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(); }
86 void __weak perf_event_print_debug(void) { }
88 static DEFINE_PER_CPU(int, perf_disable_count);
90 void perf_disable(void)
92 if (!__get_cpu_var(perf_disable_count)++)
96 void perf_enable(void)
98 if (!--__get_cpu_var(perf_disable_count))
102 static void get_ctx(struct perf_event_context *ctx)
104 WARN_ON(!atomic_inc_not_zero(&ctx->refcount));
107 static void free_ctx(struct rcu_head *head)
109 struct perf_event_context *ctx;
111 ctx = container_of(head, struct perf_event_context, rcu_head);
115 static void put_ctx(struct perf_event_context *ctx)
117 if (atomic_dec_and_test(&ctx->refcount)) {
119 put_ctx(ctx->parent_ctx);
121 put_task_struct(ctx->task);
122 call_rcu(&ctx->rcu_head, free_ctx);
126 static void unclone_ctx(struct perf_event_context *ctx)
128 if (ctx->parent_ctx) {
129 put_ctx(ctx->parent_ctx);
130 ctx->parent_ctx = NULL;
135 * If we inherit events we want to return the parent event id
138 static u64 primary_event_id(struct perf_event *event)
143 id = event->parent->id;
149 * Get the perf_event_context for a task and lock it.
150 * This has to cope with with the fact that until it is locked,
151 * the context could get moved to another task.
153 static struct perf_event_context *
154 perf_lock_task_context(struct task_struct *task, unsigned long *flags)
156 struct perf_event_context *ctx;
160 ctx = rcu_dereference(task->perf_event_ctxp);
163 * If this context is a clone of another, it might
164 * get swapped for another underneath us by
165 * perf_event_task_sched_out, though the
166 * rcu_read_lock() protects us from any context
167 * getting freed. Lock the context and check if it
168 * got swapped before we could get the lock, and retry
169 * if so. If we locked the right context, then it
170 * can't get swapped on us any more.
172 raw_spin_lock_irqsave(&ctx->lock, *flags);
173 if (ctx != rcu_dereference(task->perf_event_ctxp)) {
174 raw_spin_unlock_irqrestore(&ctx->lock, *flags);
178 if (!atomic_inc_not_zero(&ctx->refcount)) {
179 raw_spin_unlock_irqrestore(&ctx->lock, *flags);
188 * Get the context for a task and increment its pin_count so it
189 * can't get swapped to another task. This also increments its
190 * reference count so that the context can't get freed.
192 static struct perf_event_context *perf_pin_task_context(struct task_struct *task)
194 struct perf_event_context *ctx;
197 ctx = perf_lock_task_context(task, &flags);
200 raw_spin_unlock_irqrestore(&ctx->lock, flags);
205 static void perf_unpin_context(struct perf_event_context *ctx)
209 raw_spin_lock_irqsave(&ctx->lock, flags);
211 raw_spin_unlock_irqrestore(&ctx->lock, flags);
215 static inline u64 perf_clock(void)
217 return cpu_clock(raw_smp_processor_id());
221 * Update the record of the current time in a context.
223 static void update_context_time(struct perf_event_context *ctx)
225 u64 now = perf_clock();
227 ctx->time += now - ctx->timestamp;
228 ctx->timestamp = now;
232 * Update the total_time_enabled and total_time_running fields for a event.
234 static void update_event_times(struct perf_event *event)
236 struct perf_event_context *ctx = event->ctx;
239 if (event->state < PERF_EVENT_STATE_INACTIVE ||
240 event->group_leader->state < PERF_EVENT_STATE_INACTIVE)
246 run_end = event->tstamp_stopped;
248 event->total_time_enabled = run_end - event->tstamp_enabled;
250 if (event->state == PERF_EVENT_STATE_INACTIVE)
251 run_end = event->tstamp_stopped;
255 event->total_time_running = run_end - event->tstamp_running;
259 * Update total_time_enabled and total_time_running for all events in a group.
261 static void update_group_times(struct perf_event *leader)
263 struct perf_event *event;
265 update_event_times(leader);
266 list_for_each_entry(event, &leader->sibling_list, group_entry)
267 update_event_times(event);
270 static struct list_head *
271 ctx_group_list(struct perf_event *event, struct perf_event_context *ctx)
273 if (event->attr.pinned)
274 return &ctx->pinned_groups;
276 return &ctx->flexible_groups;
280 * Add a event from the lists for its context.
281 * Must be called with ctx->mutex and ctx->lock held.
284 list_add_event(struct perf_event *event, struct perf_event_context *ctx)
286 struct perf_event *group_leader = event->group_leader;
289 * Depending on whether it is a standalone or sibling event,
290 * add it straight to the context's event list, or to the group
291 * leader's sibling list:
293 if (group_leader == event) {
294 struct list_head *list;
296 if (is_software_event(event))
297 event->group_flags |= PERF_GROUP_SOFTWARE;
299 list = ctx_group_list(event, ctx);
300 list_add_tail(&event->group_entry, list);
302 if (group_leader->group_flags & PERF_GROUP_SOFTWARE &&
303 !is_software_event(event))
304 group_leader->group_flags &= ~PERF_GROUP_SOFTWARE;
306 list_add_tail(&event->group_entry, &group_leader->sibling_list);
307 group_leader->nr_siblings++;
310 list_add_rcu(&event->event_entry, &ctx->event_list);
312 if (event->attr.inherit_stat)
317 * Remove a event from the lists for its context.
318 * Must be called with ctx->mutex and ctx->lock held.
321 list_del_event(struct perf_event *event, struct perf_event_context *ctx)
323 if (list_empty(&event->group_entry))
326 if (event->attr.inherit_stat)
329 list_del_init(&event->group_entry);
330 list_del_rcu(&event->event_entry);
332 if (event->group_leader != event)
333 event->group_leader->nr_siblings--;
335 update_group_times(event);
338 * If event was in error state, then keep it
339 * that way, otherwise bogus counts will be
340 * returned on read(). The only way to get out
341 * of error state is by explicit re-enabling
344 if (event->state > PERF_EVENT_STATE_OFF)
345 event->state = PERF_EVENT_STATE_OFF;
349 perf_destroy_group(struct perf_event *event, struct perf_event_context *ctx)
351 struct perf_event *sibling, *tmp;
354 * If this was a group event with sibling events then
355 * upgrade the siblings to singleton events by adding them
356 * to the context list directly:
358 list_for_each_entry_safe(sibling, tmp, &event->sibling_list, group_entry) {
359 struct list_head *list;
361 list = ctx_group_list(event, ctx);
362 list_move_tail(&sibling->group_entry, list);
363 sibling->group_leader = sibling;
365 /* Inherit group flags from the previous leader */
366 sibling->group_flags = event->group_flags;
371 event_sched_out(struct perf_event *event,
372 struct perf_cpu_context *cpuctx,
373 struct perf_event_context *ctx)
375 if (event->state != PERF_EVENT_STATE_ACTIVE)
378 event->state = PERF_EVENT_STATE_INACTIVE;
379 if (event->pending_disable) {
380 event->pending_disable = 0;
381 event->state = PERF_EVENT_STATE_OFF;
383 event->tstamp_stopped = ctx->time;
384 event->pmu->disable(event);
387 if (!is_software_event(event))
388 cpuctx->active_oncpu--;
390 if (event->attr.exclusive || !cpuctx->active_oncpu)
391 cpuctx->exclusive = 0;
395 group_sched_out(struct perf_event *group_event,
396 struct perf_cpu_context *cpuctx,
397 struct perf_event_context *ctx)
399 struct perf_event *event;
401 if (group_event->state != PERF_EVENT_STATE_ACTIVE)
404 event_sched_out(group_event, cpuctx, ctx);
407 * Schedule out siblings (if any):
409 list_for_each_entry(event, &group_event->sibling_list, group_entry)
410 event_sched_out(event, cpuctx, ctx);
412 if (group_event->attr.exclusive)
413 cpuctx->exclusive = 0;
417 * Cross CPU call to remove a performance event
419 * We disable the event on the hardware level first. After that we
420 * remove it from the context list.
422 static void __perf_event_remove_from_context(void *info)
424 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
425 struct perf_event *event = info;
426 struct perf_event_context *ctx = event->ctx;
429 * If this is a task context, we need to check whether it is
430 * the current task context of this cpu. If not it has been
431 * scheduled out before the smp call arrived.
433 if (ctx->task && cpuctx->task_ctx != ctx)
436 raw_spin_lock(&ctx->lock);
438 * Protect the list operation against NMI by disabling the
439 * events on a global level.
443 event_sched_out(event, cpuctx, ctx);
445 list_del_event(event, ctx);
449 * Allow more per task events with respect to the
452 cpuctx->max_pertask =
453 min(perf_max_events - ctx->nr_events,
454 perf_max_events - perf_reserved_percpu);
458 raw_spin_unlock(&ctx->lock);
463 * Remove the event from a task's (or a CPU's) list of events.
465 * Must be called with ctx->mutex held.
467 * CPU events are removed with a smp call. For task events we only
468 * call when the task is on a CPU.
470 * If event->ctx is a cloned context, callers must make sure that
471 * every task struct that event->ctx->task could possibly point to
472 * remains valid. This is OK when called from perf_release since
473 * that only calls us on the top-level context, which can't be a clone.
474 * When called from perf_event_exit_task, it's OK because the
475 * context has been detached from its task.
477 static void perf_event_remove_from_context(struct perf_event *event)
479 struct perf_event_context *ctx = event->ctx;
480 struct task_struct *task = ctx->task;
484 * Per cpu events are removed via an smp call and
485 * the removal is always successful.
487 smp_call_function_single(event->cpu,
488 __perf_event_remove_from_context,
494 task_oncpu_function_call(task, __perf_event_remove_from_context,
497 raw_spin_lock_irq(&ctx->lock);
499 * If the context is active we need to retry the smp call.
501 if (ctx->nr_active && !list_empty(&event->group_entry)) {
502 raw_spin_unlock_irq(&ctx->lock);
507 * The lock prevents that this context is scheduled in so we
508 * can remove the event safely, if the call above did not
511 if (!list_empty(&event->group_entry))
512 list_del_event(event, ctx);
513 raw_spin_unlock_irq(&ctx->lock);
517 * Cross CPU call to disable a performance event
519 static void __perf_event_disable(void *info)
521 struct perf_event *event = info;
522 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
523 struct perf_event_context *ctx = event->ctx;
526 * If this is a per-task event, need to check whether this
527 * event's task is the current task on this cpu.
529 if (ctx->task && cpuctx->task_ctx != ctx)
532 raw_spin_lock(&ctx->lock);
535 * If the event is on, turn it off.
536 * If it is in error state, leave it in error state.
538 if (event->state >= PERF_EVENT_STATE_INACTIVE) {
539 update_context_time(ctx);
540 update_group_times(event);
541 if (event == event->group_leader)
542 group_sched_out(event, cpuctx, ctx);
544 event_sched_out(event, cpuctx, ctx);
545 event->state = PERF_EVENT_STATE_OFF;
548 raw_spin_unlock(&ctx->lock);
554 * If event->ctx is a cloned context, callers must make sure that
555 * every task struct that event->ctx->task could possibly point to
556 * remains valid. This condition is satisifed when called through
557 * perf_event_for_each_child or perf_event_for_each because they
558 * hold the top-level event's child_mutex, so any descendant that
559 * goes to exit will block in sync_child_event.
560 * When called from perf_pending_event it's OK because event->ctx
561 * is the current context on this CPU and preemption is disabled,
562 * hence we can't get into perf_event_task_sched_out for this context.
564 void perf_event_disable(struct perf_event *event)
566 struct perf_event_context *ctx = event->ctx;
567 struct task_struct *task = ctx->task;
571 * Disable the event on the cpu that it's on
573 smp_call_function_single(event->cpu, __perf_event_disable,
579 task_oncpu_function_call(task, __perf_event_disable, event);
581 raw_spin_lock_irq(&ctx->lock);
583 * If the event is still active, we need to retry the cross-call.
585 if (event->state == PERF_EVENT_STATE_ACTIVE) {
586 raw_spin_unlock_irq(&ctx->lock);
591 * Since we have the lock this context can't be scheduled
592 * in, so we can change the state safely.
594 if (event->state == PERF_EVENT_STATE_INACTIVE) {
595 update_group_times(event);
596 event->state = PERF_EVENT_STATE_OFF;
599 raw_spin_unlock_irq(&ctx->lock);
603 event_sched_in(struct perf_event *event,
604 struct perf_cpu_context *cpuctx,
605 struct perf_event_context *ctx)
607 if (event->state <= PERF_EVENT_STATE_OFF)
610 event->state = PERF_EVENT_STATE_ACTIVE;
611 event->oncpu = smp_processor_id();
613 * The new state must be visible before we turn it on in the hardware:
617 if (event->pmu->enable(event)) {
618 event->state = PERF_EVENT_STATE_INACTIVE;
623 event->tstamp_running += ctx->time - event->tstamp_stopped;
625 if (!is_software_event(event))
626 cpuctx->active_oncpu++;
629 if (event->attr.exclusive)
630 cpuctx->exclusive = 1;
636 group_sched_in(struct perf_event *group_event,
637 struct perf_cpu_context *cpuctx,
638 struct perf_event_context *ctx)
640 struct perf_event *event, *partial_group = NULL;
641 const struct pmu *pmu = group_event->pmu;
645 if (group_event->state == PERF_EVENT_STATE_OFF)
648 /* Check if group transaction availabe */
655 if (event_sched_in(group_event, cpuctx, ctx))
659 * Schedule in siblings as one group (if any):
661 list_for_each_entry(event, &group_event->sibling_list, group_entry) {
662 if (event_sched_in(event, cpuctx, ctx)) {
663 partial_group = event;
671 ret = pmu->commit_txn(pmu);
673 pmu->cancel_txn(pmu);
679 pmu->cancel_txn(pmu);
682 * Groups can be scheduled in as one unit only, so undo any
683 * partial group before returning:
685 list_for_each_entry(event, &group_event->sibling_list, group_entry) {
686 if (event == partial_group)
688 event_sched_out(event, cpuctx, ctx);
690 event_sched_out(group_event, cpuctx, ctx);
696 * Work out whether we can put this event group on the CPU now.
698 static int group_can_go_on(struct perf_event *event,
699 struct perf_cpu_context *cpuctx,
703 * Groups consisting entirely of software events can always go on.
705 if (event->group_flags & PERF_GROUP_SOFTWARE)
708 * If an exclusive group is already on, no other hardware
711 if (cpuctx->exclusive)
714 * If this group is exclusive and there are already
715 * events on the CPU, it can't go on.
717 if (event->attr.exclusive && cpuctx->active_oncpu)
720 * Otherwise, try to add it if all previous groups were able
726 static void add_event_to_ctx(struct perf_event *event,
727 struct perf_event_context *ctx)
729 list_add_event(event, ctx);
730 event->tstamp_enabled = ctx->time;
731 event->tstamp_running = ctx->time;
732 event->tstamp_stopped = ctx->time;
736 * Cross CPU call to install and enable a performance event
738 * Must be called with ctx->mutex held
740 static void __perf_install_in_context(void *info)
742 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
743 struct perf_event *event = info;
744 struct perf_event_context *ctx = event->ctx;
745 struct perf_event *leader = event->group_leader;
749 * If this is a task context, we need to check whether it is
750 * the current task context of this cpu. If not it has been
751 * scheduled out before the smp call arrived.
752 * Or possibly this is the right context but it isn't
753 * on this cpu because it had no events.
755 if (ctx->task && cpuctx->task_ctx != ctx) {
756 if (cpuctx->task_ctx || ctx->task != current)
758 cpuctx->task_ctx = ctx;
761 raw_spin_lock(&ctx->lock);
763 update_context_time(ctx);
766 * Protect the list operation against NMI by disabling the
767 * events on a global level. NOP for non NMI based events.
771 add_event_to_ctx(event, ctx);
773 if (event->cpu != -1 && event->cpu != smp_processor_id())
777 * Don't put the event on if it is disabled or if
778 * it is in a group and the group isn't on.
780 if (event->state != PERF_EVENT_STATE_INACTIVE ||
781 (leader != event && leader->state != PERF_EVENT_STATE_ACTIVE))
785 * An exclusive event can't go on if there are already active
786 * hardware events, and no hardware event can go on if there
787 * is already an exclusive event on.
789 if (!group_can_go_on(event, cpuctx, 1))
792 err = event_sched_in(event, cpuctx, ctx);
796 * This event couldn't go on. If it is in a group
797 * then we have to pull the whole group off.
798 * If the event group is pinned then put it in error state.
801 group_sched_out(leader, cpuctx, ctx);
802 if (leader->attr.pinned) {
803 update_group_times(leader);
804 leader->state = PERF_EVENT_STATE_ERROR;
808 if (!err && !ctx->task && cpuctx->max_pertask)
809 cpuctx->max_pertask--;
814 raw_spin_unlock(&ctx->lock);
818 * Attach a performance event to a context
820 * First we add the event to the list with the hardware enable bit
821 * in event->hw_config cleared.
823 * If the event is attached to a task which is on a CPU we use a smp
824 * call to enable it in the task context. The task might have been
825 * scheduled away, but we check this in the smp call again.
827 * Must be called with ctx->mutex held.
830 perf_install_in_context(struct perf_event_context *ctx,
831 struct perf_event *event,
834 struct task_struct *task = ctx->task;
838 * Per cpu events are installed via an smp call and
839 * the install is always successful.
841 smp_call_function_single(cpu, __perf_install_in_context,
847 task_oncpu_function_call(task, __perf_install_in_context,
850 raw_spin_lock_irq(&ctx->lock);
852 * we need to retry the smp call.
854 if (ctx->is_active && list_empty(&event->group_entry)) {
855 raw_spin_unlock_irq(&ctx->lock);
860 * The lock prevents that this context is scheduled in so we
861 * can add the event safely, if it the call above did not
864 if (list_empty(&event->group_entry))
865 add_event_to_ctx(event, ctx);
866 raw_spin_unlock_irq(&ctx->lock);
870 * Put a event into inactive state and update time fields.
871 * Enabling the leader of a group effectively enables all
872 * the group members that aren't explicitly disabled, so we
873 * have to update their ->tstamp_enabled also.
874 * Note: this works for group members as well as group leaders
875 * since the non-leader members' sibling_lists will be empty.
877 static void __perf_event_mark_enabled(struct perf_event *event,
878 struct perf_event_context *ctx)
880 struct perf_event *sub;
882 event->state = PERF_EVENT_STATE_INACTIVE;
883 event->tstamp_enabled = ctx->time - event->total_time_enabled;
884 list_for_each_entry(sub, &event->sibling_list, group_entry)
885 if (sub->state >= PERF_EVENT_STATE_INACTIVE)
886 sub->tstamp_enabled =
887 ctx->time - sub->total_time_enabled;
891 * Cross CPU call to enable a performance event
893 static void __perf_event_enable(void *info)
895 struct perf_event *event = info;
896 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
897 struct perf_event_context *ctx = event->ctx;
898 struct perf_event *leader = event->group_leader;
902 * If this is a per-task event, need to check whether this
903 * event's task is the current task on this cpu.
905 if (ctx->task && cpuctx->task_ctx != ctx) {
906 if (cpuctx->task_ctx || ctx->task != current)
908 cpuctx->task_ctx = ctx;
911 raw_spin_lock(&ctx->lock);
913 update_context_time(ctx);
915 if (event->state >= PERF_EVENT_STATE_INACTIVE)
917 __perf_event_mark_enabled(event, ctx);
919 if (event->cpu != -1 && event->cpu != smp_processor_id())
923 * If the event is in a group and isn't the group leader,
924 * then don't put it on unless the group is on.
926 if (leader != event && leader->state != PERF_EVENT_STATE_ACTIVE)
929 if (!group_can_go_on(event, cpuctx, 1)) {
934 err = group_sched_in(event, cpuctx, ctx);
936 err = event_sched_in(event, cpuctx, ctx);
942 * If this event can't go on and it's part of a
943 * group, then the whole group has to come off.
946 group_sched_out(leader, cpuctx, ctx);
947 if (leader->attr.pinned) {
948 update_group_times(leader);
949 leader->state = PERF_EVENT_STATE_ERROR;
954 raw_spin_unlock(&ctx->lock);
960 * If event->ctx is a cloned context, callers must make sure that
961 * every task struct that event->ctx->task could possibly point to
962 * remains valid. This condition is satisfied when called through
963 * perf_event_for_each_child or perf_event_for_each as described
964 * for perf_event_disable.
966 void perf_event_enable(struct perf_event *event)
968 struct perf_event_context *ctx = event->ctx;
969 struct task_struct *task = ctx->task;
973 * Enable the event on the cpu that it's on
975 smp_call_function_single(event->cpu, __perf_event_enable,
980 raw_spin_lock_irq(&ctx->lock);
981 if (event->state >= PERF_EVENT_STATE_INACTIVE)
985 * If the event is in error state, clear that first.
986 * That way, if we see the event in error state below, we
987 * know that it has gone back into error state, as distinct
988 * from the task having been scheduled away before the
989 * cross-call arrived.
991 if (event->state == PERF_EVENT_STATE_ERROR)
992 event->state = PERF_EVENT_STATE_OFF;
995 raw_spin_unlock_irq(&ctx->lock);
996 task_oncpu_function_call(task, __perf_event_enable, event);
998 raw_spin_lock_irq(&ctx->lock);
1001 * If the context is active and the event is still off,
1002 * we need to retry the cross-call.
1004 if (ctx->is_active && event->state == PERF_EVENT_STATE_OFF)
1008 * Since we have the lock this context can't be scheduled
1009 * in, so we can change the state safely.
1011 if (event->state == PERF_EVENT_STATE_OFF)
1012 __perf_event_mark_enabled(event, ctx);
1015 raw_spin_unlock_irq(&ctx->lock);
1018 static int perf_event_refresh(struct perf_event *event, int refresh)
1021 * not supported on inherited events
1023 if (event->attr.inherit)
1026 atomic_add(refresh, &event->event_limit);
1027 perf_event_enable(event);
1033 EVENT_FLEXIBLE = 0x1,
1035 EVENT_ALL = EVENT_FLEXIBLE | EVENT_PINNED,
1038 static void ctx_sched_out(struct perf_event_context *ctx,
1039 struct perf_cpu_context *cpuctx,
1040 enum event_type_t event_type)
1042 struct perf_event *event;
1044 raw_spin_lock(&ctx->lock);
1046 if (likely(!ctx->nr_events))
1048 update_context_time(ctx);
1051 if (!ctx->nr_active)
1054 if (event_type & EVENT_PINNED)
1055 list_for_each_entry(event, &ctx->pinned_groups, group_entry)
1056 group_sched_out(event, cpuctx, ctx);
1058 if (event_type & EVENT_FLEXIBLE)
1059 list_for_each_entry(event, &ctx->flexible_groups, group_entry)
1060 group_sched_out(event, cpuctx, ctx);
1065 raw_spin_unlock(&ctx->lock);
1069 * Test whether two contexts are equivalent, i.e. whether they
1070 * have both been cloned from the same version of the same context
1071 * and they both have the same number of enabled events.
1072 * If the number of enabled events is the same, then the set
1073 * of enabled events should be the same, because these are both
1074 * inherited contexts, therefore we can't access individual events
1075 * in them directly with an fd; we can only enable/disable all
1076 * events via prctl, or enable/disable all events in a family
1077 * via ioctl, which will have the same effect on both contexts.
1079 static int context_equiv(struct perf_event_context *ctx1,
1080 struct perf_event_context *ctx2)
1082 return ctx1->parent_ctx && ctx1->parent_ctx == ctx2->parent_ctx
1083 && ctx1->parent_gen == ctx2->parent_gen
1084 && !ctx1->pin_count && !ctx2->pin_count;
1087 static void __perf_event_sync_stat(struct perf_event *event,
1088 struct perf_event *next_event)
1092 if (!event->attr.inherit_stat)
1096 * Update the event value, we cannot use perf_event_read()
1097 * because we're in the middle of a context switch and have IRQs
1098 * disabled, which upsets smp_call_function_single(), however
1099 * we know the event must be on the current CPU, therefore we
1100 * don't need to use it.
1102 switch (event->state) {
1103 case PERF_EVENT_STATE_ACTIVE:
1104 event->pmu->read(event);
1107 case PERF_EVENT_STATE_INACTIVE:
1108 update_event_times(event);
1116 * In order to keep per-task stats reliable we need to flip the event
1117 * values when we flip the contexts.
1119 value = atomic64_read(&next_event->count);
1120 value = atomic64_xchg(&event->count, value);
1121 atomic64_set(&next_event->count, value);
1123 swap(event->total_time_enabled, next_event->total_time_enabled);
1124 swap(event->total_time_running, next_event->total_time_running);
1127 * Since we swizzled the values, update the user visible data too.
1129 perf_event_update_userpage(event);
1130 perf_event_update_userpage(next_event);
1133 #define list_next_entry(pos, member) \
1134 list_entry(pos->member.next, typeof(*pos), member)
1136 static void perf_event_sync_stat(struct perf_event_context *ctx,
1137 struct perf_event_context *next_ctx)
1139 struct perf_event *event, *next_event;
1144 update_context_time(ctx);
1146 event = list_first_entry(&ctx->event_list,
1147 struct perf_event, event_entry);
1149 next_event = list_first_entry(&next_ctx->event_list,
1150 struct perf_event, event_entry);
1152 while (&event->event_entry != &ctx->event_list &&
1153 &next_event->event_entry != &next_ctx->event_list) {
1155 __perf_event_sync_stat(event, next_event);
1157 event = list_next_entry(event, event_entry);
1158 next_event = list_next_entry(next_event, event_entry);
1163 * Called from scheduler to remove the events of the current task,
1164 * with interrupts disabled.
1166 * We stop each event and update the event value in event->count.
1168 * This does not protect us against NMI, but disable()
1169 * sets the disabled bit in the control field of event _before_
1170 * accessing the event control register. If a NMI hits, then it will
1171 * not restart the event.
1173 void perf_event_task_sched_out(struct task_struct *task,
1174 struct task_struct *next)
1176 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
1177 struct perf_event_context *ctx = task->perf_event_ctxp;
1178 struct perf_event_context *next_ctx;
1179 struct perf_event_context *parent;
1182 perf_sw_event(PERF_COUNT_SW_CONTEXT_SWITCHES, 1, 1, NULL, 0);
1184 if (likely(!ctx || !cpuctx->task_ctx))
1188 parent = rcu_dereference(ctx->parent_ctx);
1189 next_ctx = next->perf_event_ctxp;
1190 if (parent && next_ctx &&
1191 rcu_dereference(next_ctx->parent_ctx) == parent) {
1193 * Looks like the two contexts are clones, so we might be
1194 * able to optimize the context switch. We lock both
1195 * contexts and check that they are clones under the
1196 * lock (including re-checking that neither has been
1197 * uncloned in the meantime). It doesn't matter which
1198 * order we take the locks because no other cpu could
1199 * be trying to lock both of these tasks.
1201 raw_spin_lock(&ctx->lock);
1202 raw_spin_lock_nested(&next_ctx->lock, SINGLE_DEPTH_NESTING);
1203 if (context_equiv(ctx, next_ctx)) {
1205 * XXX do we need a memory barrier of sorts
1206 * wrt to rcu_dereference() of perf_event_ctxp
1208 task->perf_event_ctxp = next_ctx;
1209 next->perf_event_ctxp = ctx;
1211 next_ctx->task = task;
1214 perf_event_sync_stat(ctx, next_ctx);
1216 raw_spin_unlock(&next_ctx->lock);
1217 raw_spin_unlock(&ctx->lock);
1222 ctx_sched_out(ctx, cpuctx, EVENT_ALL);
1223 cpuctx->task_ctx = NULL;
1227 static void task_ctx_sched_out(struct perf_event_context *ctx,
1228 enum event_type_t event_type)
1230 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
1232 if (!cpuctx->task_ctx)
1235 if (WARN_ON_ONCE(ctx != cpuctx->task_ctx))
1238 ctx_sched_out(ctx, cpuctx, event_type);
1239 cpuctx->task_ctx = NULL;
1243 * Called with IRQs disabled
1245 static void __perf_event_task_sched_out(struct perf_event_context *ctx)
1247 task_ctx_sched_out(ctx, EVENT_ALL);
1251 * Called with IRQs disabled
1253 static void cpu_ctx_sched_out(struct perf_cpu_context *cpuctx,
1254 enum event_type_t event_type)
1256 ctx_sched_out(&cpuctx->ctx, cpuctx, event_type);
1260 ctx_pinned_sched_in(struct perf_event_context *ctx,
1261 struct perf_cpu_context *cpuctx)
1263 struct perf_event *event;
1265 list_for_each_entry(event, &ctx->pinned_groups, group_entry) {
1266 if (event->state <= PERF_EVENT_STATE_OFF)
1268 if (event->cpu != -1 && event->cpu != smp_processor_id())
1271 if (group_can_go_on(event, cpuctx, 1))
1272 group_sched_in(event, cpuctx, ctx);
1275 * If this pinned group hasn't been scheduled,
1276 * put it in error state.
1278 if (event->state == PERF_EVENT_STATE_INACTIVE) {
1279 update_group_times(event);
1280 event->state = PERF_EVENT_STATE_ERROR;
1286 ctx_flexible_sched_in(struct perf_event_context *ctx,
1287 struct perf_cpu_context *cpuctx)
1289 struct perf_event *event;
1292 list_for_each_entry(event, &ctx->flexible_groups, group_entry) {
1293 /* Ignore events in OFF or ERROR state */
1294 if (event->state <= PERF_EVENT_STATE_OFF)
1297 * Listen to the 'cpu' scheduling filter constraint
1300 if (event->cpu != -1 && event->cpu != smp_processor_id())
1303 if (group_can_go_on(event, cpuctx, can_add_hw))
1304 if (group_sched_in(event, cpuctx, ctx))
1310 ctx_sched_in(struct perf_event_context *ctx,
1311 struct perf_cpu_context *cpuctx,
1312 enum event_type_t event_type)
1314 raw_spin_lock(&ctx->lock);
1316 if (likely(!ctx->nr_events))
1319 ctx->timestamp = perf_clock();
1324 * First go through the list and put on any pinned groups
1325 * in order to give them the best chance of going on.
1327 if (event_type & EVENT_PINNED)
1328 ctx_pinned_sched_in(ctx, cpuctx);
1330 /* Then walk through the lower prio flexible groups */
1331 if (event_type & EVENT_FLEXIBLE)
1332 ctx_flexible_sched_in(ctx, cpuctx);
1336 raw_spin_unlock(&ctx->lock);
1339 static void cpu_ctx_sched_in(struct perf_cpu_context *cpuctx,
1340 enum event_type_t event_type)
1342 struct perf_event_context *ctx = &cpuctx->ctx;
1344 ctx_sched_in(ctx, cpuctx, event_type);
1347 static void task_ctx_sched_in(struct task_struct *task,
1348 enum event_type_t event_type)
1350 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
1351 struct perf_event_context *ctx = task->perf_event_ctxp;
1355 if (cpuctx->task_ctx == ctx)
1357 ctx_sched_in(ctx, cpuctx, event_type);
1358 cpuctx->task_ctx = ctx;
1361 * Called from scheduler to add the events of the current task
1362 * with interrupts disabled.
1364 * We restore the event value and then enable it.
1366 * This does not protect us against NMI, but enable()
1367 * sets the enabled bit in the control field of event _before_
1368 * accessing the event control register. If a NMI hits, then it will
1369 * keep the event running.
1371 void perf_event_task_sched_in(struct task_struct *task)
1373 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
1374 struct perf_event_context *ctx = task->perf_event_ctxp;
1379 if (cpuctx->task_ctx == ctx)
1385 * We want to keep the following priority order:
1386 * cpu pinned (that don't need to move), task pinned,
1387 * cpu flexible, task flexible.
1389 cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE);
1391 ctx_sched_in(ctx, cpuctx, EVENT_PINNED);
1392 cpu_ctx_sched_in(cpuctx, EVENT_FLEXIBLE);
1393 ctx_sched_in(ctx, cpuctx, EVENT_FLEXIBLE);
1395 cpuctx->task_ctx = ctx;
1400 #define MAX_INTERRUPTS (~0ULL)
1402 static void perf_log_throttle(struct perf_event *event, int enable);
1404 static u64 perf_calculate_period(struct perf_event *event, u64 nsec, u64 count)
1406 u64 frequency = event->attr.sample_freq;
1407 u64 sec = NSEC_PER_SEC;
1408 u64 divisor, dividend;
1410 int count_fls, nsec_fls, frequency_fls, sec_fls;
1412 count_fls = fls64(count);
1413 nsec_fls = fls64(nsec);
1414 frequency_fls = fls64(frequency);
1418 * We got @count in @nsec, with a target of sample_freq HZ
1419 * the target period becomes:
1422 * period = -------------------
1423 * @nsec * sample_freq
1428 * Reduce accuracy by one bit such that @a and @b converge
1429 * to a similar magnitude.
1431 #define REDUCE_FLS(a, b) \
1433 if (a##_fls > b##_fls) { \
1443 * Reduce accuracy until either term fits in a u64, then proceed with
1444 * the other, so that finally we can do a u64/u64 division.
1446 while (count_fls + sec_fls > 64 && nsec_fls + frequency_fls > 64) {
1447 REDUCE_FLS(nsec, frequency);
1448 REDUCE_FLS(sec, count);
1451 if (count_fls + sec_fls > 64) {
1452 divisor = nsec * frequency;
1454 while (count_fls + sec_fls > 64) {
1455 REDUCE_FLS(count, sec);
1459 dividend = count * sec;
1461 dividend = count * sec;
1463 while (nsec_fls + frequency_fls > 64) {
1464 REDUCE_FLS(nsec, frequency);
1468 divisor = nsec * frequency;
1471 return div64_u64(dividend, divisor);
1474 static void perf_event_stop(struct perf_event *event)
1476 if (!event->pmu->stop)
1477 return event->pmu->disable(event);
1479 return event->pmu->stop(event);
1482 static int perf_event_start(struct perf_event *event)
1484 if (!event->pmu->start)
1485 return event->pmu->enable(event);
1487 return event->pmu->start(event);
1490 static void perf_adjust_period(struct perf_event *event, u64 nsec, u64 count)
1492 struct hw_perf_event *hwc = &event->hw;
1493 u64 period, sample_period;
1496 period = perf_calculate_period(event, nsec, count);
1498 delta = (s64)(period - hwc->sample_period);
1499 delta = (delta + 7) / 8; /* low pass filter */
1501 sample_period = hwc->sample_period + delta;
1506 hwc->sample_period = sample_period;
1508 if (atomic64_read(&hwc->period_left) > 8*sample_period) {
1510 perf_event_stop(event);
1511 atomic64_set(&hwc->period_left, 0);
1512 perf_event_start(event);
1517 static void perf_ctx_adjust_freq(struct perf_event_context *ctx)
1519 struct perf_event *event;
1520 struct hw_perf_event *hwc;
1521 u64 interrupts, now;
1524 raw_spin_lock(&ctx->lock);
1525 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
1526 if (event->state != PERF_EVENT_STATE_ACTIVE)
1529 if (event->cpu != -1 && event->cpu != smp_processor_id())
1534 interrupts = hwc->interrupts;
1535 hwc->interrupts = 0;
1538 * unthrottle events on the tick
1540 if (interrupts == MAX_INTERRUPTS) {
1541 perf_log_throttle(event, 1);
1543 event->pmu->unthrottle(event);
1547 if (!event->attr.freq || !event->attr.sample_freq)
1551 event->pmu->read(event);
1552 now = atomic64_read(&event->count);
1553 delta = now - hwc->freq_count_stamp;
1554 hwc->freq_count_stamp = now;
1557 perf_adjust_period(event, TICK_NSEC, delta);
1560 raw_spin_unlock(&ctx->lock);
1564 * Round-robin a context's events:
1566 static void rotate_ctx(struct perf_event_context *ctx)
1568 raw_spin_lock(&ctx->lock);
1570 /* Rotate the first entry last of non-pinned groups */
1571 list_rotate_left(&ctx->flexible_groups);
1573 raw_spin_unlock(&ctx->lock);
1576 void perf_event_task_tick(struct task_struct *curr)
1578 struct perf_cpu_context *cpuctx;
1579 struct perf_event_context *ctx;
1582 if (!atomic_read(&nr_events))
1585 cpuctx = &__get_cpu_var(perf_cpu_context);
1586 if (cpuctx->ctx.nr_events &&
1587 cpuctx->ctx.nr_events != cpuctx->ctx.nr_active)
1590 ctx = curr->perf_event_ctxp;
1591 if (ctx && ctx->nr_events && ctx->nr_events != ctx->nr_active)
1594 perf_ctx_adjust_freq(&cpuctx->ctx);
1596 perf_ctx_adjust_freq(ctx);
1602 cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE);
1604 task_ctx_sched_out(ctx, EVENT_FLEXIBLE);
1606 rotate_ctx(&cpuctx->ctx);
1610 cpu_ctx_sched_in(cpuctx, EVENT_FLEXIBLE);
1612 task_ctx_sched_in(curr, EVENT_FLEXIBLE);
1616 static int event_enable_on_exec(struct perf_event *event,
1617 struct perf_event_context *ctx)
1619 if (!event->attr.enable_on_exec)
1622 event->attr.enable_on_exec = 0;
1623 if (event->state >= PERF_EVENT_STATE_INACTIVE)
1626 __perf_event_mark_enabled(event, ctx);
1632 * Enable all of a task's events that have been marked enable-on-exec.
1633 * This expects task == current.
1635 static void perf_event_enable_on_exec(struct task_struct *task)
1637 struct perf_event_context *ctx;
1638 struct perf_event *event;
1639 unsigned long flags;
1643 local_irq_save(flags);
1644 ctx = task->perf_event_ctxp;
1645 if (!ctx || !ctx->nr_events)
1648 __perf_event_task_sched_out(ctx);
1650 raw_spin_lock(&ctx->lock);
1652 list_for_each_entry(event, &ctx->pinned_groups, group_entry) {
1653 ret = event_enable_on_exec(event, ctx);
1658 list_for_each_entry(event, &ctx->flexible_groups, group_entry) {
1659 ret = event_enable_on_exec(event, ctx);
1665 * Unclone this context if we enabled any event.
1670 raw_spin_unlock(&ctx->lock);
1672 perf_event_task_sched_in(task);
1674 local_irq_restore(flags);
1678 * Cross CPU call to read the hardware event
1680 static void __perf_event_read(void *info)
1682 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
1683 struct perf_event *event = info;
1684 struct perf_event_context *ctx = event->ctx;
1687 * If this is a task context, we need to check whether it is
1688 * the current task context of this cpu. If not it has been
1689 * scheduled out before the smp call arrived. In that case
1690 * event->count would have been updated to a recent sample
1691 * when the event was scheduled out.
1693 if (ctx->task && cpuctx->task_ctx != ctx)
1696 raw_spin_lock(&ctx->lock);
1697 update_context_time(ctx);
1698 update_event_times(event);
1699 raw_spin_unlock(&ctx->lock);
1701 event->pmu->read(event);
1704 static u64 perf_event_read(struct perf_event *event)
1707 * If event is enabled and currently active on a CPU, update the
1708 * value in the event structure:
1710 if (event->state == PERF_EVENT_STATE_ACTIVE) {
1711 smp_call_function_single(event->oncpu,
1712 __perf_event_read, event, 1);
1713 } else if (event->state == PERF_EVENT_STATE_INACTIVE) {
1714 struct perf_event_context *ctx = event->ctx;
1715 unsigned long flags;
1717 raw_spin_lock_irqsave(&ctx->lock, flags);
1718 update_context_time(ctx);
1719 update_event_times(event);
1720 raw_spin_unlock_irqrestore(&ctx->lock, flags);
1723 return atomic64_read(&event->count);
1727 * Initialize the perf_event context in a task_struct:
1730 __perf_event_init_context(struct perf_event_context *ctx,
1731 struct task_struct *task)
1733 raw_spin_lock_init(&ctx->lock);
1734 mutex_init(&ctx->mutex);
1735 INIT_LIST_HEAD(&ctx->pinned_groups);
1736 INIT_LIST_HEAD(&ctx->flexible_groups);
1737 INIT_LIST_HEAD(&ctx->event_list);
1738 atomic_set(&ctx->refcount, 1);
1742 static struct perf_event_context *find_get_context(pid_t pid, int cpu)
1744 struct perf_event_context *ctx;
1745 struct perf_cpu_context *cpuctx;
1746 struct task_struct *task;
1747 unsigned long flags;
1750 if (pid == -1 && cpu != -1) {
1751 /* Must be root to operate on a CPU event: */
1752 if (perf_paranoid_cpu() && !capable(CAP_SYS_ADMIN))
1753 return ERR_PTR(-EACCES);
1755 if (cpu < 0 || cpu >= nr_cpumask_bits)
1756 return ERR_PTR(-EINVAL);
1759 * We could be clever and allow to attach a event to an
1760 * offline CPU and activate it when the CPU comes up, but
1763 if (!cpu_online(cpu))
1764 return ERR_PTR(-ENODEV);
1766 cpuctx = &per_cpu(perf_cpu_context, cpu);
1777 task = find_task_by_vpid(pid);
1779 get_task_struct(task);
1783 return ERR_PTR(-ESRCH);
1786 * Can't attach events to a dying task.
1789 if (task->flags & PF_EXITING)
1792 /* Reuse ptrace permission checks for now. */
1794 if (!ptrace_may_access(task, PTRACE_MODE_READ))
1798 ctx = perf_lock_task_context(task, &flags);
1801 raw_spin_unlock_irqrestore(&ctx->lock, flags);
1805 ctx = kzalloc(sizeof(struct perf_event_context), GFP_KERNEL);
1809 __perf_event_init_context(ctx, task);
1811 if (cmpxchg(&task->perf_event_ctxp, NULL, ctx)) {
1813 * We raced with some other task; use
1814 * the context they set.
1819 get_task_struct(task);
1822 put_task_struct(task);
1826 put_task_struct(task);
1827 return ERR_PTR(err);
1830 static void perf_event_free_filter(struct perf_event *event);
1832 static void free_event_rcu(struct rcu_head *head)
1834 struct perf_event *event;
1836 event = container_of(head, struct perf_event, rcu_head);
1838 put_pid_ns(event->ns);
1839 perf_event_free_filter(event);
1843 static void perf_pending_sync(struct perf_event *event);
1845 static void free_event(struct perf_event *event)
1847 perf_pending_sync(event);
1849 if (!event->parent) {
1850 atomic_dec(&nr_events);
1851 if (event->attr.mmap)
1852 atomic_dec(&nr_mmap_events);
1853 if (event->attr.comm)
1854 atomic_dec(&nr_comm_events);
1855 if (event->attr.task)
1856 atomic_dec(&nr_task_events);
1859 if (event->output) {
1860 fput(event->output->filp);
1861 event->output = NULL;
1865 event->destroy(event);
1867 put_ctx(event->ctx);
1868 call_rcu(&event->rcu_head, free_event_rcu);
1871 int perf_event_release_kernel(struct perf_event *event)
1873 struct perf_event_context *ctx = event->ctx;
1876 * Remove from the PMU, can't get re-enabled since we got
1877 * here because the last ref went.
1879 perf_event_disable(event);
1881 WARN_ON_ONCE(ctx->parent_ctx);
1883 * There are two ways this annotation is useful:
1885 * 1) there is a lock recursion from perf_event_exit_task
1886 * see the comment there.
1888 * 2) there is a lock-inversion with mmap_sem through
1889 * perf_event_read_group(), which takes faults while
1890 * holding ctx->mutex, however this is called after
1891 * the last filedesc died, so there is no possibility
1892 * to trigger the AB-BA case.
1894 mutex_lock_nested(&ctx->mutex, SINGLE_DEPTH_NESTING);
1895 raw_spin_lock_irq(&ctx->lock);
1896 list_del_event(event, ctx);
1897 perf_destroy_group(event, ctx);
1898 raw_spin_unlock_irq(&ctx->lock);
1899 mutex_unlock(&ctx->mutex);
1901 mutex_lock(&event->owner->perf_event_mutex);
1902 list_del_init(&event->owner_entry);
1903 mutex_unlock(&event->owner->perf_event_mutex);
1904 put_task_struct(event->owner);
1910 EXPORT_SYMBOL_GPL(perf_event_release_kernel);
1913 * Called when the last reference to the file is gone.
1915 static int perf_release(struct inode *inode, struct file *file)
1917 struct perf_event *event = file->private_data;
1919 file->private_data = NULL;
1921 return perf_event_release_kernel(event);
1924 static int perf_event_read_size(struct perf_event *event)
1926 int entry = sizeof(u64); /* value */
1930 if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1931 size += sizeof(u64);
1933 if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1934 size += sizeof(u64);
1936 if (event->attr.read_format & PERF_FORMAT_ID)
1937 entry += sizeof(u64);
1939 if (event->attr.read_format & PERF_FORMAT_GROUP) {
1940 nr += event->group_leader->nr_siblings;
1941 size += sizeof(u64);
1949 u64 perf_event_read_value(struct perf_event *event, u64 *enabled, u64 *running)
1951 struct perf_event *child;
1957 mutex_lock(&event->child_mutex);
1958 total += perf_event_read(event);
1959 *enabled += event->total_time_enabled +
1960 atomic64_read(&event->child_total_time_enabled);
1961 *running += event->total_time_running +
1962 atomic64_read(&event->child_total_time_running);
1964 list_for_each_entry(child, &event->child_list, child_list) {
1965 total += perf_event_read(child);
1966 *enabled += child->total_time_enabled;
1967 *running += child->total_time_running;
1969 mutex_unlock(&event->child_mutex);
1973 EXPORT_SYMBOL_GPL(perf_event_read_value);
1975 static int perf_event_read_group(struct perf_event *event,
1976 u64 read_format, char __user *buf)
1978 struct perf_event *leader = event->group_leader, *sub;
1979 int n = 0, size = 0, ret = -EFAULT;
1980 struct perf_event_context *ctx = leader->ctx;
1982 u64 count, enabled, running;
1984 mutex_lock(&ctx->mutex);
1985 count = perf_event_read_value(leader, &enabled, &running);
1987 values[n++] = 1 + leader->nr_siblings;
1988 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1989 values[n++] = enabled;
1990 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1991 values[n++] = running;
1992 values[n++] = count;
1993 if (read_format & PERF_FORMAT_ID)
1994 values[n++] = primary_event_id(leader);
1996 size = n * sizeof(u64);
1998 if (copy_to_user(buf, values, size))
2003 list_for_each_entry(sub, &leader->sibling_list, group_entry) {
2006 values[n++] = perf_event_read_value(sub, &enabled, &running);
2007 if (read_format & PERF_FORMAT_ID)
2008 values[n++] = primary_event_id(sub);
2010 size = n * sizeof(u64);
2012 if (copy_to_user(buf + ret, values, size)) {
2020 mutex_unlock(&ctx->mutex);
2025 static int perf_event_read_one(struct perf_event *event,
2026 u64 read_format, char __user *buf)
2028 u64 enabled, running;
2032 values[n++] = perf_event_read_value(event, &enabled, &running);
2033 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
2034 values[n++] = enabled;
2035 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
2036 values[n++] = running;
2037 if (read_format & PERF_FORMAT_ID)
2038 values[n++] = primary_event_id(event);
2040 if (copy_to_user(buf, values, n * sizeof(u64)))
2043 return n * sizeof(u64);
2047 * Read the performance event - simple non blocking version for now
2050 perf_read_hw(struct perf_event *event, char __user *buf, size_t count)
2052 u64 read_format = event->attr.read_format;
2056 * Return end-of-file for a read on a event that is in
2057 * error state (i.e. because it was pinned but it couldn't be
2058 * scheduled on to the CPU at some point).
2060 if (event->state == PERF_EVENT_STATE_ERROR)
2063 if (count < perf_event_read_size(event))
2066 WARN_ON_ONCE(event->ctx->parent_ctx);
2067 if (read_format & PERF_FORMAT_GROUP)
2068 ret = perf_event_read_group(event, read_format, buf);
2070 ret = perf_event_read_one(event, read_format, buf);
2076 perf_read(struct file *file, char __user *buf, size_t count, loff_t *ppos)
2078 struct perf_event *event = file->private_data;
2080 return perf_read_hw(event, buf, count);
2083 static unsigned int perf_poll(struct file *file, poll_table *wait)
2085 struct perf_event *event = file->private_data;
2086 struct perf_mmap_data *data;
2087 unsigned int events = POLL_HUP;
2090 data = rcu_dereference(event->data);
2092 events = atomic_xchg(&data->poll, 0);
2095 poll_wait(file, &event->waitq, wait);
2100 static void perf_event_reset(struct perf_event *event)
2102 (void)perf_event_read(event);
2103 atomic64_set(&event->count, 0);
2104 perf_event_update_userpage(event);
2108 * Holding the top-level event's child_mutex means that any
2109 * descendant process that has inherited this event will block
2110 * in sync_child_event if it goes to exit, thus satisfying the
2111 * task existence requirements of perf_event_enable/disable.
2113 static void perf_event_for_each_child(struct perf_event *event,
2114 void (*func)(struct perf_event *))
2116 struct perf_event *child;
2118 WARN_ON_ONCE(event->ctx->parent_ctx);
2119 mutex_lock(&event->child_mutex);
2121 list_for_each_entry(child, &event->child_list, child_list)
2123 mutex_unlock(&event->child_mutex);
2126 static void perf_event_for_each(struct perf_event *event,
2127 void (*func)(struct perf_event *))
2129 struct perf_event_context *ctx = event->ctx;
2130 struct perf_event *sibling;
2132 WARN_ON_ONCE(ctx->parent_ctx);
2133 mutex_lock(&ctx->mutex);
2134 event = event->group_leader;
2136 perf_event_for_each_child(event, func);
2138 list_for_each_entry(sibling, &event->sibling_list, group_entry)
2139 perf_event_for_each_child(event, func);
2140 mutex_unlock(&ctx->mutex);
2143 static int perf_event_period(struct perf_event *event, u64 __user *arg)
2145 struct perf_event_context *ctx = event->ctx;
2150 if (!event->attr.sample_period)
2153 size = copy_from_user(&value, arg, sizeof(value));
2154 if (size != sizeof(value))
2160 raw_spin_lock_irq(&ctx->lock);
2161 if (event->attr.freq) {
2162 if (value > sysctl_perf_event_sample_rate) {
2167 event->attr.sample_freq = value;
2169 event->attr.sample_period = value;
2170 event->hw.sample_period = value;
2173 raw_spin_unlock_irq(&ctx->lock);
2178 static int perf_event_set_output(struct perf_event *event, int output_fd);
2179 static int perf_event_set_filter(struct perf_event *event, void __user *arg);
2181 static long perf_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
2183 struct perf_event *event = file->private_data;
2184 void (*func)(struct perf_event *);
2188 case PERF_EVENT_IOC_ENABLE:
2189 func = perf_event_enable;
2191 case PERF_EVENT_IOC_DISABLE:
2192 func = perf_event_disable;
2194 case PERF_EVENT_IOC_RESET:
2195 func = perf_event_reset;
2198 case PERF_EVENT_IOC_REFRESH:
2199 return perf_event_refresh(event, arg);
2201 case PERF_EVENT_IOC_PERIOD:
2202 return perf_event_period(event, (u64 __user *)arg);
2204 case PERF_EVENT_IOC_SET_OUTPUT:
2205 return perf_event_set_output(event, arg);
2207 case PERF_EVENT_IOC_SET_FILTER:
2208 return perf_event_set_filter(event, (void __user *)arg);
2214 if (flags & PERF_IOC_FLAG_GROUP)
2215 perf_event_for_each(event, func);
2217 perf_event_for_each_child(event, func);
2222 int perf_event_task_enable(void)
2224 struct perf_event *event;
2226 mutex_lock(¤t->perf_event_mutex);
2227 list_for_each_entry(event, ¤t->perf_event_list, owner_entry)
2228 perf_event_for_each_child(event, perf_event_enable);
2229 mutex_unlock(¤t->perf_event_mutex);
2234 int perf_event_task_disable(void)
2236 struct perf_event *event;
2238 mutex_lock(¤t->perf_event_mutex);
2239 list_for_each_entry(event, ¤t->perf_event_list, owner_entry)
2240 perf_event_for_each_child(event, perf_event_disable);
2241 mutex_unlock(¤t->perf_event_mutex);
2246 #ifndef PERF_EVENT_INDEX_OFFSET
2247 # define PERF_EVENT_INDEX_OFFSET 0
2250 static int perf_event_index(struct perf_event *event)
2252 if (event->state != PERF_EVENT_STATE_ACTIVE)
2255 return event->hw.idx + 1 - PERF_EVENT_INDEX_OFFSET;
2259 * Callers need to ensure there can be no nesting of this function, otherwise
2260 * the seqlock logic goes bad. We can not serialize this because the arch
2261 * code calls this from NMI context.
2263 void perf_event_update_userpage(struct perf_event *event)
2265 struct perf_event_mmap_page *userpg;
2266 struct perf_mmap_data *data;
2269 data = rcu_dereference(event->data);
2273 userpg = data->user_page;
2276 * Disable preemption so as to not let the corresponding user-space
2277 * spin too long if we get preempted.
2282 userpg->index = perf_event_index(event);
2283 userpg->offset = atomic64_read(&event->count);
2284 if (event->state == PERF_EVENT_STATE_ACTIVE)
2285 userpg->offset -= atomic64_read(&event->hw.prev_count);
2287 userpg->time_enabled = event->total_time_enabled +
2288 atomic64_read(&event->child_total_time_enabled);
2290 userpg->time_running = event->total_time_running +
2291 atomic64_read(&event->child_total_time_running);
2300 static unsigned long perf_data_size(struct perf_mmap_data *data)
2302 return data->nr_pages << (PAGE_SHIFT + data->data_order);
2305 #ifndef CONFIG_PERF_USE_VMALLOC
2308 * Back perf_mmap() with regular GFP_KERNEL-0 pages.
2311 static struct page *
2312 perf_mmap_to_page(struct perf_mmap_data *data, unsigned long pgoff)
2314 if (pgoff > data->nr_pages)
2318 return virt_to_page(data->user_page);
2320 return virt_to_page(data->data_pages[pgoff - 1]);
2323 static void *perf_mmap_alloc_page(int cpu)
2328 node = (cpu == -1) ? cpu : cpu_to_node(cpu);
2329 page = alloc_pages_node(node, GFP_KERNEL | __GFP_ZERO, 0);
2333 return page_address(page);
2336 static struct perf_mmap_data *
2337 perf_mmap_data_alloc(struct perf_event *event, int nr_pages)
2339 struct perf_mmap_data *data;
2343 WARN_ON(atomic_read(&event->mmap_count));
2345 size = sizeof(struct perf_mmap_data);
2346 size += nr_pages * sizeof(void *);
2348 data = kzalloc(size, GFP_KERNEL);
2352 data->user_page = perf_mmap_alloc_page(event->cpu);
2353 if (!data->user_page)
2354 goto fail_user_page;
2356 for (i = 0; i < nr_pages; i++) {
2357 data->data_pages[i] = perf_mmap_alloc_page(event->cpu);
2358 if (!data->data_pages[i])
2359 goto fail_data_pages;
2362 data->data_order = 0;
2363 data->nr_pages = nr_pages;
2368 for (i--; i >= 0; i--)
2369 free_page((unsigned long)data->data_pages[i]);
2371 free_page((unsigned long)data->user_page);
2380 static void perf_mmap_free_page(unsigned long addr)
2382 struct page *page = virt_to_page((void *)addr);
2384 page->mapping = NULL;
2388 static void perf_mmap_data_free(struct perf_mmap_data *data)
2392 perf_mmap_free_page((unsigned long)data->user_page);
2393 for (i = 0; i < data->nr_pages; i++)
2394 perf_mmap_free_page((unsigned long)data->data_pages[i]);
2401 * Back perf_mmap() with vmalloc memory.
2403 * Required for architectures that have d-cache aliasing issues.
2406 static struct page *
2407 perf_mmap_to_page(struct perf_mmap_data *data, unsigned long pgoff)
2409 if (pgoff > (1UL << data->data_order))
2412 return vmalloc_to_page((void *)data->user_page + pgoff * PAGE_SIZE);
2415 static void perf_mmap_unmark_page(void *addr)
2417 struct page *page = vmalloc_to_page(addr);
2419 page->mapping = NULL;
2422 static void perf_mmap_data_free_work(struct work_struct *work)
2424 struct perf_mmap_data *data;
2428 data = container_of(work, struct perf_mmap_data, work);
2429 nr = 1 << data->data_order;
2431 base = data->user_page;
2432 for (i = 0; i < nr + 1; i++)
2433 perf_mmap_unmark_page(base + (i * PAGE_SIZE));
2439 static void perf_mmap_data_free(struct perf_mmap_data *data)
2441 schedule_work(&data->work);
2444 static struct perf_mmap_data *
2445 perf_mmap_data_alloc(struct perf_event *event, int nr_pages)
2447 struct perf_mmap_data *data;
2451 WARN_ON(atomic_read(&event->mmap_count));
2453 size = sizeof(struct perf_mmap_data);
2454 size += sizeof(void *);
2456 data = kzalloc(size, GFP_KERNEL);
2460 INIT_WORK(&data->work, perf_mmap_data_free_work);
2462 all_buf = vmalloc_user((nr_pages + 1) * PAGE_SIZE);
2466 data->user_page = all_buf;
2467 data->data_pages[0] = all_buf + PAGE_SIZE;
2468 data->data_order = ilog2(nr_pages);
2482 static int perf_mmap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
2484 struct perf_event *event = vma->vm_file->private_data;
2485 struct perf_mmap_data *data;
2486 int ret = VM_FAULT_SIGBUS;
2488 if (vmf->flags & FAULT_FLAG_MKWRITE) {
2489 if (vmf->pgoff == 0)
2495 data = rcu_dereference(event->data);
2499 if (vmf->pgoff && (vmf->flags & FAULT_FLAG_WRITE))
2502 vmf->page = perf_mmap_to_page(data, vmf->pgoff);
2506 get_page(vmf->page);
2507 vmf->page->mapping = vma->vm_file->f_mapping;
2508 vmf->page->index = vmf->pgoff;
2518 perf_mmap_data_init(struct perf_event *event, struct perf_mmap_data *data)
2520 long max_size = perf_data_size(data);
2522 if (event->attr.watermark) {
2523 data->watermark = min_t(long, max_size,
2524 event->attr.wakeup_watermark);
2527 if (!data->watermark)
2528 data->watermark = max_size / 2;
2531 rcu_assign_pointer(event->data, data);
2534 static void perf_mmap_data_free_rcu(struct rcu_head *rcu_head)
2536 struct perf_mmap_data *data;
2538 data = container_of(rcu_head, struct perf_mmap_data, rcu_head);
2539 perf_mmap_data_free(data);
2542 static void perf_mmap_data_release(struct perf_event *event)
2544 struct perf_mmap_data *data = event->data;
2546 WARN_ON(atomic_read(&event->mmap_count));
2548 rcu_assign_pointer(event->data, NULL);
2549 call_rcu(&data->rcu_head, perf_mmap_data_free_rcu);
2552 static void perf_mmap_open(struct vm_area_struct *vma)
2554 struct perf_event *event = vma->vm_file->private_data;
2556 atomic_inc(&event->mmap_count);
2559 static void perf_mmap_close(struct vm_area_struct *vma)
2561 struct perf_event *event = vma->vm_file->private_data;
2563 WARN_ON_ONCE(event->ctx->parent_ctx);
2564 if (atomic_dec_and_mutex_lock(&event->mmap_count, &event->mmap_mutex)) {
2565 unsigned long size = perf_data_size(event->data);
2566 struct user_struct *user = current_user();
2568 atomic_long_sub((size >> PAGE_SHIFT) + 1, &user->locked_vm);
2569 vma->vm_mm->locked_vm -= event->data->nr_locked;
2570 perf_mmap_data_release(event);
2571 mutex_unlock(&event->mmap_mutex);
2575 static const struct vm_operations_struct perf_mmap_vmops = {
2576 .open = perf_mmap_open,
2577 .close = perf_mmap_close,
2578 .fault = perf_mmap_fault,
2579 .page_mkwrite = perf_mmap_fault,
2582 static int perf_mmap(struct file *file, struct vm_area_struct *vma)
2584 struct perf_event *event = file->private_data;
2585 unsigned long user_locked, user_lock_limit;
2586 struct user_struct *user = current_user();
2587 unsigned long locked, lock_limit;
2588 struct perf_mmap_data *data;
2589 unsigned long vma_size;
2590 unsigned long nr_pages;
2591 long user_extra, extra;
2595 * Don't allow mmap() of inherited per-task counters. This would
2596 * create a performance issue due to all children writing to the
2599 if (event->cpu == -1 && event->attr.inherit)
2602 if (!(vma->vm_flags & VM_SHARED))
2605 vma_size = vma->vm_end - vma->vm_start;
2606 nr_pages = (vma_size / PAGE_SIZE) - 1;
2609 * If we have data pages ensure they're a power-of-two number, so we
2610 * can do bitmasks instead of modulo.
2612 if (nr_pages != 0 && !is_power_of_2(nr_pages))
2615 if (vma_size != PAGE_SIZE * (1 + nr_pages))
2618 if (vma->vm_pgoff != 0)
2621 WARN_ON_ONCE(event->ctx->parent_ctx);
2622 mutex_lock(&event->mmap_mutex);
2623 if (event->output) {
2628 if (atomic_inc_not_zero(&event->mmap_count)) {
2629 if (nr_pages != event->data->nr_pages)
2634 user_extra = nr_pages + 1;
2635 user_lock_limit = sysctl_perf_event_mlock >> (PAGE_SHIFT - 10);
2638 * Increase the limit linearly with more CPUs:
2640 user_lock_limit *= num_online_cpus();
2642 user_locked = atomic_long_read(&user->locked_vm) + user_extra;
2645 if (user_locked > user_lock_limit)
2646 extra = user_locked - user_lock_limit;
2648 lock_limit = rlimit(RLIMIT_MEMLOCK);
2649 lock_limit >>= PAGE_SHIFT;
2650 locked = vma->vm_mm->locked_vm + extra;
2652 if ((locked > lock_limit) && perf_paranoid_tracepoint_raw() &&
2653 !capable(CAP_IPC_LOCK)) {
2658 WARN_ON(event->data);
2660 data = perf_mmap_data_alloc(event, nr_pages);
2666 perf_mmap_data_init(event, data);
2668 atomic_set(&event->mmap_count, 1);
2669 atomic_long_add(user_extra, &user->locked_vm);
2670 vma->vm_mm->locked_vm += extra;
2671 event->data->nr_locked = extra;
2672 if (vma->vm_flags & VM_WRITE)
2673 event->data->writable = 1;
2676 mutex_unlock(&event->mmap_mutex);
2678 vma->vm_flags |= VM_RESERVED;
2679 vma->vm_ops = &perf_mmap_vmops;
2684 static int perf_fasync(int fd, struct file *filp, int on)
2686 struct inode *inode = filp->f_path.dentry->d_inode;
2687 struct perf_event *event = filp->private_data;
2690 mutex_lock(&inode->i_mutex);
2691 retval = fasync_helper(fd, filp, on, &event->fasync);
2692 mutex_unlock(&inode->i_mutex);
2700 static const struct file_operations perf_fops = {
2701 .llseek = no_llseek,
2702 .release = perf_release,
2705 .unlocked_ioctl = perf_ioctl,
2706 .compat_ioctl = perf_ioctl,
2708 .fasync = perf_fasync,
2714 * If there's data, ensure we set the poll() state and publish everything
2715 * to user-space before waking everybody up.
2718 void perf_event_wakeup(struct perf_event *event)
2720 wake_up_all(&event->waitq);
2722 if (event->pending_kill) {
2723 kill_fasync(&event->fasync, SIGIO, event->pending_kill);
2724 event->pending_kill = 0;
2731 * Handle the case where we need to wakeup up from NMI (or rq->lock) context.
2733 * The NMI bit means we cannot possibly take locks. Therefore, maintain a
2734 * single linked list and use cmpxchg() to add entries lockless.
2737 static void perf_pending_event(struct perf_pending_entry *entry)
2739 struct perf_event *event = container_of(entry,
2740 struct perf_event, pending);
2742 if (event->pending_disable) {
2743 event->pending_disable = 0;
2744 __perf_event_disable(event);
2747 if (event->pending_wakeup) {
2748 event->pending_wakeup = 0;
2749 perf_event_wakeup(event);
2753 #define PENDING_TAIL ((struct perf_pending_entry *)-1UL)
2755 static DEFINE_PER_CPU(struct perf_pending_entry *, perf_pending_head) = {
2759 static void perf_pending_queue(struct perf_pending_entry *entry,
2760 void (*func)(struct perf_pending_entry *))
2762 struct perf_pending_entry **head;
2764 if (cmpxchg(&entry->next, NULL, PENDING_TAIL) != NULL)
2769 head = &get_cpu_var(perf_pending_head);
2772 entry->next = *head;
2773 } while (cmpxchg(head, entry->next, entry) != entry->next);
2775 set_perf_event_pending();
2777 put_cpu_var(perf_pending_head);
2780 static int __perf_pending_run(void)
2782 struct perf_pending_entry *list;
2785 list = xchg(&__get_cpu_var(perf_pending_head), PENDING_TAIL);
2786 while (list != PENDING_TAIL) {
2787 void (*func)(struct perf_pending_entry *);
2788 struct perf_pending_entry *entry = list;
2795 * Ensure we observe the unqueue before we issue the wakeup,
2796 * so that we won't be waiting forever.
2797 * -- see perf_not_pending().
2808 static inline int perf_not_pending(struct perf_event *event)
2811 * If we flush on whatever cpu we run, there is a chance we don't
2815 __perf_pending_run();
2819 * Ensure we see the proper queue state before going to sleep
2820 * so that we do not miss the wakeup. -- see perf_pending_handle()
2823 return event->pending.next == NULL;
2826 static void perf_pending_sync(struct perf_event *event)
2828 wait_event(event->waitq, perf_not_pending(event));
2831 void perf_event_do_pending(void)
2833 __perf_pending_run();
2837 * Callchain support -- arch specific
2840 __weak struct perf_callchain_entry *perf_callchain(struct pt_regs *regs)
2846 void perf_arch_fetch_caller_regs(struct pt_regs *regs, unsigned long ip, int skip)
2852 * We assume there is only KVM supporting the callbacks.
2853 * Later on, we might change it to a list if there is
2854 * another virtualization implementation supporting the callbacks.
2856 struct perf_guest_info_callbacks *perf_guest_cbs;
2858 int perf_register_guest_info_callbacks(struct perf_guest_info_callbacks *cbs)
2860 perf_guest_cbs = cbs;
2863 EXPORT_SYMBOL_GPL(perf_register_guest_info_callbacks);
2865 int perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks *cbs)
2867 perf_guest_cbs = NULL;
2870 EXPORT_SYMBOL_GPL(perf_unregister_guest_info_callbacks);
2875 static bool perf_output_space(struct perf_mmap_data *data, unsigned long tail,
2876 unsigned long offset, unsigned long head)
2880 if (!data->writable)
2883 mask = perf_data_size(data) - 1;
2885 offset = (offset - tail) & mask;
2886 head = (head - tail) & mask;
2888 if ((int)(head - offset) < 0)
2894 static void perf_output_wakeup(struct perf_output_handle *handle)
2896 atomic_set(&handle->data->poll, POLL_IN);
2899 handle->event->pending_wakeup = 1;
2900 perf_pending_queue(&handle->event->pending,
2901 perf_pending_event);
2903 perf_event_wakeup(handle->event);
2907 * We need to ensure a later event_id doesn't publish a head when a former
2908 * event isn't done writing. However since we need to deal with NMIs we
2909 * cannot fully serialize things.
2911 * We only publish the head (and generate a wakeup) when the outer-most
2914 static void perf_output_get_handle(struct perf_output_handle *handle)
2916 struct perf_mmap_data *data = handle->data;
2919 local_inc(&data->nest);
2920 handle->wakeup = local_read(&data->wakeup);
2923 static void perf_output_put_handle(struct perf_output_handle *handle)
2925 struct perf_mmap_data *data = handle->data;
2929 head = local_read(&data->head);
2932 * IRQ/NMI can happen here, which means we can miss a head update.
2935 if (!local_dec_and_test(&data->nest))
2939 * Publish the known good head. Rely on the full barrier implied
2940 * by atomic_dec_and_test() order the data->head read and this
2943 data->user_page->data_head = head;
2946 * Now check if we missed an update, rely on the (compiler)
2947 * barrier in atomic_dec_and_test() to re-read data->head.
2949 if (unlikely(head != local_read(&data->head))) {
2950 local_inc(&data->nest);
2954 if (handle->wakeup != local_read(&data->wakeup))
2955 perf_output_wakeup(handle);
2961 void perf_output_copy(struct perf_output_handle *handle,
2962 const void *buf, unsigned int len)
2964 unsigned int pages_mask;
2965 unsigned long offset;
2969 offset = handle->offset;
2970 pages_mask = handle->data->nr_pages - 1;
2971 pages = handle->data->data_pages;
2974 unsigned long page_offset;
2975 unsigned long page_size;
2978 nr = (offset >> PAGE_SHIFT) & pages_mask;
2979 page_size = 1UL << (handle->data->data_order + PAGE_SHIFT);
2980 page_offset = offset & (page_size - 1);
2981 size = min_t(unsigned int, page_size - page_offset, len);
2983 memcpy(pages[nr] + page_offset, buf, size);
2990 handle->offset = offset;
2993 * Check we didn't copy past our reservation window, taking the
2994 * possible unsigned int wrap into account.
2996 WARN_ON_ONCE(((long)(handle->head - handle->offset)) < 0);
2999 int perf_output_begin(struct perf_output_handle *handle,
3000 struct perf_event *event, unsigned int size,
3001 int nmi, int sample)
3003 struct perf_event *output_event;
3004 struct perf_mmap_data *data;
3005 unsigned long tail, offset, head;
3008 struct perf_event_header header;
3015 * For inherited events we send all the output towards the parent.
3018 event = event->parent;
3020 output_event = rcu_dereference(event->output);
3022 event = output_event;
3024 data = rcu_dereference(event->data);
3028 handle->data = data;
3029 handle->event = event;
3031 handle->sample = sample;
3033 if (!data->nr_pages)
3036 have_lost = local_read(&data->lost);
3038 size += sizeof(lost_event);
3040 perf_output_get_handle(handle);
3044 * Userspace could choose to issue a mb() before updating the
3045 * tail pointer. So that all reads will be completed before the
3048 tail = ACCESS_ONCE(data->user_page->data_tail);
3050 offset = head = local_read(&data->head);
3052 if (unlikely(!perf_output_space(data, tail, offset, head)))
3054 } while (local_cmpxchg(&data->head, offset, head) != offset);
3056 handle->offset = offset;
3057 handle->head = head;
3059 if (head - tail > data->watermark)
3060 local_inc(&data->wakeup);
3063 lost_event.header.type = PERF_RECORD_LOST;
3064 lost_event.header.misc = 0;
3065 lost_event.header.size = sizeof(lost_event);
3066 lost_event.id = event->id;
3067 lost_event.lost = local_xchg(&data->lost, 0);
3069 perf_output_put(handle, lost_event);
3075 local_inc(&data->lost);
3076 perf_output_put_handle(handle);
3083 void perf_output_end(struct perf_output_handle *handle)
3085 struct perf_event *event = handle->event;
3086 struct perf_mmap_data *data = handle->data;
3088 int wakeup_events = event->attr.wakeup_events;
3090 if (handle->sample && wakeup_events) {
3091 int events = local_inc_return(&data->events);
3092 if (events >= wakeup_events) {
3093 local_sub(wakeup_events, &data->events);
3094 local_inc(&data->wakeup);
3098 perf_output_put_handle(handle);
3102 static u32 perf_event_pid(struct perf_event *event, struct task_struct *p)
3105 * only top level events have the pid namespace they were created in
3108 event = event->parent;
3110 return task_tgid_nr_ns(p, event->ns);
3113 static u32 perf_event_tid(struct perf_event *event, struct task_struct *p)
3116 * only top level events have the pid namespace they were created in
3119 event = event->parent;
3121 return task_pid_nr_ns(p, event->ns);
3124 static void perf_output_read_one(struct perf_output_handle *handle,
3125 struct perf_event *event)
3127 u64 read_format = event->attr.read_format;
3131 values[n++] = atomic64_read(&event->count);
3132 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
3133 values[n++] = event->total_time_enabled +
3134 atomic64_read(&event->child_total_time_enabled);
3136 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
3137 values[n++] = event->total_time_running +
3138 atomic64_read(&event->child_total_time_running);
3140 if (read_format & PERF_FORMAT_ID)
3141 values[n++] = primary_event_id(event);
3143 perf_output_copy(handle, values, n * sizeof(u64));
3147 * XXX PERF_FORMAT_GROUP vs inherited events seems difficult.
3149 static void perf_output_read_group(struct perf_output_handle *handle,
3150 struct perf_event *event)
3152 struct perf_event *leader = event->group_leader, *sub;
3153 u64 read_format = event->attr.read_format;
3157 values[n++] = 1 + leader->nr_siblings;
3159 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
3160 values[n++] = leader->total_time_enabled;
3162 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
3163 values[n++] = leader->total_time_running;
3165 if (leader != event)
3166 leader->pmu->read(leader);
3168 values[n++] = atomic64_read(&leader->count);
3169 if (read_format & PERF_FORMAT_ID)
3170 values[n++] = primary_event_id(leader);
3172 perf_output_copy(handle, values, n * sizeof(u64));
3174 list_for_each_entry(sub, &leader->sibling_list, group_entry) {
3178 sub->pmu->read(sub);
3180 values[n++] = atomic64_read(&sub->count);
3181 if (read_format & PERF_FORMAT_ID)
3182 values[n++] = primary_event_id(sub);
3184 perf_output_copy(handle, values, n * sizeof(u64));
3188 static void perf_output_read(struct perf_output_handle *handle,
3189 struct perf_event *event)
3191 if (event->attr.read_format & PERF_FORMAT_GROUP)
3192 perf_output_read_group(handle, event);
3194 perf_output_read_one(handle, event);
3197 void perf_output_sample(struct perf_output_handle *handle,
3198 struct perf_event_header *header,
3199 struct perf_sample_data *data,
3200 struct perf_event *event)
3202 u64 sample_type = data->type;
3204 perf_output_put(handle, *header);
3206 if (sample_type & PERF_SAMPLE_IP)
3207 perf_output_put(handle, data->ip);
3209 if (sample_type & PERF_SAMPLE_TID)
3210 perf_output_put(handle, data->tid_entry);
3212 if (sample_type & PERF_SAMPLE_TIME)
3213 perf_output_put(handle, data->time);
3215 if (sample_type & PERF_SAMPLE_ADDR)
3216 perf_output_put(handle, data->addr);
3218 if (sample_type & PERF_SAMPLE_ID)
3219 perf_output_put(handle, data->id);
3221 if (sample_type & PERF_SAMPLE_STREAM_ID)
3222 perf_output_put(handle, data->stream_id);
3224 if (sample_type & PERF_SAMPLE_CPU)
3225 perf_output_put(handle, data->cpu_entry);
3227 if (sample_type & PERF_SAMPLE_PERIOD)
3228 perf_output_put(handle, data->period);
3230 if (sample_type & PERF_SAMPLE_READ)
3231 perf_output_read(handle, event);
3233 if (sample_type & PERF_SAMPLE_CALLCHAIN) {
3234 if (data->callchain) {
3237 if (data->callchain)
3238 size += data->callchain->nr;
3240 size *= sizeof(u64);
3242 perf_output_copy(handle, data->callchain, size);
3245 perf_output_put(handle, nr);
3249 if (sample_type & PERF_SAMPLE_RAW) {
3251 perf_output_put(handle, data->raw->size);
3252 perf_output_copy(handle, data->raw->data,
3259 .size = sizeof(u32),
3262 perf_output_put(handle, raw);
3267 void perf_prepare_sample(struct perf_event_header *header,
3268 struct perf_sample_data *data,
3269 struct perf_event *event,
3270 struct pt_regs *regs)
3272 u64 sample_type = event->attr.sample_type;
3274 data->type = sample_type;
3276 header->type = PERF_RECORD_SAMPLE;
3277 header->size = sizeof(*header);
3280 header->misc |= perf_misc_flags(regs);
3282 if (sample_type & PERF_SAMPLE_IP) {
3283 data->ip = perf_instruction_pointer(regs);
3285 header->size += sizeof(data->ip);
3288 if (sample_type & PERF_SAMPLE_TID) {
3289 /* namespace issues */
3290 data->tid_entry.pid = perf_event_pid(event, current);
3291 data->tid_entry.tid = perf_event_tid(event, current);
3293 header->size += sizeof(data->tid_entry);
3296 if (sample_type & PERF_SAMPLE_TIME) {
3297 data->time = perf_clock();
3299 header->size += sizeof(data->time);
3302 if (sample_type & PERF_SAMPLE_ADDR)
3303 header->size += sizeof(data->addr);
3305 if (sample_type & PERF_SAMPLE_ID) {
3306 data->id = primary_event_id(event);
3308 header->size += sizeof(data->id);
3311 if (sample_type & PERF_SAMPLE_STREAM_ID) {
3312 data->stream_id = event->id;
3314 header->size += sizeof(data->stream_id);
3317 if (sample_type & PERF_SAMPLE_CPU) {
3318 data->cpu_entry.cpu = raw_smp_processor_id();
3319 data->cpu_entry.reserved = 0;
3321 header->size += sizeof(data->cpu_entry);
3324 if (sample_type & PERF_SAMPLE_PERIOD)
3325 header->size += sizeof(data->period);
3327 if (sample_type & PERF_SAMPLE_READ)
3328 header->size += perf_event_read_size(event);
3330 if (sample_type & PERF_SAMPLE_CALLCHAIN) {
3333 data->callchain = perf_callchain(regs);
3335 if (data->callchain)
3336 size += data->callchain->nr;
3338 header->size += size * sizeof(u64);
3341 if (sample_type & PERF_SAMPLE_RAW) {
3342 int size = sizeof(u32);
3345 size += data->raw->size;
3347 size += sizeof(u32);
3349 WARN_ON_ONCE(size & (sizeof(u64)-1));
3350 header->size += size;
3354 static void perf_event_output(struct perf_event *event, int nmi,
3355 struct perf_sample_data *data,
3356 struct pt_regs *regs)
3358 struct perf_output_handle handle;
3359 struct perf_event_header header;
3361 perf_prepare_sample(&header, data, event, regs);
3363 if (perf_output_begin(&handle, event, header.size, nmi, 1))
3366 perf_output_sample(&handle, &header, data, event);
3368 perf_output_end(&handle);
3375 struct perf_read_event {
3376 struct perf_event_header header;
3383 perf_event_read_event(struct perf_event *event,
3384 struct task_struct *task)
3386 struct perf_output_handle handle;
3387 struct perf_read_event read_event = {
3389 .type = PERF_RECORD_READ,
3391 .size = sizeof(read_event) + perf_event_read_size(event),
3393 .pid = perf_event_pid(event, task),
3394 .tid = perf_event_tid(event, task),
3398 ret = perf_output_begin(&handle, event, read_event.header.size, 0, 0);
3402 perf_output_put(&handle, read_event);
3403 perf_output_read(&handle, event);
3405 perf_output_end(&handle);
3409 * task tracking -- fork/exit
3411 * enabled by: attr.comm | attr.mmap | attr.task
3414 struct perf_task_event {
3415 struct task_struct *task;
3416 struct perf_event_context *task_ctx;
3419 struct perf_event_header header;
3429 static void perf_event_task_output(struct perf_event *event,
3430 struct perf_task_event *task_event)
3432 struct perf_output_handle handle;
3433 struct task_struct *task = task_event->task;
3436 size = task_event->event_id.header.size;
3437 ret = perf_output_begin(&handle, event, size, 0, 0);
3442 task_event->event_id.pid = perf_event_pid(event, task);
3443 task_event->event_id.ppid = perf_event_pid(event, current);
3445 task_event->event_id.tid = perf_event_tid(event, task);
3446 task_event->event_id.ptid = perf_event_tid(event, current);
3448 perf_output_put(&handle, task_event->event_id);
3450 perf_output_end(&handle);
3453 static int perf_event_task_match(struct perf_event *event)
3455 if (event->state < PERF_EVENT_STATE_INACTIVE)
3458 if (event->cpu != -1 && event->cpu != smp_processor_id())
3461 if (event->attr.comm || event->attr.mmap || event->attr.task)
3467 static void perf_event_task_ctx(struct perf_event_context *ctx,
3468 struct perf_task_event *task_event)
3470 struct perf_event *event;
3472 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
3473 if (perf_event_task_match(event))
3474 perf_event_task_output(event, task_event);
3478 static void perf_event_task_event(struct perf_task_event *task_event)
3480 struct perf_cpu_context *cpuctx;
3481 struct perf_event_context *ctx = task_event->task_ctx;
3484 cpuctx = &get_cpu_var(perf_cpu_context);
3485 perf_event_task_ctx(&cpuctx->ctx, task_event);
3487 ctx = rcu_dereference(current->perf_event_ctxp);
3489 perf_event_task_ctx(ctx, task_event);
3490 put_cpu_var(perf_cpu_context);
3494 static void perf_event_task(struct task_struct *task,
3495 struct perf_event_context *task_ctx,
3498 struct perf_task_event task_event;
3500 if (!atomic_read(&nr_comm_events) &&
3501 !atomic_read(&nr_mmap_events) &&
3502 !atomic_read(&nr_task_events))
3505 task_event = (struct perf_task_event){
3507 .task_ctx = task_ctx,
3510 .type = new ? PERF_RECORD_FORK : PERF_RECORD_EXIT,
3512 .size = sizeof(task_event.event_id),
3518 .time = perf_clock(),
3522 perf_event_task_event(&task_event);
3525 void perf_event_fork(struct task_struct *task)
3527 perf_event_task(task, NULL, 1);
3534 struct perf_comm_event {
3535 struct task_struct *task;
3540 struct perf_event_header header;
3547 static void perf_event_comm_output(struct perf_event *event,
3548 struct perf_comm_event *comm_event)
3550 struct perf_output_handle handle;
3551 int size = comm_event->event_id.header.size;
3552 int ret = perf_output_begin(&handle, event, size, 0, 0);
3557 comm_event->event_id.pid = perf_event_pid(event, comm_event->task);
3558 comm_event->event_id.tid = perf_event_tid(event, comm_event->task);
3560 perf_output_put(&handle, comm_event->event_id);
3561 perf_output_copy(&handle, comm_event->comm,
3562 comm_event->comm_size);
3563 perf_output_end(&handle);
3566 static int perf_event_comm_match(struct perf_event *event)
3568 if (event->state < PERF_EVENT_STATE_INACTIVE)
3571 if (event->cpu != -1 && event->cpu != smp_processor_id())
3574 if (event->attr.comm)
3580 static void perf_event_comm_ctx(struct perf_event_context *ctx,
3581 struct perf_comm_event *comm_event)
3583 struct perf_event *event;
3585 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
3586 if (perf_event_comm_match(event))
3587 perf_event_comm_output(event, comm_event);
3591 static void perf_event_comm_event(struct perf_comm_event *comm_event)
3593 struct perf_cpu_context *cpuctx;
3594 struct perf_event_context *ctx;
3596 char comm[TASK_COMM_LEN];
3598 memset(comm, 0, sizeof(comm));
3599 strlcpy(comm, comm_event->task->comm, sizeof(comm));
3600 size = ALIGN(strlen(comm)+1, sizeof(u64));
3602 comm_event->comm = comm;
3603 comm_event->comm_size = size;
3605 comm_event->event_id.header.size = sizeof(comm_event->event_id) + size;
3608 cpuctx = &get_cpu_var(perf_cpu_context);
3609 perf_event_comm_ctx(&cpuctx->ctx, comm_event);
3610 ctx = rcu_dereference(current->perf_event_ctxp);
3612 perf_event_comm_ctx(ctx, comm_event);
3613 put_cpu_var(perf_cpu_context);
3617 void perf_event_comm(struct task_struct *task)
3619 struct perf_comm_event comm_event;
3621 if (task->perf_event_ctxp)
3622 perf_event_enable_on_exec(task);
3624 if (!atomic_read(&nr_comm_events))
3627 comm_event = (struct perf_comm_event){
3633 .type = PERF_RECORD_COMM,
3642 perf_event_comm_event(&comm_event);
3649 struct perf_mmap_event {
3650 struct vm_area_struct *vma;
3652 const char *file_name;
3656 struct perf_event_header header;
3666 static void perf_event_mmap_output(struct perf_event *event,
3667 struct perf_mmap_event *mmap_event)
3669 struct perf_output_handle handle;
3670 int size = mmap_event->event_id.header.size;
3671 int ret = perf_output_begin(&handle, event, size, 0, 0);
3676 mmap_event->event_id.pid = perf_event_pid(event, current);
3677 mmap_event->event_id.tid = perf_event_tid(event, current);
3679 perf_output_put(&handle, mmap_event->event_id);
3680 perf_output_copy(&handle, mmap_event->file_name,
3681 mmap_event->file_size);
3682 perf_output_end(&handle);
3685 static int perf_event_mmap_match(struct perf_event *event,
3686 struct perf_mmap_event *mmap_event)
3688 if (event->state < PERF_EVENT_STATE_INACTIVE)
3691 if (event->cpu != -1 && event->cpu != smp_processor_id())
3694 if (event->attr.mmap)
3700 static void perf_event_mmap_ctx(struct perf_event_context *ctx,
3701 struct perf_mmap_event *mmap_event)
3703 struct perf_event *event;
3705 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
3706 if (perf_event_mmap_match(event, mmap_event))
3707 perf_event_mmap_output(event, mmap_event);
3711 static void perf_event_mmap_event(struct perf_mmap_event *mmap_event)
3713 struct perf_cpu_context *cpuctx;
3714 struct perf_event_context *ctx;
3715 struct vm_area_struct *vma = mmap_event->vma;
3716 struct file *file = vma->vm_file;
3722 memset(tmp, 0, sizeof(tmp));
3726 * d_path works from the end of the buffer backwards, so we
3727 * need to add enough zero bytes after the string to handle
3728 * the 64bit alignment we do later.
3730 buf = kzalloc(PATH_MAX + sizeof(u64), GFP_KERNEL);
3732 name = strncpy(tmp, "//enomem", sizeof(tmp));
3735 name = d_path(&file->f_path, buf, PATH_MAX);
3737 name = strncpy(tmp, "//toolong", sizeof(tmp));
3741 if (arch_vma_name(mmap_event->vma)) {
3742 name = strncpy(tmp, arch_vma_name(mmap_event->vma),
3748 name = strncpy(tmp, "[vdso]", sizeof(tmp));
3752 name = strncpy(tmp, "//anon", sizeof(tmp));
3757 size = ALIGN(strlen(name)+1, sizeof(u64));
3759 mmap_event->file_name = name;
3760 mmap_event->file_size = size;
3762 mmap_event->event_id.header.size = sizeof(mmap_event->event_id) + size;
3765 cpuctx = &get_cpu_var(perf_cpu_context);
3766 perf_event_mmap_ctx(&cpuctx->ctx, mmap_event);
3767 ctx = rcu_dereference(current->perf_event_ctxp);
3769 perf_event_mmap_ctx(ctx, mmap_event);
3770 put_cpu_var(perf_cpu_context);
3776 void __perf_event_mmap(struct vm_area_struct *vma)
3778 struct perf_mmap_event mmap_event;
3780 if (!atomic_read(&nr_mmap_events))
3783 mmap_event = (struct perf_mmap_event){
3789 .type = PERF_RECORD_MMAP,
3790 .misc = PERF_RECORD_MISC_USER,
3795 .start = vma->vm_start,
3796 .len = vma->vm_end - vma->vm_start,
3797 .pgoff = (u64)vma->vm_pgoff << PAGE_SHIFT,
3801 perf_event_mmap_event(&mmap_event);
3805 * IRQ throttle logging
3808 static void perf_log_throttle(struct perf_event *event, int enable)
3810 struct perf_output_handle handle;
3814 struct perf_event_header header;
3818 } throttle_event = {
3820 .type = PERF_RECORD_THROTTLE,
3822 .size = sizeof(throttle_event),
3824 .time = perf_clock(),
3825 .id = primary_event_id(event),
3826 .stream_id = event->id,
3830 throttle_event.header.type = PERF_RECORD_UNTHROTTLE;
3832 ret = perf_output_begin(&handle, event, sizeof(throttle_event), 1, 0);
3836 perf_output_put(&handle, throttle_event);
3837 perf_output_end(&handle);
3841 * Generic event overflow handling, sampling.
3844 static int __perf_event_overflow(struct perf_event *event, int nmi,
3845 int throttle, struct perf_sample_data *data,
3846 struct pt_regs *regs)
3848 int events = atomic_read(&event->event_limit);
3849 struct hw_perf_event *hwc = &event->hw;
3852 throttle = (throttle && event->pmu->unthrottle != NULL);
3857 if (hwc->interrupts != MAX_INTERRUPTS) {
3859 if (HZ * hwc->interrupts >
3860 (u64)sysctl_perf_event_sample_rate) {
3861 hwc->interrupts = MAX_INTERRUPTS;
3862 perf_log_throttle(event, 0);
3867 * Keep re-disabling events even though on the previous
3868 * pass we disabled it - just in case we raced with a
3869 * sched-in and the event got enabled again:
3875 if (event->attr.freq) {
3876 u64 now = perf_clock();
3877 s64 delta = now - hwc->freq_time_stamp;
3879 hwc->freq_time_stamp = now;
3881 if (delta > 0 && delta < 2*TICK_NSEC)
3882 perf_adjust_period(event, delta, hwc->last_period);
3886 * XXX event_limit might not quite work as expected on inherited
3890 event->pending_kill = POLL_IN;
3891 if (events && atomic_dec_and_test(&event->event_limit)) {
3893 event->pending_kill = POLL_HUP;
3895 event->pending_disable = 1;
3896 perf_pending_queue(&event->pending,
3897 perf_pending_event);
3899 perf_event_disable(event);
3902 if (event->overflow_handler)
3903 event->overflow_handler(event, nmi, data, regs);
3905 perf_event_output(event, nmi, data, regs);
3910 int perf_event_overflow(struct perf_event *event, int nmi,
3911 struct perf_sample_data *data,
3912 struct pt_regs *regs)
3914 return __perf_event_overflow(event, nmi, 1, data, regs);
3918 * Generic software event infrastructure
3922 * We directly increment event->count and keep a second value in
3923 * event->hw.period_left to count intervals. This period event
3924 * is kept in the range [-sample_period, 0] so that we can use the
3928 static u64 perf_swevent_set_period(struct perf_event *event)
3930 struct hw_perf_event *hwc = &event->hw;
3931 u64 period = hwc->last_period;
3935 hwc->last_period = hwc->sample_period;
3938 old = val = atomic64_read(&hwc->period_left);
3942 nr = div64_u64(period + val, period);
3943 offset = nr * period;
3945 if (atomic64_cmpxchg(&hwc->period_left, old, val) != old)
3951 static void perf_swevent_overflow(struct perf_event *event, u64 overflow,
3952 int nmi, struct perf_sample_data *data,
3953 struct pt_regs *regs)
3955 struct hw_perf_event *hwc = &event->hw;
3958 data->period = event->hw.last_period;
3960 overflow = perf_swevent_set_period(event);
3962 if (hwc->interrupts == MAX_INTERRUPTS)
3965 for (; overflow; overflow--) {
3966 if (__perf_event_overflow(event, nmi, throttle,
3969 * We inhibit the overflow from happening when
3970 * hwc->interrupts == MAX_INTERRUPTS.
3978 static void perf_swevent_unthrottle(struct perf_event *event)
3981 * Nothing to do, we already reset hwc->interrupts.
3985 static void perf_swevent_add(struct perf_event *event, u64 nr,
3986 int nmi, struct perf_sample_data *data,
3987 struct pt_regs *regs)
3989 struct hw_perf_event *hwc = &event->hw;
3991 atomic64_add(nr, &event->count);
3996 if (!hwc->sample_period)
3999 if (nr == 1 && hwc->sample_period == 1 && !event->attr.freq)
4000 return perf_swevent_overflow(event, 1, nmi, data, regs);
4002 if (atomic64_add_negative(nr, &hwc->period_left))
4005 perf_swevent_overflow(event, 0, nmi, data, regs);
4008 static int perf_exclude_event(struct perf_event *event,
4009 struct pt_regs *regs)
4012 if (event->attr.exclude_user && user_mode(regs))
4015 if (event->attr.exclude_kernel && !user_mode(regs))
4022 static int perf_swevent_match(struct perf_event *event,
4023 enum perf_type_id type,
4025 struct perf_sample_data *data,
4026 struct pt_regs *regs)
4028 if (event->attr.type != type)
4031 if (event->attr.config != event_id)
4034 if (perf_exclude_event(event, regs))
4040 static inline u64 swevent_hash(u64 type, u32 event_id)
4042 u64 val = event_id | (type << 32);
4044 return hash_64(val, SWEVENT_HLIST_BITS);
4047 static inline struct hlist_head *
4048 __find_swevent_head(struct swevent_hlist *hlist, u64 type, u32 event_id)
4050 u64 hash = swevent_hash(type, event_id);
4052 return &hlist->heads[hash];
4055 /* For the read side: events when they trigger */
4056 static inline struct hlist_head *
4057 find_swevent_head_rcu(struct perf_cpu_context *ctx, u64 type, u32 event_id)
4059 struct swevent_hlist *hlist;
4061 hlist = rcu_dereference(ctx->swevent_hlist);
4065 return __find_swevent_head(hlist, type, event_id);
4068 /* For the event head insertion and removal in the hlist */
4069 static inline struct hlist_head *
4070 find_swevent_head(struct perf_cpu_context *ctx, struct perf_event *event)
4072 struct swevent_hlist *hlist;
4073 u32 event_id = event->attr.config;
4074 u64 type = event->attr.type;
4077 * Event scheduling is always serialized against hlist allocation
4078 * and release. Which makes the protected version suitable here.
4079 * The context lock guarantees that.
4081 hlist = rcu_dereference_protected(ctx->swevent_hlist,
4082 lockdep_is_held(&event->ctx->lock));
4086 return __find_swevent_head(hlist, type, event_id);
4089 static void do_perf_sw_event(enum perf_type_id type, u32 event_id,
4091 struct perf_sample_data *data,
4092 struct pt_regs *regs)
4094 struct perf_cpu_context *cpuctx;
4095 struct perf_event *event;
4096 struct hlist_node *node;
4097 struct hlist_head *head;
4099 cpuctx = &__get_cpu_var(perf_cpu_context);
4103 head = find_swevent_head_rcu(cpuctx, type, event_id);
4108 hlist_for_each_entry_rcu(event, node, head, hlist_entry) {
4109 if (perf_swevent_match(event, type, event_id, data, regs))
4110 perf_swevent_add(event, nr, nmi, data, regs);
4116 int perf_swevent_get_recursion_context(void)
4118 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
4125 else if (in_softirq())
4130 if (cpuctx->recursion[rctx])
4133 cpuctx->recursion[rctx]++;
4138 EXPORT_SYMBOL_GPL(perf_swevent_get_recursion_context);
4140 void perf_swevent_put_recursion_context(int rctx)
4142 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
4144 cpuctx->recursion[rctx]--;
4146 EXPORT_SYMBOL_GPL(perf_swevent_put_recursion_context);
4149 void __perf_sw_event(u32 event_id, u64 nr, int nmi,
4150 struct pt_regs *regs, u64 addr)
4152 struct perf_sample_data data;
4155 preempt_disable_notrace();
4156 rctx = perf_swevent_get_recursion_context();
4160 perf_sample_data_init(&data, addr);
4162 do_perf_sw_event(PERF_TYPE_SOFTWARE, event_id, nr, nmi, &data, regs);
4164 perf_swevent_put_recursion_context(rctx);
4165 preempt_enable_notrace();
4168 static void perf_swevent_read(struct perf_event *event)
4172 static int perf_swevent_enable(struct perf_event *event)
4174 struct hw_perf_event *hwc = &event->hw;
4175 struct perf_cpu_context *cpuctx;
4176 struct hlist_head *head;
4178 cpuctx = &__get_cpu_var(perf_cpu_context);
4180 if (hwc->sample_period) {
4181 hwc->last_period = hwc->sample_period;
4182 perf_swevent_set_period(event);
4185 head = find_swevent_head(cpuctx, event);
4186 if (WARN_ON_ONCE(!head))
4189 hlist_add_head_rcu(&event->hlist_entry, head);
4194 static void perf_swevent_disable(struct perf_event *event)
4196 hlist_del_rcu(&event->hlist_entry);
4199 static const struct pmu perf_ops_generic = {
4200 .enable = perf_swevent_enable,
4201 .disable = perf_swevent_disable,
4202 .read = perf_swevent_read,
4203 .unthrottle = perf_swevent_unthrottle,
4207 * hrtimer based swevent callback
4210 static enum hrtimer_restart perf_swevent_hrtimer(struct hrtimer *hrtimer)
4212 enum hrtimer_restart ret = HRTIMER_RESTART;
4213 struct perf_sample_data data;
4214 struct pt_regs *regs;
4215 struct perf_event *event;
4218 event = container_of(hrtimer, struct perf_event, hw.hrtimer);
4219 event->pmu->read(event);
4221 perf_sample_data_init(&data, 0);
4222 data.period = event->hw.last_period;
4223 regs = get_irq_regs();
4225 if (regs && !perf_exclude_event(event, regs)) {
4226 if (!(event->attr.exclude_idle && current->pid == 0))
4227 if (perf_event_overflow(event, 0, &data, regs))
4228 ret = HRTIMER_NORESTART;
4231 period = max_t(u64, 10000, event->hw.sample_period);
4232 hrtimer_forward_now(hrtimer, ns_to_ktime(period));
4237 static void perf_swevent_start_hrtimer(struct perf_event *event)
4239 struct hw_perf_event *hwc = &event->hw;
4241 hrtimer_init(&hwc->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
4242 hwc->hrtimer.function = perf_swevent_hrtimer;
4243 if (hwc->sample_period) {
4246 if (hwc->remaining) {
4247 if (hwc->remaining < 0)
4250 period = hwc->remaining;
4253 period = max_t(u64, 10000, hwc->sample_period);
4255 __hrtimer_start_range_ns(&hwc->hrtimer,
4256 ns_to_ktime(period), 0,
4257 HRTIMER_MODE_REL, 0);
4261 static void perf_swevent_cancel_hrtimer(struct perf_event *event)
4263 struct hw_perf_event *hwc = &event->hw;
4265 if (hwc->sample_period) {
4266 ktime_t remaining = hrtimer_get_remaining(&hwc->hrtimer);
4267 hwc->remaining = ktime_to_ns(remaining);
4269 hrtimer_cancel(&hwc->hrtimer);
4274 * Software event: cpu wall time clock
4277 static void cpu_clock_perf_event_update(struct perf_event *event)
4279 int cpu = raw_smp_processor_id();
4283 now = cpu_clock(cpu);
4284 prev = atomic64_xchg(&event->hw.prev_count, now);
4285 atomic64_add(now - prev, &event->count);
4288 static int cpu_clock_perf_event_enable(struct perf_event *event)
4290 struct hw_perf_event *hwc = &event->hw;
4291 int cpu = raw_smp_processor_id();
4293 atomic64_set(&hwc->prev_count, cpu_clock(cpu));
4294 perf_swevent_start_hrtimer(event);
4299 static void cpu_clock_perf_event_disable(struct perf_event *event)
4301 perf_swevent_cancel_hrtimer(event);
4302 cpu_clock_perf_event_update(event);
4305 static void cpu_clock_perf_event_read(struct perf_event *event)
4307 cpu_clock_perf_event_update(event);
4310 static const struct pmu perf_ops_cpu_clock = {
4311 .enable = cpu_clock_perf_event_enable,
4312 .disable = cpu_clock_perf_event_disable,
4313 .read = cpu_clock_perf_event_read,
4317 * Software event: task time clock
4320 static void task_clock_perf_event_update(struct perf_event *event, u64 now)
4325 prev = atomic64_xchg(&event->hw.prev_count, now);
4327 atomic64_add(delta, &event->count);
4330 static int task_clock_perf_event_enable(struct perf_event *event)
4332 struct hw_perf_event *hwc = &event->hw;
4335 now = event->ctx->time;
4337 atomic64_set(&hwc->prev_count, now);
4339 perf_swevent_start_hrtimer(event);
4344 static void task_clock_perf_event_disable(struct perf_event *event)
4346 perf_swevent_cancel_hrtimer(event);
4347 task_clock_perf_event_update(event, event->ctx->time);
4351 static void task_clock_perf_event_read(struct perf_event *event)
4356 update_context_time(event->ctx);
4357 time = event->ctx->time;
4359 u64 now = perf_clock();
4360 u64 delta = now - event->ctx->timestamp;
4361 time = event->ctx->time + delta;
4364 task_clock_perf_event_update(event, time);
4367 static const struct pmu perf_ops_task_clock = {
4368 .enable = task_clock_perf_event_enable,
4369 .disable = task_clock_perf_event_disable,
4370 .read = task_clock_perf_event_read,
4373 /* Deref the hlist from the update side */
4374 static inline struct swevent_hlist *
4375 swevent_hlist_deref(struct perf_cpu_context *cpuctx)
4377 return rcu_dereference_protected(cpuctx->swevent_hlist,
4378 lockdep_is_held(&cpuctx->hlist_mutex));
4381 static void swevent_hlist_release_rcu(struct rcu_head *rcu_head)
4383 struct swevent_hlist *hlist;
4385 hlist = container_of(rcu_head, struct swevent_hlist, rcu_head);
4389 static void swevent_hlist_release(struct perf_cpu_context *cpuctx)
4391 struct swevent_hlist *hlist = swevent_hlist_deref(cpuctx);
4396 rcu_assign_pointer(cpuctx->swevent_hlist, NULL);
4397 call_rcu(&hlist->rcu_head, swevent_hlist_release_rcu);
4400 static void swevent_hlist_put_cpu(struct perf_event *event, int cpu)
4402 struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
4404 mutex_lock(&cpuctx->hlist_mutex);
4406 if (!--cpuctx->hlist_refcount)
4407 swevent_hlist_release(cpuctx);
4409 mutex_unlock(&cpuctx->hlist_mutex);
4412 static void swevent_hlist_put(struct perf_event *event)
4416 if (event->cpu != -1) {
4417 swevent_hlist_put_cpu(event, event->cpu);
4421 for_each_possible_cpu(cpu)
4422 swevent_hlist_put_cpu(event, cpu);
4425 static int swevent_hlist_get_cpu(struct perf_event *event, int cpu)
4427 struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
4430 mutex_lock(&cpuctx->hlist_mutex);
4432 if (!swevent_hlist_deref(cpuctx) && cpu_online(cpu)) {
4433 struct swevent_hlist *hlist;
4435 hlist = kzalloc(sizeof(*hlist), GFP_KERNEL);
4440 rcu_assign_pointer(cpuctx->swevent_hlist, hlist);
4442 cpuctx->hlist_refcount++;
4444 mutex_unlock(&cpuctx->hlist_mutex);
4449 static int swevent_hlist_get(struct perf_event *event)
4452 int cpu, failed_cpu;
4454 if (event->cpu != -1)
4455 return swevent_hlist_get_cpu(event, event->cpu);
4458 for_each_possible_cpu(cpu) {
4459 err = swevent_hlist_get_cpu(event, cpu);
4469 for_each_possible_cpu(cpu) {
4470 if (cpu == failed_cpu)
4472 swevent_hlist_put_cpu(event, cpu);
4479 #ifdef CONFIG_EVENT_TRACING
4481 static const struct pmu perf_ops_tracepoint = {
4482 .enable = perf_trace_enable,
4483 .disable = perf_trace_disable,
4484 .read = perf_swevent_read,
4485 .unthrottle = perf_swevent_unthrottle,
4488 static int perf_tp_filter_match(struct perf_event *event,
4489 struct perf_sample_data *data)
4491 void *record = data->raw->data;
4493 if (likely(!event->filter) || filter_match_preds(event->filter, record))
4498 static int perf_tp_event_match(struct perf_event *event,
4499 struct perf_sample_data *data,
4500 struct pt_regs *regs)
4502 if (perf_exclude_event(event, regs))
4505 if (!perf_tp_filter_match(event, data))
4511 void perf_tp_event(u64 addr, u64 count, void *record, int entry_size,
4512 struct pt_regs *regs, struct hlist_head *head)
4514 struct perf_sample_data data;
4515 struct perf_event *event;
4516 struct hlist_node *node;
4518 struct perf_raw_record raw = {
4523 perf_sample_data_init(&data, addr);
4527 hlist_for_each_entry_rcu(event, node, head, hlist_entry) {
4528 if (perf_tp_event_match(event, &data, regs))
4529 perf_swevent_add(event, count, 1, &data, regs);
4533 EXPORT_SYMBOL_GPL(perf_tp_event);
4535 static void tp_perf_event_destroy(struct perf_event *event)
4537 perf_trace_destroy(event);
4540 static const struct pmu *tp_perf_event_init(struct perf_event *event)
4545 * Raw tracepoint data is a severe data leak, only allow root to
4548 if ((event->attr.sample_type & PERF_SAMPLE_RAW) &&
4549 perf_paranoid_tracepoint_raw() &&
4550 !capable(CAP_SYS_ADMIN))
4551 return ERR_PTR(-EPERM);
4553 err = perf_trace_init(event);
4557 event->destroy = tp_perf_event_destroy;
4559 return &perf_ops_tracepoint;
4562 static int perf_event_set_filter(struct perf_event *event, void __user *arg)
4567 if (event->attr.type != PERF_TYPE_TRACEPOINT)
4570 filter_str = strndup_user(arg, PAGE_SIZE);
4571 if (IS_ERR(filter_str))
4572 return PTR_ERR(filter_str);
4574 ret = ftrace_profile_set_filter(event, event->attr.config, filter_str);
4580 static void perf_event_free_filter(struct perf_event *event)
4582 ftrace_profile_free_filter(event);
4587 static const struct pmu *tp_perf_event_init(struct perf_event *event)
4592 static int perf_event_set_filter(struct perf_event *event, void __user *arg)
4597 static void perf_event_free_filter(struct perf_event *event)
4601 #endif /* CONFIG_EVENT_TRACING */
4603 #ifdef CONFIG_HAVE_HW_BREAKPOINT
4604 static void bp_perf_event_destroy(struct perf_event *event)
4606 release_bp_slot(event);
4609 static const struct pmu *bp_perf_event_init(struct perf_event *bp)
4613 err = register_perf_hw_breakpoint(bp);
4615 return ERR_PTR(err);
4617 bp->destroy = bp_perf_event_destroy;
4619 return &perf_ops_bp;
4622 void perf_bp_event(struct perf_event *bp, void *data)
4624 struct perf_sample_data sample;
4625 struct pt_regs *regs = data;
4627 perf_sample_data_init(&sample, bp->attr.bp_addr);
4629 if (!perf_exclude_event(bp, regs))
4630 perf_swevent_add(bp, 1, 1, &sample, regs);
4633 static const struct pmu *bp_perf_event_init(struct perf_event *bp)
4638 void perf_bp_event(struct perf_event *bp, void *regs)
4643 atomic_t perf_swevent_enabled[PERF_COUNT_SW_MAX];
4645 static void sw_perf_event_destroy(struct perf_event *event)
4647 u64 event_id = event->attr.config;
4649 WARN_ON(event->parent);
4651 atomic_dec(&perf_swevent_enabled[event_id]);
4652 swevent_hlist_put(event);
4655 static const struct pmu *sw_perf_event_init(struct perf_event *event)
4657 const struct pmu *pmu = NULL;
4658 u64 event_id = event->attr.config;
4661 * Software events (currently) can't in general distinguish
4662 * between user, kernel and hypervisor events.
4663 * However, context switches and cpu migrations are considered
4664 * to be kernel events, and page faults are never hypervisor
4668 case PERF_COUNT_SW_CPU_CLOCK:
4669 pmu = &perf_ops_cpu_clock;
4672 case PERF_COUNT_SW_TASK_CLOCK:
4674 * If the user instantiates this as a per-cpu event,
4675 * use the cpu_clock event instead.
4677 if (event->ctx->task)
4678 pmu = &perf_ops_task_clock;
4680 pmu = &perf_ops_cpu_clock;
4683 case PERF_COUNT_SW_PAGE_FAULTS:
4684 case PERF_COUNT_SW_PAGE_FAULTS_MIN:
4685 case PERF_COUNT_SW_PAGE_FAULTS_MAJ:
4686 case PERF_COUNT_SW_CONTEXT_SWITCHES:
4687 case PERF_COUNT_SW_CPU_MIGRATIONS:
4688 case PERF_COUNT_SW_ALIGNMENT_FAULTS:
4689 case PERF_COUNT_SW_EMULATION_FAULTS:
4690 if (!event->parent) {
4693 err = swevent_hlist_get(event);
4695 return ERR_PTR(err);
4697 atomic_inc(&perf_swevent_enabled[event_id]);
4698 event->destroy = sw_perf_event_destroy;
4700 pmu = &perf_ops_generic;
4708 * Allocate and initialize a event structure
4710 static struct perf_event *
4711 perf_event_alloc(struct perf_event_attr *attr,
4713 struct perf_event_context *ctx,
4714 struct perf_event *group_leader,
4715 struct perf_event *parent_event,
4716 perf_overflow_handler_t overflow_handler,
4719 const struct pmu *pmu;
4720 struct perf_event *event;
4721 struct hw_perf_event *hwc;
4724 event = kzalloc(sizeof(*event), gfpflags);
4726 return ERR_PTR(-ENOMEM);
4729 * Single events are their own group leaders, with an
4730 * empty sibling list:
4733 group_leader = event;
4735 mutex_init(&event->child_mutex);
4736 INIT_LIST_HEAD(&event->child_list);
4738 INIT_LIST_HEAD(&event->group_entry);
4739 INIT_LIST_HEAD(&event->event_entry);
4740 INIT_LIST_HEAD(&event->sibling_list);
4741 init_waitqueue_head(&event->waitq);
4743 mutex_init(&event->mmap_mutex);
4746 event->attr = *attr;
4747 event->group_leader = group_leader;
4752 event->parent = parent_event;
4754 event->ns = get_pid_ns(current->nsproxy->pid_ns);
4755 event->id = atomic64_inc_return(&perf_event_id);
4757 event->state = PERF_EVENT_STATE_INACTIVE;
4759 if (!overflow_handler && parent_event)
4760 overflow_handler = parent_event->overflow_handler;
4762 event->overflow_handler = overflow_handler;
4765 event->state = PERF_EVENT_STATE_OFF;
4770 hwc->sample_period = attr->sample_period;
4771 if (attr->freq && attr->sample_freq)
4772 hwc->sample_period = 1;
4773 hwc->last_period = hwc->sample_period;
4775 atomic64_set(&hwc->period_left, hwc->sample_period);
4778 * we currently do not support PERF_FORMAT_GROUP on inherited events
4780 if (attr->inherit && (attr->read_format & PERF_FORMAT_GROUP))
4783 switch (attr->type) {
4785 case PERF_TYPE_HARDWARE:
4786 case PERF_TYPE_HW_CACHE:
4787 pmu = hw_perf_event_init(event);
4790 case PERF_TYPE_SOFTWARE:
4791 pmu = sw_perf_event_init(event);
4794 case PERF_TYPE_TRACEPOINT:
4795 pmu = tp_perf_event_init(event);
4798 case PERF_TYPE_BREAKPOINT:
4799 pmu = bp_perf_event_init(event);
4810 else if (IS_ERR(pmu))
4815 put_pid_ns(event->ns);
4817 return ERR_PTR(err);
4822 if (!event->parent) {
4823 atomic_inc(&nr_events);
4824 if (event->attr.mmap)
4825 atomic_inc(&nr_mmap_events);
4826 if (event->attr.comm)
4827 atomic_inc(&nr_comm_events);
4828 if (event->attr.task)
4829 atomic_inc(&nr_task_events);
4835 static int perf_copy_attr(struct perf_event_attr __user *uattr,
4836 struct perf_event_attr *attr)
4841 if (!access_ok(VERIFY_WRITE, uattr, PERF_ATTR_SIZE_VER0))
4845 * zero the full structure, so that a short copy will be nice.
4847 memset(attr, 0, sizeof(*attr));
4849 ret = get_user(size, &uattr->size);
4853 if (size > PAGE_SIZE) /* silly large */
4856 if (!size) /* abi compat */
4857 size = PERF_ATTR_SIZE_VER0;
4859 if (size < PERF_ATTR_SIZE_VER0)
4863 * If we're handed a bigger struct than we know of,
4864 * ensure all the unknown bits are 0 - i.e. new
4865 * user-space does not rely on any kernel feature
4866 * extensions we dont know about yet.
4868 if (size > sizeof(*attr)) {
4869 unsigned char __user *addr;
4870 unsigned char __user *end;
4873 addr = (void __user *)uattr + sizeof(*attr);
4874 end = (void __user *)uattr + size;
4876 for (; addr < end; addr++) {
4877 ret = get_user(val, addr);
4883 size = sizeof(*attr);
4886 ret = copy_from_user(attr, uattr, size);
4891 * If the type exists, the corresponding creation will verify
4894 if (attr->type >= PERF_TYPE_MAX)
4897 if (attr->__reserved_1)
4900 if (attr->sample_type & ~(PERF_SAMPLE_MAX-1))
4903 if (attr->read_format & ~(PERF_FORMAT_MAX-1))
4910 put_user(sizeof(*attr), &uattr->size);
4915 static int perf_event_set_output(struct perf_event *event, int output_fd)
4917 struct perf_event *output_event = NULL;
4918 struct file *output_file = NULL;
4919 struct perf_event *old_output;
4920 int fput_needed = 0;
4924 * Don't allow output of inherited per-task events. This would
4925 * create performance issues due to cross cpu access.
4927 if (event->cpu == -1 && event->attr.inherit)
4933 output_file = fget_light(output_fd, &fput_needed);
4937 if (output_file->f_op != &perf_fops)
4940 output_event = output_file->private_data;
4942 /* Don't chain output fds */
4943 if (output_event->output)
4946 /* Don't set an output fd when we already have an output channel */
4951 * Don't allow cross-cpu buffers
4953 if (output_event->cpu != event->cpu)
4957 * If its not a per-cpu buffer, it must be the same task.
4959 if (output_event->cpu == -1 && output_event->ctx != event->ctx)
4962 atomic_long_inc(&output_file->f_count);
4965 mutex_lock(&event->mmap_mutex);
4966 old_output = event->output;
4967 rcu_assign_pointer(event->output, output_event);
4968 mutex_unlock(&event->mmap_mutex);
4972 * we need to make sure no existing perf_output_*()
4973 * is still referencing this event.
4976 fput(old_output->filp);
4981 fput_light(output_file, fput_needed);
4986 * sys_perf_event_open - open a performance event, associate it to a task/cpu
4988 * @attr_uptr: event_id type attributes for monitoring/sampling
4991 * @group_fd: group leader event fd
4993 SYSCALL_DEFINE5(perf_event_open,
4994 struct perf_event_attr __user *, attr_uptr,
4995 pid_t, pid, int, cpu, int, group_fd, unsigned long, flags)
4997 struct perf_event *event, *group_leader;
4998 struct perf_event_attr attr;
4999 struct perf_event_context *ctx;
5000 struct file *event_file = NULL;
5001 struct file *group_file = NULL;
5002 int fput_needed = 0;
5003 int fput_needed2 = 0;
5006 /* for future expandability... */
5007 if (flags & ~(PERF_FLAG_FD_NO_GROUP | PERF_FLAG_FD_OUTPUT))
5010 err = perf_copy_attr(attr_uptr, &attr);
5014 if (!attr.exclude_kernel) {
5015 if (perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN))
5020 if (attr.sample_freq > sysctl_perf_event_sample_rate)
5025 * Get the target context (task or percpu):
5027 ctx = find_get_context(pid, cpu);
5029 return PTR_ERR(ctx);
5032 * Look up the group leader (we will attach this event to it):
5034 group_leader = NULL;
5035 if (group_fd != -1 && !(flags & PERF_FLAG_FD_NO_GROUP)) {
5037 group_file = fget_light(group_fd, &fput_needed);
5039 goto err_put_context;
5040 if (group_file->f_op != &perf_fops)
5041 goto err_put_context;
5043 group_leader = group_file->private_data;
5045 * Do not allow a recursive hierarchy (this new sibling
5046 * becoming part of another group-sibling):
5048 if (group_leader->group_leader != group_leader)
5049 goto err_put_context;
5051 * Do not allow to attach to a group in a different
5052 * task or CPU context:
5054 if (group_leader->ctx != ctx)
5055 goto err_put_context;
5057 * Only a group leader can be exclusive or pinned
5059 if (attr.exclusive || attr.pinned)
5060 goto err_put_context;
5063 event = perf_event_alloc(&attr, cpu, ctx, group_leader,
5064 NULL, NULL, GFP_KERNEL);
5065 err = PTR_ERR(event);
5067 goto err_put_context;
5069 err = anon_inode_getfd("[perf_event]", &perf_fops, event, O_RDWR);
5071 goto err_free_put_context;
5073 event_file = fget_light(err, &fput_needed2);
5075 goto err_free_put_context;
5077 if (flags & PERF_FLAG_FD_OUTPUT) {
5078 err = perf_event_set_output(event, group_fd);
5080 goto err_fput_free_put_context;
5083 event->filp = event_file;
5084 WARN_ON_ONCE(ctx->parent_ctx);
5085 mutex_lock(&ctx->mutex);
5086 perf_install_in_context(ctx, event, cpu);
5088 mutex_unlock(&ctx->mutex);
5090 event->owner = current;
5091 get_task_struct(current);
5092 mutex_lock(¤t->perf_event_mutex);
5093 list_add_tail(&event->owner_entry, ¤t->perf_event_list);
5094 mutex_unlock(¤t->perf_event_mutex);
5096 err_fput_free_put_context:
5097 fput_light(event_file, fput_needed2);
5099 err_free_put_context:
5107 fput_light(group_file, fput_needed);
5113 * perf_event_create_kernel_counter
5115 * @attr: attributes of the counter to create
5116 * @cpu: cpu in which the counter is bound
5117 * @pid: task to profile
5120 perf_event_create_kernel_counter(struct perf_event_attr *attr, int cpu,
5122 perf_overflow_handler_t overflow_handler)
5124 struct perf_event *event;
5125 struct perf_event_context *ctx;
5129 * Get the target context (task or percpu):
5132 ctx = find_get_context(pid, cpu);
5138 event = perf_event_alloc(attr, cpu, ctx, NULL,
5139 NULL, overflow_handler, GFP_KERNEL);
5140 if (IS_ERR(event)) {
5141 err = PTR_ERR(event);
5142 goto err_put_context;
5146 WARN_ON_ONCE(ctx->parent_ctx);
5147 mutex_lock(&ctx->mutex);
5148 perf_install_in_context(ctx, event, cpu);
5150 mutex_unlock(&ctx->mutex);
5152 event->owner = current;
5153 get_task_struct(current);
5154 mutex_lock(¤t->perf_event_mutex);
5155 list_add_tail(&event->owner_entry, ¤t->perf_event_list);
5156 mutex_unlock(¤t->perf_event_mutex);
5163 return ERR_PTR(err);
5165 EXPORT_SYMBOL_GPL(perf_event_create_kernel_counter);
5168 * inherit a event from parent task to child task:
5170 static struct perf_event *
5171 inherit_event(struct perf_event *parent_event,
5172 struct task_struct *parent,
5173 struct perf_event_context *parent_ctx,
5174 struct task_struct *child,
5175 struct perf_event *group_leader,
5176 struct perf_event_context *child_ctx)
5178 struct perf_event *child_event;
5181 * Instead of creating recursive hierarchies of events,
5182 * we link inherited events back to the original parent,
5183 * which has a filp for sure, which we use as the reference
5186 if (parent_event->parent)
5187 parent_event = parent_event->parent;
5189 child_event = perf_event_alloc(&parent_event->attr,
5190 parent_event->cpu, child_ctx,
5191 group_leader, parent_event,
5193 if (IS_ERR(child_event))
5198 * Make the child state follow the state of the parent event,
5199 * not its attr.disabled bit. We hold the parent's mutex,
5200 * so we won't race with perf_event_{en, dis}able_family.
5202 if (parent_event->state >= PERF_EVENT_STATE_INACTIVE)
5203 child_event->state = PERF_EVENT_STATE_INACTIVE;
5205 child_event->state = PERF_EVENT_STATE_OFF;
5207 if (parent_event->attr.freq) {
5208 u64 sample_period = parent_event->hw.sample_period;
5209 struct hw_perf_event *hwc = &child_event->hw;
5211 hwc->sample_period = sample_period;
5212 hwc->last_period = sample_period;
5214 atomic64_set(&hwc->period_left, sample_period);
5217 child_event->overflow_handler = parent_event->overflow_handler;
5220 * Link it up in the child's context:
5222 add_event_to_ctx(child_event, child_ctx);
5225 * Get a reference to the parent filp - we will fput it
5226 * when the child event exits. This is safe to do because
5227 * we are in the parent and we know that the filp still
5228 * exists and has a nonzero count:
5230 atomic_long_inc(&parent_event->filp->f_count);
5233 * Link this into the parent event's child list
5235 WARN_ON_ONCE(parent_event->ctx->parent_ctx);
5236 mutex_lock(&parent_event->child_mutex);
5237 list_add_tail(&child_event->child_list, &parent_event->child_list);
5238 mutex_unlock(&parent_event->child_mutex);
5243 static int inherit_group(struct perf_event *parent_event,
5244 struct task_struct *parent,
5245 struct perf_event_context *parent_ctx,
5246 struct task_struct *child,
5247 struct perf_event_context *child_ctx)
5249 struct perf_event *leader;
5250 struct perf_event *sub;
5251 struct perf_event *child_ctr;
5253 leader = inherit_event(parent_event, parent, parent_ctx,
5254 child, NULL, child_ctx);
5256 return PTR_ERR(leader);
5257 list_for_each_entry(sub, &parent_event->sibling_list, group_entry) {
5258 child_ctr = inherit_event(sub, parent, parent_ctx,
5259 child, leader, child_ctx);
5260 if (IS_ERR(child_ctr))
5261 return PTR_ERR(child_ctr);
5266 static void sync_child_event(struct perf_event *child_event,
5267 struct task_struct *child)
5269 struct perf_event *parent_event = child_event->parent;
5272 if (child_event->attr.inherit_stat)
5273 perf_event_read_event(child_event, child);
5275 child_val = atomic64_read(&child_event->count);
5278 * Add back the child's count to the parent's count:
5280 atomic64_add(child_val, &parent_event->count);
5281 atomic64_add(child_event->total_time_enabled,
5282 &parent_event->child_total_time_enabled);
5283 atomic64_add(child_event->total_time_running,
5284 &parent_event->child_total_time_running);
5287 * Remove this event from the parent's list
5289 WARN_ON_ONCE(parent_event->ctx->parent_ctx);
5290 mutex_lock(&parent_event->child_mutex);
5291 list_del_init(&child_event->child_list);
5292 mutex_unlock(&parent_event->child_mutex);
5295 * Release the parent event, if this was the last
5298 fput(parent_event->filp);
5302 __perf_event_exit_task(struct perf_event *child_event,
5303 struct perf_event_context *child_ctx,
5304 struct task_struct *child)
5306 struct perf_event *parent_event;
5308 perf_event_remove_from_context(child_event);
5310 parent_event = child_event->parent;
5312 * It can happen that parent exits first, and has events
5313 * that are still around due to the child reference. These
5314 * events need to be zapped - but otherwise linger.
5317 sync_child_event(child_event, child);
5318 free_event(child_event);
5323 * When a child task exits, feed back event values to parent events.
5325 void perf_event_exit_task(struct task_struct *child)
5327 struct perf_event *child_event, *tmp;
5328 struct perf_event_context *child_ctx;
5329 unsigned long flags;
5331 if (likely(!child->perf_event_ctxp)) {
5332 perf_event_task(child, NULL, 0);
5336 local_irq_save(flags);
5338 * We can't reschedule here because interrupts are disabled,
5339 * and either child is current or it is a task that can't be
5340 * scheduled, so we are now safe from rescheduling changing
5343 child_ctx = child->perf_event_ctxp;
5344 __perf_event_task_sched_out(child_ctx);
5347 * Take the context lock here so that if find_get_context is
5348 * reading child->perf_event_ctxp, we wait until it has
5349 * incremented the context's refcount before we do put_ctx below.
5351 raw_spin_lock(&child_ctx->lock);
5352 child->perf_event_ctxp = NULL;
5354 * If this context is a clone; unclone it so it can't get
5355 * swapped to another process while we're removing all
5356 * the events from it.
5358 unclone_ctx(child_ctx);
5359 update_context_time(child_ctx);
5360 raw_spin_unlock_irqrestore(&child_ctx->lock, flags);
5363 * Report the task dead after unscheduling the events so that we
5364 * won't get any samples after PERF_RECORD_EXIT. We can however still
5365 * get a few PERF_RECORD_READ events.
5367 perf_event_task(child, child_ctx, 0);
5370 * We can recurse on the same lock type through:
5372 * __perf_event_exit_task()
5373 * sync_child_event()
5374 * fput(parent_event->filp)
5376 * mutex_lock(&ctx->mutex)
5378 * But since its the parent context it won't be the same instance.
5380 mutex_lock(&child_ctx->mutex);
5383 list_for_each_entry_safe(child_event, tmp, &child_ctx->pinned_groups,
5385 __perf_event_exit_task(child_event, child_ctx, child);
5387 list_for_each_entry_safe(child_event, tmp, &child_ctx->flexible_groups,
5389 __perf_event_exit_task(child_event, child_ctx, child);
5392 * If the last event was a group event, it will have appended all
5393 * its siblings to the list, but we obtained 'tmp' before that which
5394 * will still point to the list head terminating the iteration.
5396 if (!list_empty(&child_ctx->pinned_groups) ||
5397 !list_empty(&child_ctx->flexible_groups))
5400 mutex_unlock(&child_ctx->mutex);
5405 static void perf_free_event(struct perf_event *event,
5406 struct perf_event_context *ctx)
5408 struct perf_event *parent = event->parent;
5410 if (WARN_ON_ONCE(!parent))
5413 mutex_lock(&parent->child_mutex);
5414 list_del_init(&event->child_list);
5415 mutex_unlock(&parent->child_mutex);
5419 list_del_event(event, ctx);
5424 * free an unexposed, unused context as created by inheritance by
5425 * init_task below, used by fork() in case of fail.
5427 void perf_event_free_task(struct task_struct *task)
5429 struct perf_event_context *ctx = task->perf_event_ctxp;
5430 struct perf_event *event, *tmp;
5435 mutex_lock(&ctx->mutex);
5437 list_for_each_entry_safe(event, tmp, &ctx->pinned_groups, group_entry)
5438 perf_free_event(event, ctx);
5440 list_for_each_entry_safe(event, tmp, &ctx->flexible_groups,
5442 perf_free_event(event, ctx);
5444 if (!list_empty(&ctx->pinned_groups) ||
5445 !list_empty(&ctx->flexible_groups))
5448 mutex_unlock(&ctx->mutex);
5454 inherit_task_group(struct perf_event *event, struct task_struct *parent,
5455 struct perf_event_context *parent_ctx,
5456 struct task_struct *child,
5460 struct perf_event_context *child_ctx = child->perf_event_ctxp;
5462 if (!event->attr.inherit) {
5469 * This is executed from the parent task context, so
5470 * inherit events that have been marked for cloning.
5471 * First allocate and initialize a context for the
5475 child_ctx = kzalloc(sizeof(struct perf_event_context),
5480 __perf_event_init_context(child_ctx, child);
5481 child->perf_event_ctxp = child_ctx;
5482 get_task_struct(child);
5485 ret = inherit_group(event, parent, parent_ctx,
5496 * Initialize the perf_event context in task_struct
5498 int perf_event_init_task(struct task_struct *child)
5500 struct perf_event_context *child_ctx, *parent_ctx;
5501 struct perf_event_context *cloned_ctx;
5502 struct perf_event *event;
5503 struct task_struct *parent = current;
5504 int inherited_all = 1;
5507 child->perf_event_ctxp = NULL;
5509 mutex_init(&child->perf_event_mutex);
5510 INIT_LIST_HEAD(&child->perf_event_list);
5512 if (likely(!parent->perf_event_ctxp))
5516 * If the parent's context is a clone, pin it so it won't get
5519 parent_ctx = perf_pin_task_context(parent);
5522 * No need to check if parent_ctx != NULL here; since we saw
5523 * it non-NULL earlier, the only reason for it to become NULL
5524 * is if we exit, and since we're currently in the middle of
5525 * a fork we can't be exiting at the same time.
5529 * Lock the parent list. No need to lock the child - not PID
5530 * hashed yet and not running, so nobody can access it.
5532 mutex_lock(&parent_ctx->mutex);
5535 * We dont have to disable NMIs - we are only looking at
5536 * the list, not manipulating it:
5538 list_for_each_entry(event, &parent_ctx->pinned_groups, group_entry) {
5539 ret = inherit_task_group(event, parent, parent_ctx, child,
5545 list_for_each_entry(event, &parent_ctx->flexible_groups, group_entry) {
5546 ret = inherit_task_group(event, parent, parent_ctx, child,
5552 child_ctx = child->perf_event_ctxp;
5554 if (child_ctx && inherited_all) {
5556 * Mark the child context as a clone of the parent
5557 * context, or of whatever the parent is a clone of.
5558 * Note that if the parent is a clone, it could get
5559 * uncloned at any point, but that doesn't matter
5560 * because the list of events and the generation
5561 * count can't have changed since we took the mutex.
5563 cloned_ctx = rcu_dereference(parent_ctx->parent_ctx);
5565 child_ctx->parent_ctx = cloned_ctx;
5566 child_ctx->parent_gen = parent_ctx->parent_gen;
5568 child_ctx->parent_ctx = parent_ctx;
5569 child_ctx->parent_gen = parent_ctx->generation;
5571 get_ctx(child_ctx->parent_ctx);
5574 mutex_unlock(&parent_ctx->mutex);
5576 perf_unpin_context(parent_ctx);
5581 static void __init perf_event_init_all_cpus(void)
5584 struct perf_cpu_context *cpuctx;
5586 for_each_possible_cpu(cpu) {
5587 cpuctx = &per_cpu(perf_cpu_context, cpu);
5588 mutex_init(&cpuctx->hlist_mutex);
5589 __perf_event_init_context(&cpuctx->ctx, NULL);
5593 static void __cpuinit perf_event_init_cpu(int cpu)
5595 struct perf_cpu_context *cpuctx;
5597 cpuctx = &per_cpu(perf_cpu_context, cpu);
5599 spin_lock(&perf_resource_lock);
5600 cpuctx->max_pertask = perf_max_events - perf_reserved_percpu;
5601 spin_unlock(&perf_resource_lock);
5603 mutex_lock(&cpuctx->hlist_mutex);
5604 if (cpuctx->hlist_refcount > 0) {
5605 struct swevent_hlist *hlist;
5607 hlist = kzalloc(sizeof(*hlist), GFP_KERNEL);
5608 WARN_ON_ONCE(!hlist);
5609 rcu_assign_pointer(cpuctx->swevent_hlist, hlist);
5611 mutex_unlock(&cpuctx->hlist_mutex);
5614 #ifdef CONFIG_HOTPLUG_CPU
5615 static void __perf_event_exit_cpu(void *info)
5617 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
5618 struct perf_event_context *ctx = &cpuctx->ctx;
5619 struct perf_event *event, *tmp;
5621 list_for_each_entry_safe(event, tmp, &ctx->pinned_groups, group_entry)
5622 __perf_event_remove_from_context(event);
5623 list_for_each_entry_safe(event, tmp, &ctx->flexible_groups, group_entry)
5624 __perf_event_remove_from_context(event);
5626 static void perf_event_exit_cpu(int cpu)
5628 struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
5629 struct perf_event_context *ctx = &cpuctx->ctx;
5631 mutex_lock(&cpuctx->hlist_mutex);
5632 swevent_hlist_release(cpuctx);
5633 mutex_unlock(&cpuctx->hlist_mutex);
5635 mutex_lock(&ctx->mutex);
5636 smp_call_function_single(cpu, __perf_event_exit_cpu, NULL, 1);
5637 mutex_unlock(&ctx->mutex);
5640 static inline void perf_event_exit_cpu(int cpu) { }
5643 static int __cpuinit
5644 perf_cpu_notify(struct notifier_block *self, unsigned long action, void *hcpu)
5646 unsigned int cpu = (long)hcpu;
5650 case CPU_UP_PREPARE:
5651 case CPU_UP_PREPARE_FROZEN:
5652 perf_event_init_cpu(cpu);
5655 case CPU_DOWN_PREPARE:
5656 case CPU_DOWN_PREPARE_FROZEN:
5657 perf_event_exit_cpu(cpu);
5668 * This has to have a higher priority than migration_notifier in sched.c.
5670 static struct notifier_block __cpuinitdata perf_cpu_nb = {
5671 .notifier_call = perf_cpu_notify,
5675 void __init perf_event_init(void)
5677 perf_event_init_all_cpus();
5678 perf_cpu_notify(&perf_cpu_nb, (unsigned long)CPU_UP_PREPARE,
5679 (void *)(long)smp_processor_id());
5680 perf_cpu_notify(&perf_cpu_nb, (unsigned long)CPU_ONLINE,
5681 (void *)(long)smp_processor_id());
5682 register_cpu_notifier(&perf_cpu_nb);
5685 static ssize_t perf_show_reserve_percpu(struct sysdev_class *class,
5686 struct sysdev_class_attribute *attr,
5689 return sprintf(buf, "%d\n", perf_reserved_percpu);
5693 perf_set_reserve_percpu(struct sysdev_class *class,
5694 struct sysdev_class_attribute *attr,
5698 struct perf_cpu_context *cpuctx;
5702 err = strict_strtoul(buf, 10, &val);
5705 if (val > perf_max_events)
5708 spin_lock(&perf_resource_lock);
5709 perf_reserved_percpu = val;
5710 for_each_online_cpu(cpu) {
5711 cpuctx = &per_cpu(perf_cpu_context, cpu);
5712 raw_spin_lock_irq(&cpuctx->ctx.lock);
5713 mpt = min(perf_max_events - cpuctx->ctx.nr_events,
5714 perf_max_events - perf_reserved_percpu);
5715 cpuctx->max_pertask = mpt;
5716 raw_spin_unlock_irq(&cpuctx->ctx.lock);
5718 spin_unlock(&perf_resource_lock);
5723 static ssize_t perf_show_overcommit(struct sysdev_class *class,
5724 struct sysdev_class_attribute *attr,
5727 return sprintf(buf, "%d\n", perf_overcommit);
5731 perf_set_overcommit(struct sysdev_class *class,
5732 struct sysdev_class_attribute *attr,
5733 const char *buf, size_t count)
5738 err = strict_strtoul(buf, 10, &val);
5744 spin_lock(&perf_resource_lock);
5745 perf_overcommit = val;
5746 spin_unlock(&perf_resource_lock);
5751 static SYSDEV_CLASS_ATTR(
5754 perf_show_reserve_percpu,
5755 perf_set_reserve_percpu
5758 static SYSDEV_CLASS_ATTR(
5761 perf_show_overcommit,
5765 static struct attribute *perfclass_attrs[] = {
5766 &attr_reserve_percpu.attr,
5767 &attr_overcommit.attr,
5771 static struct attribute_group perfclass_attr_group = {
5772 .attrs = perfclass_attrs,
5773 .name = "perf_events",
5776 static int __init perf_event_sysfs_init(void)
5778 return sysfs_create_group(&cpu_sysdev_class.kset.kobj,
5779 &perfclass_attr_group);
5781 device_initcall(perf_event_sysfs_init);