2 * Performance events core code:
4 * Copyright (C) 2008 Thomas Gleixner <tglx@linutronix.de>
5 * Copyright (C) 2008-2009 Red Hat, Inc., Ingo Molnar
6 * Copyright (C) 2008-2009 Red Hat, Inc., Peter Zijlstra <pzijlstr@redhat.com>
7 * Copyright © 2009 Paul Mackerras, IBM Corp. <paulus@au1.ibm.com>
9 * For licensing details see kernel-base/COPYING
14 #include <linux/cpu.h>
15 #include <linux/smp.h>
16 #include <linux/file.h>
17 #include <linux/poll.h>
18 #include <linux/slab.h>
19 #include <linux/hash.h>
20 #include <linux/sysfs.h>
21 #include <linux/dcache.h>
22 #include <linux/percpu.h>
23 #include <linux/ptrace.h>
24 #include <linux/vmstat.h>
25 #include <linux/vmalloc.h>
26 #include <linux/hardirq.h>
27 #include <linux/rculist.h>
28 #include <linux/uaccess.h>
29 #include <linux/syscalls.h>
30 #include <linux/anon_inodes.h>
31 #include <linux/kernel_stat.h>
32 #include <linux/perf_event.h>
33 #include <linux/ftrace_event.h>
34 #include <linux/hw_breakpoint.h>
36 #include <asm/irq_regs.h>
38 atomic_t perf_task_events __read_mostly;
39 static atomic_t nr_mmap_events __read_mostly;
40 static atomic_t nr_comm_events __read_mostly;
41 static atomic_t nr_task_events __read_mostly;
43 static LIST_HEAD(pmus);
44 static DEFINE_MUTEX(pmus_lock);
45 static struct srcu_struct pmus_srcu;
48 * perf event paranoia level:
49 * -1 - not paranoid at all
50 * 0 - disallow raw tracepoint access for unpriv
51 * 1 - disallow cpu events for unpriv
52 * 2 - disallow kernel profiling for unpriv
54 int sysctl_perf_event_paranoid __read_mostly = 1;
56 int sysctl_perf_event_mlock __read_mostly = 512; /* 'free' kb per user */
59 * max perf event sample rate
61 int sysctl_perf_event_sample_rate __read_mostly = 100000;
63 static atomic64_t perf_event_id;
65 void __weak perf_event_print_debug(void) { }
67 extern __weak const char *perf_pmu_name(void)
72 void perf_pmu_disable(struct pmu *pmu)
74 int *count = this_cpu_ptr(pmu->pmu_disable_count);
76 pmu->pmu_disable(pmu);
79 void perf_pmu_enable(struct pmu *pmu)
81 int *count = this_cpu_ptr(pmu->pmu_disable_count);
86 static DEFINE_PER_CPU(struct list_head, rotation_list);
89 * perf_pmu_rotate_start() and perf_rotate_context() are fully serialized
90 * because they're strictly cpu affine and rotate_start is called with IRQs
91 * disabled, while rotate_context is called from IRQ context.
93 static void perf_pmu_rotate_start(struct pmu *pmu)
95 struct perf_cpu_context *cpuctx = this_cpu_ptr(pmu->pmu_cpu_context);
96 struct list_head *head = &__get_cpu_var(rotation_list);
98 WARN_ON(!irqs_disabled());
100 if (list_empty(&cpuctx->rotation_list))
101 list_add(&cpuctx->rotation_list, head);
104 static void get_ctx(struct perf_event_context *ctx)
106 WARN_ON(!atomic_inc_not_zero(&ctx->refcount));
109 static void free_ctx(struct rcu_head *head)
111 struct perf_event_context *ctx;
113 ctx = container_of(head, struct perf_event_context, rcu_head);
117 static void put_ctx(struct perf_event_context *ctx)
119 if (atomic_dec_and_test(&ctx->refcount)) {
121 put_ctx(ctx->parent_ctx);
123 put_task_struct(ctx->task);
124 call_rcu(&ctx->rcu_head, free_ctx);
128 static void unclone_ctx(struct perf_event_context *ctx)
130 if (ctx->parent_ctx) {
131 put_ctx(ctx->parent_ctx);
132 ctx->parent_ctx = NULL;
137 * If we inherit events we want to return the parent event id
140 static u64 primary_event_id(struct perf_event *event)
145 id = event->parent->id;
151 * Get the perf_event_context for a task and lock it.
152 * This has to cope with with the fact that until it is locked,
153 * the context could get moved to another task.
155 static struct perf_event_context *
156 perf_lock_task_context(struct task_struct *task, int ctxn, unsigned long *flags)
158 struct perf_event_context *ctx;
162 ctx = rcu_dereference(task->perf_event_ctxp[ctxn]);
165 * If this context is a clone of another, it might
166 * get swapped for another underneath us by
167 * perf_event_task_sched_out, though the
168 * rcu_read_lock() protects us from any context
169 * getting freed. Lock the context and check if it
170 * got swapped before we could get the lock, and retry
171 * if so. If we locked the right context, then it
172 * can't get swapped on us any more.
174 raw_spin_lock_irqsave(&ctx->lock, *flags);
175 if (ctx != rcu_dereference(task->perf_event_ctxp[ctxn])) {
176 raw_spin_unlock_irqrestore(&ctx->lock, *flags);
180 if (!atomic_inc_not_zero(&ctx->refcount)) {
181 raw_spin_unlock_irqrestore(&ctx->lock, *flags);
190 * Get the context for a task and increment its pin_count so it
191 * can't get swapped to another task. This also increments its
192 * reference count so that the context can't get freed.
194 static struct perf_event_context *
195 perf_pin_task_context(struct task_struct *task, int ctxn)
197 struct perf_event_context *ctx;
200 ctx = perf_lock_task_context(task, ctxn, &flags);
203 raw_spin_unlock_irqrestore(&ctx->lock, flags);
208 static void perf_unpin_context(struct perf_event_context *ctx)
212 raw_spin_lock_irqsave(&ctx->lock, flags);
214 raw_spin_unlock_irqrestore(&ctx->lock, flags);
218 static inline u64 perf_clock(void)
220 return local_clock();
224 * Update the record of the current time in a context.
226 static void update_context_time(struct perf_event_context *ctx)
228 u64 now = perf_clock();
230 ctx->time += now - ctx->timestamp;
231 ctx->timestamp = now;
235 * Update the total_time_enabled and total_time_running fields for a event.
237 static void update_event_times(struct perf_event *event)
239 struct perf_event_context *ctx = event->ctx;
242 if (event->state < PERF_EVENT_STATE_INACTIVE ||
243 event->group_leader->state < PERF_EVENT_STATE_INACTIVE)
249 run_end = event->tstamp_stopped;
251 event->total_time_enabled = run_end - event->tstamp_enabled;
253 if (event->state == PERF_EVENT_STATE_INACTIVE)
254 run_end = event->tstamp_stopped;
258 event->total_time_running = run_end - event->tstamp_running;
262 * Update total_time_enabled and total_time_running for all events in a group.
264 static void update_group_times(struct perf_event *leader)
266 struct perf_event *event;
268 update_event_times(leader);
269 list_for_each_entry(event, &leader->sibling_list, group_entry)
270 update_event_times(event);
273 static struct list_head *
274 ctx_group_list(struct perf_event *event, struct perf_event_context *ctx)
276 if (event->attr.pinned)
277 return &ctx->pinned_groups;
279 return &ctx->flexible_groups;
283 * Add a event from the lists for its context.
284 * Must be called with ctx->mutex and ctx->lock held.
287 list_add_event(struct perf_event *event, struct perf_event_context *ctx)
289 WARN_ON_ONCE(event->attach_state & PERF_ATTACH_CONTEXT);
290 event->attach_state |= PERF_ATTACH_CONTEXT;
293 * If we're a stand alone event or group leader, we go to the context
294 * list, group events are kept attached to the group so that
295 * perf_group_detach can, at all times, locate all siblings.
297 if (event->group_leader == event) {
298 struct list_head *list;
300 if (is_software_event(event))
301 event->group_flags |= PERF_GROUP_SOFTWARE;
303 list = ctx_group_list(event, ctx);
304 list_add_tail(&event->group_entry, list);
307 list_add_rcu(&event->event_entry, &ctx->event_list);
309 perf_pmu_rotate_start(ctx->pmu);
311 if (event->attr.inherit_stat)
315 static void perf_group_attach(struct perf_event *event)
317 struct perf_event *group_leader = event->group_leader;
320 * We can have double attach due to group movement in perf_event_open.
322 if (event->attach_state & PERF_ATTACH_GROUP)
325 event->attach_state |= PERF_ATTACH_GROUP;
327 if (group_leader == event)
330 if (group_leader->group_flags & PERF_GROUP_SOFTWARE &&
331 !is_software_event(event))
332 group_leader->group_flags &= ~PERF_GROUP_SOFTWARE;
334 list_add_tail(&event->group_entry, &group_leader->sibling_list);
335 group_leader->nr_siblings++;
339 * Remove a event from the lists for its context.
340 * Must be called with ctx->mutex and ctx->lock held.
343 list_del_event(struct perf_event *event, struct perf_event_context *ctx)
346 * We can have double detach due to exit/hot-unplug + close.
348 if (!(event->attach_state & PERF_ATTACH_CONTEXT))
351 event->attach_state &= ~PERF_ATTACH_CONTEXT;
354 if (event->attr.inherit_stat)
357 list_del_rcu(&event->event_entry);
359 if (event->group_leader == event)
360 list_del_init(&event->group_entry);
362 update_group_times(event);
365 * If event was in error state, then keep it
366 * that way, otherwise bogus counts will be
367 * returned on read(). The only way to get out
368 * of error state is by explicit re-enabling
371 if (event->state > PERF_EVENT_STATE_OFF)
372 event->state = PERF_EVENT_STATE_OFF;
375 static void perf_group_detach(struct perf_event *event)
377 struct perf_event *sibling, *tmp;
378 struct list_head *list = NULL;
381 * We can have double detach due to exit/hot-unplug + close.
383 if (!(event->attach_state & PERF_ATTACH_GROUP))
386 event->attach_state &= ~PERF_ATTACH_GROUP;
389 * If this is a sibling, remove it from its group.
391 if (event->group_leader != event) {
392 list_del_init(&event->group_entry);
393 event->group_leader->nr_siblings--;
397 if (!list_empty(&event->group_entry))
398 list = &event->group_entry;
401 * If this was a group event with sibling events then
402 * upgrade the siblings to singleton events by adding them
403 * to whatever list we are on.
405 list_for_each_entry_safe(sibling, tmp, &event->sibling_list, group_entry) {
407 list_move_tail(&sibling->group_entry, list);
408 sibling->group_leader = sibling;
410 /* Inherit group flags from the previous leader */
411 sibling->group_flags = event->group_flags;
416 event_filter_match(struct perf_event *event)
418 return event->cpu == -1 || event->cpu == smp_processor_id();
422 event_sched_out(struct perf_event *event,
423 struct perf_cpu_context *cpuctx,
424 struct perf_event_context *ctx)
428 * An event which could not be activated because of
429 * filter mismatch still needs to have its timings
430 * maintained, otherwise bogus information is return
431 * via read() for time_enabled, time_running:
433 if (event->state == PERF_EVENT_STATE_INACTIVE
434 && !event_filter_match(event)) {
435 delta = ctx->time - event->tstamp_stopped;
436 event->tstamp_running += delta;
437 event->tstamp_stopped = ctx->time;
440 if (event->state != PERF_EVENT_STATE_ACTIVE)
443 event->state = PERF_EVENT_STATE_INACTIVE;
444 if (event->pending_disable) {
445 event->pending_disable = 0;
446 event->state = PERF_EVENT_STATE_OFF;
448 event->tstamp_stopped = ctx->time;
449 event->pmu->del(event, 0);
452 if (!is_software_event(event))
453 cpuctx->active_oncpu--;
455 if (event->attr.exclusive || !cpuctx->active_oncpu)
456 cpuctx->exclusive = 0;
460 group_sched_out(struct perf_event *group_event,
461 struct perf_cpu_context *cpuctx,
462 struct perf_event_context *ctx)
464 struct perf_event *event;
465 int state = group_event->state;
467 event_sched_out(group_event, cpuctx, ctx);
470 * Schedule out siblings (if any):
472 list_for_each_entry(event, &group_event->sibling_list, group_entry)
473 event_sched_out(event, cpuctx, ctx);
475 if (state == PERF_EVENT_STATE_ACTIVE && group_event->attr.exclusive)
476 cpuctx->exclusive = 0;
479 static inline struct perf_cpu_context *
480 __get_cpu_context(struct perf_event_context *ctx)
482 return this_cpu_ptr(ctx->pmu->pmu_cpu_context);
486 * Cross CPU call to remove a performance event
488 * We disable the event on the hardware level first. After that we
489 * remove it from the context list.
491 static void __perf_event_remove_from_context(void *info)
493 struct perf_event *event = info;
494 struct perf_event_context *ctx = event->ctx;
495 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
498 * If this is a task context, we need to check whether it is
499 * the current task context of this cpu. If not it has been
500 * scheduled out before the smp call arrived.
502 if (ctx->task && cpuctx->task_ctx != ctx)
505 raw_spin_lock(&ctx->lock);
507 event_sched_out(event, cpuctx, ctx);
509 list_del_event(event, ctx);
511 raw_spin_unlock(&ctx->lock);
516 * Remove the event from a task's (or a CPU's) list of events.
518 * Must be called with ctx->mutex held.
520 * CPU events are removed with a smp call. For task events we only
521 * call when the task is on a CPU.
523 * If event->ctx is a cloned context, callers must make sure that
524 * every task struct that event->ctx->task could possibly point to
525 * remains valid. This is OK when called from perf_release since
526 * that only calls us on the top-level context, which can't be a clone.
527 * When called from perf_event_exit_task, it's OK because the
528 * context has been detached from its task.
530 static void perf_event_remove_from_context(struct perf_event *event)
532 struct perf_event_context *ctx = event->ctx;
533 struct task_struct *task = ctx->task;
537 * Per cpu events are removed via an smp call and
538 * the removal is always successful.
540 smp_call_function_single(event->cpu,
541 __perf_event_remove_from_context,
547 task_oncpu_function_call(task, __perf_event_remove_from_context,
550 raw_spin_lock_irq(&ctx->lock);
552 * If the context is active we need to retry the smp call.
554 if (ctx->nr_active && !list_empty(&event->group_entry)) {
555 raw_spin_unlock_irq(&ctx->lock);
560 * The lock prevents that this context is scheduled in so we
561 * can remove the event safely, if the call above did not
564 if (!list_empty(&event->group_entry))
565 list_del_event(event, ctx);
566 raw_spin_unlock_irq(&ctx->lock);
570 * Cross CPU call to disable a performance event
572 static void __perf_event_disable(void *info)
574 struct perf_event *event = info;
575 struct perf_event_context *ctx = event->ctx;
576 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
579 * If this is a per-task event, need to check whether this
580 * event's task is the current task on this cpu.
582 if (ctx->task && cpuctx->task_ctx != ctx)
585 raw_spin_lock(&ctx->lock);
588 * If the event is on, turn it off.
589 * If it is in error state, leave it in error state.
591 if (event->state >= PERF_EVENT_STATE_INACTIVE) {
592 update_context_time(ctx);
593 update_group_times(event);
594 if (event == event->group_leader)
595 group_sched_out(event, cpuctx, ctx);
597 event_sched_out(event, cpuctx, ctx);
598 event->state = PERF_EVENT_STATE_OFF;
601 raw_spin_unlock(&ctx->lock);
607 * If event->ctx is a cloned context, callers must make sure that
608 * every task struct that event->ctx->task could possibly point to
609 * remains valid. This condition is satisifed when called through
610 * perf_event_for_each_child or perf_event_for_each because they
611 * hold the top-level event's child_mutex, so any descendant that
612 * goes to exit will block in sync_child_event.
613 * When called from perf_pending_event it's OK because event->ctx
614 * is the current context on this CPU and preemption is disabled,
615 * hence we can't get into perf_event_task_sched_out for this context.
617 void perf_event_disable(struct perf_event *event)
619 struct perf_event_context *ctx = event->ctx;
620 struct task_struct *task = ctx->task;
624 * Disable the event on the cpu that it's on
626 smp_call_function_single(event->cpu, __perf_event_disable,
632 task_oncpu_function_call(task, __perf_event_disable, event);
634 raw_spin_lock_irq(&ctx->lock);
636 * If the event is still active, we need to retry the cross-call.
638 if (event->state == PERF_EVENT_STATE_ACTIVE) {
639 raw_spin_unlock_irq(&ctx->lock);
644 * Since we have the lock this context can't be scheduled
645 * in, so we can change the state safely.
647 if (event->state == PERF_EVENT_STATE_INACTIVE) {
648 update_group_times(event);
649 event->state = PERF_EVENT_STATE_OFF;
652 raw_spin_unlock_irq(&ctx->lock);
656 event_sched_in(struct perf_event *event,
657 struct perf_cpu_context *cpuctx,
658 struct perf_event_context *ctx)
660 if (event->state <= PERF_EVENT_STATE_OFF)
663 event->state = PERF_EVENT_STATE_ACTIVE;
664 event->oncpu = smp_processor_id();
666 * The new state must be visible before we turn it on in the hardware:
670 if (event->pmu->add(event, PERF_EF_START)) {
671 event->state = PERF_EVENT_STATE_INACTIVE;
676 event->tstamp_running += ctx->time - event->tstamp_stopped;
678 event->shadow_ctx_time = ctx->time - ctx->timestamp;
680 if (!is_software_event(event))
681 cpuctx->active_oncpu++;
684 if (event->attr.exclusive)
685 cpuctx->exclusive = 1;
691 group_sched_in(struct perf_event *group_event,
692 struct perf_cpu_context *cpuctx,
693 struct perf_event_context *ctx)
695 struct perf_event *event, *partial_group = NULL;
696 struct pmu *pmu = group_event->pmu;
698 bool simulate = false;
700 if (group_event->state == PERF_EVENT_STATE_OFF)
705 if (event_sched_in(group_event, cpuctx, ctx)) {
706 pmu->cancel_txn(pmu);
711 * Schedule in siblings as one group (if any):
713 list_for_each_entry(event, &group_event->sibling_list, group_entry) {
714 if (event_sched_in(event, cpuctx, ctx)) {
715 partial_group = event;
720 if (!pmu->commit_txn(pmu))
725 * Groups can be scheduled in as one unit only, so undo any
726 * partial group before returning:
727 * The events up to the failed event are scheduled out normally,
728 * tstamp_stopped will be updated.
730 * The failed events and the remaining siblings need to have
731 * their timings updated as if they had gone thru event_sched_in()
732 * and event_sched_out(). This is required to get consistent timings
733 * across the group. This also takes care of the case where the group
734 * could never be scheduled by ensuring tstamp_stopped is set to mark
735 * the time the event was actually stopped, such that time delta
736 * calculation in update_event_times() is correct.
738 list_for_each_entry(event, &group_event->sibling_list, group_entry) {
739 if (event == partial_group)
743 event->tstamp_running += now - event->tstamp_stopped;
744 event->tstamp_stopped = now;
746 event_sched_out(event, cpuctx, ctx);
749 event_sched_out(group_event, cpuctx, ctx);
751 pmu->cancel_txn(pmu);
757 * Work out whether we can put this event group on the CPU now.
759 static int group_can_go_on(struct perf_event *event,
760 struct perf_cpu_context *cpuctx,
764 * Groups consisting entirely of software events can always go on.
766 if (event->group_flags & PERF_GROUP_SOFTWARE)
769 * If an exclusive group is already on, no other hardware
772 if (cpuctx->exclusive)
775 * If this group is exclusive and there are already
776 * events on the CPU, it can't go on.
778 if (event->attr.exclusive && cpuctx->active_oncpu)
781 * Otherwise, try to add it if all previous groups were able
787 static void add_event_to_ctx(struct perf_event *event,
788 struct perf_event_context *ctx)
790 list_add_event(event, ctx);
791 perf_group_attach(event);
792 event->tstamp_enabled = ctx->time;
793 event->tstamp_running = ctx->time;
794 event->tstamp_stopped = ctx->time;
798 * Cross CPU call to install and enable a performance event
800 * Must be called with ctx->mutex held
802 static void __perf_install_in_context(void *info)
804 struct perf_event *event = info;
805 struct perf_event_context *ctx = event->ctx;
806 struct perf_event *leader = event->group_leader;
807 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
811 * If this is a task context, we need to check whether it is
812 * the current task context of this cpu. If not it has been
813 * scheduled out before the smp call arrived.
814 * Or possibly this is the right context but it isn't
815 * on this cpu because it had no events.
817 if (ctx->task && cpuctx->task_ctx != ctx) {
818 if (cpuctx->task_ctx || ctx->task != current)
820 cpuctx->task_ctx = ctx;
823 raw_spin_lock(&ctx->lock);
825 update_context_time(ctx);
827 add_event_to_ctx(event, ctx);
829 if (event->cpu != -1 && event->cpu != smp_processor_id())
833 * Don't put the event on if it is disabled or if
834 * it is in a group and the group isn't on.
836 if (event->state != PERF_EVENT_STATE_INACTIVE ||
837 (leader != event && leader->state != PERF_EVENT_STATE_ACTIVE))
841 * An exclusive event can't go on if there are already active
842 * hardware events, and no hardware event can go on if there
843 * is already an exclusive event on.
845 if (!group_can_go_on(event, cpuctx, 1))
848 err = event_sched_in(event, cpuctx, ctx);
852 * This event couldn't go on. If it is in a group
853 * then we have to pull the whole group off.
854 * If the event group is pinned then put it in error state.
857 group_sched_out(leader, cpuctx, ctx);
858 if (leader->attr.pinned) {
859 update_group_times(leader);
860 leader->state = PERF_EVENT_STATE_ERROR;
865 raw_spin_unlock(&ctx->lock);
869 * Attach a performance event to a context
871 * First we add the event to the list with the hardware enable bit
872 * in event->hw_config cleared.
874 * If the event is attached to a task which is on a CPU we use a smp
875 * call to enable it in the task context. The task might have been
876 * scheduled away, but we check this in the smp call again.
878 * Must be called with ctx->mutex held.
881 perf_install_in_context(struct perf_event_context *ctx,
882 struct perf_event *event,
885 struct task_struct *task = ctx->task;
891 * Per cpu events are installed via an smp call and
892 * the install is always successful.
894 smp_call_function_single(cpu, __perf_install_in_context,
900 task_oncpu_function_call(task, __perf_install_in_context,
903 raw_spin_lock_irq(&ctx->lock);
905 * we need to retry the smp call.
907 if (ctx->is_active && list_empty(&event->group_entry)) {
908 raw_spin_unlock_irq(&ctx->lock);
913 * The lock prevents that this context is scheduled in so we
914 * can add the event safely, if it the call above did not
917 if (list_empty(&event->group_entry))
918 add_event_to_ctx(event, ctx);
919 raw_spin_unlock_irq(&ctx->lock);
923 * Put a event into inactive state and update time fields.
924 * Enabling the leader of a group effectively enables all
925 * the group members that aren't explicitly disabled, so we
926 * have to update their ->tstamp_enabled also.
927 * Note: this works for group members as well as group leaders
928 * since the non-leader members' sibling_lists will be empty.
930 static void __perf_event_mark_enabled(struct perf_event *event,
931 struct perf_event_context *ctx)
933 struct perf_event *sub;
935 event->state = PERF_EVENT_STATE_INACTIVE;
936 event->tstamp_enabled = ctx->time - event->total_time_enabled;
937 list_for_each_entry(sub, &event->sibling_list, group_entry) {
938 if (sub->state >= PERF_EVENT_STATE_INACTIVE) {
939 sub->tstamp_enabled =
940 ctx->time - sub->total_time_enabled;
946 * Cross CPU call to enable a performance event
948 static void __perf_event_enable(void *info)
950 struct perf_event *event = info;
951 struct perf_event_context *ctx = event->ctx;
952 struct perf_event *leader = event->group_leader;
953 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
957 * If this is a per-task event, need to check whether this
958 * event's task is the current task on this cpu.
960 if (ctx->task && cpuctx->task_ctx != ctx) {
961 if (cpuctx->task_ctx || ctx->task != current)
963 cpuctx->task_ctx = ctx;
966 raw_spin_lock(&ctx->lock);
968 update_context_time(ctx);
970 if (event->state >= PERF_EVENT_STATE_INACTIVE)
972 __perf_event_mark_enabled(event, ctx);
974 if (event->cpu != -1 && event->cpu != smp_processor_id())
978 * If the event is in a group and isn't the group leader,
979 * then don't put it on unless the group is on.
981 if (leader != event && leader->state != PERF_EVENT_STATE_ACTIVE)
984 if (!group_can_go_on(event, cpuctx, 1)) {
988 err = group_sched_in(event, cpuctx, ctx);
990 err = event_sched_in(event, cpuctx, ctx);
995 * If this event can't go on and it's part of a
996 * group, then the whole group has to come off.
999 group_sched_out(leader, cpuctx, ctx);
1000 if (leader->attr.pinned) {
1001 update_group_times(leader);
1002 leader->state = PERF_EVENT_STATE_ERROR;
1007 raw_spin_unlock(&ctx->lock);
1013 * If event->ctx is a cloned context, callers must make sure that
1014 * every task struct that event->ctx->task could possibly point to
1015 * remains valid. This condition is satisfied when called through
1016 * perf_event_for_each_child or perf_event_for_each as described
1017 * for perf_event_disable.
1019 void perf_event_enable(struct perf_event *event)
1021 struct perf_event_context *ctx = event->ctx;
1022 struct task_struct *task = ctx->task;
1026 * Enable the event on the cpu that it's on
1028 smp_call_function_single(event->cpu, __perf_event_enable,
1033 raw_spin_lock_irq(&ctx->lock);
1034 if (event->state >= PERF_EVENT_STATE_INACTIVE)
1038 * If the event is in error state, clear that first.
1039 * That way, if we see the event in error state below, we
1040 * know that it has gone back into error state, as distinct
1041 * from the task having been scheduled away before the
1042 * cross-call arrived.
1044 if (event->state == PERF_EVENT_STATE_ERROR)
1045 event->state = PERF_EVENT_STATE_OFF;
1048 raw_spin_unlock_irq(&ctx->lock);
1049 task_oncpu_function_call(task, __perf_event_enable, event);
1051 raw_spin_lock_irq(&ctx->lock);
1054 * If the context is active and the event is still off,
1055 * we need to retry the cross-call.
1057 if (ctx->is_active && event->state == PERF_EVENT_STATE_OFF)
1061 * Since we have the lock this context can't be scheduled
1062 * in, so we can change the state safely.
1064 if (event->state == PERF_EVENT_STATE_OFF)
1065 __perf_event_mark_enabled(event, ctx);
1068 raw_spin_unlock_irq(&ctx->lock);
1071 static int perf_event_refresh(struct perf_event *event, int refresh)
1074 * not supported on inherited events
1076 if (event->attr.inherit)
1079 atomic_add(refresh, &event->event_limit);
1080 perf_event_enable(event);
1086 EVENT_FLEXIBLE = 0x1,
1088 EVENT_ALL = EVENT_FLEXIBLE | EVENT_PINNED,
1091 static void ctx_sched_out(struct perf_event_context *ctx,
1092 struct perf_cpu_context *cpuctx,
1093 enum event_type_t event_type)
1095 struct perf_event *event;
1097 raw_spin_lock(&ctx->lock);
1098 perf_pmu_disable(ctx->pmu);
1100 if (likely(!ctx->nr_events))
1102 update_context_time(ctx);
1104 if (!ctx->nr_active)
1107 if (event_type & EVENT_PINNED) {
1108 list_for_each_entry(event, &ctx->pinned_groups, group_entry)
1109 group_sched_out(event, cpuctx, ctx);
1112 if (event_type & EVENT_FLEXIBLE) {
1113 list_for_each_entry(event, &ctx->flexible_groups, group_entry)
1114 group_sched_out(event, cpuctx, ctx);
1117 perf_pmu_enable(ctx->pmu);
1118 raw_spin_unlock(&ctx->lock);
1122 * Test whether two contexts are equivalent, i.e. whether they
1123 * have both been cloned from the same version of the same context
1124 * and they both have the same number of enabled events.
1125 * If the number of enabled events is the same, then the set
1126 * of enabled events should be the same, because these are both
1127 * inherited contexts, therefore we can't access individual events
1128 * in them directly with an fd; we can only enable/disable all
1129 * events via prctl, or enable/disable all events in a family
1130 * via ioctl, which will have the same effect on both contexts.
1132 static int context_equiv(struct perf_event_context *ctx1,
1133 struct perf_event_context *ctx2)
1135 return ctx1->parent_ctx && ctx1->parent_ctx == ctx2->parent_ctx
1136 && ctx1->parent_gen == ctx2->parent_gen
1137 && !ctx1->pin_count && !ctx2->pin_count;
1140 static void __perf_event_sync_stat(struct perf_event *event,
1141 struct perf_event *next_event)
1145 if (!event->attr.inherit_stat)
1149 * Update the event value, we cannot use perf_event_read()
1150 * because we're in the middle of a context switch and have IRQs
1151 * disabled, which upsets smp_call_function_single(), however
1152 * we know the event must be on the current CPU, therefore we
1153 * don't need to use it.
1155 switch (event->state) {
1156 case PERF_EVENT_STATE_ACTIVE:
1157 event->pmu->read(event);
1160 case PERF_EVENT_STATE_INACTIVE:
1161 update_event_times(event);
1169 * In order to keep per-task stats reliable we need to flip the event
1170 * values when we flip the contexts.
1172 value = local64_read(&next_event->count);
1173 value = local64_xchg(&event->count, value);
1174 local64_set(&next_event->count, value);
1176 swap(event->total_time_enabled, next_event->total_time_enabled);
1177 swap(event->total_time_running, next_event->total_time_running);
1180 * Since we swizzled the values, update the user visible data too.
1182 perf_event_update_userpage(event);
1183 perf_event_update_userpage(next_event);
1186 #define list_next_entry(pos, member) \
1187 list_entry(pos->member.next, typeof(*pos), member)
1189 static void perf_event_sync_stat(struct perf_event_context *ctx,
1190 struct perf_event_context *next_ctx)
1192 struct perf_event *event, *next_event;
1197 update_context_time(ctx);
1199 event = list_first_entry(&ctx->event_list,
1200 struct perf_event, event_entry);
1202 next_event = list_first_entry(&next_ctx->event_list,
1203 struct perf_event, event_entry);
1205 while (&event->event_entry != &ctx->event_list &&
1206 &next_event->event_entry != &next_ctx->event_list) {
1208 __perf_event_sync_stat(event, next_event);
1210 event = list_next_entry(event, event_entry);
1211 next_event = list_next_entry(next_event, event_entry);
1215 void perf_event_context_sched_out(struct task_struct *task, int ctxn,
1216 struct task_struct *next)
1218 struct perf_event_context *ctx = task->perf_event_ctxp[ctxn];
1219 struct perf_event_context *next_ctx;
1220 struct perf_event_context *parent;
1221 struct perf_cpu_context *cpuctx;
1227 cpuctx = __get_cpu_context(ctx);
1228 if (!cpuctx->task_ctx)
1232 parent = rcu_dereference(ctx->parent_ctx);
1233 next_ctx = next->perf_event_ctxp[ctxn];
1234 if (parent && next_ctx &&
1235 rcu_dereference(next_ctx->parent_ctx) == parent) {
1237 * Looks like the two contexts are clones, so we might be
1238 * able to optimize the context switch. We lock both
1239 * contexts and check that they are clones under the
1240 * lock (including re-checking that neither has been
1241 * uncloned in the meantime). It doesn't matter which
1242 * order we take the locks because no other cpu could
1243 * be trying to lock both of these tasks.
1245 raw_spin_lock(&ctx->lock);
1246 raw_spin_lock_nested(&next_ctx->lock, SINGLE_DEPTH_NESTING);
1247 if (context_equiv(ctx, next_ctx)) {
1249 * XXX do we need a memory barrier of sorts
1250 * wrt to rcu_dereference() of perf_event_ctxp
1252 task->perf_event_ctxp[ctxn] = next_ctx;
1253 next->perf_event_ctxp[ctxn] = ctx;
1255 next_ctx->task = task;
1258 perf_event_sync_stat(ctx, next_ctx);
1260 raw_spin_unlock(&next_ctx->lock);
1261 raw_spin_unlock(&ctx->lock);
1266 ctx_sched_out(ctx, cpuctx, EVENT_ALL);
1267 cpuctx->task_ctx = NULL;
1271 #define for_each_task_context_nr(ctxn) \
1272 for ((ctxn) = 0; (ctxn) < perf_nr_task_contexts; (ctxn)++)
1275 * Called from scheduler to remove the events of the current task,
1276 * with interrupts disabled.
1278 * We stop each event and update the event value in event->count.
1280 * This does not protect us against NMI, but disable()
1281 * sets the disabled bit in the control field of event _before_
1282 * accessing the event control register. If a NMI hits, then it will
1283 * not restart the event.
1285 void __perf_event_task_sched_out(struct task_struct *task,
1286 struct task_struct *next)
1290 perf_sw_event(PERF_COUNT_SW_CONTEXT_SWITCHES, 1, 1, NULL, 0);
1292 for_each_task_context_nr(ctxn)
1293 perf_event_context_sched_out(task, ctxn, next);
1296 static void task_ctx_sched_out(struct perf_event_context *ctx,
1297 enum event_type_t event_type)
1299 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
1301 if (!cpuctx->task_ctx)
1304 if (WARN_ON_ONCE(ctx != cpuctx->task_ctx))
1307 ctx_sched_out(ctx, cpuctx, event_type);
1308 cpuctx->task_ctx = NULL;
1312 * Called with IRQs disabled
1314 static void cpu_ctx_sched_out(struct perf_cpu_context *cpuctx,
1315 enum event_type_t event_type)
1317 ctx_sched_out(&cpuctx->ctx, cpuctx, event_type);
1321 ctx_pinned_sched_in(struct perf_event_context *ctx,
1322 struct perf_cpu_context *cpuctx)
1324 struct perf_event *event;
1326 list_for_each_entry(event, &ctx->pinned_groups, group_entry) {
1327 if (event->state <= PERF_EVENT_STATE_OFF)
1329 if (event->cpu != -1 && event->cpu != smp_processor_id())
1332 if (group_can_go_on(event, cpuctx, 1))
1333 group_sched_in(event, cpuctx, ctx);
1336 * If this pinned group hasn't been scheduled,
1337 * put it in error state.
1339 if (event->state == PERF_EVENT_STATE_INACTIVE) {
1340 update_group_times(event);
1341 event->state = PERF_EVENT_STATE_ERROR;
1347 ctx_flexible_sched_in(struct perf_event_context *ctx,
1348 struct perf_cpu_context *cpuctx)
1350 struct perf_event *event;
1353 list_for_each_entry(event, &ctx->flexible_groups, group_entry) {
1354 /* Ignore events in OFF or ERROR state */
1355 if (event->state <= PERF_EVENT_STATE_OFF)
1358 * Listen to the 'cpu' scheduling filter constraint
1361 if (event->cpu != -1 && event->cpu != smp_processor_id())
1364 if (group_can_go_on(event, cpuctx, can_add_hw)) {
1365 if (group_sched_in(event, cpuctx, ctx))
1372 ctx_sched_in(struct perf_event_context *ctx,
1373 struct perf_cpu_context *cpuctx,
1374 enum event_type_t event_type)
1376 raw_spin_lock(&ctx->lock);
1378 if (likely(!ctx->nr_events))
1381 ctx->timestamp = perf_clock();
1384 * First go through the list and put on any pinned groups
1385 * in order to give them the best chance of going on.
1387 if (event_type & EVENT_PINNED)
1388 ctx_pinned_sched_in(ctx, cpuctx);
1390 /* Then walk through the lower prio flexible groups */
1391 if (event_type & EVENT_FLEXIBLE)
1392 ctx_flexible_sched_in(ctx, cpuctx);
1395 raw_spin_unlock(&ctx->lock);
1398 static void cpu_ctx_sched_in(struct perf_cpu_context *cpuctx,
1399 enum event_type_t event_type)
1401 struct perf_event_context *ctx = &cpuctx->ctx;
1403 ctx_sched_in(ctx, cpuctx, event_type);
1406 static void task_ctx_sched_in(struct perf_event_context *ctx,
1407 enum event_type_t event_type)
1409 struct perf_cpu_context *cpuctx;
1411 cpuctx = __get_cpu_context(ctx);
1412 if (cpuctx->task_ctx == ctx)
1415 ctx_sched_in(ctx, cpuctx, event_type);
1416 cpuctx->task_ctx = ctx;
1419 void perf_event_context_sched_in(struct perf_event_context *ctx)
1421 struct perf_cpu_context *cpuctx;
1423 cpuctx = __get_cpu_context(ctx);
1424 if (cpuctx->task_ctx == ctx)
1427 perf_pmu_disable(ctx->pmu);
1429 * We want to keep the following priority order:
1430 * cpu pinned (that don't need to move), task pinned,
1431 * cpu flexible, task flexible.
1433 cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE);
1435 ctx_sched_in(ctx, cpuctx, EVENT_PINNED);
1436 cpu_ctx_sched_in(cpuctx, EVENT_FLEXIBLE);
1437 ctx_sched_in(ctx, cpuctx, EVENT_FLEXIBLE);
1439 cpuctx->task_ctx = ctx;
1442 * Since these rotations are per-cpu, we need to ensure the
1443 * cpu-context we got scheduled on is actually rotating.
1445 perf_pmu_rotate_start(ctx->pmu);
1446 perf_pmu_enable(ctx->pmu);
1450 * Called from scheduler to add the events of the current task
1451 * with interrupts disabled.
1453 * We restore the event value and then enable it.
1455 * This does not protect us against NMI, but enable()
1456 * sets the enabled bit in the control field of event _before_
1457 * accessing the event control register. If a NMI hits, then it will
1458 * keep the event running.
1460 void __perf_event_task_sched_in(struct task_struct *task)
1462 struct perf_event_context *ctx;
1465 for_each_task_context_nr(ctxn) {
1466 ctx = task->perf_event_ctxp[ctxn];
1470 perf_event_context_sched_in(ctx);
1474 #define MAX_INTERRUPTS (~0ULL)
1476 static void perf_log_throttle(struct perf_event *event, int enable);
1478 static u64 perf_calculate_period(struct perf_event *event, u64 nsec, u64 count)
1480 u64 frequency = event->attr.sample_freq;
1481 u64 sec = NSEC_PER_SEC;
1482 u64 divisor, dividend;
1484 int count_fls, nsec_fls, frequency_fls, sec_fls;
1486 count_fls = fls64(count);
1487 nsec_fls = fls64(nsec);
1488 frequency_fls = fls64(frequency);
1492 * We got @count in @nsec, with a target of sample_freq HZ
1493 * the target period becomes:
1496 * period = -------------------
1497 * @nsec * sample_freq
1502 * Reduce accuracy by one bit such that @a and @b converge
1503 * to a similar magnitude.
1505 #define REDUCE_FLS(a, b) \
1507 if (a##_fls > b##_fls) { \
1517 * Reduce accuracy until either term fits in a u64, then proceed with
1518 * the other, so that finally we can do a u64/u64 division.
1520 while (count_fls + sec_fls > 64 && nsec_fls + frequency_fls > 64) {
1521 REDUCE_FLS(nsec, frequency);
1522 REDUCE_FLS(sec, count);
1525 if (count_fls + sec_fls > 64) {
1526 divisor = nsec * frequency;
1528 while (count_fls + sec_fls > 64) {
1529 REDUCE_FLS(count, sec);
1533 dividend = count * sec;
1535 dividend = count * sec;
1537 while (nsec_fls + frequency_fls > 64) {
1538 REDUCE_FLS(nsec, frequency);
1542 divisor = nsec * frequency;
1548 return div64_u64(dividend, divisor);
1551 static void perf_adjust_period(struct perf_event *event, u64 nsec, u64 count)
1553 struct hw_perf_event *hwc = &event->hw;
1554 s64 period, sample_period;
1557 period = perf_calculate_period(event, nsec, count);
1559 delta = (s64)(period - hwc->sample_period);
1560 delta = (delta + 7) / 8; /* low pass filter */
1562 sample_period = hwc->sample_period + delta;
1567 hwc->sample_period = sample_period;
1569 if (local64_read(&hwc->period_left) > 8*sample_period) {
1570 event->pmu->stop(event, PERF_EF_UPDATE);
1571 local64_set(&hwc->period_left, 0);
1572 event->pmu->start(event, PERF_EF_RELOAD);
1576 static void perf_ctx_adjust_freq(struct perf_event_context *ctx, u64 period)
1578 struct perf_event *event;
1579 struct hw_perf_event *hwc;
1580 u64 interrupts, now;
1583 raw_spin_lock(&ctx->lock);
1584 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
1585 if (event->state != PERF_EVENT_STATE_ACTIVE)
1588 if (event->cpu != -1 && event->cpu != smp_processor_id())
1593 interrupts = hwc->interrupts;
1594 hwc->interrupts = 0;
1597 * unthrottle events on the tick
1599 if (interrupts == MAX_INTERRUPTS) {
1600 perf_log_throttle(event, 1);
1601 event->pmu->start(event, 0);
1604 if (!event->attr.freq || !event->attr.sample_freq)
1607 event->pmu->read(event);
1608 now = local64_read(&event->count);
1609 delta = now - hwc->freq_count_stamp;
1610 hwc->freq_count_stamp = now;
1613 perf_adjust_period(event, period, delta);
1615 raw_spin_unlock(&ctx->lock);
1619 * Round-robin a context's events:
1621 static void rotate_ctx(struct perf_event_context *ctx)
1623 raw_spin_lock(&ctx->lock);
1625 /* Rotate the first entry last of non-pinned groups */
1626 list_rotate_left(&ctx->flexible_groups);
1628 raw_spin_unlock(&ctx->lock);
1632 * perf_pmu_rotate_start() and perf_rotate_context() are fully serialized
1633 * because they're strictly cpu affine and rotate_start is called with IRQs
1634 * disabled, while rotate_context is called from IRQ context.
1636 static void perf_rotate_context(struct perf_cpu_context *cpuctx)
1638 u64 interval = (u64)cpuctx->jiffies_interval * TICK_NSEC;
1639 struct perf_event_context *ctx = NULL;
1640 int rotate = 0, remove = 1;
1642 if (cpuctx->ctx.nr_events) {
1644 if (cpuctx->ctx.nr_events != cpuctx->ctx.nr_active)
1648 ctx = cpuctx->task_ctx;
1649 if (ctx && ctx->nr_events) {
1651 if (ctx->nr_events != ctx->nr_active)
1655 perf_pmu_disable(cpuctx->ctx.pmu);
1656 perf_ctx_adjust_freq(&cpuctx->ctx, interval);
1658 perf_ctx_adjust_freq(ctx, interval);
1663 cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE);
1665 task_ctx_sched_out(ctx, EVENT_FLEXIBLE);
1667 rotate_ctx(&cpuctx->ctx);
1671 cpu_ctx_sched_in(cpuctx, EVENT_FLEXIBLE);
1673 task_ctx_sched_in(ctx, EVENT_FLEXIBLE);
1677 list_del_init(&cpuctx->rotation_list);
1679 perf_pmu_enable(cpuctx->ctx.pmu);
1682 void perf_event_task_tick(void)
1684 struct list_head *head = &__get_cpu_var(rotation_list);
1685 struct perf_cpu_context *cpuctx, *tmp;
1687 WARN_ON(!irqs_disabled());
1689 list_for_each_entry_safe(cpuctx, tmp, head, rotation_list) {
1690 if (cpuctx->jiffies_interval == 1 ||
1691 !(jiffies % cpuctx->jiffies_interval))
1692 perf_rotate_context(cpuctx);
1696 static int event_enable_on_exec(struct perf_event *event,
1697 struct perf_event_context *ctx)
1699 if (!event->attr.enable_on_exec)
1702 event->attr.enable_on_exec = 0;
1703 if (event->state >= PERF_EVENT_STATE_INACTIVE)
1706 __perf_event_mark_enabled(event, ctx);
1712 * Enable all of a task's events that have been marked enable-on-exec.
1713 * This expects task == current.
1715 static void perf_event_enable_on_exec(struct perf_event_context *ctx)
1717 struct perf_event *event;
1718 unsigned long flags;
1722 local_irq_save(flags);
1723 if (!ctx || !ctx->nr_events)
1726 task_ctx_sched_out(ctx, EVENT_ALL);
1728 raw_spin_lock(&ctx->lock);
1730 list_for_each_entry(event, &ctx->pinned_groups, group_entry) {
1731 ret = event_enable_on_exec(event, ctx);
1736 list_for_each_entry(event, &ctx->flexible_groups, group_entry) {
1737 ret = event_enable_on_exec(event, ctx);
1743 * Unclone this context if we enabled any event.
1748 raw_spin_unlock(&ctx->lock);
1750 perf_event_context_sched_in(ctx);
1752 local_irq_restore(flags);
1756 * Cross CPU call to read the hardware event
1758 static void __perf_event_read(void *info)
1760 struct perf_event *event = info;
1761 struct perf_event_context *ctx = event->ctx;
1762 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
1765 * If this is a task context, we need to check whether it is
1766 * the current task context of this cpu. If not it has been
1767 * scheduled out before the smp call arrived. In that case
1768 * event->count would have been updated to a recent sample
1769 * when the event was scheduled out.
1771 if (ctx->task && cpuctx->task_ctx != ctx)
1774 raw_spin_lock(&ctx->lock);
1775 update_context_time(ctx);
1776 update_event_times(event);
1777 raw_spin_unlock(&ctx->lock);
1779 event->pmu->read(event);
1782 static inline u64 perf_event_count(struct perf_event *event)
1784 return local64_read(&event->count) + atomic64_read(&event->child_count);
1787 static u64 perf_event_read(struct perf_event *event)
1790 * If event is enabled and currently active on a CPU, update the
1791 * value in the event structure:
1793 if (event->state == PERF_EVENT_STATE_ACTIVE) {
1794 smp_call_function_single(event->oncpu,
1795 __perf_event_read, event, 1);
1796 } else if (event->state == PERF_EVENT_STATE_INACTIVE) {
1797 struct perf_event_context *ctx = event->ctx;
1798 unsigned long flags;
1800 raw_spin_lock_irqsave(&ctx->lock, flags);
1802 * may read while context is not active
1803 * (e.g., thread is blocked), in that case
1804 * we cannot update context time
1807 update_context_time(ctx);
1808 update_event_times(event);
1809 raw_spin_unlock_irqrestore(&ctx->lock, flags);
1812 return perf_event_count(event);
1819 struct callchain_cpus_entries {
1820 struct rcu_head rcu_head;
1821 struct perf_callchain_entry *cpu_entries[0];
1824 static DEFINE_PER_CPU(int, callchain_recursion[PERF_NR_CONTEXTS]);
1825 static atomic_t nr_callchain_events;
1826 static DEFINE_MUTEX(callchain_mutex);
1827 struct callchain_cpus_entries *callchain_cpus_entries;
1830 __weak void perf_callchain_kernel(struct perf_callchain_entry *entry,
1831 struct pt_regs *regs)
1835 __weak void perf_callchain_user(struct perf_callchain_entry *entry,
1836 struct pt_regs *regs)
1840 static void release_callchain_buffers_rcu(struct rcu_head *head)
1842 struct callchain_cpus_entries *entries;
1845 entries = container_of(head, struct callchain_cpus_entries, rcu_head);
1847 for_each_possible_cpu(cpu)
1848 kfree(entries->cpu_entries[cpu]);
1853 static void release_callchain_buffers(void)
1855 struct callchain_cpus_entries *entries;
1857 entries = callchain_cpus_entries;
1858 rcu_assign_pointer(callchain_cpus_entries, NULL);
1859 call_rcu(&entries->rcu_head, release_callchain_buffers_rcu);
1862 static int alloc_callchain_buffers(void)
1866 struct callchain_cpus_entries *entries;
1869 * We can't use the percpu allocation API for data that can be
1870 * accessed from NMI. Use a temporary manual per cpu allocation
1871 * until that gets sorted out.
1873 size = sizeof(*entries) + sizeof(struct perf_callchain_entry *) *
1874 num_possible_cpus();
1876 entries = kzalloc(size, GFP_KERNEL);
1880 size = sizeof(struct perf_callchain_entry) * PERF_NR_CONTEXTS;
1882 for_each_possible_cpu(cpu) {
1883 entries->cpu_entries[cpu] = kmalloc_node(size, GFP_KERNEL,
1885 if (!entries->cpu_entries[cpu])
1889 rcu_assign_pointer(callchain_cpus_entries, entries);
1894 for_each_possible_cpu(cpu)
1895 kfree(entries->cpu_entries[cpu]);
1901 static int get_callchain_buffers(void)
1906 mutex_lock(&callchain_mutex);
1908 count = atomic_inc_return(&nr_callchain_events);
1909 if (WARN_ON_ONCE(count < 1)) {
1915 /* If the allocation failed, give up */
1916 if (!callchain_cpus_entries)
1921 err = alloc_callchain_buffers();
1923 release_callchain_buffers();
1925 mutex_unlock(&callchain_mutex);
1930 static void put_callchain_buffers(void)
1932 if (atomic_dec_and_mutex_lock(&nr_callchain_events, &callchain_mutex)) {
1933 release_callchain_buffers();
1934 mutex_unlock(&callchain_mutex);
1938 static int get_recursion_context(int *recursion)
1946 else if (in_softirq())
1951 if (recursion[rctx])
1960 static inline void put_recursion_context(int *recursion, int rctx)
1966 static struct perf_callchain_entry *get_callchain_entry(int *rctx)
1969 struct callchain_cpus_entries *entries;
1971 *rctx = get_recursion_context(__get_cpu_var(callchain_recursion));
1975 entries = rcu_dereference(callchain_cpus_entries);
1979 cpu = smp_processor_id();
1981 return &entries->cpu_entries[cpu][*rctx];
1985 put_callchain_entry(int rctx)
1987 put_recursion_context(__get_cpu_var(callchain_recursion), rctx);
1990 static struct perf_callchain_entry *perf_callchain(struct pt_regs *regs)
1993 struct perf_callchain_entry *entry;
1996 entry = get_callchain_entry(&rctx);
2005 if (!user_mode(regs)) {
2006 perf_callchain_store(entry, PERF_CONTEXT_KERNEL);
2007 perf_callchain_kernel(entry, regs);
2009 regs = task_pt_regs(current);
2015 perf_callchain_store(entry, PERF_CONTEXT_USER);
2016 perf_callchain_user(entry, regs);
2020 put_callchain_entry(rctx);
2026 * Initialize the perf_event context in a task_struct:
2028 static void __perf_event_init_context(struct perf_event_context *ctx)
2030 raw_spin_lock_init(&ctx->lock);
2031 mutex_init(&ctx->mutex);
2032 INIT_LIST_HEAD(&ctx->pinned_groups);
2033 INIT_LIST_HEAD(&ctx->flexible_groups);
2034 INIT_LIST_HEAD(&ctx->event_list);
2035 atomic_set(&ctx->refcount, 1);
2038 static struct perf_event_context *
2039 alloc_perf_context(struct pmu *pmu, struct task_struct *task)
2041 struct perf_event_context *ctx;
2043 ctx = kzalloc(sizeof(struct perf_event_context), GFP_KERNEL);
2047 __perf_event_init_context(ctx);
2050 get_task_struct(task);
2057 static struct task_struct *
2058 find_lively_task_by_vpid(pid_t vpid)
2060 struct task_struct *task;
2067 task = find_task_by_vpid(vpid);
2069 get_task_struct(task);
2073 return ERR_PTR(-ESRCH);
2076 * Can't attach events to a dying task.
2079 if (task->flags & PF_EXITING)
2082 /* Reuse ptrace permission checks for now. */
2084 if (!ptrace_may_access(task, PTRACE_MODE_READ))
2089 put_task_struct(task);
2090 return ERR_PTR(err);
2094 static struct perf_event_context *
2095 find_get_context(struct pmu *pmu, struct task_struct *task, int cpu)
2097 struct perf_event_context *ctx;
2098 struct perf_cpu_context *cpuctx;
2099 unsigned long flags;
2102 if (!task && cpu != -1) {
2103 /* Must be root to operate on a CPU event: */
2104 if (perf_paranoid_cpu() && !capable(CAP_SYS_ADMIN))
2105 return ERR_PTR(-EACCES);
2107 if (cpu < 0 || cpu >= nr_cpumask_bits)
2108 return ERR_PTR(-EINVAL);
2111 * We could be clever and allow to attach a event to an
2112 * offline CPU and activate it when the CPU comes up, but
2115 if (!cpu_online(cpu))
2116 return ERR_PTR(-ENODEV);
2118 cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu);
2126 ctxn = pmu->task_ctx_nr;
2131 ctx = perf_lock_task_context(task, ctxn, &flags);
2134 raw_spin_unlock_irqrestore(&ctx->lock, flags);
2138 ctx = alloc_perf_context(pmu, task);
2145 if (cmpxchg(&task->perf_event_ctxp[ctxn], NULL, ctx)) {
2147 * We raced with some other task; use
2148 * the context they set.
2150 put_task_struct(task);
2159 return ERR_PTR(err);
2162 static void perf_event_free_filter(struct perf_event *event);
2164 static void free_event_rcu(struct rcu_head *head)
2166 struct perf_event *event;
2168 event = container_of(head, struct perf_event, rcu_head);
2170 put_pid_ns(event->ns);
2171 perf_event_free_filter(event);
2175 static void perf_buffer_put(struct perf_buffer *buffer);
2177 static void free_event(struct perf_event *event)
2179 irq_work_sync(&event->pending);
2181 if (!event->parent) {
2182 if (event->attach_state & PERF_ATTACH_TASK)
2183 jump_label_dec(&perf_task_events);
2184 if (event->attr.mmap || event->attr.mmap_data)
2185 atomic_dec(&nr_mmap_events);
2186 if (event->attr.comm)
2187 atomic_dec(&nr_comm_events);
2188 if (event->attr.task)
2189 atomic_dec(&nr_task_events);
2190 if (event->attr.sample_type & PERF_SAMPLE_CALLCHAIN)
2191 put_callchain_buffers();
2194 if (event->buffer) {
2195 perf_buffer_put(event->buffer);
2196 event->buffer = NULL;
2200 event->destroy(event);
2203 put_ctx(event->ctx);
2205 call_rcu(&event->rcu_head, free_event_rcu);
2208 int perf_event_release_kernel(struct perf_event *event)
2210 struct perf_event_context *ctx = event->ctx;
2213 * Remove from the PMU, can't get re-enabled since we got
2214 * here because the last ref went.
2216 perf_event_disable(event);
2218 WARN_ON_ONCE(ctx->parent_ctx);
2220 * There are two ways this annotation is useful:
2222 * 1) there is a lock recursion from perf_event_exit_task
2223 * see the comment there.
2225 * 2) there is a lock-inversion with mmap_sem through
2226 * perf_event_read_group(), which takes faults while
2227 * holding ctx->mutex, however this is called after
2228 * the last filedesc died, so there is no possibility
2229 * to trigger the AB-BA case.
2231 mutex_lock_nested(&ctx->mutex, SINGLE_DEPTH_NESTING);
2232 raw_spin_lock_irq(&ctx->lock);
2233 perf_group_detach(event);
2234 list_del_event(event, ctx);
2235 raw_spin_unlock_irq(&ctx->lock);
2236 mutex_unlock(&ctx->mutex);
2238 mutex_lock(&event->owner->perf_event_mutex);
2239 list_del_init(&event->owner_entry);
2240 mutex_unlock(&event->owner->perf_event_mutex);
2241 put_task_struct(event->owner);
2247 EXPORT_SYMBOL_GPL(perf_event_release_kernel);
2250 * Called when the last reference to the file is gone.
2252 static int perf_release(struct inode *inode, struct file *file)
2254 struct perf_event *event = file->private_data;
2256 file->private_data = NULL;
2258 return perf_event_release_kernel(event);
2261 static int perf_event_read_size(struct perf_event *event)
2263 int entry = sizeof(u64); /* value */
2267 if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
2268 size += sizeof(u64);
2270 if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
2271 size += sizeof(u64);
2273 if (event->attr.read_format & PERF_FORMAT_ID)
2274 entry += sizeof(u64);
2276 if (event->attr.read_format & PERF_FORMAT_GROUP) {
2277 nr += event->group_leader->nr_siblings;
2278 size += sizeof(u64);
2286 u64 perf_event_read_value(struct perf_event *event, u64 *enabled, u64 *running)
2288 struct perf_event *child;
2294 mutex_lock(&event->child_mutex);
2295 total += perf_event_read(event);
2296 *enabled += event->total_time_enabled +
2297 atomic64_read(&event->child_total_time_enabled);
2298 *running += event->total_time_running +
2299 atomic64_read(&event->child_total_time_running);
2301 list_for_each_entry(child, &event->child_list, child_list) {
2302 total += perf_event_read(child);
2303 *enabled += child->total_time_enabled;
2304 *running += child->total_time_running;
2306 mutex_unlock(&event->child_mutex);
2310 EXPORT_SYMBOL_GPL(perf_event_read_value);
2312 static int perf_event_read_group(struct perf_event *event,
2313 u64 read_format, char __user *buf)
2315 struct perf_event *leader = event->group_leader, *sub;
2316 int n = 0, size = 0, ret = -EFAULT;
2317 struct perf_event_context *ctx = leader->ctx;
2319 u64 count, enabled, running;
2321 mutex_lock(&ctx->mutex);
2322 count = perf_event_read_value(leader, &enabled, &running);
2324 values[n++] = 1 + leader->nr_siblings;
2325 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
2326 values[n++] = enabled;
2327 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
2328 values[n++] = running;
2329 values[n++] = count;
2330 if (read_format & PERF_FORMAT_ID)
2331 values[n++] = primary_event_id(leader);
2333 size = n * sizeof(u64);
2335 if (copy_to_user(buf, values, size))
2340 list_for_each_entry(sub, &leader->sibling_list, group_entry) {
2343 values[n++] = perf_event_read_value(sub, &enabled, &running);
2344 if (read_format & PERF_FORMAT_ID)
2345 values[n++] = primary_event_id(sub);
2347 size = n * sizeof(u64);
2349 if (copy_to_user(buf + ret, values, size)) {
2357 mutex_unlock(&ctx->mutex);
2362 static int perf_event_read_one(struct perf_event *event,
2363 u64 read_format, char __user *buf)
2365 u64 enabled, running;
2369 values[n++] = perf_event_read_value(event, &enabled, &running);
2370 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
2371 values[n++] = enabled;
2372 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
2373 values[n++] = running;
2374 if (read_format & PERF_FORMAT_ID)
2375 values[n++] = primary_event_id(event);
2377 if (copy_to_user(buf, values, n * sizeof(u64)))
2380 return n * sizeof(u64);
2384 * Read the performance event - simple non blocking version for now
2387 perf_read_hw(struct perf_event *event, char __user *buf, size_t count)
2389 u64 read_format = event->attr.read_format;
2393 * Return end-of-file for a read on a event that is in
2394 * error state (i.e. because it was pinned but it couldn't be
2395 * scheduled on to the CPU at some point).
2397 if (event->state == PERF_EVENT_STATE_ERROR)
2400 if (count < perf_event_read_size(event))
2403 WARN_ON_ONCE(event->ctx->parent_ctx);
2404 if (read_format & PERF_FORMAT_GROUP)
2405 ret = perf_event_read_group(event, read_format, buf);
2407 ret = perf_event_read_one(event, read_format, buf);
2413 perf_read(struct file *file, char __user *buf, size_t count, loff_t *ppos)
2415 struct perf_event *event = file->private_data;
2417 return perf_read_hw(event, buf, count);
2420 static unsigned int perf_poll(struct file *file, poll_table *wait)
2422 struct perf_event *event = file->private_data;
2423 struct perf_buffer *buffer;
2424 unsigned int events = POLL_HUP;
2427 buffer = rcu_dereference(event->buffer);
2429 events = atomic_xchg(&buffer->poll, 0);
2432 poll_wait(file, &event->waitq, wait);
2437 static void perf_event_reset(struct perf_event *event)
2439 (void)perf_event_read(event);
2440 local64_set(&event->count, 0);
2441 perf_event_update_userpage(event);
2445 * Holding the top-level event's child_mutex means that any
2446 * descendant process that has inherited this event will block
2447 * in sync_child_event if it goes to exit, thus satisfying the
2448 * task existence requirements of perf_event_enable/disable.
2450 static void perf_event_for_each_child(struct perf_event *event,
2451 void (*func)(struct perf_event *))
2453 struct perf_event *child;
2455 WARN_ON_ONCE(event->ctx->parent_ctx);
2456 mutex_lock(&event->child_mutex);
2458 list_for_each_entry(child, &event->child_list, child_list)
2460 mutex_unlock(&event->child_mutex);
2463 static void perf_event_for_each(struct perf_event *event,
2464 void (*func)(struct perf_event *))
2466 struct perf_event_context *ctx = event->ctx;
2467 struct perf_event *sibling;
2469 WARN_ON_ONCE(ctx->parent_ctx);
2470 mutex_lock(&ctx->mutex);
2471 event = event->group_leader;
2473 perf_event_for_each_child(event, func);
2475 list_for_each_entry(sibling, &event->sibling_list, group_entry)
2476 perf_event_for_each_child(event, func);
2477 mutex_unlock(&ctx->mutex);
2480 static int perf_event_period(struct perf_event *event, u64 __user *arg)
2482 struct perf_event_context *ctx = event->ctx;
2486 if (!event->attr.sample_period)
2489 if (copy_from_user(&value, arg, sizeof(value)))
2495 raw_spin_lock_irq(&ctx->lock);
2496 if (event->attr.freq) {
2497 if (value > sysctl_perf_event_sample_rate) {
2502 event->attr.sample_freq = value;
2504 event->attr.sample_period = value;
2505 event->hw.sample_period = value;
2508 raw_spin_unlock_irq(&ctx->lock);
2513 static const struct file_operations perf_fops;
2515 static struct perf_event *perf_fget_light(int fd, int *fput_needed)
2519 file = fget_light(fd, fput_needed);
2521 return ERR_PTR(-EBADF);
2523 if (file->f_op != &perf_fops) {
2524 fput_light(file, *fput_needed);
2526 return ERR_PTR(-EBADF);
2529 return file->private_data;
2532 static int perf_event_set_output(struct perf_event *event,
2533 struct perf_event *output_event);
2534 static int perf_event_set_filter(struct perf_event *event, void __user *arg);
2536 static long perf_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
2538 struct perf_event *event = file->private_data;
2539 void (*func)(struct perf_event *);
2543 case PERF_EVENT_IOC_ENABLE:
2544 func = perf_event_enable;
2546 case PERF_EVENT_IOC_DISABLE:
2547 func = perf_event_disable;
2549 case PERF_EVENT_IOC_RESET:
2550 func = perf_event_reset;
2553 case PERF_EVENT_IOC_REFRESH:
2554 return perf_event_refresh(event, arg);
2556 case PERF_EVENT_IOC_PERIOD:
2557 return perf_event_period(event, (u64 __user *)arg);
2559 case PERF_EVENT_IOC_SET_OUTPUT:
2561 struct perf_event *output_event = NULL;
2562 int fput_needed = 0;
2566 output_event = perf_fget_light(arg, &fput_needed);
2567 if (IS_ERR(output_event))
2568 return PTR_ERR(output_event);
2571 ret = perf_event_set_output(event, output_event);
2573 fput_light(output_event->filp, fput_needed);
2578 case PERF_EVENT_IOC_SET_FILTER:
2579 return perf_event_set_filter(event, (void __user *)arg);
2585 if (flags & PERF_IOC_FLAG_GROUP)
2586 perf_event_for_each(event, func);
2588 perf_event_for_each_child(event, func);
2593 int perf_event_task_enable(void)
2595 struct perf_event *event;
2597 mutex_lock(¤t->perf_event_mutex);
2598 list_for_each_entry(event, ¤t->perf_event_list, owner_entry)
2599 perf_event_for_each_child(event, perf_event_enable);
2600 mutex_unlock(¤t->perf_event_mutex);
2605 int perf_event_task_disable(void)
2607 struct perf_event *event;
2609 mutex_lock(¤t->perf_event_mutex);
2610 list_for_each_entry(event, ¤t->perf_event_list, owner_entry)
2611 perf_event_for_each_child(event, perf_event_disable);
2612 mutex_unlock(¤t->perf_event_mutex);
2617 #ifndef PERF_EVENT_INDEX_OFFSET
2618 # define PERF_EVENT_INDEX_OFFSET 0
2621 static int perf_event_index(struct perf_event *event)
2623 if (event->hw.state & PERF_HES_STOPPED)
2626 if (event->state != PERF_EVENT_STATE_ACTIVE)
2629 return event->hw.idx + 1 - PERF_EVENT_INDEX_OFFSET;
2633 * Callers need to ensure there can be no nesting of this function, otherwise
2634 * the seqlock logic goes bad. We can not serialize this because the arch
2635 * code calls this from NMI context.
2637 void perf_event_update_userpage(struct perf_event *event)
2639 struct perf_event_mmap_page *userpg;
2640 struct perf_buffer *buffer;
2643 buffer = rcu_dereference(event->buffer);
2647 userpg = buffer->user_page;
2650 * Disable preemption so as to not let the corresponding user-space
2651 * spin too long if we get preempted.
2656 userpg->index = perf_event_index(event);
2657 userpg->offset = perf_event_count(event);
2658 if (event->state == PERF_EVENT_STATE_ACTIVE)
2659 userpg->offset -= local64_read(&event->hw.prev_count);
2661 userpg->time_enabled = event->total_time_enabled +
2662 atomic64_read(&event->child_total_time_enabled);
2664 userpg->time_running = event->total_time_running +
2665 atomic64_read(&event->child_total_time_running);
2674 static unsigned long perf_data_size(struct perf_buffer *buffer);
2677 perf_buffer_init(struct perf_buffer *buffer, long watermark, int flags)
2679 long max_size = perf_data_size(buffer);
2682 buffer->watermark = min(max_size, watermark);
2684 if (!buffer->watermark)
2685 buffer->watermark = max_size / 2;
2687 if (flags & PERF_BUFFER_WRITABLE)
2688 buffer->writable = 1;
2690 atomic_set(&buffer->refcount, 1);
2693 #ifndef CONFIG_PERF_USE_VMALLOC
2696 * Back perf_mmap() with regular GFP_KERNEL-0 pages.
2699 static struct page *
2700 perf_mmap_to_page(struct perf_buffer *buffer, unsigned long pgoff)
2702 if (pgoff > buffer->nr_pages)
2706 return virt_to_page(buffer->user_page);
2708 return virt_to_page(buffer->data_pages[pgoff - 1]);
2711 static void *perf_mmap_alloc_page(int cpu)
2716 node = (cpu == -1) ? cpu : cpu_to_node(cpu);
2717 page = alloc_pages_node(node, GFP_KERNEL | __GFP_ZERO, 0);
2721 return page_address(page);
2724 static struct perf_buffer *
2725 perf_buffer_alloc(int nr_pages, long watermark, int cpu, int flags)
2727 struct perf_buffer *buffer;
2731 size = sizeof(struct perf_buffer);
2732 size += nr_pages * sizeof(void *);
2734 buffer = kzalloc(size, GFP_KERNEL);
2738 buffer->user_page = perf_mmap_alloc_page(cpu);
2739 if (!buffer->user_page)
2740 goto fail_user_page;
2742 for (i = 0; i < nr_pages; i++) {
2743 buffer->data_pages[i] = perf_mmap_alloc_page(cpu);
2744 if (!buffer->data_pages[i])
2745 goto fail_data_pages;
2748 buffer->nr_pages = nr_pages;
2750 perf_buffer_init(buffer, watermark, flags);
2755 for (i--; i >= 0; i--)
2756 free_page((unsigned long)buffer->data_pages[i]);
2758 free_page((unsigned long)buffer->user_page);
2767 static void perf_mmap_free_page(unsigned long addr)
2769 struct page *page = virt_to_page((void *)addr);
2771 page->mapping = NULL;
2775 static void perf_buffer_free(struct perf_buffer *buffer)
2779 perf_mmap_free_page((unsigned long)buffer->user_page);
2780 for (i = 0; i < buffer->nr_pages; i++)
2781 perf_mmap_free_page((unsigned long)buffer->data_pages[i]);
2785 static inline int page_order(struct perf_buffer *buffer)
2793 * Back perf_mmap() with vmalloc memory.
2795 * Required for architectures that have d-cache aliasing issues.
2798 static inline int page_order(struct perf_buffer *buffer)
2800 return buffer->page_order;
2803 static struct page *
2804 perf_mmap_to_page(struct perf_buffer *buffer, unsigned long pgoff)
2806 if (pgoff > (1UL << page_order(buffer)))
2809 return vmalloc_to_page((void *)buffer->user_page + pgoff * PAGE_SIZE);
2812 static void perf_mmap_unmark_page(void *addr)
2814 struct page *page = vmalloc_to_page(addr);
2816 page->mapping = NULL;
2819 static void perf_buffer_free_work(struct work_struct *work)
2821 struct perf_buffer *buffer;
2825 buffer = container_of(work, struct perf_buffer, work);
2826 nr = 1 << page_order(buffer);
2828 base = buffer->user_page;
2829 for (i = 0; i < nr + 1; i++)
2830 perf_mmap_unmark_page(base + (i * PAGE_SIZE));
2836 static void perf_buffer_free(struct perf_buffer *buffer)
2838 schedule_work(&buffer->work);
2841 static struct perf_buffer *
2842 perf_buffer_alloc(int nr_pages, long watermark, int cpu, int flags)
2844 struct perf_buffer *buffer;
2848 size = sizeof(struct perf_buffer);
2849 size += sizeof(void *);
2851 buffer = kzalloc(size, GFP_KERNEL);
2855 INIT_WORK(&buffer->work, perf_buffer_free_work);
2857 all_buf = vmalloc_user((nr_pages + 1) * PAGE_SIZE);
2861 buffer->user_page = all_buf;
2862 buffer->data_pages[0] = all_buf + PAGE_SIZE;
2863 buffer->page_order = ilog2(nr_pages);
2864 buffer->nr_pages = 1;
2866 perf_buffer_init(buffer, watermark, flags);
2879 static unsigned long perf_data_size(struct perf_buffer *buffer)
2881 return buffer->nr_pages << (PAGE_SHIFT + page_order(buffer));
2884 static int perf_mmap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
2886 struct perf_event *event = vma->vm_file->private_data;
2887 struct perf_buffer *buffer;
2888 int ret = VM_FAULT_SIGBUS;
2890 if (vmf->flags & FAULT_FLAG_MKWRITE) {
2891 if (vmf->pgoff == 0)
2897 buffer = rcu_dereference(event->buffer);
2901 if (vmf->pgoff && (vmf->flags & FAULT_FLAG_WRITE))
2904 vmf->page = perf_mmap_to_page(buffer, vmf->pgoff);
2908 get_page(vmf->page);
2909 vmf->page->mapping = vma->vm_file->f_mapping;
2910 vmf->page->index = vmf->pgoff;
2919 static void perf_buffer_free_rcu(struct rcu_head *rcu_head)
2921 struct perf_buffer *buffer;
2923 buffer = container_of(rcu_head, struct perf_buffer, rcu_head);
2924 perf_buffer_free(buffer);
2927 static struct perf_buffer *perf_buffer_get(struct perf_event *event)
2929 struct perf_buffer *buffer;
2932 buffer = rcu_dereference(event->buffer);
2934 if (!atomic_inc_not_zero(&buffer->refcount))
2942 static void perf_buffer_put(struct perf_buffer *buffer)
2944 if (!atomic_dec_and_test(&buffer->refcount))
2947 call_rcu(&buffer->rcu_head, perf_buffer_free_rcu);
2950 static void perf_mmap_open(struct vm_area_struct *vma)
2952 struct perf_event *event = vma->vm_file->private_data;
2954 atomic_inc(&event->mmap_count);
2957 static void perf_mmap_close(struct vm_area_struct *vma)
2959 struct perf_event *event = vma->vm_file->private_data;
2961 if (atomic_dec_and_mutex_lock(&event->mmap_count, &event->mmap_mutex)) {
2962 unsigned long size = perf_data_size(event->buffer);
2963 struct user_struct *user = event->mmap_user;
2964 struct perf_buffer *buffer = event->buffer;
2966 atomic_long_sub((size >> PAGE_SHIFT) + 1, &user->locked_vm);
2967 vma->vm_mm->locked_vm -= event->mmap_locked;
2968 rcu_assign_pointer(event->buffer, NULL);
2969 mutex_unlock(&event->mmap_mutex);
2971 perf_buffer_put(buffer);
2976 static const struct vm_operations_struct perf_mmap_vmops = {
2977 .open = perf_mmap_open,
2978 .close = perf_mmap_close,
2979 .fault = perf_mmap_fault,
2980 .page_mkwrite = perf_mmap_fault,
2983 static int perf_mmap(struct file *file, struct vm_area_struct *vma)
2985 struct perf_event *event = file->private_data;
2986 unsigned long user_locked, user_lock_limit;
2987 struct user_struct *user = current_user();
2988 unsigned long locked, lock_limit;
2989 struct perf_buffer *buffer;
2990 unsigned long vma_size;
2991 unsigned long nr_pages;
2992 long user_extra, extra;
2993 int ret = 0, flags = 0;
2996 * Don't allow mmap() of inherited per-task counters. This would
2997 * create a performance issue due to all children writing to the
3000 if (event->cpu == -1 && event->attr.inherit)
3003 if (!(vma->vm_flags & VM_SHARED))
3006 vma_size = vma->vm_end - vma->vm_start;
3007 nr_pages = (vma_size / PAGE_SIZE) - 1;
3010 * If we have buffer pages ensure they're a power-of-two number, so we
3011 * can do bitmasks instead of modulo.
3013 if (nr_pages != 0 && !is_power_of_2(nr_pages))
3016 if (vma_size != PAGE_SIZE * (1 + nr_pages))
3019 if (vma->vm_pgoff != 0)
3022 WARN_ON_ONCE(event->ctx->parent_ctx);
3023 mutex_lock(&event->mmap_mutex);
3024 if (event->buffer) {
3025 if (event->buffer->nr_pages == nr_pages)
3026 atomic_inc(&event->buffer->refcount);
3032 user_extra = nr_pages + 1;
3033 user_lock_limit = sysctl_perf_event_mlock >> (PAGE_SHIFT - 10);
3036 * Increase the limit linearly with more CPUs:
3038 user_lock_limit *= num_online_cpus();
3040 user_locked = atomic_long_read(&user->locked_vm) + user_extra;
3043 if (user_locked > user_lock_limit)
3044 extra = user_locked - user_lock_limit;
3046 lock_limit = rlimit(RLIMIT_MEMLOCK);
3047 lock_limit >>= PAGE_SHIFT;
3048 locked = vma->vm_mm->locked_vm + extra;
3050 if ((locked > lock_limit) && perf_paranoid_tracepoint_raw() &&
3051 !capable(CAP_IPC_LOCK)) {
3056 WARN_ON(event->buffer);
3058 if (vma->vm_flags & VM_WRITE)
3059 flags |= PERF_BUFFER_WRITABLE;
3061 buffer = perf_buffer_alloc(nr_pages, event->attr.wakeup_watermark,
3067 rcu_assign_pointer(event->buffer, buffer);
3069 atomic_long_add(user_extra, &user->locked_vm);
3070 event->mmap_locked = extra;
3071 event->mmap_user = get_current_user();
3072 vma->vm_mm->locked_vm += event->mmap_locked;
3076 atomic_inc(&event->mmap_count);
3077 mutex_unlock(&event->mmap_mutex);
3079 vma->vm_flags |= VM_RESERVED;
3080 vma->vm_ops = &perf_mmap_vmops;
3085 static int perf_fasync(int fd, struct file *filp, int on)
3087 struct inode *inode = filp->f_path.dentry->d_inode;
3088 struct perf_event *event = filp->private_data;
3091 mutex_lock(&inode->i_mutex);
3092 retval = fasync_helper(fd, filp, on, &event->fasync);
3093 mutex_unlock(&inode->i_mutex);
3101 static const struct file_operations perf_fops = {
3102 .llseek = no_llseek,
3103 .release = perf_release,
3106 .unlocked_ioctl = perf_ioctl,
3107 .compat_ioctl = perf_ioctl,
3109 .fasync = perf_fasync,
3115 * If there's data, ensure we set the poll() state and publish everything
3116 * to user-space before waking everybody up.
3119 void perf_event_wakeup(struct perf_event *event)
3121 wake_up_all(&event->waitq);
3123 if (event->pending_kill) {
3124 kill_fasync(&event->fasync, SIGIO, event->pending_kill);
3125 event->pending_kill = 0;
3129 static void perf_pending_event(struct irq_work *entry)
3131 struct perf_event *event = container_of(entry,
3132 struct perf_event, pending);
3134 if (event->pending_disable) {
3135 event->pending_disable = 0;
3136 __perf_event_disable(event);
3139 if (event->pending_wakeup) {
3140 event->pending_wakeup = 0;
3141 perf_event_wakeup(event);
3146 * We assume there is only KVM supporting the callbacks.
3147 * Later on, we might change it to a list if there is
3148 * another virtualization implementation supporting the callbacks.
3150 struct perf_guest_info_callbacks *perf_guest_cbs;
3152 int perf_register_guest_info_callbacks(struct perf_guest_info_callbacks *cbs)
3154 perf_guest_cbs = cbs;
3157 EXPORT_SYMBOL_GPL(perf_register_guest_info_callbacks);
3159 int perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks *cbs)
3161 perf_guest_cbs = NULL;
3164 EXPORT_SYMBOL_GPL(perf_unregister_guest_info_callbacks);
3169 static bool perf_output_space(struct perf_buffer *buffer, unsigned long tail,
3170 unsigned long offset, unsigned long head)
3174 if (!buffer->writable)
3177 mask = perf_data_size(buffer) - 1;
3179 offset = (offset - tail) & mask;
3180 head = (head - tail) & mask;
3182 if ((int)(head - offset) < 0)
3188 static void perf_output_wakeup(struct perf_output_handle *handle)
3190 atomic_set(&handle->buffer->poll, POLL_IN);
3193 handle->event->pending_wakeup = 1;
3194 irq_work_queue(&handle->event->pending);
3196 perf_event_wakeup(handle->event);
3200 * We need to ensure a later event_id doesn't publish a head when a former
3201 * event isn't done writing. However since we need to deal with NMIs we
3202 * cannot fully serialize things.
3204 * We only publish the head (and generate a wakeup) when the outer-most
3207 static void perf_output_get_handle(struct perf_output_handle *handle)
3209 struct perf_buffer *buffer = handle->buffer;
3212 local_inc(&buffer->nest);
3213 handle->wakeup = local_read(&buffer->wakeup);
3216 static void perf_output_put_handle(struct perf_output_handle *handle)
3218 struct perf_buffer *buffer = handle->buffer;
3222 head = local_read(&buffer->head);
3225 * IRQ/NMI can happen here, which means we can miss a head update.
3228 if (!local_dec_and_test(&buffer->nest))
3232 * Publish the known good head. Rely on the full barrier implied
3233 * by atomic_dec_and_test() order the buffer->head read and this
3236 buffer->user_page->data_head = head;
3239 * Now check if we missed an update, rely on the (compiler)
3240 * barrier in atomic_dec_and_test() to re-read buffer->head.
3242 if (unlikely(head != local_read(&buffer->head))) {
3243 local_inc(&buffer->nest);
3247 if (handle->wakeup != local_read(&buffer->wakeup))
3248 perf_output_wakeup(handle);
3254 __always_inline void perf_output_copy(struct perf_output_handle *handle,
3255 const void *buf, unsigned int len)
3258 unsigned long size = min_t(unsigned long, handle->size, len);
3260 memcpy(handle->addr, buf, size);
3263 handle->addr += size;
3265 handle->size -= size;
3266 if (!handle->size) {
3267 struct perf_buffer *buffer = handle->buffer;
3270 handle->page &= buffer->nr_pages - 1;
3271 handle->addr = buffer->data_pages[handle->page];
3272 handle->size = PAGE_SIZE << page_order(buffer);
3277 int perf_output_begin(struct perf_output_handle *handle,
3278 struct perf_event *event, unsigned int size,
3279 int nmi, int sample)
3281 struct perf_buffer *buffer;
3282 unsigned long tail, offset, head;
3285 struct perf_event_header header;
3292 * For inherited events we send all the output towards the parent.
3295 event = event->parent;
3297 buffer = rcu_dereference(event->buffer);
3301 handle->buffer = buffer;
3302 handle->event = event;
3304 handle->sample = sample;
3306 if (!buffer->nr_pages)
3309 have_lost = local_read(&buffer->lost);
3311 size += sizeof(lost_event);
3313 perf_output_get_handle(handle);
3317 * Userspace could choose to issue a mb() before updating the
3318 * tail pointer. So that all reads will be completed before the
3321 tail = ACCESS_ONCE(buffer->user_page->data_tail);
3323 offset = head = local_read(&buffer->head);
3325 if (unlikely(!perf_output_space(buffer, tail, offset, head)))
3327 } while (local_cmpxchg(&buffer->head, offset, head) != offset);
3329 if (head - local_read(&buffer->wakeup) > buffer->watermark)
3330 local_add(buffer->watermark, &buffer->wakeup);
3332 handle->page = offset >> (PAGE_SHIFT + page_order(buffer));
3333 handle->page &= buffer->nr_pages - 1;
3334 handle->size = offset & ((PAGE_SIZE << page_order(buffer)) - 1);
3335 handle->addr = buffer->data_pages[handle->page];
3336 handle->addr += handle->size;
3337 handle->size = (PAGE_SIZE << page_order(buffer)) - handle->size;
3340 lost_event.header.type = PERF_RECORD_LOST;
3341 lost_event.header.misc = 0;
3342 lost_event.header.size = sizeof(lost_event);
3343 lost_event.id = event->id;
3344 lost_event.lost = local_xchg(&buffer->lost, 0);
3346 perf_output_put(handle, lost_event);
3352 local_inc(&buffer->lost);
3353 perf_output_put_handle(handle);
3360 void perf_output_end(struct perf_output_handle *handle)
3362 struct perf_event *event = handle->event;
3363 struct perf_buffer *buffer = handle->buffer;
3365 int wakeup_events = event->attr.wakeup_events;
3367 if (handle->sample && wakeup_events) {
3368 int events = local_inc_return(&buffer->events);
3369 if (events >= wakeup_events) {
3370 local_sub(wakeup_events, &buffer->events);
3371 local_inc(&buffer->wakeup);
3375 perf_output_put_handle(handle);
3379 static u32 perf_event_pid(struct perf_event *event, struct task_struct *p)
3382 * only top level events have the pid namespace they were created in
3385 event = event->parent;
3387 return task_tgid_nr_ns(p, event->ns);
3390 static u32 perf_event_tid(struct perf_event *event, struct task_struct *p)
3393 * only top level events have the pid namespace they were created in
3396 event = event->parent;
3398 return task_pid_nr_ns(p, event->ns);
3401 static void perf_output_read_one(struct perf_output_handle *handle,
3402 struct perf_event *event,
3403 u64 enabled, u64 running)
3405 u64 read_format = event->attr.read_format;
3409 values[n++] = perf_event_count(event);
3410 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
3411 values[n++] = enabled +
3412 atomic64_read(&event->child_total_time_enabled);
3414 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
3415 values[n++] = running +
3416 atomic64_read(&event->child_total_time_running);
3418 if (read_format & PERF_FORMAT_ID)
3419 values[n++] = primary_event_id(event);
3421 perf_output_copy(handle, values, n * sizeof(u64));
3425 * XXX PERF_FORMAT_GROUP vs inherited events seems difficult.
3427 static void perf_output_read_group(struct perf_output_handle *handle,
3428 struct perf_event *event,
3429 u64 enabled, u64 running)
3431 struct perf_event *leader = event->group_leader, *sub;
3432 u64 read_format = event->attr.read_format;
3436 values[n++] = 1 + leader->nr_siblings;
3438 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
3439 values[n++] = enabled;
3441 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
3442 values[n++] = running;
3444 if (leader != event)
3445 leader->pmu->read(leader);
3447 values[n++] = perf_event_count(leader);
3448 if (read_format & PERF_FORMAT_ID)
3449 values[n++] = primary_event_id(leader);
3451 perf_output_copy(handle, values, n * sizeof(u64));
3453 list_for_each_entry(sub, &leader->sibling_list, group_entry) {
3457 sub->pmu->read(sub);
3459 values[n++] = perf_event_count(sub);
3460 if (read_format & PERF_FORMAT_ID)
3461 values[n++] = primary_event_id(sub);
3463 perf_output_copy(handle, values, n * sizeof(u64));
3467 #define PERF_FORMAT_TOTAL_TIMES (PERF_FORMAT_TOTAL_TIME_ENABLED|\
3468 PERF_FORMAT_TOTAL_TIME_RUNNING)
3470 static void perf_output_read(struct perf_output_handle *handle,
3471 struct perf_event *event)
3473 u64 enabled = 0, running = 0, now, ctx_time;
3474 u64 read_format = event->attr.read_format;
3477 * compute total_time_enabled, total_time_running
3478 * based on snapshot values taken when the event
3479 * was last scheduled in.
3481 * we cannot simply called update_context_time()
3482 * because of locking issue as we are called in
3485 if (read_format & PERF_FORMAT_TOTAL_TIMES) {
3487 ctx_time = event->shadow_ctx_time + now;
3488 enabled = ctx_time - event->tstamp_enabled;
3489 running = ctx_time - event->tstamp_running;
3492 if (event->attr.read_format & PERF_FORMAT_GROUP)
3493 perf_output_read_group(handle, event, enabled, running);
3495 perf_output_read_one(handle, event, enabled, running);
3498 void perf_output_sample(struct perf_output_handle *handle,
3499 struct perf_event_header *header,
3500 struct perf_sample_data *data,
3501 struct perf_event *event)
3503 u64 sample_type = data->type;
3505 perf_output_put(handle, *header);
3507 if (sample_type & PERF_SAMPLE_IP)
3508 perf_output_put(handle, data->ip);
3510 if (sample_type & PERF_SAMPLE_TID)
3511 perf_output_put(handle, data->tid_entry);
3513 if (sample_type & PERF_SAMPLE_TIME)
3514 perf_output_put(handle, data->time);
3516 if (sample_type & PERF_SAMPLE_ADDR)
3517 perf_output_put(handle, data->addr);
3519 if (sample_type & PERF_SAMPLE_ID)
3520 perf_output_put(handle, data->id);
3522 if (sample_type & PERF_SAMPLE_STREAM_ID)
3523 perf_output_put(handle, data->stream_id);
3525 if (sample_type & PERF_SAMPLE_CPU)
3526 perf_output_put(handle, data->cpu_entry);
3528 if (sample_type & PERF_SAMPLE_PERIOD)
3529 perf_output_put(handle, data->period);
3531 if (sample_type & PERF_SAMPLE_READ)
3532 perf_output_read(handle, event);
3534 if (sample_type & PERF_SAMPLE_CALLCHAIN) {
3535 if (data->callchain) {
3538 if (data->callchain)
3539 size += data->callchain->nr;
3541 size *= sizeof(u64);
3543 perf_output_copy(handle, data->callchain, size);
3546 perf_output_put(handle, nr);
3550 if (sample_type & PERF_SAMPLE_RAW) {
3552 perf_output_put(handle, data->raw->size);
3553 perf_output_copy(handle, data->raw->data,
3560 .size = sizeof(u32),
3563 perf_output_put(handle, raw);
3568 void perf_prepare_sample(struct perf_event_header *header,
3569 struct perf_sample_data *data,
3570 struct perf_event *event,
3571 struct pt_regs *regs)
3573 u64 sample_type = event->attr.sample_type;
3575 data->type = sample_type;
3577 header->type = PERF_RECORD_SAMPLE;
3578 header->size = sizeof(*header);
3581 header->misc |= perf_misc_flags(regs);
3583 if (sample_type & PERF_SAMPLE_IP) {
3584 data->ip = perf_instruction_pointer(regs);
3586 header->size += sizeof(data->ip);
3589 if (sample_type & PERF_SAMPLE_TID) {
3590 /* namespace issues */
3591 data->tid_entry.pid = perf_event_pid(event, current);
3592 data->tid_entry.tid = perf_event_tid(event, current);
3594 header->size += sizeof(data->tid_entry);
3597 if (sample_type & PERF_SAMPLE_TIME) {
3598 data->time = perf_clock();
3600 header->size += sizeof(data->time);
3603 if (sample_type & PERF_SAMPLE_ADDR)
3604 header->size += sizeof(data->addr);
3606 if (sample_type & PERF_SAMPLE_ID) {
3607 data->id = primary_event_id(event);
3609 header->size += sizeof(data->id);
3612 if (sample_type & PERF_SAMPLE_STREAM_ID) {
3613 data->stream_id = event->id;
3615 header->size += sizeof(data->stream_id);
3618 if (sample_type & PERF_SAMPLE_CPU) {
3619 data->cpu_entry.cpu = raw_smp_processor_id();
3620 data->cpu_entry.reserved = 0;
3622 header->size += sizeof(data->cpu_entry);
3625 if (sample_type & PERF_SAMPLE_PERIOD)
3626 header->size += sizeof(data->period);
3628 if (sample_type & PERF_SAMPLE_READ)
3629 header->size += perf_event_read_size(event);
3631 if (sample_type & PERF_SAMPLE_CALLCHAIN) {
3634 data->callchain = perf_callchain(regs);
3636 if (data->callchain)
3637 size += data->callchain->nr;
3639 header->size += size * sizeof(u64);
3642 if (sample_type & PERF_SAMPLE_RAW) {
3643 int size = sizeof(u32);
3646 size += data->raw->size;
3648 size += sizeof(u32);
3650 WARN_ON_ONCE(size & (sizeof(u64)-1));
3651 header->size += size;
3655 static void perf_event_output(struct perf_event *event, int nmi,
3656 struct perf_sample_data *data,
3657 struct pt_regs *regs)
3659 struct perf_output_handle handle;
3660 struct perf_event_header header;
3662 /* protect the callchain buffers */
3665 perf_prepare_sample(&header, data, event, regs);
3667 if (perf_output_begin(&handle, event, header.size, nmi, 1))
3670 perf_output_sample(&handle, &header, data, event);
3672 perf_output_end(&handle);
3682 struct perf_read_event {
3683 struct perf_event_header header;
3690 perf_event_read_event(struct perf_event *event,
3691 struct task_struct *task)
3693 struct perf_output_handle handle;
3694 struct perf_read_event read_event = {
3696 .type = PERF_RECORD_READ,
3698 .size = sizeof(read_event) + perf_event_read_size(event),
3700 .pid = perf_event_pid(event, task),
3701 .tid = perf_event_tid(event, task),
3705 ret = perf_output_begin(&handle, event, read_event.header.size, 0, 0);
3709 perf_output_put(&handle, read_event);
3710 perf_output_read(&handle, event);
3712 perf_output_end(&handle);
3716 * task tracking -- fork/exit
3718 * enabled by: attr.comm | attr.mmap | attr.mmap_data | attr.task
3721 struct perf_task_event {
3722 struct task_struct *task;
3723 struct perf_event_context *task_ctx;
3726 struct perf_event_header header;
3736 static void perf_event_task_output(struct perf_event *event,
3737 struct perf_task_event *task_event)
3739 struct perf_output_handle handle;
3740 struct task_struct *task = task_event->task;
3743 size = task_event->event_id.header.size;
3744 ret = perf_output_begin(&handle, event, size, 0, 0);
3749 task_event->event_id.pid = perf_event_pid(event, task);
3750 task_event->event_id.ppid = perf_event_pid(event, current);
3752 task_event->event_id.tid = perf_event_tid(event, task);
3753 task_event->event_id.ptid = perf_event_tid(event, current);
3755 perf_output_put(&handle, task_event->event_id);
3757 perf_output_end(&handle);
3760 static int perf_event_task_match(struct perf_event *event)
3762 if (event->state < PERF_EVENT_STATE_INACTIVE)
3765 if (event->cpu != -1 && event->cpu != smp_processor_id())
3768 if (event->attr.comm || event->attr.mmap ||
3769 event->attr.mmap_data || event->attr.task)
3775 static void perf_event_task_ctx(struct perf_event_context *ctx,
3776 struct perf_task_event *task_event)
3778 struct perf_event *event;
3780 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
3781 if (perf_event_task_match(event))
3782 perf_event_task_output(event, task_event);
3786 static void perf_event_task_event(struct perf_task_event *task_event)
3788 struct perf_cpu_context *cpuctx;
3789 struct perf_event_context *ctx;
3794 list_for_each_entry_rcu(pmu, &pmus, entry) {
3795 cpuctx = get_cpu_ptr(pmu->pmu_cpu_context);
3796 perf_event_task_ctx(&cpuctx->ctx, task_event);
3798 ctx = task_event->task_ctx;
3800 ctxn = pmu->task_ctx_nr;
3803 ctx = rcu_dereference(current->perf_event_ctxp[ctxn]);
3806 perf_event_task_ctx(ctx, task_event);
3808 put_cpu_ptr(pmu->pmu_cpu_context);
3813 static void perf_event_task(struct task_struct *task,
3814 struct perf_event_context *task_ctx,
3817 struct perf_task_event task_event;
3819 if (!atomic_read(&nr_comm_events) &&
3820 !atomic_read(&nr_mmap_events) &&
3821 !atomic_read(&nr_task_events))
3824 task_event = (struct perf_task_event){
3826 .task_ctx = task_ctx,
3829 .type = new ? PERF_RECORD_FORK : PERF_RECORD_EXIT,
3831 .size = sizeof(task_event.event_id),
3837 .time = perf_clock(),
3841 perf_event_task_event(&task_event);
3844 void perf_event_fork(struct task_struct *task)
3846 perf_event_task(task, NULL, 1);
3853 struct perf_comm_event {
3854 struct task_struct *task;
3859 struct perf_event_header header;
3866 static void perf_event_comm_output(struct perf_event *event,
3867 struct perf_comm_event *comm_event)
3869 struct perf_output_handle handle;
3870 int size = comm_event->event_id.header.size;
3871 int ret = perf_output_begin(&handle, event, size, 0, 0);
3876 comm_event->event_id.pid = perf_event_pid(event, comm_event->task);
3877 comm_event->event_id.tid = perf_event_tid(event, comm_event->task);
3879 perf_output_put(&handle, comm_event->event_id);
3880 perf_output_copy(&handle, comm_event->comm,
3881 comm_event->comm_size);
3882 perf_output_end(&handle);
3885 static int perf_event_comm_match(struct perf_event *event)
3887 if (event->state < PERF_EVENT_STATE_INACTIVE)
3890 if (event->cpu != -1 && event->cpu != smp_processor_id())
3893 if (event->attr.comm)
3899 static void perf_event_comm_ctx(struct perf_event_context *ctx,
3900 struct perf_comm_event *comm_event)
3902 struct perf_event *event;
3904 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
3905 if (perf_event_comm_match(event))
3906 perf_event_comm_output(event, comm_event);
3910 static void perf_event_comm_event(struct perf_comm_event *comm_event)
3912 struct perf_cpu_context *cpuctx;
3913 struct perf_event_context *ctx;
3914 char comm[TASK_COMM_LEN];
3919 memset(comm, 0, sizeof(comm));
3920 strlcpy(comm, comm_event->task->comm, sizeof(comm));
3921 size = ALIGN(strlen(comm)+1, sizeof(u64));
3923 comm_event->comm = comm;
3924 comm_event->comm_size = size;
3926 comm_event->event_id.header.size = sizeof(comm_event->event_id) + size;
3929 list_for_each_entry_rcu(pmu, &pmus, entry) {
3930 cpuctx = get_cpu_ptr(pmu->pmu_cpu_context);
3931 perf_event_comm_ctx(&cpuctx->ctx, comm_event);
3933 ctxn = pmu->task_ctx_nr;
3937 ctx = rcu_dereference(current->perf_event_ctxp[ctxn]);
3939 perf_event_comm_ctx(ctx, comm_event);
3941 put_cpu_ptr(pmu->pmu_cpu_context);
3946 void perf_event_comm(struct task_struct *task)
3948 struct perf_comm_event comm_event;
3949 struct perf_event_context *ctx;
3952 for_each_task_context_nr(ctxn) {
3953 ctx = task->perf_event_ctxp[ctxn];
3957 perf_event_enable_on_exec(ctx);
3960 if (!atomic_read(&nr_comm_events))
3963 comm_event = (struct perf_comm_event){
3969 .type = PERF_RECORD_COMM,
3978 perf_event_comm_event(&comm_event);
3985 struct perf_mmap_event {
3986 struct vm_area_struct *vma;
3988 const char *file_name;
3992 struct perf_event_header header;
4002 static void perf_event_mmap_output(struct perf_event *event,
4003 struct perf_mmap_event *mmap_event)
4005 struct perf_output_handle handle;
4006 int size = mmap_event->event_id.header.size;
4007 int ret = perf_output_begin(&handle, event, size, 0, 0);
4012 mmap_event->event_id.pid = perf_event_pid(event, current);
4013 mmap_event->event_id.tid = perf_event_tid(event, current);
4015 perf_output_put(&handle, mmap_event->event_id);
4016 perf_output_copy(&handle, mmap_event->file_name,
4017 mmap_event->file_size);
4018 perf_output_end(&handle);
4021 static int perf_event_mmap_match(struct perf_event *event,
4022 struct perf_mmap_event *mmap_event,
4025 if (event->state < PERF_EVENT_STATE_INACTIVE)
4028 if (event->cpu != -1 && event->cpu != smp_processor_id())
4031 if ((!executable && event->attr.mmap_data) ||
4032 (executable && event->attr.mmap))
4038 static void perf_event_mmap_ctx(struct perf_event_context *ctx,
4039 struct perf_mmap_event *mmap_event,
4042 struct perf_event *event;
4044 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
4045 if (perf_event_mmap_match(event, mmap_event, executable))
4046 perf_event_mmap_output(event, mmap_event);
4050 static void perf_event_mmap_event(struct perf_mmap_event *mmap_event)
4052 struct perf_cpu_context *cpuctx;
4053 struct perf_event_context *ctx;
4054 struct vm_area_struct *vma = mmap_event->vma;
4055 struct file *file = vma->vm_file;
4063 memset(tmp, 0, sizeof(tmp));
4067 * d_path works from the end of the buffer backwards, so we
4068 * need to add enough zero bytes after the string to handle
4069 * the 64bit alignment we do later.
4071 buf = kzalloc(PATH_MAX + sizeof(u64), GFP_KERNEL);
4073 name = strncpy(tmp, "//enomem", sizeof(tmp));
4076 name = d_path(&file->f_path, buf, PATH_MAX);
4078 name = strncpy(tmp, "//toolong", sizeof(tmp));
4082 if (arch_vma_name(mmap_event->vma)) {
4083 name = strncpy(tmp, arch_vma_name(mmap_event->vma),
4089 name = strncpy(tmp, "[vdso]", sizeof(tmp));
4091 } else if (vma->vm_start <= vma->vm_mm->start_brk &&
4092 vma->vm_end >= vma->vm_mm->brk) {
4093 name = strncpy(tmp, "[heap]", sizeof(tmp));
4095 } else if (vma->vm_start <= vma->vm_mm->start_stack &&
4096 vma->vm_end >= vma->vm_mm->start_stack) {
4097 name = strncpy(tmp, "[stack]", sizeof(tmp));
4101 name = strncpy(tmp, "//anon", sizeof(tmp));
4106 size = ALIGN(strlen(name)+1, sizeof(u64));
4108 mmap_event->file_name = name;
4109 mmap_event->file_size = size;
4111 mmap_event->event_id.header.size = sizeof(mmap_event->event_id) + size;
4114 list_for_each_entry_rcu(pmu, &pmus, entry) {
4115 cpuctx = get_cpu_ptr(pmu->pmu_cpu_context);
4116 perf_event_mmap_ctx(&cpuctx->ctx, mmap_event,
4117 vma->vm_flags & VM_EXEC);
4119 ctxn = pmu->task_ctx_nr;
4123 ctx = rcu_dereference(current->perf_event_ctxp[ctxn]);
4125 perf_event_mmap_ctx(ctx, mmap_event,
4126 vma->vm_flags & VM_EXEC);
4129 put_cpu_ptr(pmu->pmu_cpu_context);
4136 void perf_event_mmap(struct vm_area_struct *vma)
4138 struct perf_mmap_event mmap_event;
4140 if (!atomic_read(&nr_mmap_events))
4143 mmap_event = (struct perf_mmap_event){
4149 .type = PERF_RECORD_MMAP,
4150 .misc = PERF_RECORD_MISC_USER,
4155 .start = vma->vm_start,
4156 .len = vma->vm_end - vma->vm_start,
4157 .pgoff = (u64)vma->vm_pgoff << PAGE_SHIFT,
4161 perf_event_mmap_event(&mmap_event);
4165 * IRQ throttle logging
4168 static void perf_log_throttle(struct perf_event *event, int enable)
4170 struct perf_output_handle handle;
4174 struct perf_event_header header;
4178 } throttle_event = {
4180 .type = PERF_RECORD_THROTTLE,
4182 .size = sizeof(throttle_event),
4184 .time = perf_clock(),
4185 .id = primary_event_id(event),
4186 .stream_id = event->id,
4190 throttle_event.header.type = PERF_RECORD_UNTHROTTLE;
4192 ret = perf_output_begin(&handle, event, sizeof(throttle_event), 1, 0);
4196 perf_output_put(&handle, throttle_event);
4197 perf_output_end(&handle);
4201 * Generic event overflow handling, sampling.
4204 static int __perf_event_overflow(struct perf_event *event, int nmi,
4205 int throttle, struct perf_sample_data *data,
4206 struct pt_regs *regs)
4208 int events = atomic_read(&event->event_limit);
4209 struct hw_perf_event *hwc = &event->hw;
4215 if (hwc->interrupts != MAX_INTERRUPTS) {
4217 if (HZ * hwc->interrupts >
4218 (u64)sysctl_perf_event_sample_rate) {
4219 hwc->interrupts = MAX_INTERRUPTS;
4220 perf_log_throttle(event, 0);
4225 * Keep re-disabling events even though on the previous
4226 * pass we disabled it - just in case we raced with a
4227 * sched-in and the event got enabled again:
4233 if (event->attr.freq) {
4234 u64 now = perf_clock();
4235 s64 delta = now - hwc->freq_time_stamp;
4237 hwc->freq_time_stamp = now;
4239 if (delta > 0 && delta < 2*TICK_NSEC)
4240 perf_adjust_period(event, delta, hwc->last_period);
4244 * XXX event_limit might not quite work as expected on inherited
4248 event->pending_kill = POLL_IN;
4249 if (events && atomic_dec_and_test(&event->event_limit)) {
4251 event->pending_kill = POLL_HUP;
4253 event->pending_disable = 1;
4254 irq_work_queue(&event->pending);
4256 perf_event_disable(event);
4259 if (event->overflow_handler)
4260 event->overflow_handler(event, nmi, data, regs);
4262 perf_event_output(event, nmi, data, regs);
4267 int perf_event_overflow(struct perf_event *event, int nmi,
4268 struct perf_sample_data *data,
4269 struct pt_regs *regs)
4271 return __perf_event_overflow(event, nmi, 1, data, regs);
4275 * Generic software event infrastructure
4278 struct swevent_htable {
4279 struct swevent_hlist *swevent_hlist;
4280 struct mutex hlist_mutex;
4283 /* Recursion avoidance in each contexts */
4284 int recursion[PERF_NR_CONTEXTS];
4287 static DEFINE_PER_CPU(struct swevent_htable, swevent_htable);
4290 * We directly increment event->count and keep a second value in
4291 * event->hw.period_left to count intervals. This period event
4292 * is kept in the range [-sample_period, 0] so that we can use the
4296 static u64 perf_swevent_set_period(struct perf_event *event)
4298 struct hw_perf_event *hwc = &event->hw;
4299 u64 period = hwc->last_period;
4303 hwc->last_period = hwc->sample_period;
4306 old = val = local64_read(&hwc->period_left);
4310 nr = div64_u64(period + val, period);
4311 offset = nr * period;
4313 if (local64_cmpxchg(&hwc->period_left, old, val) != old)
4319 static void perf_swevent_overflow(struct perf_event *event, u64 overflow,
4320 int nmi, struct perf_sample_data *data,
4321 struct pt_regs *regs)
4323 struct hw_perf_event *hwc = &event->hw;
4326 data->period = event->hw.last_period;
4328 overflow = perf_swevent_set_period(event);
4330 if (hwc->interrupts == MAX_INTERRUPTS)
4333 for (; overflow; overflow--) {
4334 if (__perf_event_overflow(event, nmi, throttle,
4337 * We inhibit the overflow from happening when
4338 * hwc->interrupts == MAX_INTERRUPTS.
4346 static void perf_swevent_event(struct perf_event *event, u64 nr,
4347 int nmi, struct perf_sample_data *data,
4348 struct pt_regs *regs)
4350 struct hw_perf_event *hwc = &event->hw;
4352 local64_add(nr, &event->count);
4357 if (!hwc->sample_period)
4360 if (nr == 1 && hwc->sample_period == 1 && !event->attr.freq)
4361 return perf_swevent_overflow(event, 1, nmi, data, regs);
4363 if (local64_add_negative(nr, &hwc->period_left))
4366 perf_swevent_overflow(event, 0, nmi, data, regs);
4369 static int perf_exclude_event(struct perf_event *event,
4370 struct pt_regs *regs)
4372 if (event->hw.state & PERF_HES_STOPPED)
4376 if (event->attr.exclude_user && user_mode(regs))
4379 if (event->attr.exclude_kernel && !user_mode(regs))
4386 static int perf_swevent_match(struct perf_event *event,
4387 enum perf_type_id type,
4389 struct perf_sample_data *data,
4390 struct pt_regs *regs)
4392 if (event->attr.type != type)
4395 if (event->attr.config != event_id)
4398 if (perf_exclude_event(event, regs))
4404 static inline u64 swevent_hash(u64 type, u32 event_id)
4406 u64 val = event_id | (type << 32);
4408 return hash_64(val, SWEVENT_HLIST_BITS);
4411 static inline struct hlist_head *
4412 __find_swevent_head(struct swevent_hlist *hlist, u64 type, u32 event_id)
4414 u64 hash = swevent_hash(type, event_id);
4416 return &hlist->heads[hash];
4419 /* For the read side: events when they trigger */
4420 static inline struct hlist_head *
4421 find_swevent_head_rcu(struct swevent_htable *swhash, u64 type, u32 event_id)
4423 struct swevent_hlist *hlist;
4425 hlist = rcu_dereference(swhash->swevent_hlist);
4429 return __find_swevent_head(hlist, type, event_id);
4432 /* For the event head insertion and removal in the hlist */
4433 static inline struct hlist_head *
4434 find_swevent_head(struct swevent_htable *swhash, struct perf_event *event)
4436 struct swevent_hlist *hlist;
4437 u32 event_id = event->attr.config;
4438 u64 type = event->attr.type;
4441 * Event scheduling is always serialized against hlist allocation
4442 * and release. Which makes the protected version suitable here.
4443 * The context lock guarantees that.
4445 hlist = rcu_dereference_protected(swhash->swevent_hlist,
4446 lockdep_is_held(&event->ctx->lock));
4450 return __find_swevent_head(hlist, type, event_id);
4453 static void do_perf_sw_event(enum perf_type_id type, u32 event_id,
4455 struct perf_sample_data *data,
4456 struct pt_regs *regs)
4458 struct swevent_htable *swhash = &__get_cpu_var(swevent_htable);
4459 struct perf_event *event;
4460 struct hlist_node *node;
4461 struct hlist_head *head;
4464 head = find_swevent_head_rcu(swhash, type, event_id);
4468 hlist_for_each_entry_rcu(event, node, head, hlist_entry) {
4469 if (perf_swevent_match(event, type, event_id, data, regs))
4470 perf_swevent_event(event, nr, nmi, data, regs);
4476 int perf_swevent_get_recursion_context(void)
4478 struct swevent_htable *swhash = &__get_cpu_var(swevent_htable);
4480 return get_recursion_context(swhash->recursion);
4482 EXPORT_SYMBOL_GPL(perf_swevent_get_recursion_context);
4484 void inline perf_swevent_put_recursion_context(int rctx)
4486 struct swevent_htable *swhash = &__get_cpu_var(swevent_htable);
4488 put_recursion_context(swhash->recursion, rctx);
4491 void __perf_sw_event(u32 event_id, u64 nr, int nmi,
4492 struct pt_regs *regs, u64 addr)
4494 struct perf_sample_data data;
4497 preempt_disable_notrace();
4498 rctx = perf_swevent_get_recursion_context();
4502 perf_sample_data_init(&data, addr);
4504 do_perf_sw_event(PERF_TYPE_SOFTWARE, event_id, nr, nmi, &data, regs);
4506 perf_swevent_put_recursion_context(rctx);
4507 preempt_enable_notrace();
4510 static void perf_swevent_read(struct perf_event *event)
4514 static int perf_swevent_add(struct perf_event *event, int flags)
4516 struct swevent_htable *swhash = &__get_cpu_var(swevent_htable);
4517 struct hw_perf_event *hwc = &event->hw;
4518 struct hlist_head *head;
4520 if (hwc->sample_period) {
4521 hwc->last_period = hwc->sample_period;
4522 perf_swevent_set_period(event);
4525 hwc->state = !(flags & PERF_EF_START);
4527 head = find_swevent_head(swhash, event);
4528 if (WARN_ON_ONCE(!head))
4531 hlist_add_head_rcu(&event->hlist_entry, head);
4536 static void perf_swevent_del(struct perf_event *event, int flags)
4538 hlist_del_rcu(&event->hlist_entry);
4541 static void perf_swevent_start(struct perf_event *event, int flags)
4543 event->hw.state = 0;
4546 static void perf_swevent_stop(struct perf_event *event, int flags)
4548 event->hw.state = PERF_HES_STOPPED;
4551 /* Deref the hlist from the update side */
4552 static inline struct swevent_hlist *
4553 swevent_hlist_deref(struct swevent_htable *swhash)
4555 return rcu_dereference_protected(swhash->swevent_hlist,
4556 lockdep_is_held(&swhash->hlist_mutex));
4559 static void swevent_hlist_release_rcu(struct rcu_head *rcu_head)
4561 struct swevent_hlist *hlist;
4563 hlist = container_of(rcu_head, struct swevent_hlist, rcu_head);
4567 static void swevent_hlist_release(struct swevent_htable *swhash)
4569 struct swevent_hlist *hlist = swevent_hlist_deref(swhash);
4574 rcu_assign_pointer(swhash->swevent_hlist, NULL);
4575 call_rcu(&hlist->rcu_head, swevent_hlist_release_rcu);
4578 static void swevent_hlist_put_cpu(struct perf_event *event, int cpu)
4580 struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
4582 mutex_lock(&swhash->hlist_mutex);
4584 if (!--swhash->hlist_refcount)
4585 swevent_hlist_release(swhash);
4587 mutex_unlock(&swhash->hlist_mutex);
4590 static void swevent_hlist_put(struct perf_event *event)
4594 if (event->cpu != -1) {
4595 swevent_hlist_put_cpu(event, event->cpu);
4599 for_each_possible_cpu(cpu)
4600 swevent_hlist_put_cpu(event, cpu);
4603 static int swevent_hlist_get_cpu(struct perf_event *event, int cpu)
4605 struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
4608 mutex_lock(&swhash->hlist_mutex);
4610 if (!swevent_hlist_deref(swhash) && cpu_online(cpu)) {
4611 struct swevent_hlist *hlist;
4613 hlist = kzalloc(sizeof(*hlist), GFP_KERNEL);
4618 rcu_assign_pointer(swhash->swevent_hlist, hlist);
4620 swhash->hlist_refcount++;
4622 mutex_unlock(&swhash->hlist_mutex);
4627 static int swevent_hlist_get(struct perf_event *event)
4630 int cpu, failed_cpu;
4632 if (event->cpu != -1)
4633 return swevent_hlist_get_cpu(event, event->cpu);
4636 for_each_possible_cpu(cpu) {
4637 err = swevent_hlist_get_cpu(event, cpu);
4647 for_each_possible_cpu(cpu) {
4648 if (cpu == failed_cpu)
4650 swevent_hlist_put_cpu(event, cpu);
4657 atomic_t perf_swevent_enabled[PERF_COUNT_SW_MAX];
4659 static void sw_perf_event_destroy(struct perf_event *event)
4661 u64 event_id = event->attr.config;
4663 WARN_ON(event->parent);
4665 jump_label_dec(&perf_swevent_enabled[event_id]);
4666 swevent_hlist_put(event);
4669 static int perf_swevent_init(struct perf_event *event)
4671 int event_id = event->attr.config;
4673 if (event->attr.type != PERF_TYPE_SOFTWARE)
4677 case PERF_COUNT_SW_CPU_CLOCK:
4678 case PERF_COUNT_SW_TASK_CLOCK:
4685 if (event_id > PERF_COUNT_SW_MAX)
4688 if (!event->parent) {
4691 err = swevent_hlist_get(event);
4695 jump_label_inc(&perf_swevent_enabled[event_id]);
4696 event->destroy = sw_perf_event_destroy;
4702 static struct pmu perf_swevent = {
4703 .task_ctx_nr = perf_sw_context,
4705 .event_init = perf_swevent_init,
4706 .add = perf_swevent_add,
4707 .del = perf_swevent_del,
4708 .start = perf_swevent_start,
4709 .stop = perf_swevent_stop,
4710 .read = perf_swevent_read,
4713 #ifdef CONFIG_EVENT_TRACING
4715 static int perf_tp_filter_match(struct perf_event *event,
4716 struct perf_sample_data *data)
4718 void *record = data->raw->data;
4720 if (likely(!event->filter) || filter_match_preds(event->filter, record))
4725 static int perf_tp_event_match(struct perf_event *event,
4726 struct perf_sample_data *data,
4727 struct pt_regs *regs)
4730 * All tracepoints are from kernel-space.
4732 if (event->attr.exclude_kernel)
4735 if (!perf_tp_filter_match(event, data))
4741 void perf_tp_event(u64 addr, u64 count, void *record, int entry_size,
4742 struct pt_regs *regs, struct hlist_head *head, int rctx)
4744 struct perf_sample_data data;
4745 struct perf_event *event;
4746 struct hlist_node *node;
4748 struct perf_raw_record raw = {
4753 perf_sample_data_init(&data, addr);
4756 hlist_for_each_entry_rcu(event, node, head, hlist_entry) {
4757 if (perf_tp_event_match(event, &data, regs))
4758 perf_swevent_event(event, count, 1, &data, regs);
4761 perf_swevent_put_recursion_context(rctx);
4763 EXPORT_SYMBOL_GPL(perf_tp_event);
4765 static void tp_perf_event_destroy(struct perf_event *event)
4767 perf_trace_destroy(event);
4770 static int perf_tp_event_init(struct perf_event *event)
4774 if (event->attr.type != PERF_TYPE_TRACEPOINT)
4778 * Raw tracepoint data is a severe data leak, only allow root to
4781 if ((event->attr.sample_type & PERF_SAMPLE_RAW) &&
4782 perf_paranoid_tracepoint_raw() &&
4783 !capable(CAP_SYS_ADMIN))
4786 err = perf_trace_init(event);
4790 event->destroy = tp_perf_event_destroy;
4795 static struct pmu perf_tracepoint = {
4796 .task_ctx_nr = perf_sw_context,
4798 .event_init = perf_tp_event_init,
4799 .add = perf_trace_add,
4800 .del = perf_trace_del,
4801 .start = perf_swevent_start,
4802 .stop = perf_swevent_stop,
4803 .read = perf_swevent_read,
4806 static inline void perf_tp_register(void)
4808 perf_pmu_register(&perf_tracepoint);
4811 static int perf_event_set_filter(struct perf_event *event, void __user *arg)
4816 if (event->attr.type != PERF_TYPE_TRACEPOINT)
4819 filter_str = strndup_user(arg, PAGE_SIZE);
4820 if (IS_ERR(filter_str))
4821 return PTR_ERR(filter_str);
4823 ret = ftrace_profile_set_filter(event, event->attr.config, filter_str);
4829 static void perf_event_free_filter(struct perf_event *event)
4831 ftrace_profile_free_filter(event);
4836 static inline void perf_tp_register(void)
4840 static int perf_event_set_filter(struct perf_event *event, void __user *arg)
4845 static void perf_event_free_filter(struct perf_event *event)
4849 #endif /* CONFIG_EVENT_TRACING */
4851 #ifdef CONFIG_HAVE_HW_BREAKPOINT
4852 void perf_bp_event(struct perf_event *bp, void *data)
4854 struct perf_sample_data sample;
4855 struct pt_regs *regs = data;
4857 perf_sample_data_init(&sample, bp->attr.bp_addr);
4859 if (!bp->hw.state && !perf_exclude_event(bp, regs))
4860 perf_swevent_event(bp, 1, 1, &sample, regs);
4865 * hrtimer based swevent callback
4868 static enum hrtimer_restart perf_swevent_hrtimer(struct hrtimer *hrtimer)
4870 enum hrtimer_restart ret = HRTIMER_RESTART;
4871 struct perf_sample_data data;
4872 struct pt_regs *regs;
4873 struct perf_event *event;
4876 event = container_of(hrtimer, struct perf_event, hw.hrtimer);
4877 event->pmu->read(event);
4879 perf_sample_data_init(&data, 0);
4880 data.period = event->hw.last_period;
4881 regs = get_irq_regs();
4883 if (regs && !perf_exclude_event(event, regs)) {
4884 if (!(event->attr.exclude_idle && current->pid == 0))
4885 if (perf_event_overflow(event, 0, &data, regs))
4886 ret = HRTIMER_NORESTART;
4889 period = max_t(u64, 10000, event->hw.sample_period);
4890 hrtimer_forward_now(hrtimer, ns_to_ktime(period));
4895 static void perf_swevent_start_hrtimer(struct perf_event *event)
4897 struct hw_perf_event *hwc = &event->hw;
4899 hrtimer_init(&hwc->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
4900 hwc->hrtimer.function = perf_swevent_hrtimer;
4901 if (hwc->sample_period) {
4902 s64 period = local64_read(&hwc->period_left);
4908 local64_set(&hwc->period_left, 0);
4910 period = max_t(u64, 10000, hwc->sample_period);
4912 __hrtimer_start_range_ns(&hwc->hrtimer,
4913 ns_to_ktime(period), 0,
4914 HRTIMER_MODE_REL_PINNED, 0);
4918 static void perf_swevent_cancel_hrtimer(struct perf_event *event)
4920 struct hw_perf_event *hwc = &event->hw;
4922 if (hwc->sample_period) {
4923 ktime_t remaining = hrtimer_get_remaining(&hwc->hrtimer);
4924 local64_set(&hwc->period_left, ktime_to_ns(remaining));
4926 hrtimer_cancel(&hwc->hrtimer);
4931 * Software event: cpu wall time clock
4934 static void cpu_clock_event_update(struct perf_event *event)
4939 now = local_clock();
4940 prev = local64_xchg(&event->hw.prev_count, now);
4941 local64_add(now - prev, &event->count);
4944 static void cpu_clock_event_start(struct perf_event *event, int flags)
4946 local64_set(&event->hw.prev_count, local_clock());
4947 perf_swevent_start_hrtimer(event);
4950 static void cpu_clock_event_stop(struct perf_event *event, int flags)
4952 perf_swevent_cancel_hrtimer(event);
4953 cpu_clock_event_update(event);
4956 static int cpu_clock_event_add(struct perf_event *event, int flags)
4958 if (flags & PERF_EF_START)
4959 cpu_clock_event_start(event, flags);
4964 static void cpu_clock_event_del(struct perf_event *event, int flags)
4966 cpu_clock_event_stop(event, flags);
4969 static void cpu_clock_event_read(struct perf_event *event)
4971 cpu_clock_event_update(event);
4974 static int cpu_clock_event_init(struct perf_event *event)
4976 if (event->attr.type != PERF_TYPE_SOFTWARE)
4979 if (event->attr.config != PERF_COUNT_SW_CPU_CLOCK)
4985 static struct pmu perf_cpu_clock = {
4986 .task_ctx_nr = perf_sw_context,
4988 .event_init = cpu_clock_event_init,
4989 .add = cpu_clock_event_add,
4990 .del = cpu_clock_event_del,
4991 .start = cpu_clock_event_start,
4992 .stop = cpu_clock_event_stop,
4993 .read = cpu_clock_event_read,
4997 * Software event: task time clock
5000 static void task_clock_event_update(struct perf_event *event, u64 now)
5005 prev = local64_xchg(&event->hw.prev_count, now);
5007 local64_add(delta, &event->count);
5010 static void task_clock_event_start(struct perf_event *event, int flags)
5012 local64_set(&event->hw.prev_count, event->ctx->time);
5013 perf_swevent_start_hrtimer(event);
5016 static void task_clock_event_stop(struct perf_event *event, int flags)
5018 perf_swevent_cancel_hrtimer(event);
5019 task_clock_event_update(event, event->ctx->time);
5022 static int task_clock_event_add(struct perf_event *event, int flags)
5024 if (flags & PERF_EF_START)
5025 task_clock_event_start(event, flags);
5030 static void task_clock_event_del(struct perf_event *event, int flags)
5032 task_clock_event_stop(event, PERF_EF_UPDATE);
5035 static void task_clock_event_read(struct perf_event *event)
5040 update_context_time(event->ctx);
5041 time = event->ctx->time;
5043 u64 now = perf_clock();
5044 u64 delta = now - event->ctx->timestamp;
5045 time = event->ctx->time + delta;
5048 task_clock_event_update(event, time);
5051 static int task_clock_event_init(struct perf_event *event)
5053 if (event->attr.type != PERF_TYPE_SOFTWARE)
5056 if (event->attr.config != PERF_COUNT_SW_TASK_CLOCK)
5062 static struct pmu perf_task_clock = {
5063 .task_ctx_nr = perf_sw_context,
5065 .event_init = task_clock_event_init,
5066 .add = task_clock_event_add,
5067 .del = task_clock_event_del,
5068 .start = task_clock_event_start,
5069 .stop = task_clock_event_stop,
5070 .read = task_clock_event_read,
5073 static void perf_pmu_nop_void(struct pmu *pmu)
5077 static int perf_pmu_nop_int(struct pmu *pmu)
5082 static void perf_pmu_start_txn(struct pmu *pmu)
5084 perf_pmu_disable(pmu);
5087 static int perf_pmu_commit_txn(struct pmu *pmu)
5089 perf_pmu_enable(pmu);
5093 static void perf_pmu_cancel_txn(struct pmu *pmu)
5095 perf_pmu_enable(pmu);
5099 * Ensures all contexts with the same task_ctx_nr have the same
5100 * pmu_cpu_context too.
5102 static void *find_pmu_context(int ctxn)
5109 list_for_each_entry(pmu, &pmus, entry) {
5110 if (pmu->task_ctx_nr == ctxn)
5111 return pmu->pmu_cpu_context;
5117 static void free_pmu_context(void * __percpu cpu_context)
5121 mutex_lock(&pmus_lock);
5123 * Like a real lame refcount.
5125 list_for_each_entry(pmu, &pmus, entry) {
5126 if (pmu->pmu_cpu_context == cpu_context)
5130 free_percpu(cpu_context);
5132 mutex_unlock(&pmus_lock);
5135 int perf_pmu_register(struct pmu *pmu)
5139 mutex_lock(&pmus_lock);
5141 pmu->pmu_disable_count = alloc_percpu(int);
5142 if (!pmu->pmu_disable_count)
5145 pmu->pmu_cpu_context = find_pmu_context(pmu->task_ctx_nr);
5146 if (pmu->pmu_cpu_context)
5147 goto got_cpu_context;
5149 pmu->pmu_cpu_context = alloc_percpu(struct perf_cpu_context);
5150 if (!pmu->pmu_cpu_context)
5153 for_each_possible_cpu(cpu) {
5154 struct perf_cpu_context *cpuctx;
5156 cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu);
5157 __perf_event_init_context(&cpuctx->ctx);
5158 cpuctx->ctx.type = cpu_context;
5159 cpuctx->ctx.pmu = pmu;
5160 cpuctx->jiffies_interval = 1;
5161 INIT_LIST_HEAD(&cpuctx->rotation_list);
5165 if (!pmu->start_txn) {
5166 if (pmu->pmu_enable) {
5168 * If we have pmu_enable/pmu_disable calls, install
5169 * transaction stubs that use that to try and batch
5170 * hardware accesses.
5172 pmu->start_txn = perf_pmu_start_txn;
5173 pmu->commit_txn = perf_pmu_commit_txn;
5174 pmu->cancel_txn = perf_pmu_cancel_txn;
5176 pmu->start_txn = perf_pmu_nop_void;
5177 pmu->commit_txn = perf_pmu_nop_int;
5178 pmu->cancel_txn = perf_pmu_nop_void;
5182 if (!pmu->pmu_enable) {
5183 pmu->pmu_enable = perf_pmu_nop_void;
5184 pmu->pmu_disable = perf_pmu_nop_void;
5187 list_add_rcu(&pmu->entry, &pmus);
5190 mutex_unlock(&pmus_lock);
5195 free_percpu(pmu->pmu_disable_count);
5199 void perf_pmu_unregister(struct pmu *pmu)
5201 mutex_lock(&pmus_lock);
5202 list_del_rcu(&pmu->entry);
5203 mutex_unlock(&pmus_lock);
5206 * We dereference the pmu list under both SRCU and regular RCU, so
5207 * synchronize against both of those.
5209 synchronize_srcu(&pmus_srcu);
5212 free_percpu(pmu->pmu_disable_count);
5213 free_pmu_context(pmu->pmu_cpu_context);
5216 struct pmu *perf_init_event(struct perf_event *event)
5218 struct pmu *pmu = NULL;
5221 idx = srcu_read_lock(&pmus_srcu);
5222 list_for_each_entry_rcu(pmu, &pmus, entry) {
5223 int ret = pmu->event_init(event);
5227 if (ret != -ENOENT) {
5232 pmu = ERR_PTR(-ENOENT);
5234 srcu_read_unlock(&pmus_srcu, idx);
5240 * Allocate and initialize a event structure
5242 static struct perf_event *
5243 perf_event_alloc(struct perf_event_attr *attr, int cpu,
5244 struct task_struct *task,
5245 struct perf_event *group_leader,
5246 struct perf_event *parent_event,
5247 perf_overflow_handler_t overflow_handler)
5250 struct perf_event *event;
5251 struct hw_perf_event *hwc;
5254 event = kzalloc(sizeof(*event), GFP_KERNEL);
5256 return ERR_PTR(-ENOMEM);
5259 * Single events are their own group leaders, with an
5260 * empty sibling list:
5263 group_leader = event;
5265 mutex_init(&event->child_mutex);
5266 INIT_LIST_HEAD(&event->child_list);
5268 INIT_LIST_HEAD(&event->group_entry);
5269 INIT_LIST_HEAD(&event->event_entry);
5270 INIT_LIST_HEAD(&event->sibling_list);
5271 init_waitqueue_head(&event->waitq);
5272 init_irq_work(&event->pending, perf_pending_event);
5274 mutex_init(&event->mmap_mutex);
5277 event->attr = *attr;
5278 event->group_leader = group_leader;
5282 event->parent = parent_event;
5284 event->ns = get_pid_ns(current->nsproxy->pid_ns);
5285 event->id = atomic64_inc_return(&perf_event_id);
5287 event->state = PERF_EVENT_STATE_INACTIVE;
5290 event->attach_state = PERF_ATTACH_TASK;
5291 #ifdef CONFIG_HAVE_HW_BREAKPOINT
5293 * hw_breakpoint is a bit difficult here..
5295 if (attr->type == PERF_TYPE_BREAKPOINT)
5296 event->hw.bp_target = task;
5300 if (!overflow_handler && parent_event)
5301 overflow_handler = parent_event->overflow_handler;
5303 event->overflow_handler = overflow_handler;
5306 event->state = PERF_EVENT_STATE_OFF;
5311 hwc->sample_period = attr->sample_period;
5312 if (attr->freq && attr->sample_freq)
5313 hwc->sample_period = 1;
5314 hwc->last_period = hwc->sample_period;
5316 local64_set(&hwc->period_left, hwc->sample_period);
5319 * we currently do not support PERF_FORMAT_GROUP on inherited events
5321 if (attr->inherit && (attr->read_format & PERF_FORMAT_GROUP))
5324 pmu = perf_init_event(event);
5330 else if (IS_ERR(pmu))
5335 put_pid_ns(event->ns);
5337 return ERR_PTR(err);
5342 if (!event->parent) {
5343 if (event->attach_state & PERF_ATTACH_TASK)
5344 jump_label_inc(&perf_task_events);
5345 if (event->attr.mmap || event->attr.mmap_data)
5346 atomic_inc(&nr_mmap_events);
5347 if (event->attr.comm)
5348 atomic_inc(&nr_comm_events);
5349 if (event->attr.task)
5350 atomic_inc(&nr_task_events);
5351 if (event->attr.sample_type & PERF_SAMPLE_CALLCHAIN) {
5352 err = get_callchain_buffers();
5355 return ERR_PTR(err);
5363 static int perf_copy_attr(struct perf_event_attr __user *uattr,
5364 struct perf_event_attr *attr)
5369 if (!access_ok(VERIFY_WRITE, uattr, PERF_ATTR_SIZE_VER0))
5373 * zero the full structure, so that a short copy will be nice.
5375 memset(attr, 0, sizeof(*attr));
5377 ret = get_user(size, &uattr->size);
5381 if (size > PAGE_SIZE) /* silly large */
5384 if (!size) /* abi compat */
5385 size = PERF_ATTR_SIZE_VER0;
5387 if (size < PERF_ATTR_SIZE_VER0)
5391 * If we're handed a bigger struct than we know of,
5392 * ensure all the unknown bits are 0 - i.e. new
5393 * user-space does not rely on any kernel feature
5394 * extensions we dont know about yet.
5396 if (size > sizeof(*attr)) {
5397 unsigned char __user *addr;
5398 unsigned char __user *end;
5401 addr = (void __user *)uattr + sizeof(*attr);
5402 end = (void __user *)uattr + size;
5404 for (; addr < end; addr++) {
5405 ret = get_user(val, addr);
5411 size = sizeof(*attr);
5414 ret = copy_from_user(attr, uattr, size);
5419 * If the type exists, the corresponding creation will verify
5422 if (attr->type >= PERF_TYPE_MAX)
5425 if (attr->__reserved_1)
5428 if (attr->sample_type & ~(PERF_SAMPLE_MAX-1))
5431 if (attr->read_format & ~(PERF_FORMAT_MAX-1))
5438 put_user(sizeof(*attr), &uattr->size);
5444 perf_event_set_output(struct perf_event *event, struct perf_event *output_event)
5446 struct perf_buffer *buffer = NULL, *old_buffer = NULL;
5452 /* don't allow circular references */
5453 if (event == output_event)
5457 * Don't allow cross-cpu buffers
5459 if (output_event->cpu != event->cpu)
5463 * If its not a per-cpu buffer, it must be the same task.
5465 if (output_event->cpu == -1 && output_event->ctx != event->ctx)
5469 mutex_lock(&event->mmap_mutex);
5470 /* Can't redirect output if we've got an active mmap() */
5471 if (atomic_read(&event->mmap_count))
5475 /* get the buffer we want to redirect to */
5476 buffer = perf_buffer_get(output_event);
5481 old_buffer = event->buffer;
5482 rcu_assign_pointer(event->buffer, buffer);
5485 mutex_unlock(&event->mmap_mutex);
5488 perf_buffer_put(old_buffer);
5494 * sys_perf_event_open - open a performance event, associate it to a task/cpu
5496 * @attr_uptr: event_id type attributes for monitoring/sampling
5499 * @group_fd: group leader event fd
5501 SYSCALL_DEFINE5(perf_event_open,
5502 struct perf_event_attr __user *, attr_uptr,
5503 pid_t, pid, int, cpu, int, group_fd, unsigned long, flags)
5505 struct perf_event *group_leader = NULL, *output_event = NULL;
5506 struct perf_event *event, *sibling;
5507 struct perf_event_attr attr;
5508 struct perf_event_context *ctx;
5509 struct file *event_file = NULL;
5510 struct file *group_file = NULL;
5511 struct task_struct *task = NULL;
5515 int fput_needed = 0;
5518 /* for future expandability... */
5519 if (flags & ~(PERF_FLAG_FD_NO_GROUP | PERF_FLAG_FD_OUTPUT))
5522 err = perf_copy_attr(attr_uptr, &attr);
5526 if (!attr.exclude_kernel) {
5527 if (perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN))
5532 if (attr.sample_freq > sysctl_perf_event_sample_rate)
5536 event_fd = get_unused_fd_flags(O_RDWR);
5540 if (group_fd != -1) {
5541 group_leader = perf_fget_light(group_fd, &fput_needed);
5542 if (IS_ERR(group_leader)) {
5543 err = PTR_ERR(group_leader);
5546 group_file = group_leader->filp;
5547 if (flags & PERF_FLAG_FD_OUTPUT)
5548 output_event = group_leader;
5549 if (flags & PERF_FLAG_FD_NO_GROUP)
5550 group_leader = NULL;
5554 task = find_lively_task_by_vpid(pid);
5556 err = PTR_ERR(task);
5561 event = perf_event_alloc(&attr, cpu, task, group_leader, NULL, NULL);
5562 if (IS_ERR(event)) {
5563 err = PTR_ERR(event);
5568 * Special case software events and allow them to be part of
5569 * any hardware group.
5574 (is_software_event(event) != is_software_event(group_leader))) {
5575 if (is_software_event(event)) {
5577 * If event and group_leader are not both a software
5578 * event, and event is, then group leader is not.
5580 * Allow the addition of software events to !software
5581 * groups, this is safe because software events never
5584 pmu = group_leader->pmu;
5585 } else if (is_software_event(group_leader) &&
5586 (group_leader->group_flags & PERF_GROUP_SOFTWARE)) {
5588 * In case the group is a pure software group, and we
5589 * try to add a hardware event, move the whole group to
5590 * the hardware context.
5597 * Get the target context (task or percpu):
5599 ctx = find_get_context(pmu, task, cpu);
5606 * Look up the group leader (we will attach this event to it):
5612 * Do not allow a recursive hierarchy (this new sibling
5613 * becoming part of another group-sibling):
5615 if (group_leader->group_leader != group_leader)
5618 * Do not allow to attach to a group in a different
5619 * task or CPU context:
5622 if (group_leader->ctx->type != ctx->type)
5625 if (group_leader->ctx != ctx)
5630 * Only a group leader can be exclusive or pinned
5632 if (attr.exclusive || attr.pinned)
5637 err = perf_event_set_output(event, output_event);
5642 event_file = anon_inode_getfile("[perf_event]", &perf_fops, event, O_RDWR);
5643 if (IS_ERR(event_file)) {
5644 err = PTR_ERR(event_file);
5649 struct perf_event_context *gctx = group_leader->ctx;
5651 mutex_lock(&gctx->mutex);
5652 perf_event_remove_from_context(group_leader);
5653 list_for_each_entry(sibling, &group_leader->sibling_list,
5655 perf_event_remove_from_context(sibling);
5658 mutex_unlock(&gctx->mutex);
5662 event->filp = event_file;
5663 WARN_ON_ONCE(ctx->parent_ctx);
5664 mutex_lock(&ctx->mutex);
5667 perf_install_in_context(ctx, group_leader, cpu);
5669 list_for_each_entry(sibling, &group_leader->sibling_list,
5671 perf_install_in_context(ctx, sibling, cpu);
5676 perf_install_in_context(ctx, event, cpu);
5678 mutex_unlock(&ctx->mutex);
5680 event->owner = current;
5681 get_task_struct(current);
5682 mutex_lock(¤t->perf_event_mutex);
5683 list_add_tail(&event->owner_entry, ¤t->perf_event_list);
5684 mutex_unlock(¤t->perf_event_mutex);
5687 * Drop the reference on the group_event after placing the
5688 * new event on the sibling_list. This ensures destruction
5689 * of the group leader will find the pointer to itself in
5690 * perf_group_detach().
5692 fput_light(group_file, fput_needed);
5693 fd_install(event_fd, event_file);
5702 put_task_struct(task);
5704 fput_light(group_file, fput_needed);
5706 put_unused_fd(event_fd);
5711 * perf_event_create_kernel_counter
5713 * @attr: attributes of the counter to create
5714 * @cpu: cpu in which the counter is bound
5715 * @task: task to profile (NULL for percpu)
5718 perf_event_create_kernel_counter(struct perf_event_attr *attr, int cpu,
5719 struct task_struct *task,
5720 perf_overflow_handler_t overflow_handler)
5722 struct perf_event_context *ctx;
5723 struct perf_event *event;
5727 * Get the target context (task or percpu):
5730 event = perf_event_alloc(attr, cpu, task, NULL, NULL, overflow_handler);
5731 if (IS_ERR(event)) {
5732 err = PTR_ERR(event);
5736 ctx = find_get_context(event->pmu, task, cpu);
5743 WARN_ON_ONCE(ctx->parent_ctx);
5744 mutex_lock(&ctx->mutex);
5745 perf_install_in_context(ctx, event, cpu);
5747 mutex_unlock(&ctx->mutex);
5749 event->owner = current;
5750 get_task_struct(current);
5751 mutex_lock(¤t->perf_event_mutex);
5752 list_add_tail(&event->owner_entry, ¤t->perf_event_list);
5753 mutex_unlock(¤t->perf_event_mutex);
5760 return ERR_PTR(err);
5762 EXPORT_SYMBOL_GPL(perf_event_create_kernel_counter);
5764 static void sync_child_event(struct perf_event *child_event,
5765 struct task_struct *child)
5767 struct perf_event *parent_event = child_event->parent;
5770 if (child_event->attr.inherit_stat)
5771 perf_event_read_event(child_event, child);
5773 child_val = perf_event_count(child_event);
5776 * Add back the child's count to the parent's count:
5778 atomic64_add(child_val, &parent_event->child_count);
5779 atomic64_add(child_event->total_time_enabled,
5780 &parent_event->child_total_time_enabled);
5781 atomic64_add(child_event->total_time_running,
5782 &parent_event->child_total_time_running);
5785 * Remove this event from the parent's list
5787 WARN_ON_ONCE(parent_event->ctx->parent_ctx);
5788 mutex_lock(&parent_event->child_mutex);
5789 list_del_init(&child_event->child_list);
5790 mutex_unlock(&parent_event->child_mutex);
5793 * Release the parent event, if this was the last
5796 fput(parent_event->filp);
5800 __perf_event_exit_task(struct perf_event *child_event,
5801 struct perf_event_context *child_ctx,
5802 struct task_struct *child)
5804 struct perf_event *parent_event;
5806 perf_event_remove_from_context(child_event);
5808 parent_event = child_event->parent;
5810 * It can happen that parent exits first, and has events
5811 * that are still around due to the child reference. These
5812 * events need to be zapped - but otherwise linger.
5815 sync_child_event(child_event, child);
5816 free_event(child_event);
5820 static void perf_event_exit_task_context(struct task_struct *child, int ctxn)
5822 struct perf_event *child_event, *tmp;
5823 struct perf_event_context *child_ctx;
5824 unsigned long flags;
5826 if (likely(!child->perf_event_ctxp[ctxn])) {
5827 perf_event_task(child, NULL, 0);
5831 local_irq_save(flags);
5833 * We can't reschedule here because interrupts are disabled,
5834 * and either child is current or it is a task that can't be
5835 * scheduled, so we are now safe from rescheduling changing
5838 child_ctx = child->perf_event_ctxp[ctxn];
5839 task_ctx_sched_out(child_ctx, EVENT_ALL);
5842 * Take the context lock here so that if find_get_context is
5843 * reading child->perf_event_ctxp, we wait until it has
5844 * incremented the context's refcount before we do put_ctx below.
5846 raw_spin_lock(&child_ctx->lock);
5847 child->perf_event_ctxp[ctxn] = NULL;
5849 * If this context is a clone; unclone it so it can't get
5850 * swapped to another process while we're removing all
5851 * the events from it.
5853 unclone_ctx(child_ctx);
5854 update_context_time(child_ctx);
5855 raw_spin_unlock_irqrestore(&child_ctx->lock, flags);
5858 * Report the task dead after unscheduling the events so that we
5859 * won't get any samples after PERF_RECORD_EXIT. We can however still
5860 * get a few PERF_RECORD_READ events.
5862 perf_event_task(child, child_ctx, 0);
5865 * We can recurse on the same lock type through:
5867 * __perf_event_exit_task()
5868 * sync_child_event()
5869 * fput(parent_event->filp)
5871 * mutex_lock(&ctx->mutex)
5873 * But since its the parent context it won't be the same instance.
5875 mutex_lock(&child_ctx->mutex);
5878 list_for_each_entry_safe(child_event, tmp, &child_ctx->pinned_groups,
5880 __perf_event_exit_task(child_event, child_ctx, child);
5882 list_for_each_entry_safe(child_event, tmp, &child_ctx->flexible_groups,
5884 __perf_event_exit_task(child_event, child_ctx, child);
5887 * If the last event was a group event, it will have appended all
5888 * its siblings to the list, but we obtained 'tmp' before that which
5889 * will still point to the list head terminating the iteration.
5891 if (!list_empty(&child_ctx->pinned_groups) ||
5892 !list_empty(&child_ctx->flexible_groups))
5895 mutex_unlock(&child_ctx->mutex);
5901 * When a child task exits, feed back event values to parent events.
5903 void perf_event_exit_task(struct task_struct *child)
5907 for_each_task_context_nr(ctxn)
5908 perf_event_exit_task_context(child, ctxn);
5911 static void perf_free_event(struct perf_event *event,
5912 struct perf_event_context *ctx)
5914 struct perf_event *parent = event->parent;
5916 if (WARN_ON_ONCE(!parent))
5919 mutex_lock(&parent->child_mutex);
5920 list_del_init(&event->child_list);
5921 mutex_unlock(&parent->child_mutex);
5925 perf_group_detach(event);
5926 list_del_event(event, ctx);
5931 * free an unexposed, unused context as created by inheritance by
5932 * perf_event_init_task below, used by fork() in case of fail.
5934 void perf_event_free_task(struct task_struct *task)
5936 struct perf_event_context *ctx;
5937 struct perf_event *event, *tmp;
5940 for_each_task_context_nr(ctxn) {
5941 ctx = task->perf_event_ctxp[ctxn];
5945 mutex_lock(&ctx->mutex);
5947 list_for_each_entry_safe(event, tmp, &ctx->pinned_groups,
5949 perf_free_event(event, ctx);
5951 list_for_each_entry_safe(event, tmp, &ctx->flexible_groups,
5953 perf_free_event(event, ctx);
5955 if (!list_empty(&ctx->pinned_groups) ||
5956 !list_empty(&ctx->flexible_groups))
5959 mutex_unlock(&ctx->mutex);
5965 void perf_event_delayed_put(struct task_struct *task)
5969 for_each_task_context_nr(ctxn)
5970 WARN_ON_ONCE(task->perf_event_ctxp[ctxn]);
5974 * inherit a event from parent task to child task:
5976 static struct perf_event *
5977 inherit_event(struct perf_event *parent_event,
5978 struct task_struct *parent,
5979 struct perf_event_context *parent_ctx,
5980 struct task_struct *child,
5981 struct perf_event *group_leader,
5982 struct perf_event_context *child_ctx)
5984 struct perf_event *child_event;
5985 unsigned long flags;
5988 * Instead of creating recursive hierarchies of events,
5989 * we link inherited events back to the original parent,
5990 * which has a filp for sure, which we use as the reference
5993 if (parent_event->parent)
5994 parent_event = parent_event->parent;
5996 child_event = perf_event_alloc(&parent_event->attr,
5999 group_leader, parent_event,
6001 if (IS_ERR(child_event))
6006 * Make the child state follow the state of the parent event,
6007 * not its attr.disabled bit. We hold the parent's mutex,
6008 * so we won't race with perf_event_{en, dis}able_family.
6010 if (parent_event->state >= PERF_EVENT_STATE_INACTIVE)
6011 child_event->state = PERF_EVENT_STATE_INACTIVE;
6013 child_event->state = PERF_EVENT_STATE_OFF;
6015 if (parent_event->attr.freq) {
6016 u64 sample_period = parent_event->hw.sample_period;
6017 struct hw_perf_event *hwc = &child_event->hw;
6019 hwc->sample_period = sample_period;
6020 hwc->last_period = sample_period;
6022 local64_set(&hwc->period_left, sample_period);
6025 child_event->ctx = child_ctx;
6026 child_event->overflow_handler = parent_event->overflow_handler;
6029 * Link it up in the child's context:
6031 raw_spin_lock_irqsave(&child_ctx->lock, flags);
6032 add_event_to_ctx(child_event, child_ctx);
6033 raw_spin_unlock_irqrestore(&child_ctx->lock, flags);
6036 * Get a reference to the parent filp - we will fput it
6037 * when the child event exits. This is safe to do because
6038 * we are in the parent and we know that the filp still
6039 * exists and has a nonzero count:
6041 atomic_long_inc(&parent_event->filp->f_count);
6044 * Link this into the parent event's child list
6046 WARN_ON_ONCE(parent_event->ctx->parent_ctx);
6047 mutex_lock(&parent_event->child_mutex);
6048 list_add_tail(&child_event->child_list, &parent_event->child_list);
6049 mutex_unlock(&parent_event->child_mutex);
6054 static int inherit_group(struct perf_event *parent_event,
6055 struct task_struct *parent,
6056 struct perf_event_context *parent_ctx,
6057 struct task_struct *child,
6058 struct perf_event_context *child_ctx)
6060 struct perf_event *leader;
6061 struct perf_event *sub;
6062 struct perf_event *child_ctr;
6064 leader = inherit_event(parent_event, parent, parent_ctx,
6065 child, NULL, child_ctx);
6067 return PTR_ERR(leader);
6068 list_for_each_entry(sub, &parent_event->sibling_list, group_entry) {
6069 child_ctr = inherit_event(sub, parent, parent_ctx,
6070 child, leader, child_ctx);
6071 if (IS_ERR(child_ctr))
6072 return PTR_ERR(child_ctr);
6078 inherit_task_group(struct perf_event *event, struct task_struct *parent,
6079 struct perf_event_context *parent_ctx,
6080 struct task_struct *child, int ctxn,
6084 struct perf_event_context *child_ctx;
6086 if (!event->attr.inherit) {
6091 child_ctx = child->perf_event_ctxp[ctxn];
6094 * This is executed from the parent task context, so
6095 * inherit events that have been marked for cloning.
6096 * First allocate and initialize a context for the
6100 child_ctx = alloc_perf_context(event->pmu, child);
6104 child->perf_event_ctxp[ctxn] = child_ctx;
6107 ret = inherit_group(event, parent, parent_ctx,
6117 * Initialize the perf_event context in task_struct
6119 int perf_event_init_context(struct task_struct *child, int ctxn)
6121 struct perf_event_context *child_ctx, *parent_ctx;
6122 struct perf_event_context *cloned_ctx;
6123 struct perf_event *event;
6124 struct task_struct *parent = current;
6125 int inherited_all = 1;
6128 child->perf_event_ctxp[ctxn] = NULL;
6130 mutex_init(&child->perf_event_mutex);
6131 INIT_LIST_HEAD(&child->perf_event_list);
6133 if (likely(!parent->perf_event_ctxp[ctxn]))
6137 * If the parent's context is a clone, pin it so it won't get
6140 parent_ctx = perf_pin_task_context(parent, ctxn);
6143 * No need to check if parent_ctx != NULL here; since we saw
6144 * it non-NULL earlier, the only reason for it to become NULL
6145 * is if we exit, and since we're currently in the middle of
6146 * a fork we can't be exiting at the same time.
6150 * Lock the parent list. No need to lock the child - not PID
6151 * hashed yet and not running, so nobody can access it.
6153 mutex_lock(&parent_ctx->mutex);
6156 * We dont have to disable NMIs - we are only looking at
6157 * the list, not manipulating it:
6159 list_for_each_entry(event, &parent_ctx->pinned_groups, group_entry) {
6160 ret = inherit_task_group(event, parent, parent_ctx,
6161 child, ctxn, &inherited_all);
6166 list_for_each_entry(event, &parent_ctx->flexible_groups, group_entry) {
6167 ret = inherit_task_group(event, parent, parent_ctx,
6168 child, ctxn, &inherited_all);
6173 child_ctx = child->perf_event_ctxp[ctxn];
6175 if (child_ctx && inherited_all) {
6177 * Mark the child context as a clone of the parent
6178 * context, or of whatever the parent is a clone of.
6179 * Note that if the parent is a clone, it could get
6180 * uncloned at any point, but that doesn't matter
6181 * because the list of events and the generation
6182 * count can't have changed since we took the mutex.
6184 cloned_ctx = rcu_dereference(parent_ctx->parent_ctx);
6186 child_ctx->parent_ctx = cloned_ctx;
6187 child_ctx->parent_gen = parent_ctx->parent_gen;
6189 child_ctx->parent_ctx = parent_ctx;
6190 child_ctx->parent_gen = parent_ctx->generation;
6192 get_ctx(child_ctx->parent_ctx);
6195 mutex_unlock(&parent_ctx->mutex);
6197 perf_unpin_context(parent_ctx);
6203 * Initialize the perf_event context in task_struct
6205 int perf_event_init_task(struct task_struct *child)
6209 for_each_task_context_nr(ctxn) {
6210 ret = perf_event_init_context(child, ctxn);
6218 static void __init perf_event_init_all_cpus(void)
6220 struct swevent_htable *swhash;
6223 for_each_possible_cpu(cpu) {
6224 swhash = &per_cpu(swevent_htable, cpu);
6225 mutex_init(&swhash->hlist_mutex);
6226 INIT_LIST_HEAD(&per_cpu(rotation_list, cpu));
6230 static void __cpuinit perf_event_init_cpu(int cpu)
6232 struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
6234 mutex_lock(&swhash->hlist_mutex);
6235 if (swhash->hlist_refcount > 0) {
6236 struct swevent_hlist *hlist;
6238 hlist = kzalloc_node(sizeof(*hlist), GFP_KERNEL, cpu_to_node(cpu));
6240 rcu_assign_pointer(swhash->swevent_hlist, hlist);
6242 mutex_unlock(&swhash->hlist_mutex);
6245 #ifdef CONFIG_HOTPLUG_CPU
6246 static void perf_pmu_rotate_stop(struct pmu *pmu)
6248 struct perf_cpu_context *cpuctx = this_cpu_ptr(pmu->pmu_cpu_context);
6250 WARN_ON(!irqs_disabled());
6252 list_del_init(&cpuctx->rotation_list);
6255 static void __perf_event_exit_context(void *__info)
6257 struct perf_event_context *ctx = __info;
6258 struct perf_event *event, *tmp;
6260 perf_pmu_rotate_stop(ctx->pmu);
6262 list_for_each_entry_safe(event, tmp, &ctx->pinned_groups, group_entry)
6263 __perf_event_remove_from_context(event);
6264 list_for_each_entry_safe(event, tmp, &ctx->flexible_groups, group_entry)
6265 __perf_event_remove_from_context(event);
6268 static void perf_event_exit_cpu_context(int cpu)
6270 struct perf_event_context *ctx;
6274 idx = srcu_read_lock(&pmus_srcu);
6275 list_for_each_entry_rcu(pmu, &pmus, entry) {
6276 ctx = &per_cpu_ptr(pmu->pmu_cpu_context, cpu)->ctx;
6278 mutex_lock(&ctx->mutex);
6279 smp_call_function_single(cpu, __perf_event_exit_context, ctx, 1);
6280 mutex_unlock(&ctx->mutex);
6282 srcu_read_unlock(&pmus_srcu, idx);
6285 static void perf_event_exit_cpu(int cpu)
6287 struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
6289 mutex_lock(&swhash->hlist_mutex);
6290 swevent_hlist_release(swhash);
6291 mutex_unlock(&swhash->hlist_mutex);
6293 perf_event_exit_cpu_context(cpu);
6296 static inline void perf_event_exit_cpu(int cpu) { }
6299 static int __cpuinit
6300 perf_cpu_notify(struct notifier_block *self, unsigned long action, void *hcpu)
6302 unsigned int cpu = (long)hcpu;
6304 switch (action & ~CPU_TASKS_FROZEN) {
6306 case CPU_UP_PREPARE:
6307 case CPU_DOWN_FAILED:
6308 perf_event_init_cpu(cpu);
6311 case CPU_UP_CANCELED:
6312 case CPU_DOWN_PREPARE:
6313 perf_event_exit_cpu(cpu);
6323 void __init perf_event_init(void)
6327 perf_event_init_all_cpus();
6328 init_srcu_struct(&pmus_srcu);
6329 perf_pmu_register(&perf_swevent);
6330 perf_pmu_register(&perf_cpu_clock);
6331 perf_pmu_register(&perf_task_clock);
6333 perf_cpu_notifier(perf_cpu_notify);
6335 ret = init_hw_breakpoint();
6336 WARN(ret, "hw_breakpoint initialization failed with: %d", ret);