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 extern __weak const char *perf_pmu_name(void)
93 static DEFINE_PER_CPU(int, perf_disable_count);
95 void perf_disable(void)
97 if (!__get_cpu_var(perf_disable_count)++)
101 void perf_enable(void)
103 if (!--__get_cpu_var(perf_disable_count))
107 static void get_ctx(struct perf_event_context *ctx)
109 WARN_ON(!atomic_inc_not_zero(&ctx->refcount));
112 static void free_ctx(struct rcu_head *head)
114 struct perf_event_context *ctx;
116 ctx = container_of(head, struct perf_event_context, rcu_head);
120 static void put_ctx(struct perf_event_context *ctx)
122 if (atomic_dec_and_test(&ctx->refcount)) {
124 put_ctx(ctx->parent_ctx);
126 put_task_struct(ctx->task);
127 call_rcu(&ctx->rcu_head, free_ctx);
131 static void unclone_ctx(struct perf_event_context *ctx)
133 if (ctx->parent_ctx) {
134 put_ctx(ctx->parent_ctx);
135 ctx->parent_ctx = NULL;
140 * If we inherit events we want to return the parent event id
143 static u64 primary_event_id(struct perf_event *event)
148 id = event->parent->id;
154 * Get the perf_event_context for a task and lock it.
155 * This has to cope with with the fact that until it is locked,
156 * the context could get moved to another task.
158 static struct perf_event_context *
159 perf_lock_task_context(struct task_struct *task, unsigned long *flags)
161 struct perf_event_context *ctx;
165 ctx = rcu_dereference(task->perf_event_ctxp);
168 * If this context is a clone of another, it might
169 * get swapped for another underneath us by
170 * perf_event_task_sched_out, though the
171 * rcu_read_lock() protects us from any context
172 * getting freed. Lock the context and check if it
173 * got swapped before we could get the lock, and retry
174 * if so. If we locked the right context, then it
175 * can't get swapped on us any more.
177 raw_spin_lock_irqsave(&ctx->lock, *flags);
178 if (ctx != rcu_dereference(task->perf_event_ctxp)) {
179 raw_spin_unlock_irqrestore(&ctx->lock, *flags);
183 if (!atomic_inc_not_zero(&ctx->refcount)) {
184 raw_spin_unlock_irqrestore(&ctx->lock, *flags);
193 * Get the context for a task and increment its pin_count so it
194 * can't get swapped to another task. This also increments its
195 * reference count so that the context can't get freed.
197 static struct perf_event_context *perf_pin_task_context(struct task_struct *task)
199 struct perf_event_context *ctx;
202 ctx = perf_lock_task_context(task, &flags);
205 raw_spin_unlock_irqrestore(&ctx->lock, flags);
210 static void perf_unpin_context(struct perf_event_context *ctx)
214 raw_spin_lock_irqsave(&ctx->lock, flags);
216 raw_spin_unlock_irqrestore(&ctx->lock, flags);
220 static inline u64 perf_clock(void)
222 return local_clock();
226 * Update the record of the current time in a context.
228 static void update_context_time(struct perf_event_context *ctx)
230 u64 now = perf_clock();
232 ctx->time += now - ctx->timestamp;
233 ctx->timestamp = now;
237 * Update the total_time_enabled and total_time_running fields for a event.
239 static void update_event_times(struct perf_event *event)
241 struct perf_event_context *ctx = event->ctx;
244 if (event->state < PERF_EVENT_STATE_INACTIVE ||
245 event->group_leader->state < PERF_EVENT_STATE_INACTIVE)
251 run_end = event->tstamp_stopped;
253 event->total_time_enabled = run_end - event->tstamp_enabled;
255 if (event->state == PERF_EVENT_STATE_INACTIVE)
256 run_end = event->tstamp_stopped;
260 event->total_time_running = run_end - event->tstamp_running;
264 * Update total_time_enabled and total_time_running for all events in a group.
266 static void update_group_times(struct perf_event *leader)
268 struct perf_event *event;
270 update_event_times(leader);
271 list_for_each_entry(event, &leader->sibling_list, group_entry)
272 update_event_times(event);
275 static struct list_head *
276 ctx_group_list(struct perf_event *event, struct perf_event_context *ctx)
278 if (event->attr.pinned)
279 return &ctx->pinned_groups;
281 return &ctx->flexible_groups;
285 * Add a event from the lists for its context.
286 * Must be called with ctx->mutex and ctx->lock held.
289 list_add_event(struct perf_event *event, struct perf_event_context *ctx)
291 WARN_ON_ONCE(event->attach_state & PERF_ATTACH_CONTEXT);
292 event->attach_state |= PERF_ATTACH_CONTEXT;
295 * If we're a stand alone event or group leader, we go to the context
296 * list, group events are kept attached to the group so that
297 * perf_group_detach can, at all times, locate all siblings.
299 if (event->group_leader == event) {
300 struct list_head *list;
302 if (is_software_event(event))
303 event->group_flags |= PERF_GROUP_SOFTWARE;
305 list = ctx_group_list(event, ctx);
306 list_add_tail(&event->group_entry, list);
309 list_add_rcu(&event->event_entry, &ctx->event_list);
311 if (event->attr.inherit_stat)
315 static void perf_group_attach(struct perf_event *event)
317 struct perf_event *group_leader = event->group_leader;
319 WARN_ON_ONCE(event->attach_state & PERF_ATTACH_GROUP);
320 event->attach_state |= PERF_ATTACH_GROUP;
322 if (group_leader == event)
325 if (group_leader->group_flags & PERF_GROUP_SOFTWARE &&
326 !is_software_event(event))
327 group_leader->group_flags &= ~PERF_GROUP_SOFTWARE;
329 list_add_tail(&event->group_entry, &group_leader->sibling_list);
330 group_leader->nr_siblings++;
334 * Remove a event from the lists for its context.
335 * Must be called with ctx->mutex and ctx->lock held.
338 list_del_event(struct perf_event *event, struct perf_event_context *ctx)
341 * We can have double detach due to exit/hot-unplug + close.
343 if (!(event->attach_state & PERF_ATTACH_CONTEXT))
346 event->attach_state &= ~PERF_ATTACH_CONTEXT;
349 if (event->attr.inherit_stat)
352 list_del_rcu(&event->event_entry);
354 if (event->group_leader == event)
355 list_del_init(&event->group_entry);
357 update_group_times(event);
360 * If event was in error state, then keep it
361 * that way, otherwise bogus counts will be
362 * returned on read(). The only way to get out
363 * of error state is by explicit re-enabling
366 if (event->state > PERF_EVENT_STATE_OFF)
367 event->state = PERF_EVENT_STATE_OFF;
370 static void perf_group_detach(struct perf_event *event)
372 struct perf_event *sibling, *tmp;
373 struct list_head *list = NULL;
376 * We can have double detach due to exit/hot-unplug + close.
378 if (!(event->attach_state & PERF_ATTACH_GROUP))
381 event->attach_state &= ~PERF_ATTACH_GROUP;
384 * If this is a sibling, remove it from its group.
386 if (event->group_leader != event) {
387 list_del_init(&event->group_entry);
388 event->group_leader->nr_siblings--;
392 if (!list_empty(&event->group_entry))
393 list = &event->group_entry;
396 * If this was a group event with sibling events then
397 * upgrade the siblings to singleton events by adding them
398 * to whatever list we are on.
400 list_for_each_entry_safe(sibling, tmp, &event->sibling_list, group_entry) {
402 list_move_tail(&sibling->group_entry, list);
403 sibling->group_leader = sibling;
405 /* Inherit group flags from the previous leader */
406 sibling->group_flags = event->group_flags;
411 event_sched_out(struct perf_event *event,
412 struct perf_cpu_context *cpuctx,
413 struct perf_event_context *ctx)
415 if (event->state != PERF_EVENT_STATE_ACTIVE)
418 event->state = PERF_EVENT_STATE_INACTIVE;
419 if (event->pending_disable) {
420 event->pending_disable = 0;
421 event->state = PERF_EVENT_STATE_OFF;
423 event->tstamp_stopped = ctx->time;
424 event->pmu->disable(event);
427 if (!is_software_event(event))
428 cpuctx->active_oncpu--;
430 if (event->attr.exclusive || !cpuctx->active_oncpu)
431 cpuctx->exclusive = 0;
435 group_sched_out(struct perf_event *group_event,
436 struct perf_cpu_context *cpuctx,
437 struct perf_event_context *ctx)
439 struct perf_event *event;
441 if (group_event->state != PERF_EVENT_STATE_ACTIVE)
444 event_sched_out(group_event, cpuctx, ctx);
447 * Schedule out siblings (if any):
449 list_for_each_entry(event, &group_event->sibling_list, group_entry)
450 event_sched_out(event, cpuctx, ctx);
452 if (group_event->attr.exclusive)
453 cpuctx->exclusive = 0;
457 * Cross CPU call to remove a performance event
459 * We disable the event on the hardware level first. After that we
460 * remove it from the context list.
462 static void __perf_event_remove_from_context(void *info)
464 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
465 struct perf_event *event = info;
466 struct perf_event_context *ctx = event->ctx;
469 * If this is a task context, we need to check whether it is
470 * the current task context of this cpu. If not it has been
471 * scheduled out before the smp call arrived.
473 if (ctx->task && cpuctx->task_ctx != ctx)
476 raw_spin_lock(&ctx->lock);
478 * Protect the list operation against NMI by disabling the
479 * events on a global level.
483 event_sched_out(event, cpuctx, ctx);
485 list_del_event(event, ctx);
489 * Allow more per task events with respect to the
492 cpuctx->max_pertask =
493 min(perf_max_events - ctx->nr_events,
494 perf_max_events - perf_reserved_percpu);
498 raw_spin_unlock(&ctx->lock);
503 * Remove the event from a task's (or a CPU's) list of events.
505 * Must be called with ctx->mutex held.
507 * CPU events are removed with a smp call. For task events we only
508 * call when the task is on a CPU.
510 * If event->ctx is a cloned context, callers must make sure that
511 * every task struct that event->ctx->task could possibly point to
512 * remains valid. This is OK when called from perf_release since
513 * that only calls us on the top-level context, which can't be a clone.
514 * When called from perf_event_exit_task, it's OK because the
515 * context has been detached from its task.
517 static void perf_event_remove_from_context(struct perf_event *event)
519 struct perf_event_context *ctx = event->ctx;
520 struct task_struct *task = ctx->task;
524 * Per cpu events are removed via an smp call and
525 * the removal is always successful.
527 smp_call_function_single(event->cpu,
528 __perf_event_remove_from_context,
534 task_oncpu_function_call(task, __perf_event_remove_from_context,
537 raw_spin_lock_irq(&ctx->lock);
539 * If the context is active we need to retry the smp call.
541 if (ctx->nr_active && !list_empty(&event->group_entry)) {
542 raw_spin_unlock_irq(&ctx->lock);
547 * The lock prevents that this context is scheduled in so we
548 * can remove the event safely, if the call above did not
551 if (!list_empty(&event->group_entry))
552 list_del_event(event, ctx);
553 raw_spin_unlock_irq(&ctx->lock);
557 * Cross CPU call to disable a performance event
559 static void __perf_event_disable(void *info)
561 struct perf_event *event = info;
562 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
563 struct perf_event_context *ctx = event->ctx;
566 * If this is a per-task event, need to check whether this
567 * event's task is the current task on this cpu.
569 if (ctx->task && cpuctx->task_ctx != ctx)
572 raw_spin_lock(&ctx->lock);
575 * If the event is on, turn it off.
576 * If it is in error state, leave it in error state.
578 if (event->state >= PERF_EVENT_STATE_INACTIVE) {
579 update_context_time(ctx);
580 update_group_times(event);
581 if (event == event->group_leader)
582 group_sched_out(event, cpuctx, ctx);
584 event_sched_out(event, cpuctx, ctx);
585 event->state = PERF_EVENT_STATE_OFF;
588 raw_spin_unlock(&ctx->lock);
594 * If event->ctx is a cloned context, callers must make sure that
595 * every task struct that event->ctx->task could possibly point to
596 * remains valid. This condition is satisifed when called through
597 * perf_event_for_each_child or perf_event_for_each because they
598 * hold the top-level event's child_mutex, so any descendant that
599 * goes to exit will block in sync_child_event.
600 * When called from perf_pending_event it's OK because event->ctx
601 * is the current context on this CPU and preemption is disabled,
602 * hence we can't get into perf_event_task_sched_out for this context.
604 void perf_event_disable(struct perf_event *event)
606 struct perf_event_context *ctx = event->ctx;
607 struct task_struct *task = ctx->task;
611 * Disable the event on the cpu that it's on
613 smp_call_function_single(event->cpu, __perf_event_disable,
619 task_oncpu_function_call(task, __perf_event_disable, event);
621 raw_spin_lock_irq(&ctx->lock);
623 * If the event is still active, we need to retry the cross-call.
625 if (event->state == PERF_EVENT_STATE_ACTIVE) {
626 raw_spin_unlock_irq(&ctx->lock);
631 * Since we have the lock this context can't be scheduled
632 * in, so we can change the state safely.
634 if (event->state == PERF_EVENT_STATE_INACTIVE) {
635 update_group_times(event);
636 event->state = PERF_EVENT_STATE_OFF;
639 raw_spin_unlock_irq(&ctx->lock);
643 event_sched_in(struct perf_event *event,
644 struct perf_cpu_context *cpuctx,
645 struct perf_event_context *ctx)
647 if (event->state <= PERF_EVENT_STATE_OFF)
650 event->state = PERF_EVENT_STATE_ACTIVE;
651 event->oncpu = smp_processor_id();
653 * The new state must be visible before we turn it on in the hardware:
657 if (event->pmu->enable(event)) {
658 event->state = PERF_EVENT_STATE_INACTIVE;
663 event->tstamp_running += ctx->time - event->tstamp_stopped;
665 if (!is_software_event(event))
666 cpuctx->active_oncpu++;
669 if (event->attr.exclusive)
670 cpuctx->exclusive = 1;
676 group_sched_in(struct perf_event *group_event,
677 struct perf_cpu_context *cpuctx,
678 struct perf_event_context *ctx)
680 struct perf_event *event, *partial_group = NULL;
681 const struct pmu *pmu = group_event->pmu;
684 if (group_event->state == PERF_EVENT_STATE_OFF)
687 /* Check if group transaction availabe */
694 if (event_sched_in(group_event, cpuctx, ctx)) {
696 pmu->cancel_txn(pmu);
701 * Schedule in siblings as one group (if any):
703 list_for_each_entry(event, &group_event->sibling_list, group_entry) {
704 if (event_sched_in(event, cpuctx, ctx)) {
705 partial_group = event;
710 if (!txn || !pmu->commit_txn(pmu))
715 * Groups can be scheduled in as one unit only, so undo any
716 * partial group before returning:
718 list_for_each_entry(event, &group_event->sibling_list, group_entry) {
719 if (event == partial_group)
721 event_sched_out(event, cpuctx, ctx);
723 event_sched_out(group_event, cpuctx, ctx);
726 pmu->cancel_txn(pmu);
732 * Work out whether we can put this event group on the CPU now.
734 static int group_can_go_on(struct perf_event *event,
735 struct perf_cpu_context *cpuctx,
739 * Groups consisting entirely of software events can always go on.
741 if (event->group_flags & PERF_GROUP_SOFTWARE)
744 * If an exclusive group is already on, no other hardware
747 if (cpuctx->exclusive)
750 * If this group is exclusive and there are already
751 * events on the CPU, it can't go on.
753 if (event->attr.exclusive && cpuctx->active_oncpu)
756 * Otherwise, try to add it if all previous groups were able
762 static void add_event_to_ctx(struct perf_event *event,
763 struct perf_event_context *ctx)
765 list_add_event(event, ctx);
766 perf_group_attach(event);
767 event->tstamp_enabled = ctx->time;
768 event->tstamp_running = ctx->time;
769 event->tstamp_stopped = ctx->time;
773 * Cross CPU call to install and enable a performance event
775 * Must be called with ctx->mutex held
777 static void __perf_install_in_context(void *info)
779 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
780 struct perf_event *event = info;
781 struct perf_event_context *ctx = event->ctx;
782 struct perf_event *leader = event->group_leader;
786 * If this is a task context, we need to check whether it is
787 * the current task context of this cpu. If not it has been
788 * scheduled out before the smp call arrived.
789 * Or possibly this is the right context but it isn't
790 * on this cpu because it had no events.
792 if (ctx->task && cpuctx->task_ctx != ctx) {
793 if (cpuctx->task_ctx || ctx->task != current)
795 cpuctx->task_ctx = ctx;
798 raw_spin_lock(&ctx->lock);
800 update_context_time(ctx);
803 * Protect the list operation against NMI by disabling the
804 * events on a global level. NOP for non NMI based events.
808 add_event_to_ctx(event, ctx);
810 if (event->cpu != -1 && event->cpu != smp_processor_id())
814 * Don't put the event on if it is disabled or if
815 * it is in a group and the group isn't on.
817 if (event->state != PERF_EVENT_STATE_INACTIVE ||
818 (leader != event && leader->state != PERF_EVENT_STATE_ACTIVE))
822 * An exclusive event can't go on if there are already active
823 * hardware events, and no hardware event can go on if there
824 * is already an exclusive event on.
826 if (!group_can_go_on(event, cpuctx, 1))
829 err = event_sched_in(event, cpuctx, ctx);
833 * This event couldn't go on. If it is in a group
834 * then we have to pull the whole group off.
835 * If the event group is pinned then put it in error state.
838 group_sched_out(leader, cpuctx, ctx);
839 if (leader->attr.pinned) {
840 update_group_times(leader);
841 leader->state = PERF_EVENT_STATE_ERROR;
845 if (!err && !ctx->task && cpuctx->max_pertask)
846 cpuctx->max_pertask--;
851 raw_spin_unlock(&ctx->lock);
855 * Attach a performance event to a context
857 * First we add the event to the list with the hardware enable bit
858 * in event->hw_config cleared.
860 * If the event is attached to a task which is on a CPU we use a smp
861 * call to enable it in the task context. The task might have been
862 * scheduled away, but we check this in the smp call again.
864 * Must be called with ctx->mutex held.
867 perf_install_in_context(struct perf_event_context *ctx,
868 struct perf_event *event,
871 struct task_struct *task = ctx->task;
875 * Per cpu events are installed via an smp call and
876 * the install is always successful.
878 smp_call_function_single(cpu, __perf_install_in_context,
884 task_oncpu_function_call(task, __perf_install_in_context,
887 raw_spin_lock_irq(&ctx->lock);
889 * we need to retry the smp call.
891 if (ctx->is_active && list_empty(&event->group_entry)) {
892 raw_spin_unlock_irq(&ctx->lock);
897 * The lock prevents that this context is scheduled in so we
898 * can add the event safely, if it the call above did not
901 if (list_empty(&event->group_entry))
902 add_event_to_ctx(event, ctx);
903 raw_spin_unlock_irq(&ctx->lock);
907 * Put a event into inactive state and update time fields.
908 * Enabling the leader of a group effectively enables all
909 * the group members that aren't explicitly disabled, so we
910 * have to update their ->tstamp_enabled also.
911 * Note: this works for group members as well as group leaders
912 * since the non-leader members' sibling_lists will be empty.
914 static void __perf_event_mark_enabled(struct perf_event *event,
915 struct perf_event_context *ctx)
917 struct perf_event *sub;
919 event->state = PERF_EVENT_STATE_INACTIVE;
920 event->tstamp_enabled = ctx->time - event->total_time_enabled;
921 list_for_each_entry(sub, &event->sibling_list, group_entry)
922 if (sub->state >= PERF_EVENT_STATE_INACTIVE)
923 sub->tstamp_enabled =
924 ctx->time - sub->total_time_enabled;
928 * Cross CPU call to enable a performance event
930 static void __perf_event_enable(void *info)
932 struct perf_event *event = info;
933 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
934 struct perf_event_context *ctx = event->ctx;
935 struct perf_event *leader = event->group_leader;
939 * If this is a per-task event, need to check whether this
940 * event's task is the current task on this cpu.
942 if (ctx->task && cpuctx->task_ctx != ctx) {
943 if (cpuctx->task_ctx || ctx->task != current)
945 cpuctx->task_ctx = ctx;
948 raw_spin_lock(&ctx->lock);
950 update_context_time(ctx);
952 if (event->state >= PERF_EVENT_STATE_INACTIVE)
954 __perf_event_mark_enabled(event, ctx);
956 if (event->cpu != -1 && event->cpu != smp_processor_id())
960 * If the event is in a group and isn't the group leader,
961 * then don't put it on unless the group is on.
963 if (leader != event && leader->state != PERF_EVENT_STATE_ACTIVE)
966 if (!group_can_go_on(event, cpuctx, 1)) {
971 err = group_sched_in(event, cpuctx, ctx);
973 err = event_sched_in(event, cpuctx, ctx);
979 * If this event can't go on and it's part of a
980 * group, then the whole group has to come off.
983 group_sched_out(leader, cpuctx, ctx);
984 if (leader->attr.pinned) {
985 update_group_times(leader);
986 leader->state = PERF_EVENT_STATE_ERROR;
991 raw_spin_unlock(&ctx->lock);
997 * If event->ctx is a cloned context, callers must make sure that
998 * every task struct that event->ctx->task could possibly point to
999 * remains valid. This condition is satisfied when called through
1000 * perf_event_for_each_child or perf_event_for_each as described
1001 * for perf_event_disable.
1003 void perf_event_enable(struct perf_event *event)
1005 struct perf_event_context *ctx = event->ctx;
1006 struct task_struct *task = ctx->task;
1010 * Enable the event on the cpu that it's on
1012 smp_call_function_single(event->cpu, __perf_event_enable,
1017 raw_spin_lock_irq(&ctx->lock);
1018 if (event->state >= PERF_EVENT_STATE_INACTIVE)
1022 * If the event is in error state, clear that first.
1023 * That way, if we see the event in error state below, we
1024 * know that it has gone back into error state, as distinct
1025 * from the task having been scheduled away before the
1026 * cross-call arrived.
1028 if (event->state == PERF_EVENT_STATE_ERROR)
1029 event->state = PERF_EVENT_STATE_OFF;
1032 raw_spin_unlock_irq(&ctx->lock);
1033 task_oncpu_function_call(task, __perf_event_enable, event);
1035 raw_spin_lock_irq(&ctx->lock);
1038 * If the context is active and the event is still off,
1039 * we need to retry the cross-call.
1041 if (ctx->is_active && event->state == PERF_EVENT_STATE_OFF)
1045 * Since we have the lock this context can't be scheduled
1046 * in, so we can change the state safely.
1048 if (event->state == PERF_EVENT_STATE_OFF)
1049 __perf_event_mark_enabled(event, ctx);
1052 raw_spin_unlock_irq(&ctx->lock);
1055 static int perf_event_refresh(struct perf_event *event, int refresh)
1058 * not supported on inherited events
1060 if (event->attr.inherit)
1063 atomic_add(refresh, &event->event_limit);
1064 perf_event_enable(event);
1070 EVENT_FLEXIBLE = 0x1,
1072 EVENT_ALL = EVENT_FLEXIBLE | EVENT_PINNED,
1075 static void ctx_sched_out(struct perf_event_context *ctx,
1076 struct perf_cpu_context *cpuctx,
1077 enum event_type_t event_type)
1079 struct perf_event *event;
1081 raw_spin_lock(&ctx->lock);
1083 if (likely(!ctx->nr_events))
1085 update_context_time(ctx);
1088 if (!ctx->nr_active)
1091 if (event_type & EVENT_PINNED)
1092 list_for_each_entry(event, &ctx->pinned_groups, group_entry)
1093 group_sched_out(event, cpuctx, ctx);
1095 if (event_type & EVENT_FLEXIBLE)
1096 list_for_each_entry(event, &ctx->flexible_groups, group_entry)
1097 group_sched_out(event, cpuctx, ctx);
1102 raw_spin_unlock(&ctx->lock);
1106 * Test whether two contexts are equivalent, i.e. whether they
1107 * have both been cloned from the same version of the same context
1108 * and they both have the same number of enabled events.
1109 * If the number of enabled events is the same, then the set
1110 * of enabled events should be the same, because these are both
1111 * inherited contexts, therefore we can't access individual events
1112 * in them directly with an fd; we can only enable/disable all
1113 * events via prctl, or enable/disable all events in a family
1114 * via ioctl, which will have the same effect on both contexts.
1116 static int context_equiv(struct perf_event_context *ctx1,
1117 struct perf_event_context *ctx2)
1119 return ctx1->parent_ctx && ctx1->parent_ctx == ctx2->parent_ctx
1120 && ctx1->parent_gen == ctx2->parent_gen
1121 && !ctx1->pin_count && !ctx2->pin_count;
1124 static void __perf_event_sync_stat(struct perf_event *event,
1125 struct perf_event *next_event)
1129 if (!event->attr.inherit_stat)
1133 * Update the event value, we cannot use perf_event_read()
1134 * because we're in the middle of a context switch and have IRQs
1135 * disabled, which upsets smp_call_function_single(), however
1136 * we know the event must be on the current CPU, therefore we
1137 * don't need to use it.
1139 switch (event->state) {
1140 case PERF_EVENT_STATE_ACTIVE:
1141 event->pmu->read(event);
1144 case PERF_EVENT_STATE_INACTIVE:
1145 update_event_times(event);
1153 * In order to keep per-task stats reliable we need to flip the event
1154 * values when we flip the contexts.
1156 value = local64_read(&next_event->count);
1157 value = local64_xchg(&event->count, value);
1158 local64_set(&next_event->count, value);
1160 swap(event->total_time_enabled, next_event->total_time_enabled);
1161 swap(event->total_time_running, next_event->total_time_running);
1164 * Since we swizzled the values, update the user visible data too.
1166 perf_event_update_userpage(event);
1167 perf_event_update_userpage(next_event);
1170 #define list_next_entry(pos, member) \
1171 list_entry(pos->member.next, typeof(*pos), member)
1173 static void perf_event_sync_stat(struct perf_event_context *ctx,
1174 struct perf_event_context *next_ctx)
1176 struct perf_event *event, *next_event;
1181 update_context_time(ctx);
1183 event = list_first_entry(&ctx->event_list,
1184 struct perf_event, event_entry);
1186 next_event = list_first_entry(&next_ctx->event_list,
1187 struct perf_event, event_entry);
1189 while (&event->event_entry != &ctx->event_list &&
1190 &next_event->event_entry != &next_ctx->event_list) {
1192 __perf_event_sync_stat(event, next_event);
1194 event = list_next_entry(event, event_entry);
1195 next_event = list_next_entry(next_event, event_entry);
1200 * Called from scheduler to remove the events of the current task,
1201 * with interrupts disabled.
1203 * We stop each event and update the event value in event->count.
1205 * This does not protect us against NMI, but disable()
1206 * sets the disabled bit in the control field of event _before_
1207 * accessing the event control register. If a NMI hits, then it will
1208 * not restart the event.
1210 void perf_event_task_sched_out(struct task_struct *task,
1211 struct task_struct *next)
1213 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
1214 struct perf_event_context *ctx = task->perf_event_ctxp;
1215 struct perf_event_context *next_ctx;
1216 struct perf_event_context *parent;
1219 perf_sw_event(PERF_COUNT_SW_CONTEXT_SWITCHES, 1, 1, NULL, 0);
1221 if (likely(!ctx || !cpuctx->task_ctx))
1225 parent = rcu_dereference(ctx->parent_ctx);
1226 next_ctx = next->perf_event_ctxp;
1227 if (parent && next_ctx &&
1228 rcu_dereference(next_ctx->parent_ctx) == parent) {
1230 * Looks like the two contexts are clones, so we might be
1231 * able to optimize the context switch. We lock both
1232 * contexts and check that they are clones under the
1233 * lock (including re-checking that neither has been
1234 * uncloned in the meantime). It doesn't matter which
1235 * order we take the locks because no other cpu could
1236 * be trying to lock both of these tasks.
1238 raw_spin_lock(&ctx->lock);
1239 raw_spin_lock_nested(&next_ctx->lock, SINGLE_DEPTH_NESTING);
1240 if (context_equiv(ctx, next_ctx)) {
1242 * XXX do we need a memory barrier of sorts
1243 * wrt to rcu_dereference() of perf_event_ctxp
1245 task->perf_event_ctxp = next_ctx;
1246 next->perf_event_ctxp = ctx;
1248 next_ctx->task = task;
1251 perf_event_sync_stat(ctx, next_ctx);
1253 raw_spin_unlock(&next_ctx->lock);
1254 raw_spin_unlock(&ctx->lock);
1259 ctx_sched_out(ctx, cpuctx, EVENT_ALL);
1260 cpuctx->task_ctx = NULL;
1264 static void task_ctx_sched_out(struct perf_event_context *ctx,
1265 enum event_type_t event_type)
1267 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
1269 if (!cpuctx->task_ctx)
1272 if (WARN_ON_ONCE(ctx != cpuctx->task_ctx))
1275 ctx_sched_out(ctx, cpuctx, event_type);
1276 cpuctx->task_ctx = NULL;
1280 * Called with IRQs disabled
1282 static void __perf_event_task_sched_out(struct perf_event_context *ctx)
1284 task_ctx_sched_out(ctx, EVENT_ALL);
1288 * Called with IRQs disabled
1290 static void cpu_ctx_sched_out(struct perf_cpu_context *cpuctx,
1291 enum event_type_t event_type)
1293 ctx_sched_out(&cpuctx->ctx, cpuctx, event_type);
1297 ctx_pinned_sched_in(struct perf_event_context *ctx,
1298 struct perf_cpu_context *cpuctx)
1300 struct perf_event *event;
1302 list_for_each_entry(event, &ctx->pinned_groups, group_entry) {
1303 if (event->state <= PERF_EVENT_STATE_OFF)
1305 if (event->cpu != -1 && event->cpu != smp_processor_id())
1308 if (group_can_go_on(event, cpuctx, 1))
1309 group_sched_in(event, cpuctx, ctx);
1312 * If this pinned group hasn't been scheduled,
1313 * put it in error state.
1315 if (event->state == PERF_EVENT_STATE_INACTIVE) {
1316 update_group_times(event);
1317 event->state = PERF_EVENT_STATE_ERROR;
1323 ctx_flexible_sched_in(struct perf_event_context *ctx,
1324 struct perf_cpu_context *cpuctx)
1326 struct perf_event *event;
1329 list_for_each_entry(event, &ctx->flexible_groups, group_entry) {
1330 /* Ignore events in OFF or ERROR state */
1331 if (event->state <= PERF_EVENT_STATE_OFF)
1334 * Listen to the 'cpu' scheduling filter constraint
1337 if (event->cpu != -1 && event->cpu != smp_processor_id())
1340 if (group_can_go_on(event, cpuctx, can_add_hw))
1341 if (group_sched_in(event, cpuctx, ctx))
1347 ctx_sched_in(struct perf_event_context *ctx,
1348 struct perf_cpu_context *cpuctx,
1349 enum event_type_t event_type)
1351 raw_spin_lock(&ctx->lock);
1353 if (likely(!ctx->nr_events))
1356 ctx->timestamp = perf_clock();
1361 * First go through the list and put on any pinned groups
1362 * in order to give them the best chance of going on.
1364 if (event_type & EVENT_PINNED)
1365 ctx_pinned_sched_in(ctx, cpuctx);
1367 /* Then walk through the lower prio flexible groups */
1368 if (event_type & EVENT_FLEXIBLE)
1369 ctx_flexible_sched_in(ctx, cpuctx);
1373 raw_spin_unlock(&ctx->lock);
1376 static void cpu_ctx_sched_in(struct perf_cpu_context *cpuctx,
1377 enum event_type_t event_type)
1379 struct perf_event_context *ctx = &cpuctx->ctx;
1381 ctx_sched_in(ctx, cpuctx, event_type);
1384 static void task_ctx_sched_in(struct task_struct *task,
1385 enum event_type_t event_type)
1387 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
1388 struct perf_event_context *ctx = task->perf_event_ctxp;
1392 if (cpuctx->task_ctx == ctx)
1394 ctx_sched_in(ctx, cpuctx, event_type);
1395 cpuctx->task_ctx = ctx;
1398 * Called from scheduler to add the events of the current task
1399 * with interrupts disabled.
1401 * We restore the event value and then enable it.
1403 * This does not protect us against NMI, but enable()
1404 * sets the enabled bit in the control field of event _before_
1405 * accessing the event control register. If a NMI hits, then it will
1406 * keep the event running.
1408 void perf_event_task_sched_in(struct task_struct *task)
1410 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
1411 struct perf_event_context *ctx = task->perf_event_ctxp;
1416 if (cpuctx->task_ctx == ctx)
1422 * We want to keep the following priority order:
1423 * cpu pinned (that don't need to move), task pinned,
1424 * cpu flexible, task flexible.
1426 cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE);
1428 ctx_sched_in(ctx, cpuctx, EVENT_PINNED);
1429 cpu_ctx_sched_in(cpuctx, EVENT_FLEXIBLE);
1430 ctx_sched_in(ctx, cpuctx, EVENT_FLEXIBLE);
1432 cpuctx->task_ctx = ctx;
1437 #define MAX_INTERRUPTS (~0ULL)
1439 static void perf_log_throttle(struct perf_event *event, int enable);
1441 static u64 perf_calculate_period(struct perf_event *event, u64 nsec, u64 count)
1443 u64 frequency = event->attr.sample_freq;
1444 u64 sec = NSEC_PER_SEC;
1445 u64 divisor, dividend;
1447 int count_fls, nsec_fls, frequency_fls, sec_fls;
1449 count_fls = fls64(count);
1450 nsec_fls = fls64(nsec);
1451 frequency_fls = fls64(frequency);
1455 * We got @count in @nsec, with a target of sample_freq HZ
1456 * the target period becomes:
1459 * period = -------------------
1460 * @nsec * sample_freq
1465 * Reduce accuracy by one bit such that @a and @b converge
1466 * to a similar magnitude.
1468 #define REDUCE_FLS(a, b) \
1470 if (a##_fls > b##_fls) { \
1480 * Reduce accuracy until either term fits in a u64, then proceed with
1481 * the other, so that finally we can do a u64/u64 division.
1483 while (count_fls + sec_fls > 64 && nsec_fls + frequency_fls > 64) {
1484 REDUCE_FLS(nsec, frequency);
1485 REDUCE_FLS(sec, count);
1488 if (count_fls + sec_fls > 64) {
1489 divisor = nsec * frequency;
1491 while (count_fls + sec_fls > 64) {
1492 REDUCE_FLS(count, sec);
1496 dividend = count * sec;
1498 dividend = count * sec;
1500 while (nsec_fls + frequency_fls > 64) {
1501 REDUCE_FLS(nsec, frequency);
1505 divisor = nsec * frequency;
1511 return div64_u64(dividend, divisor);
1514 static void perf_event_stop(struct perf_event *event)
1516 if (!event->pmu->stop)
1517 return event->pmu->disable(event);
1519 return event->pmu->stop(event);
1522 static int perf_event_start(struct perf_event *event)
1524 if (!event->pmu->start)
1525 return event->pmu->enable(event);
1527 return event->pmu->start(event);
1530 static void perf_adjust_period(struct perf_event *event, u64 nsec, u64 count)
1532 struct hw_perf_event *hwc = &event->hw;
1533 s64 period, sample_period;
1536 period = perf_calculate_period(event, nsec, count);
1538 delta = (s64)(period - hwc->sample_period);
1539 delta = (delta + 7) / 8; /* low pass filter */
1541 sample_period = hwc->sample_period + delta;
1546 hwc->sample_period = sample_period;
1548 if (local64_read(&hwc->period_left) > 8*sample_period) {
1550 perf_event_stop(event);
1551 local64_set(&hwc->period_left, 0);
1552 perf_event_start(event);
1557 static void perf_ctx_adjust_freq(struct perf_event_context *ctx)
1559 struct perf_event *event;
1560 struct hw_perf_event *hwc;
1561 u64 interrupts, now;
1564 raw_spin_lock(&ctx->lock);
1565 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
1566 if (event->state != PERF_EVENT_STATE_ACTIVE)
1569 if (event->cpu != -1 && event->cpu != smp_processor_id())
1574 interrupts = hwc->interrupts;
1575 hwc->interrupts = 0;
1578 * unthrottle events on the tick
1580 if (interrupts == MAX_INTERRUPTS) {
1581 perf_log_throttle(event, 1);
1583 event->pmu->unthrottle(event);
1587 if (!event->attr.freq || !event->attr.sample_freq)
1591 event->pmu->read(event);
1592 now = local64_read(&event->count);
1593 delta = now - hwc->freq_count_stamp;
1594 hwc->freq_count_stamp = now;
1597 perf_adjust_period(event, TICK_NSEC, delta);
1600 raw_spin_unlock(&ctx->lock);
1604 * Round-robin a context's events:
1606 static void rotate_ctx(struct perf_event_context *ctx)
1608 raw_spin_lock(&ctx->lock);
1610 /* Rotate the first entry last of non-pinned groups */
1611 list_rotate_left(&ctx->flexible_groups);
1613 raw_spin_unlock(&ctx->lock);
1616 void perf_event_task_tick(struct task_struct *curr)
1618 struct perf_cpu_context *cpuctx;
1619 struct perf_event_context *ctx;
1622 if (!atomic_read(&nr_events))
1625 cpuctx = &__get_cpu_var(perf_cpu_context);
1626 if (cpuctx->ctx.nr_events &&
1627 cpuctx->ctx.nr_events != cpuctx->ctx.nr_active)
1630 ctx = curr->perf_event_ctxp;
1631 if (ctx && ctx->nr_events && ctx->nr_events != ctx->nr_active)
1634 perf_ctx_adjust_freq(&cpuctx->ctx);
1636 perf_ctx_adjust_freq(ctx);
1642 cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE);
1644 task_ctx_sched_out(ctx, EVENT_FLEXIBLE);
1646 rotate_ctx(&cpuctx->ctx);
1650 cpu_ctx_sched_in(cpuctx, EVENT_FLEXIBLE);
1652 task_ctx_sched_in(curr, EVENT_FLEXIBLE);
1656 static int event_enable_on_exec(struct perf_event *event,
1657 struct perf_event_context *ctx)
1659 if (!event->attr.enable_on_exec)
1662 event->attr.enable_on_exec = 0;
1663 if (event->state >= PERF_EVENT_STATE_INACTIVE)
1666 __perf_event_mark_enabled(event, ctx);
1672 * Enable all of a task's events that have been marked enable-on-exec.
1673 * This expects task == current.
1675 static void perf_event_enable_on_exec(struct task_struct *task)
1677 struct perf_event_context *ctx;
1678 struct perf_event *event;
1679 unsigned long flags;
1683 local_irq_save(flags);
1684 ctx = task->perf_event_ctxp;
1685 if (!ctx || !ctx->nr_events)
1688 __perf_event_task_sched_out(ctx);
1690 raw_spin_lock(&ctx->lock);
1692 list_for_each_entry(event, &ctx->pinned_groups, group_entry) {
1693 ret = event_enable_on_exec(event, ctx);
1698 list_for_each_entry(event, &ctx->flexible_groups, group_entry) {
1699 ret = event_enable_on_exec(event, ctx);
1705 * Unclone this context if we enabled any event.
1710 raw_spin_unlock(&ctx->lock);
1712 perf_event_task_sched_in(task);
1714 local_irq_restore(flags);
1718 * Cross CPU call to read the hardware event
1720 static void __perf_event_read(void *info)
1722 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
1723 struct perf_event *event = info;
1724 struct perf_event_context *ctx = event->ctx;
1727 * If this is a task context, we need to check whether it is
1728 * the current task context of this cpu. If not it has been
1729 * scheduled out before the smp call arrived. In that case
1730 * event->count would have been updated to a recent sample
1731 * when the event was scheduled out.
1733 if (ctx->task && cpuctx->task_ctx != ctx)
1736 raw_spin_lock(&ctx->lock);
1737 update_context_time(ctx);
1738 update_event_times(event);
1739 raw_spin_unlock(&ctx->lock);
1741 event->pmu->read(event);
1744 static inline u64 perf_event_count(struct perf_event *event)
1746 return local64_read(&event->count) + atomic64_read(&event->child_count);
1749 static u64 perf_event_read(struct perf_event *event)
1752 * If event is enabled and currently active on a CPU, update the
1753 * value in the event structure:
1755 if (event->state == PERF_EVENT_STATE_ACTIVE) {
1756 smp_call_function_single(event->oncpu,
1757 __perf_event_read, event, 1);
1758 } else if (event->state == PERF_EVENT_STATE_INACTIVE) {
1759 struct perf_event_context *ctx = event->ctx;
1760 unsigned long flags;
1762 raw_spin_lock_irqsave(&ctx->lock, flags);
1763 update_context_time(ctx);
1764 update_event_times(event);
1765 raw_spin_unlock_irqrestore(&ctx->lock, flags);
1768 return perf_event_count(event);
1772 * Initialize the perf_event context in a task_struct:
1775 __perf_event_init_context(struct perf_event_context *ctx,
1776 struct task_struct *task)
1778 raw_spin_lock_init(&ctx->lock);
1779 mutex_init(&ctx->mutex);
1780 INIT_LIST_HEAD(&ctx->pinned_groups);
1781 INIT_LIST_HEAD(&ctx->flexible_groups);
1782 INIT_LIST_HEAD(&ctx->event_list);
1783 atomic_set(&ctx->refcount, 1);
1787 static struct perf_event_context *find_get_context(pid_t pid, int cpu)
1789 struct perf_event_context *ctx;
1790 struct perf_cpu_context *cpuctx;
1791 struct task_struct *task;
1792 unsigned long flags;
1795 if (pid == -1 && cpu != -1) {
1796 /* Must be root to operate on a CPU event: */
1797 if (perf_paranoid_cpu() && !capable(CAP_SYS_ADMIN))
1798 return ERR_PTR(-EACCES);
1800 if (cpu < 0 || cpu >= nr_cpumask_bits)
1801 return ERR_PTR(-EINVAL);
1804 * We could be clever and allow to attach a event to an
1805 * offline CPU and activate it when the CPU comes up, but
1808 if (!cpu_online(cpu))
1809 return ERR_PTR(-ENODEV);
1811 cpuctx = &per_cpu(perf_cpu_context, cpu);
1822 task = find_task_by_vpid(pid);
1824 get_task_struct(task);
1828 return ERR_PTR(-ESRCH);
1831 * Can't attach events to a dying task.
1834 if (task->flags & PF_EXITING)
1837 /* Reuse ptrace permission checks for now. */
1839 if (!ptrace_may_access(task, PTRACE_MODE_READ))
1843 ctx = perf_lock_task_context(task, &flags);
1846 raw_spin_unlock_irqrestore(&ctx->lock, flags);
1850 ctx = kzalloc(sizeof(struct perf_event_context), GFP_KERNEL);
1854 __perf_event_init_context(ctx, task);
1856 if (cmpxchg(&task->perf_event_ctxp, NULL, ctx)) {
1858 * We raced with some other task; use
1859 * the context they set.
1864 get_task_struct(task);
1867 put_task_struct(task);
1871 put_task_struct(task);
1872 return ERR_PTR(err);
1875 static void perf_event_free_filter(struct perf_event *event);
1877 static void free_event_rcu(struct rcu_head *head)
1879 struct perf_event *event;
1881 event = container_of(head, struct perf_event, rcu_head);
1883 put_pid_ns(event->ns);
1884 perf_event_free_filter(event);
1888 static void perf_pending_sync(struct perf_event *event);
1889 static void perf_buffer_put(struct perf_buffer *buffer);
1891 static void free_event(struct perf_event *event)
1893 perf_pending_sync(event);
1895 if (!event->parent) {
1896 atomic_dec(&nr_events);
1897 if (event->attr.mmap || event->attr.mmap_data)
1898 atomic_dec(&nr_mmap_events);
1899 if (event->attr.comm)
1900 atomic_dec(&nr_comm_events);
1901 if (event->attr.task)
1902 atomic_dec(&nr_task_events);
1905 if (event->buffer) {
1906 perf_buffer_put(event->buffer);
1907 event->buffer = NULL;
1911 event->destroy(event);
1913 put_ctx(event->ctx);
1914 call_rcu(&event->rcu_head, free_event_rcu);
1917 int perf_event_release_kernel(struct perf_event *event)
1919 struct perf_event_context *ctx = event->ctx;
1922 * Remove from the PMU, can't get re-enabled since we got
1923 * here because the last ref went.
1925 perf_event_disable(event);
1927 WARN_ON_ONCE(ctx->parent_ctx);
1929 * There are two ways this annotation is useful:
1931 * 1) there is a lock recursion from perf_event_exit_task
1932 * see the comment there.
1934 * 2) there is a lock-inversion with mmap_sem through
1935 * perf_event_read_group(), which takes faults while
1936 * holding ctx->mutex, however this is called after
1937 * the last filedesc died, so there is no possibility
1938 * to trigger the AB-BA case.
1940 mutex_lock_nested(&ctx->mutex, SINGLE_DEPTH_NESTING);
1941 raw_spin_lock_irq(&ctx->lock);
1942 perf_group_detach(event);
1943 list_del_event(event, ctx);
1944 raw_spin_unlock_irq(&ctx->lock);
1945 mutex_unlock(&ctx->mutex);
1947 mutex_lock(&event->owner->perf_event_mutex);
1948 list_del_init(&event->owner_entry);
1949 mutex_unlock(&event->owner->perf_event_mutex);
1950 put_task_struct(event->owner);
1956 EXPORT_SYMBOL_GPL(perf_event_release_kernel);
1959 * Called when the last reference to the file is gone.
1961 static int perf_release(struct inode *inode, struct file *file)
1963 struct perf_event *event = file->private_data;
1965 file->private_data = NULL;
1967 return perf_event_release_kernel(event);
1970 static int perf_event_read_size(struct perf_event *event)
1972 int entry = sizeof(u64); /* value */
1976 if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1977 size += sizeof(u64);
1979 if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1980 size += sizeof(u64);
1982 if (event->attr.read_format & PERF_FORMAT_ID)
1983 entry += sizeof(u64);
1985 if (event->attr.read_format & PERF_FORMAT_GROUP) {
1986 nr += event->group_leader->nr_siblings;
1987 size += sizeof(u64);
1995 u64 perf_event_read_value(struct perf_event *event, u64 *enabled, u64 *running)
1997 struct perf_event *child;
2003 mutex_lock(&event->child_mutex);
2004 total += perf_event_read(event);
2005 *enabled += event->total_time_enabled +
2006 atomic64_read(&event->child_total_time_enabled);
2007 *running += event->total_time_running +
2008 atomic64_read(&event->child_total_time_running);
2010 list_for_each_entry(child, &event->child_list, child_list) {
2011 total += perf_event_read(child);
2012 *enabled += child->total_time_enabled;
2013 *running += child->total_time_running;
2015 mutex_unlock(&event->child_mutex);
2019 EXPORT_SYMBOL_GPL(perf_event_read_value);
2021 static int perf_event_read_group(struct perf_event *event,
2022 u64 read_format, char __user *buf)
2024 struct perf_event *leader = event->group_leader, *sub;
2025 int n = 0, size = 0, ret = -EFAULT;
2026 struct perf_event_context *ctx = leader->ctx;
2028 u64 count, enabled, running;
2030 mutex_lock(&ctx->mutex);
2031 count = perf_event_read_value(leader, &enabled, &running);
2033 values[n++] = 1 + leader->nr_siblings;
2034 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
2035 values[n++] = enabled;
2036 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
2037 values[n++] = running;
2038 values[n++] = count;
2039 if (read_format & PERF_FORMAT_ID)
2040 values[n++] = primary_event_id(leader);
2042 size = n * sizeof(u64);
2044 if (copy_to_user(buf, values, size))
2049 list_for_each_entry(sub, &leader->sibling_list, group_entry) {
2052 values[n++] = perf_event_read_value(sub, &enabled, &running);
2053 if (read_format & PERF_FORMAT_ID)
2054 values[n++] = primary_event_id(sub);
2056 size = n * sizeof(u64);
2058 if (copy_to_user(buf + ret, values, size)) {
2066 mutex_unlock(&ctx->mutex);
2071 static int perf_event_read_one(struct perf_event *event,
2072 u64 read_format, char __user *buf)
2074 u64 enabled, running;
2078 values[n++] = perf_event_read_value(event, &enabled, &running);
2079 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
2080 values[n++] = enabled;
2081 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
2082 values[n++] = running;
2083 if (read_format & PERF_FORMAT_ID)
2084 values[n++] = primary_event_id(event);
2086 if (copy_to_user(buf, values, n * sizeof(u64)))
2089 return n * sizeof(u64);
2093 * Read the performance event - simple non blocking version for now
2096 perf_read_hw(struct perf_event *event, char __user *buf, size_t count)
2098 u64 read_format = event->attr.read_format;
2102 * Return end-of-file for a read on a event that is in
2103 * error state (i.e. because it was pinned but it couldn't be
2104 * scheduled on to the CPU at some point).
2106 if (event->state == PERF_EVENT_STATE_ERROR)
2109 if (count < perf_event_read_size(event))
2112 WARN_ON_ONCE(event->ctx->parent_ctx);
2113 if (read_format & PERF_FORMAT_GROUP)
2114 ret = perf_event_read_group(event, read_format, buf);
2116 ret = perf_event_read_one(event, read_format, buf);
2122 perf_read(struct file *file, char __user *buf, size_t count, loff_t *ppos)
2124 struct perf_event *event = file->private_data;
2126 return perf_read_hw(event, buf, count);
2129 static unsigned int perf_poll(struct file *file, poll_table *wait)
2131 struct perf_event *event = file->private_data;
2132 struct perf_buffer *buffer;
2133 unsigned int events = POLL_HUP;
2136 buffer = rcu_dereference(event->buffer);
2138 events = atomic_xchg(&buffer->poll, 0);
2141 poll_wait(file, &event->waitq, wait);
2146 static void perf_event_reset(struct perf_event *event)
2148 (void)perf_event_read(event);
2149 local64_set(&event->count, 0);
2150 perf_event_update_userpage(event);
2154 * Holding the top-level event's child_mutex means that any
2155 * descendant process that has inherited this event will block
2156 * in sync_child_event if it goes to exit, thus satisfying the
2157 * task existence requirements of perf_event_enable/disable.
2159 static void perf_event_for_each_child(struct perf_event *event,
2160 void (*func)(struct perf_event *))
2162 struct perf_event *child;
2164 WARN_ON_ONCE(event->ctx->parent_ctx);
2165 mutex_lock(&event->child_mutex);
2167 list_for_each_entry(child, &event->child_list, child_list)
2169 mutex_unlock(&event->child_mutex);
2172 static void perf_event_for_each(struct perf_event *event,
2173 void (*func)(struct perf_event *))
2175 struct perf_event_context *ctx = event->ctx;
2176 struct perf_event *sibling;
2178 WARN_ON_ONCE(ctx->parent_ctx);
2179 mutex_lock(&ctx->mutex);
2180 event = event->group_leader;
2182 perf_event_for_each_child(event, func);
2184 list_for_each_entry(sibling, &event->sibling_list, group_entry)
2185 perf_event_for_each_child(event, func);
2186 mutex_unlock(&ctx->mutex);
2189 static int perf_event_period(struct perf_event *event, u64 __user *arg)
2191 struct perf_event_context *ctx = event->ctx;
2196 if (!event->attr.sample_period)
2199 size = copy_from_user(&value, arg, sizeof(value));
2200 if (size != sizeof(value))
2206 raw_spin_lock_irq(&ctx->lock);
2207 if (event->attr.freq) {
2208 if (value > sysctl_perf_event_sample_rate) {
2213 event->attr.sample_freq = value;
2215 event->attr.sample_period = value;
2216 event->hw.sample_period = value;
2219 raw_spin_unlock_irq(&ctx->lock);
2224 static const struct file_operations perf_fops;
2226 static struct perf_event *perf_fget_light(int fd, int *fput_needed)
2230 file = fget_light(fd, fput_needed);
2232 return ERR_PTR(-EBADF);
2234 if (file->f_op != &perf_fops) {
2235 fput_light(file, *fput_needed);
2237 return ERR_PTR(-EBADF);
2240 return file->private_data;
2243 static int perf_event_set_output(struct perf_event *event,
2244 struct perf_event *output_event);
2245 static int perf_event_set_filter(struct perf_event *event, void __user *arg);
2247 static long perf_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
2249 struct perf_event *event = file->private_data;
2250 void (*func)(struct perf_event *);
2254 case PERF_EVENT_IOC_ENABLE:
2255 func = perf_event_enable;
2257 case PERF_EVENT_IOC_DISABLE:
2258 func = perf_event_disable;
2260 case PERF_EVENT_IOC_RESET:
2261 func = perf_event_reset;
2264 case PERF_EVENT_IOC_REFRESH:
2265 return perf_event_refresh(event, arg);
2267 case PERF_EVENT_IOC_PERIOD:
2268 return perf_event_period(event, (u64 __user *)arg);
2270 case PERF_EVENT_IOC_SET_OUTPUT:
2272 struct perf_event *output_event = NULL;
2273 int fput_needed = 0;
2277 output_event = perf_fget_light(arg, &fput_needed);
2278 if (IS_ERR(output_event))
2279 return PTR_ERR(output_event);
2282 ret = perf_event_set_output(event, output_event);
2284 fput_light(output_event->filp, fput_needed);
2289 case PERF_EVENT_IOC_SET_FILTER:
2290 return perf_event_set_filter(event, (void __user *)arg);
2296 if (flags & PERF_IOC_FLAG_GROUP)
2297 perf_event_for_each(event, func);
2299 perf_event_for_each_child(event, func);
2304 int perf_event_task_enable(void)
2306 struct perf_event *event;
2308 mutex_lock(¤t->perf_event_mutex);
2309 list_for_each_entry(event, ¤t->perf_event_list, owner_entry)
2310 perf_event_for_each_child(event, perf_event_enable);
2311 mutex_unlock(¤t->perf_event_mutex);
2316 int perf_event_task_disable(void)
2318 struct perf_event *event;
2320 mutex_lock(¤t->perf_event_mutex);
2321 list_for_each_entry(event, ¤t->perf_event_list, owner_entry)
2322 perf_event_for_each_child(event, perf_event_disable);
2323 mutex_unlock(¤t->perf_event_mutex);
2328 #ifndef PERF_EVENT_INDEX_OFFSET
2329 # define PERF_EVENT_INDEX_OFFSET 0
2332 static int perf_event_index(struct perf_event *event)
2334 if (event->state != PERF_EVENT_STATE_ACTIVE)
2337 return event->hw.idx + 1 - PERF_EVENT_INDEX_OFFSET;
2341 * Callers need to ensure there can be no nesting of this function, otherwise
2342 * the seqlock logic goes bad. We can not serialize this because the arch
2343 * code calls this from NMI context.
2345 void perf_event_update_userpage(struct perf_event *event)
2347 struct perf_event_mmap_page *userpg;
2348 struct perf_buffer *buffer;
2351 buffer = rcu_dereference(event->buffer);
2355 userpg = buffer->user_page;
2358 * Disable preemption so as to not let the corresponding user-space
2359 * spin too long if we get preempted.
2364 userpg->index = perf_event_index(event);
2365 userpg->offset = perf_event_count(event);
2366 if (event->state == PERF_EVENT_STATE_ACTIVE)
2367 userpg->offset -= local64_read(&event->hw.prev_count);
2369 userpg->time_enabled = event->total_time_enabled +
2370 atomic64_read(&event->child_total_time_enabled);
2372 userpg->time_running = event->total_time_running +
2373 atomic64_read(&event->child_total_time_running);
2382 static unsigned long perf_data_size(struct perf_buffer *buffer);
2385 perf_buffer_init(struct perf_buffer *buffer, long watermark, int flags)
2387 long max_size = perf_data_size(buffer);
2390 buffer->watermark = min(max_size, watermark);
2392 if (!buffer->watermark)
2393 buffer->watermark = max_size / 2;
2395 if (flags & PERF_BUFFER_WRITABLE)
2396 buffer->writable = 1;
2398 atomic_set(&buffer->refcount, 1);
2401 #ifndef CONFIG_PERF_USE_VMALLOC
2404 * Back perf_mmap() with regular GFP_KERNEL-0 pages.
2407 static struct page *
2408 perf_mmap_to_page(struct perf_buffer *buffer, unsigned long pgoff)
2410 if (pgoff > buffer->nr_pages)
2414 return virt_to_page(buffer->user_page);
2416 return virt_to_page(buffer->data_pages[pgoff - 1]);
2419 static void *perf_mmap_alloc_page(int cpu)
2424 node = (cpu == -1) ? cpu : cpu_to_node(cpu);
2425 page = alloc_pages_node(node, GFP_KERNEL | __GFP_ZERO, 0);
2429 return page_address(page);
2432 static struct perf_buffer *
2433 perf_buffer_alloc(int nr_pages, long watermark, int cpu, int flags)
2435 struct perf_buffer *buffer;
2439 size = sizeof(struct perf_buffer);
2440 size += nr_pages * sizeof(void *);
2442 buffer = kzalloc(size, GFP_KERNEL);
2446 buffer->user_page = perf_mmap_alloc_page(cpu);
2447 if (!buffer->user_page)
2448 goto fail_user_page;
2450 for (i = 0; i < nr_pages; i++) {
2451 buffer->data_pages[i] = perf_mmap_alloc_page(cpu);
2452 if (!buffer->data_pages[i])
2453 goto fail_data_pages;
2456 buffer->nr_pages = nr_pages;
2458 perf_buffer_init(buffer, watermark, flags);
2463 for (i--; i >= 0; i--)
2464 free_page((unsigned long)buffer->data_pages[i]);
2466 free_page((unsigned long)buffer->user_page);
2475 static void perf_mmap_free_page(unsigned long addr)
2477 struct page *page = virt_to_page((void *)addr);
2479 page->mapping = NULL;
2483 static void perf_buffer_free(struct perf_buffer *buffer)
2487 perf_mmap_free_page((unsigned long)buffer->user_page);
2488 for (i = 0; i < buffer->nr_pages; i++)
2489 perf_mmap_free_page((unsigned long)buffer->data_pages[i]);
2493 static inline int page_order(struct perf_buffer *buffer)
2501 * Back perf_mmap() with vmalloc memory.
2503 * Required for architectures that have d-cache aliasing issues.
2506 static inline int page_order(struct perf_buffer *buffer)
2508 return buffer->page_order;
2511 static struct page *
2512 perf_mmap_to_page(struct perf_buffer *buffer, unsigned long pgoff)
2514 if (pgoff > (1UL << page_order(buffer)))
2517 return vmalloc_to_page((void *)buffer->user_page + pgoff * PAGE_SIZE);
2520 static void perf_mmap_unmark_page(void *addr)
2522 struct page *page = vmalloc_to_page(addr);
2524 page->mapping = NULL;
2527 static void perf_buffer_free_work(struct work_struct *work)
2529 struct perf_buffer *buffer;
2533 buffer = container_of(work, struct perf_buffer, work);
2534 nr = 1 << page_order(buffer);
2536 base = buffer->user_page;
2537 for (i = 0; i < nr + 1; i++)
2538 perf_mmap_unmark_page(base + (i * PAGE_SIZE));
2544 static void perf_buffer_free(struct perf_buffer *buffer)
2546 schedule_work(&buffer->work);
2549 static struct perf_buffer *
2550 perf_buffer_alloc(int nr_pages, long watermark, int cpu, int flags)
2552 struct perf_buffer *buffer;
2556 size = sizeof(struct perf_buffer);
2557 size += sizeof(void *);
2559 buffer = kzalloc(size, GFP_KERNEL);
2563 INIT_WORK(&buffer->work, perf_buffer_free_work);
2565 all_buf = vmalloc_user((nr_pages + 1) * PAGE_SIZE);
2569 buffer->user_page = all_buf;
2570 buffer->data_pages[0] = all_buf + PAGE_SIZE;
2571 buffer->page_order = ilog2(nr_pages);
2572 buffer->nr_pages = 1;
2574 perf_buffer_init(buffer, watermark, flags);
2587 static unsigned long perf_data_size(struct perf_buffer *buffer)
2589 return buffer->nr_pages << (PAGE_SHIFT + page_order(buffer));
2592 static int perf_mmap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
2594 struct perf_event *event = vma->vm_file->private_data;
2595 struct perf_buffer *buffer;
2596 int ret = VM_FAULT_SIGBUS;
2598 if (vmf->flags & FAULT_FLAG_MKWRITE) {
2599 if (vmf->pgoff == 0)
2605 buffer = rcu_dereference(event->buffer);
2609 if (vmf->pgoff && (vmf->flags & FAULT_FLAG_WRITE))
2612 vmf->page = perf_mmap_to_page(buffer, vmf->pgoff);
2616 get_page(vmf->page);
2617 vmf->page->mapping = vma->vm_file->f_mapping;
2618 vmf->page->index = vmf->pgoff;
2627 static void perf_buffer_free_rcu(struct rcu_head *rcu_head)
2629 struct perf_buffer *buffer;
2631 buffer = container_of(rcu_head, struct perf_buffer, rcu_head);
2632 perf_buffer_free(buffer);
2635 static struct perf_buffer *perf_buffer_get(struct perf_event *event)
2637 struct perf_buffer *buffer;
2640 buffer = rcu_dereference(event->buffer);
2642 if (!atomic_inc_not_zero(&buffer->refcount))
2650 static void perf_buffer_put(struct perf_buffer *buffer)
2652 if (!atomic_dec_and_test(&buffer->refcount))
2655 call_rcu(&buffer->rcu_head, perf_buffer_free_rcu);
2658 static void perf_mmap_open(struct vm_area_struct *vma)
2660 struct perf_event *event = vma->vm_file->private_data;
2662 atomic_inc(&event->mmap_count);
2665 static void perf_mmap_close(struct vm_area_struct *vma)
2667 struct perf_event *event = vma->vm_file->private_data;
2669 if (atomic_dec_and_mutex_lock(&event->mmap_count, &event->mmap_mutex)) {
2670 unsigned long size = perf_data_size(event->buffer);
2671 struct user_struct *user = event->mmap_user;
2672 struct perf_buffer *buffer = event->buffer;
2674 atomic_long_sub((size >> PAGE_SHIFT) + 1, &user->locked_vm);
2675 vma->vm_mm->locked_vm -= event->mmap_locked;
2676 rcu_assign_pointer(event->buffer, NULL);
2677 mutex_unlock(&event->mmap_mutex);
2679 perf_buffer_put(buffer);
2684 static const struct vm_operations_struct perf_mmap_vmops = {
2685 .open = perf_mmap_open,
2686 .close = perf_mmap_close,
2687 .fault = perf_mmap_fault,
2688 .page_mkwrite = perf_mmap_fault,
2691 static int perf_mmap(struct file *file, struct vm_area_struct *vma)
2693 struct perf_event *event = file->private_data;
2694 unsigned long user_locked, user_lock_limit;
2695 struct user_struct *user = current_user();
2696 unsigned long locked, lock_limit;
2697 struct perf_buffer *buffer;
2698 unsigned long vma_size;
2699 unsigned long nr_pages;
2700 long user_extra, extra;
2701 int ret = 0, flags = 0;
2704 * Don't allow mmap() of inherited per-task counters. This would
2705 * create a performance issue due to all children writing to the
2708 if (event->cpu == -1 && event->attr.inherit)
2711 if (!(vma->vm_flags & VM_SHARED))
2714 vma_size = vma->vm_end - vma->vm_start;
2715 nr_pages = (vma_size / PAGE_SIZE) - 1;
2718 * If we have buffer pages ensure they're a power-of-two number, so we
2719 * can do bitmasks instead of modulo.
2721 if (nr_pages != 0 && !is_power_of_2(nr_pages))
2724 if (vma_size != PAGE_SIZE * (1 + nr_pages))
2727 if (vma->vm_pgoff != 0)
2730 WARN_ON_ONCE(event->ctx->parent_ctx);
2731 mutex_lock(&event->mmap_mutex);
2732 if (event->buffer) {
2733 if (event->buffer->nr_pages == nr_pages)
2734 atomic_inc(&event->buffer->refcount);
2740 user_extra = nr_pages + 1;
2741 user_lock_limit = sysctl_perf_event_mlock >> (PAGE_SHIFT - 10);
2744 * Increase the limit linearly with more CPUs:
2746 user_lock_limit *= num_online_cpus();
2748 user_locked = atomic_long_read(&user->locked_vm) + user_extra;
2751 if (user_locked > user_lock_limit)
2752 extra = user_locked - user_lock_limit;
2754 lock_limit = rlimit(RLIMIT_MEMLOCK);
2755 lock_limit >>= PAGE_SHIFT;
2756 locked = vma->vm_mm->locked_vm + extra;
2758 if ((locked > lock_limit) && perf_paranoid_tracepoint_raw() &&
2759 !capable(CAP_IPC_LOCK)) {
2764 WARN_ON(event->buffer);
2766 if (vma->vm_flags & VM_WRITE)
2767 flags |= PERF_BUFFER_WRITABLE;
2769 buffer = perf_buffer_alloc(nr_pages, event->attr.wakeup_watermark,
2775 rcu_assign_pointer(event->buffer, buffer);
2777 atomic_long_add(user_extra, &user->locked_vm);
2778 event->mmap_locked = extra;
2779 event->mmap_user = get_current_user();
2780 vma->vm_mm->locked_vm += event->mmap_locked;
2784 atomic_inc(&event->mmap_count);
2785 mutex_unlock(&event->mmap_mutex);
2787 vma->vm_flags |= VM_RESERVED;
2788 vma->vm_ops = &perf_mmap_vmops;
2793 static int perf_fasync(int fd, struct file *filp, int on)
2795 struct inode *inode = filp->f_path.dentry->d_inode;
2796 struct perf_event *event = filp->private_data;
2799 mutex_lock(&inode->i_mutex);
2800 retval = fasync_helper(fd, filp, on, &event->fasync);
2801 mutex_unlock(&inode->i_mutex);
2809 static const struct file_operations perf_fops = {
2810 .llseek = no_llseek,
2811 .release = perf_release,
2814 .unlocked_ioctl = perf_ioctl,
2815 .compat_ioctl = perf_ioctl,
2817 .fasync = perf_fasync,
2823 * If there's data, ensure we set the poll() state and publish everything
2824 * to user-space before waking everybody up.
2827 void perf_event_wakeup(struct perf_event *event)
2829 wake_up_all(&event->waitq);
2831 if (event->pending_kill) {
2832 kill_fasync(&event->fasync, SIGIO, event->pending_kill);
2833 event->pending_kill = 0;
2840 * Handle the case where we need to wakeup up from NMI (or rq->lock) context.
2842 * The NMI bit means we cannot possibly take locks. Therefore, maintain a
2843 * single linked list and use cmpxchg() to add entries lockless.
2846 static void perf_pending_event(struct perf_pending_entry *entry)
2848 struct perf_event *event = container_of(entry,
2849 struct perf_event, pending);
2851 if (event->pending_disable) {
2852 event->pending_disable = 0;
2853 __perf_event_disable(event);
2856 if (event->pending_wakeup) {
2857 event->pending_wakeup = 0;
2858 perf_event_wakeup(event);
2862 #define PENDING_TAIL ((struct perf_pending_entry *)-1UL)
2864 static DEFINE_PER_CPU(struct perf_pending_entry *, perf_pending_head) = {
2868 static void perf_pending_queue(struct perf_pending_entry *entry,
2869 void (*func)(struct perf_pending_entry *))
2871 struct perf_pending_entry **head;
2873 if (cmpxchg(&entry->next, NULL, PENDING_TAIL) != NULL)
2878 head = &get_cpu_var(perf_pending_head);
2881 entry->next = *head;
2882 } while (cmpxchg(head, entry->next, entry) != entry->next);
2884 set_perf_event_pending();
2886 put_cpu_var(perf_pending_head);
2889 static int __perf_pending_run(void)
2891 struct perf_pending_entry *list;
2894 list = xchg(&__get_cpu_var(perf_pending_head), PENDING_TAIL);
2895 while (list != PENDING_TAIL) {
2896 void (*func)(struct perf_pending_entry *);
2897 struct perf_pending_entry *entry = list;
2904 * Ensure we observe the unqueue before we issue the wakeup,
2905 * so that we won't be waiting forever.
2906 * -- see perf_not_pending().
2917 static inline int perf_not_pending(struct perf_event *event)
2920 * If we flush on whatever cpu we run, there is a chance we don't
2924 __perf_pending_run();
2928 * Ensure we see the proper queue state before going to sleep
2929 * so that we do not miss the wakeup. -- see perf_pending_handle()
2932 return event->pending.next == NULL;
2935 static void perf_pending_sync(struct perf_event *event)
2937 wait_event(event->waitq, perf_not_pending(event));
2940 void perf_event_do_pending(void)
2942 __perf_pending_run();
2946 * Callchain support -- arch specific
2949 __weak struct perf_callchain_entry *perf_callchain(struct pt_regs *regs)
2956 * We assume there is only KVM supporting the callbacks.
2957 * Later on, we might change it to a list if there is
2958 * another virtualization implementation supporting the callbacks.
2960 struct perf_guest_info_callbacks *perf_guest_cbs;
2962 int perf_register_guest_info_callbacks(struct perf_guest_info_callbacks *cbs)
2964 perf_guest_cbs = cbs;
2967 EXPORT_SYMBOL_GPL(perf_register_guest_info_callbacks);
2969 int perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks *cbs)
2971 perf_guest_cbs = NULL;
2974 EXPORT_SYMBOL_GPL(perf_unregister_guest_info_callbacks);
2979 static bool perf_output_space(struct perf_buffer *buffer, unsigned long tail,
2980 unsigned long offset, unsigned long head)
2984 if (!buffer->writable)
2987 mask = perf_data_size(buffer) - 1;
2989 offset = (offset - tail) & mask;
2990 head = (head - tail) & mask;
2992 if ((int)(head - offset) < 0)
2998 static void perf_output_wakeup(struct perf_output_handle *handle)
3000 atomic_set(&handle->buffer->poll, POLL_IN);
3003 handle->event->pending_wakeup = 1;
3004 perf_pending_queue(&handle->event->pending,
3005 perf_pending_event);
3007 perf_event_wakeup(handle->event);
3011 * We need to ensure a later event_id doesn't publish a head when a former
3012 * event isn't done writing. However since we need to deal with NMIs we
3013 * cannot fully serialize things.
3015 * We only publish the head (and generate a wakeup) when the outer-most
3018 static void perf_output_get_handle(struct perf_output_handle *handle)
3020 struct perf_buffer *buffer = handle->buffer;
3023 local_inc(&buffer->nest);
3024 handle->wakeup = local_read(&buffer->wakeup);
3027 static void perf_output_put_handle(struct perf_output_handle *handle)
3029 struct perf_buffer *buffer = handle->buffer;
3033 head = local_read(&buffer->head);
3036 * IRQ/NMI can happen here, which means we can miss a head update.
3039 if (!local_dec_and_test(&buffer->nest))
3043 * Publish the known good head. Rely on the full barrier implied
3044 * by atomic_dec_and_test() order the buffer->head read and this
3047 buffer->user_page->data_head = head;
3050 * Now check if we missed an update, rely on the (compiler)
3051 * barrier in atomic_dec_and_test() to re-read buffer->head.
3053 if (unlikely(head != local_read(&buffer->head))) {
3054 local_inc(&buffer->nest);
3058 if (handle->wakeup != local_read(&buffer->wakeup))
3059 perf_output_wakeup(handle);
3065 __always_inline void perf_output_copy(struct perf_output_handle *handle,
3066 const void *buf, unsigned int len)
3069 unsigned long size = min_t(unsigned long, handle->size, len);
3071 memcpy(handle->addr, buf, size);
3074 handle->addr += size;
3076 handle->size -= size;
3077 if (!handle->size) {
3078 struct perf_buffer *buffer = handle->buffer;
3081 handle->page &= buffer->nr_pages - 1;
3082 handle->addr = buffer->data_pages[handle->page];
3083 handle->size = PAGE_SIZE << page_order(buffer);
3088 int perf_output_begin(struct perf_output_handle *handle,
3089 struct perf_event *event, unsigned int size,
3090 int nmi, int sample)
3092 struct perf_buffer *buffer;
3093 unsigned long tail, offset, head;
3096 struct perf_event_header header;
3103 * For inherited events we send all the output towards the parent.
3106 event = event->parent;
3108 buffer = rcu_dereference(event->buffer);
3112 handle->buffer = buffer;
3113 handle->event = event;
3115 handle->sample = sample;
3117 if (!buffer->nr_pages)
3120 have_lost = local_read(&buffer->lost);
3122 size += sizeof(lost_event);
3124 perf_output_get_handle(handle);
3128 * Userspace could choose to issue a mb() before updating the
3129 * tail pointer. So that all reads will be completed before the
3132 tail = ACCESS_ONCE(buffer->user_page->data_tail);
3134 offset = head = local_read(&buffer->head);
3136 if (unlikely(!perf_output_space(buffer, tail, offset, head)))
3138 } while (local_cmpxchg(&buffer->head, offset, head) != offset);
3140 if (head - local_read(&buffer->wakeup) > buffer->watermark)
3141 local_add(buffer->watermark, &buffer->wakeup);
3143 handle->page = offset >> (PAGE_SHIFT + page_order(buffer));
3144 handle->page &= buffer->nr_pages - 1;
3145 handle->size = offset & ((PAGE_SIZE << page_order(buffer)) - 1);
3146 handle->addr = buffer->data_pages[handle->page];
3147 handle->addr += handle->size;
3148 handle->size = (PAGE_SIZE << page_order(buffer)) - handle->size;
3151 lost_event.header.type = PERF_RECORD_LOST;
3152 lost_event.header.misc = 0;
3153 lost_event.header.size = sizeof(lost_event);
3154 lost_event.id = event->id;
3155 lost_event.lost = local_xchg(&buffer->lost, 0);
3157 perf_output_put(handle, lost_event);
3163 local_inc(&buffer->lost);
3164 perf_output_put_handle(handle);
3171 void perf_output_end(struct perf_output_handle *handle)
3173 struct perf_event *event = handle->event;
3174 struct perf_buffer *buffer = handle->buffer;
3176 int wakeup_events = event->attr.wakeup_events;
3178 if (handle->sample && wakeup_events) {
3179 int events = local_inc_return(&buffer->events);
3180 if (events >= wakeup_events) {
3181 local_sub(wakeup_events, &buffer->events);
3182 local_inc(&buffer->wakeup);
3186 perf_output_put_handle(handle);
3190 static u32 perf_event_pid(struct perf_event *event, struct task_struct *p)
3193 * only top level events have the pid namespace they were created in
3196 event = event->parent;
3198 return task_tgid_nr_ns(p, event->ns);
3201 static u32 perf_event_tid(struct perf_event *event, struct task_struct *p)
3204 * only top level events have the pid namespace they were created in
3207 event = event->parent;
3209 return task_pid_nr_ns(p, event->ns);
3212 static void perf_output_read_one(struct perf_output_handle *handle,
3213 struct perf_event *event)
3215 u64 read_format = event->attr.read_format;
3219 values[n++] = perf_event_count(event);
3220 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
3221 values[n++] = event->total_time_enabled +
3222 atomic64_read(&event->child_total_time_enabled);
3224 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
3225 values[n++] = event->total_time_running +
3226 atomic64_read(&event->child_total_time_running);
3228 if (read_format & PERF_FORMAT_ID)
3229 values[n++] = primary_event_id(event);
3231 perf_output_copy(handle, values, n * sizeof(u64));
3235 * XXX PERF_FORMAT_GROUP vs inherited events seems difficult.
3237 static void perf_output_read_group(struct perf_output_handle *handle,
3238 struct perf_event *event)
3240 struct perf_event *leader = event->group_leader, *sub;
3241 u64 read_format = event->attr.read_format;
3245 values[n++] = 1 + leader->nr_siblings;
3247 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
3248 values[n++] = leader->total_time_enabled;
3250 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
3251 values[n++] = leader->total_time_running;
3253 if (leader != event)
3254 leader->pmu->read(leader);
3256 values[n++] = perf_event_count(leader);
3257 if (read_format & PERF_FORMAT_ID)
3258 values[n++] = primary_event_id(leader);
3260 perf_output_copy(handle, values, n * sizeof(u64));
3262 list_for_each_entry(sub, &leader->sibling_list, group_entry) {
3266 sub->pmu->read(sub);
3268 values[n++] = perf_event_count(sub);
3269 if (read_format & PERF_FORMAT_ID)
3270 values[n++] = primary_event_id(sub);
3272 perf_output_copy(handle, values, n * sizeof(u64));
3276 static void perf_output_read(struct perf_output_handle *handle,
3277 struct perf_event *event)
3279 if (event->attr.read_format & PERF_FORMAT_GROUP)
3280 perf_output_read_group(handle, event);
3282 perf_output_read_one(handle, event);
3285 void perf_output_sample(struct perf_output_handle *handle,
3286 struct perf_event_header *header,
3287 struct perf_sample_data *data,
3288 struct perf_event *event)
3290 u64 sample_type = data->type;
3292 perf_output_put(handle, *header);
3294 if (sample_type & PERF_SAMPLE_IP)
3295 perf_output_put(handle, data->ip);
3297 if (sample_type & PERF_SAMPLE_TID)
3298 perf_output_put(handle, data->tid_entry);
3300 if (sample_type & PERF_SAMPLE_TIME)
3301 perf_output_put(handle, data->time);
3303 if (sample_type & PERF_SAMPLE_ADDR)
3304 perf_output_put(handle, data->addr);
3306 if (sample_type & PERF_SAMPLE_ID)
3307 perf_output_put(handle, data->id);
3309 if (sample_type & PERF_SAMPLE_STREAM_ID)
3310 perf_output_put(handle, data->stream_id);
3312 if (sample_type & PERF_SAMPLE_CPU)
3313 perf_output_put(handle, data->cpu_entry);
3315 if (sample_type & PERF_SAMPLE_PERIOD)
3316 perf_output_put(handle, data->period);
3318 if (sample_type & PERF_SAMPLE_READ)
3319 perf_output_read(handle, event);
3321 if (sample_type & PERF_SAMPLE_CALLCHAIN) {
3322 if (data->callchain) {
3325 if (data->callchain)
3326 size += data->callchain->nr;
3328 size *= sizeof(u64);
3330 perf_output_copy(handle, data->callchain, size);
3333 perf_output_put(handle, nr);
3337 if (sample_type & PERF_SAMPLE_RAW) {
3339 perf_output_put(handle, data->raw->size);
3340 perf_output_copy(handle, data->raw->data,
3347 .size = sizeof(u32),
3350 perf_output_put(handle, raw);
3355 void perf_prepare_sample(struct perf_event_header *header,
3356 struct perf_sample_data *data,
3357 struct perf_event *event,
3358 struct pt_regs *regs)
3360 u64 sample_type = event->attr.sample_type;
3362 data->type = sample_type;
3364 header->type = PERF_RECORD_SAMPLE;
3365 header->size = sizeof(*header);
3368 header->misc |= perf_misc_flags(regs);
3370 if (sample_type & PERF_SAMPLE_IP) {
3371 data->ip = perf_instruction_pointer(regs);
3373 header->size += sizeof(data->ip);
3376 if (sample_type & PERF_SAMPLE_TID) {
3377 /* namespace issues */
3378 data->tid_entry.pid = perf_event_pid(event, current);
3379 data->tid_entry.tid = perf_event_tid(event, current);
3381 header->size += sizeof(data->tid_entry);
3384 if (sample_type & PERF_SAMPLE_TIME) {
3385 data->time = perf_clock();
3387 header->size += sizeof(data->time);
3390 if (sample_type & PERF_SAMPLE_ADDR)
3391 header->size += sizeof(data->addr);
3393 if (sample_type & PERF_SAMPLE_ID) {
3394 data->id = primary_event_id(event);
3396 header->size += sizeof(data->id);
3399 if (sample_type & PERF_SAMPLE_STREAM_ID) {
3400 data->stream_id = event->id;
3402 header->size += sizeof(data->stream_id);
3405 if (sample_type & PERF_SAMPLE_CPU) {
3406 data->cpu_entry.cpu = raw_smp_processor_id();
3407 data->cpu_entry.reserved = 0;
3409 header->size += sizeof(data->cpu_entry);
3412 if (sample_type & PERF_SAMPLE_PERIOD)
3413 header->size += sizeof(data->period);
3415 if (sample_type & PERF_SAMPLE_READ)
3416 header->size += perf_event_read_size(event);
3418 if (sample_type & PERF_SAMPLE_CALLCHAIN) {
3421 data->callchain = perf_callchain(regs);
3423 if (data->callchain)
3424 size += data->callchain->nr;
3426 header->size += size * sizeof(u64);
3429 if (sample_type & PERF_SAMPLE_RAW) {
3430 int size = sizeof(u32);
3433 size += data->raw->size;
3435 size += sizeof(u32);
3437 WARN_ON_ONCE(size & (sizeof(u64)-1));
3438 header->size += size;
3442 static void perf_event_output(struct perf_event *event, int nmi,
3443 struct perf_sample_data *data,
3444 struct pt_regs *regs)
3446 struct perf_output_handle handle;
3447 struct perf_event_header header;
3449 perf_prepare_sample(&header, data, event, regs);
3451 if (perf_output_begin(&handle, event, header.size, nmi, 1))
3454 perf_output_sample(&handle, &header, data, event);
3456 perf_output_end(&handle);
3463 struct perf_read_event {
3464 struct perf_event_header header;
3471 perf_event_read_event(struct perf_event *event,
3472 struct task_struct *task)
3474 struct perf_output_handle handle;
3475 struct perf_read_event read_event = {
3477 .type = PERF_RECORD_READ,
3479 .size = sizeof(read_event) + perf_event_read_size(event),
3481 .pid = perf_event_pid(event, task),
3482 .tid = perf_event_tid(event, task),
3486 ret = perf_output_begin(&handle, event, read_event.header.size, 0, 0);
3490 perf_output_put(&handle, read_event);
3491 perf_output_read(&handle, event);
3493 perf_output_end(&handle);
3497 * task tracking -- fork/exit
3499 * enabled by: attr.comm | attr.mmap | attr.mmap_data | attr.task
3502 struct perf_task_event {
3503 struct task_struct *task;
3504 struct perf_event_context *task_ctx;
3507 struct perf_event_header header;
3517 static void perf_event_task_output(struct perf_event *event,
3518 struct perf_task_event *task_event)
3520 struct perf_output_handle handle;
3521 struct task_struct *task = task_event->task;
3524 size = task_event->event_id.header.size;
3525 ret = perf_output_begin(&handle, event, size, 0, 0);
3530 task_event->event_id.pid = perf_event_pid(event, task);
3531 task_event->event_id.ppid = perf_event_pid(event, current);
3533 task_event->event_id.tid = perf_event_tid(event, task);
3534 task_event->event_id.ptid = perf_event_tid(event, current);
3536 perf_output_put(&handle, task_event->event_id);
3538 perf_output_end(&handle);
3541 static int perf_event_task_match(struct perf_event *event)
3543 if (event->state < PERF_EVENT_STATE_INACTIVE)
3546 if (event->cpu != -1 && event->cpu != smp_processor_id())
3549 if (event->attr.comm || event->attr.mmap ||
3550 event->attr.mmap_data || event->attr.task)
3556 static void perf_event_task_ctx(struct perf_event_context *ctx,
3557 struct perf_task_event *task_event)
3559 struct perf_event *event;
3561 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
3562 if (perf_event_task_match(event))
3563 perf_event_task_output(event, task_event);
3567 static void perf_event_task_event(struct perf_task_event *task_event)
3569 struct perf_cpu_context *cpuctx;
3570 struct perf_event_context *ctx = task_event->task_ctx;
3573 cpuctx = &get_cpu_var(perf_cpu_context);
3574 perf_event_task_ctx(&cpuctx->ctx, task_event);
3576 ctx = rcu_dereference(current->perf_event_ctxp);
3578 perf_event_task_ctx(ctx, task_event);
3579 put_cpu_var(perf_cpu_context);
3583 static void perf_event_task(struct task_struct *task,
3584 struct perf_event_context *task_ctx,
3587 struct perf_task_event task_event;
3589 if (!atomic_read(&nr_comm_events) &&
3590 !atomic_read(&nr_mmap_events) &&
3591 !atomic_read(&nr_task_events))
3594 task_event = (struct perf_task_event){
3596 .task_ctx = task_ctx,
3599 .type = new ? PERF_RECORD_FORK : PERF_RECORD_EXIT,
3601 .size = sizeof(task_event.event_id),
3607 .time = perf_clock(),
3611 perf_event_task_event(&task_event);
3614 void perf_event_fork(struct task_struct *task)
3616 perf_event_task(task, NULL, 1);
3623 struct perf_comm_event {
3624 struct task_struct *task;
3629 struct perf_event_header header;
3636 static void perf_event_comm_output(struct perf_event *event,
3637 struct perf_comm_event *comm_event)
3639 struct perf_output_handle handle;
3640 int size = comm_event->event_id.header.size;
3641 int ret = perf_output_begin(&handle, event, size, 0, 0);
3646 comm_event->event_id.pid = perf_event_pid(event, comm_event->task);
3647 comm_event->event_id.tid = perf_event_tid(event, comm_event->task);
3649 perf_output_put(&handle, comm_event->event_id);
3650 perf_output_copy(&handle, comm_event->comm,
3651 comm_event->comm_size);
3652 perf_output_end(&handle);
3655 static int perf_event_comm_match(struct perf_event *event)
3657 if (event->state < PERF_EVENT_STATE_INACTIVE)
3660 if (event->cpu != -1 && event->cpu != smp_processor_id())
3663 if (event->attr.comm)
3669 static void perf_event_comm_ctx(struct perf_event_context *ctx,
3670 struct perf_comm_event *comm_event)
3672 struct perf_event *event;
3674 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
3675 if (perf_event_comm_match(event))
3676 perf_event_comm_output(event, comm_event);
3680 static void perf_event_comm_event(struct perf_comm_event *comm_event)
3682 struct perf_cpu_context *cpuctx;
3683 struct perf_event_context *ctx;
3685 char comm[TASK_COMM_LEN];
3687 memset(comm, 0, sizeof(comm));
3688 strlcpy(comm, comm_event->task->comm, sizeof(comm));
3689 size = ALIGN(strlen(comm)+1, sizeof(u64));
3691 comm_event->comm = comm;
3692 comm_event->comm_size = size;
3694 comm_event->event_id.header.size = sizeof(comm_event->event_id) + size;
3697 cpuctx = &get_cpu_var(perf_cpu_context);
3698 perf_event_comm_ctx(&cpuctx->ctx, comm_event);
3699 ctx = rcu_dereference(current->perf_event_ctxp);
3701 perf_event_comm_ctx(ctx, comm_event);
3702 put_cpu_var(perf_cpu_context);
3706 void perf_event_comm(struct task_struct *task)
3708 struct perf_comm_event comm_event;
3710 if (task->perf_event_ctxp)
3711 perf_event_enable_on_exec(task);
3713 if (!atomic_read(&nr_comm_events))
3716 comm_event = (struct perf_comm_event){
3722 .type = PERF_RECORD_COMM,
3731 perf_event_comm_event(&comm_event);
3738 struct perf_mmap_event {
3739 struct vm_area_struct *vma;
3741 const char *file_name;
3745 struct perf_event_header header;
3755 static void perf_event_mmap_output(struct perf_event *event,
3756 struct perf_mmap_event *mmap_event)
3758 struct perf_output_handle handle;
3759 int size = mmap_event->event_id.header.size;
3760 int ret = perf_output_begin(&handle, event, size, 0, 0);
3765 mmap_event->event_id.pid = perf_event_pid(event, current);
3766 mmap_event->event_id.tid = perf_event_tid(event, current);
3768 perf_output_put(&handle, mmap_event->event_id);
3769 perf_output_copy(&handle, mmap_event->file_name,
3770 mmap_event->file_size);
3771 perf_output_end(&handle);
3774 static int perf_event_mmap_match(struct perf_event *event,
3775 struct perf_mmap_event *mmap_event,
3778 if (event->state < PERF_EVENT_STATE_INACTIVE)
3781 if (event->cpu != -1 && event->cpu != smp_processor_id())
3784 if ((!executable && event->attr.mmap_data) ||
3785 (executable && event->attr.mmap))
3791 static void perf_event_mmap_ctx(struct perf_event_context *ctx,
3792 struct perf_mmap_event *mmap_event,
3795 struct perf_event *event;
3797 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
3798 if (perf_event_mmap_match(event, mmap_event, executable))
3799 perf_event_mmap_output(event, mmap_event);
3803 static void perf_event_mmap_event(struct perf_mmap_event *mmap_event)
3805 struct perf_cpu_context *cpuctx;
3806 struct perf_event_context *ctx;
3807 struct vm_area_struct *vma = mmap_event->vma;
3808 struct file *file = vma->vm_file;
3814 memset(tmp, 0, sizeof(tmp));
3818 * d_path works from the end of the buffer backwards, so we
3819 * need to add enough zero bytes after the string to handle
3820 * the 64bit alignment we do later.
3822 buf = kzalloc(PATH_MAX + sizeof(u64), GFP_KERNEL);
3824 name = strncpy(tmp, "//enomem", sizeof(tmp));
3827 name = d_path(&file->f_path, buf, PATH_MAX);
3829 name = strncpy(tmp, "//toolong", sizeof(tmp));
3833 if (arch_vma_name(mmap_event->vma)) {
3834 name = strncpy(tmp, arch_vma_name(mmap_event->vma),
3840 name = strncpy(tmp, "[vdso]", sizeof(tmp));
3842 } else if (vma->vm_start <= vma->vm_mm->start_brk &&
3843 vma->vm_end >= vma->vm_mm->brk) {
3844 name = strncpy(tmp, "[heap]", sizeof(tmp));
3846 } else if (vma->vm_start <= vma->vm_mm->start_stack &&
3847 vma->vm_end >= vma->vm_mm->start_stack) {
3848 name = strncpy(tmp, "[stack]", sizeof(tmp));
3852 name = strncpy(tmp, "//anon", sizeof(tmp));
3857 size = ALIGN(strlen(name)+1, sizeof(u64));
3859 mmap_event->file_name = name;
3860 mmap_event->file_size = size;
3862 mmap_event->event_id.header.size = sizeof(mmap_event->event_id) + size;
3865 cpuctx = &get_cpu_var(perf_cpu_context);
3866 perf_event_mmap_ctx(&cpuctx->ctx, mmap_event, vma->vm_flags & VM_EXEC);
3867 ctx = rcu_dereference(current->perf_event_ctxp);
3869 perf_event_mmap_ctx(ctx, mmap_event, vma->vm_flags & VM_EXEC);
3870 put_cpu_var(perf_cpu_context);
3876 void perf_event_mmap(struct vm_area_struct *vma)
3878 struct perf_mmap_event mmap_event;
3880 if (!atomic_read(&nr_mmap_events))
3883 mmap_event = (struct perf_mmap_event){
3889 .type = PERF_RECORD_MMAP,
3890 .misc = PERF_RECORD_MISC_USER,
3895 .start = vma->vm_start,
3896 .len = vma->vm_end - vma->vm_start,
3897 .pgoff = (u64)vma->vm_pgoff << PAGE_SHIFT,
3901 perf_event_mmap_event(&mmap_event);
3905 * IRQ throttle logging
3908 static void perf_log_throttle(struct perf_event *event, int enable)
3910 struct perf_output_handle handle;
3914 struct perf_event_header header;
3918 } throttle_event = {
3920 .type = PERF_RECORD_THROTTLE,
3922 .size = sizeof(throttle_event),
3924 .time = perf_clock(),
3925 .id = primary_event_id(event),
3926 .stream_id = event->id,
3930 throttle_event.header.type = PERF_RECORD_UNTHROTTLE;
3932 ret = perf_output_begin(&handle, event, sizeof(throttle_event), 1, 0);
3936 perf_output_put(&handle, throttle_event);
3937 perf_output_end(&handle);
3941 * Generic event overflow handling, sampling.
3944 static int __perf_event_overflow(struct perf_event *event, int nmi,
3945 int throttle, struct perf_sample_data *data,
3946 struct pt_regs *regs)
3948 int events = atomic_read(&event->event_limit);
3949 struct hw_perf_event *hwc = &event->hw;
3952 throttle = (throttle && event->pmu->unthrottle != NULL);
3957 if (hwc->interrupts != MAX_INTERRUPTS) {
3959 if (HZ * hwc->interrupts >
3960 (u64)sysctl_perf_event_sample_rate) {
3961 hwc->interrupts = MAX_INTERRUPTS;
3962 perf_log_throttle(event, 0);
3967 * Keep re-disabling events even though on the previous
3968 * pass we disabled it - just in case we raced with a
3969 * sched-in and the event got enabled again:
3975 if (event->attr.freq) {
3976 u64 now = perf_clock();
3977 s64 delta = now - hwc->freq_time_stamp;
3979 hwc->freq_time_stamp = now;
3981 if (delta > 0 && delta < 2*TICK_NSEC)
3982 perf_adjust_period(event, delta, hwc->last_period);
3986 * XXX event_limit might not quite work as expected on inherited
3990 event->pending_kill = POLL_IN;
3991 if (events && atomic_dec_and_test(&event->event_limit)) {
3993 event->pending_kill = POLL_HUP;
3995 event->pending_disable = 1;
3996 perf_pending_queue(&event->pending,
3997 perf_pending_event);
3999 perf_event_disable(event);
4002 if (event->overflow_handler)
4003 event->overflow_handler(event, nmi, data, regs);
4005 perf_event_output(event, nmi, data, regs);
4010 int perf_event_overflow(struct perf_event *event, int nmi,
4011 struct perf_sample_data *data,
4012 struct pt_regs *regs)
4014 return __perf_event_overflow(event, nmi, 1, data, regs);
4018 * Generic software event infrastructure
4022 * We directly increment event->count and keep a second value in
4023 * event->hw.period_left to count intervals. This period event
4024 * is kept in the range [-sample_period, 0] so that we can use the
4028 static u64 perf_swevent_set_period(struct perf_event *event)
4030 struct hw_perf_event *hwc = &event->hw;
4031 u64 period = hwc->last_period;
4035 hwc->last_period = hwc->sample_period;
4038 old = val = local64_read(&hwc->period_left);
4042 nr = div64_u64(period + val, period);
4043 offset = nr * period;
4045 if (local64_cmpxchg(&hwc->period_left, old, val) != old)
4051 static void perf_swevent_overflow(struct perf_event *event, u64 overflow,
4052 int nmi, struct perf_sample_data *data,
4053 struct pt_regs *regs)
4055 struct hw_perf_event *hwc = &event->hw;
4058 data->period = event->hw.last_period;
4060 overflow = perf_swevent_set_period(event);
4062 if (hwc->interrupts == MAX_INTERRUPTS)
4065 for (; overflow; overflow--) {
4066 if (__perf_event_overflow(event, nmi, throttle,
4069 * We inhibit the overflow from happening when
4070 * hwc->interrupts == MAX_INTERRUPTS.
4078 static void perf_swevent_add(struct perf_event *event, u64 nr,
4079 int nmi, struct perf_sample_data *data,
4080 struct pt_regs *regs)
4082 struct hw_perf_event *hwc = &event->hw;
4084 local64_add(nr, &event->count);
4089 if (!hwc->sample_period)
4092 if (nr == 1 && hwc->sample_period == 1 && !event->attr.freq)
4093 return perf_swevent_overflow(event, 1, nmi, data, regs);
4095 if (local64_add_negative(nr, &hwc->period_left))
4098 perf_swevent_overflow(event, 0, nmi, data, regs);
4101 static int perf_exclude_event(struct perf_event *event,
4102 struct pt_regs *regs)
4105 if (event->attr.exclude_user && user_mode(regs))
4108 if (event->attr.exclude_kernel && !user_mode(regs))
4115 static int perf_swevent_match(struct perf_event *event,
4116 enum perf_type_id type,
4118 struct perf_sample_data *data,
4119 struct pt_regs *regs)
4121 if (event->attr.type != type)
4124 if (event->attr.config != event_id)
4127 if (perf_exclude_event(event, regs))
4133 static inline u64 swevent_hash(u64 type, u32 event_id)
4135 u64 val = event_id | (type << 32);
4137 return hash_64(val, SWEVENT_HLIST_BITS);
4140 static inline struct hlist_head *
4141 __find_swevent_head(struct swevent_hlist *hlist, u64 type, u32 event_id)
4143 u64 hash = swevent_hash(type, event_id);
4145 return &hlist->heads[hash];
4148 /* For the read side: events when they trigger */
4149 static inline struct hlist_head *
4150 find_swevent_head_rcu(struct perf_cpu_context *ctx, u64 type, u32 event_id)
4152 struct swevent_hlist *hlist;
4154 hlist = rcu_dereference(ctx->swevent_hlist);
4158 return __find_swevent_head(hlist, type, event_id);
4161 /* For the event head insertion and removal in the hlist */
4162 static inline struct hlist_head *
4163 find_swevent_head(struct perf_cpu_context *ctx, struct perf_event *event)
4165 struct swevent_hlist *hlist;
4166 u32 event_id = event->attr.config;
4167 u64 type = event->attr.type;
4170 * Event scheduling is always serialized against hlist allocation
4171 * and release. Which makes the protected version suitable here.
4172 * The context lock guarantees that.
4174 hlist = rcu_dereference_protected(ctx->swevent_hlist,
4175 lockdep_is_held(&event->ctx->lock));
4179 return __find_swevent_head(hlist, type, event_id);
4182 static void do_perf_sw_event(enum perf_type_id type, u32 event_id,
4184 struct perf_sample_data *data,
4185 struct pt_regs *regs)
4187 struct perf_cpu_context *cpuctx;
4188 struct perf_event *event;
4189 struct hlist_node *node;
4190 struct hlist_head *head;
4192 cpuctx = &__get_cpu_var(perf_cpu_context);
4196 head = find_swevent_head_rcu(cpuctx, type, event_id);
4201 hlist_for_each_entry_rcu(event, node, head, hlist_entry) {
4202 if (perf_swevent_match(event, type, event_id, data, regs))
4203 perf_swevent_add(event, nr, nmi, data, regs);
4209 int perf_swevent_get_recursion_context(void)
4211 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
4218 else if (in_softirq())
4223 if (cpuctx->recursion[rctx])
4226 cpuctx->recursion[rctx]++;
4231 EXPORT_SYMBOL_GPL(perf_swevent_get_recursion_context);
4233 void inline perf_swevent_put_recursion_context(int rctx)
4235 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
4237 cpuctx->recursion[rctx]--;
4240 void __perf_sw_event(u32 event_id, u64 nr, int nmi,
4241 struct pt_regs *regs, u64 addr)
4243 struct perf_sample_data data;
4246 preempt_disable_notrace();
4247 rctx = perf_swevent_get_recursion_context();
4251 perf_sample_data_init(&data, addr);
4253 do_perf_sw_event(PERF_TYPE_SOFTWARE, event_id, nr, nmi, &data, regs);
4255 perf_swevent_put_recursion_context(rctx);
4256 preempt_enable_notrace();
4259 static void perf_swevent_read(struct perf_event *event)
4263 static int perf_swevent_enable(struct perf_event *event)
4265 struct hw_perf_event *hwc = &event->hw;
4266 struct perf_cpu_context *cpuctx;
4267 struct hlist_head *head;
4269 cpuctx = &__get_cpu_var(perf_cpu_context);
4271 if (hwc->sample_period) {
4272 hwc->last_period = hwc->sample_period;
4273 perf_swevent_set_period(event);
4276 head = find_swevent_head(cpuctx, event);
4277 if (WARN_ON_ONCE(!head))
4280 hlist_add_head_rcu(&event->hlist_entry, head);
4285 static void perf_swevent_disable(struct perf_event *event)
4287 hlist_del_rcu(&event->hlist_entry);
4290 static void perf_swevent_void(struct perf_event *event)
4294 static int perf_swevent_int(struct perf_event *event)
4299 static const struct pmu perf_ops_generic = {
4300 .enable = perf_swevent_enable,
4301 .disable = perf_swevent_disable,
4302 .start = perf_swevent_int,
4303 .stop = perf_swevent_void,
4304 .read = perf_swevent_read,
4305 .unthrottle = perf_swevent_void, /* hwc->interrupts already reset */
4309 * hrtimer based swevent callback
4312 static enum hrtimer_restart perf_swevent_hrtimer(struct hrtimer *hrtimer)
4314 enum hrtimer_restart ret = HRTIMER_RESTART;
4315 struct perf_sample_data data;
4316 struct pt_regs *regs;
4317 struct perf_event *event;
4320 event = container_of(hrtimer, struct perf_event, hw.hrtimer);
4321 event->pmu->read(event);
4323 perf_sample_data_init(&data, 0);
4324 data.period = event->hw.last_period;
4325 regs = get_irq_regs();
4327 if (regs && !perf_exclude_event(event, regs)) {
4328 if (!(event->attr.exclude_idle && current->pid == 0))
4329 if (perf_event_overflow(event, 0, &data, regs))
4330 ret = HRTIMER_NORESTART;
4333 period = max_t(u64, 10000, event->hw.sample_period);
4334 hrtimer_forward_now(hrtimer, ns_to_ktime(period));
4339 static void perf_swevent_start_hrtimer(struct perf_event *event)
4341 struct hw_perf_event *hwc = &event->hw;
4343 hrtimer_init(&hwc->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
4344 hwc->hrtimer.function = perf_swevent_hrtimer;
4345 if (hwc->sample_period) {
4348 if (hwc->remaining) {
4349 if (hwc->remaining < 0)
4352 period = hwc->remaining;
4355 period = max_t(u64, 10000, hwc->sample_period);
4357 __hrtimer_start_range_ns(&hwc->hrtimer,
4358 ns_to_ktime(period), 0,
4359 HRTIMER_MODE_REL, 0);
4363 static void perf_swevent_cancel_hrtimer(struct perf_event *event)
4365 struct hw_perf_event *hwc = &event->hw;
4367 if (hwc->sample_period) {
4368 ktime_t remaining = hrtimer_get_remaining(&hwc->hrtimer);
4369 hwc->remaining = ktime_to_ns(remaining);
4371 hrtimer_cancel(&hwc->hrtimer);
4376 * Software event: cpu wall time clock
4379 static void cpu_clock_perf_event_update(struct perf_event *event)
4381 int cpu = raw_smp_processor_id();
4385 now = cpu_clock(cpu);
4386 prev = local64_xchg(&event->hw.prev_count, now);
4387 local64_add(now - prev, &event->count);
4390 static int cpu_clock_perf_event_enable(struct perf_event *event)
4392 struct hw_perf_event *hwc = &event->hw;
4393 int cpu = raw_smp_processor_id();
4395 local64_set(&hwc->prev_count, cpu_clock(cpu));
4396 perf_swevent_start_hrtimer(event);
4401 static void cpu_clock_perf_event_disable(struct perf_event *event)
4403 perf_swevent_cancel_hrtimer(event);
4404 cpu_clock_perf_event_update(event);
4407 static void cpu_clock_perf_event_read(struct perf_event *event)
4409 cpu_clock_perf_event_update(event);
4412 static const struct pmu perf_ops_cpu_clock = {
4413 .enable = cpu_clock_perf_event_enable,
4414 .disable = cpu_clock_perf_event_disable,
4415 .read = cpu_clock_perf_event_read,
4419 * Software event: task time clock
4422 static void task_clock_perf_event_update(struct perf_event *event, u64 now)
4427 prev = local64_xchg(&event->hw.prev_count, now);
4429 local64_add(delta, &event->count);
4432 static int task_clock_perf_event_enable(struct perf_event *event)
4434 struct hw_perf_event *hwc = &event->hw;
4437 now = event->ctx->time;
4439 local64_set(&hwc->prev_count, now);
4441 perf_swevent_start_hrtimer(event);
4446 static void task_clock_perf_event_disable(struct perf_event *event)
4448 perf_swevent_cancel_hrtimer(event);
4449 task_clock_perf_event_update(event, event->ctx->time);
4453 static void task_clock_perf_event_read(struct perf_event *event)
4458 update_context_time(event->ctx);
4459 time = event->ctx->time;
4461 u64 now = perf_clock();
4462 u64 delta = now - event->ctx->timestamp;
4463 time = event->ctx->time + delta;
4466 task_clock_perf_event_update(event, time);
4469 static const struct pmu perf_ops_task_clock = {
4470 .enable = task_clock_perf_event_enable,
4471 .disable = task_clock_perf_event_disable,
4472 .read = task_clock_perf_event_read,
4475 /* Deref the hlist from the update side */
4476 static inline struct swevent_hlist *
4477 swevent_hlist_deref(struct perf_cpu_context *cpuctx)
4479 return rcu_dereference_protected(cpuctx->swevent_hlist,
4480 lockdep_is_held(&cpuctx->hlist_mutex));
4483 static void swevent_hlist_release_rcu(struct rcu_head *rcu_head)
4485 struct swevent_hlist *hlist;
4487 hlist = container_of(rcu_head, struct swevent_hlist, rcu_head);
4491 static void swevent_hlist_release(struct perf_cpu_context *cpuctx)
4493 struct swevent_hlist *hlist = swevent_hlist_deref(cpuctx);
4498 rcu_assign_pointer(cpuctx->swevent_hlist, NULL);
4499 call_rcu(&hlist->rcu_head, swevent_hlist_release_rcu);
4502 static void swevent_hlist_put_cpu(struct perf_event *event, int cpu)
4504 struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
4506 mutex_lock(&cpuctx->hlist_mutex);
4508 if (!--cpuctx->hlist_refcount)
4509 swevent_hlist_release(cpuctx);
4511 mutex_unlock(&cpuctx->hlist_mutex);
4514 static void swevent_hlist_put(struct perf_event *event)
4518 if (event->cpu != -1) {
4519 swevent_hlist_put_cpu(event, event->cpu);
4523 for_each_possible_cpu(cpu)
4524 swevent_hlist_put_cpu(event, cpu);
4527 static int swevent_hlist_get_cpu(struct perf_event *event, int cpu)
4529 struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
4532 mutex_lock(&cpuctx->hlist_mutex);
4534 if (!swevent_hlist_deref(cpuctx) && cpu_online(cpu)) {
4535 struct swevent_hlist *hlist;
4537 hlist = kzalloc(sizeof(*hlist), GFP_KERNEL);
4542 rcu_assign_pointer(cpuctx->swevent_hlist, hlist);
4544 cpuctx->hlist_refcount++;
4546 mutex_unlock(&cpuctx->hlist_mutex);
4551 static int swevent_hlist_get(struct perf_event *event)
4554 int cpu, failed_cpu;
4556 if (event->cpu != -1)
4557 return swevent_hlist_get_cpu(event, event->cpu);
4560 for_each_possible_cpu(cpu) {
4561 err = swevent_hlist_get_cpu(event, cpu);
4571 for_each_possible_cpu(cpu) {
4572 if (cpu == failed_cpu)
4574 swevent_hlist_put_cpu(event, cpu);
4581 #ifdef CONFIG_EVENT_TRACING
4583 static const struct pmu perf_ops_tracepoint = {
4584 .enable = perf_trace_enable,
4585 .disable = perf_trace_disable,
4586 .start = perf_swevent_int,
4587 .stop = perf_swevent_void,
4588 .read = perf_swevent_read,
4589 .unthrottle = perf_swevent_void,
4592 static int perf_tp_filter_match(struct perf_event *event,
4593 struct perf_sample_data *data)
4595 void *record = data->raw->data;
4597 if (likely(!event->filter) || filter_match_preds(event->filter, record))
4602 static int perf_tp_event_match(struct perf_event *event,
4603 struct perf_sample_data *data,
4604 struct pt_regs *regs)
4607 * All tracepoints are from kernel-space.
4609 if (event->attr.exclude_kernel)
4612 if (!perf_tp_filter_match(event, data))
4618 void perf_tp_event(u64 addr, u64 count, void *record, int entry_size,
4619 struct pt_regs *regs, struct hlist_head *head, int rctx)
4621 struct perf_sample_data data;
4622 struct perf_event *event;
4623 struct hlist_node *node;
4625 struct perf_raw_record raw = {
4630 perf_sample_data_init(&data, addr);
4633 hlist_for_each_entry_rcu(event, node, head, hlist_entry) {
4634 if (perf_tp_event_match(event, &data, regs))
4635 perf_swevent_add(event, count, 1, &data, regs);
4638 perf_swevent_put_recursion_context(rctx);
4640 EXPORT_SYMBOL_GPL(perf_tp_event);
4642 static void tp_perf_event_destroy(struct perf_event *event)
4644 perf_trace_destroy(event);
4647 static const struct pmu *tp_perf_event_init(struct perf_event *event)
4652 * Raw tracepoint data is a severe data leak, only allow root to
4655 if ((event->attr.sample_type & PERF_SAMPLE_RAW) &&
4656 perf_paranoid_tracepoint_raw() &&
4657 !capable(CAP_SYS_ADMIN))
4658 return ERR_PTR(-EPERM);
4660 err = perf_trace_init(event);
4664 event->destroy = tp_perf_event_destroy;
4666 return &perf_ops_tracepoint;
4669 static int perf_event_set_filter(struct perf_event *event, void __user *arg)
4674 if (event->attr.type != PERF_TYPE_TRACEPOINT)
4677 filter_str = strndup_user(arg, PAGE_SIZE);
4678 if (IS_ERR(filter_str))
4679 return PTR_ERR(filter_str);
4681 ret = ftrace_profile_set_filter(event, event->attr.config, filter_str);
4687 static void perf_event_free_filter(struct perf_event *event)
4689 ftrace_profile_free_filter(event);
4694 static const struct pmu *tp_perf_event_init(struct perf_event *event)
4699 static int perf_event_set_filter(struct perf_event *event, void __user *arg)
4704 static void perf_event_free_filter(struct perf_event *event)
4708 #endif /* CONFIG_EVENT_TRACING */
4710 #ifdef CONFIG_HAVE_HW_BREAKPOINT
4711 static void bp_perf_event_destroy(struct perf_event *event)
4713 release_bp_slot(event);
4716 static const struct pmu *bp_perf_event_init(struct perf_event *bp)
4720 err = register_perf_hw_breakpoint(bp);
4722 return ERR_PTR(err);
4724 bp->destroy = bp_perf_event_destroy;
4726 return &perf_ops_bp;
4729 void perf_bp_event(struct perf_event *bp, void *data)
4731 struct perf_sample_data sample;
4732 struct pt_regs *regs = data;
4734 perf_sample_data_init(&sample, bp->attr.bp_addr);
4736 if (!perf_exclude_event(bp, regs))
4737 perf_swevent_add(bp, 1, 1, &sample, regs);
4740 static const struct pmu *bp_perf_event_init(struct perf_event *bp)
4745 void perf_bp_event(struct perf_event *bp, void *regs)
4750 atomic_t perf_swevent_enabled[PERF_COUNT_SW_MAX];
4752 static void sw_perf_event_destroy(struct perf_event *event)
4754 u64 event_id = event->attr.config;
4756 WARN_ON(event->parent);
4758 atomic_dec(&perf_swevent_enabled[event_id]);
4759 swevent_hlist_put(event);
4762 static const struct pmu *sw_perf_event_init(struct perf_event *event)
4764 const struct pmu *pmu = NULL;
4765 u64 event_id = event->attr.config;
4768 * Software events (currently) can't in general distinguish
4769 * between user, kernel and hypervisor events.
4770 * However, context switches and cpu migrations are considered
4771 * to be kernel events, and page faults are never hypervisor
4775 case PERF_COUNT_SW_CPU_CLOCK:
4776 pmu = &perf_ops_cpu_clock;
4779 case PERF_COUNT_SW_TASK_CLOCK:
4781 * If the user instantiates this as a per-cpu event,
4782 * use the cpu_clock event instead.
4784 if (event->ctx->task)
4785 pmu = &perf_ops_task_clock;
4787 pmu = &perf_ops_cpu_clock;
4790 case PERF_COUNT_SW_PAGE_FAULTS:
4791 case PERF_COUNT_SW_PAGE_FAULTS_MIN:
4792 case PERF_COUNT_SW_PAGE_FAULTS_MAJ:
4793 case PERF_COUNT_SW_CONTEXT_SWITCHES:
4794 case PERF_COUNT_SW_CPU_MIGRATIONS:
4795 case PERF_COUNT_SW_ALIGNMENT_FAULTS:
4796 case PERF_COUNT_SW_EMULATION_FAULTS:
4797 if (!event->parent) {
4800 err = swevent_hlist_get(event);
4802 return ERR_PTR(err);
4804 atomic_inc(&perf_swevent_enabled[event_id]);
4805 event->destroy = sw_perf_event_destroy;
4807 pmu = &perf_ops_generic;
4815 * Allocate and initialize a event structure
4817 static struct perf_event *
4818 perf_event_alloc(struct perf_event_attr *attr,
4820 struct perf_event_context *ctx,
4821 struct perf_event *group_leader,
4822 struct perf_event *parent_event,
4823 perf_overflow_handler_t overflow_handler,
4826 const struct pmu *pmu;
4827 struct perf_event *event;
4828 struct hw_perf_event *hwc;
4831 event = kzalloc(sizeof(*event), gfpflags);
4833 return ERR_PTR(-ENOMEM);
4836 * Single events are their own group leaders, with an
4837 * empty sibling list:
4840 group_leader = event;
4842 mutex_init(&event->child_mutex);
4843 INIT_LIST_HEAD(&event->child_list);
4845 INIT_LIST_HEAD(&event->group_entry);
4846 INIT_LIST_HEAD(&event->event_entry);
4847 INIT_LIST_HEAD(&event->sibling_list);
4848 init_waitqueue_head(&event->waitq);
4850 mutex_init(&event->mmap_mutex);
4853 event->attr = *attr;
4854 event->group_leader = group_leader;
4859 event->parent = parent_event;
4861 event->ns = get_pid_ns(current->nsproxy->pid_ns);
4862 event->id = atomic64_inc_return(&perf_event_id);
4864 event->state = PERF_EVENT_STATE_INACTIVE;
4866 if (!overflow_handler && parent_event)
4867 overflow_handler = parent_event->overflow_handler;
4869 event->overflow_handler = overflow_handler;
4872 event->state = PERF_EVENT_STATE_OFF;
4877 hwc->sample_period = attr->sample_period;
4878 if (attr->freq && attr->sample_freq)
4879 hwc->sample_period = 1;
4880 hwc->last_period = hwc->sample_period;
4882 local64_set(&hwc->period_left, hwc->sample_period);
4885 * we currently do not support PERF_FORMAT_GROUP on inherited events
4887 if (attr->inherit && (attr->read_format & PERF_FORMAT_GROUP))
4890 switch (attr->type) {
4892 case PERF_TYPE_HARDWARE:
4893 case PERF_TYPE_HW_CACHE:
4894 pmu = hw_perf_event_init(event);
4897 case PERF_TYPE_SOFTWARE:
4898 pmu = sw_perf_event_init(event);
4901 case PERF_TYPE_TRACEPOINT:
4902 pmu = tp_perf_event_init(event);
4905 case PERF_TYPE_BREAKPOINT:
4906 pmu = bp_perf_event_init(event);
4917 else if (IS_ERR(pmu))
4922 put_pid_ns(event->ns);
4924 return ERR_PTR(err);
4929 if (!event->parent) {
4930 atomic_inc(&nr_events);
4931 if (event->attr.mmap || event->attr.mmap_data)
4932 atomic_inc(&nr_mmap_events);
4933 if (event->attr.comm)
4934 atomic_inc(&nr_comm_events);
4935 if (event->attr.task)
4936 atomic_inc(&nr_task_events);
4942 static int perf_copy_attr(struct perf_event_attr __user *uattr,
4943 struct perf_event_attr *attr)
4948 if (!access_ok(VERIFY_WRITE, uattr, PERF_ATTR_SIZE_VER0))
4952 * zero the full structure, so that a short copy will be nice.
4954 memset(attr, 0, sizeof(*attr));
4956 ret = get_user(size, &uattr->size);
4960 if (size > PAGE_SIZE) /* silly large */
4963 if (!size) /* abi compat */
4964 size = PERF_ATTR_SIZE_VER0;
4966 if (size < PERF_ATTR_SIZE_VER0)
4970 * If we're handed a bigger struct than we know of,
4971 * ensure all the unknown bits are 0 - i.e. new
4972 * user-space does not rely on any kernel feature
4973 * extensions we dont know about yet.
4975 if (size > sizeof(*attr)) {
4976 unsigned char __user *addr;
4977 unsigned char __user *end;
4980 addr = (void __user *)uattr + sizeof(*attr);
4981 end = (void __user *)uattr + size;
4983 for (; addr < end; addr++) {
4984 ret = get_user(val, addr);
4990 size = sizeof(*attr);
4993 ret = copy_from_user(attr, uattr, size);
4998 * If the type exists, the corresponding creation will verify
5001 if (attr->type >= PERF_TYPE_MAX)
5004 if (attr->__reserved_1)
5007 if (attr->sample_type & ~(PERF_SAMPLE_MAX-1))
5010 if (attr->read_format & ~(PERF_FORMAT_MAX-1))
5017 put_user(sizeof(*attr), &uattr->size);
5023 perf_event_set_output(struct perf_event *event, struct perf_event *output_event)
5025 struct perf_buffer *buffer = NULL, *old_buffer = NULL;
5031 /* don't allow circular references */
5032 if (event == output_event)
5036 * Don't allow cross-cpu buffers
5038 if (output_event->cpu != event->cpu)
5042 * If its not a per-cpu buffer, it must be the same task.
5044 if (output_event->cpu == -1 && output_event->ctx != event->ctx)
5048 mutex_lock(&event->mmap_mutex);
5049 /* Can't redirect output if we've got an active mmap() */
5050 if (atomic_read(&event->mmap_count))
5054 /* get the buffer we want to redirect to */
5055 buffer = perf_buffer_get(output_event);
5060 old_buffer = event->buffer;
5061 rcu_assign_pointer(event->buffer, buffer);
5064 mutex_unlock(&event->mmap_mutex);
5067 perf_buffer_put(old_buffer);
5073 * sys_perf_event_open - open a performance event, associate it to a task/cpu
5075 * @attr_uptr: event_id type attributes for monitoring/sampling
5078 * @group_fd: group leader event fd
5080 SYSCALL_DEFINE5(perf_event_open,
5081 struct perf_event_attr __user *, attr_uptr,
5082 pid_t, pid, int, cpu, int, group_fd, unsigned long, flags)
5084 struct perf_event *event, *group_leader = NULL, *output_event = NULL;
5085 struct perf_event_attr attr;
5086 struct perf_event_context *ctx;
5087 struct file *event_file = NULL;
5088 struct file *group_file = NULL;
5090 int fput_needed = 0;
5093 /* for future expandability... */
5094 if (flags & ~(PERF_FLAG_FD_NO_GROUP | PERF_FLAG_FD_OUTPUT))
5097 err = perf_copy_attr(attr_uptr, &attr);
5101 if (!attr.exclude_kernel) {
5102 if (perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN))
5107 if (attr.sample_freq > sysctl_perf_event_sample_rate)
5111 event_fd = get_unused_fd_flags(O_RDWR);
5116 * Get the target context (task or percpu):
5118 ctx = find_get_context(pid, cpu);
5124 if (group_fd != -1) {
5125 group_leader = perf_fget_light(group_fd, &fput_needed);
5126 if (IS_ERR(group_leader)) {
5127 err = PTR_ERR(group_leader);
5128 goto err_put_context;
5130 group_file = group_leader->filp;
5131 if (flags & PERF_FLAG_FD_OUTPUT)
5132 output_event = group_leader;
5133 if (flags & PERF_FLAG_FD_NO_GROUP)
5134 group_leader = NULL;
5138 * Look up the group leader (we will attach this event to it):
5144 * Do not allow a recursive hierarchy (this new sibling
5145 * becoming part of another group-sibling):
5147 if (group_leader->group_leader != group_leader)
5148 goto err_put_context;
5150 * Do not allow to attach to a group in a different
5151 * task or CPU context:
5153 if (group_leader->ctx != ctx)
5154 goto err_put_context;
5156 * Only a group leader can be exclusive or pinned
5158 if (attr.exclusive || attr.pinned)
5159 goto err_put_context;
5162 event = perf_event_alloc(&attr, cpu, ctx, group_leader,
5163 NULL, NULL, GFP_KERNEL);
5164 if (IS_ERR(event)) {
5165 err = PTR_ERR(event);
5166 goto err_put_context;
5170 err = perf_event_set_output(event, output_event);
5172 goto err_free_put_context;
5175 event_file = anon_inode_getfile("[perf_event]", &perf_fops, event, O_RDWR);
5176 if (IS_ERR(event_file)) {
5177 err = PTR_ERR(event_file);
5178 goto err_free_put_context;
5181 event->filp = event_file;
5182 WARN_ON_ONCE(ctx->parent_ctx);
5183 mutex_lock(&ctx->mutex);
5184 perf_install_in_context(ctx, event, cpu);
5186 mutex_unlock(&ctx->mutex);
5188 event->owner = current;
5189 get_task_struct(current);
5190 mutex_lock(¤t->perf_event_mutex);
5191 list_add_tail(&event->owner_entry, ¤t->perf_event_list);
5192 mutex_unlock(¤t->perf_event_mutex);
5195 * Drop the reference on the group_event after placing the
5196 * new event on the sibling_list. This ensures destruction
5197 * of the group leader will find the pointer to itself in
5198 * perf_group_detach().
5200 fput_light(group_file, fput_needed);
5201 fd_install(event_fd, event_file);
5204 err_free_put_context:
5207 fput_light(group_file, fput_needed);
5210 put_unused_fd(event_fd);
5215 * perf_event_create_kernel_counter
5217 * @attr: attributes of the counter to create
5218 * @cpu: cpu in which the counter is bound
5219 * @pid: task to profile
5222 perf_event_create_kernel_counter(struct perf_event_attr *attr, int cpu,
5224 perf_overflow_handler_t overflow_handler)
5226 struct perf_event *event;
5227 struct perf_event_context *ctx;
5231 * Get the target context (task or percpu):
5234 ctx = find_get_context(pid, cpu);
5240 event = perf_event_alloc(attr, cpu, ctx, NULL,
5241 NULL, overflow_handler, GFP_KERNEL);
5242 if (IS_ERR(event)) {
5243 err = PTR_ERR(event);
5244 goto err_put_context;
5248 WARN_ON_ONCE(ctx->parent_ctx);
5249 mutex_lock(&ctx->mutex);
5250 perf_install_in_context(ctx, event, cpu);
5252 mutex_unlock(&ctx->mutex);
5254 event->owner = current;
5255 get_task_struct(current);
5256 mutex_lock(¤t->perf_event_mutex);
5257 list_add_tail(&event->owner_entry, ¤t->perf_event_list);
5258 mutex_unlock(¤t->perf_event_mutex);
5265 return ERR_PTR(err);
5267 EXPORT_SYMBOL_GPL(perf_event_create_kernel_counter);
5270 * inherit a event from parent task to child task:
5272 static struct perf_event *
5273 inherit_event(struct perf_event *parent_event,
5274 struct task_struct *parent,
5275 struct perf_event_context *parent_ctx,
5276 struct task_struct *child,
5277 struct perf_event *group_leader,
5278 struct perf_event_context *child_ctx)
5280 struct perf_event *child_event;
5283 * Instead of creating recursive hierarchies of events,
5284 * we link inherited events back to the original parent,
5285 * which has a filp for sure, which we use as the reference
5288 if (parent_event->parent)
5289 parent_event = parent_event->parent;
5291 child_event = perf_event_alloc(&parent_event->attr,
5292 parent_event->cpu, child_ctx,
5293 group_leader, parent_event,
5295 if (IS_ERR(child_event))
5300 * Make the child state follow the state of the parent event,
5301 * not its attr.disabled bit. We hold the parent's mutex,
5302 * so we won't race with perf_event_{en, dis}able_family.
5304 if (parent_event->state >= PERF_EVENT_STATE_INACTIVE)
5305 child_event->state = PERF_EVENT_STATE_INACTIVE;
5307 child_event->state = PERF_EVENT_STATE_OFF;
5309 if (parent_event->attr.freq) {
5310 u64 sample_period = parent_event->hw.sample_period;
5311 struct hw_perf_event *hwc = &child_event->hw;
5313 hwc->sample_period = sample_period;
5314 hwc->last_period = sample_period;
5316 local64_set(&hwc->period_left, sample_period);
5319 child_event->overflow_handler = parent_event->overflow_handler;
5322 * Link it up in the child's context:
5324 add_event_to_ctx(child_event, child_ctx);
5327 * Get a reference to the parent filp - we will fput it
5328 * when the child event exits. This is safe to do because
5329 * we are in the parent and we know that the filp still
5330 * exists and has a nonzero count:
5332 atomic_long_inc(&parent_event->filp->f_count);
5335 * Link this into the parent event's child list
5337 WARN_ON_ONCE(parent_event->ctx->parent_ctx);
5338 mutex_lock(&parent_event->child_mutex);
5339 list_add_tail(&child_event->child_list, &parent_event->child_list);
5340 mutex_unlock(&parent_event->child_mutex);
5345 static int inherit_group(struct perf_event *parent_event,
5346 struct task_struct *parent,
5347 struct perf_event_context *parent_ctx,
5348 struct task_struct *child,
5349 struct perf_event_context *child_ctx)
5351 struct perf_event *leader;
5352 struct perf_event *sub;
5353 struct perf_event *child_ctr;
5355 leader = inherit_event(parent_event, parent, parent_ctx,
5356 child, NULL, child_ctx);
5358 return PTR_ERR(leader);
5359 list_for_each_entry(sub, &parent_event->sibling_list, group_entry) {
5360 child_ctr = inherit_event(sub, parent, parent_ctx,
5361 child, leader, child_ctx);
5362 if (IS_ERR(child_ctr))
5363 return PTR_ERR(child_ctr);
5368 static void sync_child_event(struct perf_event *child_event,
5369 struct task_struct *child)
5371 struct perf_event *parent_event = child_event->parent;
5374 if (child_event->attr.inherit_stat)
5375 perf_event_read_event(child_event, child);
5377 child_val = perf_event_count(child_event);
5380 * Add back the child's count to the parent's count:
5382 atomic64_add(child_val, &parent_event->child_count);
5383 atomic64_add(child_event->total_time_enabled,
5384 &parent_event->child_total_time_enabled);
5385 atomic64_add(child_event->total_time_running,
5386 &parent_event->child_total_time_running);
5389 * Remove this event from the parent's list
5391 WARN_ON_ONCE(parent_event->ctx->parent_ctx);
5392 mutex_lock(&parent_event->child_mutex);
5393 list_del_init(&child_event->child_list);
5394 mutex_unlock(&parent_event->child_mutex);
5397 * Release the parent event, if this was the last
5400 fput(parent_event->filp);
5404 __perf_event_exit_task(struct perf_event *child_event,
5405 struct perf_event_context *child_ctx,
5406 struct task_struct *child)
5408 struct perf_event *parent_event;
5410 perf_event_remove_from_context(child_event);
5412 parent_event = child_event->parent;
5414 * It can happen that parent exits first, and has events
5415 * that are still around due to the child reference. These
5416 * events need to be zapped - but otherwise linger.
5419 sync_child_event(child_event, child);
5420 free_event(child_event);
5425 * When a child task exits, feed back event values to parent events.
5427 void perf_event_exit_task(struct task_struct *child)
5429 struct perf_event *child_event, *tmp;
5430 struct perf_event_context *child_ctx;
5431 unsigned long flags;
5433 if (likely(!child->perf_event_ctxp)) {
5434 perf_event_task(child, NULL, 0);
5438 local_irq_save(flags);
5440 * We can't reschedule here because interrupts are disabled,
5441 * and either child is current or it is a task that can't be
5442 * scheduled, so we are now safe from rescheduling changing
5445 child_ctx = child->perf_event_ctxp;
5446 __perf_event_task_sched_out(child_ctx);
5449 * Take the context lock here so that if find_get_context is
5450 * reading child->perf_event_ctxp, we wait until it has
5451 * incremented the context's refcount before we do put_ctx below.
5453 raw_spin_lock(&child_ctx->lock);
5454 child->perf_event_ctxp = NULL;
5456 * If this context is a clone; unclone it so it can't get
5457 * swapped to another process while we're removing all
5458 * the events from it.
5460 unclone_ctx(child_ctx);
5461 update_context_time(child_ctx);
5462 raw_spin_unlock_irqrestore(&child_ctx->lock, flags);
5465 * Report the task dead after unscheduling the events so that we
5466 * won't get any samples after PERF_RECORD_EXIT. We can however still
5467 * get a few PERF_RECORD_READ events.
5469 perf_event_task(child, child_ctx, 0);
5472 * We can recurse on the same lock type through:
5474 * __perf_event_exit_task()
5475 * sync_child_event()
5476 * fput(parent_event->filp)
5478 * mutex_lock(&ctx->mutex)
5480 * But since its the parent context it won't be the same instance.
5482 mutex_lock(&child_ctx->mutex);
5485 list_for_each_entry_safe(child_event, tmp, &child_ctx->pinned_groups,
5487 __perf_event_exit_task(child_event, child_ctx, child);
5489 list_for_each_entry_safe(child_event, tmp, &child_ctx->flexible_groups,
5491 __perf_event_exit_task(child_event, child_ctx, child);
5494 * If the last event was a group event, it will have appended all
5495 * its siblings to the list, but we obtained 'tmp' before that which
5496 * will still point to the list head terminating the iteration.
5498 if (!list_empty(&child_ctx->pinned_groups) ||
5499 !list_empty(&child_ctx->flexible_groups))
5502 mutex_unlock(&child_ctx->mutex);
5507 static void perf_free_event(struct perf_event *event,
5508 struct perf_event_context *ctx)
5510 struct perf_event *parent = event->parent;
5512 if (WARN_ON_ONCE(!parent))
5515 mutex_lock(&parent->child_mutex);
5516 list_del_init(&event->child_list);
5517 mutex_unlock(&parent->child_mutex);
5521 perf_group_detach(event);
5522 list_del_event(event, ctx);
5527 * free an unexposed, unused context as created by inheritance by
5528 * init_task below, used by fork() in case of fail.
5530 void perf_event_free_task(struct task_struct *task)
5532 struct perf_event_context *ctx = task->perf_event_ctxp;
5533 struct perf_event *event, *tmp;
5538 mutex_lock(&ctx->mutex);
5540 list_for_each_entry_safe(event, tmp, &ctx->pinned_groups, group_entry)
5541 perf_free_event(event, ctx);
5543 list_for_each_entry_safe(event, tmp, &ctx->flexible_groups,
5545 perf_free_event(event, ctx);
5547 if (!list_empty(&ctx->pinned_groups) ||
5548 !list_empty(&ctx->flexible_groups))
5551 mutex_unlock(&ctx->mutex);
5557 inherit_task_group(struct perf_event *event, struct task_struct *parent,
5558 struct perf_event_context *parent_ctx,
5559 struct task_struct *child,
5563 struct perf_event_context *child_ctx = child->perf_event_ctxp;
5565 if (!event->attr.inherit) {
5572 * This is executed from the parent task context, so
5573 * inherit events that have been marked for cloning.
5574 * First allocate and initialize a context for the
5578 child_ctx = kzalloc(sizeof(struct perf_event_context),
5583 __perf_event_init_context(child_ctx, child);
5584 child->perf_event_ctxp = child_ctx;
5585 get_task_struct(child);
5588 ret = inherit_group(event, parent, parent_ctx,
5599 * Initialize the perf_event context in task_struct
5601 int perf_event_init_task(struct task_struct *child)
5603 struct perf_event_context *child_ctx, *parent_ctx;
5604 struct perf_event_context *cloned_ctx;
5605 struct perf_event *event;
5606 struct task_struct *parent = current;
5607 int inherited_all = 1;
5610 child->perf_event_ctxp = NULL;
5612 mutex_init(&child->perf_event_mutex);
5613 INIT_LIST_HEAD(&child->perf_event_list);
5615 if (likely(!parent->perf_event_ctxp))
5619 * If the parent's context is a clone, pin it so it won't get
5622 parent_ctx = perf_pin_task_context(parent);
5625 * No need to check if parent_ctx != NULL here; since we saw
5626 * it non-NULL earlier, the only reason for it to become NULL
5627 * is if we exit, and since we're currently in the middle of
5628 * a fork we can't be exiting at the same time.
5632 * Lock the parent list. No need to lock the child - not PID
5633 * hashed yet and not running, so nobody can access it.
5635 mutex_lock(&parent_ctx->mutex);
5638 * We dont have to disable NMIs - we are only looking at
5639 * the list, not manipulating it:
5641 list_for_each_entry(event, &parent_ctx->pinned_groups, group_entry) {
5642 ret = inherit_task_group(event, parent, parent_ctx, child,
5648 list_for_each_entry(event, &parent_ctx->flexible_groups, group_entry) {
5649 ret = inherit_task_group(event, parent, parent_ctx, child,
5655 child_ctx = child->perf_event_ctxp;
5657 if (child_ctx && inherited_all) {
5659 * Mark the child context as a clone of the parent
5660 * context, or of whatever the parent is a clone of.
5661 * Note that if the parent is a clone, it could get
5662 * uncloned at any point, but that doesn't matter
5663 * because the list of events and the generation
5664 * count can't have changed since we took the mutex.
5666 cloned_ctx = rcu_dereference(parent_ctx->parent_ctx);
5668 child_ctx->parent_ctx = cloned_ctx;
5669 child_ctx->parent_gen = parent_ctx->parent_gen;
5671 child_ctx->parent_ctx = parent_ctx;
5672 child_ctx->parent_gen = parent_ctx->generation;
5674 get_ctx(child_ctx->parent_ctx);
5677 mutex_unlock(&parent_ctx->mutex);
5679 perf_unpin_context(parent_ctx);
5684 static void __init perf_event_init_all_cpus(void)
5687 struct perf_cpu_context *cpuctx;
5689 for_each_possible_cpu(cpu) {
5690 cpuctx = &per_cpu(perf_cpu_context, cpu);
5691 mutex_init(&cpuctx->hlist_mutex);
5692 __perf_event_init_context(&cpuctx->ctx, NULL);
5696 static void __cpuinit perf_event_init_cpu(int cpu)
5698 struct perf_cpu_context *cpuctx;
5700 cpuctx = &per_cpu(perf_cpu_context, cpu);
5702 spin_lock(&perf_resource_lock);
5703 cpuctx->max_pertask = perf_max_events - perf_reserved_percpu;
5704 spin_unlock(&perf_resource_lock);
5706 mutex_lock(&cpuctx->hlist_mutex);
5707 if (cpuctx->hlist_refcount > 0) {
5708 struct swevent_hlist *hlist;
5710 hlist = kzalloc(sizeof(*hlist), GFP_KERNEL);
5711 WARN_ON_ONCE(!hlist);
5712 rcu_assign_pointer(cpuctx->swevent_hlist, hlist);
5714 mutex_unlock(&cpuctx->hlist_mutex);
5717 #ifdef CONFIG_HOTPLUG_CPU
5718 static void __perf_event_exit_cpu(void *info)
5720 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
5721 struct perf_event_context *ctx = &cpuctx->ctx;
5722 struct perf_event *event, *tmp;
5724 list_for_each_entry_safe(event, tmp, &ctx->pinned_groups, group_entry)
5725 __perf_event_remove_from_context(event);
5726 list_for_each_entry_safe(event, tmp, &ctx->flexible_groups, group_entry)
5727 __perf_event_remove_from_context(event);
5729 static void perf_event_exit_cpu(int cpu)
5731 struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
5732 struct perf_event_context *ctx = &cpuctx->ctx;
5734 mutex_lock(&cpuctx->hlist_mutex);
5735 swevent_hlist_release(cpuctx);
5736 mutex_unlock(&cpuctx->hlist_mutex);
5738 mutex_lock(&ctx->mutex);
5739 smp_call_function_single(cpu, __perf_event_exit_cpu, NULL, 1);
5740 mutex_unlock(&ctx->mutex);
5743 static inline void perf_event_exit_cpu(int cpu) { }
5746 static int __cpuinit
5747 perf_cpu_notify(struct notifier_block *self, unsigned long action, void *hcpu)
5749 unsigned int cpu = (long)hcpu;
5753 case CPU_UP_PREPARE:
5754 case CPU_UP_PREPARE_FROZEN:
5755 perf_event_init_cpu(cpu);
5758 case CPU_DOWN_PREPARE:
5759 case CPU_DOWN_PREPARE_FROZEN:
5760 perf_event_exit_cpu(cpu);
5771 * This has to have a higher priority than migration_notifier in sched.c.
5773 static struct notifier_block __cpuinitdata perf_cpu_nb = {
5774 .notifier_call = perf_cpu_notify,
5778 void __init perf_event_init(void)
5780 perf_event_init_all_cpus();
5781 perf_cpu_notify(&perf_cpu_nb, (unsigned long)CPU_UP_PREPARE,
5782 (void *)(long)smp_processor_id());
5783 perf_cpu_notify(&perf_cpu_nb, (unsigned long)CPU_ONLINE,
5784 (void *)(long)smp_processor_id());
5785 register_cpu_notifier(&perf_cpu_nb);
5788 static ssize_t perf_show_reserve_percpu(struct sysdev_class *class,
5789 struct sysdev_class_attribute *attr,
5792 return sprintf(buf, "%d\n", perf_reserved_percpu);
5796 perf_set_reserve_percpu(struct sysdev_class *class,
5797 struct sysdev_class_attribute *attr,
5801 struct perf_cpu_context *cpuctx;
5805 err = strict_strtoul(buf, 10, &val);
5808 if (val > perf_max_events)
5811 spin_lock(&perf_resource_lock);
5812 perf_reserved_percpu = val;
5813 for_each_online_cpu(cpu) {
5814 cpuctx = &per_cpu(perf_cpu_context, cpu);
5815 raw_spin_lock_irq(&cpuctx->ctx.lock);
5816 mpt = min(perf_max_events - cpuctx->ctx.nr_events,
5817 perf_max_events - perf_reserved_percpu);
5818 cpuctx->max_pertask = mpt;
5819 raw_spin_unlock_irq(&cpuctx->ctx.lock);
5821 spin_unlock(&perf_resource_lock);
5826 static ssize_t perf_show_overcommit(struct sysdev_class *class,
5827 struct sysdev_class_attribute *attr,
5830 return sprintf(buf, "%d\n", perf_overcommit);
5834 perf_set_overcommit(struct sysdev_class *class,
5835 struct sysdev_class_attribute *attr,
5836 const char *buf, size_t count)
5841 err = strict_strtoul(buf, 10, &val);
5847 spin_lock(&perf_resource_lock);
5848 perf_overcommit = val;
5849 spin_unlock(&perf_resource_lock);
5854 static SYSDEV_CLASS_ATTR(
5857 perf_show_reserve_percpu,
5858 perf_set_reserve_percpu
5861 static SYSDEV_CLASS_ATTR(
5864 perf_show_overcommit,
5868 static struct attribute *perfclass_attrs[] = {
5869 &attr_reserve_percpu.attr,
5870 &attr_overcommit.attr,
5874 static struct attribute_group perfclass_attr_group = {
5875 .attrs = perfclass_attrs,
5876 .name = "perf_events",
5879 static int __init perf_event_sysfs_init(void)
5881 return sysfs_create_group(&cpu_sysdev_class.kset.kobj,
5882 &perfclass_attr_group);
5884 device_initcall(perf_event_sysfs_init);