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/idr.h>
17 #include <linux/file.h>
18 #include <linux/poll.h>
19 #include <linux/slab.h>
20 #include <linux/hash.h>
21 #include <linux/sysfs.h>
22 #include <linux/dcache.h>
23 #include <linux/percpu.h>
24 #include <linux/ptrace.h>
25 #include <linux/reboot.h>
26 #include <linux/vmstat.h>
27 #include <linux/device.h>
28 #include <linux/vmalloc.h>
29 #include <linux/hardirq.h>
30 #include <linux/rculist.h>
31 #include <linux/uaccess.h>
32 #include <linux/syscalls.h>
33 #include <linux/anon_inodes.h>
34 #include <linux/kernel_stat.h>
35 #include <linux/perf_event.h>
36 #include <linux/ftrace_event.h>
37 #include <linux/hw_breakpoint.h>
39 #include <asm/irq_regs.h>
44 EVENT_ALL = EVENT_FLEXIBLE | EVENT_PINNED,
47 atomic_t perf_task_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;
52 static LIST_HEAD(pmus);
53 static DEFINE_MUTEX(pmus_lock);
54 static struct srcu_struct pmus_srcu;
57 * perf event paranoia level:
58 * -1 - not paranoid at all
59 * 0 - disallow raw tracepoint access for unpriv
60 * 1 - disallow cpu events for unpriv
61 * 2 - disallow kernel profiling for unpriv
63 int sysctl_perf_event_paranoid __read_mostly = 1;
65 int sysctl_perf_event_mlock __read_mostly = 512; /* 'free' kb per user */
68 * max perf event sample rate
70 int sysctl_perf_event_sample_rate __read_mostly = 100000;
72 static atomic64_t perf_event_id;
74 static void cpu_ctx_sched_out(struct perf_cpu_context *cpuctx,
75 enum event_type_t event_type);
77 static void cpu_ctx_sched_in(struct perf_cpu_context *cpuctx,
78 enum event_type_t event_type);
80 void __weak perf_event_print_debug(void) { }
82 extern __weak const char *perf_pmu_name(void)
87 static inline u64 perf_clock(void)
92 void perf_pmu_disable(struct pmu *pmu)
94 int *count = this_cpu_ptr(pmu->pmu_disable_count);
96 pmu->pmu_disable(pmu);
99 void perf_pmu_enable(struct pmu *pmu)
101 int *count = this_cpu_ptr(pmu->pmu_disable_count);
103 pmu->pmu_enable(pmu);
106 static DEFINE_PER_CPU(struct list_head, rotation_list);
109 * perf_pmu_rotate_start() and perf_rotate_context() are fully serialized
110 * because they're strictly cpu affine and rotate_start is called with IRQs
111 * disabled, while rotate_context is called from IRQ context.
113 static void perf_pmu_rotate_start(struct pmu *pmu)
115 struct perf_cpu_context *cpuctx = this_cpu_ptr(pmu->pmu_cpu_context);
116 struct list_head *head = &__get_cpu_var(rotation_list);
118 WARN_ON(!irqs_disabled());
120 if (list_empty(&cpuctx->rotation_list))
121 list_add(&cpuctx->rotation_list, head);
124 static void get_ctx(struct perf_event_context *ctx)
126 WARN_ON(!atomic_inc_not_zero(&ctx->refcount));
129 static void free_ctx(struct rcu_head *head)
131 struct perf_event_context *ctx;
133 ctx = container_of(head, struct perf_event_context, rcu_head);
137 static void put_ctx(struct perf_event_context *ctx)
139 if (atomic_dec_and_test(&ctx->refcount)) {
141 put_ctx(ctx->parent_ctx);
143 put_task_struct(ctx->task);
144 call_rcu(&ctx->rcu_head, free_ctx);
148 static void unclone_ctx(struct perf_event_context *ctx)
150 if (ctx->parent_ctx) {
151 put_ctx(ctx->parent_ctx);
152 ctx->parent_ctx = NULL;
156 static u32 perf_event_pid(struct perf_event *event, struct task_struct *p)
159 * only top level events have the pid namespace they were created in
162 event = event->parent;
164 return task_tgid_nr_ns(p, event->ns);
167 static u32 perf_event_tid(struct perf_event *event, struct task_struct *p)
170 * only top level events have the pid namespace they were created in
173 event = event->parent;
175 return task_pid_nr_ns(p, event->ns);
179 * If we inherit events we want to return the parent event id
182 static u64 primary_event_id(struct perf_event *event)
187 id = event->parent->id;
193 * Get the perf_event_context for a task and lock it.
194 * This has to cope with with the fact that until it is locked,
195 * the context could get moved to another task.
197 static struct perf_event_context *
198 perf_lock_task_context(struct task_struct *task, int ctxn, unsigned long *flags)
200 struct perf_event_context *ctx;
204 ctx = rcu_dereference(task->perf_event_ctxp[ctxn]);
207 * If this context is a clone of another, it might
208 * get swapped for another underneath us by
209 * perf_event_task_sched_out, though the
210 * rcu_read_lock() protects us from any context
211 * getting freed. Lock the context and check if it
212 * got swapped before we could get the lock, and retry
213 * if so. If we locked the right context, then it
214 * can't get swapped on us any more.
216 raw_spin_lock_irqsave(&ctx->lock, *flags);
217 if (ctx != rcu_dereference(task->perf_event_ctxp[ctxn])) {
218 raw_spin_unlock_irqrestore(&ctx->lock, *flags);
222 if (!atomic_inc_not_zero(&ctx->refcount)) {
223 raw_spin_unlock_irqrestore(&ctx->lock, *flags);
232 * Get the context for a task and increment its pin_count so it
233 * can't get swapped to another task. This also increments its
234 * reference count so that the context can't get freed.
236 static struct perf_event_context *
237 perf_pin_task_context(struct task_struct *task, int ctxn)
239 struct perf_event_context *ctx;
242 ctx = perf_lock_task_context(task, ctxn, &flags);
245 raw_spin_unlock_irqrestore(&ctx->lock, flags);
250 static void perf_unpin_context(struct perf_event_context *ctx)
254 raw_spin_lock_irqsave(&ctx->lock, flags);
256 raw_spin_unlock_irqrestore(&ctx->lock, flags);
261 * Update the record of the current time in a context.
263 static void update_context_time(struct perf_event_context *ctx)
265 u64 now = perf_clock();
267 ctx->time += now - ctx->timestamp;
268 ctx->timestamp = now;
272 * Update the total_time_enabled and total_time_running fields for a event.
274 static void update_event_times(struct perf_event *event)
276 struct perf_event_context *ctx = event->ctx;
279 if (event->state < PERF_EVENT_STATE_INACTIVE ||
280 event->group_leader->state < PERF_EVENT_STATE_INACTIVE)
286 run_end = event->tstamp_stopped;
288 event->total_time_enabled = run_end - event->tstamp_enabled;
290 if (event->state == PERF_EVENT_STATE_INACTIVE)
291 run_end = event->tstamp_stopped;
295 event->total_time_running = run_end - event->tstamp_running;
299 * Update total_time_enabled and total_time_running for all events in a group.
301 static void update_group_times(struct perf_event *leader)
303 struct perf_event *event;
305 update_event_times(leader);
306 list_for_each_entry(event, &leader->sibling_list, group_entry)
307 update_event_times(event);
310 static struct list_head *
311 ctx_group_list(struct perf_event *event, struct perf_event_context *ctx)
313 if (event->attr.pinned)
314 return &ctx->pinned_groups;
316 return &ctx->flexible_groups;
320 * Add a event from the lists for its context.
321 * Must be called with ctx->mutex and ctx->lock held.
324 list_add_event(struct perf_event *event, struct perf_event_context *ctx)
326 WARN_ON_ONCE(event->attach_state & PERF_ATTACH_CONTEXT);
327 event->attach_state |= PERF_ATTACH_CONTEXT;
330 * If we're a stand alone event or group leader, we go to the context
331 * list, group events are kept attached to the group so that
332 * perf_group_detach can, at all times, locate all siblings.
334 if (event->group_leader == event) {
335 struct list_head *list;
337 if (is_software_event(event))
338 event->group_flags |= PERF_GROUP_SOFTWARE;
340 list = ctx_group_list(event, ctx);
341 list_add_tail(&event->group_entry, list);
344 list_add_rcu(&event->event_entry, &ctx->event_list);
346 perf_pmu_rotate_start(ctx->pmu);
348 if (event->attr.inherit_stat)
353 * Called at perf_event creation and when events are attached/detached from a
356 static void perf_event__read_size(struct perf_event *event)
358 int entry = sizeof(u64); /* value */
362 if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
365 if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
368 if (event->attr.read_format & PERF_FORMAT_ID)
369 entry += sizeof(u64);
371 if (event->attr.read_format & PERF_FORMAT_GROUP) {
372 nr += event->group_leader->nr_siblings;
377 event->read_size = size;
380 static void perf_event__header_size(struct perf_event *event)
382 struct perf_sample_data *data;
383 u64 sample_type = event->attr.sample_type;
386 perf_event__read_size(event);
388 if (sample_type & PERF_SAMPLE_IP)
389 size += sizeof(data->ip);
391 if (sample_type & PERF_SAMPLE_ADDR)
392 size += sizeof(data->addr);
394 if (sample_type & PERF_SAMPLE_PERIOD)
395 size += sizeof(data->period);
397 if (sample_type & PERF_SAMPLE_READ)
398 size += event->read_size;
400 event->header_size = size;
403 static void perf_event__id_header_size(struct perf_event *event)
405 struct perf_sample_data *data;
406 u64 sample_type = event->attr.sample_type;
409 if (sample_type & PERF_SAMPLE_TID)
410 size += sizeof(data->tid_entry);
412 if (sample_type & PERF_SAMPLE_TIME)
413 size += sizeof(data->time);
415 if (sample_type & PERF_SAMPLE_ID)
416 size += sizeof(data->id);
418 if (sample_type & PERF_SAMPLE_STREAM_ID)
419 size += sizeof(data->stream_id);
421 if (sample_type & PERF_SAMPLE_CPU)
422 size += sizeof(data->cpu_entry);
424 event->id_header_size = size;
427 static void perf_group_attach(struct perf_event *event)
429 struct perf_event *group_leader = event->group_leader, *pos;
432 * We can have double attach due to group movement in perf_event_open.
434 if (event->attach_state & PERF_ATTACH_GROUP)
437 event->attach_state |= PERF_ATTACH_GROUP;
439 if (group_leader == event)
442 if (group_leader->group_flags & PERF_GROUP_SOFTWARE &&
443 !is_software_event(event))
444 group_leader->group_flags &= ~PERF_GROUP_SOFTWARE;
446 list_add_tail(&event->group_entry, &group_leader->sibling_list);
447 group_leader->nr_siblings++;
449 perf_event__header_size(group_leader);
451 list_for_each_entry(pos, &group_leader->sibling_list, group_entry)
452 perf_event__header_size(pos);
456 * Remove a event from the lists for its context.
457 * Must be called with ctx->mutex and ctx->lock held.
460 list_del_event(struct perf_event *event, struct perf_event_context *ctx)
463 * We can have double detach due to exit/hot-unplug + close.
465 if (!(event->attach_state & PERF_ATTACH_CONTEXT))
468 event->attach_state &= ~PERF_ATTACH_CONTEXT;
471 if (event->attr.inherit_stat)
474 list_del_rcu(&event->event_entry);
476 if (event->group_leader == event)
477 list_del_init(&event->group_entry);
479 update_group_times(event);
482 * If event was in error state, then keep it
483 * that way, otherwise bogus counts will be
484 * returned on read(). The only way to get out
485 * of error state is by explicit re-enabling
488 if (event->state > PERF_EVENT_STATE_OFF)
489 event->state = PERF_EVENT_STATE_OFF;
492 static void perf_group_detach(struct perf_event *event)
494 struct perf_event *sibling, *tmp;
495 struct list_head *list = NULL;
498 * We can have double detach due to exit/hot-unplug + close.
500 if (!(event->attach_state & PERF_ATTACH_GROUP))
503 event->attach_state &= ~PERF_ATTACH_GROUP;
506 * If this is a sibling, remove it from its group.
508 if (event->group_leader != event) {
509 list_del_init(&event->group_entry);
510 event->group_leader->nr_siblings--;
514 if (!list_empty(&event->group_entry))
515 list = &event->group_entry;
518 * If this was a group event with sibling events then
519 * upgrade the siblings to singleton events by adding them
520 * to whatever list we are on.
522 list_for_each_entry_safe(sibling, tmp, &event->sibling_list, group_entry) {
524 list_move_tail(&sibling->group_entry, list);
525 sibling->group_leader = sibling;
527 /* Inherit group flags from the previous leader */
528 sibling->group_flags = event->group_flags;
532 perf_event__header_size(event->group_leader);
534 list_for_each_entry(tmp, &event->group_leader->sibling_list, group_entry)
535 perf_event__header_size(tmp);
539 event_filter_match(struct perf_event *event)
541 return event->cpu == -1 || event->cpu == smp_processor_id();
545 event_sched_out(struct perf_event *event,
546 struct perf_cpu_context *cpuctx,
547 struct perf_event_context *ctx)
551 * An event which could not be activated because of
552 * filter mismatch still needs to have its timings
553 * maintained, otherwise bogus information is return
554 * via read() for time_enabled, time_running:
556 if (event->state == PERF_EVENT_STATE_INACTIVE
557 && !event_filter_match(event)) {
558 delta = ctx->time - event->tstamp_stopped;
559 event->tstamp_running += delta;
560 event->tstamp_stopped = ctx->time;
563 if (event->state != PERF_EVENT_STATE_ACTIVE)
566 event->state = PERF_EVENT_STATE_INACTIVE;
567 if (event->pending_disable) {
568 event->pending_disable = 0;
569 event->state = PERF_EVENT_STATE_OFF;
571 event->tstamp_stopped = ctx->time;
572 event->pmu->del(event, 0);
575 if (!is_software_event(event))
576 cpuctx->active_oncpu--;
578 if (event->attr.exclusive || !cpuctx->active_oncpu)
579 cpuctx->exclusive = 0;
583 group_sched_out(struct perf_event *group_event,
584 struct perf_cpu_context *cpuctx,
585 struct perf_event_context *ctx)
587 struct perf_event *event;
588 int state = group_event->state;
590 event_sched_out(group_event, cpuctx, ctx);
593 * Schedule out siblings (if any):
595 list_for_each_entry(event, &group_event->sibling_list, group_entry)
596 event_sched_out(event, cpuctx, ctx);
598 if (state == PERF_EVENT_STATE_ACTIVE && group_event->attr.exclusive)
599 cpuctx->exclusive = 0;
602 static inline struct perf_cpu_context *
603 __get_cpu_context(struct perf_event_context *ctx)
605 return this_cpu_ptr(ctx->pmu->pmu_cpu_context);
609 * Cross CPU call to remove a performance event
611 * We disable the event on the hardware level first. After that we
612 * remove it from the context list.
614 static void __perf_event_remove_from_context(void *info)
616 struct perf_event *event = info;
617 struct perf_event_context *ctx = event->ctx;
618 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
621 * If this is a task context, we need to check whether it is
622 * the current task context of this cpu. If not it has been
623 * scheduled out before the smp call arrived.
625 if (ctx->task && cpuctx->task_ctx != ctx)
628 raw_spin_lock(&ctx->lock);
630 event_sched_out(event, cpuctx, ctx);
632 list_del_event(event, ctx);
634 raw_spin_unlock(&ctx->lock);
639 * Remove the event from a task's (or a CPU's) list of events.
641 * Must be called with ctx->mutex held.
643 * CPU events are removed with a smp call. For task events we only
644 * call when the task is on a CPU.
646 * If event->ctx is a cloned context, callers must make sure that
647 * every task struct that event->ctx->task could possibly point to
648 * remains valid. This is OK when called from perf_release since
649 * that only calls us on the top-level context, which can't be a clone.
650 * When called from perf_event_exit_task, it's OK because the
651 * context has been detached from its task.
653 static void perf_event_remove_from_context(struct perf_event *event)
655 struct perf_event_context *ctx = event->ctx;
656 struct task_struct *task = ctx->task;
660 * Per cpu events are removed via an smp call and
661 * the removal is always successful.
663 smp_call_function_single(event->cpu,
664 __perf_event_remove_from_context,
670 task_oncpu_function_call(task, __perf_event_remove_from_context,
673 raw_spin_lock_irq(&ctx->lock);
675 * If the context is active we need to retry the smp call.
677 if (ctx->nr_active && !list_empty(&event->group_entry)) {
678 raw_spin_unlock_irq(&ctx->lock);
683 * The lock prevents that this context is scheduled in so we
684 * can remove the event safely, if the call above did not
687 if (!list_empty(&event->group_entry))
688 list_del_event(event, ctx);
689 raw_spin_unlock_irq(&ctx->lock);
693 * Cross CPU call to disable a performance event
695 static void __perf_event_disable(void *info)
697 struct perf_event *event = info;
698 struct perf_event_context *ctx = event->ctx;
699 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
702 * If this is a per-task event, need to check whether this
703 * event's task is the current task on this cpu.
705 if (ctx->task && cpuctx->task_ctx != ctx)
708 raw_spin_lock(&ctx->lock);
711 * If the event is on, turn it off.
712 * If it is in error state, leave it in error state.
714 if (event->state >= PERF_EVENT_STATE_INACTIVE) {
715 update_context_time(ctx);
716 update_group_times(event);
717 if (event == event->group_leader)
718 group_sched_out(event, cpuctx, ctx);
720 event_sched_out(event, cpuctx, ctx);
721 event->state = PERF_EVENT_STATE_OFF;
724 raw_spin_unlock(&ctx->lock);
730 * If event->ctx is a cloned context, callers must make sure that
731 * every task struct that event->ctx->task could possibly point to
732 * remains valid. This condition is satisifed when called through
733 * perf_event_for_each_child or perf_event_for_each because they
734 * hold the top-level event's child_mutex, so any descendant that
735 * goes to exit will block in sync_child_event.
736 * When called from perf_pending_event it's OK because event->ctx
737 * is the current context on this CPU and preemption is disabled,
738 * hence we can't get into perf_event_task_sched_out for this context.
740 void perf_event_disable(struct perf_event *event)
742 struct perf_event_context *ctx = event->ctx;
743 struct task_struct *task = ctx->task;
747 * Disable the event on the cpu that it's on
749 smp_call_function_single(event->cpu, __perf_event_disable,
755 task_oncpu_function_call(task, __perf_event_disable, event);
757 raw_spin_lock_irq(&ctx->lock);
759 * If the event is still active, we need to retry the cross-call.
761 if (event->state == PERF_EVENT_STATE_ACTIVE) {
762 raw_spin_unlock_irq(&ctx->lock);
767 * Since we have the lock this context can't be scheduled
768 * in, so we can change the state safely.
770 if (event->state == PERF_EVENT_STATE_INACTIVE) {
771 update_group_times(event);
772 event->state = PERF_EVENT_STATE_OFF;
775 raw_spin_unlock_irq(&ctx->lock);
779 event_sched_in(struct perf_event *event,
780 struct perf_cpu_context *cpuctx,
781 struct perf_event_context *ctx)
783 if (event->state <= PERF_EVENT_STATE_OFF)
786 event->state = PERF_EVENT_STATE_ACTIVE;
787 event->oncpu = smp_processor_id();
789 * The new state must be visible before we turn it on in the hardware:
793 if (event->pmu->add(event, PERF_EF_START)) {
794 event->state = PERF_EVENT_STATE_INACTIVE;
799 event->tstamp_running += ctx->time - event->tstamp_stopped;
801 event->shadow_ctx_time = ctx->time - ctx->timestamp;
803 if (!is_software_event(event))
804 cpuctx->active_oncpu++;
807 if (event->attr.exclusive)
808 cpuctx->exclusive = 1;
814 group_sched_in(struct perf_event *group_event,
815 struct perf_cpu_context *cpuctx,
816 struct perf_event_context *ctx)
818 struct perf_event *event, *partial_group = NULL;
819 struct pmu *pmu = group_event->pmu;
821 bool simulate = false;
823 if (group_event->state == PERF_EVENT_STATE_OFF)
828 if (event_sched_in(group_event, cpuctx, ctx)) {
829 pmu->cancel_txn(pmu);
834 * Schedule in siblings as one group (if any):
836 list_for_each_entry(event, &group_event->sibling_list, group_entry) {
837 if (event_sched_in(event, cpuctx, ctx)) {
838 partial_group = event;
843 if (!pmu->commit_txn(pmu))
848 * Groups can be scheduled in as one unit only, so undo any
849 * partial group before returning:
850 * The events up to the failed event are scheduled out normally,
851 * tstamp_stopped will be updated.
853 * The failed events and the remaining siblings need to have
854 * their timings updated as if they had gone thru event_sched_in()
855 * and event_sched_out(). This is required to get consistent timings
856 * across the group. This also takes care of the case where the group
857 * could never be scheduled by ensuring tstamp_stopped is set to mark
858 * the time the event was actually stopped, such that time delta
859 * calculation in update_event_times() is correct.
861 list_for_each_entry(event, &group_event->sibling_list, group_entry) {
862 if (event == partial_group)
866 event->tstamp_running += now - event->tstamp_stopped;
867 event->tstamp_stopped = now;
869 event_sched_out(event, cpuctx, ctx);
872 event_sched_out(group_event, cpuctx, ctx);
874 pmu->cancel_txn(pmu);
880 * Work out whether we can put this event group on the CPU now.
882 static int group_can_go_on(struct perf_event *event,
883 struct perf_cpu_context *cpuctx,
887 * Groups consisting entirely of software events can always go on.
889 if (event->group_flags & PERF_GROUP_SOFTWARE)
892 * If an exclusive group is already on, no other hardware
895 if (cpuctx->exclusive)
898 * If this group is exclusive and there are already
899 * events on the CPU, it can't go on.
901 if (event->attr.exclusive && cpuctx->active_oncpu)
904 * Otherwise, try to add it if all previous groups were able
910 static void add_event_to_ctx(struct perf_event *event,
911 struct perf_event_context *ctx)
913 list_add_event(event, ctx);
914 perf_group_attach(event);
915 event->tstamp_enabled = ctx->time;
916 event->tstamp_running = ctx->time;
917 event->tstamp_stopped = ctx->time;
921 * Cross CPU call to install and enable a performance event
923 * Must be called with ctx->mutex held
925 static void __perf_install_in_context(void *info)
927 struct perf_event *event = info;
928 struct perf_event_context *ctx = event->ctx;
929 struct perf_event *leader = event->group_leader;
930 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
934 * If this is a task context, we need to check whether it is
935 * the current task context of this cpu. If not it has been
936 * scheduled out before the smp call arrived.
937 * Or possibly this is the right context but it isn't
938 * on this cpu because it had no events.
940 if (ctx->task && cpuctx->task_ctx != ctx) {
941 if (cpuctx->task_ctx || ctx->task != current)
943 cpuctx->task_ctx = ctx;
946 raw_spin_lock(&ctx->lock);
948 update_context_time(ctx);
950 add_event_to_ctx(event, ctx);
952 if (!event_filter_match(event))
956 * Don't put the event on if it is disabled or if
957 * it is in a group and the group isn't on.
959 if (event->state != PERF_EVENT_STATE_INACTIVE ||
960 (leader != event && leader->state != PERF_EVENT_STATE_ACTIVE))
964 * An exclusive event can't go on if there are already active
965 * hardware events, and no hardware event can go on if there
966 * is already an exclusive event on.
968 if (!group_can_go_on(event, cpuctx, 1))
971 err = event_sched_in(event, cpuctx, ctx);
975 * This event couldn't go on. If it is in a group
976 * then we have to pull the whole group off.
977 * If the event group is pinned then put it in error state.
980 group_sched_out(leader, cpuctx, ctx);
981 if (leader->attr.pinned) {
982 update_group_times(leader);
983 leader->state = PERF_EVENT_STATE_ERROR;
988 raw_spin_unlock(&ctx->lock);
992 * Attach a performance event to a context
994 * First we add the event to the list with the hardware enable bit
995 * in event->hw_config cleared.
997 * If the event is attached to a task which is on a CPU we use a smp
998 * call to enable it in the task context. The task might have been
999 * scheduled away, but we check this in the smp call again.
1001 * Must be called with ctx->mutex held.
1004 perf_install_in_context(struct perf_event_context *ctx,
1005 struct perf_event *event,
1008 struct task_struct *task = ctx->task;
1014 * Per cpu events are installed via an smp call and
1015 * the install is always successful.
1017 smp_call_function_single(cpu, __perf_install_in_context,
1023 task_oncpu_function_call(task, __perf_install_in_context,
1026 raw_spin_lock_irq(&ctx->lock);
1028 * we need to retry the smp call.
1030 if (ctx->is_active && list_empty(&event->group_entry)) {
1031 raw_spin_unlock_irq(&ctx->lock);
1036 * The lock prevents that this context is scheduled in so we
1037 * can add the event safely, if it the call above did not
1040 if (list_empty(&event->group_entry))
1041 add_event_to_ctx(event, ctx);
1042 raw_spin_unlock_irq(&ctx->lock);
1046 * Put a event into inactive state and update time fields.
1047 * Enabling the leader of a group effectively enables all
1048 * the group members that aren't explicitly disabled, so we
1049 * have to update their ->tstamp_enabled also.
1050 * Note: this works for group members as well as group leaders
1051 * since the non-leader members' sibling_lists will be empty.
1053 static void __perf_event_mark_enabled(struct perf_event *event,
1054 struct perf_event_context *ctx)
1056 struct perf_event *sub;
1058 event->state = PERF_EVENT_STATE_INACTIVE;
1059 event->tstamp_enabled = ctx->time - event->total_time_enabled;
1060 list_for_each_entry(sub, &event->sibling_list, group_entry) {
1061 if (sub->state >= PERF_EVENT_STATE_INACTIVE) {
1062 sub->tstamp_enabled =
1063 ctx->time - sub->total_time_enabled;
1069 * Cross CPU call to enable a performance event
1071 static void __perf_event_enable(void *info)
1073 struct perf_event *event = info;
1074 struct perf_event_context *ctx = event->ctx;
1075 struct perf_event *leader = event->group_leader;
1076 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
1080 * If this is a per-task event, need to check whether this
1081 * event's task is the current task on this cpu.
1083 if (ctx->task && cpuctx->task_ctx != ctx) {
1084 if (cpuctx->task_ctx || ctx->task != current)
1086 cpuctx->task_ctx = ctx;
1089 raw_spin_lock(&ctx->lock);
1091 update_context_time(ctx);
1093 if (event->state >= PERF_EVENT_STATE_INACTIVE)
1095 __perf_event_mark_enabled(event, ctx);
1097 if (!event_filter_match(event))
1101 * If the event is in a group and isn't the group leader,
1102 * then don't put it on unless the group is on.
1104 if (leader != event && leader->state != PERF_EVENT_STATE_ACTIVE)
1107 if (!group_can_go_on(event, cpuctx, 1)) {
1110 if (event == leader)
1111 err = group_sched_in(event, cpuctx, ctx);
1113 err = event_sched_in(event, cpuctx, ctx);
1118 * If this event can't go on and it's part of a
1119 * group, then the whole group has to come off.
1121 if (leader != event)
1122 group_sched_out(leader, cpuctx, ctx);
1123 if (leader->attr.pinned) {
1124 update_group_times(leader);
1125 leader->state = PERF_EVENT_STATE_ERROR;
1130 raw_spin_unlock(&ctx->lock);
1136 * If event->ctx is a cloned context, callers must make sure that
1137 * every task struct that event->ctx->task could possibly point to
1138 * remains valid. This condition is satisfied when called through
1139 * perf_event_for_each_child or perf_event_for_each as described
1140 * for perf_event_disable.
1142 void perf_event_enable(struct perf_event *event)
1144 struct perf_event_context *ctx = event->ctx;
1145 struct task_struct *task = ctx->task;
1149 * Enable the event on the cpu that it's on
1151 smp_call_function_single(event->cpu, __perf_event_enable,
1156 raw_spin_lock_irq(&ctx->lock);
1157 if (event->state >= PERF_EVENT_STATE_INACTIVE)
1161 * If the event is in error state, clear that first.
1162 * That way, if we see the event in error state below, we
1163 * know that it has gone back into error state, as distinct
1164 * from the task having been scheduled away before the
1165 * cross-call arrived.
1167 if (event->state == PERF_EVENT_STATE_ERROR)
1168 event->state = PERF_EVENT_STATE_OFF;
1171 raw_spin_unlock_irq(&ctx->lock);
1172 task_oncpu_function_call(task, __perf_event_enable, event);
1174 raw_spin_lock_irq(&ctx->lock);
1177 * If the context is active and the event is still off,
1178 * we need to retry the cross-call.
1180 if (ctx->is_active && event->state == PERF_EVENT_STATE_OFF)
1184 * Since we have the lock this context can't be scheduled
1185 * in, so we can change the state safely.
1187 if (event->state == PERF_EVENT_STATE_OFF)
1188 __perf_event_mark_enabled(event, ctx);
1191 raw_spin_unlock_irq(&ctx->lock);
1194 static int perf_event_refresh(struct perf_event *event, int refresh)
1197 * not supported on inherited events
1199 if (event->attr.inherit || !is_sampling_event(event))
1202 atomic_add(refresh, &event->event_limit);
1203 perf_event_enable(event);
1208 static void ctx_sched_out(struct perf_event_context *ctx,
1209 struct perf_cpu_context *cpuctx,
1210 enum event_type_t event_type)
1212 struct perf_event *event;
1214 raw_spin_lock(&ctx->lock);
1215 perf_pmu_disable(ctx->pmu);
1217 if (likely(!ctx->nr_events))
1219 update_context_time(ctx);
1221 if (!ctx->nr_active)
1224 if (event_type & EVENT_PINNED) {
1225 list_for_each_entry(event, &ctx->pinned_groups, group_entry)
1226 group_sched_out(event, cpuctx, ctx);
1229 if (event_type & EVENT_FLEXIBLE) {
1230 list_for_each_entry(event, &ctx->flexible_groups, group_entry)
1231 group_sched_out(event, cpuctx, ctx);
1234 perf_pmu_enable(ctx->pmu);
1235 raw_spin_unlock(&ctx->lock);
1239 * Test whether two contexts are equivalent, i.e. whether they
1240 * have both been cloned from the same version of the same context
1241 * and they both have the same number of enabled events.
1242 * If the number of enabled events is the same, then the set
1243 * of enabled events should be the same, because these are both
1244 * inherited contexts, therefore we can't access individual events
1245 * in them directly with an fd; we can only enable/disable all
1246 * events via prctl, or enable/disable all events in a family
1247 * via ioctl, which will have the same effect on both contexts.
1249 static int context_equiv(struct perf_event_context *ctx1,
1250 struct perf_event_context *ctx2)
1252 return ctx1->parent_ctx && ctx1->parent_ctx == ctx2->parent_ctx
1253 && ctx1->parent_gen == ctx2->parent_gen
1254 && !ctx1->pin_count && !ctx2->pin_count;
1257 static void __perf_event_sync_stat(struct perf_event *event,
1258 struct perf_event *next_event)
1262 if (!event->attr.inherit_stat)
1266 * Update the event value, we cannot use perf_event_read()
1267 * because we're in the middle of a context switch and have IRQs
1268 * disabled, which upsets smp_call_function_single(), however
1269 * we know the event must be on the current CPU, therefore we
1270 * don't need to use it.
1272 switch (event->state) {
1273 case PERF_EVENT_STATE_ACTIVE:
1274 event->pmu->read(event);
1277 case PERF_EVENT_STATE_INACTIVE:
1278 update_event_times(event);
1286 * In order to keep per-task stats reliable we need to flip the event
1287 * values when we flip the contexts.
1289 value = local64_read(&next_event->count);
1290 value = local64_xchg(&event->count, value);
1291 local64_set(&next_event->count, value);
1293 swap(event->total_time_enabled, next_event->total_time_enabled);
1294 swap(event->total_time_running, next_event->total_time_running);
1297 * Since we swizzled the values, update the user visible data too.
1299 perf_event_update_userpage(event);
1300 perf_event_update_userpage(next_event);
1303 #define list_next_entry(pos, member) \
1304 list_entry(pos->member.next, typeof(*pos), member)
1306 static void perf_event_sync_stat(struct perf_event_context *ctx,
1307 struct perf_event_context *next_ctx)
1309 struct perf_event *event, *next_event;
1314 update_context_time(ctx);
1316 event = list_first_entry(&ctx->event_list,
1317 struct perf_event, event_entry);
1319 next_event = list_first_entry(&next_ctx->event_list,
1320 struct perf_event, event_entry);
1322 while (&event->event_entry != &ctx->event_list &&
1323 &next_event->event_entry != &next_ctx->event_list) {
1325 __perf_event_sync_stat(event, next_event);
1327 event = list_next_entry(event, event_entry);
1328 next_event = list_next_entry(next_event, event_entry);
1332 void perf_event_context_sched_out(struct task_struct *task, int ctxn,
1333 struct task_struct *next)
1335 struct perf_event_context *ctx = task->perf_event_ctxp[ctxn];
1336 struct perf_event_context *next_ctx;
1337 struct perf_event_context *parent;
1338 struct perf_cpu_context *cpuctx;
1344 cpuctx = __get_cpu_context(ctx);
1345 if (!cpuctx->task_ctx)
1349 parent = rcu_dereference(ctx->parent_ctx);
1350 next_ctx = next->perf_event_ctxp[ctxn];
1351 if (parent && next_ctx &&
1352 rcu_dereference(next_ctx->parent_ctx) == parent) {
1354 * Looks like the two contexts are clones, so we might be
1355 * able to optimize the context switch. We lock both
1356 * contexts and check that they are clones under the
1357 * lock (including re-checking that neither has been
1358 * uncloned in the meantime). It doesn't matter which
1359 * order we take the locks because no other cpu could
1360 * be trying to lock both of these tasks.
1362 raw_spin_lock(&ctx->lock);
1363 raw_spin_lock_nested(&next_ctx->lock, SINGLE_DEPTH_NESTING);
1364 if (context_equiv(ctx, next_ctx)) {
1366 * XXX do we need a memory barrier of sorts
1367 * wrt to rcu_dereference() of perf_event_ctxp
1369 task->perf_event_ctxp[ctxn] = next_ctx;
1370 next->perf_event_ctxp[ctxn] = ctx;
1372 next_ctx->task = task;
1375 perf_event_sync_stat(ctx, next_ctx);
1377 raw_spin_unlock(&next_ctx->lock);
1378 raw_spin_unlock(&ctx->lock);
1383 ctx_sched_out(ctx, cpuctx, EVENT_ALL);
1384 cpuctx->task_ctx = NULL;
1388 #define for_each_task_context_nr(ctxn) \
1389 for ((ctxn) = 0; (ctxn) < perf_nr_task_contexts; (ctxn)++)
1392 * Called from scheduler to remove the events of the current task,
1393 * with interrupts disabled.
1395 * We stop each event and update the event value in event->count.
1397 * This does not protect us against NMI, but disable()
1398 * sets the disabled bit in the control field of event _before_
1399 * accessing the event control register. If a NMI hits, then it will
1400 * not restart the event.
1402 void __perf_event_task_sched_out(struct task_struct *task,
1403 struct task_struct *next)
1407 for_each_task_context_nr(ctxn)
1408 perf_event_context_sched_out(task, ctxn, next);
1411 static void task_ctx_sched_out(struct perf_event_context *ctx,
1412 enum event_type_t event_type)
1414 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
1416 if (!cpuctx->task_ctx)
1419 if (WARN_ON_ONCE(ctx != cpuctx->task_ctx))
1422 ctx_sched_out(ctx, cpuctx, event_type);
1423 cpuctx->task_ctx = NULL;
1427 * Called with IRQs disabled
1429 static void cpu_ctx_sched_out(struct perf_cpu_context *cpuctx,
1430 enum event_type_t event_type)
1432 ctx_sched_out(&cpuctx->ctx, cpuctx, event_type);
1436 ctx_pinned_sched_in(struct perf_event_context *ctx,
1437 struct perf_cpu_context *cpuctx)
1439 struct perf_event *event;
1441 list_for_each_entry(event, &ctx->pinned_groups, group_entry) {
1442 if (event->state <= PERF_EVENT_STATE_OFF)
1444 if (!event_filter_match(event))
1447 if (group_can_go_on(event, cpuctx, 1))
1448 group_sched_in(event, cpuctx, ctx);
1451 * If this pinned group hasn't been scheduled,
1452 * put it in error state.
1454 if (event->state == PERF_EVENT_STATE_INACTIVE) {
1455 update_group_times(event);
1456 event->state = PERF_EVENT_STATE_ERROR;
1462 ctx_flexible_sched_in(struct perf_event_context *ctx,
1463 struct perf_cpu_context *cpuctx)
1465 struct perf_event *event;
1468 list_for_each_entry(event, &ctx->flexible_groups, group_entry) {
1469 /* Ignore events in OFF or ERROR state */
1470 if (event->state <= PERF_EVENT_STATE_OFF)
1473 * Listen to the 'cpu' scheduling filter constraint
1476 if (!event_filter_match(event))
1479 if (group_can_go_on(event, cpuctx, can_add_hw)) {
1480 if (group_sched_in(event, cpuctx, ctx))
1487 ctx_sched_in(struct perf_event_context *ctx,
1488 struct perf_cpu_context *cpuctx,
1489 enum event_type_t event_type)
1491 raw_spin_lock(&ctx->lock);
1493 if (likely(!ctx->nr_events))
1496 ctx->timestamp = perf_clock();
1499 * First go through the list and put on any pinned groups
1500 * in order to give them the best chance of going on.
1502 if (event_type & EVENT_PINNED)
1503 ctx_pinned_sched_in(ctx, cpuctx);
1505 /* Then walk through the lower prio flexible groups */
1506 if (event_type & EVENT_FLEXIBLE)
1507 ctx_flexible_sched_in(ctx, cpuctx);
1510 raw_spin_unlock(&ctx->lock);
1513 static void cpu_ctx_sched_in(struct perf_cpu_context *cpuctx,
1514 enum event_type_t event_type)
1516 struct perf_event_context *ctx = &cpuctx->ctx;
1518 ctx_sched_in(ctx, cpuctx, event_type);
1521 static void task_ctx_sched_in(struct perf_event_context *ctx,
1522 enum event_type_t event_type)
1524 struct perf_cpu_context *cpuctx;
1526 cpuctx = __get_cpu_context(ctx);
1527 if (cpuctx->task_ctx == ctx)
1530 ctx_sched_in(ctx, cpuctx, event_type);
1531 cpuctx->task_ctx = ctx;
1534 void perf_event_context_sched_in(struct perf_event_context *ctx)
1536 struct perf_cpu_context *cpuctx;
1538 cpuctx = __get_cpu_context(ctx);
1539 if (cpuctx->task_ctx == ctx)
1542 perf_pmu_disable(ctx->pmu);
1544 * We want to keep the following priority order:
1545 * cpu pinned (that don't need to move), task pinned,
1546 * cpu flexible, task flexible.
1548 cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE);
1550 ctx_sched_in(ctx, cpuctx, EVENT_PINNED);
1551 cpu_ctx_sched_in(cpuctx, EVENT_FLEXIBLE);
1552 ctx_sched_in(ctx, cpuctx, EVENT_FLEXIBLE);
1554 cpuctx->task_ctx = ctx;
1557 * Since these rotations are per-cpu, we need to ensure the
1558 * cpu-context we got scheduled on is actually rotating.
1560 perf_pmu_rotate_start(ctx->pmu);
1561 perf_pmu_enable(ctx->pmu);
1565 * Called from scheduler to add the events of the current task
1566 * with interrupts disabled.
1568 * We restore the event value and then enable it.
1570 * This does not protect us against NMI, but enable()
1571 * sets the enabled bit in the control field of event _before_
1572 * accessing the event control register. If a NMI hits, then it will
1573 * keep the event running.
1575 void __perf_event_task_sched_in(struct task_struct *task)
1577 struct perf_event_context *ctx;
1580 for_each_task_context_nr(ctxn) {
1581 ctx = task->perf_event_ctxp[ctxn];
1585 perf_event_context_sched_in(ctx);
1589 #define MAX_INTERRUPTS (~0ULL)
1591 static void perf_log_throttle(struct perf_event *event, int enable);
1593 static u64 perf_calculate_period(struct perf_event *event, u64 nsec, u64 count)
1595 u64 frequency = event->attr.sample_freq;
1596 u64 sec = NSEC_PER_SEC;
1597 u64 divisor, dividend;
1599 int count_fls, nsec_fls, frequency_fls, sec_fls;
1601 count_fls = fls64(count);
1602 nsec_fls = fls64(nsec);
1603 frequency_fls = fls64(frequency);
1607 * We got @count in @nsec, with a target of sample_freq HZ
1608 * the target period becomes:
1611 * period = -------------------
1612 * @nsec * sample_freq
1617 * Reduce accuracy by one bit such that @a and @b converge
1618 * to a similar magnitude.
1620 #define REDUCE_FLS(a, b) \
1622 if (a##_fls > b##_fls) { \
1632 * Reduce accuracy until either term fits in a u64, then proceed with
1633 * the other, so that finally we can do a u64/u64 division.
1635 while (count_fls + sec_fls > 64 && nsec_fls + frequency_fls > 64) {
1636 REDUCE_FLS(nsec, frequency);
1637 REDUCE_FLS(sec, count);
1640 if (count_fls + sec_fls > 64) {
1641 divisor = nsec * frequency;
1643 while (count_fls + sec_fls > 64) {
1644 REDUCE_FLS(count, sec);
1648 dividend = count * sec;
1650 dividend = count * sec;
1652 while (nsec_fls + frequency_fls > 64) {
1653 REDUCE_FLS(nsec, frequency);
1657 divisor = nsec * frequency;
1663 return div64_u64(dividend, divisor);
1666 static void perf_adjust_period(struct perf_event *event, u64 nsec, u64 count)
1668 struct hw_perf_event *hwc = &event->hw;
1669 s64 period, sample_period;
1672 period = perf_calculate_period(event, nsec, count);
1674 delta = (s64)(period - hwc->sample_period);
1675 delta = (delta + 7) / 8; /* low pass filter */
1677 sample_period = hwc->sample_period + delta;
1682 hwc->sample_period = sample_period;
1684 if (local64_read(&hwc->period_left) > 8*sample_period) {
1685 event->pmu->stop(event, PERF_EF_UPDATE);
1686 local64_set(&hwc->period_left, 0);
1687 event->pmu->start(event, PERF_EF_RELOAD);
1691 static void perf_ctx_adjust_freq(struct perf_event_context *ctx, u64 period)
1693 struct perf_event *event;
1694 struct hw_perf_event *hwc;
1695 u64 interrupts, now;
1698 raw_spin_lock(&ctx->lock);
1699 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
1700 if (event->state != PERF_EVENT_STATE_ACTIVE)
1703 if (!event_filter_match(event))
1708 interrupts = hwc->interrupts;
1709 hwc->interrupts = 0;
1712 * unthrottle events on the tick
1714 if (interrupts == MAX_INTERRUPTS) {
1715 perf_log_throttle(event, 1);
1716 event->pmu->start(event, 0);
1719 if (!event->attr.freq || !event->attr.sample_freq)
1722 event->pmu->read(event);
1723 now = local64_read(&event->count);
1724 delta = now - hwc->freq_count_stamp;
1725 hwc->freq_count_stamp = now;
1728 perf_adjust_period(event, period, delta);
1730 raw_spin_unlock(&ctx->lock);
1734 * Round-robin a context's events:
1736 static void rotate_ctx(struct perf_event_context *ctx)
1738 raw_spin_lock(&ctx->lock);
1741 * Rotate the first entry last of non-pinned groups. Rotation might be
1742 * disabled by the inheritance code.
1744 if (!ctx->rotate_disable)
1745 list_rotate_left(&ctx->flexible_groups);
1747 raw_spin_unlock(&ctx->lock);
1751 * perf_pmu_rotate_start() and perf_rotate_context() are fully serialized
1752 * because they're strictly cpu affine and rotate_start is called with IRQs
1753 * disabled, while rotate_context is called from IRQ context.
1755 static void perf_rotate_context(struct perf_cpu_context *cpuctx)
1757 u64 interval = (u64)cpuctx->jiffies_interval * TICK_NSEC;
1758 struct perf_event_context *ctx = NULL;
1759 int rotate = 0, remove = 1;
1761 if (cpuctx->ctx.nr_events) {
1763 if (cpuctx->ctx.nr_events != cpuctx->ctx.nr_active)
1767 ctx = cpuctx->task_ctx;
1768 if (ctx && ctx->nr_events) {
1770 if (ctx->nr_events != ctx->nr_active)
1774 perf_pmu_disable(cpuctx->ctx.pmu);
1775 perf_ctx_adjust_freq(&cpuctx->ctx, interval);
1777 perf_ctx_adjust_freq(ctx, interval);
1782 cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE);
1784 task_ctx_sched_out(ctx, EVENT_FLEXIBLE);
1786 rotate_ctx(&cpuctx->ctx);
1790 cpu_ctx_sched_in(cpuctx, EVENT_FLEXIBLE);
1792 task_ctx_sched_in(ctx, EVENT_FLEXIBLE);
1796 list_del_init(&cpuctx->rotation_list);
1798 perf_pmu_enable(cpuctx->ctx.pmu);
1801 void perf_event_task_tick(void)
1803 struct list_head *head = &__get_cpu_var(rotation_list);
1804 struct perf_cpu_context *cpuctx, *tmp;
1806 WARN_ON(!irqs_disabled());
1808 list_for_each_entry_safe(cpuctx, tmp, head, rotation_list) {
1809 if (cpuctx->jiffies_interval == 1 ||
1810 !(jiffies % cpuctx->jiffies_interval))
1811 perf_rotate_context(cpuctx);
1815 static int event_enable_on_exec(struct perf_event *event,
1816 struct perf_event_context *ctx)
1818 if (!event->attr.enable_on_exec)
1821 event->attr.enable_on_exec = 0;
1822 if (event->state >= PERF_EVENT_STATE_INACTIVE)
1825 __perf_event_mark_enabled(event, ctx);
1831 * Enable all of a task's events that have been marked enable-on-exec.
1832 * This expects task == current.
1834 static void perf_event_enable_on_exec(struct perf_event_context *ctx)
1836 struct perf_event *event;
1837 unsigned long flags;
1841 local_irq_save(flags);
1842 if (!ctx || !ctx->nr_events)
1845 task_ctx_sched_out(ctx, EVENT_ALL);
1847 raw_spin_lock(&ctx->lock);
1849 list_for_each_entry(event, &ctx->pinned_groups, group_entry) {
1850 ret = event_enable_on_exec(event, ctx);
1855 list_for_each_entry(event, &ctx->flexible_groups, group_entry) {
1856 ret = event_enable_on_exec(event, ctx);
1862 * Unclone this context if we enabled any event.
1867 raw_spin_unlock(&ctx->lock);
1869 perf_event_context_sched_in(ctx);
1871 local_irq_restore(flags);
1875 * Cross CPU call to read the hardware event
1877 static void __perf_event_read(void *info)
1879 struct perf_event *event = info;
1880 struct perf_event_context *ctx = event->ctx;
1881 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
1884 * If this is a task context, we need to check whether it is
1885 * the current task context of this cpu. If not it has been
1886 * scheduled out before the smp call arrived. In that case
1887 * event->count would have been updated to a recent sample
1888 * when the event was scheduled out.
1890 if (ctx->task && cpuctx->task_ctx != ctx)
1893 raw_spin_lock(&ctx->lock);
1894 update_context_time(ctx);
1895 update_event_times(event);
1896 raw_spin_unlock(&ctx->lock);
1898 event->pmu->read(event);
1901 static inline u64 perf_event_count(struct perf_event *event)
1903 return local64_read(&event->count) + atomic64_read(&event->child_count);
1906 static u64 perf_event_read(struct perf_event *event)
1909 * If event is enabled and currently active on a CPU, update the
1910 * value in the event structure:
1912 if (event->state == PERF_EVENT_STATE_ACTIVE) {
1913 smp_call_function_single(event->oncpu,
1914 __perf_event_read, event, 1);
1915 } else if (event->state == PERF_EVENT_STATE_INACTIVE) {
1916 struct perf_event_context *ctx = event->ctx;
1917 unsigned long flags;
1919 raw_spin_lock_irqsave(&ctx->lock, flags);
1921 * may read while context is not active
1922 * (e.g., thread is blocked), in that case
1923 * we cannot update context time
1926 update_context_time(ctx);
1927 update_event_times(event);
1928 raw_spin_unlock_irqrestore(&ctx->lock, flags);
1931 return perf_event_count(event);
1938 struct callchain_cpus_entries {
1939 struct rcu_head rcu_head;
1940 struct perf_callchain_entry *cpu_entries[0];
1943 static DEFINE_PER_CPU(int, callchain_recursion[PERF_NR_CONTEXTS]);
1944 static atomic_t nr_callchain_events;
1945 static DEFINE_MUTEX(callchain_mutex);
1946 struct callchain_cpus_entries *callchain_cpus_entries;
1949 __weak void perf_callchain_kernel(struct perf_callchain_entry *entry,
1950 struct pt_regs *regs)
1954 __weak void perf_callchain_user(struct perf_callchain_entry *entry,
1955 struct pt_regs *regs)
1959 static void release_callchain_buffers_rcu(struct rcu_head *head)
1961 struct callchain_cpus_entries *entries;
1964 entries = container_of(head, struct callchain_cpus_entries, rcu_head);
1966 for_each_possible_cpu(cpu)
1967 kfree(entries->cpu_entries[cpu]);
1972 static void release_callchain_buffers(void)
1974 struct callchain_cpus_entries *entries;
1976 entries = callchain_cpus_entries;
1977 rcu_assign_pointer(callchain_cpus_entries, NULL);
1978 call_rcu(&entries->rcu_head, release_callchain_buffers_rcu);
1981 static int alloc_callchain_buffers(void)
1985 struct callchain_cpus_entries *entries;
1988 * We can't use the percpu allocation API for data that can be
1989 * accessed from NMI. Use a temporary manual per cpu allocation
1990 * until that gets sorted out.
1992 size = sizeof(*entries) + sizeof(struct perf_callchain_entry *) *
1993 num_possible_cpus();
1995 entries = kzalloc(size, GFP_KERNEL);
1999 size = sizeof(struct perf_callchain_entry) * PERF_NR_CONTEXTS;
2001 for_each_possible_cpu(cpu) {
2002 entries->cpu_entries[cpu] = kmalloc_node(size, GFP_KERNEL,
2004 if (!entries->cpu_entries[cpu])
2008 rcu_assign_pointer(callchain_cpus_entries, entries);
2013 for_each_possible_cpu(cpu)
2014 kfree(entries->cpu_entries[cpu]);
2020 static int get_callchain_buffers(void)
2025 mutex_lock(&callchain_mutex);
2027 count = atomic_inc_return(&nr_callchain_events);
2028 if (WARN_ON_ONCE(count < 1)) {
2034 /* If the allocation failed, give up */
2035 if (!callchain_cpus_entries)
2040 err = alloc_callchain_buffers();
2042 release_callchain_buffers();
2044 mutex_unlock(&callchain_mutex);
2049 static void put_callchain_buffers(void)
2051 if (atomic_dec_and_mutex_lock(&nr_callchain_events, &callchain_mutex)) {
2052 release_callchain_buffers();
2053 mutex_unlock(&callchain_mutex);
2057 static int get_recursion_context(int *recursion)
2065 else if (in_softirq())
2070 if (recursion[rctx])
2079 static inline void put_recursion_context(int *recursion, int rctx)
2085 static struct perf_callchain_entry *get_callchain_entry(int *rctx)
2088 struct callchain_cpus_entries *entries;
2090 *rctx = get_recursion_context(__get_cpu_var(callchain_recursion));
2094 entries = rcu_dereference(callchain_cpus_entries);
2098 cpu = smp_processor_id();
2100 return &entries->cpu_entries[cpu][*rctx];
2104 put_callchain_entry(int rctx)
2106 put_recursion_context(__get_cpu_var(callchain_recursion), rctx);
2109 static struct perf_callchain_entry *perf_callchain(struct pt_regs *regs)
2112 struct perf_callchain_entry *entry;
2115 entry = get_callchain_entry(&rctx);
2124 if (!user_mode(regs)) {
2125 perf_callchain_store(entry, PERF_CONTEXT_KERNEL);
2126 perf_callchain_kernel(entry, regs);
2128 regs = task_pt_regs(current);
2134 perf_callchain_store(entry, PERF_CONTEXT_USER);
2135 perf_callchain_user(entry, regs);
2139 put_callchain_entry(rctx);
2145 * Initialize the perf_event context in a task_struct:
2147 static void __perf_event_init_context(struct perf_event_context *ctx)
2149 raw_spin_lock_init(&ctx->lock);
2150 mutex_init(&ctx->mutex);
2151 INIT_LIST_HEAD(&ctx->pinned_groups);
2152 INIT_LIST_HEAD(&ctx->flexible_groups);
2153 INIT_LIST_HEAD(&ctx->event_list);
2154 atomic_set(&ctx->refcount, 1);
2157 static struct perf_event_context *
2158 alloc_perf_context(struct pmu *pmu, struct task_struct *task)
2160 struct perf_event_context *ctx;
2162 ctx = kzalloc(sizeof(struct perf_event_context), GFP_KERNEL);
2166 __perf_event_init_context(ctx);
2169 get_task_struct(task);
2176 static struct task_struct *
2177 find_lively_task_by_vpid(pid_t vpid)
2179 struct task_struct *task;
2186 task = find_task_by_vpid(vpid);
2188 get_task_struct(task);
2192 return ERR_PTR(-ESRCH);
2195 * Can't attach events to a dying task.
2198 if (task->flags & PF_EXITING)
2201 /* Reuse ptrace permission checks for now. */
2203 if (!ptrace_may_access(task, PTRACE_MODE_READ))
2208 put_task_struct(task);
2209 return ERR_PTR(err);
2213 static struct perf_event_context *
2214 find_get_context(struct pmu *pmu, struct task_struct *task, int cpu)
2216 struct perf_event_context *ctx;
2217 struct perf_cpu_context *cpuctx;
2218 unsigned long flags;
2221 if (!task && cpu != -1) {
2222 /* Must be root to operate on a CPU event: */
2223 if (perf_paranoid_cpu() && !capable(CAP_SYS_ADMIN))
2224 return ERR_PTR(-EACCES);
2226 if (cpu < 0 || cpu >= nr_cpumask_bits)
2227 return ERR_PTR(-EINVAL);
2230 * We could be clever and allow to attach a event to an
2231 * offline CPU and activate it when the CPU comes up, but
2234 if (!cpu_online(cpu))
2235 return ERR_PTR(-ENODEV);
2237 cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu);
2245 ctxn = pmu->task_ctx_nr;
2250 ctx = perf_lock_task_context(task, ctxn, &flags);
2253 raw_spin_unlock_irqrestore(&ctx->lock, flags);
2257 ctx = alloc_perf_context(pmu, task);
2264 if (cmpxchg(&task->perf_event_ctxp[ctxn], NULL, ctx)) {
2266 * We raced with some other task; use
2267 * the context they set.
2269 put_task_struct(task);
2278 return ERR_PTR(err);
2281 static void perf_event_free_filter(struct perf_event *event);
2283 static void free_event_rcu(struct rcu_head *head)
2285 struct perf_event *event;
2287 event = container_of(head, struct perf_event, rcu_head);
2289 put_pid_ns(event->ns);
2290 perf_event_free_filter(event);
2294 static void perf_buffer_put(struct perf_buffer *buffer);
2296 static void free_event(struct perf_event *event)
2298 irq_work_sync(&event->pending);
2300 if (!event->parent) {
2301 if (event->attach_state & PERF_ATTACH_TASK)
2302 jump_label_dec(&perf_task_events);
2303 if (event->attr.mmap || event->attr.mmap_data)
2304 atomic_dec(&nr_mmap_events);
2305 if (event->attr.comm)
2306 atomic_dec(&nr_comm_events);
2307 if (event->attr.task)
2308 atomic_dec(&nr_task_events);
2309 if (event->attr.sample_type & PERF_SAMPLE_CALLCHAIN)
2310 put_callchain_buffers();
2313 if (event->buffer) {
2314 perf_buffer_put(event->buffer);
2315 event->buffer = NULL;
2319 event->destroy(event);
2322 put_ctx(event->ctx);
2324 call_rcu(&event->rcu_head, free_event_rcu);
2327 int perf_event_release_kernel(struct perf_event *event)
2329 struct perf_event_context *ctx = event->ctx;
2332 * Remove from the PMU, can't get re-enabled since we got
2333 * here because the last ref went.
2335 perf_event_disable(event);
2337 WARN_ON_ONCE(ctx->parent_ctx);
2339 * There are two ways this annotation is useful:
2341 * 1) there is a lock recursion from perf_event_exit_task
2342 * see the comment there.
2344 * 2) there is a lock-inversion with mmap_sem through
2345 * perf_event_read_group(), which takes faults while
2346 * holding ctx->mutex, however this is called after
2347 * the last filedesc died, so there is no possibility
2348 * to trigger the AB-BA case.
2350 mutex_lock_nested(&ctx->mutex, SINGLE_DEPTH_NESTING);
2351 raw_spin_lock_irq(&ctx->lock);
2352 perf_group_detach(event);
2353 list_del_event(event, ctx);
2354 raw_spin_unlock_irq(&ctx->lock);
2355 mutex_unlock(&ctx->mutex);
2361 EXPORT_SYMBOL_GPL(perf_event_release_kernel);
2364 * Called when the last reference to the file is gone.
2366 static int perf_release(struct inode *inode, struct file *file)
2368 struct perf_event *event = file->private_data;
2369 struct task_struct *owner;
2371 file->private_data = NULL;
2374 owner = ACCESS_ONCE(event->owner);
2376 * Matches the smp_wmb() in perf_event_exit_task(). If we observe
2377 * !owner it means the list deletion is complete and we can indeed
2378 * free this event, otherwise we need to serialize on
2379 * owner->perf_event_mutex.
2381 smp_read_barrier_depends();
2384 * Since delayed_put_task_struct() also drops the last
2385 * task reference we can safely take a new reference
2386 * while holding the rcu_read_lock().
2388 get_task_struct(owner);
2393 mutex_lock(&owner->perf_event_mutex);
2395 * We have to re-check the event->owner field, if it is cleared
2396 * we raced with perf_event_exit_task(), acquiring the mutex
2397 * ensured they're done, and we can proceed with freeing the
2401 list_del_init(&event->owner_entry);
2402 mutex_unlock(&owner->perf_event_mutex);
2403 put_task_struct(owner);
2406 return perf_event_release_kernel(event);
2409 u64 perf_event_read_value(struct perf_event *event, u64 *enabled, u64 *running)
2411 struct perf_event *child;
2417 mutex_lock(&event->child_mutex);
2418 total += perf_event_read(event);
2419 *enabled += event->total_time_enabled +
2420 atomic64_read(&event->child_total_time_enabled);
2421 *running += event->total_time_running +
2422 atomic64_read(&event->child_total_time_running);
2424 list_for_each_entry(child, &event->child_list, child_list) {
2425 total += perf_event_read(child);
2426 *enabled += child->total_time_enabled;
2427 *running += child->total_time_running;
2429 mutex_unlock(&event->child_mutex);
2433 EXPORT_SYMBOL_GPL(perf_event_read_value);
2435 static int perf_event_read_group(struct perf_event *event,
2436 u64 read_format, char __user *buf)
2438 struct perf_event *leader = event->group_leader, *sub;
2439 int n = 0, size = 0, ret = -EFAULT;
2440 struct perf_event_context *ctx = leader->ctx;
2442 u64 count, enabled, running;
2444 mutex_lock(&ctx->mutex);
2445 count = perf_event_read_value(leader, &enabled, &running);
2447 values[n++] = 1 + leader->nr_siblings;
2448 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
2449 values[n++] = enabled;
2450 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
2451 values[n++] = running;
2452 values[n++] = count;
2453 if (read_format & PERF_FORMAT_ID)
2454 values[n++] = primary_event_id(leader);
2456 size = n * sizeof(u64);
2458 if (copy_to_user(buf, values, size))
2463 list_for_each_entry(sub, &leader->sibling_list, group_entry) {
2466 values[n++] = perf_event_read_value(sub, &enabled, &running);
2467 if (read_format & PERF_FORMAT_ID)
2468 values[n++] = primary_event_id(sub);
2470 size = n * sizeof(u64);
2472 if (copy_to_user(buf + ret, values, size)) {
2480 mutex_unlock(&ctx->mutex);
2485 static int perf_event_read_one(struct perf_event *event,
2486 u64 read_format, char __user *buf)
2488 u64 enabled, running;
2492 values[n++] = perf_event_read_value(event, &enabled, &running);
2493 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
2494 values[n++] = enabled;
2495 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
2496 values[n++] = running;
2497 if (read_format & PERF_FORMAT_ID)
2498 values[n++] = primary_event_id(event);
2500 if (copy_to_user(buf, values, n * sizeof(u64)))
2503 return n * sizeof(u64);
2507 * Read the performance event - simple non blocking version for now
2510 perf_read_hw(struct perf_event *event, char __user *buf, size_t count)
2512 u64 read_format = event->attr.read_format;
2516 * Return end-of-file for a read on a event that is in
2517 * error state (i.e. because it was pinned but it couldn't be
2518 * scheduled on to the CPU at some point).
2520 if (event->state == PERF_EVENT_STATE_ERROR)
2523 if (count < event->read_size)
2526 WARN_ON_ONCE(event->ctx->parent_ctx);
2527 if (read_format & PERF_FORMAT_GROUP)
2528 ret = perf_event_read_group(event, read_format, buf);
2530 ret = perf_event_read_one(event, read_format, buf);
2536 perf_read(struct file *file, char __user *buf, size_t count, loff_t *ppos)
2538 struct perf_event *event = file->private_data;
2540 return perf_read_hw(event, buf, count);
2543 static unsigned int perf_poll(struct file *file, poll_table *wait)
2545 struct perf_event *event = file->private_data;
2546 struct perf_buffer *buffer;
2547 unsigned int events = POLL_HUP;
2550 buffer = rcu_dereference(event->buffer);
2552 events = atomic_xchg(&buffer->poll, 0);
2555 poll_wait(file, &event->waitq, wait);
2560 static void perf_event_reset(struct perf_event *event)
2562 (void)perf_event_read(event);
2563 local64_set(&event->count, 0);
2564 perf_event_update_userpage(event);
2568 * Holding the top-level event's child_mutex means that any
2569 * descendant process that has inherited this event will block
2570 * in sync_child_event if it goes to exit, thus satisfying the
2571 * task existence requirements of perf_event_enable/disable.
2573 static void perf_event_for_each_child(struct perf_event *event,
2574 void (*func)(struct perf_event *))
2576 struct perf_event *child;
2578 WARN_ON_ONCE(event->ctx->parent_ctx);
2579 mutex_lock(&event->child_mutex);
2581 list_for_each_entry(child, &event->child_list, child_list)
2583 mutex_unlock(&event->child_mutex);
2586 static void perf_event_for_each(struct perf_event *event,
2587 void (*func)(struct perf_event *))
2589 struct perf_event_context *ctx = event->ctx;
2590 struct perf_event *sibling;
2592 WARN_ON_ONCE(ctx->parent_ctx);
2593 mutex_lock(&ctx->mutex);
2594 event = event->group_leader;
2596 perf_event_for_each_child(event, func);
2598 list_for_each_entry(sibling, &event->sibling_list, group_entry)
2599 perf_event_for_each_child(event, func);
2600 mutex_unlock(&ctx->mutex);
2603 static int perf_event_period(struct perf_event *event, u64 __user *arg)
2605 struct perf_event_context *ctx = event->ctx;
2609 if (!is_sampling_event(event))
2612 if (copy_from_user(&value, arg, sizeof(value)))
2618 raw_spin_lock_irq(&ctx->lock);
2619 if (event->attr.freq) {
2620 if (value > sysctl_perf_event_sample_rate) {
2625 event->attr.sample_freq = value;
2627 event->attr.sample_period = value;
2628 event->hw.sample_period = value;
2631 raw_spin_unlock_irq(&ctx->lock);
2636 static const struct file_operations perf_fops;
2638 static struct perf_event *perf_fget_light(int fd, int *fput_needed)
2642 file = fget_light(fd, fput_needed);
2644 return ERR_PTR(-EBADF);
2646 if (file->f_op != &perf_fops) {
2647 fput_light(file, *fput_needed);
2649 return ERR_PTR(-EBADF);
2652 return file->private_data;
2655 static int perf_event_set_output(struct perf_event *event,
2656 struct perf_event *output_event);
2657 static int perf_event_set_filter(struct perf_event *event, void __user *arg);
2659 static long perf_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
2661 struct perf_event *event = file->private_data;
2662 void (*func)(struct perf_event *);
2666 case PERF_EVENT_IOC_ENABLE:
2667 func = perf_event_enable;
2669 case PERF_EVENT_IOC_DISABLE:
2670 func = perf_event_disable;
2672 case PERF_EVENT_IOC_RESET:
2673 func = perf_event_reset;
2676 case PERF_EVENT_IOC_REFRESH:
2677 return perf_event_refresh(event, arg);
2679 case PERF_EVENT_IOC_PERIOD:
2680 return perf_event_period(event, (u64 __user *)arg);
2682 case PERF_EVENT_IOC_SET_OUTPUT:
2684 struct perf_event *output_event = NULL;
2685 int fput_needed = 0;
2689 output_event = perf_fget_light(arg, &fput_needed);
2690 if (IS_ERR(output_event))
2691 return PTR_ERR(output_event);
2694 ret = perf_event_set_output(event, output_event);
2696 fput_light(output_event->filp, fput_needed);
2701 case PERF_EVENT_IOC_SET_FILTER:
2702 return perf_event_set_filter(event, (void __user *)arg);
2708 if (flags & PERF_IOC_FLAG_GROUP)
2709 perf_event_for_each(event, func);
2711 perf_event_for_each_child(event, func);
2716 int perf_event_task_enable(void)
2718 struct perf_event *event;
2720 mutex_lock(¤t->perf_event_mutex);
2721 list_for_each_entry(event, ¤t->perf_event_list, owner_entry)
2722 perf_event_for_each_child(event, perf_event_enable);
2723 mutex_unlock(¤t->perf_event_mutex);
2728 int perf_event_task_disable(void)
2730 struct perf_event *event;
2732 mutex_lock(¤t->perf_event_mutex);
2733 list_for_each_entry(event, ¤t->perf_event_list, owner_entry)
2734 perf_event_for_each_child(event, perf_event_disable);
2735 mutex_unlock(¤t->perf_event_mutex);
2740 #ifndef PERF_EVENT_INDEX_OFFSET
2741 # define PERF_EVENT_INDEX_OFFSET 0
2744 static int perf_event_index(struct perf_event *event)
2746 if (event->hw.state & PERF_HES_STOPPED)
2749 if (event->state != PERF_EVENT_STATE_ACTIVE)
2752 return event->hw.idx + 1 - PERF_EVENT_INDEX_OFFSET;
2756 * Callers need to ensure there can be no nesting of this function, otherwise
2757 * the seqlock logic goes bad. We can not serialize this because the arch
2758 * code calls this from NMI context.
2760 void perf_event_update_userpage(struct perf_event *event)
2762 struct perf_event_mmap_page *userpg;
2763 struct perf_buffer *buffer;
2766 buffer = rcu_dereference(event->buffer);
2770 userpg = buffer->user_page;
2773 * Disable preemption so as to not let the corresponding user-space
2774 * spin too long if we get preempted.
2779 userpg->index = perf_event_index(event);
2780 userpg->offset = perf_event_count(event);
2781 if (event->state == PERF_EVENT_STATE_ACTIVE)
2782 userpg->offset -= local64_read(&event->hw.prev_count);
2784 userpg->time_enabled = event->total_time_enabled +
2785 atomic64_read(&event->child_total_time_enabled);
2787 userpg->time_running = event->total_time_running +
2788 atomic64_read(&event->child_total_time_running);
2797 static unsigned long perf_data_size(struct perf_buffer *buffer);
2800 perf_buffer_init(struct perf_buffer *buffer, long watermark, int flags)
2802 long max_size = perf_data_size(buffer);
2805 buffer->watermark = min(max_size, watermark);
2807 if (!buffer->watermark)
2808 buffer->watermark = max_size / 2;
2810 if (flags & PERF_BUFFER_WRITABLE)
2811 buffer->writable = 1;
2813 atomic_set(&buffer->refcount, 1);
2816 #ifndef CONFIG_PERF_USE_VMALLOC
2819 * Back perf_mmap() with regular GFP_KERNEL-0 pages.
2822 static struct page *
2823 perf_mmap_to_page(struct perf_buffer *buffer, unsigned long pgoff)
2825 if (pgoff > buffer->nr_pages)
2829 return virt_to_page(buffer->user_page);
2831 return virt_to_page(buffer->data_pages[pgoff - 1]);
2834 static void *perf_mmap_alloc_page(int cpu)
2839 node = (cpu == -1) ? cpu : cpu_to_node(cpu);
2840 page = alloc_pages_node(node, GFP_KERNEL | __GFP_ZERO, 0);
2844 return page_address(page);
2847 static struct perf_buffer *
2848 perf_buffer_alloc(int nr_pages, long watermark, int cpu, int flags)
2850 struct perf_buffer *buffer;
2854 size = sizeof(struct perf_buffer);
2855 size += nr_pages * sizeof(void *);
2857 buffer = kzalloc(size, GFP_KERNEL);
2861 buffer->user_page = perf_mmap_alloc_page(cpu);
2862 if (!buffer->user_page)
2863 goto fail_user_page;
2865 for (i = 0; i < nr_pages; i++) {
2866 buffer->data_pages[i] = perf_mmap_alloc_page(cpu);
2867 if (!buffer->data_pages[i])
2868 goto fail_data_pages;
2871 buffer->nr_pages = nr_pages;
2873 perf_buffer_init(buffer, watermark, flags);
2878 for (i--; i >= 0; i--)
2879 free_page((unsigned long)buffer->data_pages[i]);
2881 free_page((unsigned long)buffer->user_page);
2890 static void perf_mmap_free_page(unsigned long addr)
2892 struct page *page = virt_to_page((void *)addr);
2894 page->mapping = NULL;
2898 static void perf_buffer_free(struct perf_buffer *buffer)
2902 perf_mmap_free_page((unsigned long)buffer->user_page);
2903 for (i = 0; i < buffer->nr_pages; i++)
2904 perf_mmap_free_page((unsigned long)buffer->data_pages[i]);
2908 static inline int page_order(struct perf_buffer *buffer)
2916 * Back perf_mmap() with vmalloc memory.
2918 * Required for architectures that have d-cache aliasing issues.
2921 static inline int page_order(struct perf_buffer *buffer)
2923 return buffer->page_order;
2926 static struct page *
2927 perf_mmap_to_page(struct perf_buffer *buffer, unsigned long pgoff)
2929 if (pgoff > (1UL << page_order(buffer)))
2932 return vmalloc_to_page((void *)buffer->user_page + pgoff * PAGE_SIZE);
2935 static void perf_mmap_unmark_page(void *addr)
2937 struct page *page = vmalloc_to_page(addr);
2939 page->mapping = NULL;
2942 static void perf_buffer_free_work(struct work_struct *work)
2944 struct perf_buffer *buffer;
2948 buffer = container_of(work, struct perf_buffer, work);
2949 nr = 1 << page_order(buffer);
2951 base = buffer->user_page;
2952 for (i = 0; i < nr + 1; i++)
2953 perf_mmap_unmark_page(base + (i * PAGE_SIZE));
2959 static void perf_buffer_free(struct perf_buffer *buffer)
2961 schedule_work(&buffer->work);
2964 static struct perf_buffer *
2965 perf_buffer_alloc(int nr_pages, long watermark, int cpu, int flags)
2967 struct perf_buffer *buffer;
2971 size = sizeof(struct perf_buffer);
2972 size += sizeof(void *);
2974 buffer = kzalloc(size, GFP_KERNEL);
2978 INIT_WORK(&buffer->work, perf_buffer_free_work);
2980 all_buf = vmalloc_user((nr_pages + 1) * PAGE_SIZE);
2984 buffer->user_page = all_buf;
2985 buffer->data_pages[0] = all_buf + PAGE_SIZE;
2986 buffer->page_order = ilog2(nr_pages);
2987 buffer->nr_pages = 1;
2989 perf_buffer_init(buffer, watermark, flags);
3002 static unsigned long perf_data_size(struct perf_buffer *buffer)
3004 return buffer->nr_pages << (PAGE_SHIFT + page_order(buffer));
3007 static int perf_mmap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
3009 struct perf_event *event = vma->vm_file->private_data;
3010 struct perf_buffer *buffer;
3011 int ret = VM_FAULT_SIGBUS;
3013 if (vmf->flags & FAULT_FLAG_MKWRITE) {
3014 if (vmf->pgoff == 0)
3020 buffer = rcu_dereference(event->buffer);
3024 if (vmf->pgoff && (vmf->flags & FAULT_FLAG_WRITE))
3027 vmf->page = perf_mmap_to_page(buffer, vmf->pgoff);
3031 get_page(vmf->page);
3032 vmf->page->mapping = vma->vm_file->f_mapping;
3033 vmf->page->index = vmf->pgoff;
3042 static void perf_buffer_free_rcu(struct rcu_head *rcu_head)
3044 struct perf_buffer *buffer;
3046 buffer = container_of(rcu_head, struct perf_buffer, rcu_head);
3047 perf_buffer_free(buffer);
3050 static struct perf_buffer *perf_buffer_get(struct perf_event *event)
3052 struct perf_buffer *buffer;
3055 buffer = rcu_dereference(event->buffer);
3057 if (!atomic_inc_not_zero(&buffer->refcount))
3065 static void perf_buffer_put(struct perf_buffer *buffer)
3067 if (!atomic_dec_and_test(&buffer->refcount))
3070 call_rcu(&buffer->rcu_head, perf_buffer_free_rcu);
3073 static void perf_mmap_open(struct vm_area_struct *vma)
3075 struct perf_event *event = vma->vm_file->private_data;
3077 atomic_inc(&event->mmap_count);
3080 static void perf_mmap_close(struct vm_area_struct *vma)
3082 struct perf_event *event = vma->vm_file->private_data;
3084 if (atomic_dec_and_mutex_lock(&event->mmap_count, &event->mmap_mutex)) {
3085 unsigned long size = perf_data_size(event->buffer);
3086 struct user_struct *user = event->mmap_user;
3087 struct perf_buffer *buffer = event->buffer;
3089 atomic_long_sub((size >> PAGE_SHIFT) + 1, &user->locked_vm);
3090 vma->vm_mm->locked_vm -= event->mmap_locked;
3091 rcu_assign_pointer(event->buffer, NULL);
3092 mutex_unlock(&event->mmap_mutex);
3094 perf_buffer_put(buffer);
3099 static const struct vm_operations_struct perf_mmap_vmops = {
3100 .open = perf_mmap_open,
3101 .close = perf_mmap_close,
3102 .fault = perf_mmap_fault,
3103 .page_mkwrite = perf_mmap_fault,
3106 static int perf_mmap(struct file *file, struct vm_area_struct *vma)
3108 struct perf_event *event = file->private_data;
3109 unsigned long user_locked, user_lock_limit;
3110 struct user_struct *user = current_user();
3111 unsigned long locked, lock_limit;
3112 struct perf_buffer *buffer;
3113 unsigned long vma_size;
3114 unsigned long nr_pages;
3115 long user_extra, extra;
3116 int ret = 0, flags = 0;
3119 * Don't allow mmap() of inherited per-task counters. This would
3120 * create a performance issue due to all children writing to the
3123 if (event->cpu == -1 && event->attr.inherit)
3126 if (!(vma->vm_flags & VM_SHARED))
3129 vma_size = vma->vm_end - vma->vm_start;
3130 nr_pages = (vma_size / PAGE_SIZE) - 1;
3133 * If we have buffer pages ensure they're a power-of-two number, so we
3134 * can do bitmasks instead of modulo.
3136 if (nr_pages != 0 && !is_power_of_2(nr_pages))
3139 if (vma_size != PAGE_SIZE * (1 + nr_pages))
3142 if (vma->vm_pgoff != 0)
3145 WARN_ON_ONCE(event->ctx->parent_ctx);
3146 mutex_lock(&event->mmap_mutex);
3147 if (event->buffer) {
3148 if (event->buffer->nr_pages == nr_pages)
3149 atomic_inc(&event->buffer->refcount);
3155 user_extra = nr_pages + 1;
3156 user_lock_limit = sysctl_perf_event_mlock >> (PAGE_SHIFT - 10);
3159 * Increase the limit linearly with more CPUs:
3161 user_lock_limit *= num_online_cpus();
3163 user_locked = atomic_long_read(&user->locked_vm) + user_extra;
3166 if (user_locked > user_lock_limit)
3167 extra = user_locked - user_lock_limit;
3169 lock_limit = rlimit(RLIMIT_MEMLOCK);
3170 lock_limit >>= PAGE_SHIFT;
3171 locked = vma->vm_mm->locked_vm + extra;
3173 if ((locked > lock_limit) && perf_paranoid_tracepoint_raw() &&
3174 !capable(CAP_IPC_LOCK)) {
3179 WARN_ON(event->buffer);
3181 if (vma->vm_flags & VM_WRITE)
3182 flags |= PERF_BUFFER_WRITABLE;
3184 buffer = perf_buffer_alloc(nr_pages, event->attr.wakeup_watermark,
3190 rcu_assign_pointer(event->buffer, buffer);
3192 atomic_long_add(user_extra, &user->locked_vm);
3193 event->mmap_locked = extra;
3194 event->mmap_user = get_current_user();
3195 vma->vm_mm->locked_vm += event->mmap_locked;
3199 atomic_inc(&event->mmap_count);
3200 mutex_unlock(&event->mmap_mutex);
3202 vma->vm_flags |= VM_RESERVED;
3203 vma->vm_ops = &perf_mmap_vmops;
3208 static int perf_fasync(int fd, struct file *filp, int on)
3210 struct inode *inode = filp->f_path.dentry->d_inode;
3211 struct perf_event *event = filp->private_data;
3214 mutex_lock(&inode->i_mutex);
3215 retval = fasync_helper(fd, filp, on, &event->fasync);
3216 mutex_unlock(&inode->i_mutex);
3224 static const struct file_operations perf_fops = {
3225 .llseek = no_llseek,
3226 .release = perf_release,
3229 .unlocked_ioctl = perf_ioctl,
3230 .compat_ioctl = perf_ioctl,
3232 .fasync = perf_fasync,
3238 * If there's data, ensure we set the poll() state and publish everything
3239 * to user-space before waking everybody up.
3242 void perf_event_wakeup(struct perf_event *event)
3244 wake_up_all(&event->waitq);
3246 if (event->pending_kill) {
3247 kill_fasync(&event->fasync, SIGIO, event->pending_kill);
3248 event->pending_kill = 0;
3252 static void perf_pending_event(struct irq_work *entry)
3254 struct perf_event *event = container_of(entry,
3255 struct perf_event, pending);
3257 if (event->pending_disable) {
3258 event->pending_disable = 0;
3259 __perf_event_disable(event);
3262 if (event->pending_wakeup) {
3263 event->pending_wakeup = 0;
3264 perf_event_wakeup(event);
3269 * We assume there is only KVM supporting the callbacks.
3270 * Later on, we might change it to a list if there is
3271 * another virtualization implementation supporting the callbacks.
3273 struct perf_guest_info_callbacks *perf_guest_cbs;
3275 int perf_register_guest_info_callbacks(struct perf_guest_info_callbacks *cbs)
3277 perf_guest_cbs = cbs;
3280 EXPORT_SYMBOL_GPL(perf_register_guest_info_callbacks);
3282 int perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks *cbs)
3284 perf_guest_cbs = NULL;
3287 EXPORT_SYMBOL_GPL(perf_unregister_guest_info_callbacks);
3292 static bool perf_output_space(struct perf_buffer *buffer, unsigned long tail,
3293 unsigned long offset, unsigned long head)
3297 if (!buffer->writable)
3300 mask = perf_data_size(buffer) - 1;
3302 offset = (offset - tail) & mask;
3303 head = (head - tail) & mask;
3305 if ((int)(head - offset) < 0)
3311 static void perf_output_wakeup(struct perf_output_handle *handle)
3313 atomic_set(&handle->buffer->poll, POLL_IN);
3316 handle->event->pending_wakeup = 1;
3317 irq_work_queue(&handle->event->pending);
3319 perf_event_wakeup(handle->event);
3323 * We need to ensure a later event_id doesn't publish a head when a former
3324 * event isn't done writing. However since we need to deal with NMIs we
3325 * cannot fully serialize things.
3327 * We only publish the head (and generate a wakeup) when the outer-most
3330 static void perf_output_get_handle(struct perf_output_handle *handle)
3332 struct perf_buffer *buffer = handle->buffer;
3335 local_inc(&buffer->nest);
3336 handle->wakeup = local_read(&buffer->wakeup);
3339 static void perf_output_put_handle(struct perf_output_handle *handle)
3341 struct perf_buffer *buffer = handle->buffer;
3345 head = local_read(&buffer->head);
3348 * IRQ/NMI can happen here, which means we can miss a head update.
3351 if (!local_dec_and_test(&buffer->nest))
3355 * Publish the known good head. Rely on the full barrier implied
3356 * by atomic_dec_and_test() order the buffer->head read and this
3359 buffer->user_page->data_head = head;
3362 * Now check if we missed an update, rely on the (compiler)
3363 * barrier in atomic_dec_and_test() to re-read buffer->head.
3365 if (unlikely(head != local_read(&buffer->head))) {
3366 local_inc(&buffer->nest);
3370 if (handle->wakeup != local_read(&buffer->wakeup))
3371 perf_output_wakeup(handle);
3377 __always_inline void perf_output_copy(struct perf_output_handle *handle,
3378 const void *buf, unsigned int len)
3381 unsigned long size = min_t(unsigned long, handle->size, len);
3383 memcpy(handle->addr, buf, size);
3386 handle->addr += size;
3388 handle->size -= size;
3389 if (!handle->size) {
3390 struct perf_buffer *buffer = handle->buffer;
3393 handle->page &= buffer->nr_pages - 1;
3394 handle->addr = buffer->data_pages[handle->page];
3395 handle->size = PAGE_SIZE << page_order(buffer);
3400 static void __perf_event_header__init_id(struct perf_event_header *header,
3401 struct perf_sample_data *data,
3402 struct perf_event *event)
3404 u64 sample_type = event->attr.sample_type;
3406 data->type = sample_type;
3407 header->size += event->id_header_size;
3409 if (sample_type & PERF_SAMPLE_TID) {
3410 /* namespace issues */
3411 data->tid_entry.pid = perf_event_pid(event, current);
3412 data->tid_entry.tid = perf_event_tid(event, current);
3415 if (sample_type & PERF_SAMPLE_TIME)
3416 data->time = perf_clock();
3418 if (sample_type & PERF_SAMPLE_ID)
3419 data->id = primary_event_id(event);
3421 if (sample_type & PERF_SAMPLE_STREAM_ID)
3422 data->stream_id = event->id;
3424 if (sample_type & PERF_SAMPLE_CPU) {
3425 data->cpu_entry.cpu = raw_smp_processor_id();
3426 data->cpu_entry.reserved = 0;
3430 static void perf_event_header__init_id(struct perf_event_header *header,
3431 struct perf_sample_data *data,
3432 struct perf_event *event)
3434 if (event->attr.sample_id_all)
3435 __perf_event_header__init_id(header, data, event);
3438 static void __perf_event__output_id_sample(struct perf_output_handle *handle,
3439 struct perf_sample_data *data)
3441 u64 sample_type = data->type;
3443 if (sample_type & PERF_SAMPLE_TID)
3444 perf_output_put(handle, data->tid_entry);
3446 if (sample_type & PERF_SAMPLE_TIME)
3447 perf_output_put(handle, data->time);
3449 if (sample_type & PERF_SAMPLE_ID)
3450 perf_output_put(handle, data->id);
3452 if (sample_type & PERF_SAMPLE_STREAM_ID)
3453 perf_output_put(handle, data->stream_id);
3455 if (sample_type & PERF_SAMPLE_CPU)
3456 perf_output_put(handle, data->cpu_entry);
3459 static void perf_event__output_id_sample(struct perf_event *event,
3460 struct perf_output_handle *handle,
3461 struct perf_sample_data *sample)
3463 if (event->attr.sample_id_all)
3464 __perf_event__output_id_sample(handle, sample);
3467 int perf_output_begin(struct perf_output_handle *handle,
3468 struct perf_event *event, unsigned int size,
3469 int nmi, int sample)
3471 struct perf_buffer *buffer;
3472 unsigned long tail, offset, head;
3474 struct perf_sample_data sample_data;
3476 struct perf_event_header header;
3483 * For inherited events we send all the output towards the parent.
3486 event = event->parent;
3488 buffer = rcu_dereference(event->buffer);
3492 handle->buffer = buffer;
3493 handle->event = event;
3495 handle->sample = sample;
3497 if (!buffer->nr_pages)
3500 have_lost = local_read(&buffer->lost);
3502 lost_event.header.size = sizeof(lost_event);
3503 perf_event_header__init_id(&lost_event.header, &sample_data,
3505 size += lost_event.header.size;
3508 perf_output_get_handle(handle);
3512 * Userspace could choose to issue a mb() before updating the
3513 * tail pointer. So that all reads will be completed before the
3516 tail = ACCESS_ONCE(buffer->user_page->data_tail);
3518 offset = head = local_read(&buffer->head);
3520 if (unlikely(!perf_output_space(buffer, tail, offset, head)))
3522 } while (local_cmpxchg(&buffer->head, offset, head) != offset);
3524 if (head - local_read(&buffer->wakeup) > buffer->watermark)
3525 local_add(buffer->watermark, &buffer->wakeup);
3527 handle->page = offset >> (PAGE_SHIFT + page_order(buffer));
3528 handle->page &= buffer->nr_pages - 1;
3529 handle->size = offset & ((PAGE_SIZE << page_order(buffer)) - 1);
3530 handle->addr = buffer->data_pages[handle->page];
3531 handle->addr += handle->size;
3532 handle->size = (PAGE_SIZE << page_order(buffer)) - handle->size;
3535 lost_event.header.type = PERF_RECORD_LOST;
3536 lost_event.header.misc = 0;
3537 lost_event.id = event->id;
3538 lost_event.lost = local_xchg(&buffer->lost, 0);
3540 perf_output_put(handle, lost_event);
3541 perf_event__output_id_sample(event, handle, &sample_data);
3547 local_inc(&buffer->lost);
3548 perf_output_put_handle(handle);
3555 void perf_output_end(struct perf_output_handle *handle)
3557 struct perf_event *event = handle->event;
3558 struct perf_buffer *buffer = handle->buffer;
3560 int wakeup_events = event->attr.wakeup_events;
3562 if (handle->sample && wakeup_events) {
3563 int events = local_inc_return(&buffer->events);
3564 if (events >= wakeup_events) {
3565 local_sub(wakeup_events, &buffer->events);
3566 local_inc(&buffer->wakeup);
3570 perf_output_put_handle(handle);
3574 static void perf_output_read_one(struct perf_output_handle *handle,
3575 struct perf_event *event,
3576 u64 enabled, u64 running)
3578 u64 read_format = event->attr.read_format;
3582 values[n++] = perf_event_count(event);
3583 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
3584 values[n++] = enabled +
3585 atomic64_read(&event->child_total_time_enabled);
3587 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
3588 values[n++] = running +
3589 atomic64_read(&event->child_total_time_running);
3591 if (read_format & PERF_FORMAT_ID)
3592 values[n++] = primary_event_id(event);
3594 perf_output_copy(handle, values, n * sizeof(u64));
3598 * XXX PERF_FORMAT_GROUP vs inherited events seems difficult.
3600 static void perf_output_read_group(struct perf_output_handle *handle,
3601 struct perf_event *event,
3602 u64 enabled, u64 running)
3604 struct perf_event *leader = event->group_leader, *sub;
3605 u64 read_format = event->attr.read_format;
3609 values[n++] = 1 + leader->nr_siblings;
3611 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
3612 values[n++] = enabled;
3614 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
3615 values[n++] = running;
3617 if (leader != event)
3618 leader->pmu->read(leader);
3620 values[n++] = perf_event_count(leader);
3621 if (read_format & PERF_FORMAT_ID)
3622 values[n++] = primary_event_id(leader);
3624 perf_output_copy(handle, values, n * sizeof(u64));
3626 list_for_each_entry(sub, &leader->sibling_list, group_entry) {
3630 sub->pmu->read(sub);
3632 values[n++] = perf_event_count(sub);
3633 if (read_format & PERF_FORMAT_ID)
3634 values[n++] = primary_event_id(sub);
3636 perf_output_copy(handle, values, n * sizeof(u64));
3640 #define PERF_FORMAT_TOTAL_TIMES (PERF_FORMAT_TOTAL_TIME_ENABLED|\
3641 PERF_FORMAT_TOTAL_TIME_RUNNING)
3643 static void perf_output_read(struct perf_output_handle *handle,
3644 struct perf_event *event)
3646 u64 enabled = 0, running = 0, now, ctx_time;
3647 u64 read_format = event->attr.read_format;
3650 * compute total_time_enabled, total_time_running
3651 * based on snapshot values taken when the event
3652 * was last scheduled in.
3654 * we cannot simply called update_context_time()
3655 * because of locking issue as we are called in
3658 if (read_format & PERF_FORMAT_TOTAL_TIMES) {
3660 ctx_time = event->shadow_ctx_time + now;
3661 enabled = ctx_time - event->tstamp_enabled;
3662 running = ctx_time - event->tstamp_running;
3665 if (event->attr.read_format & PERF_FORMAT_GROUP)
3666 perf_output_read_group(handle, event, enabled, running);
3668 perf_output_read_one(handle, event, enabled, running);
3671 void perf_output_sample(struct perf_output_handle *handle,
3672 struct perf_event_header *header,
3673 struct perf_sample_data *data,
3674 struct perf_event *event)
3676 u64 sample_type = data->type;
3678 perf_output_put(handle, *header);
3680 if (sample_type & PERF_SAMPLE_IP)
3681 perf_output_put(handle, data->ip);
3683 if (sample_type & PERF_SAMPLE_TID)
3684 perf_output_put(handle, data->tid_entry);
3686 if (sample_type & PERF_SAMPLE_TIME)
3687 perf_output_put(handle, data->time);
3689 if (sample_type & PERF_SAMPLE_ADDR)
3690 perf_output_put(handle, data->addr);
3692 if (sample_type & PERF_SAMPLE_ID)
3693 perf_output_put(handle, data->id);
3695 if (sample_type & PERF_SAMPLE_STREAM_ID)
3696 perf_output_put(handle, data->stream_id);
3698 if (sample_type & PERF_SAMPLE_CPU)
3699 perf_output_put(handle, data->cpu_entry);
3701 if (sample_type & PERF_SAMPLE_PERIOD)
3702 perf_output_put(handle, data->period);
3704 if (sample_type & PERF_SAMPLE_READ)
3705 perf_output_read(handle, event);
3707 if (sample_type & PERF_SAMPLE_CALLCHAIN) {
3708 if (data->callchain) {
3711 if (data->callchain)
3712 size += data->callchain->nr;
3714 size *= sizeof(u64);
3716 perf_output_copy(handle, data->callchain, size);
3719 perf_output_put(handle, nr);
3723 if (sample_type & PERF_SAMPLE_RAW) {
3725 perf_output_put(handle, data->raw->size);
3726 perf_output_copy(handle, data->raw->data,
3733 .size = sizeof(u32),
3736 perf_output_put(handle, raw);
3741 void perf_prepare_sample(struct perf_event_header *header,
3742 struct perf_sample_data *data,
3743 struct perf_event *event,
3744 struct pt_regs *regs)
3746 u64 sample_type = event->attr.sample_type;
3748 header->type = PERF_RECORD_SAMPLE;
3749 header->size = sizeof(*header) + event->header_size;
3752 header->misc |= perf_misc_flags(regs);
3754 __perf_event_header__init_id(header, data, event);
3756 if (sample_type & PERF_SAMPLE_IP)
3757 data->ip = perf_instruction_pointer(regs);
3759 if (sample_type & PERF_SAMPLE_CALLCHAIN) {
3762 data->callchain = perf_callchain(regs);
3764 if (data->callchain)
3765 size += data->callchain->nr;
3767 header->size += size * sizeof(u64);
3770 if (sample_type & PERF_SAMPLE_RAW) {
3771 int size = sizeof(u32);
3774 size += data->raw->size;
3776 size += sizeof(u32);
3778 WARN_ON_ONCE(size & (sizeof(u64)-1));
3779 header->size += size;
3783 static void perf_event_output(struct perf_event *event, int nmi,
3784 struct perf_sample_data *data,
3785 struct pt_regs *regs)
3787 struct perf_output_handle handle;
3788 struct perf_event_header header;
3790 /* protect the callchain buffers */
3793 perf_prepare_sample(&header, data, event, regs);
3795 if (perf_output_begin(&handle, event, header.size, nmi, 1))
3798 perf_output_sample(&handle, &header, data, event);
3800 perf_output_end(&handle);
3810 struct perf_read_event {
3811 struct perf_event_header header;
3818 perf_event_read_event(struct perf_event *event,
3819 struct task_struct *task)
3821 struct perf_output_handle handle;
3822 struct perf_sample_data sample;
3823 struct perf_read_event read_event = {
3825 .type = PERF_RECORD_READ,
3827 .size = sizeof(read_event) + event->read_size,
3829 .pid = perf_event_pid(event, task),
3830 .tid = perf_event_tid(event, task),
3834 perf_event_header__init_id(&read_event.header, &sample, event);
3835 ret = perf_output_begin(&handle, event, read_event.header.size, 0, 0);
3839 perf_output_put(&handle, read_event);
3840 perf_output_read(&handle, event);
3841 perf_event__output_id_sample(event, &handle, &sample);
3843 perf_output_end(&handle);
3847 * task tracking -- fork/exit
3849 * enabled by: attr.comm | attr.mmap | attr.mmap_data | attr.task
3852 struct perf_task_event {
3853 struct task_struct *task;
3854 struct perf_event_context *task_ctx;
3857 struct perf_event_header header;
3867 static void perf_event_task_output(struct perf_event *event,
3868 struct perf_task_event *task_event)
3870 struct perf_output_handle handle;
3871 struct perf_sample_data sample;
3872 struct task_struct *task = task_event->task;
3873 int ret, size = task_event->event_id.header.size;
3875 perf_event_header__init_id(&task_event->event_id.header, &sample, event);
3877 ret = perf_output_begin(&handle, event,
3878 task_event->event_id.header.size, 0, 0);
3882 task_event->event_id.pid = perf_event_pid(event, task);
3883 task_event->event_id.ppid = perf_event_pid(event, current);
3885 task_event->event_id.tid = perf_event_tid(event, task);
3886 task_event->event_id.ptid = perf_event_tid(event, current);
3888 perf_output_put(&handle, task_event->event_id);
3890 perf_event__output_id_sample(event, &handle, &sample);
3892 perf_output_end(&handle);
3894 task_event->event_id.header.size = size;
3897 static int perf_event_task_match(struct perf_event *event)
3899 if (event->state < PERF_EVENT_STATE_INACTIVE)
3902 if (!event_filter_match(event))
3905 if (event->attr.comm || event->attr.mmap ||
3906 event->attr.mmap_data || event->attr.task)
3912 static void perf_event_task_ctx(struct perf_event_context *ctx,
3913 struct perf_task_event *task_event)
3915 struct perf_event *event;
3917 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
3918 if (perf_event_task_match(event))
3919 perf_event_task_output(event, task_event);
3923 static void perf_event_task_event(struct perf_task_event *task_event)
3925 struct perf_cpu_context *cpuctx;
3926 struct perf_event_context *ctx;
3931 list_for_each_entry_rcu(pmu, &pmus, entry) {
3932 cpuctx = get_cpu_ptr(pmu->pmu_cpu_context);
3933 if (cpuctx->active_pmu != pmu)
3935 perf_event_task_ctx(&cpuctx->ctx, task_event);
3937 ctx = task_event->task_ctx;
3939 ctxn = pmu->task_ctx_nr;
3942 ctx = rcu_dereference(current->perf_event_ctxp[ctxn]);
3945 perf_event_task_ctx(ctx, task_event);
3947 put_cpu_ptr(pmu->pmu_cpu_context);
3952 static void perf_event_task(struct task_struct *task,
3953 struct perf_event_context *task_ctx,
3956 struct perf_task_event task_event;
3958 if (!atomic_read(&nr_comm_events) &&
3959 !atomic_read(&nr_mmap_events) &&
3960 !atomic_read(&nr_task_events))
3963 task_event = (struct perf_task_event){
3965 .task_ctx = task_ctx,
3968 .type = new ? PERF_RECORD_FORK : PERF_RECORD_EXIT,
3970 .size = sizeof(task_event.event_id),
3976 .time = perf_clock(),
3980 perf_event_task_event(&task_event);
3983 void perf_event_fork(struct task_struct *task)
3985 perf_event_task(task, NULL, 1);
3992 struct perf_comm_event {
3993 struct task_struct *task;
3998 struct perf_event_header header;
4005 static void perf_event_comm_output(struct perf_event *event,
4006 struct perf_comm_event *comm_event)
4008 struct perf_output_handle handle;
4009 struct perf_sample_data sample;
4010 int size = comm_event->event_id.header.size;
4013 perf_event_header__init_id(&comm_event->event_id.header, &sample, event);
4014 ret = perf_output_begin(&handle, event,
4015 comm_event->event_id.header.size, 0, 0);
4020 comm_event->event_id.pid = perf_event_pid(event, comm_event->task);
4021 comm_event->event_id.tid = perf_event_tid(event, comm_event->task);
4023 perf_output_put(&handle, comm_event->event_id);
4024 perf_output_copy(&handle, comm_event->comm,
4025 comm_event->comm_size);
4027 perf_event__output_id_sample(event, &handle, &sample);
4029 perf_output_end(&handle);
4031 comm_event->event_id.header.size = size;
4034 static int perf_event_comm_match(struct perf_event *event)
4036 if (event->state < PERF_EVENT_STATE_INACTIVE)
4039 if (!event_filter_match(event))
4042 if (event->attr.comm)
4048 static void perf_event_comm_ctx(struct perf_event_context *ctx,
4049 struct perf_comm_event *comm_event)
4051 struct perf_event *event;
4053 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
4054 if (perf_event_comm_match(event))
4055 perf_event_comm_output(event, comm_event);
4059 static void perf_event_comm_event(struct perf_comm_event *comm_event)
4061 struct perf_cpu_context *cpuctx;
4062 struct perf_event_context *ctx;
4063 char comm[TASK_COMM_LEN];
4068 memset(comm, 0, sizeof(comm));
4069 strlcpy(comm, comm_event->task->comm, sizeof(comm));
4070 size = ALIGN(strlen(comm)+1, sizeof(u64));
4072 comm_event->comm = comm;
4073 comm_event->comm_size = size;
4075 comm_event->event_id.header.size = sizeof(comm_event->event_id) + size;
4077 list_for_each_entry_rcu(pmu, &pmus, entry) {
4078 cpuctx = get_cpu_ptr(pmu->pmu_cpu_context);
4079 if (cpuctx->active_pmu != pmu)
4081 perf_event_comm_ctx(&cpuctx->ctx, comm_event);
4083 ctxn = pmu->task_ctx_nr;
4087 ctx = rcu_dereference(current->perf_event_ctxp[ctxn]);
4089 perf_event_comm_ctx(ctx, comm_event);
4091 put_cpu_ptr(pmu->pmu_cpu_context);
4096 void perf_event_comm(struct task_struct *task)
4098 struct perf_comm_event comm_event;
4099 struct perf_event_context *ctx;
4102 for_each_task_context_nr(ctxn) {
4103 ctx = task->perf_event_ctxp[ctxn];
4107 perf_event_enable_on_exec(ctx);
4110 if (!atomic_read(&nr_comm_events))
4113 comm_event = (struct perf_comm_event){
4119 .type = PERF_RECORD_COMM,
4128 perf_event_comm_event(&comm_event);
4135 struct perf_mmap_event {
4136 struct vm_area_struct *vma;
4138 const char *file_name;
4142 struct perf_event_header header;
4152 static void perf_event_mmap_output(struct perf_event *event,
4153 struct perf_mmap_event *mmap_event)
4155 struct perf_output_handle handle;
4156 struct perf_sample_data sample;
4157 int size = mmap_event->event_id.header.size;
4160 perf_event_header__init_id(&mmap_event->event_id.header, &sample, event);
4161 ret = perf_output_begin(&handle, event,
4162 mmap_event->event_id.header.size, 0, 0);
4166 mmap_event->event_id.pid = perf_event_pid(event, current);
4167 mmap_event->event_id.tid = perf_event_tid(event, current);
4169 perf_output_put(&handle, mmap_event->event_id);
4170 perf_output_copy(&handle, mmap_event->file_name,
4171 mmap_event->file_size);
4173 perf_event__output_id_sample(event, &handle, &sample);
4175 perf_output_end(&handle);
4177 mmap_event->event_id.header.size = size;
4180 static int perf_event_mmap_match(struct perf_event *event,
4181 struct perf_mmap_event *mmap_event,
4184 if (event->state < PERF_EVENT_STATE_INACTIVE)
4187 if (!event_filter_match(event))
4190 if ((!executable && event->attr.mmap_data) ||
4191 (executable && event->attr.mmap))
4197 static void perf_event_mmap_ctx(struct perf_event_context *ctx,
4198 struct perf_mmap_event *mmap_event,
4201 struct perf_event *event;
4203 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
4204 if (perf_event_mmap_match(event, mmap_event, executable))
4205 perf_event_mmap_output(event, mmap_event);
4209 static void perf_event_mmap_event(struct perf_mmap_event *mmap_event)
4211 struct perf_cpu_context *cpuctx;
4212 struct perf_event_context *ctx;
4213 struct vm_area_struct *vma = mmap_event->vma;
4214 struct file *file = vma->vm_file;
4222 memset(tmp, 0, sizeof(tmp));
4226 * d_path works from the end of the buffer backwards, so we
4227 * need to add enough zero bytes after the string to handle
4228 * the 64bit alignment we do later.
4230 buf = kzalloc(PATH_MAX + sizeof(u64), GFP_KERNEL);
4232 name = strncpy(tmp, "//enomem", sizeof(tmp));
4235 name = d_path(&file->f_path, buf, PATH_MAX);
4237 name = strncpy(tmp, "//toolong", sizeof(tmp));
4241 if (arch_vma_name(mmap_event->vma)) {
4242 name = strncpy(tmp, arch_vma_name(mmap_event->vma),
4248 name = strncpy(tmp, "[vdso]", sizeof(tmp));
4250 } else if (vma->vm_start <= vma->vm_mm->start_brk &&
4251 vma->vm_end >= vma->vm_mm->brk) {
4252 name = strncpy(tmp, "[heap]", sizeof(tmp));
4254 } else if (vma->vm_start <= vma->vm_mm->start_stack &&
4255 vma->vm_end >= vma->vm_mm->start_stack) {
4256 name = strncpy(tmp, "[stack]", sizeof(tmp));
4260 name = strncpy(tmp, "//anon", sizeof(tmp));
4265 size = ALIGN(strlen(name)+1, sizeof(u64));
4267 mmap_event->file_name = name;
4268 mmap_event->file_size = size;
4270 mmap_event->event_id.header.size = sizeof(mmap_event->event_id) + size;
4273 list_for_each_entry_rcu(pmu, &pmus, entry) {
4274 cpuctx = get_cpu_ptr(pmu->pmu_cpu_context);
4275 if (cpuctx->active_pmu != pmu)
4277 perf_event_mmap_ctx(&cpuctx->ctx, mmap_event,
4278 vma->vm_flags & VM_EXEC);
4280 ctxn = pmu->task_ctx_nr;
4284 ctx = rcu_dereference(current->perf_event_ctxp[ctxn]);
4286 perf_event_mmap_ctx(ctx, mmap_event,
4287 vma->vm_flags & VM_EXEC);
4290 put_cpu_ptr(pmu->pmu_cpu_context);
4297 void perf_event_mmap(struct vm_area_struct *vma)
4299 struct perf_mmap_event mmap_event;
4301 if (!atomic_read(&nr_mmap_events))
4304 mmap_event = (struct perf_mmap_event){
4310 .type = PERF_RECORD_MMAP,
4311 .misc = PERF_RECORD_MISC_USER,
4316 .start = vma->vm_start,
4317 .len = vma->vm_end - vma->vm_start,
4318 .pgoff = (u64)vma->vm_pgoff << PAGE_SHIFT,
4322 perf_event_mmap_event(&mmap_event);
4326 * IRQ throttle logging
4329 static void perf_log_throttle(struct perf_event *event, int enable)
4331 struct perf_output_handle handle;
4332 struct perf_sample_data sample;
4336 struct perf_event_header header;
4340 } throttle_event = {
4342 .type = PERF_RECORD_THROTTLE,
4344 .size = sizeof(throttle_event),
4346 .time = perf_clock(),
4347 .id = primary_event_id(event),
4348 .stream_id = event->id,
4352 throttle_event.header.type = PERF_RECORD_UNTHROTTLE;
4354 perf_event_header__init_id(&throttle_event.header, &sample, event);
4356 ret = perf_output_begin(&handle, event,
4357 throttle_event.header.size, 1, 0);
4361 perf_output_put(&handle, throttle_event);
4362 perf_event__output_id_sample(event, &handle, &sample);
4363 perf_output_end(&handle);
4367 * Generic event overflow handling, sampling.
4370 static int __perf_event_overflow(struct perf_event *event, int nmi,
4371 int throttle, struct perf_sample_data *data,
4372 struct pt_regs *regs)
4374 int events = atomic_read(&event->event_limit);
4375 struct hw_perf_event *hwc = &event->hw;
4379 * Non-sampling counters might still use the PMI to fold short
4380 * hardware counters, ignore those.
4382 if (unlikely(!is_sampling_event(event)))
4388 if (hwc->interrupts != MAX_INTERRUPTS) {
4390 if (HZ * hwc->interrupts >
4391 (u64)sysctl_perf_event_sample_rate) {
4392 hwc->interrupts = MAX_INTERRUPTS;
4393 perf_log_throttle(event, 0);
4398 * Keep re-disabling events even though on the previous
4399 * pass we disabled it - just in case we raced with a
4400 * sched-in and the event got enabled again:
4406 if (event->attr.freq) {
4407 u64 now = perf_clock();
4408 s64 delta = now - hwc->freq_time_stamp;
4410 hwc->freq_time_stamp = now;
4412 if (delta > 0 && delta < 2*TICK_NSEC)
4413 perf_adjust_period(event, delta, hwc->last_period);
4417 * XXX event_limit might not quite work as expected on inherited
4421 event->pending_kill = POLL_IN;
4422 if (events && atomic_dec_and_test(&event->event_limit)) {
4424 event->pending_kill = POLL_HUP;
4426 event->pending_disable = 1;
4427 irq_work_queue(&event->pending);
4429 perf_event_disable(event);
4432 if (event->overflow_handler)
4433 event->overflow_handler(event, nmi, data, regs);
4435 perf_event_output(event, nmi, data, regs);
4440 int perf_event_overflow(struct perf_event *event, int nmi,
4441 struct perf_sample_data *data,
4442 struct pt_regs *regs)
4444 return __perf_event_overflow(event, nmi, 1, data, regs);
4448 * Generic software event infrastructure
4451 struct swevent_htable {
4452 struct swevent_hlist *swevent_hlist;
4453 struct mutex hlist_mutex;
4456 /* Recursion avoidance in each contexts */
4457 int recursion[PERF_NR_CONTEXTS];
4460 static DEFINE_PER_CPU(struct swevent_htable, swevent_htable);
4463 * We directly increment event->count and keep a second value in
4464 * event->hw.period_left to count intervals. This period event
4465 * is kept in the range [-sample_period, 0] so that we can use the
4469 static u64 perf_swevent_set_period(struct perf_event *event)
4471 struct hw_perf_event *hwc = &event->hw;
4472 u64 period = hwc->last_period;
4476 hwc->last_period = hwc->sample_period;
4479 old = val = local64_read(&hwc->period_left);
4483 nr = div64_u64(period + val, period);
4484 offset = nr * period;
4486 if (local64_cmpxchg(&hwc->period_left, old, val) != old)
4492 static void perf_swevent_overflow(struct perf_event *event, u64 overflow,
4493 int nmi, struct perf_sample_data *data,
4494 struct pt_regs *regs)
4496 struct hw_perf_event *hwc = &event->hw;
4499 data->period = event->hw.last_period;
4501 overflow = perf_swevent_set_period(event);
4503 if (hwc->interrupts == MAX_INTERRUPTS)
4506 for (; overflow; overflow--) {
4507 if (__perf_event_overflow(event, nmi, throttle,
4510 * We inhibit the overflow from happening when
4511 * hwc->interrupts == MAX_INTERRUPTS.
4519 static void perf_swevent_event(struct perf_event *event, u64 nr,
4520 int nmi, struct perf_sample_data *data,
4521 struct pt_regs *regs)
4523 struct hw_perf_event *hwc = &event->hw;
4525 local64_add(nr, &event->count);
4530 if (!is_sampling_event(event))
4533 if (nr == 1 && hwc->sample_period == 1 && !event->attr.freq)
4534 return perf_swevent_overflow(event, 1, nmi, data, regs);
4536 if (local64_add_negative(nr, &hwc->period_left))
4539 perf_swevent_overflow(event, 0, nmi, data, regs);
4542 static int perf_exclude_event(struct perf_event *event,
4543 struct pt_regs *regs)
4545 if (event->hw.state & PERF_HES_STOPPED)
4549 if (event->attr.exclude_user && user_mode(regs))
4552 if (event->attr.exclude_kernel && !user_mode(regs))
4559 static int perf_swevent_match(struct perf_event *event,
4560 enum perf_type_id type,
4562 struct perf_sample_data *data,
4563 struct pt_regs *regs)
4565 if (event->attr.type != type)
4568 if (event->attr.config != event_id)
4571 if (perf_exclude_event(event, regs))
4577 static inline u64 swevent_hash(u64 type, u32 event_id)
4579 u64 val = event_id | (type << 32);
4581 return hash_64(val, SWEVENT_HLIST_BITS);
4584 static inline struct hlist_head *
4585 __find_swevent_head(struct swevent_hlist *hlist, u64 type, u32 event_id)
4587 u64 hash = swevent_hash(type, event_id);
4589 return &hlist->heads[hash];
4592 /* For the read side: events when they trigger */
4593 static inline struct hlist_head *
4594 find_swevent_head_rcu(struct swevent_htable *swhash, u64 type, u32 event_id)
4596 struct swevent_hlist *hlist;
4598 hlist = rcu_dereference(swhash->swevent_hlist);
4602 return __find_swevent_head(hlist, type, event_id);
4605 /* For the event head insertion and removal in the hlist */
4606 static inline struct hlist_head *
4607 find_swevent_head(struct swevent_htable *swhash, struct perf_event *event)
4609 struct swevent_hlist *hlist;
4610 u32 event_id = event->attr.config;
4611 u64 type = event->attr.type;
4614 * Event scheduling is always serialized against hlist allocation
4615 * and release. Which makes the protected version suitable here.
4616 * The context lock guarantees that.
4618 hlist = rcu_dereference_protected(swhash->swevent_hlist,
4619 lockdep_is_held(&event->ctx->lock));
4623 return __find_swevent_head(hlist, type, event_id);
4626 static void do_perf_sw_event(enum perf_type_id type, u32 event_id,
4628 struct perf_sample_data *data,
4629 struct pt_regs *regs)
4631 struct swevent_htable *swhash = &__get_cpu_var(swevent_htable);
4632 struct perf_event *event;
4633 struct hlist_node *node;
4634 struct hlist_head *head;
4637 head = find_swevent_head_rcu(swhash, type, event_id);
4641 hlist_for_each_entry_rcu(event, node, head, hlist_entry) {
4642 if (perf_swevent_match(event, type, event_id, data, regs))
4643 perf_swevent_event(event, nr, nmi, data, regs);
4649 int perf_swevent_get_recursion_context(void)
4651 struct swevent_htable *swhash = &__get_cpu_var(swevent_htable);
4653 return get_recursion_context(swhash->recursion);
4655 EXPORT_SYMBOL_GPL(perf_swevent_get_recursion_context);
4657 void inline perf_swevent_put_recursion_context(int rctx)
4659 struct swevent_htable *swhash = &__get_cpu_var(swevent_htable);
4661 put_recursion_context(swhash->recursion, rctx);
4664 void __perf_sw_event(u32 event_id, u64 nr, int nmi,
4665 struct pt_regs *regs, u64 addr)
4667 struct perf_sample_data data;
4670 preempt_disable_notrace();
4671 rctx = perf_swevent_get_recursion_context();
4675 perf_sample_data_init(&data, addr);
4677 do_perf_sw_event(PERF_TYPE_SOFTWARE, event_id, nr, nmi, &data, regs);
4679 perf_swevent_put_recursion_context(rctx);
4680 preempt_enable_notrace();
4683 static void perf_swevent_read(struct perf_event *event)
4687 static int perf_swevent_add(struct perf_event *event, int flags)
4689 struct swevent_htable *swhash = &__get_cpu_var(swevent_htable);
4690 struct hw_perf_event *hwc = &event->hw;
4691 struct hlist_head *head;
4693 if (is_sampling_event(event)) {
4694 hwc->last_period = hwc->sample_period;
4695 perf_swevent_set_period(event);
4698 hwc->state = !(flags & PERF_EF_START);
4700 head = find_swevent_head(swhash, event);
4701 if (WARN_ON_ONCE(!head))
4704 hlist_add_head_rcu(&event->hlist_entry, head);
4709 static void perf_swevent_del(struct perf_event *event, int flags)
4711 hlist_del_rcu(&event->hlist_entry);
4714 static void perf_swevent_start(struct perf_event *event, int flags)
4716 event->hw.state = 0;
4719 static void perf_swevent_stop(struct perf_event *event, int flags)
4721 event->hw.state = PERF_HES_STOPPED;
4724 /* Deref the hlist from the update side */
4725 static inline struct swevent_hlist *
4726 swevent_hlist_deref(struct swevent_htable *swhash)
4728 return rcu_dereference_protected(swhash->swevent_hlist,
4729 lockdep_is_held(&swhash->hlist_mutex));
4732 static void swevent_hlist_release_rcu(struct rcu_head *rcu_head)
4734 struct swevent_hlist *hlist;
4736 hlist = container_of(rcu_head, struct swevent_hlist, rcu_head);
4740 static void swevent_hlist_release(struct swevent_htable *swhash)
4742 struct swevent_hlist *hlist = swevent_hlist_deref(swhash);
4747 rcu_assign_pointer(swhash->swevent_hlist, NULL);
4748 call_rcu(&hlist->rcu_head, swevent_hlist_release_rcu);
4751 static void swevent_hlist_put_cpu(struct perf_event *event, int cpu)
4753 struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
4755 mutex_lock(&swhash->hlist_mutex);
4757 if (!--swhash->hlist_refcount)
4758 swevent_hlist_release(swhash);
4760 mutex_unlock(&swhash->hlist_mutex);
4763 static void swevent_hlist_put(struct perf_event *event)
4767 if (event->cpu != -1) {
4768 swevent_hlist_put_cpu(event, event->cpu);
4772 for_each_possible_cpu(cpu)
4773 swevent_hlist_put_cpu(event, cpu);
4776 static int swevent_hlist_get_cpu(struct perf_event *event, int cpu)
4778 struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
4781 mutex_lock(&swhash->hlist_mutex);
4783 if (!swevent_hlist_deref(swhash) && cpu_online(cpu)) {
4784 struct swevent_hlist *hlist;
4786 hlist = kzalloc(sizeof(*hlist), GFP_KERNEL);
4791 rcu_assign_pointer(swhash->swevent_hlist, hlist);
4793 swhash->hlist_refcount++;
4795 mutex_unlock(&swhash->hlist_mutex);
4800 static int swevent_hlist_get(struct perf_event *event)
4803 int cpu, failed_cpu;
4805 if (event->cpu != -1)
4806 return swevent_hlist_get_cpu(event, event->cpu);
4809 for_each_possible_cpu(cpu) {
4810 err = swevent_hlist_get_cpu(event, cpu);
4820 for_each_possible_cpu(cpu) {
4821 if (cpu == failed_cpu)
4823 swevent_hlist_put_cpu(event, cpu);
4830 atomic_t perf_swevent_enabled[PERF_COUNT_SW_MAX];
4832 static void sw_perf_event_destroy(struct perf_event *event)
4834 u64 event_id = event->attr.config;
4836 WARN_ON(event->parent);
4838 jump_label_dec(&perf_swevent_enabled[event_id]);
4839 swevent_hlist_put(event);
4842 static int perf_swevent_init(struct perf_event *event)
4844 int event_id = event->attr.config;
4846 if (event->attr.type != PERF_TYPE_SOFTWARE)
4850 case PERF_COUNT_SW_CPU_CLOCK:
4851 case PERF_COUNT_SW_TASK_CLOCK:
4858 if (event_id >= PERF_COUNT_SW_MAX)
4861 if (!event->parent) {
4864 err = swevent_hlist_get(event);
4868 jump_label_inc(&perf_swevent_enabled[event_id]);
4869 event->destroy = sw_perf_event_destroy;
4875 static struct pmu perf_swevent = {
4876 .task_ctx_nr = perf_sw_context,
4878 .event_init = perf_swevent_init,
4879 .add = perf_swevent_add,
4880 .del = perf_swevent_del,
4881 .start = perf_swevent_start,
4882 .stop = perf_swevent_stop,
4883 .read = perf_swevent_read,
4886 #ifdef CONFIG_EVENT_TRACING
4888 static int perf_tp_filter_match(struct perf_event *event,
4889 struct perf_sample_data *data)
4891 void *record = data->raw->data;
4893 if (likely(!event->filter) || filter_match_preds(event->filter, record))
4898 static int perf_tp_event_match(struct perf_event *event,
4899 struct perf_sample_data *data,
4900 struct pt_regs *regs)
4903 * All tracepoints are from kernel-space.
4905 if (event->attr.exclude_kernel)
4908 if (!perf_tp_filter_match(event, data))
4914 void perf_tp_event(u64 addr, u64 count, void *record, int entry_size,
4915 struct pt_regs *regs, struct hlist_head *head, int rctx)
4917 struct perf_sample_data data;
4918 struct perf_event *event;
4919 struct hlist_node *node;
4921 struct perf_raw_record raw = {
4926 perf_sample_data_init(&data, addr);
4929 hlist_for_each_entry_rcu(event, node, head, hlist_entry) {
4930 if (perf_tp_event_match(event, &data, regs))
4931 perf_swevent_event(event, count, 1, &data, regs);
4934 perf_swevent_put_recursion_context(rctx);
4936 EXPORT_SYMBOL_GPL(perf_tp_event);
4938 static void tp_perf_event_destroy(struct perf_event *event)
4940 perf_trace_destroy(event);
4943 static int perf_tp_event_init(struct perf_event *event)
4947 if (event->attr.type != PERF_TYPE_TRACEPOINT)
4950 err = perf_trace_init(event);
4954 event->destroy = tp_perf_event_destroy;
4959 static struct pmu perf_tracepoint = {
4960 .task_ctx_nr = perf_sw_context,
4962 .event_init = perf_tp_event_init,
4963 .add = perf_trace_add,
4964 .del = perf_trace_del,
4965 .start = perf_swevent_start,
4966 .stop = perf_swevent_stop,
4967 .read = perf_swevent_read,
4970 static inline void perf_tp_register(void)
4972 perf_pmu_register(&perf_tracepoint, "tracepoint", PERF_TYPE_TRACEPOINT);
4975 static int perf_event_set_filter(struct perf_event *event, void __user *arg)
4980 if (event->attr.type != PERF_TYPE_TRACEPOINT)
4983 filter_str = strndup_user(arg, PAGE_SIZE);
4984 if (IS_ERR(filter_str))
4985 return PTR_ERR(filter_str);
4987 ret = ftrace_profile_set_filter(event, event->attr.config, filter_str);
4993 static void perf_event_free_filter(struct perf_event *event)
4995 ftrace_profile_free_filter(event);
5000 static inline void perf_tp_register(void)
5004 static int perf_event_set_filter(struct perf_event *event, void __user *arg)
5009 static void perf_event_free_filter(struct perf_event *event)
5013 #endif /* CONFIG_EVENT_TRACING */
5015 #ifdef CONFIG_HAVE_HW_BREAKPOINT
5016 void perf_bp_event(struct perf_event *bp, void *data)
5018 struct perf_sample_data sample;
5019 struct pt_regs *regs = data;
5021 perf_sample_data_init(&sample, bp->attr.bp_addr);
5023 if (!bp->hw.state && !perf_exclude_event(bp, regs))
5024 perf_swevent_event(bp, 1, 1, &sample, regs);
5029 * hrtimer based swevent callback
5032 static enum hrtimer_restart perf_swevent_hrtimer(struct hrtimer *hrtimer)
5034 enum hrtimer_restart ret = HRTIMER_RESTART;
5035 struct perf_sample_data data;
5036 struct pt_regs *regs;
5037 struct perf_event *event;
5040 event = container_of(hrtimer, struct perf_event, hw.hrtimer);
5041 event->pmu->read(event);
5043 perf_sample_data_init(&data, 0);
5044 data.period = event->hw.last_period;
5045 regs = get_irq_regs();
5047 if (regs && !perf_exclude_event(event, regs)) {
5048 if (!(event->attr.exclude_idle && current->pid == 0))
5049 if (perf_event_overflow(event, 0, &data, regs))
5050 ret = HRTIMER_NORESTART;
5053 period = max_t(u64, 10000, event->hw.sample_period);
5054 hrtimer_forward_now(hrtimer, ns_to_ktime(period));
5059 static void perf_swevent_start_hrtimer(struct perf_event *event)
5061 struct hw_perf_event *hwc = &event->hw;
5064 if (!is_sampling_event(event))
5067 hrtimer_init(&hwc->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
5068 hwc->hrtimer.function = perf_swevent_hrtimer;
5070 period = local64_read(&hwc->period_left);
5075 local64_set(&hwc->period_left, 0);
5077 period = max_t(u64, 10000, hwc->sample_period);
5079 __hrtimer_start_range_ns(&hwc->hrtimer,
5080 ns_to_ktime(period), 0,
5081 HRTIMER_MODE_REL_PINNED, 0);
5084 static void perf_swevent_cancel_hrtimer(struct perf_event *event)
5086 struct hw_perf_event *hwc = &event->hw;
5088 if (is_sampling_event(event)) {
5089 ktime_t remaining = hrtimer_get_remaining(&hwc->hrtimer);
5090 local64_set(&hwc->period_left, ktime_to_ns(remaining));
5092 hrtimer_cancel(&hwc->hrtimer);
5097 * Software event: cpu wall time clock
5100 static void cpu_clock_event_update(struct perf_event *event)
5105 now = local_clock();
5106 prev = local64_xchg(&event->hw.prev_count, now);
5107 local64_add(now - prev, &event->count);
5110 static void cpu_clock_event_start(struct perf_event *event, int flags)
5112 local64_set(&event->hw.prev_count, local_clock());
5113 perf_swevent_start_hrtimer(event);
5116 static void cpu_clock_event_stop(struct perf_event *event, int flags)
5118 perf_swevent_cancel_hrtimer(event);
5119 cpu_clock_event_update(event);
5122 static int cpu_clock_event_add(struct perf_event *event, int flags)
5124 if (flags & PERF_EF_START)
5125 cpu_clock_event_start(event, flags);
5130 static void cpu_clock_event_del(struct perf_event *event, int flags)
5132 cpu_clock_event_stop(event, flags);
5135 static void cpu_clock_event_read(struct perf_event *event)
5137 cpu_clock_event_update(event);
5140 static int cpu_clock_event_init(struct perf_event *event)
5142 if (event->attr.type != PERF_TYPE_SOFTWARE)
5145 if (event->attr.config != PERF_COUNT_SW_CPU_CLOCK)
5151 static struct pmu perf_cpu_clock = {
5152 .task_ctx_nr = perf_sw_context,
5154 .event_init = cpu_clock_event_init,
5155 .add = cpu_clock_event_add,
5156 .del = cpu_clock_event_del,
5157 .start = cpu_clock_event_start,
5158 .stop = cpu_clock_event_stop,
5159 .read = cpu_clock_event_read,
5163 * Software event: task time clock
5166 static void task_clock_event_update(struct perf_event *event, u64 now)
5171 prev = local64_xchg(&event->hw.prev_count, now);
5173 local64_add(delta, &event->count);
5176 static void task_clock_event_start(struct perf_event *event, int flags)
5178 local64_set(&event->hw.prev_count, event->ctx->time);
5179 perf_swevent_start_hrtimer(event);
5182 static void task_clock_event_stop(struct perf_event *event, int flags)
5184 perf_swevent_cancel_hrtimer(event);
5185 task_clock_event_update(event, event->ctx->time);
5188 static int task_clock_event_add(struct perf_event *event, int flags)
5190 if (flags & PERF_EF_START)
5191 task_clock_event_start(event, flags);
5196 static void task_clock_event_del(struct perf_event *event, int flags)
5198 task_clock_event_stop(event, PERF_EF_UPDATE);
5201 static void task_clock_event_read(struct perf_event *event)
5206 update_context_time(event->ctx);
5207 time = event->ctx->time;
5209 u64 now = perf_clock();
5210 u64 delta = now - event->ctx->timestamp;
5211 time = event->ctx->time + delta;
5214 task_clock_event_update(event, time);
5217 static int task_clock_event_init(struct perf_event *event)
5219 if (event->attr.type != PERF_TYPE_SOFTWARE)
5222 if (event->attr.config != PERF_COUNT_SW_TASK_CLOCK)
5228 static struct pmu perf_task_clock = {
5229 .task_ctx_nr = perf_sw_context,
5231 .event_init = task_clock_event_init,
5232 .add = task_clock_event_add,
5233 .del = task_clock_event_del,
5234 .start = task_clock_event_start,
5235 .stop = task_clock_event_stop,
5236 .read = task_clock_event_read,
5239 static void perf_pmu_nop_void(struct pmu *pmu)
5243 static int perf_pmu_nop_int(struct pmu *pmu)
5248 static void perf_pmu_start_txn(struct pmu *pmu)
5250 perf_pmu_disable(pmu);
5253 static int perf_pmu_commit_txn(struct pmu *pmu)
5255 perf_pmu_enable(pmu);
5259 static void perf_pmu_cancel_txn(struct pmu *pmu)
5261 perf_pmu_enable(pmu);
5265 * Ensures all contexts with the same task_ctx_nr have the same
5266 * pmu_cpu_context too.
5268 static void *find_pmu_context(int ctxn)
5275 list_for_each_entry(pmu, &pmus, entry) {
5276 if (pmu->task_ctx_nr == ctxn)
5277 return pmu->pmu_cpu_context;
5283 static void update_pmu_context(struct pmu *pmu, struct pmu *old_pmu)
5287 for_each_possible_cpu(cpu) {
5288 struct perf_cpu_context *cpuctx;
5290 cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu);
5292 if (cpuctx->active_pmu == old_pmu)
5293 cpuctx->active_pmu = pmu;
5297 static void free_pmu_context(struct pmu *pmu)
5301 mutex_lock(&pmus_lock);
5303 * Like a real lame refcount.
5305 list_for_each_entry(i, &pmus, entry) {
5306 if (i->pmu_cpu_context == pmu->pmu_cpu_context) {
5307 update_pmu_context(i, pmu);
5312 free_percpu(pmu->pmu_cpu_context);
5314 mutex_unlock(&pmus_lock);
5316 static struct idr pmu_idr;
5319 type_show(struct device *dev, struct device_attribute *attr, char *page)
5321 struct pmu *pmu = dev_get_drvdata(dev);
5323 return snprintf(page, PAGE_SIZE-1, "%d\n", pmu->type);
5326 static struct device_attribute pmu_dev_attrs[] = {
5331 static int pmu_bus_running;
5332 static struct bus_type pmu_bus = {
5333 .name = "event_source",
5334 .dev_attrs = pmu_dev_attrs,
5337 static void pmu_dev_release(struct device *dev)
5342 static int pmu_dev_alloc(struct pmu *pmu)
5346 pmu->dev = kzalloc(sizeof(struct device), GFP_KERNEL);
5350 device_initialize(pmu->dev);
5351 ret = dev_set_name(pmu->dev, "%s", pmu->name);
5355 dev_set_drvdata(pmu->dev, pmu);
5356 pmu->dev->bus = &pmu_bus;
5357 pmu->dev->release = pmu_dev_release;
5358 ret = device_add(pmu->dev);
5366 put_device(pmu->dev);
5370 int perf_pmu_register(struct pmu *pmu, char *name, int type)
5374 mutex_lock(&pmus_lock);
5376 pmu->pmu_disable_count = alloc_percpu(int);
5377 if (!pmu->pmu_disable_count)
5386 int err = idr_pre_get(&pmu_idr, GFP_KERNEL);
5390 err = idr_get_new_above(&pmu_idr, pmu, PERF_TYPE_MAX, &type);
5398 if (pmu_bus_running) {
5399 ret = pmu_dev_alloc(pmu);
5405 pmu->pmu_cpu_context = find_pmu_context(pmu->task_ctx_nr);
5406 if (pmu->pmu_cpu_context)
5407 goto got_cpu_context;
5409 pmu->pmu_cpu_context = alloc_percpu(struct perf_cpu_context);
5410 if (!pmu->pmu_cpu_context)
5413 for_each_possible_cpu(cpu) {
5414 struct perf_cpu_context *cpuctx;
5416 cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu);
5417 __perf_event_init_context(&cpuctx->ctx);
5418 cpuctx->ctx.type = cpu_context;
5419 cpuctx->ctx.pmu = pmu;
5420 cpuctx->jiffies_interval = 1;
5421 INIT_LIST_HEAD(&cpuctx->rotation_list);
5422 cpuctx->active_pmu = pmu;
5426 if (!pmu->start_txn) {
5427 if (pmu->pmu_enable) {
5429 * If we have pmu_enable/pmu_disable calls, install
5430 * transaction stubs that use that to try and batch
5431 * hardware accesses.
5433 pmu->start_txn = perf_pmu_start_txn;
5434 pmu->commit_txn = perf_pmu_commit_txn;
5435 pmu->cancel_txn = perf_pmu_cancel_txn;
5437 pmu->start_txn = perf_pmu_nop_void;
5438 pmu->commit_txn = perf_pmu_nop_int;
5439 pmu->cancel_txn = perf_pmu_nop_void;
5443 if (!pmu->pmu_enable) {
5444 pmu->pmu_enable = perf_pmu_nop_void;
5445 pmu->pmu_disable = perf_pmu_nop_void;
5448 list_add_rcu(&pmu->entry, &pmus);
5451 mutex_unlock(&pmus_lock);
5456 device_del(pmu->dev);
5457 put_device(pmu->dev);
5460 if (pmu->type >= PERF_TYPE_MAX)
5461 idr_remove(&pmu_idr, pmu->type);
5464 free_percpu(pmu->pmu_disable_count);
5468 void perf_pmu_unregister(struct pmu *pmu)
5470 mutex_lock(&pmus_lock);
5471 list_del_rcu(&pmu->entry);
5472 mutex_unlock(&pmus_lock);
5475 * We dereference the pmu list under both SRCU and regular RCU, so
5476 * synchronize against both of those.
5478 synchronize_srcu(&pmus_srcu);
5481 free_percpu(pmu->pmu_disable_count);
5482 if (pmu->type >= PERF_TYPE_MAX)
5483 idr_remove(&pmu_idr, pmu->type);
5484 device_del(pmu->dev);
5485 put_device(pmu->dev);
5486 free_pmu_context(pmu);
5489 struct pmu *perf_init_event(struct perf_event *event)
5491 struct pmu *pmu = NULL;
5494 idx = srcu_read_lock(&pmus_srcu);
5497 pmu = idr_find(&pmu_idr, event->attr.type);
5502 list_for_each_entry_rcu(pmu, &pmus, entry) {
5503 int ret = pmu->event_init(event);
5507 if (ret != -ENOENT) {
5512 pmu = ERR_PTR(-ENOENT);
5514 srcu_read_unlock(&pmus_srcu, idx);
5520 * Allocate and initialize a event structure
5522 static struct perf_event *
5523 perf_event_alloc(struct perf_event_attr *attr, int cpu,
5524 struct task_struct *task,
5525 struct perf_event *group_leader,
5526 struct perf_event *parent_event,
5527 perf_overflow_handler_t overflow_handler)
5530 struct perf_event *event;
5531 struct hw_perf_event *hwc;
5534 event = kzalloc(sizeof(*event), GFP_KERNEL);
5536 return ERR_PTR(-ENOMEM);
5539 * Single events are their own group leaders, with an
5540 * empty sibling list:
5543 group_leader = event;
5545 mutex_init(&event->child_mutex);
5546 INIT_LIST_HEAD(&event->child_list);
5548 INIT_LIST_HEAD(&event->group_entry);
5549 INIT_LIST_HEAD(&event->event_entry);
5550 INIT_LIST_HEAD(&event->sibling_list);
5551 init_waitqueue_head(&event->waitq);
5552 init_irq_work(&event->pending, perf_pending_event);
5554 mutex_init(&event->mmap_mutex);
5557 event->attr = *attr;
5558 event->group_leader = group_leader;
5562 event->parent = parent_event;
5564 event->ns = get_pid_ns(current->nsproxy->pid_ns);
5565 event->id = atomic64_inc_return(&perf_event_id);
5567 event->state = PERF_EVENT_STATE_INACTIVE;
5570 event->attach_state = PERF_ATTACH_TASK;
5571 #ifdef CONFIG_HAVE_HW_BREAKPOINT
5573 * hw_breakpoint is a bit difficult here..
5575 if (attr->type == PERF_TYPE_BREAKPOINT)
5576 event->hw.bp_target = task;
5580 if (!overflow_handler && parent_event)
5581 overflow_handler = parent_event->overflow_handler;
5583 event->overflow_handler = overflow_handler;
5586 event->state = PERF_EVENT_STATE_OFF;
5591 hwc->sample_period = attr->sample_period;
5592 if (attr->freq && attr->sample_freq)
5593 hwc->sample_period = 1;
5594 hwc->last_period = hwc->sample_period;
5596 local64_set(&hwc->period_left, hwc->sample_period);
5599 * we currently do not support PERF_FORMAT_GROUP on inherited events
5601 if (attr->inherit && (attr->read_format & PERF_FORMAT_GROUP))
5604 pmu = perf_init_event(event);
5610 else if (IS_ERR(pmu))
5615 put_pid_ns(event->ns);
5617 return ERR_PTR(err);
5622 if (!event->parent) {
5623 if (event->attach_state & PERF_ATTACH_TASK)
5624 jump_label_inc(&perf_task_events);
5625 if (event->attr.mmap || event->attr.mmap_data)
5626 atomic_inc(&nr_mmap_events);
5627 if (event->attr.comm)
5628 atomic_inc(&nr_comm_events);
5629 if (event->attr.task)
5630 atomic_inc(&nr_task_events);
5631 if (event->attr.sample_type & PERF_SAMPLE_CALLCHAIN) {
5632 err = get_callchain_buffers();
5635 return ERR_PTR(err);
5643 static int perf_copy_attr(struct perf_event_attr __user *uattr,
5644 struct perf_event_attr *attr)
5649 if (!access_ok(VERIFY_WRITE, uattr, PERF_ATTR_SIZE_VER0))
5653 * zero the full structure, so that a short copy will be nice.
5655 memset(attr, 0, sizeof(*attr));
5657 ret = get_user(size, &uattr->size);
5661 if (size > PAGE_SIZE) /* silly large */
5664 if (!size) /* abi compat */
5665 size = PERF_ATTR_SIZE_VER0;
5667 if (size < PERF_ATTR_SIZE_VER0)
5671 * If we're handed a bigger struct than we know of,
5672 * ensure all the unknown bits are 0 - i.e. new
5673 * user-space does not rely on any kernel feature
5674 * extensions we dont know about yet.
5676 if (size > sizeof(*attr)) {
5677 unsigned char __user *addr;
5678 unsigned char __user *end;
5681 addr = (void __user *)uattr + sizeof(*attr);
5682 end = (void __user *)uattr + size;
5684 for (; addr < end; addr++) {
5685 ret = get_user(val, addr);
5691 size = sizeof(*attr);
5694 ret = copy_from_user(attr, uattr, size);
5699 * If the type exists, the corresponding creation will verify
5702 if (attr->type >= PERF_TYPE_MAX)
5705 if (attr->__reserved_1)
5708 if (attr->sample_type & ~(PERF_SAMPLE_MAX-1))
5711 if (attr->read_format & ~(PERF_FORMAT_MAX-1))
5718 put_user(sizeof(*attr), &uattr->size);
5724 perf_event_set_output(struct perf_event *event, struct perf_event *output_event)
5726 struct perf_buffer *buffer = NULL, *old_buffer = NULL;
5732 /* don't allow circular references */
5733 if (event == output_event)
5737 * Don't allow cross-cpu buffers
5739 if (output_event->cpu != event->cpu)
5743 * If its not a per-cpu buffer, it must be the same task.
5745 if (output_event->cpu == -1 && output_event->ctx != event->ctx)
5749 mutex_lock(&event->mmap_mutex);
5750 /* Can't redirect output if we've got an active mmap() */
5751 if (atomic_read(&event->mmap_count))
5755 /* get the buffer we want to redirect to */
5756 buffer = perf_buffer_get(output_event);
5761 old_buffer = event->buffer;
5762 rcu_assign_pointer(event->buffer, buffer);
5765 mutex_unlock(&event->mmap_mutex);
5768 perf_buffer_put(old_buffer);
5774 * sys_perf_event_open - open a performance event, associate it to a task/cpu
5776 * @attr_uptr: event_id type attributes for monitoring/sampling
5779 * @group_fd: group leader event fd
5781 SYSCALL_DEFINE5(perf_event_open,
5782 struct perf_event_attr __user *, attr_uptr,
5783 pid_t, pid, int, cpu, int, group_fd, unsigned long, flags)
5785 struct perf_event *group_leader = NULL, *output_event = NULL;
5786 struct perf_event *event, *sibling;
5787 struct perf_event_attr attr;
5788 struct perf_event_context *ctx;
5789 struct file *event_file = NULL;
5790 struct file *group_file = NULL;
5791 struct task_struct *task = NULL;
5795 int fput_needed = 0;
5798 /* for future expandability... */
5799 if (flags & ~(PERF_FLAG_FD_NO_GROUP | PERF_FLAG_FD_OUTPUT))
5802 err = perf_copy_attr(attr_uptr, &attr);
5806 if (!attr.exclude_kernel) {
5807 if (perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN))
5812 if (attr.sample_freq > sysctl_perf_event_sample_rate)
5816 event_fd = get_unused_fd_flags(O_RDWR);
5820 if (group_fd != -1) {
5821 group_leader = perf_fget_light(group_fd, &fput_needed);
5822 if (IS_ERR(group_leader)) {
5823 err = PTR_ERR(group_leader);
5826 group_file = group_leader->filp;
5827 if (flags & PERF_FLAG_FD_OUTPUT)
5828 output_event = group_leader;
5829 if (flags & PERF_FLAG_FD_NO_GROUP)
5830 group_leader = NULL;
5834 task = find_lively_task_by_vpid(pid);
5836 err = PTR_ERR(task);
5841 event = perf_event_alloc(&attr, cpu, task, group_leader, NULL, NULL);
5842 if (IS_ERR(event)) {
5843 err = PTR_ERR(event);
5848 * Special case software events and allow them to be part of
5849 * any hardware group.
5854 (is_software_event(event) != is_software_event(group_leader))) {
5855 if (is_software_event(event)) {
5857 * If event and group_leader are not both a software
5858 * event, and event is, then group leader is not.
5860 * Allow the addition of software events to !software
5861 * groups, this is safe because software events never
5864 pmu = group_leader->pmu;
5865 } else if (is_software_event(group_leader) &&
5866 (group_leader->group_flags & PERF_GROUP_SOFTWARE)) {
5868 * In case the group is a pure software group, and we
5869 * try to add a hardware event, move the whole group to
5870 * the hardware context.
5877 * Get the target context (task or percpu):
5879 ctx = find_get_context(pmu, task, cpu);
5886 * Look up the group leader (we will attach this event to it):
5892 * Do not allow a recursive hierarchy (this new sibling
5893 * becoming part of another group-sibling):
5895 if (group_leader->group_leader != group_leader)
5898 * Do not allow to attach to a group in a different
5899 * task or CPU context:
5902 if (group_leader->ctx->type != ctx->type)
5905 if (group_leader->ctx != ctx)
5910 * Only a group leader can be exclusive or pinned
5912 if (attr.exclusive || attr.pinned)
5917 err = perf_event_set_output(event, output_event);
5922 event_file = anon_inode_getfile("[perf_event]", &perf_fops, event, O_RDWR);
5923 if (IS_ERR(event_file)) {
5924 err = PTR_ERR(event_file);
5929 struct perf_event_context *gctx = group_leader->ctx;
5931 mutex_lock(&gctx->mutex);
5932 perf_event_remove_from_context(group_leader);
5933 list_for_each_entry(sibling, &group_leader->sibling_list,
5935 perf_event_remove_from_context(sibling);
5938 mutex_unlock(&gctx->mutex);
5942 event->filp = event_file;
5943 WARN_ON_ONCE(ctx->parent_ctx);
5944 mutex_lock(&ctx->mutex);
5947 perf_install_in_context(ctx, group_leader, cpu);
5949 list_for_each_entry(sibling, &group_leader->sibling_list,
5951 perf_install_in_context(ctx, sibling, cpu);
5956 perf_install_in_context(ctx, event, cpu);
5958 mutex_unlock(&ctx->mutex);
5960 event->owner = current;
5962 mutex_lock(¤t->perf_event_mutex);
5963 list_add_tail(&event->owner_entry, ¤t->perf_event_list);
5964 mutex_unlock(¤t->perf_event_mutex);
5967 * Precalculate sample_data sizes
5969 perf_event__header_size(event);
5970 perf_event__id_header_size(event);
5973 * Drop the reference on the group_event after placing the
5974 * new event on the sibling_list. This ensures destruction
5975 * of the group leader will find the pointer to itself in
5976 * perf_group_detach().
5978 fput_light(group_file, fput_needed);
5979 fd_install(event_fd, event_file);
5988 put_task_struct(task);
5990 fput_light(group_file, fput_needed);
5992 put_unused_fd(event_fd);
5997 * perf_event_create_kernel_counter
5999 * @attr: attributes of the counter to create
6000 * @cpu: cpu in which the counter is bound
6001 * @task: task to profile (NULL for percpu)
6004 perf_event_create_kernel_counter(struct perf_event_attr *attr, int cpu,
6005 struct task_struct *task,
6006 perf_overflow_handler_t overflow_handler)
6008 struct perf_event_context *ctx;
6009 struct perf_event *event;
6013 * Get the target context (task or percpu):
6016 event = perf_event_alloc(attr, cpu, task, NULL, NULL, overflow_handler);
6017 if (IS_ERR(event)) {
6018 err = PTR_ERR(event);
6022 ctx = find_get_context(event->pmu, task, cpu);
6029 WARN_ON_ONCE(ctx->parent_ctx);
6030 mutex_lock(&ctx->mutex);
6031 perf_install_in_context(ctx, event, cpu);
6033 mutex_unlock(&ctx->mutex);
6040 return ERR_PTR(err);
6042 EXPORT_SYMBOL_GPL(perf_event_create_kernel_counter);
6044 static void sync_child_event(struct perf_event *child_event,
6045 struct task_struct *child)
6047 struct perf_event *parent_event = child_event->parent;
6050 if (child_event->attr.inherit_stat)
6051 perf_event_read_event(child_event, child);
6053 child_val = perf_event_count(child_event);
6056 * Add back the child's count to the parent's count:
6058 atomic64_add(child_val, &parent_event->child_count);
6059 atomic64_add(child_event->total_time_enabled,
6060 &parent_event->child_total_time_enabled);
6061 atomic64_add(child_event->total_time_running,
6062 &parent_event->child_total_time_running);
6065 * Remove this event from the parent's list
6067 WARN_ON_ONCE(parent_event->ctx->parent_ctx);
6068 mutex_lock(&parent_event->child_mutex);
6069 list_del_init(&child_event->child_list);
6070 mutex_unlock(&parent_event->child_mutex);
6073 * Release the parent event, if this was the last
6076 fput(parent_event->filp);
6080 __perf_event_exit_task(struct perf_event *child_event,
6081 struct perf_event_context *child_ctx,
6082 struct task_struct *child)
6084 struct perf_event *parent_event;
6086 perf_event_remove_from_context(child_event);
6088 parent_event = child_event->parent;
6090 * It can happen that parent exits first, and has events
6091 * that are still around due to the child reference. These
6092 * events need to be zapped - but otherwise linger.
6095 sync_child_event(child_event, child);
6096 free_event(child_event);
6100 static void perf_event_exit_task_context(struct task_struct *child, int ctxn)
6102 struct perf_event *child_event, *tmp;
6103 struct perf_event_context *child_ctx;
6104 unsigned long flags;
6106 if (likely(!child->perf_event_ctxp[ctxn])) {
6107 perf_event_task(child, NULL, 0);
6111 local_irq_save(flags);
6113 * We can't reschedule here because interrupts are disabled,
6114 * and either child is current or it is a task that can't be
6115 * scheduled, so we are now safe from rescheduling changing
6118 child_ctx = child->perf_event_ctxp[ctxn];
6119 task_ctx_sched_out(child_ctx, EVENT_ALL);
6122 * Take the context lock here so that if find_get_context is
6123 * reading child->perf_event_ctxp, we wait until it has
6124 * incremented the context's refcount before we do put_ctx below.
6126 raw_spin_lock(&child_ctx->lock);
6127 child->perf_event_ctxp[ctxn] = NULL;
6129 * If this context is a clone; unclone it so it can't get
6130 * swapped to another process while we're removing all
6131 * the events from it.
6133 unclone_ctx(child_ctx);
6134 update_context_time(child_ctx);
6135 raw_spin_unlock_irqrestore(&child_ctx->lock, flags);
6138 * Report the task dead after unscheduling the events so that we
6139 * won't get any samples after PERF_RECORD_EXIT. We can however still
6140 * get a few PERF_RECORD_READ events.
6142 perf_event_task(child, child_ctx, 0);
6145 * We can recurse on the same lock type through:
6147 * __perf_event_exit_task()
6148 * sync_child_event()
6149 * fput(parent_event->filp)
6151 * mutex_lock(&ctx->mutex)
6153 * But since its the parent context it won't be the same instance.
6155 mutex_lock(&child_ctx->mutex);
6158 list_for_each_entry_safe(child_event, tmp, &child_ctx->pinned_groups,
6160 __perf_event_exit_task(child_event, child_ctx, child);
6162 list_for_each_entry_safe(child_event, tmp, &child_ctx->flexible_groups,
6164 __perf_event_exit_task(child_event, child_ctx, child);
6167 * If the last event was a group event, it will have appended all
6168 * its siblings to the list, but we obtained 'tmp' before that which
6169 * will still point to the list head terminating the iteration.
6171 if (!list_empty(&child_ctx->pinned_groups) ||
6172 !list_empty(&child_ctx->flexible_groups))
6175 mutex_unlock(&child_ctx->mutex);
6181 * When a child task exits, feed back event values to parent events.
6183 void perf_event_exit_task(struct task_struct *child)
6185 struct perf_event *event, *tmp;
6188 mutex_lock(&child->perf_event_mutex);
6189 list_for_each_entry_safe(event, tmp, &child->perf_event_list,
6191 list_del_init(&event->owner_entry);
6194 * Ensure the list deletion is visible before we clear
6195 * the owner, closes a race against perf_release() where
6196 * we need to serialize on the owner->perf_event_mutex.
6199 event->owner = NULL;
6201 mutex_unlock(&child->perf_event_mutex);
6203 for_each_task_context_nr(ctxn)
6204 perf_event_exit_task_context(child, ctxn);
6207 static void perf_free_event(struct perf_event *event,
6208 struct perf_event_context *ctx)
6210 struct perf_event *parent = event->parent;
6212 if (WARN_ON_ONCE(!parent))
6215 mutex_lock(&parent->child_mutex);
6216 list_del_init(&event->child_list);
6217 mutex_unlock(&parent->child_mutex);
6221 perf_group_detach(event);
6222 list_del_event(event, ctx);
6227 * free an unexposed, unused context as created by inheritance by
6228 * perf_event_init_task below, used by fork() in case of fail.
6230 void perf_event_free_task(struct task_struct *task)
6232 struct perf_event_context *ctx;
6233 struct perf_event *event, *tmp;
6236 for_each_task_context_nr(ctxn) {
6237 ctx = task->perf_event_ctxp[ctxn];
6241 mutex_lock(&ctx->mutex);
6243 list_for_each_entry_safe(event, tmp, &ctx->pinned_groups,
6245 perf_free_event(event, ctx);
6247 list_for_each_entry_safe(event, tmp, &ctx->flexible_groups,
6249 perf_free_event(event, ctx);
6251 if (!list_empty(&ctx->pinned_groups) ||
6252 !list_empty(&ctx->flexible_groups))
6255 mutex_unlock(&ctx->mutex);
6261 void perf_event_delayed_put(struct task_struct *task)
6265 for_each_task_context_nr(ctxn)
6266 WARN_ON_ONCE(task->perf_event_ctxp[ctxn]);
6270 * inherit a event from parent task to child task:
6272 static struct perf_event *
6273 inherit_event(struct perf_event *parent_event,
6274 struct task_struct *parent,
6275 struct perf_event_context *parent_ctx,
6276 struct task_struct *child,
6277 struct perf_event *group_leader,
6278 struct perf_event_context *child_ctx)
6280 struct perf_event *child_event;
6281 unsigned long flags;
6284 * Instead of creating recursive hierarchies of events,
6285 * we link inherited events back to the original parent,
6286 * which has a filp for sure, which we use as the reference
6289 if (parent_event->parent)
6290 parent_event = parent_event->parent;
6292 child_event = perf_event_alloc(&parent_event->attr,
6295 group_leader, parent_event,
6297 if (IS_ERR(child_event))
6302 * Make the child state follow the state of the parent event,
6303 * not its attr.disabled bit. We hold the parent's mutex,
6304 * so we won't race with perf_event_{en, dis}able_family.
6306 if (parent_event->state >= PERF_EVENT_STATE_INACTIVE)
6307 child_event->state = PERF_EVENT_STATE_INACTIVE;
6309 child_event->state = PERF_EVENT_STATE_OFF;
6311 if (parent_event->attr.freq) {
6312 u64 sample_period = parent_event->hw.sample_period;
6313 struct hw_perf_event *hwc = &child_event->hw;
6315 hwc->sample_period = sample_period;
6316 hwc->last_period = sample_period;
6318 local64_set(&hwc->period_left, sample_period);
6321 child_event->ctx = child_ctx;
6322 child_event->overflow_handler = parent_event->overflow_handler;
6325 * Precalculate sample_data sizes
6327 perf_event__header_size(child_event);
6328 perf_event__id_header_size(child_event);
6331 * Link it up in the child's context:
6333 raw_spin_lock_irqsave(&child_ctx->lock, flags);
6334 add_event_to_ctx(child_event, child_ctx);
6335 raw_spin_unlock_irqrestore(&child_ctx->lock, flags);
6338 * Get a reference to the parent filp - we will fput it
6339 * when the child event exits. This is safe to do because
6340 * we are in the parent and we know that the filp still
6341 * exists and has a nonzero count:
6343 atomic_long_inc(&parent_event->filp->f_count);
6346 * Link this into the parent event's child list
6348 WARN_ON_ONCE(parent_event->ctx->parent_ctx);
6349 mutex_lock(&parent_event->child_mutex);
6350 list_add_tail(&child_event->child_list, &parent_event->child_list);
6351 mutex_unlock(&parent_event->child_mutex);
6356 static int inherit_group(struct perf_event *parent_event,
6357 struct task_struct *parent,
6358 struct perf_event_context *parent_ctx,
6359 struct task_struct *child,
6360 struct perf_event_context *child_ctx)
6362 struct perf_event *leader;
6363 struct perf_event *sub;
6364 struct perf_event *child_ctr;
6366 leader = inherit_event(parent_event, parent, parent_ctx,
6367 child, NULL, child_ctx);
6369 return PTR_ERR(leader);
6370 list_for_each_entry(sub, &parent_event->sibling_list, group_entry) {
6371 child_ctr = inherit_event(sub, parent, parent_ctx,
6372 child, leader, child_ctx);
6373 if (IS_ERR(child_ctr))
6374 return PTR_ERR(child_ctr);
6380 inherit_task_group(struct perf_event *event, struct task_struct *parent,
6381 struct perf_event_context *parent_ctx,
6382 struct task_struct *child, int ctxn,
6386 struct perf_event_context *child_ctx;
6388 if (!event->attr.inherit) {
6393 child_ctx = child->perf_event_ctxp[ctxn];
6396 * This is executed from the parent task context, so
6397 * inherit events that have been marked for cloning.
6398 * First allocate and initialize a context for the
6402 child_ctx = alloc_perf_context(event->pmu, child);
6406 child->perf_event_ctxp[ctxn] = child_ctx;
6409 ret = inherit_group(event, parent, parent_ctx,
6419 * Initialize the perf_event context in task_struct
6421 int perf_event_init_context(struct task_struct *child, int ctxn)
6423 struct perf_event_context *child_ctx, *parent_ctx;
6424 struct perf_event_context *cloned_ctx;
6425 struct perf_event *event;
6426 struct task_struct *parent = current;
6427 int inherited_all = 1;
6428 unsigned long flags;
6431 child->perf_event_ctxp[ctxn] = NULL;
6433 mutex_init(&child->perf_event_mutex);
6434 INIT_LIST_HEAD(&child->perf_event_list);
6436 if (likely(!parent->perf_event_ctxp[ctxn]))
6440 * If the parent's context is a clone, pin it so it won't get
6443 parent_ctx = perf_pin_task_context(parent, ctxn);
6446 * No need to check if parent_ctx != NULL here; since we saw
6447 * it non-NULL earlier, the only reason for it to become NULL
6448 * is if we exit, and since we're currently in the middle of
6449 * a fork we can't be exiting at the same time.
6453 * Lock the parent list. No need to lock the child - not PID
6454 * hashed yet and not running, so nobody can access it.
6456 mutex_lock(&parent_ctx->mutex);
6459 * We dont have to disable NMIs - we are only looking at
6460 * the list, not manipulating it:
6462 list_for_each_entry(event, &parent_ctx->pinned_groups, group_entry) {
6463 ret = inherit_task_group(event, parent, parent_ctx,
6464 child, ctxn, &inherited_all);
6470 * We can't hold ctx->lock when iterating the ->flexible_group list due
6471 * to allocations, but we need to prevent rotation because
6472 * rotate_ctx() will change the list from interrupt context.
6474 raw_spin_lock_irqsave(&parent_ctx->lock, flags);
6475 parent_ctx->rotate_disable = 1;
6476 raw_spin_unlock_irqrestore(&parent_ctx->lock, flags);
6478 list_for_each_entry(event, &parent_ctx->flexible_groups, group_entry) {
6479 ret = inherit_task_group(event, parent, parent_ctx,
6480 child, ctxn, &inherited_all);
6485 raw_spin_lock_irqsave(&parent_ctx->lock, flags);
6486 parent_ctx->rotate_disable = 0;
6487 raw_spin_unlock_irqrestore(&parent_ctx->lock, flags);
6489 child_ctx = child->perf_event_ctxp[ctxn];
6491 if (child_ctx && inherited_all) {
6493 * Mark the child context as a clone of the parent
6494 * context, or of whatever the parent is a clone of.
6495 * Note that if the parent is a clone, it could get
6496 * uncloned at any point, but that doesn't matter
6497 * because the list of events and the generation
6498 * count can't have changed since we took the mutex.
6500 cloned_ctx = rcu_dereference(parent_ctx->parent_ctx);
6502 child_ctx->parent_ctx = cloned_ctx;
6503 child_ctx->parent_gen = parent_ctx->parent_gen;
6505 child_ctx->parent_ctx = parent_ctx;
6506 child_ctx->parent_gen = parent_ctx->generation;
6508 get_ctx(child_ctx->parent_ctx);
6511 mutex_unlock(&parent_ctx->mutex);
6513 perf_unpin_context(parent_ctx);
6519 * Initialize the perf_event context in task_struct
6521 int perf_event_init_task(struct task_struct *child)
6525 for_each_task_context_nr(ctxn) {
6526 ret = perf_event_init_context(child, ctxn);
6534 static void __init perf_event_init_all_cpus(void)
6536 struct swevent_htable *swhash;
6539 for_each_possible_cpu(cpu) {
6540 swhash = &per_cpu(swevent_htable, cpu);
6541 mutex_init(&swhash->hlist_mutex);
6542 INIT_LIST_HEAD(&per_cpu(rotation_list, cpu));
6546 static void __cpuinit perf_event_init_cpu(int cpu)
6548 struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
6550 mutex_lock(&swhash->hlist_mutex);
6551 if (swhash->hlist_refcount > 0) {
6552 struct swevent_hlist *hlist;
6554 hlist = kzalloc_node(sizeof(*hlist), GFP_KERNEL, cpu_to_node(cpu));
6556 rcu_assign_pointer(swhash->swevent_hlist, hlist);
6558 mutex_unlock(&swhash->hlist_mutex);
6561 #if defined CONFIG_HOTPLUG_CPU || defined CONFIG_KEXEC
6562 static void perf_pmu_rotate_stop(struct pmu *pmu)
6564 struct perf_cpu_context *cpuctx = this_cpu_ptr(pmu->pmu_cpu_context);
6566 WARN_ON(!irqs_disabled());
6568 list_del_init(&cpuctx->rotation_list);
6571 static void __perf_event_exit_context(void *__info)
6573 struct perf_event_context *ctx = __info;
6574 struct perf_event *event, *tmp;
6576 perf_pmu_rotate_stop(ctx->pmu);
6578 list_for_each_entry_safe(event, tmp, &ctx->pinned_groups, group_entry)
6579 __perf_event_remove_from_context(event);
6580 list_for_each_entry_safe(event, tmp, &ctx->flexible_groups, group_entry)
6581 __perf_event_remove_from_context(event);
6584 static void perf_event_exit_cpu_context(int cpu)
6586 struct perf_event_context *ctx;
6590 idx = srcu_read_lock(&pmus_srcu);
6591 list_for_each_entry_rcu(pmu, &pmus, entry) {
6592 ctx = &per_cpu_ptr(pmu->pmu_cpu_context, cpu)->ctx;
6594 mutex_lock(&ctx->mutex);
6595 smp_call_function_single(cpu, __perf_event_exit_context, ctx, 1);
6596 mutex_unlock(&ctx->mutex);
6598 srcu_read_unlock(&pmus_srcu, idx);
6601 static void perf_event_exit_cpu(int cpu)
6603 struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
6605 mutex_lock(&swhash->hlist_mutex);
6606 swevent_hlist_release(swhash);
6607 mutex_unlock(&swhash->hlist_mutex);
6609 perf_event_exit_cpu_context(cpu);
6612 static inline void perf_event_exit_cpu(int cpu) { }
6616 perf_reboot(struct notifier_block *notifier, unsigned long val, void *v)
6620 for_each_online_cpu(cpu)
6621 perf_event_exit_cpu(cpu);
6627 * Run the perf reboot notifier at the very last possible moment so that
6628 * the generic watchdog code runs as long as possible.
6630 static struct notifier_block perf_reboot_notifier = {
6631 .notifier_call = perf_reboot,
6632 .priority = INT_MIN,
6635 static int __cpuinit
6636 perf_cpu_notify(struct notifier_block *self, unsigned long action, void *hcpu)
6638 unsigned int cpu = (long)hcpu;
6640 switch (action & ~CPU_TASKS_FROZEN) {
6642 case CPU_UP_PREPARE:
6643 case CPU_DOWN_FAILED:
6644 perf_event_init_cpu(cpu);
6647 case CPU_UP_CANCELED:
6648 case CPU_DOWN_PREPARE:
6649 perf_event_exit_cpu(cpu);
6659 void __init perf_event_init(void)
6665 perf_event_init_all_cpus();
6666 init_srcu_struct(&pmus_srcu);
6667 perf_pmu_register(&perf_swevent, "software", PERF_TYPE_SOFTWARE);
6668 perf_pmu_register(&perf_cpu_clock, NULL, -1);
6669 perf_pmu_register(&perf_task_clock, NULL, -1);
6671 perf_cpu_notifier(perf_cpu_notify);
6672 register_reboot_notifier(&perf_reboot_notifier);
6674 ret = init_hw_breakpoint();
6675 WARN(ret, "hw_breakpoint initialization failed with: %d", ret);
6678 static int __init perf_event_sysfs_init(void)
6683 mutex_lock(&pmus_lock);
6685 ret = bus_register(&pmu_bus);
6689 list_for_each_entry(pmu, &pmus, entry) {
6690 if (!pmu->name || pmu->type < 0)
6693 ret = pmu_dev_alloc(pmu);
6694 WARN(ret, "Failed to register pmu: %s, reason %d\n", pmu->name, ret);
6696 pmu_bus_running = 1;
6700 mutex_unlock(&pmus_lock);
6704 device_initcall(perf_event_sysfs_init);