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;
271 static u64 perf_event_time(struct perf_event *event)
273 struct perf_event_context *ctx = event->ctx;
274 return ctx ? ctx->time : 0;
278 * Update the total_time_enabled and total_time_running fields for a event.
280 static void update_event_times(struct perf_event *event)
282 struct perf_event_context *ctx = event->ctx;
285 if (event->state < PERF_EVENT_STATE_INACTIVE ||
286 event->group_leader->state < PERF_EVENT_STATE_INACTIVE)
290 run_end = perf_event_time(event);
292 run_end = event->tstamp_stopped;
294 event->total_time_enabled = run_end - event->tstamp_enabled;
296 if (event->state == PERF_EVENT_STATE_INACTIVE)
297 run_end = event->tstamp_stopped;
299 run_end = perf_event_time(event);
301 event->total_time_running = run_end - event->tstamp_running;
305 * Update total_time_enabled and total_time_running for all events in a group.
307 static void update_group_times(struct perf_event *leader)
309 struct perf_event *event;
311 update_event_times(leader);
312 list_for_each_entry(event, &leader->sibling_list, group_entry)
313 update_event_times(event);
316 static struct list_head *
317 ctx_group_list(struct perf_event *event, struct perf_event_context *ctx)
319 if (event->attr.pinned)
320 return &ctx->pinned_groups;
322 return &ctx->flexible_groups;
326 * Add a event from the lists for its context.
327 * Must be called with ctx->mutex and ctx->lock held.
330 list_add_event(struct perf_event *event, struct perf_event_context *ctx)
332 WARN_ON_ONCE(event->attach_state & PERF_ATTACH_CONTEXT);
333 event->attach_state |= PERF_ATTACH_CONTEXT;
336 * If we're a stand alone event or group leader, we go to the context
337 * list, group events are kept attached to the group so that
338 * perf_group_detach can, at all times, locate all siblings.
340 if (event->group_leader == event) {
341 struct list_head *list;
343 if (is_software_event(event))
344 event->group_flags |= PERF_GROUP_SOFTWARE;
346 list = ctx_group_list(event, ctx);
347 list_add_tail(&event->group_entry, list);
350 list_add_rcu(&event->event_entry, &ctx->event_list);
352 perf_pmu_rotate_start(ctx->pmu);
354 if (event->attr.inherit_stat)
359 * Called at perf_event creation and when events are attached/detached from a
362 static void perf_event__read_size(struct perf_event *event)
364 int entry = sizeof(u64); /* value */
368 if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
371 if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
374 if (event->attr.read_format & PERF_FORMAT_ID)
375 entry += sizeof(u64);
377 if (event->attr.read_format & PERF_FORMAT_GROUP) {
378 nr += event->group_leader->nr_siblings;
383 event->read_size = size;
386 static void perf_event__header_size(struct perf_event *event)
388 struct perf_sample_data *data;
389 u64 sample_type = event->attr.sample_type;
392 perf_event__read_size(event);
394 if (sample_type & PERF_SAMPLE_IP)
395 size += sizeof(data->ip);
397 if (sample_type & PERF_SAMPLE_ADDR)
398 size += sizeof(data->addr);
400 if (sample_type & PERF_SAMPLE_PERIOD)
401 size += sizeof(data->period);
403 if (sample_type & PERF_SAMPLE_READ)
404 size += event->read_size;
406 event->header_size = size;
409 static void perf_event__id_header_size(struct perf_event *event)
411 struct perf_sample_data *data;
412 u64 sample_type = event->attr.sample_type;
415 if (sample_type & PERF_SAMPLE_TID)
416 size += sizeof(data->tid_entry);
418 if (sample_type & PERF_SAMPLE_TIME)
419 size += sizeof(data->time);
421 if (sample_type & PERF_SAMPLE_ID)
422 size += sizeof(data->id);
424 if (sample_type & PERF_SAMPLE_STREAM_ID)
425 size += sizeof(data->stream_id);
427 if (sample_type & PERF_SAMPLE_CPU)
428 size += sizeof(data->cpu_entry);
430 event->id_header_size = size;
433 static void perf_group_attach(struct perf_event *event)
435 struct perf_event *group_leader = event->group_leader, *pos;
438 * We can have double attach due to group movement in perf_event_open.
440 if (event->attach_state & PERF_ATTACH_GROUP)
443 event->attach_state |= PERF_ATTACH_GROUP;
445 if (group_leader == event)
448 if (group_leader->group_flags & PERF_GROUP_SOFTWARE &&
449 !is_software_event(event))
450 group_leader->group_flags &= ~PERF_GROUP_SOFTWARE;
452 list_add_tail(&event->group_entry, &group_leader->sibling_list);
453 group_leader->nr_siblings++;
455 perf_event__header_size(group_leader);
457 list_for_each_entry(pos, &group_leader->sibling_list, group_entry)
458 perf_event__header_size(pos);
462 * Remove a event from the lists for its context.
463 * Must be called with ctx->mutex and ctx->lock held.
466 list_del_event(struct perf_event *event, struct perf_event_context *ctx)
469 * We can have double detach due to exit/hot-unplug + close.
471 if (!(event->attach_state & PERF_ATTACH_CONTEXT))
474 event->attach_state &= ~PERF_ATTACH_CONTEXT;
477 if (event->attr.inherit_stat)
480 list_del_rcu(&event->event_entry);
482 if (event->group_leader == event)
483 list_del_init(&event->group_entry);
485 update_group_times(event);
488 * If event was in error state, then keep it
489 * that way, otherwise bogus counts will be
490 * returned on read(). The only way to get out
491 * of error state is by explicit re-enabling
494 if (event->state > PERF_EVENT_STATE_OFF)
495 event->state = PERF_EVENT_STATE_OFF;
498 static void perf_group_detach(struct perf_event *event)
500 struct perf_event *sibling, *tmp;
501 struct list_head *list = NULL;
504 * We can have double detach due to exit/hot-unplug + close.
506 if (!(event->attach_state & PERF_ATTACH_GROUP))
509 event->attach_state &= ~PERF_ATTACH_GROUP;
512 * If this is a sibling, remove it from its group.
514 if (event->group_leader != event) {
515 list_del_init(&event->group_entry);
516 event->group_leader->nr_siblings--;
520 if (!list_empty(&event->group_entry))
521 list = &event->group_entry;
524 * If this was a group event with sibling events then
525 * upgrade the siblings to singleton events by adding them
526 * to whatever list we are on.
528 list_for_each_entry_safe(sibling, tmp, &event->sibling_list, group_entry) {
530 list_move_tail(&sibling->group_entry, list);
531 sibling->group_leader = sibling;
533 /* Inherit group flags from the previous leader */
534 sibling->group_flags = event->group_flags;
538 perf_event__header_size(event->group_leader);
540 list_for_each_entry(tmp, &event->group_leader->sibling_list, group_entry)
541 perf_event__header_size(tmp);
545 event_filter_match(struct perf_event *event)
547 return event->cpu == -1 || event->cpu == smp_processor_id();
551 event_sched_out(struct perf_event *event,
552 struct perf_cpu_context *cpuctx,
553 struct perf_event_context *ctx)
555 u64 tstamp = perf_event_time(event);
558 * An event which could not be activated because of
559 * filter mismatch still needs to have its timings
560 * maintained, otherwise bogus information is return
561 * via read() for time_enabled, time_running:
563 if (event->state == PERF_EVENT_STATE_INACTIVE
564 && !event_filter_match(event)) {
565 delta = ctx->time - event->tstamp_stopped;
566 event->tstamp_running += delta;
567 event->tstamp_stopped = tstamp;
570 if (event->state != PERF_EVENT_STATE_ACTIVE)
573 event->state = PERF_EVENT_STATE_INACTIVE;
574 if (event->pending_disable) {
575 event->pending_disable = 0;
576 event->state = PERF_EVENT_STATE_OFF;
578 event->tstamp_stopped = tstamp;
579 event->pmu->del(event, 0);
582 if (!is_software_event(event))
583 cpuctx->active_oncpu--;
585 if (event->attr.exclusive || !cpuctx->active_oncpu)
586 cpuctx->exclusive = 0;
590 group_sched_out(struct perf_event *group_event,
591 struct perf_cpu_context *cpuctx,
592 struct perf_event_context *ctx)
594 struct perf_event *event;
595 int state = group_event->state;
597 event_sched_out(group_event, cpuctx, ctx);
600 * Schedule out siblings (if any):
602 list_for_each_entry(event, &group_event->sibling_list, group_entry)
603 event_sched_out(event, cpuctx, ctx);
605 if (state == PERF_EVENT_STATE_ACTIVE && group_event->attr.exclusive)
606 cpuctx->exclusive = 0;
609 static inline struct perf_cpu_context *
610 __get_cpu_context(struct perf_event_context *ctx)
612 return this_cpu_ptr(ctx->pmu->pmu_cpu_context);
616 * Cross CPU call to remove a performance event
618 * We disable the event on the hardware level first. After that we
619 * remove it from the context list.
621 static void __perf_event_remove_from_context(void *info)
623 struct perf_event *event = info;
624 struct perf_event_context *ctx = event->ctx;
625 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
628 * If this is a task context, we need to check whether it is
629 * the current task context of this cpu. If not it has been
630 * scheduled out before the smp call arrived.
632 if (ctx->task && cpuctx->task_ctx != ctx)
635 raw_spin_lock(&ctx->lock);
637 event_sched_out(event, cpuctx, ctx);
639 list_del_event(event, ctx);
641 raw_spin_unlock(&ctx->lock);
646 * Remove the event from a task's (or a CPU's) list of events.
648 * Must be called with ctx->mutex held.
650 * CPU events are removed with a smp call. For task events we only
651 * call when the task is on a CPU.
653 * If event->ctx is a cloned context, callers must make sure that
654 * every task struct that event->ctx->task could possibly point to
655 * remains valid. This is OK when called from perf_release since
656 * that only calls us on the top-level context, which can't be a clone.
657 * When called from perf_event_exit_task, it's OK because the
658 * context has been detached from its task.
660 static void perf_event_remove_from_context(struct perf_event *event)
662 struct perf_event_context *ctx = event->ctx;
663 struct task_struct *task = ctx->task;
667 * Per cpu events are removed via an smp call and
668 * the removal is always successful.
670 smp_call_function_single(event->cpu,
671 __perf_event_remove_from_context,
677 task_oncpu_function_call(task, __perf_event_remove_from_context,
680 raw_spin_lock_irq(&ctx->lock);
682 * If the context is active we need to retry the smp call.
684 if (ctx->nr_active && !list_empty(&event->group_entry)) {
685 raw_spin_unlock_irq(&ctx->lock);
690 * The lock prevents that this context is scheduled in so we
691 * can remove the event safely, if the call above did not
694 if (!list_empty(&event->group_entry))
695 list_del_event(event, ctx);
696 raw_spin_unlock_irq(&ctx->lock);
700 * Cross CPU call to disable a performance event
702 static void __perf_event_disable(void *info)
704 struct perf_event *event = info;
705 struct perf_event_context *ctx = event->ctx;
706 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
709 * If this is a per-task event, need to check whether this
710 * event's task is the current task on this cpu.
712 if (ctx->task && cpuctx->task_ctx != ctx)
715 raw_spin_lock(&ctx->lock);
718 * If the event is on, turn it off.
719 * If it is in error state, leave it in error state.
721 if (event->state >= PERF_EVENT_STATE_INACTIVE) {
722 update_context_time(ctx);
723 update_group_times(event);
724 if (event == event->group_leader)
725 group_sched_out(event, cpuctx, ctx);
727 event_sched_out(event, cpuctx, ctx);
728 event->state = PERF_EVENT_STATE_OFF;
731 raw_spin_unlock(&ctx->lock);
737 * If event->ctx is a cloned context, callers must make sure that
738 * every task struct that event->ctx->task could possibly point to
739 * remains valid. This condition is satisifed when called through
740 * perf_event_for_each_child or perf_event_for_each because they
741 * hold the top-level event's child_mutex, so any descendant that
742 * goes to exit will block in sync_child_event.
743 * When called from perf_pending_event it's OK because event->ctx
744 * is the current context on this CPU and preemption is disabled,
745 * hence we can't get into perf_event_task_sched_out for this context.
747 void perf_event_disable(struct perf_event *event)
749 struct perf_event_context *ctx = event->ctx;
750 struct task_struct *task = ctx->task;
754 * Disable the event on the cpu that it's on
756 smp_call_function_single(event->cpu, __perf_event_disable,
762 task_oncpu_function_call(task, __perf_event_disable, event);
764 raw_spin_lock_irq(&ctx->lock);
766 * If the event is still active, we need to retry the cross-call.
768 if (event->state == PERF_EVENT_STATE_ACTIVE) {
769 raw_spin_unlock_irq(&ctx->lock);
774 * Since we have the lock this context can't be scheduled
775 * in, so we can change the state safely.
777 if (event->state == PERF_EVENT_STATE_INACTIVE) {
778 update_group_times(event);
779 event->state = PERF_EVENT_STATE_OFF;
782 raw_spin_unlock_irq(&ctx->lock);
786 event_sched_in(struct perf_event *event,
787 struct perf_cpu_context *cpuctx,
788 struct perf_event_context *ctx)
790 u64 tstamp = perf_event_time(event);
792 if (event->state <= PERF_EVENT_STATE_OFF)
795 event->state = PERF_EVENT_STATE_ACTIVE;
796 event->oncpu = smp_processor_id();
798 * The new state must be visible before we turn it on in the hardware:
802 if (event->pmu->add(event, PERF_EF_START)) {
803 event->state = PERF_EVENT_STATE_INACTIVE;
808 event->tstamp_running += tstamp - event->tstamp_stopped;
810 event->shadow_ctx_time = tstamp - ctx->timestamp;
812 if (!is_software_event(event))
813 cpuctx->active_oncpu++;
816 if (event->attr.exclusive)
817 cpuctx->exclusive = 1;
823 group_sched_in(struct perf_event *group_event,
824 struct perf_cpu_context *cpuctx,
825 struct perf_event_context *ctx)
827 struct perf_event *event, *partial_group = NULL;
828 struct pmu *pmu = group_event->pmu;
830 bool simulate = false;
832 if (group_event->state == PERF_EVENT_STATE_OFF)
837 if (event_sched_in(group_event, cpuctx, ctx)) {
838 pmu->cancel_txn(pmu);
843 * Schedule in siblings as one group (if any):
845 list_for_each_entry(event, &group_event->sibling_list, group_entry) {
846 if (event_sched_in(event, cpuctx, ctx)) {
847 partial_group = event;
852 if (!pmu->commit_txn(pmu))
857 * Groups can be scheduled in as one unit only, so undo any
858 * partial group before returning:
859 * The events up to the failed event are scheduled out normally,
860 * tstamp_stopped will be updated.
862 * The failed events and the remaining siblings need to have
863 * their timings updated as if they had gone thru event_sched_in()
864 * and event_sched_out(). This is required to get consistent timings
865 * across the group. This also takes care of the case where the group
866 * could never be scheduled by ensuring tstamp_stopped is set to mark
867 * the time the event was actually stopped, such that time delta
868 * calculation in update_event_times() is correct.
870 list_for_each_entry(event, &group_event->sibling_list, group_entry) {
871 if (event == partial_group)
875 event->tstamp_running += now - event->tstamp_stopped;
876 event->tstamp_stopped = now;
878 event_sched_out(event, cpuctx, ctx);
881 event_sched_out(group_event, cpuctx, ctx);
883 pmu->cancel_txn(pmu);
889 * Work out whether we can put this event group on the CPU now.
891 static int group_can_go_on(struct perf_event *event,
892 struct perf_cpu_context *cpuctx,
896 * Groups consisting entirely of software events can always go on.
898 if (event->group_flags & PERF_GROUP_SOFTWARE)
901 * If an exclusive group is already on, no other hardware
904 if (cpuctx->exclusive)
907 * If this group is exclusive and there are already
908 * events on the CPU, it can't go on.
910 if (event->attr.exclusive && cpuctx->active_oncpu)
913 * Otherwise, try to add it if all previous groups were able
919 static void add_event_to_ctx(struct perf_event *event,
920 struct perf_event_context *ctx)
922 u64 tstamp = perf_event_time(event);
924 list_add_event(event, ctx);
925 perf_group_attach(event);
926 event->tstamp_enabled = tstamp;
927 event->tstamp_running = tstamp;
928 event->tstamp_stopped = tstamp;
932 * Cross CPU call to install and enable a performance event
934 * Must be called with ctx->mutex held
936 static void __perf_install_in_context(void *info)
938 struct perf_event *event = info;
939 struct perf_event_context *ctx = event->ctx;
940 struct perf_event *leader = event->group_leader;
941 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
945 * If this is a task context, we need to check whether it is
946 * the current task context of this cpu. If not it has been
947 * scheduled out before the smp call arrived.
948 * Or possibly this is the right context but it isn't
949 * on this cpu because it had no events.
951 if (ctx->task && cpuctx->task_ctx != ctx) {
952 if (cpuctx->task_ctx || ctx->task != current)
954 cpuctx->task_ctx = ctx;
957 raw_spin_lock(&ctx->lock);
959 update_context_time(ctx);
961 add_event_to_ctx(event, ctx);
963 if (!event_filter_match(event))
967 * Don't put the event on if it is disabled or if
968 * it is in a group and the group isn't on.
970 if (event->state != PERF_EVENT_STATE_INACTIVE ||
971 (leader != event && leader->state != PERF_EVENT_STATE_ACTIVE))
975 * An exclusive event can't go on if there are already active
976 * hardware events, and no hardware event can go on if there
977 * is already an exclusive event on.
979 if (!group_can_go_on(event, cpuctx, 1))
982 err = event_sched_in(event, cpuctx, ctx);
986 * This event couldn't go on. If it is in a group
987 * then we have to pull the whole group off.
988 * If the event group is pinned then put it in error state.
991 group_sched_out(leader, cpuctx, ctx);
992 if (leader->attr.pinned) {
993 update_group_times(leader);
994 leader->state = PERF_EVENT_STATE_ERROR;
999 raw_spin_unlock(&ctx->lock);
1003 * Attach a performance event to a context
1005 * First we add the event to the list with the hardware enable bit
1006 * in event->hw_config cleared.
1008 * If the event is attached to a task which is on a CPU we use a smp
1009 * call to enable it in the task context. The task might have been
1010 * scheduled away, but we check this in the smp call again.
1012 * Must be called with ctx->mutex held.
1015 perf_install_in_context(struct perf_event_context *ctx,
1016 struct perf_event *event,
1019 struct task_struct *task = ctx->task;
1025 * Per cpu events are installed via an smp call and
1026 * the install is always successful.
1028 smp_call_function_single(cpu, __perf_install_in_context,
1034 task_oncpu_function_call(task, __perf_install_in_context,
1037 raw_spin_lock_irq(&ctx->lock);
1039 * we need to retry the smp call.
1041 if (ctx->is_active && list_empty(&event->group_entry)) {
1042 raw_spin_unlock_irq(&ctx->lock);
1047 * The lock prevents that this context is scheduled in so we
1048 * can add the event safely, if it the call above did not
1051 if (list_empty(&event->group_entry))
1052 add_event_to_ctx(event, ctx);
1053 raw_spin_unlock_irq(&ctx->lock);
1057 * Put a event into inactive state and update time fields.
1058 * Enabling the leader of a group effectively enables all
1059 * the group members that aren't explicitly disabled, so we
1060 * have to update their ->tstamp_enabled also.
1061 * Note: this works for group members as well as group leaders
1062 * since the non-leader members' sibling_lists will be empty.
1064 static void __perf_event_mark_enabled(struct perf_event *event,
1065 struct perf_event_context *ctx)
1067 struct perf_event *sub;
1068 u64 tstamp = perf_event_time(event);
1070 event->state = PERF_EVENT_STATE_INACTIVE;
1071 event->tstamp_enabled = tstamp - event->total_time_enabled;
1072 list_for_each_entry(sub, &event->sibling_list, group_entry) {
1073 if (sub->state >= PERF_EVENT_STATE_INACTIVE)
1074 sub->tstamp_enabled = tstamp - sub->total_time_enabled;
1079 * Cross CPU call to enable a performance event
1081 static void __perf_event_enable(void *info)
1083 struct perf_event *event = info;
1084 struct perf_event_context *ctx = event->ctx;
1085 struct perf_event *leader = event->group_leader;
1086 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
1090 * If this is a per-task event, need to check whether this
1091 * event's task is the current task on this cpu.
1093 if (ctx->task && cpuctx->task_ctx != ctx) {
1094 if (cpuctx->task_ctx || ctx->task != current)
1096 cpuctx->task_ctx = ctx;
1099 raw_spin_lock(&ctx->lock);
1101 update_context_time(ctx);
1103 if (event->state >= PERF_EVENT_STATE_INACTIVE)
1105 __perf_event_mark_enabled(event, ctx);
1107 if (!event_filter_match(event))
1111 * If the event is in a group and isn't the group leader,
1112 * then don't put it on unless the group is on.
1114 if (leader != event && leader->state != PERF_EVENT_STATE_ACTIVE)
1117 if (!group_can_go_on(event, cpuctx, 1)) {
1120 if (event == leader)
1121 err = group_sched_in(event, cpuctx, ctx);
1123 err = event_sched_in(event, cpuctx, ctx);
1128 * If this event can't go on and it's part of a
1129 * group, then the whole group has to come off.
1131 if (leader != event)
1132 group_sched_out(leader, cpuctx, ctx);
1133 if (leader->attr.pinned) {
1134 update_group_times(leader);
1135 leader->state = PERF_EVENT_STATE_ERROR;
1140 raw_spin_unlock(&ctx->lock);
1146 * If event->ctx is a cloned context, callers must make sure that
1147 * every task struct that event->ctx->task could possibly point to
1148 * remains valid. This condition is satisfied when called through
1149 * perf_event_for_each_child or perf_event_for_each as described
1150 * for perf_event_disable.
1152 void perf_event_enable(struct perf_event *event)
1154 struct perf_event_context *ctx = event->ctx;
1155 struct task_struct *task = ctx->task;
1159 * Enable the event on the cpu that it's on
1161 smp_call_function_single(event->cpu, __perf_event_enable,
1166 raw_spin_lock_irq(&ctx->lock);
1167 if (event->state >= PERF_EVENT_STATE_INACTIVE)
1171 * If the event is in error state, clear that first.
1172 * That way, if we see the event in error state below, we
1173 * know that it has gone back into error state, as distinct
1174 * from the task having been scheduled away before the
1175 * cross-call arrived.
1177 if (event->state == PERF_EVENT_STATE_ERROR)
1178 event->state = PERF_EVENT_STATE_OFF;
1181 raw_spin_unlock_irq(&ctx->lock);
1182 task_oncpu_function_call(task, __perf_event_enable, event);
1184 raw_spin_lock_irq(&ctx->lock);
1187 * If the context is active and the event is still off,
1188 * we need to retry the cross-call.
1190 if (ctx->is_active && event->state == PERF_EVENT_STATE_OFF)
1194 * Since we have the lock this context can't be scheduled
1195 * in, so we can change the state safely.
1197 if (event->state == PERF_EVENT_STATE_OFF)
1198 __perf_event_mark_enabled(event, ctx);
1201 raw_spin_unlock_irq(&ctx->lock);
1204 static int perf_event_refresh(struct perf_event *event, int refresh)
1207 * not supported on inherited events
1209 if (event->attr.inherit || !is_sampling_event(event))
1212 atomic_add(refresh, &event->event_limit);
1213 perf_event_enable(event);
1218 static void ctx_sched_out(struct perf_event_context *ctx,
1219 struct perf_cpu_context *cpuctx,
1220 enum event_type_t event_type)
1222 struct perf_event *event;
1224 raw_spin_lock(&ctx->lock);
1225 perf_pmu_disable(ctx->pmu);
1227 if (likely(!ctx->nr_events))
1229 update_context_time(ctx);
1231 if (!ctx->nr_active)
1234 if (event_type & EVENT_PINNED) {
1235 list_for_each_entry(event, &ctx->pinned_groups, group_entry)
1236 group_sched_out(event, cpuctx, ctx);
1239 if (event_type & EVENT_FLEXIBLE) {
1240 list_for_each_entry(event, &ctx->flexible_groups, group_entry)
1241 group_sched_out(event, cpuctx, ctx);
1244 perf_pmu_enable(ctx->pmu);
1245 raw_spin_unlock(&ctx->lock);
1249 * Test whether two contexts are equivalent, i.e. whether they
1250 * have both been cloned from the same version of the same context
1251 * and they both have the same number of enabled events.
1252 * If the number of enabled events is the same, then the set
1253 * of enabled events should be the same, because these are both
1254 * inherited contexts, therefore we can't access individual events
1255 * in them directly with an fd; we can only enable/disable all
1256 * events via prctl, or enable/disable all events in a family
1257 * via ioctl, which will have the same effect on both contexts.
1259 static int context_equiv(struct perf_event_context *ctx1,
1260 struct perf_event_context *ctx2)
1262 return ctx1->parent_ctx && ctx1->parent_ctx == ctx2->parent_ctx
1263 && ctx1->parent_gen == ctx2->parent_gen
1264 && !ctx1->pin_count && !ctx2->pin_count;
1267 static void __perf_event_sync_stat(struct perf_event *event,
1268 struct perf_event *next_event)
1272 if (!event->attr.inherit_stat)
1276 * Update the event value, we cannot use perf_event_read()
1277 * because we're in the middle of a context switch and have IRQs
1278 * disabled, which upsets smp_call_function_single(), however
1279 * we know the event must be on the current CPU, therefore we
1280 * don't need to use it.
1282 switch (event->state) {
1283 case PERF_EVENT_STATE_ACTIVE:
1284 event->pmu->read(event);
1287 case PERF_EVENT_STATE_INACTIVE:
1288 update_event_times(event);
1296 * In order to keep per-task stats reliable we need to flip the event
1297 * values when we flip the contexts.
1299 value = local64_read(&next_event->count);
1300 value = local64_xchg(&event->count, value);
1301 local64_set(&next_event->count, value);
1303 swap(event->total_time_enabled, next_event->total_time_enabled);
1304 swap(event->total_time_running, next_event->total_time_running);
1307 * Since we swizzled the values, update the user visible data too.
1309 perf_event_update_userpage(event);
1310 perf_event_update_userpage(next_event);
1313 #define list_next_entry(pos, member) \
1314 list_entry(pos->member.next, typeof(*pos), member)
1316 static void perf_event_sync_stat(struct perf_event_context *ctx,
1317 struct perf_event_context *next_ctx)
1319 struct perf_event *event, *next_event;
1324 update_context_time(ctx);
1326 event = list_first_entry(&ctx->event_list,
1327 struct perf_event, event_entry);
1329 next_event = list_first_entry(&next_ctx->event_list,
1330 struct perf_event, event_entry);
1332 while (&event->event_entry != &ctx->event_list &&
1333 &next_event->event_entry != &next_ctx->event_list) {
1335 __perf_event_sync_stat(event, next_event);
1337 event = list_next_entry(event, event_entry);
1338 next_event = list_next_entry(next_event, event_entry);
1342 void perf_event_context_sched_out(struct task_struct *task, int ctxn,
1343 struct task_struct *next)
1345 struct perf_event_context *ctx = task->perf_event_ctxp[ctxn];
1346 struct perf_event_context *next_ctx;
1347 struct perf_event_context *parent;
1348 struct perf_cpu_context *cpuctx;
1354 cpuctx = __get_cpu_context(ctx);
1355 if (!cpuctx->task_ctx)
1359 parent = rcu_dereference(ctx->parent_ctx);
1360 next_ctx = next->perf_event_ctxp[ctxn];
1361 if (parent && next_ctx &&
1362 rcu_dereference(next_ctx->parent_ctx) == parent) {
1364 * Looks like the two contexts are clones, so we might be
1365 * able to optimize the context switch. We lock both
1366 * contexts and check that they are clones under the
1367 * lock (including re-checking that neither has been
1368 * uncloned in the meantime). It doesn't matter which
1369 * order we take the locks because no other cpu could
1370 * be trying to lock both of these tasks.
1372 raw_spin_lock(&ctx->lock);
1373 raw_spin_lock_nested(&next_ctx->lock, SINGLE_DEPTH_NESTING);
1374 if (context_equiv(ctx, next_ctx)) {
1376 * XXX do we need a memory barrier of sorts
1377 * wrt to rcu_dereference() of perf_event_ctxp
1379 task->perf_event_ctxp[ctxn] = next_ctx;
1380 next->perf_event_ctxp[ctxn] = ctx;
1382 next_ctx->task = task;
1385 perf_event_sync_stat(ctx, next_ctx);
1387 raw_spin_unlock(&next_ctx->lock);
1388 raw_spin_unlock(&ctx->lock);
1393 ctx_sched_out(ctx, cpuctx, EVENT_ALL);
1394 cpuctx->task_ctx = NULL;
1398 #define for_each_task_context_nr(ctxn) \
1399 for ((ctxn) = 0; (ctxn) < perf_nr_task_contexts; (ctxn)++)
1402 * Called from scheduler to remove the events of the current task,
1403 * with interrupts disabled.
1405 * We stop each event and update the event value in event->count.
1407 * This does not protect us against NMI, but disable()
1408 * sets the disabled bit in the control field of event _before_
1409 * accessing the event control register. If a NMI hits, then it will
1410 * not restart the event.
1412 void __perf_event_task_sched_out(struct task_struct *task,
1413 struct task_struct *next)
1417 for_each_task_context_nr(ctxn)
1418 perf_event_context_sched_out(task, ctxn, next);
1421 static void task_ctx_sched_out(struct perf_event_context *ctx,
1422 enum event_type_t event_type)
1424 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
1426 if (!cpuctx->task_ctx)
1429 if (WARN_ON_ONCE(ctx != cpuctx->task_ctx))
1432 ctx_sched_out(ctx, cpuctx, event_type);
1433 cpuctx->task_ctx = NULL;
1437 * Called with IRQs disabled
1439 static void cpu_ctx_sched_out(struct perf_cpu_context *cpuctx,
1440 enum event_type_t event_type)
1442 ctx_sched_out(&cpuctx->ctx, cpuctx, event_type);
1446 ctx_pinned_sched_in(struct perf_event_context *ctx,
1447 struct perf_cpu_context *cpuctx)
1449 struct perf_event *event;
1451 list_for_each_entry(event, &ctx->pinned_groups, group_entry) {
1452 if (event->state <= PERF_EVENT_STATE_OFF)
1454 if (!event_filter_match(event))
1457 if (group_can_go_on(event, cpuctx, 1))
1458 group_sched_in(event, cpuctx, ctx);
1461 * If this pinned group hasn't been scheduled,
1462 * put it in error state.
1464 if (event->state == PERF_EVENT_STATE_INACTIVE) {
1465 update_group_times(event);
1466 event->state = PERF_EVENT_STATE_ERROR;
1472 ctx_flexible_sched_in(struct perf_event_context *ctx,
1473 struct perf_cpu_context *cpuctx)
1475 struct perf_event *event;
1478 list_for_each_entry(event, &ctx->flexible_groups, group_entry) {
1479 /* Ignore events in OFF or ERROR state */
1480 if (event->state <= PERF_EVENT_STATE_OFF)
1483 * Listen to the 'cpu' scheduling filter constraint
1486 if (!event_filter_match(event))
1489 if (group_can_go_on(event, cpuctx, can_add_hw)) {
1490 if (group_sched_in(event, cpuctx, ctx))
1497 ctx_sched_in(struct perf_event_context *ctx,
1498 struct perf_cpu_context *cpuctx,
1499 enum event_type_t event_type)
1501 raw_spin_lock(&ctx->lock);
1503 if (likely(!ctx->nr_events))
1506 ctx->timestamp = perf_clock();
1509 * First go through the list and put on any pinned groups
1510 * in order to give them the best chance of going on.
1512 if (event_type & EVENT_PINNED)
1513 ctx_pinned_sched_in(ctx, cpuctx);
1515 /* Then walk through the lower prio flexible groups */
1516 if (event_type & EVENT_FLEXIBLE)
1517 ctx_flexible_sched_in(ctx, cpuctx);
1520 raw_spin_unlock(&ctx->lock);
1523 static void cpu_ctx_sched_in(struct perf_cpu_context *cpuctx,
1524 enum event_type_t event_type)
1526 struct perf_event_context *ctx = &cpuctx->ctx;
1528 ctx_sched_in(ctx, cpuctx, event_type);
1531 static void task_ctx_sched_in(struct perf_event_context *ctx,
1532 enum event_type_t event_type)
1534 struct perf_cpu_context *cpuctx;
1536 cpuctx = __get_cpu_context(ctx);
1537 if (cpuctx->task_ctx == ctx)
1540 ctx_sched_in(ctx, cpuctx, event_type);
1541 cpuctx->task_ctx = ctx;
1544 void perf_event_context_sched_in(struct perf_event_context *ctx)
1546 struct perf_cpu_context *cpuctx;
1548 cpuctx = __get_cpu_context(ctx);
1549 if (cpuctx->task_ctx == ctx)
1552 perf_pmu_disable(ctx->pmu);
1554 * We want to keep the following priority order:
1555 * cpu pinned (that don't need to move), task pinned,
1556 * cpu flexible, task flexible.
1558 cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE);
1560 ctx_sched_in(ctx, cpuctx, EVENT_PINNED);
1561 cpu_ctx_sched_in(cpuctx, EVENT_FLEXIBLE);
1562 ctx_sched_in(ctx, cpuctx, EVENT_FLEXIBLE);
1564 cpuctx->task_ctx = ctx;
1567 * Since these rotations are per-cpu, we need to ensure the
1568 * cpu-context we got scheduled on is actually rotating.
1570 perf_pmu_rotate_start(ctx->pmu);
1571 perf_pmu_enable(ctx->pmu);
1575 * Called from scheduler to add the events of the current task
1576 * with interrupts disabled.
1578 * We restore the event value and then enable it.
1580 * This does not protect us against NMI, but enable()
1581 * sets the enabled bit in the control field of event _before_
1582 * accessing the event control register. If a NMI hits, then it will
1583 * keep the event running.
1585 void __perf_event_task_sched_in(struct task_struct *task)
1587 struct perf_event_context *ctx;
1590 for_each_task_context_nr(ctxn) {
1591 ctx = task->perf_event_ctxp[ctxn];
1595 perf_event_context_sched_in(ctx);
1599 #define MAX_INTERRUPTS (~0ULL)
1601 static void perf_log_throttle(struct perf_event *event, int enable);
1603 static u64 perf_calculate_period(struct perf_event *event, u64 nsec, u64 count)
1605 u64 frequency = event->attr.sample_freq;
1606 u64 sec = NSEC_PER_SEC;
1607 u64 divisor, dividend;
1609 int count_fls, nsec_fls, frequency_fls, sec_fls;
1611 count_fls = fls64(count);
1612 nsec_fls = fls64(nsec);
1613 frequency_fls = fls64(frequency);
1617 * We got @count in @nsec, with a target of sample_freq HZ
1618 * the target period becomes:
1621 * period = -------------------
1622 * @nsec * sample_freq
1627 * Reduce accuracy by one bit such that @a and @b converge
1628 * to a similar magnitude.
1630 #define REDUCE_FLS(a, b) \
1632 if (a##_fls > b##_fls) { \
1642 * Reduce accuracy until either term fits in a u64, then proceed with
1643 * the other, so that finally we can do a u64/u64 division.
1645 while (count_fls + sec_fls > 64 && nsec_fls + frequency_fls > 64) {
1646 REDUCE_FLS(nsec, frequency);
1647 REDUCE_FLS(sec, count);
1650 if (count_fls + sec_fls > 64) {
1651 divisor = nsec * frequency;
1653 while (count_fls + sec_fls > 64) {
1654 REDUCE_FLS(count, sec);
1658 dividend = count * sec;
1660 dividend = count * sec;
1662 while (nsec_fls + frequency_fls > 64) {
1663 REDUCE_FLS(nsec, frequency);
1667 divisor = nsec * frequency;
1673 return div64_u64(dividend, divisor);
1676 static void perf_adjust_period(struct perf_event *event, u64 nsec, u64 count)
1678 struct hw_perf_event *hwc = &event->hw;
1679 s64 period, sample_period;
1682 period = perf_calculate_period(event, nsec, count);
1684 delta = (s64)(period - hwc->sample_period);
1685 delta = (delta + 7) / 8; /* low pass filter */
1687 sample_period = hwc->sample_period + delta;
1692 hwc->sample_period = sample_period;
1694 if (local64_read(&hwc->period_left) > 8*sample_period) {
1695 event->pmu->stop(event, PERF_EF_UPDATE);
1696 local64_set(&hwc->period_left, 0);
1697 event->pmu->start(event, PERF_EF_RELOAD);
1701 static void perf_ctx_adjust_freq(struct perf_event_context *ctx, u64 period)
1703 struct perf_event *event;
1704 struct hw_perf_event *hwc;
1705 u64 interrupts, now;
1708 raw_spin_lock(&ctx->lock);
1709 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
1710 if (event->state != PERF_EVENT_STATE_ACTIVE)
1713 if (!event_filter_match(event))
1718 interrupts = hwc->interrupts;
1719 hwc->interrupts = 0;
1722 * unthrottle events on the tick
1724 if (interrupts == MAX_INTERRUPTS) {
1725 perf_log_throttle(event, 1);
1726 event->pmu->start(event, 0);
1729 if (!event->attr.freq || !event->attr.sample_freq)
1732 event->pmu->read(event);
1733 now = local64_read(&event->count);
1734 delta = now - hwc->freq_count_stamp;
1735 hwc->freq_count_stamp = now;
1738 perf_adjust_period(event, period, delta);
1740 raw_spin_unlock(&ctx->lock);
1744 * Round-robin a context's events:
1746 static void rotate_ctx(struct perf_event_context *ctx)
1748 raw_spin_lock(&ctx->lock);
1751 * Rotate the first entry last of non-pinned groups. Rotation might be
1752 * disabled by the inheritance code.
1754 if (!ctx->rotate_disable)
1755 list_rotate_left(&ctx->flexible_groups);
1757 raw_spin_unlock(&ctx->lock);
1761 * perf_pmu_rotate_start() and perf_rotate_context() are fully serialized
1762 * because they're strictly cpu affine and rotate_start is called with IRQs
1763 * disabled, while rotate_context is called from IRQ context.
1765 static void perf_rotate_context(struct perf_cpu_context *cpuctx)
1767 u64 interval = (u64)cpuctx->jiffies_interval * TICK_NSEC;
1768 struct perf_event_context *ctx = NULL;
1769 int rotate = 0, remove = 1;
1771 if (cpuctx->ctx.nr_events) {
1773 if (cpuctx->ctx.nr_events != cpuctx->ctx.nr_active)
1777 ctx = cpuctx->task_ctx;
1778 if (ctx && ctx->nr_events) {
1780 if (ctx->nr_events != ctx->nr_active)
1784 perf_pmu_disable(cpuctx->ctx.pmu);
1785 perf_ctx_adjust_freq(&cpuctx->ctx, interval);
1787 perf_ctx_adjust_freq(ctx, interval);
1792 cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE);
1794 task_ctx_sched_out(ctx, EVENT_FLEXIBLE);
1796 rotate_ctx(&cpuctx->ctx);
1800 cpu_ctx_sched_in(cpuctx, EVENT_FLEXIBLE);
1802 task_ctx_sched_in(ctx, EVENT_FLEXIBLE);
1806 list_del_init(&cpuctx->rotation_list);
1808 perf_pmu_enable(cpuctx->ctx.pmu);
1811 void perf_event_task_tick(void)
1813 struct list_head *head = &__get_cpu_var(rotation_list);
1814 struct perf_cpu_context *cpuctx, *tmp;
1816 WARN_ON(!irqs_disabled());
1818 list_for_each_entry_safe(cpuctx, tmp, head, rotation_list) {
1819 if (cpuctx->jiffies_interval == 1 ||
1820 !(jiffies % cpuctx->jiffies_interval))
1821 perf_rotate_context(cpuctx);
1825 static int event_enable_on_exec(struct perf_event *event,
1826 struct perf_event_context *ctx)
1828 if (!event->attr.enable_on_exec)
1831 event->attr.enable_on_exec = 0;
1832 if (event->state >= PERF_EVENT_STATE_INACTIVE)
1835 __perf_event_mark_enabled(event, ctx);
1841 * Enable all of a task's events that have been marked enable-on-exec.
1842 * This expects task == current.
1844 static void perf_event_enable_on_exec(struct perf_event_context *ctx)
1846 struct perf_event *event;
1847 unsigned long flags;
1851 local_irq_save(flags);
1852 if (!ctx || !ctx->nr_events)
1855 task_ctx_sched_out(ctx, EVENT_ALL);
1857 raw_spin_lock(&ctx->lock);
1859 list_for_each_entry(event, &ctx->pinned_groups, group_entry) {
1860 ret = event_enable_on_exec(event, ctx);
1865 list_for_each_entry(event, &ctx->flexible_groups, group_entry) {
1866 ret = event_enable_on_exec(event, ctx);
1872 * Unclone this context if we enabled any event.
1877 raw_spin_unlock(&ctx->lock);
1879 perf_event_context_sched_in(ctx);
1881 local_irq_restore(flags);
1885 * Cross CPU call to read the hardware event
1887 static void __perf_event_read(void *info)
1889 struct perf_event *event = info;
1890 struct perf_event_context *ctx = event->ctx;
1891 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
1894 * If this is a task context, we need to check whether it is
1895 * the current task context of this cpu. If not it has been
1896 * scheduled out before the smp call arrived. In that case
1897 * event->count would have been updated to a recent sample
1898 * when the event was scheduled out.
1900 if (ctx->task && cpuctx->task_ctx != ctx)
1903 raw_spin_lock(&ctx->lock);
1904 update_context_time(ctx);
1905 update_event_times(event);
1906 raw_spin_unlock(&ctx->lock);
1908 event->pmu->read(event);
1911 static inline u64 perf_event_count(struct perf_event *event)
1913 return local64_read(&event->count) + atomic64_read(&event->child_count);
1916 static u64 perf_event_read(struct perf_event *event)
1919 * If event is enabled and currently active on a CPU, update the
1920 * value in the event structure:
1922 if (event->state == PERF_EVENT_STATE_ACTIVE) {
1923 smp_call_function_single(event->oncpu,
1924 __perf_event_read, event, 1);
1925 } else if (event->state == PERF_EVENT_STATE_INACTIVE) {
1926 struct perf_event_context *ctx = event->ctx;
1927 unsigned long flags;
1929 raw_spin_lock_irqsave(&ctx->lock, flags);
1931 * may read while context is not active
1932 * (e.g., thread is blocked), in that case
1933 * we cannot update context time
1936 update_context_time(ctx);
1937 update_event_times(event);
1938 raw_spin_unlock_irqrestore(&ctx->lock, flags);
1941 return perf_event_count(event);
1948 struct callchain_cpus_entries {
1949 struct rcu_head rcu_head;
1950 struct perf_callchain_entry *cpu_entries[0];
1953 static DEFINE_PER_CPU(int, callchain_recursion[PERF_NR_CONTEXTS]);
1954 static atomic_t nr_callchain_events;
1955 static DEFINE_MUTEX(callchain_mutex);
1956 struct callchain_cpus_entries *callchain_cpus_entries;
1959 __weak void perf_callchain_kernel(struct perf_callchain_entry *entry,
1960 struct pt_regs *regs)
1964 __weak void perf_callchain_user(struct perf_callchain_entry *entry,
1965 struct pt_regs *regs)
1969 static void release_callchain_buffers_rcu(struct rcu_head *head)
1971 struct callchain_cpus_entries *entries;
1974 entries = container_of(head, struct callchain_cpus_entries, rcu_head);
1976 for_each_possible_cpu(cpu)
1977 kfree(entries->cpu_entries[cpu]);
1982 static void release_callchain_buffers(void)
1984 struct callchain_cpus_entries *entries;
1986 entries = callchain_cpus_entries;
1987 rcu_assign_pointer(callchain_cpus_entries, NULL);
1988 call_rcu(&entries->rcu_head, release_callchain_buffers_rcu);
1991 static int alloc_callchain_buffers(void)
1995 struct callchain_cpus_entries *entries;
1998 * We can't use the percpu allocation API for data that can be
1999 * accessed from NMI. Use a temporary manual per cpu allocation
2000 * until that gets sorted out.
2002 size = sizeof(*entries) + sizeof(struct perf_callchain_entry *) *
2003 num_possible_cpus();
2005 entries = kzalloc(size, GFP_KERNEL);
2009 size = sizeof(struct perf_callchain_entry) * PERF_NR_CONTEXTS;
2011 for_each_possible_cpu(cpu) {
2012 entries->cpu_entries[cpu] = kmalloc_node(size, GFP_KERNEL,
2014 if (!entries->cpu_entries[cpu])
2018 rcu_assign_pointer(callchain_cpus_entries, entries);
2023 for_each_possible_cpu(cpu)
2024 kfree(entries->cpu_entries[cpu]);
2030 static int get_callchain_buffers(void)
2035 mutex_lock(&callchain_mutex);
2037 count = atomic_inc_return(&nr_callchain_events);
2038 if (WARN_ON_ONCE(count < 1)) {
2044 /* If the allocation failed, give up */
2045 if (!callchain_cpus_entries)
2050 err = alloc_callchain_buffers();
2052 release_callchain_buffers();
2054 mutex_unlock(&callchain_mutex);
2059 static void put_callchain_buffers(void)
2061 if (atomic_dec_and_mutex_lock(&nr_callchain_events, &callchain_mutex)) {
2062 release_callchain_buffers();
2063 mutex_unlock(&callchain_mutex);
2067 static int get_recursion_context(int *recursion)
2075 else if (in_softirq())
2080 if (recursion[rctx])
2089 static inline void put_recursion_context(int *recursion, int rctx)
2095 static struct perf_callchain_entry *get_callchain_entry(int *rctx)
2098 struct callchain_cpus_entries *entries;
2100 *rctx = get_recursion_context(__get_cpu_var(callchain_recursion));
2104 entries = rcu_dereference(callchain_cpus_entries);
2108 cpu = smp_processor_id();
2110 return &entries->cpu_entries[cpu][*rctx];
2114 put_callchain_entry(int rctx)
2116 put_recursion_context(__get_cpu_var(callchain_recursion), rctx);
2119 static struct perf_callchain_entry *perf_callchain(struct pt_regs *regs)
2122 struct perf_callchain_entry *entry;
2125 entry = get_callchain_entry(&rctx);
2134 if (!user_mode(regs)) {
2135 perf_callchain_store(entry, PERF_CONTEXT_KERNEL);
2136 perf_callchain_kernel(entry, regs);
2138 regs = task_pt_regs(current);
2144 perf_callchain_store(entry, PERF_CONTEXT_USER);
2145 perf_callchain_user(entry, regs);
2149 put_callchain_entry(rctx);
2155 * Initialize the perf_event context in a task_struct:
2157 static void __perf_event_init_context(struct perf_event_context *ctx)
2159 raw_spin_lock_init(&ctx->lock);
2160 mutex_init(&ctx->mutex);
2161 INIT_LIST_HEAD(&ctx->pinned_groups);
2162 INIT_LIST_HEAD(&ctx->flexible_groups);
2163 INIT_LIST_HEAD(&ctx->event_list);
2164 atomic_set(&ctx->refcount, 1);
2167 static struct perf_event_context *
2168 alloc_perf_context(struct pmu *pmu, struct task_struct *task)
2170 struct perf_event_context *ctx;
2172 ctx = kzalloc(sizeof(struct perf_event_context), GFP_KERNEL);
2176 __perf_event_init_context(ctx);
2179 get_task_struct(task);
2186 static struct task_struct *
2187 find_lively_task_by_vpid(pid_t vpid)
2189 struct task_struct *task;
2196 task = find_task_by_vpid(vpid);
2198 get_task_struct(task);
2202 return ERR_PTR(-ESRCH);
2205 * Can't attach events to a dying task.
2208 if (task->flags & PF_EXITING)
2211 /* Reuse ptrace permission checks for now. */
2213 if (!ptrace_may_access(task, PTRACE_MODE_READ))
2218 put_task_struct(task);
2219 return ERR_PTR(err);
2223 static struct perf_event_context *
2224 find_get_context(struct pmu *pmu, struct task_struct *task, int cpu)
2226 struct perf_event_context *ctx;
2227 struct perf_cpu_context *cpuctx;
2228 unsigned long flags;
2232 /* Must be root to operate on a CPU event: */
2233 if (perf_paranoid_cpu() && !capable(CAP_SYS_ADMIN))
2234 return ERR_PTR(-EACCES);
2237 * We could be clever and allow to attach a event to an
2238 * offline CPU and activate it when the CPU comes up, but
2241 if (!cpu_online(cpu))
2242 return ERR_PTR(-ENODEV);
2244 cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu);
2252 ctxn = pmu->task_ctx_nr;
2257 ctx = perf_lock_task_context(task, ctxn, &flags);
2260 raw_spin_unlock_irqrestore(&ctx->lock, flags);
2264 ctx = alloc_perf_context(pmu, task);
2271 if (cmpxchg(&task->perf_event_ctxp[ctxn], NULL, ctx)) {
2273 * We raced with some other task; use
2274 * the context they set.
2276 put_task_struct(task);
2285 return ERR_PTR(err);
2288 static void perf_event_free_filter(struct perf_event *event);
2290 static void free_event_rcu(struct rcu_head *head)
2292 struct perf_event *event;
2294 event = container_of(head, struct perf_event, rcu_head);
2296 put_pid_ns(event->ns);
2297 perf_event_free_filter(event);
2301 static void perf_buffer_put(struct perf_buffer *buffer);
2303 static void free_event(struct perf_event *event)
2305 irq_work_sync(&event->pending);
2307 if (!event->parent) {
2308 if (event->attach_state & PERF_ATTACH_TASK)
2309 jump_label_dec(&perf_task_events);
2310 if (event->attr.mmap || event->attr.mmap_data)
2311 atomic_dec(&nr_mmap_events);
2312 if (event->attr.comm)
2313 atomic_dec(&nr_comm_events);
2314 if (event->attr.task)
2315 atomic_dec(&nr_task_events);
2316 if (event->attr.sample_type & PERF_SAMPLE_CALLCHAIN)
2317 put_callchain_buffers();
2320 if (event->buffer) {
2321 perf_buffer_put(event->buffer);
2322 event->buffer = NULL;
2326 event->destroy(event);
2329 put_ctx(event->ctx);
2331 call_rcu(&event->rcu_head, free_event_rcu);
2334 int perf_event_release_kernel(struct perf_event *event)
2336 struct perf_event_context *ctx = event->ctx;
2339 * Remove from the PMU, can't get re-enabled since we got
2340 * here because the last ref went.
2342 perf_event_disable(event);
2344 WARN_ON_ONCE(ctx->parent_ctx);
2346 * There are two ways this annotation is useful:
2348 * 1) there is a lock recursion from perf_event_exit_task
2349 * see the comment there.
2351 * 2) there is a lock-inversion with mmap_sem through
2352 * perf_event_read_group(), which takes faults while
2353 * holding ctx->mutex, however this is called after
2354 * the last filedesc died, so there is no possibility
2355 * to trigger the AB-BA case.
2357 mutex_lock_nested(&ctx->mutex, SINGLE_DEPTH_NESTING);
2358 raw_spin_lock_irq(&ctx->lock);
2359 perf_group_detach(event);
2360 list_del_event(event, ctx);
2361 raw_spin_unlock_irq(&ctx->lock);
2362 mutex_unlock(&ctx->mutex);
2368 EXPORT_SYMBOL_GPL(perf_event_release_kernel);
2371 * Called when the last reference to the file is gone.
2373 static int perf_release(struct inode *inode, struct file *file)
2375 struct perf_event *event = file->private_data;
2376 struct task_struct *owner;
2378 file->private_data = NULL;
2381 owner = ACCESS_ONCE(event->owner);
2383 * Matches the smp_wmb() in perf_event_exit_task(). If we observe
2384 * !owner it means the list deletion is complete and we can indeed
2385 * free this event, otherwise we need to serialize on
2386 * owner->perf_event_mutex.
2388 smp_read_barrier_depends();
2391 * Since delayed_put_task_struct() also drops the last
2392 * task reference we can safely take a new reference
2393 * while holding the rcu_read_lock().
2395 get_task_struct(owner);
2400 mutex_lock(&owner->perf_event_mutex);
2402 * We have to re-check the event->owner field, if it is cleared
2403 * we raced with perf_event_exit_task(), acquiring the mutex
2404 * ensured they're done, and we can proceed with freeing the
2408 list_del_init(&event->owner_entry);
2409 mutex_unlock(&owner->perf_event_mutex);
2410 put_task_struct(owner);
2413 return perf_event_release_kernel(event);
2416 u64 perf_event_read_value(struct perf_event *event, u64 *enabled, u64 *running)
2418 struct perf_event *child;
2424 mutex_lock(&event->child_mutex);
2425 total += perf_event_read(event);
2426 *enabled += event->total_time_enabled +
2427 atomic64_read(&event->child_total_time_enabled);
2428 *running += event->total_time_running +
2429 atomic64_read(&event->child_total_time_running);
2431 list_for_each_entry(child, &event->child_list, child_list) {
2432 total += perf_event_read(child);
2433 *enabled += child->total_time_enabled;
2434 *running += child->total_time_running;
2436 mutex_unlock(&event->child_mutex);
2440 EXPORT_SYMBOL_GPL(perf_event_read_value);
2442 static int perf_event_read_group(struct perf_event *event,
2443 u64 read_format, char __user *buf)
2445 struct perf_event *leader = event->group_leader, *sub;
2446 int n = 0, size = 0, ret = -EFAULT;
2447 struct perf_event_context *ctx = leader->ctx;
2449 u64 count, enabled, running;
2451 mutex_lock(&ctx->mutex);
2452 count = perf_event_read_value(leader, &enabled, &running);
2454 values[n++] = 1 + leader->nr_siblings;
2455 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
2456 values[n++] = enabled;
2457 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
2458 values[n++] = running;
2459 values[n++] = count;
2460 if (read_format & PERF_FORMAT_ID)
2461 values[n++] = primary_event_id(leader);
2463 size = n * sizeof(u64);
2465 if (copy_to_user(buf, values, size))
2470 list_for_each_entry(sub, &leader->sibling_list, group_entry) {
2473 values[n++] = perf_event_read_value(sub, &enabled, &running);
2474 if (read_format & PERF_FORMAT_ID)
2475 values[n++] = primary_event_id(sub);
2477 size = n * sizeof(u64);
2479 if (copy_to_user(buf + ret, values, size)) {
2487 mutex_unlock(&ctx->mutex);
2492 static int perf_event_read_one(struct perf_event *event,
2493 u64 read_format, char __user *buf)
2495 u64 enabled, running;
2499 values[n++] = perf_event_read_value(event, &enabled, &running);
2500 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
2501 values[n++] = enabled;
2502 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
2503 values[n++] = running;
2504 if (read_format & PERF_FORMAT_ID)
2505 values[n++] = primary_event_id(event);
2507 if (copy_to_user(buf, values, n * sizeof(u64)))
2510 return n * sizeof(u64);
2514 * Read the performance event - simple non blocking version for now
2517 perf_read_hw(struct perf_event *event, char __user *buf, size_t count)
2519 u64 read_format = event->attr.read_format;
2523 * Return end-of-file for a read on a event that is in
2524 * error state (i.e. because it was pinned but it couldn't be
2525 * scheduled on to the CPU at some point).
2527 if (event->state == PERF_EVENT_STATE_ERROR)
2530 if (count < event->read_size)
2533 WARN_ON_ONCE(event->ctx->parent_ctx);
2534 if (read_format & PERF_FORMAT_GROUP)
2535 ret = perf_event_read_group(event, read_format, buf);
2537 ret = perf_event_read_one(event, read_format, buf);
2543 perf_read(struct file *file, char __user *buf, size_t count, loff_t *ppos)
2545 struct perf_event *event = file->private_data;
2547 return perf_read_hw(event, buf, count);
2550 static unsigned int perf_poll(struct file *file, poll_table *wait)
2552 struct perf_event *event = file->private_data;
2553 struct perf_buffer *buffer;
2554 unsigned int events = POLL_HUP;
2557 buffer = rcu_dereference(event->buffer);
2559 events = atomic_xchg(&buffer->poll, 0);
2562 poll_wait(file, &event->waitq, wait);
2567 static void perf_event_reset(struct perf_event *event)
2569 (void)perf_event_read(event);
2570 local64_set(&event->count, 0);
2571 perf_event_update_userpage(event);
2575 * Holding the top-level event's child_mutex means that any
2576 * descendant process that has inherited this event will block
2577 * in sync_child_event if it goes to exit, thus satisfying the
2578 * task existence requirements of perf_event_enable/disable.
2580 static void perf_event_for_each_child(struct perf_event *event,
2581 void (*func)(struct perf_event *))
2583 struct perf_event *child;
2585 WARN_ON_ONCE(event->ctx->parent_ctx);
2586 mutex_lock(&event->child_mutex);
2588 list_for_each_entry(child, &event->child_list, child_list)
2590 mutex_unlock(&event->child_mutex);
2593 static void perf_event_for_each(struct perf_event *event,
2594 void (*func)(struct perf_event *))
2596 struct perf_event_context *ctx = event->ctx;
2597 struct perf_event *sibling;
2599 WARN_ON_ONCE(ctx->parent_ctx);
2600 mutex_lock(&ctx->mutex);
2601 event = event->group_leader;
2603 perf_event_for_each_child(event, func);
2605 list_for_each_entry(sibling, &event->sibling_list, group_entry)
2606 perf_event_for_each_child(event, func);
2607 mutex_unlock(&ctx->mutex);
2610 static int perf_event_period(struct perf_event *event, u64 __user *arg)
2612 struct perf_event_context *ctx = event->ctx;
2616 if (!is_sampling_event(event))
2619 if (copy_from_user(&value, arg, sizeof(value)))
2625 raw_spin_lock_irq(&ctx->lock);
2626 if (event->attr.freq) {
2627 if (value > sysctl_perf_event_sample_rate) {
2632 event->attr.sample_freq = value;
2634 event->attr.sample_period = value;
2635 event->hw.sample_period = value;
2638 raw_spin_unlock_irq(&ctx->lock);
2643 static const struct file_operations perf_fops;
2645 static struct perf_event *perf_fget_light(int fd, int *fput_needed)
2649 file = fget_light(fd, fput_needed);
2651 return ERR_PTR(-EBADF);
2653 if (file->f_op != &perf_fops) {
2654 fput_light(file, *fput_needed);
2656 return ERR_PTR(-EBADF);
2659 return file->private_data;
2662 static int perf_event_set_output(struct perf_event *event,
2663 struct perf_event *output_event);
2664 static int perf_event_set_filter(struct perf_event *event, void __user *arg);
2666 static long perf_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
2668 struct perf_event *event = file->private_data;
2669 void (*func)(struct perf_event *);
2673 case PERF_EVENT_IOC_ENABLE:
2674 func = perf_event_enable;
2676 case PERF_EVENT_IOC_DISABLE:
2677 func = perf_event_disable;
2679 case PERF_EVENT_IOC_RESET:
2680 func = perf_event_reset;
2683 case PERF_EVENT_IOC_REFRESH:
2684 return perf_event_refresh(event, arg);
2686 case PERF_EVENT_IOC_PERIOD:
2687 return perf_event_period(event, (u64 __user *)arg);
2689 case PERF_EVENT_IOC_SET_OUTPUT:
2691 struct perf_event *output_event = NULL;
2692 int fput_needed = 0;
2696 output_event = perf_fget_light(arg, &fput_needed);
2697 if (IS_ERR(output_event))
2698 return PTR_ERR(output_event);
2701 ret = perf_event_set_output(event, output_event);
2703 fput_light(output_event->filp, fput_needed);
2708 case PERF_EVENT_IOC_SET_FILTER:
2709 return perf_event_set_filter(event, (void __user *)arg);
2715 if (flags & PERF_IOC_FLAG_GROUP)
2716 perf_event_for_each(event, func);
2718 perf_event_for_each_child(event, func);
2723 int perf_event_task_enable(void)
2725 struct perf_event *event;
2727 mutex_lock(¤t->perf_event_mutex);
2728 list_for_each_entry(event, ¤t->perf_event_list, owner_entry)
2729 perf_event_for_each_child(event, perf_event_enable);
2730 mutex_unlock(¤t->perf_event_mutex);
2735 int perf_event_task_disable(void)
2737 struct perf_event *event;
2739 mutex_lock(¤t->perf_event_mutex);
2740 list_for_each_entry(event, ¤t->perf_event_list, owner_entry)
2741 perf_event_for_each_child(event, perf_event_disable);
2742 mutex_unlock(¤t->perf_event_mutex);
2747 #ifndef PERF_EVENT_INDEX_OFFSET
2748 # define PERF_EVENT_INDEX_OFFSET 0
2751 static int perf_event_index(struct perf_event *event)
2753 if (event->hw.state & PERF_HES_STOPPED)
2756 if (event->state != PERF_EVENT_STATE_ACTIVE)
2759 return event->hw.idx + 1 - PERF_EVENT_INDEX_OFFSET;
2763 * Callers need to ensure there can be no nesting of this function, otherwise
2764 * the seqlock logic goes bad. We can not serialize this because the arch
2765 * code calls this from NMI context.
2767 void perf_event_update_userpage(struct perf_event *event)
2769 struct perf_event_mmap_page *userpg;
2770 struct perf_buffer *buffer;
2773 buffer = rcu_dereference(event->buffer);
2777 userpg = buffer->user_page;
2780 * Disable preemption so as to not let the corresponding user-space
2781 * spin too long if we get preempted.
2786 userpg->index = perf_event_index(event);
2787 userpg->offset = perf_event_count(event);
2788 if (event->state == PERF_EVENT_STATE_ACTIVE)
2789 userpg->offset -= local64_read(&event->hw.prev_count);
2791 userpg->time_enabled = event->total_time_enabled +
2792 atomic64_read(&event->child_total_time_enabled);
2794 userpg->time_running = event->total_time_running +
2795 atomic64_read(&event->child_total_time_running);
2804 static unsigned long perf_data_size(struct perf_buffer *buffer);
2807 perf_buffer_init(struct perf_buffer *buffer, long watermark, int flags)
2809 long max_size = perf_data_size(buffer);
2812 buffer->watermark = min(max_size, watermark);
2814 if (!buffer->watermark)
2815 buffer->watermark = max_size / 2;
2817 if (flags & PERF_BUFFER_WRITABLE)
2818 buffer->writable = 1;
2820 atomic_set(&buffer->refcount, 1);
2823 #ifndef CONFIG_PERF_USE_VMALLOC
2826 * Back perf_mmap() with regular GFP_KERNEL-0 pages.
2829 static struct page *
2830 perf_mmap_to_page(struct perf_buffer *buffer, unsigned long pgoff)
2832 if (pgoff > buffer->nr_pages)
2836 return virt_to_page(buffer->user_page);
2838 return virt_to_page(buffer->data_pages[pgoff - 1]);
2841 static void *perf_mmap_alloc_page(int cpu)
2846 node = (cpu == -1) ? cpu : cpu_to_node(cpu);
2847 page = alloc_pages_node(node, GFP_KERNEL | __GFP_ZERO, 0);
2851 return page_address(page);
2854 static struct perf_buffer *
2855 perf_buffer_alloc(int nr_pages, long watermark, int cpu, int flags)
2857 struct perf_buffer *buffer;
2861 size = sizeof(struct perf_buffer);
2862 size += nr_pages * sizeof(void *);
2864 buffer = kzalloc(size, GFP_KERNEL);
2868 buffer->user_page = perf_mmap_alloc_page(cpu);
2869 if (!buffer->user_page)
2870 goto fail_user_page;
2872 for (i = 0; i < nr_pages; i++) {
2873 buffer->data_pages[i] = perf_mmap_alloc_page(cpu);
2874 if (!buffer->data_pages[i])
2875 goto fail_data_pages;
2878 buffer->nr_pages = nr_pages;
2880 perf_buffer_init(buffer, watermark, flags);
2885 for (i--; i >= 0; i--)
2886 free_page((unsigned long)buffer->data_pages[i]);
2888 free_page((unsigned long)buffer->user_page);
2897 static void perf_mmap_free_page(unsigned long addr)
2899 struct page *page = virt_to_page((void *)addr);
2901 page->mapping = NULL;
2905 static void perf_buffer_free(struct perf_buffer *buffer)
2909 perf_mmap_free_page((unsigned long)buffer->user_page);
2910 for (i = 0; i < buffer->nr_pages; i++)
2911 perf_mmap_free_page((unsigned long)buffer->data_pages[i]);
2915 static inline int page_order(struct perf_buffer *buffer)
2923 * Back perf_mmap() with vmalloc memory.
2925 * Required for architectures that have d-cache aliasing issues.
2928 static inline int page_order(struct perf_buffer *buffer)
2930 return buffer->page_order;
2933 static struct page *
2934 perf_mmap_to_page(struct perf_buffer *buffer, unsigned long pgoff)
2936 if (pgoff > (1UL << page_order(buffer)))
2939 return vmalloc_to_page((void *)buffer->user_page + pgoff * PAGE_SIZE);
2942 static void perf_mmap_unmark_page(void *addr)
2944 struct page *page = vmalloc_to_page(addr);
2946 page->mapping = NULL;
2949 static void perf_buffer_free_work(struct work_struct *work)
2951 struct perf_buffer *buffer;
2955 buffer = container_of(work, struct perf_buffer, work);
2956 nr = 1 << page_order(buffer);
2958 base = buffer->user_page;
2959 for (i = 0; i < nr + 1; i++)
2960 perf_mmap_unmark_page(base + (i * PAGE_SIZE));
2966 static void perf_buffer_free(struct perf_buffer *buffer)
2968 schedule_work(&buffer->work);
2971 static struct perf_buffer *
2972 perf_buffer_alloc(int nr_pages, long watermark, int cpu, int flags)
2974 struct perf_buffer *buffer;
2978 size = sizeof(struct perf_buffer);
2979 size += sizeof(void *);
2981 buffer = kzalloc(size, GFP_KERNEL);
2985 INIT_WORK(&buffer->work, perf_buffer_free_work);
2987 all_buf = vmalloc_user((nr_pages + 1) * PAGE_SIZE);
2991 buffer->user_page = all_buf;
2992 buffer->data_pages[0] = all_buf + PAGE_SIZE;
2993 buffer->page_order = ilog2(nr_pages);
2994 buffer->nr_pages = 1;
2996 perf_buffer_init(buffer, watermark, flags);
3009 static unsigned long perf_data_size(struct perf_buffer *buffer)
3011 return buffer->nr_pages << (PAGE_SHIFT + page_order(buffer));
3014 static int perf_mmap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
3016 struct perf_event *event = vma->vm_file->private_data;
3017 struct perf_buffer *buffer;
3018 int ret = VM_FAULT_SIGBUS;
3020 if (vmf->flags & FAULT_FLAG_MKWRITE) {
3021 if (vmf->pgoff == 0)
3027 buffer = rcu_dereference(event->buffer);
3031 if (vmf->pgoff && (vmf->flags & FAULT_FLAG_WRITE))
3034 vmf->page = perf_mmap_to_page(buffer, vmf->pgoff);
3038 get_page(vmf->page);
3039 vmf->page->mapping = vma->vm_file->f_mapping;
3040 vmf->page->index = vmf->pgoff;
3049 static void perf_buffer_free_rcu(struct rcu_head *rcu_head)
3051 struct perf_buffer *buffer;
3053 buffer = container_of(rcu_head, struct perf_buffer, rcu_head);
3054 perf_buffer_free(buffer);
3057 static struct perf_buffer *perf_buffer_get(struct perf_event *event)
3059 struct perf_buffer *buffer;
3062 buffer = rcu_dereference(event->buffer);
3064 if (!atomic_inc_not_zero(&buffer->refcount))
3072 static void perf_buffer_put(struct perf_buffer *buffer)
3074 if (!atomic_dec_and_test(&buffer->refcount))
3077 call_rcu(&buffer->rcu_head, perf_buffer_free_rcu);
3080 static void perf_mmap_open(struct vm_area_struct *vma)
3082 struct perf_event *event = vma->vm_file->private_data;
3084 atomic_inc(&event->mmap_count);
3087 static void perf_mmap_close(struct vm_area_struct *vma)
3089 struct perf_event *event = vma->vm_file->private_data;
3091 if (atomic_dec_and_mutex_lock(&event->mmap_count, &event->mmap_mutex)) {
3092 unsigned long size = perf_data_size(event->buffer);
3093 struct user_struct *user = event->mmap_user;
3094 struct perf_buffer *buffer = event->buffer;
3096 atomic_long_sub((size >> PAGE_SHIFT) + 1, &user->locked_vm);
3097 vma->vm_mm->locked_vm -= event->mmap_locked;
3098 rcu_assign_pointer(event->buffer, NULL);
3099 mutex_unlock(&event->mmap_mutex);
3101 perf_buffer_put(buffer);
3106 static const struct vm_operations_struct perf_mmap_vmops = {
3107 .open = perf_mmap_open,
3108 .close = perf_mmap_close,
3109 .fault = perf_mmap_fault,
3110 .page_mkwrite = perf_mmap_fault,
3113 static int perf_mmap(struct file *file, struct vm_area_struct *vma)
3115 struct perf_event *event = file->private_data;
3116 unsigned long user_locked, user_lock_limit;
3117 struct user_struct *user = current_user();
3118 unsigned long locked, lock_limit;
3119 struct perf_buffer *buffer;
3120 unsigned long vma_size;
3121 unsigned long nr_pages;
3122 long user_extra, extra;
3123 int ret = 0, flags = 0;
3126 * Don't allow mmap() of inherited per-task counters. This would
3127 * create a performance issue due to all children writing to the
3130 if (event->cpu == -1 && event->attr.inherit)
3133 if (!(vma->vm_flags & VM_SHARED))
3136 vma_size = vma->vm_end - vma->vm_start;
3137 nr_pages = (vma_size / PAGE_SIZE) - 1;
3140 * If we have buffer pages ensure they're a power-of-two number, so we
3141 * can do bitmasks instead of modulo.
3143 if (nr_pages != 0 && !is_power_of_2(nr_pages))
3146 if (vma_size != PAGE_SIZE * (1 + nr_pages))
3149 if (vma->vm_pgoff != 0)
3152 WARN_ON_ONCE(event->ctx->parent_ctx);
3153 mutex_lock(&event->mmap_mutex);
3154 if (event->buffer) {
3155 if (event->buffer->nr_pages == nr_pages)
3156 atomic_inc(&event->buffer->refcount);
3162 user_extra = nr_pages + 1;
3163 user_lock_limit = sysctl_perf_event_mlock >> (PAGE_SHIFT - 10);
3166 * Increase the limit linearly with more CPUs:
3168 user_lock_limit *= num_online_cpus();
3170 user_locked = atomic_long_read(&user->locked_vm) + user_extra;
3173 if (user_locked > user_lock_limit)
3174 extra = user_locked - user_lock_limit;
3176 lock_limit = rlimit(RLIMIT_MEMLOCK);
3177 lock_limit >>= PAGE_SHIFT;
3178 locked = vma->vm_mm->locked_vm + extra;
3180 if ((locked > lock_limit) && perf_paranoid_tracepoint_raw() &&
3181 !capable(CAP_IPC_LOCK)) {
3186 WARN_ON(event->buffer);
3188 if (vma->vm_flags & VM_WRITE)
3189 flags |= PERF_BUFFER_WRITABLE;
3191 buffer = perf_buffer_alloc(nr_pages, event->attr.wakeup_watermark,
3197 rcu_assign_pointer(event->buffer, buffer);
3199 atomic_long_add(user_extra, &user->locked_vm);
3200 event->mmap_locked = extra;
3201 event->mmap_user = get_current_user();
3202 vma->vm_mm->locked_vm += event->mmap_locked;
3206 atomic_inc(&event->mmap_count);
3207 mutex_unlock(&event->mmap_mutex);
3209 vma->vm_flags |= VM_RESERVED;
3210 vma->vm_ops = &perf_mmap_vmops;
3215 static int perf_fasync(int fd, struct file *filp, int on)
3217 struct inode *inode = filp->f_path.dentry->d_inode;
3218 struct perf_event *event = filp->private_data;
3221 mutex_lock(&inode->i_mutex);
3222 retval = fasync_helper(fd, filp, on, &event->fasync);
3223 mutex_unlock(&inode->i_mutex);
3231 static const struct file_operations perf_fops = {
3232 .llseek = no_llseek,
3233 .release = perf_release,
3236 .unlocked_ioctl = perf_ioctl,
3237 .compat_ioctl = perf_ioctl,
3239 .fasync = perf_fasync,
3245 * If there's data, ensure we set the poll() state and publish everything
3246 * to user-space before waking everybody up.
3249 void perf_event_wakeup(struct perf_event *event)
3251 wake_up_all(&event->waitq);
3253 if (event->pending_kill) {
3254 kill_fasync(&event->fasync, SIGIO, event->pending_kill);
3255 event->pending_kill = 0;
3259 static void perf_pending_event(struct irq_work *entry)
3261 struct perf_event *event = container_of(entry,
3262 struct perf_event, pending);
3264 if (event->pending_disable) {
3265 event->pending_disable = 0;
3266 __perf_event_disable(event);
3269 if (event->pending_wakeup) {
3270 event->pending_wakeup = 0;
3271 perf_event_wakeup(event);
3276 * We assume there is only KVM supporting the callbacks.
3277 * Later on, we might change it to a list if there is
3278 * another virtualization implementation supporting the callbacks.
3280 struct perf_guest_info_callbacks *perf_guest_cbs;
3282 int perf_register_guest_info_callbacks(struct perf_guest_info_callbacks *cbs)
3284 perf_guest_cbs = cbs;
3287 EXPORT_SYMBOL_GPL(perf_register_guest_info_callbacks);
3289 int perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks *cbs)
3291 perf_guest_cbs = NULL;
3294 EXPORT_SYMBOL_GPL(perf_unregister_guest_info_callbacks);
3299 static bool perf_output_space(struct perf_buffer *buffer, unsigned long tail,
3300 unsigned long offset, unsigned long head)
3304 if (!buffer->writable)
3307 mask = perf_data_size(buffer) - 1;
3309 offset = (offset - tail) & mask;
3310 head = (head - tail) & mask;
3312 if ((int)(head - offset) < 0)
3318 static void perf_output_wakeup(struct perf_output_handle *handle)
3320 atomic_set(&handle->buffer->poll, POLL_IN);
3323 handle->event->pending_wakeup = 1;
3324 irq_work_queue(&handle->event->pending);
3326 perf_event_wakeup(handle->event);
3330 * We need to ensure a later event_id doesn't publish a head when a former
3331 * event isn't done writing. However since we need to deal with NMIs we
3332 * cannot fully serialize things.
3334 * We only publish the head (and generate a wakeup) when the outer-most
3337 static void perf_output_get_handle(struct perf_output_handle *handle)
3339 struct perf_buffer *buffer = handle->buffer;
3342 local_inc(&buffer->nest);
3343 handle->wakeup = local_read(&buffer->wakeup);
3346 static void perf_output_put_handle(struct perf_output_handle *handle)
3348 struct perf_buffer *buffer = handle->buffer;
3352 head = local_read(&buffer->head);
3355 * IRQ/NMI can happen here, which means we can miss a head update.
3358 if (!local_dec_and_test(&buffer->nest))
3362 * Publish the known good head. Rely on the full barrier implied
3363 * by atomic_dec_and_test() order the buffer->head read and this
3366 buffer->user_page->data_head = head;
3369 * Now check if we missed an update, rely on the (compiler)
3370 * barrier in atomic_dec_and_test() to re-read buffer->head.
3372 if (unlikely(head != local_read(&buffer->head))) {
3373 local_inc(&buffer->nest);
3377 if (handle->wakeup != local_read(&buffer->wakeup))
3378 perf_output_wakeup(handle);
3384 __always_inline void perf_output_copy(struct perf_output_handle *handle,
3385 const void *buf, unsigned int len)
3388 unsigned long size = min_t(unsigned long, handle->size, len);
3390 memcpy(handle->addr, buf, size);
3393 handle->addr += size;
3395 handle->size -= size;
3396 if (!handle->size) {
3397 struct perf_buffer *buffer = handle->buffer;
3400 handle->page &= buffer->nr_pages - 1;
3401 handle->addr = buffer->data_pages[handle->page];
3402 handle->size = PAGE_SIZE << page_order(buffer);
3407 static void __perf_event_header__init_id(struct perf_event_header *header,
3408 struct perf_sample_data *data,
3409 struct perf_event *event)
3411 u64 sample_type = event->attr.sample_type;
3413 data->type = sample_type;
3414 header->size += event->id_header_size;
3416 if (sample_type & PERF_SAMPLE_TID) {
3417 /* namespace issues */
3418 data->tid_entry.pid = perf_event_pid(event, current);
3419 data->tid_entry.tid = perf_event_tid(event, current);
3422 if (sample_type & PERF_SAMPLE_TIME)
3423 data->time = perf_clock();
3425 if (sample_type & PERF_SAMPLE_ID)
3426 data->id = primary_event_id(event);
3428 if (sample_type & PERF_SAMPLE_STREAM_ID)
3429 data->stream_id = event->id;
3431 if (sample_type & PERF_SAMPLE_CPU) {
3432 data->cpu_entry.cpu = raw_smp_processor_id();
3433 data->cpu_entry.reserved = 0;
3437 static void perf_event_header__init_id(struct perf_event_header *header,
3438 struct perf_sample_data *data,
3439 struct perf_event *event)
3441 if (event->attr.sample_id_all)
3442 __perf_event_header__init_id(header, data, event);
3445 static void __perf_event__output_id_sample(struct perf_output_handle *handle,
3446 struct perf_sample_data *data)
3448 u64 sample_type = data->type;
3450 if (sample_type & PERF_SAMPLE_TID)
3451 perf_output_put(handle, data->tid_entry);
3453 if (sample_type & PERF_SAMPLE_TIME)
3454 perf_output_put(handle, data->time);
3456 if (sample_type & PERF_SAMPLE_ID)
3457 perf_output_put(handle, data->id);
3459 if (sample_type & PERF_SAMPLE_STREAM_ID)
3460 perf_output_put(handle, data->stream_id);
3462 if (sample_type & PERF_SAMPLE_CPU)
3463 perf_output_put(handle, data->cpu_entry);
3466 static void perf_event__output_id_sample(struct perf_event *event,
3467 struct perf_output_handle *handle,
3468 struct perf_sample_data *sample)
3470 if (event->attr.sample_id_all)
3471 __perf_event__output_id_sample(handle, sample);
3474 int perf_output_begin(struct perf_output_handle *handle,
3475 struct perf_event *event, unsigned int size,
3476 int nmi, int sample)
3478 struct perf_buffer *buffer;
3479 unsigned long tail, offset, head;
3481 struct perf_sample_data sample_data;
3483 struct perf_event_header header;
3490 * For inherited events we send all the output towards the parent.
3493 event = event->parent;
3495 buffer = rcu_dereference(event->buffer);
3499 handle->buffer = buffer;
3500 handle->event = event;
3502 handle->sample = sample;
3504 if (!buffer->nr_pages)
3507 have_lost = local_read(&buffer->lost);
3509 lost_event.header.size = sizeof(lost_event);
3510 perf_event_header__init_id(&lost_event.header, &sample_data,
3512 size += lost_event.header.size;
3515 perf_output_get_handle(handle);
3519 * Userspace could choose to issue a mb() before updating the
3520 * tail pointer. So that all reads will be completed before the
3523 tail = ACCESS_ONCE(buffer->user_page->data_tail);
3525 offset = head = local_read(&buffer->head);
3527 if (unlikely(!perf_output_space(buffer, tail, offset, head)))
3529 } while (local_cmpxchg(&buffer->head, offset, head) != offset);
3531 if (head - local_read(&buffer->wakeup) > buffer->watermark)
3532 local_add(buffer->watermark, &buffer->wakeup);
3534 handle->page = offset >> (PAGE_SHIFT + page_order(buffer));
3535 handle->page &= buffer->nr_pages - 1;
3536 handle->size = offset & ((PAGE_SIZE << page_order(buffer)) - 1);
3537 handle->addr = buffer->data_pages[handle->page];
3538 handle->addr += handle->size;
3539 handle->size = (PAGE_SIZE << page_order(buffer)) - handle->size;
3542 lost_event.header.type = PERF_RECORD_LOST;
3543 lost_event.header.misc = 0;
3544 lost_event.id = event->id;
3545 lost_event.lost = local_xchg(&buffer->lost, 0);
3547 perf_output_put(handle, lost_event);
3548 perf_event__output_id_sample(event, handle, &sample_data);
3554 local_inc(&buffer->lost);
3555 perf_output_put_handle(handle);
3562 void perf_output_end(struct perf_output_handle *handle)
3564 struct perf_event *event = handle->event;
3565 struct perf_buffer *buffer = handle->buffer;
3567 int wakeup_events = event->attr.wakeup_events;
3569 if (handle->sample && wakeup_events) {
3570 int events = local_inc_return(&buffer->events);
3571 if (events >= wakeup_events) {
3572 local_sub(wakeup_events, &buffer->events);
3573 local_inc(&buffer->wakeup);
3577 perf_output_put_handle(handle);
3581 static void perf_output_read_one(struct perf_output_handle *handle,
3582 struct perf_event *event,
3583 u64 enabled, u64 running)
3585 u64 read_format = event->attr.read_format;
3589 values[n++] = perf_event_count(event);
3590 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
3591 values[n++] = enabled +
3592 atomic64_read(&event->child_total_time_enabled);
3594 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
3595 values[n++] = running +
3596 atomic64_read(&event->child_total_time_running);
3598 if (read_format & PERF_FORMAT_ID)
3599 values[n++] = primary_event_id(event);
3601 perf_output_copy(handle, values, n * sizeof(u64));
3605 * XXX PERF_FORMAT_GROUP vs inherited events seems difficult.
3607 static void perf_output_read_group(struct perf_output_handle *handle,
3608 struct perf_event *event,
3609 u64 enabled, u64 running)
3611 struct perf_event *leader = event->group_leader, *sub;
3612 u64 read_format = event->attr.read_format;
3616 values[n++] = 1 + leader->nr_siblings;
3618 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
3619 values[n++] = enabled;
3621 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
3622 values[n++] = running;
3624 if (leader != event)
3625 leader->pmu->read(leader);
3627 values[n++] = perf_event_count(leader);
3628 if (read_format & PERF_FORMAT_ID)
3629 values[n++] = primary_event_id(leader);
3631 perf_output_copy(handle, values, n * sizeof(u64));
3633 list_for_each_entry(sub, &leader->sibling_list, group_entry) {
3637 sub->pmu->read(sub);
3639 values[n++] = perf_event_count(sub);
3640 if (read_format & PERF_FORMAT_ID)
3641 values[n++] = primary_event_id(sub);
3643 perf_output_copy(handle, values, n * sizeof(u64));
3647 #define PERF_FORMAT_TOTAL_TIMES (PERF_FORMAT_TOTAL_TIME_ENABLED|\
3648 PERF_FORMAT_TOTAL_TIME_RUNNING)
3650 static void perf_output_read(struct perf_output_handle *handle,
3651 struct perf_event *event)
3653 u64 enabled = 0, running = 0, now, ctx_time;
3654 u64 read_format = event->attr.read_format;
3657 * compute total_time_enabled, total_time_running
3658 * based on snapshot values taken when the event
3659 * was last scheduled in.
3661 * we cannot simply called update_context_time()
3662 * because of locking issue as we are called in
3665 if (read_format & PERF_FORMAT_TOTAL_TIMES) {
3667 ctx_time = event->shadow_ctx_time + now;
3668 enabled = ctx_time - event->tstamp_enabled;
3669 running = ctx_time - event->tstamp_running;
3672 if (event->attr.read_format & PERF_FORMAT_GROUP)
3673 perf_output_read_group(handle, event, enabled, running);
3675 perf_output_read_one(handle, event, enabled, running);
3678 void perf_output_sample(struct perf_output_handle *handle,
3679 struct perf_event_header *header,
3680 struct perf_sample_data *data,
3681 struct perf_event *event)
3683 u64 sample_type = data->type;
3685 perf_output_put(handle, *header);
3687 if (sample_type & PERF_SAMPLE_IP)
3688 perf_output_put(handle, data->ip);
3690 if (sample_type & PERF_SAMPLE_TID)
3691 perf_output_put(handle, data->tid_entry);
3693 if (sample_type & PERF_SAMPLE_TIME)
3694 perf_output_put(handle, data->time);
3696 if (sample_type & PERF_SAMPLE_ADDR)
3697 perf_output_put(handle, data->addr);
3699 if (sample_type & PERF_SAMPLE_ID)
3700 perf_output_put(handle, data->id);
3702 if (sample_type & PERF_SAMPLE_STREAM_ID)
3703 perf_output_put(handle, data->stream_id);
3705 if (sample_type & PERF_SAMPLE_CPU)
3706 perf_output_put(handle, data->cpu_entry);
3708 if (sample_type & PERF_SAMPLE_PERIOD)
3709 perf_output_put(handle, data->period);
3711 if (sample_type & PERF_SAMPLE_READ)
3712 perf_output_read(handle, event);
3714 if (sample_type & PERF_SAMPLE_CALLCHAIN) {
3715 if (data->callchain) {
3718 if (data->callchain)
3719 size += data->callchain->nr;
3721 size *= sizeof(u64);
3723 perf_output_copy(handle, data->callchain, size);
3726 perf_output_put(handle, nr);
3730 if (sample_type & PERF_SAMPLE_RAW) {
3732 perf_output_put(handle, data->raw->size);
3733 perf_output_copy(handle, data->raw->data,
3740 .size = sizeof(u32),
3743 perf_output_put(handle, raw);
3748 void perf_prepare_sample(struct perf_event_header *header,
3749 struct perf_sample_data *data,
3750 struct perf_event *event,
3751 struct pt_regs *regs)
3753 u64 sample_type = event->attr.sample_type;
3755 header->type = PERF_RECORD_SAMPLE;
3756 header->size = sizeof(*header) + event->header_size;
3759 header->misc |= perf_misc_flags(regs);
3761 __perf_event_header__init_id(header, data, event);
3763 if (sample_type & PERF_SAMPLE_IP)
3764 data->ip = perf_instruction_pointer(regs);
3766 if (sample_type & PERF_SAMPLE_CALLCHAIN) {
3769 data->callchain = perf_callchain(regs);
3771 if (data->callchain)
3772 size += data->callchain->nr;
3774 header->size += size * sizeof(u64);
3777 if (sample_type & PERF_SAMPLE_RAW) {
3778 int size = sizeof(u32);
3781 size += data->raw->size;
3783 size += sizeof(u32);
3785 WARN_ON_ONCE(size & (sizeof(u64)-1));
3786 header->size += size;
3790 static void perf_event_output(struct perf_event *event, int nmi,
3791 struct perf_sample_data *data,
3792 struct pt_regs *regs)
3794 struct perf_output_handle handle;
3795 struct perf_event_header header;
3797 /* protect the callchain buffers */
3800 perf_prepare_sample(&header, data, event, regs);
3802 if (perf_output_begin(&handle, event, header.size, nmi, 1))
3805 perf_output_sample(&handle, &header, data, event);
3807 perf_output_end(&handle);
3817 struct perf_read_event {
3818 struct perf_event_header header;
3825 perf_event_read_event(struct perf_event *event,
3826 struct task_struct *task)
3828 struct perf_output_handle handle;
3829 struct perf_sample_data sample;
3830 struct perf_read_event read_event = {
3832 .type = PERF_RECORD_READ,
3834 .size = sizeof(read_event) + event->read_size,
3836 .pid = perf_event_pid(event, task),
3837 .tid = perf_event_tid(event, task),
3841 perf_event_header__init_id(&read_event.header, &sample, event);
3842 ret = perf_output_begin(&handle, event, read_event.header.size, 0, 0);
3846 perf_output_put(&handle, read_event);
3847 perf_output_read(&handle, event);
3848 perf_event__output_id_sample(event, &handle, &sample);
3850 perf_output_end(&handle);
3854 * task tracking -- fork/exit
3856 * enabled by: attr.comm | attr.mmap | attr.mmap_data | attr.task
3859 struct perf_task_event {
3860 struct task_struct *task;
3861 struct perf_event_context *task_ctx;
3864 struct perf_event_header header;
3874 static void perf_event_task_output(struct perf_event *event,
3875 struct perf_task_event *task_event)
3877 struct perf_output_handle handle;
3878 struct perf_sample_data sample;
3879 struct task_struct *task = task_event->task;
3880 int ret, size = task_event->event_id.header.size;
3882 perf_event_header__init_id(&task_event->event_id.header, &sample, event);
3884 ret = perf_output_begin(&handle, event,
3885 task_event->event_id.header.size, 0, 0);
3889 task_event->event_id.pid = perf_event_pid(event, task);
3890 task_event->event_id.ppid = perf_event_pid(event, current);
3892 task_event->event_id.tid = perf_event_tid(event, task);
3893 task_event->event_id.ptid = perf_event_tid(event, current);
3895 perf_output_put(&handle, task_event->event_id);
3897 perf_event__output_id_sample(event, &handle, &sample);
3899 perf_output_end(&handle);
3901 task_event->event_id.header.size = size;
3904 static int perf_event_task_match(struct perf_event *event)
3906 if (event->state < PERF_EVENT_STATE_INACTIVE)
3909 if (!event_filter_match(event))
3912 if (event->attr.comm || event->attr.mmap ||
3913 event->attr.mmap_data || event->attr.task)
3919 static void perf_event_task_ctx(struct perf_event_context *ctx,
3920 struct perf_task_event *task_event)
3922 struct perf_event *event;
3924 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
3925 if (perf_event_task_match(event))
3926 perf_event_task_output(event, task_event);
3930 static void perf_event_task_event(struct perf_task_event *task_event)
3932 struct perf_cpu_context *cpuctx;
3933 struct perf_event_context *ctx;
3938 list_for_each_entry_rcu(pmu, &pmus, entry) {
3939 cpuctx = get_cpu_ptr(pmu->pmu_cpu_context);
3940 if (cpuctx->active_pmu != pmu)
3942 perf_event_task_ctx(&cpuctx->ctx, task_event);
3944 ctx = task_event->task_ctx;
3946 ctxn = pmu->task_ctx_nr;
3949 ctx = rcu_dereference(current->perf_event_ctxp[ctxn]);
3952 perf_event_task_ctx(ctx, task_event);
3954 put_cpu_ptr(pmu->pmu_cpu_context);
3959 static void perf_event_task(struct task_struct *task,
3960 struct perf_event_context *task_ctx,
3963 struct perf_task_event task_event;
3965 if (!atomic_read(&nr_comm_events) &&
3966 !atomic_read(&nr_mmap_events) &&
3967 !atomic_read(&nr_task_events))
3970 task_event = (struct perf_task_event){
3972 .task_ctx = task_ctx,
3975 .type = new ? PERF_RECORD_FORK : PERF_RECORD_EXIT,
3977 .size = sizeof(task_event.event_id),
3983 .time = perf_clock(),
3987 perf_event_task_event(&task_event);
3990 void perf_event_fork(struct task_struct *task)
3992 perf_event_task(task, NULL, 1);
3999 struct perf_comm_event {
4000 struct task_struct *task;
4005 struct perf_event_header header;
4012 static void perf_event_comm_output(struct perf_event *event,
4013 struct perf_comm_event *comm_event)
4015 struct perf_output_handle handle;
4016 struct perf_sample_data sample;
4017 int size = comm_event->event_id.header.size;
4020 perf_event_header__init_id(&comm_event->event_id.header, &sample, event);
4021 ret = perf_output_begin(&handle, event,
4022 comm_event->event_id.header.size, 0, 0);
4027 comm_event->event_id.pid = perf_event_pid(event, comm_event->task);
4028 comm_event->event_id.tid = perf_event_tid(event, comm_event->task);
4030 perf_output_put(&handle, comm_event->event_id);
4031 perf_output_copy(&handle, comm_event->comm,
4032 comm_event->comm_size);
4034 perf_event__output_id_sample(event, &handle, &sample);
4036 perf_output_end(&handle);
4038 comm_event->event_id.header.size = size;
4041 static int perf_event_comm_match(struct perf_event *event)
4043 if (event->state < PERF_EVENT_STATE_INACTIVE)
4046 if (!event_filter_match(event))
4049 if (event->attr.comm)
4055 static void perf_event_comm_ctx(struct perf_event_context *ctx,
4056 struct perf_comm_event *comm_event)
4058 struct perf_event *event;
4060 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
4061 if (perf_event_comm_match(event))
4062 perf_event_comm_output(event, comm_event);
4066 static void perf_event_comm_event(struct perf_comm_event *comm_event)
4068 struct perf_cpu_context *cpuctx;
4069 struct perf_event_context *ctx;
4070 char comm[TASK_COMM_LEN];
4075 memset(comm, 0, sizeof(comm));
4076 strlcpy(comm, comm_event->task->comm, sizeof(comm));
4077 size = ALIGN(strlen(comm)+1, sizeof(u64));
4079 comm_event->comm = comm;
4080 comm_event->comm_size = size;
4082 comm_event->event_id.header.size = sizeof(comm_event->event_id) + size;
4084 list_for_each_entry_rcu(pmu, &pmus, entry) {
4085 cpuctx = get_cpu_ptr(pmu->pmu_cpu_context);
4086 if (cpuctx->active_pmu != pmu)
4088 perf_event_comm_ctx(&cpuctx->ctx, comm_event);
4090 ctxn = pmu->task_ctx_nr;
4094 ctx = rcu_dereference(current->perf_event_ctxp[ctxn]);
4096 perf_event_comm_ctx(ctx, comm_event);
4098 put_cpu_ptr(pmu->pmu_cpu_context);
4103 void perf_event_comm(struct task_struct *task)
4105 struct perf_comm_event comm_event;
4106 struct perf_event_context *ctx;
4109 for_each_task_context_nr(ctxn) {
4110 ctx = task->perf_event_ctxp[ctxn];
4114 perf_event_enable_on_exec(ctx);
4117 if (!atomic_read(&nr_comm_events))
4120 comm_event = (struct perf_comm_event){
4126 .type = PERF_RECORD_COMM,
4135 perf_event_comm_event(&comm_event);
4142 struct perf_mmap_event {
4143 struct vm_area_struct *vma;
4145 const char *file_name;
4149 struct perf_event_header header;
4159 static void perf_event_mmap_output(struct perf_event *event,
4160 struct perf_mmap_event *mmap_event)
4162 struct perf_output_handle handle;
4163 struct perf_sample_data sample;
4164 int size = mmap_event->event_id.header.size;
4167 perf_event_header__init_id(&mmap_event->event_id.header, &sample, event);
4168 ret = perf_output_begin(&handle, event,
4169 mmap_event->event_id.header.size, 0, 0);
4173 mmap_event->event_id.pid = perf_event_pid(event, current);
4174 mmap_event->event_id.tid = perf_event_tid(event, current);
4176 perf_output_put(&handle, mmap_event->event_id);
4177 perf_output_copy(&handle, mmap_event->file_name,
4178 mmap_event->file_size);
4180 perf_event__output_id_sample(event, &handle, &sample);
4182 perf_output_end(&handle);
4184 mmap_event->event_id.header.size = size;
4187 static int perf_event_mmap_match(struct perf_event *event,
4188 struct perf_mmap_event *mmap_event,
4191 if (event->state < PERF_EVENT_STATE_INACTIVE)
4194 if (!event_filter_match(event))
4197 if ((!executable && event->attr.mmap_data) ||
4198 (executable && event->attr.mmap))
4204 static void perf_event_mmap_ctx(struct perf_event_context *ctx,
4205 struct perf_mmap_event *mmap_event,
4208 struct perf_event *event;
4210 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
4211 if (perf_event_mmap_match(event, mmap_event, executable))
4212 perf_event_mmap_output(event, mmap_event);
4216 static void perf_event_mmap_event(struct perf_mmap_event *mmap_event)
4218 struct perf_cpu_context *cpuctx;
4219 struct perf_event_context *ctx;
4220 struct vm_area_struct *vma = mmap_event->vma;
4221 struct file *file = vma->vm_file;
4229 memset(tmp, 0, sizeof(tmp));
4233 * d_path works from the end of the buffer backwards, so we
4234 * need to add enough zero bytes after the string to handle
4235 * the 64bit alignment we do later.
4237 buf = kzalloc(PATH_MAX + sizeof(u64), GFP_KERNEL);
4239 name = strncpy(tmp, "//enomem", sizeof(tmp));
4242 name = d_path(&file->f_path, buf, PATH_MAX);
4244 name = strncpy(tmp, "//toolong", sizeof(tmp));
4248 if (arch_vma_name(mmap_event->vma)) {
4249 name = strncpy(tmp, arch_vma_name(mmap_event->vma),
4255 name = strncpy(tmp, "[vdso]", sizeof(tmp));
4257 } else if (vma->vm_start <= vma->vm_mm->start_brk &&
4258 vma->vm_end >= vma->vm_mm->brk) {
4259 name = strncpy(tmp, "[heap]", sizeof(tmp));
4261 } else if (vma->vm_start <= vma->vm_mm->start_stack &&
4262 vma->vm_end >= vma->vm_mm->start_stack) {
4263 name = strncpy(tmp, "[stack]", sizeof(tmp));
4267 name = strncpy(tmp, "//anon", sizeof(tmp));
4272 size = ALIGN(strlen(name)+1, sizeof(u64));
4274 mmap_event->file_name = name;
4275 mmap_event->file_size = size;
4277 mmap_event->event_id.header.size = sizeof(mmap_event->event_id) + size;
4280 list_for_each_entry_rcu(pmu, &pmus, entry) {
4281 cpuctx = get_cpu_ptr(pmu->pmu_cpu_context);
4282 if (cpuctx->active_pmu != pmu)
4284 perf_event_mmap_ctx(&cpuctx->ctx, mmap_event,
4285 vma->vm_flags & VM_EXEC);
4287 ctxn = pmu->task_ctx_nr;
4291 ctx = rcu_dereference(current->perf_event_ctxp[ctxn]);
4293 perf_event_mmap_ctx(ctx, mmap_event,
4294 vma->vm_flags & VM_EXEC);
4297 put_cpu_ptr(pmu->pmu_cpu_context);
4304 void perf_event_mmap(struct vm_area_struct *vma)
4306 struct perf_mmap_event mmap_event;
4308 if (!atomic_read(&nr_mmap_events))
4311 mmap_event = (struct perf_mmap_event){
4317 .type = PERF_RECORD_MMAP,
4318 .misc = PERF_RECORD_MISC_USER,
4323 .start = vma->vm_start,
4324 .len = vma->vm_end - vma->vm_start,
4325 .pgoff = (u64)vma->vm_pgoff << PAGE_SHIFT,
4329 perf_event_mmap_event(&mmap_event);
4333 * IRQ throttle logging
4336 static void perf_log_throttle(struct perf_event *event, int enable)
4338 struct perf_output_handle handle;
4339 struct perf_sample_data sample;
4343 struct perf_event_header header;
4347 } throttle_event = {
4349 .type = PERF_RECORD_THROTTLE,
4351 .size = sizeof(throttle_event),
4353 .time = perf_clock(),
4354 .id = primary_event_id(event),
4355 .stream_id = event->id,
4359 throttle_event.header.type = PERF_RECORD_UNTHROTTLE;
4361 perf_event_header__init_id(&throttle_event.header, &sample, event);
4363 ret = perf_output_begin(&handle, event,
4364 throttle_event.header.size, 1, 0);
4368 perf_output_put(&handle, throttle_event);
4369 perf_event__output_id_sample(event, &handle, &sample);
4370 perf_output_end(&handle);
4374 * Generic event overflow handling, sampling.
4377 static int __perf_event_overflow(struct perf_event *event, int nmi,
4378 int throttle, struct perf_sample_data *data,
4379 struct pt_regs *regs)
4381 int events = atomic_read(&event->event_limit);
4382 struct hw_perf_event *hwc = &event->hw;
4386 * Non-sampling counters might still use the PMI to fold short
4387 * hardware counters, ignore those.
4389 if (unlikely(!is_sampling_event(event)))
4395 if (hwc->interrupts != MAX_INTERRUPTS) {
4397 if (HZ * hwc->interrupts >
4398 (u64)sysctl_perf_event_sample_rate) {
4399 hwc->interrupts = MAX_INTERRUPTS;
4400 perf_log_throttle(event, 0);
4405 * Keep re-disabling events even though on the previous
4406 * pass we disabled it - just in case we raced with a
4407 * sched-in and the event got enabled again:
4413 if (event->attr.freq) {
4414 u64 now = perf_clock();
4415 s64 delta = now - hwc->freq_time_stamp;
4417 hwc->freq_time_stamp = now;
4419 if (delta > 0 && delta < 2*TICK_NSEC)
4420 perf_adjust_period(event, delta, hwc->last_period);
4424 * XXX event_limit might not quite work as expected on inherited
4428 event->pending_kill = POLL_IN;
4429 if (events && atomic_dec_and_test(&event->event_limit)) {
4431 event->pending_kill = POLL_HUP;
4433 event->pending_disable = 1;
4434 irq_work_queue(&event->pending);
4436 perf_event_disable(event);
4439 if (event->overflow_handler)
4440 event->overflow_handler(event, nmi, data, regs);
4442 perf_event_output(event, nmi, data, regs);
4447 int perf_event_overflow(struct perf_event *event, int nmi,
4448 struct perf_sample_data *data,
4449 struct pt_regs *regs)
4451 return __perf_event_overflow(event, nmi, 1, data, regs);
4455 * Generic software event infrastructure
4458 struct swevent_htable {
4459 struct swevent_hlist *swevent_hlist;
4460 struct mutex hlist_mutex;
4463 /* Recursion avoidance in each contexts */
4464 int recursion[PERF_NR_CONTEXTS];
4467 static DEFINE_PER_CPU(struct swevent_htable, swevent_htable);
4470 * We directly increment event->count and keep a second value in
4471 * event->hw.period_left to count intervals. This period event
4472 * is kept in the range [-sample_period, 0] so that we can use the
4476 static u64 perf_swevent_set_period(struct perf_event *event)
4478 struct hw_perf_event *hwc = &event->hw;
4479 u64 period = hwc->last_period;
4483 hwc->last_period = hwc->sample_period;
4486 old = val = local64_read(&hwc->period_left);
4490 nr = div64_u64(period + val, period);
4491 offset = nr * period;
4493 if (local64_cmpxchg(&hwc->period_left, old, val) != old)
4499 static void perf_swevent_overflow(struct perf_event *event, u64 overflow,
4500 int nmi, struct perf_sample_data *data,
4501 struct pt_regs *regs)
4503 struct hw_perf_event *hwc = &event->hw;
4506 data->period = event->hw.last_period;
4508 overflow = perf_swevent_set_period(event);
4510 if (hwc->interrupts == MAX_INTERRUPTS)
4513 for (; overflow; overflow--) {
4514 if (__perf_event_overflow(event, nmi, throttle,
4517 * We inhibit the overflow from happening when
4518 * hwc->interrupts == MAX_INTERRUPTS.
4526 static void perf_swevent_event(struct perf_event *event, u64 nr,
4527 int nmi, struct perf_sample_data *data,
4528 struct pt_regs *regs)
4530 struct hw_perf_event *hwc = &event->hw;
4532 local64_add(nr, &event->count);
4537 if (!is_sampling_event(event))
4540 if (nr == 1 && hwc->sample_period == 1 && !event->attr.freq)
4541 return perf_swevent_overflow(event, 1, nmi, data, regs);
4543 if (local64_add_negative(nr, &hwc->period_left))
4546 perf_swevent_overflow(event, 0, nmi, data, regs);
4549 static int perf_exclude_event(struct perf_event *event,
4550 struct pt_regs *regs)
4552 if (event->hw.state & PERF_HES_STOPPED)
4556 if (event->attr.exclude_user && user_mode(regs))
4559 if (event->attr.exclude_kernel && !user_mode(regs))
4566 static int perf_swevent_match(struct perf_event *event,
4567 enum perf_type_id type,
4569 struct perf_sample_data *data,
4570 struct pt_regs *regs)
4572 if (event->attr.type != type)
4575 if (event->attr.config != event_id)
4578 if (perf_exclude_event(event, regs))
4584 static inline u64 swevent_hash(u64 type, u32 event_id)
4586 u64 val = event_id | (type << 32);
4588 return hash_64(val, SWEVENT_HLIST_BITS);
4591 static inline struct hlist_head *
4592 __find_swevent_head(struct swevent_hlist *hlist, u64 type, u32 event_id)
4594 u64 hash = swevent_hash(type, event_id);
4596 return &hlist->heads[hash];
4599 /* For the read side: events when they trigger */
4600 static inline struct hlist_head *
4601 find_swevent_head_rcu(struct swevent_htable *swhash, u64 type, u32 event_id)
4603 struct swevent_hlist *hlist;
4605 hlist = rcu_dereference(swhash->swevent_hlist);
4609 return __find_swevent_head(hlist, type, event_id);
4612 /* For the event head insertion and removal in the hlist */
4613 static inline struct hlist_head *
4614 find_swevent_head(struct swevent_htable *swhash, struct perf_event *event)
4616 struct swevent_hlist *hlist;
4617 u32 event_id = event->attr.config;
4618 u64 type = event->attr.type;
4621 * Event scheduling is always serialized against hlist allocation
4622 * and release. Which makes the protected version suitable here.
4623 * The context lock guarantees that.
4625 hlist = rcu_dereference_protected(swhash->swevent_hlist,
4626 lockdep_is_held(&event->ctx->lock));
4630 return __find_swevent_head(hlist, type, event_id);
4633 static void do_perf_sw_event(enum perf_type_id type, u32 event_id,
4635 struct perf_sample_data *data,
4636 struct pt_regs *regs)
4638 struct swevent_htable *swhash = &__get_cpu_var(swevent_htable);
4639 struct perf_event *event;
4640 struct hlist_node *node;
4641 struct hlist_head *head;
4644 head = find_swevent_head_rcu(swhash, type, event_id);
4648 hlist_for_each_entry_rcu(event, node, head, hlist_entry) {
4649 if (perf_swevent_match(event, type, event_id, data, regs))
4650 perf_swevent_event(event, nr, nmi, data, regs);
4656 int perf_swevent_get_recursion_context(void)
4658 struct swevent_htable *swhash = &__get_cpu_var(swevent_htable);
4660 return get_recursion_context(swhash->recursion);
4662 EXPORT_SYMBOL_GPL(perf_swevent_get_recursion_context);
4664 inline void perf_swevent_put_recursion_context(int rctx)
4666 struct swevent_htable *swhash = &__get_cpu_var(swevent_htable);
4668 put_recursion_context(swhash->recursion, rctx);
4671 void __perf_sw_event(u32 event_id, u64 nr, int nmi,
4672 struct pt_regs *regs, u64 addr)
4674 struct perf_sample_data data;
4677 preempt_disable_notrace();
4678 rctx = perf_swevent_get_recursion_context();
4682 perf_sample_data_init(&data, addr);
4684 do_perf_sw_event(PERF_TYPE_SOFTWARE, event_id, nr, nmi, &data, regs);
4686 perf_swevent_put_recursion_context(rctx);
4687 preempt_enable_notrace();
4690 static void perf_swevent_read(struct perf_event *event)
4694 static int perf_swevent_add(struct perf_event *event, int flags)
4696 struct swevent_htable *swhash = &__get_cpu_var(swevent_htable);
4697 struct hw_perf_event *hwc = &event->hw;
4698 struct hlist_head *head;
4700 if (is_sampling_event(event)) {
4701 hwc->last_period = hwc->sample_period;
4702 perf_swevent_set_period(event);
4705 hwc->state = !(flags & PERF_EF_START);
4707 head = find_swevent_head(swhash, event);
4708 if (WARN_ON_ONCE(!head))
4711 hlist_add_head_rcu(&event->hlist_entry, head);
4716 static void perf_swevent_del(struct perf_event *event, int flags)
4718 hlist_del_rcu(&event->hlist_entry);
4721 static void perf_swevent_start(struct perf_event *event, int flags)
4723 event->hw.state = 0;
4726 static void perf_swevent_stop(struct perf_event *event, int flags)
4728 event->hw.state = PERF_HES_STOPPED;
4731 /* Deref the hlist from the update side */
4732 static inline struct swevent_hlist *
4733 swevent_hlist_deref(struct swevent_htable *swhash)
4735 return rcu_dereference_protected(swhash->swevent_hlist,
4736 lockdep_is_held(&swhash->hlist_mutex));
4739 static void swevent_hlist_release_rcu(struct rcu_head *rcu_head)
4741 struct swevent_hlist *hlist;
4743 hlist = container_of(rcu_head, struct swevent_hlist, rcu_head);
4747 static void swevent_hlist_release(struct swevent_htable *swhash)
4749 struct swevent_hlist *hlist = swevent_hlist_deref(swhash);
4754 rcu_assign_pointer(swhash->swevent_hlist, NULL);
4755 call_rcu(&hlist->rcu_head, swevent_hlist_release_rcu);
4758 static void swevent_hlist_put_cpu(struct perf_event *event, int cpu)
4760 struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
4762 mutex_lock(&swhash->hlist_mutex);
4764 if (!--swhash->hlist_refcount)
4765 swevent_hlist_release(swhash);
4767 mutex_unlock(&swhash->hlist_mutex);
4770 static void swevent_hlist_put(struct perf_event *event)
4774 if (event->cpu != -1) {
4775 swevent_hlist_put_cpu(event, event->cpu);
4779 for_each_possible_cpu(cpu)
4780 swevent_hlist_put_cpu(event, cpu);
4783 static int swevent_hlist_get_cpu(struct perf_event *event, int cpu)
4785 struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
4788 mutex_lock(&swhash->hlist_mutex);
4790 if (!swevent_hlist_deref(swhash) && cpu_online(cpu)) {
4791 struct swevent_hlist *hlist;
4793 hlist = kzalloc(sizeof(*hlist), GFP_KERNEL);
4798 rcu_assign_pointer(swhash->swevent_hlist, hlist);
4800 swhash->hlist_refcount++;
4802 mutex_unlock(&swhash->hlist_mutex);
4807 static int swevent_hlist_get(struct perf_event *event)
4810 int cpu, failed_cpu;
4812 if (event->cpu != -1)
4813 return swevent_hlist_get_cpu(event, event->cpu);
4816 for_each_possible_cpu(cpu) {
4817 err = swevent_hlist_get_cpu(event, cpu);
4827 for_each_possible_cpu(cpu) {
4828 if (cpu == failed_cpu)
4830 swevent_hlist_put_cpu(event, cpu);
4837 atomic_t perf_swevent_enabled[PERF_COUNT_SW_MAX];
4839 static void sw_perf_event_destroy(struct perf_event *event)
4841 u64 event_id = event->attr.config;
4843 WARN_ON(event->parent);
4845 jump_label_dec(&perf_swevent_enabled[event_id]);
4846 swevent_hlist_put(event);
4849 static int perf_swevent_init(struct perf_event *event)
4851 int event_id = event->attr.config;
4853 if (event->attr.type != PERF_TYPE_SOFTWARE)
4857 case PERF_COUNT_SW_CPU_CLOCK:
4858 case PERF_COUNT_SW_TASK_CLOCK:
4865 if (event_id >= PERF_COUNT_SW_MAX)
4868 if (!event->parent) {
4871 err = swevent_hlist_get(event);
4875 jump_label_inc(&perf_swevent_enabled[event_id]);
4876 event->destroy = sw_perf_event_destroy;
4882 static struct pmu perf_swevent = {
4883 .task_ctx_nr = perf_sw_context,
4885 .event_init = perf_swevent_init,
4886 .add = perf_swevent_add,
4887 .del = perf_swevent_del,
4888 .start = perf_swevent_start,
4889 .stop = perf_swevent_stop,
4890 .read = perf_swevent_read,
4893 #ifdef CONFIG_EVENT_TRACING
4895 static int perf_tp_filter_match(struct perf_event *event,
4896 struct perf_sample_data *data)
4898 void *record = data->raw->data;
4900 if (likely(!event->filter) || filter_match_preds(event->filter, record))
4905 static int perf_tp_event_match(struct perf_event *event,
4906 struct perf_sample_data *data,
4907 struct pt_regs *regs)
4910 * All tracepoints are from kernel-space.
4912 if (event->attr.exclude_kernel)
4915 if (!perf_tp_filter_match(event, data))
4921 void perf_tp_event(u64 addr, u64 count, void *record, int entry_size,
4922 struct pt_regs *regs, struct hlist_head *head, int rctx)
4924 struct perf_sample_data data;
4925 struct perf_event *event;
4926 struct hlist_node *node;
4928 struct perf_raw_record raw = {
4933 perf_sample_data_init(&data, addr);
4936 hlist_for_each_entry_rcu(event, node, head, hlist_entry) {
4937 if (perf_tp_event_match(event, &data, regs))
4938 perf_swevent_event(event, count, 1, &data, regs);
4941 perf_swevent_put_recursion_context(rctx);
4943 EXPORT_SYMBOL_GPL(perf_tp_event);
4945 static void tp_perf_event_destroy(struct perf_event *event)
4947 perf_trace_destroy(event);
4950 static int perf_tp_event_init(struct perf_event *event)
4954 if (event->attr.type != PERF_TYPE_TRACEPOINT)
4957 err = perf_trace_init(event);
4961 event->destroy = tp_perf_event_destroy;
4966 static struct pmu perf_tracepoint = {
4967 .task_ctx_nr = perf_sw_context,
4969 .event_init = perf_tp_event_init,
4970 .add = perf_trace_add,
4971 .del = perf_trace_del,
4972 .start = perf_swevent_start,
4973 .stop = perf_swevent_stop,
4974 .read = perf_swevent_read,
4977 static inline void perf_tp_register(void)
4979 perf_pmu_register(&perf_tracepoint, "tracepoint", PERF_TYPE_TRACEPOINT);
4982 static int perf_event_set_filter(struct perf_event *event, void __user *arg)
4987 if (event->attr.type != PERF_TYPE_TRACEPOINT)
4990 filter_str = strndup_user(arg, PAGE_SIZE);
4991 if (IS_ERR(filter_str))
4992 return PTR_ERR(filter_str);
4994 ret = ftrace_profile_set_filter(event, event->attr.config, filter_str);
5000 static void perf_event_free_filter(struct perf_event *event)
5002 ftrace_profile_free_filter(event);
5007 static inline void perf_tp_register(void)
5011 static int perf_event_set_filter(struct perf_event *event, void __user *arg)
5016 static void perf_event_free_filter(struct perf_event *event)
5020 #endif /* CONFIG_EVENT_TRACING */
5022 #ifdef CONFIG_HAVE_HW_BREAKPOINT
5023 void perf_bp_event(struct perf_event *bp, void *data)
5025 struct perf_sample_data sample;
5026 struct pt_regs *regs = data;
5028 perf_sample_data_init(&sample, bp->attr.bp_addr);
5030 if (!bp->hw.state && !perf_exclude_event(bp, regs))
5031 perf_swevent_event(bp, 1, 1, &sample, regs);
5036 * hrtimer based swevent callback
5039 static enum hrtimer_restart perf_swevent_hrtimer(struct hrtimer *hrtimer)
5041 enum hrtimer_restart ret = HRTIMER_RESTART;
5042 struct perf_sample_data data;
5043 struct pt_regs *regs;
5044 struct perf_event *event;
5047 event = container_of(hrtimer, struct perf_event, hw.hrtimer);
5048 event->pmu->read(event);
5050 perf_sample_data_init(&data, 0);
5051 data.period = event->hw.last_period;
5052 regs = get_irq_regs();
5054 if (regs && !perf_exclude_event(event, regs)) {
5055 if (!(event->attr.exclude_idle && current->pid == 0))
5056 if (perf_event_overflow(event, 0, &data, regs))
5057 ret = HRTIMER_NORESTART;
5060 period = max_t(u64, 10000, event->hw.sample_period);
5061 hrtimer_forward_now(hrtimer, ns_to_ktime(period));
5066 static void perf_swevent_start_hrtimer(struct perf_event *event)
5068 struct hw_perf_event *hwc = &event->hw;
5071 if (!is_sampling_event(event))
5074 hrtimer_init(&hwc->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
5075 hwc->hrtimer.function = perf_swevent_hrtimer;
5077 period = local64_read(&hwc->period_left);
5082 local64_set(&hwc->period_left, 0);
5084 period = max_t(u64, 10000, hwc->sample_period);
5086 __hrtimer_start_range_ns(&hwc->hrtimer,
5087 ns_to_ktime(period), 0,
5088 HRTIMER_MODE_REL_PINNED, 0);
5091 static void perf_swevent_cancel_hrtimer(struct perf_event *event)
5093 struct hw_perf_event *hwc = &event->hw;
5095 if (is_sampling_event(event)) {
5096 ktime_t remaining = hrtimer_get_remaining(&hwc->hrtimer);
5097 local64_set(&hwc->period_left, ktime_to_ns(remaining));
5099 hrtimer_cancel(&hwc->hrtimer);
5104 * Software event: cpu wall time clock
5107 static void cpu_clock_event_update(struct perf_event *event)
5112 now = local_clock();
5113 prev = local64_xchg(&event->hw.prev_count, now);
5114 local64_add(now - prev, &event->count);
5117 static void cpu_clock_event_start(struct perf_event *event, int flags)
5119 local64_set(&event->hw.prev_count, local_clock());
5120 perf_swevent_start_hrtimer(event);
5123 static void cpu_clock_event_stop(struct perf_event *event, int flags)
5125 perf_swevent_cancel_hrtimer(event);
5126 cpu_clock_event_update(event);
5129 static int cpu_clock_event_add(struct perf_event *event, int flags)
5131 if (flags & PERF_EF_START)
5132 cpu_clock_event_start(event, flags);
5137 static void cpu_clock_event_del(struct perf_event *event, int flags)
5139 cpu_clock_event_stop(event, flags);
5142 static void cpu_clock_event_read(struct perf_event *event)
5144 cpu_clock_event_update(event);
5147 static int cpu_clock_event_init(struct perf_event *event)
5149 if (event->attr.type != PERF_TYPE_SOFTWARE)
5152 if (event->attr.config != PERF_COUNT_SW_CPU_CLOCK)
5158 static struct pmu perf_cpu_clock = {
5159 .task_ctx_nr = perf_sw_context,
5161 .event_init = cpu_clock_event_init,
5162 .add = cpu_clock_event_add,
5163 .del = cpu_clock_event_del,
5164 .start = cpu_clock_event_start,
5165 .stop = cpu_clock_event_stop,
5166 .read = cpu_clock_event_read,
5170 * Software event: task time clock
5173 static void task_clock_event_update(struct perf_event *event, u64 now)
5178 prev = local64_xchg(&event->hw.prev_count, now);
5180 local64_add(delta, &event->count);
5183 static void task_clock_event_start(struct perf_event *event, int flags)
5185 local64_set(&event->hw.prev_count, event->ctx->time);
5186 perf_swevent_start_hrtimer(event);
5189 static void task_clock_event_stop(struct perf_event *event, int flags)
5191 perf_swevent_cancel_hrtimer(event);
5192 task_clock_event_update(event, event->ctx->time);
5195 static int task_clock_event_add(struct perf_event *event, int flags)
5197 if (flags & PERF_EF_START)
5198 task_clock_event_start(event, flags);
5203 static void task_clock_event_del(struct perf_event *event, int flags)
5205 task_clock_event_stop(event, PERF_EF_UPDATE);
5208 static void task_clock_event_read(struct perf_event *event)
5213 update_context_time(event->ctx);
5214 time = event->ctx->time;
5216 u64 now = perf_clock();
5217 u64 delta = now - event->ctx->timestamp;
5218 time = event->ctx->time + delta;
5221 task_clock_event_update(event, time);
5224 static int task_clock_event_init(struct perf_event *event)
5226 if (event->attr.type != PERF_TYPE_SOFTWARE)
5229 if (event->attr.config != PERF_COUNT_SW_TASK_CLOCK)
5235 static struct pmu perf_task_clock = {
5236 .task_ctx_nr = perf_sw_context,
5238 .event_init = task_clock_event_init,
5239 .add = task_clock_event_add,
5240 .del = task_clock_event_del,
5241 .start = task_clock_event_start,
5242 .stop = task_clock_event_stop,
5243 .read = task_clock_event_read,
5246 static void perf_pmu_nop_void(struct pmu *pmu)
5250 static int perf_pmu_nop_int(struct pmu *pmu)
5255 static void perf_pmu_start_txn(struct pmu *pmu)
5257 perf_pmu_disable(pmu);
5260 static int perf_pmu_commit_txn(struct pmu *pmu)
5262 perf_pmu_enable(pmu);
5266 static void perf_pmu_cancel_txn(struct pmu *pmu)
5268 perf_pmu_enable(pmu);
5272 * Ensures all contexts with the same task_ctx_nr have the same
5273 * pmu_cpu_context too.
5275 static void *find_pmu_context(int ctxn)
5282 list_for_each_entry(pmu, &pmus, entry) {
5283 if (pmu->task_ctx_nr == ctxn)
5284 return pmu->pmu_cpu_context;
5290 static void update_pmu_context(struct pmu *pmu, struct pmu *old_pmu)
5294 for_each_possible_cpu(cpu) {
5295 struct perf_cpu_context *cpuctx;
5297 cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu);
5299 if (cpuctx->active_pmu == old_pmu)
5300 cpuctx->active_pmu = pmu;
5304 static void free_pmu_context(struct pmu *pmu)
5308 mutex_lock(&pmus_lock);
5310 * Like a real lame refcount.
5312 list_for_each_entry(i, &pmus, entry) {
5313 if (i->pmu_cpu_context == pmu->pmu_cpu_context) {
5314 update_pmu_context(i, pmu);
5319 free_percpu(pmu->pmu_cpu_context);
5321 mutex_unlock(&pmus_lock);
5323 static struct idr pmu_idr;
5326 type_show(struct device *dev, struct device_attribute *attr, char *page)
5328 struct pmu *pmu = dev_get_drvdata(dev);
5330 return snprintf(page, PAGE_SIZE-1, "%d\n", pmu->type);
5333 static struct device_attribute pmu_dev_attrs[] = {
5338 static int pmu_bus_running;
5339 static struct bus_type pmu_bus = {
5340 .name = "event_source",
5341 .dev_attrs = pmu_dev_attrs,
5344 static void pmu_dev_release(struct device *dev)
5349 static int pmu_dev_alloc(struct pmu *pmu)
5353 pmu->dev = kzalloc(sizeof(struct device), GFP_KERNEL);
5357 device_initialize(pmu->dev);
5358 ret = dev_set_name(pmu->dev, "%s", pmu->name);
5362 dev_set_drvdata(pmu->dev, pmu);
5363 pmu->dev->bus = &pmu_bus;
5364 pmu->dev->release = pmu_dev_release;
5365 ret = device_add(pmu->dev);
5373 put_device(pmu->dev);
5377 int perf_pmu_register(struct pmu *pmu, char *name, int type)
5381 mutex_lock(&pmus_lock);
5383 pmu->pmu_disable_count = alloc_percpu(int);
5384 if (!pmu->pmu_disable_count)
5393 int err = idr_pre_get(&pmu_idr, GFP_KERNEL);
5397 err = idr_get_new_above(&pmu_idr, pmu, PERF_TYPE_MAX, &type);
5405 if (pmu_bus_running) {
5406 ret = pmu_dev_alloc(pmu);
5412 pmu->pmu_cpu_context = find_pmu_context(pmu->task_ctx_nr);
5413 if (pmu->pmu_cpu_context)
5414 goto got_cpu_context;
5416 pmu->pmu_cpu_context = alloc_percpu(struct perf_cpu_context);
5417 if (!pmu->pmu_cpu_context)
5420 for_each_possible_cpu(cpu) {
5421 struct perf_cpu_context *cpuctx;
5423 cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu);
5424 __perf_event_init_context(&cpuctx->ctx);
5425 cpuctx->ctx.type = cpu_context;
5426 cpuctx->ctx.pmu = pmu;
5427 cpuctx->jiffies_interval = 1;
5428 INIT_LIST_HEAD(&cpuctx->rotation_list);
5429 cpuctx->active_pmu = pmu;
5433 if (!pmu->start_txn) {
5434 if (pmu->pmu_enable) {
5436 * If we have pmu_enable/pmu_disable calls, install
5437 * transaction stubs that use that to try and batch
5438 * hardware accesses.
5440 pmu->start_txn = perf_pmu_start_txn;
5441 pmu->commit_txn = perf_pmu_commit_txn;
5442 pmu->cancel_txn = perf_pmu_cancel_txn;
5444 pmu->start_txn = perf_pmu_nop_void;
5445 pmu->commit_txn = perf_pmu_nop_int;
5446 pmu->cancel_txn = perf_pmu_nop_void;
5450 if (!pmu->pmu_enable) {
5451 pmu->pmu_enable = perf_pmu_nop_void;
5452 pmu->pmu_disable = perf_pmu_nop_void;
5455 list_add_rcu(&pmu->entry, &pmus);
5458 mutex_unlock(&pmus_lock);
5463 device_del(pmu->dev);
5464 put_device(pmu->dev);
5467 if (pmu->type >= PERF_TYPE_MAX)
5468 idr_remove(&pmu_idr, pmu->type);
5471 free_percpu(pmu->pmu_disable_count);
5475 void perf_pmu_unregister(struct pmu *pmu)
5477 mutex_lock(&pmus_lock);
5478 list_del_rcu(&pmu->entry);
5479 mutex_unlock(&pmus_lock);
5482 * We dereference the pmu list under both SRCU and regular RCU, so
5483 * synchronize against both of those.
5485 synchronize_srcu(&pmus_srcu);
5488 free_percpu(pmu->pmu_disable_count);
5489 if (pmu->type >= PERF_TYPE_MAX)
5490 idr_remove(&pmu_idr, pmu->type);
5491 device_del(pmu->dev);
5492 put_device(pmu->dev);
5493 free_pmu_context(pmu);
5496 struct pmu *perf_init_event(struct perf_event *event)
5498 struct pmu *pmu = NULL;
5501 idx = srcu_read_lock(&pmus_srcu);
5504 pmu = idr_find(&pmu_idr, event->attr.type);
5509 list_for_each_entry_rcu(pmu, &pmus, entry) {
5510 int ret = pmu->event_init(event);
5514 if (ret != -ENOENT) {
5519 pmu = ERR_PTR(-ENOENT);
5521 srcu_read_unlock(&pmus_srcu, idx);
5527 * Allocate and initialize a event structure
5529 static struct perf_event *
5530 perf_event_alloc(struct perf_event_attr *attr, int cpu,
5531 struct task_struct *task,
5532 struct perf_event *group_leader,
5533 struct perf_event *parent_event,
5534 perf_overflow_handler_t overflow_handler)
5537 struct perf_event *event;
5538 struct hw_perf_event *hwc;
5541 if ((unsigned)cpu >= nr_cpu_ids) {
5542 if (!task || cpu != -1)
5543 return ERR_PTR(-EINVAL);
5546 event = kzalloc(sizeof(*event), GFP_KERNEL);
5548 return ERR_PTR(-ENOMEM);
5551 * Single events are their own group leaders, with an
5552 * empty sibling list:
5555 group_leader = event;
5557 mutex_init(&event->child_mutex);
5558 INIT_LIST_HEAD(&event->child_list);
5560 INIT_LIST_HEAD(&event->group_entry);
5561 INIT_LIST_HEAD(&event->event_entry);
5562 INIT_LIST_HEAD(&event->sibling_list);
5563 init_waitqueue_head(&event->waitq);
5564 init_irq_work(&event->pending, perf_pending_event);
5566 mutex_init(&event->mmap_mutex);
5569 event->attr = *attr;
5570 event->group_leader = group_leader;
5574 event->parent = parent_event;
5576 event->ns = get_pid_ns(current->nsproxy->pid_ns);
5577 event->id = atomic64_inc_return(&perf_event_id);
5579 event->state = PERF_EVENT_STATE_INACTIVE;
5582 event->attach_state = PERF_ATTACH_TASK;
5583 #ifdef CONFIG_HAVE_HW_BREAKPOINT
5585 * hw_breakpoint is a bit difficult here..
5587 if (attr->type == PERF_TYPE_BREAKPOINT)
5588 event->hw.bp_target = task;
5592 if (!overflow_handler && parent_event)
5593 overflow_handler = parent_event->overflow_handler;
5595 event->overflow_handler = overflow_handler;
5598 event->state = PERF_EVENT_STATE_OFF;
5603 hwc->sample_period = attr->sample_period;
5604 if (attr->freq && attr->sample_freq)
5605 hwc->sample_period = 1;
5606 hwc->last_period = hwc->sample_period;
5608 local64_set(&hwc->period_left, hwc->sample_period);
5611 * we currently do not support PERF_FORMAT_GROUP on inherited events
5613 if (attr->inherit && (attr->read_format & PERF_FORMAT_GROUP))
5616 pmu = perf_init_event(event);
5622 else if (IS_ERR(pmu))
5627 put_pid_ns(event->ns);
5629 return ERR_PTR(err);
5634 if (!event->parent) {
5635 if (event->attach_state & PERF_ATTACH_TASK)
5636 jump_label_inc(&perf_task_events);
5637 if (event->attr.mmap || event->attr.mmap_data)
5638 atomic_inc(&nr_mmap_events);
5639 if (event->attr.comm)
5640 atomic_inc(&nr_comm_events);
5641 if (event->attr.task)
5642 atomic_inc(&nr_task_events);
5643 if (event->attr.sample_type & PERF_SAMPLE_CALLCHAIN) {
5644 err = get_callchain_buffers();
5647 return ERR_PTR(err);
5655 static int perf_copy_attr(struct perf_event_attr __user *uattr,
5656 struct perf_event_attr *attr)
5661 if (!access_ok(VERIFY_WRITE, uattr, PERF_ATTR_SIZE_VER0))
5665 * zero the full structure, so that a short copy will be nice.
5667 memset(attr, 0, sizeof(*attr));
5669 ret = get_user(size, &uattr->size);
5673 if (size > PAGE_SIZE) /* silly large */
5676 if (!size) /* abi compat */
5677 size = PERF_ATTR_SIZE_VER0;
5679 if (size < PERF_ATTR_SIZE_VER0)
5683 * If we're handed a bigger struct than we know of,
5684 * ensure all the unknown bits are 0 - i.e. new
5685 * user-space does not rely on any kernel feature
5686 * extensions we dont know about yet.
5688 if (size > sizeof(*attr)) {
5689 unsigned char __user *addr;
5690 unsigned char __user *end;
5693 addr = (void __user *)uattr + sizeof(*attr);
5694 end = (void __user *)uattr + size;
5696 for (; addr < end; addr++) {
5697 ret = get_user(val, addr);
5703 size = sizeof(*attr);
5706 ret = copy_from_user(attr, uattr, size);
5711 * If the type exists, the corresponding creation will verify
5714 if (attr->type >= PERF_TYPE_MAX)
5717 if (attr->__reserved_1)
5720 if (attr->sample_type & ~(PERF_SAMPLE_MAX-1))
5723 if (attr->read_format & ~(PERF_FORMAT_MAX-1))
5730 put_user(sizeof(*attr), &uattr->size);
5736 perf_event_set_output(struct perf_event *event, struct perf_event *output_event)
5738 struct perf_buffer *buffer = NULL, *old_buffer = NULL;
5744 /* don't allow circular references */
5745 if (event == output_event)
5749 * Don't allow cross-cpu buffers
5751 if (output_event->cpu != event->cpu)
5755 * If its not a per-cpu buffer, it must be the same task.
5757 if (output_event->cpu == -1 && output_event->ctx != event->ctx)
5761 mutex_lock(&event->mmap_mutex);
5762 /* Can't redirect output if we've got an active mmap() */
5763 if (atomic_read(&event->mmap_count))
5767 /* get the buffer we want to redirect to */
5768 buffer = perf_buffer_get(output_event);
5773 old_buffer = event->buffer;
5774 rcu_assign_pointer(event->buffer, buffer);
5777 mutex_unlock(&event->mmap_mutex);
5780 perf_buffer_put(old_buffer);
5786 * sys_perf_event_open - open a performance event, associate it to a task/cpu
5788 * @attr_uptr: event_id type attributes for monitoring/sampling
5791 * @group_fd: group leader event fd
5793 SYSCALL_DEFINE5(perf_event_open,
5794 struct perf_event_attr __user *, attr_uptr,
5795 pid_t, pid, int, cpu, int, group_fd, unsigned long, flags)
5797 struct perf_event *group_leader = NULL, *output_event = NULL;
5798 struct perf_event *event, *sibling;
5799 struct perf_event_attr attr;
5800 struct perf_event_context *ctx;
5801 struct file *event_file = NULL;
5802 struct file *group_file = NULL;
5803 struct task_struct *task = NULL;
5807 int fput_needed = 0;
5810 /* for future expandability... */
5811 if (flags & ~(PERF_FLAG_FD_NO_GROUP | PERF_FLAG_FD_OUTPUT))
5814 err = perf_copy_attr(attr_uptr, &attr);
5818 if (!attr.exclude_kernel) {
5819 if (perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN))
5824 if (attr.sample_freq > sysctl_perf_event_sample_rate)
5828 event_fd = get_unused_fd_flags(O_RDWR);
5832 if (group_fd != -1) {
5833 group_leader = perf_fget_light(group_fd, &fput_needed);
5834 if (IS_ERR(group_leader)) {
5835 err = PTR_ERR(group_leader);
5838 group_file = group_leader->filp;
5839 if (flags & PERF_FLAG_FD_OUTPUT)
5840 output_event = group_leader;
5841 if (flags & PERF_FLAG_FD_NO_GROUP)
5842 group_leader = NULL;
5846 task = find_lively_task_by_vpid(pid);
5848 err = PTR_ERR(task);
5853 event = perf_event_alloc(&attr, cpu, task, group_leader, NULL, NULL);
5854 if (IS_ERR(event)) {
5855 err = PTR_ERR(event);
5860 * Special case software events and allow them to be part of
5861 * any hardware group.
5866 (is_software_event(event) != is_software_event(group_leader))) {
5867 if (is_software_event(event)) {
5869 * If event and group_leader are not both a software
5870 * event, and event is, then group leader is not.
5872 * Allow the addition of software events to !software
5873 * groups, this is safe because software events never
5876 pmu = group_leader->pmu;
5877 } else if (is_software_event(group_leader) &&
5878 (group_leader->group_flags & PERF_GROUP_SOFTWARE)) {
5880 * In case the group is a pure software group, and we
5881 * try to add a hardware event, move the whole group to
5882 * the hardware context.
5889 * Get the target context (task or percpu):
5891 ctx = find_get_context(pmu, task, cpu);
5898 * Look up the group leader (we will attach this event to it):
5904 * Do not allow a recursive hierarchy (this new sibling
5905 * becoming part of another group-sibling):
5907 if (group_leader->group_leader != group_leader)
5910 * Do not allow to attach to a group in a different
5911 * task or CPU context:
5914 if (group_leader->ctx->type != ctx->type)
5917 if (group_leader->ctx != ctx)
5922 * Only a group leader can be exclusive or pinned
5924 if (attr.exclusive || attr.pinned)
5929 err = perf_event_set_output(event, output_event);
5934 event_file = anon_inode_getfile("[perf_event]", &perf_fops, event, O_RDWR);
5935 if (IS_ERR(event_file)) {
5936 err = PTR_ERR(event_file);
5941 struct perf_event_context *gctx = group_leader->ctx;
5943 mutex_lock(&gctx->mutex);
5944 perf_event_remove_from_context(group_leader);
5945 list_for_each_entry(sibling, &group_leader->sibling_list,
5947 perf_event_remove_from_context(sibling);
5950 mutex_unlock(&gctx->mutex);
5954 event->filp = event_file;
5955 WARN_ON_ONCE(ctx->parent_ctx);
5956 mutex_lock(&ctx->mutex);
5959 perf_install_in_context(ctx, group_leader, cpu);
5961 list_for_each_entry(sibling, &group_leader->sibling_list,
5963 perf_install_in_context(ctx, sibling, cpu);
5968 perf_install_in_context(ctx, event, cpu);
5970 mutex_unlock(&ctx->mutex);
5972 event->owner = current;
5974 mutex_lock(¤t->perf_event_mutex);
5975 list_add_tail(&event->owner_entry, ¤t->perf_event_list);
5976 mutex_unlock(¤t->perf_event_mutex);
5979 * Precalculate sample_data sizes
5981 perf_event__header_size(event);
5982 perf_event__id_header_size(event);
5985 * Drop the reference on the group_event after placing the
5986 * new event on the sibling_list. This ensures destruction
5987 * of the group leader will find the pointer to itself in
5988 * perf_group_detach().
5990 fput_light(group_file, fput_needed);
5991 fd_install(event_fd, event_file);
6000 put_task_struct(task);
6002 fput_light(group_file, fput_needed);
6004 put_unused_fd(event_fd);
6009 * perf_event_create_kernel_counter
6011 * @attr: attributes of the counter to create
6012 * @cpu: cpu in which the counter is bound
6013 * @task: task to profile (NULL for percpu)
6016 perf_event_create_kernel_counter(struct perf_event_attr *attr, int cpu,
6017 struct task_struct *task,
6018 perf_overflow_handler_t overflow_handler)
6020 struct perf_event_context *ctx;
6021 struct perf_event *event;
6025 * Get the target context (task or percpu):
6028 event = perf_event_alloc(attr, cpu, task, NULL, NULL, overflow_handler);
6029 if (IS_ERR(event)) {
6030 err = PTR_ERR(event);
6034 ctx = find_get_context(event->pmu, task, cpu);
6041 WARN_ON_ONCE(ctx->parent_ctx);
6042 mutex_lock(&ctx->mutex);
6043 perf_install_in_context(ctx, event, cpu);
6045 mutex_unlock(&ctx->mutex);
6052 return ERR_PTR(err);
6054 EXPORT_SYMBOL_GPL(perf_event_create_kernel_counter);
6056 static void sync_child_event(struct perf_event *child_event,
6057 struct task_struct *child)
6059 struct perf_event *parent_event = child_event->parent;
6062 if (child_event->attr.inherit_stat)
6063 perf_event_read_event(child_event, child);
6065 child_val = perf_event_count(child_event);
6068 * Add back the child's count to the parent's count:
6070 atomic64_add(child_val, &parent_event->child_count);
6071 atomic64_add(child_event->total_time_enabled,
6072 &parent_event->child_total_time_enabled);
6073 atomic64_add(child_event->total_time_running,
6074 &parent_event->child_total_time_running);
6077 * Remove this event from the parent's list
6079 WARN_ON_ONCE(parent_event->ctx->parent_ctx);
6080 mutex_lock(&parent_event->child_mutex);
6081 list_del_init(&child_event->child_list);
6082 mutex_unlock(&parent_event->child_mutex);
6085 * Release the parent event, if this was the last
6088 fput(parent_event->filp);
6092 __perf_event_exit_task(struct perf_event *child_event,
6093 struct perf_event_context *child_ctx,
6094 struct task_struct *child)
6096 struct perf_event *parent_event;
6098 perf_event_remove_from_context(child_event);
6100 parent_event = child_event->parent;
6102 * It can happen that parent exits first, and has events
6103 * that are still around due to the child reference. These
6104 * events need to be zapped - but otherwise linger.
6107 sync_child_event(child_event, child);
6108 free_event(child_event);
6112 static void perf_event_exit_task_context(struct task_struct *child, int ctxn)
6114 struct perf_event *child_event, *tmp;
6115 struct perf_event_context *child_ctx;
6116 unsigned long flags;
6118 if (likely(!child->perf_event_ctxp[ctxn])) {
6119 perf_event_task(child, NULL, 0);
6123 local_irq_save(flags);
6125 * We can't reschedule here because interrupts are disabled,
6126 * and either child is current or it is a task that can't be
6127 * scheduled, so we are now safe from rescheduling changing
6130 child_ctx = child->perf_event_ctxp[ctxn];
6131 task_ctx_sched_out(child_ctx, EVENT_ALL);
6134 * Take the context lock here so that if find_get_context is
6135 * reading child->perf_event_ctxp, we wait until it has
6136 * incremented the context's refcount before we do put_ctx below.
6138 raw_spin_lock(&child_ctx->lock);
6139 child->perf_event_ctxp[ctxn] = NULL;
6141 * If this context is a clone; unclone it so it can't get
6142 * swapped to another process while we're removing all
6143 * the events from it.
6145 unclone_ctx(child_ctx);
6146 update_context_time(child_ctx);
6147 raw_spin_unlock_irqrestore(&child_ctx->lock, flags);
6150 * Report the task dead after unscheduling the events so that we
6151 * won't get any samples after PERF_RECORD_EXIT. We can however still
6152 * get a few PERF_RECORD_READ events.
6154 perf_event_task(child, child_ctx, 0);
6157 * We can recurse on the same lock type through:
6159 * __perf_event_exit_task()
6160 * sync_child_event()
6161 * fput(parent_event->filp)
6163 * mutex_lock(&ctx->mutex)
6165 * But since its the parent context it won't be the same instance.
6167 mutex_lock(&child_ctx->mutex);
6170 list_for_each_entry_safe(child_event, tmp, &child_ctx->pinned_groups,
6172 __perf_event_exit_task(child_event, child_ctx, child);
6174 list_for_each_entry_safe(child_event, tmp, &child_ctx->flexible_groups,
6176 __perf_event_exit_task(child_event, child_ctx, child);
6179 * If the last event was a group event, it will have appended all
6180 * its siblings to the list, but we obtained 'tmp' before that which
6181 * will still point to the list head terminating the iteration.
6183 if (!list_empty(&child_ctx->pinned_groups) ||
6184 !list_empty(&child_ctx->flexible_groups))
6187 mutex_unlock(&child_ctx->mutex);
6193 * When a child task exits, feed back event values to parent events.
6195 void perf_event_exit_task(struct task_struct *child)
6197 struct perf_event *event, *tmp;
6200 mutex_lock(&child->perf_event_mutex);
6201 list_for_each_entry_safe(event, tmp, &child->perf_event_list,
6203 list_del_init(&event->owner_entry);
6206 * Ensure the list deletion is visible before we clear
6207 * the owner, closes a race against perf_release() where
6208 * we need to serialize on the owner->perf_event_mutex.
6211 event->owner = NULL;
6213 mutex_unlock(&child->perf_event_mutex);
6215 for_each_task_context_nr(ctxn)
6216 perf_event_exit_task_context(child, ctxn);
6219 static void perf_free_event(struct perf_event *event,
6220 struct perf_event_context *ctx)
6222 struct perf_event *parent = event->parent;
6224 if (WARN_ON_ONCE(!parent))
6227 mutex_lock(&parent->child_mutex);
6228 list_del_init(&event->child_list);
6229 mutex_unlock(&parent->child_mutex);
6233 perf_group_detach(event);
6234 list_del_event(event, ctx);
6239 * free an unexposed, unused context as created by inheritance by
6240 * perf_event_init_task below, used by fork() in case of fail.
6242 void perf_event_free_task(struct task_struct *task)
6244 struct perf_event_context *ctx;
6245 struct perf_event *event, *tmp;
6248 for_each_task_context_nr(ctxn) {
6249 ctx = task->perf_event_ctxp[ctxn];
6253 mutex_lock(&ctx->mutex);
6255 list_for_each_entry_safe(event, tmp, &ctx->pinned_groups,
6257 perf_free_event(event, ctx);
6259 list_for_each_entry_safe(event, tmp, &ctx->flexible_groups,
6261 perf_free_event(event, ctx);
6263 if (!list_empty(&ctx->pinned_groups) ||
6264 !list_empty(&ctx->flexible_groups))
6267 mutex_unlock(&ctx->mutex);
6273 void perf_event_delayed_put(struct task_struct *task)
6277 for_each_task_context_nr(ctxn)
6278 WARN_ON_ONCE(task->perf_event_ctxp[ctxn]);
6282 * inherit a event from parent task to child task:
6284 static struct perf_event *
6285 inherit_event(struct perf_event *parent_event,
6286 struct task_struct *parent,
6287 struct perf_event_context *parent_ctx,
6288 struct task_struct *child,
6289 struct perf_event *group_leader,
6290 struct perf_event_context *child_ctx)
6292 struct perf_event *child_event;
6293 unsigned long flags;
6296 * Instead of creating recursive hierarchies of events,
6297 * we link inherited events back to the original parent,
6298 * which has a filp for sure, which we use as the reference
6301 if (parent_event->parent)
6302 parent_event = parent_event->parent;
6304 child_event = perf_event_alloc(&parent_event->attr,
6307 group_leader, parent_event,
6309 if (IS_ERR(child_event))
6314 * Make the child state follow the state of the parent event,
6315 * not its attr.disabled bit. We hold the parent's mutex,
6316 * so we won't race with perf_event_{en, dis}able_family.
6318 if (parent_event->state >= PERF_EVENT_STATE_INACTIVE)
6319 child_event->state = PERF_EVENT_STATE_INACTIVE;
6321 child_event->state = PERF_EVENT_STATE_OFF;
6323 if (parent_event->attr.freq) {
6324 u64 sample_period = parent_event->hw.sample_period;
6325 struct hw_perf_event *hwc = &child_event->hw;
6327 hwc->sample_period = sample_period;
6328 hwc->last_period = sample_period;
6330 local64_set(&hwc->period_left, sample_period);
6333 child_event->ctx = child_ctx;
6334 child_event->overflow_handler = parent_event->overflow_handler;
6337 * Precalculate sample_data sizes
6339 perf_event__header_size(child_event);
6340 perf_event__id_header_size(child_event);
6343 * Link it up in the child's context:
6345 raw_spin_lock_irqsave(&child_ctx->lock, flags);
6346 add_event_to_ctx(child_event, child_ctx);
6347 raw_spin_unlock_irqrestore(&child_ctx->lock, flags);
6350 * Get a reference to the parent filp - we will fput it
6351 * when the child event exits. This is safe to do because
6352 * we are in the parent and we know that the filp still
6353 * exists and has a nonzero count:
6355 atomic_long_inc(&parent_event->filp->f_count);
6358 * Link this into the parent event's child list
6360 WARN_ON_ONCE(parent_event->ctx->parent_ctx);
6361 mutex_lock(&parent_event->child_mutex);
6362 list_add_tail(&child_event->child_list, &parent_event->child_list);
6363 mutex_unlock(&parent_event->child_mutex);
6368 static int inherit_group(struct perf_event *parent_event,
6369 struct task_struct *parent,
6370 struct perf_event_context *parent_ctx,
6371 struct task_struct *child,
6372 struct perf_event_context *child_ctx)
6374 struct perf_event *leader;
6375 struct perf_event *sub;
6376 struct perf_event *child_ctr;
6378 leader = inherit_event(parent_event, parent, parent_ctx,
6379 child, NULL, child_ctx);
6381 return PTR_ERR(leader);
6382 list_for_each_entry(sub, &parent_event->sibling_list, group_entry) {
6383 child_ctr = inherit_event(sub, parent, parent_ctx,
6384 child, leader, child_ctx);
6385 if (IS_ERR(child_ctr))
6386 return PTR_ERR(child_ctr);
6392 inherit_task_group(struct perf_event *event, struct task_struct *parent,
6393 struct perf_event_context *parent_ctx,
6394 struct task_struct *child, int ctxn,
6398 struct perf_event_context *child_ctx;
6400 if (!event->attr.inherit) {
6405 child_ctx = child->perf_event_ctxp[ctxn];
6408 * This is executed from the parent task context, so
6409 * inherit events that have been marked for cloning.
6410 * First allocate and initialize a context for the
6414 child_ctx = alloc_perf_context(event->pmu, child);
6418 child->perf_event_ctxp[ctxn] = child_ctx;
6421 ret = inherit_group(event, parent, parent_ctx,
6431 * Initialize the perf_event context in task_struct
6433 int perf_event_init_context(struct task_struct *child, int ctxn)
6435 struct perf_event_context *child_ctx, *parent_ctx;
6436 struct perf_event_context *cloned_ctx;
6437 struct perf_event *event;
6438 struct task_struct *parent = current;
6439 int inherited_all = 1;
6440 unsigned long flags;
6443 child->perf_event_ctxp[ctxn] = NULL;
6445 mutex_init(&child->perf_event_mutex);
6446 INIT_LIST_HEAD(&child->perf_event_list);
6448 if (likely(!parent->perf_event_ctxp[ctxn]))
6452 * If the parent's context is a clone, pin it so it won't get
6455 parent_ctx = perf_pin_task_context(parent, ctxn);
6458 * No need to check if parent_ctx != NULL here; since we saw
6459 * it non-NULL earlier, the only reason for it to become NULL
6460 * is if we exit, and since we're currently in the middle of
6461 * a fork we can't be exiting at the same time.
6465 * Lock the parent list. No need to lock the child - not PID
6466 * hashed yet and not running, so nobody can access it.
6468 mutex_lock(&parent_ctx->mutex);
6471 * We dont have to disable NMIs - we are only looking at
6472 * the list, not manipulating it:
6474 list_for_each_entry(event, &parent_ctx->pinned_groups, group_entry) {
6475 ret = inherit_task_group(event, parent, parent_ctx,
6476 child, ctxn, &inherited_all);
6482 * We can't hold ctx->lock when iterating the ->flexible_group list due
6483 * to allocations, but we need to prevent rotation because
6484 * rotate_ctx() will change the list from interrupt context.
6486 raw_spin_lock_irqsave(&parent_ctx->lock, flags);
6487 parent_ctx->rotate_disable = 1;
6488 raw_spin_unlock_irqrestore(&parent_ctx->lock, flags);
6490 list_for_each_entry(event, &parent_ctx->flexible_groups, group_entry) {
6491 ret = inherit_task_group(event, parent, parent_ctx,
6492 child, ctxn, &inherited_all);
6497 raw_spin_lock_irqsave(&parent_ctx->lock, flags);
6498 parent_ctx->rotate_disable = 0;
6500 child_ctx = child->perf_event_ctxp[ctxn];
6502 if (child_ctx && inherited_all) {
6504 * Mark the child context as a clone of the parent
6505 * context, or of whatever the parent is a clone of.
6507 * Note that if the parent is a clone, the holding of
6508 * parent_ctx->lock avoids it from being uncloned.
6510 cloned_ctx = parent_ctx->parent_ctx;
6512 child_ctx->parent_ctx = cloned_ctx;
6513 child_ctx->parent_gen = parent_ctx->parent_gen;
6515 child_ctx->parent_ctx = parent_ctx;
6516 child_ctx->parent_gen = parent_ctx->generation;
6518 get_ctx(child_ctx->parent_ctx);
6521 raw_spin_unlock_irqrestore(&parent_ctx->lock, flags);
6522 mutex_unlock(&parent_ctx->mutex);
6524 perf_unpin_context(parent_ctx);
6530 * Initialize the perf_event context in task_struct
6532 int perf_event_init_task(struct task_struct *child)
6536 for_each_task_context_nr(ctxn) {
6537 ret = perf_event_init_context(child, ctxn);
6545 static void __init perf_event_init_all_cpus(void)
6547 struct swevent_htable *swhash;
6550 for_each_possible_cpu(cpu) {
6551 swhash = &per_cpu(swevent_htable, cpu);
6552 mutex_init(&swhash->hlist_mutex);
6553 INIT_LIST_HEAD(&per_cpu(rotation_list, cpu));
6557 static void __cpuinit perf_event_init_cpu(int cpu)
6559 struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
6561 mutex_lock(&swhash->hlist_mutex);
6562 if (swhash->hlist_refcount > 0) {
6563 struct swevent_hlist *hlist;
6565 hlist = kzalloc_node(sizeof(*hlist), GFP_KERNEL, cpu_to_node(cpu));
6567 rcu_assign_pointer(swhash->swevent_hlist, hlist);
6569 mutex_unlock(&swhash->hlist_mutex);
6572 #if defined CONFIG_HOTPLUG_CPU || defined CONFIG_KEXEC
6573 static void perf_pmu_rotate_stop(struct pmu *pmu)
6575 struct perf_cpu_context *cpuctx = this_cpu_ptr(pmu->pmu_cpu_context);
6577 WARN_ON(!irqs_disabled());
6579 list_del_init(&cpuctx->rotation_list);
6582 static void __perf_event_exit_context(void *__info)
6584 struct perf_event_context *ctx = __info;
6585 struct perf_event *event, *tmp;
6587 perf_pmu_rotate_stop(ctx->pmu);
6589 list_for_each_entry_safe(event, tmp, &ctx->pinned_groups, group_entry)
6590 __perf_event_remove_from_context(event);
6591 list_for_each_entry_safe(event, tmp, &ctx->flexible_groups, group_entry)
6592 __perf_event_remove_from_context(event);
6595 static void perf_event_exit_cpu_context(int cpu)
6597 struct perf_event_context *ctx;
6601 idx = srcu_read_lock(&pmus_srcu);
6602 list_for_each_entry_rcu(pmu, &pmus, entry) {
6603 ctx = &per_cpu_ptr(pmu->pmu_cpu_context, cpu)->ctx;
6605 mutex_lock(&ctx->mutex);
6606 smp_call_function_single(cpu, __perf_event_exit_context, ctx, 1);
6607 mutex_unlock(&ctx->mutex);
6609 srcu_read_unlock(&pmus_srcu, idx);
6612 static void perf_event_exit_cpu(int cpu)
6614 struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
6616 mutex_lock(&swhash->hlist_mutex);
6617 swevent_hlist_release(swhash);
6618 mutex_unlock(&swhash->hlist_mutex);
6620 perf_event_exit_cpu_context(cpu);
6623 static inline void perf_event_exit_cpu(int cpu) { }
6627 perf_reboot(struct notifier_block *notifier, unsigned long val, void *v)
6631 for_each_online_cpu(cpu)
6632 perf_event_exit_cpu(cpu);
6638 * Run the perf reboot notifier at the very last possible moment so that
6639 * the generic watchdog code runs as long as possible.
6641 static struct notifier_block perf_reboot_notifier = {
6642 .notifier_call = perf_reboot,
6643 .priority = INT_MIN,
6646 static int __cpuinit
6647 perf_cpu_notify(struct notifier_block *self, unsigned long action, void *hcpu)
6649 unsigned int cpu = (long)hcpu;
6651 switch (action & ~CPU_TASKS_FROZEN) {
6653 case CPU_UP_PREPARE:
6654 case CPU_DOWN_FAILED:
6655 perf_event_init_cpu(cpu);
6658 case CPU_UP_CANCELED:
6659 case CPU_DOWN_PREPARE:
6660 perf_event_exit_cpu(cpu);
6670 void __init perf_event_init(void)
6676 perf_event_init_all_cpus();
6677 init_srcu_struct(&pmus_srcu);
6678 perf_pmu_register(&perf_swevent, "software", PERF_TYPE_SOFTWARE);
6679 perf_pmu_register(&perf_cpu_clock, NULL, -1);
6680 perf_pmu_register(&perf_task_clock, NULL, -1);
6682 perf_cpu_notifier(perf_cpu_notify);
6683 register_reboot_notifier(&perf_reboot_notifier);
6685 ret = init_hw_breakpoint();
6686 WARN(ret, "hw_breakpoint initialization failed with: %d", ret);
6689 static int __init perf_event_sysfs_init(void)
6694 mutex_lock(&pmus_lock);
6696 ret = bus_register(&pmu_bus);
6700 list_for_each_entry(pmu, &pmus, entry) {
6701 if (!pmu->name || pmu->type < 0)
6704 ret = pmu_dev_alloc(pmu);
6705 WARN(ret, "Failed to register pmu: %s, reason %d\n", pmu->name, ret);
6707 pmu_bus_running = 1;
6711 mutex_unlock(&pmus_lock);
6715 device_initcall(perf_event_sysfs_init);