2 * Performance events core code:
4 * Copyright (C) 2008 Thomas Gleixner <tglx@linutronix.de>
5 * Copyright (C) 2008-2009 Red Hat, Inc., Ingo Molnar
6 * Copyright (C) 2008-2009 Red Hat, Inc., Peter Zijlstra <pzijlstr@redhat.com>
7 * Copyright © 2009 Paul Mackerras, IBM Corp. <paulus@au1.ibm.com>
9 * For licensing details see kernel-base/COPYING
14 #include <linux/cpu.h>
15 #include <linux/smp.h>
16 #include <linux/file.h>
17 #include <linux/poll.h>
18 #include <linux/slab.h>
19 #include <linux/hash.h>
20 #include <linux/sysfs.h>
21 #include <linux/dcache.h>
22 #include <linux/percpu.h>
23 #include <linux/ptrace.h>
24 #include <linux/vmstat.h>
25 #include <linux/vmalloc.h>
26 #include <linux/hardirq.h>
27 #include <linux/rculist.h>
28 #include <linux/uaccess.h>
29 #include <linux/syscalls.h>
30 #include <linux/anon_inodes.h>
31 #include <linux/kernel_stat.h>
32 #include <linux/perf_event.h>
33 #include <linux/ftrace_event.h>
34 #include <linux/hw_breakpoint.h>
36 #include <asm/irq_regs.h>
39 * Each CPU has a list of per CPU events:
41 static DEFINE_PER_CPU(struct perf_cpu_context, perf_cpu_context);
43 int perf_max_events __read_mostly = 1;
44 static int perf_reserved_percpu __read_mostly;
45 static int perf_overcommit __read_mostly = 1;
47 static atomic_t nr_events __read_mostly;
48 static atomic_t nr_mmap_events __read_mostly;
49 static atomic_t nr_comm_events __read_mostly;
50 static atomic_t nr_task_events __read_mostly;
53 * perf event paranoia level:
54 * -1 - not paranoid at all
55 * 0 - disallow raw tracepoint access for unpriv
56 * 1 - disallow cpu events for unpriv
57 * 2 - disallow kernel profiling for unpriv
59 int sysctl_perf_event_paranoid __read_mostly = 1;
61 int sysctl_perf_event_mlock __read_mostly = 512; /* 'free' kb per user */
64 * max perf event sample rate
66 int sysctl_perf_event_sample_rate __read_mostly = 100000;
68 static atomic64_t perf_event_id;
71 * Lock for (sysadmin-configurable) event reservations:
73 static DEFINE_SPINLOCK(perf_resource_lock);
76 * Architecture provided APIs - weak aliases:
78 extern __weak const struct pmu *hw_perf_event_init(struct perf_event *event)
83 void __weak hw_perf_disable(void) { barrier(); }
84 void __weak hw_perf_enable(void) { barrier(); }
87 hw_perf_group_sched_in(struct perf_event *group_leader,
88 struct perf_cpu_context *cpuctx,
89 struct perf_event_context *ctx)
94 void __weak perf_event_print_debug(void) { }
96 static DEFINE_PER_CPU(int, perf_disable_count);
98 void perf_disable(void)
100 if (!__get_cpu_var(perf_disable_count)++)
104 void perf_enable(void)
106 if (!--__get_cpu_var(perf_disable_count))
110 static void get_ctx(struct perf_event_context *ctx)
112 WARN_ON(!atomic_inc_not_zero(&ctx->refcount));
115 static void free_ctx(struct rcu_head *head)
117 struct perf_event_context *ctx;
119 ctx = container_of(head, struct perf_event_context, rcu_head);
123 static void put_ctx(struct perf_event_context *ctx)
125 if (atomic_dec_and_test(&ctx->refcount)) {
127 put_ctx(ctx->parent_ctx);
129 put_task_struct(ctx->task);
130 call_rcu(&ctx->rcu_head, free_ctx);
134 static void unclone_ctx(struct perf_event_context *ctx)
136 if (ctx->parent_ctx) {
137 put_ctx(ctx->parent_ctx);
138 ctx->parent_ctx = NULL;
143 * If we inherit events we want to return the parent event id
146 static u64 primary_event_id(struct perf_event *event)
151 id = event->parent->id;
157 * Get the perf_event_context for a task and lock it.
158 * This has to cope with with the fact that until it is locked,
159 * the context could get moved to another task.
161 static struct perf_event_context *
162 perf_lock_task_context(struct task_struct *task, unsigned long *flags)
164 struct perf_event_context *ctx;
168 ctx = rcu_dereference(task->perf_event_ctxp);
171 * If this context is a clone of another, it might
172 * get swapped for another underneath us by
173 * perf_event_task_sched_out, though the
174 * rcu_read_lock() protects us from any context
175 * getting freed. Lock the context and check if it
176 * got swapped before we could get the lock, and retry
177 * if so. If we locked the right context, then it
178 * can't get swapped on us any more.
180 raw_spin_lock_irqsave(&ctx->lock, *flags);
181 if (ctx != rcu_dereference(task->perf_event_ctxp)) {
182 raw_spin_unlock_irqrestore(&ctx->lock, *flags);
186 if (!atomic_inc_not_zero(&ctx->refcount)) {
187 raw_spin_unlock_irqrestore(&ctx->lock, *flags);
196 * Get the context for a task and increment its pin_count so it
197 * can't get swapped to another task. This also increments its
198 * reference count so that the context can't get freed.
200 static struct perf_event_context *perf_pin_task_context(struct task_struct *task)
202 struct perf_event_context *ctx;
205 ctx = perf_lock_task_context(task, &flags);
208 raw_spin_unlock_irqrestore(&ctx->lock, flags);
213 static void perf_unpin_context(struct perf_event_context *ctx)
217 raw_spin_lock_irqsave(&ctx->lock, flags);
219 raw_spin_unlock_irqrestore(&ctx->lock, flags);
223 static inline u64 perf_clock(void)
225 return cpu_clock(raw_smp_processor_id());
229 * Update the record of the current time in a context.
231 static void update_context_time(struct perf_event_context *ctx)
233 u64 now = perf_clock();
235 ctx->time += now - ctx->timestamp;
236 ctx->timestamp = now;
240 * Update the total_time_enabled and total_time_running fields for a event.
242 static void update_event_times(struct perf_event *event)
244 struct perf_event_context *ctx = event->ctx;
247 if (event->state < PERF_EVENT_STATE_INACTIVE ||
248 event->group_leader->state < PERF_EVENT_STATE_INACTIVE)
254 run_end = event->tstamp_stopped;
256 event->total_time_enabled = run_end - event->tstamp_enabled;
258 if (event->state == PERF_EVENT_STATE_INACTIVE)
259 run_end = event->tstamp_stopped;
263 event->total_time_running = run_end - event->tstamp_running;
266 static struct list_head *
267 ctx_group_list(struct perf_event *event, struct perf_event_context *ctx)
269 if (event->attr.pinned)
270 return &ctx->pinned_groups;
272 return &ctx->flexible_groups;
276 * Add a event from the lists for its context.
277 * Must be called with ctx->mutex and ctx->lock held.
280 list_add_event(struct perf_event *event, struct perf_event_context *ctx)
282 struct perf_event *group_leader = event->group_leader;
285 * Depending on whether it is a standalone or sibling event,
286 * add it straight to the context's event list, or to the group
287 * leader's sibling list:
289 if (group_leader == event) {
290 struct list_head *list;
292 if (is_software_event(event))
293 event->group_flags |= PERF_GROUP_SOFTWARE;
295 list = ctx_group_list(event, ctx);
296 list_add_tail(&event->group_entry, list);
298 if (group_leader->group_flags & PERF_GROUP_SOFTWARE &&
299 !is_software_event(event))
300 group_leader->group_flags &= ~PERF_GROUP_SOFTWARE;
302 list_add_tail(&event->group_entry, &group_leader->sibling_list);
303 group_leader->nr_siblings++;
306 list_add_rcu(&event->event_entry, &ctx->event_list);
308 if (event->attr.inherit_stat)
313 * Remove a event from the lists for its context.
314 * Must be called with ctx->mutex and ctx->lock held.
317 list_del_event(struct perf_event *event, struct perf_event_context *ctx)
319 struct perf_event *sibling, *tmp;
321 if (list_empty(&event->group_entry))
324 if (event->attr.inherit_stat)
327 list_del_init(&event->group_entry);
328 list_del_rcu(&event->event_entry);
330 if (event->group_leader != event)
331 event->group_leader->nr_siblings--;
333 update_event_times(event);
336 * If event was in error state, then keep it
337 * that way, otherwise bogus counts will be
338 * returned on read(). The only way to get out
339 * of error state is by explicit re-enabling
342 if (event->state > PERF_EVENT_STATE_OFF)
343 event->state = PERF_EVENT_STATE_OFF;
346 * If this was a group event with sibling events then
347 * upgrade the siblings to singleton events by adding them
348 * to the context list directly:
350 list_for_each_entry_safe(sibling, tmp, &event->sibling_list, group_entry) {
351 struct list_head *list;
353 list = ctx_group_list(event, ctx);
354 list_move_tail(&sibling->group_entry, list);
355 sibling->group_leader = sibling;
357 /* Inherit group flags from the previous leader */
358 sibling->group_flags = event->group_flags;
363 event_sched_out(struct perf_event *event,
364 struct perf_cpu_context *cpuctx,
365 struct perf_event_context *ctx)
367 if (event->state != PERF_EVENT_STATE_ACTIVE)
370 event->state = PERF_EVENT_STATE_INACTIVE;
371 if (event->pending_disable) {
372 event->pending_disable = 0;
373 event->state = PERF_EVENT_STATE_OFF;
375 event->tstamp_stopped = ctx->time;
376 event->pmu->disable(event);
379 if (!is_software_event(event))
380 cpuctx->active_oncpu--;
382 if (event->attr.exclusive || !cpuctx->active_oncpu)
383 cpuctx->exclusive = 0;
387 group_sched_out(struct perf_event *group_event,
388 struct perf_cpu_context *cpuctx,
389 struct perf_event_context *ctx)
391 struct perf_event *event;
393 if (group_event->state != PERF_EVENT_STATE_ACTIVE)
396 event_sched_out(group_event, cpuctx, ctx);
399 * Schedule out siblings (if any):
401 list_for_each_entry(event, &group_event->sibling_list, group_entry)
402 event_sched_out(event, cpuctx, ctx);
404 if (group_event->attr.exclusive)
405 cpuctx->exclusive = 0;
409 * Cross CPU call to remove a performance event
411 * We disable the event on the hardware level first. After that we
412 * remove it from the context list.
414 static void __perf_event_remove_from_context(void *info)
416 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
417 struct perf_event *event = info;
418 struct perf_event_context *ctx = event->ctx;
421 * If this is a task context, we need to check whether it is
422 * the current task context of this cpu. If not it has been
423 * scheduled out before the smp call arrived.
425 if (ctx->task && cpuctx->task_ctx != ctx)
428 raw_spin_lock(&ctx->lock);
430 * Protect the list operation against NMI by disabling the
431 * events on a global level.
435 event_sched_out(event, cpuctx, ctx);
437 list_del_event(event, ctx);
441 * Allow more per task events with respect to the
444 cpuctx->max_pertask =
445 min(perf_max_events - ctx->nr_events,
446 perf_max_events - perf_reserved_percpu);
450 raw_spin_unlock(&ctx->lock);
455 * Remove the event from a task's (or a CPU's) list of events.
457 * Must be called with ctx->mutex held.
459 * CPU events are removed with a smp call. For task events we only
460 * call when the task is on a CPU.
462 * If event->ctx is a cloned context, callers must make sure that
463 * every task struct that event->ctx->task could possibly point to
464 * remains valid. This is OK when called from perf_release since
465 * that only calls us on the top-level context, which can't be a clone.
466 * When called from perf_event_exit_task, it's OK because the
467 * context has been detached from its task.
469 static void perf_event_remove_from_context(struct perf_event *event)
471 struct perf_event_context *ctx = event->ctx;
472 struct task_struct *task = ctx->task;
476 * Per cpu events are removed via an smp call and
477 * the removal is always successful.
479 smp_call_function_single(event->cpu,
480 __perf_event_remove_from_context,
486 task_oncpu_function_call(task, __perf_event_remove_from_context,
489 raw_spin_lock_irq(&ctx->lock);
491 * If the context is active we need to retry the smp call.
493 if (ctx->nr_active && !list_empty(&event->group_entry)) {
494 raw_spin_unlock_irq(&ctx->lock);
499 * The lock prevents that this context is scheduled in so we
500 * can remove the event safely, if the call above did not
503 if (!list_empty(&event->group_entry))
504 list_del_event(event, ctx);
505 raw_spin_unlock_irq(&ctx->lock);
509 * Update total_time_enabled and total_time_running for all events in a group.
511 static void update_group_times(struct perf_event *leader)
513 struct perf_event *event;
515 update_event_times(leader);
516 list_for_each_entry(event, &leader->sibling_list, group_entry)
517 update_event_times(event);
521 * Cross CPU call to disable a performance event
523 static void __perf_event_disable(void *info)
525 struct perf_event *event = info;
526 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
527 struct perf_event_context *ctx = event->ctx;
530 * If this is a per-task event, need to check whether this
531 * event's task is the current task on this cpu.
533 if (ctx->task && cpuctx->task_ctx != ctx)
536 raw_spin_lock(&ctx->lock);
539 * If the event is on, turn it off.
540 * If it is in error state, leave it in error state.
542 if (event->state >= PERF_EVENT_STATE_INACTIVE) {
543 update_context_time(ctx);
544 update_group_times(event);
545 if (event == event->group_leader)
546 group_sched_out(event, cpuctx, ctx);
548 event_sched_out(event, cpuctx, ctx);
549 event->state = PERF_EVENT_STATE_OFF;
552 raw_spin_unlock(&ctx->lock);
558 * If event->ctx is a cloned context, callers must make sure that
559 * every task struct that event->ctx->task could possibly point to
560 * remains valid. This condition is satisifed when called through
561 * perf_event_for_each_child or perf_event_for_each because they
562 * hold the top-level event's child_mutex, so any descendant that
563 * goes to exit will block in sync_child_event.
564 * When called from perf_pending_event it's OK because event->ctx
565 * is the current context on this CPU and preemption is disabled,
566 * hence we can't get into perf_event_task_sched_out for this context.
568 void perf_event_disable(struct perf_event *event)
570 struct perf_event_context *ctx = event->ctx;
571 struct task_struct *task = ctx->task;
575 * Disable the event on the cpu that it's on
577 smp_call_function_single(event->cpu, __perf_event_disable,
583 task_oncpu_function_call(task, __perf_event_disable, event);
585 raw_spin_lock_irq(&ctx->lock);
587 * If the event is still active, we need to retry the cross-call.
589 if (event->state == PERF_EVENT_STATE_ACTIVE) {
590 raw_spin_unlock_irq(&ctx->lock);
595 * Since we have the lock this context can't be scheduled
596 * in, so we can change the state safely.
598 if (event->state == PERF_EVENT_STATE_INACTIVE) {
599 update_group_times(event);
600 event->state = PERF_EVENT_STATE_OFF;
603 raw_spin_unlock_irq(&ctx->lock);
607 event_sched_in(struct perf_event *event,
608 struct perf_cpu_context *cpuctx,
609 struct perf_event_context *ctx)
611 if (event->state <= PERF_EVENT_STATE_OFF)
614 event->state = PERF_EVENT_STATE_ACTIVE;
615 event->oncpu = smp_processor_id();
617 * The new state must be visible before we turn it on in the hardware:
621 if (event->pmu->enable(event)) {
622 event->state = PERF_EVENT_STATE_INACTIVE;
627 event->tstamp_running += ctx->time - event->tstamp_stopped;
629 if (!is_software_event(event))
630 cpuctx->active_oncpu++;
633 if (event->attr.exclusive)
634 cpuctx->exclusive = 1;
640 group_sched_in(struct perf_event *group_event,
641 struct perf_cpu_context *cpuctx,
642 struct perf_event_context *ctx)
644 struct perf_event *event, *partial_group;
647 if (group_event->state == PERF_EVENT_STATE_OFF)
650 ret = hw_perf_group_sched_in(group_event, cpuctx, ctx);
652 return ret < 0 ? ret : 0;
654 if (event_sched_in(group_event, cpuctx, ctx))
658 * Schedule in siblings as one group (if any):
660 list_for_each_entry(event, &group_event->sibling_list, group_entry) {
661 if (event_sched_in(event, cpuctx, ctx)) {
662 partial_group = event;
671 * Groups can be scheduled in as one unit only, so undo any
672 * partial group before returning:
674 list_for_each_entry(event, &group_event->sibling_list, group_entry) {
675 if (event == partial_group)
677 event_sched_out(event, cpuctx, ctx);
679 event_sched_out(group_event, cpuctx, ctx);
685 * Work out whether we can put this event group on the CPU now.
687 static int group_can_go_on(struct perf_event *event,
688 struct perf_cpu_context *cpuctx,
692 * Groups consisting entirely of software events can always go on.
694 if (event->group_flags & PERF_GROUP_SOFTWARE)
697 * If an exclusive group is already on, no other hardware
700 if (cpuctx->exclusive)
703 * If this group is exclusive and there are already
704 * events on the CPU, it can't go on.
706 if (event->attr.exclusive && cpuctx->active_oncpu)
709 * Otherwise, try to add it if all previous groups were able
715 static void add_event_to_ctx(struct perf_event *event,
716 struct perf_event_context *ctx)
718 list_add_event(event, ctx);
719 event->tstamp_enabled = ctx->time;
720 event->tstamp_running = ctx->time;
721 event->tstamp_stopped = ctx->time;
725 * Cross CPU call to install and enable a performance event
727 * Must be called with ctx->mutex held
729 static void __perf_install_in_context(void *info)
731 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
732 struct perf_event *event = info;
733 struct perf_event_context *ctx = event->ctx;
734 struct perf_event *leader = event->group_leader;
738 * If this is a task context, we need to check whether it is
739 * the current task context of this cpu. If not it has been
740 * scheduled out before the smp call arrived.
741 * Or possibly this is the right context but it isn't
742 * on this cpu because it had no events.
744 if (ctx->task && cpuctx->task_ctx != ctx) {
745 if (cpuctx->task_ctx || ctx->task != current)
747 cpuctx->task_ctx = ctx;
750 raw_spin_lock(&ctx->lock);
752 update_context_time(ctx);
755 * Protect the list operation against NMI by disabling the
756 * events on a global level. NOP for non NMI based events.
760 add_event_to_ctx(event, ctx);
762 if (event->cpu != -1 && event->cpu != smp_processor_id())
766 * Don't put the event on if it is disabled or if
767 * it is in a group and the group isn't on.
769 if (event->state != PERF_EVENT_STATE_INACTIVE ||
770 (leader != event && leader->state != PERF_EVENT_STATE_ACTIVE))
774 * An exclusive event can't go on if there are already active
775 * hardware events, and no hardware event can go on if there
776 * is already an exclusive event on.
778 if (!group_can_go_on(event, cpuctx, 1))
781 err = event_sched_in(event, cpuctx, ctx);
785 * This event couldn't go on. If it is in a group
786 * then we have to pull the whole group off.
787 * If the event group is pinned then put it in error state.
790 group_sched_out(leader, cpuctx, ctx);
791 if (leader->attr.pinned) {
792 update_group_times(leader);
793 leader->state = PERF_EVENT_STATE_ERROR;
797 if (!err && !ctx->task && cpuctx->max_pertask)
798 cpuctx->max_pertask--;
803 raw_spin_unlock(&ctx->lock);
807 * Attach a performance event to a context
809 * First we add the event to the list with the hardware enable bit
810 * in event->hw_config cleared.
812 * If the event is attached to a task which is on a CPU we use a smp
813 * call to enable it in the task context. The task might have been
814 * scheduled away, but we check this in the smp call again.
816 * Must be called with ctx->mutex held.
819 perf_install_in_context(struct perf_event_context *ctx,
820 struct perf_event *event,
823 struct task_struct *task = ctx->task;
827 * Per cpu events are installed via an smp call and
828 * the install is always successful.
830 smp_call_function_single(cpu, __perf_install_in_context,
836 task_oncpu_function_call(task, __perf_install_in_context,
839 raw_spin_lock_irq(&ctx->lock);
841 * we need to retry the smp call.
843 if (ctx->is_active && list_empty(&event->group_entry)) {
844 raw_spin_unlock_irq(&ctx->lock);
849 * The lock prevents that this context is scheduled in so we
850 * can add the event safely, if it the call above did not
853 if (list_empty(&event->group_entry))
854 add_event_to_ctx(event, ctx);
855 raw_spin_unlock_irq(&ctx->lock);
859 * Put a event into inactive state and update time fields.
860 * Enabling the leader of a group effectively enables all
861 * the group members that aren't explicitly disabled, so we
862 * have to update their ->tstamp_enabled also.
863 * Note: this works for group members as well as group leaders
864 * since the non-leader members' sibling_lists will be empty.
866 static void __perf_event_mark_enabled(struct perf_event *event,
867 struct perf_event_context *ctx)
869 struct perf_event *sub;
871 event->state = PERF_EVENT_STATE_INACTIVE;
872 event->tstamp_enabled = ctx->time - event->total_time_enabled;
873 list_for_each_entry(sub, &event->sibling_list, group_entry)
874 if (sub->state >= PERF_EVENT_STATE_INACTIVE)
875 sub->tstamp_enabled =
876 ctx->time - sub->total_time_enabled;
880 * Cross CPU call to enable a performance event
882 static void __perf_event_enable(void *info)
884 struct perf_event *event = info;
885 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
886 struct perf_event_context *ctx = event->ctx;
887 struct perf_event *leader = event->group_leader;
891 * If this is a per-task event, need to check whether this
892 * event's task is the current task on this cpu.
894 if (ctx->task && cpuctx->task_ctx != ctx) {
895 if (cpuctx->task_ctx || ctx->task != current)
897 cpuctx->task_ctx = ctx;
900 raw_spin_lock(&ctx->lock);
902 update_context_time(ctx);
904 if (event->state >= PERF_EVENT_STATE_INACTIVE)
906 __perf_event_mark_enabled(event, ctx);
908 if (event->cpu != -1 && event->cpu != smp_processor_id())
912 * If the event is in a group and isn't the group leader,
913 * then don't put it on unless the group is on.
915 if (leader != event && leader->state != PERF_EVENT_STATE_ACTIVE)
918 if (!group_can_go_on(event, cpuctx, 1)) {
923 err = group_sched_in(event, cpuctx, ctx);
925 err = event_sched_in(event, cpuctx, ctx);
931 * If this event can't go on and it's part of a
932 * group, then the whole group has to come off.
935 group_sched_out(leader, cpuctx, ctx);
936 if (leader->attr.pinned) {
937 update_group_times(leader);
938 leader->state = PERF_EVENT_STATE_ERROR;
943 raw_spin_unlock(&ctx->lock);
949 * If event->ctx is a cloned context, callers must make sure that
950 * every task struct that event->ctx->task could possibly point to
951 * remains valid. This condition is satisfied when called through
952 * perf_event_for_each_child or perf_event_for_each as described
953 * for perf_event_disable.
955 void perf_event_enable(struct perf_event *event)
957 struct perf_event_context *ctx = event->ctx;
958 struct task_struct *task = ctx->task;
962 * Enable the event on the cpu that it's on
964 smp_call_function_single(event->cpu, __perf_event_enable,
969 raw_spin_lock_irq(&ctx->lock);
970 if (event->state >= PERF_EVENT_STATE_INACTIVE)
974 * If the event is in error state, clear that first.
975 * That way, if we see the event in error state below, we
976 * know that it has gone back into error state, as distinct
977 * from the task having been scheduled away before the
978 * cross-call arrived.
980 if (event->state == PERF_EVENT_STATE_ERROR)
981 event->state = PERF_EVENT_STATE_OFF;
984 raw_spin_unlock_irq(&ctx->lock);
985 task_oncpu_function_call(task, __perf_event_enable, event);
987 raw_spin_lock_irq(&ctx->lock);
990 * If the context is active and the event is still off,
991 * we need to retry the cross-call.
993 if (ctx->is_active && event->state == PERF_EVENT_STATE_OFF)
997 * Since we have the lock this context can't be scheduled
998 * in, so we can change the state safely.
1000 if (event->state == PERF_EVENT_STATE_OFF)
1001 __perf_event_mark_enabled(event, ctx);
1004 raw_spin_unlock_irq(&ctx->lock);
1007 static int perf_event_refresh(struct perf_event *event, int refresh)
1010 * not supported on inherited events
1012 if (event->attr.inherit)
1015 atomic_add(refresh, &event->event_limit);
1016 perf_event_enable(event);
1022 EVENT_FLEXIBLE = 0x1,
1024 EVENT_ALL = EVENT_FLEXIBLE | EVENT_PINNED,
1027 static void ctx_sched_out(struct perf_event_context *ctx,
1028 struct perf_cpu_context *cpuctx,
1029 enum event_type_t event_type)
1031 struct perf_event *event;
1033 raw_spin_lock(&ctx->lock);
1035 if (likely(!ctx->nr_events))
1037 update_context_time(ctx);
1040 if (!ctx->nr_active)
1043 if (event_type & EVENT_PINNED)
1044 list_for_each_entry(event, &ctx->pinned_groups, group_entry)
1045 group_sched_out(event, cpuctx, ctx);
1047 if (event_type & EVENT_FLEXIBLE)
1048 list_for_each_entry(event, &ctx->flexible_groups, group_entry)
1049 group_sched_out(event, cpuctx, ctx);
1054 raw_spin_unlock(&ctx->lock);
1058 * Test whether two contexts are equivalent, i.e. whether they
1059 * have both been cloned from the same version of the same context
1060 * and they both have the same number of enabled events.
1061 * If the number of enabled events is the same, then the set
1062 * of enabled events should be the same, because these are both
1063 * inherited contexts, therefore we can't access individual events
1064 * in them directly with an fd; we can only enable/disable all
1065 * events via prctl, or enable/disable all events in a family
1066 * via ioctl, which will have the same effect on both contexts.
1068 static int context_equiv(struct perf_event_context *ctx1,
1069 struct perf_event_context *ctx2)
1071 return ctx1->parent_ctx && ctx1->parent_ctx == ctx2->parent_ctx
1072 && ctx1->parent_gen == ctx2->parent_gen
1073 && !ctx1->pin_count && !ctx2->pin_count;
1076 static void __perf_event_sync_stat(struct perf_event *event,
1077 struct perf_event *next_event)
1081 if (!event->attr.inherit_stat)
1085 * Update the event value, we cannot use perf_event_read()
1086 * because we're in the middle of a context switch and have IRQs
1087 * disabled, which upsets smp_call_function_single(), however
1088 * we know the event must be on the current CPU, therefore we
1089 * don't need to use it.
1091 switch (event->state) {
1092 case PERF_EVENT_STATE_ACTIVE:
1093 event->pmu->read(event);
1096 case PERF_EVENT_STATE_INACTIVE:
1097 update_event_times(event);
1105 * In order to keep per-task stats reliable we need to flip the event
1106 * values when we flip the contexts.
1108 value = atomic64_read(&next_event->count);
1109 value = atomic64_xchg(&event->count, value);
1110 atomic64_set(&next_event->count, value);
1112 swap(event->total_time_enabled, next_event->total_time_enabled);
1113 swap(event->total_time_running, next_event->total_time_running);
1116 * Since we swizzled the values, update the user visible data too.
1118 perf_event_update_userpage(event);
1119 perf_event_update_userpage(next_event);
1122 #define list_next_entry(pos, member) \
1123 list_entry(pos->member.next, typeof(*pos), member)
1125 static void perf_event_sync_stat(struct perf_event_context *ctx,
1126 struct perf_event_context *next_ctx)
1128 struct perf_event *event, *next_event;
1133 update_context_time(ctx);
1135 event = list_first_entry(&ctx->event_list,
1136 struct perf_event, event_entry);
1138 next_event = list_first_entry(&next_ctx->event_list,
1139 struct perf_event, event_entry);
1141 while (&event->event_entry != &ctx->event_list &&
1142 &next_event->event_entry != &next_ctx->event_list) {
1144 __perf_event_sync_stat(event, next_event);
1146 event = list_next_entry(event, event_entry);
1147 next_event = list_next_entry(next_event, event_entry);
1152 * Called from scheduler to remove the events of the current task,
1153 * with interrupts disabled.
1155 * We stop each event and update the event value in event->count.
1157 * This does not protect us against NMI, but disable()
1158 * sets the disabled bit in the control field of event _before_
1159 * accessing the event control register. If a NMI hits, then it will
1160 * not restart the event.
1162 void perf_event_task_sched_out(struct task_struct *task,
1163 struct task_struct *next)
1165 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
1166 struct perf_event_context *ctx = task->perf_event_ctxp;
1167 struct perf_event_context *next_ctx;
1168 struct perf_event_context *parent;
1171 perf_sw_event(PERF_COUNT_SW_CONTEXT_SWITCHES, 1, 1, NULL, 0);
1173 if (likely(!ctx || !cpuctx->task_ctx))
1177 parent = rcu_dereference(ctx->parent_ctx);
1178 next_ctx = next->perf_event_ctxp;
1179 if (parent && next_ctx &&
1180 rcu_dereference(next_ctx->parent_ctx) == parent) {
1182 * Looks like the two contexts are clones, so we might be
1183 * able to optimize the context switch. We lock both
1184 * contexts and check that they are clones under the
1185 * lock (including re-checking that neither has been
1186 * uncloned in the meantime). It doesn't matter which
1187 * order we take the locks because no other cpu could
1188 * be trying to lock both of these tasks.
1190 raw_spin_lock(&ctx->lock);
1191 raw_spin_lock_nested(&next_ctx->lock, SINGLE_DEPTH_NESTING);
1192 if (context_equiv(ctx, next_ctx)) {
1194 * XXX do we need a memory barrier of sorts
1195 * wrt to rcu_dereference() of perf_event_ctxp
1197 task->perf_event_ctxp = next_ctx;
1198 next->perf_event_ctxp = ctx;
1200 next_ctx->task = task;
1203 perf_event_sync_stat(ctx, next_ctx);
1205 raw_spin_unlock(&next_ctx->lock);
1206 raw_spin_unlock(&ctx->lock);
1211 ctx_sched_out(ctx, cpuctx, EVENT_ALL);
1212 cpuctx->task_ctx = NULL;
1216 static void task_ctx_sched_out(struct perf_event_context *ctx,
1217 enum event_type_t event_type)
1219 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
1221 if (!cpuctx->task_ctx)
1224 if (WARN_ON_ONCE(ctx != cpuctx->task_ctx))
1227 ctx_sched_out(ctx, cpuctx, event_type);
1228 cpuctx->task_ctx = NULL;
1232 * Called with IRQs disabled
1234 static void __perf_event_task_sched_out(struct perf_event_context *ctx)
1236 task_ctx_sched_out(ctx, EVENT_ALL);
1240 * Called with IRQs disabled
1242 static void cpu_ctx_sched_out(struct perf_cpu_context *cpuctx,
1243 enum event_type_t event_type)
1245 ctx_sched_out(&cpuctx->ctx, cpuctx, event_type);
1249 ctx_pinned_sched_in(struct perf_event_context *ctx,
1250 struct perf_cpu_context *cpuctx)
1252 struct perf_event *event;
1254 list_for_each_entry(event, &ctx->pinned_groups, group_entry) {
1255 if (event->state <= PERF_EVENT_STATE_OFF)
1257 if (event->cpu != -1 && event->cpu != smp_processor_id())
1260 if (group_can_go_on(event, cpuctx, 1))
1261 group_sched_in(event, cpuctx, ctx);
1264 * If this pinned group hasn't been scheduled,
1265 * put it in error state.
1267 if (event->state == PERF_EVENT_STATE_INACTIVE) {
1268 update_group_times(event);
1269 event->state = PERF_EVENT_STATE_ERROR;
1275 ctx_flexible_sched_in(struct perf_event_context *ctx,
1276 struct perf_cpu_context *cpuctx)
1278 struct perf_event *event;
1281 list_for_each_entry(event, &ctx->flexible_groups, group_entry) {
1282 /* Ignore events in OFF or ERROR state */
1283 if (event->state <= PERF_EVENT_STATE_OFF)
1286 * Listen to the 'cpu' scheduling filter constraint
1289 if (event->cpu != -1 && event->cpu != smp_processor_id())
1292 if (group_can_go_on(event, cpuctx, can_add_hw))
1293 if (group_sched_in(event, cpuctx, ctx))
1299 ctx_sched_in(struct perf_event_context *ctx,
1300 struct perf_cpu_context *cpuctx,
1301 enum event_type_t event_type)
1303 raw_spin_lock(&ctx->lock);
1305 if (likely(!ctx->nr_events))
1308 ctx->timestamp = perf_clock();
1313 * First go through the list and put on any pinned groups
1314 * in order to give them the best chance of going on.
1316 if (event_type & EVENT_PINNED)
1317 ctx_pinned_sched_in(ctx, cpuctx);
1319 /* Then walk through the lower prio flexible groups */
1320 if (event_type & EVENT_FLEXIBLE)
1321 ctx_flexible_sched_in(ctx, cpuctx);
1325 raw_spin_unlock(&ctx->lock);
1328 static void cpu_ctx_sched_in(struct perf_cpu_context *cpuctx,
1329 enum event_type_t event_type)
1331 struct perf_event_context *ctx = &cpuctx->ctx;
1333 ctx_sched_in(ctx, cpuctx, event_type);
1336 static void task_ctx_sched_in(struct task_struct *task,
1337 enum event_type_t event_type)
1339 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
1340 struct perf_event_context *ctx = task->perf_event_ctxp;
1344 if (cpuctx->task_ctx == ctx)
1346 ctx_sched_in(ctx, cpuctx, event_type);
1347 cpuctx->task_ctx = ctx;
1350 * Called from scheduler to add the events of the current task
1351 * with interrupts disabled.
1353 * We restore the event value and then enable it.
1355 * This does not protect us against NMI, but enable()
1356 * sets the enabled bit in the control field of event _before_
1357 * accessing the event control register. If a NMI hits, then it will
1358 * keep the event running.
1360 void perf_event_task_sched_in(struct task_struct *task)
1362 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
1363 struct perf_event_context *ctx = task->perf_event_ctxp;
1368 if (cpuctx->task_ctx == ctx)
1374 * We want to keep the following priority order:
1375 * cpu pinned (that don't need to move), task pinned,
1376 * cpu flexible, task flexible.
1378 cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE);
1380 ctx_sched_in(ctx, cpuctx, EVENT_PINNED);
1381 cpu_ctx_sched_in(cpuctx, EVENT_FLEXIBLE);
1382 ctx_sched_in(ctx, cpuctx, EVENT_FLEXIBLE);
1384 cpuctx->task_ctx = ctx;
1389 #define MAX_INTERRUPTS (~0ULL)
1391 static void perf_log_throttle(struct perf_event *event, int enable);
1393 static u64 perf_calculate_period(struct perf_event *event, u64 nsec, u64 count)
1395 u64 frequency = event->attr.sample_freq;
1396 u64 sec = NSEC_PER_SEC;
1397 u64 divisor, dividend;
1399 int count_fls, nsec_fls, frequency_fls, sec_fls;
1401 count_fls = fls64(count);
1402 nsec_fls = fls64(nsec);
1403 frequency_fls = fls64(frequency);
1407 * We got @count in @nsec, with a target of sample_freq HZ
1408 * the target period becomes:
1411 * period = -------------------
1412 * @nsec * sample_freq
1417 * Reduce accuracy by one bit such that @a and @b converge
1418 * to a similar magnitude.
1420 #define REDUCE_FLS(a, b) \
1422 if (a##_fls > b##_fls) { \
1432 * Reduce accuracy until either term fits in a u64, then proceed with
1433 * the other, so that finally we can do a u64/u64 division.
1435 while (count_fls + sec_fls > 64 && nsec_fls + frequency_fls > 64) {
1436 REDUCE_FLS(nsec, frequency);
1437 REDUCE_FLS(sec, count);
1440 if (count_fls + sec_fls > 64) {
1441 divisor = nsec * frequency;
1443 while (count_fls + sec_fls > 64) {
1444 REDUCE_FLS(count, sec);
1448 dividend = count * sec;
1450 dividend = count * sec;
1452 while (nsec_fls + frequency_fls > 64) {
1453 REDUCE_FLS(nsec, frequency);
1457 divisor = nsec * frequency;
1460 return div64_u64(dividend, divisor);
1463 static void perf_event_stop(struct perf_event *event)
1465 if (!event->pmu->stop)
1466 return event->pmu->disable(event);
1468 return event->pmu->stop(event);
1471 static int perf_event_start(struct perf_event *event)
1473 if (!event->pmu->start)
1474 return event->pmu->enable(event);
1476 return event->pmu->start(event);
1479 static void perf_adjust_period(struct perf_event *event, u64 nsec, u64 count)
1481 struct hw_perf_event *hwc = &event->hw;
1482 u64 period, sample_period;
1485 period = perf_calculate_period(event, nsec, count);
1487 delta = (s64)(period - hwc->sample_period);
1488 delta = (delta + 7) / 8; /* low pass filter */
1490 sample_period = hwc->sample_period + delta;
1495 hwc->sample_period = sample_period;
1497 if (atomic64_read(&hwc->period_left) > 8*sample_period) {
1499 perf_event_stop(event);
1500 atomic64_set(&hwc->period_left, 0);
1501 perf_event_start(event);
1506 static void perf_ctx_adjust_freq(struct perf_event_context *ctx)
1508 struct perf_event *event;
1509 struct hw_perf_event *hwc;
1510 u64 interrupts, now;
1513 raw_spin_lock(&ctx->lock);
1514 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
1515 if (event->state != PERF_EVENT_STATE_ACTIVE)
1518 if (event->cpu != -1 && event->cpu != smp_processor_id())
1523 interrupts = hwc->interrupts;
1524 hwc->interrupts = 0;
1527 * unthrottle events on the tick
1529 if (interrupts == MAX_INTERRUPTS) {
1530 perf_log_throttle(event, 1);
1532 event->pmu->unthrottle(event);
1536 if (!event->attr.freq || !event->attr.sample_freq)
1540 event->pmu->read(event);
1541 now = atomic64_read(&event->count);
1542 delta = now - hwc->freq_count_stamp;
1543 hwc->freq_count_stamp = now;
1546 perf_adjust_period(event, TICK_NSEC, delta);
1549 raw_spin_unlock(&ctx->lock);
1553 * Round-robin a context's events:
1555 static void rotate_ctx(struct perf_event_context *ctx)
1557 raw_spin_lock(&ctx->lock);
1559 /* Rotate the first entry last of non-pinned groups */
1560 list_rotate_left(&ctx->flexible_groups);
1562 raw_spin_unlock(&ctx->lock);
1565 void perf_event_task_tick(struct task_struct *curr)
1567 struct perf_cpu_context *cpuctx;
1568 struct perf_event_context *ctx;
1571 if (!atomic_read(&nr_events))
1574 cpuctx = &__get_cpu_var(perf_cpu_context);
1575 if (cpuctx->ctx.nr_events &&
1576 cpuctx->ctx.nr_events != cpuctx->ctx.nr_active)
1579 ctx = curr->perf_event_ctxp;
1580 if (ctx && ctx->nr_events && ctx->nr_events != ctx->nr_active)
1583 perf_ctx_adjust_freq(&cpuctx->ctx);
1585 perf_ctx_adjust_freq(ctx);
1591 cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE);
1593 task_ctx_sched_out(ctx, EVENT_FLEXIBLE);
1595 rotate_ctx(&cpuctx->ctx);
1599 cpu_ctx_sched_in(cpuctx, EVENT_FLEXIBLE);
1601 task_ctx_sched_in(curr, EVENT_FLEXIBLE);
1605 static int event_enable_on_exec(struct perf_event *event,
1606 struct perf_event_context *ctx)
1608 if (!event->attr.enable_on_exec)
1611 event->attr.enable_on_exec = 0;
1612 if (event->state >= PERF_EVENT_STATE_INACTIVE)
1615 __perf_event_mark_enabled(event, ctx);
1621 * Enable all of a task's events that have been marked enable-on-exec.
1622 * This expects task == current.
1624 static void perf_event_enable_on_exec(struct task_struct *task)
1626 struct perf_event_context *ctx;
1627 struct perf_event *event;
1628 unsigned long flags;
1632 local_irq_save(flags);
1633 ctx = task->perf_event_ctxp;
1634 if (!ctx || !ctx->nr_events)
1637 __perf_event_task_sched_out(ctx);
1639 raw_spin_lock(&ctx->lock);
1641 list_for_each_entry(event, &ctx->pinned_groups, group_entry) {
1642 ret = event_enable_on_exec(event, ctx);
1647 list_for_each_entry(event, &ctx->flexible_groups, group_entry) {
1648 ret = event_enable_on_exec(event, ctx);
1654 * Unclone this context if we enabled any event.
1659 raw_spin_unlock(&ctx->lock);
1661 perf_event_task_sched_in(task);
1663 local_irq_restore(flags);
1667 * Cross CPU call to read the hardware event
1669 static void __perf_event_read(void *info)
1671 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
1672 struct perf_event *event = info;
1673 struct perf_event_context *ctx = event->ctx;
1676 * If this is a task context, we need to check whether it is
1677 * the current task context of this cpu. If not it has been
1678 * scheduled out before the smp call arrived. In that case
1679 * event->count would have been updated to a recent sample
1680 * when the event was scheduled out.
1682 if (ctx->task && cpuctx->task_ctx != ctx)
1685 raw_spin_lock(&ctx->lock);
1686 update_context_time(ctx);
1687 update_event_times(event);
1688 raw_spin_unlock(&ctx->lock);
1690 event->pmu->read(event);
1693 static u64 perf_event_read(struct perf_event *event)
1696 * If event is enabled and currently active on a CPU, update the
1697 * value in the event structure:
1699 if (event->state == PERF_EVENT_STATE_ACTIVE) {
1700 smp_call_function_single(event->oncpu,
1701 __perf_event_read, event, 1);
1702 } else if (event->state == PERF_EVENT_STATE_INACTIVE) {
1703 struct perf_event_context *ctx = event->ctx;
1704 unsigned long flags;
1706 raw_spin_lock_irqsave(&ctx->lock, flags);
1707 update_context_time(ctx);
1708 update_event_times(event);
1709 raw_spin_unlock_irqrestore(&ctx->lock, flags);
1712 return atomic64_read(&event->count);
1716 * Initialize the perf_event context in a task_struct:
1719 __perf_event_init_context(struct perf_event_context *ctx,
1720 struct task_struct *task)
1722 raw_spin_lock_init(&ctx->lock);
1723 mutex_init(&ctx->mutex);
1724 INIT_LIST_HEAD(&ctx->pinned_groups);
1725 INIT_LIST_HEAD(&ctx->flexible_groups);
1726 INIT_LIST_HEAD(&ctx->event_list);
1727 atomic_set(&ctx->refcount, 1);
1731 static struct perf_event_context *find_get_context(pid_t pid, int cpu)
1733 struct perf_event_context *ctx;
1734 struct perf_cpu_context *cpuctx;
1735 struct task_struct *task;
1736 unsigned long flags;
1739 if (pid == -1 && cpu != -1) {
1740 /* Must be root to operate on a CPU event: */
1741 if (perf_paranoid_cpu() && !capable(CAP_SYS_ADMIN))
1742 return ERR_PTR(-EACCES);
1744 if (cpu < 0 || cpu >= nr_cpumask_bits)
1745 return ERR_PTR(-EINVAL);
1748 * We could be clever and allow to attach a event to an
1749 * offline CPU and activate it when the CPU comes up, but
1752 if (!cpu_online(cpu))
1753 return ERR_PTR(-ENODEV);
1755 cpuctx = &per_cpu(perf_cpu_context, cpu);
1766 task = find_task_by_vpid(pid);
1768 get_task_struct(task);
1772 return ERR_PTR(-ESRCH);
1775 * Can't attach events to a dying task.
1778 if (task->flags & PF_EXITING)
1781 /* Reuse ptrace permission checks for now. */
1783 if (!ptrace_may_access(task, PTRACE_MODE_READ))
1787 ctx = perf_lock_task_context(task, &flags);
1790 raw_spin_unlock_irqrestore(&ctx->lock, flags);
1794 ctx = kzalloc(sizeof(struct perf_event_context), GFP_KERNEL);
1798 __perf_event_init_context(ctx, task);
1800 if (cmpxchg(&task->perf_event_ctxp, NULL, ctx)) {
1802 * We raced with some other task; use
1803 * the context they set.
1808 get_task_struct(task);
1811 put_task_struct(task);
1815 put_task_struct(task);
1816 return ERR_PTR(err);
1819 static void perf_event_free_filter(struct perf_event *event);
1821 static void free_event_rcu(struct rcu_head *head)
1823 struct perf_event *event;
1825 event = container_of(head, struct perf_event, rcu_head);
1827 put_pid_ns(event->ns);
1828 perf_event_free_filter(event);
1832 static void perf_pending_sync(struct perf_event *event);
1834 static void free_event(struct perf_event *event)
1836 perf_pending_sync(event);
1838 if (!event->parent) {
1839 atomic_dec(&nr_events);
1840 if (event->attr.mmap)
1841 atomic_dec(&nr_mmap_events);
1842 if (event->attr.comm)
1843 atomic_dec(&nr_comm_events);
1844 if (event->attr.task)
1845 atomic_dec(&nr_task_events);
1848 if (event->output) {
1849 fput(event->output->filp);
1850 event->output = NULL;
1854 event->destroy(event);
1856 put_ctx(event->ctx);
1857 call_rcu(&event->rcu_head, free_event_rcu);
1860 int perf_event_release_kernel(struct perf_event *event)
1862 struct perf_event_context *ctx = event->ctx;
1864 WARN_ON_ONCE(ctx->parent_ctx);
1865 mutex_lock(&ctx->mutex);
1866 perf_event_remove_from_context(event);
1867 mutex_unlock(&ctx->mutex);
1869 mutex_lock(&event->owner->perf_event_mutex);
1870 list_del_init(&event->owner_entry);
1871 mutex_unlock(&event->owner->perf_event_mutex);
1872 put_task_struct(event->owner);
1878 EXPORT_SYMBOL_GPL(perf_event_release_kernel);
1881 * Called when the last reference to the file is gone.
1883 static int perf_release(struct inode *inode, struct file *file)
1885 struct perf_event *event = file->private_data;
1887 file->private_data = NULL;
1889 return perf_event_release_kernel(event);
1892 static int perf_event_read_size(struct perf_event *event)
1894 int entry = sizeof(u64); /* value */
1898 if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1899 size += sizeof(u64);
1901 if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1902 size += sizeof(u64);
1904 if (event->attr.read_format & PERF_FORMAT_ID)
1905 entry += sizeof(u64);
1907 if (event->attr.read_format & PERF_FORMAT_GROUP) {
1908 nr += event->group_leader->nr_siblings;
1909 size += sizeof(u64);
1917 u64 perf_event_read_value(struct perf_event *event, u64 *enabled, u64 *running)
1919 struct perf_event *child;
1925 mutex_lock(&event->child_mutex);
1926 total += perf_event_read(event);
1927 *enabled += event->total_time_enabled +
1928 atomic64_read(&event->child_total_time_enabled);
1929 *running += event->total_time_running +
1930 atomic64_read(&event->child_total_time_running);
1932 list_for_each_entry(child, &event->child_list, child_list) {
1933 total += perf_event_read(child);
1934 *enabled += child->total_time_enabled;
1935 *running += child->total_time_running;
1937 mutex_unlock(&event->child_mutex);
1941 EXPORT_SYMBOL_GPL(perf_event_read_value);
1943 static int perf_event_read_group(struct perf_event *event,
1944 u64 read_format, char __user *buf)
1946 struct perf_event *leader = event->group_leader, *sub;
1947 int n = 0, size = 0, ret = -EFAULT;
1948 struct perf_event_context *ctx = leader->ctx;
1950 u64 count, enabled, running;
1952 mutex_lock(&ctx->mutex);
1953 count = perf_event_read_value(leader, &enabled, &running);
1955 values[n++] = 1 + leader->nr_siblings;
1956 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1957 values[n++] = enabled;
1958 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1959 values[n++] = running;
1960 values[n++] = count;
1961 if (read_format & PERF_FORMAT_ID)
1962 values[n++] = primary_event_id(leader);
1964 size = n * sizeof(u64);
1966 if (copy_to_user(buf, values, size))
1971 list_for_each_entry(sub, &leader->sibling_list, group_entry) {
1974 values[n++] = perf_event_read_value(sub, &enabled, &running);
1975 if (read_format & PERF_FORMAT_ID)
1976 values[n++] = primary_event_id(sub);
1978 size = n * sizeof(u64);
1980 if (copy_to_user(buf + ret, values, size)) {
1988 mutex_unlock(&ctx->mutex);
1993 static int perf_event_read_one(struct perf_event *event,
1994 u64 read_format, char __user *buf)
1996 u64 enabled, running;
2000 values[n++] = perf_event_read_value(event, &enabled, &running);
2001 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
2002 values[n++] = enabled;
2003 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
2004 values[n++] = running;
2005 if (read_format & PERF_FORMAT_ID)
2006 values[n++] = primary_event_id(event);
2008 if (copy_to_user(buf, values, n * sizeof(u64)))
2011 return n * sizeof(u64);
2015 * Read the performance event - simple non blocking version for now
2018 perf_read_hw(struct perf_event *event, char __user *buf, size_t count)
2020 u64 read_format = event->attr.read_format;
2024 * Return end-of-file for a read on a event that is in
2025 * error state (i.e. because it was pinned but it couldn't be
2026 * scheduled on to the CPU at some point).
2028 if (event->state == PERF_EVENT_STATE_ERROR)
2031 if (count < perf_event_read_size(event))
2034 WARN_ON_ONCE(event->ctx->parent_ctx);
2035 if (read_format & PERF_FORMAT_GROUP)
2036 ret = perf_event_read_group(event, read_format, buf);
2038 ret = perf_event_read_one(event, read_format, buf);
2044 perf_read(struct file *file, char __user *buf, size_t count, loff_t *ppos)
2046 struct perf_event *event = file->private_data;
2048 return perf_read_hw(event, buf, count);
2051 static unsigned int perf_poll(struct file *file, poll_table *wait)
2053 struct perf_event *event = file->private_data;
2054 struct perf_mmap_data *data;
2055 unsigned int events = POLL_HUP;
2058 data = rcu_dereference(event->data);
2060 events = atomic_xchg(&data->poll, 0);
2063 poll_wait(file, &event->waitq, wait);
2068 static void perf_event_reset(struct perf_event *event)
2070 (void)perf_event_read(event);
2071 atomic64_set(&event->count, 0);
2072 perf_event_update_userpage(event);
2076 * Holding the top-level event's child_mutex means that any
2077 * descendant process that has inherited this event will block
2078 * in sync_child_event if it goes to exit, thus satisfying the
2079 * task existence requirements of perf_event_enable/disable.
2081 static void perf_event_for_each_child(struct perf_event *event,
2082 void (*func)(struct perf_event *))
2084 struct perf_event *child;
2086 WARN_ON_ONCE(event->ctx->parent_ctx);
2087 mutex_lock(&event->child_mutex);
2089 list_for_each_entry(child, &event->child_list, child_list)
2091 mutex_unlock(&event->child_mutex);
2094 static void perf_event_for_each(struct perf_event *event,
2095 void (*func)(struct perf_event *))
2097 struct perf_event_context *ctx = event->ctx;
2098 struct perf_event *sibling;
2100 WARN_ON_ONCE(ctx->parent_ctx);
2101 mutex_lock(&ctx->mutex);
2102 event = event->group_leader;
2104 perf_event_for_each_child(event, func);
2106 list_for_each_entry(sibling, &event->sibling_list, group_entry)
2107 perf_event_for_each_child(event, func);
2108 mutex_unlock(&ctx->mutex);
2111 static int perf_event_period(struct perf_event *event, u64 __user *arg)
2113 struct perf_event_context *ctx = event->ctx;
2118 if (!event->attr.sample_period)
2121 size = copy_from_user(&value, arg, sizeof(value));
2122 if (size != sizeof(value))
2128 raw_spin_lock_irq(&ctx->lock);
2129 if (event->attr.freq) {
2130 if (value > sysctl_perf_event_sample_rate) {
2135 event->attr.sample_freq = value;
2137 event->attr.sample_period = value;
2138 event->hw.sample_period = value;
2141 raw_spin_unlock_irq(&ctx->lock);
2146 static int perf_event_set_output(struct perf_event *event, int output_fd);
2147 static int perf_event_set_filter(struct perf_event *event, void __user *arg);
2149 static long perf_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
2151 struct perf_event *event = file->private_data;
2152 void (*func)(struct perf_event *);
2156 case PERF_EVENT_IOC_ENABLE:
2157 func = perf_event_enable;
2159 case PERF_EVENT_IOC_DISABLE:
2160 func = perf_event_disable;
2162 case PERF_EVENT_IOC_RESET:
2163 func = perf_event_reset;
2166 case PERF_EVENT_IOC_REFRESH:
2167 return perf_event_refresh(event, arg);
2169 case PERF_EVENT_IOC_PERIOD:
2170 return perf_event_period(event, (u64 __user *)arg);
2172 case PERF_EVENT_IOC_SET_OUTPUT:
2173 return perf_event_set_output(event, arg);
2175 case PERF_EVENT_IOC_SET_FILTER:
2176 return perf_event_set_filter(event, (void __user *)arg);
2182 if (flags & PERF_IOC_FLAG_GROUP)
2183 perf_event_for_each(event, func);
2185 perf_event_for_each_child(event, func);
2190 int perf_event_task_enable(void)
2192 struct perf_event *event;
2194 mutex_lock(¤t->perf_event_mutex);
2195 list_for_each_entry(event, ¤t->perf_event_list, owner_entry)
2196 perf_event_for_each_child(event, perf_event_enable);
2197 mutex_unlock(¤t->perf_event_mutex);
2202 int perf_event_task_disable(void)
2204 struct perf_event *event;
2206 mutex_lock(¤t->perf_event_mutex);
2207 list_for_each_entry(event, ¤t->perf_event_list, owner_entry)
2208 perf_event_for_each_child(event, perf_event_disable);
2209 mutex_unlock(¤t->perf_event_mutex);
2214 #ifndef PERF_EVENT_INDEX_OFFSET
2215 # define PERF_EVENT_INDEX_OFFSET 0
2218 static int perf_event_index(struct perf_event *event)
2220 if (event->state != PERF_EVENT_STATE_ACTIVE)
2223 return event->hw.idx + 1 - PERF_EVENT_INDEX_OFFSET;
2227 * Callers need to ensure there can be no nesting of this function, otherwise
2228 * the seqlock logic goes bad. We can not serialize this because the arch
2229 * code calls this from NMI context.
2231 void perf_event_update_userpage(struct perf_event *event)
2233 struct perf_event_mmap_page *userpg;
2234 struct perf_mmap_data *data;
2237 data = rcu_dereference(event->data);
2241 userpg = data->user_page;
2244 * Disable preemption so as to not let the corresponding user-space
2245 * spin too long if we get preempted.
2250 userpg->index = perf_event_index(event);
2251 userpg->offset = atomic64_read(&event->count);
2252 if (event->state == PERF_EVENT_STATE_ACTIVE)
2253 userpg->offset -= atomic64_read(&event->hw.prev_count);
2255 userpg->time_enabled = event->total_time_enabled +
2256 atomic64_read(&event->child_total_time_enabled);
2258 userpg->time_running = event->total_time_running +
2259 atomic64_read(&event->child_total_time_running);
2268 static unsigned long perf_data_size(struct perf_mmap_data *data)
2270 return data->nr_pages << (PAGE_SHIFT + data->data_order);
2273 #ifndef CONFIG_PERF_USE_VMALLOC
2276 * Back perf_mmap() with regular GFP_KERNEL-0 pages.
2279 static struct page *
2280 perf_mmap_to_page(struct perf_mmap_data *data, unsigned long pgoff)
2282 if (pgoff > data->nr_pages)
2286 return virt_to_page(data->user_page);
2288 return virt_to_page(data->data_pages[pgoff - 1]);
2291 static struct perf_mmap_data *
2292 perf_mmap_data_alloc(struct perf_event *event, int nr_pages)
2294 struct perf_mmap_data *data;
2298 WARN_ON(atomic_read(&event->mmap_count));
2300 size = sizeof(struct perf_mmap_data);
2301 size += nr_pages * sizeof(void *);
2303 data = kzalloc(size, GFP_KERNEL);
2307 data->user_page = (void *)get_zeroed_page(GFP_KERNEL);
2308 if (!data->user_page)
2309 goto fail_user_page;
2311 for (i = 0; i < nr_pages; i++) {
2312 data->data_pages[i] = (void *)get_zeroed_page(GFP_KERNEL);
2313 if (!data->data_pages[i])
2314 goto fail_data_pages;
2317 data->data_order = 0;
2318 data->nr_pages = nr_pages;
2323 for (i--; i >= 0; i--)
2324 free_page((unsigned long)data->data_pages[i]);
2326 free_page((unsigned long)data->user_page);
2335 static void perf_mmap_free_page(unsigned long addr)
2337 struct page *page = virt_to_page((void *)addr);
2339 page->mapping = NULL;
2343 static void perf_mmap_data_free(struct perf_mmap_data *data)
2347 perf_mmap_free_page((unsigned long)data->user_page);
2348 for (i = 0; i < data->nr_pages; i++)
2349 perf_mmap_free_page((unsigned long)data->data_pages[i]);
2356 * Back perf_mmap() with vmalloc memory.
2358 * Required for architectures that have d-cache aliasing issues.
2361 static struct page *
2362 perf_mmap_to_page(struct perf_mmap_data *data, unsigned long pgoff)
2364 if (pgoff > (1UL << data->data_order))
2367 return vmalloc_to_page((void *)data->user_page + pgoff * PAGE_SIZE);
2370 static void perf_mmap_unmark_page(void *addr)
2372 struct page *page = vmalloc_to_page(addr);
2374 page->mapping = NULL;
2377 static void perf_mmap_data_free_work(struct work_struct *work)
2379 struct perf_mmap_data *data;
2383 data = container_of(work, struct perf_mmap_data, work);
2384 nr = 1 << data->data_order;
2386 base = data->user_page;
2387 for (i = 0; i < nr + 1; i++)
2388 perf_mmap_unmark_page(base + (i * PAGE_SIZE));
2394 static void perf_mmap_data_free(struct perf_mmap_data *data)
2396 schedule_work(&data->work);
2399 static struct perf_mmap_data *
2400 perf_mmap_data_alloc(struct perf_event *event, int nr_pages)
2402 struct perf_mmap_data *data;
2406 WARN_ON(atomic_read(&event->mmap_count));
2408 size = sizeof(struct perf_mmap_data);
2409 size += sizeof(void *);
2411 data = kzalloc(size, GFP_KERNEL);
2415 INIT_WORK(&data->work, perf_mmap_data_free_work);
2417 all_buf = vmalloc_user((nr_pages + 1) * PAGE_SIZE);
2421 data->user_page = all_buf;
2422 data->data_pages[0] = all_buf + PAGE_SIZE;
2423 data->data_order = ilog2(nr_pages);
2437 static int perf_mmap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
2439 struct perf_event *event = vma->vm_file->private_data;
2440 struct perf_mmap_data *data;
2441 int ret = VM_FAULT_SIGBUS;
2443 if (vmf->flags & FAULT_FLAG_MKWRITE) {
2444 if (vmf->pgoff == 0)
2450 data = rcu_dereference(event->data);
2454 if (vmf->pgoff && (vmf->flags & FAULT_FLAG_WRITE))
2457 vmf->page = perf_mmap_to_page(data, vmf->pgoff);
2461 get_page(vmf->page);
2462 vmf->page->mapping = vma->vm_file->f_mapping;
2463 vmf->page->index = vmf->pgoff;
2473 perf_mmap_data_init(struct perf_event *event, struct perf_mmap_data *data)
2475 long max_size = perf_data_size(data);
2477 atomic_set(&data->lock, -1);
2479 if (event->attr.watermark) {
2480 data->watermark = min_t(long, max_size,
2481 event->attr.wakeup_watermark);
2484 if (!data->watermark)
2485 data->watermark = max_size / 2;
2488 rcu_assign_pointer(event->data, data);
2491 static void perf_mmap_data_free_rcu(struct rcu_head *rcu_head)
2493 struct perf_mmap_data *data;
2495 data = container_of(rcu_head, struct perf_mmap_data, rcu_head);
2496 perf_mmap_data_free(data);
2499 static void perf_mmap_data_release(struct perf_event *event)
2501 struct perf_mmap_data *data = event->data;
2503 WARN_ON(atomic_read(&event->mmap_count));
2505 rcu_assign_pointer(event->data, NULL);
2506 call_rcu(&data->rcu_head, perf_mmap_data_free_rcu);
2509 static void perf_mmap_open(struct vm_area_struct *vma)
2511 struct perf_event *event = vma->vm_file->private_data;
2513 atomic_inc(&event->mmap_count);
2516 static void perf_mmap_close(struct vm_area_struct *vma)
2518 struct perf_event *event = vma->vm_file->private_data;
2520 WARN_ON_ONCE(event->ctx->parent_ctx);
2521 if (atomic_dec_and_mutex_lock(&event->mmap_count, &event->mmap_mutex)) {
2522 unsigned long size = perf_data_size(event->data);
2523 struct user_struct *user = current_user();
2525 atomic_long_sub((size >> PAGE_SHIFT) + 1, &user->locked_vm);
2526 vma->vm_mm->locked_vm -= event->data->nr_locked;
2527 perf_mmap_data_release(event);
2528 mutex_unlock(&event->mmap_mutex);
2532 static const struct vm_operations_struct perf_mmap_vmops = {
2533 .open = perf_mmap_open,
2534 .close = perf_mmap_close,
2535 .fault = perf_mmap_fault,
2536 .page_mkwrite = perf_mmap_fault,
2539 static int perf_mmap(struct file *file, struct vm_area_struct *vma)
2541 struct perf_event *event = file->private_data;
2542 unsigned long user_locked, user_lock_limit;
2543 struct user_struct *user = current_user();
2544 unsigned long locked, lock_limit;
2545 struct perf_mmap_data *data;
2546 unsigned long vma_size;
2547 unsigned long nr_pages;
2548 long user_extra, extra;
2551 if (!(vma->vm_flags & VM_SHARED))
2554 vma_size = vma->vm_end - vma->vm_start;
2555 nr_pages = (vma_size / PAGE_SIZE) - 1;
2558 * If we have data pages ensure they're a power-of-two number, so we
2559 * can do bitmasks instead of modulo.
2561 if (nr_pages != 0 && !is_power_of_2(nr_pages))
2564 if (vma_size != PAGE_SIZE * (1 + nr_pages))
2567 if (vma->vm_pgoff != 0)
2570 WARN_ON_ONCE(event->ctx->parent_ctx);
2571 mutex_lock(&event->mmap_mutex);
2572 if (event->output) {
2577 if (atomic_inc_not_zero(&event->mmap_count)) {
2578 if (nr_pages != event->data->nr_pages)
2583 user_extra = nr_pages + 1;
2584 user_lock_limit = sysctl_perf_event_mlock >> (PAGE_SHIFT - 10);
2587 * Increase the limit linearly with more CPUs:
2589 user_lock_limit *= num_online_cpus();
2591 user_locked = atomic_long_read(&user->locked_vm) + user_extra;
2594 if (user_locked > user_lock_limit)
2595 extra = user_locked - user_lock_limit;
2597 lock_limit = rlimit(RLIMIT_MEMLOCK);
2598 lock_limit >>= PAGE_SHIFT;
2599 locked = vma->vm_mm->locked_vm + extra;
2601 if ((locked > lock_limit) && perf_paranoid_tracepoint_raw() &&
2602 !capable(CAP_IPC_LOCK)) {
2607 WARN_ON(event->data);
2609 data = perf_mmap_data_alloc(event, nr_pages);
2615 perf_mmap_data_init(event, data);
2617 atomic_set(&event->mmap_count, 1);
2618 atomic_long_add(user_extra, &user->locked_vm);
2619 vma->vm_mm->locked_vm += extra;
2620 event->data->nr_locked = extra;
2621 if (vma->vm_flags & VM_WRITE)
2622 event->data->writable = 1;
2625 mutex_unlock(&event->mmap_mutex);
2627 vma->vm_flags |= VM_RESERVED;
2628 vma->vm_ops = &perf_mmap_vmops;
2633 static int perf_fasync(int fd, struct file *filp, int on)
2635 struct inode *inode = filp->f_path.dentry->d_inode;
2636 struct perf_event *event = filp->private_data;
2639 mutex_lock(&inode->i_mutex);
2640 retval = fasync_helper(fd, filp, on, &event->fasync);
2641 mutex_unlock(&inode->i_mutex);
2649 static const struct file_operations perf_fops = {
2650 .llseek = no_llseek,
2651 .release = perf_release,
2654 .unlocked_ioctl = perf_ioctl,
2655 .compat_ioctl = perf_ioctl,
2657 .fasync = perf_fasync,
2663 * If there's data, ensure we set the poll() state and publish everything
2664 * to user-space before waking everybody up.
2667 void perf_event_wakeup(struct perf_event *event)
2669 wake_up_all(&event->waitq);
2671 if (event->pending_kill) {
2672 kill_fasync(&event->fasync, SIGIO, event->pending_kill);
2673 event->pending_kill = 0;
2680 * Handle the case where we need to wakeup up from NMI (or rq->lock) context.
2682 * The NMI bit means we cannot possibly take locks. Therefore, maintain a
2683 * single linked list and use cmpxchg() to add entries lockless.
2686 static void perf_pending_event(struct perf_pending_entry *entry)
2688 struct perf_event *event = container_of(entry,
2689 struct perf_event, pending);
2691 if (event->pending_disable) {
2692 event->pending_disable = 0;
2693 __perf_event_disable(event);
2696 if (event->pending_wakeup) {
2697 event->pending_wakeup = 0;
2698 perf_event_wakeup(event);
2702 #define PENDING_TAIL ((struct perf_pending_entry *)-1UL)
2704 static DEFINE_PER_CPU(struct perf_pending_entry *, perf_pending_head) = {
2708 static void perf_pending_queue(struct perf_pending_entry *entry,
2709 void (*func)(struct perf_pending_entry *))
2711 struct perf_pending_entry **head;
2713 if (cmpxchg(&entry->next, NULL, PENDING_TAIL) != NULL)
2718 head = &get_cpu_var(perf_pending_head);
2721 entry->next = *head;
2722 } while (cmpxchg(head, entry->next, entry) != entry->next);
2724 set_perf_event_pending();
2726 put_cpu_var(perf_pending_head);
2729 static int __perf_pending_run(void)
2731 struct perf_pending_entry *list;
2734 list = xchg(&__get_cpu_var(perf_pending_head), PENDING_TAIL);
2735 while (list != PENDING_TAIL) {
2736 void (*func)(struct perf_pending_entry *);
2737 struct perf_pending_entry *entry = list;
2744 * Ensure we observe the unqueue before we issue the wakeup,
2745 * so that we won't be waiting forever.
2746 * -- see perf_not_pending().
2757 static inline int perf_not_pending(struct perf_event *event)
2760 * If we flush on whatever cpu we run, there is a chance we don't
2764 __perf_pending_run();
2768 * Ensure we see the proper queue state before going to sleep
2769 * so that we do not miss the wakeup. -- see perf_pending_handle()
2772 return event->pending.next == NULL;
2775 static void perf_pending_sync(struct perf_event *event)
2777 wait_event(event->waitq, perf_not_pending(event));
2780 void perf_event_do_pending(void)
2782 __perf_pending_run();
2786 * Callchain support -- arch specific
2789 __weak struct perf_callchain_entry *perf_callchain(struct pt_regs *regs)
2795 void perf_arch_fetch_caller_regs(struct pt_regs *regs, unsigned long ip, int skip)
2801 * We assume there is only KVM supporting the callbacks.
2802 * Later on, we might change it to a list if there is
2803 * another virtualization implementation supporting the callbacks.
2805 struct perf_guest_info_callbacks *perf_guest_cbs;
2807 int perf_register_guest_info_callbacks(struct perf_guest_info_callbacks *cbs)
2809 perf_guest_cbs = cbs;
2812 EXPORT_SYMBOL_GPL(perf_register_guest_info_callbacks);
2814 int perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks *cbs)
2816 perf_guest_cbs = NULL;
2819 EXPORT_SYMBOL_GPL(perf_unregister_guest_info_callbacks);
2824 static bool perf_output_space(struct perf_mmap_data *data, unsigned long tail,
2825 unsigned long offset, unsigned long head)
2829 if (!data->writable)
2832 mask = perf_data_size(data) - 1;
2834 offset = (offset - tail) & mask;
2835 head = (head - tail) & mask;
2837 if ((int)(head - offset) < 0)
2843 static void perf_output_wakeup(struct perf_output_handle *handle)
2845 atomic_set(&handle->data->poll, POLL_IN);
2848 handle->event->pending_wakeup = 1;
2849 perf_pending_queue(&handle->event->pending,
2850 perf_pending_event);
2852 perf_event_wakeup(handle->event);
2856 * Curious locking construct.
2858 * We need to ensure a later event_id doesn't publish a head when a former
2859 * event_id isn't done writing. However since we need to deal with NMIs we
2860 * cannot fully serialize things.
2862 * What we do is serialize between CPUs so we only have to deal with NMI
2863 * nesting on a single CPU.
2865 * We only publish the head (and generate a wakeup) when the outer-most
2866 * event_id completes.
2868 static void perf_output_lock(struct perf_output_handle *handle)
2870 struct perf_mmap_data *data = handle->data;
2871 int cur, cpu = get_cpu();
2876 cur = atomic_cmpxchg(&data->lock, -1, cpu);
2888 static void perf_output_unlock(struct perf_output_handle *handle)
2890 struct perf_mmap_data *data = handle->data;
2894 data->done_head = data->head;
2896 if (!handle->locked)
2901 * The xchg implies a full barrier that ensures all writes are done
2902 * before we publish the new head, matched by a rmb() in userspace when
2903 * reading this position.
2905 while ((head = atomic_long_xchg(&data->done_head, 0)))
2906 data->user_page->data_head = head;
2909 * NMI can happen here, which means we can miss a done_head update.
2912 cpu = atomic_xchg(&data->lock, -1);
2913 WARN_ON_ONCE(cpu != smp_processor_id());
2916 * Therefore we have to validate we did not indeed do so.
2918 if (unlikely(atomic_long_read(&data->done_head))) {
2920 * Since we had it locked, we can lock it again.
2922 while (atomic_cmpxchg(&data->lock, -1, cpu) != -1)
2928 if (atomic_xchg(&data->wakeup, 0))
2929 perf_output_wakeup(handle);
2934 void perf_output_copy(struct perf_output_handle *handle,
2935 const void *buf, unsigned int len)
2937 unsigned int pages_mask;
2938 unsigned long offset;
2942 offset = handle->offset;
2943 pages_mask = handle->data->nr_pages - 1;
2944 pages = handle->data->data_pages;
2947 unsigned long page_offset;
2948 unsigned long page_size;
2951 nr = (offset >> PAGE_SHIFT) & pages_mask;
2952 page_size = 1UL << (handle->data->data_order + PAGE_SHIFT);
2953 page_offset = offset & (page_size - 1);
2954 size = min_t(unsigned int, page_size - page_offset, len);
2956 memcpy(pages[nr] + page_offset, buf, size);
2963 handle->offset = offset;
2966 * Check we didn't copy past our reservation window, taking the
2967 * possible unsigned int wrap into account.
2969 WARN_ON_ONCE(((long)(handle->head - handle->offset)) < 0);
2972 int perf_output_begin(struct perf_output_handle *handle,
2973 struct perf_event *event, unsigned int size,
2974 int nmi, int sample)
2976 struct perf_event *output_event;
2977 struct perf_mmap_data *data;
2978 unsigned long tail, offset, head;
2981 struct perf_event_header header;
2988 * For inherited events we send all the output towards the parent.
2991 event = event->parent;
2993 output_event = rcu_dereference(event->output);
2995 event = output_event;
2997 data = rcu_dereference(event->data);
3001 handle->data = data;
3002 handle->event = event;
3004 handle->sample = sample;
3006 if (!data->nr_pages)
3009 have_lost = atomic_read(&data->lost);
3011 size += sizeof(lost_event);
3013 perf_output_lock(handle);
3017 * Userspace could choose to issue a mb() before updating the
3018 * tail pointer. So that all reads will be completed before the
3021 tail = ACCESS_ONCE(data->user_page->data_tail);
3023 offset = head = atomic_long_read(&data->head);
3025 if (unlikely(!perf_output_space(data, tail, offset, head)))
3027 } while (atomic_long_cmpxchg(&data->head, offset, head) != offset);
3029 handle->offset = offset;
3030 handle->head = head;
3032 if (head - tail > data->watermark)
3033 atomic_set(&data->wakeup, 1);
3036 lost_event.header.type = PERF_RECORD_LOST;
3037 lost_event.header.misc = 0;
3038 lost_event.header.size = sizeof(lost_event);
3039 lost_event.id = event->id;
3040 lost_event.lost = atomic_xchg(&data->lost, 0);
3042 perf_output_put(handle, lost_event);
3048 atomic_inc(&data->lost);
3049 perf_output_unlock(handle);
3056 void perf_output_end(struct perf_output_handle *handle)
3058 struct perf_event *event = handle->event;
3059 struct perf_mmap_data *data = handle->data;
3061 int wakeup_events = event->attr.wakeup_events;
3063 if (handle->sample && wakeup_events) {
3064 int events = atomic_inc_return(&data->events);
3065 if (events >= wakeup_events) {
3066 atomic_sub(wakeup_events, &data->events);
3067 atomic_set(&data->wakeup, 1);
3071 perf_output_unlock(handle);
3075 static u32 perf_event_pid(struct perf_event *event, struct task_struct *p)
3078 * only top level events have the pid namespace they were created in
3081 event = event->parent;
3083 return task_tgid_nr_ns(p, event->ns);
3086 static u32 perf_event_tid(struct perf_event *event, struct task_struct *p)
3089 * only top level events have the pid namespace they were created in
3092 event = event->parent;
3094 return task_pid_nr_ns(p, event->ns);
3097 static void perf_output_read_one(struct perf_output_handle *handle,
3098 struct perf_event *event)
3100 u64 read_format = event->attr.read_format;
3104 values[n++] = atomic64_read(&event->count);
3105 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
3106 values[n++] = event->total_time_enabled +
3107 atomic64_read(&event->child_total_time_enabled);
3109 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
3110 values[n++] = event->total_time_running +
3111 atomic64_read(&event->child_total_time_running);
3113 if (read_format & PERF_FORMAT_ID)
3114 values[n++] = primary_event_id(event);
3116 perf_output_copy(handle, values, n * sizeof(u64));
3120 * XXX PERF_FORMAT_GROUP vs inherited events seems difficult.
3122 static void perf_output_read_group(struct perf_output_handle *handle,
3123 struct perf_event *event)
3125 struct perf_event *leader = event->group_leader, *sub;
3126 u64 read_format = event->attr.read_format;
3130 values[n++] = 1 + leader->nr_siblings;
3132 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
3133 values[n++] = leader->total_time_enabled;
3135 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
3136 values[n++] = leader->total_time_running;
3138 if (leader != event)
3139 leader->pmu->read(leader);
3141 values[n++] = atomic64_read(&leader->count);
3142 if (read_format & PERF_FORMAT_ID)
3143 values[n++] = primary_event_id(leader);
3145 perf_output_copy(handle, values, n * sizeof(u64));
3147 list_for_each_entry(sub, &leader->sibling_list, group_entry) {
3151 sub->pmu->read(sub);
3153 values[n++] = atomic64_read(&sub->count);
3154 if (read_format & PERF_FORMAT_ID)
3155 values[n++] = primary_event_id(sub);
3157 perf_output_copy(handle, values, n * sizeof(u64));
3161 static void perf_output_read(struct perf_output_handle *handle,
3162 struct perf_event *event)
3164 if (event->attr.read_format & PERF_FORMAT_GROUP)
3165 perf_output_read_group(handle, event);
3167 perf_output_read_one(handle, event);
3170 void perf_output_sample(struct perf_output_handle *handle,
3171 struct perf_event_header *header,
3172 struct perf_sample_data *data,
3173 struct perf_event *event)
3175 u64 sample_type = data->type;
3177 perf_output_put(handle, *header);
3179 if (sample_type & PERF_SAMPLE_IP)
3180 perf_output_put(handle, data->ip);
3182 if (sample_type & PERF_SAMPLE_TID)
3183 perf_output_put(handle, data->tid_entry);
3185 if (sample_type & PERF_SAMPLE_TIME)
3186 perf_output_put(handle, data->time);
3188 if (sample_type & PERF_SAMPLE_ADDR)
3189 perf_output_put(handle, data->addr);
3191 if (sample_type & PERF_SAMPLE_ID)
3192 perf_output_put(handle, data->id);
3194 if (sample_type & PERF_SAMPLE_STREAM_ID)
3195 perf_output_put(handle, data->stream_id);
3197 if (sample_type & PERF_SAMPLE_CPU)
3198 perf_output_put(handle, data->cpu_entry);
3200 if (sample_type & PERF_SAMPLE_PERIOD)
3201 perf_output_put(handle, data->period);
3203 if (sample_type & PERF_SAMPLE_READ)
3204 perf_output_read(handle, event);
3206 if (sample_type & PERF_SAMPLE_CALLCHAIN) {
3207 if (data->callchain) {
3210 if (data->callchain)
3211 size += data->callchain->nr;
3213 size *= sizeof(u64);
3215 perf_output_copy(handle, data->callchain, size);
3218 perf_output_put(handle, nr);
3222 if (sample_type & PERF_SAMPLE_RAW) {
3224 perf_output_put(handle, data->raw->size);
3225 perf_output_copy(handle, data->raw->data,
3232 .size = sizeof(u32),
3235 perf_output_put(handle, raw);
3240 void perf_prepare_sample(struct perf_event_header *header,
3241 struct perf_sample_data *data,
3242 struct perf_event *event,
3243 struct pt_regs *regs)
3245 u64 sample_type = event->attr.sample_type;
3247 data->type = sample_type;
3249 header->type = PERF_RECORD_SAMPLE;
3250 header->size = sizeof(*header);
3253 header->misc |= perf_misc_flags(regs);
3255 if (sample_type & PERF_SAMPLE_IP) {
3256 data->ip = perf_instruction_pointer(regs);
3258 header->size += sizeof(data->ip);
3261 if (sample_type & PERF_SAMPLE_TID) {
3262 /* namespace issues */
3263 data->tid_entry.pid = perf_event_pid(event, current);
3264 data->tid_entry.tid = perf_event_tid(event, current);
3266 header->size += sizeof(data->tid_entry);
3269 if (sample_type & PERF_SAMPLE_TIME) {
3270 data->time = perf_clock();
3272 header->size += sizeof(data->time);
3275 if (sample_type & PERF_SAMPLE_ADDR)
3276 header->size += sizeof(data->addr);
3278 if (sample_type & PERF_SAMPLE_ID) {
3279 data->id = primary_event_id(event);
3281 header->size += sizeof(data->id);
3284 if (sample_type & PERF_SAMPLE_STREAM_ID) {
3285 data->stream_id = event->id;
3287 header->size += sizeof(data->stream_id);
3290 if (sample_type & PERF_SAMPLE_CPU) {
3291 data->cpu_entry.cpu = raw_smp_processor_id();
3292 data->cpu_entry.reserved = 0;
3294 header->size += sizeof(data->cpu_entry);
3297 if (sample_type & PERF_SAMPLE_PERIOD)
3298 header->size += sizeof(data->period);
3300 if (sample_type & PERF_SAMPLE_READ)
3301 header->size += perf_event_read_size(event);
3303 if (sample_type & PERF_SAMPLE_CALLCHAIN) {
3306 data->callchain = perf_callchain(regs);
3308 if (data->callchain)
3309 size += data->callchain->nr;
3311 header->size += size * sizeof(u64);
3314 if (sample_type & PERF_SAMPLE_RAW) {
3315 int size = sizeof(u32);
3318 size += data->raw->size;
3320 size += sizeof(u32);
3322 WARN_ON_ONCE(size & (sizeof(u64)-1));
3323 header->size += size;
3327 static void perf_event_output(struct perf_event *event, int nmi,
3328 struct perf_sample_data *data,
3329 struct pt_regs *regs)
3331 struct perf_output_handle handle;
3332 struct perf_event_header header;
3334 perf_prepare_sample(&header, data, event, regs);
3336 if (perf_output_begin(&handle, event, header.size, nmi, 1))
3339 perf_output_sample(&handle, &header, data, event);
3341 perf_output_end(&handle);
3348 struct perf_read_event {
3349 struct perf_event_header header;
3356 perf_event_read_event(struct perf_event *event,
3357 struct task_struct *task)
3359 struct perf_output_handle handle;
3360 struct perf_read_event read_event = {
3362 .type = PERF_RECORD_READ,
3364 .size = sizeof(read_event) + perf_event_read_size(event),
3366 .pid = perf_event_pid(event, task),
3367 .tid = perf_event_tid(event, task),
3371 ret = perf_output_begin(&handle, event, read_event.header.size, 0, 0);
3375 perf_output_put(&handle, read_event);
3376 perf_output_read(&handle, event);
3378 perf_output_end(&handle);
3382 * task tracking -- fork/exit
3384 * enabled by: attr.comm | attr.mmap | attr.task
3387 struct perf_task_event {
3388 struct task_struct *task;
3389 struct perf_event_context *task_ctx;
3392 struct perf_event_header header;
3402 static void perf_event_task_output(struct perf_event *event,
3403 struct perf_task_event *task_event)
3405 struct perf_output_handle handle;
3406 struct task_struct *task = task_event->task;
3407 unsigned long flags;
3411 * If this CPU attempts to acquire an rq lock held by a CPU spinning
3412 * in perf_output_lock() from interrupt context, it's game over.
3414 local_irq_save(flags);
3416 size = task_event->event_id.header.size;
3417 ret = perf_output_begin(&handle, event, size, 0, 0);
3420 local_irq_restore(flags);
3424 task_event->event_id.pid = perf_event_pid(event, task);
3425 task_event->event_id.ppid = perf_event_pid(event, current);
3427 task_event->event_id.tid = perf_event_tid(event, task);
3428 task_event->event_id.ptid = perf_event_tid(event, current);
3430 perf_output_put(&handle, task_event->event_id);
3432 perf_output_end(&handle);
3433 local_irq_restore(flags);
3436 static int perf_event_task_match(struct perf_event *event)
3438 if (event->state < PERF_EVENT_STATE_INACTIVE)
3441 if (event->cpu != -1 && event->cpu != smp_processor_id())
3444 if (event->attr.comm || event->attr.mmap || event->attr.task)
3450 static void perf_event_task_ctx(struct perf_event_context *ctx,
3451 struct perf_task_event *task_event)
3453 struct perf_event *event;
3455 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
3456 if (perf_event_task_match(event))
3457 perf_event_task_output(event, task_event);
3461 static void perf_event_task_event(struct perf_task_event *task_event)
3463 struct perf_cpu_context *cpuctx;
3464 struct perf_event_context *ctx = task_event->task_ctx;
3467 cpuctx = &get_cpu_var(perf_cpu_context);
3468 perf_event_task_ctx(&cpuctx->ctx, task_event);
3470 ctx = rcu_dereference(current->perf_event_ctxp);
3472 perf_event_task_ctx(ctx, task_event);
3473 put_cpu_var(perf_cpu_context);
3477 static void perf_event_task(struct task_struct *task,
3478 struct perf_event_context *task_ctx,
3481 struct perf_task_event task_event;
3483 if (!atomic_read(&nr_comm_events) &&
3484 !atomic_read(&nr_mmap_events) &&
3485 !atomic_read(&nr_task_events))
3488 task_event = (struct perf_task_event){
3490 .task_ctx = task_ctx,
3493 .type = new ? PERF_RECORD_FORK : PERF_RECORD_EXIT,
3495 .size = sizeof(task_event.event_id),
3501 .time = perf_clock(),
3505 perf_event_task_event(&task_event);
3508 void perf_event_fork(struct task_struct *task)
3510 perf_event_task(task, NULL, 1);
3517 struct perf_comm_event {
3518 struct task_struct *task;
3523 struct perf_event_header header;
3530 static void perf_event_comm_output(struct perf_event *event,
3531 struct perf_comm_event *comm_event)
3533 struct perf_output_handle handle;
3534 int size = comm_event->event_id.header.size;
3535 int ret = perf_output_begin(&handle, event, size, 0, 0);
3540 comm_event->event_id.pid = perf_event_pid(event, comm_event->task);
3541 comm_event->event_id.tid = perf_event_tid(event, comm_event->task);
3543 perf_output_put(&handle, comm_event->event_id);
3544 perf_output_copy(&handle, comm_event->comm,
3545 comm_event->comm_size);
3546 perf_output_end(&handle);
3549 static int perf_event_comm_match(struct perf_event *event)
3551 if (event->state < PERF_EVENT_STATE_INACTIVE)
3554 if (event->cpu != -1 && event->cpu != smp_processor_id())
3557 if (event->attr.comm)
3563 static void perf_event_comm_ctx(struct perf_event_context *ctx,
3564 struct perf_comm_event *comm_event)
3566 struct perf_event *event;
3568 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
3569 if (perf_event_comm_match(event))
3570 perf_event_comm_output(event, comm_event);
3574 static void perf_event_comm_event(struct perf_comm_event *comm_event)
3576 struct perf_cpu_context *cpuctx;
3577 struct perf_event_context *ctx;
3579 char comm[TASK_COMM_LEN];
3581 memset(comm, 0, sizeof(comm));
3582 strlcpy(comm, comm_event->task->comm, sizeof(comm));
3583 size = ALIGN(strlen(comm)+1, sizeof(u64));
3585 comm_event->comm = comm;
3586 comm_event->comm_size = size;
3588 comm_event->event_id.header.size = sizeof(comm_event->event_id) + size;
3591 cpuctx = &get_cpu_var(perf_cpu_context);
3592 perf_event_comm_ctx(&cpuctx->ctx, comm_event);
3593 ctx = rcu_dereference(current->perf_event_ctxp);
3595 perf_event_comm_ctx(ctx, comm_event);
3596 put_cpu_var(perf_cpu_context);
3600 void perf_event_comm(struct task_struct *task)
3602 struct perf_comm_event comm_event;
3604 if (task->perf_event_ctxp)
3605 perf_event_enable_on_exec(task);
3607 if (!atomic_read(&nr_comm_events))
3610 comm_event = (struct perf_comm_event){
3616 .type = PERF_RECORD_COMM,
3625 perf_event_comm_event(&comm_event);
3632 struct perf_mmap_event {
3633 struct vm_area_struct *vma;
3635 const char *file_name;
3639 struct perf_event_header header;
3649 static void perf_event_mmap_output(struct perf_event *event,
3650 struct perf_mmap_event *mmap_event)
3652 struct perf_output_handle handle;
3653 int size = mmap_event->event_id.header.size;
3654 int ret = perf_output_begin(&handle, event, size, 0, 0);
3659 mmap_event->event_id.pid = perf_event_pid(event, current);
3660 mmap_event->event_id.tid = perf_event_tid(event, current);
3662 perf_output_put(&handle, mmap_event->event_id);
3663 perf_output_copy(&handle, mmap_event->file_name,
3664 mmap_event->file_size);
3665 perf_output_end(&handle);
3668 static int perf_event_mmap_match(struct perf_event *event,
3669 struct perf_mmap_event *mmap_event)
3671 if (event->state < PERF_EVENT_STATE_INACTIVE)
3674 if (event->cpu != -1 && event->cpu != smp_processor_id())
3677 if (event->attr.mmap)
3683 static void perf_event_mmap_ctx(struct perf_event_context *ctx,
3684 struct perf_mmap_event *mmap_event)
3686 struct perf_event *event;
3688 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
3689 if (perf_event_mmap_match(event, mmap_event))
3690 perf_event_mmap_output(event, mmap_event);
3694 static void perf_event_mmap_event(struct perf_mmap_event *mmap_event)
3696 struct perf_cpu_context *cpuctx;
3697 struct perf_event_context *ctx;
3698 struct vm_area_struct *vma = mmap_event->vma;
3699 struct file *file = vma->vm_file;
3705 memset(tmp, 0, sizeof(tmp));
3709 * d_path works from the end of the buffer backwards, so we
3710 * need to add enough zero bytes after the string to handle
3711 * the 64bit alignment we do later.
3713 buf = kzalloc(PATH_MAX + sizeof(u64), GFP_KERNEL);
3715 name = strncpy(tmp, "//enomem", sizeof(tmp));
3718 name = d_path(&file->f_path, buf, PATH_MAX);
3720 name = strncpy(tmp, "//toolong", sizeof(tmp));
3724 if (arch_vma_name(mmap_event->vma)) {
3725 name = strncpy(tmp, arch_vma_name(mmap_event->vma),
3731 name = strncpy(tmp, "[vdso]", sizeof(tmp));
3735 name = strncpy(tmp, "//anon", sizeof(tmp));
3740 size = ALIGN(strlen(name)+1, sizeof(u64));
3742 mmap_event->file_name = name;
3743 mmap_event->file_size = size;
3745 mmap_event->event_id.header.size = sizeof(mmap_event->event_id) + size;
3748 cpuctx = &get_cpu_var(perf_cpu_context);
3749 perf_event_mmap_ctx(&cpuctx->ctx, mmap_event);
3750 ctx = rcu_dereference(current->perf_event_ctxp);
3752 perf_event_mmap_ctx(ctx, mmap_event);
3753 put_cpu_var(perf_cpu_context);
3759 void __perf_event_mmap(struct vm_area_struct *vma)
3761 struct perf_mmap_event mmap_event;
3763 if (!atomic_read(&nr_mmap_events))
3766 mmap_event = (struct perf_mmap_event){
3772 .type = PERF_RECORD_MMAP,
3773 .misc = PERF_RECORD_MISC_USER,
3778 .start = vma->vm_start,
3779 .len = vma->vm_end - vma->vm_start,
3780 .pgoff = (u64)vma->vm_pgoff << PAGE_SHIFT,
3784 perf_event_mmap_event(&mmap_event);
3788 * IRQ throttle logging
3791 static void perf_log_throttle(struct perf_event *event, int enable)
3793 struct perf_output_handle handle;
3797 struct perf_event_header header;
3801 } throttle_event = {
3803 .type = PERF_RECORD_THROTTLE,
3805 .size = sizeof(throttle_event),
3807 .time = perf_clock(),
3808 .id = primary_event_id(event),
3809 .stream_id = event->id,
3813 throttle_event.header.type = PERF_RECORD_UNTHROTTLE;
3815 ret = perf_output_begin(&handle, event, sizeof(throttle_event), 1, 0);
3819 perf_output_put(&handle, throttle_event);
3820 perf_output_end(&handle);
3824 * Generic event overflow handling, sampling.
3827 static int __perf_event_overflow(struct perf_event *event, int nmi,
3828 int throttle, struct perf_sample_data *data,
3829 struct pt_regs *regs)
3831 int events = atomic_read(&event->event_limit);
3832 struct hw_perf_event *hwc = &event->hw;
3835 throttle = (throttle && event->pmu->unthrottle != NULL);
3840 if (hwc->interrupts != MAX_INTERRUPTS) {
3842 if (HZ * hwc->interrupts >
3843 (u64)sysctl_perf_event_sample_rate) {
3844 hwc->interrupts = MAX_INTERRUPTS;
3845 perf_log_throttle(event, 0);
3850 * Keep re-disabling events even though on the previous
3851 * pass we disabled it - just in case we raced with a
3852 * sched-in and the event got enabled again:
3858 if (event->attr.freq) {
3859 u64 now = perf_clock();
3860 s64 delta = now - hwc->freq_time_stamp;
3862 hwc->freq_time_stamp = now;
3864 if (delta > 0 && delta < 2*TICK_NSEC)
3865 perf_adjust_period(event, delta, hwc->last_period);
3869 * XXX event_limit might not quite work as expected on inherited
3873 event->pending_kill = POLL_IN;
3874 if (events && atomic_dec_and_test(&event->event_limit)) {
3876 event->pending_kill = POLL_HUP;
3878 event->pending_disable = 1;
3879 perf_pending_queue(&event->pending,
3880 perf_pending_event);
3882 perf_event_disable(event);
3885 if (event->overflow_handler)
3886 event->overflow_handler(event, nmi, data, regs);
3888 perf_event_output(event, nmi, data, regs);
3893 int perf_event_overflow(struct perf_event *event, int nmi,
3894 struct perf_sample_data *data,
3895 struct pt_regs *regs)
3897 return __perf_event_overflow(event, nmi, 1, data, regs);
3901 * Generic software event infrastructure
3905 * We directly increment event->count and keep a second value in
3906 * event->hw.period_left to count intervals. This period event
3907 * is kept in the range [-sample_period, 0] so that we can use the
3911 static u64 perf_swevent_set_period(struct perf_event *event)
3913 struct hw_perf_event *hwc = &event->hw;
3914 u64 period = hwc->last_period;
3918 hwc->last_period = hwc->sample_period;
3921 old = val = atomic64_read(&hwc->period_left);
3925 nr = div64_u64(period + val, period);
3926 offset = nr * period;
3928 if (atomic64_cmpxchg(&hwc->period_left, old, val) != old)
3934 static void perf_swevent_overflow(struct perf_event *event, u64 overflow,
3935 int nmi, struct perf_sample_data *data,
3936 struct pt_regs *regs)
3938 struct hw_perf_event *hwc = &event->hw;
3941 data->period = event->hw.last_period;
3943 overflow = perf_swevent_set_period(event);
3945 if (hwc->interrupts == MAX_INTERRUPTS)
3948 for (; overflow; overflow--) {
3949 if (__perf_event_overflow(event, nmi, throttle,
3952 * We inhibit the overflow from happening when
3953 * hwc->interrupts == MAX_INTERRUPTS.
3961 static void perf_swevent_unthrottle(struct perf_event *event)
3964 * Nothing to do, we already reset hwc->interrupts.
3968 static void perf_swevent_add(struct perf_event *event, u64 nr,
3969 int nmi, struct perf_sample_data *data,
3970 struct pt_regs *regs)
3972 struct hw_perf_event *hwc = &event->hw;
3974 atomic64_add(nr, &event->count);
3979 if (!hwc->sample_period)
3982 if (nr == 1 && hwc->sample_period == 1 && !event->attr.freq)
3983 return perf_swevent_overflow(event, 1, nmi, data, regs);
3985 if (atomic64_add_negative(nr, &hwc->period_left))
3988 perf_swevent_overflow(event, 0, nmi, data, regs);
3991 static int perf_tp_event_match(struct perf_event *event,
3992 struct perf_sample_data *data);
3994 static int perf_exclude_event(struct perf_event *event,
3995 struct pt_regs *regs)
3998 if (event->attr.exclude_user && user_mode(regs))
4001 if (event->attr.exclude_kernel && !user_mode(regs))
4008 static int perf_swevent_match(struct perf_event *event,
4009 enum perf_type_id type,
4011 struct perf_sample_data *data,
4012 struct pt_regs *regs)
4014 if (event->attr.type != type)
4017 if (event->attr.config != event_id)
4020 if (perf_exclude_event(event, regs))
4023 if (event->attr.type == PERF_TYPE_TRACEPOINT &&
4024 !perf_tp_event_match(event, data))
4030 static inline u64 swevent_hash(u64 type, u32 event_id)
4032 u64 val = event_id | (type << 32);
4034 return hash_64(val, SWEVENT_HLIST_BITS);
4037 static struct hlist_head *
4038 find_swevent_head(struct perf_cpu_context *ctx, u64 type, u32 event_id)
4041 struct swevent_hlist *hlist;
4043 hash = swevent_hash(type, event_id);
4045 hlist = rcu_dereference(ctx->swevent_hlist);
4049 return &hlist->heads[hash];
4052 static void do_perf_sw_event(enum perf_type_id type, u32 event_id,
4054 struct perf_sample_data *data,
4055 struct pt_regs *regs)
4057 struct perf_cpu_context *cpuctx;
4058 struct perf_event *event;
4059 struct hlist_node *node;
4060 struct hlist_head *head;
4062 cpuctx = &__get_cpu_var(perf_cpu_context);
4066 head = find_swevent_head(cpuctx, type, event_id);
4071 hlist_for_each_entry_rcu(event, node, head, hlist_entry) {
4072 if (perf_swevent_match(event, type, event_id, data, regs))
4073 perf_swevent_add(event, nr, nmi, data, regs);
4079 int perf_swevent_get_recursion_context(void)
4081 struct perf_cpu_context *cpuctx = &get_cpu_var(perf_cpu_context);
4088 else if (in_softirq())
4093 if (cpuctx->recursion[rctx]) {
4094 put_cpu_var(perf_cpu_context);
4098 cpuctx->recursion[rctx]++;
4103 EXPORT_SYMBOL_GPL(perf_swevent_get_recursion_context);
4105 void perf_swevent_put_recursion_context(int rctx)
4107 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
4109 cpuctx->recursion[rctx]--;
4110 put_cpu_var(perf_cpu_context);
4112 EXPORT_SYMBOL_GPL(perf_swevent_put_recursion_context);
4115 void __perf_sw_event(u32 event_id, u64 nr, int nmi,
4116 struct pt_regs *regs, u64 addr)
4118 struct perf_sample_data data;
4121 rctx = perf_swevent_get_recursion_context();
4125 perf_sample_data_init(&data, addr);
4127 do_perf_sw_event(PERF_TYPE_SOFTWARE, event_id, nr, nmi, &data, regs);
4129 perf_swevent_put_recursion_context(rctx);
4132 static void perf_swevent_read(struct perf_event *event)
4136 static int perf_swevent_enable(struct perf_event *event)
4138 struct hw_perf_event *hwc = &event->hw;
4139 struct perf_cpu_context *cpuctx;
4140 struct hlist_head *head;
4142 cpuctx = &__get_cpu_var(perf_cpu_context);
4144 if (hwc->sample_period) {
4145 hwc->last_period = hwc->sample_period;
4146 perf_swevent_set_period(event);
4149 head = find_swevent_head(cpuctx, event->attr.type, event->attr.config);
4150 if (WARN_ON_ONCE(!head))
4153 hlist_add_head_rcu(&event->hlist_entry, head);
4158 static void perf_swevent_disable(struct perf_event *event)
4160 hlist_del_rcu(&event->hlist_entry);
4163 static const struct pmu perf_ops_generic = {
4164 .enable = perf_swevent_enable,
4165 .disable = perf_swevent_disable,
4166 .read = perf_swevent_read,
4167 .unthrottle = perf_swevent_unthrottle,
4171 * hrtimer based swevent callback
4174 static enum hrtimer_restart perf_swevent_hrtimer(struct hrtimer *hrtimer)
4176 enum hrtimer_restart ret = HRTIMER_RESTART;
4177 struct perf_sample_data data;
4178 struct pt_regs *regs;
4179 struct perf_event *event;
4182 event = container_of(hrtimer, struct perf_event, hw.hrtimer);
4183 event->pmu->read(event);
4185 perf_sample_data_init(&data, 0);
4186 data.period = event->hw.last_period;
4187 regs = get_irq_regs();
4189 if (regs && !perf_exclude_event(event, regs)) {
4190 if (!(event->attr.exclude_idle && current->pid == 0))
4191 if (perf_event_overflow(event, 0, &data, regs))
4192 ret = HRTIMER_NORESTART;
4195 period = max_t(u64, 10000, event->hw.sample_period);
4196 hrtimer_forward_now(hrtimer, ns_to_ktime(period));
4201 static void perf_swevent_start_hrtimer(struct perf_event *event)
4203 struct hw_perf_event *hwc = &event->hw;
4205 hrtimer_init(&hwc->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
4206 hwc->hrtimer.function = perf_swevent_hrtimer;
4207 if (hwc->sample_period) {
4210 if (hwc->remaining) {
4211 if (hwc->remaining < 0)
4214 period = hwc->remaining;
4217 period = max_t(u64, 10000, hwc->sample_period);
4219 __hrtimer_start_range_ns(&hwc->hrtimer,
4220 ns_to_ktime(period), 0,
4221 HRTIMER_MODE_REL, 0);
4225 static void perf_swevent_cancel_hrtimer(struct perf_event *event)
4227 struct hw_perf_event *hwc = &event->hw;
4229 if (hwc->sample_period) {
4230 ktime_t remaining = hrtimer_get_remaining(&hwc->hrtimer);
4231 hwc->remaining = ktime_to_ns(remaining);
4233 hrtimer_cancel(&hwc->hrtimer);
4238 * Software event: cpu wall time clock
4241 static void cpu_clock_perf_event_update(struct perf_event *event)
4243 int cpu = raw_smp_processor_id();
4247 now = cpu_clock(cpu);
4248 prev = atomic64_xchg(&event->hw.prev_count, now);
4249 atomic64_add(now - prev, &event->count);
4252 static int cpu_clock_perf_event_enable(struct perf_event *event)
4254 struct hw_perf_event *hwc = &event->hw;
4255 int cpu = raw_smp_processor_id();
4257 atomic64_set(&hwc->prev_count, cpu_clock(cpu));
4258 perf_swevent_start_hrtimer(event);
4263 static void cpu_clock_perf_event_disable(struct perf_event *event)
4265 perf_swevent_cancel_hrtimer(event);
4266 cpu_clock_perf_event_update(event);
4269 static void cpu_clock_perf_event_read(struct perf_event *event)
4271 cpu_clock_perf_event_update(event);
4274 static const struct pmu perf_ops_cpu_clock = {
4275 .enable = cpu_clock_perf_event_enable,
4276 .disable = cpu_clock_perf_event_disable,
4277 .read = cpu_clock_perf_event_read,
4281 * Software event: task time clock
4284 static void task_clock_perf_event_update(struct perf_event *event, u64 now)
4289 prev = atomic64_xchg(&event->hw.prev_count, now);
4291 atomic64_add(delta, &event->count);
4294 static int task_clock_perf_event_enable(struct perf_event *event)
4296 struct hw_perf_event *hwc = &event->hw;
4299 now = event->ctx->time;
4301 atomic64_set(&hwc->prev_count, now);
4303 perf_swevent_start_hrtimer(event);
4308 static void task_clock_perf_event_disable(struct perf_event *event)
4310 perf_swevent_cancel_hrtimer(event);
4311 task_clock_perf_event_update(event, event->ctx->time);
4315 static void task_clock_perf_event_read(struct perf_event *event)
4320 update_context_time(event->ctx);
4321 time = event->ctx->time;
4323 u64 now = perf_clock();
4324 u64 delta = now - event->ctx->timestamp;
4325 time = event->ctx->time + delta;
4328 task_clock_perf_event_update(event, time);
4331 static const struct pmu perf_ops_task_clock = {
4332 .enable = task_clock_perf_event_enable,
4333 .disable = task_clock_perf_event_disable,
4334 .read = task_clock_perf_event_read,
4337 static void swevent_hlist_release_rcu(struct rcu_head *rcu_head)
4339 struct swevent_hlist *hlist;
4341 hlist = container_of(rcu_head, struct swevent_hlist, rcu_head);
4345 static void swevent_hlist_release(struct perf_cpu_context *cpuctx)
4347 struct swevent_hlist *hlist;
4349 if (!cpuctx->swevent_hlist)
4352 hlist = cpuctx->swevent_hlist;
4353 rcu_assign_pointer(cpuctx->swevent_hlist, NULL);
4354 call_rcu(&hlist->rcu_head, swevent_hlist_release_rcu);
4357 static void swevent_hlist_put_cpu(struct perf_event *event, int cpu)
4359 struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
4361 mutex_lock(&cpuctx->hlist_mutex);
4363 if (!--cpuctx->hlist_refcount)
4364 swevent_hlist_release(cpuctx);
4366 mutex_unlock(&cpuctx->hlist_mutex);
4369 static void swevent_hlist_put(struct perf_event *event)
4373 if (event->cpu != -1) {
4374 swevent_hlist_put_cpu(event, event->cpu);
4378 for_each_possible_cpu(cpu)
4379 swevent_hlist_put_cpu(event, cpu);
4382 static int swevent_hlist_get_cpu(struct perf_event *event, int cpu)
4384 struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
4387 mutex_lock(&cpuctx->hlist_mutex);
4389 if (!cpuctx->swevent_hlist && cpu_online(cpu)) {
4390 struct swevent_hlist *hlist;
4392 hlist = kzalloc(sizeof(*hlist), GFP_KERNEL);
4397 rcu_assign_pointer(cpuctx->swevent_hlist, hlist);
4399 cpuctx->hlist_refcount++;
4401 mutex_unlock(&cpuctx->hlist_mutex);
4406 static int swevent_hlist_get(struct perf_event *event)
4409 int cpu, failed_cpu;
4411 if (event->cpu != -1)
4412 return swevent_hlist_get_cpu(event, event->cpu);
4415 for_each_possible_cpu(cpu) {
4416 err = swevent_hlist_get_cpu(event, cpu);
4426 for_each_possible_cpu(cpu) {
4427 if (cpu == failed_cpu)
4429 swevent_hlist_put_cpu(event, cpu);
4436 #ifdef CONFIG_EVENT_TRACING
4438 void perf_tp_event(int event_id, u64 addr, u64 count, void *record,
4439 int entry_size, struct pt_regs *regs)
4441 struct perf_sample_data data;
4442 struct perf_raw_record raw = {
4447 perf_sample_data_init(&data, addr);
4450 /* Trace events already protected against recursion */
4451 do_perf_sw_event(PERF_TYPE_TRACEPOINT, event_id, count, 1,
4454 EXPORT_SYMBOL_GPL(perf_tp_event);
4456 static int perf_tp_event_match(struct perf_event *event,
4457 struct perf_sample_data *data)
4459 void *record = data->raw->data;
4461 if (likely(!event->filter) || filter_match_preds(event->filter, record))
4466 static void tp_perf_event_destroy(struct perf_event *event)
4468 perf_trace_disable(event->attr.config);
4469 swevent_hlist_put(event);
4472 static const struct pmu *tp_perf_event_init(struct perf_event *event)
4477 * Raw tracepoint data is a severe data leak, only allow root to
4480 if ((event->attr.sample_type & PERF_SAMPLE_RAW) &&
4481 perf_paranoid_tracepoint_raw() &&
4482 !capable(CAP_SYS_ADMIN))
4483 return ERR_PTR(-EPERM);
4485 if (perf_trace_enable(event->attr.config))
4488 event->destroy = tp_perf_event_destroy;
4489 err = swevent_hlist_get(event);
4491 perf_trace_disable(event->attr.config);
4492 return ERR_PTR(err);
4495 return &perf_ops_generic;
4498 static int perf_event_set_filter(struct perf_event *event, void __user *arg)
4503 if (event->attr.type != PERF_TYPE_TRACEPOINT)
4506 filter_str = strndup_user(arg, PAGE_SIZE);
4507 if (IS_ERR(filter_str))
4508 return PTR_ERR(filter_str);
4510 ret = ftrace_profile_set_filter(event, event->attr.config, filter_str);
4516 static void perf_event_free_filter(struct perf_event *event)
4518 ftrace_profile_free_filter(event);
4523 static int perf_tp_event_match(struct perf_event *event,
4524 struct perf_sample_data *data)
4529 static const struct pmu *tp_perf_event_init(struct perf_event *event)
4534 static int perf_event_set_filter(struct perf_event *event, void __user *arg)
4539 static void perf_event_free_filter(struct perf_event *event)
4543 #endif /* CONFIG_EVENT_TRACING */
4545 #ifdef CONFIG_HAVE_HW_BREAKPOINT
4546 static void bp_perf_event_destroy(struct perf_event *event)
4548 release_bp_slot(event);
4551 static const struct pmu *bp_perf_event_init(struct perf_event *bp)
4555 err = register_perf_hw_breakpoint(bp);
4557 return ERR_PTR(err);
4559 bp->destroy = bp_perf_event_destroy;
4561 return &perf_ops_bp;
4564 void perf_bp_event(struct perf_event *bp, void *data)
4566 struct perf_sample_data sample;
4567 struct pt_regs *regs = data;
4569 perf_sample_data_init(&sample, bp->attr.bp_addr);
4571 if (!perf_exclude_event(bp, regs))
4572 perf_swevent_add(bp, 1, 1, &sample, regs);
4575 static const struct pmu *bp_perf_event_init(struct perf_event *bp)
4580 void perf_bp_event(struct perf_event *bp, void *regs)
4585 atomic_t perf_swevent_enabled[PERF_COUNT_SW_MAX];
4587 static void sw_perf_event_destroy(struct perf_event *event)
4589 u64 event_id = event->attr.config;
4591 WARN_ON(event->parent);
4593 atomic_dec(&perf_swevent_enabled[event_id]);
4594 swevent_hlist_put(event);
4597 static const struct pmu *sw_perf_event_init(struct perf_event *event)
4599 const struct pmu *pmu = NULL;
4600 u64 event_id = event->attr.config;
4603 * Software events (currently) can't in general distinguish
4604 * between user, kernel and hypervisor events.
4605 * However, context switches and cpu migrations are considered
4606 * to be kernel events, and page faults are never hypervisor
4610 case PERF_COUNT_SW_CPU_CLOCK:
4611 pmu = &perf_ops_cpu_clock;
4614 case PERF_COUNT_SW_TASK_CLOCK:
4616 * If the user instantiates this as a per-cpu event,
4617 * use the cpu_clock event instead.
4619 if (event->ctx->task)
4620 pmu = &perf_ops_task_clock;
4622 pmu = &perf_ops_cpu_clock;
4625 case PERF_COUNT_SW_PAGE_FAULTS:
4626 case PERF_COUNT_SW_PAGE_FAULTS_MIN:
4627 case PERF_COUNT_SW_PAGE_FAULTS_MAJ:
4628 case PERF_COUNT_SW_CONTEXT_SWITCHES:
4629 case PERF_COUNT_SW_CPU_MIGRATIONS:
4630 case PERF_COUNT_SW_ALIGNMENT_FAULTS:
4631 case PERF_COUNT_SW_EMULATION_FAULTS:
4632 if (!event->parent) {
4635 err = swevent_hlist_get(event);
4637 return ERR_PTR(err);
4639 atomic_inc(&perf_swevent_enabled[event_id]);
4640 event->destroy = sw_perf_event_destroy;
4642 pmu = &perf_ops_generic;
4650 * Allocate and initialize a event structure
4652 static struct perf_event *
4653 perf_event_alloc(struct perf_event_attr *attr,
4655 struct perf_event_context *ctx,
4656 struct perf_event *group_leader,
4657 struct perf_event *parent_event,
4658 perf_overflow_handler_t overflow_handler,
4661 const struct pmu *pmu;
4662 struct perf_event *event;
4663 struct hw_perf_event *hwc;
4666 event = kzalloc(sizeof(*event), gfpflags);
4668 return ERR_PTR(-ENOMEM);
4671 * Single events are their own group leaders, with an
4672 * empty sibling list:
4675 group_leader = event;
4677 mutex_init(&event->child_mutex);
4678 INIT_LIST_HEAD(&event->child_list);
4680 INIT_LIST_HEAD(&event->group_entry);
4681 INIT_LIST_HEAD(&event->event_entry);
4682 INIT_LIST_HEAD(&event->sibling_list);
4683 init_waitqueue_head(&event->waitq);
4685 mutex_init(&event->mmap_mutex);
4688 event->attr = *attr;
4689 event->group_leader = group_leader;
4694 event->parent = parent_event;
4696 event->ns = get_pid_ns(current->nsproxy->pid_ns);
4697 event->id = atomic64_inc_return(&perf_event_id);
4699 event->state = PERF_EVENT_STATE_INACTIVE;
4701 if (!overflow_handler && parent_event)
4702 overflow_handler = parent_event->overflow_handler;
4704 event->overflow_handler = overflow_handler;
4707 event->state = PERF_EVENT_STATE_OFF;
4712 hwc->sample_period = attr->sample_period;
4713 if (attr->freq && attr->sample_freq)
4714 hwc->sample_period = 1;
4715 hwc->last_period = hwc->sample_period;
4717 atomic64_set(&hwc->period_left, hwc->sample_period);
4720 * we currently do not support PERF_FORMAT_GROUP on inherited events
4722 if (attr->inherit && (attr->read_format & PERF_FORMAT_GROUP))
4725 switch (attr->type) {
4727 case PERF_TYPE_HARDWARE:
4728 case PERF_TYPE_HW_CACHE:
4729 pmu = hw_perf_event_init(event);
4732 case PERF_TYPE_SOFTWARE:
4733 pmu = sw_perf_event_init(event);
4736 case PERF_TYPE_TRACEPOINT:
4737 pmu = tp_perf_event_init(event);
4740 case PERF_TYPE_BREAKPOINT:
4741 pmu = bp_perf_event_init(event);
4752 else if (IS_ERR(pmu))
4757 put_pid_ns(event->ns);
4759 return ERR_PTR(err);
4764 if (!event->parent) {
4765 atomic_inc(&nr_events);
4766 if (event->attr.mmap)
4767 atomic_inc(&nr_mmap_events);
4768 if (event->attr.comm)
4769 atomic_inc(&nr_comm_events);
4770 if (event->attr.task)
4771 atomic_inc(&nr_task_events);
4777 static int perf_copy_attr(struct perf_event_attr __user *uattr,
4778 struct perf_event_attr *attr)
4783 if (!access_ok(VERIFY_WRITE, uattr, PERF_ATTR_SIZE_VER0))
4787 * zero the full structure, so that a short copy will be nice.
4789 memset(attr, 0, sizeof(*attr));
4791 ret = get_user(size, &uattr->size);
4795 if (size > PAGE_SIZE) /* silly large */
4798 if (!size) /* abi compat */
4799 size = PERF_ATTR_SIZE_VER0;
4801 if (size < PERF_ATTR_SIZE_VER0)
4805 * If we're handed a bigger struct than we know of,
4806 * ensure all the unknown bits are 0 - i.e. new
4807 * user-space does not rely on any kernel feature
4808 * extensions we dont know about yet.
4810 if (size > sizeof(*attr)) {
4811 unsigned char __user *addr;
4812 unsigned char __user *end;
4815 addr = (void __user *)uattr + sizeof(*attr);
4816 end = (void __user *)uattr + size;
4818 for (; addr < end; addr++) {
4819 ret = get_user(val, addr);
4825 size = sizeof(*attr);
4828 ret = copy_from_user(attr, uattr, size);
4833 * If the type exists, the corresponding creation will verify
4836 if (attr->type >= PERF_TYPE_MAX)
4839 if (attr->__reserved_1)
4842 if (attr->sample_type & ~(PERF_SAMPLE_MAX-1))
4845 if (attr->read_format & ~(PERF_FORMAT_MAX-1))
4852 put_user(sizeof(*attr), &uattr->size);
4857 static int perf_event_set_output(struct perf_event *event, int output_fd)
4859 struct perf_event *output_event = NULL;
4860 struct file *output_file = NULL;
4861 struct perf_event *old_output;
4862 int fput_needed = 0;
4868 output_file = fget_light(output_fd, &fput_needed);
4872 if (output_file->f_op != &perf_fops)
4875 output_event = output_file->private_data;
4877 /* Don't chain output fds */
4878 if (output_event->output)
4881 /* Don't set an output fd when we already have an output channel */
4885 atomic_long_inc(&output_file->f_count);
4888 mutex_lock(&event->mmap_mutex);
4889 old_output = event->output;
4890 rcu_assign_pointer(event->output, output_event);
4891 mutex_unlock(&event->mmap_mutex);
4895 * we need to make sure no existing perf_output_*()
4896 * is still referencing this event.
4899 fput(old_output->filp);
4904 fput_light(output_file, fput_needed);
4909 * sys_perf_event_open - open a performance event, associate it to a task/cpu
4911 * @attr_uptr: event_id type attributes for monitoring/sampling
4914 * @group_fd: group leader event fd
4916 SYSCALL_DEFINE5(perf_event_open,
4917 struct perf_event_attr __user *, attr_uptr,
4918 pid_t, pid, int, cpu, int, group_fd, unsigned long, flags)
4920 struct perf_event *event, *group_leader;
4921 struct perf_event_attr attr;
4922 struct perf_event_context *ctx;
4923 struct file *event_file = NULL;
4924 struct file *group_file = NULL;
4925 int fput_needed = 0;
4926 int fput_needed2 = 0;
4929 /* for future expandability... */
4930 if (flags & ~(PERF_FLAG_FD_NO_GROUP | PERF_FLAG_FD_OUTPUT))
4933 err = perf_copy_attr(attr_uptr, &attr);
4937 if (!attr.exclude_kernel) {
4938 if (perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN))
4943 if (attr.sample_freq > sysctl_perf_event_sample_rate)
4948 * Get the target context (task or percpu):
4950 ctx = find_get_context(pid, cpu);
4952 return PTR_ERR(ctx);
4955 * Look up the group leader (we will attach this event to it):
4957 group_leader = NULL;
4958 if (group_fd != -1 && !(flags & PERF_FLAG_FD_NO_GROUP)) {
4960 group_file = fget_light(group_fd, &fput_needed);
4962 goto err_put_context;
4963 if (group_file->f_op != &perf_fops)
4964 goto err_put_context;
4966 group_leader = group_file->private_data;
4968 * Do not allow a recursive hierarchy (this new sibling
4969 * becoming part of another group-sibling):
4971 if (group_leader->group_leader != group_leader)
4972 goto err_put_context;
4974 * Do not allow to attach to a group in a different
4975 * task or CPU context:
4977 if (group_leader->ctx != ctx)
4978 goto err_put_context;
4980 * Only a group leader can be exclusive or pinned
4982 if (attr.exclusive || attr.pinned)
4983 goto err_put_context;
4986 event = perf_event_alloc(&attr, cpu, ctx, group_leader,
4987 NULL, NULL, GFP_KERNEL);
4988 err = PTR_ERR(event);
4990 goto err_put_context;
4992 err = anon_inode_getfd("[perf_event]", &perf_fops, event, O_RDWR);
4994 goto err_free_put_context;
4996 event_file = fget_light(err, &fput_needed2);
4998 goto err_free_put_context;
5000 if (flags & PERF_FLAG_FD_OUTPUT) {
5001 err = perf_event_set_output(event, group_fd);
5003 goto err_fput_free_put_context;
5006 event->filp = event_file;
5007 WARN_ON_ONCE(ctx->parent_ctx);
5008 mutex_lock(&ctx->mutex);
5009 perf_install_in_context(ctx, event, cpu);
5011 mutex_unlock(&ctx->mutex);
5013 event->owner = current;
5014 get_task_struct(current);
5015 mutex_lock(¤t->perf_event_mutex);
5016 list_add_tail(&event->owner_entry, ¤t->perf_event_list);
5017 mutex_unlock(¤t->perf_event_mutex);
5019 err_fput_free_put_context:
5020 fput_light(event_file, fput_needed2);
5022 err_free_put_context:
5030 fput_light(group_file, fput_needed);
5036 * perf_event_create_kernel_counter
5038 * @attr: attributes of the counter to create
5039 * @cpu: cpu in which the counter is bound
5040 * @pid: task to profile
5043 perf_event_create_kernel_counter(struct perf_event_attr *attr, int cpu,
5045 perf_overflow_handler_t overflow_handler)
5047 struct perf_event *event;
5048 struct perf_event_context *ctx;
5052 * Get the target context (task or percpu):
5055 ctx = find_get_context(pid, cpu);
5061 event = perf_event_alloc(attr, cpu, ctx, NULL,
5062 NULL, overflow_handler, GFP_KERNEL);
5063 if (IS_ERR(event)) {
5064 err = PTR_ERR(event);
5065 goto err_put_context;
5069 WARN_ON_ONCE(ctx->parent_ctx);
5070 mutex_lock(&ctx->mutex);
5071 perf_install_in_context(ctx, event, cpu);
5073 mutex_unlock(&ctx->mutex);
5075 event->owner = current;
5076 get_task_struct(current);
5077 mutex_lock(¤t->perf_event_mutex);
5078 list_add_tail(&event->owner_entry, ¤t->perf_event_list);
5079 mutex_unlock(¤t->perf_event_mutex);
5086 return ERR_PTR(err);
5088 EXPORT_SYMBOL_GPL(perf_event_create_kernel_counter);
5091 * inherit a event from parent task to child task:
5093 static struct perf_event *
5094 inherit_event(struct perf_event *parent_event,
5095 struct task_struct *parent,
5096 struct perf_event_context *parent_ctx,
5097 struct task_struct *child,
5098 struct perf_event *group_leader,
5099 struct perf_event_context *child_ctx)
5101 struct perf_event *child_event;
5104 * Instead of creating recursive hierarchies of events,
5105 * we link inherited events back to the original parent,
5106 * which has a filp for sure, which we use as the reference
5109 if (parent_event->parent)
5110 parent_event = parent_event->parent;
5112 child_event = perf_event_alloc(&parent_event->attr,
5113 parent_event->cpu, child_ctx,
5114 group_leader, parent_event,
5116 if (IS_ERR(child_event))
5121 * Make the child state follow the state of the parent event,
5122 * not its attr.disabled bit. We hold the parent's mutex,
5123 * so we won't race with perf_event_{en, dis}able_family.
5125 if (parent_event->state >= PERF_EVENT_STATE_INACTIVE)
5126 child_event->state = PERF_EVENT_STATE_INACTIVE;
5128 child_event->state = PERF_EVENT_STATE_OFF;
5130 if (parent_event->attr.freq) {
5131 u64 sample_period = parent_event->hw.sample_period;
5132 struct hw_perf_event *hwc = &child_event->hw;
5134 hwc->sample_period = sample_period;
5135 hwc->last_period = sample_period;
5137 atomic64_set(&hwc->period_left, sample_period);
5140 child_event->overflow_handler = parent_event->overflow_handler;
5143 * Link it up in the child's context:
5145 add_event_to_ctx(child_event, child_ctx);
5148 * Get a reference to the parent filp - we will fput it
5149 * when the child event exits. This is safe to do because
5150 * we are in the parent and we know that the filp still
5151 * exists and has a nonzero count:
5153 atomic_long_inc(&parent_event->filp->f_count);
5156 * Link this into the parent event's child list
5158 WARN_ON_ONCE(parent_event->ctx->parent_ctx);
5159 mutex_lock(&parent_event->child_mutex);
5160 list_add_tail(&child_event->child_list, &parent_event->child_list);
5161 mutex_unlock(&parent_event->child_mutex);
5166 static int inherit_group(struct perf_event *parent_event,
5167 struct task_struct *parent,
5168 struct perf_event_context *parent_ctx,
5169 struct task_struct *child,
5170 struct perf_event_context *child_ctx)
5172 struct perf_event *leader;
5173 struct perf_event *sub;
5174 struct perf_event *child_ctr;
5176 leader = inherit_event(parent_event, parent, parent_ctx,
5177 child, NULL, child_ctx);
5179 return PTR_ERR(leader);
5180 list_for_each_entry(sub, &parent_event->sibling_list, group_entry) {
5181 child_ctr = inherit_event(sub, parent, parent_ctx,
5182 child, leader, child_ctx);
5183 if (IS_ERR(child_ctr))
5184 return PTR_ERR(child_ctr);
5189 static void sync_child_event(struct perf_event *child_event,
5190 struct task_struct *child)
5192 struct perf_event *parent_event = child_event->parent;
5195 if (child_event->attr.inherit_stat)
5196 perf_event_read_event(child_event, child);
5198 child_val = atomic64_read(&child_event->count);
5201 * Add back the child's count to the parent's count:
5203 atomic64_add(child_val, &parent_event->count);
5204 atomic64_add(child_event->total_time_enabled,
5205 &parent_event->child_total_time_enabled);
5206 atomic64_add(child_event->total_time_running,
5207 &parent_event->child_total_time_running);
5210 * Remove this event from the parent's list
5212 WARN_ON_ONCE(parent_event->ctx->parent_ctx);
5213 mutex_lock(&parent_event->child_mutex);
5214 list_del_init(&child_event->child_list);
5215 mutex_unlock(&parent_event->child_mutex);
5218 * Release the parent event, if this was the last
5221 fput(parent_event->filp);
5225 __perf_event_exit_task(struct perf_event *child_event,
5226 struct perf_event_context *child_ctx,
5227 struct task_struct *child)
5229 struct perf_event *parent_event;
5231 perf_event_remove_from_context(child_event);
5233 parent_event = child_event->parent;
5235 * It can happen that parent exits first, and has events
5236 * that are still around due to the child reference. These
5237 * events need to be zapped - but otherwise linger.
5240 sync_child_event(child_event, child);
5241 free_event(child_event);
5246 * When a child task exits, feed back event values to parent events.
5248 void perf_event_exit_task(struct task_struct *child)
5250 struct perf_event *child_event, *tmp;
5251 struct perf_event_context *child_ctx;
5252 unsigned long flags;
5254 if (likely(!child->perf_event_ctxp)) {
5255 perf_event_task(child, NULL, 0);
5259 local_irq_save(flags);
5261 * We can't reschedule here because interrupts are disabled,
5262 * and either child is current or it is a task that can't be
5263 * scheduled, so we are now safe from rescheduling changing
5266 child_ctx = child->perf_event_ctxp;
5267 __perf_event_task_sched_out(child_ctx);
5270 * Take the context lock here so that if find_get_context is
5271 * reading child->perf_event_ctxp, we wait until it has
5272 * incremented the context's refcount before we do put_ctx below.
5274 raw_spin_lock(&child_ctx->lock);
5275 child->perf_event_ctxp = NULL;
5277 * If this context is a clone; unclone it so it can't get
5278 * swapped to another process while we're removing all
5279 * the events from it.
5281 unclone_ctx(child_ctx);
5282 update_context_time(child_ctx);
5283 raw_spin_unlock_irqrestore(&child_ctx->lock, flags);
5286 * Report the task dead after unscheduling the events so that we
5287 * won't get any samples after PERF_RECORD_EXIT. We can however still
5288 * get a few PERF_RECORD_READ events.
5290 perf_event_task(child, child_ctx, 0);
5293 * We can recurse on the same lock type through:
5295 * __perf_event_exit_task()
5296 * sync_child_event()
5297 * fput(parent_event->filp)
5299 * mutex_lock(&ctx->mutex)
5301 * But since its the parent context it won't be the same instance.
5303 mutex_lock_nested(&child_ctx->mutex, SINGLE_DEPTH_NESTING);
5306 list_for_each_entry_safe(child_event, tmp, &child_ctx->pinned_groups,
5308 __perf_event_exit_task(child_event, child_ctx, child);
5310 list_for_each_entry_safe(child_event, tmp, &child_ctx->flexible_groups,
5312 __perf_event_exit_task(child_event, child_ctx, child);
5315 * If the last event was a group event, it will have appended all
5316 * its siblings to the list, but we obtained 'tmp' before that which
5317 * will still point to the list head terminating the iteration.
5319 if (!list_empty(&child_ctx->pinned_groups) ||
5320 !list_empty(&child_ctx->flexible_groups))
5323 mutex_unlock(&child_ctx->mutex);
5328 static void perf_free_event(struct perf_event *event,
5329 struct perf_event_context *ctx)
5331 struct perf_event *parent = event->parent;
5333 if (WARN_ON_ONCE(!parent))
5336 mutex_lock(&parent->child_mutex);
5337 list_del_init(&event->child_list);
5338 mutex_unlock(&parent->child_mutex);
5342 list_del_event(event, ctx);
5347 * free an unexposed, unused context as created by inheritance by
5348 * init_task below, used by fork() in case of fail.
5350 void perf_event_free_task(struct task_struct *task)
5352 struct perf_event_context *ctx = task->perf_event_ctxp;
5353 struct perf_event *event, *tmp;
5358 mutex_lock(&ctx->mutex);
5360 list_for_each_entry_safe(event, tmp, &ctx->pinned_groups, group_entry)
5361 perf_free_event(event, ctx);
5363 list_for_each_entry_safe(event, tmp, &ctx->flexible_groups,
5365 perf_free_event(event, ctx);
5367 if (!list_empty(&ctx->pinned_groups) ||
5368 !list_empty(&ctx->flexible_groups))
5371 mutex_unlock(&ctx->mutex);
5377 inherit_task_group(struct perf_event *event, struct task_struct *parent,
5378 struct perf_event_context *parent_ctx,
5379 struct task_struct *child,
5383 struct perf_event_context *child_ctx = child->perf_event_ctxp;
5385 if (!event->attr.inherit) {
5392 * This is executed from the parent task context, so
5393 * inherit events that have been marked for cloning.
5394 * First allocate and initialize a context for the
5398 child_ctx = kzalloc(sizeof(struct perf_event_context),
5403 __perf_event_init_context(child_ctx, child);
5404 child->perf_event_ctxp = child_ctx;
5405 get_task_struct(child);
5408 ret = inherit_group(event, parent, parent_ctx,
5419 * Initialize the perf_event context in task_struct
5421 int perf_event_init_task(struct task_struct *child)
5423 struct perf_event_context *child_ctx, *parent_ctx;
5424 struct perf_event_context *cloned_ctx;
5425 struct perf_event *event;
5426 struct task_struct *parent = current;
5427 int inherited_all = 1;
5430 child->perf_event_ctxp = NULL;
5432 mutex_init(&child->perf_event_mutex);
5433 INIT_LIST_HEAD(&child->perf_event_list);
5435 if (likely(!parent->perf_event_ctxp))
5439 * If the parent's context is a clone, pin it so it won't get
5442 parent_ctx = perf_pin_task_context(parent);
5445 * No need to check if parent_ctx != NULL here; since we saw
5446 * it non-NULL earlier, the only reason for it to become NULL
5447 * is if we exit, and since we're currently in the middle of
5448 * a fork we can't be exiting at the same time.
5452 * Lock the parent list. No need to lock the child - not PID
5453 * hashed yet and not running, so nobody can access it.
5455 mutex_lock(&parent_ctx->mutex);
5458 * We dont have to disable NMIs - we are only looking at
5459 * the list, not manipulating it:
5461 list_for_each_entry(event, &parent_ctx->pinned_groups, group_entry) {
5462 ret = inherit_task_group(event, parent, parent_ctx, child,
5468 list_for_each_entry(event, &parent_ctx->flexible_groups, group_entry) {
5469 ret = inherit_task_group(event, parent, parent_ctx, child,
5475 child_ctx = child->perf_event_ctxp;
5477 if (child_ctx && inherited_all) {
5479 * Mark the child context as a clone of the parent
5480 * context, or of whatever the parent is a clone of.
5481 * Note that if the parent is a clone, it could get
5482 * uncloned at any point, but that doesn't matter
5483 * because the list of events and the generation
5484 * count can't have changed since we took the mutex.
5486 cloned_ctx = rcu_dereference(parent_ctx->parent_ctx);
5488 child_ctx->parent_ctx = cloned_ctx;
5489 child_ctx->parent_gen = parent_ctx->parent_gen;
5491 child_ctx->parent_ctx = parent_ctx;
5492 child_ctx->parent_gen = parent_ctx->generation;
5494 get_ctx(child_ctx->parent_ctx);
5497 mutex_unlock(&parent_ctx->mutex);
5499 perf_unpin_context(parent_ctx);
5504 static void __init perf_event_init_all_cpus(void)
5507 struct perf_cpu_context *cpuctx;
5509 for_each_possible_cpu(cpu) {
5510 cpuctx = &per_cpu(perf_cpu_context, cpu);
5511 mutex_init(&cpuctx->hlist_mutex);
5512 __perf_event_init_context(&cpuctx->ctx, NULL);
5516 static void __cpuinit perf_event_init_cpu(int cpu)
5518 struct perf_cpu_context *cpuctx;
5520 cpuctx = &per_cpu(perf_cpu_context, cpu);
5522 spin_lock(&perf_resource_lock);
5523 cpuctx->max_pertask = perf_max_events - perf_reserved_percpu;
5524 spin_unlock(&perf_resource_lock);
5526 mutex_lock(&cpuctx->hlist_mutex);
5527 if (cpuctx->hlist_refcount > 0) {
5528 struct swevent_hlist *hlist;
5530 hlist = kzalloc(sizeof(*hlist), GFP_KERNEL);
5531 WARN_ON_ONCE(!hlist);
5532 rcu_assign_pointer(cpuctx->swevent_hlist, hlist);
5534 mutex_unlock(&cpuctx->hlist_mutex);
5537 #ifdef CONFIG_HOTPLUG_CPU
5538 static void __perf_event_exit_cpu(void *info)
5540 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
5541 struct perf_event_context *ctx = &cpuctx->ctx;
5542 struct perf_event *event, *tmp;
5544 list_for_each_entry_safe(event, tmp, &ctx->pinned_groups, group_entry)
5545 __perf_event_remove_from_context(event);
5546 list_for_each_entry_safe(event, tmp, &ctx->flexible_groups, group_entry)
5547 __perf_event_remove_from_context(event);
5549 static void perf_event_exit_cpu(int cpu)
5551 struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
5552 struct perf_event_context *ctx = &cpuctx->ctx;
5554 mutex_lock(&cpuctx->hlist_mutex);
5555 swevent_hlist_release(cpuctx);
5556 mutex_unlock(&cpuctx->hlist_mutex);
5558 mutex_lock(&ctx->mutex);
5559 smp_call_function_single(cpu, __perf_event_exit_cpu, NULL, 1);
5560 mutex_unlock(&ctx->mutex);
5563 static inline void perf_event_exit_cpu(int cpu) { }
5566 static int __cpuinit
5567 perf_cpu_notify(struct notifier_block *self, unsigned long action, void *hcpu)
5569 unsigned int cpu = (long)hcpu;
5573 case CPU_UP_PREPARE:
5574 case CPU_UP_PREPARE_FROZEN:
5575 perf_event_init_cpu(cpu);
5578 case CPU_DOWN_PREPARE:
5579 case CPU_DOWN_PREPARE_FROZEN:
5580 perf_event_exit_cpu(cpu);
5591 * This has to have a higher priority than migration_notifier in sched.c.
5593 static struct notifier_block __cpuinitdata perf_cpu_nb = {
5594 .notifier_call = perf_cpu_notify,
5598 void __init perf_event_init(void)
5600 perf_event_init_all_cpus();
5601 perf_cpu_notify(&perf_cpu_nb, (unsigned long)CPU_UP_PREPARE,
5602 (void *)(long)smp_processor_id());
5603 perf_cpu_notify(&perf_cpu_nb, (unsigned long)CPU_ONLINE,
5604 (void *)(long)smp_processor_id());
5605 register_cpu_notifier(&perf_cpu_nb);
5608 static ssize_t perf_show_reserve_percpu(struct sysdev_class *class,
5609 struct sysdev_class_attribute *attr,
5612 return sprintf(buf, "%d\n", perf_reserved_percpu);
5616 perf_set_reserve_percpu(struct sysdev_class *class,
5617 struct sysdev_class_attribute *attr,
5621 struct perf_cpu_context *cpuctx;
5625 err = strict_strtoul(buf, 10, &val);
5628 if (val > perf_max_events)
5631 spin_lock(&perf_resource_lock);
5632 perf_reserved_percpu = val;
5633 for_each_online_cpu(cpu) {
5634 cpuctx = &per_cpu(perf_cpu_context, cpu);
5635 raw_spin_lock_irq(&cpuctx->ctx.lock);
5636 mpt = min(perf_max_events - cpuctx->ctx.nr_events,
5637 perf_max_events - perf_reserved_percpu);
5638 cpuctx->max_pertask = mpt;
5639 raw_spin_unlock_irq(&cpuctx->ctx.lock);
5641 spin_unlock(&perf_resource_lock);
5646 static ssize_t perf_show_overcommit(struct sysdev_class *class,
5647 struct sysdev_class_attribute *attr,
5650 return sprintf(buf, "%d\n", perf_overcommit);
5654 perf_set_overcommit(struct sysdev_class *class,
5655 struct sysdev_class_attribute *attr,
5656 const char *buf, size_t count)
5661 err = strict_strtoul(buf, 10, &val);
5667 spin_lock(&perf_resource_lock);
5668 perf_overcommit = val;
5669 spin_unlock(&perf_resource_lock);
5674 static SYSDEV_CLASS_ATTR(
5677 perf_show_reserve_percpu,
5678 perf_set_reserve_percpu
5681 static SYSDEV_CLASS_ATTR(
5684 perf_show_overcommit,
5688 static struct attribute *perfclass_attrs[] = {
5689 &attr_reserve_percpu.attr,
5690 &attr_overcommit.attr,
5694 static struct attribute_group perfclass_attr_group = {
5695 .attrs = perfclass_attrs,
5696 .name = "perf_events",
5699 static int __init perf_event_sysfs_init(void)
5701 return sysfs_create_group(&cpu_sysdev_class.kset.kobj,
5702 &perfclass_attr_group);
5704 device_initcall(perf_event_sysfs_init);