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>
32 #include <asm/irq_regs.h>
35 * Each CPU has a list of per CPU events:
37 DEFINE_PER_CPU(struct perf_cpu_context, perf_cpu_context);
39 int perf_max_events __read_mostly = 1;
40 static int perf_reserved_percpu __read_mostly;
41 static int perf_overcommit __read_mostly = 1;
43 static atomic_t nr_events __read_mostly;
44 static atomic_t nr_mmap_events __read_mostly;
45 static atomic_t nr_comm_events __read_mostly;
46 static atomic_t nr_task_events __read_mostly;
49 * perf event paranoia level:
50 * -1 - not paranoid at all
51 * 0 - disallow raw tracepoint access for unpriv
52 * 1 - disallow cpu events for unpriv
53 * 2 - disallow kernel profiling for unpriv
55 int sysctl_perf_event_paranoid __read_mostly = 1;
57 static inline bool perf_paranoid_tracepoint_raw(void)
59 return sysctl_perf_event_paranoid > -1;
62 static inline bool perf_paranoid_cpu(void)
64 return sysctl_perf_event_paranoid > 0;
67 static inline bool perf_paranoid_kernel(void)
69 return sysctl_perf_event_paranoid > 1;
72 int sysctl_perf_event_mlock __read_mostly = 512; /* 'free' kb per user */
75 * max perf event sample rate
77 int sysctl_perf_event_sample_rate __read_mostly = 100000;
79 static atomic64_t perf_event_id;
82 * Lock for (sysadmin-configurable) event reservations:
84 static DEFINE_SPINLOCK(perf_resource_lock);
87 * Architecture provided APIs - weak aliases:
89 extern __weak const struct pmu *hw_perf_event_init(struct perf_event *event)
94 void __weak hw_perf_disable(void) { barrier(); }
95 void __weak hw_perf_enable(void) { barrier(); }
97 void __weak hw_perf_event_setup(int cpu) { barrier(); }
98 void __weak hw_perf_event_setup_online(int cpu) { barrier(); }
101 hw_perf_group_sched_in(struct perf_event *group_leader,
102 struct perf_cpu_context *cpuctx,
103 struct perf_event_context *ctx, int cpu)
108 void __weak perf_event_print_debug(void) { }
110 static DEFINE_PER_CPU(int, perf_disable_count);
112 void __perf_disable(void)
114 __get_cpu_var(perf_disable_count)++;
117 bool __perf_enable(void)
119 return !--__get_cpu_var(perf_disable_count);
122 void perf_disable(void)
128 void perf_enable(void)
134 static void get_ctx(struct perf_event_context *ctx)
136 WARN_ON(!atomic_inc_not_zero(&ctx->refcount));
139 static void free_ctx(struct rcu_head *head)
141 struct perf_event_context *ctx;
143 ctx = container_of(head, struct perf_event_context, rcu_head);
147 static void put_ctx(struct perf_event_context *ctx)
149 if (atomic_dec_and_test(&ctx->refcount)) {
151 put_ctx(ctx->parent_ctx);
153 put_task_struct(ctx->task);
154 call_rcu(&ctx->rcu_head, free_ctx);
158 static void unclone_ctx(struct perf_event_context *ctx)
160 if (ctx->parent_ctx) {
161 put_ctx(ctx->parent_ctx);
162 ctx->parent_ctx = NULL;
167 * If we inherit events we want to return the parent event id
170 static u64 primary_event_id(struct perf_event *event)
175 id = event->parent->id;
181 * Get the perf_event_context for a task and lock it.
182 * This has to cope with with the fact that until it is locked,
183 * the context could get moved to another task.
185 static struct perf_event_context *
186 perf_lock_task_context(struct task_struct *task, unsigned long *flags)
188 struct perf_event_context *ctx;
192 ctx = rcu_dereference(task->perf_event_ctxp);
195 * If this context is a clone of another, it might
196 * get swapped for another underneath us by
197 * perf_event_task_sched_out, though the
198 * rcu_read_lock() protects us from any context
199 * getting freed. Lock the context and check if it
200 * got swapped before we could get the lock, and retry
201 * if so. If we locked the right context, then it
202 * can't get swapped on us any more.
204 spin_lock_irqsave(&ctx->lock, *flags);
205 if (ctx != rcu_dereference(task->perf_event_ctxp)) {
206 spin_unlock_irqrestore(&ctx->lock, *flags);
210 if (!atomic_inc_not_zero(&ctx->refcount)) {
211 spin_unlock_irqrestore(&ctx->lock, *flags);
220 * Get the context for a task and increment its pin_count so it
221 * can't get swapped to another task. This also increments its
222 * reference count so that the context can't get freed.
224 static struct perf_event_context *perf_pin_task_context(struct task_struct *task)
226 struct perf_event_context *ctx;
229 ctx = perf_lock_task_context(task, &flags);
232 spin_unlock_irqrestore(&ctx->lock, flags);
237 static void perf_unpin_context(struct perf_event_context *ctx)
241 spin_lock_irqsave(&ctx->lock, flags);
243 spin_unlock_irqrestore(&ctx->lock, flags);
248 * Add a event from the lists for its context.
249 * Must be called with ctx->mutex and ctx->lock held.
252 list_add_event(struct perf_event *event, struct perf_event_context *ctx)
254 struct perf_event *group_leader = event->group_leader;
257 * Depending on whether it is a standalone or sibling event,
258 * add it straight to the context's event list, or to the group
259 * leader's sibling list:
261 if (group_leader == event)
262 list_add_tail(&event->group_entry, &ctx->group_list);
264 list_add_tail(&event->group_entry, &group_leader->sibling_list);
265 group_leader->nr_siblings++;
268 list_add_rcu(&event->event_entry, &ctx->event_list);
270 if (event->attr.inherit_stat)
275 * Remove a event from the lists for its context.
276 * Must be called with ctx->mutex and ctx->lock held.
279 list_del_event(struct perf_event *event, struct perf_event_context *ctx)
281 struct perf_event *sibling, *tmp;
283 if (list_empty(&event->group_entry))
286 if (event->attr.inherit_stat)
289 list_del_init(&event->group_entry);
290 list_del_rcu(&event->event_entry);
292 if (event->group_leader != event)
293 event->group_leader->nr_siblings--;
296 * If this was a group event with sibling events then
297 * upgrade the siblings to singleton events by adding them
298 * to the context list directly:
300 list_for_each_entry_safe(sibling, tmp, &event->sibling_list, group_entry) {
302 list_move_tail(&sibling->group_entry, &ctx->group_list);
303 sibling->group_leader = sibling;
308 event_sched_out(struct perf_event *event,
309 struct perf_cpu_context *cpuctx,
310 struct perf_event_context *ctx)
312 if (event->state != PERF_EVENT_STATE_ACTIVE)
315 event->state = PERF_EVENT_STATE_INACTIVE;
316 if (event->pending_disable) {
317 event->pending_disable = 0;
318 event->state = PERF_EVENT_STATE_OFF;
320 event->tstamp_stopped = ctx->time;
321 event->pmu->disable(event);
324 if (!is_software_event(event))
325 cpuctx->active_oncpu--;
327 if (event->attr.exclusive || !cpuctx->active_oncpu)
328 cpuctx->exclusive = 0;
332 group_sched_out(struct perf_event *group_event,
333 struct perf_cpu_context *cpuctx,
334 struct perf_event_context *ctx)
336 struct perf_event *event;
338 if (group_event->state != PERF_EVENT_STATE_ACTIVE)
341 event_sched_out(group_event, cpuctx, ctx);
344 * Schedule out siblings (if any):
346 list_for_each_entry(event, &group_event->sibling_list, group_entry)
347 event_sched_out(event, cpuctx, ctx);
349 if (group_event->attr.exclusive)
350 cpuctx->exclusive = 0;
354 * Cross CPU call to remove a performance event
356 * We disable the event on the hardware level first. After that we
357 * remove it from the context list.
359 static void __perf_event_remove_from_context(void *info)
361 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
362 struct perf_event *event = info;
363 struct perf_event_context *ctx = event->ctx;
366 * If this is a task context, we need to check whether it is
367 * the current task context of this cpu. If not it has been
368 * scheduled out before the smp call arrived.
370 if (ctx->task && cpuctx->task_ctx != ctx)
373 spin_lock(&ctx->lock);
375 * Protect the list operation against NMI by disabling the
376 * events on a global level.
380 event_sched_out(event, cpuctx, ctx);
382 list_del_event(event, ctx);
386 * Allow more per task events with respect to the
389 cpuctx->max_pertask =
390 min(perf_max_events - ctx->nr_events,
391 perf_max_events - perf_reserved_percpu);
395 spin_unlock(&ctx->lock);
400 * Remove the event from a task's (or a CPU's) list of events.
402 * Must be called with ctx->mutex held.
404 * CPU events are removed with a smp call. For task events we only
405 * call when the task is on a CPU.
407 * If event->ctx is a cloned context, callers must make sure that
408 * every task struct that event->ctx->task could possibly point to
409 * remains valid. This is OK when called from perf_release since
410 * that only calls us on the top-level context, which can't be a clone.
411 * When called from perf_event_exit_task, it's OK because the
412 * context has been detached from its task.
414 static void perf_event_remove_from_context(struct perf_event *event)
416 struct perf_event_context *ctx = event->ctx;
417 struct task_struct *task = ctx->task;
421 * Per cpu events are removed via an smp call and
422 * the removal is always sucessful.
424 smp_call_function_single(event->cpu,
425 __perf_event_remove_from_context,
431 task_oncpu_function_call(task, __perf_event_remove_from_context,
434 spin_lock_irq(&ctx->lock);
436 * If the context is active we need to retry the smp call.
438 if (ctx->nr_active && !list_empty(&event->group_entry)) {
439 spin_unlock_irq(&ctx->lock);
444 * The lock prevents that this context is scheduled in so we
445 * can remove the event safely, if the call above did not
448 if (!list_empty(&event->group_entry)) {
449 list_del_event(event, ctx);
451 spin_unlock_irq(&ctx->lock);
454 static inline u64 perf_clock(void)
456 return cpu_clock(smp_processor_id());
460 * Update the record of the current time in a context.
462 static void update_context_time(struct perf_event_context *ctx)
464 u64 now = perf_clock();
466 ctx->time += now - ctx->timestamp;
467 ctx->timestamp = now;
471 * Update the total_time_enabled and total_time_running fields for a event.
473 static void update_event_times(struct perf_event *event)
475 struct perf_event_context *ctx = event->ctx;
478 if (event->state < PERF_EVENT_STATE_INACTIVE ||
479 event->group_leader->state < PERF_EVENT_STATE_INACTIVE)
482 event->total_time_enabled = ctx->time - event->tstamp_enabled;
484 if (event->state == PERF_EVENT_STATE_INACTIVE)
485 run_end = event->tstamp_stopped;
489 event->total_time_running = run_end - event->tstamp_running;
493 * Update total_time_enabled and total_time_running for all events in a group.
495 static void update_group_times(struct perf_event *leader)
497 struct perf_event *event;
499 update_event_times(leader);
500 list_for_each_entry(event, &leader->sibling_list, group_entry)
501 update_event_times(event);
505 * Cross CPU call to disable a performance event
507 static void __perf_event_disable(void *info)
509 struct perf_event *event = info;
510 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
511 struct perf_event_context *ctx = event->ctx;
514 * If this is a per-task event, need to check whether this
515 * event's task is the current task on this cpu.
517 if (ctx->task && cpuctx->task_ctx != ctx)
520 spin_lock(&ctx->lock);
523 * If the event is on, turn it off.
524 * If it is in error state, leave it in error state.
526 if (event->state >= PERF_EVENT_STATE_INACTIVE) {
527 update_context_time(ctx);
528 update_group_times(event);
529 if (event == event->group_leader)
530 group_sched_out(event, cpuctx, ctx);
532 event_sched_out(event, cpuctx, ctx);
533 event->state = PERF_EVENT_STATE_OFF;
536 spin_unlock(&ctx->lock);
542 * If event->ctx is a cloned context, callers must make sure that
543 * every task struct that event->ctx->task could possibly point to
544 * remains valid. This condition is satisifed when called through
545 * perf_event_for_each_child or perf_event_for_each because they
546 * hold the top-level event's child_mutex, so any descendant that
547 * goes to exit will block in sync_child_event.
548 * When called from perf_pending_event it's OK because event->ctx
549 * is the current context on this CPU and preemption is disabled,
550 * hence we can't get into perf_event_task_sched_out for this context.
552 static void perf_event_disable(struct perf_event *event)
554 struct perf_event_context *ctx = event->ctx;
555 struct task_struct *task = ctx->task;
559 * Disable the event on the cpu that it's on
561 smp_call_function_single(event->cpu, __perf_event_disable,
567 task_oncpu_function_call(task, __perf_event_disable, event);
569 spin_lock_irq(&ctx->lock);
571 * If the event is still active, we need to retry the cross-call.
573 if (event->state == PERF_EVENT_STATE_ACTIVE) {
574 spin_unlock_irq(&ctx->lock);
579 * Since we have the lock this context can't be scheduled
580 * in, so we can change the state safely.
582 if (event->state == PERF_EVENT_STATE_INACTIVE) {
583 update_group_times(event);
584 event->state = PERF_EVENT_STATE_OFF;
587 spin_unlock_irq(&ctx->lock);
591 event_sched_in(struct perf_event *event,
592 struct perf_cpu_context *cpuctx,
593 struct perf_event_context *ctx,
596 if (event->state <= PERF_EVENT_STATE_OFF)
599 event->state = PERF_EVENT_STATE_ACTIVE;
600 event->oncpu = cpu; /* TODO: put 'cpu' into cpuctx->cpu */
602 * The new state must be visible before we turn it on in the hardware:
606 if (event->pmu->enable(event)) {
607 event->state = PERF_EVENT_STATE_INACTIVE;
612 event->tstamp_running += ctx->time - event->tstamp_stopped;
614 if (!is_software_event(event))
615 cpuctx->active_oncpu++;
618 if (event->attr.exclusive)
619 cpuctx->exclusive = 1;
625 group_sched_in(struct perf_event *group_event,
626 struct perf_cpu_context *cpuctx,
627 struct perf_event_context *ctx,
630 struct perf_event *event, *partial_group;
633 if (group_event->state == PERF_EVENT_STATE_OFF)
636 ret = hw_perf_group_sched_in(group_event, cpuctx, ctx, cpu);
638 return ret < 0 ? ret : 0;
640 if (event_sched_in(group_event, cpuctx, ctx, cpu))
644 * Schedule in siblings as one group (if any):
646 list_for_each_entry(event, &group_event->sibling_list, group_entry) {
647 if (event_sched_in(event, cpuctx, ctx, cpu)) {
648 partial_group = event;
657 * Groups can be scheduled in as one unit only, so undo any
658 * partial group before returning:
660 list_for_each_entry(event, &group_event->sibling_list, group_entry) {
661 if (event == partial_group)
663 event_sched_out(event, cpuctx, ctx);
665 event_sched_out(group_event, cpuctx, ctx);
671 * Return 1 for a group consisting entirely of software events,
672 * 0 if the group contains any hardware events.
674 static int is_software_only_group(struct perf_event *leader)
676 struct perf_event *event;
678 if (!is_software_event(leader))
681 list_for_each_entry(event, &leader->sibling_list, group_entry)
682 if (!is_software_event(event))
689 * Work out whether we can put this event group on the CPU now.
691 static int group_can_go_on(struct perf_event *event,
692 struct perf_cpu_context *cpuctx,
696 * Groups consisting entirely of software events can always go on.
698 if (is_software_only_group(event))
701 * If an exclusive group is already on, no other hardware
704 if (cpuctx->exclusive)
707 * If this group is exclusive and there are already
708 * events on the CPU, it can't go on.
710 if (event->attr.exclusive && cpuctx->active_oncpu)
713 * Otherwise, try to add it if all previous groups were able
719 static void add_event_to_ctx(struct perf_event *event,
720 struct perf_event_context *ctx)
722 list_add_event(event, ctx);
723 event->tstamp_enabled = ctx->time;
724 event->tstamp_running = ctx->time;
725 event->tstamp_stopped = ctx->time;
729 * Cross CPU call to install and enable a performance event
731 * Must be called with ctx->mutex held
733 static void __perf_install_in_context(void *info)
735 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
736 struct perf_event *event = info;
737 struct perf_event_context *ctx = event->ctx;
738 struct perf_event *leader = event->group_leader;
739 int cpu = smp_processor_id();
743 * If this is a task context, we need to check whether it is
744 * the current task context of this cpu. If not it has been
745 * scheduled out before the smp call arrived.
746 * Or possibly this is the right context but it isn't
747 * on this cpu because it had no events.
749 if (ctx->task && cpuctx->task_ctx != ctx) {
750 if (cpuctx->task_ctx || ctx->task != current)
752 cpuctx->task_ctx = ctx;
755 spin_lock(&ctx->lock);
757 update_context_time(ctx);
760 * Protect the list operation against NMI by disabling the
761 * events on a global level. NOP for non NMI based events.
765 add_event_to_ctx(event, ctx);
768 * Don't put the event on if it is disabled or if
769 * it is in a group and the group isn't on.
771 if (event->state != PERF_EVENT_STATE_INACTIVE ||
772 (leader != event && leader->state != PERF_EVENT_STATE_ACTIVE))
776 * An exclusive event can't go on if there are already active
777 * hardware events, and no hardware event can go on if there
778 * is already an exclusive event on.
780 if (!group_can_go_on(event, cpuctx, 1))
783 err = event_sched_in(event, cpuctx, ctx, cpu);
787 * This event couldn't go on. If it is in a group
788 * then we have to pull the whole group off.
789 * If the event group is pinned then put it in error state.
792 group_sched_out(leader, cpuctx, ctx);
793 if (leader->attr.pinned) {
794 update_group_times(leader);
795 leader->state = PERF_EVENT_STATE_ERROR;
799 if (!err && !ctx->task && cpuctx->max_pertask)
800 cpuctx->max_pertask--;
805 spin_unlock(&ctx->lock);
809 * Attach a performance event to a context
811 * First we add the event to the list with the hardware enable bit
812 * in event->hw_config cleared.
814 * If the event is attached to a task which is on a CPU we use a smp
815 * call to enable it in the task context. The task might have been
816 * scheduled away, but we check this in the smp call again.
818 * Must be called with ctx->mutex held.
821 perf_install_in_context(struct perf_event_context *ctx,
822 struct perf_event *event,
825 struct task_struct *task = ctx->task;
829 * Per cpu events are installed via an smp call and
830 * the install is always sucessful.
832 smp_call_function_single(cpu, __perf_install_in_context,
838 task_oncpu_function_call(task, __perf_install_in_context,
841 spin_lock_irq(&ctx->lock);
843 * we need to retry the smp call.
845 if (ctx->is_active && list_empty(&event->group_entry)) {
846 spin_unlock_irq(&ctx->lock);
851 * The lock prevents that this context is scheduled in so we
852 * can add the event safely, if it the call above did not
855 if (list_empty(&event->group_entry))
856 add_event_to_ctx(event, ctx);
857 spin_unlock_irq(&ctx->lock);
861 * Put a event into inactive state and update time fields.
862 * Enabling the leader of a group effectively enables all
863 * the group members that aren't explicitly disabled, so we
864 * have to update their ->tstamp_enabled also.
865 * Note: this works for group members as well as group leaders
866 * since the non-leader members' sibling_lists will be empty.
868 static void __perf_event_mark_enabled(struct perf_event *event,
869 struct perf_event_context *ctx)
871 struct perf_event *sub;
873 event->state = PERF_EVENT_STATE_INACTIVE;
874 event->tstamp_enabled = ctx->time - event->total_time_enabled;
875 list_for_each_entry(sub, &event->sibling_list, group_entry)
876 if (sub->state >= PERF_EVENT_STATE_INACTIVE)
877 sub->tstamp_enabled =
878 ctx->time - sub->total_time_enabled;
882 * Cross CPU call to enable a performance event
884 static void __perf_event_enable(void *info)
886 struct perf_event *event = info;
887 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
888 struct perf_event_context *ctx = event->ctx;
889 struct perf_event *leader = event->group_leader;
893 * If this is a per-task event, need to check whether this
894 * event's task is the current task on this cpu.
896 if (ctx->task && cpuctx->task_ctx != ctx) {
897 if (cpuctx->task_ctx || ctx->task != current)
899 cpuctx->task_ctx = ctx;
902 spin_lock(&ctx->lock);
904 update_context_time(ctx);
906 if (event->state >= PERF_EVENT_STATE_INACTIVE)
908 __perf_event_mark_enabled(event, ctx);
911 * If the event is in a group and isn't the group leader,
912 * then don't put it on unless the group is on.
914 if (leader != event && leader->state != PERF_EVENT_STATE_ACTIVE)
917 if (!group_can_go_on(event, cpuctx, 1)) {
922 err = group_sched_in(event, cpuctx, ctx,
925 err = event_sched_in(event, cpuctx, ctx,
932 * If this event can't go on and it's part of a
933 * group, then the whole group has to come off.
936 group_sched_out(leader, cpuctx, ctx);
937 if (leader->attr.pinned) {
938 update_group_times(leader);
939 leader->state = PERF_EVENT_STATE_ERROR;
944 spin_unlock(&ctx->lock);
950 * If event->ctx is a cloned context, callers must make sure that
951 * every task struct that event->ctx->task could possibly point to
952 * remains valid. This condition is satisfied when called through
953 * perf_event_for_each_child or perf_event_for_each as described
954 * for perf_event_disable.
956 static void perf_event_enable(struct perf_event *event)
958 struct perf_event_context *ctx = event->ctx;
959 struct task_struct *task = ctx->task;
963 * Enable the event on the cpu that it's on
965 smp_call_function_single(event->cpu, __perf_event_enable,
970 spin_lock_irq(&ctx->lock);
971 if (event->state >= PERF_EVENT_STATE_INACTIVE)
975 * If the event is in error state, clear that first.
976 * That way, if we see the event in error state below, we
977 * know that it has gone back into error state, as distinct
978 * from the task having been scheduled away before the
979 * cross-call arrived.
981 if (event->state == PERF_EVENT_STATE_ERROR)
982 event->state = PERF_EVENT_STATE_OFF;
985 spin_unlock_irq(&ctx->lock);
986 task_oncpu_function_call(task, __perf_event_enable, event);
988 spin_lock_irq(&ctx->lock);
991 * If the context is active and the event is still off,
992 * we need to retry the cross-call.
994 if (ctx->is_active && event->state == PERF_EVENT_STATE_OFF)
998 * Since we have the lock this context can't be scheduled
999 * in, so we can change the state safely.
1001 if (event->state == PERF_EVENT_STATE_OFF)
1002 __perf_event_mark_enabled(event, ctx);
1005 spin_unlock_irq(&ctx->lock);
1008 static int perf_event_refresh(struct perf_event *event, int refresh)
1011 * not supported on inherited events
1013 if (event->attr.inherit)
1016 atomic_add(refresh, &event->event_limit);
1017 perf_event_enable(event);
1022 void __perf_event_sched_out(struct perf_event_context *ctx,
1023 struct perf_cpu_context *cpuctx)
1025 struct perf_event *event;
1027 spin_lock(&ctx->lock);
1029 if (likely(!ctx->nr_events))
1031 update_context_time(ctx);
1035 list_for_each_entry(event, &ctx->group_list, group_entry)
1036 group_sched_out(event, cpuctx, ctx);
1040 spin_unlock(&ctx->lock);
1044 * Test whether two contexts are equivalent, i.e. whether they
1045 * have both been cloned from the same version of the same context
1046 * and they both have the same number of enabled events.
1047 * If the number of enabled events is the same, then the set
1048 * of enabled events should be the same, because these are both
1049 * inherited contexts, therefore we can't access individual events
1050 * in them directly with an fd; we can only enable/disable all
1051 * events via prctl, or enable/disable all events in a family
1052 * via ioctl, which will have the same effect on both contexts.
1054 static int context_equiv(struct perf_event_context *ctx1,
1055 struct perf_event_context *ctx2)
1057 return ctx1->parent_ctx && ctx1->parent_ctx == ctx2->parent_ctx
1058 && ctx1->parent_gen == ctx2->parent_gen
1059 && !ctx1->pin_count && !ctx2->pin_count;
1062 static void __perf_event_read(void *event);
1064 static void __perf_event_sync_stat(struct perf_event *event,
1065 struct perf_event *next_event)
1069 if (!event->attr.inherit_stat)
1073 * Update the event value, we cannot use perf_event_read()
1074 * because we're in the middle of a context switch and have IRQs
1075 * disabled, which upsets smp_call_function_single(), however
1076 * we know the event must be on the current CPU, therefore we
1077 * don't need to use it.
1079 switch (event->state) {
1080 case PERF_EVENT_STATE_ACTIVE:
1081 __perf_event_read(event);
1084 case PERF_EVENT_STATE_INACTIVE:
1085 update_event_times(event);
1093 * In order to keep per-task stats reliable we need to flip the event
1094 * values when we flip the contexts.
1096 value = atomic64_read(&next_event->count);
1097 value = atomic64_xchg(&event->count, value);
1098 atomic64_set(&next_event->count, value);
1100 swap(event->total_time_enabled, next_event->total_time_enabled);
1101 swap(event->total_time_running, next_event->total_time_running);
1104 * Since we swizzled the values, update the user visible data too.
1106 perf_event_update_userpage(event);
1107 perf_event_update_userpage(next_event);
1110 #define list_next_entry(pos, member) \
1111 list_entry(pos->member.next, typeof(*pos), member)
1113 static void perf_event_sync_stat(struct perf_event_context *ctx,
1114 struct perf_event_context *next_ctx)
1116 struct perf_event *event, *next_event;
1121 event = list_first_entry(&ctx->event_list,
1122 struct perf_event, event_entry);
1124 next_event = list_first_entry(&next_ctx->event_list,
1125 struct perf_event, event_entry);
1127 while (&event->event_entry != &ctx->event_list &&
1128 &next_event->event_entry != &next_ctx->event_list) {
1130 __perf_event_sync_stat(event, next_event);
1132 event = list_next_entry(event, event_entry);
1133 next_event = list_next_entry(next_event, event_entry);
1138 * Called from scheduler to remove the events of the current task,
1139 * with interrupts disabled.
1141 * We stop each event and update the event value in event->count.
1143 * This does not protect us against NMI, but disable()
1144 * sets the disabled bit in the control field of event _before_
1145 * accessing the event control register. If a NMI hits, then it will
1146 * not restart the event.
1148 void perf_event_task_sched_out(struct task_struct *task,
1149 struct task_struct *next, int cpu)
1151 struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
1152 struct perf_event_context *ctx = task->perf_event_ctxp;
1153 struct perf_event_context *next_ctx;
1154 struct perf_event_context *parent;
1155 struct pt_regs *regs;
1158 regs = task_pt_regs(task);
1159 perf_sw_event(PERF_COUNT_SW_CONTEXT_SWITCHES, 1, 1, regs, 0);
1161 if (likely(!ctx || !cpuctx->task_ctx))
1164 update_context_time(ctx);
1167 parent = rcu_dereference(ctx->parent_ctx);
1168 next_ctx = next->perf_event_ctxp;
1169 if (parent && next_ctx &&
1170 rcu_dereference(next_ctx->parent_ctx) == parent) {
1172 * Looks like the two contexts are clones, so we might be
1173 * able to optimize the context switch. We lock both
1174 * contexts and check that they are clones under the
1175 * lock (including re-checking that neither has been
1176 * uncloned in the meantime). It doesn't matter which
1177 * order we take the locks because no other cpu could
1178 * be trying to lock both of these tasks.
1180 spin_lock(&ctx->lock);
1181 spin_lock_nested(&next_ctx->lock, SINGLE_DEPTH_NESTING);
1182 if (context_equiv(ctx, next_ctx)) {
1184 * XXX do we need a memory barrier of sorts
1185 * wrt to rcu_dereference() of perf_event_ctxp
1187 task->perf_event_ctxp = next_ctx;
1188 next->perf_event_ctxp = ctx;
1190 next_ctx->task = task;
1193 perf_event_sync_stat(ctx, next_ctx);
1195 spin_unlock(&next_ctx->lock);
1196 spin_unlock(&ctx->lock);
1201 __perf_event_sched_out(ctx, cpuctx);
1202 cpuctx->task_ctx = NULL;
1207 * Called with IRQs disabled
1209 static void __perf_event_task_sched_out(struct perf_event_context *ctx)
1211 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
1213 if (!cpuctx->task_ctx)
1216 if (WARN_ON_ONCE(ctx != cpuctx->task_ctx))
1219 __perf_event_sched_out(ctx, cpuctx);
1220 cpuctx->task_ctx = NULL;
1224 * Called with IRQs disabled
1226 static void perf_event_cpu_sched_out(struct perf_cpu_context *cpuctx)
1228 __perf_event_sched_out(&cpuctx->ctx, cpuctx);
1232 __perf_event_sched_in(struct perf_event_context *ctx,
1233 struct perf_cpu_context *cpuctx, int cpu)
1235 struct perf_event *event;
1238 spin_lock(&ctx->lock);
1240 if (likely(!ctx->nr_events))
1243 ctx->timestamp = perf_clock();
1248 * First go through the list and put on any pinned groups
1249 * in order to give them the best chance of going on.
1251 list_for_each_entry(event, &ctx->group_list, group_entry) {
1252 if (event->state <= PERF_EVENT_STATE_OFF ||
1253 !event->attr.pinned)
1255 if (event->cpu != -1 && event->cpu != cpu)
1258 if (group_can_go_on(event, cpuctx, 1))
1259 group_sched_in(event, cpuctx, ctx, cpu);
1262 * If this pinned group hasn't been scheduled,
1263 * put it in error state.
1265 if (event->state == PERF_EVENT_STATE_INACTIVE) {
1266 update_group_times(event);
1267 event->state = PERF_EVENT_STATE_ERROR;
1271 list_for_each_entry(event, &ctx->group_list, group_entry) {
1273 * Ignore events in OFF or ERROR state, and
1274 * ignore pinned events since we did them already.
1276 if (event->state <= PERF_EVENT_STATE_OFF ||
1281 * Listen to the 'cpu' scheduling filter constraint
1284 if (event->cpu != -1 && event->cpu != cpu)
1287 if (group_can_go_on(event, cpuctx, can_add_hw))
1288 if (group_sched_in(event, cpuctx, ctx, cpu))
1293 spin_unlock(&ctx->lock);
1297 * Called from scheduler to add the events of the current task
1298 * with interrupts disabled.
1300 * We restore the event value and then enable it.
1302 * This does not protect us against NMI, but enable()
1303 * sets the enabled bit in the control field of event _before_
1304 * accessing the event control register. If a NMI hits, then it will
1305 * keep the event running.
1307 void perf_event_task_sched_in(struct task_struct *task, int cpu)
1309 struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
1310 struct perf_event_context *ctx = task->perf_event_ctxp;
1314 if (cpuctx->task_ctx == ctx)
1316 __perf_event_sched_in(ctx, cpuctx, cpu);
1317 cpuctx->task_ctx = ctx;
1320 static void perf_event_cpu_sched_in(struct perf_cpu_context *cpuctx, int cpu)
1322 struct perf_event_context *ctx = &cpuctx->ctx;
1324 __perf_event_sched_in(ctx, cpuctx, cpu);
1327 #define MAX_INTERRUPTS (~0ULL)
1329 static void perf_log_throttle(struct perf_event *event, int enable);
1331 static void perf_adjust_period(struct perf_event *event, u64 events)
1333 struct hw_perf_event *hwc = &event->hw;
1334 u64 period, sample_period;
1337 events *= hwc->sample_period;
1338 period = div64_u64(events, event->attr.sample_freq);
1340 delta = (s64)(period - hwc->sample_period);
1341 delta = (delta + 7) / 8; /* low pass filter */
1343 sample_period = hwc->sample_period + delta;
1348 hwc->sample_period = sample_period;
1351 static void perf_ctx_adjust_freq(struct perf_event_context *ctx)
1353 struct perf_event *event;
1354 struct hw_perf_event *hwc;
1355 u64 interrupts, freq;
1357 spin_lock(&ctx->lock);
1358 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
1359 if (event->state != PERF_EVENT_STATE_ACTIVE)
1364 interrupts = hwc->interrupts;
1365 hwc->interrupts = 0;
1368 * unthrottle events on the tick
1370 if (interrupts == MAX_INTERRUPTS) {
1371 perf_log_throttle(event, 1);
1372 event->pmu->unthrottle(event);
1373 interrupts = 2*sysctl_perf_event_sample_rate/HZ;
1376 if (!event->attr.freq || !event->attr.sample_freq)
1380 * if the specified freq < HZ then we need to skip ticks
1382 if (event->attr.sample_freq < HZ) {
1383 freq = event->attr.sample_freq;
1385 hwc->freq_count += freq;
1386 hwc->freq_interrupts += interrupts;
1388 if (hwc->freq_count < HZ)
1391 interrupts = hwc->freq_interrupts;
1392 hwc->freq_interrupts = 0;
1393 hwc->freq_count -= HZ;
1397 perf_adjust_period(event, freq * interrupts);
1400 * In order to avoid being stalled by an (accidental) huge
1401 * sample period, force reset the sample period if we didn't
1402 * get any events in this freq period.
1406 event->pmu->disable(event);
1407 atomic64_set(&hwc->period_left, 0);
1408 event->pmu->enable(event);
1412 spin_unlock(&ctx->lock);
1416 * Round-robin a context's events:
1418 static void rotate_ctx(struct perf_event_context *ctx)
1420 struct perf_event *event;
1422 if (!ctx->nr_events)
1425 spin_lock(&ctx->lock);
1427 * Rotate the first entry last (works just fine for group events too):
1430 list_for_each_entry(event, &ctx->group_list, group_entry) {
1431 list_move_tail(&event->group_entry, &ctx->group_list);
1436 spin_unlock(&ctx->lock);
1439 void perf_event_task_tick(struct task_struct *curr, int cpu)
1441 struct perf_cpu_context *cpuctx;
1442 struct perf_event_context *ctx;
1444 if (!atomic_read(&nr_events))
1447 cpuctx = &per_cpu(perf_cpu_context, cpu);
1448 ctx = curr->perf_event_ctxp;
1450 perf_ctx_adjust_freq(&cpuctx->ctx);
1452 perf_ctx_adjust_freq(ctx);
1454 perf_event_cpu_sched_out(cpuctx);
1456 __perf_event_task_sched_out(ctx);
1458 rotate_ctx(&cpuctx->ctx);
1462 perf_event_cpu_sched_in(cpuctx, cpu);
1464 perf_event_task_sched_in(curr, cpu);
1468 * Enable all of a task's events that have been marked enable-on-exec.
1469 * This expects task == current.
1471 static void perf_event_enable_on_exec(struct task_struct *task)
1473 struct perf_event_context *ctx;
1474 struct perf_event *event;
1475 unsigned long flags;
1478 local_irq_save(flags);
1479 ctx = task->perf_event_ctxp;
1480 if (!ctx || !ctx->nr_events)
1483 __perf_event_task_sched_out(ctx);
1485 spin_lock(&ctx->lock);
1487 list_for_each_entry(event, &ctx->group_list, group_entry) {
1488 if (!event->attr.enable_on_exec)
1490 event->attr.enable_on_exec = 0;
1491 if (event->state >= PERF_EVENT_STATE_INACTIVE)
1493 __perf_event_mark_enabled(event, ctx);
1498 * Unclone this context if we enabled any event.
1503 spin_unlock(&ctx->lock);
1505 perf_event_task_sched_in(task, smp_processor_id());
1507 local_irq_restore(flags);
1511 * Cross CPU call to read the hardware event
1513 static void __perf_event_read(void *info)
1515 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
1516 struct perf_event *event = info;
1517 struct perf_event_context *ctx = event->ctx;
1518 unsigned long flags;
1521 * If this is a task context, we need to check whether it is
1522 * the current task context of this cpu. If not it has been
1523 * scheduled out before the smp call arrived. In that case
1524 * event->count would have been updated to a recent sample
1525 * when the event was scheduled out.
1527 if (ctx->task && cpuctx->task_ctx != ctx)
1530 local_irq_save(flags);
1532 update_context_time(ctx);
1533 event->pmu->read(event);
1534 update_event_times(event);
1535 local_irq_restore(flags);
1538 static u64 perf_event_read(struct perf_event *event)
1541 * If event is enabled and currently active on a CPU, update the
1542 * value in the event structure:
1544 if (event->state == PERF_EVENT_STATE_ACTIVE) {
1545 smp_call_function_single(event->oncpu,
1546 __perf_event_read, event, 1);
1547 } else if (event->state == PERF_EVENT_STATE_INACTIVE) {
1548 update_event_times(event);
1551 return atomic64_read(&event->count);
1555 * Initialize the perf_event context in a task_struct:
1558 __perf_event_init_context(struct perf_event_context *ctx,
1559 struct task_struct *task)
1561 memset(ctx, 0, sizeof(*ctx));
1562 spin_lock_init(&ctx->lock);
1563 mutex_init(&ctx->mutex);
1564 INIT_LIST_HEAD(&ctx->group_list);
1565 INIT_LIST_HEAD(&ctx->event_list);
1566 atomic_set(&ctx->refcount, 1);
1570 static struct perf_event_context *find_get_context(pid_t pid, int cpu)
1572 struct perf_event_context *ctx;
1573 struct perf_cpu_context *cpuctx;
1574 struct task_struct *task;
1575 unsigned long flags;
1579 * If cpu is not a wildcard then this is a percpu event:
1582 /* Must be root to operate on a CPU event: */
1583 if (perf_paranoid_cpu() && !capable(CAP_SYS_ADMIN))
1584 return ERR_PTR(-EACCES);
1586 if (cpu < 0 || cpu >= nr_cpumask_bits)
1587 return ERR_PTR(-EINVAL);
1590 * We could be clever and allow to attach a event to an
1591 * offline CPU and activate it when the CPU comes up, but
1594 if (!cpu_isset(cpu, cpu_online_map))
1595 return ERR_PTR(-ENODEV);
1597 cpuctx = &per_cpu(perf_cpu_context, cpu);
1608 task = find_task_by_vpid(pid);
1610 get_task_struct(task);
1614 return ERR_PTR(-ESRCH);
1617 * Can't attach events to a dying task.
1620 if (task->flags & PF_EXITING)
1623 /* Reuse ptrace permission checks for now. */
1625 if (!ptrace_may_access(task, PTRACE_MODE_READ))
1629 ctx = perf_lock_task_context(task, &flags);
1632 spin_unlock_irqrestore(&ctx->lock, flags);
1636 ctx = kmalloc(sizeof(struct perf_event_context), GFP_KERNEL);
1640 __perf_event_init_context(ctx, task);
1642 if (cmpxchg(&task->perf_event_ctxp, NULL, ctx)) {
1644 * We raced with some other task; use
1645 * the context they set.
1650 get_task_struct(task);
1653 put_task_struct(task);
1657 put_task_struct(task);
1658 return ERR_PTR(err);
1661 static void free_event_rcu(struct rcu_head *head)
1663 struct perf_event *event;
1665 event = container_of(head, struct perf_event, rcu_head);
1667 put_pid_ns(event->ns);
1671 static void perf_pending_sync(struct perf_event *event);
1673 static void free_event(struct perf_event *event)
1675 perf_pending_sync(event);
1677 if (!event->parent) {
1678 atomic_dec(&nr_events);
1679 if (event->attr.mmap)
1680 atomic_dec(&nr_mmap_events);
1681 if (event->attr.comm)
1682 atomic_dec(&nr_comm_events);
1683 if (event->attr.task)
1684 atomic_dec(&nr_task_events);
1687 if (event->output) {
1688 fput(event->output->filp);
1689 event->output = NULL;
1693 event->destroy(event);
1695 put_ctx(event->ctx);
1696 call_rcu(&event->rcu_head, free_event_rcu);
1700 * Called when the last reference to the file is gone.
1702 static int perf_release(struct inode *inode, struct file *file)
1704 struct perf_event *event = file->private_data;
1705 struct perf_event_context *ctx = event->ctx;
1707 file->private_data = NULL;
1709 WARN_ON_ONCE(ctx->parent_ctx);
1710 mutex_lock(&ctx->mutex);
1711 perf_event_remove_from_context(event);
1712 mutex_unlock(&ctx->mutex);
1714 mutex_lock(&event->owner->perf_event_mutex);
1715 list_del_init(&event->owner_entry);
1716 mutex_unlock(&event->owner->perf_event_mutex);
1717 put_task_struct(event->owner);
1724 static int perf_event_read_size(struct perf_event *event)
1726 int entry = sizeof(u64); /* value */
1730 if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1731 size += sizeof(u64);
1733 if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1734 size += sizeof(u64);
1736 if (event->attr.read_format & PERF_FORMAT_ID)
1737 entry += sizeof(u64);
1739 if (event->attr.read_format & PERF_FORMAT_GROUP) {
1740 nr += event->group_leader->nr_siblings;
1741 size += sizeof(u64);
1749 static u64 perf_event_read_value(struct perf_event *event)
1751 struct perf_event *child;
1754 total += perf_event_read(event);
1755 list_for_each_entry(child, &event->child_list, child_list)
1756 total += perf_event_read(child);
1761 static int perf_event_read_entry(struct perf_event *event,
1762 u64 read_format, char __user *buf)
1764 int n = 0, count = 0;
1767 values[n++] = perf_event_read_value(event);
1768 if (read_format & PERF_FORMAT_ID)
1769 values[n++] = primary_event_id(event);
1771 count = n * sizeof(u64);
1773 if (copy_to_user(buf, values, count))
1779 static int perf_event_read_group(struct perf_event *event,
1780 u64 read_format, char __user *buf)
1782 struct perf_event *leader = event->group_leader, *sub;
1783 int n = 0, size = 0, err = -EFAULT;
1786 values[n++] = 1 + leader->nr_siblings;
1787 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
1788 values[n++] = leader->total_time_enabled +
1789 atomic64_read(&leader->child_total_time_enabled);
1791 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
1792 values[n++] = leader->total_time_running +
1793 atomic64_read(&leader->child_total_time_running);
1796 size = n * sizeof(u64);
1798 if (copy_to_user(buf, values, size))
1801 err = perf_event_read_entry(leader, read_format, buf + size);
1807 list_for_each_entry(sub, &leader->sibling_list, group_entry) {
1808 err = perf_event_read_entry(sub, read_format,
1819 static int perf_event_read_one(struct perf_event *event,
1820 u64 read_format, char __user *buf)
1825 values[n++] = perf_event_read_value(event);
1826 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
1827 values[n++] = event->total_time_enabled +
1828 atomic64_read(&event->child_total_time_enabled);
1830 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
1831 values[n++] = event->total_time_running +
1832 atomic64_read(&event->child_total_time_running);
1834 if (read_format & PERF_FORMAT_ID)
1835 values[n++] = primary_event_id(event);
1837 if (copy_to_user(buf, values, n * sizeof(u64)))
1840 return n * sizeof(u64);
1844 * Read the performance event - simple non blocking version for now
1847 perf_read_hw(struct perf_event *event, char __user *buf, size_t count)
1849 u64 read_format = event->attr.read_format;
1853 * Return end-of-file for a read on a event that is in
1854 * error state (i.e. because it was pinned but it couldn't be
1855 * scheduled on to the CPU at some point).
1857 if (event->state == PERF_EVENT_STATE_ERROR)
1860 if (count < perf_event_read_size(event))
1863 WARN_ON_ONCE(event->ctx->parent_ctx);
1864 mutex_lock(&event->child_mutex);
1865 if (read_format & PERF_FORMAT_GROUP)
1866 ret = perf_event_read_group(event, read_format, buf);
1868 ret = perf_event_read_one(event, read_format, buf);
1869 mutex_unlock(&event->child_mutex);
1875 perf_read(struct file *file, char __user *buf, size_t count, loff_t *ppos)
1877 struct perf_event *event = file->private_data;
1879 return perf_read_hw(event, buf, count);
1882 static unsigned int perf_poll(struct file *file, poll_table *wait)
1884 struct perf_event *event = file->private_data;
1885 struct perf_mmap_data *data;
1886 unsigned int events = POLL_HUP;
1889 data = rcu_dereference(event->data);
1891 events = atomic_xchg(&data->poll, 0);
1894 poll_wait(file, &event->waitq, wait);
1899 static void perf_event_reset(struct perf_event *event)
1901 (void)perf_event_read(event);
1902 atomic64_set(&event->count, 0);
1903 perf_event_update_userpage(event);
1907 * Holding the top-level event's child_mutex means that any
1908 * descendant process that has inherited this event will block
1909 * in sync_child_event if it goes to exit, thus satisfying the
1910 * task existence requirements of perf_event_enable/disable.
1912 static void perf_event_for_each_child(struct perf_event *event,
1913 void (*func)(struct perf_event *))
1915 struct perf_event *child;
1917 WARN_ON_ONCE(event->ctx->parent_ctx);
1918 mutex_lock(&event->child_mutex);
1920 list_for_each_entry(child, &event->child_list, child_list)
1922 mutex_unlock(&event->child_mutex);
1925 static void perf_event_for_each(struct perf_event *event,
1926 void (*func)(struct perf_event *))
1928 struct perf_event_context *ctx = event->ctx;
1929 struct perf_event *sibling;
1931 WARN_ON_ONCE(ctx->parent_ctx);
1932 mutex_lock(&ctx->mutex);
1933 event = event->group_leader;
1935 perf_event_for_each_child(event, func);
1937 list_for_each_entry(sibling, &event->sibling_list, group_entry)
1938 perf_event_for_each_child(event, func);
1939 mutex_unlock(&ctx->mutex);
1942 static int perf_event_period(struct perf_event *event, u64 __user *arg)
1944 struct perf_event_context *ctx = event->ctx;
1949 if (!event->attr.sample_period)
1952 size = copy_from_user(&value, arg, sizeof(value));
1953 if (size != sizeof(value))
1959 spin_lock_irq(&ctx->lock);
1960 if (event->attr.freq) {
1961 if (value > sysctl_perf_event_sample_rate) {
1966 event->attr.sample_freq = value;
1968 event->attr.sample_period = value;
1969 event->hw.sample_period = value;
1972 spin_unlock_irq(&ctx->lock);
1977 int perf_event_set_output(struct perf_event *event, int output_fd);
1979 static long perf_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
1981 struct perf_event *event = file->private_data;
1982 void (*func)(struct perf_event *);
1986 case PERF_EVENT_IOC_ENABLE:
1987 func = perf_event_enable;
1989 case PERF_EVENT_IOC_DISABLE:
1990 func = perf_event_disable;
1992 case PERF_EVENT_IOC_RESET:
1993 func = perf_event_reset;
1996 case PERF_EVENT_IOC_REFRESH:
1997 return perf_event_refresh(event, arg);
1999 case PERF_EVENT_IOC_PERIOD:
2000 return perf_event_period(event, (u64 __user *)arg);
2002 case PERF_EVENT_IOC_SET_OUTPUT:
2003 return perf_event_set_output(event, arg);
2009 if (flags & PERF_IOC_FLAG_GROUP)
2010 perf_event_for_each(event, func);
2012 perf_event_for_each_child(event, func);
2017 int perf_event_task_enable(void)
2019 struct perf_event *event;
2021 mutex_lock(¤t->perf_event_mutex);
2022 list_for_each_entry(event, ¤t->perf_event_list, owner_entry)
2023 perf_event_for_each_child(event, perf_event_enable);
2024 mutex_unlock(¤t->perf_event_mutex);
2029 int perf_event_task_disable(void)
2031 struct perf_event *event;
2033 mutex_lock(¤t->perf_event_mutex);
2034 list_for_each_entry(event, ¤t->perf_event_list, owner_entry)
2035 perf_event_for_each_child(event, perf_event_disable);
2036 mutex_unlock(¤t->perf_event_mutex);
2041 #ifndef PERF_EVENT_INDEX_OFFSET
2042 # define PERF_EVENT_INDEX_OFFSET 0
2045 static int perf_event_index(struct perf_event *event)
2047 if (event->state != PERF_EVENT_STATE_ACTIVE)
2050 return event->hw.idx + 1 - PERF_EVENT_INDEX_OFFSET;
2054 * Callers need to ensure there can be no nesting of this function, otherwise
2055 * the seqlock logic goes bad. We can not serialize this because the arch
2056 * code calls this from NMI context.
2058 void perf_event_update_userpage(struct perf_event *event)
2060 struct perf_event_mmap_page *userpg;
2061 struct perf_mmap_data *data;
2064 data = rcu_dereference(event->data);
2068 userpg = data->user_page;
2071 * Disable preemption so as to not let the corresponding user-space
2072 * spin too long if we get preempted.
2077 userpg->index = perf_event_index(event);
2078 userpg->offset = atomic64_read(&event->count);
2079 if (event->state == PERF_EVENT_STATE_ACTIVE)
2080 userpg->offset -= atomic64_read(&event->hw.prev_count);
2082 userpg->time_enabled = event->total_time_enabled +
2083 atomic64_read(&event->child_total_time_enabled);
2085 userpg->time_running = event->total_time_running +
2086 atomic64_read(&event->child_total_time_running);
2095 static unsigned long perf_data_size(struct perf_mmap_data *data)
2097 return data->nr_pages << (PAGE_SHIFT + data->data_order);
2100 #ifndef CONFIG_PERF_USE_VMALLOC
2103 * Back perf_mmap() with regular GFP_KERNEL-0 pages.
2106 static struct page *
2107 perf_mmap_to_page(struct perf_mmap_data *data, unsigned long pgoff)
2109 if (pgoff > data->nr_pages)
2113 return virt_to_page(data->user_page);
2115 return virt_to_page(data->data_pages[pgoff - 1]);
2118 static struct perf_mmap_data *
2119 perf_mmap_data_alloc(struct perf_event *event, int nr_pages)
2121 struct perf_mmap_data *data;
2125 WARN_ON(atomic_read(&event->mmap_count));
2127 size = sizeof(struct perf_mmap_data);
2128 size += nr_pages * sizeof(void *);
2130 data = kzalloc(size, GFP_KERNEL);
2134 data->user_page = (void *)get_zeroed_page(GFP_KERNEL);
2135 if (!data->user_page)
2136 goto fail_user_page;
2138 for (i = 0; i < nr_pages; i++) {
2139 data->data_pages[i] = (void *)get_zeroed_page(GFP_KERNEL);
2140 if (!data->data_pages[i])
2141 goto fail_data_pages;
2144 data->data_order = 0;
2145 data->nr_pages = nr_pages;
2150 for (i--; i >= 0; i--)
2151 free_page((unsigned long)data->data_pages[i]);
2153 free_page((unsigned long)data->user_page);
2162 static void perf_mmap_free_page(unsigned long addr)
2164 struct page *page = virt_to_page((void *)addr);
2166 page->mapping = NULL;
2170 static void perf_mmap_data_free(struct perf_mmap_data *data)
2174 perf_mmap_free_page((unsigned long)data->user_page);
2175 for (i = 0; i < data->nr_pages; i++)
2176 perf_mmap_free_page((unsigned long)data->data_pages[i]);
2183 * Back perf_mmap() with vmalloc memory.
2185 * Required for architectures that have d-cache aliasing issues.
2188 static struct page *
2189 perf_mmap_to_page(struct perf_mmap_data *data, unsigned long pgoff)
2191 if (pgoff > (1UL << data->data_order))
2194 return vmalloc_to_page((void *)data->user_page + pgoff * PAGE_SIZE);
2197 static void perf_mmap_unmark_page(void *addr)
2199 struct page *page = vmalloc_to_page(addr);
2201 page->mapping = NULL;
2204 static void perf_mmap_data_free_work(struct work_struct *work)
2206 struct perf_mmap_data *data;
2210 data = container_of(work, struct perf_mmap_data, work);
2211 nr = 1 << data->data_order;
2213 base = data->user_page;
2214 for (i = 0; i < nr + 1; i++)
2215 perf_mmap_unmark_page(base + (i * PAGE_SIZE));
2221 static void perf_mmap_data_free(struct perf_mmap_data *data)
2223 schedule_work(&data->work);
2226 static struct perf_mmap_data *
2227 perf_mmap_data_alloc(struct perf_event *event, int nr_pages)
2229 struct perf_mmap_data *data;
2233 WARN_ON(atomic_read(&event->mmap_count));
2235 size = sizeof(struct perf_mmap_data);
2236 size += sizeof(void *);
2238 data = kzalloc(size, GFP_KERNEL);
2242 INIT_WORK(&data->work, perf_mmap_data_free_work);
2244 all_buf = vmalloc_user((nr_pages + 1) * PAGE_SIZE);
2248 data->user_page = all_buf;
2249 data->data_pages[0] = all_buf + PAGE_SIZE;
2250 data->data_order = ilog2(nr_pages);
2264 static int perf_mmap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
2266 struct perf_event *event = vma->vm_file->private_data;
2267 struct perf_mmap_data *data;
2268 int ret = VM_FAULT_SIGBUS;
2270 if (vmf->flags & FAULT_FLAG_MKWRITE) {
2271 if (vmf->pgoff == 0)
2277 data = rcu_dereference(event->data);
2281 if (vmf->pgoff && (vmf->flags & FAULT_FLAG_WRITE))
2284 vmf->page = perf_mmap_to_page(data, vmf->pgoff);
2288 get_page(vmf->page);
2289 vmf->page->mapping = vma->vm_file->f_mapping;
2290 vmf->page->index = vmf->pgoff;
2300 perf_mmap_data_init(struct perf_event *event, struct perf_mmap_data *data)
2302 long max_size = perf_data_size(data);
2304 atomic_set(&data->lock, -1);
2306 if (event->attr.watermark) {
2307 data->watermark = min_t(long, max_size,
2308 event->attr.wakeup_watermark);
2311 if (!data->watermark)
2312 data->watermark = max_t(long, PAGE_SIZE, max_size / 2);
2315 rcu_assign_pointer(event->data, data);
2318 static void perf_mmap_data_free_rcu(struct rcu_head *rcu_head)
2320 struct perf_mmap_data *data;
2322 data = container_of(rcu_head, struct perf_mmap_data, rcu_head);
2323 perf_mmap_data_free(data);
2326 static void perf_mmap_data_release(struct perf_event *event)
2328 struct perf_mmap_data *data = event->data;
2330 WARN_ON(atomic_read(&event->mmap_count));
2332 rcu_assign_pointer(event->data, NULL);
2333 call_rcu(&data->rcu_head, perf_mmap_data_free_rcu);
2336 static void perf_mmap_open(struct vm_area_struct *vma)
2338 struct perf_event *event = vma->vm_file->private_data;
2340 atomic_inc(&event->mmap_count);
2343 static void perf_mmap_close(struct vm_area_struct *vma)
2345 struct perf_event *event = vma->vm_file->private_data;
2347 WARN_ON_ONCE(event->ctx->parent_ctx);
2348 if (atomic_dec_and_mutex_lock(&event->mmap_count, &event->mmap_mutex)) {
2349 unsigned long size = perf_data_size(event->data);
2350 struct user_struct *user = current_user();
2352 atomic_long_sub((size >> PAGE_SHIFT) + 1, &user->locked_vm);
2353 vma->vm_mm->locked_vm -= event->data->nr_locked;
2354 perf_mmap_data_release(event);
2355 mutex_unlock(&event->mmap_mutex);
2359 static const struct vm_operations_struct perf_mmap_vmops = {
2360 .open = perf_mmap_open,
2361 .close = perf_mmap_close,
2362 .fault = perf_mmap_fault,
2363 .page_mkwrite = perf_mmap_fault,
2366 static int perf_mmap(struct file *file, struct vm_area_struct *vma)
2368 struct perf_event *event = file->private_data;
2369 unsigned long user_locked, user_lock_limit;
2370 struct user_struct *user = current_user();
2371 unsigned long locked, lock_limit;
2372 struct perf_mmap_data *data;
2373 unsigned long vma_size;
2374 unsigned long nr_pages;
2375 long user_extra, extra;
2378 if (!(vma->vm_flags & VM_SHARED))
2381 vma_size = vma->vm_end - vma->vm_start;
2382 nr_pages = (vma_size / PAGE_SIZE) - 1;
2385 * If we have data pages ensure they're a power-of-two number, so we
2386 * can do bitmasks instead of modulo.
2388 if (nr_pages != 0 && !is_power_of_2(nr_pages))
2391 if (vma_size != PAGE_SIZE * (1 + nr_pages))
2394 if (vma->vm_pgoff != 0)
2397 WARN_ON_ONCE(event->ctx->parent_ctx);
2398 mutex_lock(&event->mmap_mutex);
2399 if (event->output) {
2404 if (atomic_inc_not_zero(&event->mmap_count)) {
2405 if (nr_pages != event->data->nr_pages)
2410 user_extra = nr_pages + 1;
2411 user_lock_limit = sysctl_perf_event_mlock >> (PAGE_SHIFT - 10);
2414 * Increase the limit linearly with more CPUs:
2416 user_lock_limit *= num_online_cpus();
2418 user_locked = atomic_long_read(&user->locked_vm) + user_extra;
2421 if (user_locked > user_lock_limit)
2422 extra = user_locked - user_lock_limit;
2424 lock_limit = current->signal->rlim[RLIMIT_MEMLOCK].rlim_cur;
2425 lock_limit >>= PAGE_SHIFT;
2426 locked = vma->vm_mm->locked_vm + extra;
2428 if ((locked > lock_limit) && perf_paranoid_tracepoint_raw() &&
2429 !capable(CAP_IPC_LOCK)) {
2434 WARN_ON(event->data);
2436 data = perf_mmap_data_alloc(event, nr_pages);
2442 perf_mmap_data_init(event, data);
2444 atomic_set(&event->mmap_count, 1);
2445 atomic_long_add(user_extra, &user->locked_vm);
2446 vma->vm_mm->locked_vm += extra;
2447 event->data->nr_locked = extra;
2448 if (vma->vm_flags & VM_WRITE)
2449 event->data->writable = 1;
2452 mutex_unlock(&event->mmap_mutex);
2454 vma->vm_flags |= VM_RESERVED;
2455 vma->vm_ops = &perf_mmap_vmops;
2460 static int perf_fasync(int fd, struct file *filp, int on)
2462 struct inode *inode = filp->f_path.dentry->d_inode;
2463 struct perf_event *event = filp->private_data;
2466 mutex_lock(&inode->i_mutex);
2467 retval = fasync_helper(fd, filp, on, &event->fasync);
2468 mutex_unlock(&inode->i_mutex);
2476 static const struct file_operations perf_fops = {
2477 .release = perf_release,
2480 .unlocked_ioctl = perf_ioctl,
2481 .compat_ioctl = perf_ioctl,
2483 .fasync = perf_fasync,
2489 * If there's data, ensure we set the poll() state and publish everything
2490 * to user-space before waking everybody up.
2493 void perf_event_wakeup(struct perf_event *event)
2495 wake_up_all(&event->waitq);
2497 if (event->pending_kill) {
2498 kill_fasync(&event->fasync, SIGIO, event->pending_kill);
2499 event->pending_kill = 0;
2506 * Handle the case where we need to wakeup up from NMI (or rq->lock) context.
2508 * The NMI bit means we cannot possibly take locks. Therefore, maintain a
2509 * single linked list and use cmpxchg() to add entries lockless.
2512 static void perf_pending_event(struct perf_pending_entry *entry)
2514 struct perf_event *event = container_of(entry,
2515 struct perf_event, pending);
2517 if (event->pending_disable) {
2518 event->pending_disable = 0;
2519 __perf_event_disable(event);
2522 if (event->pending_wakeup) {
2523 event->pending_wakeup = 0;
2524 perf_event_wakeup(event);
2528 #define PENDING_TAIL ((struct perf_pending_entry *)-1UL)
2530 static DEFINE_PER_CPU(struct perf_pending_entry *, perf_pending_head) = {
2534 static void perf_pending_queue(struct perf_pending_entry *entry,
2535 void (*func)(struct perf_pending_entry *))
2537 struct perf_pending_entry **head;
2539 if (cmpxchg(&entry->next, NULL, PENDING_TAIL) != NULL)
2544 head = &get_cpu_var(perf_pending_head);
2547 entry->next = *head;
2548 } while (cmpxchg(head, entry->next, entry) != entry->next);
2550 set_perf_event_pending();
2552 put_cpu_var(perf_pending_head);
2555 static int __perf_pending_run(void)
2557 struct perf_pending_entry *list;
2560 list = xchg(&__get_cpu_var(perf_pending_head), PENDING_TAIL);
2561 while (list != PENDING_TAIL) {
2562 void (*func)(struct perf_pending_entry *);
2563 struct perf_pending_entry *entry = list;
2570 * Ensure we observe the unqueue before we issue the wakeup,
2571 * so that we won't be waiting forever.
2572 * -- see perf_not_pending().
2583 static inline int perf_not_pending(struct perf_event *event)
2586 * If we flush on whatever cpu we run, there is a chance we don't
2590 __perf_pending_run();
2594 * Ensure we see the proper queue state before going to sleep
2595 * so that we do not miss the wakeup. -- see perf_pending_handle()
2598 return event->pending.next == NULL;
2601 static void perf_pending_sync(struct perf_event *event)
2603 wait_event(event->waitq, perf_not_pending(event));
2606 void perf_event_do_pending(void)
2608 __perf_pending_run();
2612 * Callchain support -- arch specific
2615 __weak struct perf_callchain_entry *perf_callchain(struct pt_regs *regs)
2623 static bool perf_output_space(struct perf_mmap_data *data, unsigned long tail,
2624 unsigned long offset, unsigned long head)
2628 if (!data->writable)
2631 mask = perf_data_size(data) - 1;
2633 offset = (offset - tail) & mask;
2634 head = (head - tail) & mask;
2636 if ((int)(head - offset) < 0)
2642 static void perf_output_wakeup(struct perf_output_handle *handle)
2644 atomic_set(&handle->data->poll, POLL_IN);
2647 handle->event->pending_wakeup = 1;
2648 perf_pending_queue(&handle->event->pending,
2649 perf_pending_event);
2651 perf_event_wakeup(handle->event);
2655 * Curious locking construct.
2657 * We need to ensure a later event_id doesn't publish a head when a former
2658 * event_id isn't done writing. However since we need to deal with NMIs we
2659 * cannot fully serialize things.
2661 * What we do is serialize between CPUs so we only have to deal with NMI
2662 * nesting on a single CPU.
2664 * We only publish the head (and generate a wakeup) when the outer-most
2665 * event_id completes.
2667 static void perf_output_lock(struct perf_output_handle *handle)
2669 struct perf_mmap_data *data = handle->data;
2674 local_irq_save(handle->flags);
2675 cpu = smp_processor_id();
2677 if (in_nmi() && atomic_read(&data->lock) == cpu)
2680 while (atomic_cmpxchg(&data->lock, -1, cpu) != -1)
2686 static void perf_output_unlock(struct perf_output_handle *handle)
2688 struct perf_mmap_data *data = handle->data;
2692 data->done_head = data->head;
2694 if (!handle->locked)
2699 * The xchg implies a full barrier that ensures all writes are done
2700 * before we publish the new head, matched by a rmb() in userspace when
2701 * reading this position.
2703 while ((head = atomic_long_xchg(&data->done_head, 0)))
2704 data->user_page->data_head = head;
2707 * NMI can happen here, which means we can miss a done_head update.
2710 cpu = atomic_xchg(&data->lock, -1);
2711 WARN_ON_ONCE(cpu != smp_processor_id());
2714 * Therefore we have to validate we did not indeed do so.
2716 if (unlikely(atomic_long_read(&data->done_head))) {
2718 * Since we had it locked, we can lock it again.
2720 while (atomic_cmpxchg(&data->lock, -1, cpu) != -1)
2726 if (atomic_xchg(&data->wakeup, 0))
2727 perf_output_wakeup(handle);
2729 local_irq_restore(handle->flags);
2732 void perf_output_copy(struct perf_output_handle *handle,
2733 const void *buf, unsigned int len)
2735 unsigned int pages_mask;
2736 unsigned long offset;
2740 offset = handle->offset;
2741 pages_mask = handle->data->nr_pages - 1;
2742 pages = handle->data->data_pages;
2745 unsigned long page_offset;
2746 unsigned long page_size;
2749 nr = (offset >> PAGE_SHIFT) & pages_mask;
2750 page_size = 1UL << (handle->data->data_order + PAGE_SHIFT);
2751 page_offset = offset & (page_size - 1);
2752 size = min_t(unsigned int, page_size - page_offset, len);
2754 memcpy(pages[nr] + page_offset, buf, size);
2761 handle->offset = offset;
2764 * Check we didn't copy past our reservation window, taking the
2765 * possible unsigned int wrap into account.
2767 WARN_ON_ONCE(((long)(handle->head - handle->offset)) < 0);
2770 int perf_output_begin(struct perf_output_handle *handle,
2771 struct perf_event *event, unsigned int size,
2772 int nmi, int sample)
2774 struct perf_event *output_event;
2775 struct perf_mmap_data *data;
2776 unsigned long tail, offset, head;
2779 struct perf_event_header header;
2786 * For inherited events we send all the output towards the parent.
2789 event = event->parent;
2791 output_event = rcu_dereference(event->output);
2793 event = output_event;
2795 data = rcu_dereference(event->data);
2799 handle->data = data;
2800 handle->event = event;
2802 handle->sample = sample;
2804 if (!data->nr_pages)
2807 have_lost = atomic_read(&data->lost);
2809 size += sizeof(lost_event);
2811 perf_output_lock(handle);
2815 * Userspace could choose to issue a mb() before updating the
2816 * tail pointer. So that all reads will be completed before the
2819 tail = ACCESS_ONCE(data->user_page->data_tail);
2821 offset = head = atomic_long_read(&data->head);
2823 if (unlikely(!perf_output_space(data, tail, offset, head)))
2825 } while (atomic_long_cmpxchg(&data->head, offset, head) != offset);
2827 handle->offset = offset;
2828 handle->head = head;
2830 if (head - tail > data->watermark)
2831 atomic_set(&data->wakeup, 1);
2834 lost_event.header.type = PERF_RECORD_LOST;
2835 lost_event.header.misc = 0;
2836 lost_event.header.size = sizeof(lost_event);
2837 lost_event.id = event->id;
2838 lost_event.lost = atomic_xchg(&data->lost, 0);
2840 perf_output_put(handle, lost_event);
2846 atomic_inc(&data->lost);
2847 perf_output_unlock(handle);
2854 void perf_output_end(struct perf_output_handle *handle)
2856 struct perf_event *event = handle->event;
2857 struct perf_mmap_data *data = handle->data;
2859 int wakeup_events = event->attr.wakeup_events;
2861 if (handle->sample && wakeup_events) {
2862 int events = atomic_inc_return(&data->events);
2863 if (events >= wakeup_events) {
2864 atomic_sub(wakeup_events, &data->events);
2865 atomic_set(&data->wakeup, 1);
2869 perf_output_unlock(handle);
2873 static u32 perf_event_pid(struct perf_event *event, struct task_struct *p)
2876 * only top level events have the pid namespace they were created in
2879 event = event->parent;
2881 return task_tgid_nr_ns(p, event->ns);
2884 static u32 perf_event_tid(struct perf_event *event, struct task_struct *p)
2887 * only top level events have the pid namespace they were created in
2890 event = event->parent;
2892 return task_pid_nr_ns(p, event->ns);
2895 static void perf_output_read_one(struct perf_output_handle *handle,
2896 struct perf_event *event)
2898 u64 read_format = event->attr.read_format;
2902 values[n++] = atomic64_read(&event->count);
2903 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
2904 values[n++] = event->total_time_enabled +
2905 atomic64_read(&event->child_total_time_enabled);
2907 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
2908 values[n++] = event->total_time_running +
2909 atomic64_read(&event->child_total_time_running);
2911 if (read_format & PERF_FORMAT_ID)
2912 values[n++] = primary_event_id(event);
2914 perf_output_copy(handle, values, n * sizeof(u64));
2918 * XXX PERF_FORMAT_GROUP vs inherited events seems difficult.
2920 static void perf_output_read_group(struct perf_output_handle *handle,
2921 struct perf_event *event)
2923 struct perf_event *leader = event->group_leader, *sub;
2924 u64 read_format = event->attr.read_format;
2928 values[n++] = 1 + leader->nr_siblings;
2930 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
2931 values[n++] = leader->total_time_enabled;
2933 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
2934 values[n++] = leader->total_time_running;
2936 if (leader != event)
2937 leader->pmu->read(leader);
2939 values[n++] = atomic64_read(&leader->count);
2940 if (read_format & PERF_FORMAT_ID)
2941 values[n++] = primary_event_id(leader);
2943 perf_output_copy(handle, values, n * sizeof(u64));
2945 list_for_each_entry(sub, &leader->sibling_list, group_entry) {
2949 sub->pmu->read(sub);
2951 values[n++] = atomic64_read(&sub->count);
2952 if (read_format & PERF_FORMAT_ID)
2953 values[n++] = primary_event_id(sub);
2955 perf_output_copy(handle, values, n * sizeof(u64));
2959 static void perf_output_read(struct perf_output_handle *handle,
2960 struct perf_event *event)
2962 if (event->attr.read_format & PERF_FORMAT_GROUP)
2963 perf_output_read_group(handle, event);
2965 perf_output_read_one(handle, event);
2968 void perf_output_sample(struct perf_output_handle *handle,
2969 struct perf_event_header *header,
2970 struct perf_sample_data *data,
2971 struct perf_event *event)
2973 u64 sample_type = data->type;
2975 perf_output_put(handle, *header);
2977 if (sample_type & PERF_SAMPLE_IP)
2978 perf_output_put(handle, data->ip);
2980 if (sample_type & PERF_SAMPLE_TID)
2981 perf_output_put(handle, data->tid_entry);
2983 if (sample_type & PERF_SAMPLE_TIME)
2984 perf_output_put(handle, data->time);
2986 if (sample_type & PERF_SAMPLE_ADDR)
2987 perf_output_put(handle, data->addr);
2989 if (sample_type & PERF_SAMPLE_ID)
2990 perf_output_put(handle, data->id);
2992 if (sample_type & PERF_SAMPLE_STREAM_ID)
2993 perf_output_put(handle, data->stream_id);
2995 if (sample_type & PERF_SAMPLE_CPU)
2996 perf_output_put(handle, data->cpu_entry);
2998 if (sample_type & PERF_SAMPLE_PERIOD)
2999 perf_output_put(handle, data->period);
3001 if (sample_type & PERF_SAMPLE_READ)
3002 perf_output_read(handle, event);
3004 if (sample_type & PERF_SAMPLE_CALLCHAIN) {
3005 if (data->callchain) {
3008 if (data->callchain)
3009 size += data->callchain->nr;
3011 size *= sizeof(u64);
3013 perf_output_copy(handle, data->callchain, size);
3016 perf_output_put(handle, nr);
3020 if (sample_type & PERF_SAMPLE_RAW) {
3022 perf_output_put(handle, data->raw->size);
3023 perf_output_copy(handle, data->raw->data,
3030 .size = sizeof(u32),
3033 perf_output_put(handle, raw);
3038 void perf_prepare_sample(struct perf_event_header *header,
3039 struct perf_sample_data *data,
3040 struct perf_event *event,
3041 struct pt_regs *regs)
3043 u64 sample_type = event->attr.sample_type;
3045 data->type = sample_type;
3047 header->type = PERF_RECORD_SAMPLE;
3048 header->size = sizeof(*header);
3051 header->misc |= perf_misc_flags(regs);
3053 if (sample_type & PERF_SAMPLE_IP) {
3054 data->ip = perf_instruction_pointer(regs);
3056 header->size += sizeof(data->ip);
3059 if (sample_type & PERF_SAMPLE_TID) {
3060 /* namespace issues */
3061 data->tid_entry.pid = perf_event_pid(event, current);
3062 data->tid_entry.tid = perf_event_tid(event, current);
3064 header->size += sizeof(data->tid_entry);
3067 if (sample_type & PERF_SAMPLE_TIME) {
3068 data->time = perf_clock();
3070 header->size += sizeof(data->time);
3073 if (sample_type & PERF_SAMPLE_ADDR)
3074 header->size += sizeof(data->addr);
3076 if (sample_type & PERF_SAMPLE_ID) {
3077 data->id = primary_event_id(event);
3079 header->size += sizeof(data->id);
3082 if (sample_type & PERF_SAMPLE_STREAM_ID) {
3083 data->stream_id = event->id;
3085 header->size += sizeof(data->stream_id);
3088 if (sample_type & PERF_SAMPLE_CPU) {
3089 data->cpu_entry.cpu = raw_smp_processor_id();
3090 data->cpu_entry.reserved = 0;
3092 header->size += sizeof(data->cpu_entry);
3095 if (sample_type & PERF_SAMPLE_PERIOD)
3096 header->size += sizeof(data->period);
3098 if (sample_type & PERF_SAMPLE_READ)
3099 header->size += perf_event_read_size(event);
3101 if (sample_type & PERF_SAMPLE_CALLCHAIN) {
3104 data->callchain = perf_callchain(regs);
3106 if (data->callchain)
3107 size += data->callchain->nr;
3109 header->size += size * sizeof(u64);
3112 if (sample_type & PERF_SAMPLE_RAW) {
3113 int size = sizeof(u32);
3116 size += data->raw->size;
3118 size += sizeof(u32);
3120 WARN_ON_ONCE(size & (sizeof(u64)-1));
3121 header->size += size;
3125 static void perf_event_output(struct perf_event *event, int nmi,
3126 struct perf_sample_data *data,
3127 struct pt_regs *regs)
3129 struct perf_output_handle handle;
3130 struct perf_event_header header;
3132 perf_prepare_sample(&header, data, event, regs);
3134 if (perf_output_begin(&handle, event, header.size, nmi, 1))
3137 perf_output_sample(&handle, &header, data, event);
3139 perf_output_end(&handle);
3146 struct perf_read_event {
3147 struct perf_event_header header;
3154 perf_event_read_event(struct perf_event *event,
3155 struct task_struct *task)
3157 struct perf_output_handle handle;
3158 struct perf_read_event read_event = {
3160 .type = PERF_RECORD_READ,
3162 .size = sizeof(read_event) + perf_event_read_size(event),
3164 .pid = perf_event_pid(event, task),
3165 .tid = perf_event_tid(event, task),
3169 ret = perf_output_begin(&handle, event, read_event.header.size, 0, 0);
3173 perf_output_put(&handle, read_event);
3174 perf_output_read(&handle, event);
3176 perf_output_end(&handle);
3180 * task tracking -- fork/exit
3182 * enabled by: attr.comm | attr.mmap | attr.task
3185 struct perf_task_event {
3186 struct task_struct *task;
3187 struct perf_event_context *task_ctx;
3190 struct perf_event_header header;
3200 static void perf_event_task_output(struct perf_event *event,
3201 struct perf_task_event *task_event)
3203 struct perf_output_handle handle;
3205 struct task_struct *task = task_event->task;
3208 size = task_event->event_id.header.size;
3209 ret = perf_output_begin(&handle, event, size, 0, 0);
3214 task_event->event_id.pid = perf_event_pid(event, task);
3215 task_event->event_id.ppid = perf_event_pid(event, current);
3217 task_event->event_id.tid = perf_event_tid(event, task);
3218 task_event->event_id.ptid = perf_event_tid(event, current);
3220 task_event->event_id.time = perf_clock();
3222 perf_output_put(&handle, task_event->event_id);
3224 perf_output_end(&handle);
3227 static int perf_event_task_match(struct perf_event *event)
3229 if (event->attr.comm || event->attr.mmap || event->attr.task)
3235 static void perf_event_task_ctx(struct perf_event_context *ctx,
3236 struct perf_task_event *task_event)
3238 struct perf_event *event;
3240 if (system_state != SYSTEM_RUNNING || list_empty(&ctx->event_list))
3244 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
3245 if (perf_event_task_match(event))
3246 perf_event_task_output(event, task_event);
3251 static void perf_event_task_event(struct perf_task_event *task_event)
3253 struct perf_cpu_context *cpuctx;
3254 struct perf_event_context *ctx = task_event->task_ctx;
3256 cpuctx = &get_cpu_var(perf_cpu_context);
3257 perf_event_task_ctx(&cpuctx->ctx, task_event);
3258 put_cpu_var(perf_cpu_context);
3262 ctx = rcu_dereference(task_event->task->perf_event_ctxp);
3264 perf_event_task_ctx(ctx, task_event);
3268 static void perf_event_task(struct task_struct *task,
3269 struct perf_event_context *task_ctx,
3272 struct perf_task_event task_event;
3274 if (!atomic_read(&nr_comm_events) &&
3275 !atomic_read(&nr_mmap_events) &&
3276 !atomic_read(&nr_task_events))
3279 task_event = (struct perf_task_event){
3281 .task_ctx = task_ctx,
3284 .type = new ? PERF_RECORD_FORK : PERF_RECORD_EXIT,
3286 .size = sizeof(task_event.event_id),
3295 perf_event_task_event(&task_event);
3298 void perf_event_fork(struct task_struct *task)
3300 perf_event_task(task, NULL, 1);
3307 struct perf_comm_event {
3308 struct task_struct *task;
3313 struct perf_event_header header;
3320 static void perf_event_comm_output(struct perf_event *event,
3321 struct perf_comm_event *comm_event)
3323 struct perf_output_handle handle;
3324 int size = comm_event->event_id.header.size;
3325 int ret = perf_output_begin(&handle, event, size, 0, 0);
3330 comm_event->event_id.pid = perf_event_pid(event, comm_event->task);
3331 comm_event->event_id.tid = perf_event_tid(event, comm_event->task);
3333 perf_output_put(&handle, comm_event->event_id);
3334 perf_output_copy(&handle, comm_event->comm,
3335 comm_event->comm_size);
3336 perf_output_end(&handle);
3339 static int perf_event_comm_match(struct perf_event *event)
3341 if (event->attr.comm)
3347 static void perf_event_comm_ctx(struct perf_event_context *ctx,
3348 struct perf_comm_event *comm_event)
3350 struct perf_event *event;
3352 if (system_state != SYSTEM_RUNNING || list_empty(&ctx->event_list))
3356 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
3357 if (perf_event_comm_match(event))
3358 perf_event_comm_output(event, comm_event);
3363 static void perf_event_comm_event(struct perf_comm_event *comm_event)
3365 struct perf_cpu_context *cpuctx;
3366 struct perf_event_context *ctx;
3368 char comm[TASK_COMM_LEN];
3370 memset(comm, 0, sizeof(comm));
3371 strncpy(comm, comm_event->task->comm, sizeof(comm));
3372 size = ALIGN(strlen(comm)+1, sizeof(u64));
3374 comm_event->comm = comm;
3375 comm_event->comm_size = size;
3377 comm_event->event_id.header.size = sizeof(comm_event->event_id) + size;
3379 cpuctx = &get_cpu_var(perf_cpu_context);
3380 perf_event_comm_ctx(&cpuctx->ctx, comm_event);
3381 put_cpu_var(perf_cpu_context);
3385 * doesn't really matter which of the child contexts the
3386 * events ends up in.
3388 ctx = rcu_dereference(current->perf_event_ctxp);
3390 perf_event_comm_ctx(ctx, comm_event);
3394 void perf_event_comm(struct task_struct *task)
3396 struct perf_comm_event comm_event;
3398 if (task->perf_event_ctxp)
3399 perf_event_enable_on_exec(task);
3401 if (!atomic_read(&nr_comm_events))
3404 comm_event = (struct perf_comm_event){
3410 .type = PERF_RECORD_COMM,
3419 perf_event_comm_event(&comm_event);
3426 struct perf_mmap_event {
3427 struct vm_area_struct *vma;
3429 const char *file_name;
3433 struct perf_event_header header;
3443 static void perf_event_mmap_output(struct perf_event *event,
3444 struct perf_mmap_event *mmap_event)
3446 struct perf_output_handle handle;
3447 int size = mmap_event->event_id.header.size;
3448 int ret = perf_output_begin(&handle, event, size, 0, 0);
3453 mmap_event->event_id.pid = perf_event_pid(event, current);
3454 mmap_event->event_id.tid = perf_event_tid(event, current);
3456 perf_output_put(&handle, mmap_event->event_id);
3457 perf_output_copy(&handle, mmap_event->file_name,
3458 mmap_event->file_size);
3459 perf_output_end(&handle);
3462 static int perf_event_mmap_match(struct perf_event *event,
3463 struct perf_mmap_event *mmap_event)
3465 if (event->attr.mmap)
3471 static void perf_event_mmap_ctx(struct perf_event_context *ctx,
3472 struct perf_mmap_event *mmap_event)
3474 struct perf_event *event;
3476 if (system_state != SYSTEM_RUNNING || list_empty(&ctx->event_list))
3480 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
3481 if (perf_event_mmap_match(event, mmap_event))
3482 perf_event_mmap_output(event, mmap_event);
3487 static void perf_event_mmap_event(struct perf_mmap_event *mmap_event)
3489 struct perf_cpu_context *cpuctx;
3490 struct perf_event_context *ctx;
3491 struct vm_area_struct *vma = mmap_event->vma;
3492 struct file *file = vma->vm_file;
3498 memset(tmp, 0, sizeof(tmp));
3502 * d_path works from the end of the buffer backwards, so we
3503 * need to add enough zero bytes after the string to handle
3504 * the 64bit alignment we do later.
3506 buf = kzalloc(PATH_MAX + sizeof(u64), GFP_KERNEL);
3508 name = strncpy(tmp, "//enomem", sizeof(tmp));
3511 name = d_path(&file->f_path, buf, PATH_MAX);
3513 name = strncpy(tmp, "//toolong", sizeof(tmp));
3517 if (arch_vma_name(mmap_event->vma)) {
3518 name = strncpy(tmp, arch_vma_name(mmap_event->vma),
3524 name = strncpy(tmp, "[vdso]", sizeof(tmp));
3528 name = strncpy(tmp, "//anon", sizeof(tmp));
3533 size = ALIGN(strlen(name)+1, sizeof(u64));
3535 mmap_event->file_name = name;
3536 mmap_event->file_size = size;
3538 mmap_event->event_id.header.size = sizeof(mmap_event->event_id) + size;
3540 cpuctx = &get_cpu_var(perf_cpu_context);
3541 perf_event_mmap_ctx(&cpuctx->ctx, mmap_event);
3542 put_cpu_var(perf_cpu_context);
3546 * doesn't really matter which of the child contexts the
3547 * events ends up in.
3549 ctx = rcu_dereference(current->perf_event_ctxp);
3551 perf_event_mmap_ctx(ctx, mmap_event);
3557 void __perf_event_mmap(struct vm_area_struct *vma)
3559 struct perf_mmap_event mmap_event;
3561 if (!atomic_read(&nr_mmap_events))
3564 mmap_event = (struct perf_mmap_event){
3570 .type = PERF_RECORD_MMAP,
3576 .start = vma->vm_start,
3577 .len = vma->vm_end - vma->vm_start,
3578 .pgoff = vma->vm_pgoff,
3582 perf_event_mmap_event(&mmap_event);
3586 * IRQ throttle logging
3589 static void perf_log_throttle(struct perf_event *event, int enable)
3591 struct perf_output_handle handle;
3595 struct perf_event_header header;
3599 } throttle_event = {
3601 .type = PERF_RECORD_THROTTLE,
3603 .size = sizeof(throttle_event),
3605 .time = perf_clock(),
3606 .id = primary_event_id(event),
3607 .stream_id = event->id,
3611 throttle_event.header.type = PERF_RECORD_UNTHROTTLE;
3613 ret = perf_output_begin(&handle, event, sizeof(throttle_event), 1, 0);
3617 perf_output_put(&handle, throttle_event);
3618 perf_output_end(&handle);
3622 * Generic event overflow handling, sampling.
3625 static int __perf_event_overflow(struct perf_event *event, int nmi,
3626 int throttle, struct perf_sample_data *data,
3627 struct pt_regs *regs)
3629 int events = atomic_read(&event->event_limit);
3630 struct hw_perf_event *hwc = &event->hw;
3633 throttle = (throttle && event->pmu->unthrottle != NULL);
3638 if (hwc->interrupts != MAX_INTERRUPTS) {
3640 if (HZ * hwc->interrupts >
3641 (u64)sysctl_perf_event_sample_rate) {
3642 hwc->interrupts = MAX_INTERRUPTS;
3643 perf_log_throttle(event, 0);
3648 * Keep re-disabling events even though on the previous
3649 * pass we disabled it - just in case we raced with a
3650 * sched-in and the event got enabled again:
3656 if (event->attr.freq) {
3657 u64 now = perf_clock();
3658 s64 delta = now - hwc->freq_stamp;
3660 hwc->freq_stamp = now;
3662 if (delta > 0 && delta < TICK_NSEC)
3663 perf_adjust_period(event, NSEC_PER_SEC / (int)delta);
3667 * XXX event_limit might not quite work as expected on inherited
3671 event->pending_kill = POLL_IN;
3672 if (events && atomic_dec_and_test(&event->event_limit)) {
3674 event->pending_kill = POLL_HUP;
3676 event->pending_disable = 1;
3677 perf_pending_queue(&event->pending,
3678 perf_pending_event);
3680 perf_event_disable(event);
3683 perf_event_output(event, nmi, data, regs);
3687 int perf_event_overflow(struct perf_event *event, int nmi,
3688 struct perf_sample_data *data,
3689 struct pt_regs *regs)
3691 return __perf_event_overflow(event, nmi, 1, data, regs);
3695 * Generic software event infrastructure
3699 * We directly increment event->count and keep a second value in
3700 * event->hw.period_left to count intervals. This period event
3701 * is kept in the range [-sample_period, 0] so that we can use the
3705 static u64 perf_swevent_set_period(struct perf_event *event)
3707 struct hw_perf_event *hwc = &event->hw;
3708 u64 period = hwc->last_period;
3712 hwc->last_period = hwc->sample_period;
3715 old = val = atomic64_read(&hwc->period_left);
3719 nr = div64_u64(period + val, period);
3720 offset = nr * period;
3722 if (atomic64_cmpxchg(&hwc->period_left, old, val) != old)
3728 static void perf_swevent_overflow(struct perf_event *event,
3729 int nmi, struct perf_sample_data *data,
3730 struct pt_regs *regs)
3732 struct hw_perf_event *hwc = &event->hw;
3736 data->period = event->hw.last_period;
3737 overflow = perf_swevent_set_period(event);
3739 if (hwc->interrupts == MAX_INTERRUPTS)
3742 for (; overflow; overflow--) {
3743 if (__perf_event_overflow(event, nmi, throttle,
3746 * We inhibit the overflow from happening when
3747 * hwc->interrupts == MAX_INTERRUPTS.
3755 static void perf_swevent_unthrottle(struct perf_event *event)
3758 * Nothing to do, we already reset hwc->interrupts.
3762 static void perf_swevent_add(struct perf_event *event, u64 nr,
3763 int nmi, struct perf_sample_data *data,
3764 struct pt_regs *regs)
3766 struct hw_perf_event *hwc = &event->hw;
3768 atomic64_add(nr, &event->count);
3770 if (!hwc->sample_period)
3776 if (!atomic64_add_negative(nr, &hwc->period_left))
3777 perf_swevent_overflow(event, nmi, data, regs);
3780 static int perf_swevent_is_counting(struct perf_event *event)
3783 * The event is active, we're good!
3785 if (event->state == PERF_EVENT_STATE_ACTIVE)
3789 * The event is off/error, not counting.
3791 if (event->state != PERF_EVENT_STATE_INACTIVE)
3795 * The event is inactive, if the context is active
3796 * we're part of a group that didn't make it on the 'pmu',
3799 if (event->ctx->is_active)
3803 * We're inactive and the context is too, this means the
3804 * task is scheduled out, we're counting events that happen
3805 * to us, like migration events.
3810 static int perf_swevent_match(struct perf_event *event,
3811 enum perf_type_id type,
3812 u32 event_id, struct pt_regs *regs)
3814 if (!perf_swevent_is_counting(event))
3817 if (event->attr.type != type)
3819 if (event->attr.config != event_id)
3823 if (event->attr.exclude_user && user_mode(regs))
3826 if (event->attr.exclude_kernel && !user_mode(regs))
3833 static void perf_swevent_ctx_event(struct perf_event_context *ctx,
3834 enum perf_type_id type,
3835 u32 event_id, u64 nr, int nmi,
3836 struct perf_sample_data *data,
3837 struct pt_regs *regs)
3839 struct perf_event *event;
3841 if (system_state != SYSTEM_RUNNING || list_empty(&ctx->event_list))
3845 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
3846 if (perf_swevent_match(event, type, event_id, regs))
3847 perf_swevent_add(event, nr, nmi, data, regs);
3852 static int *perf_swevent_recursion_context(struct perf_cpu_context *cpuctx)
3855 return &cpuctx->recursion[3];
3858 return &cpuctx->recursion[2];
3861 return &cpuctx->recursion[1];
3863 return &cpuctx->recursion[0];
3866 static void do_perf_sw_event(enum perf_type_id type, u32 event_id,
3868 struct perf_sample_data *data,
3869 struct pt_regs *regs)
3871 struct perf_cpu_context *cpuctx = &get_cpu_var(perf_cpu_context);
3872 int *recursion = perf_swevent_recursion_context(cpuctx);
3873 struct perf_event_context *ctx;
3881 perf_swevent_ctx_event(&cpuctx->ctx, type, event_id,
3882 nr, nmi, data, regs);
3885 * doesn't really matter which of the child contexts the
3886 * events ends up in.
3888 ctx = rcu_dereference(current->perf_event_ctxp);
3890 perf_swevent_ctx_event(ctx, type, event_id, nr, nmi, data, regs);
3897 put_cpu_var(perf_cpu_context);
3900 void __perf_sw_event(u32 event_id, u64 nr, int nmi,
3901 struct pt_regs *regs, u64 addr)
3903 struct perf_sample_data data = {
3907 do_perf_sw_event(PERF_TYPE_SOFTWARE, event_id, nr, nmi,
3911 static void perf_swevent_read(struct perf_event *event)
3915 static int perf_swevent_enable(struct perf_event *event)
3917 struct hw_perf_event *hwc = &event->hw;
3919 if (hwc->sample_period) {
3920 hwc->last_period = hwc->sample_period;
3921 perf_swevent_set_period(event);
3926 static void perf_swevent_disable(struct perf_event *event)
3930 static const struct pmu perf_ops_generic = {
3931 .enable = perf_swevent_enable,
3932 .disable = perf_swevent_disable,
3933 .read = perf_swevent_read,
3934 .unthrottle = perf_swevent_unthrottle,
3938 * hrtimer based swevent callback
3941 static enum hrtimer_restart perf_swevent_hrtimer(struct hrtimer *hrtimer)
3943 enum hrtimer_restart ret = HRTIMER_RESTART;
3944 struct perf_sample_data data;
3945 struct pt_regs *regs;
3946 struct perf_event *event;
3949 event = container_of(hrtimer, struct perf_event, hw.hrtimer);
3950 event->pmu->read(event);
3953 data.period = event->hw.last_period;
3954 regs = get_irq_regs();
3956 * In case we exclude kernel IPs or are somehow not in interrupt
3957 * context, provide the next best thing, the user IP.
3959 if ((event->attr.exclude_kernel || !regs) &&
3960 !event->attr.exclude_user)
3961 regs = task_pt_regs(current);
3964 if (!(event->attr.exclude_idle && current->pid == 0))
3965 if (perf_event_overflow(event, 0, &data, regs))
3966 ret = HRTIMER_NORESTART;
3969 period = max_t(u64, 10000, event->hw.sample_period);
3970 hrtimer_forward_now(hrtimer, ns_to_ktime(period));
3975 static void perf_swevent_start_hrtimer(struct perf_event *event)
3977 struct hw_perf_event *hwc = &event->hw;
3979 hrtimer_init(&hwc->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
3980 hwc->hrtimer.function = perf_swevent_hrtimer;
3981 if (hwc->sample_period) {
3984 if (hwc->remaining) {
3985 if (hwc->remaining < 0)
3988 period = hwc->remaining;
3991 period = max_t(u64, 10000, hwc->sample_period);
3993 __hrtimer_start_range_ns(&hwc->hrtimer,
3994 ns_to_ktime(period), 0,
3995 HRTIMER_MODE_REL, 0);
3999 static void perf_swevent_cancel_hrtimer(struct perf_event *event)
4001 struct hw_perf_event *hwc = &event->hw;
4003 if (hwc->sample_period) {
4004 ktime_t remaining = hrtimer_get_remaining(&hwc->hrtimer);
4005 hwc->remaining = ktime_to_ns(remaining);
4007 hrtimer_cancel(&hwc->hrtimer);
4012 * Software event: cpu wall time clock
4015 static void cpu_clock_perf_event_update(struct perf_event *event)
4017 int cpu = raw_smp_processor_id();
4021 now = cpu_clock(cpu);
4022 prev = atomic64_read(&event->hw.prev_count);
4023 atomic64_set(&event->hw.prev_count, now);
4024 atomic64_add(now - prev, &event->count);
4027 static int cpu_clock_perf_event_enable(struct perf_event *event)
4029 struct hw_perf_event *hwc = &event->hw;
4030 int cpu = raw_smp_processor_id();
4032 atomic64_set(&hwc->prev_count, cpu_clock(cpu));
4033 perf_swevent_start_hrtimer(event);
4038 static void cpu_clock_perf_event_disable(struct perf_event *event)
4040 perf_swevent_cancel_hrtimer(event);
4041 cpu_clock_perf_event_update(event);
4044 static void cpu_clock_perf_event_read(struct perf_event *event)
4046 cpu_clock_perf_event_update(event);
4049 static const struct pmu perf_ops_cpu_clock = {
4050 .enable = cpu_clock_perf_event_enable,
4051 .disable = cpu_clock_perf_event_disable,
4052 .read = cpu_clock_perf_event_read,
4056 * Software event: task time clock
4059 static void task_clock_perf_event_update(struct perf_event *event, u64 now)
4064 prev = atomic64_xchg(&event->hw.prev_count, now);
4066 atomic64_add(delta, &event->count);
4069 static int task_clock_perf_event_enable(struct perf_event *event)
4071 struct hw_perf_event *hwc = &event->hw;
4074 now = event->ctx->time;
4076 atomic64_set(&hwc->prev_count, now);
4078 perf_swevent_start_hrtimer(event);
4083 static void task_clock_perf_event_disable(struct perf_event *event)
4085 perf_swevent_cancel_hrtimer(event);
4086 task_clock_perf_event_update(event, event->ctx->time);
4090 static void task_clock_perf_event_read(struct perf_event *event)
4095 update_context_time(event->ctx);
4096 time = event->ctx->time;
4098 u64 now = perf_clock();
4099 u64 delta = now - event->ctx->timestamp;
4100 time = event->ctx->time + delta;
4103 task_clock_perf_event_update(event, time);
4106 static const struct pmu perf_ops_task_clock = {
4107 .enable = task_clock_perf_event_enable,
4108 .disable = task_clock_perf_event_disable,
4109 .read = task_clock_perf_event_read,
4112 #ifdef CONFIG_EVENT_PROFILE
4113 void perf_tp_event(int event_id, u64 addr, u64 count, void *record,
4116 struct perf_raw_record raw = {
4121 struct perf_sample_data data = {
4126 struct pt_regs *regs = get_irq_regs();
4129 regs = task_pt_regs(current);
4131 do_perf_sw_event(PERF_TYPE_TRACEPOINT, event_id, count, 1,
4134 EXPORT_SYMBOL_GPL(perf_tp_event);
4136 extern int ftrace_profile_enable(int);
4137 extern void ftrace_profile_disable(int);
4139 static void tp_perf_event_destroy(struct perf_event *event)
4141 ftrace_profile_disable(event->attr.config);
4144 static const struct pmu *tp_perf_event_init(struct perf_event *event)
4147 * Raw tracepoint data is a severe data leak, only allow root to
4150 if ((event->attr.sample_type & PERF_SAMPLE_RAW) &&
4151 perf_paranoid_tracepoint_raw() &&
4152 !capable(CAP_SYS_ADMIN))
4153 return ERR_PTR(-EPERM);
4155 if (ftrace_profile_enable(event->attr.config))
4158 event->destroy = tp_perf_event_destroy;
4160 return &perf_ops_generic;
4163 static const struct pmu *tp_perf_event_init(struct perf_event *event)
4169 atomic_t perf_swevent_enabled[PERF_COUNT_SW_MAX];
4171 static void sw_perf_event_destroy(struct perf_event *event)
4173 u64 event_id = event->attr.config;
4175 WARN_ON(event->parent);
4177 atomic_dec(&perf_swevent_enabled[event_id]);
4180 static const struct pmu *sw_perf_event_init(struct perf_event *event)
4182 const struct pmu *pmu = NULL;
4183 u64 event_id = event->attr.config;
4186 * Software events (currently) can't in general distinguish
4187 * between user, kernel and hypervisor events.
4188 * However, context switches and cpu migrations are considered
4189 * to be kernel events, and page faults are never hypervisor
4193 case PERF_COUNT_SW_CPU_CLOCK:
4194 pmu = &perf_ops_cpu_clock;
4197 case PERF_COUNT_SW_TASK_CLOCK:
4199 * If the user instantiates this as a per-cpu event,
4200 * use the cpu_clock event instead.
4202 if (event->ctx->task)
4203 pmu = &perf_ops_task_clock;
4205 pmu = &perf_ops_cpu_clock;
4208 case PERF_COUNT_SW_PAGE_FAULTS:
4209 case PERF_COUNT_SW_PAGE_FAULTS_MIN:
4210 case PERF_COUNT_SW_PAGE_FAULTS_MAJ:
4211 case PERF_COUNT_SW_CONTEXT_SWITCHES:
4212 case PERF_COUNT_SW_CPU_MIGRATIONS:
4213 if (!event->parent) {
4214 atomic_inc(&perf_swevent_enabled[event_id]);
4215 event->destroy = sw_perf_event_destroy;
4217 pmu = &perf_ops_generic;
4225 * Allocate and initialize a event structure
4227 static struct perf_event *
4228 perf_event_alloc(struct perf_event_attr *attr,
4230 struct perf_event_context *ctx,
4231 struct perf_event *group_leader,
4232 struct perf_event *parent_event,
4235 const struct pmu *pmu;
4236 struct perf_event *event;
4237 struct hw_perf_event *hwc;
4240 event = kzalloc(sizeof(*event), gfpflags);
4242 return ERR_PTR(-ENOMEM);
4245 * Single events are their own group leaders, with an
4246 * empty sibling list:
4249 group_leader = event;
4251 mutex_init(&event->child_mutex);
4252 INIT_LIST_HEAD(&event->child_list);
4254 INIT_LIST_HEAD(&event->group_entry);
4255 INIT_LIST_HEAD(&event->event_entry);
4256 INIT_LIST_HEAD(&event->sibling_list);
4257 init_waitqueue_head(&event->waitq);
4259 mutex_init(&event->mmap_mutex);
4262 event->attr = *attr;
4263 event->group_leader = group_leader;
4268 event->parent = parent_event;
4270 event->ns = get_pid_ns(current->nsproxy->pid_ns);
4271 event->id = atomic64_inc_return(&perf_event_id);
4273 event->state = PERF_EVENT_STATE_INACTIVE;
4276 event->state = PERF_EVENT_STATE_OFF;
4281 hwc->sample_period = attr->sample_period;
4282 if (attr->freq && attr->sample_freq)
4283 hwc->sample_period = 1;
4284 hwc->last_period = hwc->sample_period;
4286 atomic64_set(&hwc->period_left, hwc->sample_period);
4289 * we currently do not support PERF_FORMAT_GROUP on inherited events
4291 if (attr->inherit && (attr->read_format & PERF_FORMAT_GROUP))
4294 switch (attr->type) {
4296 case PERF_TYPE_HARDWARE:
4297 case PERF_TYPE_HW_CACHE:
4298 pmu = hw_perf_event_init(event);
4301 case PERF_TYPE_SOFTWARE:
4302 pmu = sw_perf_event_init(event);
4305 case PERF_TYPE_TRACEPOINT:
4306 pmu = tp_perf_event_init(event);
4316 else if (IS_ERR(pmu))
4321 put_pid_ns(event->ns);
4323 return ERR_PTR(err);
4328 if (!event->parent) {
4329 atomic_inc(&nr_events);
4330 if (event->attr.mmap)
4331 atomic_inc(&nr_mmap_events);
4332 if (event->attr.comm)
4333 atomic_inc(&nr_comm_events);
4334 if (event->attr.task)
4335 atomic_inc(&nr_task_events);
4341 static int perf_copy_attr(struct perf_event_attr __user *uattr,
4342 struct perf_event_attr *attr)
4347 if (!access_ok(VERIFY_WRITE, uattr, PERF_ATTR_SIZE_VER0))
4351 * zero the full structure, so that a short copy will be nice.
4353 memset(attr, 0, sizeof(*attr));
4355 ret = get_user(size, &uattr->size);
4359 if (size > PAGE_SIZE) /* silly large */
4362 if (!size) /* abi compat */
4363 size = PERF_ATTR_SIZE_VER0;
4365 if (size < PERF_ATTR_SIZE_VER0)
4369 * If we're handed a bigger struct than we know of,
4370 * ensure all the unknown bits are 0 - i.e. new
4371 * user-space does not rely on any kernel feature
4372 * extensions we dont know about yet.
4374 if (size > sizeof(*attr)) {
4375 unsigned char __user *addr;
4376 unsigned char __user *end;
4379 addr = (void __user *)uattr + sizeof(*attr);
4380 end = (void __user *)uattr + size;
4382 for (; addr < end; addr++) {
4383 ret = get_user(val, addr);
4389 size = sizeof(*attr);
4392 ret = copy_from_user(attr, uattr, size);
4397 * If the type exists, the corresponding creation will verify
4400 if (attr->type >= PERF_TYPE_MAX)
4403 if (attr->__reserved_1 || attr->__reserved_2 || attr->__reserved_3)
4406 if (attr->sample_type & ~(PERF_SAMPLE_MAX-1))
4409 if (attr->read_format & ~(PERF_FORMAT_MAX-1))
4416 put_user(sizeof(*attr), &uattr->size);
4421 int perf_event_set_output(struct perf_event *event, int output_fd)
4423 struct perf_event *output_event = NULL;
4424 struct file *output_file = NULL;
4425 struct perf_event *old_output;
4426 int fput_needed = 0;
4432 output_file = fget_light(output_fd, &fput_needed);
4436 if (output_file->f_op != &perf_fops)
4439 output_event = output_file->private_data;
4441 /* Don't chain output fds */
4442 if (output_event->output)
4445 /* Don't set an output fd when we already have an output channel */
4449 atomic_long_inc(&output_file->f_count);
4452 mutex_lock(&event->mmap_mutex);
4453 old_output = event->output;
4454 rcu_assign_pointer(event->output, output_event);
4455 mutex_unlock(&event->mmap_mutex);
4459 * we need to make sure no existing perf_output_*()
4460 * is still referencing this event.
4463 fput(old_output->filp);
4468 fput_light(output_file, fput_needed);
4473 * sys_perf_event_open - open a performance event, associate it to a task/cpu
4475 * @attr_uptr: event_id type attributes for monitoring/sampling
4478 * @group_fd: group leader event fd
4480 SYSCALL_DEFINE5(perf_event_open,
4481 struct perf_event_attr __user *, attr_uptr,
4482 pid_t, pid, int, cpu, int, group_fd, unsigned long, flags)
4484 struct perf_event *event, *group_leader;
4485 struct perf_event_attr attr;
4486 struct perf_event_context *ctx;
4487 struct file *event_file = NULL;
4488 struct file *group_file = NULL;
4489 int fput_needed = 0;
4490 int fput_needed2 = 0;
4493 /* for future expandability... */
4494 if (flags & ~(PERF_FLAG_FD_NO_GROUP | PERF_FLAG_FD_OUTPUT))
4497 err = perf_copy_attr(attr_uptr, &attr);
4501 if (!attr.exclude_kernel) {
4502 if (perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN))
4507 if (attr.sample_freq > sysctl_perf_event_sample_rate)
4512 * Get the target context (task or percpu):
4514 ctx = find_get_context(pid, cpu);
4516 return PTR_ERR(ctx);
4519 * Look up the group leader (we will attach this event to it):
4521 group_leader = NULL;
4522 if (group_fd != -1 && !(flags & PERF_FLAG_FD_NO_GROUP)) {
4524 group_file = fget_light(group_fd, &fput_needed);
4526 goto err_put_context;
4527 if (group_file->f_op != &perf_fops)
4528 goto err_put_context;
4530 group_leader = group_file->private_data;
4532 * Do not allow a recursive hierarchy (this new sibling
4533 * becoming part of another group-sibling):
4535 if (group_leader->group_leader != group_leader)
4536 goto err_put_context;
4538 * Do not allow to attach to a group in a different
4539 * task or CPU context:
4541 if (group_leader->ctx != ctx)
4542 goto err_put_context;
4544 * Only a group leader can be exclusive or pinned
4546 if (attr.exclusive || attr.pinned)
4547 goto err_put_context;
4550 event = perf_event_alloc(&attr, cpu, ctx, group_leader,
4552 err = PTR_ERR(event);
4554 goto err_put_context;
4556 err = anon_inode_getfd("[perf_event]", &perf_fops, event, 0);
4558 goto err_free_put_context;
4560 event_file = fget_light(err, &fput_needed2);
4562 goto err_free_put_context;
4564 if (flags & PERF_FLAG_FD_OUTPUT) {
4565 err = perf_event_set_output(event, group_fd);
4567 goto err_fput_free_put_context;
4570 event->filp = event_file;
4571 WARN_ON_ONCE(ctx->parent_ctx);
4572 mutex_lock(&ctx->mutex);
4573 perf_install_in_context(ctx, event, cpu);
4575 mutex_unlock(&ctx->mutex);
4577 event->owner = current;
4578 get_task_struct(current);
4579 mutex_lock(¤t->perf_event_mutex);
4580 list_add_tail(&event->owner_entry, ¤t->perf_event_list);
4581 mutex_unlock(¤t->perf_event_mutex);
4583 err_fput_free_put_context:
4584 fput_light(event_file, fput_needed2);
4586 err_free_put_context:
4594 fput_light(group_file, fput_needed);
4600 * inherit a event from parent task to child task:
4602 static struct perf_event *
4603 inherit_event(struct perf_event *parent_event,
4604 struct task_struct *parent,
4605 struct perf_event_context *parent_ctx,
4606 struct task_struct *child,
4607 struct perf_event *group_leader,
4608 struct perf_event_context *child_ctx)
4610 struct perf_event *child_event;
4613 * Instead of creating recursive hierarchies of events,
4614 * we link inherited events back to the original parent,
4615 * which has a filp for sure, which we use as the reference
4618 if (parent_event->parent)
4619 parent_event = parent_event->parent;
4621 child_event = perf_event_alloc(&parent_event->attr,
4622 parent_event->cpu, child_ctx,
4623 group_leader, parent_event,
4625 if (IS_ERR(child_event))
4630 * Make the child state follow the state of the parent event,
4631 * not its attr.disabled bit. We hold the parent's mutex,
4632 * so we won't race with perf_event_{en, dis}able_family.
4634 if (parent_event->state >= PERF_EVENT_STATE_INACTIVE)
4635 child_event->state = PERF_EVENT_STATE_INACTIVE;
4637 child_event->state = PERF_EVENT_STATE_OFF;
4639 if (parent_event->attr.freq)
4640 child_event->hw.sample_period = parent_event->hw.sample_period;
4643 * Link it up in the child's context:
4645 add_event_to_ctx(child_event, child_ctx);
4648 * Get a reference to the parent filp - we will fput it
4649 * when the child event exits. This is safe to do because
4650 * we are in the parent and we know that the filp still
4651 * exists and has a nonzero count:
4653 atomic_long_inc(&parent_event->filp->f_count);
4656 * Link this into the parent event's child list
4658 WARN_ON_ONCE(parent_event->ctx->parent_ctx);
4659 mutex_lock(&parent_event->child_mutex);
4660 list_add_tail(&child_event->child_list, &parent_event->child_list);
4661 mutex_unlock(&parent_event->child_mutex);
4666 static int inherit_group(struct perf_event *parent_event,
4667 struct task_struct *parent,
4668 struct perf_event_context *parent_ctx,
4669 struct task_struct *child,
4670 struct perf_event_context *child_ctx)
4672 struct perf_event *leader;
4673 struct perf_event *sub;
4674 struct perf_event *child_ctr;
4676 leader = inherit_event(parent_event, parent, parent_ctx,
4677 child, NULL, child_ctx);
4679 return PTR_ERR(leader);
4680 list_for_each_entry(sub, &parent_event->sibling_list, group_entry) {
4681 child_ctr = inherit_event(sub, parent, parent_ctx,
4682 child, leader, child_ctx);
4683 if (IS_ERR(child_ctr))
4684 return PTR_ERR(child_ctr);
4689 static void sync_child_event(struct perf_event *child_event,
4690 struct task_struct *child)
4692 struct perf_event *parent_event = child_event->parent;
4695 if (child_event->attr.inherit_stat)
4696 perf_event_read_event(child_event, child);
4698 child_val = atomic64_read(&child_event->count);
4701 * Add back the child's count to the parent's count:
4703 atomic64_add(child_val, &parent_event->count);
4704 atomic64_add(child_event->total_time_enabled,
4705 &parent_event->child_total_time_enabled);
4706 atomic64_add(child_event->total_time_running,
4707 &parent_event->child_total_time_running);
4710 * Remove this event from the parent's list
4712 WARN_ON_ONCE(parent_event->ctx->parent_ctx);
4713 mutex_lock(&parent_event->child_mutex);
4714 list_del_init(&child_event->child_list);
4715 mutex_unlock(&parent_event->child_mutex);
4718 * Release the parent event, if this was the last
4721 fput(parent_event->filp);
4725 __perf_event_exit_task(struct perf_event *child_event,
4726 struct perf_event_context *child_ctx,
4727 struct task_struct *child)
4729 struct perf_event *parent_event;
4731 update_event_times(child_event);
4732 perf_event_remove_from_context(child_event);
4734 parent_event = child_event->parent;
4736 * It can happen that parent exits first, and has events
4737 * that are still around due to the child reference. These
4738 * events need to be zapped - but otherwise linger.
4741 sync_child_event(child_event, child);
4742 free_event(child_event);
4747 * When a child task exits, feed back event values to parent events.
4749 void perf_event_exit_task(struct task_struct *child)
4751 struct perf_event *child_event, *tmp;
4752 struct perf_event_context *child_ctx;
4753 unsigned long flags;
4755 if (likely(!child->perf_event_ctxp)) {
4756 perf_event_task(child, NULL, 0);
4760 local_irq_save(flags);
4762 * We can't reschedule here because interrupts are disabled,
4763 * and either child is current or it is a task that can't be
4764 * scheduled, so we are now safe from rescheduling changing
4767 child_ctx = child->perf_event_ctxp;
4768 __perf_event_task_sched_out(child_ctx);
4771 * Take the context lock here so that if find_get_context is
4772 * reading child->perf_event_ctxp, we wait until it has
4773 * incremented the context's refcount before we do put_ctx below.
4775 spin_lock(&child_ctx->lock);
4776 child->perf_event_ctxp = NULL;
4778 * If this context is a clone; unclone it so it can't get
4779 * swapped to another process while we're removing all
4780 * the events from it.
4782 unclone_ctx(child_ctx);
4783 spin_unlock_irqrestore(&child_ctx->lock, flags);
4786 * Report the task dead after unscheduling the events so that we
4787 * won't get any samples after PERF_RECORD_EXIT. We can however still
4788 * get a few PERF_RECORD_READ events.
4790 perf_event_task(child, child_ctx, 0);
4793 * We can recurse on the same lock type through:
4795 * __perf_event_exit_task()
4796 * sync_child_event()
4797 * fput(parent_event->filp)
4799 * mutex_lock(&ctx->mutex)
4801 * But since its the parent context it won't be the same instance.
4803 mutex_lock_nested(&child_ctx->mutex, SINGLE_DEPTH_NESTING);
4806 list_for_each_entry_safe(child_event, tmp, &child_ctx->group_list,
4808 __perf_event_exit_task(child_event, child_ctx, child);
4811 * If the last event was a group event, it will have appended all
4812 * its siblings to the list, but we obtained 'tmp' before that which
4813 * will still point to the list head terminating the iteration.
4815 if (!list_empty(&child_ctx->group_list))
4818 mutex_unlock(&child_ctx->mutex);
4824 * free an unexposed, unused context as created by inheritance by
4825 * init_task below, used by fork() in case of fail.
4827 void perf_event_free_task(struct task_struct *task)
4829 struct perf_event_context *ctx = task->perf_event_ctxp;
4830 struct perf_event *event, *tmp;
4835 mutex_lock(&ctx->mutex);
4837 list_for_each_entry_safe(event, tmp, &ctx->group_list, group_entry) {
4838 struct perf_event *parent = event->parent;
4840 if (WARN_ON_ONCE(!parent))
4843 mutex_lock(&parent->child_mutex);
4844 list_del_init(&event->child_list);
4845 mutex_unlock(&parent->child_mutex);
4849 list_del_event(event, ctx);
4853 if (!list_empty(&ctx->group_list))
4856 mutex_unlock(&ctx->mutex);
4862 * Initialize the perf_event context in task_struct
4864 int perf_event_init_task(struct task_struct *child)
4866 struct perf_event_context *child_ctx, *parent_ctx;
4867 struct perf_event_context *cloned_ctx;
4868 struct perf_event *event;
4869 struct task_struct *parent = current;
4870 int inherited_all = 1;
4873 child->perf_event_ctxp = NULL;
4875 mutex_init(&child->perf_event_mutex);
4876 INIT_LIST_HEAD(&child->perf_event_list);
4878 if (likely(!parent->perf_event_ctxp))
4882 * This is executed from the parent task context, so inherit
4883 * events that have been marked for cloning.
4884 * First allocate and initialize a context for the child.
4887 child_ctx = kmalloc(sizeof(struct perf_event_context), GFP_KERNEL);
4891 __perf_event_init_context(child_ctx, child);
4892 child->perf_event_ctxp = child_ctx;
4893 get_task_struct(child);
4896 * If the parent's context is a clone, pin it so it won't get
4899 parent_ctx = perf_pin_task_context(parent);
4902 * No need to check if parent_ctx != NULL here; since we saw
4903 * it non-NULL earlier, the only reason for it to become NULL
4904 * is if we exit, and since we're currently in the middle of
4905 * a fork we can't be exiting at the same time.
4909 * Lock the parent list. No need to lock the child - not PID
4910 * hashed yet and not running, so nobody can access it.
4912 mutex_lock(&parent_ctx->mutex);
4915 * We dont have to disable NMIs - we are only looking at
4916 * the list, not manipulating it:
4918 list_for_each_entry(event, &parent_ctx->group_list, group_entry) {
4920 if (!event->attr.inherit) {
4925 ret = inherit_group(event, parent, parent_ctx,
4933 if (inherited_all) {
4935 * Mark the child context as a clone of the parent
4936 * context, or of whatever the parent is a clone of.
4937 * Note that if the parent is a clone, it could get
4938 * uncloned at any point, but that doesn't matter
4939 * because the list of events and the generation
4940 * count can't have changed since we took the mutex.
4942 cloned_ctx = rcu_dereference(parent_ctx->parent_ctx);
4944 child_ctx->parent_ctx = cloned_ctx;
4945 child_ctx->parent_gen = parent_ctx->parent_gen;
4947 child_ctx->parent_ctx = parent_ctx;
4948 child_ctx->parent_gen = parent_ctx->generation;
4950 get_ctx(child_ctx->parent_ctx);
4953 mutex_unlock(&parent_ctx->mutex);
4955 perf_unpin_context(parent_ctx);
4960 static void __cpuinit perf_event_init_cpu(int cpu)
4962 struct perf_cpu_context *cpuctx;
4964 cpuctx = &per_cpu(perf_cpu_context, cpu);
4965 __perf_event_init_context(&cpuctx->ctx, NULL);
4967 spin_lock(&perf_resource_lock);
4968 cpuctx->max_pertask = perf_max_events - perf_reserved_percpu;
4969 spin_unlock(&perf_resource_lock);
4971 hw_perf_event_setup(cpu);
4974 #ifdef CONFIG_HOTPLUG_CPU
4975 static void __perf_event_exit_cpu(void *info)
4977 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
4978 struct perf_event_context *ctx = &cpuctx->ctx;
4979 struct perf_event *event, *tmp;
4981 list_for_each_entry_safe(event, tmp, &ctx->group_list, group_entry)
4982 __perf_event_remove_from_context(event);
4984 static void perf_event_exit_cpu(int cpu)
4986 struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
4987 struct perf_event_context *ctx = &cpuctx->ctx;
4989 mutex_lock(&ctx->mutex);
4990 smp_call_function_single(cpu, __perf_event_exit_cpu, NULL, 1);
4991 mutex_unlock(&ctx->mutex);
4994 static inline void perf_event_exit_cpu(int cpu) { }
4997 static int __cpuinit
4998 perf_cpu_notify(struct notifier_block *self, unsigned long action, void *hcpu)
5000 unsigned int cpu = (long)hcpu;
5004 case CPU_UP_PREPARE:
5005 case CPU_UP_PREPARE_FROZEN:
5006 perf_event_init_cpu(cpu);
5010 case CPU_ONLINE_FROZEN:
5011 hw_perf_event_setup_online(cpu);
5014 case CPU_DOWN_PREPARE:
5015 case CPU_DOWN_PREPARE_FROZEN:
5016 perf_event_exit_cpu(cpu);
5027 * This has to have a higher priority than migration_notifier in sched.c.
5029 static struct notifier_block __cpuinitdata perf_cpu_nb = {
5030 .notifier_call = perf_cpu_notify,
5034 void __init perf_event_init(void)
5036 perf_cpu_notify(&perf_cpu_nb, (unsigned long)CPU_UP_PREPARE,
5037 (void *)(long)smp_processor_id());
5038 perf_cpu_notify(&perf_cpu_nb, (unsigned long)CPU_ONLINE,
5039 (void *)(long)smp_processor_id());
5040 register_cpu_notifier(&perf_cpu_nb);
5043 static ssize_t perf_show_reserve_percpu(struct sysdev_class *class, char *buf)
5045 return sprintf(buf, "%d\n", perf_reserved_percpu);
5049 perf_set_reserve_percpu(struct sysdev_class *class,
5053 struct perf_cpu_context *cpuctx;
5057 err = strict_strtoul(buf, 10, &val);
5060 if (val > perf_max_events)
5063 spin_lock(&perf_resource_lock);
5064 perf_reserved_percpu = val;
5065 for_each_online_cpu(cpu) {
5066 cpuctx = &per_cpu(perf_cpu_context, cpu);
5067 spin_lock_irq(&cpuctx->ctx.lock);
5068 mpt = min(perf_max_events - cpuctx->ctx.nr_events,
5069 perf_max_events - perf_reserved_percpu);
5070 cpuctx->max_pertask = mpt;
5071 spin_unlock_irq(&cpuctx->ctx.lock);
5073 spin_unlock(&perf_resource_lock);
5078 static ssize_t perf_show_overcommit(struct sysdev_class *class, char *buf)
5080 return sprintf(buf, "%d\n", perf_overcommit);
5084 perf_set_overcommit(struct sysdev_class *class, const char *buf, size_t count)
5089 err = strict_strtoul(buf, 10, &val);
5095 spin_lock(&perf_resource_lock);
5096 perf_overcommit = val;
5097 spin_unlock(&perf_resource_lock);
5102 static SYSDEV_CLASS_ATTR(
5105 perf_show_reserve_percpu,
5106 perf_set_reserve_percpu
5109 static SYSDEV_CLASS_ATTR(
5112 perf_show_overcommit,
5116 static struct attribute *perfclass_attrs[] = {
5117 &attr_reserve_percpu.attr,
5118 &attr_overcommit.attr,
5122 static struct attribute_group perfclass_attr_group = {
5123 .attrs = perfclass_attrs,
5124 .name = "perf_events",
5127 static int __init perf_event_sysfs_init(void)
5129 return sysfs_create_group(&cpu_sysdev_class.kset.kobj,
5130 &perfclass_attr_group);
5132 device_initcall(perf_event_sysfs_init);