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 static inline bool perf_paranoid_tracepoint_raw(void)
61 return sysctl_perf_event_paranoid > -1;
64 static inline bool perf_paranoid_cpu(void)
66 return sysctl_perf_event_paranoid > 0;
69 static inline bool perf_paranoid_kernel(void)
71 return sysctl_perf_event_paranoid > 1;
74 int sysctl_perf_event_mlock __read_mostly = 512; /* 'free' kb per user */
77 * max perf event sample rate
79 int sysctl_perf_event_sample_rate __read_mostly = 100000;
81 static atomic64_t perf_event_id;
84 * Lock for (sysadmin-configurable) event reservations:
86 static DEFINE_SPINLOCK(perf_resource_lock);
89 * Architecture provided APIs - weak aliases:
91 extern __weak const struct pmu *hw_perf_event_init(struct perf_event *event)
96 void __weak hw_perf_disable(void) { barrier(); }
97 void __weak hw_perf_enable(void) { barrier(); }
99 void __weak hw_perf_event_setup(int cpu) { barrier(); }
100 void __weak hw_perf_event_setup_online(int cpu) { barrier(); }
103 hw_perf_group_sched_in(struct perf_event *group_leader,
104 struct perf_cpu_context *cpuctx,
105 struct perf_event_context *ctx, int cpu)
110 void __weak perf_event_print_debug(void) { }
112 static DEFINE_PER_CPU(int, perf_disable_count);
114 void __perf_disable(void)
116 __get_cpu_var(perf_disable_count)++;
119 bool __perf_enable(void)
121 return !--__get_cpu_var(perf_disable_count);
124 void perf_disable(void)
130 void perf_enable(void)
136 static void get_ctx(struct perf_event_context *ctx)
138 WARN_ON(!atomic_inc_not_zero(&ctx->refcount));
141 static void free_ctx(struct rcu_head *head)
143 struct perf_event_context *ctx;
145 ctx = container_of(head, struct perf_event_context, rcu_head);
149 static void put_ctx(struct perf_event_context *ctx)
151 if (atomic_dec_and_test(&ctx->refcount)) {
153 put_ctx(ctx->parent_ctx);
155 put_task_struct(ctx->task);
156 call_rcu(&ctx->rcu_head, free_ctx);
160 static void unclone_ctx(struct perf_event_context *ctx)
162 if (ctx->parent_ctx) {
163 put_ctx(ctx->parent_ctx);
164 ctx->parent_ctx = NULL;
169 * If we inherit events we want to return the parent event id
172 static u64 primary_event_id(struct perf_event *event)
177 id = event->parent->id;
183 * Get the perf_event_context for a task and lock it.
184 * This has to cope with with the fact that until it is locked,
185 * the context could get moved to another task.
187 static struct perf_event_context *
188 perf_lock_task_context(struct task_struct *task, unsigned long *flags)
190 struct perf_event_context *ctx;
194 ctx = rcu_dereference(task->perf_event_ctxp);
197 * If this context is a clone of another, it might
198 * get swapped for another underneath us by
199 * perf_event_task_sched_out, though the
200 * rcu_read_lock() protects us from any context
201 * getting freed. Lock the context and check if it
202 * got swapped before we could get the lock, and retry
203 * if so. If we locked the right context, then it
204 * can't get swapped on us any more.
206 raw_spin_lock_irqsave(&ctx->lock, *flags);
207 if (ctx != rcu_dereference(task->perf_event_ctxp)) {
208 raw_spin_unlock_irqrestore(&ctx->lock, *flags);
212 if (!atomic_inc_not_zero(&ctx->refcount)) {
213 raw_spin_unlock_irqrestore(&ctx->lock, *flags);
222 * Get the context for a task and increment its pin_count so it
223 * can't get swapped to another task. This also increments its
224 * reference count so that the context can't get freed.
226 static struct perf_event_context *perf_pin_task_context(struct task_struct *task)
228 struct perf_event_context *ctx;
231 ctx = perf_lock_task_context(task, &flags);
234 raw_spin_unlock_irqrestore(&ctx->lock, flags);
239 static void perf_unpin_context(struct perf_event_context *ctx)
243 raw_spin_lock_irqsave(&ctx->lock, flags);
245 raw_spin_unlock_irqrestore(&ctx->lock, flags);
249 static inline u64 perf_clock(void)
251 return cpu_clock(smp_processor_id());
255 * Update the record of the current time in a context.
257 static void update_context_time(struct perf_event_context *ctx)
259 u64 now = perf_clock();
261 ctx->time += now - ctx->timestamp;
262 ctx->timestamp = now;
266 * Update the total_time_enabled and total_time_running fields for a event.
268 static void update_event_times(struct perf_event *event)
270 struct perf_event_context *ctx = event->ctx;
273 if (event->state < PERF_EVENT_STATE_INACTIVE ||
274 event->group_leader->state < PERF_EVENT_STATE_INACTIVE)
280 run_end = event->tstamp_stopped;
282 event->total_time_enabled = run_end - event->tstamp_enabled;
284 if (event->state == PERF_EVENT_STATE_INACTIVE)
285 run_end = event->tstamp_stopped;
289 event->total_time_running = run_end - event->tstamp_running;
292 static struct list_head *
293 ctx_group_list(struct perf_event *event, struct perf_event_context *ctx)
295 if (event->attr.pinned)
296 return &ctx->pinned_groups;
298 return &ctx->flexible_groups;
302 * Add a event from the lists for its context.
303 * Must be called with ctx->mutex and ctx->lock held.
306 list_add_event(struct perf_event *event, struct perf_event_context *ctx)
308 struct perf_event *group_leader = event->group_leader;
311 * Depending on whether it is a standalone or sibling event,
312 * add it straight to the context's event list, or to the group
313 * leader's sibling list:
315 if (group_leader == event) {
316 struct list_head *list;
318 list = ctx_group_list(event, ctx);
319 list_add_tail(&event->group_entry, list);
321 list_add_tail(&event->group_entry, &group_leader->sibling_list);
322 group_leader->nr_siblings++;
325 list_add_rcu(&event->event_entry, &ctx->event_list);
327 if (event->attr.inherit_stat)
332 * Remove a event from the lists for its context.
333 * Must be called with ctx->mutex and ctx->lock held.
336 list_del_event(struct perf_event *event, struct perf_event_context *ctx)
338 struct perf_event *sibling, *tmp;
340 if (list_empty(&event->group_entry))
343 if (event->attr.inherit_stat)
346 list_del_init(&event->group_entry);
347 list_del_rcu(&event->event_entry);
349 if (event->group_leader != event)
350 event->group_leader->nr_siblings--;
352 update_event_times(event);
355 * If event was in error state, then keep it
356 * that way, otherwise bogus counts will be
357 * returned on read(). The only way to get out
358 * of error state is by explicit re-enabling
361 if (event->state > PERF_EVENT_STATE_OFF)
362 event->state = PERF_EVENT_STATE_OFF;
365 * If this was a group event with sibling events then
366 * upgrade the siblings to singleton events by adding them
367 * to the context list directly:
369 list_for_each_entry_safe(sibling, tmp, &event->sibling_list, group_entry) {
370 struct list_head *list;
372 list = ctx_group_list(event, ctx);
373 list_move_tail(&sibling->group_entry, list);
374 sibling->group_leader = sibling;
379 event_sched_out(struct perf_event *event,
380 struct perf_cpu_context *cpuctx,
381 struct perf_event_context *ctx)
383 if (event->state != PERF_EVENT_STATE_ACTIVE)
386 event->state = PERF_EVENT_STATE_INACTIVE;
387 if (event->pending_disable) {
388 event->pending_disable = 0;
389 event->state = PERF_EVENT_STATE_OFF;
391 event->tstamp_stopped = ctx->time;
392 event->pmu->disable(event);
395 if (!is_software_event(event))
396 cpuctx->active_oncpu--;
398 if (event->attr.exclusive || !cpuctx->active_oncpu)
399 cpuctx->exclusive = 0;
403 group_sched_out(struct perf_event *group_event,
404 struct perf_cpu_context *cpuctx,
405 struct perf_event_context *ctx)
407 struct perf_event *event;
409 if (group_event->state != PERF_EVENT_STATE_ACTIVE)
412 event_sched_out(group_event, cpuctx, ctx);
415 * Schedule out siblings (if any):
417 list_for_each_entry(event, &group_event->sibling_list, group_entry)
418 event_sched_out(event, cpuctx, ctx);
420 if (group_event->attr.exclusive)
421 cpuctx->exclusive = 0;
425 * Cross CPU call to remove a performance event
427 * We disable the event on the hardware level first. After that we
428 * remove it from the context list.
430 static void __perf_event_remove_from_context(void *info)
432 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
433 struct perf_event *event = info;
434 struct perf_event_context *ctx = event->ctx;
437 * If this is a task context, we need to check whether it is
438 * the current task context of this cpu. If not it has been
439 * scheduled out before the smp call arrived.
441 if (ctx->task && cpuctx->task_ctx != ctx)
444 raw_spin_lock(&ctx->lock);
446 * Protect the list operation against NMI by disabling the
447 * events on a global level.
451 event_sched_out(event, cpuctx, ctx);
453 list_del_event(event, ctx);
457 * Allow more per task events with respect to the
460 cpuctx->max_pertask =
461 min(perf_max_events - ctx->nr_events,
462 perf_max_events - perf_reserved_percpu);
466 raw_spin_unlock(&ctx->lock);
471 * Remove the event from a task's (or a CPU's) list of events.
473 * Must be called with ctx->mutex held.
475 * CPU events are removed with a smp call. For task events we only
476 * call when the task is on a CPU.
478 * If event->ctx is a cloned context, callers must make sure that
479 * every task struct that event->ctx->task could possibly point to
480 * remains valid. This is OK when called from perf_release since
481 * that only calls us on the top-level context, which can't be a clone.
482 * When called from perf_event_exit_task, it's OK because the
483 * context has been detached from its task.
485 static void perf_event_remove_from_context(struct perf_event *event)
487 struct perf_event_context *ctx = event->ctx;
488 struct task_struct *task = ctx->task;
492 * Per cpu events are removed via an smp call and
493 * the removal is always successful.
495 smp_call_function_single(event->cpu,
496 __perf_event_remove_from_context,
502 task_oncpu_function_call(task, __perf_event_remove_from_context,
505 raw_spin_lock_irq(&ctx->lock);
507 * If the context is active we need to retry the smp call.
509 if (ctx->nr_active && !list_empty(&event->group_entry)) {
510 raw_spin_unlock_irq(&ctx->lock);
515 * The lock prevents that this context is scheduled in so we
516 * can remove the event safely, if the call above did not
519 if (!list_empty(&event->group_entry))
520 list_del_event(event, ctx);
521 raw_spin_unlock_irq(&ctx->lock);
525 * Update total_time_enabled and total_time_running for all events in a group.
527 static void update_group_times(struct perf_event *leader)
529 struct perf_event *event;
531 update_event_times(leader);
532 list_for_each_entry(event, &leader->sibling_list, group_entry)
533 update_event_times(event);
537 * Cross CPU call to disable a performance event
539 static void __perf_event_disable(void *info)
541 struct perf_event *event = info;
542 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
543 struct perf_event_context *ctx = event->ctx;
546 * If this is a per-task event, need to check whether this
547 * event's task is the current task on this cpu.
549 if (ctx->task && cpuctx->task_ctx != ctx)
552 raw_spin_lock(&ctx->lock);
555 * If the event is on, turn it off.
556 * If it is in error state, leave it in error state.
558 if (event->state >= PERF_EVENT_STATE_INACTIVE) {
559 update_context_time(ctx);
560 update_group_times(event);
561 if (event == event->group_leader)
562 group_sched_out(event, cpuctx, ctx);
564 event_sched_out(event, cpuctx, ctx);
565 event->state = PERF_EVENT_STATE_OFF;
568 raw_spin_unlock(&ctx->lock);
574 * If event->ctx is a cloned context, callers must make sure that
575 * every task struct that event->ctx->task could possibly point to
576 * remains valid. This condition is satisifed when called through
577 * perf_event_for_each_child or perf_event_for_each because they
578 * hold the top-level event's child_mutex, so any descendant that
579 * goes to exit will block in sync_child_event.
580 * When called from perf_pending_event it's OK because event->ctx
581 * is the current context on this CPU and preemption is disabled,
582 * hence we can't get into perf_event_task_sched_out for this context.
584 void perf_event_disable(struct perf_event *event)
586 struct perf_event_context *ctx = event->ctx;
587 struct task_struct *task = ctx->task;
591 * Disable the event on the cpu that it's on
593 smp_call_function_single(event->cpu, __perf_event_disable,
599 task_oncpu_function_call(task, __perf_event_disable, event);
601 raw_spin_lock_irq(&ctx->lock);
603 * If the event is still active, we need to retry the cross-call.
605 if (event->state == PERF_EVENT_STATE_ACTIVE) {
606 raw_spin_unlock_irq(&ctx->lock);
611 * Since we have the lock this context can't be scheduled
612 * in, so we can change the state safely.
614 if (event->state == PERF_EVENT_STATE_INACTIVE) {
615 update_group_times(event);
616 event->state = PERF_EVENT_STATE_OFF;
619 raw_spin_unlock_irq(&ctx->lock);
623 event_sched_in(struct perf_event *event,
624 struct perf_cpu_context *cpuctx,
625 struct perf_event_context *ctx,
628 if (event->state <= PERF_EVENT_STATE_OFF)
631 event->state = PERF_EVENT_STATE_ACTIVE;
632 event->oncpu = cpu; /* TODO: put 'cpu' into cpuctx->cpu */
634 * The new state must be visible before we turn it on in the hardware:
638 if (event->pmu->enable(event)) {
639 event->state = PERF_EVENT_STATE_INACTIVE;
644 event->tstamp_running += ctx->time - event->tstamp_stopped;
646 if (!is_software_event(event))
647 cpuctx->active_oncpu++;
650 if (event->attr.exclusive)
651 cpuctx->exclusive = 1;
657 group_sched_in(struct perf_event *group_event,
658 struct perf_cpu_context *cpuctx,
659 struct perf_event_context *ctx,
662 struct perf_event *event, *partial_group;
665 if (group_event->state == PERF_EVENT_STATE_OFF)
668 ret = hw_perf_group_sched_in(group_event, cpuctx, ctx, cpu);
670 return ret < 0 ? ret : 0;
672 if (event_sched_in(group_event, cpuctx, ctx, cpu))
676 * Schedule in siblings as one group (if any):
678 list_for_each_entry(event, &group_event->sibling_list, group_entry) {
679 if (event_sched_in(event, cpuctx, ctx, cpu)) {
680 partial_group = event;
689 * Groups can be scheduled in as one unit only, so undo any
690 * partial group before returning:
692 list_for_each_entry(event, &group_event->sibling_list, group_entry) {
693 if (event == partial_group)
695 event_sched_out(event, cpuctx, ctx);
697 event_sched_out(group_event, cpuctx, ctx);
703 * Return 1 for a group consisting entirely of software events,
704 * 0 if the group contains any hardware events.
706 static int is_software_only_group(struct perf_event *leader)
708 struct perf_event *event;
710 if (!is_software_event(leader))
713 list_for_each_entry(event, &leader->sibling_list, group_entry)
714 if (!is_software_event(event))
721 * Work out whether we can put this event group on the CPU now.
723 static int group_can_go_on(struct perf_event *event,
724 struct perf_cpu_context *cpuctx,
728 * Groups consisting entirely of software events can always go on.
730 if (is_software_only_group(event))
733 * If an exclusive group is already on, no other hardware
736 if (cpuctx->exclusive)
739 * If this group is exclusive and there are already
740 * events on the CPU, it can't go on.
742 if (event->attr.exclusive && cpuctx->active_oncpu)
745 * Otherwise, try to add it if all previous groups were able
751 static void add_event_to_ctx(struct perf_event *event,
752 struct perf_event_context *ctx)
754 list_add_event(event, ctx);
755 event->tstamp_enabled = ctx->time;
756 event->tstamp_running = ctx->time;
757 event->tstamp_stopped = ctx->time;
761 * Cross CPU call to install and enable a performance event
763 * Must be called with ctx->mutex held
765 static void __perf_install_in_context(void *info)
767 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
768 struct perf_event *event = info;
769 struct perf_event_context *ctx = event->ctx;
770 struct perf_event *leader = event->group_leader;
771 int cpu = smp_processor_id();
775 * If this is a task context, we need to check whether it is
776 * the current task context of this cpu. If not it has been
777 * scheduled out before the smp call arrived.
778 * Or possibly this is the right context but it isn't
779 * on this cpu because it had no events.
781 if (ctx->task && cpuctx->task_ctx != ctx) {
782 if (cpuctx->task_ctx || ctx->task != current)
784 cpuctx->task_ctx = ctx;
787 raw_spin_lock(&ctx->lock);
789 update_context_time(ctx);
792 * Protect the list operation against NMI by disabling the
793 * events on a global level. NOP for non NMI based events.
797 add_event_to_ctx(event, ctx);
799 if (event->cpu != -1 && event->cpu != smp_processor_id())
803 * Don't put the event on if it is disabled or if
804 * it is in a group and the group isn't on.
806 if (event->state != PERF_EVENT_STATE_INACTIVE ||
807 (leader != event && leader->state != PERF_EVENT_STATE_ACTIVE))
811 * An exclusive event can't go on if there are already active
812 * hardware events, and no hardware event can go on if there
813 * is already an exclusive event on.
815 if (!group_can_go_on(event, cpuctx, 1))
818 err = event_sched_in(event, cpuctx, ctx, cpu);
822 * This event couldn't go on. If it is in a group
823 * then we have to pull the whole group off.
824 * If the event group is pinned then put it in error state.
827 group_sched_out(leader, cpuctx, ctx);
828 if (leader->attr.pinned) {
829 update_group_times(leader);
830 leader->state = PERF_EVENT_STATE_ERROR;
834 if (!err && !ctx->task && cpuctx->max_pertask)
835 cpuctx->max_pertask--;
840 raw_spin_unlock(&ctx->lock);
844 * Attach a performance event to a context
846 * First we add the event to the list with the hardware enable bit
847 * in event->hw_config cleared.
849 * If the event is attached to a task which is on a CPU we use a smp
850 * call to enable it in the task context. The task might have been
851 * scheduled away, but we check this in the smp call again.
853 * Must be called with ctx->mutex held.
856 perf_install_in_context(struct perf_event_context *ctx,
857 struct perf_event *event,
860 struct task_struct *task = ctx->task;
864 * Per cpu events are installed via an smp call and
865 * the install is always successful.
867 smp_call_function_single(cpu, __perf_install_in_context,
873 task_oncpu_function_call(task, __perf_install_in_context,
876 raw_spin_lock_irq(&ctx->lock);
878 * we need to retry the smp call.
880 if (ctx->is_active && list_empty(&event->group_entry)) {
881 raw_spin_unlock_irq(&ctx->lock);
886 * The lock prevents that this context is scheduled in so we
887 * can add the event safely, if it the call above did not
890 if (list_empty(&event->group_entry))
891 add_event_to_ctx(event, ctx);
892 raw_spin_unlock_irq(&ctx->lock);
896 * Put a event into inactive state and update time fields.
897 * Enabling the leader of a group effectively enables all
898 * the group members that aren't explicitly disabled, so we
899 * have to update their ->tstamp_enabled also.
900 * Note: this works for group members as well as group leaders
901 * since the non-leader members' sibling_lists will be empty.
903 static void __perf_event_mark_enabled(struct perf_event *event,
904 struct perf_event_context *ctx)
906 struct perf_event *sub;
908 event->state = PERF_EVENT_STATE_INACTIVE;
909 event->tstamp_enabled = ctx->time - event->total_time_enabled;
910 list_for_each_entry(sub, &event->sibling_list, group_entry)
911 if (sub->state >= PERF_EVENT_STATE_INACTIVE)
912 sub->tstamp_enabled =
913 ctx->time - sub->total_time_enabled;
917 * Cross CPU call to enable a performance event
919 static void __perf_event_enable(void *info)
921 struct perf_event *event = info;
922 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
923 struct perf_event_context *ctx = event->ctx;
924 struct perf_event *leader = event->group_leader;
928 * If this is a per-task event, need to check whether this
929 * event's task is the current task on this cpu.
931 if (ctx->task && cpuctx->task_ctx != ctx) {
932 if (cpuctx->task_ctx || ctx->task != current)
934 cpuctx->task_ctx = ctx;
937 raw_spin_lock(&ctx->lock);
939 update_context_time(ctx);
941 if (event->state >= PERF_EVENT_STATE_INACTIVE)
943 __perf_event_mark_enabled(event, ctx);
945 if (event->cpu != -1 && event->cpu != smp_processor_id())
949 * If the event is in a group and isn't the group leader,
950 * then don't put it on unless the group is on.
952 if (leader != event && leader->state != PERF_EVENT_STATE_ACTIVE)
955 if (!group_can_go_on(event, cpuctx, 1)) {
960 err = group_sched_in(event, cpuctx, ctx,
963 err = event_sched_in(event, cpuctx, ctx,
970 * If this event can't go on and it's part of a
971 * group, then the whole group has to come off.
974 group_sched_out(leader, cpuctx, ctx);
975 if (leader->attr.pinned) {
976 update_group_times(leader);
977 leader->state = PERF_EVENT_STATE_ERROR;
982 raw_spin_unlock(&ctx->lock);
988 * If event->ctx is a cloned context, callers must make sure that
989 * every task struct that event->ctx->task could possibly point to
990 * remains valid. This condition is satisfied when called through
991 * perf_event_for_each_child or perf_event_for_each as described
992 * for perf_event_disable.
994 void perf_event_enable(struct perf_event *event)
996 struct perf_event_context *ctx = event->ctx;
997 struct task_struct *task = ctx->task;
1001 * Enable the event on the cpu that it's on
1003 smp_call_function_single(event->cpu, __perf_event_enable,
1008 raw_spin_lock_irq(&ctx->lock);
1009 if (event->state >= PERF_EVENT_STATE_INACTIVE)
1013 * If the event is in error state, clear that first.
1014 * That way, if we see the event in error state below, we
1015 * know that it has gone back into error state, as distinct
1016 * from the task having been scheduled away before the
1017 * cross-call arrived.
1019 if (event->state == PERF_EVENT_STATE_ERROR)
1020 event->state = PERF_EVENT_STATE_OFF;
1023 raw_spin_unlock_irq(&ctx->lock);
1024 task_oncpu_function_call(task, __perf_event_enable, event);
1026 raw_spin_lock_irq(&ctx->lock);
1029 * If the context is active and the event is still off,
1030 * we need to retry the cross-call.
1032 if (ctx->is_active && event->state == PERF_EVENT_STATE_OFF)
1036 * Since we have the lock this context can't be scheduled
1037 * in, so we can change the state safely.
1039 if (event->state == PERF_EVENT_STATE_OFF)
1040 __perf_event_mark_enabled(event, ctx);
1043 raw_spin_unlock_irq(&ctx->lock);
1046 static int perf_event_refresh(struct perf_event *event, int refresh)
1049 * not supported on inherited events
1051 if (event->attr.inherit)
1054 atomic_add(refresh, &event->event_limit);
1055 perf_event_enable(event);
1060 void __perf_event_sched_out(struct perf_event_context *ctx,
1061 struct perf_cpu_context *cpuctx)
1063 struct perf_event *event;
1065 raw_spin_lock(&ctx->lock);
1067 if (likely(!ctx->nr_events))
1069 update_context_time(ctx);
1072 if (ctx->nr_active) {
1073 list_for_each_entry(event, &ctx->pinned_groups, group_entry)
1074 group_sched_out(event, cpuctx, ctx);
1076 list_for_each_entry(event, &ctx->flexible_groups, group_entry)
1077 group_sched_out(event, cpuctx, ctx);
1081 raw_spin_unlock(&ctx->lock);
1085 * Test whether two contexts are equivalent, i.e. whether they
1086 * have both been cloned from the same version of the same context
1087 * and they both have the same number of enabled events.
1088 * If the number of enabled events is the same, then the set
1089 * of enabled events should be the same, because these are both
1090 * inherited contexts, therefore we can't access individual events
1091 * in them directly with an fd; we can only enable/disable all
1092 * events via prctl, or enable/disable all events in a family
1093 * via ioctl, which will have the same effect on both contexts.
1095 static int context_equiv(struct perf_event_context *ctx1,
1096 struct perf_event_context *ctx2)
1098 return ctx1->parent_ctx && ctx1->parent_ctx == ctx2->parent_ctx
1099 && ctx1->parent_gen == ctx2->parent_gen
1100 && !ctx1->pin_count && !ctx2->pin_count;
1103 static void __perf_event_sync_stat(struct perf_event *event,
1104 struct perf_event *next_event)
1108 if (!event->attr.inherit_stat)
1112 * Update the event value, we cannot use perf_event_read()
1113 * because we're in the middle of a context switch and have IRQs
1114 * disabled, which upsets smp_call_function_single(), however
1115 * we know the event must be on the current CPU, therefore we
1116 * don't need to use it.
1118 switch (event->state) {
1119 case PERF_EVENT_STATE_ACTIVE:
1120 event->pmu->read(event);
1123 case PERF_EVENT_STATE_INACTIVE:
1124 update_event_times(event);
1132 * In order to keep per-task stats reliable we need to flip the event
1133 * values when we flip the contexts.
1135 value = atomic64_read(&next_event->count);
1136 value = atomic64_xchg(&event->count, value);
1137 atomic64_set(&next_event->count, value);
1139 swap(event->total_time_enabled, next_event->total_time_enabled);
1140 swap(event->total_time_running, next_event->total_time_running);
1143 * Since we swizzled the values, update the user visible data too.
1145 perf_event_update_userpage(event);
1146 perf_event_update_userpage(next_event);
1149 #define list_next_entry(pos, member) \
1150 list_entry(pos->member.next, typeof(*pos), member)
1152 static void perf_event_sync_stat(struct perf_event_context *ctx,
1153 struct perf_event_context *next_ctx)
1155 struct perf_event *event, *next_event;
1160 update_context_time(ctx);
1162 event = list_first_entry(&ctx->event_list,
1163 struct perf_event, event_entry);
1165 next_event = list_first_entry(&next_ctx->event_list,
1166 struct perf_event, event_entry);
1168 while (&event->event_entry != &ctx->event_list &&
1169 &next_event->event_entry != &next_ctx->event_list) {
1171 __perf_event_sync_stat(event, next_event);
1173 event = list_next_entry(event, event_entry);
1174 next_event = list_next_entry(next_event, event_entry);
1179 * Called from scheduler to remove the events of the current task,
1180 * with interrupts disabled.
1182 * We stop each event and update the event value in event->count.
1184 * This does not protect us against NMI, but disable()
1185 * sets the disabled bit in the control field of event _before_
1186 * accessing the event control register. If a NMI hits, then it will
1187 * not restart the event.
1189 void perf_event_task_sched_out(struct task_struct *task,
1190 struct task_struct *next)
1192 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
1193 struct perf_event_context *ctx = task->perf_event_ctxp;
1194 struct perf_event_context *next_ctx;
1195 struct perf_event_context *parent;
1196 struct pt_regs *regs;
1199 regs = task_pt_regs(task);
1200 perf_sw_event(PERF_COUNT_SW_CONTEXT_SWITCHES, 1, 1, regs, 0);
1202 if (likely(!ctx || !cpuctx->task_ctx))
1206 parent = rcu_dereference(ctx->parent_ctx);
1207 next_ctx = next->perf_event_ctxp;
1208 if (parent && next_ctx &&
1209 rcu_dereference(next_ctx->parent_ctx) == parent) {
1211 * Looks like the two contexts are clones, so we might be
1212 * able to optimize the context switch. We lock both
1213 * contexts and check that they are clones under the
1214 * lock (including re-checking that neither has been
1215 * uncloned in the meantime). It doesn't matter which
1216 * order we take the locks because no other cpu could
1217 * be trying to lock both of these tasks.
1219 raw_spin_lock(&ctx->lock);
1220 raw_spin_lock_nested(&next_ctx->lock, SINGLE_DEPTH_NESTING);
1221 if (context_equiv(ctx, next_ctx)) {
1223 * XXX do we need a memory barrier of sorts
1224 * wrt to rcu_dereference() of perf_event_ctxp
1226 task->perf_event_ctxp = next_ctx;
1227 next->perf_event_ctxp = ctx;
1229 next_ctx->task = task;
1232 perf_event_sync_stat(ctx, next_ctx);
1234 raw_spin_unlock(&next_ctx->lock);
1235 raw_spin_unlock(&ctx->lock);
1240 __perf_event_sched_out(ctx, cpuctx);
1241 cpuctx->task_ctx = NULL;
1246 * Called with IRQs disabled
1248 static void __perf_event_task_sched_out(struct perf_event_context *ctx)
1250 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
1252 if (!cpuctx->task_ctx)
1255 if (WARN_ON_ONCE(ctx != cpuctx->task_ctx))
1258 __perf_event_sched_out(ctx, cpuctx);
1259 cpuctx->task_ctx = NULL;
1263 * Called with IRQs disabled
1265 static void perf_event_cpu_sched_out(struct perf_cpu_context *cpuctx)
1267 __perf_event_sched_out(&cpuctx->ctx, cpuctx);
1271 __perf_event_sched_in(struct perf_event_context *ctx,
1272 struct perf_cpu_context *cpuctx)
1274 int cpu = smp_processor_id();
1275 struct perf_event *event;
1278 raw_spin_lock(&ctx->lock);
1280 if (likely(!ctx->nr_events))
1283 ctx->timestamp = perf_clock();
1288 * First go through the list and put on any pinned groups
1289 * in order to give them the best chance of going on.
1291 list_for_each_entry(event, &ctx->pinned_groups, group_entry) {
1292 if (event->state <= PERF_EVENT_STATE_OFF)
1294 if (event->cpu != -1 && event->cpu != cpu)
1297 if (group_can_go_on(event, cpuctx, 1))
1298 group_sched_in(event, cpuctx, ctx, cpu);
1301 * If this pinned group hasn't been scheduled,
1302 * put it in error state.
1304 if (event->state == PERF_EVENT_STATE_INACTIVE) {
1305 update_group_times(event);
1306 event->state = PERF_EVENT_STATE_ERROR;
1310 list_for_each_entry(event, &ctx->flexible_groups, group_entry) {
1311 /* Ignore events in OFF or ERROR state */
1312 if (event->state <= PERF_EVENT_STATE_OFF)
1315 * Listen to the 'cpu' scheduling filter constraint
1318 if (event->cpu != -1 && event->cpu != cpu)
1321 if (group_can_go_on(event, cpuctx, can_add_hw))
1322 if (group_sched_in(event, cpuctx, ctx, cpu))
1327 raw_spin_unlock(&ctx->lock);
1331 * Called from scheduler to add the events of the current task
1332 * with interrupts disabled.
1334 * We restore the event value and then enable it.
1336 * This does not protect us against NMI, but enable()
1337 * sets the enabled bit in the control field of event _before_
1338 * accessing the event control register. If a NMI hits, then it will
1339 * keep the event running.
1341 void perf_event_task_sched_in(struct task_struct *task)
1343 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
1344 struct perf_event_context *ctx = task->perf_event_ctxp;
1348 if (cpuctx->task_ctx == ctx)
1350 __perf_event_sched_in(ctx, cpuctx);
1351 cpuctx->task_ctx = ctx;
1354 static void perf_event_cpu_sched_in(struct perf_cpu_context *cpuctx)
1356 struct perf_event_context *ctx = &cpuctx->ctx;
1358 __perf_event_sched_in(ctx, cpuctx);
1361 #define MAX_INTERRUPTS (~0ULL)
1363 static void perf_log_throttle(struct perf_event *event, int enable);
1365 static void perf_adjust_period(struct perf_event *event, u64 events)
1367 struct hw_perf_event *hwc = &event->hw;
1368 u64 period, sample_period;
1371 events *= hwc->sample_period;
1372 period = div64_u64(events, event->attr.sample_freq);
1374 delta = (s64)(period - hwc->sample_period);
1375 delta = (delta + 7) / 8; /* low pass filter */
1377 sample_period = hwc->sample_period + delta;
1382 hwc->sample_period = sample_period;
1385 static void perf_ctx_adjust_freq(struct perf_event_context *ctx)
1387 struct perf_event *event;
1388 struct hw_perf_event *hwc;
1389 u64 interrupts, freq;
1391 raw_spin_lock(&ctx->lock);
1392 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
1393 if (event->state != PERF_EVENT_STATE_ACTIVE)
1396 if (event->cpu != -1 && event->cpu != smp_processor_id())
1401 interrupts = hwc->interrupts;
1402 hwc->interrupts = 0;
1405 * unthrottle events on the tick
1407 if (interrupts == MAX_INTERRUPTS) {
1408 perf_log_throttle(event, 1);
1409 event->pmu->unthrottle(event);
1410 interrupts = 2*sysctl_perf_event_sample_rate/HZ;
1413 if (!event->attr.freq || !event->attr.sample_freq)
1417 * if the specified freq < HZ then we need to skip ticks
1419 if (event->attr.sample_freq < HZ) {
1420 freq = event->attr.sample_freq;
1422 hwc->freq_count += freq;
1423 hwc->freq_interrupts += interrupts;
1425 if (hwc->freq_count < HZ)
1428 interrupts = hwc->freq_interrupts;
1429 hwc->freq_interrupts = 0;
1430 hwc->freq_count -= HZ;
1434 perf_adjust_period(event, freq * interrupts);
1437 * In order to avoid being stalled by an (accidental) huge
1438 * sample period, force reset the sample period if we didn't
1439 * get any events in this freq period.
1443 event->pmu->disable(event);
1444 atomic64_set(&hwc->period_left, 0);
1445 event->pmu->enable(event);
1449 raw_spin_unlock(&ctx->lock);
1453 * Round-robin a context's events:
1455 static void rotate_ctx(struct perf_event_context *ctx)
1457 if (!ctx->nr_events)
1460 raw_spin_lock(&ctx->lock);
1462 /* Rotate the first entry last of non-pinned groups */
1465 list_rotate_left(&ctx->flexible_groups);
1469 raw_spin_unlock(&ctx->lock);
1472 void perf_event_task_tick(struct task_struct *curr)
1474 struct perf_cpu_context *cpuctx;
1475 struct perf_event_context *ctx;
1477 if (!atomic_read(&nr_events))
1480 cpuctx = &__get_cpu_var(perf_cpu_context);
1481 ctx = curr->perf_event_ctxp;
1483 perf_ctx_adjust_freq(&cpuctx->ctx);
1485 perf_ctx_adjust_freq(ctx);
1487 perf_event_cpu_sched_out(cpuctx);
1489 __perf_event_task_sched_out(ctx);
1491 rotate_ctx(&cpuctx->ctx);
1495 perf_event_cpu_sched_in(cpuctx);
1497 perf_event_task_sched_in(curr);
1500 static int event_enable_on_exec(struct perf_event *event,
1501 struct perf_event_context *ctx)
1503 if (!event->attr.enable_on_exec)
1506 event->attr.enable_on_exec = 0;
1507 if (event->state >= PERF_EVENT_STATE_INACTIVE)
1510 __perf_event_mark_enabled(event, ctx);
1516 * Enable all of a task's events that have been marked enable-on-exec.
1517 * This expects task == current.
1519 static void perf_event_enable_on_exec(struct task_struct *task)
1521 struct perf_event_context *ctx;
1522 struct perf_event *event;
1523 unsigned long flags;
1527 local_irq_save(flags);
1528 ctx = task->perf_event_ctxp;
1529 if (!ctx || !ctx->nr_events)
1532 __perf_event_task_sched_out(ctx);
1534 raw_spin_lock(&ctx->lock);
1536 list_for_each_entry(event, &ctx->pinned_groups, group_entry) {
1537 ret = event_enable_on_exec(event, ctx);
1542 list_for_each_entry(event, &ctx->flexible_groups, group_entry) {
1543 ret = event_enable_on_exec(event, ctx);
1549 * Unclone this context if we enabled any event.
1554 raw_spin_unlock(&ctx->lock);
1556 perf_event_task_sched_in(task);
1558 local_irq_restore(flags);
1562 * Cross CPU call to read the hardware event
1564 static void __perf_event_read(void *info)
1566 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
1567 struct perf_event *event = info;
1568 struct perf_event_context *ctx = event->ctx;
1571 * If this is a task context, we need to check whether it is
1572 * the current task context of this cpu. If not it has been
1573 * scheduled out before the smp call arrived. In that case
1574 * event->count would have been updated to a recent sample
1575 * when the event was scheduled out.
1577 if (ctx->task && cpuctx->task_ctx != ctx)
1580 raw_spin_lock(&ctx->lock);
1581 update_context_time(ctx);
1582 update_event_times(event);
1583 raw_spin_unlock(&ctx->lock);
1585 event->pmu->read(event);
1588 static u64 perf_event_read(struct perf_event *event)
1591 * If event is enabled and currently active on a CPU, update the
1592 * value in the event structure:
1594 if (event->state == PERF_EVENT_STATE_ACTIVE) {
1595 smp_call_function_single(event->oncpu,
1596 __perf_event_read, event, 1);
1597 } else if (event->state == PERF_EVENT_STATE_INACTIVE) {
1598 struct perf_event_context *ctx = event->ctx;
1599 unsigned long flags;
1601 raw_spin_lock_irqsave(&ctx->lock, flags);
1602 update_context_time(ctx);
1603 update_event_times(event);
1604 raw_spin_unlock_irqrestore(&ctx->lock, flags);
1607 return atomic64_read(&event->count);
1611 * Initialize the perf_event context in a task_struct:
1614 __perf_event_init_context(struct perf_event_context *ctx,
1615 struct task_struct *task)
1617 raw_spin_lock_init(&ctx->lock);
1618 mutex_init(&ctx->mutex);
1619 INIT_LIST_HEAD(&ctx->pinned_groups);
1620 INIT_LIST_HEAD(&ctx->flexible_groups);
1621 INIT_LIST_HEAD(&ctx->event_list);
1622 atomic_set(&ctx->refcount, 1);
1626 static struct perf_event_context *find_get_context(pid_t pid, int cpu)
1628 struct perf_event_context *ctx;
1629 struct perf_cpu_context *cpuctx;
1630 struct task_struct *task;
1631 unsigned long flags;
1634 if (pid == -1 && cpu != -1) {
1635 /* Must be root to operate on a CPU event: */
1636 if (perf_paranoid_cpu() && !capable(CAP_SYS_ADMIN))
1637 return ERR_PTR(-EACCES);
1639 if (cpu < 0 || cpu >= nr_cpumask_bits)
1640 return ERR_PTR(-EINVAL);
1643 * We could be clever and allow to attach a event to an
1644 * offline CPU and activate it when the CPU comes up, but
1647 if (!cpu_online(cpu))
1648 return ERR_PTR(-ENODEV);
1650 cpuctx = &per_cpu(perf_cpu_context, cpu);
1661 task = find_task_by_vpid(pid);
1663 get_task_struct(task);
1667 return ERR_PTR(-ESRCH);
1670 * Can't attach events to a dying task.
1673 if (task->flags & PF_EXITING)
1676 /* Reuse ptrace permission checks for now. */
1678 if (!ptrace_may_access(task, PTRACE_MODE_READ))
1682 ctx = perf_lock_task_context(task, &flags);
1685 raw_spin_unlock_irqrestore(&ctx->lock, flags);
1689 ctx = kzalloc(sizeof(struct perf_event_context), GFP_KERNEL);
1693 __perf_event_init_context(ctx, task);
1695 if (cmpxchg(&task->perf_event_ctxp, NULL, ctx)) {
1697 * We raced with some other task; use
1698 * the context they set.
1703 get_task_struct(task);
1706 put_task_struct(task);
1710 put_task_struct(task);
1711 return ERR_PTR(err);
1714 static void perf_event_free_filter(struct perf_event *event);
1716 static void free_event_rcu(struct rcu_head *head)
1718 struct perf_event *event;
1720 event = container_of(head, struct perf_event, rcu_head);
1722 put_pid_ns(event->ns);
1723 perf_event_free_filter(event);
1727 static void perf_pending_sync(struct perf_event *event);
1729 static void free_event(struct perf_event *event)
1731 perf_pending_sync(event);
1733 if (!event->parent) {
1734 atomic_dec(&nr_events);
1735 if (event->attr.mmap)
1736 atomic_dec(&nr_mmap_events);
1737 if (event->attr.comm)
1738 atomic_dec(&nr_comm_events);
1739 if (event->attr.task)
1740 atomic_dec(&nr_task_events);
1743 if (event->output) {
1744 fput(event->output->filp);
1745 event->output = NULL;
1749 event->destroy(event);
1751 put_ctx(event->ctx);
1752 call_rcu(&event->rcu_head, free_event_rcu);
1755 int perf_event_release_kernel(struct perf_event *event)
1757 struct perf_event_context *ctx = event->ctx;
1759 WARN_ON_ONCE(ctx->parent_ctx);
1760 mutex_lock(&ctx->mutex);
1761 perf_event_remove_from_context(event);
1762 mutex_unlock(&ctx->mutex);
1764 mutex_lock(&event->owner->perf_event_mutex);
1765 list_del_init(&event->owner_entry);
1766 mutex_unlock(&event->owner->perf_event_mutex);
1767 put_task_struct(event->owner);
1773 EXPORT_SYMBOL_GPL(perf_event_release_kernel);
1776 * Called when the last reference to the file is gone.
1778 static int perf_release(struct inode *inode, struct file *file)
1780 struct perf_event *event = file->private_data;
1782 file->private_data = NULL;
1784 return perf_event_release_kernel(event);
1787 static int perf_event_read_size(struct perf_event *event)
1789 int entry = sizeof(u64); /* value */
1793 if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1794 size += sizeof(u64);
1796 if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1797 size += sizeof(u64);
1799 if (event->attr.read_format & PERF_FORMAT_ID)
1800 entry += sizeof(u64);
1802 if (event->attr.read_format & PERF_FORMAT_GROUP) {
1803 nr += event->group_leader->nr_siblings;
1804 size += sizeof(u64);
1812 u64 perf_event_read_value(struct perf_event *event, u64 *enabled, u64 *running)
1814 struct perf_event *child;
1820 mutex_lock(&event->child_mutex);
1821 total += perf_event_read(event);
1822 *enabled += event->total_time_enabled +
1823 atomic64_read(&event->child_total_time_enabled);
1824 *running += event->total_time_running +
1825 atomic64_read(&event->child_total_time_running);
1827 list_for_each_entry(child, &event->child_list, child_list) {
1828 total += perf_event_read(child);
1829 *enabled += child->total_time_enabled;
1830 *running += child->total_time_running;
1832 mutex_unlock(&event->child_mutex);
1836 EXPORT_SYMBOL_GPL(perf_event_read_value);
1838 static int perf_event_read_group(struct perf_event *event,
1839 u64 read_format, char __user *buf)
1841 struct perf_event *leader = event->group_leader, *sub;
1842 int n = 0, size = 0, ret = -EFAULT;
1843 struct perf_event_context *ctx = leader->ctx;
1845 u64 count, enabled, running;
1847 mutex_lock(&ctx->mutex);
1848 count = perf_event_read_value(leader, &enabled, &running);
1850 values[n++] = 1 + leader->nr_siblings;
1851 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1852 values[n++] = enabled;
1853 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1854 values[n++] = running;
1855 values[n++] = count;
1856 if (read_format & PERF_FORMAT_ID)
1857 values[n++] = primary_event_id(leader);
1859 size = n * sizeof(u64);
1861 if (copy_to_user(buf, values, size))
1866 list_for_each_entry(sub, &leader->sibling_list, group_entry) {
1869 values[n++] = perf_event_read_value(sub, &enabled, &running);
1870 if (read_format & PERF_FORMAT_ID)
1871 values[n++] = primary_event_id(sub);
1873 size = n * sizeof(u64);
1875 if (copy_to_user(buf + ret, values, size)) {
1883 mutex_unlock(&ctx->mutex);
1888 static int perf_event_read_one(struct perf_event *event,
1889 u64 read_format, char __user *buf)
1891 u64 enabled, running;
1895 values[n++] = perf_event_read_value(event, &enabled, &running);
1896 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1897 values[n++] = enabled;
1898 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1899 values[n++] = running;
1900 if (read_format & PERF_FORMAT_ID)
1901 values[n++] = primary_event_id(event);
1903 if (copy_to_user(buf, values, n * sizeof(u64)))
1906 return n * sizeof(u64);
1910 * Read the performance event - simple non blocking version for now
1913 perf_read_hw(struct perf_event *event, char __user *buf, size_t count)
1915 u64 read_format = event->attr.read_format;
1919 * Return end-of-file for a read on a event that is in
1920 * error state (i.e. because it was pinned but it couldn't be
1921 * scheduled on to the CPU at some point).
1923 if (event->state == PERF_EVENT_STATE_ERROR)
1926 if (count < perf_event_read_size(event))
1929 WARN_ON_ONCE(event->ctx->parent_ctx);
1930 if (read_format & PERF_FORMAT_GROUP)
1931 ret = perf_event_read_group(event, read_format, buf);
1933 ret = perf_event_read_one(event, read_format, buf);
1939 perf_read(struct file *file, char __user *buf, size_t count, loff_t *ppos)
1941 struct perf_event *event = file->private_data;
1943 return perf_read_hw(event, buf, count);
1946 static unsigned int perf_poll(struct file *file, poll_table *wait)
1948 struct perf_event *event = file->private_data;
1949 struct perf_mmap_data *data;
1950 unsigned int events = POLL_HUP;
1953 data = rcu_dereference(event->data);
1955 events = atomic_xchg(&data->poll, 0);
1958 poll_wait(file, &event->waitq, wait);
1963 static void perf_event_reset(struct perf_event *event)
1965 (void)perf_event_read(event);
1966 atomic64_set(&event->count, 0);
1967 perf_event_update_userpage(event);
1971 * Holding the top-level event's child_mutex means that any
1972 * descendant process that has inherited this event will block
1973 * in sync_child_event if it goes to exit, thus satisfying the
1974 * task existence requirements of perf_event_enable/disable.
1976 static void perf_event_for_each_child(struct perf_event *event,
1977 void (*func)(struct perf_event *))
1979 struct perf_event *child;
1981 WARN_ON_ONCE(event->ctx->parent_ctx);
1982 mutex_lock(&event->child_mutex);
1984 list_for_each_entry(child, &event->child_list, child_list)
1986 mutex_unlock(&event->child_mutex);
1989 static void perf_event_for_each(struct perf_event *event,
1990 void (*func)(struct perf_event *))
1992 struct perf_event_context *ctx = event->ctx;
1993 struct perf_event *sibling;
1995 WARN_ON_ONCE(ctx->parent_ctx);
1996 mutex_lock(&ctx->mutex);
1997 event = event->group_leader;
1999 perf_event_for_each_child(event, func);
2001 list_for_each_entry(sibling, &event->sibling_list, group_entry)
2002 perf_event_for_each_child(event, func);
2003 mutex_unlock(&ctx->mutex);
2006 static int perf_event_period(struct perf_event *event, u64 __user *arg)
2008 struct perf_event_context *ctx = event->ctx;
2013 if (!event->attr.sample_period)
2016 size = copy_from_user(&value, arg, sizeof(value));
2017 if (size != sizeof(value))
2023 raw_spin_lock_irq(&ctx->lock);
2024 if (event->attr.freq) {
2025 if (value > sysctl_perf_event_sample_rate) {
2030 event->attr.sample_freq = value;
2032 event->attr.sample_period = value;
2033 event->hw.sample_period = value;
2036 raw_spin_unlock_irq(&ctx->lock);
2041 static int perf_event_set_output(struct perf_event *event, int output_fd);
2042 static int perf_event_set_filter(struct perf_event *event, void __user *arg);
2044 static long perf_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
2046 struct perf_event *event = file->private_data;
2047 void (*func)(struct perf_event *);
2051 case PERF_EVENT_IOC_ENABLE:
2052 func = perf_event_enable;
2054 case PERF_EVENT_IOC_DISABLE:
2055 func = perf_event_disable;
2057 case PERF_EVENT_IOC_RESET:
2058 func = perf_event_reset;
2061 case PERF_EVENT_IOC_REFRESH:
2062 return perf_event_refresh(event, arg);
2064 case PERF_EVENT_IOC_PERIOD:
2065 return perf_event_period(event, (u64 __user *)arg);
2067 case PERF_EVENT_IOC_SET_OUTPUT:
2068 return perf_event_set_output(event, arg);
2070 case PERF_EVENT_IOC_SET_FILTER:
2071 return perf_event_set_filter(event, (void __user *)arg);
2077 if (flags & PERF_IOC_FLAG_GROUP)
2078 perf_event_for_each(event, func);
2080 perf_event_for_each_child(event, func);
2085 int perf_event_task_enable(void)
2087 struct perf_event *event;
2089 mutex_lock(¤t->perf_event_mutex);
2090 list_for_each_entry(event, ¤t->perf_event_list, owner_entry)
2091 perf_event_for_each_child(event, perf_event_enable);
2092 mutex_unlock(¤t->perf_event_mutex);
2097 int perf_event_task_disable(void)
2099 struct perf_event *event;
2101 mutex_lock(¤t->perf_event_mutex);
2102 list_for_each_entry(event, ¤t->perf_event_list, owner_entry)
2103 perf_event_for_each_child(event, perf_event_disable);
2104 mutex_unlock(¤t->perf_event_mutex);
2109 #ifndef PERF_EVENT_INDEX_OFFSET
2110 # define PERF_EVENT_INDEX_OFFSET 0
2113 static int perf_event_index(struct perf_event *event)
2115 if (event->state != PERF_EVENT_STATE_ACTIVE)
2118 return event->hw.idx + 1 - PERF_EVENT_INDEX_OFFSET;
2122 * Callers need to ensure there can be no nesting of this function, otherwise
2123 * the seqlock logic goes bad. We can not serialize this because the arch
2124 * code calls this from NMI context.
2126 void perf_event_update_userpage(struct perf_event *event)
2128 struct perf_event_mmap_page *userpg;
2129 struct perf_mmap_data *data;
2132 data = rcu_dereference(event->data);
2136 userpg = data->user_page;
2139 * Disable preemption so as to not let the corresponding user-space
2140 * spin too long if we get preempted.
2145 userpg->index = perf_event_index(event);
2146 userpg->offset = atomic64_read(&event->count);
2147 if (event->state == PERF_EVENT_STATE_ACTIVE)
2148 userpg->offset -= atomic64_read(&event->hw.prev_count);
2150 userpg->time_enabled = event->total_time_enabled +
2151 atomic64_read(&event->child_total_time_enabled);
2153 userpg->time_running = event->total_time_running +
2154 atomic64_read(&event->child_total_time_running);
2163 static unsigned long perf_data_size(struct perf_mmap_data *data)
2165 return data->nr_pages << (PAGE_SHIFT + data->data_order);
2168 #ifndef CONFIG_PERF_USE_VMALLOC
2171 * Back perf_mmap() with regular GFP_KERNEL-0 pages.
2174 static struct page *
2175 perf_mmap_to_page(struct perf_mmap_data *data, unsigned long pgoff)
2177 if (pgoff > data->nr_pages)
2181 return virt_to_page(data->user_page);
2183 return virt_to_page(data->data_pages[pgoff - 1]);
2186 static struct perf_mmap_data *
2187 perf_mmap_data_alloc(struct perf_event *event, int nr_pages)
2189 struct perf_mmap_data *data;
2193 WARN_ON(atomic_read(&event->mmap_count));
2195 size = sizeof(struct perf_mmap_data);
2196 size += nr_pages * sizeof(void *);
2198 data = kzalloc(size, GFP_KERNEL);
2202 data->user_page = (void *)get_zeroed_page(GFP_KERNEL);
2203 if (!data->user_page)
2204 goto fail_user_page;
2206 for (i = 0; i < nr_pages; i++) {
2207 data->data_pages[i] = (void *)get_zeroed_page(GFP_KERNEL);
2208 if (!data->data_pages[i])
2209 goto fail_data_pages;
2212 data->data_order = 0;
2213 data->nr_pages = nr_pages;
2218 for (i--; i >= 0; i--)
2219 free_page((unsigned long)data->data_pages[i]);
2221 free_page((unsigned long)data->user_page);
2230 static void perf_mmap_free_page(unsigned long addr)
2232 struct page *page = virt_to_page((void *)addr);
2234 page->mapping = NULL;
2238 static void perf_mmap_data_free(struct perf_mmap_data *data)
2242 perf_mmap_free_page((unsigned long)data->user_page);
2243 for (i = 0; i < data->nr_pages; i++)
2244 perf_mmap_free_page((unsigned long)data->data_pages[i]);
2251 * Back perf_mmap() with vmalloc memory.
2253 * Required for architectures that have d-cache aliasing issues.
2256 static struct page *
2257 perf_mmap_to_page(struct perf_mmap_data *data, unsigned long pgoff)
2259 if (pgoff > (1UL << data->data_order))
2262 return vmalloc_to_page((void *)data->user_page + pgoff * PAGE_SIZE);
2265 static void perf_mmap_unmark_page(void *addr)
2267 struct page *page = vmalloc_to_page(addr);
2269 page->mapping = NULL;
2272 static void perf_mmap_data_free_work(struct work_struct *work)
2274 struct perf_mmap_data *data;
2278 data = container_of(work, struct perf_mmap_data, work);
2279 nr = 1 << data->data_order;
2281 base = data->user_page;
2282 for (i = 0; i < nr + 1; i++)
2283 perf_mmap_unmark_page(base + (i * PAGE_SIZE));
2289 static void perf_mmap_data_free(struct perf_mmap_data *data)
2291 schedule_work(&data->work);
2294 static struct perf_mmap_data *
2295 perf_mmap_data_alloc(struct perf_event *event, int nr_pages)
2297 struct perf_mmap_data *data;
2301 WARN_ON(atomic_read(&event->mmap_count));
2303 size = sizeof(struct perf_mmap_data);
2304 size += sizeof(void *);
2306 data = kzalloc(size, GFP_KERNEL);
2310 INIT_WORK(&data->work, perf_mmap_data_free_work);
2312 all_buf = vmalloc_user((nr_pages + 1) * PAGE_SIZE);
2316 data->user_page = all_buf;
2317 data->data_pages[0] = all_buf + PAGE_SIZE;
2318 data->data_order = ilog2(nr_pages);
2332 static int perf_mmap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
2334 struct perf_event *event = vma->vm_file->private_data;
2335 struct perf_mmap_data *data;
2336 int ret = VM_FAULT_SIGBUS;
2338 if (vmf->flags & FAULT_FLAG_MKWRITE) {
2339 if (vmf->pgoff == 0)
2345 data = rcu_dereference(event->data);
2349 if (vmf->pgoff && (vmf->flags & FAULT_FLAG_WRITE))
2352 vmf->page = perf_mmap_to_page(data, vmf->pgoff);
2356 get_page(vmf->page);
2357 vmf->page->mapping = vma->vm_file->f_mapping;
2358 vmf->page->index = vmf->pgoff;
2368 perf_mmap_data_init(struct perf_event *event, struct perf_mmap_data *data)
2370 long max_size = perf_data_size(data);
2372 atomic_set(&data->lock, -1);
2374 if (event->attr.watermark) {
2375 data->watermark = min_t(long, max_size,
2376 event->attr.wakeup_watermark);
2379 if (!data->watermark)
2380 data->watermark = max_size / 2;
2383 rcu_assign_pointer(event->data, data);
2386 static void perf_mmap_data_free_rcu(struct rcu_head *rcu_head)
2388 struct perf_mmap_data *data;
2390 data = container_of(rcu_head, struct perf_mmap_data, rcu_head);
2391 perf_mmap_data_free(data);
2394 static void perf_mmap_data_release(struct perf_event *event)
2396 struct perf_mmap_data *data = event->data;
2398 WARN_ON(atomic_read(&event->mmap_count));
2400 rcu_assign_pointer(event->data, NULL);
2401 call_rcu(&data->rcu_head, perf_mmap_data_free_rcu);
2404 static void perf_mmap_open(struct vm_area_struct *vma)
2406 struct perf_event *event = vma->vm_file->private_data;
2408 atomic_inc(&event->mmap_count);
2411 static void perf_mmap_close(struct vm_area_struct *vma)
2413 struct perf_event *event = vma->vm_file->private_data;
2415 WARN_ON_ONCE(event->ctx->parent_ctx);
2416 if (atomic_dec_and_mutex_lock(&event->mmap_count, &event->mmap_mutex)) {
2417 unsigned long size = perf_data_size(event->data);
2418 struct user_struct *user = current_user();
2420 atomic_long_sub((size >> PAGE_SHIFT) + 1, &user->locked_vm);
2421 vma->vm_mm->locked_vm -= event->data->nr_locked;
2422 perf_mmap_data_release(event);
2423 mutex_unlock(&event->mmap_mutex);
2427 static const struct vm_operations_struct perf_mmap_vmops = {
2428 .open = perf_mmap_open,
2429 .close = perf_mmap_close,
2430 .fault = perf_mmap_fault,
2431 .page_mkwrite = perf_mmap_fault,
2434 static int perf_mmap(struct file *file, struct vm_area_struct *vma)
2436 struct perf_event *event = file->private_data;
2437 unsigned long user_locked, user_lock_limit;
2438 struct user_struct *user = current_user();
2439 unsigned long locked, lock_limit;
2440 struct perf_mmap_data *data;
2441 unsigned long vma_size;
2442 unsigned long nr_pages;
2443 long user_extra, extra;
2446 if (!(vma->vm_flags & VM_SHARED))
2449 vma_size = vma->vm_end - vma->vm_start;
2450 nr_pages = (vma_size / PAGE_SIZE) - 1;
2453 * If we have data pages ensure they're a power-of-two number, so we
2454 * can do bitmasks instead of modulo.
2456 if (nr_pages != 0 && !is_power_of_2(nr_pages))
2459 if (vma_size != PAGE_SIZE * (1 + nr_pages))
2462 if (vma->vm_pgoff != 0)
2465 WARN_ON_ONCE(event->ctx->parent_ctx);
2466 mutex_lock(&event->mmap_mutex);
2467 if (event->output) {
2472 if (atomic_inc_not_zero(&event->mmap_count)) {
2473 if (nr_pages != event->data->nr_pages)
2478 user_extra = nr_pages + 1;
2479 user_lock_limit = sysctl_perf_event_mlock >> (PAGE_SHIFT - 10);
2482 * Increase the limit linearly with more CPUs:
2484 user_lock_limit *= num_online_cpus();
2486 user_locked = atomic_long_read(&user->locked_vm) + user_extra;
2489 if (user_locked > user_lock_limit)
2490 extra = user_locked - user_lock_limit;
2492 lock_limit = current->signal->rlim[RLIMIT_MEMLOCK].rlim_cur;
2493 lock_limit >>= PAGE_SHIFT;
2494 locked = vma->vm_mm->locked_vm + extra;
2496 if ((locked > lock_limit) && perf_paranoid_tracepoint_raw() &&
2497 !capable(CAP_IPC_LOCK)) {
2502 WARN_ON(event->data);
2504 data = perf_mmap_data_alloc(event, nr_pages);
2510 perf_mmap_data_init(event, data);
2512 atomic_set(&event->mmap_count, 1);
2513 atomic_long_add(user_extra, &user->locked_vm);
2514 vma->vm_mm->locked_vm += extra;
2515 event->data->nr_locked = extra;
2516 if (vma->vm_flags & VM_WRITE)
2517 event->data->writable = 1;
2520 mutex_unlock(&event->mmap_mutex);
2522 vma->vm_flags |= VM_RESERVED;
2523 vma->vm_ops = &perf_mmap_vmops;
2528 static int perf_fasync(int fd, struct file *filp, int on)
2530 struct inode *inode = filp->f_path.dentry->d_inode;
2531 struct perf_event *event = filp->private_data;
2534 mutex_lock(&inode->i_mutex);
2535 retval = fasync_helper(fd, filp, on, &event->fasync);
2536 mutex_unlock(&inode->i_mutex);
2544 static const struct file_operations perf_fops = {
2545 .release = perf_release,
2548 .unlocked_ioctl = perf_ioctl,
2549 .compat_ioctl = perf_ioctl,
2551 .fasync = perf_fasync,
2557 * If there's data, ensure we set the poll() state and publish everything
2558 * to user-space before waking everybody up.
2561 void perf_event_wakeup(struct perf_event *event)
2563 wake_up_all(&event->waitq);
2565 if (event->pending_kill) {
2566 kill_fasync(&event->fasync, SIGIO, event->pending_kill);
2567 event->pending_kill = 0;
2574 * Handle the case where we need to wakeup up from NMI (or rq->lock) context.
2576 * The NMI bit means we cannot possibly take locks. Therefore, maintain a
2577 * single linked list and use cmpxchg() to add entries lockless.
2580 static void perf_pending_event(struct perf_pending_entry *entry)
2582 struct perf_event *event = container_of(entry,
2583 struct perf_event, pending);
2585 if (event->pending_disable) {
2586 event->pending_disable = 0;
2587 __perf_event_disable(event);
2590 if (event->pending_wakeup) {
2591 event->pending_wakeup = 0;
2592 perf_event_wakeup(event);
2596 #define PENDING_TAIL ((struct perf_pending_entry *)-1UL)
2598 static DEFINE_PER_CPU(struct perf_pending_entry *, perf_pending_head) = {
2602 static void perf_pending_queue(struct perf_pending_entry *entry,
2603 void (*func)(struct perf_pending_entry *))
2605 struct perf_pending_entry **head;
2607 if (cmpxchg(&entry->next, NULL, PENDING_TAIL) != NULL)
2612 head = &get_cpu_var(perf_pending_head);
2615 entry->next = *head;
2616 } while (cmpxchg(head, entry->next, entry) != entry->next);
2618 set_perf_event_pending();
2620 put_cpu_var(perf_pending_head);
2623 static int __perf_pending_run(void)
2625 struct perf_pending_entry *list;
2628 list = xchg(&__get_cpu_var(perf_pending_head), PENDING_TAIL);
2629 while (list != PENDING_TAIL) {
2630 void (*func)(struct perf_pending_entry *);
2631 struct perf_pending_entry *entry = list;
2638 * Ensure we observe the unqueue before we issue the wakeup,
2639 * so that we won't be waiting forever.
2640 * -- see perf_not_pending().
2651 static inline int perf_not_pending(struct perf_event *event)
2654 * If we flush on whatever cpu we run, there is a chance we don't
2658 __perf_pending_run();
2662 * Ensure we see the proper queue state before going to sleep
2663 * so that we do not miss the wakeup. -- see perf_pending_handle()
2666 return event->pending.next == NULL;
2669 static void perf_pending_sync(struct perf_event *event)
2671 wait_event(event->waitq, perf_not_pending(event));
2674 void perf_event_do_pending(void)
2676 __perf_pending_run();
2680 * Callchain support -- arch specific
2683 __weak struct perf_callchain_entry *perf_callchain(struct pt_regs *regs)
2691 static bool perf_output_space(struct perf_mmap_data *data, unsigned long tail,
2692 unsigned long offset, unsigned long head)
2696 if (!data->writable)
2699 mask = perf_data_size(data) - 1;
2701 offset = (offset - tail) & mask;
2702 head = (head - tail) & mask;
2704 if ((int)(head - offset) < 0)
2710 static void perf_output_wakeup(struct perf_output_handle *handle)
2712 atomic_set(&handle->data->poll, POLL_IN);
2715 handle->event->pending_wakeup = 1;
2716 perf_pending_queue(&handle->event->pending,
2717 perf_pending_event);
2719 perf_event_wakeup(handle->event);
2723 * Curious locking construct.
2725 * We need to ensure a later event_id doesn't publish a head when a former
2726 * event_id isn't done writing. However since we need to deal with NMIs we
2727 * cannot fully serialize things.
2729 * What we do is serialize between CPUs so we only have to deal with NMI
2730 * nesting on a single CPU.
2732 * We only publish the head (and generate a wakeup) when the outer-most
2733 * event_id completes.
2735 static void perf_output_lock(struct perf_output_handle *handle)
2737 struct perf_mmap_data *data = handle->data;
2738 int cur, cpu = get_cpu();
2743 cur = atomic_cmpxchg(&data->lock, -1, cpu);
2755 static void perf_output_unlock(struct perf_output_handle *handle)
2757 struct perf_mmap_data *data = handle->data;
2761 data->done_head = data->head;
2763 if (!handle->locked)
2768 * The xchg implies a full barrier that ensures all writes are done
2769 * before we publish the new head, matched by a rmb() in userspace when
2770 * reading this position.
2772 while ((head = atomic_long_xchg(&data->done_head, 0)))
2773 data->user_page->data_head = head;
2776 * NMI can happen here, which means we can miss a done_head update.
2779 cpu = atomic_xchg(&data->lock, -1);
2780 WARN_ON_ONCE(cpu != smp_processor_id());
2783 * Therefore we have to validate we did not indeed do so.
2785 if (unlikely(atomic_long_read(&data->done_head))) {
2787 * Since we had it locked, we can lock it again.
2789 while (atomic_cmpxchg(&data->lock, -1, cpu) != -1)
2795 if (atomic_xchg(&data->wakeup, 0))
2796 perf_output_wakeup(handle);
2801 void perf_output_copy(struct perf_output_handle *handle,
2802 const void *buf, unsigned int len)
2804 unsigned int pages_mask;
2805 unsigned long offset;
2809 offset = handle->offset;
2810 pages_mask = handle->data->nr_pages - 1;
2811 pages = handle->data->data_pages;
2814 unsigned long page_offset;
2815 unsigned long page_size;
2818 nr = (offset >> PAGE_SHIFT) & pages_mask;
2819 page_size = 1UL << (handle->data->data_order + PAGE_SHIFT);
2820 page_offset = offset & (page_size - 1);
2821 size = min_t(unsigned int, page_size - page_offset, len);
2823 memcpy(pages[nr] + page_offset, buf, size);
2830 handle->offset = offset;
2833 * Check we didn't copy past our reservation window, taking the
2834 * possible unsigned int wrap into account.
2836 WARN_ON_ONCE(((long)(handle->head - handle->offset)) < 0);
2839 int perf_output_begin(struct perf_output_handle *handle,
2840 struct perf_event *event, unsigned int size,
2841 int nmi, int sample)
2843 struct perf_event *output_event;
2844 struct perf_mmap_data *data;
2845 unsigned long tail, offset, head;
2848 struct perf_event_header header;
2855 * For inherited events we send all the output towards the parent.
2858 event = event->parent;
2860 output_event = rcu_dereference(event->output);
2862 event = output_event;
2864 data = rcu_dereference(event->data);
2868 handle->data = data;
2869 handle->event = event;
2871 handle->sample = sample;
2873 if (!data->nr_pages)
2876 have_lost = atomic_read(&data->lost);
2878 size += sizeof(lost_event);
2880 perf_output_lock(handle);
2884 * Userspace could choose to issue a mb() before updating the
2885 * tail pointer. So that all reads will be completed before the
2888 tail = ACCESS_ONCE(data->user_page->data_tail);
2890 offset = head = atomic_long_read(&data->head);
2892 if (unlikely(!perf_output_space(data, tail, offset, head)))
2894 } while (atomic_long_cmpxchg(&data->head, offset, head) != offset);
2896 handle->offset = offset;
2897 handle->head = head;
2899 if (head - tail > data->watermark)
2900 atomic_set(&data->wakeup, 1);
2903 lost_event.header.type = PERF_RECORD_LOST;
2904 lost_event.header.misc = 0;
2905 lost_event.header.size = sizeof(lost_event);
2906 lost_event.id = event->id;
2907 lost_event.lost = atomic_xchg(&data->lost, 0);
2909 perf_output_put(handle, lost_event);
2915 atomic_inc(&data->lost);
2916 perf_output_unlock(handle);
2923 void perf_output_end(struct perf_output_handle *handle)
2925 struct perf_event *event = handle->event;
2926 struct perf_mmap_data *data = handle->data;
2928 int wakeup_events = event->attr.wakeup_events;
2930 if (handle->sample && wakeup_events) {
2931 int events = atomic_inc_return(&data->events);
2932 if (events >= wakeup_events) {
2933 atomic_sub(wakeup_events, &data->events);
2934 atomic_set(&data->wakeup, 1);
2938 perf_output_unlock(handle);
2942 static u32 perf_event_pid(struct perf_event *event, struct task_struct *p)
2945 * only top level events have the pid namespace they were created in
2948 event = event->parent;
2950 return task_tgid_nr_ns(p, event->ns);
2953 static u32 perf_event_tid(struct perf_event *event, struct task_struct *p)
2956 * only top level events have the pid namespace they were created in
2959 event = event->parent;
2961 return task_pid_nr_ns(p, event->ns);
2964 static void perf_output_read_one(struct perf_output_handle *handle,
2965 struct perf_event *event)
2967 u64 read_format = event->attr.read_format;
2971 values[n++] = atomic64_read(&event->count);
2972 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
2973 values[n++] = event->total_time_enabled +
2974 atomic64_read(&event->child_total_time_enabled);
2976 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
2977 values[n++] = event->total_time_running +
2978 atomic64_read(&event->child_total_time_running);
2980 if (read_format & PERF_FORMAT_ID)
2981 values[n++] = primary_event_id(event);
2983 perf_output_copy(handle, values, n * sizeof(u64));
2987 * XXX PERF_FORMAT_GROUP vs inherited events seems difficult.
2989 static void perf_output_read_group(struct perf_output_handle *handle,
2990 struct perf_event *event)
2992 struct perf_event *leader = event->group_leader, *sub;
2993 u64 read_format = event->attr.read_format;
2997 values[n++] = 1 + leader->nr_siblings;
2999 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
3000 values[n++] = leader->total_time_enabled;
3002 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
3003 values[n++] = leader->total_time_running;
3005 if (leader != event)
3006 leader->pmu->read(leader);
3008 values[n++] = atomic64_read(&leader->count);
3009 if (read_format & PERF_FORMAT_ID)
3010 values[n++] = primary_event_id(leader);
3012 perf_output_copy(handle, values, n * sizeof(u64));
3014 list_for_each_entry(sub, &leader->sibling_list, group_entry) {
3018 sub->pmu->read(sub);
3020 values[n++] = atomic64_read(&sub->count);
3021 if (read_format & PERF_FORMAT_ID)
3022 values[n++] = primary_event_id(sub);
3024 perf_output_copy(handle, values, n * sizeof(u64));
3028 static void perf_output_read(struct perf_output_handle *handle,
3029 struct perf_event *event)
3031 if (event->attr.read_format & PERF_FORMAT_GROUP)
3032 perf_output_read_group(handle, event);
3034 perf_output_read_one(handle, event);
3037 void perf_output_sample(struct perf_output_handle *handle,
3038 struct perf_event_header *header,
3039 struct perf_sample_data *data,
3040 struct perf_event *event)
3042 u64 sample_type = data->type;
3044 perf_output_put(handle, *header);
3046 if (sample_type & PERF_SAMPLE_IP)
3047 perf_output_put(handle, data->ip);
3049 if (sample_type & PERF_SAMPLE_TID)
3050 perf_output_put(handle, data->tid_entry);
3052 if (sample_type & PERF_SAMPLE_TIME)
3053 perf_output_put(handle, data->time);
3055 if (sample_type & PERF_SAMPLE_ADDR)
3056 perf_output_put(handle, data->addr);
3058 if (sample_type & PERF_SAMPLE_ID)
3059 perf_output_put(handle, data->id);
3061 if (sample_type & PERF_SAMPLE_STREAM_ID)
3062 perf_output_put(handle, data->stream_id);
3064 if (sample_type & PERF_SAMPLE_CPU)
3065 perf_output_put(handle, data->cpu_entry);
3067 if (sample_type & PERF_SAMPLE_PERIOD)
3068 perf_output_put(handle, data->period);
3070 if (sample_type & PERF_SAMPLE_READ)
3071 perf_output_read(handle, event);
3073 if (sample_type & PERF_SAMPLE_CALLCHAIN) {
3074 if (data->callchain) {
3077 if (data->callchain)
3078 size += data->callchain->nr;
3080 size *= sizeof(u64);
3082 perf_output_copy(handle, data->callchain, size);
3085 perf_output_put(handle, nr);
3089 if (sample_type & PERF_SAMPLE_RAW) {
3091 perf_output_put(handle, data->raw->size);
3092 perf_output_copy(handle, data->raw->data,
3099 .size = sizeof(u32),
3102 perf_output_put(handle, raw);
3107 void perf_prepare_sample(struct perf_event_header *header,
3108 struct perf_sample_data *data,
3109 struct perf_event *event,
3110 struct pt_regs *regs)
3112 u64 sample_type = event->attr.sample_type;
3114 data->type = sample_type;
3116 header->type = PERF_RECORD_SAMPLE;
3117 header->size = sizeof(*header);
3120 header->misc |= perf_misc_flags(regs);
3122 if (sample_type & PERF_SAMPLE_IP) {
3123 data->ip = perf_instruction_pointer(regs);
3125 header->size += sizeof(data->ip);
3128 if (sample_type & PERF_SAMPLE_TID) {
3129 /* namespace issues */
3130 data->tid_entry.pid = perf_event_pid(event, current);
3131 data->tid_entry.tid = perf_event_tid(event, current);
3133 header->size += sizeof(data->tid_entry);
3136 if (sample_type & PERF_SAMPLE_TIME) {
3137 data->time = perf_clock();
3139 header->size += sizeof(data->time);
3142 if (sample_type & PERF_SAMPLE_ADDR)
3143 header->size += sizeof(data->addr);
3145 if (sample_type & PERF_SAMPLE_ID) {
3146 data->id = primary_event_id(event);
3148 header->size += sizeof(data->id);
3151 if (sample_type & PERF_SAMPLE_STREAM_ID) {
3152 data->stream_id = event->id;
3154 header->size += sizeof(data->stream_id);
3157 if (sample_type & PERF_SAMPLE_CPU) {
3158 data->cpu_entry.cpu = raw_smp_processor_id();
3159 data->cpu_entry.reserved = 0;
3161 header->size += sizeof(data->cpu_entry);
3164 if (sample_type & PERF_SAMPLE_PERIOD)
3165 header->size += sizeof(data->period);
3167 if (sample_type & PERF_SAMPLE_READ)
3168 header->size += perf_event_read_size(event);
3170 if (sample_type & PERF_SAMPLE_CALLCHAIN) {
3173 data->callchain = perf_callchain(regs);
3175 if (data->callchain)
3176 size += data->callchain->nr;
3178 header->size += size * sizeof(u64);
3181 if (sample_type & PERF_SAMPLE_RAW) {
3182 int size = sizeof(u32);
3185 size += data->raw->size;
3187 size += sizeof(u32);
3189 WARN_ON_ONCE(size & (sizeof(u64)-1));
3190 header->size += size;
3194 static void perf_event_output(struct perf_event *event, int nmi,
3195 struct perf_sample_data *data,
3196 struct pt_regs *regs)
3198 struct perf_output_handle handle;
3199 struct perf_event_header header;
3201 perf_prepare_sample(&header, data, event, regs);
3203 if (perf_output_begin(&handle, event, header.size, nmi, 1))
3206 perf_output_sample(&handle, &header, data, event);
3208 perf_output_end(&handle);
3215 struct perf_read_event {
3216 struct perf_event_header header;
3223 perf_event_read_event(struct perf_event *event,
3224 struct task_struct *task)
3226 struct perf_output_handle handle;
3227 struct perf_read_event read_event = {
3229 .type = PERF_RECORD_READ,
3231 .size = sizeof(read_event) + perf_event_read_size(event),
3233 .pid = perf_event_pid(event, task),
3234 .tid = perf_event_tid(event, task),
3238 ret = perf_output_begin(&handle, event, read_event.header.size, 0, 0);
3242 perf_output_put(&handle, read_event);
3243 perf_output_read(&handle, event);
3245 perf_output_end(&handle);
3249 * task tracking -- fork/exit
3251 * enabled by: attr.comm | attr.mmap | attr.task
3254 struct perf_task_event {
3255 struct task_struct *task;
3256 struct perf_event_context *task_ctx;
3259 struct perf_event_header header;
3269 static void perf_event_task_output(struct perf_event *event,
3270 struct perf_task_event *task_event)
3272 struct perf_output_handle handle;
3274 struct task_struct *task = task_event->task;
3277 size = task_event->event_id.header.size;
3278 ret = perf_output_begin(&handle, event, size, 0, 0);
3283 task_event->event_id.pid = perf_event_pid(event, task);
3284 task_event->event_id.ppid = perf_event_pid(event, current);
3286 task_event->event_id.tid = perf_event_tid(event, task);
3287 task_event->event_id.ptid = perf_event_tid(event, current);
3289 task_event->event_id.time = perf_clock();
3291 perf_output_put(&handle, task_event->event_id);
3293 perf_output_end(&handle);
3296 static int perf_event_task_match(struct perf_event *event)
3298 if (event->cpu != -1 && event->cpu != smp_processor_id())
3301 if (event->attr.comm || event->attr.mmap || event->attr.task)
3307 static void perf_event_task_ctx(struct perf_event_context *ctx,
3308 struct perf_task_event *task_event)
3310 struct perf_event *event;
3312 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
3313 if (perf_event_task_match(event))
3314 perf_event_task_output(event, task_event);
3318 static void perf_event_task_event(struct perf_task_event *task_event)
3320 struct perf_cpu_context *cpuctx;
3321 struct perf_event_context *ctx = task_event->task_ctx;
3324 cpuctx = &get_cpu_var(perf_cpu_context);
3325 perf_event_task_ctx(&cpuctx->ctx, task_event);
3327 ctx = rcu_dereference(task_event->task->perf_event_ctxp);
3329 perf_event_task_ctx(ctx, task_event);
3330 put_cpu_var(perf_cpu_context);
3334 static void perf_event_task(struct task_struct *task,
3335 struct perf_event_context *task_ctx,
3338 struct perf_task_event task_event;
3340 if (!atomic_read(&nr_comm_events) &&
3341 !atomic_read(&nr_mmap_events) &&
3342 !atomic_read(&nr_task_events))
3345 task_event = (struct perf_task_event){
3347 .task_ctx = task_ctx,
3350 .type = new ? PERF_RECORD_FORK : PERF_RECORD_EXIT,
3352 .size = sizeof(task_event.event_id),
3361 perf_event_task_event(&task_event);
3364 void perf_event_fork(struct task_struct *task)
3366 perf_event_task(task, NULL, 1);
3373 struct perf_comm_event {
3374 struct task_struct *task;
3379 struct perf_event_header header;
3386 static void perf_event_comm_output(struct perf_event *event,
3387 struct perf_comm_event *comm_event)
3389 struct perf_output_handle handle;
3390 int size = comm_event->event_id.header.size;
3391 int ret = perf_output_begin(&handle, event, size, 0, 0);
3396 comm_event->event_id.pid = perf_event_pid(event, comm_event->task);
3397 comm_event->event_id.tid = perf_event_tid(event, comm_event->task);
3399 perf_output_put(&handle, comm_event->event_id);
3400 perf_output_copy(&handle, comm_event->comm,
3401 comm_event->comm_size);
3402 perf_output_end(&handle);
3405 static int perf_event_comm_match(struct perf_event *event)
3407 if (event->cpu != -1 && event->cpu != smp_processor_id())
3410 if (event->attr.comm)
3416 static void perf_event_comm_ctx(struct perf_event_context *ctx,
3417 struct perf_comm_event *comm_event)
3419 struct perf_event *event;
3421 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
3422 if (perf_event_comm_match(event))
3423 perf_event_comm_output(event, comm_event);
3427 static void perf_event_comm_event(struct perf_comm_event *comm_event)
3429 struct perf_cpu_context *cpuctx;
3430 struct perf_event_context *ctx;
3432 char comm[TASK_COMM_LEN];
3434 memset(comm, 0, sizeof(comm));
3435 strlcpy(comm, comm_event->task->comm, sizeof(comm));
3436 size = ALIGN(strlen(comm)+1, sizeof(u64));
3438 comm_event->comm = comm;
3439 comm_event->comm_size = size;
3441 comm_event->event_id.header.size = sizeof(comm_event->event_id) + size;
3444 cpuctx = &get_cpu_var(perf_cpu_context);
3445 perf_event_comm_ctx(&cpuctx->ctx, comm_event);
3446 ctx = rcu_dereference(current->perf_event_ctxp);
3448 perf_event_comm_ctx(ctx, comm_event);
3449 put_cpu_var(perf_cpu_context);
3453 void perf_event_comm(struct task_struct *task)
3455 struct perf_comm_event comm_event;
3457 if (task->perf_event_ctxp)
3458 perf_event_enable_on_exec(task);
3460 if (!atomic_read(&nr_comm_events))
3463 comm_event = (struct perf_comm_event){
3469 .type = PERF_RECORD_COMM,
3478 perf_event_comm_event(&comm_event);
3485 struct perf_mmap_event {
3486 struct vm_area_struct *vma;
3488 const char *file_name;
3492 struct perf_event_header header;
3502 static void perf_event_mmap_output(struct perf_event *event,
3503 struct perf_mmap_event *mmap_event)
3505 struct perf_output_handle handle;
3506 int size = mmap_event->event_id.header.size;
3507 int ret = perf_output_begin(&handle, event, size, 0, 0);
3512 mmap_event->event_id.pid = perf_event_pid(event, current);
3513 mmap_event->event_id.tid = perf_event_tid(event, current);
3515 perf_output_put(&handle, mmap_event->event_id);
3516 perf_output_copy(&handle, mmap_event->file_name,
3517 mmap_event->file_size);
3518 perf_output_end(&handle);
3521 static int perf_event_mmap_match(struct perf_event *event,
3522 struct perf_mmap_event *mmap_event)
3524 if (event->cpu != -1 && event->cpu != smp_processor_id())
3527 if (event->attr.mmap)
3533 static void perf_event_mmap_ctx(struct perf_event_context *ctx,
3534 struct perf_mmap_event *mmap_event)
3536 struct perf_event *event;
3538 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
3539 if (perf_event_mmap_match(event, mmap_event))
3540 perf_event_mmap_output(event, mmap_event);
3544 static void perf_event_mmap_event(struct perf_mmap_event *mmap_event)
3546 struct perf_cpu_context *cpuctx;
3547 struct perf_event_context *ctx;
3548 struct vm_area_struct *vma = mmap_event->vma;
3549 struct file *file = vma->vm_file;
3555 memset(tmp, 0, sizeof(tmp));
3559 * d_path works from the end of the buffer backwards, so we
3560 * need to add enough zero bytes after the string to handle
3561 * the 64bit alignment we do later.
3563 buf = kzalloc(PATH_MAX + sizeof(u64), GFP_KERNEL);
3565 name = strncpy(tmp, "//enomem", sizeof(tmp));
3568 name = d_path(&file->f_path, buf, PATH_MAX);
3570 name = strncpy(tmp, "//toolong", sizeof(tmp));
3574 if (arch_vma_name(mmap_event->vma)) {
3575 name = strncpy(tmp, arch_vma_name(mmap_event->vma),
3581 name = strncpy(tmp, "[vdso]", sizeof(tmp));
3585 name = strncpy(tmp, "//anon", sizeof(tmp));
3590 size = ALIGN(strlen(name)+1, sizeof(u64));
3592 mmap_event->file_name = name;
3593 mmap_event->file_size = size;
3595 mmap_event->event_id.header.size = sizeof(mmap_event->event_id) + size;
3598 cpuctx = &get_cpu_var(perf_cpu_context);
3599 perf_event_mmap_ctx(&cpuctx->ctx, mmap_event);
3600 ctx = rcu_dereference(current->perf_event_ctxp);
3602 perf_event_mmap_ctx(ctx, mmap_event);
3603 put_cpu_var(perf_cpu_context);
3609 void __perf_event_mmap(struct vm_area_struct *vma)
3611 struct perf_mmap_event mmap_event;
3613 if (!atomic_read(&nr_mmap_events))
3616 mmap_event = (struct perf_mmap_event){
3622 .type = PERF_RECORD_MMAP,
3628 .start = vma->vm_start,
3629 .len = vma->vm_end - vma->vm_start,
3630 .pgoff = vma->vm_pgoff,
3634 perf_event_mmap_event(&mmap_event);
3638 * IRQ throttle logging
3641 static void perf_log_throttle(struct perf_event *event, int enable)
3643 struct perf_output_handle handle;
3647 struct perf_event_header header;
3651 } throttle_event = {
3653 .type = PERF_RECORD_THROTTLE,
3655 .size = sizeof(throttle_event),
3657 .time = perf_clock(),
3658 .id = primary_event_id(event),
3659 .stream_id = event->id,
3663 throttle_event.header.type = PERF_RECORD_UNTHROTTLE;
3665 ret = perf_output_begin(&handle, event, sizeof(throttle_event), 1, 0);
3669 perf_output_put(&handle, throttle_event);
3670 perf_output_end(&handle);
3674 * Generic event overflow handling, sampling.
3677 static int __perf_event_overflow(struct perf_event *event, int nmi,
3678 int throttle, struct perf_sample_data *data,
3679 struct pt_regs *regs)
3681 int events = atomic_read(&event->event_limit);
3682 struct hw_perf_event *hwc = &event->hw;
3685 throttle = (throttle && event->pmu->unthrottle != NULL);
3690 if (hwc->interrupts != MAX_INTERRUPTS) {
3692 if (HZ * hwc->interrupts >
3693 (u64)sysctl_perf_event_sample_rate) {
3694 hwc->interrupts = MAX_INTERRUPTS;
3695 perf_log_throttle(event, 0);
3700 * Keep re-disabling events even though on the previous
3701 * pass we disabled it - just in case we raced with a
3702 * sched-in and the event got enabled again:
3708 if (event->attr.freq) {
3709 u64 now = perf_clock();
3710 s64 delta = now - hwc->freq_stamp;
3712 hwc->freq_stamp = now;
3714 if (delta > 0 && delta < TICK_NSEC)
3715 perf_adjust_period(event, NSEC_PER_SEC / (int)delta);
3719 * XXX event_limit might not quite work as expected on inherited
3723 event->pending_kill = POLL_IN;
3724 if (events && atomic_dec_and_test(&event->event_limit)) {
3726 event->pending_kill = POLL_HUP;
3728 event->pending_disable = 1;
3729 perf_pending_queue(&event->pending,
3730 perf_pending_event);
3732 perf_event_disable(event);
3735 if (event->overflow_handler)
3736 event->overflow_handler(event, nmi, data, regs);
3738 perf_event_output(event, nmi, data, regs);
3743 int perf_event_overflow(struct perf_event *event, int nmi,
3744 struct perf_sample_data *data,
3745 struct pt_regs *regs)
3747 return __perf_event_overflow(event, nmi, 1, data, regs);
3751 * Generic software event infrastructure
3755 * We directly increment event->count and keep a second value in
3756 * event->hw.period_left to count intervals. This period event
3757 * is kept in the range [-sample_period, 0] so that we can use the
3761 static u64 perf_swevent_set_period(struct perf_event *event)
3763 struct hw_perf_event *hwc = &event->hw;
3764 u64 period = hwc->last_period;
3768 hwc->last_period = hwc->sample_period;
3771 old = val = atomic64_read(&hwc->period_left);
3775 nr = div64_u64(period + val, period);
3776 offset = nr * period;
3778 if (atomic64_cmpxchg(&hwc->period_left, old, val) != old)
3784 static void perf_swevent_overflow(struct perf_event *event, u64 overflow,
3785 int nmi, struct perf_sample_data *data,
3786 struct pt_regs *regs)
3788 struct hw_perf_event *hwc = &event->hw;
3791 data->period = event->hw.last_period;
3793 overflow = perf_swevent_set_period(event);
3795 if (hwc->interrupts == MAX_INTERRUPTS)
3798 for (; overflow; overflow--) {
3799 if (__perf_event_overflow(event, nmi, throttle,
3802 * We inhibit the overflow from happening when
3803 * hwc->interrupts == MAX_INTERRUPTS.
3811 static void perf_swevent_unthrottle(struct perf_event *event)
3814 * Nothing to do, we already reset hwc->interrupts.
3818 static void perf_swevent_add(struct perf_event *event, u64 nr,
3819 int nmi, struct perf_sample_data *data,
3820 struct pt_regs *regs)
3822 struct hw_perf_event *hwc = &event->hw;
3824 atomic64_add(nr, &event->count);
3829 if (!hwc->sample_period)
3832 if (nr == 1 && hwc->sample_period == 1 && !event->attr.freq)
3833 return perf_swevent_overflow(event, 1, nmi, data, regs);
3835 if (atomic64_add_negative(nr, &hwc->period_left))
3838 perf_swevent_overflow(event, 0, nmi, data, regs);
3841 static int perf_swevent_is_counting(struct perf_event *event)
3844 * The event is active, we're good!
3846 if (event->state == PERF_EVENT_STATE_ACTIVE)
3850 * The event is off/error, not counting.
3852 if (event->state != PERF_EVENT_STATE_INACTIVE)
3856 * The event is inactive, if the context is active
3857 * we're part of a group that didn't make it on the 'pmu',
3860 if (event->ctx->is_active)
3864 * We're inactive and the context is too, this means the
3865 * task is scheduled out, we're counting events that happen
3866 * to us, like migration events.
3871 static int perf_tp_event_match(struct perf_event *event,
3872 struct perf_sample_data *data);
3874 static int perf_exclude_event(struct perf_event *event,
3875 struct pt_regs *regs)
3878 if (event->attr.exclude_user && user_mode(regs))
3881 if (event->attr.exclude_kernel && !user_mode(regs))
3888 static int perf_swevent_match(struct perf_event *event,
3889 enum perf_type_id type,
3891 struct perf_sample_data *data,
3892 struct pt_regs *regs)
3894 if (event->cpu != -1 && event->cpu != smp_processor_id())
3897 if (!perf_swevent_is_counting(event))
3900 if (event->attr.type != type)
3903 if (event->attr.config != event_id)
3906 if (perf_exclude_event(event, regs))
3909 if (event->attr.type == PERF_TYPE_TRACEPOINT &&
3910 !perf_tp_event_match(event, data))
3916 static void perf_swevent_ctx_event(struct perf_event_context *ctx,
3917 enum perf_type_id type,
3918 u32 event_id, u64 nr, int nmi,
3919 struct perf_sample_data *data,
3920 struct pt_regs *regs)
3922 struct perf_event *event;
3924 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
3925 if (perf_swevent_match(event, type, event_id, data, regs))
3926 perf_swevent_add(event, nr, nmi, data, regs);
3930 int perf_swevent_get_recursion_context(void)
3932 struct perf_cpu_context *cpuctx = &get_cpu_var(perf_cpu_context);
3939 else if (in_softirq())
3944 if (cpuctx->recursion[rctx]) {
3945 put_cpu_var(perf_cpu_context);
3949 cpuctx->recursion[rctx]++;
3954 EXPORT_SYMBOL_GPL(perf_swevent_get_recursion_context);
3956 void perf_swevent_put_recursion_context(int rctx)
3958 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
3960 cpuctx->recursion[rctx]--;
3961 put_cpu_var(perf_cpu_context);
3963 EXPORT_SYMBOL_GPL(perf_swevent_put_recursion_context);
3965 static void do_perf_sw_event(enum perf_type_id type, u32 event_id,
3967 struct perf_sample_data *data,
3968 struct pt_regs *regs)
3970 struct perf_cpu_context *cpuctx;
3971 struct perf_event_context *ctx;
3973 cpuctx = &__get_cpu_var(perf_cpu_context);
3975 perf_swevent_ctx_event(&cpuctx->ctx, type, event_id,
3976 nr, nmi, data, regs);
3978 * doesn't really matter which of the child contexts the
3979 * events ends up in.
3981 ctx = rcu_dereference(current->perf_event_ctxp);
3983 perf_swevent_ctx_event(ctx, type, event_id, nr, nmi, data, regs);
3987 void __perf_sw_event(u32 event_id, u64 nr, int nmi,
3988 struct pt_regs *regs, u64 addr)
3990 struct perf_sample_data data;
3993 rctx = perf_swevent_get_recursion_context();
4000 do_perf_sw_event(PERF_TYPE_SOFTWARE, event_id, nr, nmi, &data, regs);
4002 perf_swevent_put_recursion_context(rctx);
4005 static void perf_swevent_read(struct perf_event *event)
4009 static int perf_swevent_enable(struct perf_event *event)
4011 struct hw_perf_event *hwc = &event->hw;
4013 if (hwc->sample_period) {
4014 hwc->last_period = hwc->sample_period;
4015 perf_swevent_set_period(event);
4020 static void perf_swevent_disable(struct perf_event *event)
4024 static const struct pmu perf_ops_generic = {
4025 .enable = perf_swevent_enable,
4026 .disable = perf_swevent_disable,
4027 .read = perf_swevent_read,
4028 .unthrottle = perf_swevent_unthrottle,
4032 * hrtimer based swevent callback
4035 static enum hrtimer_restart perf_swevent_hrtimer(struct hrtimer *hrtimer)
4037 enum hrtimer_restart ret = HRTIMER_RESTART;
4038 struct perf_sample_data data;
4039 struct pt_regs *regs;
4040 struct perf_event *event;
4043 event = container_of(hrtimer, struct perf_event, hw.hrtimer);
4044 event->pmu->read(event);
4048 data.period = event->hw.last_period;
4049 regs = get_irq_regs();
4051 * In case we exclude kernel IPs or are somehow not in interrupt
4052 * context, provide the next best thing, the user IP.
4054 if ((event->attr.exclude_kernel || !regs) &&
4055 !event->attr.exclude_user)
4056 regs = task_pt_regs(current);
4059 if (!(event->attr.exclude_idle && current->pid == 0))
4060 if (perf_event_overflow(event, 0, &data, regs))
4061 ret = HRTIMER_NORESTART;
4064 period = max_t(u64, 10000, event->hw.sample_period);
4065 hrtimer_forward_now(hrtimer, ns_to_ktime(period));
4070 static void perf_swevent_start_hrtimer(struct perf_event *event)
4072 struct hw_perf_event *hwc = &event->hw;
4074 hrtimer_init(&hwc->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
4075 hwc->hrtimer.function = perf_swevent_hrtimer;
4076 if (hwc->sample_period) {
4079 if (hwc->remaining) {
4080 if (hwc->remaining < 0)
4083 period = hwc->remaining;
4086 period = max_t(u64, 10000, hwc->sample_period);
4088 __hrtimer_start_range_ns(&hwc->hrtimer,
4089 ns_to_ktime(period), 0,
4090 HRTIMER_MODE_REL, 0);
4094 static void perf_swevent_cancel_hrtimer(struct perf_event *event)
4096 struct hw_perf_event *hwc = &event->hw;
4098 if (hwc->sample_period) {
4099 ktime_t remaining = hrtimer_get_remaining(&hwc->hrtimer);
4100 hwc->remaining = ktime_to_ns(remaining);
4102 hrtimer_cancel(&hwc->hrtimer);
4107 * Software event: cpu wall time clock
4110 static void cpu_clock_perf_event_update(struct perf_event *event)
4112 int cpu = raw_smp_processor_id();
4116 now = cpu_clock(cpu);
4117 prev = atomic64_xchg(&event->hw.prev_count, now);
4118 atomic64_add(now - prev, &event->count);
4121 static int cpu_clock_perf_event_enable(struct perf_event *event)
4123 struct hw_perf_event *hwc = &event->hw;
4124 int cpu = raw_smp_processor_id();
4126 atomic64_set(&hwc->prev_count, cpu_clock(cpu));
4127 perf_swevent_start_hrtimer(event);
4132 static void cpu_clock_perf_event_disable(struct perf_event *event)
4134 perf_swevent_cancel_hrtimer(event);
4135 cpu_clock_perf_event_update(event);
4138 static void cpu_clock_perf_event_read(struct perf_event *event)
4140 cpu_clock_perf_event_update(event);
4143 static const struct pmu perf_ops_cpu_clock = {
4144 .enable = cpu_clock_perf_event_enable,
4145 .disable = cpu_clock_perf_event_disable,
4146 .read = cpu_clock_perf_event_read,
4150 * Software event: task time clock
4153 static void task_clock_perf_event_update(struct perf_event *event, u64 now)
4158 prev = atomic64_xchg(&event->hw.prev_count, now);
4160 atomic64_add(delta, &event->count);
4163 static int task_clock_perf_event_enable(struct perf_event *event)
4165 struct hw_perf_event *hwc = &event->hw;
4168 now = event->ctx->time;
4170 atomic64_set(&hwc->prev_count, now);
4172 perf_swevent_start_hrtimer(event);
4177 static void task_clock_perf_event_disable(struct perf_event *event)
4179 perf_swevent_cancel_hrtimer(event);
4180 task_clock_perf_event_update(event, event->ctx->time);
4184 static void task_clock_perf_event_read(struct perf_event *event)
4189 update_context_time(event->ctx);
4190 time = event->ctx->time;
4192 u64 now = perf_clock();
4193 u64 delta = now - event->ctx->timestamp;
4194 time = event->ctx->time + delta;
4197 task_clock_perf_event_update(event, time);
4200 static const struct pmu perf_ops_task_clock = {
4201 .enable = task_clock_perf_event_enable,
4202 .disable = task_clock_perf_event_disable,
4203 .read = task_clock_perf_event_read,
4206 #ifdef CONFIG_EVENT_TRACING
4208 void perf_tp_event(int event_id, u64 addr, u64 count, void *record,
4211 struct perf_raw_record raw = {
4216 struct perf_sample_data data = {
4221 struct pt_regs *regs = get_irq_regs();
4224 regs = task_pt_regs(current);
4226 /* Trace events already protected against recursion */
4227 do_perf_sw_event(PERF_TYPE_TRACEPOINT, event_id, count, 1,
4230 EXPORT_SYMBOL_GPL(perf_tp_event);
4232 static int perf_tp_event_match(struct perf_event *event,
4233 struct perf_sample_data *data)
4235 void *record = data->raw->data;
4237 if (likely(!event->filter) || filter_match_preds(event->filter, record))
4242 static void tp_perf_event_destroy(struct perf_event *event)
4244 ftrace_profile_disable(event->attr.config);
4247 static const struct pmu *tp_perf_event_init(struct perf_event *event)
4250 * Raw tracepoint data is a severe data leak, only allow root to
4253 if ((event->attr.sample_type & PERF_SAMPLE_RAW) &&
4254 perf_paranoid_tracepoint_raw() &&
4255 !capable(CAP_SYS_ADMIN))
4256 return ERR_PTR(-EPERM);
4258 if (ftrace_profile_enable(event->attr.config))
4261 event->destroy = tp_perf_event_destroy;
4263 return &perf_ops_generic;
4266 static int perf_event_set_filter(struct perf_event *event, void __user *arg)
4271 if (event->attr.type != PERF_TYPE_TRACEPOINT)
4274 filter_str = strndup_user(arg, PAGE_SIZE);
4275 if (IS_ERR(filter_str))
4276 return PTR_ERR(filter_str);
4278 ret = ftrace_profile_set_filter(event, event->attr.config, filter_str);
4284 static void perf_event_free_filter(struct perf_event *event)
4286 ftrace_profile_free_filter(event);
4291 static int perf_tp_event_match(struct perf_event *event,
4292 struct perf_sample_data *data)
4297 static const struct pmu *tp_perf_event_init(struct perf_event *event)
4302 static int perf_event_set_filter(struct perf_event *event, void __user *arg)
4307 static void perf_event_free_filter(struct perf_event *event)
4311 #endif /* CONFIG_EVENT_TRACING */
4313 #ifdef CONFIG_HAVE_HW_BREAKPOINT
4314 static void bp_perf_event_destroy(struct perf_event *event)
4316 release_bp_slot(event);
4319 static const struct pmu *bp_perf_event_init(struct perf_event *bp)
4323 err = register_perf_hw_breakpoint(bp);
4325 return ERR_PTR(err);
4327 bp->destroy = bp_perf_event_destroy;
4329 return &perf_ops_bp;
4332 void perf_bp_event(struct perf_event *bp, void *data)
4334 struct perf_sample_data sample;
4335 struct pt_regs *regs = data;
4338 sample.addr = bp->attr.bp_addr;
4340 if (!perf_exclude_event(bp, regs))
4341 perf_swevent_add(bp, 1, 1, &sample, regs);
4344 static const struct pmu *bp_perf_event_init(struct perf_event *bp)
4349 void perf_bp_event(struct perf_event *bp, void *regs)
4354 atomic_t perf_swevent_enabled[PERF_COUNT_SW_MAX];
4356 static void sw_perf_event_destroy(struct perf_event *event)
4358 u64 event_id = event->attr.config;
4360 WARN_ON(event->parent);
4362 atomic_dec(&perf_swevent_enabled[event_id]);
4365 static const struct pmu *sw_perf_event_init(struct perf_event *event)
4367 const struct pmu *pmu = NULL;
4368 u64 event_id = event->attr.config;
4371 * Software events (currently) can't in general distinguish
4372 * between user, kernel and hypervisor events.
4373 * However, context switches and cpu migrations are considered
4374 * to be kernel events, and page faults are never hypervisor
4378 case PERF_COUNT_SW_CPU_CLOCK:
4379 pmu = &perf_ops_cpu_clock;
4382 case PERF_COUNT_SW_TASK_CLOCK:
4384 * If the user instantiates this as a per-cpu event,
4385 * use the cpu_clock event instead.
4387 if (event->ctx->task)
4388 pmu = &perf_ops_task_clock;
4390 pmu = &perf_ops_cpu_clock;
4393 case PERF_COUNT_SW_PAGE_FAULTS:
4394 case PERF_COUNT_SW_PAGE_FAULTS_MIN:
4395 case PERF_COUNT_SW_PAGE_FAULTS_MAJ:
4396 case PERF_COUNT_SW_CONTEXT_SWITCHES:
4397 case PERF_COUNT_SW_CPU_MIGRATIONS:
4398 case PERF_COUNT_SW_ALIGNMENT_FAULTS:
4399 case PERF_COUNT_SW_EMULATION_FAULTS:
4400 if (!event->parent) {
4401 atomic_inc(&perf_swevent_enabled[event_id]);
4402 event->destroy = sw_perf_event_destroy;
4404 pmu = &perf_ops_generic;
4412 * Allocate and initialize a event structure
4414 static struct perf_event *
4415 perf_event_alloc(struct perf_event_attr *attr,
4417 struct perf_event_context *ctx,
4418 struct perf_event *group_leader,
4419 struct perf_event *parent_event,
4420 perf_overflow_handler_t overflow_handler,
4423 const struct pmu *pmu;
4424 struct perf_event *event;
4425 struct hw_perf_event *hwc;
4428 event = kzalloc(sizeof(*event), gfpflags);
4430 return ERR_PTR(-ENOMEM);
4433 * Single events are their own group leaders, with an
4434 * empty sibling list:
4437 group_leader = event;
4439 mutex_init(&event->child_mutex);
4440 INIT_LIST_HEAD(&event->child_list);
4442 INIT_LIST_HEAD(&event->group_entry);
4443 INIT_LIST_HEAD(&event->event_entry);
4444 INIT_LIST_HEAD(&event->sibling_list);
4445 init_waitqueue_head(&event->waitq);
4447 mutex_init(&event->mmap_mutex);
4450 event->attr = *attr;
4451 event->group_leader = group_leader;
4456 event->parent = parent_event;
4458 event->ns = get_pid_ns(current->nsproxy->pid_ns);
4459 event->id = atomic64_inc_return(&perf_event_id);
4461 event->state = PERF_EVENT_STATE_INACTIVE;
4463 if (!overflow_handler && parent_event)
4464 overflow_handler = parent_event->overflow_handler;
4466 event->overflow_handler = overflow_handler;
4469 event->state = PERF_EVENT_STATE_OFF;
4474 hwc->sample_period = attr->sample_period;
4475 if (attr->freq && attr->sample_freq)
4476 hwc->sample_period = 1;
4477 hwc->last_period = hwc->sample_period;
4479 atomic64_set(&hwc->period_left, hwc->sample_period);
4482 * we currently do not support PERF_FORMAT_GROUP on inherited events
4484 if (attr->inherit && (attr->read_format & PERF_FORMAT_GROUP))
4487 switch (attr->type) {
4489 case PERF_TYPE_HARDWARE:
4490 case PERF_TYPE_HW_CACHE:
4491 pmu = hw_perf_event_init(event);
4494 case PERF_TYPE_SOFTWARE:
4495 pmu = sw_perf_event_init(event);
4498 case PERF_TYPE_TRACEPOINT:
4499 pmu = tp_perf_event_init(event);
4502 case PERF_TYPE_BREAKPOINT:
4503 pmu = bp_perf_event_init(event);
4514 else if (IS_ERR(pmu))
4519 put_pid_ns(event->ns);
4521 return ERR_PTR(err);
4526 if (!event->parent) {
4527 atomic_inc(&nr_events);
4528 if (event->attr.mmap)
4529 atomic_inc(&nr_mmap_events);
4530 if (event->attr.comm)
4531 atomic_inc(&nr_comm_events);
4532 if (event->attr.task)
4533 atomic_inc(&nr_task_events);
4539 static int perf_copy_attr(struct perf_event_attr __user *uattr,
4540 struct perf_event_attr *attr)
4545 if (!access_ok(VERIFY_WRITE, uattr, PERF_ATTR_SIZE_VER0))
4549 * zero the full structure, so that a short copy will be nice.
4551 memset(attr, 0, sizeof(*attr));
4553 ret = get_user(size, &uattr->size);
4557 if (size > PAGE_SIZE) /* silly large */
4560 if (!size) /* abi compat */
4561 size = PERF_ATTR_SIZE_VER0;
4563 if (size < PERF_ATTR_SIZE_VER0)
4567 * If we're handed a bigger struct than we know of,
4568 * ensure all the unknown bits are 0 - i.e. new
4569 * user-space does not rely on any kernel feature
4570 * extensions we dont know about yet.
4572 if (size > sizeof(*attr)) {
4573 unsigned char __user *addr;
4574 unsigned char __user *end;
4577 addr = (void __user *)uattr + sizeof(*attr);
4578 end = (void __user *)uattr + size;
4580 for (; addr < end; addr++) {
4581 ret = get_user(val, addr);
4587 size = sizeof(*attr);
4590 ret = copy_from_user(attr, uattr, size);
4595 * If the type exists, the corresponding creation will verify
4598 if (attr->type >= PERF_TYPE_MAX)
4601 if (attr->__reserved_1 || attr->__reserved_2)
4604 if (attr->sample_type & ~(PERF_SAMPLE_MAX-1))
4607 if (attr->read_format & ~(PERF_FORMAT_MAX-1))
4614 put_user(sizeof(*attr), &uattr->size);
4619 static int perf_event_set_output(struct perf_event *event, int output_fd)
4621 struct perf_event *output_event = NULL;
4622 struct file *output_file = NULL;
4623 struct perf_event *old_output;
4624 int fput_needed = 0;
4630 output_file = fget_light(output_fd, &fput_needed);
4634 if (output_file->f_op != &perf_fops)
4637 output_event = output_file->private_data;
4639 /* Don't chain output fds */
4640 if (output_event->output)
4643 /* Don't set an output fd when we already have an output channel */
4647 atomic_long_inc(&output_file->f_count);
4650 mutex_lock(&event->mmap_mutex);
4651 old_output = event->output;
4652 rcu_assign_pointer(event->output, output_event);
4653 mutex_unlock(&event->mmap_mutex);
4657 * we need to make sure no existing perf_output_*()
4658 * is still referencing this event.
4661 fput(old_output->filp);
4666 fput_light(output_file, fput_needed);
4671 * sys_perf_event_open - open a performance event, associate it to a task/cpu
4673 * @attr_uptr: event_id type attributes for monitoring/sampling
4676 * @group_fd: group leader event fd
4678 SYSCALL_DEFINE5(perf_event_open,
4679 struct perf_event_attr __user *, attr_uptr,
4680 pid_t, pid, int, cpu, int, group_fd, unsigned long, flags)
4682 struct perf_event *event, *group_leader;
4683 struct perf_event_attr attr;
4684 struct perf_event_context *ctx;
4685 struct file *event_file = NULL;
4686 struct file *group_file = NULL;
4687 int fput_needed = 0;
4688 int fput_needed2 = 0;
4691 /* for future expandability... */
4692 if (flags & ~(PERF_FLAG_FD_NO_GROUP | PERF_FLAG_FD_OUTPUT))
4695 err = perf_copy_attr(attr_uptr, &attr);
4699 if (!attr.exclude_kernel) {
4700 if (perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN))
4705 if (attr.sample_freq > sysctl_perf_event_sample_rate)
4710 * Get the target context (task or percpu):
4712 ctx = find_get_context(pid, cpu);
4714 return PTR_ERR(ctx);
4717 * Look up the group leader (we will attach this event to it):
4719 group_leader = NULL;
4720 if (group_fd != -1 && !(flags & PERF_FLAG_FD_NO_GROUP)) {
4722 group_file = fget_light(group_fd, &fput_needed);
4724 goto err_put_context;
4725 if (group_file->f_op != &perf_fops)
4726 goto err_put_context;
4728 group_leader = group_file->private_data;
4730 * Do not allow a recursive hierarchy (this new sibling
4731 * becoming part of another group-sibling):
4733 if (group_leader->group_leader != group_leader)
4734 goto err_put_context;
4736 * Do not allow to attach to a group in a different
4737 * task or CPU context:
4739 if (group_leader->ctx != ctx)
4740 goto err_put_context;
4742 * Only a group leader can be exclusive or pinned
4744 if (attr.exclusive || attr.pinned)
4745 goto err_put_context;
4748 event = perf_event_alloc(&attr, cpu, ctx, group_leader,
4749 NULL, NULL, GFP_KERNEL);
4750 err = PTR_ERR(event);
4752 goto err_put_context;
4754 err = anon_inode_getfd("[perf_event]", &perf_fops, event, O_RDWR);
4756 goto err_free_put_context;
4758 event_file = fget_light(err, &fput_needed2);
4760 goto err_free_put_context;
4762 if (flags & PERF_FLAG_FD_OUTPUT) {
4763 err = perf_event_set_output(event, group_fd);
4765 goto err_fput_free_put_context;
4768 event->filp = event_file;
4769 WARN_ON_ONCE(ctx->parent_ctx);
4770 mutex_lock(&ctx->mutex);
4771 perf_install_in_context(ctx, event, cpu);
4773 mutex_unlock(&ctx->mutex);
4775 event->owner = current;
4776 get_task_struct(current);
4777 mutex_lock(¤t->perf_event_mutex);
4778 list_add_tail(&event->owner_entry, ¤t->perf_event_list);
4779 mutex_unlock(¤t->perf_event_mutex);
4781 err_fput_free_put_context:
4782 fput_light(event_file, fput_needed2);
4784 err_free_put_context:
4792 fput_light(group_file, fput_needed);
4798 * perf_event_create_kernel_counter
4800 * @attr: attributes of the counter to create
4801 * @cpu: cpu in which the counter is bound
4802 * @pid: task to profile
4805 perf_event_create_kernel_counter(struct perf_event_attr *attr, int cpu,
4807 perf_overflow_handler_t overflow_handler)
4809 struct perf_event *event;
4810 struct perf_event_context *ctx;
4814 * Get the target context (task or percpu):
4817 ctx = find_get_context(pid, cpu);
4823 event = perf_event_alloc(attr, cpu, ctx, NULL,
4824 NULL, overflow_handler, GFP_KERNEL);
4825 if (IS_ERR(event)) {
4826 err = PTR_ERR(event);
4827 goto err_put_context;
4831 WARN_ON_ONCE(ctx->parent_ctx);
4832 mutex_lock(&ctx->mutex);
4833 perf_install_in_context(ctx, event, cpu);
4835 mutex_unlock(&ctx->mutex);
4837 event->owner = current;
4838 get_task_struct(current);
4839 mutex_lock(¤t->perf_event_mutex);
4840 list_add_tail(&event->owner_entry, ¤t->perf_event_list);
4841 mutex_unlock(¤t->perf_event_mutex);
4848 return ERR_PTR(err);
4850 EXPORT_SYMBOL_GPL(perf_event_create_kernel_counter);
4853 * inherit a event from parent task to child task:
4855 static struct perf_event *
4856 inherit_event(struct perf_event *parent_event,
4857 struct task_struct *parent,
4858 struct perf_event_context *parent_ctx,
4859 struct task_struct *child,
4860 struct perf_event *group_leader,
4861 struct perf_event_context *child_ctx)
4863 struct perf_event *child_event;
4866 * Instead of creating recursive hierarchies of events,
4867 * we link inherited events back to the original parent,
4868 * which has a filp for sure, which we use as the reference
4871 if (parent_event->parent)
4872 parent_event = parent_event->parent;
4874 child_event = perf_event_alloc(&parent_event->attr,
4875 parent_event->cpu, child_ctx,
4876 group_leader, parent_event,
4878 if (IS_ERR(child_event))
4883 * Make the child state follow the state of the parent event,
4884 * not its attr.disabled bit. We hold the parent's mutex,
4885 * so we won't race with perf_event_{en, dis}able_family.
4887 if (parent_event->state >= PERF_EVENT_STATE_INACTIVE)
4888 child_event->state = PERF_EVENT_STATE_INACTIVE;
4890 child_event->state = PERF_EVENT_STATE_OFF;
4892 if (parent_event->attr.freq)
4893 child_event->hw.sample_period = parent_event->hw.sample_period;
4895 child_event->overflow_handler = parent_event->overflow_handler;
4898 * Link it up in the child's context:
4900 add_event_to_ctx(child_event, child_ctx);
4903 * Get a reference to the parent filp - we will fput it
4904 * when the child event exits. This is safe to do because
4905 * we are in the parent and we know that the filp still
4906 * exists and has a nonzero count:
4908 atomic_long_inc(&parent_event->filp->f_count);
4911 * Link this into the parent event's child list
4913 WARN_ON_ONCE(parent_event->ctx->parent_ctx);
4914 mutex_lock(&parent_event->child_mutex);
4915 list_add_tail(&child_event->child_list, &parent_event->child_list);
4916 mutex_unlock(&parent_event->child_mutex);
4921 static int inherit_group(struct perf_event *parent_event,
4922 struct task_struct *parent,
4923 struct perf_event_context *parent_ctx,
4924 struct task_struct *child,
4925 struct perf_event_context *child_ctx)
4927 struct perf_event *leader;
4928 struct perf_event *sub;
4929 struct perf_event *child_ctr;
4931 leader = inherit_event(parent_event, parent, parent_ctx,
4932 child, NULL, child_ctx);
4934 return PTR_ERR(leader);
4935 list_for_each_entry(sub, &parent_event->sibling_list, group_entry) {
4936 child_ctr = inherit_event(sub, parent, parent_ctx,
4937 child, leader, child_ctx);
4938 if (IS_ERR(child_ctr))
4939 return PTR_ERR(child_ctr);
4944 static void sync_child_event(struct perf_event *child_event,
4945 struct task_struct *child)
4947 struct perf_event *parent_event = child_event->parent;
4950 if (child_event->attr.inherit_stat)
4951 perf_event_read_event(child_event, child);
4953 child_val = atomic64_read(&child_event->count);
4956 * Add back the child's count to the parent's count:
4958 atomic64_add(child_val, &parent_event->count);
4959 atomic64_add(child_event->total_time_enabled,
4960 &parent_event->child_total_time_enabled);
4961 atomic64_add(child_event->total_time_running,
4962 &parent_event->child_total_time_running);
4965 * Remove this event from the parent's list
4967 WARN_ON_ONCE(parent_event->ctx->parent_ctx);
4968 mutex_lock(&parent_event->child_mutex);
4969 list_del_init(&child_event->child_list);
4970 mutex_unlock(&parent_event->child_mutex);
4973 * Release the parent event, if this was the last
4976 fput(parent_event->filp);
4980 __perf_event_exit_task(struct perf_event *child_event,
4981 struct perf_event_context *child_ctx,
4982 struct task_struct *child)
4984 struct perf_event *parent_event;
4986 perf_event_remove_from_context(child_event);
4988 parent_event = child_event->parent;
4990 * It can happen that parent exits first, and has events
4991 * that are still around due to the child reference. These
4992 * events need to be zapped - but otherwise linger.
4995 sync_child_event(child_event, child);
4996 free_event(child_event);
5001 * When a child task exits, feed back event values to parent events.
5003 void perf_event_exit_task(struct task_struct *child)
5005 struct perf_event *child_event, *tmp;
5006 struct perf_event_context *child_ctx;
5007 unsigned long flags;
5009 if (likely(!child->perf_event_ctxp)) {
5010 perf_event_task(child, NULL, 0);
5014 local_irq_save(flags);
5016 * We can't reschedule here because interrupts are disabled,
5017 * and either child is current or it is a task that can't be
5018 * scheduled, so we are now safe from rescheduling changing
5021 child_ctx = child->perf_event_ctxp;
5022 __perf_event_task_sched_out(child_ctx);
5025 * Take the context lock here so that if find_get_context is
5026 * reading child->perf_event_ctxp, we wait until it has
5027 * incremented the context's refcount before we do put_ctx below.
5029 raw_spin_lock(&child_ctx->lock);
5030 child->perf_event_ctxp = NULL;
5032 * If this context is a clone; unclone it so it can't get
5033 * swapped to another process while we're removing all
5034 * the events from it.
5036 unclone_ctx(child_ctx);
5037 update_context_time(child_ctx);
5038 raw_spin_unlock_irqrestore(&child_ctx->lock, flags);
5041 * Report the task dead after unscheduling the events so that we
5042 * won't get any samples after PERF_RECORD_EXIT. We can however still
5043 * get a few PERF_RECORD_READ events.
5045 perf_event_task(child, child_ctx, 0);
5048 * We can recurse on the same lock type through:
5050 * __perf_event_exit_task()
5051 * sync_child_event()
5052 * fput(parent_event->filp)
5054 * mutex_lock(&ctx->mutex)
5056 * But since its the parent context it won't be the same instance.
5058 mutex_lock_nested(&child_ctx->mutex, SINGLE_DEPTH_NESTING);
5061 list_for_each_entry_safe(child_event, tmp, &child_ctx->pinned_groups,
5063 __perf_event_exit_task(child_event, child_ctx, child);
5065 list_for_each_entry_safe(child_event, tmp, &child_ctx->flexible_groups,
5067 __perf_event_exit_task(child_event, child_ctx, child);
5070 * If the last event was a group event, it will have appended all
5071 * its siblings to the list, but we obtained 'tmp' before that which
5072 * will still point to the list head terminating the iteration.
5074 if (!list_empty(&child_ctx->pinned_groups) ||
5075 !list_empty(&child_ctx->flexible_groups))
5078 mutex_unlock(&child_ctx->mutex);
5083 static void perf_free_event(struct perf_event *event,
5084 struct perf_event_context *ctx)
5086 struct perf_event *parent = event->parent;
5088 if (WARN_ON_ONCE(!parent))
5091 mutex_lock(&parent->child_mutex);
5092 list_del_init(&event->child_list);
5093 mutex_unlock(&parent->child_mutex);
5097 list_del_event(event, ctx);
5102 * free an unexposed, unused context as created by inheritance by
5103 * init_task below, used by fork() in case of fail.
5105 void perf_event_free_task(struct task_struct *task)
5107 struct perf_event_context *ctx = task->perf_event_ctxp;
5108 struct perf_event *event, *tmp;
5113 mutex_lock(&ctx->mutex);
5115 list_for_each_entry_safe(event, tmp, &ctx->pinned_groups, group_entry)
5116 perf_free_event(event, ctx);
5118 list_for_each_entry_safe(event, tmp, &ctx->flexible_groups,
5120 perf_free_event(event, ctx);
5122 if (!list_empty(&ctx->pinned_groups) ||
5123 !list_empty(&ctx->flexible_groups))
5126 mutex_unlock(&ctx->mutex);
5132 inherit_task_group(struct perf_event *event, struct task_struct *parent,
5133 struct perf_event_context *parent_ctx,
5134 struct task_struct *child,
5138 struct perf_event_context *child_ctx = child->perf_event_ctxp;
5140 if (!event->attr.inherit) {
5147 * This is executed from the parent task context, so
5148 * inherit events that have been marked for cloning.
5149 * First allocate and initialize a context for the
5153 child_ctx = kzalloc(sizeof(struct perf_event_context),
5158 __perf_event_init_context(child_ctx, child);
5159 child->perf_event_ctxp = child_ctx;
5160 get_task_struct(child);
5163 ret = inherit_group(event, parent, parent_ctx,
5174 * Initialize the perf_event context in task_struct
5176 int perf_event_init_task(struct task_struct *child)
5178 struct perf_event_context *child_ctx, *parent_ctx;
5179 struct perf_event_context *cloned_ctx;
5180 struct perf_event *event;
5181 struct task_struct *parent = current;
5182 int inherited_all = 1;
5185 child->perf_event_ctxp = NULL;
5187 mutex_init(&child->perf_event_mutex);
5188 INIT_LIST_HEAD(&child->perf_event_list);
5190 if (likely(!parent->perf_event_ctxp))
5194 * If the parent's context is a clone, pin it so it won't get
5197 parent_ctx = perf_pin_task_context(parent);
5200 * No need to check if parent_ctx != NULL here; since we saw
5201 * it non-NULL earlier, the only reason for it to become NULL
5202 * is if we exit, and since we're currently in the middle of
5203 * a fork we can't be exiting at the same time.
5207 * Lock the parent list. No need to lock the child - not PID
5208 * hashed yet and not running, so nobody can access it.
5210 mutex_lock(&parent_ctx->mutex);
5213 * We dont have to disable NMIs - we are only looking at
5214 * the list, not manipulating it:
5216 list_for_each_entry(event, &parent_ctx->pinned_groups, group_entry) {
5217 ret = inherit_task_group(event, parent, parent_ctx, child,
5223 list_for_each_entry(event, &parent_ctx->flexible_groups, group_entry) {
5224 ret = inherit_task_group(event, parent, parent_ctx, child,
5230 child_ctx = child->perf_event_ctxp;
5232 if (child_ctx && inherited_all) {
5234 * Mark the child context as a clone of the parent
5235 * context, or of whatever the parent is a clone of.
5236 * Note that if the parent is a clone, it could get
5237 * uncloned at any point, but that doesn't matter
5238 * because the list of events and the generation
5239 * count can't have changed since we took the mutex.
5241 cloned_ctx = rcu_dereference(parent_ctx->parent_ctx);
5243 child_ctx->parent_ctx = cloned_ctx;
5244 child_ctx->parent_gen = parent_ctx->parent_gen;
5246 child_ctx->parent_ctx = parent_ctx;
5247 child_ctx->parent_gen = parent_ctx->generation;
5249 get_ctx(child_ctx->parent_ctx);
5252 mutex_unlock(&parent_ctx->mutex);
5254 perf_unpin_context(parent_ctx);
5259 static void __cpuinit perf_event_init_cpu(int cpu)
5261 struct perf_cpu_context *cpuctx;
5263 cpuctx = &per_cpu(perf_cpu_context, cpu);
5264 __perf_event_init_context(&cpuctx->ctx, NULL);
5266 spin_lock(&perf_resource_lock);
5267 cpuctx->max_pertask = perf_max_events - perf_reserved_percpu;
5268 spin_unlock(&perf_resource_lock);
5270 hw_perf_event_setup(cpu);
5273 #ifdef CONFIG_HOTPLUG_CPU
5274 static void __perf_event_exit_cpu(void *info)
5276 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
5277 struct perf_event_context *ctx = &cpuctx->ctx;
5278 struct perf_event *event, *tmp;
5280 list_for_each_entry_safe(event, tmp, &ctx->pinned_groups, group_entry)
5281 __perf_event_remove_from_context(event);
5282 list_for_each_entry_safe(event, tmp, &ctx->flexible_groups, group_entry)
5283 __perf_event_remove_from_context(event);
5285 static void perf_event_exit_cpu(int cpu)
5287 struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
5288 struct perf_event_context *ctx = &cpuctx->ctx;
5290 mutex_lock(&ctx->mutex);
5291 smp_call_function_single(cpu, __perf_event_exit_cpu, NULL, 1);
5292 mutex_unlock(&ctx->mutex);
5295 static inline void perf_event_exit_cpu(int cpu) { }
5298 static int __cpuinit
5299 perf_cpu_notify(struct notifier_block *self, unsigned long action, void *hcpu)
5301 unsigned int cpu = (long)hcpu;
5305 case CPU_UP_PREPARE:
5306 case CPU_UP_PREPARE_FROZEN:
5307 perf_event_init_cpu(cpu);
5311 case CPU_ONLINE_FROZEN:
5312 hw_perf_event_setup_online(cpu);
5315 case CPU_DOWN_PREPARE:
5316 case CPU_DOWN_PREPARE_FROZEN:
5317 perf_event_exit_cpu(cpu);
5328 * This has to have a higher priority than migration_notifier in sched.c.
5330 static struct notifier_block __cpuinitdata perf_cpu_nb = {
5331 .notifier_call = perf_cpu_notify,
5335 void __init perf_event_init(void)
5337 perf_cpu_notify(&perf_cpu_nb, (unsigned long)CPU_UP_PREPARE,
5338 (void *)(long)smp_processor_id());
5339 perf_cpu_notify(&perf_cpu_nb, (unsigned long)CPU_ONLINE,
5340 (void *)(long)smp_processor_id());
5341 register_cpu_notifier(&perf_cpu_nb);
5344 static ssize_t perf_show_reserve_percpu(struct sysdev_class *class, char *buf)
5346 return sprintf(buf, "%d\n", perf_reserved_percpu);
5350 perf_set_reserve_percpu(struct sysdev_class *class,
5354 struct perf_cpu_context *cpuctx;
5358 err = strict_strtoul(buf, 10, &val);
5361 if (val > perf_max_events)
5364 spin_lock(&perf_resource_lock);
5365 perf_reserved_percpu = val;
5366 for_each_online_cpu(cpu) {
5367 cpuctx = &per_cpu(perf_cpu_context, cpu);
5368 raw_spin_lock_irq(&cpuctx->ctx.lock);
5369 mpt = min(perf_max_events - cpuctx->ctx.nr_events,
5370 perf_max_events - perf_reserved_percpu);
5371 cpuctx->max_pertask = mpt;
5372 raw_spin_unlock_irq(&cpuctx->ctx.lock);
5374 spin_unlock(&perf_resource_lock);
5379 static ssize_t perf_show_overcommit(struct sysdev_class *class, char *buf)
5381 return sprintf(buf, "%d\n", perf_overcommit);
5385 perf_set_overcommit(struct sysdev_class *class, const char *buf, size_t count)
5390 err = strict_strtoul(buf, 10, &val);
5396 spin_lock(&perf_resource_lock);
5397 perf_overcommit = val;
5398 spin_unlock(&perf_resource_lock);
5403 static SYSDEV_CLASS_ATTR(
5406 perf_show_reserve_percpu,
5407 perf_set_reserve_percpu
5410 static SYSDEV_CLASS_ATTR(
5413 perf_show_overcommit,
5417 static struct attribute *perfclass_attrs[] = {
5418 &attr_reserve_percpu.attr,
5419 &attr_overcommit.attr,
5423 static struct attribute_group perfclass_attr_group = {
5424 .attrs = perfclass_attrs,
5425 .name = "perf_events",
5428 static int __init perf_event_sysfs_init(void)
5430 return sysfs_create_group(&cpu_sysdev_class.kset.kobj,
5431 &perfclass_attr_group);
5433 device_initcall(perf_event_sysfs_init);