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/sysfs.h>
19 #include <linux/dcache.h>
20 #include <linux/percpu.h>
21 #include <linux/ptrace.h>
22 #include <linux/vmstat.h>
23 #include <linux/vmalloc.h>
24 #include <linux/hardirq.h>
25 #include <linux/rculist.h>
26 #include <linux/uaccess.h>
27 #include <linux/syscalls.h>
28 #include <linux/anon_inodes.h>
29 #include <linux/kernel_stat.h>
30 #include <linux/perf_event.h>
31 #include <linux/ftrace_event.h>
32 #include <linux/hw_breakpoint.h>
34 #include <asm/irq_regs.h>
37 * Each CPU has a list of per CPU events:
39 static DEFINE_PER_CPU(struct perf_cpu_context, perf_cpu_context);
41 int perf_max_events __read_mostly = 1;
42 static int perf_reserved_percpu __read_mostly;
43 static int perf_overcommit __read_mostly = 1;
45 static atomic_t nr_events __read_mostly;
46 static atomic_t nr_mmap_events __read_mostly;
47 static atomic_t nr_comm_events __read_mostly;
48 static atomic_t nr_task_events __read_mostly;
51 * perf event paranoia level:
52 * -1 - not paranoid at all
53 * 0 - disallow raw tracepoint access for unpriv
54 * 1 - disallow cpu events for unpriv
55 * 2 - disallow kernel profiling for unpriv
57 int sysctl_perf_event_paranoid __read_mostly = 1;
59 int sysctl_perf_event_mlock __read_mostly = 512; /* 'free' kb per user */
62 * max perf event sample rate
64 int sysctl_perf_event_sample_rate __read_mostly = 100000;
66 static atomic64_t perf_event_id;
69 * Lock for (sysadmin-configurable) event reservations:
71 static DEFINE_SPINLOCK(perf_resource_lock);
74 * Architecture provided APIs - weak aliases:
76 extern __weak const struct pmu *hw_perf_event_init(struct perf_event *event)
81 void __weak hw_perf_disable(void) { barrier(); }
82 void __weak hw_perf_enable(void) { barrier(); }
84 void __weak hw_perf_event_setup(int cpu) { barrier(); }
85 void __weak hw_perf_event_setup_online(int cpu) { barrier(); }
86 void __weak hw_perf_event_setup_offline(int cpu) { barrier(); }
89 hw_perf_group_sched_in(struct perf_event *group_leader,
90 struct perf_cpu_context *cpuctx,
91 struct perf_event_context *ctx)
96 void __weak perf_event_print_debug(void) { }
98 static DEFINE_PER_CPU(int, perf_disable_count);
100 void __perf_disable(void)
102 __get_cpu_var(perf_disable_count)++;
105 bool __perf_enable(void)
107 return !--__get_cpu_var(perf_disable_count);
110 void perf_disable(void)
116 void perf_enable(void)
122 static void get_ctx(struct perf_event_context *ctx)
124 WARN_ON(!atomic_inc_not_zero(&ctx->refcount));
127 static void free_ctx(struct rcu_head *head)
129 struct perf_event_context *ctx;
131 ctx = container_of(head, struct perf_event_context, rcu_head);
135 static void put_ctx(struct perf_event_context *ctx)
137 if (atomic_dec_and_test(&ctx->refcount)) {
139 put_ctx(ctx->parent_ctx);
141 put_task_struct(ctx->task);
142 call_rcu(&ctx->rcu_head, free_ctx);
146 static void unclone_ctx(struct perf_event_context *ctx)
148 if (ctx->parent_ctx) {
149 put_ctx(ctx->parent_ctx);
150 ctx->parent_ctx = NULL;
155 * If we inherit events we want to return the parent event id
158 static u64 primary_event_id(struct perf_event *event)
163 id = event->parent->id;
169 * Get the perf_event_context for a task and lock it.
170 * This has to cope with with the fact that until it is locked,
171 * the context could get moved to another task.
173 static struct perf_event_context *
174 perf_lock_task_context(struct task_struct *task, unsigned long *flags)
176 struct perf_event_context *ctx;
180 ctx = rcu_dereference(task->perf_event_ctxp);
183 * If this context is a clone of another, it might
184 * get swapped for another underneath us by
185 * perf_event_task_sched_out, though the
186 * rcu_read_lock() protects us from any context
187 * getting freed. Lock the context and check if it
188 * got swapped before we could get the lock, and retry
189 * if so. If we locked the right context, then it
190 * can't get swapped on us any more.
192 raw_spin_lock_irqsave(&ctx->lock, *flags);
193 if (ctx != rcu_dereference(task->perf_event_ctxp)) {
194 raw_spin_unlock_irqrestore(&ctx->lock, *flags);
198 if (!atomic_inc_not_zero(&ctx->refcount)) {
199 raw_spin_unlock_irqrestore(&ctx->lock, *flags);
208 * Get the context for a task and increment its pin_count so it
209 * can't get swapped to another task. This also increments its
210 * reference count so that the context can't get freed.
212 static struct perf_event_context *perf_pin_task_context(struct task_struct *task)
214 struct perf_event_context *ctx;
217 ctx = perf_lock_task_context(task, &flags);
220 raw_spin_unlock_irqrestore(&ctx->lock, flags);
225 static void perf_unpin_context(struct perf_event_context *ctx)
229 raw_spin_lock_irqsave(&ctx->lock, flags);
231 raw_spin_unlock_irqrestore(&ctx->lock, flags);
235 static inline u64 perf_clock(void)
237 return cpu_clock(raw_smp_processor_id());
241 * Update the record of the current time in a context.
243 static void update_context_time(struct perf_event_context *ctx)
245 u64 now = perf_clock();
247 ctx->time += now - ctx->timestamp;
248 ctx->timestamp = now;
252 * Update the total_time_enabled and total_time_running fields for a event.
254 static void update_event_times(struct perf_event *event)
256 struct perf_event_context *ctx = event->ctx;
259 if (event->state < PERF_EVENT_STATE_INACTIVE ||
260 event->group_leader->state < PERF_EVENT_STATE_INACTIVE)
266 run_end = event->tstamp_stopped;
268 event->total_time_enabled = run_end - event->tstamp_enabled;
270 if (event->state == PERF_EVENT_STATE_INACTIVE)
271 run_end = event->tstamp_stopped;
275 event->total_time_running = run_end - event->tstamp_running;
278 static struct list_head *
279 ctx_group_list(struct perf_event *event, struct perf_event_context *ctx)
281 if (event->attr.pinned)
282 return &ctx->pinned_groups;
284 return &ctx->flexible_groups;
288 * Add a event from the lists for its context.
289 * Must be called with ctx->mutex and ctx->lock held.
292 list_add_event(struct perf_event *event, struct perf_event_context *ctx)
294 struct perf_event *group_leader = event->group_leader;
297 * Depending on whether it is a standalone or sibling event,
298 * add it straight to the context's event list, or to the group
299 * leader's sibling list:
301 if (group_leader == event) {
302 struct list_head *list;
304 if (is_software_event(event))
305 event->group_flags |= PERF_GROUP_SOFTWARE;
307 list = ctx_group_list(event, ctx);
308 list_add_tail(&event->group_entry, list);
310 if (group_leader->group_flags & PERF_GROUP_SOFTWARE &&
311 !is_software_event(event))
312 group_leader->group_flags &= ~PERF_GROUP_SOFTWARE;
314 list_add_tail(&event->group_entry, &group_leader->sibling_list);
315 group_leader->nr_siblings++;
318 list_add_rcu(&event->event_entry, &ctx->event_list);
320 if (event->attr.inherit_stat)
325 * Remove a event from the lists for its context.
326 * Must be called with ctx->mutex and ctx->lock held.
329 list_del_event(struct perf_event *event, struct perf_event_context *ctx)
331 struct perf_event *sibling, *tmp;
333 if (list_empty(&event->group_entry))
336 if (event->attr.inherit_stat)
339 list_del_init(&event->group_entry);
340 list_del_rcu(&event->event_entry);
342 if (event->group_leader != event)
343 event->group_leader->nr_siblings--;
345 update_event_times(event);
348 * If event was in error state, then keep it
349 * that way, otherwise bogus counts will be
350 * returned on read(). The only way to get out
351 * of error state is by explicit re-enabling
354 if (event->state > PERF_EVENT_STATE_OFF)
355 event->state = PERF_EVENT_STATE_OFF;
358 * If this was a group event with sibling events then
359 * upgrade the siblings to singleton events by adding them
360 * to the context list directly:
362 list_for_each_entry_safe(sibling, tmp, &event->sibling_list, group_entry) {
363 struct list_head *list;
365 list = ctx_group_list(event, ctx);
366 list_move_tail(&sibling->group_entry, list);
367 sibling->group_leader = sibling;
369 /* Inherit group flags from the previous leader */
370 sibling->group_flags = event->group_flags;
375 event_sched_out(struct perf_event *event,
376 struct perf_cpu_context *cpuctx,
377 struct perf_event_context *ctx)
379 if (event->state != PERF_EVENT_STATE_ACTIVE)
382 event->state = PERF_EVENT_STATE_INACTIVE;
383 if (event->pending_disable) {
384 event->pending_disable = 0;
385 event->state = PERF_EVENT_STATE_OFF;
387 event->tstamp_stopped = ctx->time;
388 event->pmu->disable(event);
391 if (!is_software_event(event))
392 cpuctx->active_oncpu--;
394 if (event->attr.exclusive || !cpuctx->active_oncpu)
395 cpuctx->exclusive = 0;
399 group_sched_out(struct perf_event *group_event,
400 struct perf_cpu_context *cpuctx,
401 struct perf_event_context *ctx)
403 struct perf_event *event;
405 if (group_event->state != PERF_EVENT_STATE_ACTIVE)
408 event_sched_out(group_event, cpuctx, ctx);
411 * Schedule out siblings (if any):
413 list_for_each_entry(event, &group_event->sibling_list, group_entry)
414 event_sched_out(event, cpuctx, ctx);
416 if (group_event->attr.exclusive)
417 cpuctx->exclusive = 0;
421 * Cross CPU call to remove a performance event
423 * We disable the event on the hardware level first. After that we
424 * remove it from the context list.
426 static void __perf_event_remove_from_context(void *info)
428 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
429 struct perf_event *event = info;
430 struct perf_event_context *ctx = event->ctx;
433 * If this is a task context, we need to check whether it is
434 * the current task context of this cpu. If not it has been
435 * scheduled out before the smp call arrived.
437 if (ctx->task && cpuctx->task_ctx != ctx)
440 raw_spin_lock(&ctx->lock);
442 * Protect the list operation against NMI by disabling the
443 * events on a global level.
447 event_sched_out(event, cpuctx, ctx);
449 list_del_event(event, ctx);
453 * Allow more per task events with respect to the
456 cpuctx->max_pertask =
457 min(perf_max_events - ctx->nr_events,
458 perf_max_events - perf_reserved_percpu);
462 raw_spin_unlock(&ctx->lock);
467 * Remove the event from a task's (or a CPU's) list of events.
469 * Must be called with ctx->mutex held.
471 * CPU events are removed with a smp call. For task events we only
472 * call when the task is on a CPU.
474 * If event->ctx is a cloned context, callers must make sure that
475 * every task struct that event->ctx->task could possibly point to
476 * remains valid. This is OK when called from perf_release since
477 * that only calls us on the top-level context, which can't be a clone.
478 * When called from perf_event_exit_task, it's OK because the
479 * context has been detached from its task.
481 static void perf_event_remove_from_context(struct perf_event *event)
483 struct perf_event_context *ctx = event->ctx;
484 struct task_struct *task = ctx->task;
488 * Per cpu events are removed via an smp call and
489 * the removal is always successful.
491 smp_call_function_single(event->cpu,
492 __perf_event_remove_from_context,
498 task_oncpu_function_call(task, __perf_event_remove_from_context,
501 raw_spin_lock_irq(&ctx->lock);
503 * If the context is active we need to retry the smp call.
505 if (ctx->nr_active && !list_empty(&event->group_entry)) {
506 raw_spin_unlock_irq(&ctx->lock);
511 * The lock prevents that this context is scheduled in so we
512 * can remove the event safely, if the call above did not
515 if (!list_empty(&event->group_entry))
516 list_del_event(event, ctx);
517 raw_spin_unlock_irq(&ctx->lock);
521 * Update total_time_enabled and total_time_running for all events in a group.
523 static void update_group_times(struct perf_event *leader)
525 struct perf_event *event;
527 update_event_times(leader);
528 list_for_each_entry(event, &leader->sibling_list, group_entry)
529 update_event_times(event);
533 * Cross CPU call to disable a performance event
535 static void __perf_event_disable(void *info)
537 struct perf_event *event = info;
538 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
539 struct perf_event_context *ctx = event->ctx;
542 * If this is a per-task event, need to check whether this
543 * event's task is the current task on this cpu.
545 if (ctx->task && cpuctx->task_ctx != ctx)
548 raw_spin_lock(&ctx->lock);
551 * If the event is on, turn it off.
552 * If it is in error state, leave it in error state.
554 if (event->state >= PERF_EVENT_STATE_INACTIVE) {
555 update_context_time(ctx);
556 update_group_times(event);
557 if (event == event->group_leader)
558 group_sched_out(event, cpuctx, ctx);
560 event_sched_out(event, cpuctx, ctx);
561 event->state = PERF_EVENT_STATE_OFF;
564 raw_spin_unlock(&ctx->lock);
570 * If event->ctx is a cloned context, callers must make sure that
571 * every task struct that event->ctx->task could possibly point to
572 * remains valid. This condition is satisifed when called through
573 * perf_event_for_each_child or perf_event_for_each because they
574 * hold the top-level event's child_mutex, so any descendant that
575 * goes to exit will block in sync_child_event.
576 * When called from perf_pending_event it's OK because event->ctx
577 * is the current context on this CPU and preemption is disabled,
578 * hence we can't get into perf_event_task_sched_out for this context.
580 void perf_event_disable(struct perf_event *event)
582 struct perf_event_context *ctx = event->ctx;
583 struct task_struct *task = ctx->task;
587 * Disable the event on the cpu that it's on
589 smp_call_function_single(event->cpu, __perf_event_disable,
595 task_oncpu_function_call(task, __perf_event_disable, event);
597 raw_spin_lock_irq(&ctx->lock);
599 * If the event is still active, we need to retry the cross-call.
601 if (event->state == PERF_EVENT_STATE_ACTIVE) {
602 raw_spin_unlock_irq(&ctx->lock);
607 * Since we have the lock this context can't be scheduled
608 * in, so we can change the state safely.
610 if (event->state == PERF_EVENT_STATE_INACTIVE) {
611 update_group_times(event);
612 event->state = PERF_EVENT_STATE_OFF;
615 raw_spin_unlock_irq(&ctx->lock);
619 event_sched_in(struct perf_event *event,
620 struct perf_cpu_context *cpuctx,
621 struct perf_event_context *ctx)
623 if (event->state <= PERF_EVENT_STATE_OFF)
626 event->state = PERF_EVENT_STATE_ACTIVE;
627 event->oncpu = smp_processor_id();
629 * The new state must be visible before we turn it on in the hardware:
633 if (event->pmu->enable(event)) {
634 event->state = PERF_EVENT_STATE_INACTIVE;
639 event->tstamp_running += ctx->time - event->tstamp_stopped;
641 if (!is_software_event(event))
642 cpuctx->active_oncpu++;
645 if (event->attr.exclusive)
646 cpuctx->exclusive = 1;
652 group_sched_in(struct perf_event *group_event,
653 struct perf_cpu_context *cpuctx,
654 struct perf_event_context *ctx)
656 struct perf_event *event, *partial_group;
659 if (group_event->state == PERF_EVENT_STATE_OFF)
662 ret = hw_perf_group_sched_in(group_event, cpuctx, ctx);
664 return ret < 0 ? ret : 0;
666 if (event_sched_in(group_event, cpuctx, ctx))
670 * Schedule in siblings as one group (if any):
672 list_for_each_entry(event, &group_event->sibling_list, group_entry) {
673 if (event_sched_in(event, cpuctx, ctx)) {
674 partial_group = event;
683 * Groups can be scheduled in as one unit only, so undo any
684 * partial group before returning:
686 list_for_each_entry(event, &group_event->sibling_list, group_entry) {
687 if (event == partial_group)
689 event_sched_out(event, cpuctx, ctx);
691 event_sched_out(group_event, cpuctx, ctx);
697 * Work out whether we can put this event group on the CPU now.
699 static int group_can_go_on(struct perf_event *event,
700 struct perf_cpu_context *cpuctx,
704 * Groups consisting entirely of software events can always go on.
706 if (event->group_flags & PERF_GROUP_SOFTWARE)
709 * If an exclusive group is already on, no other hardware
712 if (cpuctx->exclusive)
715 * If this group is exclusive and there are already
716 * events on the CPU, it can't go on.
718 if (event->attr.exclusive && cpuctx->active_oncpu)
721 * Otherwise, try to add it if all previous groups were able
727 static void add_event_to_ctx(struct perf_event *event,
728 struct perf_event_context *ctx)
730 list_add_event(event, ctx);
731 event->tstamp_enabled = ctx->time;
732 event->tstamp_running = ctx->time;
733 event->tstamp_stopped = ctx->time;
737 * Cross CPU call to install and enable a performance event
739 * Must be called with ctx->mutex held
741 static void __perf_install_in_context(void *info)
743 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
744 struct perf_event *event = info;
745 struct perf_event_context *ctx = event->ctx;
746 struct perf_event *leader = event->group_leader;
750 * If this is a task context, we need to check whether it is
751 * the current task context of this cpu. If not it has been
752 * scheduled out before the smp call arrived.
753 * Or possibly this is the right context but it isn't
754 * on this cpu because it had no events.
756 if (ctx->task && cpuctx->task_ctx != ctx) {
757 if (cpuctx->task_ctx || ctx->task != current)
759 cpuctx->task_ctx = ctx;
762 raw_spin_lock(&ctx->lock);
764 update_context_time(ctx);
767 * Protect the list operation against NMI by disabling the
768 * events on a global level. NOP for non NMI based events.
772 add_event_to_ctx(event, ctx);
774 if (event->cpu != -1 && event->cpu != smp_processor_id())
778 * Don't put the event on if it is disabled or if
779 * it is in a group and the group isn't on.
781 if (event->state != PERF_EVENT_STATE_INACTIVE ||
782 (leader != event && leader->state != PERF_EVENT_STATE_ACTIVE))
786 * An exclusive event can't go on if there are already active
787 * hardware events, and no hardware event can go on if there
788 * is already an exclusive event on.
790 if (!group_can_go_on(event, cpuctx, 1))
793 err = event_sched_in(event, cpuctx, ctx);
797 * This event couldn't go on. If it is in a group
798 * then we have to pull the whole group off.
799 * If the event group is pinned then put it in error state.
802 group_sched_out(leader, cpuctx, ctx);
803 if (leader->attr.pinned) {
804 update_group_times(leader);
805 leader->state = PERF_EVENT_STATE_ERROR;
809 if (!err && !ctx->task && cpuctx->max_pertask)
810 cpuctx->max_pertask--;
815 raw_spin_unlock(&ctx->lock);
819 * Attach a performance event to a context
821 * First we add the event to the list with the hardware enable bit
822 * in event->hw_config cleared.
824 * If the event is attached to a task which is on a CPU we use a smp
825 * call to enable it in the task context. The task might have been
826 * scheduled away, but we check this in the smp call again.
828 * Must be called with ctx->mutex held.
831 perf_install_in_context(struct perf_event_context *ctx,
832 struct perf_event *event,
835 struct task_struct *task = ctx->task;
839 * Per cpu events are installed via an smp call and
840 * the install is always successful.
842 smp_call_function_single(cpu, __perf_install_in_context,
848 task_oncpu_function_call(task, __perf_install_in_context,
851 raw_spin_lock_irq(&ctx->lock);
853 * we need to retry the smp call.
855 if (ctx->is_active && list_empty(&event->group_entry)) {
856 raw_spin_unlock_irq(&ctx->lock);
861 * The lock prevents that this context is scheduled in so we
862 * can add the event safely, if it the call above did not
865 if (list_empty(&event->group_entry))
866 add_event_to_ctx(event, ctx);
867 raw_spin_unlock_irq(&ctx->lock);
871 * Put a event into inactive state and update time fields.
872 * Enabling the leader of a group effectively enables all
873 * the group members that aren't explicitly disabled, so we
874 * have to update their ->tstamp_enabled also.
875 * Note: this works for group members as well as group leaders
876 * since the non-leader members' sibling_lists will be empty.
878 static void __perf_event_mark_enabled(struct perf_event *event,
879 struct perf_event_context *ctx)
881 struct perf_event *sub;
883 event->state = PERF_EVENT_STATE_INACTIVE;
884 event->tstamp_enabled = ctx->time - event->total_time_enabled;
885 list_for_each_entry(sub, &event->sibling_list, group_entry)
886 if (sub->state >= PERF_EVENT_STATE_INACTIVE)
887 sub->tstamp_enabled =
888 ctx->time - sub->total_time_enabled;
892 * Cross CPU call to enable a performance event
894 static void __perf_event_enable(void *info)
896 struct perf_event *event = info;
897 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
898 struct perf_event_context *ctx = event->ctx;
899 struct perf_event *leader = event->group_leader;
903 * If this is a per-task event, need to check whether this
904 * event's task is the current task on this cpu.
906 if (ctx->task && cpuctx->task_ctx != ctx) {
907 if (cpuctx->task_ctx || ctx->task != current)
909 cpuctx->task_ctx = ctx;
912 raw_spin_lock(&ctx->lock);
914 update_context_time(ctx);
916 if (event->state >= PERF_EVENT_STATE_INACTIVE)
918 __perf_event_mark_enabled(event, ctx);
920 if (event->cpu != -1 && event->cpu != smp_processor_id())
924 * If the event is in a group and isn't the group leader,
925 * then don't put it on unless the group is on.
927 if (leader != event && leader->state != PERF_EVENT_STATE_ACTIVE)
930 if (!group_can_go_on(event, cpuctx, 1)) {
935 err = group_sched_in(event, cpuctx, ctx);
937 err = event_sched_in(event, cpuctx, ctx);
943 * If this event can't go on and it's part of a
944 * group, then the whole group has to come off.
947 group_sched_out(leader, cpuctx, ctx);
948 if (leader->attr.pinned) {
949 update_group_times(leader);
950 leader->state = PERF_EVENT_STATE_ERROR;
955 raw_spin_unlock(&ctx->lock);
961 * If event->ctx is a cloned context, callers must make sure that
962 * every task struct that event->ctx->task could possibly point to
963 * remains valid. This condition is satisfied when called through
964 * perf_event_for_each_child or perf_event_for_each as described
965 * for perf_event_disable.
967 void perf_event_enable(struct perf_event *event)
969 struct perf_event_context *ctx = event->ctx;
970 struct task_struct *task = ctx->task;
974 * Enable the event on the cpu that it's on
976 smp_call_function_single(event->cpu, __perf_event_enable,
981 raw_spin_lock_irq(&ctx->lock);
982 if (event->state >= PERF_EVENT_STATE_INACTIVE)
986 * If the event is in error state, clear that first.
987 * That way, if we see the event in error state below, we
988 * know that it has gone back into error state, as distinct
989 * from the task having been scheduled away before the
990 * cross-call arrived.
992 if (event->state == PERF_EVENT_STATE_ERROR)
993 event->state = PERF_EVENT_STATE_OFF;
996 raw_spin_unlock_irq(&ctx->lock);
997 task_oncpu_function_call(task, __perf_event_enable, event);
999 raw_spin_lock_irq(&ctx->lock);
1002 * If the context is active and the event is still off,
1003 * we need to retry the cross-call.
1005 if (ctx->is_active && event->state == PERF_EVENT_STATE_OFF)
1009 * Since we have the lock this context can't be scheduled
1010 * in, so we can change the state safely.
1012 if (event->state == PERF_EVENT_STATE_OFF)
1013 __perf_event_mark_enabled(event, ctx);
1016 raw_spin_unlock_irq(&ctx->lock);
1019 static int perf_event_refresh(struct perf_event *event, int refresh)
1022 * not supported on inherited events
1024 if (event->attr.inherit)
1027 atomic_add(refresh, &event->event_limit);
1028 perf_event_enable(event);
1034 EVENT_FLEXIBLE = 0x1,
1036 EVENT_ALL = EVENT_FLEXIBLE | EVENT_PINNED,
1039 static void ctx_sched_out(struct perf_event_context *ctx,
1040 struct perf_cpu_context *cpuctx,
1041 enum event_type_t event_type)
1043 struct perf_event *event;
1045 raw_spin_lock(&ctx->lock);
1047 if (likely(!ctx->nr_events))
1049 update_context_time(ctx);
1052 if (!ctx->nr_active)
1055 if (event_type & EVENT_PINNED)
1056 list_for_each_entry(event, &ctx->pinned_groups, group_entry)
1057 group_sched_out(event, cpuctx, ctx);
1059 if (event_type & EVENT_FLEXIBLE)
1060 list_for_each_entry(event, &ctx->flexible_groups, group_entry)
1061 group_sched_out(event, cpuctx, ctx);
1066 raw_spin_unlock(&ctx->lock);
1070 * Test whether two contexts are equivalent, i.e. whether they
1071 * have both been cloned from the same version of the same context
1072 * and they both have the same number of enabled events.
1073 * If the number of enabled events is the same, then the set
1074 * of enabled events should be the same, because these are both
1075 * inherited contexts, therefore we can't access individual events
1076 * in them directly with an fd; we can only enable/disable all
1077 * events via prctl, or enable/disable all events in a family
1078 * via ioctl, which will have the same effect on both contexts.
1080 static int context_equiv(struct perf_event_context *ctx1,
1081 struct perf_event_context *ctx2)
1083 return ctx1->parent_ctx && ctx1->parent_ctx == ctx2->parent_ctx
1084 && ctx1->parent_gen == ctx2->parent_gen
1085 && !ctx1->pin_count && !ctx2->pin_count;
1088 static void __perf_event_sync_stat(struct perf_event *event,
1089 struct perf_event *next_event)
1093 if (!event->attr.inherit_stat)
1097 * Update the event value, we cannot use perf_event_read()
1098 * because we're in the middle of a context switch and have IRQs
1099 * disabled, which upsets smp_call_function_single(), however
1100 * we know the event must be on the current CPU, therefore we
1101 * don't need to use it.
1103 switch (event->state) {
1104 case PERF_EVENT_STATE_ACTIVE:
1105 event->pmu->read(event);
1108 case PERF_EVENT_STATE_INACTIVE:
1109 update_event_times(event);
1117 * In order to keep per-task stats reliable we need to flip the event
1118 * values when we flip the contexts.
1120 value = atomic64_read(&next_event->count);
1121 value = atomic64_xchg(&event->count, value);
1122 atomic64_set(&next_event->count, value);
1124 swap(event->total_time_enabled, next_event->total_time_enabled);
1125 swap(event->total_time_running, next_event->total_time_running);
1128 * Since we swizzled the values, update the user visible data too.
1130 perf_event_update_userpage(event);
1131 perf_event_update_userpage(next_event);
1134 #define list_next_entry(pos, member) \
1135 list_entry(pos->member.next, typeof(*pos), member)
1137 static void perf_event_sync_stat(struct perf_event_context *ctx,
1138 struct perf_event_context *next_ctx)
1140 struct perf_event *event, *next_event;
1145 update_context_time(ctx);
1147 event = list_first_entry(&ctx->event_list,
1148 struct perf_event, event_entry);
1150 next_event = list_first_entry(&next_ctx->event_list,
1151 struct perf_event, event_entry);
1153 while (&event->event_entry != &ctx->event_list &&
1154 &next_event->event_entry != &next_ctx->event_list) {
1156 __perf_event_sync_stat(event, next_event);
1158 event = list_next_entry(event, event_entry);
1159 next_event = list_next_entry(next_event, event_entry);
1164 * Called from scheduler to remove the events of the current task,
1165 * with interrupts disabled.
1167 * We stop each event and update the event value in event->count.
1169 * This does not protect us against NMI, but disable()
1170 * sets the disabled bit in the control field of event _before_
1171 * accessing the event control register. If a NMI hits, then it will
1172 * not restart the event.
1174 void perf_event_task_sched_out(struct task_struct *task,
1175 struct task_struct *next)
1177 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
1178 struct perf_event_context *ctx = task->perf_event_ctxp;
1179 struct perf_event_context *next_ctx;
1180 struct perf_event_context *parent;
1181 struct pt_regs *regs;
1184 regs = task_pt_regs(task);
1185 perf_sw_event(PERF_COUNT_SW_CONTEXT_SWITCHES, 1, 1, regs, 0);
1187 if (likely(!ctx || !cpuctx->task_ctx))
1191 parent = rcu_dereference(ctx->parent_ctx);
1192 next_ctx = next->perf_event_ctxp;
1193 if (parent && next_ctx &&
1194 rcu_dereference(next_ctx->parent_ctx) == parent) {
1196 * Looks like the two contexts are clones, so we might be
1197 * able to optimize the context switch. We lock both
1198 * contexts and check that they are clones under the
1199 * lock (including re-checking that neither has been
1200 * uncloned in the meantime). It doesn't matter which
1201 * order we take the locks because no other cpu could
1202 * be trying to lock both of these tasks.
1204 raw_spin_lock(&ctx->lock);
1205 raw_spin_lock_nested(&next_ctx->lock, SINGLE_DEPTH_NESTING);
1206 if (context_equiv(ctx, next_ctx)) {
1208 * XXX do we need a memory barrier of sorts
1209 * wrt to rcu_dereference() of perf_event_ctxp
1211 task->perf_event_ctxp = next_ctx;
1212 next->perf_event_ctxp = ctx;
1214 next_ctx->task = task;
1217 perf_event_sync_stat(ctx, next_ctx);
1219 raw_spin_unlock(&next_ctx->lock);
1220 raw_spin_unlock(&ctx->lock);
1225 ctx_sched_out(ctx, cpuctx, EVENT_ALL);
1226 cpuctx->task_ctx = NULL;
1230 static void task_ctx_sched_out(struct perf_event_context *ctx,
1231 enum event_type_t event_type)
1233 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
1235 if (!cpuctx->task_ctx)
1238 if (WARN_ON_ONCE(ctx != cpuctx->task_ctx))
1241 ctx_sched_out(ctx, cpuctx, event_type);
1242 cpuctx->task_ctx = NULL;
1246 * Called with IRQs disabled
1248 static void __perf_event_task_sched_out(struct perf_event_context *ctx)
1250 task_ctx_sched_out(ctx, EVENT_ALL);
1254 * Called with IRQs disabled
1256 static void cpu_ctx_sched_out(struct perf_cpu_context *cpuctx,
1257 enum event_type_t event_type)
1259 ctx_sched_out(&cpuctx->ctx, cpuctx, event_type);
1263 ctx_pinned_sched_in(struct perf_event_context *ctx,
1264 struct perf_cpu_context *cpuctx)
1266 struct perf_event *event;
1268 list_for_each_entry(event, &ctx->pinned_groups, group_entry) {
1269 if (event->state <= PERF_EVENT_STATE_OFF)
1271 if (event->cpu != -1 && event->cpu != smp_processor_id())
1274 if (group_can_go_on(event, cpuctx, 1))
1275 group_sched_in(event, cpuctx, ctx);
1278 * If this pinned group hasn't been scheduled,
1279 * put it in error state.
1281 if (event->state == PERF_EVENT_STATE_INACTIVE) {
1282 update_group_times(event);
1283 event->state = PERF_EVENT_STATE_ERROR;
1289 ctx_flexible_sched_in(struct perf_event_context *ctx,
1290 struct perf_cpu_context *cpuctx)
1292 struct perf_event *event;
1295 list_for_each_entry(event, &ctx->flexible_groups, group_entry) {
1296 /* Ignore events in OFF or ERROR state */
1297 if (event->state <= PERF_EVENT_STATE_OFF)
1300 * Listen to the 'cpu' scheduling filter constraint
1303 if (event->cpu != -1 && event->cpu != smp_processor_id())
1306 if (group_can_go_on(event, cpuctx, can_add_hw))
1307 if (group_sched_in(event, cpuctx, ctx))
1313 ctx_sched_in(struct perf_event_context *ctx,
1314 struct perf_cpu_context *cpuctx,
1315 enum event_type_t event_type)
1317 raw_spin_lock(&ctx->lock);
1319 if (likely(!ctx->nr_events))
1322 ctx->timestamp = perf_clock();
1327 * First go through the list and put on any pinned groups
1328 * in order to give them the best chance of going on.
1330 if (event_type & EVENT_PINNED)
1331 ctx_pinned_sched_in(ctx, cpuctx);
1333 /* Then walk through the lower prio flexible groups */
1334 if (event_type & EVENT_FLEXIBLE)
1335 ctx_flexible_sched_in(ctx, cpuctx);
1339 raw_spin_unlock(&ctx->lock);
1342 static void cpu_ctx_sched_in(struct perf_cpu_context *cpuctx,
1343 enum event_type_t event_type)
1345 struct perf_event_context *ctx = &cpuctx->ctx;
1347 ctx_sched_in(ctx, cpuctx, event_type);
1350 static void task_ctx_sched_in(struct task_struct *task,
1351 enum event_type_t event_type)
1353 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
1354 struct perf_event_context *ctx = task->perf_event_ctxp;
1358 if (cpuctx->task_ctx == ctx)
1360 ctx_sched_in(ctx, cpuctx, event_type);
1361 cpuctx->task_ctx = ctx;
1364 * Called from scheduler to add the events of the current task
1365 * with interrupts disabled.
1367 * We restore the event value and then enable it.
1369 * This does not protect us against NMI, but enable()
1370 * sets the enabled bit in the control field of event _before_
1371 * accessing the event control register. If a NMI hits, then it will
1372 * keep the event running.
1374 void perf_event_task_sched_in(struct task_struct *task)
1376 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
1377 struct perf_event_context *ctx = task->perf_event_ctxp;
1382 if (cpuctx->task_ctx == ctx)
1386 * We want to keep the following priority order:
1387 * cpu pinned (that don't need to move), task pinned,
1388 * cpu flexible, task flexible.
1390 cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE);
1392 ctx_sched_in(ctx, cpuctx, EVENT_PINNED);
1393 cpu_ctx_sched_in(cpuctx, EVENT_FLEXIBLE);
1394 ctx_sched_in(ctx, cpuctx, EVENT_FLEXIBLE);
1396 cpuctx->task_ctx = ctx;
1399 #define MAX_INTERRUPTS (~0ULL)
1401 static void perf_log_throttle(struct perf_event *event, int enable);
1403 static u64 perf_calculate_period(struct perf_event *event, u64 nsec, u64 count)
1405 u64 frequency = event->attr.sample_freq;
1406 u64 sec = NSEC_PER_SEC;
1407 u64 divisor, dividend;
1409 int count_fls, nsec_fls, frequency_fls, sec_fls;
1411 count_fls = fls64(count);
1412 nsec_fls = fls64(nsec);
1413 frequency_fls = fls64(frequency);
1417 * We got @count in @nsec, with a target of sample_freq HZ
1418 * the target period becomes:
1421 * period = -------------------
1422 * @nsec * sample_freq
1427 * Reduce accuracy by one bit such that @a and @b converge
1428 * to a similar magnitude.
1430 #define REDUCE_FLS(a, b) \
1432 if (a##_fls > b##_fls) { \
1442 * Reduce accuracy until either term fits in a u64, then proceed with
1443 * the other, so that finally we can do a u64/u64 division.
1445 while (count_fls + sec_fls > 64 && nsec_fls + frequency_fls > 64) {
1446 REDUCE_FLS(nsec, frequency);
1447 REDUCE_FLS(sec, count);
1450 if (count_fls + sec_fls > 64) {
1451 divisor = nsec * frequency;
1453 while (count_fls + sec_fls > 64) {
1454 REDUCE_FLS(count, sec);
1458 dividend = count * sec;
1460 dividend = count * sec;
1462 while (nsec_fls + frequency_fls > 64) {
1463 REDUCE_FLS(nsec, frequency);
1467 divisor = nsec * frequency;
1470 return div64_u64(dividend, divisor);
1473 static void perf_event_stop(struct perf_event *event)
1475 if (!event->pmu->stop)
1476 return event->pmu->disable(event);
1478 return event->pmu->stop(event);
1481 static int perf_event_start(struct perf_event *event)
1483 if (!event->pmu->start)
1484 return event->pmu->enable(event);
1486 return event->pmu->start(event);
1489 static void perf_adjust_period(struct perf_event *event, u64 nsec, u64 count)
1491 struct hw_perf_event *hwc = &event->hw;
1492 u64 period, sample_period;
1495 period = perf_calculate_period(event, nsec, count);
1497 delta = (s64)(period - hwc->sample_period);
1498 delta = (delta + 7) / 8; /* low pass filter */
1500 sample_period = hwc->sample_period + delta;
1505 hwc->sample_period = sample_period;
1507 if (atomic64_read(&hwc->period_left) > 8*sample_period) {
1509 perf_event_stop(event);
1510 atomic64_set(&hwc->period_left, 0);
1511 perf_event_start(event);
1516 static void perf_ctx_adjust_freq(struct perf_event_context *ctx)
1518 struct perf_event *event;
1519 struct hw_perf_event *hwc;
1520 u64 interrupts, now;
1523 raw_spin_lock(&ctx->lock);
1524 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
1525 if (event->state != PERF_EVENT_STATE_ACTIVE)
1528 if (event->cpu != -1 && event->cpu != smp_processor_id())
1533 interrupts = hwc->interrupts;
1534 hwc->interrupts = 0;
1537 * unthrottle events on the tick
1539 if (interrupts == MAX_INTERRUPTS) {
1540 perf_log_throttle(event, 1);
1541 event->pmu->unthrottle(event);
1544 if (!event->attr.freq || !event->attr.sample_freq)
1547 event->pmu->read(event);
1548 now = atomic64_read(&event->count);
1549 delta = now - hwc->freq_count_stamp;
1550 hwc->freq_count_stamp = now;
1553 perf_adjust_period(event, TICK_NSEC, delta);
1555 raw_spin_unlock(&ctx->lock);
1559 * Round-robin a context's events:
1561 static void rotate_ctx(struct perf_event_context *ctx)
1563 if (!ctx->nr_events)
1566 raw_spin_lock(&ctx->lock);
1568 /* Rotate the first entry last of non-pinned groups */
1569 list_rotate_left(&ctx->flexible_groups);
1571 raw_spin_unlock(&ctx->lock);
1574 void perf_event_task_tick(struct task_struct *curr)
1576 struct perf_cpu_context *cpuctx;
1577 struct perf_event_context *ctx;
1579 if (!atomic_read(&nr_events))
1582 cpuctx = &__get_cpu_var(perf_cpu_context);
1583 ctx = curr->perf_event_ctxp;
1587 perf_ctx_adjust_freq(&cpuctx->ctx);
1589 perf_ctx_adjust_freq(ctx);
1591 cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE);
1593 task_ctx_sched_out(ctx, EVENT_FLEXIBLE);
1595 rotate_ctx(&cpuctx->ctx);
1599 cpu_ctx_sched_in(cpuctx, EVENT_FLEXIBLE);
1601 task_ctx_sched_in(curr, EVENT_FLEXIBLE);
1606 static int event_enable_on_exec(struct perf_event *event,
1607 struct perf_event_context *ctx)
1609 if (!event->attr.enable_on_exec)
1612 event->attr.enable_on_exec = 0;
1613 if (event->state >= PERF_EVENT_STATE_INACTIVE)
1616 __perf_event_mark_enabled(event, ctx);
1622 * Enable all of a task's events that have been marked enable-on-exec.
1623 * This expects task == current.
1625 static void perf_event_enable_on_exec(struct task_struct *task)
1627 struct perf_event_context *ctx;
1628 struct perf_event *event;
1629 unsigned long flags;
1633 local_irq_save(flags);
1634 ctx = task->perf_event_ctxp;
1635 if (!ctx || !ctx->nr_events)
1638 __perf_event_task_sched_out(ctx);
1640 raw_spin_lock(&ctx->lock);
1642 list_for_each_entry(event, &ctx->pinned_groups, group_entry) {
1643 ret = event_enable_on_exec(event, ctx);
1648 list_for_each_entry(event, &ctx->flexible_groups, group_entry) {
1649 ret = event_enable_on_exec(event, ctx);
1655 * Unclone this context if we enabled any event.
1660 raw_spin_unlock(&ctx->lock);
1662 perf_event_task_sched_in(task);
1664 local_irq_restore(flags);
1668 * Cross CPU call to read the hardware event
1670 static void __perf_event_read(void *info)
1672 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
1673 struct perf_event *event = info;
1674 struct perf_event_context *ctx = event->ctx;
1677 * If this is a task context, we need to check whether it is
1678 * the current task context of this cpu. If not it has been
1679 * scheduled out before the smp call arrived. In that case
1680 * event->count would have been updated to a recent sample
1681 * when the event was scheduled out.
1683 if (ctx->task && cpuctx->task_ctx != ctx)
1686 raw_spin_lock(&ctx->lock);
1687 update_context_time(ctx);
1688 update_event_times(event);
1689 raw_spin_unlock(&ctx->lock);
1691 event->pmu->read(event);
1694 static u64 perf_event_read(struct perf_event *event)
1697 * If event is enabled and currently active on a CPU, update the
1698 * value in the event structure:
1700 if (event->state == PERF_EVENT_STATE_ACTIVE) {
1701 smp_call_function_single(event->oncpu,
1702 __perf_event_read, event, 1);
1703 } else if (event->state == PERF_EVENT_STATE_INACTIVE) {
1704 struct perf_event_context *ctx = event->ctx;
1705 unsigned long flags;
1707 raw_spin_lock_irqsave(&ctx->lock, flags);
1708 update_context_time(ctx);
1709 update_event_times(event);
1710 raw_spin_unlock_irqrestore(&ctx->lock, flags);
1713 return atomic64_read(&event->count);
1717 * Initialize the perf_event context in a task_struct:
1720 __perf_event_init_context(struct perf_event_context *ctx,
1721 struct task_struct *task)
1723 raw_spin_lock_init(&ctx->lock);
1724 mutex_init(&ctx->mutex);
1725 INIT_LIST_HEAD(&ctx->pinned_groups);
1726 INIT_LIST_HEAD(&ctx->flexible_groups);
1727 INIT_LIST_HEAD(&ctx->event_list);
1728 atomic_set(&ctx->refcount, 1);
1732 static struct perf_event_context *find_get_context(pid_t pid, int cpu)
1734 struct perf_event_context *ctx;
1735 struct perf_cpu_context *cpuctx;
1736 struct task_struct *task;
1737 unsigned long flags;
1740 if (pid == -1 && cpu != -1) {
1741 /* Must be root to operate on a CPU event: */
1742 if (perf_paranoid_cpu() && !capable(CAP_SYS_ADMIN))
1743 return ERR_PTR(-EACCES);
1745 if (cpu < 0 || cpu >= nr_cpumask_bits)
1746 return ERR_PTR(-EINVAL);
1749 * We could be clever and allow to attach a event to an
1750 * offline CPU and activate it when the CPU comes up, but
1753 if (!cpu_online(cpu))
1754 return ERR_PTR(-ENODEV);
1756 cpuctx = &per_cpu(perf_cpu_context, cpu);
1767 task = find_task_by_vpid(pid);
1769 get_task_struct(task);
1773 return ERR_PTR(-ESRCH);
1776 * Can't attach events to a dying task.
1779 if (task->flags & PF_EXITING)
1782 /* Reuse ptrace permission checks for now. */
1784 if (!ptrace_may_access(task, PTRACE_MODE_READ))
1788 ctx = perf_lock_task_context(task, &flags);
1791 raw_spin_unlock_irqrestore(&ctx->lock, flags);
1795 ctx = kzalloc(sizeof(struct perf_event_context), GFP_KERNEL);
1799 __perf_event_init_context(ctx, task);
1801 if (cmpxchg(&task->perf_event_ctxp, NULL, ctx)) {
1803 * We raced with some other task; use
1804 * the context they set.
1809 get_task_struct(task);
1812 put_task_struct(task);
1816 put_task_struct(task);
1817 return ERR_PTR(err);
1820 static void perf_event_free_filter(struct perf_event *event);
1822 static void free_event_rcu(struct rcu_head *head)
1824 struct perf_event *event;
1826 event = container_of(head, struct perf_event, rcu_head);
1828 put_pid_ns(event->ns);
1829 perf_event_free_filter(event);
1833 static void perf_pending_sync(struct perf_event *event);
1835 static void free_event(struct perf_event *event)
1837 perf_pending_sync(event);
1839 if (!event->parent) {
1840 atomic_dec(&nr_events);
1841 if (event->attr.mmap)
1842 atomic_dec(&nr_mmap_events);
1843 if (event->attr.comm)
1844 atomic_dec(&nr_comm_events);
1845 if (event->attr.task)
1846 atomic_dec(&nr_task_events);
1849 if (event->output) {
1850 fput(event->output->filp);
1851 event->output = NULL;
1855 event->destroy(event);
1857 put_ctx(event->ctx);
1858 call_rcu(&event->rcu_head, free_event_rcu);
1861 int perf_event_release_kernel(struct perf_event *event)
1863 struct perf_event_context *ctx = event->ctx;
1865 WARN_ON_ONCE(ctx->parent_ctx);
1866 mutex_lock(&ctx->mutex);
1867 perf_event_remove_from_context(event);
1868 mutex_unlock(&ctx->mutex);
1870 mutex_lock(&event->owner->perf_event_mutex);
1871 list_del_init(&event->owner_entry);
1872 mutex_unlock(&event->owner->perf_event_mutex);
1873 put_task_struct(event->owner);
1879 EXPORT_SYMBOL_GPL(perf_event_release_kernel);
1882 * Called when the last reference to the file is gone.
1884 static int perf_release(struct inode *inode, struct file *file)
1886 struct perf_event *event = file->private_data;
1888 file->private_data = NULL;
1890 return perf_event_release_kernel(event);
1893 static int perf_event_read_size(struct perf_event *event)
1895 int entry = sizeof(u64); /* value */
1899 if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1900 size += sizeof(u64);
1902 if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1903 size += sizeof(u64);
1905 if (event->attr.read_format & PERF_FORMAT_ID)
1906 entry += sizeof(u64);
1908 if (event->attr.read_format & PERF_FORMAT_GROUP) {
1909 nr += event->group_leader->nr_siblings;
1910 size += sizeof(u64);
1918 u64 perf_event_read_value(struct perf_event *event, u64 *enabled, u64 *running)
1920 struct perf_event *child;
1926 mutex_lock(&event->child_mutex);
1927 total += perf_event_read(event);
1928 *enabled += event->total_time_enabled +
1929 atomic64_read(&event->child_total_time_enabled);
1930 *running += event->total_time_running +
1931 atomic64_read(&event->child_total_time_running);
1933 list_for_each_entry(child, &event->child_list, child_list) {
1934 total += perf_event_read(child);
1935 *enabled += child->total_time_enabled;
1936 *running += child->total_time_running;
1938 mutex_unlock(&event->child_mutex);
1942 EXPORT_SYMBOL_GPL(perf_event_read_value);
1944 static int perf_event_read_group(struct perf_event *event,
1945 u64 read_format, char __user *buf)
1947 struct perf_event *leader = event->group_leader, *sub;
1948 int n = 0, size = 0, ret = -EFAULT;
1949 struct perf_event_context *ctx = leader->ctx;
1951 u64 count, enabled, running;
1953 mutex_lock(&ctx->mutex);
1954 count = perf_event_read_value(leader, &enabled, &running);
1956 values[n++] = 1 + leader->nr_siblings;
1957 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1958 values[n++] = enabled;
1959 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1960 values[n++] = running;
1961 values[n++] = count;
1962 if (read_format & PERF_FORMAT_ID)
1963 values[n++] = primary_event_id(leader);
1965 size = n * sizeof(u64);
1967 if (copy_to_user(buf, values, size))
1972 list_for_each_entry(sub, &leader->sibling_list, group_entry) {
1975 values[n++] = perf_event_read_value(sub, &enabled, &running);
1976 if (read_format & PERF_FORMAT_ID)
1977 values[n++] = primary_event_id(sub);
1979 size = n * sizeof(u64);
1981 if (copy_to_user(buf + ret, values, size)) {
1989 mutex_unlock(&ctx->mutex);
1994 static int perf_event_read_one(struct perf_event *event,
1995 u64 read_format, char __user *buf)
1997 u64 enabled, running;
2001 values[n++] = perf_event_read_value(event, &enabled, &running);
2002 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
2003 values[n++] = enabled;
2004 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
2005 values[n++] = running;
2006 if (read_format & PERF_FORMAT_ID)
2007 values[n++] = primary_event_id(event);
2009 if (copy_to_user(buf, values, n * sizeof(u64)))
2012 return n * sizeof(u64);
2016 * Read the performance event - simple non blocking version for now
2019 perf_read_hw(struct perf_event *event, char __user *buf, size_t count)
2021 u64 read_format = event->attr.read_format;
2025 * Return end-of-file for a read on a event that is in
2026 * error state (i.e. because it was pinned but it couldn't be
2027 * scheduled on to the CPU at some point).
2029 if (event->state == PERF_EVENT_STATE_ERROR)
2032 if (count < perf_event_read_size(event))
2035 WARN_ON_ONCE(event->ctx->parent_ctx);
2036 if (read_format & PERF_FORMAT_GROUP)
2037 ret = perf_event_read_group(event, read_format, buf);
2039 ret = perf_event_read_one(event, read_format, buf);
2045 perf_read(struct file *file, char __user *buf, size_t count, loff_t *ppos)
2047 struct perf_event *event = file->private_data;
2049 return perf_read_hw(event, buf, count);
2052 static unsigned int perf_poll(struct file *file, poll_table *wait)
2054 struct perf_event *event = file->private_data;
2055 struct perf_mmap_data *data;
2056 unsigned int events = POLL_HUP;
2059 data = rcu_dereference(event->data);
2061 events = atomic_xchg(&data->poll, 0);
2064 poll_wait(file, &event->waitq, wait);
2069 static void perf_event_reset(struct perf_event *event)
2071 (void)perf_event_read(event);
2072 atomic64_set(&event->count, 0);
2073 perf_event_update_userpage(event);
2077 * Holding the top-level event's child_mutex means that any
2078 * descendant process that has inherited this event will block
2079 * in sync_child_event if it goes to exit, thus satisfying the
2080 * task existence requirements of perf_event_enable/disable.
2082 static void perf_event_for_each_child(struct perf_event *event,
2083 void (*func)(struct perf_event *))
2085 struct perf_event *child;
2087 WARN_ON_ONCE(event->ctx->parent_ctx);
2088 mutex_lock(&event->child_mutex);
2090 list_for_each_entry(child, &event->child_list, child_list)
2092 mutex_unlock(&event->child_mutex);
2095 static void perf_event_for_each(struct perf_event *event,
2096 void (*func)(struct perf_event *))
2098 struct perf_event_context *ctx = event->ctx;
2099 struct perf_event *sibling;
2101 WARN_ON_ONCE(ctx->parent_ctx);
2102 mutex_lock(&ctx->mutex);
2103 event = event->group_leader;
2105 perf_event_for_each_child(event, func);
2107 list_for_each_entry(sibling, &event->sibling_list, group_entry)
2108 perf_event_for_each_child(event, func);
2109 mutex_unlock(&ctx->mutex);
2112 static int perf_event_period(struct perf_event *event, u64 __user *arg)
2114 struct perf_event_context *ctx = event->ctx;
2119 if (!event->attr.sample_period)
2122 size = copy_from_user(&value, arg, sizeof(value));
2123 if (size != sizeof(value))
2129 raw_spin_lock_irq(&ctx->lock);
2130 if (event->attr.freq) {
2131 if (value > sysctl_perf_event_sample_rate) {
2136 event->attr.sample_freq = value;
2138 event->attr.sample_period = value;
2139 event->hw.sample_period = value;
2142 raw_spin_unlock_irq(&ctx->lock);
2147 static int perf_event_set_output(struct perf_event *event, int output_fd);
2148 static int perf_event_set_filter(struct perf_event *event, void __user *arg);
2150 static long perf_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
2152 struct perf_event *event = file->private_data;
2153 void (*func)(struct perf_event *);
2157 case PERF_EVENT_IOC_ENABLE:
2158 func = perf_event_enable;
2160 case PERF_EVENT_IOC_DISABLE:
2161 func = perf_event_disable;
2163 case PERF_EVENT_IOC_RESET:
2164 func = perf_event_reset;
2167 case PERF_EVENT_IOC_REFRESH:
2168 return perf_event_refresh(event, arg);
2170 case PERF_EVENT_IOC_PERIOD:
2171 return perf_event_period(event, (u64 __user *)arg);
2173 case PERF_EVENT_IOC_SET_OUTPUT:
2174 return perf_event_set_output(event, arg);
2176 case PERF_EVENT_IOC_SET_FILTER:
2177 return perf_event_set_filter(event, (void __user *)arg);
2183 if (flags & PERF_IOC_FLAG_GROUP)
2184 perf_event_for_each(event, func);
2186 perf_event_for_each_child(event, func);
2191 int perf_event_task_enable(void)
2193 struct perf_event *event;
2195 mutex_lock(¤t->perf_event_mutex);
2196 list_for_each_entry(event, ¤t->perf_event_list, owner_entry)
2197 perf_event_for_each_child(event, perf_event_enable);
2198 mutex_unlock(¤t->perf_event_mutex);
2203 int perf_event_task_disable(void)
2205 struct perf_event *event;
2207 mutex_lock(¤t->perf_event_mutex);
2208 list_for_each_entry(event, ¤t->perf_event_list, owner_entry)
2209 perf_event_for_each_child(event, perf_event_disable);
2210 mutex_unlock(¤t->perf_event_mutex);
2215 #ifndef PERF_EVENT_INDEX_OFFSET
2216 # define PERF_EVENT_INDEX_OFFSET 0
2219 static int perf_event_index(struct perf_event *event)
2221 if (event->state != PERF_EVENT_STATE_ACTIVE)
2224 return event->hw.idx + 1 - PERF_EVENT_INDEX_OFFSET;
2228 * Callers need to ensure there can be no nesting of this function, otherwise
2229 * the seqlock logic goes bad. We can not serialize this because the arch
2230 * code calls this from NMI context.
2232 void perf_event_update_userpage(struct perf_event *event)
2234 struct perf_event_mmap_page *userpg;
2235 struct perf_mmap_data *data;
2238 data = rcu_dereference(event->data);
2242 userpg = data->user_page;
2245 * Disable preemption so as to not let the corresponding user-space
2246 * spin too long if we get preempted.
2251 userpg->index = perf_event_index(event);
2252 userpg->offset = atomic64_read(&event->count);
2253 if (event->state == PERF_EVENT_STATE_ACTIVE)
2254 userpg->offset -= atomic64_read(&event->hw.prev_count);
2256 userpg->time_enabled = event->total_time_enabled +
2257 atomic64_read(&event->child_total_time_enabled);
2259 userpg->time_running = event->total_time_running +
2260 atomic64_read(&event->child_total_time_running);
2269 static unsigned long perf_data_size(struct perf_mmap_data *data)
2271 return data->nr_pages << (PAGE_SHIFT + data->data_order);
2274 #ifndef CONFIG_PERF_USE_VMALLOC
2277 * Back perf_mmap() with regular GFP_KERNEL-0 pages.
2280 static struct page *
2281 perf_mmap_to_page(struct perf_mmap_data *data, unsigned long pgoff)
2283 if (pgoff > data->nr_pages)
2287 return virt_to_page(data->user_page);
2289 return virt_to_page(data->data_pages[pgoff - 1]);
2292 static struct perf_mmap_data *
2293 perf_mmap_data_alloc(struct perf_event *event, int nr_pages)
2295 struct perf_mmap_data *data;
2299 WARN_ON(atomic_read(&event->mmap_count));
2301 size = sizeof(struct perf_mmap_data);
2302 size += nr_pages * sizeof(void *);
2304 data = kzalloc(size, GFP_KERNEL);
2308 data->user_page = (void *)get_zeroed_page(GFP_KERNEL);
2309 if (!data->user_page)
2310 goto fail_user_page;
2312 for (i = 0; i < nr_pages; i++) {
2313 data->data_pages[i] = (void *)get_zeroed_page(GFP_KERNEL);
2314 if (!data->data_pages[i])
2315 goto fail_data_pages;
2318 data->data_order = 0;
2319 data->nr_pages = nr_pages;
2324 for (i--; i >= 0; i--)
2325 free_page((unsigned long)data->data_pages[i]);
2327 free_page((unsigned long)data->user_page);
2336 static void perf_mmap_free_page(unsigned long addr)
2338 struct page *page = virt_to_page((void *)addr);
2340 page->mapping = NULL;
2344 static void perf_mmap_data_free(struct perf_mmap_data *data)
2348 perf_mmap_free_page((unsigned long)data->user_page);
2349 for (i = 0; i < data->nr_pages; i++)
2350 perf_mmap_free_page((unsigned long)data->data_pages[i]);
2357 * Back perf_mmap() with vmalloc memory.
2359 * Required for architectures that have d-cache aliasing issues.
2362 static struct page *
2363 perf_mmap_to_page(struct perf_mmap_data *data, unsigned long pgoff)
2365 if (pgoff > (1UL << data->data_order))
2368 return vmalloc_to_page((void *)data->user_page + pgoff * PAGE_SIZE);
2371 static void perf_mmap_unmark_page(void *addr)
2373 struct page *page = vmalloc_to_page(addr);
2375 page->mapping = NULL;
2378 static void perf_mmap_data_free_work(struct work_struct *work)
2380 struct perf_mmap_data *data;
2384 data = container_of(work, struct perf_mmap_data, work);
2385 nr = 1 << data->data_order;
2387 base = data->user_page;
2388 for (i = 0; i < nr + 1; i++)
2389 perf_mmap_unmark_page(base + (i * PAGE_SIZE));
2395 static void perf_mmap_data_free(struct perf_mmap_data *data)
2397 schedule_work(&data->work);
2400 static struct perf_mmap_data *
2401 perf_mmap_data_alloc(struct perf_event *event, int nr_pages)
2403 struct perf_mmap_data *data;
2407 WARN_ON(atomic_read(&event->mmap_count));
2409 size = sizeof(struct perf_mmap_data);
2410 size += sizeof(void *);
2412 data = kzalloc(size, GFP_KERNEL);
2416 INIT_WORK(&data->work, perf_mmap_data_free_work);
2418 all_buf = vmalloc_user((nr_pages + 1) * PAGE_SIZE);
2422 data->user_page = all_buf;
2423 data->data_pages[0] = all_buf + PAGE_SIZE;
2424 data->data_order = ilog2(nr_pages);
2438 static int perf_mmap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
2440 struct perf_event *event = vma->vm_file->private_data;
2441 struct perf_mmap_data *data;
2442 int ret = VM_FAULT_SIGBUS;
2444 if (vmf->flags & FAULT_FLAG_MKWRITE) {
2445 if (vmf->pgoff == 0)
2451 data = rcu_dereference(event->data);
2455 if (vmf->pgoff && (vmf->flags & FAULT_FLAG_WRITE))
2458 vmf->page = perf_mmap_to_page(data, vmf->pgoff);
2462 get_page(vmf->page);
2463 vmf->page->mapping = vma->vm_file->f_mapping;
2464 vmf->page->index = vmf->pgoff;
2474 perf_mmap_data_init(struct perf_event *event, struct perf_mmap_data *data)
2476 long max_size = perf_data_size(data);
2478 atomic_set(&data->lock, -1);
2480 if (event->attr.watermark) {
2481 data->watermark = min_t(long, max_size,
2482 event->attr.wakeup_watermark);
2485 if (!data->watermark)
2486 data->watermark = max_size / 2;
2489 rcu_assign_pointer(event->data, data);
2492 static void perf_mmap_data_free_rcu(struct rcu_head *rcu_head)
2494 struct perf_mmap_data *data;
2496 data = container_of(rcu_head, struct perf_mmap_data, rcu_head);
2497 perf_mmap_data_free(data);
2500 static void perf_mmap_data_release(struct perf_event *event)
2502 struct perf_mmap_data *data = event->data;
2504 WARN_ON(atomic_read(&event->mmap_count));
2506 rcu_assign_pointer(event->data, NULL);
2507 call_rcu(&data->rcu_head, perf_mmap_data_free_rcu);
2510 static void perf_mmap_open(struct vm_area_struct *vma)
2512 struct perf_event *event = vma->vm_file->private_data;
2514 atomic_inc(&event->mmap_count);
2517 static void perf_mmap_close(struct vm_area_struct *vma)
2519 struct perf_event *event = vma->vm_file->private_data;
2521 WARN_ON_ONCE(event->ctx->parent_ctx);
2522 if (atomic_dec_and_mutex_lock(&event->mmap_count, &event->mmap_mutex)) {
2523 unsigned long size = perf_data_size(event->data);
2524 struct user_struct *user = current_user();
2526 atomic_long_sub((size >> PAGE_SHIFT) + 1, &user->locked_vm);
2527 vma->vm_mm->locked_vm -= event->data->nr_locked;
2528 perf_mmap_data_release(event);
2529 mutex_unlock(&event->mmap_mutex);
2533 static const struct vm_operations_struct perf_mmap_vmops = {
2534 .open = perf_mmap_open,
2535 .close = perf_mmap_close,
2536 .fault = perf_mmap_fault,
2537 .page_mkwrite = perf_mmap_fault,
2540 static int perf_mmap(struct file *file, struct vm_area_struct *vma)
2542 struct perf_event *event = file->private_data;
2543 unsigned long user_locked, user_lock_limit;
2544 struct user_struct *user = current_user();
2545 unsigned long locked, lock_limit;
2546 struct perf_mmap_data *data;
2547 unsigned long vma_size;
2548 unsigned long nr_pages;
2549 long user_extra, extra;
2552 if (!(vma->vm_flags & VM_SHARED))
2555 vma_size = vma->vm_end - vma->vm_start;
2556 nr_pages = (vma_size / PAGE_SIZE) - 1;
2559 * If we have data pages ensure they're a power-of-two number, so we
2560 * can do bitmasks instead of modulo.
2562 if (nr_pages != 0 && !is_power_of_2(nr_pages))
2565 if (vma_size != PAGE_SIZE * (1 + nr_pages))
2568 if (vma->vm_pgoff != 0)
2571 WARN_ON_ONCE(event->ctx->parent_ctx);
2572 mutex_lock(&event->mmap_mutex);
2573 if (event->output) {
2578 if (atomic_inc_not_zero(&event->mmap_count)) {
2579 if (nr_pages != event->data->nr_pages)
2584 user_extra = nr_pages + 1;
2585 user_lock_limit = sysctl_perf_event_mlock >> (PAGE_SHIFT - 10);
2588 * Increase the limit linearly with more CPUs:
2590 user_lock_limit *= num_online_cpus();
2592 user_locked = atomic_long_read(&user->locked_vm) + user_extra;
2595 if (user_locked > user_lock_limit)
2596 extra = user_locked - user_lock_limit;
2598 lock_limit = rlimit(RLIMIT_MEMLOCK);
2599 lock_limit >>= PAGE_SHIFT;
2600 locked = vma->vm_mm->locked_vm + extra;
2602 if ((locked > lock_limit) && perf_paranoid_tracepoint_raw() &&
2603 !capable(CAP_IPC_LOCK)) {
2608 WARN_ON(event->data);
2610 data = perf_mmap_data_alloc(event, nr_pages);
2616 perf_mmap_data_init(event, data);
2618 atomic_set(&event->mmap_count, 1);
2619 atomic_long_add(user_extra, &user->locked_vm);
2620 vma->vm_mm->locked_vm += extra;
2621 event->data->nr_locked = extra;
2622 if (vma->vm_flags & VM_WRITE)
2623 event->data->writable = 1;
2626 mutex_unlock(&event->mmap_mutex);
2628 vma->vm_flags |= VM_RESERVED;
2629 vma->vm_ops = &perf_mmap_vmops;
2634 static int perf_fasync(int fd, struct file *filp, int on)
2636 struct inode *inode = filp->f_path.dentry->d_inode;
2637 struct perf_event *event = filp->private_data;
2640 mutex_lock(&inode->i_mutex);
2641 retval = fasync_helper(fd, filp, on, &event->fasync);
2642 mutex_unlock(&inode->i_mutex);
2650 static const struct file_operations perf_fops = {
2651 .release = perf_release,
2654 .unlocked_ioctl = perf_ioctl,
2655 .compat_ioctl = perf_ioctl,
2657 .fasync = perf_fasync,
2663 * If there's data, ensure we set the poll() state and publish everything
2664 * to user-space before waking everybody up.
2667 void perf_event_wakeup(struct perf_event *event)
2669 wake_up_all(&event->waitq);
2671 if (event->pending_kill) {
2672 kill_fasync(&event->fasync, SIGIO, event->pending_kill);
2673 event->pending_kill = 0;
2680 * Handle the case where we need to wakeup up from NMI (or rq->lock) context.
2682 * The NMI bit means we cannot possibly take locks. Therefore, maintain a
2683 * single linked list and use cmpxchg() to add entries lockless.
2686 static void perf_pending_event(struct perf_pending_entry *entry)
2688 struct perf_event *event = container_of(entry,
2689 struct perf_event, pending);
2691 if (event->pending_disable) {
2692 event->pending_disable = 0;
2693 __perf_event_disable(event);
2696 if (event->pending_wakeup) {
2697 event->pending_wakeup = 0;
2698 perf_event_wakeup(event);
2702 #define PENDING_TAIL ((struct perf_pending_entry *)-1UL)
2704 static DEFINE_PER_CPU(struct perf_pending_entry *, perf_pending_head) = {
2708 static void perf_pending_queue(struct perf_pending_entry *entry,
2709 void (*func)(struct perf_pending_entry *))
2711 struct perf_pending_entry **head;
2713 if (cmpxchg(&entry->next, NULL, PENDING_TAIL) != NULL)
2718 head = &get_cpu_var(perf_pending_head);
2721 entry->next = *head;
2722 } while (cmpxchg(head, entry->next, entry) != entry->next);
2724 set_perf_event_pending();
2726 put_cpu_var(perf_pending_head);
2729 static int __perf_pending_run(void)
2731 struct perf_pending_entry *list;
2734 list = xchg(&__get_cpu_var(perf_pending_head), PENDING_TAIL);
2735 while (list != PENDING_TAIL) {
2736 void (*func)(struct perf_pending_entry *);
2737 struct perf_pending_entry *entry = list;
2744 * Ensure we observe the unqueue before we issue the wakeup,
2745 * so that we won't be waiting forever.
2746 * -- see perf_not_pending().
2757 static inline int perf_not_pending(struct perf_event *event)
2760 * If we flush on whatever cpu we run, there is a chance we don't
2764 __perf_pending_run();
2768 * Ensure we see the proper queue state before going to sleep
2769 * so that we do not miss the wakeup. -- see perf_pending_handle()
2772 return event->pending.next == NULL;
2775 static void perf_pending_sync(struct perf_event *event)
2777 wait_event(event->waitq, perf_not_pending(event));
2780 void perf_event_do_pending(void)
2782 __perf_pending_run();
2786 * Callchain support -- arch specific
2789 __weak struct perf_callchain_entry *perf_callchain(struct pt_regs *regs)
2797 static bool perf_output_space(struct perf_mmap_data *data, unsigned long tail,
2798 unsigned long offset, unsigned long head)
2802 if (!data->writable)
2805 mask = perf_data_size(data) - 1;
2807 offset = (offset - tail) & mask;
2808 head = (head - tail) & mask;
2810 if ((int)(head - offset) < 0)
2816 static void perf_output_wakeup(struct perf_output_handle *handle)
2818 atomic_set(&handle->data->poll, POLL_IN);
2821 handle->event->pending_wakeup = 1;
2822 perf_pending_queue(&handle->event->pending,
2823 perf_pending_event);
2825 perf_event_wakeup(handle->event);
2829 * Curious locking construct.
2831 * We need to ensure a later event_id doesn't publish a head when a former
2832 * event_id isn't done writing. However since we need to deal with NMIs we
2833 * cannot fully serialize things.
2835 * What we do is serialize between CPUs so we only have to deal with NMI
2836 * nesting on a single CPU.
2838 * We only publish the head (and generate a wakeup) when the outer-most
2839 * event_id completes.
2841 static void perf_output_lock(struct perf_output_handle *handle)
2843 struct perf_mmap_data *data = handle->data;
2844 int cur, cpu = get_cpu();
2849 cur = atomic_cmpxchg(&data->lock, -1, cpu);
2861 static void perf_output_unlock(struct perf_output_handle *handle)
2863 struct perf_mmap_data *data = handle->data;
2867 data->done_head = data->head;
2869 if (!handle->locked)
2874 * The xchg implies a full barrier that ensures all writes are done
2875 * before we publish the new head, matched by a rmb() in userspace when
2876 * reading this position.
2878 while ((head = atomic_long_xchg(&data->done_head, 0)))
2879 data->user_page->data_head = head;
2882 * NMI can happen here, which means we can miss a done_head update.
2885 cpu = atomic_xchg(&data->lock, -1);
2886 WARN_ON_ONCE(cpu != smp_processor_id());
2889 * Therefore we have to validate we did not indeed do so.
2891 if (unlikely(atomic_long_read(&data->done_head))) {
2893 * Since we had it locked, we can lock it again.
2895 while (atomic_cmpxchg(&data->lock, -1, cpu) != -1)
2901 if (atomic_xchg(&data->wakeup, 0))
2902 perf_output_wakeup(handle);
2907 void perf_output_copy(struct perf_output_handle *handle,
2908 const void *buf, unsigned int len)
2910 unsigned int pages_mask;
2911 unsigned long offset;
2915 offset = handle->offset;
2916 pages_mask = handle->data->nr_pages - 1;
2917 pages = handle->data->data_pages;
2920 unsigned long page_offset;
2921 unsigned long page_size;
2924 nr = (offset >> PAGE_SHIFT) & pages_mask;
2925 page_size = 1UL << (handle->data->data_order + PAGE_SHIFT);
2926 page_offset = offset & (page_size - 1);
2927 size = min_t(unsigned int, page_size - page_offset, len);
2929 memcpy(pages[nr] + page_offset, buf, size);
2936 handle->offset = offset;
2939 * Check we didn't copy past our reservation window, taking the
2940 * possible unsigned int wrap into account.
2942 WARN_ON_ONCE(((long)(handle->head - handle->offset)) < 0);
2945 int perf_output_begin(struct perf_output_handle *handle,
2946 struct perf_event *event, unsigned int size,
2947 int nmi, int sample)
2949 struct perf_event *output_event;
2950 struct perf_mmap_data *data;
2951 unsigned long tail, offset, head;
2954 struct perf_event_header header;
2961 * For inherited events we send all the output towards the parent.
2964 event = event->parent;
2966 output_event = rcu_dereference(event->output);
2968 event = output_event;
2970 data = rcu_dereference(event->data);
2974 handle->data = data;
2975 handle->event = event;
2977 handle->sample = sample;
2979 if (!data->nr_pages)
2982 have_lost = atomic_read(&data->lost);
2984 size += sizeof(lost_event);
2986 perf_output_lock(handle);
2990 * Userspace could choose to issue a mb() before updating the
2991 * tail pointer. So that all reads will be completed before the
2994 tail = ACCESS_ONCE(data->user_page->data_tail);
2996 offset = head = atomic_long_read(&data->head);
2998 if (unlikely(!perf_output_space(data, tail, offset, head)))
3000 } while (atomic_long_cmpxchg(&data->head, offset, head) != offset);
3002 handle->offset = offset;
3003 handle->head = head;
3005 if (head - tail > data->watermark)
3006 atomic_set(&data->wakeup, 1);
3009 lost_event.header.type = PERF_RECORD_LOST;
3010 lost_event.header.misc = 0;
3011 lost_event.header.size = sizeof(lost_event);
3012 lost_event.id = event->id;
3013 lost_event.lost = atomic_xchg(&data->lost, 0);
3015 perf_output_put(handle, lost_event);
3021 atomic_inc(&data->lost);
3022 perf_output_unlock(handle);
3029 void perf_output_end(struct perf_output_handle *handle)
3031 struct perf_event *event = handle->event;
3032 struct perf_mmap_data *data = handle->data;
3034 int wakeup_events = event->attr.wakeup_events;
3036 if (handle->sample && wakeup_events) {
3037 int events = atomic_inc_return(&data->events);
3038 if (events >= wakeup_events) {
3039 atomic_sub(wakeup_events, &data->events);
3040 atomic_set(&data->wakeup, 1);
3044 perf_output_unlock(handle);
3048 static u32 perf_event_pid(struct perf_event *event, struct task_struct *p)
3051 * only top level events have the pid namespace they were created in
3054 event = event->parent;
3056 return task_tgid_nr_ns(p, event->ns);
3059 static u32 perf_event_tid(struct perf_event *event, struct task_struct *p)
3062 * only top level events have the pid namespace they were created in
3065 event = event->parent;
3067 return task_pid_nr_ns(p, event->ns);
3070 static void perf_output_read_one(struct perf_output_handle *handle,
3071 struct perf_event *event)
3073 u64 read_format = event->attr.read_format;
3077 values[n++] = atomic64_read(&event->count);
3078 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
3079 values[n++] = event->total_time_enabled +
3080 atomic64_read(&event->child_total_time_enabled);
3082 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
3083 values[n++] = event->total_time_running +
3084 atomic64_read(&event->child_total_time_running);
3086 if (read_format & PERF_FORMAT_ID)
3087 values[n++] = primary_event_id(event);
3089 perf_output_copy(handle, values, n * sizeof(u64));
3093 * XXX PERF_FORMAT_GROUP vs inherited events seems difficult.
3095 static void perf_output_read_group(struct perf_output_handle *handle,
3096 struct perf_event *event)
3098 struct perf_event *leader = event->group_leader, *sub;
3099 u64 read_format = event->attr.read_format;
3103 values[n++] = 1 + leader->nr_siblings;
3105 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
3106 values[n++] = leader->total_time_enabled;
3108 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
3109 values[n++] = leader->total_time_running;
3111 if (leader != event)
3112 leader->pmu->read(leader);
3114 values[n++] = atomic64_read(&leader->count);
3115 if (read_format & PERF_FORMAT_ID)
3116 values[n++] = primary_event_id(leader);
3118 perf_output_copy(handle, values, n * sizeof(u64));
3120 list_for_each_entry(sub, &leader->sibling_list, group_entry) {
3124 sub->pmu->read(sub);
3126 values[n++] = atomic64_read(&sub->count);
3127 if (read_format & PERF_FORMAT_ID)
3128 values[n++] = primary_event_id(sub);
3130 perf_output_copy(handle, values, n * sizeof(u64));
3134 static void perf_output_read(struct perf_output_handle *handle,
3135 struct perf_event *event)
3137 if (event->attr.read_format & PERF_FORMAT_GROUP)
3138 perf_output_read_group(handle, event);
3140 perf_output_read_one(handle, event);
3143 void perf_output_sample(struct perf_output_handle *handle,
3144 struct perf_event_header *header,
3145 struct perf_sample_data *data,
3146 struct perf_event *event)
3148 u64 sample_type = data->type;
3150 perf_output_put(handle, *header);
3152 if (sample_type & PERF_SAMPLE_IP)
3153 perf_output_put(handle, data->ip);
3155 if (sample_type & PERF_SAMPLE_TID)
3156 perf_output_put(handle, data->tid_entry);
3158 if (sample_type & PERF_SAMPLE_TIME)
3159 perf_output_put(handle, data->time);
3161 if (sample_type & PERF_SAMPLE_ADDR)
3162 perf_output_put(handle, data->addr);
3164 if (sample_type & PERF_SAMPLE_ID)
3165 perf_output_put(handle, data->id);
3167 if (sample_type & PERF_SAMPLE_STREAM_ID)
3168 perf_output_put(handle, data->stream_id);
3170 if (sample_type & PERF_SAMPLE_CPU)
3171 perf_output_put(handle, data->cpu_entry);
3173 if (sample_type & PERF_SAMPLE_PERIOD)
3174 perf_output_put(handle, data->period);
3176 if (sample_type & PERF_SAMPLE_READ)
3177 perf_output_read(handle, event);
3179 if (sample_type & PERF_SAMPLE_CALLCHAIN) {
3180 if (data->callchain) {
3183 if (data->callchain)
3184 size += data->callchain->nr;
3186 size *= sizeof(u64);
3188 perf_output_copy(handle, data->callchain, size);
3191 perf_output_put(handle, nr);
3195 if (sample_type & PERF_SAMPLE_RAW) {
3197 perf_output_put(handle, data->raw->size);
3198 perf_output_copy(handle, data->raw->data,
3205 .size = sizeof(u32),
3208 perf_output_put(handle, raw);
3213 void perf_prepare_sample(struct perf_event_header *header,
3214 struct perf_sample_data *data,
3215 struct perf_event *event,
3216 struct pt_regs *regs)
3218 u64 sample_type = event->attr.sample_type;
3220 data->type = sample_type;
3222 header->type = PERF_RECORD_SAMPLE;
3223 header->size = sizeof(*header);
3226 header->misc |= perf_misc_flags(regs);
3228 if (sample_type & PERF_SAMPLE_IP) {
3229 data->ip = perf_instruction_pointer(regs);
3231 header->size += sizeof(data->ip);
3234 if (sample_type & PERF_SAMPLE_TID) {
3235 /* namespace issues */
3236 data->tid_entry.pid = perf_event_pid(event, current);
3237 data->tid_entry.tid = perf_event_tid(event, current);
3239 header->size += sizeof(data->tid_entry);
3242 if (sample_type & PERF_SAMPLE_TIME) {
3243 data->time = perf_clock();
3245 header->size += sizeof(data->time);
3248 if (sample_type & PERF_SAMPLE_ADDR)
3249 header->size += sizeof(data->addr);
3251 if (sample_type & PERF_SAMPLE_ID) {
3252 data->id = primary_event_id(event);
3254 header->size += sizeof(data->id);
3257 if (sample_type & PERF_SAMPLE_STREAM_ID) {
3258 data->stream_id = event->id;
3260 header->size += sizeof(data->stream_id);
3263 if (sample_type & PERF_SAMPLE_CPU) {
3264 data->cpu_entry.cpu = raw_smp_processor_id();
3265 data->cpu_entry.reserved = 0;
3267 header->size += sizeof(data->cpu_entry);
3270 if (sample_type & PERF_SAMPLE_PERIOD)
3271 header->size += sizeof(data->period);
3273 if (sample_type & PERF_SAMPLE_READ)
3274 header->size += perf_event_read_size(event);
3276 if (sample_type & PERF_SAMPLE_CALLCHAIN) {
3279 data->callchain = perf_callchain(regs);
3281 if (data->callchain)
3282 size += data->callchain->nr;
3284 header->size += size * sizeof(u64);
3287 if (sample_type & PERF_SAMPLE_RAW) {
3288 int size = sizeof(u32);
3291 size += data->raw->size;
3293 size += sizeof(u32);
3295 WARN_ON_ONCE(size & (sizeof(u64)-1));
3296 header->size += size;
3300 static void perf_event_output(struct perf_event *event, int nmi,
3301 struct perf_sample_data *data,
3302 struct pt_regs *regs)
3304 struct perf_output_handle handle;
3305 struct perf_event_header header;
3307 perf_prepare_sample(&header, data, event, regs);
3309 if (perf_output_begin(&handle, event, header.size, nmi, 1))
3312 perf_output_sample(&handle, &header, data, event);
3314 perf_output_end(&handle);
3321 struct perf_read_event {
3322 struct perf_event_header header;
3329 perf_event_read_event(struct perf_event *event,
3330 struct task_struct *task)
3332 struct perf_output_handle handle;
3333 struct perf_read_event read_event = {
3335 .type = PERF_RECORD_READ,
3337 .size = sizeof(read_event) + perf_event_read_size(event),
3339 .pid = perf_event_pid(event, task),
3340 .tid = perf_event_tid(event, task),
3344 ret = perf_output_begin(&handle, event, read_event.header.size, 0, 0);
3348 perf_output_put(&handle, read_event);
3349 perf_output_read(&handle, event);
3351 perf_output_end(&handle);
3355 * task tracking -- fork/exit
3357 * enabled by: attr.comm | attr.mmap | attr.task
3360 struct perf_task_event {
3361 struct task_struct *task;
3362 struct perf_event_context *task_ctx;
3365 struct perf_event_header header;
3375 static void perf_event_task_output(struct perf_event *event,
3376 struct perf_task_event *task_event)
3378 struct perf_output_handle handle;
3380 struct task_struct *task = task_event->task;
3383 size = task_event->event_id.header.size;
3384 ret = perf_output_begin(&handle, event, size, 0, 0);
3389 task_event->event_id.pid = perf_event_pid(event, task);
3390 task_event->event_id.ppid = perf_event_pid(event, current);
3392 task_event->event_id.tid = perf_event_tid(event, task);
3393 task_event->event_id.ptid = perf_event_tid(event, current);
3395 perf_output_put(&handle, task_event->event_id);
3397 perf_output_end(&handle);
3400 static int perf_event_task_match(struct perf_event *event)
3402 if (event->state < PERF_EVENT_STATE_INACTIVE)
3405 if (event->cpu != -1 && event->cpu != smp_processor_id())
3408 if (event->attr.comm || event->attr.mmap || event->attr.task)
3414 static void perf_event_task_ctx(struct perf_event_context *ctx,
3415 struct perf_task_event *task_event)
3417 struct perf_event *event;
3419 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
3420 if (perf_event_task_match(event))
3421 perf_event_task_output(event, task_event);
3425 static void perf_event_task_event(struct perf_task_event *task_event)
3427 struct perf_cpu_context *cpuctx;
3428 struct perf_event_context *ctx = task_event->task_ctx;
3431 cpuctx = &get_cpu_var(perf_cpu_context);
3432 perf_event_task_ctx(&cpuctx->ctx, task_event);
3434 ctx = rcu_dereference(current->perf_event_ctxp);
3436 perf_event_task_ctx(ctx, task_event);
3437 put_cpu_var(perf_cpu_context);
3441 static void perf_event_task(struct task_struct *task,
3442 struct perf_event_context *task_ctx,
3445 struct perf_task_event task_event;
3447 if (!atomic_read(&nr_comm_events) &&
3448 !atomic_read(&nr_mmap_events) &&
3449 !atomic_read(&nr_task_events))
3452 task_event = (struct perf_task_event){
3454 .task_ctx = task_ctx,
3457 .type = new ? PERF_RECORD_FORK : PERF_RECORD_EXIT,
3459 .size = sizeof(task_event.event_id),
3465 .time = perf_clock(),
3469 perf_event_task_event(&task_event);
3472 void perf_event_fork(struct task_struct *task)
3474 perf_event_task(task, NULL, 1);
3481 struct perf_comm_event {
3482 struct task_struct *task;
3487 struct perf_event_header header;
3494 static void perf_event_comm_output(struct perf_event *event,
3495 struct perf_comm_event *comm_event)
3497 struct perf_output_handle handle;
3498 int size = comm_event->event_id.header.size;
3499 int ret = perf_output_begin(&handle, event, size, 0, 0);
3504 comm_event->event_id.pid = perf_event_pid(event, comm_event->task);
3505 comm_event->event_id.tid = perf_event_tid(event, comm_event->task);
3507 perf_output_put(&handle, comm_event->event_id);
3508 perf_output_copy(&handle, comm_event->comm,
3509 comm_event->comm_size);
3510 perf_output_end(&handle);
3513 static int perf_event_comm_match(struct perf_event *event)
3515 if (event->state < PERF_EVENT_STATE_INACTIVE)
3518 if (event->cpu != -1 && event->cpu != smp_processor_id())
3521 if (event->attr.comm)
3527 static void perf_event_comm_ctx(struct perf_event_context *ctx,
3528 struct perf_comm_event *comm_event)
3530 struct perf_event *event;
3532 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
3533 if (perf_event_comm_match(event))
3534 perf_event_comm_output(event, comm_event);
3538 static void perf_event_comm_event(struct perf_comm_event *comm_event)
3540 struct perf_cpu_context *cpuctx;
3541 struct perf_event_context *ctx;
3543 char comm[TASK_COMM_LEN];
3545 memset(comm, 0, sizeof(comm));
3546 strlcpy(comm, comm_event->task->comm, sizeof(comm));
3547 size = ALIGN(strlen(comm)+1, sizeof(u64));
3549 comm_event->comm = comm;
3550 comm_event->comm_size = size;
3552 comm_event->event_id.header.size = sizeof(comm_event->event_id) + size;
3555 cpuctx = &get_cpu_var(perf_cpu_context);
3556 perf_event_comm_ctx(&cpuctx->ctx, comm_event);
3557 ctx = rcu_dereference(current->perf_event_ctxp);
3559 perf_event_comm_ctx(ctx, comm_event);
3560 put_cpu_var(perf_cpu_context);
3564 void perf_event_comm(struct task_struct *task)
3566 struct perf_comm_event comm_event;
3568 if (task->perf_event_ctxp)
3569 perf_event_enable_on_exec(task);
3571 if (!atomic_read(&nr_comm_events))
3574 comm_event = (struct perf_comm_event){
3580 .type = PERF_RECORD_COMM,
3589 perf_event_comm_event(&comm_event);
3596 struct perf_mmap_event {
3597 struct vm_area_struct *vma;
3599 const char *file_name;
3603 struct perf_event_header header;
3613 static void perf_event_mmap_output(struct perf_event *event,
3614 struct perf_mmap_event *mmap_event)
3616 struct perf_output_handle handle;
3617 int size = mmap_event->event_id.header.size;
3618 int ret = perf_output_begin(&handle, event, size, 0, 0);
3623 mmap_event->event_id.pid = perf_event_pid(event, current);
3624 mmap_event->event_id.tid = perf_event_tid(event, current);
3626 perf_output_put(&handle, mmap_event->event_id);
3627 perf_output_copy(&handle, mmap_event->file_name,
3628 mmap_event->file_size);
3629 perf_output_end(&handle);
3632 static int perf_event_mmap_match(struct perf_event *event,
3633 struct perf_mmap_event *mmap_event)
3635 if (event->state < PERF_EVENT_STATE_INACTIVE)
3638 if (event->cpu != -1 && event->cpu != smp_processor_id())
3641 if (event->attr.mmap)
3647 static void perf_event_mmap_ctx(struct perf_event_context *ctx,
3648 struct perf_mmap_event *mmap_event)
3650 struct perf_event *event;
3652 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
3653 if (perf_event_mmap_match(event, mmap_event))
3654 perf_event_mmap_output(event, mmap_event);
3658 static void perf_event_mmap_event(struct perf_mmap_event *mmap_event)
3660 struct perf_cpu_context *cpuctx;
3661 struct perf_event_context *ctx;
3662 struct vm_area_struct *vma = mmap_event->vma;
3663 struct file *file = vma->vm_file;
3669 memset(tmp, 0, sizeof(tmp));
3673 * d_path works from the end of the buffer backwards, so we
3674 * need to add enough zero bytes after the string to handle
3675 * the 64bit alignment we do later.
3677 buf = kzalloc(PATH_MAX + sizeof(u64), GFP_KERNEL);
3679 name = strncpy(tmp, "//enomem", sizeof(tmp));
3682 name = d_path(&file->f_path, buf, PATH_MAX);
3684 name = strncpy(tmp, "//toolong", sizeof(tmp));
3688 if (arch_vma_name(mmap_event->vma)) {
3689 name = strncpy(tmp, arch_vma_name(mmap_event->vma),
3695 name = strncpy(tmp, "[vdso]", sizeof(tmp));
3699 name = strncpy(tmp, "//anon", sizeof(tmp));
3704 size = ALIGN(strlen(name)+1, sizeof(u64));
3706 mmap_event->file_name = name;
3707 mmap_event->file_size = size;
3709 mmap_event->event_id.header.size = sizeof(mmap_event->event_id) + size;
3712 cpuctx = &get_cpu_var(perf_cpu_context);
3713 perf_event_mmap_ctx(&cpuctx->ctx, mmap_event);
3714 ctx = rcu_dereference(current->perf_event_ctxp);
3716 perf_event_mmap_ctx(ctx, mmap_event);
3717 put_cpu_var(perf_cpu_context);
3723 void __perf_event_mmap(struct vm_area_struct *vma)
3725 struct perf_mmap_event mmap_event;
3727 if (!atomic_read(&nr_mmap_events))
3730 mmap_event = (struct perf_mmap_event){
3736 .type = PERF_RECORD_MMAP,
3742 .start = vma->vm_start,
3743 .len = vma->vm_end - vma->vm_start,
3744 .pgoff = (u64)vma->vm_pgoff << PAGE_SHIFT,
3748 perf_event_mmap_event(&mmap_event);
3752 * IRQ throttle logging
3755 static void perf_log_throttle(struct perf_event *event, int enable)
3757 struct perf_output_handle handle;
3761 struct perf_event_header header;
3765 } throttle_event = {
3767 .type = PERF_RECORD_THROTTLE,
3769 .size = sizeof(throttle_event),
3771 .time = perf_clock(),
3772 .id = primary_event_id(event),
3773 .stream_id = event->id,
3777 throttle_event.header.type = PERF_RECORD_UNTHROTTLE;
3779 ret = perf_output_begin(&handle, event, sizeof(throttle_event), 1, 0);
3783 perf_output_put(&handle, throttle_event);
3784 perf_output_end(&handle);
3788 * Generic event overflow handling, sampling.
3791 static int __perf_event_overflow(struct perf_event *event, int nmi,
3792 int throttle, struct perf_sample_data *data,
3793 struct pt_regs *regs)
3795 int events = atomic_read(&event->event_limit);
3796 struct hw_perf_event *hwc = &event->hw;
3799 throttle = (throttle && event->pmu->unthrottle != NULL);
3804 if (hwc->interrupts != MAX_INTERRUPTS) {
3806 if (HZ * hwc->interrupts >
3807 (u64)sysctl_perf_event_sample_rate) {
3808 hwc->interrupts = MAX_INTERRUPTS;
3809 perf_log_throttle(event, 0);
3814 * Keep re-disabling events even though on the previous
3815 * pass we disabled it - just in case we raced with a
3816 * sched-in and the event got enabled again:
3822 if (event->attr.freq) {
3823 u64 now = perf_clock();
3824 s64 delta = now - hwc->freq_time_stamp;
3826 hwc->freq_time_stamp = now;
3828 if (delta > 0 && delta < 2*TICK_NSEC)
3829 perf_adjust_period(event, delta, hwc->last_period);
3833 * XXX event_limit might not quite work as expected on inherited
3837 event->pending_kill = POLL_IN;
3838 if (events && atomic_dec_and_test(&event->event_limit)) {
3840 event->pending_kill = POLL_HUP;
3842 event->pending_disable = 1;
3843 perf_pending_queue(&event->pending,
3844 perf_pending_event);
3846 perf_event_disable(event);
3849 if (event->overflow_handler)
3850 event->overflow_handler(event, nmi, data, regs);
3852 perf_event_output(event, nmi, data, regs);
3857 int perf_event_overflow(struct perf_event *event, int nmi,
3858 struct perf_sample_data *data,
3859 struct pt_regs *regs)
3861 return __perf_event_overflow(event, nmi, 1, data, regs);
3865 * Generic software event infrastructure
3869 * We directly increment event->count and keep a second value in
3870 * event->hw.period_left to count intervals. This period event
3871 * is kept in the range [-sample_period, 0] so that we can use the
3875 static u64 perf_swevent_set_period(struct perf_event *event)
3877 struct hw_perf_event *hwc = &event->hw;
3878 u64 period = hwc->last_period;
3882 hwc->last_period = hwc->sample_period;
3885 old = val = atomic64_read(&hwc->period_left);
3889 nr = div64_u64(period + val, period);
3890 offset = nr * period;
3892 if (atomic64_cmpxchg(&hwc->period_left, old, val) != old)
3898 static void perf_swevent_overflow(struct perf_event *event, u64 overflow,
3899 int nmi, struct perf_sample_data *data,
3900 struct pt_regs *regs)
3902 struct hw_perf_event *hwc = &event->hw;
3905 data->period = event->hw.last_period;
3907 overflow = perf_swevent_set_period(event);
3909 if (hwc->interrupts == MAX_INTERRUPTS)
3912 for (; overflow; overflow--) {
3913 if (__perf_event_overflow(event, nmi, throttle,
3916 * We inhibit the overflow from happening when
3917 * hwc->interrupts == MAX_INTERRUPTS.
3925 static void perf_swevent_unthrottle(struct perf_event *event)
3928 * Nothing to do, we already reset hwc->interrupts.
3932 static void perf_swevent_add(struct perf_event *event, u64 nr,
3933 int nmi, struct perf_sample_data *data,
3934 struct pt_regs *regs)
3936 struct hw_perf_event *hwc = &event->hw;
3938 atomic64_add(nr, &event->count);
3943 if (!hwc->sample_period)
3946 if (nr == 1 && hwc->sample_period == 1 && !event->attr.freq)
3947 return perf_swevent_overflow(event, 1, nmi, data, regs);
3949 if (atomic64_add_negative(nr, &hwc->period_left))
3952 perf_swevent_overflow(event, 0, nmi, data, regs);
3955 static int perf_swevent_is_counting(struct perf_event *event)
3958 * The event is active, we're good!
3960 if (event->state == PERF_EVENT_STATE_ACTIVE)
3964 * The event is off/error, not counting.
3966 if (event->state != PERF_EVENT_STATE_INACTIVE)
3970 * The event is inactive, if the context is active
3971 * we're part of a group that didn't make it on the 'pmu',
3974 if (event->ctx->is_active)
3978 * We're inactive and the context is too, this means the
3979 * task is scheduled out, we're counting events that happen
3980 * to us, like migration events.
3985 static int perf_tp_event_match(struct perf_event *event,
3986 struct perf_sample_data *data);
3988 static int perf_exclude_event(struct perf_event *event,
3989 struct pt_regs *regs)
3992 if (event->attr.exclude_user && user_mode(regs))
3995 if (event->attr.exclude_kernel && !user_mode(regs))
4002 static int perf_swevent_match(struct perf_event *event,
4003 enum perf_type_id type,
4005 struct perf_sample_data *data,
4006 struct pt_regs *regs)
4008 if (event->cpu != -1 && event->cpu != smp_processor_id())
4011 if (!perf_swevent_is_counting(event))
4014 if (event->attr.type != type)
4017 if (event->attr.config != event_id)
4020 if (perf_exclude_event(event, regs))
4023 if (event->attr.type == PERF_TYPE_TRACEPOINT &&
4024 !perf_tp_event_match(event, data))
4030 static void perf_swevent_ctx_event(struct perf_event_context *ctx,
4031 enum perf_type_id type,
4032 u32 event_id, u64 nr, int nmi,
4033 struct perf_sample_data *data,
4034 struct pt_regs *regs)
4036 struct perf_event *event;
4038 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
4039 if (perf_swevent_match(event, type, event_id, data, regs))
4040 perf_swevent_add(event, nr, nmi, data, regs);
4044 int perf_swevent_get_recursion_context(void)
4046 struct perf_cpu_context *cpuctx = &get_cpu_var(perf_cpu_context);
4053 else if (in_softirq())
4058 if (cpuctx->recursion[rctx]) {
4059 put_cpu_var(perf_cpu_context);
4063 cpuctx->recursion[rctx]++;
4068 EXPORT_SYMBOL_GPL(perf_swevent_get_recursion_context);
4070 void perf_swevent_put_recursion_context(int rctx)
4072 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
4074 cpuctx->recursion[rctx]--;
4075 put_cpu_var(perf_cpu_context);
4077 EXPORT_SYMBOL_GPL(perf_swevent_put_recursion_context);
4079 static void do_perf_sw_event(enum perf_type_id type, u32 event_id,
4081 struct perf_sample_data *data,
4082 struct pt_regs *regs)
4084 struct perf_cpu_context *cpuctx;
4085 struct perf_event_context *ctx;
4087 cpuctx = &__get_cpu_var(perf_cpu_context);
4089 perf_swevent_ctx_event(&cpuctx->ctx, type, event_id,
4090 nr, nmi, data, regs);
4092 * doesn't really matter which of the child contexts the
4093 * events ends up in.
4095 ctx = rcu_dereference(current->perf_event_ctxp);
4097 perf_swevent_ctx_event(ctx, type, event_id, nr, nmi, data, regs);
4101 void __perf_sw_event(u32 event_id, u64 nr, int nmi,
4102 struct pt_regs *regs, u64 addr)
4104 struct perf_sample_data data;
4107 rctx = perf_swevent_get_recursion_context();
4111 perf_sample_data_init(&data, addr);
4113 do_perf_sw_event(PERF_TYPE_SOFTWARE, event_id, nr, nmi, &data, regs);
4115 perf_swevent_put_recursion_context(rctx);
4118 static void perf_swevent_read(struct perf_event *event)
4122 static int perf_swevent_enable(struct perf_event *event)
4124 struct hw_perf_event *hwc = &event->hw;
4126 if (hwc->sample_period) {
4127 hwc->last_period = hwc->sample_period;
4128 perf_swevent_set_period(event);
4133 static void perf_swevent_disable(struct perf_event *event)
4137 static const struct pmu perf_ops_generic = {
4138 .enable = perf_swevent_enable,
4139 .disable = perf_swevent_disable,
4140 .read = perf_swevent_read,
4141 .unthrottle = perf_swevent_unthrottle,
4145 * hrtimer based swevent callback
4148 static enum hrtimer_restart perf_swevent_hrtimer(struct hrtimer *hrtimer)
4150 enum hrtimer_restart ret = HRTIMER_RESTART;
4151 struct perf_sample_data data;
4152 struct pt_regs *regs;
4153 struct perf_event *event;
4156 event = container_of(hrtimer, struct perf_event, hw.hrtimer);
4157 event->pmu->read(event);
4159 perf_sample_data_init(&data, 0);
4160 data.period = event->hw.last_period;
4161 regs = get_irq_regs();
4163 * In case we exclude kernel IPs or are somehow not in interrupt
4164 * context, provide the next best thing, the user IP.
4166 if ((event->attr.exclude_kernel || !regs) &&
4167 !event->attr.exclude_user)
4168 regs = task_pt_regs(current);
4171 if (!(event->attr.exclude_idle && current->pid == 0))
4172 if (perf_event_overflow(event, 0, &data, regs))
4173 ret = HRTIMER_NORESTART;
4176 period = max_t(u64, 10000, event->hw.sample_period);
4177 hrtimer_forward_now(hrtimer, ns_to_ktime(period));
4182 static void perf_swevent_start_hrtimer(struct perf_event *event)
4184 struct hw_perf_event *hwc = &event->hw;
4186 hrtimer_init(&hwc->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
4187 hwc->hrtimer.function = perf_swevent_hrtimer;
4188 if (hwc->sample_period) {
4191 if (hwc->remaining) {
4192 if (hwc->remaining < 0)
4195 period = hwc->remaining;
4198 period = max_t(u64, 10000, hwc->sample_period);
4200 __hrtimer_start_range_ns(&hwc->hrtimer,
4201 ns_to_ktime(period), 0,
4202 HRTIMER_MODE_REL, 0);
4206 static void perf_swevent_cancel_hrtimer(struct perf_event *event)
4208 struct hw_perf_event *hwc = &event->hw;
4210 if (hwc->sample_period) {
4211 ktime_t remaining = hrtimer_get_remaining(&hwc->hrtimer);
4212 hwc->remaining = ktime_to_ns(remaining);
4214 hrtimer_cancel(&hwc->hrtimer);
4219 * Software event: cpu wall time clock
4222 static void cpu_clock_perf_event_update(struct perf_event *event)
4224 int cpu = raw_smp_processor_id();
4228 now = cpu_clock(cpu);
4229 prev = atomic64_xchg(&event->hw.prev_count, now);
4230 atomic64_add(now - prev, &event->count);
4233 static int cpu_clock_perf_event_enable(struct perf_event *event)
4235 struct hw_perf_event *hwc = &event->hw;
4236 int cpu = raw_smp_processor_id();
4238 atomic64_set(&hwc->prev_count, cpu_clock(cpu));
4239 perf_swevent_start_hrtimer(event);
4244 static void cpu_clock_perf_event_disable(struct perf_event *event)
4246 perf_swevent_cancel_hrtimer(event);
4247 cpu_clock_perf_event_update(event);
4250 static void cpu_clock_perf_event_read(struct perf_event *event)
4252 cpu_clock_perf_event_update(event);
4255 static const struct pmu perf_ops_cpu_clock = {
4256 .enable = cpu_clock_perf_event_enable,
4257 .disable = cpu_clock_perf_event_disable,
4258 .read = cpu_clock_perf_event_read,
4262 * Software event: task time clock
4265 static void task_clock_perf_event_update(struct perf_event *event, u64 now)
4270 prev = atomic64_xchg(&event->hw.prev_count, now);
4272 atomic64_add(delta, &event->count);
4275 static int task_clock_perf_event_enable(struct perf_event *event)
4277 struct hw_perf_event *hwc = &event->hw;
4280 now = event->ctx->time;
4282 atomic64_set(&hwc->prev_count, now);
4284 perf_swevent_start_hrtimer(event);
4289 static void task_clock_perf_event_disable(struct perf_event *event)
4291 perf_swevent_cancel_hrtimer(event);
4292 task_clock_perf_event_update(event, event->ctx->time);
4296 static void task_clock_perf_event_read(struct perf_event *event)
4301 update_context_time(event->ctx);
4302 time = event->ctx->time;
4304 u64 now = perf_clock();
4305 u64 delta = now - event->ctx->timestamp;
4306 time = event->ctx->time + delta;
4309 task_clock_perf_event_update(event, time);
4312 static const struct pmu perf_ops_task_clock = {
4313 .enable = task_clock_perf_event_enable,
4314 .disable = task_clock_perf_event_disable,
4315 .read = task_clock_perf_event_read,
4318 #ifdef CONFIG_EVENT_TRACING
4320 void perf_tp_event(int event_id, u64 addr, u64 count, void *record,
4323 struct pt_regs *regs = get_irq_regs();
4324 struct perf_sample_data data;
4325 struct perf_raw_record raw = {
4330 perf_sample_data_init(&data, addr);
4334 regs = task_pt_regs(current);
4336 /* Trace events already protected against recursion */
4337 do_perf_sw_event(PERF_TYPE_TRACEPOINT, event_id, count, 1,
4340 EXPORT_SYMBOL_GPL(perf_tp_event);
4342 static int perf_tp_event_match(struct perf_event *event,
4343 struct perf_sample_data *data)
4345 void *record = data->raw->data;
4347 if (likely(!event->filter) || filter_match_preds(event->filter, record))
4352 static void tp_perf_event_destroy(struct perf_event *event)
4354 ftrace_profile_disable(event->attr.config);
4357 static const struct pmu *tp_perf_event_init(struct perf_event *event)
4360 * Raw tracepoint data is a severe data leak, only allow root to
4363 if ((event->attr.sample_type & PERF_SAMPLE_RAW) &&
4364 perf_paranoid_tracepoint_raw() &&
4365 !capable(CAP_SYS_ADMIN))
4366 return ERR_PTR(-EPERM);
4368 if (ftrace_profile_enable(event->attr.config))
4371 event->destroy = tp_perf_event_destroy;
4373 return &perf_ops_generic;
4376 static int perf_event_set_filter(struct perf_event *event, void __user *arg)
4381 if (event->attr.type != PERF_TYPE_TRACEPOINT)
4384 filter_str = strndup_user(arg, PAGE_SIZE);
4385 if (IS_ERR(filter_str))
4386 return PTR_ERR(filter_str);
4388 ret = ftrace_profile_set_filter(event, event->attr.config, filter_str);
4394 static void perf_event_free_filter(struct perf_event *event)
4396 ftrace_profile_free_filter(event);
4401 static int perf_tp_event_match(struct perf_event *event,
4402 struct perf_sample_data *data)
4407 static const struct pmu *tp_perf_event_init(struct perf_event *event)
4412 static int perf_event_set_filter(struct perf_event *event, void __user *arg)
4417 static void perf_event_free_filter(struct perf_event *event)
4421 #endif /* CONFIG_EVENT_TRACING */
4423 #ifdef CONFIG_HAVE_HW_BREAKPOINT
4424 static void bp_perf_event_destroy(struct perf_event *event)
4426 release_bp_slot(event);
4429 static const struct pmu *bp_perf_event_init(struct perf_event *bp)
4433 err = register_perf_hw_breakpoint(bp);
4435 return ERR_PTR(err);
4437 bp->destroy = bp_perf_event_destroy;
4439 return &perf_ops_bp;
4442 void perf_bp_event(struct perf_event *bp, void *data)
4444 struct perf_sample_data sample;
4445 struct pt_regs *regs = data;
4447 perf_sample_data_init(&sample, bp->attr.bp_addr);
4449 if (!perf_exclude_event(bp, regs))
4450 perf_swevent_add(bp, 1, 1, &sample, regs);
4453 static const struct pmu *bp_perf_event_init(struct perf_event *bp)
4458 void perf_bp_event(struct perf_event *bp, void *regs)
4463 atomic_t perf_swevent_enabled[PERF_COUNT_SW_MAX];
4465 static void sw_perf_event_destroy(struct perf_event *event)
4467 u64 event_id = event->attr.config;
4469 WARN_ON(event->parent);
4471 atomic_dec(&perf_swevent_enabled[event_id]);
4474 static const struct pmu *sw_perf_event_init(struct perf_event *event)
4476 const struct pmu *pmu = NULL;
4477 u64 event_id = event->attr.config;
4480 * Software events (currently) can't in general distinguish
4481 * between user, kernel and hypervisor events.
4482 * However, context switches and cpu migrations are considered
4483 * to be kernel events, and page faults are never hypervisor
4487 case PERF_COUNT_SW_CPU_CLOCK:
4488 pmu = &perf_ops_cpu_clock;
4491 case PERF_COUNT_SW_TASK_CLOCK:
4493 * If the user instantiates this as a per-cpu event,
4494 * use the cpu_clock event instead.
4496 if (event->ctx->task)
4497 pmu = &perf_ops_task_clock;
4499 pmu = &perf_ops_cpu_clock;
4502 case PERF_COUNT_SW_PAGE_FAULTS:
4503 case PERF_COUNT_SW_PAGE_FAULTS_MIN:
4504 case PERF_COUNT_SW_PAGE_FAULTS_MAJ:
4505 case PERF_COUNT_SW_CONTEXT_SWITCHES:
4506 case PERF_COUNT_SW_CPU_MIGRATIONS:
4507 case PERF_COUNT_SW_ALIGNMENT_FAULTS:
4508 case PERF_COUNT_SW_EMULATION_FAULTS:
4509 if (!event->parent) {
4510 atomic_inc(&perf_swevent_enabled[event_id]);
4511 event->destroy = sw_perf_event_destroy;
4513 pmu = &perf_ops_generic;
4521 * Allocate and initialize a event structure
4523 static struct perf_event *
4524 perf_event_alloc(struct perf_event_attr *attr,
4526 struct perf_event_context *ctx,
4527 struct perf_event *group_leader,
4528 struct perf_event *parent_event,
4529 perf_overflow_handler_t overflow_handler,
4532 const struct pmu *pmu;
4533 struct perf_event *event;
4534 struct hw_perf_event *hwc;
4537 event = kzalloc(sizeof(*event), gfpflags);
4539 return ERR_PTR(-ENOMEM);
4542 * Single events are their own group leaders, with an
4543 * empty sibling list:
4546 group_leader = event;
4548 mutex_init(&event->child_mutex);
4549 INIT_LIST_HEAD(&event->child_list);
4551 INIT_LIST_HEAD(&event->group_entry);
4552 INIT_LIST_HEAD(&event->event_entry);
4553 INIT_LIST_HEAD(&event->sibling_list);
4554 init_waitqueue_head(&event->waitq);
4556 mutex_init(&event->mmap_mutex);
4559 event->attr = *attr;
4560 event->group_leader = group_leader;
4565 event->parent = parent_event;
4567 event->ns = get_pid_ns(current->nsproxy->pid_ns);
4568 event->id = atomic64_inc_return(&perf_event_id);
4570 event->state = PERF_EVENT_STATE_INACTIVE;
4572 if (!overflow_handler && parent_event)
4573 overflow_handler = parent_event->overflow_handler;
4575 event->overflow_handler = overflow_handler;
4578 event->state = PERF_EVENT_STATE_OFF;
4583 hwc->sample_period = attr->sample_period;
4584 if (attr->freq && attr->sample_freq)
4585 hwc->sample_period = 1;
4586 hwc->last_period = hwc->sample_period;
4588 atomic64_set(&hwc->period_left, hwc->sample_period);
4591 * we currently do not support PERF_FORMAT_GROUP on inherited events
4593 if (attr->inherit && (attr->read_format & PERF_FORMAT_GROUP))
4596 switch (attr->type) {
4598 case PERF_TYPE_HARDWARE:
4599 case PERF_TYPE_HW_CACHE:
4600 pmu = hw_perf_event_init(event);
4603 case PERF_TYPE_SOFTWARE:
4604 pmu = sw_perf_event_init(event);
4607 case PERF_TYPE_TRACEPOINT:
4608 pmu = tp_perf_event_init(event);
4611 case PERF_TYPE_BREAKPOINT:
4612 pmu = bp_perf_event_init(event);
4623 else if (IS_ERR(pmu))
4628 put_pid_ns(event->ns);
4630 return ERR_PTR(err);
4635 if (!event->parent) {
4636 atomic_inc(&nr_events);
4637 if (event->attr.mmap)
4638 atomic_inc(&nr_mmap_events);
4639 if (event->attr.comm)
4640 atomic_inc(&nr_comm_events);
4641 if (event->attr.task)
4642 atomic_inc(&nr_task_events);
4648 static int perf_copy_attr(struct perf_event_attr __user *uattr,
4649 struct perf_event_attr *attr)
4654 if (!access_ok(VERIFY_WRITE, uattr, PERF_ATTR_SIZE_VER0))
4658 * zero the full structure, so that a short copy will be nice.
4660 memset(attr, 0, sizeof(*attr));
4662 ret = get_user(size, &uattr->size);
4666 if (size > PAGE_SIZE) /* silly large */
4669 if (!size) /* abi compat */
4670 size = PERF_ATTR_SIZE_VER0;
4672 if (size < PERF_ATTR_SIZE_VER0)
4676 * If we're handed a bigger struct than we know of,
4677 * ensure all the unknown bits are 0 - i.e. new
4678 * user-space does not rely on any kernel feature
4679 * extensions we dont know about yet.
4681 if (size > sizeof(*attr)) {
4682 unsigned char __user *addr;
4683 unsigned char __user *end;
4686 addr = (void __user *)uattr + sizeof(*attr);
4687 end = (void __user *)uattr + size;
4689 for (; addr < end; addr++) {
4690 ret = get_user(val, addr);
4696 size = sizeof(*attr);
4699 ret = copy_from_user(attr, uattr, size);
4704 * If the type exists, the corresponding creation will verify
4707 if (attr->type >= PERF_TYPE_MAX)
4710 if (attr->__reserved_1)
4713 if (attr->sample_type & ~(PERF_SAMPLE_MAX-1))
4716 if (attr->read_format & ~(PERF_FORMAT_MAX-1))
4723 put_user(sizeof(*attr), &uattr->size);
4728 static int perf_event_set_output(struct perf_event *event, int output_fd)
4730 struct perf_event *output_event = NULL;
4731 struct file *output_file = NULL;
4732 struct perf_event *old_output;
4733 int fput_needed = 0;
4739 output_file = fget_light(output_fd, &fput_needed);
4743 if (output_file->f_op != &perf_fops)
4746 output_event = output_file->private_data;
4748 /* Don't chain output fds */
4749 if (output_event->output)
4752 /* Don't set an output fd when we already have an output channel */
4756 atomic_long_inc(&output_file->f_count);
4759 mutex_lock(&event->mmap_mutex);
4760 old_output = event->output;
4761 rcu_assign_pointer(event->output, output_event);
4762 mutex_unlock(&event->mmap_mutex);
4766 * we need to make sure no existing perf_output_*()
4767 * is still referencing this event.
4770 fput(old_output->filp);
4775 fput_light(output_file, fput_needed);
4780 * sys_perf_event_open - open a performance event, associate it to a task/cpu
4782 * @attr_uptr: event_id type attributes for monitoring/sampling
4785 * @group_fd: group leader event fd
4787 SYSCALL_DEFINE5(perf_event_open,
4788 struct perf_event_attr __user *, attr_uptr,
4789 pid_t, pid, int, cpu, int, group_fd, unsigned long, flags)
4791 struct perf_event *event, *group_leader;
4792 struct perf_event_attr attr;
4793 struct perf_event_context *ctx;
4794 struct file *event_file = NULL;
4795 struct file *group_file = NULL;
4796 int fput_needed = 0;
4797 int fput_needed2 = 0;
4800 /* for future expandability... */
4801 if (flags & ~(PERF_FLAG_FD_NO_GROUP | PERF_FLAG_FD_OUTPUT))
4804 err = perf_copy_attr(attr_uptr, &attr);
4808 if (!attr.exclude_kernel) {
4809 if (perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN))
4814 if (attr.sample_freq > sysctl_perf_event_sample_rate)
4819 * Get the target context (task or percpu):
4821 ctx = find_get_context(pid, cpu);
4823 return PTR_ERR(ctx);
4826 * Look up the group leader (we will attach this event to it):
4828 group_leader = NULL;
4829 if (group_fd != -1 && !(flags & PERF_FLAG_FD_NO_GROUP)) {
4831 group_file = fget_light(group_fd, &fput_needed);
4833 goto err_put_context;
4834 if (group_file->f_op != &perf_fops)
4835 goto err_put_context;
4837 group_leader = group_file->private_data;
4839 * Do not allow a recursive hierarchy (this new sibling
4840 * becoming part of another group-sibling):
4842 if (group_leader->group_leader != group_leader)
4843 goto err_put_context;
4845 * Do not allow to attach to a group in a different
4846 * task or CPU context:
4848 if (group_leader->ctx != ctx)
4849 goto err_put_context;
4851 * Only a group leader can be exclusive or pinned
4853 if (attr.exclusive || attr.pinned)
4854 goto err_put_context;
4857 event = perf_event_alloc(&attr, cpu, ctx, group_leader,
4858 NULL, NULL, GFP_KERNEL);
4859 err = PTR_ERR(event);
4861 goto err_put_context;
4863 err = anon_inode_getfd("[perf_event]", &perf_fops, event, O_RDWR);
4865 goto err_free_put_context;
4867 event_file = fget_light(err, &fput_needed2);
4869 goto err_free_put_context;
4871 if (flags & PERF_FLAG_FD_OUTPUT) {
4872 err = perf_event_set_output(event, group_fd);
4874 goto err_fput_free_put_context;
4877 event->filp = event_file;
4878 WARN_ON_ONCE(ctx->parent_ctx);
4879 mutex_lock(&ctx->mutex);
4880 perf_install_in_context(ctx, event, cpu);
4882 mutex_unlock(&ctx->mutex);
4884 event->owner = current;
4885 get_task_struct(current);
4886 mutex_lock(¤t->perf_event_mutex);
4887 list_add_tail(&event->owner_entry, ¤t->perf_event_list);
4888 mutex_unlock(¤t->perf_event_mutex);
4890 err_fput_free_put_context:
4891 fput_light(event_file, fput_needed2);
4893 err_free_put_context:
4901 fput_light(group_file, fput_needed);
4907 * perf_event_create_kernel_counter
4909 * @attr: attributes of the counter to create
4910 * @cpu: cpu in which the counter is bound
4911 * @pid: task to profile
4914 perf_event_create_kernel_counter(struct perf_event_attr *attr, int cpu,
4916 perf_overflow_handler_t overflow_handler)
4918 struct perf_event *event;
4919 struct perf_event_context *ctx;
4923 * Get the target context (task or percpu):
4926 ctx = find_get_context(pid, cpu);
4932 event = perf_event_alloc(attr, cpu, ctx, NULL,
4933 NULL, overflow_handler, GFP_KERNEL);
4934 if (IS_ERR(event)) {
4935 err = PTR_ERR(event);
4936 goto err_put_context;
4940 WARN_ON_ONCE(ctx->parent_ctx);
4941 mutex_lock(&ctx->mutex);
4942 perf_install_in_context(ctx, event, cpu);
4944 mutex_unlock(&ctx->mutex);
4946 event->owner = current;
4947 get_task_struct(current);
4948 mutex_lock(¤t->perf_event_mutex);
4949 list_add_tail(&event->owner_entry, ¤t->perf_event_list);
4950 mutex_unlock(¤t->perf_event_mutex);
4957 return ERR_PTR(err);
4959 EXPORT_SYMBOL_GPL(perf_event_create_kernel_counter);
4962 * inherit a event from parent task to child task:
4964 static struct perf_event *
4965 inherit_event(struct perf_event *parent_event,
4966 struct task_struct *parent,
4967 struct perf_event_context *parent_ctx,
4968 struct task_struct *child,
4969 struct perf_event *group_leader,
4970 struct perf_event_context *child_ctx)
4972 struct perf_event *child_event;
4975 * Instead of creating recursive hierarchies of events,
4976 * we link inherited events back to the original parent,
4977 * which has a filp for sure, which we use as the reference
4980 if (parent_event->parent)
4981 parent_event = parent_event->parent;
4983 child_event = perf_event_alloc(&parent_event->attr,
4984 parent_event->cpu, child_ctx,
4985 group_leader, parent_event,
4987 if (IS_ERR(child_event))
4992 * Make the child state follow the state of the parent event,
4993 * not its attr.disabled bit. We hold the parent's mutex,
4994 * so we won't race with perf_event_{en, dis}able_family.
4996 if (parent_event->state >= PERF_EVENT_STATE_INACTIVE)
4997 child_event->state = PERF_EVENT_STATE_INACTIVE;
4999 child_event->state = PERF_EVENT_STATE_OFF;
5001 if (parent_event->attr.freq) {
5002 u64 sample_period = parent_event->hw.sample_period;
5003 struct hw_perf_event *hwc = &child_event->hw;
5005 hwc->sample_period = sample_period;
5006 hwc->last_period = sample_period;
5008 atomic64_set(&hwc->period_left, sample_period);
5011 child_event->overflow_handler = parent_event->overflow_handler;
5014 * Link it up in the child's context:
5016 add_event_to_ctx(child_event, child_ctx);
5019 * Get a reference to the parent filp - we will fput it
5020 * when the child event exits. This is safe to do because
5021 * we are in the parent and we know that the filp still
5022 * exists and has a nonzero count:
5024 atomic_long_inc(&parent_event->filp->f_count);
5027 * Link this into the parent event's child list
5029 WARN_ON_ONCE(parent_event->ctx->parent_ctx);
5030 mutex_lock(&parent_event->child_mutex);
5031 list_add_tail(&child_event->child_list, &parent_event->child_list);
5032 mutex_unlock(&parent_event->child_mutex);
5037 static int inherit_group(struct perf_event *parent_event,
5038 struct task_struct *parent,
5039 struct perf_event_context *parent_ctx,
5040 struct task_struct *child,
5041 struct perf_event_context *child_ctx)
5043 struct perf_event *leader;
5044 struct perf_event *sub;
5045 struct perf_event *child_ctr;
5047 leader = inherit_event(parent_event, parent, parent_ctx,
5048 child, NULL, child_ctx);
5050 return PTR_ERR(leader);
5051 list_for_each_entry(sub, &parent_event->sibling_list, group_entry) {
5052 child_ctr = inherit_event(sub, parent, parent_ctx,
5053 child, leader, child_ctx);
5054 if (IS_ERR(child_ctr))
5055 return PTR_ERR(child_ctr);
5060 static void sync_child_event(struct perf_event *child_event,
5061 struct task_struct *child)
5063 struct perf_event *parent_event = child_event->parent;
5066 if (child_event->attr.inherit_stat)
5067 perf_event_read_event(child_event, child);
5069 child_val = atomic64_read(&child_event->count);
5072 * Add back the child's count to the parent's count:
5074 atomic64_add(child_val, &parent_event->count);
5075 atomic64_add(child_event->total_time_enabled,
5076 &parent_event->child_total_time_enabled);
5077 atomic64_add(child_event->total_time_running,
5078 &parent_event->child_total_time_running);
5081 * Remove this event from the parent's list
5083 WARN_ON_ONCE(parent_event->ctx->parent_ctx);
5084 mutex_lock(&parent_event->child_mutex);
5085 list_del_init(&child_event->child_list);
5086 mutex_unlock(&parent_event->child_mutex);
5089 * Release the parent event, if this was the last
5092 fput(parent_event->filp);
5096 __perf_event_exit_task(struct perf_event *child_event,
5097 struct perf_event_context *child_ctx,
5098 struct task_struct *child)
5100 struct perf_event *parent_event;
5102 perf_event_remove_from_context(child_event);
5104 parent_event = child_event->parent;
5106 * It can happen that parent exits first, and has events
5107 * that are still around due to the child reference. These
5108 * events need to be zapped - but otherwise linger.
5111 sync_child_event(child_event, child);
5112 free_event(child_event);
5117 * When a child task exits, feed back event values to parent events.
5119 void perf_event_exit_task(struct task_struct *child)
5121 struct perf_event *child_event, *tmp;
5122 struct perf_event_context *child_ctx;
5123 unsigned long flags;
5125 if (likely(!child->perf_event_ctxp)) {
5126 perf_event_task(child, NULL, 0);
5130 local_irq_save(flags);
5132 * We can't reschedule here because interrupts are disabled,
5133 * and either child is current or it is a task that can't be
5134 * scheduled, so we are now safe from rescheduling changing
5137 child_ctx = child->perf_event_ctxp;
5138 __perf_event_task_sched_out(child_ctx);
5141 * Take the context lock here so that if find_get_context is
5142 * reading child->perf_event_ctxp, we wait until it has
5143 * incremented the context's refcount before we do put_ctx below.
5145 raw_spin_lock(&child_ctx->lock);
5146 child->perf_event_ctxp = NULL;
5148 * If this context is a clone; unclone it so it can't get
5149 * swapped to another process while we're removing all
5150 * the events from it.
5152 unclone_ctx(child_ctx);
5153 update_context_time(child_ctx);
5154 raw_spin_unlock_irqrestore(&child_ctx->lock, flags);
5157 * Report the task dead after unscheduling the events so that we
5158 * won't get any samples after PERF_RECORD_EXIT. We can however still
5159 * get a few PERF_RECORD_READ events.
5161 perf_event_task(child, child_ctx, 0);
5164 * We can recurse on the same lock type through:
5166 * __perf_event_exit_task()
5167 * sync_child_event()
5168 * fput(parent_event->filp)
5170 * mutex_lock(&ctx->mutex)
5172 * But since its the parent context it won't be the same instance.
5174 mutex_lock_nested(&child_ctx->mutex, SINGLE_DEPTH_NESTING);
5177 list_for_each_entry_safe(child_event, tmp, &child_ctx->pinned_groups,
5179 __perf_event_exit_task(child_event, child_ctx, child);
5181 list_for_each_entry_safe(child_event, tmp, &child_ctx->flexible_groups,
5183 __perf_event_exit_task(child_event, child_ctx, child);
5186 * If the last event was a group event, it will have appended all
5187 * its siblings to the list, but we obtained 'tmp' before that which
5188 * will still point to the list head terminating the iteration.
5190 if (!list_empty(&child_ctx->pinned_groups) ||
5191 !list_empty(&child_ctx->flexible_groups))
5194 mutex_unlock(&child_ctx->mutex);
5199 static void perf_free_event(struct perf_event *event,
5200 struct perf_event_context *ctx)
5202 struct perf_event *parent = event->parent;
5204 if (WARN_ON_ONCE(!parent))
5207 mutex_lock(&parent->child_mutex);
5208 list_del_init(&event->child_list);
5209 mutex_unlock(&parent->child_mutex);
5213 list_del_event(event, ctx);
5218 * free an unexposed, unused context as created by inheritance by
5219 * init_task below, used by fork() in case of fail.
5221 void perf_event_free_task(struct task_struct *task)
5223 struct perf_event_context *ctx = task->perf_event_ctxp;
5224 struct perf_event *event, *tmp;
5229 mutex_lock(&ctx->mutex);
5231 list_for_each_entry_safe(event, tmp, &ctx->pinned_groups, group_entry)
5232 perf_free_event(event, ctx);
5234 list_for_each_entry_safe(event, tmp, &ctx->flexible_groups,
5236 perf_free_event(event, ctx);
5238 if (!list_empty(&ctx->pinned_groups) ||
5239 !list_empty(&ctx->flexible_groups))
5242 mutex_unlock(&ctx->mutex);
5248 inherit_task_group(struct perf_event *event, struct task_struct *parent,
5249 struct perf_event_context *parent_ctx,
5250 struct task_struct *child,
5254 struct perf_event_context *child_ctx = child->perf_event_ctxp;
5256 if (!event->attr.inherit) {
5263 * This is executed from the parent task context, so
5264 * inherit events that have been marked for cloning.
5265 * First allocate and initialize a context for the
5269 child_ctx = kzalloc(sizeof(struct perf_event_context),
5274 __perf_event_init_context(child_ctx, child);
5275 child->perf_event_ctxp = child_ctx;
5276 get_task_struct(child);
5279 ret = inherit_group(event, parent, parent_ctx,
5290 * Initialize the perf_event context in task_struct
5292 int perf_event_init_task(struct task_struct *child)
5294 struct perf_event_context *child_ctx, *parent_ctx;
5295 struct perf_event_context *cloned_ctx;
5296 struct perf_event *event;
5297 struct task_struct *parent = current;
5298 int inherited_all = 1;
5301 child->perf_event_ctxp = NULL;
5303 mutex_init(&child->perf_event_mutex);
5304 INIT_LIST_HEAD(&child->perf_event_list);
5306 if (likely(!parent->perf_event_ctxp))
5310 * If the parent's context is a clone, pin it so it won't get
5313 parent_ctx = perf_pin_task_context(parent);
5316 * No need to check if parent_ctx != NULL here; since we saw
5317 * it non-NULL earlier, the only reason for it to become NULL
5318 * is if we exit, and since we're currently in the middle of
5319 * a fork we can't be exiting at the same time.
5323 * Lock the parent list. No need to lock the child - not PID
5324 * hashed yet and not running, so nobody can access it.
5326 mutex_lock(&parent_ctx->mutex);
5329 * We dont have to disable NMIs - we are only looking at
5330 * the list, not manipulating it:
5332 list_for_each_entry(event, &parent_ctx->pinned_groups, group_entry) {
5333 ret = inherit_task_group(event, parent, parent_ctx, child,
5339 list_for_each_entry(event, &parent_ctx->flexible_groups, group_entry) {
5340 ret = inherit_task_group(event, parent, parent_ctx, child,
5346 child_ctx = child->perf_event_ctxp;
5348 if (child_ctx && inherited_all) {
5350 * Mark the child context as a clone of the parent
5351 * context, or of whatever the parent is a clone of.
5352 * Note that if the parent is a clone, it could get
5353 * uncloned at any point, but that doesn't matter
5354 * because the list of events and the generation
5355 * count can't have changed since we took the mutex.
5357 cloned_ctx = rcu_dereference(parent_ctx->parent_ctx);
5359 child_ctx->parent_ctx = cloned_ctx;
5360 child_ctx->parent_gen = parent_ctx->parent_gen;
5362 child_ctx->parent_ctx = parent_ctx;
5363 child_ctx->parent_gen = parent_ctx->generation;
5365 get_ctx(child_ctx->parent_ctx);
5368 mutex_unlock(&parent_ctx->mutex);
5370 perf_unpin_context(parent_ctx);
5375 static void __cpuinit perf_event_init_cpu(int cpu)
5377 struct perf_cpu_context *cpuctx;
5379 cpuctx = &per_cpu(perf_cpu_context, cpu);
5380 __perf_event_init_context(&cpuctx->ctx, NULL);
5382 spin_lock(&perf_resource_lock);
5383 cpuctx->max_pertask = perf_max_events - perf_reserved_percpu;
5384 spin_unlock(&perf_resource_lock);
5386 hw_perf_event_setup(cpu);
5389 #ifdef CONFIG_HOTPLUG_CPU
5390 static void __perf_event_exit_cpu(void *info)
5392 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
5393 struct perf_event_context *ctx = &cpuctx->ctx;
5394 struct perf_event *event, *tmp;
5396 list_for_each_entry_safe(event, tmp, &ctx->pinned_groups, group_entry)
5397 __perf_event_remove_from_context(event);
5398 list_for_each_entry_safe(event, tmp, &ctx->flexible_groups, group_entry)
5399 __perf_event_remove_from_context(event);
5401 static void perf_event_exit_cpu(int cpu)
5403 struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
5404 struct perf_event_context *ctx = &cpuctx->ctx;
5406 mutex_lock(&ctx->mutex);
5407 smp_call_function_single(cpu, __perf_event_exit_cpu, NULL, 1);
5408 mutex_unlock(&ctx->mutex);
5411 static inline void perf_event_exit_cpu(int cpu) { }
5414 static int __cpuinit
5415 perf_cpu_notify(struct notifier_block *self, unsigned long action, void *hcpu)
5417 unsigned int cpu = (long)hcpu;
5421 case CPU_UP_PREPARE:
5422 case CPU_UP_PREPARE_FROZEN:
5423 perf_event_init_cpu(cpu);
5427 case CPU_ONLINE_FROZEN:
5428 hw_perf_event_setup_online(cpu);
5431 case CPU_DOWN_PREPARE:
5432 case CPU_DOWN_PREPARE_FROZEN:
5433 perf_event_exit_cpu(cpu);
5437 hw_perf_event_setup_offline(cpu);
5448 * This has to have a higher priority than migration_notifier in sched.c.
5450 static struct notifier_block __cpuinitdata perf_cpu_nb = {
5451 .notifier_call = perf_cpu_notify,
5455 void __init perf_event_init(void)
5457 perf_cpu_notify(&perf_cpu_nb, (unsigned long)CPU_UP_PREPARE,
5458 (void *)(long)smp_processor_id());
5459 perf_cpu_notify(&perf_cpu_nb, (unsigned long)CPU_ONLINE,
5460 (void *)(long)smp_processor_id());
5461 register_cpu_notifier(&perf_cpu_nb);
5464 static ssize_t perf_show_reserve_percpu(struct sysdev_class *class,
5465 struct sysdev_class_attribute *attr,
5468 return sprintf(buf, "%d\n", perf_reserved_percpu);
5472 perf_set_reserve_percpu(struct sysdev_class *class,
5473 struct sysdev_class_attribute *attr,
5477 struct perf_cpu_context *cpuctx;
5481 err = strict_strtoul(buf, 10, &val);
5484 if (val > perf_max_events)
5487 spin_lock(&perf_resource_lock);
5488 perf_reserved_percpu = val;
5489 for_each_online_cpu(cpu) {
5490 cpuctx = &per_cpu(perf_cpu_context, cpu);
5491 raw_spin_lock_irq(&cpuctx->ctx.lock);
5492 mpt = min(perf_max_events - cpuctx->ctx.nr_events,
5493 perf_max_events - perf_reserved_percpu);
5494 cpuctx->max_pertask = mpt;
5495 raw_spin_unlock_irq(&cpuctx->ctx.lock);
5497 spin_unlock(&perf_resource_lock);
5502 static ssize_t perf_show_overcommit(struct sysdev_class *class,
5503 struct sysdev_class_attribute *attr,
5506 return sprintf(buf, "%d\n", perf_overcommit);
5510 perf_set_overcommit(struct sysdev_class *class,
5511 struct sysdev_class_attribute *attr,
5512 const char *buf, size_t count)
5517 err = strict_strtoul(buf, 10, &val);
5523 spin_lock(&perf_resource_lock);
5524 perf_overcommit = val;
5525 spin_unlock(&perf_resource_lock);
5530 static SYSDEV_CLASS_ATTR(
5533 perf_show_reserve_percpu,
5534 perf_set_reserve_percpu
5537 static SYSDEV_CLASS_ATTR(
5540 perf_show_overcommit,
5544 static struct attribute *perfclass_attrs[] = {
5545 &attr_reserve_percpu.attr,
5546 &attr_overcommit.attr,
5550 static struct attribute_group perfclass_attr_group = {
5551 .attrs = perfclass_attrs,
5552 .name = "perf_events",
5555 static int __init perf_event_sysfs_init(void)
5557 return sysfs_create_group(&cpu_sysdev_class.kset.kobj,
5558 &perfclass_attr_group);
5560 device_initcall(perf_event_sysfs_init);