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 /* Minimum for 128 pages + 1 for the user control page */
62 int sysctl_perf_event_mlock __read_mostly = 516; /* 'free' kb per user */
65 * max perf event sample rate
67 int sysctl_perf_event_sample_rate __read_mostly = 100000;
69 static atomic64_t perf_event_id;
72 * Lock for (sysadmin-configurable) event reservations:
74 static DEFINE_SPINLOCK(perf_resource_lock);
77 * Architecture provided APIs - weak aliases:
79 extern __weak const struct pmu *hw_perf_event_init(struct perf_event *event)
84 void __weak hw_perf_disable(void) { barrier(); }
85 void __weak hw_perf_enable(void) { barrier(); }
87 void __weak perf_event_print_debug(void) { }
89 static DEFINE_PER_CPU(int, perf_disable_count);
91 void perf_disable(void)
93 if (!__get_cpu_var(perf_disable_count)++)
97 void perf_enable(void)
99 if (!--__get_cpu_var(perf_disable_count))
103 static void get_ctx(struct perf_event_context *ctx)
105 WARN_ON(!atomic_inc_not_zero(&ctx->refcount));
108 static void free_ctx(struct rcu_head *head)
110 struct perf_event_context *ctx;
112 ctx = container_of(head, struct perf_event_context, rcu_head);
116 static void put_ctx(struct perf_event_context *ctx)
118 if (atomic_dec_and_test(&ctx->refcount)) {
120 put_ctx(ctx->parent_ctx);
122 put_task_struct(ctx->task);
123 call_rcu(&ctx->rcu_head, free_ctx);
127 static void unclone_ctx(struct perf_event_context *ctx)
129 if (ctx->parent_ctx) {
130 put_ctx(ctx->parent_ctx);
131 ctx->parent_ctx = NULL;
136 * If we inherit events we want to return the parent event id
139 static u64 primary_event_id(struct perf_event *event)
144 id = event->parent->id;
150 * Get the perf_event_context for a task and lock it.
151 * This has to cope with with the fact that until it is locked,
152 * the context could get moved to another task.
154 static struct perf_event_context *
155 perf_lock_task_context(struct task_struct *task, unsigned long *flags)
157 struct perf_event_context *ctx;
161 ctx = rcu_dereference(task->perf_event_ctxp);
164 * If this context is a clone of another, it might
165 * get swapped for another underneath us by
166 * perf_event_task_sched_out, though the
167 * rcu_read_lock() protects us from any context
168 * getting freed. Lock the context and check if it
169 * got swapped before we could get the lock, and retry
170 * if so. If we locked the right context, then it
171 * can't get swapped on us any more.
173 raw_spin_lock_irqsave(&ctx->lock, *flags);
174 if (ctx != rcu_dereference(task->perf_event_ctxp)) {
175 raw_spin_unlock_irqrestore(&ctx->lock, *flags);
179 if (!atomic_inc_not_zero(&ctx->refcount)) {
180 raw_spin_unlock_irqrestore(&ctx->lock, *flags);
189 * Get the context for a task and increment its pin_count so it
190 * can't get swapped to another task. This also increments its
191 * reference count so that the context can't get freed.
193 static struct perf_event_context *perf_pin_task_context(struct task_struct *task)
195 struct perf_event_context *ctx;
198 ctx = perf_lock_task_context(task, &flags);
201 raw_spin_unlock_irqrestore(&ctx->lock, flags);
206 static void perf_unpin_context(struct perf_event_context *ctx)
210 raw_spin_lock_irqsave(&ctx->lock, flags);
212 raw_spin_unlock_irqrestore(&ctx->lock, flags);
216 static inline u64 perf_clock(void)
218 return cpu_clock(raw_smp_processor_id());
222 * Update the record of the current time in a context.
224 static void update_context_time(struct perf_event_context *ctx)
226 u64 now = perf_clock();
228 ctx->time += now - ctx->timestamp;
229 ctx->timestamp = now;
233 * Update the total_time_enabled and total_time_running fields for a event.
235 static void update_event_times(struct perf_event *event)
237 struct perf_event_context *ctx = event->ctx;
240 if (event->state < PERF_EVENT_STATE_INACTIVE ||
241 event->group_leader->state < PERF_EVENT_STATE_INACTIVE)
247 run_end = event->tstamp_stopped;
249 event->total_time_enabled = run_end - event->tstamp_enabled;
251 if (event->state == PERF_EVENT_STATE_INACTIVE)
252 run_end = event->tstamp_stopped;
256 event->total_time_running = run_end - event->tstamp_running;
260 * Update total_time_enabled and total_time_running for all events in a group.
262 static void update_group_times(struct perf_event *leader)
264 struct perf_event *event;
266 update_event_times(leader);
267 list_for_each_entry(event, &leader->sibling_list, group_entry)
268 update_event_times(event);
271 static struct list_head *
272 ctx_group_list(struct perf_event *event, struct perf_event_context *ctx)
274 if (event->attr.pinned)
275 return &ctx->pinned_groups;
277 return &ctx->flexible_groups;
281 * Add a event from the lists for its context.
282 * Must be called with ctx->mutex and ctx->lock held.
285 list_add_event(struct perf_event *event, struct perf_event_context *ctx)
287 WARN_ON_ONCE(event->attach_state & PERF_ATTACH_CONTEXT);
288 event->attach_state |= PERF_ATTACH_CONTEXT;
291 * If we're a stand alone event or group leader, we go to the context
292 * list, group events are kept attached to the group so that
293 * perf_group_detach can, at all times, locate all siblings.
295 if (event->group_leader == event) {
296 struct list_head *list;
298 if (is_software_event(event))
299 event->group_flags |= PERF_GROUP_SOFTWARE;
301 list = ctx_group_list(event, ctx);
302 list_add_tail(&event->group_entry, list);
305 list_add_rcu(&event->event_entry, &ctx->event_list);
307 if (event->attr.inherit_stat)
311 static void perf_group_attach(struct perf_event *event)
313 struct perf_event *group_leader = event->group_leader;
315 WARN_ON_ONCE(event->attach_state & PERF_ATTACH_GROUP);
316 event->attach_state |= PERF_ATTACH_GROUP;
318 if (group_leader == event)
321 if (group_leader->group_flags & PERF_GROUP_SOFTWARE &&
322 !is_software_event(event))
323 group_leader->group_flags &= ~PERF_GROUP_SOFTWARE;
325 list_add_tail(&event->group_entry, &group_leader->sibling_list);
326 group_leader->nr_siblings++;
330 * Remove a event from the lists for its context.
331 * Must be called with ctx->mutex and ctx->lock held.
334 list_del_event(struct perf_event *event, struct perf_event_context *ctx)
337 * We can have double detach due to exit/hot-unplug + close.
339 if (!(event->attach_state & PERF_ATTACH_CONTEXT))
342 event->attach_state &= ~PERF_ATTACH_CONTEXT;
345 if (event->attr.inherit_stat)
348 list_del_rcu(&event->event_entry);
350 if (event->group_leader == event)
351 list_del_init(&event->group_entry);
353 update_group_times(event);
356 * If event was in error state, then keep it
357 * that way, otherwise bogus counts will be
358 * returned on read(). The only way to get out
359 * of error state is by explicit re-enabling
362 if (event->state > PERF_EVENT_STATE_OFF)
363 event->state = PERF_EVENT_STATE_OFF;
366 static void perf_group_detach(struct perf_event *event)
368 struct perf_event *sibling, *tmp;
369 struct list_head *list = NULL;
372 * We can have double detach due to exit/hot-unplug + close.
374 if (!(event->attach_state & PERF_ATTACH_GROUP))
377 event->attach_state &= ~PERF_ATTACH_GROUP;
380 * If this is a sibling, remove it from its group.
382 if (event->group_leader != event) {
383 list_del_init(&event->group_entry);
384 event->group_leader->nr_siblings--;
388 if (!list_empty(&event->group_entry))
389 list = &event->group_entry;
392 * If this was a group event with sibling events then
393 * upgrade the siblings to singleton events by adding them
394 * to whatever list we are on.
396 list_for_each_entry_safe(sibling, tmp, &event->sibling_list, group_entry) {
398 list_move_tail(&sibling->group_entry, list);
399 sibling->group_leader = sibling;
401 /* Inherit group flags from the previous leader */
402 sibling->group_flags = event->group_flags;
407 event_sched_out(struct perf_event *event,
408 struct perf_cpu_context *cpuctx,
409 struct perf_event_context *ctx)
411 if (event->state != PERF_EVENT_STATE_ACTIVE)
414 event->state = PERF_EVENT_STATE_INACTIVE;
415 if (event->pending_disable) {
416 event->pending_disable = 0;
417 event->state = PERF_EVENT_STATE_OFF;
419 event->tstamp_stopped = ctx->time;
420 event->pmu->disable(event);
423 if (!is_software_event(event))
424 cpuctx->active_oncpu--;
426 if (event->attr.exclusive || !cpuctx->active_oncpu)
427 cpuctx->exclusive = 0;
431 group_sched_out(struct perf_event *group_event,
432 struct perf_cpu_context *cpuctx,
433 struct perf_event_context *ctx)
435 struct perf_event *event;
437 if (group_event->state != PERF_EVENT_STATE_ACTIVE)
440 event_sched_out(group_event, cpuctx, ctx);
443 * Schedule out siblings (if any):
445 list_for_each_entry(event, &group_event->sibling_list, group_entry)
446 event_sched_out(event, cpuctx, ctx);
448 if (group_event->attr.exclusive)
449 cpuctx->exclusive = 0;
453 * Cross CPU call to remove a performance event
455 * We disable the event on the hardware level first. After that we
456 * remove it from the context list.
458 static void __perf_event_remove_from_context(void *info)
460 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
461 struct perf_event *event = info;
462 struct perf_event_context *ctx = event->ctx;
465 * If this is a task context, we need to check whether it is
466 * the current task context of this cpu. If not it has been
467 * scheduled out before the smp call arrived.
469 if (ctx->task && cpuctx->task_ctx != ctx)
472 raw_spin_lock(&ctx->lock);
474 * Protect the list operation against NMI by disabling the
475 * events on a global level.
479 event_sched_out(event, cpuctx, ctx);
481 list_del_event(event, ctx);
485 * Allow more per task events with respect to the
488 cpuctx->max_pertask =
489 min(perf_max_events - ctx->nr_events,
490 perf_max_events - perf_reserved_percpu);
494 raw_spin_unlock(&ctx->lock);
499 * Remove the event from a task's (or a CPU's) list of events.
501 * Must be called with ctx->mutex held.
503 * CPU events are removed with a smp call. For task events we only
504 * call when the task is on a CPU.
506 * If event->ctx is a cloned context, callers must make sure that
507 * every task struct that event->ctx->task could possibly point to
508 * remains valid. This is OK when called from perf_release since
509 * that only calls us on the top-level context, which can't be a clone.
510 * When called from perf_event_exit_task, it's OK because the
511 * context has been detached from its task.
513 static void perf_event_remove_from_context(struct perf_event *event)
515 struct perf_event_context *ctx = event->ctx;
516 struct task_struct *task = ctx->task;
520 * Per cpu events are removed via an smp call and
521 * the removal is always successful.
523 smp_call_function_single(event->cpu,
524 __perf_event_remove_from_context,
530 task_oncpu_function_call(task, __perf_event_remove_from_context,
533 raw_spin_lock_irq(&ctx->lock);
535 * If the context is active we need to retry the smp call.
537 if (ctx->nr_active && !list_empty(&event->group_entry)) {
538 raw_spin_unlock_irq(&ctx->lock);
543 * The lock prevents that this context is scheduled in so we
544 * can remove the event safely, if the call above did not
547 if (!list_empty(&event->group_entry))
548 list_del_event(event, ctx);
549 raw_spin_unlock_irq(&ctx->lock);
553 * Cross CPU call to disable a performance event
555 static void __perf_event_disable(void *info)
557 struct perf_event *event = info;
558 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
559 struct perf_event_context *ctx = event->ctx;
562 * If this is a per-task event, need to check whether this
563 * event's task is the current task on this cpu.
565 if (ctx->task && cpuctx->task_ctx != ctx)
568 raw_spin_lock(&ctx->lock);
571 * If the event is on, turn it off.
572 * If it is in error state, leave it in error state.
574 if (event->state >= PERF_EVENT_STATE_INACTIVE) {
575 update_context_time(ctx);
576 update_group_times(event);
577 if (event == event->group_leader)
578 group_sched_out(event, cpuctx, ctx);
580 event_sched_out(event, cpuctx, ctx);
581 event->state = PERF_EVENT_STATE_OFF;
584 raw_spin_unlock(&ctx->lock);
590 * If event->ctx is a cloned context, callers must make sure that
591 * every task struct that event->ctx->task could possibly point to
592 * remains valid. This condition is satisifed when called through
593 * perf_event_for_each_child or perf_event_for_each because they
594 * hold the top-level event's child_mutex, so any descendant that
595 * goes to exit will block in sync_child_event.
596 * When called from perf_pending_event it's OK because event->ctx
597 * is the current context on this CPU and preemption is disabled,
598 * hence we can't get into perf_event_task_sched_out for this context.
600 void perf_event_disable(struct perf_event *event)
602 struct perf_event_context *ctx = event->ctx;
603 struct task_struct *task = ctx->task;
607 * Disable the event on the cpu that it's on
609 smp_call_function_single(event->cpu, __perf_event_disable,
615 task_oncpu_function_call(task, __perf_event_disable, event);
617 raw_spin_lock_irq(&ctx->lock);
619 * If the event is still active, we need to retry the cross-call.
621 if (event->state == PERF_EVENT_STATE_ACTIVE) {
622 raw_spin_unlock_irq(&ctx->lock);
627 * Since we have the lock this context can't be scheduled
628 * in, so we can change the state safely.
630 if (event->state == PERF_EVENT_STATE_INACTIVE) {
631 update_group_times(event);
632 event->state = PERF_EVENT_STATE_OFF;
635 raw_spin_unlock_irq(&ctx->lock);
639 event_sched_in(struct perf_event *event,
640 struct perf_cpu_context *cpuctx,
641 struct perf_event_context *ctx)
643 if (event->state <= PERF_EVENT_STATE_OFF)
646 event->state = PERF_EVENT_STATE_ACTIVE;
647 event->oncpu = smp_processor_id();
649 * The new state must be visible before we turn it on in the hardware:
653 if (event->pmu->enable(event)) {
654 event->state = PERF_EVENT_STATE_INACTIVE;
659 event->tstamp_running += ctx->time - event->tstamp_stopped;
661 if (!is_software_event(event))
662 cpuctx->active_oncpu++;
665 if (event->attr.exclusive)
666 cpuctx->exclusive = 1;
672 group_sched_in(struct perf_event *group_event,
673 struct perf_cpu_context *cpuctx,
674 struct perf_event_context *ctx)
676 struct perf_event *event, *partial_group = NULL;
677 const struct pmu *pmu = group_event->pmu;
681 if (group_event->state == PERF_EVENT_STATE_OFF)
684 /* Check if group transaction availabe */
691 if (event_sched_in(group_event, cpuctx, ctx)) {
693 pmu->cancel_txn(pmu);
698 * Schedule in siblings as one group (if any):
700 list_for_each_entry(event, &group_event->sibling_list, group_entry) {
701 if (event_sched_in(event, cpuctx, ctx)) {
702 partial_group = event;
710 ret = pmu->commit_txn(pmu);
712 pmu->cancel_txn(pmu);
718 * Groups can be scheduled in as one unit only, so undo any
719 * partial group before returning:
721 list_for_each_entry(event, &group_event->sibling_list, group_entry) {
722 if (event == partial_group)
724 event_sched_out(event, cpuctx, ctx);
726 event_sched_out(group_event, cpuctx, ctx);
729 pmu->cancel_txn(pmu);
735 * Work out whether we can put this event group on the CPU now.
737 static int group_can_go_on(struct perf_event *event,
738 struct perf_cpu_context *cpuctx,
742 * Groups consisting entirely of software events can always go on.
744 if (event->group_flags & PERF_GROUP_SOFTWARE)
747 * If an exclusive group is already on, no other hardware
750 if (cpuctx->exclusive)
753 * If this group is exclusive and there are already
754 * events on the CPU, it can't go on.
756 if (event->attr.exclusive && cpuctx->active_oncpu)
759 * Otherwise, try to add it if all previous groups were able
765 static void add_event_to_ctx(struct perf_event *event,
766 struct perf_event_context *ctx)
768 list_add_event(event, ctx);
769 perf_group_attach(event);
770 event->tstamp_enabled = ctx->time;
771 event->tstamp_running = ctx->time;
772 event->tstamp_stopped = ctx->time;
776 * Cross CPU call to install and enable a performance event
778 * Must be called with ctx->mutex held
780 static void __perf_install_in_context(void *info)
782 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
783 struct perf_event *event = info;
784 struct perf_event_context *ctx = event->ctx;
785 struct perf_event *leader = event->group_leader;
789 * If this is a task context, we need to check whether it is
790 * the current task context of this cpu. If not it has been
791 * scheduled out before the smp call arrived.
792 * Or possibly this is the right context but it isn't
793 * on this cpu because it had no events.
795 if (ctx->task && cpuctx->task_ctx != ctx) {
796 if (cpuctx->task_ctx || ctx->task != current)
798 cpuctx->task_ctx = ctx;
801 raw_spin_lock(&ctx->lock);
803 update_context_time(ctx);
806 * Protect the list operation against NMI by disabling the
807 * events on a global level. NOP for non NMI based events.
811 add_event_to_ctx(event, ctx);
813 if (event->cpu != -1 && event->cpu != smp_processor_id())
817 * Don't put the event on if it is disabled or if
818 * it is in a group and the group isn't on.
820 if (event->state != PERF_EVENT_STATE_INACTIVE ||
821 (leader != event && leader->state != PERF_EVENT_STATE_ACTIVE))
825 * An exclusive event can't go on if there are already active
826 * hardware events, and no hardware event can go on if there
827 * is already an exclusive event on.
829 if (!group_can_go_on(event, cpuctx, 1))
832 err = event_sched_in(event, cpuctx, ctx);
836 * This event couldn't go on. If it is in a group
837 * then we have to pull the whole group off.
838 * If the event group is pinned then put it in error state.
841 group_sched_out(leader, cpuctx, ctx);
842 if (leader->attr.pinned) {
843 update_group_times(leader);
844 leader->state = PERF_EVENT_STATE_ERROR;
848 if (!err && !ctx->task && cpuctx->max_pertask)
849 cpuctx->max_pertask--;
854 raw_spin_unlock(&ctx->lock);
858 * Attach a performance event to a context
860 * First we add the event to the list with the hardware enable bit
861 * in event->hw_config cleared.
863 * If the event is attached to a task which is on a CPU we use a smp
864 * call to enable it in the task context. The task might have been
865 * scheduled away, but we check this in the smp call again.
867 * Must be called with ctx->mutex held.
870 perf_install_in_context(struct perf_event_context *ctx,
871 struct perf_event *event,
874 struct task_struct *task = ctx->task;
878 * Per cpu events are installed via an smp call and
879 * the install is always successful.
881 smp_call_function_single(cpu, __perf_install_in_context,
887 task_oncpu_function_call(task, __perf_install_in_context,
890 raw_spin_lock_irq(&ctx->lock);
892 * we need to retry the smp call.
894 if (ctx->is_active && list_empty(&event->group_entry)) {
895 raw_spin_unlock_irq(&ctx->lock);
900 * The lock prevents that this context is scheduled in so we
901 * can add the event safely, if it the call above did not
904 if (list_empty(&event->group_entry))
905 add_event_to_ctx(event, ctx);
906 raw_spin_unlock_irq(&ctx->lock);
910 * Put a event into inactive state and update time fields.
911 * Enabling the leader of a group effectively enables all
912 * the group members that aren't explicitly disabled, so we
913 * have to update their ->tstamp_enabled also.
914 * Note: this works for group members as well as group leaders
915 * since the non-leader members' sibling_lists will be empty.
917 static void __perf_event_mark_enabled(struct perf_event *event,
918 struct perf_event_context *ctx)
920 struct perf_event *sub;
922 event->state = PERF_EVENT_STATE_INACTIVE;
923 event->tstamp_enabled = ctx->time - event->total_time_enabled;
924 list_for_each_entry(sub, &event->sibling_list, group_entry)
925 if (sub->state >= PERF_EVENT_STATE_INACTIVE)
926 sub->tstamp_enabled =
927 ctx->time - sub->total_time_enabled;
931 * Cross CPU call to enable a performance event
933 static void __perf_event_enable(void *info)
935 struct perf_event *event = info;
936 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
937 struct perf_event_context *ctx = event->ctx;
938 struct perf_event *leader = event->group_leader;
942 * If this is a per-task event, need to check whether this
943 * event's task is the current task on this cpu.
945 if (ctx->task && cpuctx->task_ctx != ctx) {
946 if (cpuctx->task_ctx || ctx->task != current)
948 cpuctx->task_ctx = ctx;
951 raw_spin_lock(&ctx->lock);
953 update_context_time(ctx);
955 if (event->state >= PERF_EVENT_STATE_INACTIVE)
957 __perf_event_mark_enabled(event, ctx);
959 if (event->cpu != -1 && event->cpu != smp_processor_id())
963 * If the event is in a group and isn't the group leader,
964 * then don't put it on unless the group is on.
966 if (leader != event && leader->state != PERF_EVENT_STATE_ACTIVE)
969 if (!group_can_go_on(event, cpuctx, 1)) {
974 err = group_sched_in(event, cpuctx, ctx);
976 err = event_sched_in(event, cpuctx, ctx);
982 * If this event can't go on and it's part of a
983 * group, then the whole group has to come off.
986 group_sched_out(leader, cpuctx, ctx);
987 if (leader->attr.pinned) {
988 update_group_times(leader);
989 leader->state = PERF_EVENT_STATE_ERROR;
994 raw_spin_unlock(&ctx->lock);
1000 * If event->ctx is a cloned context, callers must make sure that
1001 * every task struct that event->ctx->task could possibly point to
1002 * remains valid. This condition is satisfied when called through
1003 * perf_event_for_each_child or perf_event_for_each as described
1004 * for perf_event_disable.
1006 void perf_event_enable(struct perf_event *event)
1008 struct perf_event_context *ctx = event->ctx;
1009 struct task_struct *task = ctx->task;
1013 * Enable the event on the cpu that it's on
1015 smp_call_function_single(event->cpu, __perf_event_enable,
1020 raw_spin_lock_irq(&ctx->lock);
1021 if (event->state >= PERF_EVENT_STATE_INACTIVE)
1025 * If the event is in error state, clear that first.
1026 * That way, if we see the event in error state below, we
1027 * know that it has gone back into error state, as distinct
1028 * from the task having been scheduled away before the
1029 * cross-call arrived.
1031 if (event->state == PERF_EVENT_STATE_ERROR)
1032 event->state = PERF_EVENT_STATE_OFF;
1035 raw_spin_unlock_irq(&ctx->lock);
1036 task_oncpu_function_call(task, __perf_event_enable, event);
1038 raw_spin_lock_irq(&ctx->lock);
1041 * If the context is active and the event is still off,
1042 * we need to retry the cross-call.
1044 if (ctx->is_active && event->state == PERF_EVENT_STATE_OFF)
1048 * Since we have the lock this context can't be scheduled
1049 * in, so we can change the state safely.
1051 if (event->state == PERF_EVENT_STATE_OFF)
1052 __perf_event_mark_enabled(event, ctx);
1055 raw_spin_unlock_irq(&ctx->lock);
1058 static int perf_event_refresh(struct perf_event *event, int refresh)
1061 * not supported on inherited events
1063 if (event->attr.inherit)
1066 atomic_add(refresh, &event->event_limit);
1067 perf_event_enable(event);
1073 EVENT_FLEXIBLE = 0x1,
1075 EVENT_ALL = EVENT_FLEXIBLE | EVENT_PINNED,
1078 static void ctx_sched_out(struct perf_event_context *ctx,
1079 struct perf_cpu_context *cpuctx,
1080 enum event_type_t event_type)
1082 struct perf_event *event;
1084 raw_spin_lock(&ctx->lock);
1086 if (likely(!ctx->nr_events))
1088 update_context_time(ctx);
1091 if (!ctx->nr_active)
1094 if (event_type & EVENT_PINNED)
1095 list_for_each_entry(event, &ctx->pinned_groups, group_entry)
1096 group_sched_out(event, cpuctx, ctx);
1098 if (event_type & EVENT_FLEXIBLE)
1099 list_for_each_entry(event, &ctx->flexible_groups, group_entry)
1100 group_sched_out(event, cpuctx, ctx);
1105 raw_spin_unlock(&ctx->lock);
1109 * Test whether two contexts are equivalent, i.e. whether they
1110 * have both been cloned from the same version of the same context
1111 * and they both have the same number of enabled events.
1112 * If the number of enabled events is the same, then the set
1113 * of enabled events should be the same, because these are both
1114 * inherited contexts, therefore we can't access individual events
1115 * in them directly with an fd; we can only enable/disable all
1116 * events via prctl, or enable/disable all events in a family
1117 * via ioctl, which will have the same effect on both contexts.
1119 static int context_equiv(struct perf_event_context *ctx1,
1120 struct perf_event_context *ctx2)
1122 return ctx1->parent_ctx && ctx1->parent_ctx == ctx2->parent_ctx
1123 && ctx1->parent_gen == ctx2->parent_gen
1124 && !ctx1->pin_count && !ctx2->pin_count;
1127 static void __perf_event_sync_stat(struct perf_event *event,
1128 struct perf_event *next_event)
1132 if (!event->attr.inherit_stat)
1136 * Update the event value, we cannot use perf_event_read()
1137 * because we're in the middle of a context switch and have IRQs
1138 * disabled, which upsets smp_call_function_single(), however
1139 * we know the event must be on the current CPU, therefore we
1140 * don't need to use it.
1142 switch (event->state) {
1143 case PERF_EVENT_STATE_ACTIVE:
1144 event->pmu->read(event);
1147 case PERF_EVENT_STATE_INACTIVE:
1148 update_event_times(event);
1156 * In order to keep per-task stats reliable we need to flip the event
1157 * values when we flip the contexts.
1159 value = atomic64_read(&next_event->count);
1160 value = atomic64_xchg(&event->count, value);
1161 atomic64_set(&next_event->count, value);
1163 swap(event->total_time_enabled, next_event->total_time_enabled);
1164 swap(event->total_time_running, next_event->total_time_running);
1167 * Since we swizzled the values, update the user visible data too.
1169 perf_event_update_userpage(event);
1170 perf_event_update_userpage(next_event);
1173 #define list_next_entry(pos, member) \
1174 list_entry(pos->member.next, typeof(*pos), member)
1176 static void perf_event_sync_stat(struct perf_event_context *ctx,
1177 struct perf_event_context *next_ctx)
1179 struct perf_event *event, *next_event;
1184 update_context_time(ctx);
1186 event = list_first_entry(&ctx->event_list,
1187 struct perf_event, event_entry);
1189 next_event = list_first_entry(&next_ctx->event_list,
1190 struct perf_event, event_entry);
1192 while (&event->event_entry != &ctx->event_list &&
1193 &next_event->event_entry != &next_ctx->event_list) {
1195 __perf_event_sync_stat(event, next_event);
1197 event = list_next_entry(event, event_entry);
1198 next_event = list_next_entry(next_event, event_entry);
1203 * Called from scheduler to remove the events of the current task,
1204 * with interrupts disabled.
1206 * We stop each event and update the event value in event->count.
1208 * This does not protect us against NMI, but disable()
1209 * sets the disabled bit in the control field of event _before_
1210 * accessing the event control register. If a NMI hits, then it will
1211 * not restart the event.
1213 void perf_event_task_sched_out(struct task_struct *task,
1214 struct task_struct *next)
1216 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
1217 struct perf_event_context *ctx = task->perf_event_ctxp;
1218 struct perf_event_context *next_ctx;
1219 struct perf_event_context *parent;
1222 perf_sw_event(PERF_COUNT_SW_CONTEXT_SWITCHES, 1, 1, NULL, 0);
1224 if (likely(!ctx || !cpuctx->task_ctx))
1228 parent = rcu_dereference(ctx->parent_ctx);
1229 next_ctx = next->perf_event_ctxp;
1230 if (parent && next_ctx &&
1231 rcu_dereference(next_ctx->parent_ctx) == parent) {
1233 * Looks like the two contexts are clones, so we might be
1234 * able to optimize the context switch. We lock both
1235 * contexts and check that they are clones under the
1236 * lock (including re-checking that neither has been
1237 * uncloned in the meantime). It doesn't matter which
1238 * order we take the locks because no other cpu could
1239 * be trying to lock both of these tasks.
1241 raw_spin_lock(&ctx->lock);
1242 raw_spin_lock_nested(&next_ctx->lock, SINGLE_DEPTH_NESTING);
1243 if (context_equiv(ctx, next_ctx)) {
1245 * XXX do we need a memory barrier of sorts
1246 * wrt to rcu_dereference() of perf_event_ctxp
1248 task->perf_event_ctxp = next_ctx;
1249 next->perf_event_ctxp = ctx;
1251 next_ctx->task = task;
1254 perf_event_sync_stat(ctx, next_ctx);
1256 raw_spin_unlock(&next_ctx->lock);
1257 raw_spin_unlock(&ctx->lock);
1262 ctx_sched_out(ctx, cpuctx, EVENT_ALL);
1263 cpuctx->task_ctx = NULL;
1267 static void task_ctx_sched_out(struct perf_event_context *ctx,
1268 enum event_type_t event_type)
1270 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
1272 if (!cpuctx->task_ctx)
1275 if (WARN_ON_ONCE(ctx != cpuctx->task_ctx))
1278 ctx_sched_out(ctx, cpuctx, event_type);
1279 cpuctx->task_ctx = NULL;
1283 * Called with IRQs disabled
1285 static void __perf_event_task_sched_out(struct perf_event_context *ctx)
1287 task_ctx_sched_out(ctx, EVENT_ALL);
1291 * Called with IRQs disabled
1293 static void cpu_ctx_sched_out(struct perf_cpu_context *cpuctx,
1294 enum event_type_t event_type)
1296 ctx_sched_out(&cpuctx->ctx, cpuctx, event_type);
1300 ctx_pinned_sched_in(struct perf_event_context *ctx,
1301 struct perf_cpu_context *cpuctx)
1303 struct perf_event *event;
1305 list_for_each_entry(event, &ctx->pinned_groups, group_entry) {
1306 if (event->state <= PERF_EVENT_STATE_OFF)
1308 if (event->cpu != -1 && event->cpu != smp_processor_id())
1311 if (group_can_go_on(event, cpuctx, 1))
1312 group_sched_in(event, cpuctx, ctx);
1315 * If this pinned group hasn't been scheduled,
1316 * put it in error state.
1318 if (event->state == PERF_EVENT_STATE_INACTIVE) {
1319 update_group_times(event);
1320 event->state = PERF_EVENT_STATE_ERROR;
1326 ctx_flexible_sched_in(struct perf_event_context *ctx,
1327 struct perf_cpu_context *cpuctx)
1329 struct perf_event *event;
1332 list_for_each_entry(event, &ctx->flexible_groups, group_entry) {
1333 /* Ignore events in OFF or ERROR state */
1334 if (event->state <= PERF_EVENT_STATE_OFF)
1337 * Listen to the 'cpu' scheduling filter constraint
1340 if (event->cpu != -1 && event->cpu != smp_processor_id())
1343 if (group_can_go_on(event, cpuctx, can_add_hw))
1344 if (group_sched_in(event, cpuctx, ctx))
1350 ctx_sched_in(struct perf_event_context *ctx,
1351 struct perf_cpu_context *cpuctx,
1352 enum event_type_t event_type)
1354 raw_spin_lock(&ctx->lock);
1356 if (likely(!ctx->nr_events))
1359 ctx->timestamp = perf_clock();
1364 * First go through the list and put on any pinned groups
1365 * in order to give them the best chance of going on.
1367 if (event_type & EVENT_PINNED)
1368 ctx_pinned_sched_in(ctx, cpuctx);
1370 /* Then walk through the lower prio flexible groups */
1371 if (event_type & EVENT_FLEXIBLE)
1372 ctx_flexible_sched_in(ctx, cpuctx);
1376 raw_spin_unlock(&ctx->lock);
1379 static void cpu_ctx_sched_in(struct perf_cpu_context *cpuctx,
1380 enum event_type_t event_type)
1382 struct perf_event_context *ctx = &cpuctx->ctx;
1384 ctx_sched_in(ctx, cpuctx, event_type);
1387 static void task_ctx_sched_in(struct task_struct *task,
1388 enum event_type_t event_type)
1390 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
1391 struct perf_event_context *ctx = task->perf_event_ctxp;
1395 if (cpuctx->task_ctx == ctx)
1397 ctx_sched_in(ctx, cpuctx, event_type);
1398 cpuctx->task_ctx = ctx;
1401 * Called from scheduler to add the events of the current task
1402 * with interrupts disabled.
1404 * We restore the event value and then enable it.
1406 * This does not protect us against NMI, but enable()
1407 * sets the enabled bit in the control field of event _before_
1408 * accessing the event control register. If a NMI hits, then it will
1409 * keep the event running.
1411 void perf_event_task_sched_in(struct task_struct *task)
1413 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
1414 struct perf_event_context *ctx = task->perf_event_ctxp;
1419 if (cpuctx->task_ctx == ctx)
1425 * We want to keep the following priority order:
1426 * cpu pinned (that don't need to move), task pinned,
1427 * cpu flexible, task flexible.
1429 cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE);
1431 ctx_sched_in(ctx, cpuctx, EVENT_PINNED);
1432 cpu_ctx_sched_in(cpuctx, EVENT_FLEXIBLE);
1433 ctx_sched_in(ctx, cpuctx, EVENT_FLEXIBLE);
1435 cpuctx->task_ctx = ctx;
1440 #define MAX_INTERRUPTS (~0ULL)
1442 static void perf_log_throttle(struct perf_event *event, int enable);
1444 static u64 perf_calculate_period(struct perf_event *event, u64 nsec, u64 count)
1446 u64 frequency = event->attr.sample_freq;
1447 u64 sec = NSEC_PER_SEC;
1448 u64 divisor, dividend;
1450 int count_fls, nsec_fls, frequency_fls, sec_fls;
1452 count_fls = fls64(count);
1453 nsec_fls = fls64(nsec);
1454 frequency_fls = fls64(frequency);
1458 * We got @count in @nsec, with a target of sample_freq HZ
1459 * the target period becomes:
1462 * period = -------------------
1463 * @nsec * sample_freq
1468 * Reduce accuracy by one bit such that @a and @b converge
1469 * to a similar magnitude.
1471 #define REDUCE_FLS(a, b) \
1473 if (a##_fls > b##_fls) { \
1483 * Reduce accuracy until either term fits in a u64, then proceed with
1484 * the other, so that finally we can do a u64/u64 division.
1486 while (count_fls + sec_fls > 64 && nsec_fls + frequency_fls > 64) {
1487 REDUCE_FLS(nsec, frequency);
1488 REDUCE_FLS(sec, count);
1491 if (count_fls + sec_fls > 64) {
1492 divisor = nsec * frequency;
1494 while (count_fls + sec_fls > 64) {
1495 REDUCE_FLS(count, sec);
1499 dividend = count * sec;
1501 dividend = count * sec;
1503 while (nsec_fls + frequency_fls > 64) {
1504 REDUCE_FLS(nsec, frequency);
1508 divisor = nsec * frequency;
1514 return div64_u64(dividend, divisor);
1517 static void perf_event_stop(struct perf_event *event)
1519 if (!event->pmu->stop)
1520 return event->pmu->disable(event);
1522 return event->pmu->stop(event);
1525 static int perf_event_start(struct perf_event *event)
1527 if (!event->pmu->start)
1528 return event->pmu->enable(event);
1530 return event->pmu->start(event);
1533 static void perf_adjust_period(struct perf_event *event, u64 nsec, u64 count)
1535 struct hw_perf_event *hwc = &event->hw;
1536 s64 period, sample_period;
1539 period = perf_calculate_period(event, nsec, count);
1541 delta = (s64)(period - hwc->sample_period);
1542 delta = (delta + 7) / 8; /* low pass filter */
1544 sample_period = hwc->sample_period + delta;
1549 hwc->sample_period = sample_period;
1551 if (atomic64_read(&hwc->period_left) > 8*sample_period) {
1553 perf_event_stop(event);
1554 atomic64_set(&hwc->period_left, 0);
1555 perf_event_start(event);
1560 static void perf_ctx_adjust_freq(struct perf_event_context *ctx)
1562 struct perf_event *event;
1563 struct hw_perf_event *hwc;
1564 u64 interrupts, now;
1567 raw_spin_lock(&ctx->lock);
1568 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
1569 if (event->state != PERF_EVENT_STATE_ACTIVE)
1572 if (event->cpu != -1 && event->cpu != smp_processor_id())
1577 interrupts = hwc->interrupts;
1578 hwc->interrupts = 0;
1581 * unthrottle events on the tick
1583 if (interrupts == MAX_INTERRUPTS) {
1584 perf_log_throttle(event, 1);
1586 event->pmu->unthrottle(event);
1590 if (!event->attr.freq || !event->attr.sample_freq)
1594 event->pmu->read(event);
1595 now = atomic64_read(&event->count);
1596 delta = now - hwc->freq_count_stamp;
1597 hwc->freq_count_stamp = now;
1600 perf_adjust_period(event, TICK_NSEC, delta);
1603 raw_spin_unlock(&ctx->lock);
1607 * Round-robin a context's events:
1609 static void rotate_ctx(struct perf_event_context *ctx)
1611 raw_spin_lock(&ctx->lock);
1614 * Rotate the first entry last of non-pinned groups. Rotation might be
1615 * disabled by the inheritance code.
1617 if (!ctx->rotate_disable)
1618 list_rotate_left(&ctx->flexible_groups);
1620 raw_spin_unlock(&ctx->lock);
1623 void perf_event_task_tick(struct task_struct *curr)
1625 struct perf_cpu_context *cpuctx;
1626 struct perf_event_context *ctx;
1629 if (!atomic_read(&nr_events))
1632 cpuctx = &__get_cpu_var(perf_cpu_context);
1633 if (cpuctx->ctx.nr_events &&
1634 cpuctx->ctx.nr_events != cpuctx->ctx.nr_active)
1637 ctx = curr->perf_event_ctxp;
1638 if (ctx && ctx->nr_events && ctx->nr_events != ctx->nr_active)
1641 perf_ctx_adjust_freq(&cpuctx->ctx);
1643 perf_ctx_adjust_freq(ctx);
1649 cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE);
1651 task_ctx_sched_out(ctx, EVENT_FLEXIBLE);
1653 rotate_ctx(&cpuctx->ctx);
1657 cpu_ctx_sched_in(cpuctx, EVENT_FLEXIBLE);
1659 task_ctx_sched_in(curr, EVENT_FLEXIBLE);
1663 static int event_enable_on_exec(struct perf_event *event,
1664 struct perf_event_context *ctx)
1666 if (!event->attr.enable_on_exec)
1669 event->attr.enable_on_exec = 0;
1670 if (event->state >= PERF_EVENT_STATE_INACTIVE)
1673 __perf_event_mark_enabled(event, ctx);
1679 * Enable all of a task's events that have been marked enable-on-exec.
1680 * This expects task == current.
1682 static void perf_event_enable_on_exec(struct task_struct *task)
1684 struct perf_event_context *ctx;
1685 struct perf_event *event;
1686 unsigned long flags;
1690 local_irq_save(flags);
1691 ctx = task->perf_event_ctxp;
1692 if (!ctx || !ctx->nr_events)
1695 __perf_event_task_sched_out(ctx);
1697 raw_spin_lock(&ctx->lock);
1699 list_for_each_entry(event, &ctx->pinned_groups, group_entry) {
1700 ret = event_enable_on_exec(event, ctx);
1705 list_for_each_entry(event, &ctx->flexible_groups, group_entry) {
1706 ret = event_enable_on_exec(event, ctx);
1712 * Unclone this context if we enabled any event.
1717 raw_spin_unlock(&ctx->lock);
1719 perf_event_task_sched_in(task);
1721 local_irq_restore(flags);
1725 * Cross CPU call to read the hardware event
1727 static void __perf_event_read(void *info)
1729 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
1730 struct perf_event *event = info;
1731 struct perf_event_context *ctx = event->ctx;
1734 * If this is a task context, we need to check whether it is
1735 * the current task context of this cpu. If not it has been
1736 * scheduled out before the smp call arrived. In that case
1737 * event->count would have been updated to a recent sample
1738 * when the event was scheduled out.
1740 if (ctx->task && cpuctx->task_ctx != ctx)
1743 raw_spin_lock(&ctx->lock);
1744 update_context_time(ctx);
1745 update_event_times(event);
1746 raw_spin_unlock(&ctx->lock);
1748 event->pmu->read(event);
1751 static u64 perf_event_read(struct perf_event *event)
1754 * If event is enabled and currently active on a CPU, update the
1755 * value in the event structure:
1757 if (event->state == PERF_EVENT_STATE_ACTIVE) {
1758 smp_call_function_single(event->oncpu,
1759 __perf_event_read, event, 1);
1760 } else if (event->state == PERF_EVENT_STATE_INACTIVE) {
1761 struct perf_event_context *ctx = event->ctx;
1762 unsigned long flags;
1764 raw_spin_lock_irqsave(&ctx->lock, flags);
1766 * may read while context is not active
1767 * (e.g., thread is blocked), in that case
1768 * we cannot update context time
1771 update_context_time(ctx);
1772 update_event_times(event);
1773 raw_spin_unlock_irqrestore(&ctx->lock, flags);
1776 return atomic64_read(&event->count);
1780 * Initialize the perf_event context in a task_struct:
1783 __perf_event_init_context(struct perf_event_context *ctx,
1784 struct task_struct *task)
1786 raw_spin_lock_init(&ctx->lock);
1787 mutex_init(&ctx->mutex);
1788 INIT_LIST_HEAD(&ctx->pinned_groups);
1789 INIT_LIST_HEAD(&ctx->flexible_groups);
1790 INIT_LIST_HEAD(&ctx->event_list);
1791 atomic_set(&ctx->refcount, 1);
1795 static struct perf_event_context *find_get_context(pid_t pid, int cpu)
1797 struct perf_event_context *ctx;
1798 struct perf_cpu_context *cpuctx;
1799 struct task_struct *task;
1800 unsigned long flags;
1803 if (pid == -1 && cpu != -1) {
1804 /* Must be root to operate on a CPU event: */
1805 if (perf_paranoid_cpu() && !capable(CAP_SYS_ADMIN))
1806 return ERR_PTR(-EACCES);
1808 if (cpu < 0 || cpu >= nr_cpumask_bits)
1809 return ERR_PTR(-EINVAL);
1812 * We could be clever and allow to attach a event to an
1813 * offline CPU and activate it when the CPU comes up, but
1816 if (!cpu_online(cpu))
1817 return ERR_PTR(-ENODEV);
1819 cpuctx = &per_cpu(perf_cpu_context, cpu);
1830 task = find_task_by_vpid(pid);
1832 get_task_struct(task);
1836 return ERR_PTR(-ESRCH);
1839 * Can't attach events to a dying task.
1842 if (task->flags & PF_EXITING)
1845 /* Reuse ptrace permission checks for now. */
1847 if (!ptrace_may_access(task, PTRACE_MODE_READ))
1851 ctx = perf_lock_task_context(task, &flags);
1854 raw_spin_unlock_irqrestore(&ctx->lock, flags);
1858 ctx = kzalloc(sizeof(struct perf_event_context), GFP_KERNEL);
1862 __perf_event_init_context(ctx, task);
1864 if (cmpxchg(&task->perf_event_ctxp, NULL, ctx)) {
1866 * We raced with some other task; use
1867 * the context they set.
1872 get_task_struct(task);
1875 put_task_struct(task);
1879 put_task_struct(task);
1880 return ERR_PTR(err);
1883 static void perf_event_free_filter(struct perf_event *event);
1885 static void free_event_rcu(struct rcu_head *head)
1887 struct perf_event *event;
1889 event = container_of(head, struct perf_event, rcu_head);
1891 put_pid_ns(event->ns);
1892 perf_event_free_filter(event);
1896 static void perf_pending_sync(struct perf_event *event);
1897 static void perf_mmap_data_put(struct perf_mmap_data *data);
1899 static void free_event(struct perf_event *event)
1901 perf_pending_sync(event);
1903 if (!event->parent) {
1904 atomic_dec(&nr_events);
1905 if (event->attr.mmap)
1906 atomic_dec(&nr_mmap_events);
1907 if (event->attr.comm)
1908 atomic_dec(&nr_comm_events);
1909 if (event->attr.task)
1910 atomic_dec(&nr_task_events);
1914 perf_mmap_data_put(event->data);
1919 event->destroy(event);
1921 put_ctx(event->ctx);
1922 call_rcu(&event->rcu_head, free_event_rcu);
1925 int perf_event_release_kernel(struct perf_event *event)
1927 struct perf_event_context *ctx = event->ctx;
1930 * Remove from the PMU, can't get re-enabled since we got
1931 * here because the last ref went.
1933 perf_event_disable(event);
1935 WARN_ON_ONCE(ctx->parent_ctx);
1937 * There are two ways this annotation is useful:
1939 * 1) there is a lock recursion from perf_event_exit_task
1940 * see the comment there.
1942 * 2) there is a lock-inversion with mmap_sem through
1943 * perf_event_read_group(), which takes faults while
1944 * holding ctx->mutex, however this is called after
1945 * the last filedesc died, so there is no possibility
1946 * to trigger the AB-BA case.
1948 mutex_lock_nested(&ctx->mutex, SINGLE_DEPTH_NESTING);
1949 raw_spin_lock_irq(&ctx->lock);
1950 perf_group_detach(event);
1951 list_del_event(event, ctx);
1952 raw_spin_unlock_irq(&ctx->lock);
1953 mutex_unlock(&ctx->mutex);
1955 mutex_lock(&event->owner->perf_event_mutex);
1956 list_del_init(&event->owner_entry);
1957 mutex_unlock(&event->owner->perf_event_mutex);
1958 put_task_struct(event->owner);
1964 EXPORT_SYMBOL_GPL(perf_event_release_kernel);
1967 * Called when the last reference to the file is gone.
1969 static int perf_release(struct inode *inode, struct file *file)
1971 struct perf_event *event = file->private_data;
1973 file->private_data = NULL;
1975 return perf_event_release_kernel(event);
1978 static int perf_event_read_size(struct perf_event *event)
1980 int entry = sizeof(u64); /* value */
1984 if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1985 size += sizeof(u64);
1987 if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1988 size += sizeof(u64);
1990 if (event->attr.read_format & PERF_FORMAT_ID)
1991 entry += sizeof(u64);
1993 if (event->attr.read_format & PERF_FORMAT_GROUP) {
1994 nr += event->group_leader->nr_siblings;
1995 size += sizeof(u64);
2003 u64 perf_event_read_value(struct perf_event *event, u64 *enabled, u64 *running)
2005 struct perf_event *child;
2011 mutex_lock(&event->child_mutex);
2012 total += perf_event_read(event);
2013 *enabled += event->total_time_enabled +
2014 atomic64_read(&event->child_total_time_enabled);
2015 *running += event->total_time_running +
2016 atomic64_read(&event->child_total_time_running);
2018 list_for_each_entry(child, &event->child_list, child_list) {
2019 total += perf_event_read(child);
2020 *enabled += child->total_time_enabled;
2021 *running += child->total_time_running;
2023 mutex_unlock(&event->child_mutex);
2027 EXPORT_SYMBOL_GPL(perf_event_read_value);
2029 static int perf_event_read_group(struct perf_event *event,
2030 u64 read_format, char __user *buf)
2032 struct perf_event *leader = event->group_leader, *sub;
2033 int n = 0, size = 0, ret = -EFAULT;
2034 struct perf_event_context *ctx = leader->ctx;
2036 u64 count, enabled, running;
2038 mutex_lock(&ctx->mutex);
2039 count = perf_event_read_value(leader, &enabled, &running);
2041 values[n++] = 1 + leader->nr_siblings;
2042 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
2043 values[n++] = enabled;
2044 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
2045 values[n++] = running;
2046 values[n++] = count;
2047 if (read_format & PERF_FORMAT_ID)
2048 values[n++] = primary_event_id(leader);
2050 size = n * sizeof(u64);
2052 if (copy_to_user(buf, values, size))
2057 list_for_each_entry(sub, &leader->sibling_list, group_entry) {
2060 values[n++] = perf_event_read_value(sub, &enabled, &running);
2061 if (read_format & PERF_FORMAT_ID)
2062 values[n++] = primary_event_id(sub);
2064 size = n * sizeof(u64);
2066 if (copy_to_user(buf + ret, values, size)) {
2074 mutex_unlock(&ctx->mutex);
2079 static int perf_event_read_one(struct perf_event *event,
2080 u64 read_format, char __user *buf)
2082 u64 enabled, running;
2086 values[n++] = perf_event_read_value(event, &enabled, &running);
2087 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
2088 values[n++] = enabled;
2089 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
2090 values[n++] = running;
2091 if (read_format & PERF_FORMAT_ID)
2092 values[n++] = primary_event_id(event);
2094 if (copy_to_user(buf, values, n * sizeof(u64)))
2097 return n * sizeof(u64);
2101 * Read the performance event - simple non blocking version for now
2104 perf_read_hw(struct perf_event *event, char __user *buf, size_t count)
2106 u64 read_format = event->attr.read_format;
2110 * Return end-of-file for a read on a event that is in
2111 * error state (i.e. because it was pinned but it couldn't be
2112 * scheduled on to the CPU at some point).
2114 if (event->state == PERF_EVENT_STATE_ERROR)
2117 if (count < perf_event_read_size(event))
2120 WARN_ON_ONCE(event->ctx->parent_ctx);
2121 if (read_format & PERF_FORMAT_GROUP)
2122 ret = perf_event_read_group(event, read_format, buf);
2124 ret = perf_event_read_one(event, read_format, buf);
2130 perf_read(struct file *file, char __user *buf, size_t count, loff_t *ppos)
2132 struct perf_event *event = file->private_data;
2134 return perf_read_hw(event, buf, count);
2137 static unsigned int perf_poll(struct file *file, poll_table *wait)
2139 struct perf_event *event = file->private_data;
2140 struct perf_mmap_data *data;
2141 unsigned int events = POLL_HUP;
2144 data = rcu_dereference(event->data);
2146 events = atomic_xchg(&data->poll, 0);
2149 poll_wait(file, &event->waitq, wait);
2154 static void perf_event_reset(struct perf_event *event)
2156 (void)perf_event_read(event);
2157 atomic64_set(&event->count, 0);
2158 perf_event_update_userpage(event);
2162 * Holding the top-level event's child_mutex means that any
2163 * descendant process that has inherited this event will block
2164 * in sync_child_event if it goes to exit, thus satisfying the
2165 * task existence requirements of perf_event_enable/disable.
2167 static void perf_event_for_each_child(struct perf_event *event,
2168 void (*func)(struct perf_event *))
2170 struct perf_event *child;
2172 WARN_ON_ONCE(event->ctx->parent_ctx);
2173 mutex_lock(&event->child_mutex);
2175 list_for_each_entry(child, &event->child_list, child_list)
2177 mutex_unlock(&event->child_mutex);
2180 static void perf_event_for_each(struct perf_event *event,
2181 void (*func)(struct perf_event *))
2183 struct perf_event_context *ctx = event->ctx;
2184 struct perf_event *sibling;
2186 WARN_ON_ONCE(ctx->parent_ctx);
2187 mutex_lock(&ctx->mutex);
2188 event = event->group_leader;
2190 perf_event_for_each_child(event, func);
2192 list_for_each_entry(sibling, &event->sibling_list, group_entry)
2193 perf_event_for_each_child(event, func);
2194 mutex_unlock(&ctx->mutex);
2197 static int perf_event_period(struct perf_event *event, u64 __user *arg)
2199 struct perf_event_context *ctx = event->ctx;
2204 if (!event->attr.sample_period)
2207 size = copy_from_user(&value, arg, sizeof(value));
2208 if (size != sizeof(value))
2214 raw_spin_lock_irq(&ctx->lock);
2215 if (event->attr.freq) {
2216 if (value > sysctl_perf_event_sample_rate) {
2221 event->attr.sample_freq = value;
2223 event->attr.sample_period = value;
2224 event->hw.sample_period = value;
2227 raw_spin_unlock_irq(&ctx->lock);
2232 static const struct file_operations perf_fops;
2234 static struct perf_event *perf_fget_light(int fd, int *fput_needed)
2238 file = fget_light(fd, fput_needed);
2240 return ERR_PTR(-EBADF);
2242 if (file->f_op != &perf_fops) {
2243 fput_light(file, *fput_needed);
2245 return ERR_PTR(-EBADF);
2248 return file->private_data;
2251 static int perf_event_set_output(struct perf_event *event,
2252 struct perf_event *output_event);
2253 static int perf_event_set_filter(struct perf_event *event, void __user *arg);
2255 static long perf_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
2257 struct perf_event *event = file->private_data;
2258 void (*func)(struct perf_event *);
2262 case PERF_EVENT_IOC_ENABLE:
2263 func = perf_event_enable;
2265 case PERF_EVENT_IOC_DISABLE:
2266 func = perf_event_disable;
2268 case PERF_EVENT_IOC_RESET:
2269 func = perf_event_reset;
2272 case PERF_EVENT_IOC_REFRESH:
2273 return perf_event_refresh(event, arg);
2275 case PERF_EVENT_IOC_PERIOD:
2276 return perf_event_period(event, (u64 __user *)arg);
2278 case PERF_EVENT_IOC_SET_OUTPUT:
2280 struct perf_event *output_event = NULL;
2281 int fput_needed = 0;
2285 output_event = perf_fget_light(arg, &fput_needed);
2286 if (IS_ERR(output_event))
2287 return PTR_ERR(output_event);
2290 ret = perf_event_set_output(event, output_event);
2292 fput_light(output_event->filp, fput_needed);
2297 case PERF_EVENT_IOC_SET_FILTER:
2298 return perf_event_set_filter(event, (void __user *)arg);
2304 if (flags & PERF_IOC_FLAG_GROUP)
2305 perf_event_for_each(event, func);
2307 perf_event_for_each_child(event, func);
2312 int perf_event_task_enable(void)
2314 struct perf_event *event;
2316 mutex_lock(¤t->perf_event_mutex);
2317 list_for_each_entry(event, ¤t->perf_event_list, owner_entry)
2318 perf_event_for_each_child(event, perf_event_enable);
2319 mutex_unlock(¤t->perf_event_mutex);
2324 int perf_event_task_disable(void)
2326 struct perf_event *event;
2328 mutex_lock(¤t->perf_event_mutex);
2329 list_for_each_entry(event, ¤t->perf_event_list, owner_entry)
2330 perf_event_for_each_child(event, perf_event_disable);
2331 mutex_unlock(¤t->perf_event_mutex);
2336 #ifndef PERF_EVENT_INDEX_OFFSET
2337 # define PERF_EVENT_INDEX_OFFSET 0
2340 static int perf_event_index(struct perf_event *event)
2342 if (event->state != PERF_EVENT_STATE_ACTIVE)
2345 return event->hw.idx + 1 - PERF_EVENT_INDEX_OFFSET;
2349 * Callers need to ensure there can be no nesting of this function, otherwise
2350 * the seqlock logic goes bad. We can not serialize this because the arch
2351 * code calls this from NMI context.
2353 void perf_event_update_userpage(struct perf_event *event)
2355 struct perf_event_mmap_page *userpg;
2356 struct perf_mmap_data *data;
2359 data = rcu_dereference(event->data);
2363 userpg = data->user_page;
2366 * Disable preemption so as to not let the corresponding user-space
2367 * spin too long if we get preempted.
2372 userpg->index = perf_event_index(event);
2373 userpg->offset = atomic64_read(&event->count);
2374 if (event->state == PERF_EVENT_STATE_ACTIVE)
2375 userpg->offset -= atomic64_read(&event->hw.prev_count);
2377 userpg->time_enabled = event->total_time_enabled +
2378 atomic64_read(&event->child_total_time_enabled);
2380 userpg->time_running = event->total_time_running +
2381 atomic64_read(&event->child_total_time_running);
2390 #ifndef CONFIG_PERF_USE_VMALLOC
2393 * Back perf_mmap() with regular GFP_KERNEL-0 pages.
2396 static struct page *
2397 perf_mmap_to_page(struct perf_mmap_data *data, unsigned long pgoff)
2399 if (pgoff > data->nr_pages)
2403 return virt_to_page(data->user_page);
2405 return virt_to_page(data->data_pages[pgoff - 1]);
2408 static void *perf_mmap_alloc_page(int cpu)
2413 node = (cpu == -1) ? cpu : cpu_to_node(cpu);
2414 page = alloc_pages_node(node, GFP_KERNEL | __GFP_ZERO, 0);
2418 return page_address(page);
2421 static struct perf_mmap_data *
2422 perf_mmap_data_alloc(struct perf_event *event, int nr_pages)
2424 struct perf_mmap_data *data;
2428 size = sizeof(struct perf_mmap_data);
2429 size += nr_pages * sizeof(void *);
2431 data = kzalloc(size, GFP_KERNEL);
2435 data->user_page = perf_mmap_alloc_page(event->cpu);
2436 if (!data->user_page)
2437 goto fail_user_page;
2439 for (i = 0; i < nr_pages; i++) {
2440 data->data_pages[i] = perf_mmap_alloc_page(event->cpu);
2441 if (!data->data_pages[i])
2442 goto fail_data_pages;
2445 data->nr_pages = nr_pages;
2450 for (i--; i >= 0; i--)
2451 free_page((unsigned long)data->data_pages[i]);
2453 free_page((unsigned long)data->user_page);
2462 static void perf_mmap_free_page(unsigned long addr)
2464 struct page *page = virt_to_page((void *)addr);
2466 page->mapping = NULL;
2470 static void perf_mmap_data_free(struct perf_mmap_data *data)
2474 perf_mmap_free_page((unsigned long)data->user_page);
2475 for (i = 0; i < data->nr_pages; i++)
2476 perf_mmap_free_page((unsigned long)data->data_pages[i]);
2480 static inline int page_order(struct perf_mmap_data *data)
2488 * Back perf_mmap() with vmalloc memory.
2490 * Required for architectures that have d-cache aliasing issues.
2493 static inline int page_order(struct perf_mmap_data *data)
2495 return data->page_order;
2498 static struct page *
2499 perf_mmap_to_page(struct perf_mmap_data *data, unsigned long pgoff)
2501 if (pgoff > (1UL << page_order(data)))
2504 return vmalloc_to_page((void *)data->user_page + pgoff * PAGE_SIZE);
2507 static void perf_mmap_unmark_page(void *addr)
2509 struct page *page = vmalloc_to_page(addr);
2511 page->mapping = NULL;
2514 static void perf_mmap_data_free_work(struct work_struct *work)
2516 struct perf_mmap_data *data;
2520 data = container_of(work, struct perf_mmap_data, work);
2521 nr = 1 << page_order(data);
2523 base = data->user_page;
2524 for (i = 0; i < nr + 1; i++)
2525 perf_mmap_unmark_page(base + (i * PAGE_SIZE));
2531 static void perf_mmap_data_free(struct perf_mmap_data *data)
2533 schedule_work(&data->work);
2536 static struct perf_mmap_data *
2537 perf_mmap_data_alloc(struct perf_event *event, int nr_pages)
2539 struct perf_mmap_data *data;
2543 size = sizeof(struct perf_mmap_data);
2544 size += sizeof(void *);
2546 data = kzalloc(size, GFP_KERNEL);
2550 INIT_WORK(&data->work, perf_mmap_data_free_work);
2552 all_buf = vmalloc_user((nr_pages + 1) * PAGE_SIZE);
2556 data->user_page = all_buf;
2557 data->data_pages[0] = all_buf + PAGE_SIZE;
2558 data->page_order = ilog2(nr_pages);
2572 static unsigned long perf_data_size(struct perf_mmap_data *data)
2574 return data->nr_pages << (PAGE_SHIFT + page_order(data));
2577 static int perf_mmap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
2579 struct perf_event *event = vma->vm_file->private_data;
2580 struct perf_mmap_data *data;
2581 int ret = VM_FAULT_SIGBUS;
2583 if (vmf->flags & FAULT_FLAG_MKWRITE) {
2584 if (vmf->pgoff == 0)
2590 data = rcu_dereference(event->data);
2594 if (vmf->pgoff && (vmf->flags & FAULT_FLAG_WRITE))
2597 vmf->page = perf_mmap_to_page(data, vmf->pgoff);
2601 get_page(vmf->page);
2602 vmf->page->mapping = vma->vm_file->f_mapping;
2603 vmf->page->index = vmf->pgoff;
2613 perf_mmap_data_init(struct perf_event *event, struct perf_mmap_data *data)
2615 long max_size = perf_data_size(data);
2617 if (event->attr.watermark) {
2618 data->watermark = min_t(long, max_size,
2619 event->attr.wakeup_watermark);
2622 if (!data->watermark)
2623 data->watermark = max_size / 2;
2625 atomic_set(&data->refcount, 1);
2626 rcu_assign_pointer(event->data, data);
2629 static void perf_mmap_data_free_rcu(struct rcu_head *rcu_head)
2631 struct perf_mmap_data *data;
2633 data = container_of(rcu_head, struct perf_mmap_data, rcu_head);
2634 perf_mmap_data_free(data);
2637 static struct perf_mmap_data *perf_mmap_data_get(struct perf_event *event)
2639 struct perf_mmap_data *data;
2642 data = rcu_dereference(event->data);
2644 if (!atomic_inc_not_zero(&data->refcount))
2652 static void perf_mmap_data_put(struct perf_mmap_data *data)
2654 if (!atomic_dec_and_test(&data->refcount))
2657 call_rcu(&data->rcu_head, perf_mmap_data_free_rcu);
2660 static void perf_mmap_open(struct vm_area_struct *vma)
2662 struct perf_event *event = vma->vm_file->private_data;
2664 atomic_inc(&event->mmap_count);
2667 static void perf_mmap_close(struct vm_area_struct *vma)
2669 struct perf_event *event = vma->vm_file->private_data;
2671 if (atomic_dec_and_mutex_lock(&event->mmap_count, &event->mmap_mutex)) {
2672 unsigned long size = perf_data_size(event->data);
2673 struct user_struct *user = event->mmap_user;
2674 struct perf_mmap_data *data = event->data;
2676 atomic_long_sub((size >> PAGE_SHIFT) + 1, &user->locked_vm);
2677 vma->vm_mm->locked_vm -= event->mmap_locked;
2678 rcu_assign_pointer(event->data, NULL);
2679 mutex_unlock(&event->mmap_mutex);
2681 perf_mmap_data_put(data);
2686 static const struct vm_operations_struct perf_mmap_vmops = {
2687 .open = perf_mmap_open,
2688 .close = perf_mmap_close,
2689 .fault = perf_mmap_fault,
2690 .page_mkwrite = perf_mmap_fault,
2693 static int perf_mmap(struct file *file, struct vm_area_struct *vma)
2695 struct perf_event *event = file->private_data;
2696 unsigned long user_locked, user_lock_limit;
2697 struct user_struct *user = current_user();
2698 unsigned long locked, lock_limit;
2699 struct perf_mmap_data *data;
2700 unsigned long vma_size;
2701 unsigned long nr_pages;
2702 long user_extra, extra;
2706 * Don't allow mmap() of inherited per-task counters. This would
2707 * create a performance issue due to all children writing to the
2710 if (event->cpu == -1 && event->attr.inherit)
2713 if (!(vma->vm_flags & VM_SHARED))
2716 vma_size = vma->vm_end - vma->vm_start;
2717 nr_pages = (vma_size / PAGE_SIZE) - 1;
2720 * If we have data pages ensure they're a power-of-two number, so we
2721 * can do bitmasks instead of modulo.
2723 if (nr_pages != 0 && !is_power_of_2(nr_pages))
2726 if (vma_size != PAGE_SIZE * (1 + nr_pages))
2729 if (vma->vm_pgoff != 0)
2732 WARN_ON_ONCE(event->ctx->parent_ctx);
2733 mutex_lock(&event->mmap_mutex);
2735 if (event->data->nr_pages == nr_pages)
2736 atomic_inc(&event->data->refcount);
2742 user_extra = nr_pages + 1;
2743 user_lock_limit = sysctl_perf_event_mlock >> (PAGE_SHIFT - 10);
2746 * Increase the limit linearly with more CPUs:
2748 user_lock_limit *= num_online_cpus();
2750 user_locked = atomic_long_read(&user->locked_vm) + user_extra;
2753 if (user_locked > user_lock_limit)
2754 extra = user_locked - user_lock_limit;
2756 lock_limit = rlimit(RLIMIT_MEMLOCK);
2757 lock_limit >>= PAGE_SHIFT;
2758 locked = vma->vm_mm->locked_vm + extra;
2760 if ((locked > lock_limit) && perf_paranoid_tracepoint_raw() &&
2761 !capable(CAP_IPC_LOCK)) {
2766 WARN_ON(event->data);
2768 data = perf_mmap_data_alloc(event, nr_pages);
2774 perf_mmap_data_init(event, data);
2775 if (vma->vm_flags & VM_WRITE)
2776 event->data->writable = 1;
2778 atomic_long_add(user_extra, &user->locked_vm);
2779 event->mmap_locked = extra;
2780 event->mmap_user = get_current_user();
2781 vma->vm_mm->locked_vm += event->mmap_locked;
2785 atomic_inc(&event->mmap_count);
2786 mutex_unlock(&event->mmap_mutex);
2788 vma->vm_flags |= VM_RESERVED;
2789 vma->vm_ops = &perf_mmap_vmops;
2794 static int perf_fasync(int fd, struct file *filp, int on)
2796 struct inode *inode = filp->f_path.dentry->d_inode;
2797 struct perf_event *event = filp->private_data;
2800 mutex_lock(&inode->i_mutex);
2801 retval = fasync_helper(fd, filp, on, &event->fasync);
2802 mutex_unlock(&inode->i_mutex);
2810 static const struct file_operations perf_fops = {
2811 .llseek = no_llseek,
2812 .release = perf_release,
2815 .unlocked_ioctl = perf_ioctl,
2816 .compat_ioctl = perf_ioctl,
2818 .fasync = perf_fasync,
2824 * If there's data, ensure we set the poll() state and publish everything
2825 * to user-space before waking everybody up.
2828 void perf_event_wakeup(struct perf_event *event)
2830 wake_up_all(&event->waitq);
2832 if (event->pending_kill) {
2833 kill_fasync(&event->fasync, SIGIO, event->pending_kill);
2834 event->pending_kill = 0;
2841 * Handle the case where we need to wakeup up from NMI (or rq->lock) context.
2843 * The NMI bit means we cannot possibly take locks. Therefore, maintain a
2844 * single linked list and use cmpxchg() to add entries lockless.
2847 static void perf_pending_event(struct perf_pending_entry *entry)
2849 struct perf_event *event = container_of(entry,
2850 struct perf_event, pending);
2852 if (event->pending_disable) {
2853 event->pending_disable = 0;
2854 __perf_event_disable(event);
2857 if (event->pending_wakeup) {
2858 event->pending_wakeup = 0;
2859 perf_event_wakeup(event);
2863 #define PENDING_TAIL ((struct perf_pending_entry *)-1UL)
2865 static DEFINE_PER_CPU(struct perf_pending_entry *, perf_pending_head) = {
2869 static void perf_pending_queue(struct perf_pending_entry *entry,
2870 void (*func)(struct perf_pending_entry *))
2872 struct perf_pending_entry **head;
2874 if (cmpxchg(&entry->next, NULL, PENDING_TAIL) != NULL)
2879 head = &get_cpu_var(perf_pending_head);
2882 entry->next = *head;
2883 } while (cmpxchg(head, entry->next, entry) != entry->next);
2885 set_perf_event_pending();
2887 put_cpu_var(perf_pending_head);
2890 static int __perf_pending_run(void)
2892 struct perf_pending_entry *list;
2895 list = xchg(&__get_cpu_var(perf_pending_head), PENDING_TAIL);
2896 while (list != PENDING_TAIL) {
2897 void (*func)(struct perf_pending_entry *);
2898 struct perf_pending_entry *entry = list;
2905 * Ensure we observe the unqueue before we issue the wakeup,
2906 * so that we won't be waiting forever.
2907 * -- see perf_not_pending().
2918 static inline int perf_not_pending(struct perf_event *event)
2921 * If we flush on whatever cpu we run, there is a chance we don't
2925 __perf_pending_run();
2929 * Ensure we see the proper queue state before going to sleep
2930 * so that we do not miss the wakeup. -- see perf_pending_handle()
2933 return event->pending.next == NULL;
2936 static void perf_pending_sync(struct perf_event *event)
2938 wait_event(event->waitq, perf_not_pending(event));
2941 void perf_event_do_pending(void)
2943 __perf_pending_run();
2947 * Callchain support -- arch specific
2950 __weak struct perf_callchain_entry *perf_callchain(struct pt_regs *regs)
2956 void perf_arch_fetch_caller_regs(struct pt_regs *regs, unsigned long ip, int skip)
2962 * We assume there is only KVM supporting the callbacks.
2963 * Later on, we might change it to a list if there is
2964 * another virtualization implementation supporting the callbacks.
2966 struct perf_guest_info_callbacks *perf_guest_cbs;
2968 int perf_register_guest_info_callbacks(struct perf_guest_info_callbacks *cbs)
2970 perf_guest_cbs = cbs;
2973 EXPORT_SYMBOL_GPL(perf_register_guest_info_callbacks);
2975 int perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks *cbs)
2977 perf_guest_cbs = NULL;
2980 EXPORT_SYMBOL_GPL(perf_unregister_guest_info_callbacks);
2985 static bool perf_output_space(struct perf_mmap_data *data, unsigned long tail,
2986 unsigned long offset, unsigned long head)
2990 if (!data->writable)
2993 mask = perf_data_size(data) - 1;
2995 offset = (offset - tail) & mask;
2996 head = (head - tail) & mask;
2998 if ((int)(head - offset) < 0)
3004 static void perf_output_wakeup(struct perf_output_handle *handle)
3006 atomic_set(&handle->data->poll, POLL_IN);
3009 handle->event->pending_wakeup = 1;
3010 perf_pending_queue(&handle->event->pending,
3011 perf_pending_event);
3013 perf_event_wakeup(handle->event);
3017 * We need to ensure a later event_id doesn't publish a head when a former
3018 * event isn't done writing. However since we need to deal with NMIs we
3019 * cannot fully serialize things.
3021 * We only publish the head (and generate a wakeup) when the outer-most
3024 static void perf_output_get_handle(struct perf_output_handle *handle)
3026 struct perf_mmap_data *data = handle->data;
3029 local_inc(&data->nest);
3030 handle->wakeup = local_read(&data->wakeup);
3033 static void perf_output_put_handle(struct perf_output_handle *handle)
3035 struct perf_mmap_data *data = handle->data;
3039 head = local_read(&data->head);
3042 * IRQ/NMI can happen here, which means we can miss a head update.
3045 if (!local_dec_and_test(&data->nest))
3049 * Publish the known good head. Rely on the full barrier implied
3050 * by atomic_dec_and_test() order the data->head read and this
3053 data->user_page->data_head = head;
3056 * Now check if we missed an update, rely on the (compiler)
3057 * barrier in atomic_dec_and_test() to re-read data->head.
3059 if (unlikely(head != local_read(&data->head))) {
3060 local_inc(&data->nest);
3064 if (handle->wakeup != local_read(&data->wakeup))
3065 perf_output_wakeup(handle);
3071 __always_inline void perf_output_copy(struct perf_output_handle *handle,
3072 const void *buf, unsigned int len)
3075 unsigned long size = min_t(unsigned long, handle->size, len);
3077 memcpy(handle->addr, buf, size);
3080 handle->addr += size;
3082 handle->size -= size;
3083 if (!handle->size) {
3084 struct perf_mmap_data *data = handle->data;
3087 handle->page &= data->nr_pages - 1;
3088 handle->addr = data->data_pages[handle->page];
3089 handle->size = PAGE_SIZE << page_order(data);
3094 int perf_output_begin(struct perf_output_handle *handle,
3095 struct perf_event *event, unsigned int size,
3096 int nmi, int sample)
3098 struct perf_mmap_data *data;
3099 unsigned long tail, offset, head;
3102 struct perf_event_header header;
3109 * For inherited events we send all the output towards the parent.
3112 event = event->parent;
3114 data = rcu_dereference(event->data);
3118 handle->data = data;
3119 handle->event = event;
3121 handle->sample = sample;
3123 if (!data->nr_pages)
3126 have_lost = local_read(&data->lost);
3128 size += sizeof(lost_event);
3130 perf_output_get_handle(handle);
3134 * Userspace could choose to issue a mb() before updating the
3135 * tail pointer. So that all reads will be completed before the
3138 tail = ACCESS_ONCE(data->user_page->data_tail);
3140 offset = head = local_read(&data->head);
3142 if (unlikely(!perf_output_space(data, tail, offset, head)))
3144 } while (local_cmpxchg(&data->head, offset, head) != offset);
3146 if (head - local_read(&data->wakeup) > data->watermark)
3147 local_add(data->watermark, &data->wakeup);
3149 handle->page = offset >> (PAGE_SHIFT + page_order(data));
3150 handle->page &= data->nr_pages - 1;
3151 handle->size = offset & ((PAGE_SIZE << page_order(data)) - 1);
3152 handle->addr = data->data_pages[handle->page];
3153 handle->addr += handle->size;
3154 handle->size = (PAGE_SIZE << page_order(data)) - handle->size;
3157 lost_event.header.type = PERF_RECORD_LOST;
3158 lost_event.header.misc = 0;
3159 lost_event.header.size = sizeof(lost_event);
3160 lost_event.id = event->id;
3161 lost_event.lost = local_xchg(&data->lost, 0);
3163 perf_output_put(handle, lost_event);
3169 local_inc(&data->lost);
3170 perf_output_put_handle(handle);
3177 void perf_output_end(struct perf_output_handle *handle)
3179 struct perf_event *event = handle->event;
3180 struct perf_mmap_data *data = handle->data;
3182 int wakeup_events = event->attr.wakeup_events;
3184 if (handle->sample && wakeup_events) {
3185 int events = local_inc_return(&data->events);
3186 if (events >= wakeup_events) {
3187 local_sub(wakeup_events, &data->events);
3188 local_inc(&data->wakeup);
3192 perf_output_put_handle(handle);
3196 static u32 perf_event_pid(struct perf_event *event, struct task_struct *p)
3199 * only top level events have the pid namespace they were created in
3202 event = event->parent;
3204 return task_tgid_nr_ns(p, event->ns);
3207 static u32 perf_event_tid(struct perf_event *event, struct task_struct *p)
3210 * only top level events have the pid namespace they were created in
3213 event = event->parent;
3215 return task_pid_nr_ns(p, event->ns);
3218 static void perf_output_read_one(struct perf_output_handle *handle,
3219 struct perf_event *event)
3221 u64 read_format = event->attr.read_format;
3225 values[n++] = atomic64_read(&event->count);
3226 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
3227 values[n++] = event->total_time_enabled +
3228 atomic64_read(&event->child_total_time_enabled);
3230 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
3231 values[n++] = event->total_time_running +
3232 atomic64_read(&event->child_total_time_running);
3234 if (read_format & PERF_FORMAT_ID)
3235 values[n++] = primary_event_id(event);
3237 perf_output_copy(handle, values, n * sizeof(u64));
3241 * XXX PERF_FORMAT_GROUP vs inherited events seems difficult.
3243 static void perf_output_read_group(struct perf_output_handle *handle,
3244 struct perf_event *event)
3246 struct perf_event *leader = event->group_leader, *sub;
3247 u64 read_format = event->attr.read_format;
3251 values[n++] = 1 + leader->nr_siblings;
3253 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
3254 values[n++] = leader->total_time_enabled;
3256 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
3257 values[n++] = leader->total_time_running;
3259 if (leader != event)
3260 leader->pmu->read(leader);
3262 values[n++] = atomic64_read(&leader->count);
3263 if (read_format & PERF_FORMAT_ID)
3264 values[n++] = primary_event_id(leader);
3266 perf_output_copy(handle, values, n * sizeof(u64));
3268 list_for_each_entry(sub, &leader->sibling_list, group_entry) {
3272 sub->pmu->read(sub);
3274 values[n++] = atomic64_read(&sub->count);
3275 if (read_format & PERF_FORMAT_ID)
3276 values[n++] = primary_event_id(sub);
3278 perf_output_copy(handle, values, n * sizeof(u64));
3282 static void perf_output_read(struct perf_output_handle *handle,
3283 struct perf_event *event)
3285 if (event->attr.read_format & PERF_FORMAT_GROUP)
3286 perf_output_read_group(handle, event);
3288 perf_output_read_one(handle, event);
3291 void perf_output_sample(struct perf_output_handle *handle,
3292 struct perf_event_header *header,
3293 struct perf_sample_data *data,
3294 struct perf_event *event)
3296 u64 sample_type = data->type;
3298 perf_output_put(handle, *header);
3300 if (sample_type & PERF_SAMPLE_IP)
3301 perf_output_put(handle, data->ip);
3303 if (sample_type & PERF_SAMPLE_TID)
3304 perf_output_put(handle, data->tid_entry);
3306 if (sample_type & PERF_SAMPLE_TIME)
3307 perf_output_put(handle, data->time);
3309 if (sample_type & PERF_SAMPLE_ADDR)
3310 perf_output_put(handle, data->addr);
3312 if (sample_type & PERF_SAMPLE_ID)
3313 perf_output_put(handle, data->id);
3315 if (sample_type & PERF_SAMPLE_STREAM_ID)
3316 perf_output_put(handle, data->stream_id);
3318 if (sample_type & PERF_SAMPLE_CPU)
3319 perf_output_put(handle, data->cpu_entry);
3321 if (sample_type & PERF_SAMPLE_PERIOD)
3322 perf_output_put(handle, data->period);
3324 if (sample_type & PERF_SAMPLE_READ)
3325 perf_output_read(handle, event);
3327 if (sample_type & PERF_SAMPLE_CALLCHAIN) {
3328 if (data->callchain) {
3331 if (data->callchain)
3332 size += data->callchain->nr;
3334 size *= sizeof(u64);
3336 perf_output_copy(handle, data->callchain, size);
3339 perf_output_put(handle, nr);
3343 if (sample_type & PERF_SAMPLE_RAW) {
3345 perf_output_put(handle, data->raw->size);
3346 perf_output_copy(handle, data->raw->data,
3353 .size = sizeof(u32),
3356 perf_output_put(handle, raw);
3361 void perf_prepare_sample(struct perf_event_header *header,
3362 struct perf_sample_data *data,
3363 struct perf_event *event,
3364 struct pt_regs *regs)
3366 u64 sample_type = event->attr.sample_type;
3368 data->type = sample_type;
3370 header->type = PERF_RECORD_SAMPLE;
3371 header->size = sizeof(*header);
3374 header->misc |= perf_misc_flags(regs);
3376 if (sample_type & PERF_SAMPLE_IP) {
3377 data->ip = perf_instruction_pointer(regs);
3379 header->size += sizeof(data->ip);
3382 if (sample_type & PERF_SAMPLE_TID) {
3383 /* namespace issues */
3384 data->tid_entry.pid = perf_event_pid(event, current);
3385 data->tid_entry.tid = perf_event_tid(event, current);
3387 header->size += sizeof(data->tid_entry);
3390 if (sample_type & PERF_SAMPLE_TIME) {
3391 data->time = perf_clock();
3393 header->size += sizeof(data->time);
3396 if (sample_type & PERF_SAMPLE_ADDR)
3397 header->size += sizeof(data->addr);
3399 if (sample_type & PERF_SAMPLE_ID) {
3400 data->id = primary_event_id(event);
3402 header->size += sizeof(data->id);
3405 if (sample_type & PERF_SAMPLE_STREAM_ID) {
3406 data->stream_id = event->id;
3408 header->size += sizeof(data->stream_id);
3411 if (sample_type & PERF_SAMPLE_CPU) {
3412 data->cpu_entry.cpu = raw_smp_processor_id();
3413 data->cpu_entry.reserved = 0;
3415 header->size += sizeof(data->cpu_entry);
3418 if (sample_type & PERF_SAMPLE_PERIOD)
3419 header->size += sizeof(data->period);
3421 if (sample_type & PERF_SAMPLE_READ)
3422 header->size += perf_event_read_size(event);
3424 if (sample_type & PERF_SAMPLE_CALLCHAIN) {
3427 data->callchain = perf_callchain(regs);
3429 if (data->callchain)
3430 size += data->callchain->nr;
3432 header->size += size * sizeof(u64);
3435 if (sample_type & PERF_SAMPLE_RAW) {
3436 int size = sizeof(u32);
3439 size += data->raw->size;
3441 size += sizeof(u32);
3443 WARN_ON_ONCE(size & (sizeof(u64)-1));
3444 header->size += size;
3448 static void perf_event_output(struct perf_event *event, int nmi,
3449 struct perf_sample_data *data,
3450 struct pt_regs *regs)
3452 struct perf_output_handle handle;
3453 struct perf_event_header header;
3455 perf_prepare_sample(&header, data, event, regs);
3457 if (perf_output_begin(&handle, event, header.size, nmi, 1))
3460 perf_output_sample(&handle, &header, data, event);
3462 perf_output_end(&handle);
3469 struct perf_read_event {
3470 struct perf_event_header header;
3477 perf_event_read_event(struct perf_event *event,
3478 struct task_struct *task)
3480 struct perf_output_handle handle;
3481 struct perf_read_event read_event = {
3483 .type = PERF_RECORD_READ,
3485 .size = sizeof(read_event) + perf_event_read_size(event),
3487 .pid = perf_event_pid(event, task),
3488 .tid = perf_event_tid(event, task),
3492 ret = perf_output_begin(&handle, event, read_event.header.size, 0, 0);
3496 perf_output_put(&handle, read_event);
3497 perf_output_read(&handle, event);
3499 perf_output_end(&handle);
3503 * task tracking -- fork/exit
3505 * enabled by: attr.comm | attr.mmap | attr.task
3508 struct perf_task_event {
3509 struct task_struct *task;
3510 struct perf_event_context *task_ctx;
3513 struct perf_event_header header;
3523 static void perf_event_task_output(struct perf_event *event,
3524 struct perf_task_event *task_event)
3526 struct perf_output_handle handle;
3527 struct task_struct *task = task_event->task;
3530 size = task_event->event_id.header.size;
3531 ret = perf_output_begin(&handle, event, size, 0, 0);
3536 task_event->event_id.pid = perf_event_pid(event, task);
3537 task_event->event_id.ppid = perf_event_pid(event, current);
3539 task_event->event_id.tid = perf_event_tid(event, task);
3540 task_event->event_id.ptid = perf_event_tid(event, current);
3542 perf_output_put(&handle, task_event->event_id);
3544 perf_output_end(&handle);
3547 static int perf_event_task_match(struct perf_event *event)
3549 if (event->state < PERF_EVENT_STATE_INACTIVE)
3552 if (event->cpu != -1 && event->cpu != smp_processor_id())
3555 if (event->attr.comm || event->attr.mmap || event->attr.task)
3561 static void perf_event_task_ctx(struct perf_event_context *ctx,
3562 struct perf_task_event *task_event)
3564 struct perf_event *event;
3566 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
3567 if (perf_event_task_match(event))
3568 perf_event_task_output(event, task_event);
3572 static void perf_event_task_event(struct perf_task_event *task_event)
3574 struct perf_cpu_context *cpuctx;
3575 struct perf_event_context *ctx = task_event->task_ctx;
3578 cpuctx = &get_cpu_var(perf_cpu_context);
3579 perf_event_task_ctx(&cpuctx->ctx, task_event);
3581 ctx = rcu_dereference(current->perf_event_ctxp);
3583 perf_event_task_ctx(ctx, task_event);
3584 put_cpu_var(perf_cpu_context);
3588 static void perf_event_task(struct task_struct *task,
3589 struct perf_event_context *task_ctx,
3592 struct perf_task_event task_event;
3594 if (!atomic_read(&nr_comm_events) &&
3595 !atomic_read(&nr_mmap_events) &&
3596 !atomic_read(&nr_task_events))
3599 task_event = (struct perf_task_event){
3601 .task_ctx = task_ctx,
3604 .type = new ? PERF_RECORD_FORK : PERF_RECORD_EXIT,
3606 .size = sizeof(task_event.event_id),
3612 .time = perf_clock(),
3616 perf_event_task_event(&task_event);
3619 void perf_event_fork(struct task_struct *task)
3621 perf_event_task(task, NULL, 1);
3628 struct perf_comm_event {
3629 struct task_struct *task;
3634 struct perf_event_header header;
3641 static void perf_event_comm_output(struct perf_event *event,
3642 struct perf_comm_event *comm_event)
3644 struct perf_output_handle handle;
3645 int size = comm_event->event_id.header.size;
3646 int ret = perf_output_begin(&handle, event, size, 0, 0);
3651 comm_event->event_id.pid = perf_event_pid(event, comm_event->task);
3652 comm_event->event_id.tid = perf_event_tid(event, comm_event->task);
3654 perf_output_put(&handle, comm_event->event_id);
3655 perf_output_copy(&handle, comm_event->comm,
3656 comm_event->comm_size);
3657 perf_output_end(&handle);
3660 static int perf_event_comm_match(struct perf_event *event)
3662 if (event->state < PERF_EVENT_STATE_INACTIVE)
3665 if (event->cpu != -1 && event->cpu != smp_processor_id())
3668 if (event->attr.comm)
3674 static void perf_event_comm_ctx(struct perf_event_context *ctx,
3675 struct perf_comm_event *comm_event)
3677 struct perf_event *event;
3679 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
3680 if (perf_event_comm_match(event))
3681 perf_event_comm_output(event, comm_event);
3685 static void perf_event_comm_event(struct perf_comm_event *comm_event)
3687 struct perf_cpu_context *cpuctx;
3688 struct perf_event_context *ctx;
3690 char comm[TASK_COMM_LEN];
3692 memset(comm, 0, sizeof(comm));
3693 strlcpy(comm, comm_event->task->comm, sizeof(comm));
3694 size = ALIGN(strlen(comm)+1, sizeof(u64));
3696 comm_event->comm = comm;
3697 comm_event->comm_size = size;
3699 comm_event->event_id.header.size = sizeof(comm_event->event_id) + size;
3702 cpuctx = &get_cpu_var(perf_cpu_context);
3703 perf_event_comm_ctx(&cpuctx->ctx, comm_event);
3704 ctx = rcu_dereference(current->perf_event_ctxp);
3706 perf_event_comm_ctx(ctx, comm_event);
3707 put_cpu_var(perf_cpu_context);
3711 void perf_event_comm(struct task_struct *task)
3713 struct perf_comm_event comm_event;
3715 if (task->perf_event_ctxp)
3716 perf_event_enable_on_exec(task);
3718 if (!atomic_read(&nr_comm_events))
3721 comm_event = (struct perf_comm_event){
3727 .type = PERF_RECORD_COMM,
3736 perf_event_comm_event(&comm_event);
3743 struct perf_mmap_event {
3744 struct vm_area_struct *vma;
3746 const char *file_name;
3750 struct perf_event_header header;
3760 static void perf_event_mmap_output(struct perf_event *event,
3761 struct perf_mmap_event *mmap_event)
3763 struct perf_output_handle handle;
3764 int size = mmap_event->event_id.header.size;
3765 int ret = perf_output_begin(&handle, event, size, 0, 0);
3770 mmap_event->event_id.pid = perf_event_pid(event, current);
3771 mmap_event->event_id.tid = perf_event_tid(event, current);
3773 perf_output_put(&handle, mmap_event->event_id);
3774 perf_output_copy(&handle, mmap_event->file_name,
3775 mmap_event->file_size);
3776 perf_output_end(&handle);
3779 static int perf_event_mmap_match(struct perf_event *event,
3780 struct perf_mmap_event *mmap_event)
3782 if (event->state < PERF_EVENT_STATE_INACTIVE)
3785 if (event->cpu != -1 && event->cpu != smp_processor_id())
3788 if (event->attr.mmap)
3794 static void perf_event_mmap_ctx(struct perf_event_context *ctx,
3795 struct perf_mmap_event *mmap_event)
3797 struct perf_event *event;
3799 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
3800 if (perf_event_mmap_match(event, mmap_event))
3801 perf_event_mmap_output(event, mmap_event);
3805 static void perf_event_mmap_event(struct perf_mmap_event *mmap_event)
3807 struct perf_cpu_context *cpuctx;
3808 struct perf_event_context *ctx;
3809 struct vm_area_struct *vma = mmap_event->vma;
3810 struct file *file = vma->vm_file;
3816 memset(tmp, 0, sizeof(tmp));
3820 * d_path works from the end of the buffer backwards, so we
3821 * need to add enough zero bytes after the string to handle
3822 * the 64bit alignment we do later.
3824 buf = kzalloc(PATH_MAX + sizeof(u64), GFP_KERNEL);
3826 name = strncpy(tmp, "//enomem", sizeof(tmp));
3829 name = d_path(&file->f_path, buf, PATH_MAX);
3831 name = strncpy(tmp, "//toolong", sizeof(tmp));
3835 if (arch_vma_name(mmap_event->vma)) {
3836 name = strncpy(tmp, arch_vma_name(mmap_event->vma),
3842 name = strncpy(tmp, "[vdso]", sizeof(tmp));
3846 name = strncpy(tmp, "//anon", sizeof(tmp));
3851 size = ALIGN(strlen(name)+1, sizeof(u64));
3853 mmap_event->file_name = name;
3854 mmap_event->file_size = size;
3856 mmap_event->event_id.header.size = sizeof(mmap_event->event_id) + size;
3859 cpuctx = &get_cpu_var(perf_cpu_context);
3860 perf_event_mmap_ctx(&cpuctx->ctx, mmap_event);
3861 ctx = rcu_dereference(current->perf_event_ctxp);
3863 perf_event_mmap_ctx(ctx, mmap_event);
3864 put_cpu_var(perf_cpu_context);
3870 void __perf_event_mmap(struct vm_area_struct *vma)
3872 struct perf_mmap_event mmap_event;
3874 if (!atomic_read(&nr_mmap_events))
3877 mmap_event = (struct perf_mmap_event){
3883 .type = PERF_RECORD_MMAP,
3884 .misc = PERF_RECORD_MISC_USER,
3889 .start = vma->vm_start,
3890 .len = vma->vm_end - vma->vm_start,
3891 .pgoff = (u64)vma->vm_pgoff << PAGE_SHIFT,
3895 perf_event_mmap_event(&mmap_event);
3899 * IRQ throttle logging
3902 static void perf_log_throttle(struct perf_event *event, int enable)
3904 struct perf_output_handle handle;
3908 struct perf_event_header header;
3912 } throttle_event = {
3914 .type = PERF_RECORD_THROTTLE,
3916 .size = sizeof(throttle_event),
3918 .time = perf_clock(),
3919 .id = primary_event_id(event),
3920 .stream_id = event->id,
3924 throttle_event.header.type = PERF_RECORD_UNTHROTTLE;
3926 ret = perf_output_begin(&handle, event, sizeof(throttle_event), 1, 0);
3930 perf_output_put(&handle, throttle_event);
3931 perf_output_end(&handle);
3935 * Generic event overflow handling, sampling.
3938 static int __perf_event_overflow(struct perf_event *event, int nmi,
3939 int throttle, struct perf_sample_data *data,
3940 struct pt_regs *regs)
3942 int events = atomic_read(&event->event_limit);
3943 struct hw_perf_event *hwc = &event->hw;
3946 throttle = (throttle && event->pmu->unthrottle != NULL);
3951 if (hwc->interrupts != MAX_INTERRUPTS) {
3953 if (HZ * hwc->interrupts >
3954 (u64)sysctl_perf_event_sample_rate) {
3955 hwc->interrupts = MAX_INTERRUPTS;
3956 perf_log_throttle(event, 0);
3961 * Keep re-disabling events even though on the previous
3962 * pass we disabled it - just in case we raced with a
3963 * sched-in and the event got enabled again:
3969 if (event->attr.freq) {
3970 u64 now = perf_clock();
3971 s64 delta = now - hwc->freq_time_stamp;
3973 hwc->freq_time_stamp = now;
3975 if (delta > 0 && delta < 2*TICK_NSEC)
3976 perf_adjust_period(event, delta, hwc->last_period);
3980 * XXX event_limit might not quite work as expected on inherited
3984 event->pending_kill = POLL_IN;
3985 if (events && atomic_dec_and_test(&event->event_limit)) {
3987 event->pending_kill = POLL_HUP;
3989 event->pending_disable = 1;
3990 perf_pending_queue(&event->pending,
3991 perf_pending_event);
3993 perf_event_disable(event);
3996 if (event->overflow_handler)
3997 event->overflow_handler(event, nmi, data, regs);
3999 perf_event_output(event, nmi, data, regs);
4004 int perf_event_overflow(struct perf_event *event, int nmi,
4005 struct perf_sample_data *data,
4006 struct pt_regs *regs)
4008 return __perf_event_overflow(event, nmi, 1, data, regs);
4012 * Generic software event infrastructure
4016 * We directly increment event->count and keep a second value in
4017 * event->hw.period_left to count intervals. This period event
4018 * is kept in the range [-sample_period, 0] so that we can use the
4022 static u64 perf_swevent_set_period(struct perf_event *event)
4024 struct hw_perf_event *hwc = &event->hw;
4025 u64 period = hwc->last_period;
4029 hwc->last_period = hwc->sample_period;
4032 old = val = atomic64_read(&hwc->period_left);
4036 nr = div64_u64(period + val, period);
4037 offset = nr * period;
4039 if (atomic64_cmpxchg(&hwc->period_left, old, val) != old)
4045 static void perf_swevent_overflow(struct perf_event *event, u64 overflow,
4046 int nmi, struct perf_sample_data *data,
4047 struct pt_regs *regs)
4049 struct hw_perf_event *hwc = &event->hw;
4052 data->period = event->hw.last_period;
4054 overflow = perf_swevent_set_period(event);
4056 if (hwc->interrupts == MAX_INTERRUPTS)
4059 for (; overflow; overflow--) {
4060 if (__perf_event_overflow(event, nmi, throttle,
4063 * We inhibit the overflow from happening when
4064 * hwc->interrupts == MAX_INTERRUPTS.
4072 static void perf_swevent_add(struct perf_event *event, u64 nr,
4073 int nmi, struct perf_sample_data *data,
4074 struct pt_regs *regs)
4076 struct hw_perf_event *hwc = &event->hw;
4078 atomic64_add(nr, &event->count);
4083 if (!hwc->sample_period)
4086 if (nr == 1 && hwc->sample_period == 1 && !event->attr.freq)
4087 return perf_swevent_overflow(event, 1, nmi, data, regs);
4089 if (atomic64_add_negative(nr, &hwc->period_left))
4092 perf_swevent_overflow(event, 0, nmi, data, regs);
4095 static int perf_exclude_event(struct perf_event *event,
4096 struct pt_regs *regs)
4099 if (event->attr.exclude_user && user_mode(regs))
4102 if (event->attr.exclude_kernel && !user_mode(regs))
4109 static int perf_swevent_match(struct perf_event *event,
4110 enum perf_type_id type,
4112 struct perf_sample_data *data,
4113 struct pt_regs *regs)
4115 if (event->attr.type != type)
4118 if (event->attr.config != event_id)
4121 if (perf_exclude_event(event, regs))
4127 static inline u64 swevent_hash(u64 type, u32 event_id)
4129 u64 val = event_id | (type << 32);
4131 return hash_64(val, SWEVENT_HLIST_BITS);
4134 static inline struct hlist_head *
4135 __find_swevent_head(struct swevent_hlist *hlist, u64 type, u32 event_id)
4137 u64 hash = swevent_hash(type, event_id);
4139 return &hlist->heads[hash];
4142 /* For the read side: events when they trigger */
4143 static inline struct hlist_head *
4144 find_swevent_head_rcu(struct perf_cpu_context *ctx, u64 type, u32 event_id)
4146 struct swevent_hlist *hlist;
4148 hlist = rcu_dereference(ctx->swevent_hlist);
4152 return __find_swevent_head(hlist, type, event_id);
4155 /* For the event head insertion and removal in the hlist */
4156 static inline struct hlist_head *
4157 find_swevent_head(struct perf_cpu_context *ctx, struct perf_event *event)
4159 struct swevent_hlist *hlist;
4160 u32 event_id = event->attr.config;
4161 u64 type = event->attr.type;
4164 * Event scheduling is always serialized against hlist allocation
4165 * and release. Which makes the protected version suitable here.
4166 * The context lock guarantees that.
4168 hlist = rcu_dereference_protected(ctx->swevent_hlist,
4169 lockdep_is_held(&event->ctx->lock));
4173 return __find_swevent_head(hlist, type, event_id);
4176 static void do_perf_sw_event(enum perf_type_id type, u32 event_id,
4178 struct perf_sample_data *data,
4179 struct pt_regs *regs)
4181 struct perf_cpu_context *cpuctx;
4182 struct perf_event *event;
4183 struct hlist_node *node;
4184 struct hlist_head *head;
4186 cpuctx = &__get_cpu_var(perf_cpu_context);
4190 head = find_swevent_head_rcu(cpuctx, type, event_id);
4195 hlist_for_each_entry_rcu(event, node, head, hlist_entry) {
4196 if (perf_swevent_match(event, type, event_id, data, regs))
4197 perf_swevent_add(event, nr, nmi, data, regs);
4203 int perf_swevent_get_recursion_context(void)
4205 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
4212 else if (in_softirq())
4217 if (cpuctx->recursion[rctx])
4220 cpuctx->recursion[rctx]++;
4225 EXPORT_SYMBOL_GPL(perf_swevent_get_recursion_context);
4227 void perf_swevent_put_recursion_context(int rctx)
4229 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
4231 cpuctx->recursion[rctx]--;
4233 EXPORT_SYMBOL_GPL(perf_swevent_put_recursion_context);
4236 void __perf_sw_event(u32 event_id, u64 nr, int nmi,
4237 struct pt_regs *regs, u64 addr)
4239 struct perf_sample_data data;
4242 preempt_disable_notrace();
4243 rctx = perf_swevent_get_recursion_context();
4247 perf_sample_data_init(&data, addr);
4249 do_perf_sw_event(PERF_TYPE_SOFTWARE, event_id, nr, nmi, &data, regs);
4251 perf_swevent_put_recursion_context(rctx);
4252 preempt_enable_notrace();
4255 static void perf_swevent_read(struct perf_event *event)
4259 static int perf_swevent_enable(struct perf_event *event)
4261 struct hw_perf_event *hwc = &event->hw;
4262 struct perf_cpu_context *cpuctx;
4263 struct hlist_head *head;
4265 cpuctx = &__get_cpu_var(perf_cpu_context);
4267 if (hwc->sample_period) {
4268 hwc->last_period = hwc->sample_period;
4269 perf_swevent_set_period(event);
4272 head = find_swevent_head(cpuctx, event);
4273 if (WARN_ON_ONCE(!head))
4276 hlist_add_head_rcu(&event->hlist_entry, head);
4281 static void perf_swevent_disable(struct perf_event *event)
4283 hlist_del_rcu(&event->hlist_entry);
4286 static void perf_swevent_void(struct perf_event *event)
4290 static int perf_swevent_int(struct perf_event *event)
4295 static const struct pmu perf_ops_generic = {
4296 .enable = perf_swevent_enable,
4297 .disable = perf_swevent_disable,
4298 .start = perf_swevent_int,
4299 .stop = perf_swevent_void,
4300 .read = perf_swevent_read,
4301 .unthrottle = perf_swevent_void, /* hwc->interrupts already reset */
4305 * hrtimer based swevent callback
4308 static enum hrtimer_restart perf_swevent_hrtimer(struct hrtimer *hrtimer)
4310 enum hrtimer_restart ret = HRTIMER_RESTART;
4311 struct perf_sample_data data;
4312 struct pt_regs *regs;
4313 struct perf_event *event;
4316 event = container_of(hrtimer, struct perf_event, hw.hrtimer);
4317 event->pmu->read(event);
4319 perf_sample_data_init(&data, 0);
4320 data.period = event->hw.last_period;
4321 regs = get_irq_regs();
4323 if (regs && !perf_exclude_event(event, regs)) {
4324 if (!(event->attr.exclude_idle && current->pid == 0))
4325 if (perf_event_overflow(event, 0, &data, regs))
4326 ret = HRTIMER_NORESTART;
4329 period = max_t(u64, 10000, event->hw.sample_period);
4330 hrtimer_forward_now(hrtimer, ns_to_ktime(period));
4335 static void perf_swevent_start_hrtimer(struct perf_event *event)
4337 struct hw_perf_event *hwc = &event->hw;
4339 hrtimer_init(&hwc->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
4340 hwc->hrtimer.function = perf_swevent_hrtimer;
4341 if (hwc->sample_period) {
4344 if (hwc->remaining) {
4345 if (hwc->remaining < 0)
4348 period = hwc->remaining;
4351 period = max_t(u64, 10000, hwc->sample_period);
4353 __hrtimer_start_range_ns(&hwc->hrtimer,
4354 ns_to_ktime(period), 0,
4355 HRTIMER_MODE_REL, 0);
4359 static void perf_swevent_cancel_hrtimer(struct perf_event *event)
4361 struct hw_perf_event *hwc = &event->hw;
4363 if (hwc->sample_period) {
4364 ktime_t remaining = hrtimer_get_remaining(&hwc->hrtimer);
4365 hwc->remaining = ktime_to_ns(remaining);
4367 hrtimer_cancel(&hwc->hrtimer);
4372 * Software event: cpu wall time clock
4375 static void cpu_clock_perf_event_update(struct perf_event *event)
4377 int cpu = raw_smp_processor_id();
4381 now = cpu_clock(cpu);
4382 prev = atomic64_xchg(&event->hw.prev_count, now);
4383 atomic64_add(now - prev, &event->count);
4386 static int cpu_clock_perf_event_enable(struct perf_event *event)
4388 struct hw_perf_event *hwc = &event->hw;
4389 int cpu = raw_smp_processor_id();
4391 atomic64_set(&hwc->prev_count, cpu_clock(cpu));
4392 perf_swevent_start_hrtimer(event);
4397 static void cpu_clock_perf_event_disable(struct perf_event *event)
4399 perf_swevent_cancel_hrtimer(event);
4400 cpu_clock_perf_event_update(event);
4403 static void cpu_clock_perf_event_read(struct perf_event *event)
4405 cpu_clock_perf_event_update(event);
4408 static const struct pmu perf_ops_cpu_clock = {
4409 .enable = cpu_clock_perf_event_enable,
4410 .disable = cpu_clock_perf_event_disable,
4411 .read = cpu_clock_perf_event_read,
4415 * Software event: task time clock
4418 static void task_clock_perf_event_update(struct perf_event *event, u64 now)
4423 prev = atomic64_xchg(&event->hw.prev_count, now);
4425 atomic64_add(delta, &event->count);
4428 static int task_clock_perf_event_enable(struct perf_event *event)
4430 struct hw_perf_event *hwc = &event->hw;
4433 now = event->ctx->time;
4435 atomic64_set(&hwc->prev_count, now);
4437 perf_swevent_start_hrtimer(event);
4442 static void task_clock_perf_event_disable(struct perf_event *event)
4444 perf_swevent_cancel_hrtimer(event);
4445 task_clock_perf_event_update(event, event->ctx->time);
4449 static void task_clock_perf_event_read(struct perf_event *event)
4454 update_context_time(event->ctx);
4455 time = event->ctx->time;
4457 u64 now = perf_clock();
4458 u64 delta = now - event->ctx->timestamp;
4459 time = event->ctx->time + delta;
4462 task_clock_perf_event_update(event, time);
4465 static const struct pmu perf_ops_task_clock = {
4466 .enable = task_clock_perf_event_enable,
4467 .disable = task_clock_perf_event_disable,
4468 .read = task_clock_perf_event_read,
4471 /* Deref the hlist from the update side */
4472 static inline struct swevent_hlist *
4473 swevent_hlist_deref(struct perf_cpu_context *cpuctx)
4475 return rcu_dereference_protected(cpuctx->swevent_hlist,
4476 lockdep_is_held(&cpuctx->hlist_mutex));
4479 static void swevent_hlist_release_rcu(struct rcu_head *rcu_head)
4481 struct swevent_hlist *hlist;
4483 hlist = container_of(rcu_head, struct swevent_hlist, rcu_head);
4487 static void swevent_hlist_release(struct perf_cpu_context *cpuctx)
4489 struct swevent_hlist *hlist = swevent_hlist_deref(cpuctx);
4494 rcu_assign_pointer(cpuctx->swevent_hlist, NULL);
4495 call_rcu(&hlist->rcu_head, swevent_hlist_release_rcu);
4498 static void swevent_hlist_put_cpu(struct perf_event *event, int cpu)
4500 struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
4502 mutex_lock(&cpuctx->hlist_mutex);
4504 if (!--cpuctx->hlist_refcount)
4505 swevent_hlist_release(cpuctx);
4507 mutex_unlock(&cpuctx->hlist_mutex);
4510 static void swevent_hlist_put(struct perf_event *event)
4514 if (event->cpu != -1) {
4515 swevent_hlist_put_cpu(event, event->cpu);
4519 for_each_possible_cpu(cpu)
4520 swevent_hlist_put_cpu(event, cpu);
4523 static int swevent_hlist_get_cpu(struct perf_event *event, int cpu)
4525 struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
4528 mutex_lock(&cpuctx->hlist_mutex);
4530 if (!swevent_hlist_deref(cpuctx) && cpu_online(cpu)) {
4531 struct swevent_hlist *hlist;
4533 hlist = kzalloc(sizeof(*hlist), GFP_KERNEL);
4538 rcu_assign_pointer(cpuctx->swevent_hlist, hlist);
4540 cpuctx->hlist_refcount++;
4542 mutex_unlock(&cpuctx->hlist_mutex);
4547 static int swevent_hlist_get(struct perf_event *event)
4550 int cpu, failed_cpu;
4552 if (event->cpu != -1)
4553 return swevent_hlist_get_cpu(event, event->cpu);
4556 for_each_possible_cpu(cpu) {
4557 err = swevent_hlist_get_cpu(event, cpu);
4567 for_each_possible_cpu(cpu) {
4568 if (cpu == failed_cpu)
4570 swevent_hlist_put_cpu(event, cpu);
4577 #ifdef CONFIG_EVENT_TRACING
4579 static const struct pmu perf_ops_tracepoint = {
4580 .enable = perf_trace_enable,
4581 .disable = perf_trace_disable,
4582 .start = perf_swevent_int,
4583 .stop = perf_swevent_void,
4584 .read = perf_swevent_read,
4585 .unthrottle = perf_swevent_void,
4588 static int perf_tp_filter_match(struct perf_event *event,
4589 struct perf_sample_data *data)
4591 void *record = data->raw->data;
4593 if (likely(!event->filter) || filter_match_preds(event->filter, record))
4598 static int perf_tp_event_match(struct perf_event *event,
4599 struct perf_sample_data *data,
4600 struct pt_regs *regs)
4603 * All tracepoints are from kernel-space.
4605 if (event->attr.exclude_kernel)
4608 if (!perf_tp_filter_match(event, data))
4614 void perf_tp_event(u64 addr, u64 count, void *record, int entry_size,
4615 struct pt_regs *regs, struct hlist_head *head)
4617 struct perf_sample_data data;
4618 struct perf_event *event;
4619 struct hlist_node *node;
4621 struct perf_raw_record raw = {
4626 perf_sample_data_init(&data, addr);
4630 hlist_for_each_entry_rcu(event, node, head, hlist_entry) {
4631 if (perf_tp_event_match(event, &data, regs))
4632 perf_swevent_add(event, count, 1, &data, regs);
4636 EXPORT_SYMBOL_GPL(perf_tp_event);
4638 static void tp_perf_event_destroy(struct perf_event *event)
4640 perf_trace_destroy(event);
4643 static const struct pmu *tp_perf_event_init(struct perf_event *event)
4648 * Raw tracepoint data is a severe data leak, only allow root to
4651 if ((event->attr.sample_type & PERF_SAMPLE_RAW) &&
4652 perf_paranoid_tracepoint_raw() &&
4653 !capable(CAP_SYS_ADMIN))
4654 return ERR_PTR(-EPERM);
4656 err = perf_trace_init(event);
4660 event->destroy = tp_perf_event_destroy;
4662 return &perf_ops_tracepoint;
4665 static int perf_event_set_filter(struct perf_event *event, void __user *arg)
4670 if (event->attr.type != PERF_TYPE_TRACEPOINT)
4673 filter_str = strndup_user(arg, PAGE_SIZE);
4674 if (IS_ERR(filter_str))
4675 return PTR_ERR(filter_str);
4677 ret = ftrace_profile_set_filter(event, event->attr.config, filter_str);
4683 static void perf_event_free_filter(struct perf_event *event)
4685 ftrace_profile_free_filter(event);
4690 static const struct pmu *tp_perf_event_init(struct perf_event *event)
4695 static int perf_event_set_filter(struct perf_event *event, void __user *arg)
4700 static void perf_event_free_filter(struct perf_event *event)
4704 #endif /* CONFIG_EVENT_TRACING */
4706 #ifdef CONFIG_HAVE_HW_BREAKPOINT
4707 static void bp_perf_event_destroy(struct perf_event *event)
4709 release_bp_slot(event);
4712 static const struct pmu *bp_perf_event_init(struct perf_event *bp)
4716 err = register_perf_hw_breakpoint(bp);
4718 return ERR_PTR(err);
4720 bp->destroy = bp_perf_event_destroy;
4722 return &perf_ops_bp;
4725 void perf_bp_event(struct perf_event *bp, void *data)
4727 struct perf_sample_data sample;
4728 struct pt_regs *regs = data;
4730 perf_sample_data_init(&sample, bp->attr.bp_addr);
4732 if (!perf_exclude_event(bp, regs))
4733 perf_swevent_add(bp, 1, 1, &sample, regs);
4736 static const struct pmu *bp_perf_event_init(struct perf_event *bp)
4741 void perf_bp_event(struct perf_event *bp, void *regs)
4746 atomic_t perf_swevent_enabled[PERF_COUNT_SW_MAX];
4748 static void sw_perf_event_destroy(struct perf_event *event)
4750 u64 event_id = event->attr.config;
4752 WARN_ON(event->parent);
4754 atomic_dec(&perf_swevent_enabled[event_id]);
4755 swevent_hlist_put(event);
4758 static const struct pmu *sw_perf_event_init(struct perf_event *event)
4760 const struct pmu *pmu = NULL;
4761 u64 event_id = event->attr.config;
4764 * Software events (currently) can't in general distinguish
4765 * between user, kernel and hypervisor events.
4766 * However, context switches and cpu migrations are considered
4767 * to be kernel events, and page faults are never hypervisor
4771 case PERF_COUNT_SW_CPU_CLOCK:
4772 pmu = &perf_ops_cpu_clock;
4775 case PERF_COUNT_SW_TASK_CLOCK:
4777 * If the user instantiates this as a per-cpu event,
4778 * use the cpu_clock event instead.
4780 if (event->ctx->task)
4781 pmu = &perf_ops_task_clock;
4783 pmu = &perf_ops_cpu_clock;
4786 case PERF_COUNT_SW_PAGE_FAULTS:
4787 case PERF_COUNT_SW_PAGE_FAULTS_MIN:
4788 case PERF_COUNT_SW_PAGE_FAULTS_MAJ:
4789 case PERF_COUNT_SW_CONTEXT_SWITCHES:
4790 case PERF_COUNT_SW_CPU_MIGRATIONS:
4791 case PERF_COUNT_SW_ALIGNMENT_FAULTS:
4792 case PERF_COUNT_SW_EMULATION_FAULTS:
4793 if (!event->parent) {
4796 err = swevent_hlist_get(event);
4798 return ERR_PTR(err);
4800 atomic_inc(&perf_swevent_enabled[event_id]);
4801 event->destroy = sw_perf_event_destroy;
4803 pmu = &perf_ops_generic;
4811 * Allocate and initialize a event structure
4813 static struct perf_event *
4814 perf_event_alloc(struct perf_event_attr *attr,
4816 struct perf_event_context *ctx,
4817 struct perf_event *group_leader,
4818 struct perf_event *parent_event,
4819 perf_overflow_handler_t overflow_handler,
4822 const struct pmu *pmu;
4823 struct perf_event *event;
4824 struct hw_perf_event *hwc;
4827 event = kzalloc(sizeof(*event), gfpflags);
4829 return ERR_PTR(-ENOMEM);
4832 * Single events are their own group leaders, with an
4833 * empty sibling list:
4836 group_leader = event;
4838 mutex_init(&event->child_mutex);
4839 INIT_LIST_HEAD(&event->child_list);
4841 INIT_LIST_HEAD(&event->group_entry);
4842 INIT_LIST_HEAD(&event->event_entry);
4843 INIT_LIST_HEAD(&event->sibling_list);
4844 init_waitqueue_head(&event->waitq);
4846 mutex_init(&event->mmap_mutex);
4849 event->attr = *attr;
4850 event->group_leader = group_leader;
4855 event->parent = parent_event;
4857 event->ns = get_pid_ns(current->nsproxy->pid_ns);
4858 event->id = atomic64_inc_return(&perf_event_id);
4860 event->state = PERF_EVENT_STATE_INACTIVE;
4862 if (!overflow_handler && parent_event)
4863 overflow_handler = parent_event->overflow_handler;
4865 event->overflow_handler = overflow_handler;
4868 event->state = PERF_EVENT_STATE_OFF;
4873 hwc->sample_period = attr->sample_period;
4874 if (attr->freq && attr->sample_freq)
4875 hwc->sample_period = 1;
4876 hwc->last_period = hwc->sample_period;
4878 atomic64_set(&hwc->period_left, hwc->sample_period);
4881 * we currently do not support PERF_FORMAT_GROUP on inherited events
4883 if (attr->inherit && (attr->read_format & PERF_FORMAT_GROUP))
4886 switch (attr->type) {
4888 case PERF_TYPE_HARDWARE:
4889 case PERF_TYPE_HW_CACHE:
4890 pmu = hw_perf_event_init(event);
4893 case PERF_TYPE_SOFTWARE:
4894 pmu = sw_perf_event_init(event);
4897 case PERF_TYPE_TRACEPOINT:
4898 pmu = tp_perf_event_init(event);
4901 case PERF_TYPE_BREAKPOINT:
4902 pmu = bp_perf_event_init(event);
4913 else if (IS_ERR(pmu))
4918 put_pid_ns(event->ns);
4920 return ERR_PTR(err);
4925 if (!event->parent) {
4926 atomic_inc(&nr_events);
4927 if (event->attr.mmap)
4928 atomic_inc(&nr_mmap_events);
4929 if (event->attr.comm)
4930 atomic_inc(&nr_comm_events);
4931 if (event->attr.task)
4932 atomic_inc(&nr_task_events);
4938 static int perf_copy_attr(struct perf_event_attr __user *uattr,
4939 struct perf_event_attr *attr)
4944 if (!access_ok(VERIFY_WRITE, uattr, PERF_ATTR_SIZE_VER0))
4948 * zero the full structure, so that a short copy will be nice.
4950 memset(attr, 0, sizeof(*attr));
4952 ret = get_user(size, &uattr->size);
4956 if (size > PAGE_SIZE) /* silly large */
4959 if (!size) /* abi compat */
4960 size = PERF_ATTR_SIZE_VER0;
4962 if (size < PERF_ATTR_SIZE_VER0)
4966 * If we're handed a bigger struct than we know of,
4967 * ensure all the unknown bits are 0 - i.e. new
4968 * user-space does not rely on any kernel feature
4969 * extensions we dont know about yet.
4971 if (size > sizeof(*attr)) {
4972 unsigned char __user *addr;
4973 unsigned char __user *end;
4976 addr = (void __user *)uattr + sizeof(*attr);
4977 end = (void __user *)uattr + size;
4979 for (; addr < end; addr++) {
4980 ret = get_user(val, addr);
4986 size = sizeof(*attr);
4989 ret = copy_from_user(attr, uattr, size);
4994 * If the type exists, the corresponding creation will verify
4997 if (attr->type >= PERF_TYPE_MAX)
5000 if (attr->__reserved_1)
5003 if (attr->sample_type & ~(PERF_SAMPLE_MAX-1))
5006 if (attr->read_format & ~(PERF_FORMAT_MAX-1))
5013 put_user(sizeof(*attr), &uattr->size);
5019 perf_event_set_output(struct perf_event *event, struct perf_event *output_event)
5021 struct perf_mmap_data *data = NULL, *old_data = NULL;
5027 /* don't allow circular references */
5028 if (event == output_event)
5032 * Don't allow cross-cpu buffers
5034 if (output_event->cpu != event->cpu)
5038 * If its not a per-cpu buffer, it must be the same task.
5040 if (output_event->cpu == -1 && output_event->ctx != event->ctx)
5044 mutex_lock(&event->mmap_mutex);
5045 /* Can't redirect output if we've got an active mmap() */
5046 if (atomic_read(&event->mmap_count))
5050 /* get the buffer we want to redirect to */
5051 data = perf_mmap_data_get(output_event);
5056 old_data = event->data;
5057 rcu_assign_pointer(event->data, data);
5060 mutex_unlock(&event->mmap_mutex);
5063 perf_mmap_data_put(old_data);
5069 * sys_perf_event_open - open a performance event, associate it to a task/cpu
5071 * @attr_uptr: event_id type attributes for monitoring/sampling
5074 * @group_fd: group leader event fd
5076 SYSCALL_DEFINE5(perf_event_open,
5077 struct perf_event_attr __user *, attr_uptr,
5078 pid_t, pid, int, cpu, int, group_fd, unsigned long, flags)
5080 struct perf_event *event, *group_leader = NULL, *output_event = NULL;
5081 struct perf_event_attr attr;
5082 struct perf_event_context *ctx;
5083 struct file *event_file = NULL;
5084 struct file *group_file = NULL;
5086 int fput_needed = 0;
5089 /* for future expandability... */
5090 if (flags & ~(PERF_FLAG_FD_NO_GROUP | PERF_FLAG_FD_OUTPUT))
5093 err = perf_copy_attr(attr_uptr, &attr);
5097 if (!attr.exclude_kernel) {
5098 if (perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN))
5103 if (attr.sample_freq > sysctl_perf_event_sample_rate)
5107 event_fd = get_unused_fd_flags(O_RDWR);
5112 * Get the target context (task or percpu):
5114 ctx = find_get_context(pid, cpu);
5120 if (group_fd != -1) {
5121 group_leader = perf_fget_light(group_fd, &fput_needed);
5122 if (IS_ERR(group_leader)) {
5123 err = PTR_ERR(group_leader);
5124 goto err_put_context;
5126 group_file = group_leader->filp;
5127 if (flags & PERF_FLAG_FD_OUTPUT)
5128 output_event = group_leader;
5129 if (flags & PERF_FLAG_FD_NO_GROUP)
5130 group_leader = NULL;
5134 * Look up the group leader (we will attach this event to it):
5140 * Do not allow a recursive hierarchy (this new sibling
5141 * becoming part of another group-sibling):
5143 if (group_leader->group_leader != group_leader)
5144 goto err_put_context;
5146 * Do not allow to attach to a group in a different
5147 * task or CPU context:
5149 if (group_leader->ctx != ctx)
5150 goto err_put_context;
5152 * Only a group leader can be exclusive or pinned
5154 if (attr.exclusive || attr.pinned)
5155 goto err_put_context;
5158 event = perf_event_alloc(&attr, cpu, ctx, group_leader,
5159 NULL, NULL, GFP_KERNEL);
5160 if (IS_ERR(event)) {
5161 err = PTR_ERR(event);
5162 goto err_put_context;
5166 err = perf_event_set_output(event, output_event);
5168 goto err_free_put_context;
5171 event_file = anon_inode_getfile("[perf_event]", &perf_fops, event, O_RDWR);
5172 if (IS_ERR(event_file)) {
5173 err = PTR_ERR(event_file);
5174 goto err_free_put_context;
5177 event->filp = event_file;
5178 WARN_ON_ONCE(ctx->parent_ctx);
5179 mutex_lock(&ctx->mutex);
5180 perf_install_in_context(ctx, event, cpu);
5182 mutex_unlock(&ctx->mutex);
5184 event->owner = current;
5185 get_task_struct(current);
5186 mutex_lock(¤t->perf_event_mutex);
5187 list_add_tail(&event->owner_entry, ¤t->perf_event_list);
5188 mutex_unlock(¤t->perf_event_mutex);
5191 * Drop the reference on the group_event after placing the
5192 * new event on the sibling_list. This ensures destruction
5193 * of the group leader will find the pointer to itself in
5194 * perf_group_detach().
5196 fput_light(group_file, fput_needed);
5197 fd_install(event_fd, event_file);
5200 err_free_put_context:
5203 fput_light(group_file, fput_needed);
5206 put_unused_fd(event_fd);
5211 * perf_event_create_kernel_counter
5213 * @attr: attributes of the counter to create
5214 * @cpu: cpu in which the counter is bound
5215 * @pid: task to profile
5218 perf_event_create_kernel_counter(struct perf_event_attr *attr, int cpu,
5220 perf_overflow_handler_t overflow_handler)
5222 struct perf_event *event;
5223 struct perf_event_context *ctx;
5227 * Get the target context (task or percpu):
5230 ctx = find_get_context(pid, cpu);
5236 event = perf_event_alloc(attr, cpu, ctx, NULL,
5237 NULL, overflow_handler, GFP_KERNEL);
5238 if (IS_ERR(event)) {
5239 err = PTR_ERR(event);
5240 goto err_put_context;
5244 WARN_ON_ONCE(ctx->parent_ctx);
5245 mutex_lock(&ctx->mutex);
5246 perf_install_in_context(ctx, event, cpu);
5248 mutex_unlock(&ctx->mutex);
5250 event->owner = current;
5251 get_task_struct(current);
5252 mutex_lock(¤t->perf_event_mutex);
5253 list_add_tail(&event->owner_entry, ¤t->perf_event_list);
5254 mutex_unlock(¤t->perf_event_mutex);
5261 return ERR_PTR(err);
5263 EXPORT_SYMBOL_GPL(perf_event_create_kernel_counter);
5266 * inherit a event from parent task to child task:
5268 static struct perf_event *
5269 inherit_event(struct perf_event *parent_event,
5270 struct task_struct *parent,
5271 struct perf_event_context *parent_ctx,
5272 struct task_struct *child,
5273 struct perf_event *group_leader,
5274 struct perf_event_context *child_ctx)
5276 struct perf_event *child_event;
5279 * Instead of creating recursive hierarchies of events,
5280 * we link inherited events back to the original parent,
5281 * which has a filp for sure, which we use as the reference
5284 if (parent_event->parent)
5285 parent_event = parent_event->parent;
5287 child_event = perf_event_alloc(&parent_event->attr,
5288 parent_event->cpu, child_ctx,
5289 group_leader, parent_event,
5291 if (IS_ERR(child_event))
5296 * Make the child state follow the state of the parent event,
5297 * not its attr.disabled bit. We hold the parent's mutex,
5298 * so we won't race with perf_event_{en, dis}able_family.
5300 if (parent_event->state >= PERF_EVENT_STATE_INACTIVE)
5301 child_event->state = PERF_EVENT_STATE_INACTIVE;
5303 child_event->state = PERF_EVENT_STATE_OFF;
5305 if (parent_event->attr.freq) {
5306 u64 sample_period = parent_event->hw.sample_period;
5307 struct hw_perf_event *hwc = &child_event->hw;
5309 hwc->sample_period = sample_period;
5310 hwc->last_period = sample_period;
5312 atomic64_set(&hwc->period_left, sample_period);
5315 child_event->overflow_handler = parent_event->overflow_handler;
5318 * Link it up in the child's context:
5320 add_event_to_ctx(child_event, child_ctx);
5323 * Get a reference to the parent filp - we will fput it
5324 * when the child event exits. This is safe to do because
5325 * we are in the parent and we know that the filp still
5326 * exists and has a nonzero count:
5328 atomic_long_inc(&parent_event->filp->f_count);
5331 * Link this into the parent event's child list
5333 WARN_ON_ONCE(parent_event->ctx->parent_ctx);
5334 mutex_lock(&parent_event->child_mutex);
5335 list_add_tail(&child_event->child_list, &parent_event->child_list);
5336 mutex_unlock(&parent_event->child_mutex);
5341 static int inherit_group(struct perf_event *parent_event,
5342 struct task_struct *parent,
5343 struct perf_event_context *parent_ctx,
5344 struct task_struct *child,
5345 struct perf_event_context *child_ctx)
5347 struct perf_event *leader;
5348 struct perf_event *sub;
5349 struct perf_event *child_ctr;
5351 leader = inherit_event(parent_event, parent, parent_ctx,
5352 child, NULL, child_ctx);
5354 return PTR_ERR(leader);
5355 list_for_each_entry(sub, &parent_event->sibling_list, group_entry) {
5356 child_ctr = inherit_event(sub, parent, parent_ctx,
5357 child, leader, child_ctx);
5358 if (IS_ERR(child_ctr))
5359 return PTR_ERR(child_ctr);
5364 static void sync_child_event(struct perf_event *child_event,
5365 struct task_struct *child)
5367 struct perf_event *parent_event = child_event->parent;
5370 if (child_event->attr.inherit_stat)
5371 perf_event_read_event(child_event, child);
5373 child_val = atomic64_read(&child_event->count);
5376 * Add back the child's count to the parent's count:
5378 atomic64_add(child_val, &parent_event->count);
5379 atomic64_add(child_event->total_time_enabled,
5380 &parent_event->child_total_time_enabled);
5381 atomic64_add(child_event->total_time_running,
5382 &parent_event->child_total_time_running);
5385 * Remove this event from the parent's list
5387 WARN_ON_ONCE(parent_event->ctx->parent_ctx);
5388 mutex_lock(&parent_event->child_mutex);
5389 list_del_init(&child_event->child_list);
5390 mutex_unlock(&parent_event->child_mutex);
5393 * Release the parent event, if this was the last
5396 fput(parent_event->filp);
5400 __perf_event_exit_task(struct perf_event *child_event,
5401 struct perf_event_context *child_ctx,
5402 struct task_struct *child)
5404 if (child_event->parent) {
5405 raw_spin_lock_irq(&child_ctx->lock);
5406 perf_group_detach(child_event);
5407 raw_spin_unlock_irq(&child_ctx->lock);
5410 perf_event_remove_from_context(child_event);
5413 * It can happen that the parent exits first, and has events
5414 * that are still around due to the child reference. These
5415 * events need to be zapped.
5417 if (child_event->parent) {
5418 sync_child_event(child_event, child);
5419 free_event(child_event);
5424 * When a child task exits, feed back event values to parent events.
5426 void perf_event_exit_task(struct task_struct *child)
5428 struct perf_event *child_event, *tmp;
5429 struct perf_event_context *child_ctx;
5430 unsigned long flags;
5432 if (likely(!child->perf_event_ctxp)) {
5433 perf_event_task(child, NULL, 0);
5437 local_irq_save(flags);
5439 * We can't reschedule here because interrupts are disabled,
5440 * and either child is current or it is a task that can't be
5441 * scheduled, so we are now safe from rescheduling changing
5444 child_ctx = child->perf_event_ctxp;
5445 __perf_event_task_sched_out(child_ctx);
5448 * Take the context lock here so that if find_get_context is
5449 * reading child->perf_event_ctxp, we wait until it has
5450 * incremented the context's refcount before we do put_ctx below.
5452 raw_spin_lock(&child_ctx->lock);
5453 child->perf_event_ctxp = NULL;
5455 * If this context is a clone; unclone it so it can't get
5456 * swapped to another process while we're removing all
5457 * the events from it.
5459 unclone_ctx(child_ctx);
5460 update_context_time(child_ctx);
5461 raw_spin_unlock_irqrestore(&child_ctx->lock, flags);
5464 * Report the task dead after unscheduling the events so that we
5465 * won't get any samples after PERF_RECORD_EXIT. We can however still
5466 * get a few PERF_RECORD_READ events.
5468 perf_event_task(child, child_ctx, 0);
5471 * We can recurse on the same lock type through:
5473 * __perf_event_exit_task()
5474 * sync_child_event()
5475 * fput(parent_event->filp)
5477 * mutex_lock(&ctx->mutex)
5479 * But since its the parent context it won't be the same instance.
5481 mutex_lock(&child_ctx->mutex);
5484 list_for_each_entry_safe(child_event, tmp, &child_ctx->pinned_groups,
5486 __perf_event_exit_task(child_event, child_ctx, child);
5488 list_for_each_entry_safe(child_event, tmp, &child_ctx->flexible_groups,
5490 __perf_event_exit_task(child_event, child_ctx, child);
5493 * If the last event was a group event, it will have appended all
5494 * its siblings to the list, but we obtained 'tmp' before that which
5495 * will still point to the list head terminating the iteration.
5497 if (!list_empty(&child_ctx->pinned_groups) ||
5498 !list_empty(&child_ctx->flexible_groups))
5501 mutex_unlock(&child_ctx->mutex);
5506 static void perf_free_event(struct perf_event *event,
5507 struct perf_event_context *ctx)
5509 struct perf_event *parent = event->parent;
5511 if (WARN_ON_ONCE(!parent))
5514 mutex_lock(&parent->child_mutex);
5515 list_del_init(&event->child_list);
5516 mutex_unlock(&parent->child_mutex);
5520 perf_group_detach(event);
5521 list_del_event(event, ctx);
5526 * free an unexposed, unused context as created by inheritance by
5527 * init_task below, used by fork() in case of fail.
5529 void perf_event_free_task(struct task_struct *task)
5531 struct perf_event_context *ctx = task->perf_event_ctxp;
5532 struct perf_event *event, *tmp;
5537 mutex_lock(&ctx->mutex);
5539 list_for_each_entry_safe(event, tmp, &ctx->pinned_groups, group_entry)
5540 perf_free_event(event, ctx);
5542 list_for_each_entry_safe(event, tmp, &ctx->flexible_groups,
5544 perf_free_event(event, ctx);
5546 if (!list_empty(&ctx->pinned_groups) ||
5547 !list_empty(&ctx->flexible_groups))
5550 mutex_unlock(&ctx->mutex);
5556 inherit_task_group(struct perf_event *event, struct task_struct *parent,
5557 struct perf_event_context *parent_ctx,
5558 struct task_struct *child,
5562 struct perf_event_context *child_ctx = child->perf_event_ctxp;
5564 if (!event->attr.inherit) {
5571 * This is executed from the parent task context, so
5572 * inherit events that have been marked for cloning.
5573 * First allocate and initialize a context for the
5577 child_ctx = kzalloc(sizeof(struct perf_event_context),
5582 __perf_event_init_context(child_ctx, child);
5583 child->perf_event_ctxp = child_ctx;
5584 get_task_struct(child);
5587 ret = inherit_group(event, parent, parent_ctx,
5598 * Initialize the perf_event context in task_struct
5600 int perf_event_init_task(struct task_struct *child)
5602 struct perf_event_context *child_ctx, *parent_ctx;
5603 struct perf_event_context *cloned_ctx;
5604 struct perf_event *event;
5605 struct task_struct *parent = current;
5606 int inherited_all = 1;
5607 unsigned long flags;
5610 child->perf_event_ctxp = NULL;
5612 mutex_init(&child->perf_event_mutex);
5613 INIT_LIST_HEAD(&child->perf_event_list);
5615 if (likely(!parent->perf_event_ctxp))
5619 * If the parent's context is a clone, pin it so it won't get
5622 parent_ctx = perf_pin_task_context(parent);
5625 * No need to check if parent_ctx != NULL here; since we saw
5626 * it non-NULL earlier, the only reason for it to become NULL
5627 * is if we exit, and since we're currently in the middle of
5628 * a fork we can't be exiting at the same time.
5632 * Lock the parent list. No need to lock the child - not PID
5633 * hashed yet and not running, so nobody can access it.
5635 mutex_lock(&parent_ctx->mutex);
5638 * We dont have to disable NMIs - we are only looking at
5639 * the list, not manipulating it:
5641 list_for_each_entry(event, &parent_ctx->pinned_groups, group_entry) {
5642 ret = inherit_task_group(event, parent, parent_ctx, child,
5649 * We can't hold ctx->lock when iterating the ->flexible_group list due
5650 * to allocations, but we need to prevent rotation because
5651 * rotate_ctx() will change the list from interrupt context.
5653 raw_spin_lock_irqsave(&parent_ctx->lock, flags);
5654 parent_ctx->rotate_disable = 1;
5655 raw_spin_unlock_irqrestore(&parent_ctx->lock, flags);
5657 list_for_each_entry(event, &parent_ctx->flexible_groups, group_entry) {
5658 ret = inherit_task_group(event, parent, parent_ctx, child,
5664 raw_spin_lock_irqsave(&parent_ctx->lock, flags);
5665 parent_ctx->rotate_disable = 0;
5666 raw_spin_unlock_irqrestore(&parent_ctx->lock, flags);
5668 child_ctx = child->perf_event_ctxp;
5670 if (child_ctx && inherited_all) {
5672 * Mark the child context as a clone of the parent
5673 * context, or of whatever the parent is a clone of.
5674 * Note that if the parent is a clone, it could get
5675 * uncloned at any point, but that doesn't matter
5676 * because the list of events and the generation
5677 * count can't have changed since we took the mutex.
5679 cloned_ctx = rcu_dereference(parent_ctx->parent_ctx);
5681 child_ctx->parent_ctx = cloned_ctx;
5682 child_ctx->parent_gen = parent_ctx->parent_gen;
5684 child_ctx->parent_ctx = parent_ctx;
5685 child_ctx->parent_gen = parent_ctx->generation;
5687 get_ctx(child_ctx->parent_ctx);
5690 mutex_unlock(&parent_ctx->mutex);
5692 perf_unpin_context(parent_ctx);
5697 static void __init perf_event_init_all_cpus(void)
5700 struct perf_cpu_context *cpuctx;
5702 for_each_possible_cpu(cpu) {
5703 cpuctx = &per_cpu(perf_cpu_context, cpu);
5704 mutex_init(&cpuctx->hlist_mutex);
5705 __perf_event_init_context(&cpuctx->ctx, NULL);
5709 static void __cpuinit perf_event_init_cpu(int cpu)
5711 struct perf_cpu_context *cpuctx;
5713 cpuctx = &per_cpu(perf_cpu_context, cpu);
5715 spin_lock(&perf_resource_lock);
5716 cpuctx->max_pertask = perf_max_events - perf_reserved_percpu;
5717 spin_unlock(&perf_resource_lock);
5719 mutex_lock(&cpuctx->hlist_mutex);
5720 if (cpuctx->hlist_refcount > 0) {
5721 struct swevent_hlist *hlist;
5723 hlist = kzalloc(sizeof(*hlist), GFP_KERNEL);
5724 WARN_ON_ONCE(!hlist);
5725 rcu_assign_pointer(cpuctx->swevent_hlist, hlist);
5727 mutex_unlock(&cpuctx->hlist_mutex);
5730 #ifdef CONFIG_HOTPLUG_CPU
5731 static void __perf_event_exit_cpu(void *info)
5733 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
5734 struct perf_event_context *ctx = &cpuctx->ctx;
5735 struct perf_event *event, *tmp;
5737 list_for_each_entry_safe(event, tmp, &ctx->pinned_groups, group_entry)
5738 __perf_event_remove_from_context(event);
5739 list_for_each_entry_safe(event, tmp, &ctx->flexible_groups, group_entry)
5740 __perf_event_remove_from_context(event);
5742 static void perf_event_exit_cpu(int cpu)
5744 struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
5745 struct perf_event_context *ctx = &cpuctx->ctx;
5747 mutex_lock(&cpuctx->hlist_mutex);
5748 swevent_hlist_release(cpuctx);
5749 mutex_unlock(&cpuctx->hlist_mutex);
5751 mutex_lock(&ctx->mutex);
5752 smp_call_function_single(cpu, __perf_event_exit_cpu, NULL, 1);
5753 mutex_unlock(&ctx->mutex);
5756 static inline void perf_event_exit_cpu(int cpu) { }
5759 static int __cpuinit
5760 perf_cpu_notify(struct notifier_block *self, unsigned long action, void *hcpu)
5762 unsigned int cpu = (long)hcpu;
5766 case CPU_UP_PREPARE:
5767 case CPU_UP_PREPARE_FROZEN:
5768 perf_event_init_cpu(cpu);
5771 case CPU_DOWN_PREPARE:
5772 case CPU_DOWN_PREPARE_FROZEN:
5773 perf_event_exit_cpu(cpu);
5784 * This has to have a higher priority than migration_notifier in sched.c.
5786 static struct notifier_block __cpuinitdata perf_cpu_nb = {
5787 .notifier_call = perf_cpu_notify,
5791 void __init perf_event_init(void)
5793 perf_event_init_all_cpus();
5794 perf_cpu_notify(&perf_cpu_nb, (unsigned long)CPU_UP_PREPARE,
5795 (void *)(long)smp_processor_id());
5796 perf_cpu_notify(&perf_cpu_nb, (unsigned long)CPU_ONLINE,
5797 (void *)(long)smp_processor_id());
5798 register_cpu_notifier(&perf_cpu_nb);
5801 static ssize_t perf_show_reserve_percpu(struct sysdev_class *class,
5802 struct sysdev_class_attribute *attr,
5805 return sprintf(buf, "%d\n", perf_reserved_percpu);
5809 perf_set_reserve_percpu(struct sysdev_class *class,
5810 struct sysdev_class_attribute *attr,
5814 struct perf_cpu_context *cpuctx;
5818 err = strict_strtoul(buf, 10, &val);
5821 if (val > perf_max_events)
5824 spin_lock(&perf_resource_lock);
5825 perf_reserved_percpu = val;
5826 for_each_online_cpu(cpu) {
5827 cpuctx = &per_cpu(perf_cpu_context, cpu);
5828 raw_spin_lock_irq(&cpuctx->ctx.lock);
5829 mpt = min(perf_max_events - cpuctx->ctx.nr_events,
5830 perf_max_events - perf_reserved_percpu);
5831 cpuctx->max_pertask = mpt;
5832 raw_spin_unlock_irq(&cpuctx->ctx.lock);
5834 spin_unlock(&perf_resource_lock);
5839 static ssize_t perf_show_overcommit(struct sysdev_class *class,
5840 struct sysdev_class_attribute *attr,
5843 return sprintf(buf, "%d\n", perf_overcommit);
5847 perf_set_overcommit(struct sysdev_class *class,
5848 struct sysdev_class_attribute *attr,
5849 const char *buf, size_t count)
5854 err = strict_strtoul(buf, 10, &val);
5860 spin_lock(&perf_resource_lock);
5861 perf_overcommit = val;
5862 spin_unlock(&perf_resource_lock);
5867 static SYSDEV_CLASS_ATTR(
5870 perf_show_reserve_percpu,
5871 perf_set_reserve_percpu
5874 static SYSDEV_CLASS_ATTR(
5877 perf_show_overcommit,
5881 static struct attribute *perfclass_attrs[] = {
5882 &attr_reserve_percpu.attr,
5883 &attr_overcommit.attr,
5887 static struct attribute_group perfclass_attr_group = {
5888 .attrs = perfclass_attrs,
5889 .name = "perf_events",
5892 static int __init perf_event_sysfs_init(void)
5894 return sysfs_create_group(&cpu_sysdev_class.kset.kobj,
5895 &perfclass_attr_group);
5897 device_initcall(perf_event_sysfs_init);