data->tid_entry.pid = perf_event_pid(event, current);
3288 data->tid_entry.tid = perf_event_tid(event, current);
3290 header->size += sizeof(data->tid_entry);
3293 if (sample_type & PERF_SAMPLE_TIME) {
3294 data->time = perf_clock();
3296 header->size += sizeof(data->time);
3299 if (sample_type & PERF_SAMPLE_ADDR)
3300 header->size += sizeof(data->addr);
3302 if (sample_type & PERF_SAMPLE_ID) {
3303 data->id = primary_event_id(event);
3305 header->size += sizeof(data->id);
3308 if (sample_type & PERF_SAMPLE_STREAM_ID) {
3309 data->stream_id = event->id;
3311 header->size += sizeof(data->stream_id);
3314 if (sample_type & PERF_SAMPLE_CPU) {
3315 data->cpu_entry.cpu = raw_smp_processor_id();
3316 data->cpu_entry.reserved = 0;
3318 header->size += sizeof(data->cpu_entry);
3321 if (sample_type & PERF_SAMPLE_PERIOD)
3322 header->size += sizeof(data->period);
3324 if (sample_type & PERF_SAMPLE_READ)
3325 header->size += perf_event_read_size(event);
3327 if (sample_type & PERF_SAMPLE_CALLCHAIN) {
3330 data->callchain = perf_callchain(regs);
3332 if (data->callchain)
3333 size += data->callchain->nr;
3335 header->size += size * sizeof(u64);
3338 if (sample_type & PERF_SAMPLE_RAW) {
3339 int size = sizeof(u32);
3342 size += data->raw->size;
3344 size += sizeof(u32);
3346 WARN_ON_ONCE(size & (sizeof(u64)-1));
3347 header->size += size;
3351 static void perf_event_output(struct perf_event *event, int nmi,
3352 struct perf_sample_data *data,
3353 struct pt_regs *regs)
3355 struct perf_output_handle handle;
3356 struct perf_event_header header;
3358 perf_prepare_sample(&header, data, event, regs);
3360 if (perf_output_begin(&handle, event, header.size, nmi, 1))
3363 perf_output_sample(&handle, &header, data, event);
3365 perf_output_end(&handle);
3372 struct perf_read_event {
3373 struct perf_event_header header;
3380 perf_event_read_event(struct perf_event *event,
3381 struct task_struct *task)
3383 struct perf_output_handle handle;
3384 struct perf_read_event read_event = {
3386 .type = PERF_RECORD_READ,
3388 .size = sizeof(read_event) + perf_event_read_size(event),
3390 .pid = perf_event_pid(event, task),
3391 .tid = perf_event_tid(event, task),
3395 ret = perf_output_begin(&handle, event, read_event.header.size, 0, 0);
3399 perf_output_put(&handle, read_event);
3400 perf_output_read(&handle, event);
3402 perf_output_end(&handle);
3406 * task tracking -- fork/exit
3408 * enabled by: attr.comm | attr.mmap | attr.task
3411 struct perf_task_event {
3412 struct task_struct *task;
3413 struct perf_event_context *task_ctx;
3416 struct perf_event_header header;
3426 static void perf_event_task_output(struct perf_event *event,
3427 struct perf_task_event *task_event)
3429 struct perf_output_handle handle;
3430 struct task_struct *task = task_event->task;
3431 unsigned long flags;
3435 * If this CPU attempts to acquire an rq lock held by a CPU spinning
3436 * in perf_output_lock() from interrupt context, it's game over.
3438 local_irq_save(flags);
3440 size = task_event->event_id.header.size;
3441 ret = perf_output_begin(&handle, event, size, 0, 0);
3444 local_irq_restore(flags);
3448 task_event->event_id.pid = perf_event_pid(event, task);
3449 task_event->event_id.ppid = perf_event_pid(event, current);
3451 task_event->event_id.tid = perf_event_tid(event, task);
3452 task_event->event_id.ptid = perf_event_tid(event, current);
3454 perf_output_put(&handle, task_event->event_id);
3456 perf_output_end(&handle);
3457 local_irq_restore(flags);
3460 static int perf_event_task_match(struct perf_event *event)
3462 if (event->state < PERF_EVENT_STATE_INACTIVE)
3465 if (event->cpu != -1 && event->cpu != smp_processor_id())
3468 if (event->attr.comm || event->attr.mmap || event->attr.task)
3474 static void perf_event_task_ctx(struct perf_event_context *ctx,
3475 struct perf_task_event *task_event)
3477 struct perf_event *event;
3479 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
3480 if (perf_event_task_match(event))
3481 perf_event_task_output(event, task_event);
3485 static void perf_event_task_event(struct perf_task_event *task_event)
3487 struct perf_cpu_context *cpuctx;
3488 struct perf_event_context *ctx = task_event->task_ctx;
3491 cpuctx = &get_cpu_var(perf_cpu_context);
3492 perf_event_task_ctx(&cpuctx->ctx, task_event);
3494 ctx = rcu_dereference(current->perf_event_ctxp);
3496 perf_event_task_ctx(ctx, task_event);
3497 put_cpu_var(perf_cpu_context);
3501 static void perf_event_task(struct task_struct *task,
3502 struct perf_event_context *task_ctx,
3505 struct perf_task_event task_event;
3507 if (!atomic_read(&nr_comm_events) &&
3508 !atomic_read(&nr_mmap_events) &&
3509 !atomic_read(&nr_task_events))
3512 task_event = (struct perf_task_event){
3514 .task_ctx = task_ctx,
3517 .type = new ? PERF_RECORD_FORK : PERF_RECORD_EXIT,
3519 .size = sizeof(task_event.event_id),
3525 .time = perf_clock(),
3529 perf_event_task_event(&task_event);
3532 void perf_event_fork(struct task_struct *task)
3534 perf_event_task(task, NULL, 1);
3541 struct perf_comm_event {
3542 struct task_struct *task;
3547 struct perf_event_header header;
3554 static void perf_event_comm_output(struct perf_event *event,
3555 struct perf_comm_event *comm_event)
3557 struct perf_output_handle handle;
3558 int size = comm_event->event_id.header.size;
3559 int ret = perf_output_begin(&handle, event, size, 0, 0);
3564 comm_event->event_id.pid = perf_event_pid(event, comm_event->task);
3565 comm_event->event_id.tid = perf_event_tid(event, comm_event->task);
3567 perf_output_put(&handle, comm_event->event_id);
3568 perf_output_copy(&handle, comm_event->comm,
3569 comm_event->comm_size);
3570 perf_output_end(&handle);
3573 static int perf_event_comm_match(struct perf_event *event)
3575 if (event->state < PERF_EVENT_STATE_INACTIVE)
3578 if (event->cpu != -1 && event->cpu != smp_processor_id())
3581 if (event->attr.comm)
3587 static void perf_event_comm_ctx(struct perf_event_context *ctx,
3588 struct perf_comm_event *comm_event)
3590 struct perf_event *event;
3592 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
3593 if (perf_event_comm_match(event))
3594 perf_event_comm_output(event, comm_event);
3598 static void perf_event_comm_event(struct perf_comm_event *comm_event)
3600 struct perf_cpu_context *cpuctx;
3601 struct perf_event_context *ctx;
3603 char comm[TASK_COMM_LEN];
3605 memset(comm, 0, sizeof(comm));
3606 strlcpy(comm, comm_event->task->comm, sizeof(comm));
3607 size = ALIGN(strlen(comm)+1, sizeof(u64));
3609 comm_event->comm = comm;
3610 comm_event->comm_size = size;
3612 comm_event->event_id.header.size = sizeof(comm_event->event_id) + size;
3615 cpuctx = &get_cpu_var(perf_cpu_context);
3616 perf_event_comm_ctx(&cpuctx->ctx, comm_event);
3617 ctx = rcu_dereference(current->perf_event_ctxp);
3619 perf_event_comm_ctx(ctx, comm_event);
3620 put_cpu_var(perf_cpu_context);
3624 void perf_event_comm(struct task_struct *task)
3626 struct perf_comm_event comm_event;
3628 if (task->perf_event_ctxp)
3629 perf_event_enable_on_exec(task);
3631 if (!atomic_read(&nr_comm_events))
3634 comm_event = (struct perf_comm_event){
3640 .type = PERF_RECORD_COMM,
3649 perf_event_comm_event(&comm_event);
3656 struct perf_mmap_event {
3657 struct vm_area_struct *vma;
3659 const char *file_name;
3663 struct perf_event_header header;
3673 static void perf_event_mmap_output(struct perf_event *event,
3674 struct perf_mmap_event *mmap_event)
3676 struct perf_output_handle handle;
3677 int size = mmap_event->event_id.header.size;
3678 int ret = perf_output_begin(&handle, event, size, 0, 0);
3683 mmap_event->event_id.pid = perf_event_pid(event, current);
3684 mmap_event->event_id.tid = perf_event_tid(event, current);
3686 perf_output_put(&handle, mmap_event->event_id);
3687 perf_output_copy(&handle, mmap_event->file_name,
3688 mmap_event->file_size);
3689 perf_output_end(&handle);
3692 static int perf_event_mmap_match(struct perf_event *event,
3693 struct perf_mmap_event *mmap_event)
3695 if (event->state < PERF_EVENT_STATE_INACTIVE)
3698 if (event->cpu != -1 && event->cpu != smp_processor_id())
3701 if (event->attr.mmap)
3707 static void perf_event_mmap_ctx(struct perf_event_context *ctx,
3708 struct perf_mmap_event *mmap_event)
3710 struct perf_event *event;
3712 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
3713 if (perf_event_mmap_match(event, mmap_event))
3714 perf_event_mmap_output(event, mmap_event);
3718 static void perf_event_mmap_event(struct perf_mmap_event *mmap_event)
3720 struct perf_cpu_context *cpuctx;
3721 struct perf_event_context *ctx;
3722 struct vm_area_struct *vma = mmap_event->vma;
3723 struct file *file = vma->vm_file;
3729 memset(tmp, 0, sizeof(tmp));
3733 * d_path works from the end of the buffer backwards, so we
3734 * need to add enough zero bytes after the string to handle
3735 * the 64bit alignment we do later.
3737 buf = kzalloc(PATH_MAX + sizeof(u64), GFP_KERNEL);
3739 name = strncpy(tmp, "//enomem", sizeof(tmp));
3742 name = d_path(&file->f_path, buf, PATH_MAX);
3744 name = strncpy(tmp, "//toolong", sizeof(tmp));
3748 if (arch_vma_name(mmap_event->vma)) {
3749 name = strncpy(tmp, arch_vma_name(mmap_event->vma),
3755 name = strncpy(tmp, "[vdso]", sizeof(tmp));
3759 name = strncpy(tmp, "//anon", sizeof(tmp));
3764 size = ALIGN(strlen(name)+1, sizeof(u64));
3766 mmap_event->file_name = name;
3767 mmap_event->file_size = size;
3769 mmap_event->event_id.header.size = sizeof(mmap_event->event_id) + size;
3772 cpuctx = &get_cpu_var(perf_cpu_context);
3773 perf_event_mmap_ctx(&cpuctx->ctx, mmap_event);
3774 ctx = rcu_dereference(current->perf_event_ctxp);
3776 perf_event_mmap_ctx(ctx, mmap_event);
3777 put_cpu_var(perf_cpu_context);
3783 void __perf_event_mmap(struct vm_area_struct *vma)
3785 struct perf_mmap_event mmap_event;
3787 if (!atomic_read(&nr_mmap_events))
3790 mmap_event = (struct perf_mmap_event){
3796 .type = PERF_RECORD_MMAP,
3797 .misc = PERF_RECORD_MISC_USER,
3802 .start = vma->vm_start,
3803 .len = vma->vm_end - vma->vm_start,
3804 .pgoff = (u64)vma->vm_pgoff << PAGE_SHIFT,
3808 perf_event_mmap_event(&mmap_event);
3812 * IRQ throttle logging
3815 static void perf_log_throttle(struct perf_event *event, int enable)
3817 struct perf_output_handle handle;
3821 struct perf_event_header header;
3825 } throttle_event = {
3827 .type = PERF_RECORD_THROTTLE,
3829 .size = sizeof(throttle_event),
3831 .time = perf_clock(),
3832 .id = primary_event_id(event),
3833 .stream_id = event->id,
3837 throttle_event.header.type = PERF_RECORD_UNTHROTTLE;
3839 ret = perf_output_begin(&handle, event, sizeof(throttle_event), 1, 0);
3843 perf_output_put(&handle, throttle_event);
3844 perf_output_end(&handle);
3848 * Generic event overflow handling, sampling.
3851 static int __perf_event_overflow(struct perf_event *event, int nmi,
3852 int throttle, struct perf_sample_data *data,
3853 struct pt_regs *regs)
3855 int events = atomic_read(&event->event_limit);
3856 struct hw_perf_event *hwc = &event->hw;
3859 throttle = (throttle && event->pmu->unthrottle != NULL);
3864 if (hwc->interrupts != MAX_INTERRUPTS) {
3866 if (HZ * hwc->interrupts >
3867 (u64)sysctl_perf_event_sample_rate) {
3868 hwc->interrupts = MAX_INTERRUPTS;
3869 perf_log_throttle(event, 0);
3874 * Keep re-disabling events even though on the previous
3875 * pass we disabled it - just in case we raced with a
3876 * sched-in and the event got enabled again:
3882 if (event->attr.freq) {
3883 u64 now = perf_clock();
3884 s64 delta = now - hwc->freq_time_stamp;
3886 hwc->freq_time_stamp = now;
3888 if (delta > 0 && delta < 2*TICK_NSEC)
3889 perf_adjust_period(event, delta, hwc->last_period);
3893 * XXX event_limit might not quite work as expected on inherited
3897 event->pending_kill = POLL_IN;
3898 if (events && atomic_dec_and_test(&event->event_limit)) {
3900 event->pending_kill = POLL_HUP;
3902 event->pending_disable = 1;
3903 perf_pending_queue(&event->pending,
3904 perf_pending_event);
3906 perf_event_disable(event);
3909 if (event->overflow_handler)
3910 event->overflow_handler(event, nmi, data, regs);
3912 perf_event_output(event, nmi, data, regs);
3917 int perf_event_overflow(struct perf_event *event, int nmi,
3918 struct perf_sample_data *data,
3919 struct pt_regs *regs)
3921 return __perf_event_overflow(event, nmi, 1, data, regs);
3925 * Generic software event infrastructure
3929 * We directly increment event->count and keep a second value in
3930 * event->hw.period_left to count intervals. This period event
3931 * is kept in the range [-sample_period, 0] so that we can use the
3935 static u64 perf_swevent_set_period(struct perf_event *event)
3937 struct hw_perf_event *hwc = &event->hw;
3938 u64 period = hwc->last_period;
3942 hwc->last_period = hwc->sample_period;
3945 old = val = atomic64_read(&hwc->period_left);
3949 nr = div64_u64(period + val, period);
3950 offset = nr * period;
3952 if (atomic64_cmpxchg(&hwc->period_left, old, val) != old)
3958 static void perf_swevent_overflow(struct perf_event *event, u64 overflow,
3959 int nmi, struct perf_sample_data *data,
3960 struct pt_regs *regs)
3962 struct hw_perf_event *hwc = &event->hw;
3965 data->period = event->hw.last_period;
3967 overflow = perf_swevent_set_period(event);
3969 if (hwc->interrupts == MAX_INTERRUPTS)
3972 for (; overflow; overflow--) {
3973 if (__perf_event_overflow(event, nmi, throttle,
3976 * We inhibit the overflow from happening when
3977 * hwc->interrupts == MAX_INTERRUPTS.
3985 static void perf_swevent_unthrottle(struct perf_event *event)
3988 * Nothing to do, we already reset hwc->interrupts.
3992 static void perf_swevent_add(struct perf_event *event, u64 nr,
3993 int nmi, struct perf_sample_data *data,
3994 struct pt_regs *regs)
3996 struct hw_perf_event *hwc = &event->hw;
3998 atomic64_add(nr, &event->count);
4003 if (!hwc->sample_period)
4006 if (nr == 1 && hwc->sample_period == 1 && !event->attr.freq)
4007 return perf_swevent_overflow(event, 1, nmi, data, regs);
4009 if (atomic64_add_negative(nr, &hwc->period_left))
4012 perf_swevent_overflow(event, 0, nmi, data, regs);
4015 static int perf_tp_event_match(struct perf_event *event,
4016 struct perf_sample_data *data);
4018 static int perf_exclude_event(struct perf_event *event,
4019 struct pt_regs *regs)
4022 if (event->attr.exclude_user && user_mode(regs))
4025 if (event->attr.exclude_kernel && !user_mode(regs))
4032 static int perf_swevent_match(struct perf_event *event,
4033 enum perf_type_id type,
4035 struct perf_sample_data *data,
4036 struct pt_regs *regs)
4038 if (event->attr.type != type)
4041 if (event->attr.config != event_id)
4044 if (perf_exclude_event(event, regs))
4047 if (event->attr.type == PERF_TYPE_TRACEPOINT &&
4048 !perf_tp_event_match(event, data))
4054 static inline u64 swevent_hash(u64 type, u32 event_id)
4056 u64 val = event_id | (type << 32);
4058 return hash_64(val, SWEVENT_HLIST_BITS);
4061 static struct hlist_head *
4062 find_swevent_head(struct perf_cpu_context *ctx, u64 type, u32 event_id)
4065 struct swevent_hlist *hlist;
4067 hash = swevent_hash(type, event_id);
4069 hlist = rcu_dereference(ctx->swevent_hlist);
4073 return &hlist->heads[hash];
4076 static void do_perf_sw_event(enum perf_type_id type, u32 event_id,
4078 struct perf_sample_data *data,
4079 struct pt_regs *regs)
4081 struct perf_cpu_context *cpuctx;
4082 struct perf_event *event;
4083 struct hlist_node *node;
4084 struct hlist_head *head;
4086 cpuctx = &__get_cpu_var(perf_cpu_context);
4090 head = find_swevent_head(cpuctx, type, event_id);
4095 hlist_for_each_entry_rcu(event, node, head, hlist_entry) {
4096 if (perf_swevent_match(event, type, event_id, data, regs))
4097 perf_swevent_add(event, nr, nmi, data, regs);
4103 int perf_swevent_get_recursion_context(void)
4105 struct perf_cpu_context *cpuctx = &get_cpu_var(perf_cpu_context);
4112 else if (in_softirq())
4117 if (cpuctx->recursion[rctx]) {
4118 put_cpu_var(perf_cpu_context);
4122 cpuctx->recursion[rctx]++;
4127 EXPORT_SYMBOL_GPL(perf_swevent_get_recursion_context);
4129 void perf_swevent_put_recursion_context(int rctx)
4131 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
4133 cpuctx->recursion[rctx]--;
4134 put_cpu_var(perf_cpu_context);
4136 EXPORT_SYMBOL_GPL(perf_swevent_put_recursion_context);
4139 void __perf_sw_event(u32 event_id, u64 nr, int nmi,
4140 struct pt_regs *regs, u64 addr)
4142 struct perf_sample_data data;
4145 rctx = perf_swevent_get_recursion_context();
4149 perf_sample_data_init(&data, addr);
4151 do_perf_sw_event(PERF_TYPE_SOFTWARE, event_id, nr, nmi, &data, regs);
4153 perf_swevent_put_recursion_context(rctx);
4156 static void perf_swevent_read(struct perf_event *event)
4160 static int perf_swevent_enable(struct perf_event *event)
4162 struct hw_perf_event *hwc = &event->hw;
4163 struct perf_cpu_context *cpuctx;
4164 struct hlist_head *head;
4166 cpuctx = &__get_cpu_var(perf_cpu_context);
4168 if (hwc->sample_period) {
4169 hwc->last_period = hwc->sample_period;
4170 perf_swevent_set_period(event);
4173 head = find_swevent_head(cpuctx, event->attr.type, event->attr.config);
4174 if (WARN_ON_ONCE(!head))
4177 hlist_add_head_rcu(&event->hlist_entry, head);
4182 static void perf_swevent_disable(struct perf_event *event)
4184 hlist_del_rcu(&event->hlist_entry);
4187 static const struct pmu perf_ops_generic = {
4188 .enable = perf_swevent_enable,
4189 .disable = perf_swevent_disable,
4190 .read = perf_swevent_read,
4191 .unthrottle = perf_swevent_unthrottle,
4195 * hrtimer based swevent callback
4198 static enum hrtimer_restart perf_swevent_hrtimer(struct hrtimer *hrtimer)
4200 enum hrtimer_restart ret = HRTIMER_RESTART;
4201 struct perf_sample_data data;
4202 struct pt_regs *regs;
4203 struct perf_event *event;
4206 event = container_of(hrtimer, struct perf_event, hw.hrtimer);
4207 event->pmu->read(event);
4209 perf_sample_data_init(&data, 0);
4210 data.period = event->hw.last_period;
4211 regs = get_irq_regs();
4213 if (regs && !perf_exclude_event(event, regs)) {
4214 if (!(event->attr.exclude_idle && current->pid == 0))
4215 if (perf_event_overflow(event, 0, &data, regs))
4216 ret = HRTIMER_NORESTART;
4219 period = max_t(u64, 10000, event->hw.sample_period);
4220 hrtimer_forward_now(hrtimer, ns_to_ktime(period));
4225 static void perf_swevent_start_hrtimer(struct perf_event *event)
4227 struct hw_perf_event *hwc = &event->hw;
4229 hrtimer_init(&hwc->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
4230 hwc->hrtimer.function = perf_swevent_hrtimer;
4231 if (hwc->sample_period) {
4234 if (hwc->remaining) {
4235 if (hwc->remaining < 0)
4238 period = hwc->remaining;
4241 period = max_t(u64, 10000, hwc->sample_period);
4243 __hrtimer_start_range_ns(&hwc->hrtimer,
4244 ns_to_ktime(period), 0,
4245 HRTIMER_MODE_REL, 0);
4249 static void perf_swevent_cancel_hrtimer(struct perf_event *event)
4251 struct hw_perf_event *hwc = &event->hw;
4253 if (hwc->sample_period) {
4254 ktime_t remaining = hrtimer_get_remaining(&hwc->hrtimer);
4255 hwc->remaining = ktime_to_ns(remaining);
4257 hrtimer_cancel(&hwc->hrtimer);
4262 * Software event: cpu wall time clock
4265 static void cpu_clock_perf_event_update(struct perf_event *event)
4267 int cpu = raw_smp_processor_id();
4271 now = cpu_clock(cpu);
4272 prev = atomic64_xchg(&event->hw.prev_count, now);
4273 atomic64_add(now - prev, &event->count);
4276 static int cpu_clock_perf_event_enable(struct perf_event *event)
4278 struct hw_perf_event *hwc = &event->hw;
4279 int cpu = raw_smp_processor_id();
4281 atomic64_set(&hwc->prev_count, cpu_clock(cpu));
4282 perf_swevent_start_hrtimer(event);
4287 static void cpu_clock_perf_event_disable(struct perf_event *event)
4289 perf_swevent_cancel_hrtimer(event);
4290 cpu_clock_perf_event_update(event);
4293 static void cpu_clock_perf_event_read(struct perf_event *event)
4295 cpu_clock_perf_event_update(event);
4298 static const struct pmu perf_ops_cpu_clock = {
4299 .enable = cpu_clock_perf_event_enable,
4300 .disable = cpu_clock_perf_event_disable,
4301 .read = cpu_clock_perf_event_read,
4305 * Software event: task time clock
4308 static void task_clock_perf_event_update(struct perf_event *event, u64 now)
4313 prev = atomic64_xchg(&event->hw.prev_count, now);
4315 atomic64_add(delta, &event->count);
4318 static int task_clock_perf_event_enable(struct perf_event *event)
4320 struct hw_perf_event *hwc = &event->hw;
4323 now = event->ctx->time;
4325 atomic64_set(&hwc->prev_count, now);
4327 perf_swevent_start_hrtimer(event);
4332 static void task_clock_perf_event_disable(struct perf_event *event)
4334 perf_swevent_cancel_hrtimer(event);
4335 task_clock_perf_event_update(event, event->ctx->time);
4339 static void task_clock_perf_event_read(struct perf_event *event)
4344 update_context_time(event->ctx);
4345 time = event->ctx->time;
4347 u64 now = perf_clock();
4348 u64 delta = now - event->ctx->timestamp;
4349 time = event->ctx->time + delta;
4352 task_clock_perf_event_update(event, time);
4355 static const struct pmu perf_ops_task_clock = {
4356 .enable = task_clock_perf_event_enable,
4357 .disable = task_clock_perf_event_disable,
4358 .read = task_clock_perf_event_read,
4361 static void swevent_hlist_release_rcu(struct rcu_head *rcu_head)
4363 struct swevent_hlist *hlist;
4365 hlist = container_of(rcu_head, struct swevent_hlist, rcu_head);
4369 static void swevent_hlist_release(struct perf_cpu_context *cpuctx)
4371 struct swevent_hlist *hlist;
4373 if (!cpuctx->swevent_hlist)
4376 hlist = cpuctx->swevent_hlist;
4377 rcu_assign_pointer(cpuctx->swevent_hlist, NULL);
4378 call_rcu(&hlist->rcu_head, swevent_hlist_release_rcu);
4381 static void swevent_hlist_put_cpu(struct perf_event *event, int cpu)
4383 struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
4385 mutex_lock(&cpuctx->hlist_mutex);
4387 if (!--cpuctx->hlist_refcount)
4388 swevent_hlist_release(cpuctx);
4390 mutex_unlock(&cpuctx->hlist_mutex);
4393 static void swevent_hlist_put(struct perf_event *event)
4397 if (event->cpu != -1) {
4398 swevent_hlist_put_cpu(event, event->cpu);
4402 for_each_possible_cpu(cpu)
4403 swevent_hlist_put_cpu(event, cpu);
4406 static int swevent_hlist_get_cpu(struct perf_event *event, int cpu)
4408 struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
4411 mutex_lock(&cpuctx->hlist_mutex);
4413 if (!cpuctx->swevent_hlist && cpu_online(cpu)) {
4414 struct swevent_hlist *hlist;
4416 hlist = kzalloc(sizeof(*hlist), GFP_KERNEL);
4421 rcu_assign_pointer(cpuctx->swevent_hlist, hlist);
4423 cpuctx->hlist_refcount++;
4425 mutex_unlock(&cpuctx->hlist_mutex);
4430 static int swevent_hlist_get(struct perf_event *event)
4433 int cpu, failed_cpu;
4435 if (event->cpu != -1)
4436 return swevent_hlist_get_cpu(event, event->cpu);
4439 for_each_possible_cpu(cpu) {
4440 err = swevent_hlist_get_cpu(event, cpu);
4450 for_each_possible_cpu(cpu) {
4451 if (cpu == failed_cpu)
4453 swevent_hlist_put_cpu(event, cpu);
4460 #ifdef CONFIG_EVENT_TRACING
4462 void perf_tp_event(int event_id, u64 addr, u64 count, void *record,
4463 int entry_size, struct pt_regs *regs)
4465 struct perf_sample_data data;
4466 struct perf_raw_record raw = {
4471 perf_sample_data_init(&data, addr);
4474 /* Trace events already protected against recursion */
4475 do_perf_sw_event(PERF_TYPE_TRACEPOINT, event_id, count, 1,
4478 EXPORT_SYMBOL_GPL(perf_tp_event);
4480 static int perf_tp_event_match(struct perf_event *event,
4481 struct perf_sample_data *data)
4483 void *record = data->raw->data;
4485 if (likely(!event->filter) || filter_match_preds(event->filter, record))
4490 static void tp_perf_event_destroy(struct perf_event *event)
4492 perf_trace_disable(event->attr.config);
4493 swevent_hlist_put(event);
4496 static const struct pmu *tp_perf_event_init(struct perf_event *event)
4501 * Raw tracepoint data is a severe data leak, only allow root to
4504 if ((event->attr.sample_type & PERF_SAMPLE_RAW) &&
4505 perf_paranoid_tracepoint_raw() &&
4506 !capable(CAP_SYS_ADMIN))
4507 return ERR_PTR(-EPERM);
4509 if (perf_trace_enable(event->attr.config))
4512 event->destroy = tp_perf_event_destroy;
4513 err = swevent_hlist_get(event);
4515 perf_trace_disable(event->attr.config);
4516 return ERR_PTR(err);
4519 return &perf_ops_generic;
4522 static int perf_event_set_filter(struct perf_event *event, void __user *arg)
4527 if (event->attr.type != PERF_TYPE_TRACEPOINT)
4530 filter_str = strndup_user(arg, PAGE_SIZE);
4531 if (IS_ERR(filter_str))
4532 return PTR_ERR(filter_str);
4534 ret = ftrace_profile_set_filter(event, event->attr.config, filter_str);
4540 static void perf_event_free_filter(struct perf_event *event)
4542 ftrace_profile_free_filter(event);
4547 static int perf_tp_event_match(struct perf_event *event,
4548 struct perf_sample_data *data)
4553 static const struct pmu *tp_perf_event_init(struct perf_event *event)
4558 static int perf_event_set_filter(struct perf_event *event, void __user *arg)
4563 static void perf_event_free_filter(struct perf_event *event)
4567 #endif /* CONFIG_EVENT_TRACING */
4569 #ifdef CONFIG_HAVE_HW_BREAKPOINT
4570 static void bp_perf_event_destroy(struct perf_event *event)
4572 release_bp_slot(event);
4575 static const struct pmu *bp_perf_event_init(struct perf_event *bp)
4579 err = register_perf_hw_breakpoint(bp);
4581 return ERR_PTR(err);
4583 bp->destroy = bp_perf_event_destroy;
4585 return &perf_ops_bp;
4588 void perf_bp_event(struct perf_event *bp, void *data)
4590 struct perf_sample_data sample;
4591 struct pt_regs *regs = data;
4593 perf_sample_data_init(&sample, bp->attr.bp_addr);
4595 if (!perf_exclude_event(bp, regs))
4596 perf_swevent_add(bp, 1, 1, &sample, regs);
4599 static const struct pmu *bp_perf_event_init(struct perf_event *bp)
4604 void perf_bp_event(struct perf_event *bp, void *regs)
4609 atomic_t perf_swevent_enabled[PERF_COUNT_SW_MAX];
4611 static void sw_perf_event_destroy(struct perf_event *event)
4613 u64 event_id = event->attr.config;
4615 WARN_ON(event->parent);
4617 atomic_dec(&perf_swevent_enabled[event_id]);
4618 swevent_hlist_put(event);
4621 static const struct pmu *sw_perf_event_init(struct perf_event *event)
4623 const struct pmu *pmu = NULL;
4624 u64 event_id = event->attr.config;
4627 * Software events (currently) can't in general distinguish
4628 * between user, kernel and hypervisor events.
4629 * However, context switches and cpu migrations are considered
4630 * to be kernel events, and page faults are never hypervisor
4634 case PERF_COUNT_SW_CPU_CLOCK:
4635 pmu = &perf_ops_cpu_clock;
4638 case PERF_COUNT_SW_TASK_CLOCK:
4640 * If the user instantiates this as a per-cpu event,
4641 * use the cpu_clock event instead.
4643 if (event->ctx->task)
4644 pmu = &perf_ops_task_clock;
4646 pmu = &perf_ops_cpu_clock;
4649 case PERF_COUNT_SW_PAGE_FAULTS:
4650 case PERF_COUNT_SW_PAGE_FAULTS_MIN:
4651 case PERF_COUNT_SW_PAGE_FAULTS_MAJ:
4652 case PERF_COUNT_SW_CONTEXT_SWITCHES:
4653 case PERF_COUNT_SW_CPU_MIGRATIONS:
4654 case PERF_COUNT_SW_ALIGNMENT_FAULTS:
4655 case PERF_COUNT_SW_EMULATION_FAULTS:
4656 if (!event->parent) {
4659 err = swevent_hlist_get(event);
4661 return ERR_PTR(err);
4663 atomic_inc(&perf_swevent_enabled[event_id]);
4664 event->destroy = sw_perf_event_destroy;
4666 pmu = &perf_ops_generic;
4674 * Allocate and initialize a event structure
4676 static struct perf_event *
4677 perf_event_alloc(struct perf_event_attr *attr,
4679 struct perf_event_context *ctx,
4680 struct perf_event *group_leader,
4681 struct perf_event *parent_event,
4682 perf_overflow_handler_t overflow_handler,
4685 const struct pmu *pmu;
4686 struct perf_event *event;
4687 struct hw_perf_event *hwc;
4690 event = kzalloc(sizeof(*event), gfpflags);
4692 return ERR_PTR(-ENOMEM);
4695 * Single events are their own group leaders, with an
4696 * empty sibling list:
4699 group_leader = event;
4701 mutex_init(&event->child_mutex);
4702 INIT_LIST_HEAD(&event->child_list);
4704 INIT_LIST_HEAD(&event->group_entry);
4705 INIT_LIST_HEAD(&event->event_entry);
4706 INIT_LIST_HEAD(&event->sibling_list);
4707 init_waitqueue_head(&event->waitq);
4709 mutex_init(&event->mmap_mutex);
4712 event->attr = *attr;
4713 event->group_leader = group_leader;
4718 event->parent = parent_event;
4720 event->ns = get_pid_ns(current->nsproxy->pid_ns);
4721 event->id = atomic64_inc_return(&perf_event_id);
4723 event->state = PERF_EVENT_STATE_INACTIVE;
4725 if (!overflow_handler && parent_event)
4726 overflow_handler = parent_event->overflow_handler;
4728 event->overflow_handler = overflow_handler;
4731 event->state = PERF_EVENT_STATE_OFF;
4736 hwc->sample_period = attr->sample_period;
4737 if (attr->freq && attr->sample_freq)
4738 hwc->sample_period = 1;
4739 hwc->last_period = hwc->sample_period;
4741 atomic64_set(&hwc->period_left, hwc->sample_period);
4744 * we currently do not support PERF_FORMAT_GROUP on inherited events
4746 if (attr->inherit && (attr->read_format & PERF_FORMAT_GROUP))
4749 switch (attr->type) {
4751 case PERF_TYPE_HARDWARE:
4752 case PERF_TYPE_HW_CACHE:
4753 pmu = hw_perf_event_init(event);
4756 case PERF_TYPE_SOFTWARE:
4757 pmu = sw_perf_event_init(event);
4760 case PERF_TYPE_TRACEPOINT:
4761 pmu = tp_perf_event_init(event);
4764 case PERF_TYPE_BREAKPOINT:
4765 pmu = bp_perf_event_init(event);
4776 else if (IS_ERR(pmu))
4781 put_pid_ns(event->ns);
4783 return ERR_PTR(err);
4788 if (!event->parent) {
4789 atomic_inc(&nr_events);
4790 if (event->attr.mmap)
4791 atomic_inc(&nr_mmap_events);
4792 if (event->attr.comm)
4793 atomic_inc(&nr_comm_events);
4794 if (event->attr.task)
4795 atomic_inc(&nr_task_events);
4801 static int perf_copy_attr(struct perf_event_attr __user *uattr,
4802 struct perf_event_attr *attr)
4807 if (!access_ok(VERIFY_WRITE, uattr, PERF_ATTR_SIZE_VER0))
4811 * zero the full structure, so that a short copy will be nice.
4813 memset(attr, 0, sizeof(*attr));
4815 ret = get_user(size, &uattr->size);
4819 if (size > PAGE_SIZE) /* silly large */
4822 if (!size) /* abi compat */
4823 size = PERF_ATTR_SIZE_VER0;
4825 if (size < PERF_ATTR_SIZE_VER0)
4829 * If we're handed a bigger struct than we know of,
4830 * ensure all the unknown bits are 0 - i.e. new
4831 * user-space does not rely on any kernel feature
4832 * extensions we dont know about yet.
4834 if (size > sizeof(*attr)) {
4835 unsigned char __user *addr;
4836 unsigned char __user *end;
4839 addr = (void __user *)uattr + sizeof(*attr);
4840 end = (void __user *)uattr + size;
4842 for (; addr < end; addr++) {
4843 ret = get_user(val, addr);
4849 size = sizeof(*attr);
4852 ret = copy_from_user(attr, uattr, size);
4857 * If the type exists, the corresponding creation will verify
4860 if (attr->type >= PERF_TYPE_MAX)
4863 if (attr->__reserved_1)
4866 if (attr->sample_type & ~(PERF_SAMPLE_MAX-1))
4869 if (attr->read_format & ~(PERF_FORMAT_MAX-1))
4876 put_user(sizeof(*attr), &uattr->size);
4881 static int perf_event_set_output(struct perf_event *event, int output_fd)
4883 struct perf_event *output_event = NULL;
4884 struct file *output_file = NULL;
4885 struct perf_event *old_output;
4886 int fput_needed = 0;
4892 output_file = fget_light(output_fd, &fput_needed);
4896 if (output_file->f_op != &perf_fops)
4899 output_event = output_file->private_data;
4901 /* Don't chain output fds */
4902 if (output_event->output)
4905 /* Don't set an output fd when we already have an output channel */
4909 atomic_long_inc(&output_file->f_count);
4912 mutex_lock(&event->mmap_mutex);
4913 old_output = event->output;
4914 rcu_assign_pointer(event->output, output_event);
4915 mutex_unlock(&event->mmap_mutex);
4919 * we need to make sure no existing perf_output_*()
4920 * is still referencing this event.
4923 fput(old_output->filp);
4928 fput_light(output_file, fput_needed);
4933 * sys_perf_event_open - open a performance event, associate it to a task/cpu
4935 * @attr_uptr: event_id type attributes for monitoring/sampling
4938 * @group_fd: group leader event fd
4940 SYSCALL_DEFINE5(perf_event_open,
4941 struct perf_event_attr __user *, attr_uptr,
4942 pid_t, pid, int, cpu, int, group_fd, unsigned long, flags)
4944 struct perf_event *event, *group_leader;
4945 struct perf_event_attr attr;
4946 struct perf_event_context *ctx;
4947 struct file *event_file = NULL;
4948 struct file *group_file = NULL;
4949 int fput_needed = 0;
4950 int fput_needed2 = 0;
4953 /* for future expandability... */
4954 if (flags & ~(PERF_FLAG_FD_NO_GROUP | PERF_FLAG_FD_OUTPUT))
4957 err = perf_copy_attr(attr_uptr, &attr);
4961 if (!attr.exclude_kernel) {
4962 if (perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN))
4967 if (attr.sample_freq > sysctl_perf_event_sample_rate)
4972 * Get the target context (task or percpu):
4974 ctx = find_get_context(pid, cpu);
4976 return PTR_ERR(ctx);
4979 * Look up the group leader (we will attach this event to it):
4981 group_leader = NULL;
4982 if (group_fd != -1 && !(flags & PERF_FLAG_FD_NO_GROUP)) {
4984 group_file = fget_light(group_fd, &fput_needed);
4986 goto err_put_context;
4987 if (group_file->f_op != &perf_fops)
4988 goto err_put_context;
4990 group_leader = group_file->private_data;
4992 * Do not allow a recursive hierarchy (this new sibling
4993 * becoming part of another group-sibling):
4995 if (group_leader->group_leader != group_leader)
4996 goto err_put_context;
4998 * Do not allow to attach to a group in a different
4999 * task or CPU context:
5001 if (group_leader->ctx != ctx)
5002 goto err_put_context;
5004 * Only a group leader can be exclusive or pinned
5006 if (attr.exclusive || attr.pinned)
5007 goto err_put_context;
5010 event = perf_event_alloc(&attr, cpu, ctx, group_leader,
5011 NULL, NULL, GFP_KERNEL);
5012 err = PTR_ERR(event);
5014 goto err_put_context;
5016 err = anon_inode_getfd("[perf_event]", &perf_fops, event, O_RDWR);
5018 goto err_free_put_context;
5020 event_file = fget_light(err, &fput_needed2);
5022 goto err_free_put_context;
5024 if (flags & PERF_FLAG_FD_OUTPUT) {
5025 err = perf_event_set_output(event, group_fd);
5027 goto err_fput_free_put_context;
5030 event->filp = event_file;
5031 WARN_ON_ONCE(ctx->parent_ctx);
5032 mutex_lock(&ctx->mutex);
5033 perf_install_in_context(ctx, event, cpu);
5035 mutex_unlock(&ctx->mutex);
5037 event->owner = current;
5038 get_task_struct(current);
5039 mutex_lock(¤t->perf_event_mutex);
5040 list_add_tail(&event->owner_entry, ¤t->perf_event_list);
5041 mutex_unlock(¤t->perf_event_mutex);
5043 err_fput_free_put_context:
5044 fput_light(event_file, fput_needed2);
5046 err_free_put_context:
5054 fput_light(group_file, fput_needed);
5060 * perf_event_create_kernel_counter
5062 * @attr: attributes of the counter to create
5063 * @cpu: cpu in which the counter is bound
5064 * @pid: task to profile
5067 perf_event_create_kernel_counter(struct perf_event_attr *attr, int cpu,
5069 perf_overflow_handler_t overflow_handler)
5071 struct perf_event *event;
5072 struct perf_event_context *ctx;
5076 * Get the target context (task or percpu):
5079 ctx = find_get_context(pid, cpu);
5085 event = perf_event_alloc(attr, cpu, ctx, NULL,
5086 NULL, overflow_handler, GFP_KERNEL);
5087 if (IS_ERR(event)) {
5088 err = PTR_ERR(event);
5089 goto err_put_context;
5093 WARN_ON_ONCE(ctx->parent_ctx);
5094 mutex_lock(&ctx->mutex);
5095 perf_install_in_context(ctx, event, cpu);
5097 mutex_unlock(&ctx->mutex);
5099 event->owner = current;
5100 get_task_struct(current);
5101 mutex_lock(¤t->perf_event_mutex);
5102 list_add_tail(&event->owner_entry, ¤t->perf_event_list);
5103 mutex_unlock(¤t->perf_event_mutex);
5110 return ERR_PTR(err);
5112 EXPORT_SYMBOL_GPL(perf_event_create_kernel_counter);
5115 * inherit a event from parent task to child task:
5117 static struct perf_event *
5118 inherit_event(struct perf_event *parent_event,
5119 struct task_struct *parent,
5120 struct perf_event_context *parent_ctx,
5121 struct task_struct *child,
5122 struct perf_event *group_leader,
5123 struct perf_event_context *child_ctx)
5125 struct perf_event *child_event;
5128 * Instead of creating recursive hierarchies of events,
5129 * we link inherited events back to the original parent,
5130 * which has a filp for sure, which we use as the reference
5133 if (parent_event->parent)
5134 parent_event = parent_event->parent;
5136 child_event = perf_event_alloc(&parent_event->attr,
5137 parent_event->cpu, child_ctx,
5138 group_leader, parent_event,
5140 if (IS_ERR(child_event))
5145 * Make the child state follow the state of the parent event,
5146 * not its attr.disabled bit. We hold the parent's mutex,
5147 * so we won't race with perf_event_{en, dis}able_family.
5149 if (parent_event->state >= PERF_EVENT_STATE_INACTIVE)
5150 child_event->state = PERF_EVENT_STATE_INACTIVE;
5152 child_event->state = PERF_EVENT_STATE_OFF;
5154 if (parent_event->attr.freq) {
5155 u64 sample_period = parent_event->hw.sample_period;
5156 struct hw_perf_event *hwc = &child_event->hw;
5158 hwc->sample_period = sample_period;
5159 hwc->last_period = sample_period;
5161 atomic64_set(&hwc->period_left, sample_period);
5164 child_event->overflow_handler = parent_event->overflow_handler;
5167 * Link it up in the child's context:
5169 add_event_to_ctx(child_event, child_ctx);
5172 * Get a reference to the parent filp - we will fput it
5173 * when the child event exits. This is safe to do because
5174 * we are in the parent and we know that the filp still
5175 * exists and has a nonzero count:
5177 atomic_long_inc(&parent_event->filp->f_count);
5180 * Link this into the parent event's child list
5182 WARN_ON_ONCE(parent_event->ctx->parent_ctx);
5183 mutex_lock(&parent_event->child_mutex);
5184 list_add_tail(&child_event->child_list, &parent_event->child_list);
5185 mutex_unlock(&parent_event->child_mutex);
5190 static int inherit_group(struct perf_event *parent_event,
5191 struct task_struct *parent,
5192 struct perf_event_context *parent_ctx,
5193 struct task_struct *child,
5194 struct perf_event_context *child_ctx)
5196 struct perf_event *leader;
5197 struct perf_event *sub;
5198 struct perf_event *child_ctr;
5200 leader = inherit_event(parent_event, parent, parent_ctx,
5201 child, NULL, child_ctx);
5203 return PTR_ERR(leader);
5204 list_for_each_entry(sub, &parent_event->sibling_list, group_entry) {
5205 child_ctr = inherit_event(sub, parent, parent_ctx,
5206 child, leader, child_ctx);
5207 if (IS_ERR(child_ctr))
5208 return PTR_ERR(child_ctr);
5213 static void sync_child_event(struct perf_event *child_event,
5214 struct task_struct *child)
5216 struct perf_event *parent_event = child_event->parent;
5219 if (child_event->attr.inherit_stat)
5220 perf_event_read_event(child_event, child);
5222 child_val = atomic64_read(&child_event->count);
5225 * Add back the child's count to the parent's count:
5227 atomic64_add(child_val, &parent_event->count);
5228 atomic64_add(child_event->total_time_enabled,
5229 &parent_event->child_total_time_enabled);
5230 atomic64_add(child_event->total_time_running,
5231 &parent_event->child_total_time_running);
5234 * Remove this event from the parent's list
5236 WARN_ON_ONCE(parent_event->ctx->parent_ctx);
5237 mutex_lock(&parent_event->child_mutex);
5238 list_del_init(&child_event->child_list);
5239 mutex_unlock(&parent_event->child_mutex);
5242 * Release the parent event, if this was the last
5245 fput(parent_event->filp);
5249 __perf_event_exit_task(struct perf_event *child_event,
5250 struct perf_event_context *child_ctx,
5251 struct task_struct *child)
5253 struct perf_event *parent_event;
5255 perf_event_remove_from_context(child_event);
5257 parent_event = child_event->parent;
5259 * It can happen that parent exits first, and has events
5260 * that are still around due to the child reference. These
5261 * events need to be zapped - but otherwise linger.
5264 sync_child_event(child_event, child);
5265 free_event(child_event);
5270 * When a child task exits, feed back event values to parent events.
5272 void perf_event_exit_task(struct task_struct *child)
5274 struct perf_event *child_event, *tmp;
5275 struct perf_event_context *child_ctx;
5276 unsigned long flags;
5278 if (likely(!child->perf_event_ctxp)) {
5279 perf_event_task(child, NULL, 0);
5283 local_irq_save(flags);
5285 * We can't reschedule here because interrupts are disabled,
5286 * and either child is current or it is a task that can't be
5287 * scheduled, so we are now safe from rescheduling changing
5290 child_ctx = child->perf_event_ctxp;
5291 __perf_event_task_sched_out(child_ctx);
5294 * Take the context lock here so that if find_get_context is
5295 * reading child->perf_event_ctxp, we wait until it has
5296 * incremented the context's refcount before we do put_ctx below.
5298 raw_spin_lock(&child_ctx->lock);
5299 child->perf_event_ctxp = NULL;
5301 * If this context is a clone; unclone it so it can't get
5302 * swapped to another process while we're removing all
5303 * the events from it.
5305 unclone_ctx(child_ctx);
5306 update_context_time(child_ctx);
5307 raw_spin_unlock_irqrestore(&child_ctx->lock, flags);
5310 * Report the task dead after unscheduling the events so that we
5311 * won't get any samples after PERF_RECORD_EXIT. We can however still
5312 * get a few PERF_RECORD_READ events.
5314 perf_event_task(child, child_ctx, 0);
5317 * We can recurse on the same lock type through:
5319 * __perf_event_exit_task()
5320 * sync_child_event()
5321 * fput(parent_event->filp)
5323 * mutex_lock(&ctx->mutex)
5325 * But since its the parent context it won't be the same instance.
5327 mutex_lock(&child_ctx->mutex);
5330 list_for_each_entry_safe(child_event, tmp, &child_ctx->pinned_groups,
5332 __perf_event_exit_task(child_event, child_ctx, child);
5334 list_for_each_entry_safe(child_event, tmp, &child_ctx->flexible_groups,
5336 __perf_event_exit_task(child_event, child_ctx, child);
5339 * If the last event was a group event, it will have appended all
5340 * its siblings to the list, but we obtained 'tmp' before that which
5341 * will still point to the list head terminating the iteration.
5343 if (!list_empty(&child_ctx->pinned_groups) ||
5344 !list_empty(&child_ctx->flexible_groups))
5347 mutex_unlock(&child_ctx->mutex);
5352 static void perf_free_event(struct perf_event *event,
5353 struct perf_event_context *ctx)
5355 struct perf_event *parent = event->parent;
5357 if (WARN_ON_ONCE(!parent))
5360 mutex_lock(&parent->child_mutex);
5361 list_del_init(&event->child_list);
5362 mutex_unlock(&parent->child_mutex);
5366 list_del_event(event, ctx);
5371 * free an unexposed, unused context as created by inheritance by
5372 * init_task below, used by fork() in case of fail.
5374 void perf_event_free_task(struct task_struct *task)
5376 struct perf_event_context *ctx = task->perf_event_ctxp;
5377 struct perf_event *event, *tmp;
5382 mutex_lock(&ctx->mutex);
5384 list_for_each_entry_safe(event, tmp, &ctx->pinned_groups, group_entry)
5385 perf_free_event(event, ctx);
5387 list_for_each_entry_safe(event, tmp, &ctx->flexible_groups,
5389 perf_free_event(event, ctx);
5391 if (!list_empty(&ctx->pinned_groups) ||
5392 !list_empty(&ctx->flexible_groups))
5395 mutex_unlock(&ctx->mutex);
5401 inherit_task_group(struct perf_event *event, struct task_struct *parent,
5402 struct perf_event_context *parent_ctx,
5403 struct task_struct *child,
5407 struct perf_event_context *child_ctx = child->perf_event_ctxp;
5409 if (!event->attr.inherit) {
5416 * This is executed from the parent task context, so
5417 * inherit events that have been marked for cloning.
5418 * First allocate and initialize a context for the
5422 child_ctx = kzalloc(sizeof(struct perf_event_context),
5427 __perf_event_init_context(child_ctx, child);
5428 child->perf_event_ctxp = child_ctx;
5429 get_task_struct(child);
5432 ret = inherit_group(event, parent, parent_ctx,
5443 * Initialize the perf_event context in task_struct
5445 int perf_event_init_task(struct task_struct *child)
5447 struct perf_event_context *child_ctx, *parent_ctx;
5448 struct perf_event_context *cloned_ctx;
5449 struct perf_event *event;
5450 struct task_struct *parent = current;
5451 int inherited_all = 1;
5454 child->perf_event_ctxp = NULL;
5456 mutex_init(&child->perf_event_mutex);
5457 INIT_LIST_HEAD(&child->perf_event_list);
5459 if (likely(!parent->perf_event_ctxp))
5463 * If the parent's context is a clone, pin it so it won't get
5466 parent_ctx = perf_pin_task_context(parent);
5469 * No need to check if parent_ctx != NULL here; since we saw
5470 * it non-NULL earlier, the only reason for it to become NULL
5471 * is if we exit, and since we're currently in the middle of
5472 * a fork we can't be exiting at the same time.
5476 * Lock the parent list. No need to lock the child - not PID
5477 * hashed yet and not running, so nobody can access it.
5479 mutex_lock(&parent_ctx->mutex);
5482 * We dont have to disable NMIs - we are only looking at
5483 * the list, not manipulating it:
5485 list_for_each_entry(event, &parent_ctx->pinned_groups, group_entry) {
5486 ret = inherit_task_group(event, parent, parent_ctx, child,
5492 list_for_each_entry(event, &parent_ctx->flexible_groups, group_entry) {
5493 ret = inherit_task_group(event, parent, parent_ctx, child,
5499 child_ctx = child->perf_event_ctxp;
5501 if (child_ctx && inherited_all) {
5503 * Mark the child context as a clone of the parent
5504 * context, or of whatever the parent is a clone of.
5505 * Note that if the parent is a clone, it could get
5506 * uncloned at any point, but that doesn't matter
5507 * because the list of events and the generation
5508 * count can't have changed since we took the mutex.
5510 cloned_ctx = rcu_dereference(parent_ctx->parent_ctx);
5512 child_ctx->parent_ctx = cloned_ctx;
5513 child_ctx->parent_gen = parent_ctx->parent_gen;
5515 child_ctx->parent_ctx = parent_ctx;
5516 child_ctx->parent_gen = parent_ctx->generation;
5518 get_ctx(child_ctx->parent_ctx);
5521 mutex_unlock(&parent_ctx->mutex);
5523 perf_unpin_context(parent_ctx);
5528 static void __init perf_event_init_all_cpus(void)
5531 struct perf_cpu_context *cpuctx;
5533 for_each_possible_cpu(cpu) {
5534 cpuctx = &per_cpu(perf_cpu_context, cpu);
5535 mutex_init(&cpuctx->hlist_mutex);
5536 __perf_event_init_context(&cpuctx->ctx, NULL);
5540 static void __cpuinit perf_event_init_cpu(int cpu)
5542 struct perf_cpu_context *cpuctx;
5544 cpuctx = &per_cpu(perf_cpu_context, cpu);
5546 spin_lock(&perf_resource_lock);
5547 cpuctx->max_pertask = perf_max_events - perf_reserved_percpu;
5548 spin_unlock(&perf_resource_lock);
5550 mutex_lock(&cpuctx->hlist_mutex);
5551 if (cpuctx->hlist_refcount > 0) {
5552 struct swevent_hlist *hlist;
5554 hlist = kzalloc(sizeof(*hlist), GFP_KERNEL);
5555 WARN_ON_ONCE(!hlist);
5556 rcu_assign_pointer(cpuctx->swevent_hlist, hlist);
5558 mutex_unlock(&cpuctx->hlist_mutex);
5561 #ifdef CONFIG_HOTPLUG_CPU
5562 static void __perf_event_exit_cpu(void *info)
5564 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
5565 struct perf_event_context *ctx = &cpuctx->ctx;
5566 struct perf_event *event, *tmp;
5568 list_for_each_entry_safe(event, tmp, &ctx->pinned_groups, group_entry)
5569 __perf_event_remove_from_context(event);
5570 list_for_each_entry_safe(event, tmp, &ctx->flexible_groups, group_entry)
5571 __perf_event_remove_from_context(event);
5573 static void perf_event_exit_cpu(int cpu)
5575 struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
5576 struct perf_event_context *ctx = &cpuctx->ctx;
5578 mutex_lock(&cpuctx->hlist_mutex);
5579 swevent_hlist_release(cpuctx);
5580 mutex_unlock(&cpuctx->hlist_mutex);
5582 mutex_lock(&ctx->mutex);
5583 smp_call_function_single(cpu, __perf_event_exit_cpu, NULL, 1);
5584 mutex_unlock(&ctx->mutex);
5587 static inline void perf_event_exit_cpu(int cpu) { }
5590 static int __cpuinit
5591 perf_cpu_notify(struct notifier_block *self, unsigned long action, void *hcpu)
5593 unsigned int cpu = (long)hcpu;
5597 case CPU_UP_PREPARE:
5598 case CPU_UP_PREPARE_FROZEN:
5599 perf_event_init_cpu(cpu);
5602 case CPU_DOWN_PREPARE:
5603 case CPU_DOWN_PREPARE_FROZEN:
5604 perf_event_exit_cpu(cpu);
5615 * This has to have a higher priority than migration_notifier in sched.c.
5617 static struct notifier_block __cpuinitdata perf_cpu_nb = {
5618 .notifier_call = perf_cpu_notify,
5622 void __init perf_event_init(void)
5624 perf_event_init_all_cpus();
5625 perf_cpu_notify(&perf_cpu_nb, (unsigned long)CPU_UP_PREPARE,
5626 (void *)(long)smp_processor_id());
5627 perf_cpu_notify(&perf_cpu_nb, (unsigned long)CPU_ONLINE,
5628 (void *)(long)smp_processor_id());
5629 register_cpu_notifier(&perf_cpu_nb);
5632 static ssize_t perf_show_reserve_percpu(struct sysdev_class *class,
5633 struct sysdev_class_attribute *attr,
5636 return sprintf(buf, "%d\n", perf_reserved_percpu);
5640 perf_set_reserve_percpu(struct sysdev_class *class,
5641 struct sysdev_class_attribute *attr,
5645 struct perf_cpu_context *cpuctx;
5649 err = strict_strtoul(buf, 10, &val);
5652 if (val > perf_max_events)
5655 spin_lock(&perf_resource_lock);
5656 perf_reserved_percpu = val;
5657 for_each_online_cpu(cpu) {
5658 cpuctx = &per_cpu(perf_cpu_context, cpu);
5659 raw_spin_lock_irq(&cpuctx->ctx.lock);
5660 mpt = min(perf_max_events - cpuctx->ctx.nr_events,
5661 perf_max_events - perf_reserved_percpu);
5662 cpuctx->max_pertask = mpt;
5663 raw_spin_unlock_irq(&cpuctx->ctx.lock);
5665 spin_unlock(&perf_resource_lock);
5670 static ssize_t perf_show_overcommit(struct sysdev_class *class,
5671 struct sysdev_class_attribute *attr,
5674 return sprintf(buf, "%d\n", perf_overcommit);
5678 perf_set_overcommit(struct sysdev_class *class,
5679 struct sysdev_class_attribute *attr,
5680 const char *buf, size_t count)
5685 err = strict_strtoul(buf, 10, &val);
5691 spin_lock(&perf_resource_lock);
5692 perf_overcommit = val;
5693 spin_unlock(&perf_resource_lock);
5698 static SYSDEV_CLASS_ATTR(
5701 perf_show_reserve_percpu,
5702 perf_set_reserve_percpu
5705 static SYSDEV_CLASS_ATTR(
5708 perf_show_overcommit,
5712 static struct attribute *perfclass_attrs[] = {
5713 &attr_reserve_percpu.attr,
5714 &attr_overcommit.attr,
5718 static struct attribute_group perfclass_attr_group = {
5719 .attrs = perfclass_attrs,
5720 .name = "perf_events",
5723 static int __init perf_event_sysfs_init(void)
5725 return sysfs_create_group(&cpu_sysdev_class.kset.kobj,
5726 &perfclass_attr_group);
5728 device_initcall(perf_event_sysfs_init);